mm: vmscan: fix misused nr_reclaimed in shrink_mem_cgroup_zone()
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60
61 #include <asm/io.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
64 #include <asm/tlb.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
67
68 #include "internal.h"
69
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
73 struct page *mem_map;
74
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
77 #endif
78
79 unsigned long num_physpages;
80 /*
81  * A number of key systems in x86 including ioremap() rely on the assumption
82  * that high_memory defines the upper bound on direct map memory, then end
83  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
84  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85  * and ZONE_HIGHMEM.
86  */
87 void * high_memory;
88
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
91
92 /*
93  * Randomize the address space (stacks, mmaps, brk, etc.).
94  *
95  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96  *   as ancient (libc5 based) binaries can segfault. )
97  */
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
100                                         1;
101 #else
102                                         2;
103 #endif
104
105 static int __init disable_randmaps(char *s)
106 {
107         randomize_va_space = 0;
108         return 1;
109 }
110 __setup("norandmaps", disable_randmaps);
111
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
114
115 /*
116  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117  */
118 static int __init init_zero_pfn(void)
119 {
120         zero_pfn = page_to_pfn(ZERO_PAGE(0));
121         return 0;
122 }
123 core_initcall(init_zero_pfn);
124
125
126 #if defined(SPLIT_RSS_COUNTING)
127
128 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
129 {
130         int i;
131
132         for (i = 0; i < NR_MM_COUNTERS; i++) {
133                 if (task->rss_stat.count[i]) {
134                         add_mm_counter(mm, i, task->rss_stat.count[i]);
135                         task->rss_stat.count[i] = 0;
136                 }
137         }
138         task->rss_stat.events = 0;
139 }
140
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142 {
143         struct task_struct *task = current;
144
145         if (likely(task->mm == mm))
146                 task->rss_stat.count[member] += val;
147         else
148                 add_mm_counter(mm, member, val);
149 }
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH  (64)
155 static void check_sync_rss_stat(struct task_struct *task)
156 {
157         if (unlikely(task != current))
158                 return;
159         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160                 __sync_task_rss_stat(task, task->mm);
161 }
162
163 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
164 {
165         __sync_task_rss_stat(task, mm);
166 }
167 #else /* SPLIT_RSS_COUNTING */
168
169 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
170 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
171
172 static void check_sync_rss_stat(struct task_struct *task)
173 {
174 }
175
176 #endif /* SPLIT_RSS_COUNTING */
177
178 #ifdef HAVE_GENERIC_MMU_GATHER
179
180 static int tlb_next_batch(struct mmu_gather *tlb)
181 {
182         struct mmu_gather_batch *batch;
183
184         batch = tlb->active;
185         if (batch->next) {
186                 tlb->active = batch->next;
187                 return 1;
188         }
189
190         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
191         if (!batch)
192                 return 0;
193
194         batch->next = NULL;
195         batch->nr   = 0;
196         batch->max  = MAX_GATHER_BATCH;
197
198         tlb->active->next = batch;
199         tlb->active = batch;
200
201         return 1;
202 }
203
204 /* tlb_gather_mmu
205  *      Called to initialize an (on-stack) mmu_gather structure for page-table
206  *      tear-down from @mm. The @fullmm argument is used when @mm is without
207  *      users and we're going to destroy the full address space (exit/execve).
208  */
209 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
210 {
211         tlb->mm = mm;
212
213         tlb->fullmm     = fullmm;
214         tlb->need_flush = 0;
215         tlb->fast_mode  = (num_possible_cpus() == 1);
216         tlb->local.next = NULL;
217         tlb->local.nr   = 0;
218         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
219         tlb->active     = &tlb->local;
220
221 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
222         tlb->batch = NULL;
223 #endif
224 }
225
226 void tlb_flush_mmu(struct mmu_gather *tlb)
227 {
228         struct mmu_gather_batch *batch;
229
230         if (!tlb->need_flush)
231                 return;
232         tlb->need_flush = 0;
233         tlb_flush(tlb);
234 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
235         tlb_table_flush(tlb);
236 #endif
237
238         if (tlb_fast_mode(tlb))
239                 return;
240
241         for (batch = &tlb->local; batch; batch = batch->next) {
242                 free_pages_and_swap_cache(batch->pages, batch->nr);
243                 batch->nr = 0;
244         }
245         tlb->active = &tlb->local;
246 }
247
248 /* tlb_finish_mmu
249  *      Called at the end of the shootdown operation to free up any resources
250  *      that were required.
251  */
252 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
253 {
254         struct mmu_gather_batch *batch, *next;
255
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(current, 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 int 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         int ret;
1044
1045         /*
1046          * Don't copy ptes where a page fault will fill them correctly.
1047          * Fork becomes much lighter when there are big shared or private
1048          * readonly mappings. The tradeoff is that copy_page_range is more
1049          * efficient than faulting.
1050          */
1051         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1052                 if (!vma->anon_vma)
1053                         return 0;
1054         }
1055
1056         if (is_vm_hugetlb_page(vma))
1057                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1058
1059         if (unlikely(is_pfn_mapping(vma))) {
1060                 /*
1061                  * We do not free on error cases below as remove_vma
1062                  * gets called on error from higher level routine
1063                  */
1064                 ret = track_pfn_vma_copy(vma);
1065                 if (ret)
1066                         return ret;
1067         }
1068
1069         /*
1070          * We need to invalidate the secondary MMU mappings only when
1071          * there could be a permission downgrade on the ptes of the
1072          * parent mm. And a permission downgrade will only happen if
1073          * is_cow_mapping() returns true.
1074          */
1075         if (is_cow_mapping(vma->vm_flags))
1076                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1077
1078         ret = 0;
1079         dst_pgd = pgd_offset(dst_mm, addr);
1080         src_pgd = pgd_offset(src_mm, addr);
1081         do {
1082                 next = pgd_addr_end(addr, end);
1083                 if (pgd_none_or_clear_bad(src_pgd))
1084                         continue;
1085                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1086                                             vma, addr, next))) {
1087                         ret = -ENOMEM;
1088                         break;
1089                 }
1090         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1091
1092         if (is_cow_mapping(vma->vm_flags))
1093                 mmu_notifier_invalidate_range_end(src_mm,
1094                                                   vma->vm_start, end);
1095         return ret;
1096 }
1097
1098 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1099                                 struct vm_area_struct *vma, pmd_t *pmd,
1100                                 unsigned long addr, unsigned long end,
1101                                 struct zap_details *details)
1102 {
1103         struct mm_struct *mm = tlb->mm;
1104         int force_flush = 0;
1105         int rss[NR_MM_COUNTERS];
1106         spinlock_t *ptl;
1107         pte_t *start_pte;
1108         pte_t *pte;
1109
1110 again:
1111         init_rss_vec(rss);
1112         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1113         pte = start_pte;
1114         arch_enter_lazy_mmu_mode();
1115         do {
1116                 pte_t ptent = *pte;
1117                 if (pte_none(ptent)) {
1118                         continue;
1119                 }
1120
1121                 if (pte_present(ptent)) {
1122                         struct page *page;
1123
1124                         page = vm_normal_page(vma, addr, ptent);
1125                         if (unlikely(details) && page) {
1126                                 /*
1127                                  * unmap_shared_mapping_pages() wants to
1128                                  * invalidate cache without truncating:
1129                                  * unmap shared but keep private pages.
1130                                  */
1131                                 if (details->check_mapping &&
1132                                     details->check_mapping != page->mapping)
1133                                         continue;
1134                                 /*
1135                                  * Each page->index must be checked when
1136                                  * invalidating or truncating nonlinear.
1137                                  */
1138                                 if (details->nonlinear_vma &&
1139                                     (page->index < details->first_index ||
1140                                      page->index > details->last_index))
1141                                         continue;
1142                         }
1143                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1144                                                         tlb->fullmm);
1145                         tlb_remove_tlb_entry(tlb, pte, addr);
1146                         if (unlikely(!page))
1147                                 continue;
1148                         if (unlikely(details) && details->nonlinear_vma
1149                             && linear_page_index(details->nonlinear_vma,
1150                                                 addr) != page->index)
1151                                 set_pte_at(mm, addr, pte,
1152                                            pgoff_to_pte(page->index));
1153                         if (PageAnon(page))
1154                                 rss[MM_ANONPAGES]--;
1155                         else {
1156                                 if (pte_dirty(ptent))
1157                                         set_page_dirty(page);
1158                                 if (pte_young(ptent) &&
1159                                     likely(!VM_SequentialReadHint(vma)))
1160                                         mark_page_accessed(page);
1161                                 rss[MM_FILEPAGES]--;
1162                         }
1163                         page_remove_rmap(page);
1164                         if (unlikely(page_mapcount(page) < 0))
1165                                 print_bad_pte(vma, addr, ptent, page);
1166                         force_flush = !__tlb_remove_page(tlb, page);
1167                         if (force_flush)
1168                                 break;
1169                         continue;
1170                 }
1171                 /*
1172                  * If details->check_mapping, we leave swap entries;
1173                  * if details->nonlinear_vma, we leave file entries.
1174                  */
1175                 if (unlikely(details))
1176                         continue;
1177                 if (pte_file(ptent)) {
1178                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1179                                 print_bad_pte(vma, addr, ptent, NULL);
1180                 } else {
1181                         swp_entry_t entry = pte_to_swp_entry(ptent);
1182
1183                         if (!non_swap_entry(entry))
1184                                 rss[MM_SWAPENTS]--;
1185                         else if (is_migration_entry(entry)) {
1186                                 struct page *page;
1187
1188                                 page = migration_entry_to_page(entry);
1189
1190                                 if (PageAnon(page))
1191                                         rss[MM_ANONPAGES]--;
1192                                 else
1193                                         rss[MM_FILEPAGES]--;
1194                         }
1195                         if (unlikely(!free_swap_and_cache(entry)))
1196                                 print_bad_pte(vma, addr, ptent, NULL);
1197                 }
1198                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1199         } while (pte++, addr += PAGE_SIZE, addr != end);
1200
1201         add_mm_rss_vec(mm, rss);
1202         arch_leave_lazy_mmu_mode();
1203         pte_unmap_unlock(start_pte, ptl);
1204
1205         /*
1206          * mmu_gather ran out of room to batch pages, we break out of
1207          * the PTE lock to avoid doing the potential expensive TLB invalidate
1208          * and page-free while holding it.
1209          */
1210         if (force_flush) {
1211                 force_flush = 0;
1212                 tlb_flush_mmu(tlb);
1213                 if (addr != end)
1214                         goto again;
1215         }
1216
1217         return addr;
1218 }
1219
1220 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1221                                 struct vm_area_struct *vma, pud_t *pud,
1222                                 unsigned long addr, unsigned long end,
1223                                 struct zap_details *details)
1224 {
1225         pmd_t *pmd;
1226         unsigned long next;
1227
1228         pmd = pmd_offset(pud, addr);
1229         do {
1230                 next = pmd_addr_end(addr, end);
1231                 if (pmd_trans_huge(*pmd)) {
1232                         if (next - addr != HPAGE_PMD_SIZE) {
1233                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1234                                 split_huge_page_pmd(vma->vm_mm, pmd);
1235                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1236                                 goto next;
1237                         /* fall through */
1238                 }
1239                 /*
1240                  * Here there can be other concurrent MADV_DONTNEED or
1241                  * trans huge page faults running, and if the pmd is
1242                  * none or trans huge it can change under us. This is
1243                  * because MADV_DONTNEED holds the mmap_sem in read
1244                  * mode.
