mm: sched: Adapt the scanning rate if a NUMA hinting fault does not migrate
[linux-3.10.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61
62 #include <asm/io.h>
63 #include <asm/pgalloc.h>
64 #include <asm/uaccess.h>
65 #include <asm/tlb.h>
66 #include <asm/tlbflush.h>
67 #include <asm/pgtable.h>
68
69 #include "internal.h"
70
71 #ifndef CONFIG_NEED_MULTIPLE_NODES
72 /* use the per-pgdat data instead for discontigmem - mbligh */
73 unsigned long max_mapnr;
74 struct page *mem_map;
75
76 EXPORT_SYMBOL(max_mapnr);
77 EXPORT_SYMBOL(mem_map);
78 #endif
79
80 unsigned long num_physpages;
81 /*
82  * A number of key systems in x86 including ioremap() rely on the assumption
83  * that high_memory defines the upper bound on direct map memory, then end
84  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
85  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
86  * and ZONE_HIGHMEM.
87  */
88 void * high_memory;
89
90 EXPORT_SYMBOL(num_physpages);
91 EXPORT_SYMBOL(high_memory);
92
93 /*
94  * Randomize the address space (stacks, mmaps, brk, etc.).
95  *
96  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
97  *   as ancient (libc5 based) binaries can segfault. )
98  */
99 int randomize_va_space __read_mostly =
100 #ifdef CONFIG_COMPAT_BRK
101                                         1;
102 #else
103                                         2;
104 #endif
105
106 static int __init disable_randmaps(char *s)
107 {
108         randomize_va_space = 0;
109         return 1;
110 }
111 __setup("norandmaps", disable_randmaps);
112
113 unsigned long zero_pfn __read_mostly;
114 unsigned long highest_memmap_pfn __read_mostly;
115
116 /*
117  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
118  */
119 static int __init init_zero_pfn(void)
120 {
121         zero_pfn = page_to_pfn(ZERO_PAGE(0));
122         return 0;
123 }
124 core_initcall(init_zero_pfn);
125
126
127 #if defined(SPLIT_RSS_COUNTING)
128
129 void sync_mm_rss(struct mm_struct *mm)
130 {
131         int i;
132
133         for (i = 0; i < NR_MM_COUNTERS; i++) {
134                 if (current->rss_stat.count[i]) {
135                         add_mm_counter(mm, i, current->rss_stat.count[i]);
136                         current->rss_stat.count[i] = 0;
137                 }
138         }
139         current->rss_stat.events = 0;
140 }
141
142 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
143 {
144         struct task_struct *task = current;
145
146         if (likely(task->mm == mm))
147                 task->rss_stat.count[member] += val;
148         else
149                 add_mm_counter(mm, member, val);
150 }
151 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
152 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
153
154 /* sync counter once per 64 page faults */
155 #define TASK_RSS_EVENTS_THRESH  (64)
156 static void check_sync_rss_stat(struct task_struct *task)
157 {
158         if (unlikely(task != current))
159                 return;
160         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
161                 sync_mm_rss(task->mm);
162 }
163 #else /* SPLIT_RSS_COUNTING */
164
165 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
166 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
167
168 static void check_sync_rss_stat(struct task_struct *task)
169 {
170 }
171
172 #endif /* SPLIT_RSS_COUNTING */
173
174 #ifdef HAVE_GENERIC_MMU_GATHER
175
176 static int tlb_next_batch(struct mmu_gather *tlb)
177 {
178         struct mmu_gather_batch *batch;
179
180         batch = tlb->active;
181         if (batch->next) {
182                 tlb->active = batch->next;
183                 return 1;
184         }
185
186         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
187         if (!batch)
188                 return 0;
189
190         batch->next = NULL;
191         batch->nr   = 0;
192         batch->max  = MAX_GATHER_BATCH;
193
194         tlb->active->next = batch;
195         tlb->active = batch;
196
197         return 1;
198 }
199
200 /* tlb_gather_mmu
201  *      Called to initialize an (on-stack) mmu_gather structure for page-table
202  *      tear-down from @mm. The @fullmm argument is used when @mm is without
203  *      users and we're going to destroy the full address space (exit/execve).
204  */
205 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
206 {
207         tlb->mm = mm;
208
209         tlb->fullmm     = fullmm;
210         tlb->start      = -1UL;
211         tlb->end        = 0;
212         tlb->need_flush = 0;
213         tlb->fast_mode  = (num_possible_cpus() == 1);
214         tlb->local.next = NULL;
215         tlb->local.nr   = 0;
216         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
217         tlb->active     = &tlb->local;
218
219 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
220         tlb->batch = NULL;
221 #endif
222 }
223
224 void tlb_flush_mmu(struct mmu_gather *tlb)
225 {
226         struct mmu_gather_batch *batch;
227
228         if (!tlb->need_flush)
229                 return;
230         tlb->need_flush = 0;
231         tlb_flush(tlb);
232 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
233         tlb_table_flush(tlb);
234 #endif
235
236         if (tlb_fast_mode(tlb))
237                 return;
238
239         for (batch = &tlb->local; batch; batch = batch->next) {
240                 free_pages_and_swap_cache(batch->pages, batch->nr);
241                 batch->nr = 0;
242         }
243         tlb->active = &tlb->local;
244 }
245
246 /* tlb_finish_mmu
247  *      Called at the end of the shootdown operation to free up any resources
248  *      that were required.
249  */
250 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
251 {
252         struct mmu_gather_batch *batch, *next;
253
254         tlb->start = start;
255         tlb->end   = end;
256         tlb_flush_mmu(tlb);
257
258         /* keep the page table cache within bounds */
259         check_pgt_cache();
260
261         for (batch = tlb->local.next; batch; batch = next) {
262                 next = batch->next;
263                 free_pages((unsigned long)batch, 0);
264         }
265         tlb->local.next = NULL;
266 }
267
268 /* __tlb_remove_page
269  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
270  *      handling the additional races in SMP caused by other CPUs caching valid
271  *      mappings in their TLBs. Returns the number of free page slots left.
272  *      When out of page slots we must call tlb_flush_mmu().
273  */
274 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
275 {
276         struct mmu_gather_batch *batch;
277
278         VM_BUG_ON(!tlb->need_flush);
279
280         if (tlb_fast_mode(tlb)) {
281                 free_page_and_swap_cache(page);
282                 return 1; /* avoid calling tlb_flush_mmu() */
283         }
284
285         batch = tlb->active;
286         batch->pages[batch->nr++] = page;
287         if (batch->nr == batch->max) {
288                 if (!tlb_next_batch(tlb))
289                         return 0;
290                 batch = tlb->active;
291         }
292         VM_BUG_ON(batch->nr > batch->max);
293
294         return batch->max - batch->nr;
295 }
296
297 #endif /* HAVE_GENERIC_MMU_GATHER */
298
299 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
300
301 /*
302  * See the comment near struct mmu_table_batch.
303  */
304
305 static void tlb_remove_table_smp_sync(void *arg)
306 {
307         /* Simply deliver the interrupt */
308 }
309
310 static void tlb_remove_table_one(void *table)
311 {
312         /*
313          * This isn't an RCU grace period and hence the page-tables cannot be
314          * assumed to be actually RCU-freed.
315          *
316          * It is however sufficient for software page-table walkers that rely on
317          * IRQ disabling. See the comment near struct mmu_table_batch.
318          */
319         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
320         __tlb_remove_table(table);
321 }
322
323 static void tlb_remove_table_rcu(struct rcu_head *head)
324 {
325         struct mmu_table_batch *batch;
326         int i;
327
328         batch = container_of(head, struct mmu_table_batch, rcu);
329
330         for (i = 0; i < batch->nr; i++)
331                 __tlb_remove_table(batch->tables[i]);
332
333         free_page((unsigned long)batch);
334 }
335
336 void tlb_table_flush(struct mmu_gather *tlb)
337 {
338         struct mmu_table_batch **batch = &tlb->batch;
339
340         if (*batch) {
341                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
342                 *batch = NULL;
343         }
344 }
345
346 void tlb_remove_table(struct mmu_gather *tlb, void *table)
347 {
348         struct mmu_table_batch **batch = &tlb->batch;
349
350         tlb->need_flush = 1;
351
352         /*
353          * When there's less then two users of this mm there cannot be a
354          * concurrent page-table walk.
355          */
356         if (atomic_read(&tlb->mm->mm_users) < 2) {
357                 __tlb_remove_table(table);
358                 return;
359         }
360
361         if (*batch == NULL) {
362                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
363                 if (*batch == NULL) {
364                         tlb_remove_table_one(table);
365                         return;
366                 }
367                 (*batch)->nr = 0;
368         }
369         (*batch)->tables[(*batch)->nr++] = table;
370         if ((*batch)->nr == MAX_TABLE_BATCH)
371                 tlb_table_flush(tlb);
372 }
373
374 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
375
376 /*
377  * If a p?d_bad entry is found while walking page tables, report
378  * the error, before resetting entry to p?d_none.  Usually (but
379  * very seldom) called out from the p?d_none_or_clear_bad macros.
380  */
381
382 void pgd_clear_bad(pgd_t *pgd)
383 {
384         pgd_ERROR(*pgd);
385         pgd_clear(pgd);
386 }
387
388 void pud_clear_bad(pud_t *pud)
389 {
390         pud_ERROR(*pud);
391         pud_clear(pud);
392 }
393
394 void pmd_clear_bad(pmd_t *pmd)
395 {
396         pmd_ERROR(*pmd);
397         pmd_clear(pmd);
398 }
399
400 /*
401  * Note: this doesn't free the actual pages themselves. That
402  * has been handled earlier when unmapping all the memory regions.
403  */
404 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
405                            unsigned long addr)
406 {
407         pgtable_t token = pmd_pgtable(*pmd);
408         pmd_clear(pmd);
409         pte_free_tlb(tlb, token, addr);
410         tlb->mm->nr_ptes--;
411 }
412
413 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
414                                 unsigned long addr, unsigned long end,
415                                 unsigned long floor, unsigned long ceiling)
416 {
417         pmd_t *pmd;
418         unsigned long next;
419         unsigned long start;
420
421         start = addr;
422         pmd = pmd_offset(pud, addr);
423         do {
424                 next = pmd_addr_end(addr, end);
425                 if (pmd_none_or_clear_bad(pmd))
426                         continue;
427                 free_pte_range(tlb, pmd, addr);
428         } while (pmd++, addr = next, addr != end);
429
430         start &= PUD_MASK;
431         if (start < floor)
432                 return;
433         if (ceiling) {
434                 ceiling &= PUD_MASK;
435                 if (!ceiling)
436                         return;
437         }
438         if (end - 1 > ceiling - 1)
439                 return;
440
441         pmd = pmd_offset(pud, start);
442         pud_clear(pud);
443         pmd_free_tlb(tlb, pmd, start);
444 }
445
446 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
447                                 unsigned long addr, unsigned long end,
448                                 unsigned long floor, unsigned long ceiling)
449 {
450         pud_t *pud;
451         unsigned long next;
452         unsigned long start;
453
454         start = addr;
455         pud = pud_offset(pgd, addr);
456         do {
457                 next = pud_addr_end(addr, end);
458                 if (pud_none_or_clear_bad(pud))
459                         continue;
460                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
461         } while (pud++, addr = next, addr != end);
462
463         start &= PGDIR_MASK;
464         if (start < floor)
465                 return;
466         if (ceiling) {
467                 ceiling &= PGDIR_MASK;
468                 if (!ceiling)
469                         return;
470         }
471         if (end - 1 > ceiling - 1)
472                 return;
473
474         pud = pud_offset(pgd, start);
475         pgd_clear(pgd);
476         pud_free_tlb(tlb, pud, start);
477 }
478
479 /*
480  * This function frees user-level page tables of a process.
481  *
482  * Must be called with pagetable lock held.
483  */
484 void free_pgd_range(struct mmu_gather *tlb,
485                         unsigned long addr, unsigned long end,
486                         unsigned long floor, unsigned long ceiling)
487 {
488         pgd_t *pgd;
489         unsigned long next;
490
491         /*
492          * The next few lines have given us lots of grief...
493          *
494          * Why are we testing PMD* at this top level?  Because often
495          * there will be no work to do at all, and we'd prefer not to
496          * go all the way down to the bottom just to discover that.
497          *
498          * Why all these "- 1"s?  Because 0 represents both the bottom
499          * of the address space and the top of it (using -1 for the
500          * top wouldn't help much: the masks would do the wrong thing).
501          * The rule is that addr 0 and floor 0 refer to the bottom of
502          * the address space, but end 0 and ceiling 0 refer to the top
503          * Comparisons need to use "end - 1" and "ceiling - 1" (though
504          * that end 0 case should be mythical).
505          *
506          * Wherever addr is brought up or ceiling brought down, we must
507          * be careful to reject "the opposite 0" before it confuses the
508          * subsequent tests.  But what about where end is brought down
509          * by PMD_SIZE below? no, end can't go down to 0 there.
510          *
511          * Whereas we round start (addr) and ceiling down, by different
512          * masks at different levels, in order to test whether a table
513          * now has no other vmas using it, so can be freed, we don't
514          * bother to round floor or end up - the tests don't need that.
515          */
516
517         addr &= PMD_MASK;
518         if (addr < floor) {
519                 addr += PMD_SIZE;
520                 if (!addr)
521                         return;
522         }
523         if (ceiling) {
524                 ceiling &= PMD_MASK;
525                 if (!ceiling)
526                         return;
527         }
528         if (end - 1 > ceiling - 1)
529                 end -= PMD_SIZE;
530         if (addr > end - 1)
531                 return;
532
533         pgd = pgd_offset(tlb->mm, addr);
534         do {
535                 next = pgd_addr_end(addr, end);
536                 if (pgd_none_or_clear_bad(pgd))
537                         continue;
538                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
539         } while (pgd++, addr = next, addr != end);
540 }
541
542 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
543                 unsigned long floor, unsigned long ceiling)
544 {
545         while (vma) {
546                 struct vm_area_struct *next = vma->vm_next;
547                 unsigned long addr = vma->vm_start;
548
549                 /*
550                  * Hide vma from rmap and truncate_pagecache before freeing
551                  * pgtables
552                  */
553                 unlink_anon_vmas(vma);
554                 unlink_file_vma(vma);
555
556                 if (is_vm_hugetlb_page(vma)) {
557                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
558                                 floor, next? next->vm_start: ceiling);
559                 } else {
560                         /*
561                          * Optimization: gather nearby vmas into one call down
562                          */
563                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
564                                && !is_vm_hugetlb_page(next)) {
565                                 vma = next;
566                                 next = vma->vm_next;
567                                 unlink_anon_vmas(vma);
568                                 unlink_file_vma(vma);
569                         }
570                         free_pgd_range(tlb, addr, vma->vm_end,
571                                 floor, next? next->vm_start: ceiling);
572                 }
573                 vma = next;
574         }
575 }
576
577 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
578                 pmd_t *pmd, unsigned long address)
579 {
580         pgtable_t new = pte_alloc_one(mm, address);
581         int wait_split_huge_page;
582         if (!new)
583                 return -ENOMEM;
584
585         /*
586          * Ensure all pte setup (eg. pte page lock and page clearing) are
587          * visible before the pte is made visible to other CPUs by being
588          * put into page tables.
589          *
590          * The other side of the story is the pointer chasing in the page
591          * table walking code (when walking the page table without locking;
592          * ie. most of the time). Fortunately, these data accesses consist
593          * of a chain of data-dependent loads, meaning most CPUs (alpha
594          * being the notable exception) will already guarantee loads are
595          * seen in-order. See the alpha page table accessors for the
596          * smp_read_barrier_depends() barriers in page table walking code.
597          */
598         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
599
600         spin_lock(&mm->page_table_lock);
601         wait_split_huge_page = 0;
602         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
603                 mm->nr_ptes++;
604                 pmd_populate(mm, pmd, new);
605                 new = NULL;
606         } else if (unlikely(pmd_trans_splitting(*pmd)))
607                 wait_split_huge_page = 1;
608         spin_unlock(&mm->page_table_lock);
609         if (new)
610                 pte_free(mm, new);
611         if (wait_split_huge_page)
612                 wait_split_huge_page(vma->anon_vma, pmd);
613         return 0;
614 }
615
616 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
617 {
618         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
619         if (!new)
620                 return -ENOMEM;
621
622         smp_wmb(); /* See comment in __pte_alloc */
623
624         spin_lock(&init_mm.page_table_lock);
625         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
626                 pmd_populate_kernel(&init_mm, pmd, new);
627                 new = NULL;
628         } else
629                 VM_BUG_ON(pmd_trans_splitting(*pmd));
630         spin_unlock(&init_mm.page_table_lock);
631         if (new)
632                 pte_free_kernel(&init_mm, new);
633         return 0;
634 }
635
636 static inline void init_rss_vec(int *rss)
637 {
638         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
639 }
640
641 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
642 {
643         int i;
644
645         if (current->mm == mm)
646                 sync_mm_rss(mm);
647         for (i = 0; i < NR_MM_COUNTERS; i++)
648                 if (rss[i])
649                         add_mm_counter(mm, i, rss[i]);
650 }
651
652 /*
653  * This function is called to print an error when a bad pte
654  * is found. For example, we might have a PFN-mapped pte in
655  * a region that doesn't allow it.
656  *
657  * The calling function must still handle the error.
658  */
659 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
660                           pte_t pte, struct page *page)
661 {
662         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
663         pud_t *pud = pud_offset(pgd, addr);
664         pmd_t *pmd = pmd_offset(pud, addr);
665         struct address_space *mapping;
666         pgoff_t index;
667         static unsigned long resume;
668         static unsigned long nr_shown;
669         static unsigned long nr_unshown;
670
671         /*
672          * Allow a burst of 60 reports, then keep quiet for that minute;
673          * or allow a steady drip of one report per second.
674          */
675         if (nr_shown == 60) {
676                 if (time_before(jiffies, resume)) {
677                         nr_unshown++;
678                         return;
679                 }
680                 if (nr_unshown) {
681                         printk(KERN_ALERT
682                                 "BUG: Bad page map: %lu messages suppressed\n",
683                                 nr_unshown);
684                         nr_unshown = 0;
685                 }
686                 nr_shown = 0;
687         }
688         if (nr_shown++ == 0)
689                 resume = jiffies + 60 * HZ;
690
691         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
692         index = linear_page_index(vma, addr);
693
694         printk(KERN_ALERT
695                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
696                 current->comm,
697                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
698         if (page)
699                 dump_page(page);
700         printk(KERN_ALERT
701                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
702                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
703         /*
704          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
705          */
706         if (vma->vm_ops)
707                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
708                                 (unsigned long)vma->vm_ops->fault);
709         if (vma->vm_file && vma->vm_file->f_op)
710                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
711                                 (unsigned long)vma->vm_file->f_op->mmap);
712         dump_stack();
713         add_taint(TAINT_BAD_PAGE);
714 }
715
716 static inline bool is_cow_mapping(vm_flags_t flags)
717 {
718         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
719 }
720
721 #ifndef is_zero_pfn
722 static inline int is_zero_pfn(unsigned long pfn)
723 {
724         return pfn == zero_pfn;
725 }
726 #endif
727
728 #ifndef my_zero_pfn
729 static inline unsigned long my_zero_pfn(unsigned long addr)
730 {
731         return zero_pfn;
732 }
733 #endif
734
735 /*
736  * vm_normal_page -- This function gets the "struct page" associated with a pte.
737  *
738  * "Special" mappings do not wish to be associated with a "struct page" (either
739  * it doesn't exist, or it exists but they don't want to touch it). In this
740  * case, NULL is returned here. "Normal" mappings do have a struct page.
741  *
742  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
743  * pte bit, in which case this function is trivial. Secondly, an architecture
744  * may not have a spare pte bit, which requires a more complicated scheme,
745  * described below.
746  *
747  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
748  * special mapping (even if there are underlying and valid "struct pages").
749  * COWed pages of a VM_PFNMAP are always normal.
750  *
751  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
752  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
753  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
754  * mapping will always honor the rule
755  *
756  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
757  *
758  * And for normal mappings this is false.
759  *
760  * This restricts such mappings to be a linear translation from virtual address
761  * to pfn. To get around this restriction, we allow arbitrary mappings so long
762  * as the vma is not a COW mapping; in that case, we know that all ptes are
763  * special (because none can have been COWed).
764  *
765  *
766  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
767  *
768  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
769  * page" backing, however the difference is that _all_ pages with a struct
770  * page (that is, those where pfn_valid is true) are refcounted and considered
771  * normal pages by the VM. The disadvantage is that pages are refcounted
772  * (which can be slower and simply not an option for some PFNMAP users). The
773  * advantage is that we don't have to follow the strict linearity rule of
774  * PFNMAP mappings in order to support COWable mappings.
775  *
776  */
777 #ifdef __HAVE_ARCH_PTE_SPECIAL
778 # define HAVE_PTE_SPECIAL 1
779 #else
780 # define HAVE_PTE_SPECIAL 0
781 #endif
782 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
783                                 pte_t pte)
784 {
785         unsigned long pfn = pte_pfn(pte);
786
787         if (HAVE_PTE_SPECIAL) {
788                 if (likely(!pte_special(pte)))
789                         goto check_pfn;
790                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
791                         return NULL;
792                 if (!is_zero_pfn(pfn))
793                         print_bad_pte(vma, addr, pte, NULL);
794                 return NULL;
795         }
796
797         /* !HAVE_PTE_SPECIAL case follows: */
798
799         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
800                 if (vma->vm_flags & VM_MIXEDMAP) {
801                         if (!pfn_valid(pfn))
802                                 return NULL;
803                         goto out;
804                 } else {
805                         unsigned long off;
806                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
807                         if (pfn == vma->vm_pgoff + off)
808                                 return NULL;
809                         if (!is_cow_mapping(vma->vm_flags))
810                                 return NULL;
811                 }
812         }
813
814         if (is_zero_pfn(pfn))
815                 return NULL;
816 check_pfn:
817         if (unlikely(pfn > highest_memmap_pfn)) {
818                 print_bad_pte(vma, addr, pte, NULL);
819                 return NULL;
820         }
821
822         /*
823          * NOTE! We still have PageReserved() pages in the page tables.
824          * eg. VDSO mappings can cause them to exist.
825          */
826 out:
827         return pfn_to_page(pfn);
828 }
829
830 /*
831  * copy one vm_area from one task to the other. Assumes the page tables
832  * already present in the new task to be cleared in the whole range
833  * covered by this vma.
834  */
835
836 static inline unsigned long
837 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
838                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
839                 unsigned long addr, int *rss)
840 {
841         unsigned long vm_flags = vma->vm_flags;
842         pte_t pte = *src_pte;
843         struct page *page;
844
845         /* pte contains position in swap or file, so copy. */
846         if (unlikely(!pte_present(pte))) {
847                 if (!pte_file(pte)) {
848                         swp_entry_t entry = pte_to_swp_entry(pte);
849
850                         if (swap_duplicate(entry) < 0)
851                                 return entry.val;
852
853                         /* make sure dst_mm is on swapoff's mmlist. */
854                         if (unlikely(list_empty(&dst_mm->mmlist))) {
855                                 spin_lock(&mmlist_lock);
856                                 if (list_empty(&dst_mm->mmlist))
857                                         list_add(&dst_mm->mmlist,
858                                                  &src_mm->mmlist);
859                                 spin_unlock(&mmlist_lock);
860                         }
861                         if (likely(!non_swap_entry(entry)))
862                                 rss[MM_SWAPENTS]++;
863                         else if (is_migration_entry(entry)) {
864                                 page = migration_entry_to_page(entry);
865
866                                 if (PageAnon(page))
867                                         rss[MM_ANONPAGES]++;
868                                 else
869                                         rss[MM_FILEPAGES]++;
870
871                                 if (is_write_migration_entry(entry) &&
872                                     is_cow_mapping(vm_flags)) {
873                                         /*
874                                          * COW mappings require pages in both
875                                          * parent and child to be set to read.
876                                          */
877                                         make_migration_entry_read(&entry);
878                                         pte = swp_entry_to_pte(entry);
879                                         set_pte_at(src_mm, addr, src_pte, pte);
880                                 }
881                         }
882                 }
883                 goto out_set_pte;
884         }
885
886         /*
887          * If it's a COW mapping, write protect it both
888          * in the parent and the child
889          */
890         if (is_cow_mapping(vm_flags)) {
891                 ptep_set_wrprotect(src_mm, addr, src_pte);
892                 pte = pte_wrprotect(pte);
893         }
894
895         /*
896          * If it's a shared mapping, mark it clean in
897          * the child
898          */
899         if (vm_flags & VM_SHARED)
900                 pte = pte_mkclean(pte);
901         pte = pte_mkold(pte);
902
903         page = vm_normal_page(vma, addr, pte);
904         if (page) {
905                 get_page(page);
906                 page_dup_rmap(page);
907                 if (PageAnon(page))
908                         rss[MM_ANONPAGES]++;
909                 else
910                         rss[MM_FILEPAGES]++;
911         }
912
913 out_set_pte:
914         set_pte_at(dst_mm, addr, dst_pte, pte);
915         return 0;
916 }
917
918 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
919                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
920                    unsigned long addr, unsigned long end)
921 {
922         pte_t *orig_src_pte, *orig_dst_pte;
923         pte_t *src_pte, *dst_pte;
924         spinlock_t *src_ptl, *dst_ptl;
925         int progress = 0;
926         int rss[NR_MM_COUNTERS];
927         swp_entry_t entry = (swp_entry_t){0};
928
929 again:
930         init_rss_vec(rss);
931
932         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
933         if (!dst_pte)
934                 return -ENOMEM;
935         src_pte = pte_offset_map(src_pmd, addr);
936         src_ptl = pte_lockptr(src_mm, src_pmd);
937         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
938         orig_src_pte = src_pte;
939         orig_dst_pte = dst_pte;
940         arch_enter_lazy_mmu_mode();
941
942         do {
943                 /*
944                  * We are holding two locks at this point - either of them
945                  * could generate latencies in another task on another CPU.
946                  */
947                 if (progress >= 32) {
948                         progress = 0;
949                         if (need_resched() ||
950                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
951                                 break;
952                 }
953                 if (pte_none(*src_pte)) {
954                         progress++;
955                         continue;
956                 }
957                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
958                                                         vma, addr, rss);
959                 if (entry.val)
960                         break;
961                 progress += 8;
962         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
963
964         arch_leave_lazy_mmu_mode();
965         spin_unlock(src_ptl);
966         pte_unmap(orig_src_pte);
967         add_mm_rss_vec(dst_mm, rss);
968         pte_unmap_unlock(orig_dst_pte, dst_ptl);
969         cond_resched();
970
971         if (entry.val) {
972                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
973                         return -ENOMEM;
974                 progress = 0;
975         }
976         if (addr != end)
977                 goto again;
978         return 0;
979 }
980
981 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
982                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
983                 unsigned long addr, unsigned long end)
984 {
985         pmd_t *src_pmd, *dst_pmd;
986         unsigned long next;
987
988         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
989         if (!dst_pmd)
990                 return -ENOMEM;
991         src_pmd = pmd_offset(src_pud, addr);
992         do {
993                 next = pmd_addr_end(addr, end);
994                 if (pmd_trans_huge(*src_pmd)) {
995                         int err;
996                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
997                         err = copy_huge_pmd(dst_mm, src_mm,
998                                             dst_pmd, src_pmd, addr, vma);
999                         if (err == -ENOMEM)
1000                                 return -ENOMEM;
1001                         if (!err)
1002                                 continue;
1003                         /* fall through */
1004                 }
1005                 if (pmd_none_or_clear_bad(src_pmd))
1006                         continue;
1007                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1008                                                 vma, addr, next))
1009                         return -ENOMEM;
1010         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1011         return 0;
1012 }
1013
1014 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1015                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1016                 unsigned long addr, unsigned long end)
1017 {
1018         pud_t *src_pud, *dst_pud;
1019         unsigned long next;
1020
1021         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1022         if (!dst_pud)
1023                 return -ENOMEM;
1024         src_pud = pud_offset(src_pgd, addr);
1025         do {
1026                 next = pud_addr_end(addr, end);
1027                 if (pud_none_or_clear_bad(src_pud))
1028                         continue;
1029                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1030                                                 vma, addr, next))
1031                         return -ENOMEM;
1032         } while (dst_pud++, src_pud++, addr = next, addr != end);
1033         return 0;
1034 }
1035
1036 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1037                 struct vm_area_struct *vma)
1038 {
1039         pgd_t *src_pgd, *dst_pgd;
1040         unsigned long next;
1041         unsigned long addr = vma->vm_start;
1042         unsigned long end = vma->vm_end;
1043         unsigned long mmun_start;       /* For mmu_notifiers */
1044         unsigned long mmun_end;         /* For mmu_notifiers */
1045         bool is_cow;
1046         int ret;
1047
1048         /*
1049          * Don't copy ptes where a page fault will fill them correctly.
1050          * Fork becomes much lighter when there are big shared or private
1051          * readonly mappings. The tradeoff is that copy_page_range is more
1052          * efficient than faulting.
1053          */
1054         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1055                                VM_PFNMAP | VM_MIXEDMAP))) {
1056                 if (!vma->anon_vma)
1057                         return 0;
1058         }
1059
1060         if (is_vm_hugetlb_page(vma))
1061                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1062
1063         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1064                 /*
1065                  * We do not free on error cases below as remove_vma
1066                  * gets called on error from higher level routine
1067                  */
1068                 ret = track_pfn_copy(vma);
1069                 if (ret)
1070                         return ret;
1071         }
1072
1073         /*
1074          * We need to invalidate the secondary MMU mappings only when
1075          * there could be a permission downgrade on the ptes of the
1076          * parent mm. And a permission downgrade will only happen if
1077          * is_cow_mapping() returns true.
1078          */
1079         is_cow = is_cow_mapping(vma->vm_flags);
1080         mmun_start = addr;
1081         mmun_end   = end;
1082         if (is_cow)
1083                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1084                                                     mmun_end);
1085
1086         ret = 0;
1087         dst_pgd = pgd_offset(dst_mm, addr);
1088         src_pgd = pgd_offset(src_mm, addr);
1089         do {
1090                 next = pgd_addr_end(addr, end);
1091                 if (pgd_none_or_clear_bad(src_pgd))
1092                         continue;
1093                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1094                                             vma, addr, next))) {
1095                         ret = -ENOMEM;
1096                         break;
1097                 }
1098         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1099
1100         if (is_cow)
1101                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1102         return ret;
1103 }
1104
1105 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1106                                 struct vm_area_struct *vma, pmd_t *pmd,
1107                                 unsigned long addr, unsigned long end,
1108                                 struct zap_details *details)
1109 {
1110         struct mm_struct *mm = tlb->mm;
1111         int force_flush = 0;
1112         int rss[NR_MM_COUNTERS];
1113         spinlock_t *ptl;
1114         pte_t *start_pte;
1115         pte_t *pte;
1116
1117 again:
1118         init_rss_vec(rss);
1119         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1120         pte = start_pte;
1121         arch_enter_lazy_mmu_mode();
1122         do {
1123                 pte_t ptent = *pte;
1124                 if (pte_none(ptent)) {
1125                         continue;
1126                 }
1127
1128                 if (pte_present(ptent)) {
1129                         struct page *page;
1130
1131                         page = vm_normal_page(vma, addr, ptent);
1132                         if (unlikely(details) && page) {
1133                                 /*
1134                                  * unmap_shared_mapping_pages() wants to
1135                                  * invalidate cache without truncating:
1136                                  * unmap shared but keep private pages.
1137                                  */
1138                                 if (details->check_mapping &&
1139                                     details->check_mapping != page->mapping)
1140                                         continue;
1141                                 /*
1142                                  * Each page->index must be checked when
1143                                  * invalidating or truncating nonlinear.
1144                                  */
1145                                 if (details->nonlinear_vma &&
1146                                     (page->index < details->first_index ||
1147                                      page->index > details->last_index))
1148                                         continue;
1149                         }
1150                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1151                                                         tlb->fullmm);
1152                         tlb_remove_tlb_entry(tlb, pte, addr);
1153                         if (unlikely(!page))
1154                                 continue;
1155                         if (unlikely(details) && details->nonlinear_vma
1156                             && linear_page_index(details->nonlinear_vma,
1157                                                 addr) != page->index)
1158                                 set_pte_at(mm, addr, pte,
1159                                            pgoff_to_pte(page->index));
1160                         if (PageAnon(page))
1161                                 rss[MM_ANONPAGES]--;
1162                         else {
1163                                 if (pte_dirty(ptent))
1164                                         set_page_dirty(page);
1165                                 if (pte_young(ptent) &&
1166                                     likely(!VM_SequentialReadHint(vma)))
1167                                         mark_page_accessed(page);
1168                                 rss[MM_FILEPAGES]--;
1169                         }
1170                         page_remove_rmap(page);
1171                         if (unlikely(page_mapcount(page) < 0))
1172                                 print_bad_pte(vma, addr, ptent, page);
1173                         force_flush = !__tlb_remove_page(tlb, page);
1174                         if (force_flush)
1175                                 break;
1176                         continue;
1177                 }
1178                 /*
1179                  * If details->check_mapping, we leave swap entries;
1180                  * if details->nonlinear_vma, we leave file entries.