1245                  */
1246                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1247                         goto next;
1248                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1249 next:
1250                 cond_resched();
1251         } while (pmd++, addr = next, addr != end);
1252
1253         return addr;
1254 }
1255
1256 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1257                                 struct vm_area_struct *vma, pgd_t *pgd,
1258                                 unsigned long addr, unsigned long end,
1259                                 struct zap_details *details)
1260 {
1261         pud_t *pud;
1262         unsigned long next;
1263
1264         pud = pud_offset(pgd, addr);
1265         do {
1266                 next = pud_addr_end(addr, end);
1267                 if (pud_none_or_clear_bad(pud))
1268                         continue;
1269                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1270         } while (pud++, addr = next, addr != end);
1271
1272         return addr;
1273 }
1274
1275 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1276                                 struct vm_area_struct *vma,
1277                                 unsigned long addr, unsigned long end,
1278                                 struct zap_details *details)
1279 {
1280         pgd_t *pgd;
1281         unsigned long next;
1282
1283         if (details && !details->check_mapping && !details->nonlinear_vma)
1284                 details = NULL;
1285
1286         BUG_ON(addr >= end);
1287         mem_cgroup_uncharge_start();
1288         tlb_start_vma(tlb, vma);
1289         pgd = pgd_offset(vma->vm_mm, addr);
1290         do {
1291                 next = pgd_addr_end(addr, end);
1292                 if (pgd_none_or_clear_bad(pgd))
1293                         continue;
1294                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1295         } while (pgd++, addr = next, addr != end);
1296         tlb_end_vma(tlb, vma);
1297         mem_cgroup_uncharge_end();
1298
1299         return addr;
1300 }
1301
1302 /**
1303  * unmap_vmas - unmap a range of memory covered by a list of vma's
1304  * @tlb: address of the caller's struct mmu_gather
1305  * @vma: the starting vma
1306  * @start_addr: virtual address at which to start unmapping
1307  * @end_addr: virtual address at which to end unmapping
1308  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1309  * @details: details of nonlinear truncation or shared cache invalidation
1310  *
1311  * Returns the end address of the unmapping (restart addr if interrupted).
1312  *
1313  * Unmap all pages in the vma list.
1314  *
1315  * Only addresses between `start' and `end' will be unmapped.
1316  *
1317  * The VMA list must be sorted in ascending virtual address order.
1318  *
1319  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1320  * range after unmap_vmas() returns.  So the only responsibility here is to
1321  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1322  * drops the lock and schedules.
1323  */
1324 unsigned long unmap_vmas(struct mmu_gather *tlb,
1325                 struct vm_area_struct *vma, unsigned long start_addr,
1326                 unsigned long end_addr, unsigned long *nr_accounted,
1327                 struct zap_details *details)
1328 {
1329         unsigned long start = start_addr;
1330         struct mm_struct *mm = vma->vm_mm;
1331
1332         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1333         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1334                 unsigned long end;
1335
1336                 start = max(vma->vm_start, start_addr);
1337                 if (start >= vma->vm_end)
1338                         continue;
1339                 end = min(vma->vm_end, end_addr);
1340                 if (end <= vma->vm_start)
1341                         continue;
1342
1343                 if (vma->vm_flags & VM_ACCOUNT)
1344                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1345
1346                 if (unlikely(is_pfn_mapping(vma)))
1347                         untrack_pfn_vma(vma, 0, 0);
1348
1349                 while (start != end) {
1350                         if (unlikely(is_vm_hugetlb_page(vma))) {
1351                                 /*
1352                                  * It is undesirable to test vma->vm_file as it
1353                                  * should be non-null for valid hugetlb area.
1354                                  * However, vm_file will be NULL in the error
1355                                  * cleanup path of do_mmap_pgoff. When
1356                                  * hugetlbfs ->mmap method fails,
1357                                  * do_mmap_pgoff() nullifies vma->vm_file
1358                                  * before calling this function to clean up.
1359                                  * Since no pte has actually been setup, it is
1360                                  * safe to do nothing in this case.
1361                                  */
1362                                 if (vma->vm_file)
1363                                         unmap_hugepage_range(vma, start, end, NULL);
1364
1365                                 start = end;
1366                         } else
1367                                 start = unmap_page_range(tlb, vma, start, end, details);
1368                 }
1369         }
1370
1371         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1372         return start;   /* which is now the end (or restart) address */
1373 }
1374
1375 /**
1376  * zap_page_range - remove user pages in a given range
1377  * @vma: vm_area_struct holding the applicable pages
1378  * @address: starting address of pages to zap
1379  * @size: number of bytes to zap
1380  * @details: details of nonlinear truncation or shared cache invalidation
1381  */
1382 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1383                 unsigned long size, struct zap_details *details)
1384 {
1385         struct mm_struct *mm = vma->vm_mm;
1386         struct mmu_gather tlb;
1387         unsigned long end = address + size;
1388         unsigned long nr_accounted = 0;
1389
1390         lru_add_drain();
1391         tlb_gather_mmu(&tlb, mm, 0);
1392         update_hiwater_rss(mm);
1393         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1394         tlb_finish_mmu(&tlb, address, end);
1395         return end;
1396 }
1397
1398 /**
1399  * zap_vma_ptes - remove ptes mapping the vma
1400  * @vma: vm_area_struct holding ptes to be zapped
1401  * @address: starting address of pages to zap
1402  * @size: number of bytes to zap
1403  *
1404  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1405  *
1406  * The entire address range must be fully contained within the vma.
1407  *
1408  * Returns 0 if successful.
1409  */
1410 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1411                 unsigned long size)
1412 {
1413         if (address < vma->vm_start || address + size > vma->vm_end ||
1414                         !(vma->vm_flags & VM_PFNMAP))
1415                 return -1;
1416         zap_page_range(vma, address, size, NULL);
1417         return 0;
1418 }
1419 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1420
1421 /**
1422  * follow_page - look up a page descriptor from a user-virtual address
1423  * @vma: vm_area_struct mapping @address
1424  * @address: virtual address to look up
1425  * @flags: flags modifying lookup behaviour
1426  *
1427  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1428  *
1429  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1430  * an error pointer if there is a mapping to something not represented
1431  * by a page descriptor (see also vm_normal_page()).
1432  */
1433 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1434                         unsigned int flags)
1435 {
1436         pgd_t *pgd;
1437         pud_t *pud;
1438         pmd_t *pmd;
1439         pte_t *ptep, pte;
1440         spinlock_t *ptl;
1441         struct page *page;
1442         struct mm_struct *mm = vma->vm_mm;
1443
1444         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1445         if (!IS_ERR(page)) {
1446                 BUG_ON(flags & FOLL_GET);
1447                 goto out;
1448         }
1449
1450         page = NULL;
1451         pgd = pgd_offset(mm, address);
1452         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1453                 goto no_page_table;
1454
1455         pud = pud_offset(pgd, address);
1456         if (pud_none(*pud))
1457                 goto no_page_table;
1458         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1459                 BUG_ON(flags & FOLL_GET);
1460                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1461                 goto out;
1462         }
1463         if (unlikely(pud_bad(*pud)))
1464                 goto no_page_table;
1465
1466         pmd = pmd_offset(pud, address);
1467         if (pmd_none(*pmd))
1468                 goto no_page_table;
1469         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1470                 BUG_ON(flags & FOLL_GET);
1471                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1472                 goto out;
1473         }
1474         if (pmd_trans_huge(*pmd)) {
1475                 if (flags & FOLL_SPLIT) {
1476                         split_huge_page_pmd(mm, pmd);
1477                         goto split_fallthrough;
1478                 }
1479                 spin_lock(&mm->page_table_lock);
1480                 if (likely(pmd_trans_huge(*pmd))) {
1481                         if (unlikely(pmd_trans_splitting(*pmd))) {
1482                                 spin_unlock(&mm->page_table_lock);
1483                                 wait_split_huge_page(vma->anon_vma, pmd);
1484                         } else {
1485                                 page = follow_trans_huge_pmd(mm, address,
1486                                                              pmd, flags);
1487                                 spin_unlock(&mm->page_table_lock);
1488                                 goto out;
1489                         }
1490                 } else
1491                         spin_unlock(&mm->page_table_lock);
1492                 /* fall through */
1493         }
1494 split_fallthrough:
1495         if (unlikely(pmd_bad(*pmd)))
1496                 goto no_page_table;
1497
1498         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1499
1500         pte = *ptep;
1501         if (!pte_present(pte))
1502                 goto no_page;
1503         if ((flags & FOLL_WRITE) && !pte_write(pte))
1504                 goto unlock;
1505
1506         page = vm_normal_page(vma, address, pte);
1507         if (unlikely(!page)) {
1508                 if ((flags & FOLL_DUMP) ||
1509                     !is_zero_pfn(pte_pfn(pte)))
1510                         goto bad_page;
1511                 page = pte_page(pte);
1512         }
1513
1514         if (flags & FOLL_GET)
1515                 get_page_foll(page);
1516         if (flags & FOLL_TOUCH) {
1517                 if ((flags & FOLL_WRITE) &&
1518                     !pte_dirty(pte) && !PageDirty(page))
1519                         set_page_dirty(page);
1520                 /*
1521                  * pte_mkyoung() would be more correct here, but atomic care
1522                  * is needed to avoid losing the dirty bit: it is easier to use
1523                  * mark_page_accessed().
1524                  */
1525                 mark_page_accessed(page);
1526         }
1527         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1528                 /*
1529                  * The preliminary mapping check is mainly to avoid the
1530                  * pointless overhead of lock_page on the ZERO_PAGE
1531                  * which might bounce very badly if there is contention.
1532                  *
1533                  * If the page is already locked, we don't need to
1534                  * handle it now - vmscan will handle it later if and
1535                  * when it attempts to reclaim the page.
1536                  */
1537                 if (page->mapping && trylock_page(page)) {
1538                         lru_add_drain();  /* push cached pages to LRU */
1539                         /*
1540                          * Because we lock page here and migration is
1541                          * blocked by the pte's page reference, we need
1542                          * only check for file-cache page truncation.
1543                          */
1544                         if (page->mapping)
1545                                 mlock_vma_page(page);
1546                         unlock_page(page);
1547                 }
1548         }
1549 unlock:
1550         pte_unmap_unlock(ptep, ptl);
1551 out:
1552         return page;
1553
1554 bad_page:
1555         pte_unmap_unlock(ptep, ptl);
1556         return ERR_PTR(-EFAULT);
1557
1558 no_page:
1559         pte_unmap_unlock(ptep, ptl);
1560         if (!pte_none(pte))
1561                 return page;
1562
1563 no_page_table:
1564         /*
1565          * When core dumping an enormous anonymous area that nobody
1566          * has touched so far, we don't want to allocate unnecessary pages or
1567          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1568          * then get_dump_page() will return NULL to leave a hole in the dump.