1181                  */
1182                 if (unlikely(details))
1183                         continue;
1184                 if (pte_file(ptent)) {
1185                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1186                                 print_bad_pte(vma, addr, ptent, NULL);
1187                 } else {
1188                         swp_entry_t entry = pte_to_swp_entry(ptent);
1189
1190                         if (!non_swap_entry(entry))
1191                                 rss[MM_SWAPENTS]--;
1192                         else if (is_migration_entry(entry)) {
1193                                 struct page *page;
1194
1195                                 page = migration_entry_to_page(entry);
1196
1197                                 if (PageAnon(page))
1198                                         rss[MM_ANONPAGES]--;
1199                                 else
1200                                         rss[MM_FILEPAGES]--;
1201                         }
1202                         if (unlikely(!free_swap_and_cache(entry)))
1203                                 print_bad_pte(vma, addr, ptent, NULL);
1204                 }
1205                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1206         } while (pte++, addr += PAGE_SIZE, addr != end);
1207
1208         add_mm_rss_vec(mm, rss);
1209         arch_leave_lazy_mmu_mode();
1210         pte_unmap_unlock(start_pte, ptl);
1211
1212         /*
1213          * mmu_gather ran out of room to batch pages, we break out of
1214          * the PTE lock to avoid doing the potential expensive TLB invalidate
1215          * and page-free while holding it.
1216          */
1217         if (force_flush) {
1218                 force_flush = 0;
1219
1220 #ifdef HAVE_GENERIC_MMU_GATHER
1221                 tlb->start = addr;
1222                 tlb->end = end;
1223 #endif
1224                 tlb_flush_mmu(tlb);
1225                 if (addr != end)
1226                         goto again;
1227         }
1228
1229         return addr;
1230 }
1231
1232 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1233                                 struct vm_area_struct *vma, pud_t *pud,
1234                                 unsigned long addr, unsigned long end,
1235                                 struct zap_details *details)
1236 {
1237         pmd_t *pmd;
1238         unsigned long next;
1239
1240         pmd = pmd_offset(pud, addr);
1241         do {
1242                 next = pmd_addr_end(addr, end);
1243                 if (pmd_trans_huge(*pmd)) {
1244                         if (next - addr != HPAGE_PMD_SIZE) {
1245 #ifdef CONFIG_DEBUG_VM
1246                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1247                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1248                                                 __func__, addr, end,
1249                                                 vma->vm_start,
1250                                                 vma->vm_end);
1251                                         BUG();
1252                                 }
1253 #endif
1254                                 split_huge_page_pmd(vma->vm_mm, pmd);
1255                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1256                                 goto next;
1257                         /* fall through */
1258                 }
1259                 /*
1260                  * Here there can be other concurrent MADV_DONTNEED or
1261                  * trans huge page faults running, and if the pmd is
1262                  * none or trans huge it can change under us. This is
1263                  * because MADV_DONTNEED holds the mmap_sem in read
1264                  * mode.
1265                  */
1266                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1267                         goto next;
1268                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1269 next:
1270                 cond_resched();
1271         } while (pmd++, addr = next, addr != end);
1272
1273         return addr;
1274 }
1275
1276 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1277                                 struct vm_area_struct *vma, pgd_t *pgd,
1278                                 unsigned long addr, unsigned long end,
1279                                 struct zap_details *details)
1280 {
1281         pud_t *pud;
1282         unsigned long next;
1283
1284         pud = pud_offset(pgd, addr);
1285         do {
1286                 next = pud_addr_end(addr, end);
1287                 if (pud_none_or_clear_bad(pud))
1288                         continue;
1289                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1290         } while (pud++, addr = next, addr != end);
1291
1292         return addr;
1293 }
1294
1295 static void unmap_page_range(struct mmu_gather *tlb,
1296                              struct vm_area_struct *vma,
1297                              unsigned long addr, unsigned long end,
1298                              struct zap_details *details)
1299 {
1300         pgd_t *pgd;
1301         unsigned long next;
1302
1303         if (details && !details->check_mapping && !details->nonlinear_vma)
1304                 details = NULL;
1305
1306         BUG_ON(addr >= end);
1307         mem_cgroup_uncharge_start();
1308         tlb_start_vma(tlb, vma);
1309         pgd = pgd_offset(vma->vm_mm, addr);
1310         do {
1311                 next = pgd_addr_end(addr, end);
1312                 if (pgd_none_or_clear_bad(pgd))
1313                         continue;
1314                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1315         } while (pgd++, addr = next, addr != end);
1316         tlb_end_vma(tlb, vma);
1317         mem_cgroup_uncharge_end();
1318 }
1319
1320
1321 static void unmap_single_vma(struct mmu_gather *tlb,
1322                 struct vm_area_struct *vma, unsigned long start_addr,
1323                 unsigned long end_addr,
1324                 struct zap_details *details)
1325 {
1326         unsigned long start = max(vma->vm_start, start_addr);
1327         unsigned long end;
1328
1329         if (start >= vma->vm_end)
1330                 return;
1331         end = min(vma->vm_end, end_addr);
1332         if (end <= vma->vm_start)
1333                 return;
1334
1335         if (vma->vm_file)
1336                 uprobe_munmap(vma, start, end);
1337
1338         if (unlikely(vma->vm_flags & VM_PFNMAP))
1339                 untrack_pfn(vma, 0, 0);
1340
1341         if (start != end) {
1342                 if (unlikely(is_vm_hugetlb_page(vma))) {
1343                         /*
1344                          * It is undesirable to test vma->vm_file as it
1345                          * should be non-null for valid hugetlb area.
1346                          * However, vm_file will be NULL in the error
1347                          * cleanup path of do_mmap_pgoff. When
1348                          * hugetlbfs ->mmap method fails,
1349                          * do_mmap_pgoff() nullifies vma->vm_file
1350                          * before calling this function to clean up.
1351                          * Since no pte has actually been setup, it is
1352                          * safe to do nothing in this case.
1353                          */
1354                         if (vma->vm_file) {
1355                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1356                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1357                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1358                         }
1359                 } else
1360                         unmap_page_range(tlb, vma, start, end, details);
1361         }
1362 }
1363
1364 /**
1365  * unmap_vmas - unmap a range of memory covered by a list of vma's
1366  * @tlb: address of the caller's struct mmu_gather
1367  * @vma: the starting vma
1368  * @start_addr: virtual address at which to start unmapping
1369  * @end_addr: virtual address at which to end unmapping
1370  *
1371  * Unmap all pages in the vma list.
1372  *
1373  * Only addresses between `start' and `end' will be unmapped.
1374  *
1375  * The VMA list must be sorted in ascending virtual address order.
1376  *
1377  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1378  * range after unmap_vmas() returns.  So the only responsibility here is to
1379  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1380  * drops the lock and schedules.
1381  */
1382 void unmap_vmas(struct mmu_gather *tlb,
1383                 struct vm_area_struct *vma, unsigned long start_addr,
1384                 unsigned long end_addr)
1385 {
1386         struct mm_struct *mm = vma->vm_mm;
1387
1388         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1389         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1390                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1391         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1392 }
1393
1394 /**
1395  * zap_page_range - remove user pages in a given range
1396  * @vma: vm_area_struct holding the applicable pages
1397  * @start: starting address of pages to zap
1398  * @size: number of bytes to zap
1399  * @details: details of nonlinear truncation or shared cache invalidation
1400  *
1401  * Caller must protect the VMA list
1402  */
1403 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1404                 unsigned long size, struct zap_details *details)
1405 {
1406         struct mm_struct *mm = vma->vm_mm;
1407         struct mmu_gather tlb;
1408         unsigned long end = start + size;
1409
1410         lru_add_drain();
1411         tlb_gather_mmu(&tlb, mm, 0);
1412         update_hiwater_rss(mm);
1413         mmu_notifier_invalidate_range_start(mm, start, end);
1414         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1415                 unmap_single_vma(&tlb, vma, start, end, details);
1416         mmu_notifier_invalidate_range_end(mm, start, end);
1417         tlb_finish_mmu(&tlb, start, end);
1418 }
1419
1420 /**
1421  * zap_page_range_single - remove user pages in a given range
1422  * @vma: vm_area_struct holding the applicable pages
1423  * @address: starting address of pages to zap
1424  * @size: number of bytes to zap
1425  * @details: details of nonlinear truncation or shared cache invalidation
1426  *
1427  * The range must fit into one VMA.
1428  */
1429 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1430                 unsigned long size, struct zap_details *details)
1431 {
1432         struct mm_struct *mm = vma->vm_mm;
1433         struct mmu_gather tlb;
1434         unsigned long end = address + size;
1435
1436         lru_add_drain();
1437         tlb_gather_mmu(&tlb, mm, 0);
1438         update_hiwater_rss(mm);
1439         mmu_notifier_invalidate_range_start(mm, address, end);
1440         unmap_single_vma(&tlb, vma, address, end, details);
1441         mmu_notifier_invalidate_range_end(mm, address, end);
1442         tlb_finish_mmu(&tlb, address, end);
1443 }
1444
1445 /**
1446  * zap_vma_ptes - remove ptes mapping the vma
1447  * @vma: vm_area_struct holding ptes to be zapped
1448  * @address: starting address of pages to zap
1449  * @size: number of bytes to zap
1450  *
1451  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1452  *
1453  * The entire address range must be fully contained within the vma.
1454  *
1455  * Returns 0 if successful.
1456  */
1457 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1458                 unsigned long size)
1459 {
1460         if (address < vma->vm_start || address + size > vma->vm_end ||
1461                         !(vma->vm_flags & VM_PFNMAP))
1462                 return -1;
1463         zap_page_range_single(vma, address, size, NULL);
1464         return 0;
1465 }
1466 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1467
1468 /**
1469  * follow_page - look up a page descriptor from a user-virtual address
1470  * @vma: vm_area_struct mapping @address
1471  * @address: virtual address to look up
1472  * @flags: flags modifying lookup behaviour
1473  *
1474  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1475  *
1476  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1477  * an error pointer if there is a mapping to something not represented
1478  * by a page descriptor (see also vm_normal_page()).
1479  */
1480 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1481                         unsigned int flags)
1482 {
1483         pgd_t *pgd;
1484         pud_t *pud;
1485         pmd_t *pmd;
1486         pte_t *ptep, pte;
1487         spinlock_t *ptl;
1488         struct page *page;
1489         struct mm_struct *mm = vma->vm_mm;
1490
1491         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1492         if (!IS_ERR(page)) {
1493                 BUG_ON(flags & FOLL_GET);
1494                 goto out;
1495         }
1496
1497         page = NULL;
1498         pgd = pgd_offset(mm, address);
1499         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1500                 goto no_page_table;
1501
1502         pud = pud_offset(pgd, address);
1503         if (pud_none(*pud))
1504                 goto no_page_table;
1505         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1506                 BUG_ON(flags & FOLL_GET);
1507                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1508                 goto out;
1509         }
1510         if (unlikely(pud_bad(*pud)))
1511                 goto no_page_table;
1512
1513         pmd = pmd_offset(pud, address);
1514         if (pmd_none(*pmd))
1515                 goto no_page_table;
1516         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1517                 BUG_ON(flags & FOLL_GET);
1518                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1519                 goto out;
1520         }
1521         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1522                 goto no_page_table;
1523         if (pmd_trans_huge(*pmd)) {
1524                 if (flags & FOLL_SPLIT) {
1525                         split_huge_page_pmd(mm, pmd);
1526                         goto split_fallthrough;
1527                 }
1528                 spin_lock(&mm->page_table_lock);
1529                 if (likely(pmd_trans_huge(*pmd))) {
1530                         if (unlikely(pmd_trans_splitting(*pmd))) {
1531                                 spin_unlock(&mm->page_table_lock);
1532                                 wait_split_huge_page(vma->anon_vma, pmd);
1533                         } else {
1534                                 page = follow_trans_huge_pmd(vma, address,
1535                                                              pmd, flags);
1536                                 spin_unlock(&mm->page_table_lock);
1537                                 goto out;
1538                         }
1539                 } else
1540                         spin_unlock(&mm->page_table_lock);
1541                 /* fall through */
1542         }
1543 split_fallthrough:
1544         if (unlikely(pmd_bad(*pmd)))
1545                 goto no_page_table;
1546
1547         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1548
1549         pte = *ptep;
1550         if (!pte_present(pte))
1551                 goto no_page;
1552         if ((flags & FOLL_NUMA) && pte_numa(pte))
1553                 goto no_page;
1554         if ((flags & FOLL_WRITE) && !pte_write(pte))
1555                 goto unlock;
1556
1557         page = vm_normal_page(vma, address, pte);
1558         if (unlikely(!page)) {
1559                 if ((flags & FOLL_DUMP) ||
1560                     !is_zero_pfn(pte_pfn(pte)))
1561                         goto bad_page;
1562                 page = pte_page(pte);
1563         }
1564
1565         if (flags & FOLL_GET)
1566                 get_page_foll(page);
1567         if (flags & FOLL_TOUCH) {
1568                 if ((flags & FOLL_WRITE) &&
1569                     !pte_dirty(pte) && !PageDirty(page))
1570                         set_page_dirty(page);
1571                 /*
1572                  * pte_mkyoung() would be more correct here, but atomic care
1573                  * is needed to avoid losing the dirty bit: it is easier to use
1574                  * mark_page_accessed().
1575                  */
1576                 mark_page_accessed(page);
1577         }
1578         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1579                 /*
1580                  * The preliminary mapping check is mainly to avoid the
1581                  * pointless overhead of lock_page on the ZERO_PAGE
1582                  * which might bounce very badly if there is contention.
1583                  *
1584                  * If the page is already locked, we don't need to
1585                  * handle it now - vmscan will handle it later if and
1586                  * when it attempts to reclaim the page.
1587                  */
1588                 if (page->mapping && trylock_page(page)) {
1589                         lru_add_drain();  /* push cached pages to LRU */
1590                         /*
1591                          * Because we lock page here, and migration is
1592                          * blocked by the pte's page reference, and we
1593                          * know the page is still mapped, we don't even
1594                          * need to check for file-cache page truncation.
1595                          */
1596                         mlock_vma_page(page);
1597                         unlock_page(page);
1598                 }
1599         }
1600 unlock:
1601         pte_unmap_unlock(ptep, ptl);
1602 out:
1603         return page;
1604
1605 bad_page:
1606         pte_unmap_unlock(ptep, ptl);
1607         return ERR_PTR(-EFAULT);
1608
1609 no_page:
1610         pte_unmap_unlock(ptep, ptl);
1611         if (!pte_none(pte))
1612                 return page;
1613
1614 no_page_table:
1615         /*
1616          * When core dumping an enormous anonymous area that nobody
1617          * has touched so far, we don't want to allocate unnecessary pages or
1618          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1619          * then get_dump_page() will return NULL to leave a hole in the dump.
1620          * But we can only make this optimization where a hole would surely
1621          * be zero-filled if handle_mm_fault() actually did handle it.
1622          */
1623         if ((flags & FOLL_DUMP) &&
1624             (!vma->vm_ops || !vma->vm_ops->fault))
1625                 return ERR_PTR(-EFAULT);
1626         return page;
1627 }
1628
1629 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1630 {
1631         return stack_guard_page_start(vma, addr) ||
1632                stack_guard_page_end(vma, addr+PAGE_SIZE);
1633 }
1634
1635 /**
1636  * __get_user_pages() - pin user pages in memory
1637  * @tsk:        task_struct of target task
1638  * @mm:         mm_struct of target mm
1639  * @start:      starting user address
1640  * @nr_pages:   number of pages from start to pin
1641  * @gup_flags:  flags modifying pin behaviour
1642  * @pages:      array that receives pointers to the pages pinned.
1643  *              Should be at least nr_pages long. Or NULL, if caller
1644  *              only intends to ensure the pages are faulted in.
1645  * @vmas:       array of pointers to vmas corresponding to each page.
1646  *              Or NULL if the caller does not require them.
1647  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1648  *
1649  * Returns number of pages pinned. This may be fewer than the number
1650  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1651  * were pinned, returns -errno. Each page returned must be released
1652  * with a put_page() call when it is finished with. vmas will only
1653  * remain valid while mmap_sem is held.
1654  *
1655  * Must be called with mmap_sem held for read or write.
1656  *
1657  * __get_user_pages walks a process's page tables and takes a reference to
1658  * each struct page that each user address corresponds to at a given
1659  * instant. That is, it takes the page that would be accessed if a user
1660  * thread accesses the given user virtual address at that instant.
1661  *
1662  * This does not guarantee that the page exists in the user mappings when
1663  * __get_user_pages returns, and there may even be a completely different
1664  * page there in some cases (eg. if mmapped pagecache has been invalidated
1665  * and subsequently re faulted). However it does guarantee that the page
1666  * won't be freed completely. And mostly callers simply care that the page
1667  * contains data that was valid *at some point in time*. Typically, an IO
1668  * or similar operation cannot guarantee anything stronger anyway because
1669  * locks can't be held over the syscall boundary.
1670  *
1671  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1672  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1673  * appropriate) must be called after the page is finished with, and
1674  * before put_page is called.
1675  *
1676  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1677  * or mmap_sem contention, and if waiting is needed to pin all pages,
1678  * *@nonblocking will be set to 0.
1679  *
1680  * In most cases, get_user_pages or get_user_pages_fast should be used
1681  * instead of __get_user_pages. __get_user_pages should be used only if
1682  * you need some special @gup_flags.
1683  */
1684 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1685                      unsigned long start, int nr_pages, unsigned int gup_flags,
1686                      struct page **pages, struct vm_area_struct **vmas,
1687                      int *nonblocking)
1688 {
1689         int i;
1690         unsigned long vm_flags;
1691
1692         if (nr_pages <= 0)
1693                 return 0;
1694
1695         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1696
1697         /* 
1698          * Require read or write permissions.
1699          * If FOLL_FORCE is set, we only require the "MAY" flags.
1700          */
1701         vm_flags  = (gup_flags & FOLL_WRITE) ?
1702                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1703         vm_flags &= (gup_flags & FOLL_FORCE) ?
1704                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1705
1706         /*
1707          * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1708          * would be called on PROT_NONE ranges. We must never invoke
1709          * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1710          * page faults would unprotect the PROT_NONE ranges if
1711          * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1712          * bitflag. So to avoid that, don't set FOLL_NUMA if
1713          * FOLL_FORCE is set.
1714          */
1715         if (!(gup_flags & FOLL_FORCE))
1716                 gup_flags |= FOLL_NUMA;
1717
1718         i = 0;
1719
1720         do {
1721                 struct vm_area_struct *vma;
1722
1723                 vma = find_extend_vma(mm, start);
1724                 if (!vma && in_gate_area(mm, start)) {
1725                         unsigned long pg = start & PAGE_MASK;
1726                         pgd_t *pgd;
1727                         pud_t *pud;
1728                         pmd_t *pmd;
1729                         pte_t *pte;
1730
1731                         /* user gate pages are read-only */
1732                         if (gup_flags & FOLL_WRITE)
1733                                 return i ? : -EFAULT;
1734                         if (pg > TASK_SIZE)
1735                                 pgd = pgd_offset_k(pg);
1736                         else
1737                                 pgd = pgd_offset_gate(mm, pg);
1738                         BUG_ON(pgd_none(*pgd));
1739                         pud = pud_offset(pgd, pg);
1740                         BUG_ON(pud_none(*pud));
1741                         pmd = pmd_offset(pud, pg);
1742                         if (pmd_none(*pmd))
1743                                 return i ? : -EFAULT;
1744                         VM_BUG_ON(pmd_trans_huge(*pmd));
1745                         pte = pte_offset_map(pmd, pg);
1746                         if (pte_none(*pte)) {
1747                                 pte_unmap(pte);
1748                                 return i ? : -EFAULT;
1749                         }
1750                         vma = get_gate_vma(mm);
1751                         if (pages) {
1752                                 struct page *page;
1753
1754                                 page = vm_normal_page(vma, start, *pte);
1755                                 if (!page) {
1756                                         if (!(gup_flags & FOLL_DUMP) &&
1757                                              is_zero_pfn(pte_pfn(*pte)))
1758                                                 page = pte_page(*pte);
1759                                         else {
1760                                                 pte_unmap(pte);
1761                                                 return i ? : -EFAULT;
1762                                         }
1763                                 }
1764                                 pages[i] = page;
1765                                 get_page(page);
1766                         }
1767                         pte_unmap(pte);
1768                         goto next_page;
1769                 }
1770
1771                 if (!vma ||
1772                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1773                     !(vm_flags & vma->vm_flags))
1774                         return i ? : -EFAULT;
1775
1776                 if (is_vm_hugetlb_page(vma)) {
1777                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1778                                         &start, &nr_pages, i, gup_flags);
1779                         continue;
1780                 }
1781
1782                 do {
1783                         struct page *page;
1784                         unsigned int foll_flags = gup_flags;
1785
1786                         /*
1787                          * If we have a pending SIGKILL, don't keep faulting
1788                          * pages and potentially allocating memory.
1789                          */
1790                         if (unlikely(fatal_signal_pending(current)))
1791                                 return i ? i : -ERESTARTSYS;
1792
1793                         cond_resched();
1794                         while (!(page = follow_page(vma, start, foll_flags))) {
1795                                 int ret;
1796                                 unsigned int fault_flags = 0;
1797
1798                                 /* For mlock, just skip the stack guard page. */
1799                                 if (foll_flags & FOLL_MLOCK) {
1800                                         if (stack_guard_page(vma, start))
1801                                                 goto next_page;
1802                                 }
1803                                 if (foll_flags & FOLL_WRITE)
1804                                         fault_flags |= FAULT_FLAG_WRITE;
1805                                 if (nonblocking)
1806                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1807                                 if (foll_flags & FOLL_NOWAIT)
1808                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1809
1810                                 ret = handle_mm_fault(mm, vma, start,
1811                                                         fault_flags);
1812
1813                                 if (ret & VM_FAULT_ERROR) {
1814                                         if (ret & VM_FAULT_OOM)
1815                                                 return i ? i : -ENOMEM;
1816                                         if (ret & (VM_FAULT_HWPOISON |
1817                                                    VM_FAULT_HWPOISON_LARGE)) {
1818                                                 if (i)
1819                                                         return i;
1820                                                 else if (gup_flags & FOLL_HWPOISON)
1821                                                         return -EHWPOISON;
1822                                                 else
1823                                                         return -EFAULT;
1824                                         }
1825                                         if (ret & VM_FAULT_SIGBUS)
1826                                                 return i ? i : -EFAULT;
1827                                         BUG();
1828                                 }
1829
1830                                 if (tsk) {
1831                                         if (ret & VM_FAULT_MAJOR)
1832                                                 tsk->maj_flt++;
1833                                         else
1834                                                 tsk->min_flt++;
1835                                 }
1836
1837                                 if (ret & VM_FAULT_RETRY) {
1838                                         if (nonblocking)
1839                                                 *nonblocking = 0;
1840                                         return i;
1841                                 }
1842
1843                                 /*
1844                                  * The VM_FAULT_WRITE bit tells us that
1845                                  * do_wp_page has broken COW when necessary,
1846                                  * even if maybe_mkwrite decided not to set
1847                                  * pte_write. We can thus safely do subsequent
1848                                  * page lookups as if they were reads. But only
1849                                  * do so when looping for pte_write is futile:
1850                                  * in some cases userspace may also be wanting
1851                                  * to write to the gotten user page, which a
1852                                  * read fault here might prevent (a readonly
1853                                  * page might get reCOWed by userspace write).
1854                                  */
1855                                 if ((ret & VM_FAULT_WRITE) &&
1856                                     !(vma->vm_flags & VM_WRITE))
1857                                         foll_flags &= ~FOLL_WRITE;
1858
1859                                 cond_resched();
1860                         }
1861                         if (IS_ERR(page))
1862                                 return i ? i : PTR_ERR(page);
1863                         if (pages) {
1864                                 pages[i] = page;
1865
1866                                 flush_anon_page(vma, page, start);
1867                                 flush_dcache_page(page);
1868                         }
1869 next_page:
1870                         if (vmas)
1871                                 vmas[i] = vma;
1872                         i++;
1873                         start += PAGE_SIZE;
1874                         nr_pages--;
1875                 } while (nr_pages && start < vma->vm_end);
1876         } while (nr_pages);
1877         return i;
1878 }
1879 EXPORT_SYMBOL(__get_user_pages);
1880
1881 /*
1882  * fixup_user_fault() - manually resolve a user page fault
1883  * @tsk:        the task_struct to use for page fault accounting, or
1884  *              NULL if faults are not to be recorded.
1885  * @mm:         mm_struct of target mm
1886  * @address:    user address
1887  * @fault_flags:flags to pass down to handle_mm_fault()
1888  *
1889  * This is meant to be called in the specific scenario where for locking reasons
1890  * we try to access user memory in atomic context (within a pagefault_disable()
1891  * section), this returns -EFAULT, and we want to resolve the user fault before
1892  * trying again.
1893  *
1894  * Typically this is meant to be used by the futex code.
1895  *
1896  * The main difference with get_user_pages() is that this function will
1897  * unconditionally call handle_mm_fault() which will in turn perform all the
1898  * necessary SW fixup of the dirty and young bits in the PTE, while
1899  * handle_mm_fault() only guarantees to update these in the struct page.
1900  *
1901  * This is important for some architectures where those bits also gate the
1902  * access permission to the page because they are maintained in software.  On
1903  * such architectures, gup() will not be enough to make a subsequent access
1904  * succeed.
1905  *
1906  * This should be called with the mm_sem held for read.
1907  */
1908 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1909                      unsigned long address, unsigned int fault_flags)
1910 {
1911         struct vm_area_struct *vma;
1912         int ret;
1913
1914         vma = find_extend_vma(mm, address);
1915         if (!vma || address < vma->vm_start)
1916                 return -EFAULT;
1917
1918         ret = handle_mm_fault(mm, vma, address, fault_flags);
1919         if (ret & VM_FAULT_ERROR) {
1920                 if (ret & VM_FAULT_OOM)
1921                         return -ENOMEM;
1922                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1923                         return -EHWPOISON;
1924                 if (ret & VM_FAULT_SIGBUS)
1925                         return -EFAULT;
1926                 BUG();
1927         }
1928         if (tsk) {
1929                 if (ret & VM_FAULT_MAJOR)
1930                         tsk->maj_flt++;
1931                 else
1932                         tsk->min_flt++;
1933         }
1934         return 0;
1935 }
1936
1937 /*
1938  * get_user_pages() - pin user pages in memory
1939  * @tsk:        the task_struct to use for page fault accounting, or
1940  *              NULL if faults are not to be recorded.
1941  * @mm:         mm_struct of target mm
1942  * @start:      starting user address
1943  * @nr_pages:   number of pages from start to pin
1944  * @write:      whether pages will be written to by the caller
1945  * @force:      whether to force write access even if user mapping is
1946  *              readonly. This will result in the page being COWed even
1947  *              in MAP_SHARED mappings. You do not want this.
1948  * @pages:      array that receives pointers to the pages pinned.
1949  *              Should be at least nr_pages long. Or NULL, if caller
1950  *              only intends to ensure the pages are faulted in.
1951  * @vmas:       array of pointers to vmas corresponding to each page.
1952  *              Or NULL if the caller does not require them.
1953  *
1954  * Returns number of pages pinned. This may be fewer than the number
1955  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1956  * were pinned, returns -errno. Each page returned must be released
1957  * with a put_page() call when it is finished with. vmas will only
1958  * remain valid while mmap_sem is held.
1959  *
1960  * Must be called with mmap_sem held for read or write.
1961  *
1962  * get_user_pages walks a process's page tables and takes a reference to
1963  * each struct page that each user address corresponds to at a given
1964  * instant. That is, it takes the page that would be accessed if a user
1965  * thread accesses the given user virtual address at that instant.
1966  *
1967  * This does not guarantee that the page exists in the user mappings when
1968  * get_user_pages returns, and there may even be a completely different
1969  * page there in some cases (eg. if mmapped pagecache has been invalidated
1970  * and subsequently re faulted). However it does guarantee that the page
1971  * won't be freed completely. And mostly callers simply care that the page
1972  * contains data that was valid *at some point in time*. Typically, an IO
1973  * or similar operation cannot guarantee anything stronger anyway because
1974  * locks can't be held over the syscall boundary.
1975  *
1976  * If write=0, the page must not be written to. If the page is written to,
1977  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1978  * after the page is finished with, and before put_page is called.
1979  *
1980  * get_user_pages is typically used for fewer-copy IO operations, to get a
1981  * handle on the memory by some means other than accesses via the user virtual
1982  * addresses. The pages may be submitted for DMA to devices or accessed via
1983  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1984  * use the correct cache flushing APIs.
1985  *
1986  * See also get_user_pages_fast, for performance critical applications.
1987  */
1988 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1989                 unsigned long start, int nr_pages, int write, int force,
1990                 struct page **pages, struct vm_area_struct **vmas)
1991 {
1992         int flags = FOLL_TOUCH;
1993
1994         if (pages)
1995                 flags |= FOLL_GET;
1996         if (write)
1997                 flags |= FOLL_WRITE;
1998         if (force)
1999                 flags |= FOLL_FORCE;
2000
2001         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2002                                 NULL);
2003 }
2004 EXPORT_SYMBOL(get_user_pages);
2005
2006 /**
2007  * get_dump_page() - pin user page in memory while writing it to core dump
2008  * @addr: user address
2009  *
2010  * Returns struct page pointer of user page pinned for dump,
2011  * to be freed afterwards by page_cache_release() or put_page().
2012  *
2013  * Returns NULL on any kind of failure - a hole must then be inserted into
2014  * the corefile, to preserve alignment with its headers; and also returns
2015  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2016  * allowing a hole to be left in the corefile to save diskspace.
2017  *
2018  * Called without mmap_sem, but after all other threads have been killed.
2019  */
2020 #ifdef CONFIG_ELF_CORE
2021 struct page *get_dump_page(unsigned long addr)
2022 {
2023         struct vm_area_struct *vma;
2024         struct page *page;
2025
2026         if (__get_user_pages(current, current->mm, addr, 1,
2027                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2028                              NULL) < 1)
2029                 return NULL;
2030         flush_cache_page(vma, addr, page_to_pfn(page));
2031         return page;
2032 }
2033 #endif /* CONFIG_ELF_CORE */
2034
2035 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2036                         spinlock_t **ptl)
2037 {
2038         pgd_t * pgd = pgd_offset(mm, addr);
2039         pud_t * pud = pud_alloc(mm, pgd, addr);
2040         if (pud) {
2041                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2042                 if (pmd) {
2043                         VM_BUG_ON(pmd_trans_huge(*pmd));
2044                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2045                 }
2046         }
2047         return NULL;
2048 }
2049
2050 /*
2051  * This is the old fallback for page remapping.
2052  *
2053  * For historical reasons, it only allows reserved pages. Only
2054  * old drivers should use this, and they needed to mark their
2055  * pages reserved for the old functions anyway.
2056  */
2057 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2058                         struct page *page, pgprot_t prot)
2059 {
2060         struct mm_struct *mm = vma->vm_mm;
2061         int retval;
2062         pte_t *pte;
2063         spinlock_t *ptl;
2064
2065         retval = -EINVAL;
2066         if (PageAnon(page))
2067                 goto out;
2068         retval = -ENOMEM;
2069         flush_dcache_page(page);
2070         pte = get_locked_pte(mm, addr, &ptl);
2071         if (!pte)
2072                 goto out;
2073         retval = -EBUSY;
2074         if (!pte_none(*pte))
2075                 goto out_unlock;
2076
2077         /* Ok, finally just insert the thing.. */
2078         get_page(page);
2079         inc_mm_counter_fast(mm, MM_FILEPAGES);
2080         page_add_file_rmap(page);
2081         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2082
2083         retval = 0;
2084         pte_unmap_unlock(pte, ptl);
2085         return retval;
2086 out_unlock:
2087         pte_unmap_unlock(pte, ptl);
2088 out:
2089         return retval;
2090 }
2091
2092 /**
2093  * vm_insert_page - insert single page into user vma
2094  * @vma: user vma to map to
2095  * @addr: target user address of this page
2096  * @page: source kernel page
2097  *
2098  * This allows drivers to insert individual pages they've allocated
2099  * into a user vma.
2100  *
2101  * The page has to be a nice clean _individual_ kernel allocation.
2102  * If you allocate a compound page, you need to have marked it as
2103  * such (__GFP_COMP), or manually just split the page up yourself
2104  * (see split_page()).
2105  *
2106  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2107  * took an arbitrary page protection parameter. This doesn't allow
2108  * that. Your vma protection will have to be set up correctly, which
2109  * means that if you want a shared writable mapping, you'd better
2110  * ask for a shared writable mapping!
2111  *
2112  * The page does not need to be reserved.
2113  *
2114  * Usually this function is called from f_op->mmap() handler
2115  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2116  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2117  * function from other places, for example from page-fault handler.