1569          * But we can only make this optimization where a hole would surely
1570          * be zero-filled if handle_mm_fault() actually did handle it.
1571          */
1572         if ((flags & FOLL_DUMP) &&
1573             (!vma->vm_ops || !vma->vm_ops->fault))
1574                 return ERR_PTR(-EFAULT);
1575         return page;
1576 }
1577
1578 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1579 {
1580         return stack_guard_page_start(vma, addr) ||
1581                stack_guard_page_end(vma, addr+PAGE_SIZE);
1582 }
1583
1584 /**
1585  * __get_user_pages() - pin user pages in memory
1586  * @tsk:        task_struct of target task
1587  * @mm:         mm_struct of target mm
1588  * @start:      starting user address
1589  * @nr_pages:   number of pages from start to pin
1590  * @gup_flags:  flags modifying pin behaviour
1591  * @pages:      array that receives pointers to the pages pinned.
1592  *              Should be at least nr_pages long. Or NULL, if caller
1593  *              only intends to ensure the pages are faulted in.
1594  * @vmas:       array of pointers to vmas corresponding to each page.
1595  *              Or NULL if the caller does not require them.
1596  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1597  *
1598  * Returns number of pages pinned. This may be fewer than the number
1599  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1600  * were pinned, returns -errno. Each page returned must be released
1601  * with a put_page() call when it is finished with. vmas will only
1602  * remain valid while mmap_sem is held.
1603  *
1604  * Must be called with mmap_sem held for read or write.
1605  *
1606  * __get_user_pages walks a process's page tables and takes a reference to
1607  * each struct page that each user address corresponds to at a given
1608  * instant. That is, it takes the page that would be accessed if a user
1609  * thread accesses the given user virtual address at that instant.
1610  *
1611  * This does not guarantee that the page exists in the user mappings when
1612  * __get_user_pages returns, and there may even be a completely different
1613  * page there in some cases (eg. if mmapped pagecache has been invalidated
1614  * and subsequently re faulted). However it does guarantee that the page
1615  * won't be freed completely. And mostly callers simply care that the page
1616  * contains data that was valid *at some point in time*. Typically, an IO
1617  * or similar operation cannot guarantee anything stronger anyway because
1618  * locks can't be held over the syscall boundary.
1619  *
1620  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1621  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1622  * appropriate) must be called after the page is finished with, and
1623  * before put_page is called.
1624  *
1625  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1626  * or mmap_sem contention, and if waiting is needed to pin all pages,
1627  * *@nonblocking will be set to 0.
1628  *
1629  * In most cases, get_user_pages or get_user_pages_fast should be used
1630  * instead of __get_user_pages. __get_user_pages should be used only if
1631  * you need some special @gup_flags.
1632  */
1633 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1634                      unsigned long start, int nr_pages, unsigned int gup_flags,
1635                      struct page **pages, struct vm_area_struct **vmas,
1636                      int *nonblocking)
1637 {
1638         int i;
1639         unsigned long vm_flags;
1640
1641         if (nr_pages <= 0)
1642                 return 0;
1643
1644         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1645
1646         /* 
1647          * Require read or write permissions.
1648          * If FOLL_FORCE is set, we only require the "MAY" flags.
1649          */
1650         vm_flags  = (gup_flags & FOLL_WRITE) ?
1651                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1652         vm_flags &= (gup_flags & FOLL_FORCE) ?
1653                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1654         i = 0;
1655
1656         do {
1657                 struct vm_area_struct *vma;
1658
1659                 vma = find_extend_vma(mm, start);
1660                 if (!vma && in_gate_area(mm, start)) {
1661                         unsigned long pg = start & PAGE_MASK;
1662                         pgd_t *pgd;
1663                         pud_t *pud;
1664                         pmd_t *pmd;
1665                         pte_t *pte;
1666
1667                         /* user gate pages are read-only */
1668                         if (gup_flags & FOLL_WRITE)
1669                                 return i ? : -EFAULT;
1670                         if (pg > TASK_SIZE)
1671                                 pgd = pgd_offset_k(pg);
1672                         else
1673                                 pgd = pgd_offset_gate(mm, pg);
1674                         BUG_ON(pgd_none(*pgd));
1675                         pud = pud_offset(pgd, pg);
1676                         BUG_ON(pud_none(*pud));
1677                         pmd = pmd_offset(pud, pg);
1678                         if (pmd_none(*pmd))
1679                                 return i ? : -EFAULT;
1680                         VM_BUG_ON(pmd_trans_huge(*pmd));
1681                         pte = pte_offset_map(pmd, pg);
1682                         if (pte_none(*pte)) {
1683                                 pte_unmap(pte);
1684                                 return i ? : -EFAULT;
1685                         }
1686                         vma = get_gate_vma(mm);
1687                         if (pages) {
1688                                 struct page *page;
1689
1690                                 page = vm_normal_page(vma, start, *pte);
1691                                 if (!page) {
1692                                         if (!(gup_flags & FOLL_DUMP) &&
1693                                              is_zero_pfn(pte_pfn(*pte)))
1694                                                 page = pte_page(*pte);
1695                                         else {
1696                                                 pte_unmap(pte);
1697                                                 return i ? : -EFAULT;
1698                                         }
1699                                 }
1700                                 pages[i] = page;
1701                                 get_page(page);
1702                         }
1703                         pte_unmap(pte);
1704                         goto next_page;
1705                 }
1706
1707                 if (!vma ||
1708                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1709                     !(vm_flags & vma->vm_flags))
1710                         return i ? : -EFAULT;
1711
1712                 if (is_vm_hugetlb_page(vma)) {
1713                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1714                                         &start, &nr_pages, i, gup_flags);
1715                         continue;
1716                 }
1717
1718                 do {
1719                         struct page *page;
1720                         unsigned int foll_flags = gup_flags;
1721
1722                         /*
1723                          * If we have a pending SIGKILL, don't keep faulting
1724                          * pages and potentially allocating memory.
1725                          */
1726                         if (unlikely(fatal_signal_pending(current)))
1727                                 return i ? i : -ERESTARTSYS;
1728
1729                         cond_resched();
1730                         while (!(page = follow_page(vma, start, foll_flags))) {
1731                                 int ret;
1732                                 unsigned int fault_flags = 0;
1733
1734                                 /* For mlock, just skip the stack guard page. */
1735                                 if (foll_flags & FOLL_MLOCK) {
1736                                         if (stack_guard_page(vma, start))
1737                                                 goto next_page;
1738                                 }
1739                                 if (foll_flags & FOLL_WRITE)
1740                                         fault_flags |= FAULT_FLAG_WRITE;
1741                                 if (nonblocking)
1742                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1743                                 if (foll_flags & FOLL_NOWAIT)
1744                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1745
1746                                 ret = handle_mm_fault(mm, vma, start,
1747                                                         fault_flags);
1748
1749                                 if (ret & VM_FAULT_ERROR) {
1750                                         if (ret & VM_FAULT_OOM)
1751                                                 return i ? i : -ENOMEM;
1752                                         if (ret & (VM_FAULT_HWPOISON |
1753                                                    VM_FAULT_HWPOISON_LARGE)) {
1754                                                 if (i)
1755                                                         return i;
1756                                                 else if (gup_flags & FOLL_HWPOISON)
1757                                                         return -EHWPOISON;
1758                                                 else
1759                                                         return -EFAULT;
1760                                         }
1761                                         if (ret & VM_FAULT_SIGBUS)
1762                                                 return i ? i : -EFAULT;
1763                                         BUG();
1764                                 }
1765
1766                                 if (tsk) {
1767                                         if (ret & VM_FAULT_MAJOR)
1768                                                 tsk->maj_flt++;
1769                                         else
1770                                                 tsk->min_flt++;
1771                                 }
1772
1773                                 if (ret & VM_FAULT_RETRY) {
1774                                         if (nonblocking)
1775                                                 *nonblocking = 0;
1776                                         return i;
1777                                 }
1778
1779                                 /*
1780                                  * The VM_FAULT_WRITE bit tells us that
1781                                  * do_wp_page has broken COW when necessary,
1782                                  * even if maybe_mkwrite decided not to set
1783                                  * pte_write. We can thus safely do subsequent
1784                                  * page lookups as if they were reads. But only
1785                                  * do so when looping for pte_write is futile:
1786                                  * in some cases userspace may also be wanting
1787                                  * to write to the gotten user page, which a
1788                                  * read fault here might prevent (a readonly
1789                                  * page might get reCOWed by userspace write).
1790                                  */
1791                                 if ((ret & VM_FAULT_WRITE) &&
1792                                     !(vma->vm_flags & VM_WRITE))
1793                                         foll_flags &= ~FOLL_WRITE;
1794
1795                                 cond_resched();
1796                         }
1797                         if (IS_ERR(page))
1798                                 return i ? i : PTR_ERR(page);
1799                         if (pages) {
1800                                 pages[i] = page;
1801
1802                                 flush_anon_page(vma, page, start);
1803                                 flush_dcache_page(page);
1804                         }
1805 next_page:
1806                         if (vmas)
1807                                 vmas[i] = vma;
1808                         i++;
1809                         start += PAGE_SIZE;
1810                         nr_pages--;
1811                 } while (nr_pages && start < vma->vm_end);
1812         } while (nr_pages);
1813         return i;
1814 }
1815 EXPORT_SYMBOL(__get_user_pages);
1816
1817 /*
1818  * fixup_user_fault() - manually resolve a user page fault
1819  * @tsk:        the task_struct to use for page fault accounting, or
1820  *              NULL if faults are not to be recorded.
1821  * @mm:         mm_struct of target mm
1822  * @address:    user address
1823  * @fault_flags:flags to pass down to handle_mm_fault()
1824  *
1825  * This is meant to be called in the specific scenario where for locking reasons
1826  * we try to access user memory in atomic context (within a pagefault_disable()
1827  * section), this returns -EFAULT, and we want to resolve the user fault before
1828  * trying again.
1829  *
1830  * Typically this is meant to be used by the futex code.
1831  *
1832  * The main difference with get_user_pages() is that this function will
1833  * unconditionally call handle_mm_fault() which will in turn perform all the
1834  * necessary SW fixup of the dirty and young bits in the PTE, while
1835  * handle_mm_fault() only guarantees to update these in the struct page.
1836  *
1837  * This is important for some architectures where those bits also gate the
1838  * access permission to the page because they are maintained in software.  On
1839  * such architectures, gup() will not be enough to make a subsequent access
1840  * succeed.
1841  *
1842  * This should be called with the mm_sem held for read.