2118  */
2119 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2120                         struct page *page)
2121 {
2122         if (addr < vma->vm_start || addr >= vma->vm_end)
2123                 return -EFAULT;
2124         if (!page_count(page))
2125                 return -EINVAL;
2126         if (!(vma->vm_flags & VM_MIXEDMAP)) {
2127                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2128                 BUG_ON(vma->vm_flags & VM_PFNMAP);
2129                 vma->vm_flags |= VM_MIXEDMAP;
2130         }
2131         return insert_page(vma, addr, page, vma->vm_page_prot);
2132 }
2133 EXPORT_SYMBOL(vm_insert_page);
2134
2135 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2136                         unsigned long pfn, pgprot_t prot)
2137 {
2138         struct mm_struct *mm = vma->vm_mm;
2139         int retval;
2140         pte_t *pte, entry;
2141         spinlock_t *ptl;
2142
2143         retval = -ENOMEM;
2144         pte = get_locked_pte(mm, addr, &ptl);
2145         if (!pte)
2146                 goto out;
2147         retval = -EBUSY;
2148         if (!pte_none(*pte))
2149                 goto out_unlock;
2150
2151         /* Ok, finally just insert the thing.. */
2152         entry = pte_mkspecial(pfn_pte(pfn, prot));
2153         set_pte_at(mm, addr, pte, entry);
2154         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2155
2156         retval = 0;
2157 out_unlock:
2158         pte_unmap_unlock(pte, ptl);
2159 out:
2160         return retval;
2161 }
2162
2163 /**
2164  * vm_insert_pfn - insert single pfn into user vma
2165  * @vma: user vma to map to
2166  * @addr: target user address of this page
2167  * @pfn: source kernel pfn
2168  *
2169  * Similar to vm_insert_page, this allows drivers to insert individual pages
2170  * they've allocated into a user vma. Same comments apply.
2171  *
2172  * This function should only be called from a vm_ops->fault handler, and
2173  * in that case the handler should return NULL.
2174  *
2175  * vma cannot be a COW mapping.
2176  *
2177  * As this is called only for pages that do not currently exist, we
2178  * do not need to flush old virtual caches or the TLB.
2179  */
2180 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2181                         unsigned long pfn)
2182 {
2183         int ret;
2184         pgprot_t pgprot = vma->vm_page_prot;
2185         /*
2186          * Technically, architectures with pte_special can avoid all these
2187          * restrictions (same for remap_pfn_range).  However we would like
2188          * consistency in testing and feature parity among all, so we should
2189          * try to keep these invariants in place for everybody.
2190          */
2191         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2192         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2193                                                 (VM_PFNMAP|VM_MIXEDMAP));
2194         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2195         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2196
2197         if (addr < vma->vm_start || addr >= vma->vm_end)
2198                 return -EFAULT;
2199         if (track_pfn_insert(vma, &pgprot, pfn))
2200                 return -EINVAL;
2201
2202         ret = insert_pfn(vma, addr, pfn, pgprot);
2203
2204         return ret;
2205 }
2206 EXPORT_SYMBOL(vm_insert_pfn);
2207
2208 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2209                         unsigned long pfn)
2210 {
2211         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2212
2213         if (addr < vma->vm_start || addr >= vma->vm_end)
2214                 return -EFAULT;
2215
2216         /*
2217          * If we don't have pte special, then we have to use the pfn_valid()
2218          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2219          * refcount the page if pfn_valid is true (hence insert_page rather
2220          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2221          * without pte special, it would there be refcounted as a normal page.
2222          */
2223         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2224                 struct page *page;
2225
2226                 page = pfn_to_page(pfn);
2227                 return insert_page(vma, addr, page, vma->vm_page_prot);
2228         }
2229         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2230 }
2231 EXPORT_SYMBOL(vm_insert_mixed);
2232
2233 /*
2234  * maps a range of physical memory into the requested pages. the old
2235  * mappings are removed. any references to nonexistent pages results
2236  * in null mappings (currently treated as "copy-on-access")
2237  */
2238 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2239                         unsigned long addr, unsigned long end,
2240                         unsigned long pfn, pgprot_t prot)
2241 {
2242         pte_t *pte;
2243         spinlock_t *ptl;
2244
2245         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2246         if (!pte)
2247                 return -ENOMEM;
2248         arch_enter_lazy_mmu_mode();
2249         do {
2250                 BUG_ON(!pte_none(*pte));
2251                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2252                 pfn++;
2253         } while (pte++, addr += PAGE_SIZE, addr != end);
2254         arch_leave_lazy_mmu_mode();
2255         pte_unmap_unlock(pte - 1, ptl);
2256         return 0;
2257 }
2258
2259 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2260                         unsigned long addr, unsigned long end,
2261                         unsigned long pfn, pgprot_t prot)
2262 {
2263         pmd_t *pmd;
2264         unsigned long next;
2265
2266         pfn -= addr >> PAGE_SHIFT;
2267         pmd = pmd_alloc(mm, pud, addr);
2268         if (!pmd)
2269                 return -ENOMEM;
2270         VM_BUG_ON(pmd_trans_huge(*pmd));
2271         do {
2272                 next = pmd_addr_end(addr, end);
2273                 if (remap_pte_range(mm, pmd, addr, next,
2274                                 pfn + (addr >> PAGE_SHIFT), prot))
2275                         return -ENOMEM;
2276         } while (pmd++, addr = next, addr != end);
2277         return 0;
2278 }
2279
2280 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2281                         unsigned long addr, unsigned long end,
2282                         unsigned long pfn, pgprot_t prot)
2283 {
2284         pud_t *pud;
2285         unsigned long next;
2286
2287         pfn -= addr >> PAGE_SHIFT;
2288         pud = pud_alloc(mm, pgd, addr);
2289         if (!pud)
2290                 return -ENOMEM;
2291         do {
2292                 next = pud_addr_end(addr, end);
2293                 if (remap_pmd_range(mm, pud, addr, next,
2294                                 pfn + (addr >> PAGE_SHIFT), prot))
2295                         return -ENOMEM;
2296         } while (pud++, addr = next, addr != end);
2297         return 0;
2298 }
2299
2300 /**
2301  * remap_pfn_range - remap kernel memory to userspace
2302  * @vma: user vma to map to
2303  * @addr: target user address to start at
2304  * @pfn: physical address of kernel memory
2305  * @size: size of map area
2306  * @prot: page protection flags for this mapping
2307  *
2308  *  Note: this is only safe if the mm semaphore is held when called.
2309  */
2310 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2311                     unsigned long pfn, unsigned long size, pgprot_t prot)
2312 {
2313         pgd_t *pgd;
2314         unsigned long next;
2315         unsigned long end = addr + PAGE_ALIGN(size);
2316         struct mm_struct *mm = vma->vm_mm;
2317         int err;
2318
2319         /*
2320          * Physically remapped pages are special. Tell the
2321          * rest of the world about it:
2322          *   VM_IO tells people not to look at these pages
2323          *      (accesses can have side effects).
2324          *   VM_PFNMAP tells the core MM that the base pages are just
2325          *      raw PFN mappings, and do not have a "struct page" associated
2326          *      with them.
2327          *   VM_DONTEXPAND
2328          *      Disable vma merging and expanding with mremap().
2329          *   VM_DONTDUMP
2330          *      Omit vma from core dump, even when VM_IO turned off.
2331          *
2332          * There's a horrible special case to handle copy-on-write
2333          * behaviour that some programs depend on. We mark the "original"
2334          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2335          * See vm_normal_page() for details.
2336          */
2337         if (is_cow_mapping(vma->vm_flags)) {
2338                 if (addr != vma->vm_start || end != vma->vm_end)
2339                         return -EINVAL;
2340                 vma->vm_pgoff = pfn;
2341         }
2342
2343         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2344         if (err)
2345                 return -EINVAL;
2346
2347         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2348
2349         BUG_ON(addr >= end);
2350         pfn -= addr >> PAGE_SHIFT;
2351         pgd = pgd_offset(mm, addr);
2352         flush_cache_range(vma, addr, end);
2353         do {
2354                 next = pgd_addr_end(addr, end);
2355                 err = remap_pud_range(mm, pgd, addr, next,
2356                                 pfn + (addr >> PAGE_SHIFT), prot);
2357                 if (err)
2358                         break;
2359         } while (pgd++, addr = next, addr != end);
2360
2361         if (err)
2362                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2363
2364         return err;
2365 }
2366 EXPORT_SYMBOL(remap_pfn_range);
2367
2368 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2369                                      unsigned long addr, unsigned long end,
2370                                      pte_fn_t fn, void *data)
2371 {
2372         pte_t *pte;
2373         int err;
2374         pgtable_t token;
2375         spinlock_t *uninitialized_var(ptl);
2376
2377         pte = (mm == &init_mm) ?
2378                 pte_alloc_kernel(pmd, addr) :
2379                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2380         if (!pte)
2381                 return -ENOMEM;
2382
2383         BUG_ON(pmd_huge(*pmd));
2384
2385         arch_enter_lazy_mmu_mode();
2386
2387         token = pmd_pgtable(*pmd);
2388
2389         do {
2390                 err = fn(pte++, token, addr, data);
2391                 if (err)
2392                         break;
2393         } while (addr += PAGE_SIZE, addr != end);
2394
2395         arch_leave_lazy_mmu_mode();
2396
2397         if (mm != &init_mm)
2398                 pte_unmap_unlock(pte-1, ptl);
2399         return err;
2400 }
2401
2402 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2403                                      unsigned long addr, unsigned long end,
2404                                      pte_fn_t fn, void *data)
2405 {
2406         pmd_t *pmd;
2407         unsigned long next;
2408         int err;
2409
2410         BUG_ON(pud_huge(*pud));
2411
2412         pmd = pmd_alloc(mm, pud, addr);
2413         if (!pmd)
2414                 return -ENOMEM;
2415         do {
2416                 next = pmd_addr_end(addr, end);
2417                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2418                 if (err)
2419                         break;
2420         } while (pmd++, addr = next, addr != end);
2421         return err;
2422 }
2423
2424 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2425                                      unsigned long addr, unsigned long end,
2426                                      pte_fn_t fn, void *data)
2427 {
2428         pud_t *pud;
2429         unsigned long next;
2430         int err;
2431
2432         pud = pud_alloc(mm, pgd, addr);
2433         if (!pud)
2434                 return -ENOMEM;
2435         do {
2436                 next = pud_addr_end(addr, end);
2437                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2438                 if (err)
2439                         break;
2440         } while (pud++, addr = next, addr != end);
2441         return err;
2442 }
2443
2444 /*
2445  * Scan a region of virtual memory, filling in page tables as necessary
2446  * and calling a provided function on each leaf page table.
2447  */
2448 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2449                         unsigned long size, pte_fn_t fn, void *data)
2450 {
2451         pgd_t *pgd;
2452         unsigned long next;
2453         unsigned long end = addr + size;
2454         int err;
2455
2456         BUG_ON(addr >= end);
2457         pgd = pgd_offset(mm, addr);
2458         do {
2459                 next = pgd_addr_end(addr, end);
2460                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2461                 if (err)
2462                         break;
2463         } while (pgd++, addr = next, addr != end);
2464
2465         return err;
2466 }
2467 EXPORT_SYMBOL_GPL(apply_to_page_range);
2468
2469 /*
2470  * handle_pte_fault chooses page fault handler according to an entry
2471  * which was read non-atomically.  Before making any commitment, on
2472  * those architectures or configurations (e.g. i386 with PAE) which
2473  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2474  * must check under lock before unmapping the pte and proceeding
2475  * (but do_wp_page is only called after already making such a check;
2476  * and do_anonymous_page can safely check later on).
2477  */
2478 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2479                                 pte_t *page_table, pte_t orig_pte)
2480 {
2481         int same = 1;
2482 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2483         if (sizeof(pte_t) > sizeof(unsigned long)) {
2484                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2485                 spin_lock(ptl);
2486                 same = pte_same(*page_table, orig_pte);
2487                 spin_unlock(ptl);
2488         }
2489 #endif
2490         pte_unmap(page_table);
2491         return same;
2492 }
2493
2494 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2495 {
2496         /*
2497          * If the source page was a PFN mapping, we don't have
2498          * a "struct page" for it. We do a best-effort copy by
2499          * just copying from the original user address. If that
2500          * fails, we just zero-fill it. Live with it.
2501          */
2502         if (unlikely(!src)) {
2503                 void *kaddr = kmap_atomic(dst);
2504                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2505
2506                 /*
2507                  * This really shouldn't fail, because the page is there
2508                  * in the page tables. But it might just be unreadable,
2509                  * in which case we just give up and fill the result with
2510                  * zeroes.
2511                  */
2512                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2513                         clear_page(kaddr);
2514                 kunmap_atomic(kaddr);
2515                 flush_dcache_page(dst);
2516         } else
2517                 copy_user_highpage(dst, src, va, vma);
2518 }
2519
2520 /*
2521  * This routine handles present pages, when users try to write
2522  * to a shared page. It is done by copying the page to a new address
2523  * and decrementing the shared-page counter for the old page.
2524  *
2525  * Note that this routine assumes that the protection checks have been
2526  * done by the caller (the low-level page fault routine in most cases).
2527  * Thus we can safely just mark it writable once we've done any necessary
2528  * COW.
2529  *
2530  * We also mark the page dirty at this point even though the page will
2531  * change only once the write actually happens. This avoids a few races,
2532  * and potentially makes it more efficient.
2533  *
2534  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2535  * but allow concurrent faults), with pte both mapped and locked.
2536  * We return with mmap_sem still held, but pte unmapped and unlocked.
2537  */
2538 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2539                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2540                 spinlock_t *ptl, pte_t orig_pte)
2541         __releases(ptl)
2542 {
2543         struct page *old_page, *new_page = NULL;
2544         pte_t entry;
2545         int ret = 0;
2546         int page_mkwrite = 0;
2547         struct page *dirty_page = NULL;
2548         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2549         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2550
2551         old_page = vm_normal_page(vma, address, orig_pte);
2552         if (!old_page) {
2553                 /*
2554                  * VM_MIXEDMAP !pfn_valid() case
2555                  *
2556                  * We should not cow pages in a shared writeable mapping.
2557                  * Just mark the pages writable as we can't do any dirty
2558                  * accounting on raw pfn maps.
2559                  */
2560                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2561                                      (VM_WRITE|VM_SHARED))
2562                         goto reuse;
2563                 goto gotten;
2564         }
2565
2566         /*
2567          * Take out anonymous pages first, anonymous shared vmas are
2568          * not dirty accountable.
2569          */
2570         if (PageAnon(old_page) && !PageKsm(old_page)) {
2571                 if (!trylock_page(old_page)) {
2572                         page_cache_get(old_page);
2573                         pte_unmap_unlock(page_table, ptl);
2574                         lock_page(old_page);
2575                         page_table = pte_offset_map_lock(mm, pmd, address,
2576                                                          &ptl);
2577                         if (!pte_same(*page_table, orig_pte)) {
2578                                 unlock_page(old_page);
2579                                 goto unlock;
2580                         }
2581                         page_cache_release(old_page);
2582                 }
2583                 if (reuse_swap_page(old_page)) {
2584                         /*
2585                          * The page is all ours.  Move it to our anon_vma so
2586                          * the rmap code will not search our parent or siblings.
2587                          * Protected against the rmap code by the page lock.
2588                          */
2589                         page_move_anon_rmap(old_page, vma, address);
2590                         unlock_page(old_page);
2591                         goto reuse;
2592                 }
2593                 unlock_page(old_page);
2594         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2595                                         (VM_WRITE|VM_SHARED))) {
2596                 /*
2597                  * Only catch write-faults on shared writable pages,
2598                  * read-only shared pages can get COWed by
2599                  * get_user_pages(.write=1, .force=1).
2600                  */
2601                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2602                         struct vm_fault vmf;
2603                         int tmp;
2604
2605                         vmf.virtual_address = (void __user *)(address &
2606                                                                 PAGE_MASK);
2607                         vmf.pgoff = old_page->index;
2608                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2609                         vmf.page = old_page;
2610
2611                         /*
2612                          * Notify the address space that the page is about to
2613                          * become writable so that it can prohibit this or wait
2614                          * for the page to get into an appropriate state.
2615                          *
2616                          * We do this without the lock held, so that it can
2617                          * sleep if it needs to.
2618                          */
2619                         page_cache_get(old_page);
2620                         pte_unmap_unlock(page_table, ptl);
2621
2622                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2623                         if (unlikely(tmp &
2624                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2625                                 ret = tmp;
2626                                 goto unwritable_page;
2627                         }
2628                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2629                                 lock_page(old_page);
2630                                 if (!old_page->mapping) {
2631                                         ret = 0; /* retry the fault */
2632                                         unlock_page(old_page);
2633                                         goto unwritable_page;
2634                                 }
2635                         } else
2636                                 VM_BUG_ON(!PageLocked(old_page));
2637
2638                         /*
2639                          * Since we dropped the lock we need to revalidate
2640                          * the PTE as someone else may have changed it.  If
2641                          * they did, we just return, as we can count on the
2642                          * MMU to tell us if they didn't also make it writable.
2643                          */
2644                         page_table = pte_offset_map_lock(mm, pmd, address,
2645                                                          &ptl);
2646                         if (!pte_same(*page_table, orig_pte)) {
2647                                 unlock_page(old_page);
2648                                 goto unlock;
2649                         }
2650
2651                         page_mkwrite = 1;
2652                 }
2653                 dirty_page = old_page;
2654                 get_page(dirty_page);
2655
2656 reuse:
2657                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2658                 entry = pte_mkyoung(orig_pte);
2659                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2660                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2661                         update_mmu_cache(vma, address, page_table);
2662                 pte_unmap_unlock(page_table, ptl);
2663                 ret |= VM_FAULT_WRITE;
2664
2665                 if (!dirty_page)
2666                         return ret;
2667
2668                 /*
2669                  * Yes, Virginia, this is actually required to prevent a race
2670                  * with clear_page_dirty_for_io() from clearing the page dirty
2671                  * bit after it clear all dirty ptes, but before a racing
2672                  * do_wp_page installs a dirty pte.
2673                  *
2674                  * __do_fault is protected similarly.
2675                  */
2676                 if (!page_mkwrite) {
2677                         wait_on_page_locked(dirty_page);
2678                         set_page_dirty_balance(dirty_page, page_mkwrite);
2679                         /* file_update_time outside page_lock */
2680                         if (vma->vm_file)
2681                                 file_update_time(vma->vm_file);
2682                 }
2683                 put_page(dirty_page);
2684                 if (page_mkwrite) {
2685                         struct address_space *mapping = dirty_page->mapping;
2686
2687                         set_page_dirty(dirty_page);
2688                         unlock_page(dirty_page);
2689                         page_cache_release(dirty_page);
2690                         if (mapping)    {
2691                                 /*
2692                                  * Some device drivers do not set page.mapping
2693                                  * but still dirty their pages
2694                                  */
2695                                 balance_dirty_pages_ratelimited(mapping);
2696                         }
2697                 }
2698
2699                 return ret;
2700         }
2701
2702         /*
2703          * Ok, we need to copy. Oh, well..
2704          */
2705         page_cache_get(old_page);
2706 gotten:
2707         pte_unmap_unlock(page_table, ptl);
2708
2709         if (unlikely(anon_vma_prepare(vma)))
2710                 goto oom;
2711
2712         if (is_zero_pfn(pte_pfn(orig_pte))) {
2713                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2714                 if (!new_page)
2715                         goto oom;
2716         } else {
2717                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2718                 if (!new_page)
2719                         goto oom;
2720                 cow_user_page(new_page, old_page, address, vma);
2721         }
2722         __SetPageUptodate(new_page);
2723
2724         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2725                 goto oom_free_new;
2726
2727         mmun_start  = address & PAGE_MASK;
2728         mmun_end    = mmun_start + PAGE_SIZE;
2729         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2730
2731         /*
2732          * Re-check the pte - we dropped the lock
2733          */
2734         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2735         if (likely(pte_same(*page_table, orig_pte))) {
2736                 if (old_page) {
2737                         if (!PageAnon(old_page)) {
2738                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2739                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2740                         }
2741                 } else
2742                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2743                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2744                 entry = mk_pte(new_page, vma->vm_page_prot);
2745                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2746                 /*
2747                  * Clear the pte entry and flush it first, before updating the
2748                  * pte with the new entry. This will avoid a race condition
2749                  * seen in the presence of one thread doing SMC and another
2750                  * thread doing COW.
2751                  */
2752                 ptep_clear_flush(vma, address, page_table);
2753                 page_add_new_anon_rmap(new_page, vma, address);
2754                 /*
2755                  * We call the notify macro here because, when using secondary
2756                  * mmu page tables (such as kvm shadow page tables), we want the
2757                  * new page to be mapped directly into the secondary page table.
2758                  */
2759                 set_pte_at_notify(mm, address, page_table, entry);
2760                 update_mmu_cache(vma, address, page_table);
2761                 if (old_page) {
2762                         /*
2763                          * Only after switching the pte to the new page may
2764                          * we remove the mapcount here. Otherwise another
2765                          * process may come and find the rmap count decremented
2766                          * before the pte is switched to the new page, and
2767                          * "reuse" the old page writing into it while our pte
2768                          * here still points into it and can be read by other
2769                          * threads.
2770                          *
2771                          * The critical issue is to order this
2772                          * page_remove_rmap with the ptp_clear_flush above.
2773                          * Those stores are ordered by (if nothing else,)
2774                          * the barrier present in the atomic_add_negative
2775                          * in page_remove_rmap.
2776                          *
2777                          * Then the TLB flush in ptep_clear_flush ensures that
2778                          * no process can access the old page before the
2779                          * decremented mapcount is visible. And the old page
2780                          * cannot be reused until after the decremented
2781                          * mapcount is visible. So transitively, TLBs to
2782                          * old page will be flushed before it can be reused.
2783                          */
2784                         page_remove_rmap(old_page);
2785                 }
2786
2787                 /* Free the old page.. */
2788                 new_page = old_page;
2789                 ret |= VM_FAULT_WRITE;
2790         } else
2791                 mem_cgroup_uncharge_page(new_page);
2792
2793         if (new_page)
2794                 page_cache_release(new_page);
2795 unlock:
2796         pte_unmap_unlock(page_table, ptl);
2797         if (mmun_end > mmun_start)
2798                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2799         if (old_page) {
2800                 /*
2801                  * Don't let another task, with possibly unlocked vma,
2802                  * keep the mlocked page.
2803                  */
2804                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2805                         lock_page(old_page);    /* LRU manipulation */
2806                         munlock_vma_page(old_page);
2807                         unlock_page(old_page);
2808                 }
2809                 page_cache_release(old_page);
2810         }
2811         return ret;
2812 oom_free_new:
2813         page_cache_release(new_page);
2814 oom:
2815         if (old_page) {
2816                 if (page_mkwrite) {
2817                         unlock_page(old_page);
2818                         page_cache_release(old_page);
2819                 }
2820                 page_cache_release(old_page);
2821         }
2822         return VM_FAULT_OOM;
2823
2824 unwritable_page:
2825         page_cache_release(old_page);
2826         return ret;
2827 }
2828
2829 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2830                 unsigned long start_addr, unsigned long end_addr,
2831                 struct zap_details *details)
2832 {
2833         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2834 }
2835
2836 static inline void unmap_mapping_range_tree(struct rb_root *root,
2837                                             struct zap_details *details)
2838 {
2839         struct vm_area_struct *vma;
2840         pgoff_t vba, vea, zba, zea;
2841
2842         vma_interval_tree_foreach(vma, root,
2843                         details->first_index, details->last_index) {
2844
2845                 vba = vma->vm_pgoff;
2846                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2847                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2848                 zba = details->first_index;
2849                 if (zba < vba)
2850                         zba = vba;
2851                 zea = details->last_index;
2852                 if (zea > vea)
2853                         zea = vea;
2854
2855                 unmap_mapping_range_vma(vma,
2856                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2857                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2858                                 details);
2859         }
2860 }
2861
2862 static inline void unmap_mapping_range_list(struct list_head *head,
2863                                             struct zap_details *details)
2864 {
2865         struct vm_area_struct *vma;
2866
2867         /*
2868          * In nonlinear VMAs there is no correspondence between virtual address
2869          * offset and file offset.  So we must perform an exhaustive search
2870          * across *all* the pages in each nonlinear VMA, not just the pages
2871          * whose virtual address lies outside the file truncation point.
2872          */
2873         list_for_each_entry(vma, head, shared.nonlinear) {
2874                 details->nonlinear_vma = vma;
2875                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2876         }
2877 }
2878
2879 /**
2880  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2881  * @mapping: the address space containing mmaps to be unmapped.
2882  * @holebegin: byte in first page to unmap, relative to the start of
2883  * the underlying file.  This will be rounded down to a PAGE_SIZE
2884  * boundary.  Note that this is different from truncate_pagecache(), which
2885  * must keep the partial page.  In contrast, we must get rid of
2886  * partial pages.
2887  * @holelen: size of prospective hole in bytes.  This will be rounded
2888  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2889  * end of the file.
2890  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2891  * but 0 when invalidating pagecache, don't throw away private data.
2892  */
2893 void unmap_mapping_range(struct address_space *mapping,
2894                 loff_t const holebegin, loff_t const holelen, int even_cows)
2895 {
2896         struct zap_details details;
2897         pgoff_t hba = holebegin >> PAGE_SHIFT;
2898         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2899
2900         /* Check for overflow. */
2901         if (sizeof(holelen) > sizeof(hlen)) {
2902                 long long holeend =
2903                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2904                 if (holeend & ~(long long)ULONG_MAX)
2905                         hlen = ULONG_MAX - hba + 1;
2906         }
2907
2908         details.check_mapping = even_cows? NULL: mapping;
2909         details.nonlinear_vma = NULL;
2910         details.first_index = hba;
2911         details.last_index = hba + hlen - 1;
2912         if (details.last_index < details.first_index)
2913                 details.last_index = ULONG_MAX;
2914
2915
2916         mutex_lock(&mapping->i_mmap_mutex);
2917         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2918                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2919         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2920                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2921         mutex_unlock(&mapping->i_mmap_mutex);
2922 }
2923 EXPORT_SYMBOL(unmap_mapping_range);
2924
2925 /*
2926  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2927  * but allow concurrent faults), and pte mapped but not yet locked.
2928  * We return with mmap_sem still held, but pte unmapped and unlocked.
2929  */
2930 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2931                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2932                 unsigned int flags, pte_t orig_pte)
2933 {
2934         spinlock_t *ptl;
2935         struct page *page, *swapcache = NULL;
2936         swp_entry_t entry;
2937         pte_t pte;
2938         int locked;
2939         struct mem_cgroup *ptr;
2940         int exclusive = 0;
2941         int ret = 0;
2942
2943         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2944                 goto out;
2945
2946         entry = pte_to_swp_entry(orig_pte);
2947         if (unlikely(non_swap_entry(entry))) {
2948                 if (is_migration_entry(entry)) {
2949                         migration_entry_wait(mm, pmd, address);
2950                 } else if (is_hwpoison_entry(entry)) {
2951                         ret = VM_FAULT_HWPOISON;
2952                 } else {
2953                         print_bad_pte(vma, address, orig_pte, NULL);
2954                         ret = VM_FAULT_SIGBUS;
2955                 }
2956                 goto out;
2957         }
2958         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2959         page = lookup_swap_cache(entry);
2960         if (!page) {
2961                 page = swapin_readahead(entry,
2962                                         GFP_HIGHUSER_MOVABLE, vma, address);
2963                 if (!page) {
2964                         /*
2965                          * Back out if somebody else faulted in this pte
2966                          * while we released the pte lock.
2967                          */
2968                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2969                         if (likely(pte_same(*page_table, orig_pte)))
2970                                 ret = VM_FAULT_OOM;
2971                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2972                         goto unlock;
2973                 }
2974
2975                 /* Had to read the page from swap area: Major fault */
2976                 ret = VM_FAULT_MAJOR;
2977                 count_vm_event(PGMAJFAULT);
2978                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2979         } else if (PageHWPoison(page)) {
2980                 /*
2981                  * hwpoisoned dirty swapcache pages are kept for killing
2982                  * owner processes (which may be unknown at hwpoison time)
2983                  */
2984                 ret = VM_FAULT_HWPOISON;
2985                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2986                 goto out_release;
2987         }
2988
2989         locked = lock_page_or_retry(page, mm, flags);
2990
2991         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2992         if (!locked) {
2993                 ret |= VM_FAULT_RETRY;
2994                 goto out_release;
2995         }
2996
2997         /*
2998          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2999          * release the swapcache from under us.  The page pin, and pte_same
3000          * test below, are not enough to exclude that.  Even if it is still
3001          * swapcache, we need to check that the page's swap has not changed.
3002          */
3003         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3004                 goto out_page;
3005
3006         if (ksm_might_need_to_copy(page, vma, address)) {
3007                 swapcache = page;
3008                 page = ksm_does_need_to_copy(page, vma, address);
3009
3010                 if (unlikely(!page)) {
3011                         ret = VM_FAULT_OOM;
3012                         page = swapcache;
3013                         swapcache = NULL;
3014                         goto out_page;
3015                 }
3016         }
3017
3018         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3019                 ret = VM_FAULT_OOM;
3020                 goto out_page;
3021         }
3022
3023         /*
3024          * Back out if somebody else already faulted in this pte.
3025          */
3026         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3027         if (unlikely(!pte_same(*page_table, orig_pte)))
3028                 goto out_nomap;
3029
3030         if (unlikely(!PageUptodate(page))) {
3031                 ret = VM_FAULT_SIGBUS;
3032                 goto out_nomap;
3033         }
3034
3035         /*
3036          * The page isn't present yet, go ahead with the fault.
3037          *
3038          * Be careful about the sequence of operations here.
3039          * To get its accounting right, reuse_swap_page() must be called
3040          * while the page is counted on swap but not yet in mapcount i.e.
3041          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3042          * must be called after the swap_free(), or it will never succeed.
3043          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3044          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3045          * in page->private. In this case, a record in swap_cgroup  is silently
3046          * discarded at swap_free().
3047          */
3048
3049         inc_mm_counter_fast(mm, MM_ANONPAGES);
3050         dec_mm_counter_fast(mm, MM_SWAPENTS);
3051         pte = mk_pte(page, vma->vm_page_prot);
3052         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3053                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3054                 flags &= ~FAULT_FLAG_WRITE;
3055                 ret |= VM_FAULT_WRITE;
3056                 exclusive = 1;
3057         }
3058         flush_icache_page(vma, page);
3059         set_pte_at(mm, address, page_table, pte);
3060         do_page_add_anon_rmap(page, vma, address, exclusive);
3061         /* It's better to call commit-charge after rmap is established */
3062         mem_cgroup_commit_charge_swapin(page, ptr);
3063
3064         swap_free(entry);
3065         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3066                 try_to_free_swap(page);
3067         unlock_page(page);
3068         if (swapcache) {
3069                 /*
3070                  * Hold the lock to avoid the swap entry to be reused
3071                  * until we take the PT lock for the pte_same() check
3072                  * (to avoid false positives from pte_same). For
3073                  * further safety release the lock after the swap_free
3074                  * so that the swap count won't change under a
3075                  * parallel locked swapcache.
3076                  */
3077                 unlock_page(swapcache);
3078                 page_cache_release(swapcache);
3079         }
3080
3081         if (flags & FAULT_FLAG_WRITE) {
3082                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3083                 if (ret & VM_FAULT_ERROR)
3084                         ret &= VM_FAULT_ERROR;
3085                 goto out;
3086         }
3087
3088         /* No need to invalidate - it was non-present before */
3089         update_mmu_cache(vma, address, page_table);
3090 unlock:
3091         pte_unmap_unlock(page_table, ptl);
3092 out:
3093         return ret;
3094 out_nomap:
3095         mem_cgroup_cancel_charge_swapin(ptr);
3096         pte_unmap_unlock(page_table, ptl);
3097 out_page:
3098         unlock_page(page);
3099 out_release:
3100         page_cache_release(page);
3101         if (swapcache) {
3102                 unlock_page(swapcache);
3103                 page_cache_release(swapcache);
3104         }
3105         return ret;
3106 }
3107
3108 /*
3109  * This is like a special single-page "expand_{down|up}wards()",
3110  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3111  * doesn't hit another vma.
3112  */
3113 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3114 {
3115         address &= PAGE_MASK;
3116         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3117                 struct vm_area_struct *prev = vma->vm_prev;
3118
3119                 /*
3120                  * Is there a mapping abutting this one below?
3121                  *
3122                  * That's only ok if it's the same stack mapping
3123                  * that has gotten split..
3124                  */
3125                 if (prev && prev->vm_end == address)
3126                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3127
3128                 expand_downwards(vma, address - PAGE_SIZE);
3129         }
3130         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3131                 struct vm_area_struct *next = vma->vm_next;
3132
3133                 /* As VM_GROWSDOWN but s/below/above/ */
3134                 if (next && next->vm_start == address + PAGE_SIZE)
3135                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3136
3137                 expand_upwards(vma, address + PAGE_SIZE);
3138         }
3139         return 0;
3140 }
3141
3142 /*
3143  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3144  * but allow concurrent faults), and pte mapped but not yet locked.
3145  * We return with mmap_sem still held, but pte unmapped and unlocked.