1843  */
1844 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1845                      unsigned long address, unsigned int fault_flags)
1846 {
1847         struct vm_area_struct *vma;
1848         int ret;
1849
1850         vma = find_extend_vma(mm, address);
1851         if (!vma || address < vma->vm_start)
1852                 return -EFAULT;
1853
1854         ret = handle_mm_fault(mm, vma, address, fault_flags);
1855         if (ret & VM_FAULT_ERROR) {
1856                 if (ret & VM_FAULT_OOM)
1857                         return -ENOMEM;
1858                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1859                         return -EHWPOISON;
1860                 if (ret & VM_FAULT_SIGBUS)
1861                         return -EFAULT;
1862                 BUG();
1863         }
1864         if (tsk) {
1865                 if (ret & VM_FAULT_MAJOR)
1866                         tsk->maj_flt++;
1867                 else
1868                         tsk->min_flt++;
1869         }
1870         return 0;
1871 }
1872
1873 /*
1874  * get_user_pages() - pin user pages in memory
1875  * @tsk:        the task_struct to use for page fault accounting, or
1876  *              NULL if faults are not to be recorded.
1877  * @mm:         mm_struct of target mm
1878  * @start:      starting user address
1879  * @nr_pages:   number of pages from start to pin
1880  * @write:      whether pages will be written to by the caller
1881  * @force:      whether to force write access even if user mapping is
1882  *              readonly. This will result in the page being COWed even
1883  *              in MAP_SHARED mappings. You do not want this.
1884  * @pages:      array that receives pointers to the pages pinned.
1885  *              Should be at least nr_pages long. Or NULL, if caller
1886  *              only intends to ensure the pages are faulted in.
1887  * @vmas:       array of pointers to vmas corresponding to each page.
1888  *              Or NULL if the caller does not require them.
1889  *
1890  * Returns number of pages pinned. This may be fewer than the number
1891  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1892  * were pinned, returns -errno. Each page returned must be released
1893  * with a put_page() call when it is finished with. vmas will only
1894  * remain valid while mmap_sem is held.
1895  *
1896  * Must be called with mmap_sem held for read or write.
1897  *
1898  * get_user_pages walks a process's page tables and takes a reference to
1899  * each struct page that each user address corresponds to at a given
1900  * instant. That is, it takes the page that would be accessed if a user
1901  * thread accesses the given user virtual address at that instant.
1902  *
1903  * This does not guarantee that the page exists in the user mappings when
1904  * get_user_pages returns, and there may even be a completely different
1905  * page there in some cases (eg. if mmapped pagecache has been invalidated
1906  * and subsequently re faulted). However it does guarantee that the page
1907  * won't be freed completely. And mostly callers simply care that the page
1908  * contains data that was valid *at some point in time*. Typically, an IO
1909  * or similar operation cannot guarantee anything stronger anyway because
1910  * locks can't be held over the syscall boundary.
1911  *
1912  * If write=0, the page must not be written to. If the page is written to,
1913  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1914  * after the page is finished with, and before put_page is called.
1915  *
1916  * get_user_pages is typically used for fewer-copy IO operations, to get a
1917  * handle on the memory by some means other than accesses via the user virtual
1918  * addresses. The pages may be submitted for DMA to devices or accessed via
1919  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1920  * use the correct cache flushing APIs.
1921  *
1922  * See also get_user_pages_fast, for performance critical applications.
1923  */
1924 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1925                 unsigned long start, int nr_pages, int write, int force,
1926                 struct page **pages, struct vm_area_struct **vmas)
1927 {
1928         int flags = FOLL_TOUCH;
1929
1930         if (pages)
1931                 flags |= FOLL_GET;
1932         if (write)
1933                 flags |= FOLL_WRITE;
1934         if (force)
1935                 flags |= FOLL_FORCE;
1936
1937         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1938                                 NULL);
1939 }
1940 EXPORT_SYMBOL(get_user_pages);
1941
1942 /**
1943  * get_dump_page() - pin user page in memory while writing it to core dump
1944  * @addr: user address
1945  *
1946  * Returns struct page pointer of user page pinned for dump,
1947  * to be freed afterwards by page_cache_release() or put_page().
1948  *
1949  * Returns NULL on any kind of failure - a hole must then be inserted into
1950  * the corefile, to preserve alignment with its headers; and also returns
1951  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1952  * allowing a hole to be left in the corefile to save diskspace.
1953  *
1954  * Called without mmap_sem, but after all other threads have been killed.
1955  */
1956 #ifdef CONFIG_ELF_CORE
1957 struct page *get_dump_page(unsigned long addr)
1958 {
1959         struct vm_area_struct *vma;
1960         struct page *page;
1961
1962         if (__get_user_pages(current, current->mm, addr, 1,
1963                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1964                              NULL) < 1)
1965                 return NULL;
1966         flush_cache_page(vma, addr, page_to_pfn(page));
1967         return page;
1968 }
1969 #endif /* CONFIG_ELF_CORE */
1970
1971 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1972                         spinlock_t **ptl)
1973 {
1974         pgd_t * pgd = pgd_offset(mm, addr);
1975         pud_t * pud = pud_alloc(mm, pgd, addr);
1976         if (pud) {
1977                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1978                 if (pmd) {
1979                         VM_BUG_ON(pmd_trans_huge(*pmd));
1980                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1981                 }
1982         }
1983         return NULL;
1984 }
1985
1986 /*
1987  * This is the old fallback for page remapping.
1988  *
1989  * For historical reasons, it only allows reserved pages. Only
1990  * old drivers should use this, and they needed to mark their
1991  * pages reserved for the old functions anyway.
1992  */
1993 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1994                         struct page *page, pgprot_t prot)
1995 {
1996         struct mm_struct *mm = vma->vm_mm;
1997         int retval;
1998         pte_t *pte;
1999         spinlock_t *ptl;
2000
2001         retval = -EINVAL;
2002         if (PageAnon(page))
2003                 goto out;
2004         retval = -ENOMEM;
2005         flush_dcache_page(page);
2006         pte = get_locked_pte(mm, addr, &ptl);
2007         if (!pte)
2008                 goto out;
2009         retval = -EBUSY;
2010         if (!pte_none(*pte))
2011                 goto out_unlock;
2012
2013         /* Ok, finally just insert the thing.. */
2014         get_page(page);
2015         inc_mm_counter_fast(mm, MM_FILEPAGES);
2016         page_add_file_rmap(page);
2017         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2018
2019         retval = 0;
2020         pte_unmap_unlock(pte, ptl);
2021         return retval;
2022 out_unlock:
2023         pte_unmap_unlock(pte, ptl);
2024 out:
2025         return retval;
2026 }
2027
2028 /**
2029  * vm_insert_page - insert single page into user vma
2030  * @vma: user vma to map to
2031  * @addr: target user address of this page
2032  * @page: source kernel page
2033  *
2034  * This allows drivers to insert individual pages they've allocated
2035  * into a user vma.
2036  *
2037  * The page has to be a nice clean _individual_ kernel allocation.
2038  * If you allocate a compound page, you need to have marked it as
2039  * such (__GFP_COMP), or manually just split the page up yourself
2040  * (see split_page()).
2041  *
2042  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2043  * took an arbitrary page protection parameter. This doesn't allow
2044  * that. Your vma protection will have to be set up correctly, which
2045  * means that if you want a shared writable mapping, you'd better
2046  * ask for a shared writable mapping!
2047  *
2048  * The page does not need to be reserved.
2049  */
2050 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2051                         struct page *page)
2052 {
2053         if (addr < vma->vm_start || addr >= vma->vm_end)
2054                 return -EFAULT;
2055         if (!page_count(page))
2056                 return -EINVAL;
2057         vma->vm_flags |= VM_INSERTPAGE;
2058         return insert_page(vma, addr, page, vma->vm_page_prot);
2059 }
2060 EXPORT_SYMBOL(vm_insert_page);
2061
2062 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2063                         unsigned long pfn, pgprot_t prot)
2064 {
2065         struct mm_struct *mm = vma->vm_mm;
2066         int retval;
2067         pte_t *pte, entry;
2068         spinlock_t *ptl;
2069
2070         retval = -ENOMEM;
2071         pte = get_locked_pte(mm, addr, &ptl);
2072         if (!pte)
2073                 goto out;
2074         retval = -EBUSY;
2075         if (!pte_none(*pte))
2076                 goto out_unlock;
2077
2078         /* Ok, finally just insert the thing.. */
2079         entry = pte_mkspecial(pfn_pte(pfn, prot));
2080         set_pte_at(mm, addr, pte, entry);
2081         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2082
2083         retval = 0;
2084 out_unlock:
2085         pte_unmap_unlock(pte, ptl);
2086 out:
2087         return retval;
2088 }
2089
2090 /**
2091  * vm_insert_pfn - insert single pfn into user vma
2092  * @vma: user vma to map to
2093  * @addr: target user address of this page
2094  * @pfn: source kernel pfn
2095  *
2096  * Similar to vm_inert_page, this allows drivers to insert individual pages
2097  * they've allocated into a user vma. Same comments apply.
2098  *
2099  * This function should only be called from a vm_ops->fault handler, and
2100  * in that case the handler should return NULL.
2101  *
2102  * vma cannot be a COW mapping.
2103  *
2104  * As this is called only for pages that do not currently exist, we
2105  * do not need to flush old virtual caches or the TLB.
2106  */
2107 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2108                         unsigned long pfn)
2109 {
2110         int ret;
2111         pgprot_t pgprot = vma->vm_page_prot;
2112         /*
2113          * Technically, architectures with pte_special can avoid all these
2114          * restrictions (same for remap_pfn_range).  However we would like
2115          * consistency in testing and feature parity among all, so we should
2116          * try to keep these invariants in place for everybody.
2117          */
2118         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2119         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2120                                                 (VM_PFNMAP|VM_MIXEDMAP));
2121         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2122         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2123
2124         if (addr < vma->vm_start || addr >= vma->vm_end)
2125                 return -EFAULT;
2126         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2127                 return -EINVAL;
2128
2129         ret = insert_pfn(vma, addr, pfn, pgprot);
2130
2131         if (ret)
2132                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2133
2134         return ret;
2135 }
2136 EXPORT_SYMBOL(vm_insert_pfn);
2137
2138 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2139                         unsigned long pfn)
2140 {
2141         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2142
2143         if (addr < vma->vm_start || addr >= vma->vm_end)
2144                 return -EFAULT;
2145
2146         /*
2147          * If we don't have pte special, then we have to use the pfn_valid()
2148          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2149          * refcount the page if pfn_valid is true (hence insert_page rather
2150          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2151          * without pte special, it would there be refcounted as a normal page.
2152          */
2153         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2154                 struct page *page;
2155
2156                 page = pfn_to_page(pfn);
2157                 return insert_page(vma, addr, page, vma->vm_page_prot);
2158         }
2159         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2160 }
2161 EXPORT_SYMBOL(vm_insert_mixed);
2162
2163 /*
2164  * maps a range of physical memory into the requested pages. the old
2165  * mappings are removed. any references to nonexistent pages results
2166  * in null mappings (currently treated as "copy-on-access")
2167  */
2168 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2169                         unsigned long addr, unsigned long end,
2170                         unsigned long pfn, pgprot_t prot)
2171 {
2172         pte_t *pte;
2173         spinlock_t *ptl;
2174
2175         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2176         if (!pte)
2177                 return -ENOMEM;
2178         arch_enter_lazy_mmu_mode();
2179         do {
2180                 BUG_ON(!pte_none(*pte));
2181                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2182                 pfn++;
2183         } while (pte++, addr += PAGE_SIZE, addr != end);
2184         arch_leave_lazy_mmu_mode();
2185         pte_unmap_unlock(pte - 1, ptl);
2186         return 0;
2187 }
2188
2189 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2190                         unsigned long addr, unsigned long end,
2191                         unsigned long pfn, pgprot_t prot)
2192 {
2193         pmd_t *pmd;
2194         unsigned long next;
2195
2196         pfn -= addr >> PAGE_SHIFT;
2197         pmd = pmd_alloc(mm, pud, addr);
2198         if (!pmd)
2199                 return -ENOMEM;
2200         VM_BUG_ON(pmd_trans_huge(*pmd));
2201         do {
2202                 next = pmd_addr_end(addr, end);
2203                 if (remap_pte_range(mm, pmd, addr, next,
2204                                 pfn + (addr >> PAGE_SHIFT), prot))
2205                         return -ENOMEM;
2206         } while (pmd++, addr = next, addr != end);
2207         return 0;
2208 }
2209
2210 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2211                         unsigned long addr, unsigned long end,
2212                         unsigned long pfn, pgprot_t prot)
2213 {
2214         pud_t *pud;
2215         unsigned long next;
2216
2217         pfn -= addr >> PAGE_SHIFT;
2218         pud = pud_alloc(mm, pgd, addr);
2219         if (!pud)
2220                 return -ENOMEM;
2221         do {
2222                 next = pud_addr_end(addr, end);
2223                 if (remap_pmd_range(mm, pud, addr, next,
2224                                 pfn + (addr >> PAGE_SHIFT), prot))
2225                         return -ENOMEM;
2226         } while (pud++, addr = next, addr != end);
2227         return 0;
2228 }
2229
2230 /**
2231  * remap_pfn_range - remap kernel memory to userspace
2232  * @vma: user vma to map to
2233  * @addr: target user address to start at
2234  * @pfn: physical address of kernel memory
2235  * @size: size of map area
2236  * @prot: page protection flags for this mapping
2237  *
2238  *  Note: this is only safe if the mm semaphore is held when called.
2239  */
2240 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2241                     unsigned long pfn, unsigned long size, pgprot_t prot)
2242 {
2243         pgd_t *pgd;
2244         unsigned long next;
2245         unsigned long end = addr + PAGE_ALIGN(size);
2246         struct mm_struct *mm = vma->vm_mm;
2247         int err;
2248
2249         /*
2250          * Physically remapped pages are special. Tell the
2251          * rest of the world about it:
2252          *   VM_IO tells people not to look at these pages
2253          *      (accesses can have side effects).
2254          *   VM_RESERVED is specified all over the place, because
2255          *      in 2.4 it kept swapout's vma scan off this vma; but
2256          *      in 2.6 the LRU scan won't even find its pages, so this
2257          *      flag means no more than count its pages in reserved_vm,
2258          *      and omit it from core dump, even when VM_IO turned off.
2259          *   VM_PFNMAP tells the core MM that the base pages are just
2260          *      raw PFN mappings, and do not have a "struct page" associated
2261          *      with them.
2262          *
2263          * There's a horrible special case to handle copy-on-write
2264          * behaviour that some programs depend on. We mark the "original"
2265          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2266          */
2267         if (addr == vma->vm_start && end == vma->vm_end) {
2268                 vma->vm_pgoff = pfn;
2269                 vma->vm_flags |= VM_PFN_AT_MMAP;
2270         } else if (is_cow_mapping(vma->vm_flags))
2271                 return -EINVAL;
2272
2273         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2274
2275         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2276         if (err) {
2277                 /*
2278                  * To indicate that track_pfn related cleanup is not
2279                  * needed from higher level routine calling unmap_vmas
2280                  */
2281                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2282                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2283                 return -EINVAL;
2284         }
2285
2286         BUG_ON(addr >= end);
2287         pfn -= addr >> PAGE_SHIFT;
2288         pgd = pgd_offset(mm, addr);
2289         flush_cache_range(vma, addr, end);
2290         do {
2291                 next = pgd_addr_end(addr, end);
2292                 err = remap_pud_range(mm, pgd, addr, next,
2293                                 pfn + (addr >> PAGE_SHIFT), prot);
2294                 if (err)
2295                         break;
2296         } while (pgd++, addr = next, addr != end);
2297
2298         if (err)
2299                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2300
2301         return err;
2302 }
2303 EXPORT_SYMBOL(remap_pfn_range);
2304
2305 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2306                                      unsigned long addr, unsigned long end,
2307                                      pte_fn_t fn, void *data)
2308 {
2309         pte_t *pte;
2310         int err;
2311         pgtable_t token;
2312         spinlock_t *uninitialized_var(ptl);
2313
2314         pte = (mm == &init_mm) ?
2315                 pte_alloc_kernel(pmd, addr) :
2316                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2317         if (!pte)
2318                 return -ENOMEM;
2319
2320         BUG_ON(pmd_huge(*pmd));
2321
2322         arch_enter_lazy_mmu_mode();
2323
2324         token = pmd_pgtable(*pmd);
2325
2326         do {
2327                 err = fn(pte++, token, addr, data);
2328                 if (err)
2329                         break;
2330         } while (addr += PAGE_SIZE, addr != end);
2331
2332         arch_leave_lazy_mmu_mode();
2333
2334         if (mm != &init_mm)
2335                 pte_unmap_unlock(pte-1, ptl);
2336         return err;
2337 }
2338
2339 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2340                                      unsigned long addr, unsigned long end,
2341                                      pte_fn_t fn, void *data)
2342 {
2343         pmd_t *pmd;
2344         unsigned long next;
2345         int err;
2346
2347         BUG_ON(pud_huge(*pud));
2348
2349         pmd = pmd_alloc(mm, pud, addr);
2350         if (!pmd)
2351                 return -ENOMEM;
2352         do {
2353                 next = pmd_addr_end(addr, end);
2354                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2355                 if (err)
2356                         break;
2357         } while (pmd++, addr = next, addr != end);
2358         return err;
2359 }
2360
2361 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2362                                      unsigned long addr, unsigned long end,
2363                                      pte_fn_t fn, void *data)
2364 {
2365         pud_t *pud;
2366         unsigned long next;
2367         int err;
2368
2369         pud = pud_alloc(mm, pgd, addr);
2370         if (!pud)
2371                 return -ENOMEM;
2372         do {
2373                 next = pud_addr_end(addr, end);
2374                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2375                 if (err)
2376                         break;
2377         } while (pud++, addr = next, addr != end);
2378         return err;
2379 }
2380
2381 /*
2382  * Scan a region of virtual memory, filling in page tables as necessary
2383  * and calling a provided function on each leaf page table.
2384  */
2385 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2386                         unsigned long size, pte_fn_t fn, void *data)
2387 {
2388         pgd_t *pgd;
2389         unsigned long next;
2390         unsigned long end = addr + size;
2391         int err;
2392
2393         BUG_ON(addr >= end);
2394         pgd = pgd_offset(mm, addr);
2395         do {
2396                 next = pgd_addr_end(addr, end);
2397                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2398                 if (err)
2399                         break;
2400         } while (pgd++, addr = next, addr != end);
2401
2402         return err;
2403 }
2404 EXPORT_SYMBOL_GPL(apply_to_page_range);
2405
2406 /*
2407  * handle_pte_fault chooses page fault handler according to an entry
2408  * which was read non-atomically.  Before making any commitment, on
2409  * those architectures or configurations (e.g. i386 with PAE) which
2410  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2411  * must check under lock before unmapping the pte and proceeding
2412  * (but do_wp_page is only called after already making such a check;
2413  * and do_anonymous_page can safely check later on).
2414  */
2415 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2416                                 pte_t *page_table, pte_t orig_pte)
2417 {
2418         int same = 1;
2419 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2420         if (sizeof(pte_t) > sizeof(unsigned long)) {
2421                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2422                 spin_lock(ptl);
2423                 same = pte_same(*page_table, orig_pte);
2424                 spin_unlock(ptl);
2425         }
2426 #endif
2427         pte_unmap(page_table);
2428         return same;
2429 }
2430
2431 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2432 {
2433         /*
2434          * If the source page was a PFN mapping, we don't have
2435          * a "struct page" for it. We do a best-effort copy by
2436          * just copying from the original user address. If that
2437          * fails, we just zero-fill it. Live with it.
2438          */
2439         if (unlikely(!src)) {
2440                 void *kaddr = kmap_atomic(dst);
2441                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2442
2443                 /*
2444                  * This really shouldn't fail, because the page is there
2445                  * in the page tables. But it might just be unreadable,
2446                  * in which case we just give up and fill the result with
2447                  * zeroes.
2448                  */
2449                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2450                         clear_page(kaddr);
2451                 kunmap_atomic(kaddr);
2452                 flush_dcache_page(dst);
2453         } else
2454                 copy_user_highpage(dst, src, va, vma);
2455 }
2456
2457 /*
2458  * This routine handles present pages, when users try to write
2459  * to a shared page. It is done by copying the page to a new address
2460  * and decrementing the shared-page counter for the old page.
2461  *
2462  * Note that this routine assumes that the protection checks have been
2463  * done by the caller (the low-level page fault routine in most cases).
2464  * Thus we can safely just mark it writable once we've done any necessary
2465  * COW.
2466  *
2467  * We also mark the page dirty at this point even though the page will
2468  * change only once the write actually happens. This avoids a few races,
2469  * and potentially makes it more efficient.
2470  *
2471  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2472  * but allow concurrent faults), with pte both mapped and locked.
2473  * We return with mmap_sem still held, but pte unmapped and unlocked.
2474  */
2475 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2476                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2477                 spinlock_t *ptl, pte_t orig_pte)
2478         __releases(ptl)
2479 {
2480         struct page *old_page, *new_page;
2481         pte_t entry;
2482         int ret = 0;
2483         int page_mkwrite = 0;
2484         struct page *dirty_page = NULL;
2485
2486         old_page = vm_normal_page(vma, address, orig_pte);
2487         if (!old_page) {
2488                 /*
2489                  * VM_MIXEDMAP !pfn_valid() case
2490                  *
2491                  * We should not cow pages in a shared writeable mapping.
2492                  * Just mark the pages writable as we can't do any dirty
2493                  * accounting on raw pfn maps.
2494                  */
2495                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2496                                      (VM_WRITE|VM_SHARED))
2497                         goto reuse;
2498                 goto gotten;
2499         }
2500
2501         /*
2502          * Take out anonymous pages first, anonymous shared vmas are
2503          * not dirty accountable.