3146  */
3147 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3148                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3149                 unsigned int flags)
3150 {
3151         struct page *page;
3152         spinlock_t *ptl;
3153         pte_t entry;
3154
3155         pte_unmap(page_table);
3156
3157         /* Check if we need to add a guard page to the stack */
3158         if (check_stack_guard_page(vma, address) < 0)
3159                 return VM_FAULT_SIGBUS;
3160
3161         /* Use the zero-page for reads */
3162         if (!(flags & FAULT_FLAG_WRITE)) {
3163                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3164                                                 vma->vm_page_prot));
3165                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3166                 if (!pte_none(*page_table))
3167                         goto unlock;
3168                 goto setpte;
3169         }
3170
3171         /* Allocate our own private page. */
3172         if (unlikely(anon_vma_prepare(vma)))
3173                 goto oom;
3174         page = alloc_zeroed_user_highpage_movable(vma, address);
3175         if (!page)
3176                 goto oom;
3177         __SetPageUptodate(page);
3178
3179         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3180                 goto oom_free_page;
3181
3182         entry = mk_pte(page, vma->vm_page_prot);
3183         if (vma->vm_flags & VM_WRITE)
3184                 entry = pte_mkwrite(pte_mkdirty(entry));
3185
3186         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3187         if (!pte_none(*page_table))
3188                 goto release;
3189
3190         inc_mm_counter_fast(mm, MM_ANONPAGES);
3191         page_add_new_anon_rmap(page, vma, address);
3192 setpte:
3193         set_pte_at(mm, address, page_table, entry);
3194
3195         /* No need to invalidate - it was non-present before */
3196         update_mmu_cache(vma, address, page_table);
3197 unlock:
3198         pte_unmap_unlock(page_table, ptl);
3199         return 0;
3200 release:
3201         mem_cgroup_uncharge_page(page);
3202         page_cache_release(page);
3203         goto unlock;
3204 oom_free_page:
3205         page_cache_release(page);
3206 oom:
3207         return VM_FAULT_OOM;
3208 }
3209
3210 /*
3211  * __do_fault() tries to create a new page mapping. It aggressively
3212  * tries to share with existing pages, but makes a separate copy if
3213  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3214  * the next page fault.
3215  *
3216  * As this is called only for pages that do not currently exist, we
3217  * do not need to flush old virtual caches or the TLB.
3218  *
3219  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3220  * but allow concurrent faults), and pte neither mapped nor locked.
3221  * We return with mmap_sem still held, but pte unmapped and unlocked.
3222  */
3223 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3224                 unsigned long address, pmd_t *pmd,
3225                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3226 {
3227         pte_t *page_table;
3228         spinlock_t *ptl;
3229         struct page *page;
3230         struct page *cow_page;
3231         pte_t entry;
3232         int anon = 0;
3233         struct page *dirty_page = NULL;
3234         struct vm_fault vmf;
3235         int ret;
3236         int page_mkwrite = 0;
3237
3238         /*
3239          * If we do COW later, allocate page befor taking lock_page()
3240          * on the file cache page. This will reduce lock holding time.
3241          */
3242         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3243
3244                 if (unlikely(anon_vma_prepare(vma)))
3245                         return VM_FAULT_OOM;
3246
3247                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3248                 if (!cow_page)
3249                         return VM_FAULT_OOM;
3250
3251                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3252                         page_cache_release(cow_page);
3253                         return VM_FAULT_OOM;
3254                 }
3255         } else
3256                 cow_page = NULL;
3257
3258         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3259         vmf.pgoff = pgoff;
3260         vmf.flags = flags;
3261         vmf.page = NULL;
3262
3263         ret = vma->vm_ops->fault(vma, &vmf);
3264         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3265                             VM_FAULT_RETRY)))
3266                 goto uncharge_out;
3267
3268         if (unlikely(PageHWPoison(vmf.page))) {
3269                 if (ret & VM_FAULT_LOCKED)
3270                         unlock_page(vmf.page);
3271                 ret = VM_FAULT_HWPOISON;
3272                 goto uncharge_out;
3273         }
3274
3275         /*
3276          * For consistency in subsequent calls, make the faulted page always
3277          * locked.
3278          */
3279         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3280                 lock_page(vmf.page);
3281         else
3282                 VM_BUG_ON(!PageLocked(vmf.page));
3283
3284         /*
3285          * Should we do an early C-O-W break?
3286          */
3287         page = vmf.page;
3288         if (flags & FAULT_FLAG_WRITE) {
3289                 if (!(vma->vm_flags & VM_SHARED)) {
3290                         page = cow_page;
3291                         anon = 1;
3292                         copy_user_highpage(page, vmf.page, address, vma);
3293                         __SetPageUptodate(page);
3294                 } else {
3295                         /*
3296                          * If the page will be shareable, see if the backing
3297                          * address space wants to know that the page is about
3298                          * to become writable
3299                          */
3300                         if (vma->vm_ops->page_mkwrite) {
3301                                 int tmp;
3302
3303                                 unlock_page(page);
3304                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3305                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3306                                 if (unlikely(tmp &
3307                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3308                                         ret = tmp;
3309                                         goto unwritable_page;
3310                                 }
3311                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3312                                         lock_page(page);
3313                                         if (!page->mapping) {
3314                                                 ret = 0; /* retry the fault */
3315                                                 unlock_page(page);
3316                                                 goto unwritable_page;
3317                                         }
3318                                 } else
3319                                         VM_BUG_ON(!PageLocked(page));
3320                                 page_mkwrite = 1;
3321                         }
3322                 }
3323
3324         }
3325
3326         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3327
3328         /*
3329          * This silly early PAGE_DIRTY setting removes a race
3330          * due to the bad i386 page protection. But it's valid
3331          * for other architectures too.
3332          *
3333          * Note that if FAULT_FLAG_WRITE is set, we either now have
3334          * an exclusive copy of the page, or this is a shared mapping,
3335          * so we can make it writable and dirty to avoid having to
3336          * handle that later.
3337          */
3338         /* Only go through if we didn't race with anybody else... */
3339         if (likely(pte_same(*page_table, orig_pte))) {
3340                 flush_icache_page(vma, page);
3341                 entry = mk_pte(page, vma->vm_page_prot);
3342                 if (flags & FAULT_FLAG_WRITE)
3343                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3344                 if (anon) {
3345                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3346                         page_add_new_anon_rmap(page, vma, address);
3347                 } else {
3348                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3349                         page_add_file_rmap(page);
3350                         if (flags & FAULT_FLAG_WRITE) {
3351                                 dirty_page = page;
3352                                 get_page(dirty_page);
3353                         }
3354                 }
3355                 set_pte_at(mm, address, page_table, entry);
3356
3357                 /* no need to invalidate: a not-present page won't be cached */
3358                 update_mmu_cache(vma, address, page_table);
3359         } else {
3360                 if (cow_page)
3361                         mem_cgroup_uncharge_page(cow_page);
3362                 if (anon)
3363                         page_cache_release(page);
3364                 else
3365                         anon = 1; /* no anon but release faulted_page */
3366         }
3367
3368         pte_unmap_unlock(page_table, ptl);
3369
3370         if (dirty_page) {
3371                 struct address_space *mapping = page->mapping;
3372                 int dirtied = 0;
3373
3374                 if (set_page_dirty(dirty_page))
3375                         dirtied = 1;
3376                 unlock_page(dirty_page);
3377                 put_page(dirty_page);
3378                 if ((dirtied || page_mkwrite) && mapping) {
3379                         /*
3380                          * Some device drivers do not set page.mapping but still
3381                          * dirty their pages
3382                          */
3383                         balance_dirty_pages_ratelimited(mapping);
3384                 }
3385
3386                 /* file_update_time outside page_lock */
3387                 if (vma->vm_file && !page_mkwrite)
3388                         file_update_time(vma->vm_file);
3389         } else {
3390                 unlock_page(vmf.page);
3391                 if (anon)
3392                         page_cache_release(vmf.page);
3393         }
3394
3395         return ret;
3396
3397 unwritable_page:
3398         page_cache_release(page);
3399         return ret;
3400 uncharge_out:
3401         /* fs's fault handler get error */
3402         if (cow_page) {
3403                 mem_cgroup_uncharge_page(cow_page);
3404                 page_cache_release(cow_page);
3405         }
3406         return ret;
3407 }
3408
3409 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3410                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3411                 unsigned int flags, pte_t orig_pte)
3412 {
3413         pgoff_t pgoff = (((address & PAGE_MASK)
3414                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3415
3416         pte_unmap(page_table);
3417         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3418 }
3419
3420 /*
3421  * Fault of a previously existing named mapping. Repopulate the pte
3422  * from the encoded file_pte if possible. This enables swappable
3423  * nonlinear vmas.
3424  *
3425  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3426  * but allow concurrent faults), and pte mapped but not yet locked.
3427  * We return with mmap_sem still held, but pte unmapped and unlocked.
3428  */
3429 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3430                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3431                 unsigned int flags, pte_t orig_pte)
3432 {
3433         pgoff_t pgoff;
3434
3435         flags |= FAULT_FLAG_NONLINEAR;
3436
3437         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3438                 return 0;
3439
3440         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3441                 /*
3442                  * Page table corrupted: show pte and kill process.
3443                  */
3444                 print_bad_pte(vma, address, orig_pte, NULL);
3445                 return VM_FAULT_SIGBUS;
3446         }
3447
3448         pgoff = pte_to_pgoff(orig_pte);
3449         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3450 }
3451
3452 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3453                                 unsigned long addr, int current_nid)
3454 {
3455         get_page(page);
3456
3457         count_vm_numa_event(NUMA_HINT_FAULTS);
3458         if (current_nid == numa_node_id())
3459                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3460
3461         return mpol_misplaced(page, vma, addr);
3462 }
3463
3464 int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3465                    unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3466 {
3467         struct page *page = NULL;
3468         spinlock_t *ptl;
3469         int current_nid = -1;
3470         int target_nid;
3471         bool migrated = false;
3472
3473         /*
3474         * The "pte" at this point cannot be used safely without
3475         * validation through pte_unmap_same(). It's of NUMA type but
3476         * the pfn may be screwed if the read is non atomic.
3477         *
3478         * ptep_modify_prot_start is not called as this is clearing
3479         * the _PAGE_NUMA bit and it is not really expected that there
3480         * would be concurrent hardware modifications to the PTE.
3481         */
3482         ptl = pte_lockptr(mm, pmd);
3483         spin_lock(ptl);
3484         if (unlikely(!pte_same(*ptep, pte))) {
3485                 pte_unmap_unlock(ptep, ptl);
3486                 goto out;
3487         }
3488
3489         pte = pte_mknonnuma(pte);
3490         set_pte_at(mm, addr, ptep, pte);
3491         update_mmu_cache(vma, addr, ptep);
3492
3493         page = vm_normal_page(vma, addr, pte);
3494         if (!page) {
3495                 pte_unmap_unlock(ptep, ptl);
3496                 return 0;
3497         }
3498
3499         current_nid = page_to_nid(page);
3500         target_nid = numa_migrate_prep(page, vma, addr, current_nid);
3501         pte_unmap_unlock(ptep, ptl);
3502         if (target_nid == -1) {
3503                 /*
3504                  * Account for the fault against the current node if it not
3505                  * being replaced regardless of where the page is located.
3506                  */
3507                 current_nid = numa_node_id();
3508                 put_page(page);
3509                 goto out;
3510         }
3511
3512         /* Migrate to the requested node */
3513         migrated = migrate_misplaced_page(page, target_nid);
3514         if (migrated)
3515                 current_nid = target_nid;
3516
3517 out:
3518         if (current_nid != -1)
3519                 task_numa_fault(current_nid, 1, migrated);
3520         return 0;
3521 }
3522
3523 /* NUMA hinting page fault entry point for regular pmds */
3524 #ifdef CONFIG_NUMA_BALANCING
3525 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3526                      unsigned long addr, pmd_t *pmdp)
3527 {
3528         pmd_t pmd;
3529         pte_t *pte, *orig_pte;
3530         unsigned long _addr = addr & PMD_MASK;
3531         unsigned long offset;
3532         spinlock_t *ptl;
3533         bool numa = false;
3534         int local_nid = numa_node_id();
3535
3536         spin_lock(&mm->page_table_lock);
3537         pmd = *pmdp;
3538         if (pmd_numa(pmd)) {
3539                 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3540                 numa = true;
3541         }
3542         spin_unlock(&mm->page_table_lock);
3543
3544         if (!numa)
3545                 return 0;
3546
3547         /* we're in a page fault so some vma must be in the range */
3548         BUG_ON(!vma);
3549         BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3550         offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3551         VM_BUG_ON(offset >= PMD_SIZE);
3552         orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3553         pte += offset >> PAGE_SHIFT;
3554         for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3555                 pte_t pteval = *pte;
3556                 struct page *page;
3557                 int curr_nid = local_nid;
3558                 int target_nid;
3559                 bool migrated;
3560                 if (!pte_present(pteval))
3561                         continue;
3562                 if (!pte_numa(pteval))
3563                         continue;
3564                 if (addr >= vma->vm_end) {
3565                         vma = find_vma(mm, addr);
3566                         /* there's a pte present so there must be a vma */
3567                         BUG_ON(!vma);
3568                         BUG_ON(addr < vma->vm_start);
3569                 }
3570                 if (pte_numa(pteval)) {
3571                         pteval = pte_mknonnuma(pteval);
3572                         set_pte_at(mm, addr, pte, pteval);
3573                 }
3574                 page = vm_normal_page(vma, addr, pteval);
3575                 if (unlikely(!page))
3576                         continue;
3577                 /* only check non-shared pages */
3578                 if (unlikely(page_mapcount(page) != 1))
3579                         continue;
3580
3581                 /*
3582                  * Note that the NUMA fault is later accounted to either
3583                  * the node that is currently running or where the page is
3584                  * migrated to.
3585                  */
3586                 curr_nid = local_nid;
3587                 target_nid = numa_migrate_prep(page, vma, addr,
3588                                                page_to_nid(page));
3589                 if (target_nid == -1) {
3590                         put_page(page);
3591                         continue;
3592                 }
3593
3594                 /* Migrate to the requested node */
3595                 pte_unmap_unlock(pte, ptl);
3596                 migrated = migrate_misplaced_page(page, target_nid);
3597                 if (migrated)
3598                         curr_nid = target_nid;
3599                 task_numa_fault(curr_nid, 1, migrated);
3600
3601                 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
3602         }
3603         pte_unmap_unlock(orig_pte, ptl);
3604
3605         return 0;
3606 }
3607 #else
3608 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3609                      unsigned long addr, pmd_t *pmdp)
3610 {
3611         BUG();
3612 }
3613 #endif /* CONFIG_NUMA_BALANCING */
3614
3615 /*
3616  * These routines also need to handle stuff like marking pages dirty
3617  * and/or accessed for architectures that don't do it in hardware (most
3618  * RISC architectures).  The early dirtying is also good on the i386.
3619  *
3620  * There is also a hook called "update_mmu_cache()" that architectures
3621  * with external mmu caches can use to update those (ie the Sparc or
3622  * PowerPC hashed page tables that act as extended TLBs).
3623  *
3624  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3625  * but allow concurrent faults), and pte mapped but not yet locked.
3626  * We return with mmap_sem still held, but pte unmapped and unlocked.
3627  */
3628 int handle_pte_fault(struct mm_struct *mm,
3629                      struct vm_area_struct *vma, unsigned long address,
3630                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3631 {
3632         pte_t entry;
3633         spinlock_t *ptl;
3634
3635         entry = *pte;
3636         if (!pte_present(entry)) {
3637                 if (pte_none(entry)) {
3638                         if (vma->vm_ops) {
3639                                 if (likely(vma->vm_ops->fault))
3640                                         return do_linear_fault(mm, vma, address,
3641                                                 pte, pmd, flags, entry);
3642                         }
3643                         return do_anonymous_page(mm, vma, address,
3644                                                  pte, pmd, flags);
3645                 }
3646                 if (pte_file(entry))
3647                         return do_nonlinear_fault(mm, vma, address,
3648                                         pte, pmd, flags, entry);
3649                 return do_swap_page(mm, vma, address,
3650                                         pte, pmd, flags, entry);
3651         }
3652
3653         if (pte_numa(entry))
3654                 return do_numa_page(mm, vma, address, entry, pte, pmd);
3655
3656         ptl = pte_lockptr(mm, pmd);
3657         spin_lock(ptl);
3658         if (unlikely(!pte_same(*pte, entry)))
3659                 goto unlock;
3660         if (flags & FAULT_FLAG_WRITE) {
3661                 if (!pte_write(entry))
3662                         return do_wp_page(mm, vma, address,
3663                                         pte, pmd, ptl, entry);
3664                 entry = pte_mkdirty(entry);
3665         }
3666         entry = pte_mkyoung(entry);
3667         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3668                 update_mmu_cache(vma, address, pte);
3669         } else {
3670                 /*
3671                  * This is needed only for protection faults but the arch code
3672                  * is not yet telling us if this is a protection fault or not.
3673                  * This still avoids useless tlb flushes for .text page faults
3674                  * with threads.