2504          */
2505         if (PageAnon(old_page) && !PageKsm(old_page)) {
2506                 if (!trylock_page(old_page)) {
2507                         page_cache_get(old_page);
2508                         pte_unmap_unlock(page_table, ptl);
2509                         lock_page(old_page);
2510                         page_table = pte_offset_map_lock(mm, pmd, address,
2511                                                          &ptl);
2512                         if (!pte_same(*page_table, orig_pte)) {
2513                                 unlock_page(old_page);
2514                                 goto unlock;
2515                         }
2516                         page_cache_release(old_page);
2517                 }
2518                 if (reuse_swap_page(old_page)) {
2519                         /*
2520                          * The page is all ours.  Move it to our anon_vma so
2521                          * the rmap code will not search our parent or siblings.
2522                          * Protected against the rmap code by the page lock.
2523                          */
2524                         page_move_anon_rmap(old_page, vma, address);
2525                         unlock_page(old_page);
2526                         goto reuse;
2527                 }
2528                 unlock_page(old_page);
2529         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2530                                         (VM_WRITE|VM_SHARED))) {
2531                 /*
2532                  * Only catch write-faults on shared writable pages,
2533                  * read-only shared pages can get COWed by
2534                  * get_user_pages(.write=1, .force=1).
2535                  */
2536                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2537                         struct vm_fault vmf;
2538                         int tmp;
2539
2540                         vmf.virtual_address = (void __user *)(address &
2541                                                                 PAGE_MASK);
2542                         vmf.pgoff = old_page->index;
2543                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2544                         vmf.page = old_page;
2545
2546                         /*
2547                          * Notify the address space that the page is about to
2548                          * become writable so that it can prohibit this or wait
2549                          * for the page to get into an appropriate state.
2550                          *
2551                          * We do this without the lock held, so that it can
2552                          * sleep if it needs to.
2553                          */
2554                         page_cache_get(old_page);
2555                         pte_unmap_unlock(page_table, ptl);
2556
2557                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2558                         if (unlikely(tmp &
2559                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2560                                 ret = tmp;
2561                                 goto unwritable_page;
2562                         }
2563                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2564                                 lock_page(old_page);
2565                                 if (!old_page->mapping) {
2566                                         ret = 0; /* retry the fault */
2567                                         unlock_page(old_page);
2568                                         goto unwritable_page;
2569                                 }
2570                         } else
2571                                 VM_BUG_ON(!PageLocked(old_page));
2572
2573                         /*
2574                          * Since we dropped the lock we need to revalidate
2575                          * the PTE as someone else may have changed it.  If
2576                          * they did, we just return, as we can count on the
2577                          * MMU to tell us if they didn't also make it writable.
2578                          */
2579                         page_table = pte_offset_map_lock(mm, pmd, address,
2580                                                          &ptl);
2581                         if (!pte_same(*page_table, orig_pte)) {
2582                                 unlock_page(old_page);
2583                                 goto unlock;
2584                         }
2585
2586                         page_mkwrite = 1;
2587                 }
2588                 dirty_page = old_page;
2589                 get_page(dirty_page);
2590
2591 reuse:
2592                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2593                 entry = pte_mkyoung(orig_pte);
2594                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2595                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2596                         update_mmu_cache(vma, address, page_table);
2597                 pte_unmap_unlock(page_table, ptl);
2598                 ret |= VM_FAULT_WRITE;
2599
2600                 if (!dirty_page)
2601                         return ret;
2602
2603                 /*
2604                  * Yes, Virginia, this is actually required to prevent a race
2605                  * with clear_page_dirty_for_io() from clearing the page dirty
2606                  * bit after it clear all dirty ptes, but before a racing
2607                  * do_wp_page installs a dirty pte.
2608                  *
2609                  * __do_fault is protected similarly.
2610                  */
2611                 if (!page_mkwrite) {
2612                         wait_on_page_locked(dirty_page);
2613                         set_page_dirty_balance(dirty_page, page_mkwrite);
2614                 }
2615                 put_page(dirty_page);
2616                 if (page_mkwrite) {
2617                         struct address_space *mapping = dirty_page->mapping;
2618
2619                         set_page_dirty(dirty_page);
2620                         unlock_page(dirty_page);
2621                         page_cache_release(dirty_page);
2622                         if (mapping)    {
2623                                 /*
2624                                  * Some device drivers do not set page.mapping
2625                                  * but still dirty their pages
2626                                  */
2627                                 balance_dirty_pages_ratelimited(mapping);
2628                         }
2629                 }
2630
2631                 /* file_update_time outside page_lock */
2632                 if (vma->vm_file)
2633                         file_update_time(vma->vm_file);
2634
2635                 return ret;
2636         }
2637
2638         /*
2639          * Ok, we need to copy. Oh, well..
2640          */
2641         page_cache_get(old_page);
2642 gotten:
2643         pte_unmap_unlock(page_table, ptl);
2644
2645         if (unlikely(anon_vma_prepare(vma)))
2646                 goto oom;
2647
2648         if (is_zero_pfn(pte_pfn(orig_pte))) {
2649                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2650                 if (!new_page)
2651                         goto oom;
2652         } else {
2653                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2654                 if (!new_page)
2655                         goto oom;
2656                 cow_user_page(new_page, old_page, address, vma);
2657         }
2658         __SetPageUptodate(new_page);
2659
2660         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2661                 goto oom_free_new;
2662
2663         /*
2664          * Re-check the pte - we dropped the lock
2665          */
2666         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2667         if (likely(pte_same(*page_table, orig_pte))) {
2668                 if (old_page) {
2669                         if (!PageAnon(old_page)) {
2670                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2671                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2672                         }
2673                 } else
2674                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2675                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2676                 entry = mk_pte(new_page, vma->vm_page_prot);
2677                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2678                 /*
2679                  * Clear the pte entry and flush it first, before updating the
2680                  * pte with the new entry. This will avoid a race condition
2681                  * seen in the presence of one thread doing SMC and another
2682                  * thread doing COW.
2683                  */
2684                 ptep_clear_flush(vma, address, page_table);
2685                 page_add_new_anon_rmap(new_page, vma, address);
2686                 /*
2687                  * We call the notify macro here because, when using secondary
2688                  * mmu page tables (such as kvm shadow page tables), we want the
2689                  * new page to be mapped directly into the secondary page table.
2690                  */
2691                 set_pte_at_notify(mm, address, page_table, entry);
2692                 update_mmu_cache(vma, address, page_table);
2693                 if (old_page) {
2694                         /*
2695                          * Only after switching the pte to the new page may
2696                          * we remove the mapcount here. Otherwise another
2697                          * process may come and find the rmap count decremented
2698                          * before the pte is switched to the new page, and
2699                          * "reuse" the old page writing into it while our pte
2700                          * here still points into it and can be read by other
2701                          * threads.
2702                          *
2703                          * The critical issue is to order this
2704                          * page_remove_rmap with the ptp_clear_flush above.
2705                          * Those stores are ordered by (if nothing else,)
2706                          * the barrier present in the atomic_add_negative
2707                          * in page_remove_rmap.
2708                          *
2709                          * Then the TLB flush in ptep_clear_flush ensures that
2710                          * no process can access the old page before the
2711                          * decremented mapcount is visible. And the old page
2712                          * cannot be reused until after the decremented
2713                          * mapcount is visible. So transitively, TLBs to
2714                          * old page will be flushed before it can be reused.
2715                          */
2716                         page_remove_rmap(old_page);
2717                 }
2718
2719                 /* Free the old page.. */
2720                 new_page = old_page;
2721                 ret |= VM_FAULT_WRITE;
2722         } else
2723                 mem_cgroup_uncharge_page(new_page);
2724
2725         if (new_page)
2726                 page_cache_release(new_page);
2727 unlock:
2728         pte_unmap_unlock(page_table, ptl);
2729         if (old_page) {
2730                 /*
2731                  * Don't let another task, with possibly unlocked vma,
2732                  * keep the mlocked page.
2733                  */
2734                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2735                         lock_page(old_page);    /* LRU manipulation */
2736                         munlock_vma_page(old_page);
2737                         unlock_page(old_page);
2738                 }
2739                 page_cache_release(old_page);
2740         }
2741         return ret;
2742 oom_free_new:
2743         page_cache_release(new_page);
2744 oom:
2745         if (old_page) {
2746                 if (page_mkwrite) {
2747                         unlock_page(old_page);
2748                         page_cache_release(old_page);
2749                 }
2750                 page_cache_release(old_page);
2751         }
2752         return VM_FAULT_OOM;
2753
2754 unwritable_page:
2755         page_cache_release(old_page);
2756         return ret;
2757 }
2758
2759 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2760                 unsigned long start_addr, unsigned long end_addr,
2761                 struct zap_details *details)
2762 {
2763         zap_page_range(vma, start_addr, end_addr - start_addr, details);
2764 }
2765
2766 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2767                                             struct zap_details *details)
2768 {
2769         struct vm_area_struct *vma;
2770         struct prio_tree_iter iter;
2771         pgoff_t vba, vea, zba, zea;
2772
2773         vma_prio_tree_foreach(vma, &iter, root,
2774                         details->first_index, details->last_index) {
2775
2776                 vba = vma->vm_pgoff;
2777                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2778                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2779                 zba = details->first_index;
2780                 if (zba < vba)
2781                         zba = vba;
2782                 zea = details->last_index;
2783                 if (zea > vea)
2784                         zea = vea;
2785
2786                 unmap_mapping_range_vma(vma,
2787                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2788                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2789                                 details);
2790         }
2791 }
2792
2793 static inline void unmap_mapping_range_list(struct list_head *head,
2794                                             struct zap_details *details)
2795 {
2796         struct vm_area_struct *vma;
2797
2798         /*
2799          * In nonlinear VMAs there is no correspondence between virtual address
2800          * offset and file offset.  So we must perform an exhaustive search
2801          * across *all* the pages in each nonlinear VMA, not just the pages
2802          * whose virtual address lies outside the file truncation point.
2803          */
2804         list_for_each_entry(vma, head, shared.vm_set.list) {
2805                 details->nonlinear_vma = vma;
2806                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2807         }
2808 }
2809
2810 /**
2811  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2812  * @mapping: the address space containing mmaps to be unmapped.
2813  * @holebegin: byte in first page to unmap, relative to the start of
2814  * the underlying file.  This will be rounded down to a PAGE_SIZE
2815  * boundary.  Note that this is different from truncate_pagecache(), which
2816  * must keep the partial page.  In contrast, we must get rid of
2817  * partial pages.
2818  * @holelen: size of prospective hole in bytes.  This will be rounded
2819  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2820  * end of the file.
2821  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2822  * but 0 when invalidating pagecache, don't throw away private data.
2823  */
2824 void unmap_mapping_range(struct address_space *mapping,
2825                 loff_t const holebegin, loff_t const holelen, int even_cows)
2826 {
2827         struct zap_details details;
2828         pgoff_t hba = holebegin >> PAGE_SHIFT;
2829         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2830
2831         /* Check for overflow. */
2832         if (sizeof(holelen) > sizeof(hlen)) {
2833                 long long holeend =
2834                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2835                 if (holeend & ~(long long)ULONG_MAX)
2836                         hlen = ULONG_MAX - hba + 1;
2837         }
2838
2839         details.check_mapping = even_cows? NULL: mapping;
2840         details.nonlinear_vma = NULL;
2841         details.first_index = hba;
2842         details.last_index = hba + hlen - 1;
2843         if (details.last_index < details.first_index)
2844                 details.last_index = ULONG_MAX;
2845
2846
2847         mutex_lock(&mapping->i_mmap_mutex);
2848         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2849                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2850         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2851                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2852         mutex_unlock(&mapping->i_mmap_mutex);
2853 }
2854 EXPORT_SYMBOL(unmap_mapping_range);
2855
2856 /*
2857  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2858  * but allow concurrent faults), and pte mapped but not yet locked.
2859  * We return with mmap_sem still held, but pte unmapped and unlocked.
2860  */
2861 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2862                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2863                 unsigned int flags, pte_t orig_pte)
2864 {
2865         spinlock_t *ptl;
2866         struct page *page, *swapcache = NULL;
2867         swp_entry_t entry;
2868         pte_t pte;
2869         int locked;
2870         struct mem_cgroup *ptr;
2871         int exclusive = 0;
2872         int ret = 0;
2873
2874         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2875                 goto out;
2876
2877         entry = pte_to_swp_entry(orig_pte);
2878         if (unlikely(non_swap_entry(entry))) {
2879                 if (is_migration_entry(entry)) {
2880                         migration_entry_wait(mm, pmd, address);
2881                 } else if (is_hwpoison_entry(entry)) {
2882                         ret = VM_FAULT_HWPOISON;
2883                 } else {
2884                         print_bad_pte(vma, address, orig_pte, NULL);
2885                         ret = VM_FAULT_SIGBUS;
2886                 }
2887                 goto out;
2888         }
2889         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2890         page = lookup_swap_cache(entry);
2891         if (!page) {
2892                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2893                 page = swapin_readahead(entry,
2894                                         GFP_HIGHUSER_MOVABLE, vma, address);
2895                 if (!page) {
2896                         /*
2897                          * Back out if somebody else faulted in this pte
2898                          * while we released the pte lock.
2899                          */
2900                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2901                         if (likely(pte_same(*page_table, orig_pte)))
2902                                 ret = VM_FAULT_OOM;
2903                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2904                         goto unlock;
2905                 }
2906
2907                 /* Had to read the page from swap area: Major fault */
2908                 ret = VM_FAULT_MAJOR;
2909                 count_vm_event(PGMAJFAULT);
2910                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2911         } else if (PageHWPoison(page)) {
2912                 /*
2913                  * hwpoisoned dirty swapcache pages are kept for killing
2914                  * owner processes (which may be unknown at hwpoison time)
2915                  */
2916                 ret = VM_FAULT_HWPOISON;
2917                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2918                 goto out_release;
2919         }
2920
2921         locked = lock_page_or_retry(page, mm, flags);
2922         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2923         if (!locked) {
2924                 ret |= VM_FAULT_RETRY;
2925                 goto out_release;
2926         }
2927
2928         /*
2929          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2930          * release the swapcache from under us.  The page pin, and pte_same
2931          * test below, are not enough to exclude that.  Even if it is still
2932          * swapcache, we need to check that the page's swap has not changed.
2933          */
2934         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2935                 goto out_page;
2936
2937         if (ksm_might_need_to_copy(page, vma, address)) {
2938                 swapcache = page;
2939                 page = ksm_does_need_to_copy(page, vma, address);
2940
2941                 if (unlikely(!page)) {
2942                         ret = VM_FAULT_OOM;
2943                         page = swapcache;
2944                         swapcache = NULL;
2945                         goto out_page;
2946                 }
2947         }
2948
2949         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2950                 ret = VM_FAULT_OOM;
2951                 goto out_page;
2952         }
2953
2954         /*
2955          * Back out if somebody else already faulted in this pte.
2956          */
2957         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2958         if (unlikely(!pte_same(*page_table, orig_pte)))
2959                 goto out_nomap;
2960
2961         if (unlikely(!PageUptodate(page))) {
2962                 ret = VM_FAULT_SIGBUS;
2963                 goto out_nomap;
2964         }
2965
2966         /*
2967          * The page isn't present yet, go ahead with the fault.
2968          *
2969          * Be careful about the sequence of operations here.
2970          * To get its accounting right, reuse_swap_page() must be called
2971          * while the page is counted on swap but not yet in mapcount i.e.
2972          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2973          * must be called after the swap_free(), or it will never succeed.
2974          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2975          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2976          * in page->private. In this case, a record in swap_cgroup  is silently
2977          * discarded at swap_free().
2978          */
2979
2980         inc_mm_counter_fast(mm, MM_ANONPAGES);
2981         dec_mm_counter_fast(mm, MM_SWAPENTS);
2982         pte = mk_pte(page, vma->vm_page_prot);
2983         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2984                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2985                 flags &= ~FAULT_FLAG_WRITE;
2986                 ret |= VM_FAULT_WRITE;
2987                 exclusive = 1;
2988         }
2989         flush_icache_page(vma, page);
2990         set_pte_at(mm, address, page_table, pte);
2991         do_page_add_anon_rmap(page, vma, address, exclusive);
2992         /* It's better to call commit-charge after rmap is established */
2993         mem_cgroup_commit_charge_swapin(page, ptr);
2994
2995         swap_free(entry);
2996         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2997                 try_to_free_swap(page);
2998         unlock_page(page);
2999         if (swapcache) {
3000                 /*
3001                  * Hold the lock to avoid the swap entry to be reused
3002                  * until we take the PT lock for the pte_same() check
3003                  * (to avoid false positives from pte_same). For
3004                  * further safety release the lock after the swap_free
3005                  * so that the swap count won't change under a
3006                  * parallel locked swapcache.
3007                  */
3008                 unlock_page(swapcache);
3009                 page_cache_release(swapcache);
3010         }
3011
3012         if (flags & FAULT_FLAG_WRITE) {
3013                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3014                 if (ret & VM_FAULT_ERROR)
3015                         ret &= VM_FAULT_ERROR;
3016                 goto out;
3017         }
3018
3019         /* No need to invalidate - it was non-present before */
3020         update_mmu_cache(vma, address, page_table);
3021 unlock:
3022         pte_unmap_unlock(page_table, ptl);
3023 out:
3024         return ret;
3025 out_nomap:
3026         mem_cgroup_cancel_charge_swapin(ptr);
3027         pte_unmap_unlock(page_table, ptl);
3028 out_page:
3029         unlock_page(page);
3030 out_release:
3031         page_cache_release(page);
3032         if (swapcache) {
3033                 unlock_page(swapcache);
3034                 page_cache_release(swapcache);
3035         }
3036         return ret;
3037 }
3038
3039 /*
3040  * This is like a special single-page "expand_{down|up}wards()",
3041  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3042  * doesn't hit another vma.
3043  */
3044 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3045 {
3046         address &= PAGE_MASK;
3047         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3048                 struct vm_area_struct *prev = vma->vm_prev;
3049
3050                 /*
3051                  * Is there a mapping abutting this one below?
3052                  *
3053                  * That's only ok if it's the same stack mapping
3054                  * that has gotten split..
3055                  */
3056                 if (prev && prev->vm_end == address)
3057                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3058
3059                 expand_downwards(vma, address - PAGE_SIZE);
3060         }
3061         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3062                 struct vm_area_struct *next = vma->vm_next;
3063
3064                 /* As VM_GROWSDOWN but s/below/above/ */
3065                 if (next && next->vm_start == address + PAGE_SIZE)
3066                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3067
3068                 expand_upwards(vma, address + PAGE_SIZE);
3069         }
3070         return 0;
3071 }
3072
3073 /*
3074  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3075  * but allow concurrent faults), and pte mapped but not yet locked.
3076  * We return with mmap_sem still held, but pte unmapped and unlocked.
3077  */
3078 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3079                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3080                 unsigned int flags)
3081 {
3082         struct page *page;
3083         spinlock_t *ptl;
3084         pte_t entry;
3085
3086         pte_unmap(page_table);
3087
3088         /* Check if we need to add a guard page to the stack */
3089         if (check_stack_guard_page(vma, address) < 0)
3090                 return VM_FAULT_SIGBUS;
3091
3092         /* Use the zero-page for reads */
3093         if (!(flags & FAULT_FLAG_WRITE)) {
3094                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3095                                                 vma->vm_page_prot));
3096                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3097                 if (!pte_none(*page_table))
3098                         goto unlock;
3099                 goto setpte;
3100         }
3101
3102         /* Allocate our own private page. */
3103         if (unlikely(anon_vma_prepare(vma)))
3104                 goto oom;
3105         page = alloc_zeroed_user_highpage_movable(vma, address);
3106         if (!page)
3107                 goto oom;
3108         __SetPageUptodate(page);
3109
3110         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3111                 goto oom_free_page;
3112
3113         entry = mk_pte(page, vma->vm_page_prot);
3114         if (vma->vm_flags & VM_WRITE)
3115                 entry = pte_mkwrite(pte_mkdirty(entry));
3116
3117         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3118         if (!pte_none(*page_table))
3119                 goto release;
3120
3121         inc_mm_counter_fast(mm, MM_ANONPAGES);
3122         page_add_new_anon_rmap(page, vma, address);
3123 setpte:
3124         set_pte_at(mm, address, page_table, entry);
3125
3126         /* No need to invalidate - it was non-present before */
3127         update_mmu_cache(vma, address, page_table);
3128 unlock:
3129         pte_unmap_unlock(page_table, ptl);
3130         return 0;
3131 release:
3132         mem_cgroup_uncharge_page(page);
3133         page_cache_release(page);
3134         goto unlock;
3135 oom_free_page:
3136         page_cache_release(page);
3137 oom:
3138         return VM_FAULT_OOM;
3139 }
3140
3141 /*
3142  * __do_fault() tries to create a new page mapping. It aggressively
3143  * tries to share with existing pages, but makes a separate copy if
3144  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3145  * the next page fault.
3146  *
3147  * As this is called only for pages that do not currently exist, we
3148  * do not need to flush old virtual caches or the TLB.
3149  *
3150  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3151  * but allow concurrent faults), and pte neither mapped nor locked.
3152  * We return with mmap_sem still held, but pte unmapped and unlocked.
3153  */
3154 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3155                 unsigned long address, pmd_t *pmd,
3156                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3157 {
3158         pte_t *page_table;
3159         spinlock_t *ptl;
3160         struct page *page;
3161         struct page *cow_page;
3162         pte_t entry;
3163         int anon = 0;
3164         struct page *dirty_page = NULL;
3165         struct vm_fault vmf;
3166         int ret;
3167         int page_mkwrite = 0;
3168
3169         /*
3170          * If we do COW later, allocate page befor taking lock_page()
3171          * on the file cache page. This will reduce lock holding time.
3172          */
3173         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3174
3175                 if (unlikely(anon_vma_prepare(vma)))
3176                         return VM_FAULT_OOM;
3177
3178                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3179                 if (!cow_page)
3180                         return VM_FAULT_OOM;
3181
3182                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3183                         page_cache_release(cow_page);
3184                         return VM_FAULT_OOM;
3185                 }
3186         } else
3187                 cow_page = NULL;
3188
3189         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3190         vmf.pgoff = pgoff;
3191         vmf.flags = flags;
3192         vmf.page = NULL;
3193
3194         ret = vma->vm_ops->fault(vma, &vmf);
3195         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3196                             VM_FAULT_RETRY)))
3197                 goto uncharge_out;
3198
3199         if (unlikely(PageHWPoison(vmf.page))) {
3200                 if (ret & VM_FAULT_LOCKED)
3201                         unlock_page(vmf.page);
3202                 ret = VM_FAULT_HWPOISON;
3203                 goto uncharge_out;
3204         }
3205
3206         /*
3207          * For consistency in subsequent calls, make the faulted page always
3208          * locked.
3209          */
3210         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3211                 lock_page(vmf.page);
3212         else
3213                 VM_BUG_ON(!PageLocked(vmf.page));
3214
3215         /*
3216          * Should we do an early C-O-W break?
3217          */
3218         page = vmf.page;
3219         if (flags & FAULT_FLAG_WRITE) {
3220                 if (!(vma->vm_flags & VM_SHARED)) {
3221                         page = cow_page;
3222                         anon = 1;
3223                         copy_user_highpage(page, vmf.page, address, vma);
3224                         __SetPageUptodate(page);
3225                 } else {
3226                         /*
3227                          * If the page will be shareable, see if the backing
3228                          * address space wants to know that the page is about
3229                          * to become writable
3230                          */
3231                         if (vma->vm_ops->page_mkwrite) {
3232                                 int tmp;
3233
3234                                 unlock_page(page);
3235                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3236                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3237                                 if (unlikely(tmp &
3238                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3239                                         ret = tmp;
3240                                         goto unwritable_page;
3241                                 }
3242                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3243                                         lock_page(page);
3244                                         if (!page->mapping) {
3245                                                 ret = 0; /* retry the fault */
3246                                                 unlock_page(page);
3247                                                 goto unwritable_page;
3248                                         }
3249                                 } else
3250                                         VM_BUG_ON(!PageLocked(page));
3251                                 page_mkwrite = 1;
3252                         }
3253                 }
3254
3255         }
3256
3257         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3258
3259         /*
3260          * This silly early PAGE_DIRTY setting removes a race
3261          * due to the bad i386 page protection. But it's valid
3262          * for other architectures too.
3263          *
3264          * Note that if FAULT_FLAG_WRITE is set, we either now have
3265          * an exclusive copy of the page, or this is a shared mapping,
3266          * so we can make it writable and dirty to avoid having to
3267          * handle that later.
3268          */
3269         /* Only go through if we didn't race with anybody else... */
3270         if (likely(pte_same(*page_table, orig_pte))) {
3271                 flush_icache_page(vma, page);
3272                 entry = mk_pte(page, vma->vm_page_prot);
3273                 if (flags & FAULT_FLAG_WRITE)
3274                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3275                 if (anon) {
3276                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3277                         page_add_new_anon_rmap(page, vma, address);
3278                 } else {
3279                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3280                         page_add_file_rmap(page);
3281                         if (flags & FAULT_FLAG_WRITE) {
3282                                 dirty_page = page;
3283                                 get_page(dirty_page);
3284                         }
3285                 }
3286                 set_pte_at(mm, address, page_table, entry);
3287
3288                 /* no need to invalidate: a not-present page won't be cached */
3289                 update_mmu_cache(vma, address, page_table);
3290         } else {
3291                 if (cow_page)
3292                         mem_cgroup_uncharge_page(cow_page);
3293                 if (anon)
3294                         page_cache_release(page);
3295                 else
3296                         anon = 1; /* no anon but release faulted_page */
3297         }
3298
3299         pte_unmap_unlock(page_table, ptl);
3300
3301         if (dirty_page) {
3302                 struct address_space *mapping = page->mapping;
3303
3304                 if (set_page_dirty(dirty_page))
3305                         page_mkwrite = 1;
3306                 unlock_page(dirty_page);
3307                 put_page(dirty_page);
3308                 if (page_mkwrite && mapping) {
3309                         /*
3310                          * Some device drivers do not set page.mapping but still
3311                          * dirty their pages
3312                          */
3313                         balance_dirty_pages_ratelimited(mapping);
3314                 }
3315
3316                 /* file_update_time outside page_lock */
3317                 if (vma->vm_file)
3318                         file_update_time(vma->vm_file);
3319         } else {
3320                 unlock_page(vmf.page);
3321                 if (anon)
3322                         page_cache_release(vmf.page);
3323         }
3324
3325         return ret;
3326
3327 unwritable_page:
3328         page_cache_release(page);
3329         return ret;
3330 uncharge_out:
3331         /* fs's fault handler get error */
3332         if (cow_page) {
3333                 mem_cgroup_uncharge_page(cow_page);
3334                 page_cache_release(cow_page);
3335         }
3336         return ret;
3337 }
3338
3339 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3340                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3341                 unsigned int flags, pte_t orig_pte)
3342 {
3343         pgoff_t pgoff = (((address & PAGE_MASK)
3344                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3345
3346         pte_unmap(page_table);
3347         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3348 }
3349
3350 /*
3351  * Fault of a previously existing named mapping. Repopulate the pte
3352  * from the encoded file_pte if possible. This enables swappable
3353  * nonlinear vmas.
3354  *
3355  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3356  * but allow concurrent faults), and pte mapped but not yet locked.
3357  * We return with mmap_sem still held, but pte unmapped and unlocked.
3358  */
3359 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3360                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3361                 unsigned int flags, pte_t orig_pte)
3362 {
3363         pgoff_t pgoff;
3364
3365         flags |= FAULT_FLAG_NONLINEAR;
3366
3367         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3368                 return 0;
3369
3370         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3371                 /*
3372                  * Page table corrupted: show pte and kill process.
3373                  */
3374                 print_bad_pte(vma, address, orig_pte, NULL);
3375                 return VM_FAULT_SIGBUS;
3376         }
3377
3378         pgoff = pte_to_pgoff(orig_pte);
3379         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3380 }
3381
3382 /*
3383  * These routines also need to handle stuff like marking pages dirty
3384  * and/or accessed for architectures that don't do it in hardware (most
3385  * RISC architectures).  The early dirtying is also good on the i386.
3386  *
3387  * There is also a hook called "update_mmu_cache()" that architectures
3388  * with external mmu caches can use to update those (ie the Sparc or
3389  * PowerPC hashed page tables that act as extended TLBs).
3390  *
3391  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3392  * but allow concurrent faults), and pte mapped but not yet locked.
3393  * We return with mmap_sem still held, but pte unmapped and unlocked.
3394  */
3395 int handle_pte_fault(struct mm_struct *mm,
3396                      struct vm_area_struct *vma, unsigned long address,
3397                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3398 {
3399         pte_t entry;
3400         spinlock_t *ptl;
3401
3402         entry = *pte;
3403         if (!pte_present(entry)) {
3404                 if (pte_none(entry)) {
3405                         if (vma->vm_ops) {
3406                                 if (likely(vma->vm_ops->fault))
3407                                         return do_linear_fault(mm, vma, address,
3408                                                 pte, pmd, flags, entry);
3409                         }
3410                         return do_anonymous_page(mm, vma, address,
3411                                                  pte, pmd, flags);
3412                 }
3413                 if (pte_file(entry))
3414                         return do_nonlinear_fault(mm, vma, address,
3415                                         pte, pmd, flags, entry);
3416                 return do_swap_page(mm, vma, address,
3417                                         pte, pmd, flags, entry);
3418         }
3419
3420         ptl = pte_lockptr(mm, pmd);
3421         spin_lock(ptl);
3422         if (unlikely(!pte_same(*pte, entry)))
3423                 goto unlock;
3424         if (flags & FAULT_FLAG_WRITE) {
3425                 if (!pte_write(entry))
3426                         return do_wp_page(mm, vma, address,
3427                                         pte, pmd, ptl, entry);
3428                 entry = pte_mkdirty(entry);
3429         }
3430         entry = pte_mkyoung(entry);
3431         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3432                 update_mmu_cache(vma, address, pte);
3433         } else {
3434                 /*
3435                  * This is needed only for protection faults but the arch code
3436                  * is not yet telling us if this is a protection fault or not.
3437                  * This still avoids useless tlb flushes for .text page faults
3438                  * with threads.
3439                  */
3440                 if (flags & FAULT_FLAG_WRITE)
3441                         flush_tlb_fix_spurious_fault(vma, address);
3442         }
3443 unlock:
3444         pte_unmap_unlock(pte, ptl);
3445         return 0;
3446 }
3447
3448 /*
3449  * By the time we get here, we already hold the mm semaphore
3450  */
3451 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3452                 unsigned long address, unsigned int flags)
3453 {
3454         pgd_t *pgd;
3455         pud_t *pud;
3456         pmd_t *pmd;
3457         pte_t *pte;
3458
3459         __set_current_state(TASK_RUNNING);
3460
3461         count_vm_event(PGFAULT);
3462         mem_cgroup_count_vm_event(mm, PGFAULT);
3463
3464         /* do counter updates before entering really critical section. */
3465         check_sync_rss_stat(current);
3466
3467         if (unlikely(is_vm_hugetlb_page(vma)))
3468                 return hugetlb_fault(mm, vma, address, flags);
3469
3470         pgd = pgd_offset(mm, address);
3471         pud = pud_alloc(mm, pgd, address);
3472         if (!pud)
3473                 return VM_FAULT_OOM;
3474         pmd = pmd_alloc(mm, pud, address);
3475         if (!pmd)
3476                 return VM_FAULT_OOM;
3477         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3478                 if (!vma->vm_ops)
3479                         return do_huge_pmd_anonymous_page(mm, vma, address,
3480                                                           pmd, flags);
3481         } else {
3482                 pmd_t orig_pmd = *pmd;
3483                 barrier();
3484                 if (pmd_trans_huge(orig_pmd)) {
3485                         if (flags & FAULT_FLAG_WRITE &&
3486                             !pmd_write(orig_pmd) &&
3487                             !pmd_trans_splitting(orig_pmd))
3488                                 return do_huge_pmd_wp_page(mm, vma, address,
3489                                                            pmd, orig_pmd);
3490                         return 0;
3491                 }
3492         }
3493
3494         /*
3495          * Use __pte_alloc instead of pte_alloc_map, because we can't
3496          * run pte_offset_map on the pmd, if an huge pmd could
3497          * materialize from under us from a different thread.
3498          */
3499         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3500                 return VM_FAULT_OOM;
3501         /* if an huge pmd materialized from under us just retry later */
3