3675                  */
3676                 if (flags & FAULT_FLAG_WRITE)
3677                         flush_tlb_fix_spurious_fault(vma, address);
3678         }
3679 unlock:
3680         pte_unmap_unlock(pte, ptl);
3681         return 0;
3682 }
3683
3684 /*
3685  * By the time we get here, we already hold the mm semaphore
3686  */
3687 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3688                 unsigned long address, unsigned int flags)
3689 {
3690         pgd_t *pgd;
3691         pud_t *pud;
3692         pmd_t *pmd;
3693         pte_t *pte;
3694
3695         __set_current_state(TASK_RUNNING);
3696
3697         count_vm_event(PGFAULT);
3698         mem_cgroup_count_vm_event(mm, PGFAULT);
3699
3700         /* do counter updates before entering really critical section. */
3701         check_sync_rss_stat(current);
3702
3703         if (unlikely(is_vm_hugetlb_page(vma)))
3704                 return hugetlb_fault(mm, vma, address, flags);
3705
3706 retry:
3707         pgd = pgd_offset(mm, address);
3708         pud = pud_alloc(mm, pgd, address);
3709         if (!pud)
3710                 return VM_FAULT_OOM;
3711         pmd = pmd_alloc(mm, pud, address);
3712         if (!pmd)
3713                 return VM_FAULT_OOM;
3714         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3715                 if (!vma->vm_ops)
3716                         return do_huge_pmd_anonymous_page(mm, vma, address,
3717                                                           pmd, flags);
3718         } else {
3719                 pmd_t orig_pmd = *pmd;
3720                 int ret;
3721
3722                 barrier();
3723                 if (pmd_trans_huge(orig_pmd)) {
3724                         if (pmd_numa(*pmd))
3725                                 return do_huge_pmd_numa_page(mm, vma, address,
3726                                                              orig_pmd, pmd);
3727
3728                         if ((flags & FAULT_FLAG_WRITE) && !pmd_write(orig_pmd)) {
3729                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3730                                                           orig_pmd);
3731                                 /*
3732                                  * If COW results in an oom, the huge pmd will
3733                                  * have been split, so retry the fault on the
3734                                  * pte for a smaller charge.
3735                                  */
3736                                 if (unlikely(ret & VM_FAULT_OOM))
3737                                         goto retry;
3738                                 return ret;
3739                         }
3740
3741                         return 0;
3742                 }
3743         }
3744
3745         if (pmd_numa(*pmd))
3746                 return do_pmd_numa_page(mm, vma, address, pmd);
3747
3748         /*
3749          * Use __pte_alloc instead of pte_alloc_map, because we can't
3750          * run pte_offset_map on the pmd, if an huge pmd could
3751          * materialize from under us from a different thread.
3752          */
3753         if (unlikely(pmd_none(*pmd)) &&
3754             unlikely(__pte_alloc(mm, vma, pmd, address)))
3755                 return VM_FAULT_OOM;
3756         /* if an huge pmd materialized from under us just retry later */
3757         if (unlikely(pmd_trans_huge(*pmd)))
3758                 return 0;
3759         /*
3760          * A regular pmd is established and it can't morph into a huge pmd
3761          * from under us anymore at this point because we hold the mmap_sem
3762          * read mode and khugepaged takes it in write mode. So now it's
3763          * safe to run pte_offset_map().
3764          */
3765         pte = pte_offset_map(pmd, address);
3766
3767         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3768 }
3769
3770 #ifndef __PAGETABLE_PUD_FOLDED
3771 /*
3772  * Allocate page upper directory.
3773  * We've already handled the fast-path in-line.
3774  */
3775 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3776 {
3777         pud_t *new = pud_alloc_one(mm, address);
3778         if (!new)
3779                 return -ENOMEM;
3780
3781         smp_wmb(); /* See comment in __pte_alloc */
3782
3783         spin_lock(&mm->page_table_lock);
3784         if (pgd_present(*pgd))          /* Another has populated it */
3785                 pud_free(mm, new);
3786         else
3787                 pgd_populate(mm, pgd, new);
3788         spin_unlock(&mm->page_table_lock);
3789         return 0;
3790 }
3791 #endif /* __PAGETABLE_PUD_FOLDED */
3792
3793 #ifndef __PAGETABLE_PMD_FOLDED
3794 /*
3795  * Allocate page middle directory.
3796  * We've already handled the fast-path in-line.
3797  */
3798 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3799 {
3800         pmd_t *new = pmd_alloc_one(mm, address);
3801         if (!new)
3802                 return -ENOMEM;
3803
3804         smp_wmb(); /* See comment in __pte_alloc */
3805
3806         spin_lock(&mm->page_table_lock);
3807 #ifndef __ARCH_HAS_4LEVEL_HACK
3808         if (pud_present(*pud))          /* Another has populated it */
3809                 pmd_free(mm, new);
3810         else
3811                 pud_populate(mm, pud, new);
3812 #else
3813         if (pgd_present(*pud))          /* Another has populated it */
3814                 pmd_free(mm, new);
3815         else
3816                 pgd_populate(mm, pud, new);
3817 #endif /* __ARCH_HAS_4LEVEL_HACK */
3818         spin_unlock(&mm->page_table_lock);
3819         return 0;
3820 }
3821 #endif /* __PAGETABLE_PMD_FOLDED */
3822
3823 int make_pages_present(unsigned long addr, unsigned long end)
3824 {
3825         int ret, len, write;
3826         struct vm_area_struct * vma;
3827
3828         vma = find_vma(current->mm, addr);
3829         if (!vma)
3830                 return -ENOMEM;
3831         /*
3832          * We want to touch writable mappings with a write fault in order
3833          * to break COW, except for shared mappings because these don't COW
3834          * and we would not want to dirty them for nothing.
3835          */
3836         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3837         BUG_ON(addr >= end);
3838         BUG_ON(end > vma->vm_end);
3839         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3840         ret = get_user_pages(current, current->mm, addr,
3841                         len, write, 0, NULL, NULL);
3842         if (ret < 0)
3843                 return ret;
3844         return ret == len ? 0 : -EFAULT;
3845 }
3846
3847 #if !defined(__HAVE_ARCH_GATE_AREA)
3848
3849 #if defined(AT_SYSINFO_EHDR)
3850 static struct vm_area_struct gate_vma;
3851
3852 static int __init gate_vma_init(void)
3853 {
3854         gate_vma.vm_mm = NULL;
3855         gate_vma.vm_start = FIXADDR_USER_START;
3856         gate_vma.vm_end = FIXADDR_USER_END;
3857         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3858         gate_vma.vm_page_prot = __P101;
3859
3860         return 0;
3861 }
3862 __initcall(gate_vma_init);
3863 #endif
3864
3865 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3866 {
3867 #ifdef AT_SYSINFO_EHDR
3868         return &gate_vma;
3869 #else
3870         return NULL;
3871 #endif
3872 }
3873
3874 int in_gate_area_no_mm(unsigned long addr)
3875 {
3876 #ifdef AT_SYSINFO_EHDR
3877         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3878                 return 1;
3879 #endif
3880         return 0;
3881 }
3882
3883 #endif  /* __HAVE_ARCH_GATE_AREA */
3884
3885 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3886                 pte_t **ptepp, spinlock_t **ptlp)
3887 {
3888         pgd_t *pgd;
3889         pud_t *pud;
3890         pmd_t *pmd;
3891         pte_t *ptep;
3892
3893         pgd = pgd_offset(mm, address);
3894         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3895                 goto out;
3896
3897         pud = pud_offset(pgd, address);
3898         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3899                 goto out;
3900
3901         pmd = pmd_offset(pud, address);
3902         VM_BUG_ON(pmd_trans_huge(*pmd));
3903         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3904                 goto out;
3905
3906         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3907         if (pmd_huge(*pmd))
3908                 goto out;
3909
3910         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3911         if (!ptep)
3912                 goto out;
3913         if (!pte_present(*ptep))
3914                 goto unlock;
3915         *ptepp = ptep;
3916         return 0;
3917 unlock:
3918         pte_unmap_unlock(ptep, *ptlp);
3919 out:
3920         return -EINVAL;
3921 }
3922
3923 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3924                              pte_t **ptepp, spinlock_t **ptlp)
3925 {
3926         int res;
3927
3928         /* (void) is needed to make gcc happy */
3929         (void) __cond_lock(*ptlp,
3930                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3931         return res;
3932 }
3933
3934 /**
3935  * follow_pfn - look up PFN at a user virtual address
3936  * @vma: memory mapping
3937  * @address: user virtual address
3938  * @pfn: location to store found PFN
3939  *
3940  * Only IO mappings and raw PFN mappings are allowed.
3941  *
3942  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3943  */
3944 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3945         unsigned long *pfn)
3946 {
3947         int ret = -EINVAL;
3948         spinlock_t *ptl;
3949         pte_t *ptep;
3950
3951         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3952                 return ret;
3953
3954         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3955         if (ret)
3956                 return ret;
3957         *pfn = pte_pfn(*ptep);
3958         pte_unmap_unlock(ptep, ptl);
3959         return 0;
3960 }
3961 EXPORT_SYMBOL(follow_pfn);
3962
3963 #ifdef CONFIG_HAVE_IOREMAP_PROT
3964 int follow_phys(struct vm_area_struct *vma,
3965                 unsigned long address, unsigned int flags,
3966                 unsigned long *prot, resource_size_t *phys)
3967 {
3968         int ret = -EINVAL;
3969         pte_t *ptep, pte;
3970         spinlock_t *ptl;
3971
3972         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3973                 goto out;
3974
3975         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3976                 goto out;
3977         pte = *ptep;
3978
3979         if ((flags & FOLL_WRITE) && !pte_write(pte))
3980                 goto unlock;
3981
3982         *prot = pgprot_val(pte_pgprot(pte));
3983         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3984
3985         ret = 0;
3986 unlock:
3987         pte_unmap_unlock(ptep, ptl);
3988 out:
3989         return ret;
3990 }
3991
3992 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3993                         void *buf, int len, int write)
3994 {
3995         resource_size_t phys_addr;
3996         unsigned long prot = 0;
3997         void __iomem *maddr;
3998         int offset = addr & (PAGE_SIZE-1);
3999
4000         if (follow_phys(vma, addr, write, &prot, &phys_addr))
4001                 return -EINVAL;
4002
4003         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4004         if (write)
4005                 memcpy_toio(maddr + offset, buf, len);
4006         else
4007                 memcpy_fromio(buf, maddr + offset, len);
4008         iounmap(maddr);
4009
4010         return len;
4011 }
4012 #endif
4013
4014 /*
4015  * Access another process' address space as given in mm.  If non-NULL, use the
4016  * given task for page fault accounting.
4017  */
4018 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4019                 unsigned long addr, void *buf, int len, int write)
4020 {
4021         struct vm_area_struct *vma;
4022         void *old_buf = buf;
4023
4024         down_read(&mm->mmap_sem);
4025         /* ignore errors, just check how much was successfully transferred */
4026         while (len) {
4027                 int bytes, ret, offset;
4028                 void *maddr;
4029                 struct page *page = NULL;
4030
4031                 ret = get_user_pages(tsk, mm, addr, 1,
4032                                 write, 1, &page, &vma);
4033                 if (ret <= 0) {
4034                         /*
4035                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
4036                          * we can access using slightly different code.
4037                          */
4038 #ifdef CONFIG_HAVE_IOREMAP_PROT
4039                         vma = find_vma(mm, addr);
4040                         if (!vma || vma->vm_start > addr)
4041                                 break;
4042                         if (vma->vm_ops && vma->vm_ops->access)
4043                                 ret = vma->vm_ops->access(vma, addr, buf,
4044                                                           len, write);
4045                         if (ret <= 0)
4046 #endif
4047                                 break;
4048                         bytes = ret;
4049                 } else {
4050                         bytes = len;
4051                         offset = addr & (PAGE_SIZE-1);
4052                         if (bytes > PAGE_SIZE-offset)
4053                                 bytes = PAGE_SIZE-offset;
4054
4055                         maddr = kmap(page);
4056                         if (write) {
4057                                 copy_to_user_page(vma, page, addr,
4058                                                   maddr + offset, buf, bytes);
4059                                 set_page_dirty_lock(page);
4060                         } else {
4061                                 copy_from_user_page(vma, page, addr,
4062                                                     buf, maddr + offset, bytes);
4063                         }
4064                         kunmap(page);
4065                         page_cache_release(page);
4066                 }
4067                 len -= bytes;
4068                 buf += bytes;
4069                 addr += bytes;
4070         }
4071         up_read(&mm->mmap_sem);
4072
4073         return buf - old_buf;
4074 }
4075
4076 /**
4077  * access_remote_vm - access another process' address space
4078  * @mm:         the mm_struct of the target address space
4079  * @addr:       start address to access
4080  * @buf:        source or destination buffer
4081  * @len:        number of bytes to transfer
4082  * @write:      whether the access is a write
4083  *
4084  * The caller must hold a reference on @mm.
4085  */
4086 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4087                 void *buf, int len, int write)
4088 {
4089         return __access_remote_vm(NULL, mm, addr, buf, len, write);
4090 }
4091
4092 /*
4093  * Access another process' address space.
4094  * Source/target buffer must be kernel space,
4095  * Do not walk the page table directly, use get_user_pages
4096  */
4097 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4098                 void *buf, int len, int write)
4099 {
4100         struct mm_struct *mm;
4101         int ret;
4102
4103         mm = get_task_mm(tsk);
4104         if (!mm)
4105                 return 0;
4106
4107         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4108         mmput(mm);
4109
4110         return ret;
4111 }
4112
4113 /*
4114  * Print the name of a VMA.
4115  */
4116 void print_vma_addr(char *prefix, unsigned long ip)
4117 {
4118         struct mm_struct *mm = current->mm;
4119         struct vm_area_struct *vma;
4120
4121         /*
4122          * Do not print if we are in atomic
4123          * contexts (in exception stacks, etc.):
4124          */
4125         if (preempt_count())
4126                 return;
4127
4128         down_read(&mm->mmap_sem);
4129         vma = find_vma(mm, ip);
4130         if (vma && vma->vm_file) {
4131                 struct file *f = vma->vm_file;
4132                 char *buf = (char *)__get_free_page(GFP_KERNEL);
4133                 if (buf) {
4134                         char *p, *s;
4135
4136                         p = d_path(&f->f_path, buf, PAGE_SIZE);
4137                         if (IS_ERR(p))
4138                                 p = "?";
4139                         s = strrchr(p, '/');
4140                         if (s)
4141                                 p = s+1;
4142                         printk("%s%s[%lx+%lx]", prefix, p,
4143                                         vma->vm_start,
4144                                         vma->vm_end - vma->vm_start);
4145                         free_page((unsigned long)buf);
4146                 }
4147         }
4148         up_read(&mm->mmap_sem);
4149 }
4150
4151 #ifdef CONFIG_PROVE_LOCKING
4152 void might_fault(void)
4153 {
4154         /*
4155          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4156          * holding the mmap_sem, this is safe because kernel memory doesn't
4157          * get paged out, therefore we'll never actually fault, and the
4158          * below annotations will generate false positives.
4159          */
4160         if (segment_eq(get_fs(), KERNEL_DS))
4161                 return;
4162
4163         might_sleep();
4164         /*
4165          * it would be nicer only to annotate paths which are not under
4166          * pagefault_disable, however that requires a larger audit and
4167          * providing helpers like get_user_atomic.
4168          */
4169         if (!in_atomic() && current->mm)
4170                 might_lock_read(&current->mm->mmap_sem);
4171 }
4172 EXPORT_SYMBOL(might_fault);
4173 #endif
4174
4175 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4176 static void clear_gigantic_page(struct page *page,
4177                                 unsigned long addr,
4178                                 unsigned int pages_per_huge_page)
4179 {
4180         int i;
4181         struct page *p = page;
4182
4183         might_sleep();
4184         for (i = 0; i < pages_per_huge_page;
4185              i++, p = mem_map_next(p, page, i)) {
4186                 cond_resched();
4187                 clear_user_highpage(p, addr + i * PAGE_SIZE);
4188         }
4189 }
4190 void clear_huge_page(struct page *page,
4191                      unsigned long addr, unsigned int pages_per_huge_page)
4192 {
4193         int i;
4194
4195         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4196                 clear_gigantic_page(page, addr, pages_per_huge_page);
4197                 return;
4198         }
4199
4200         might_sleep();
4201         for (i = 0; i < pages_per_huge_page; i++) {
4202                 cond_resched();
4203                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4204         }
4205 }
4206
4207 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4208                                     unsigned long addr,
4209                                     struct vm_area_struct *vma,
4210                                     unsigned int pages_per_huge_page)
4211 {
4212         int i;
4213         struct page *dst_base = dst;
4214         struct page *src_base = src;
4215
4216         for (i = 0; i < pages_per_huge_page; ) {
4217                 cond_resched();
4218                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4219
4220                 i++;
4221                 dst = mem_map_next(dst, dst_base, i);
4222                 src = mem_map_next(src, src_base, i);
4223         }
4224 }
4225
4226 void copy_user_huge_page(struct page *dst, struct page *src,
4227                          unsigned long addr, struct vm_area_struct *vma,
4228                          unsigned int pages_per_huge_page)
4229 {
4230         int i;
4231
4232         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4233                 copy_user_gigantic_page(dst, src, addr, vma,
4234                                         pages_per_huge_page);
4235                 return;
4236         }
4237
4238         might_sleep();
4239         for (i = 0; i < pages_per_huge_page; i++) {
4240                 cond_resched();
4241                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4242         }
4243 }
4244 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */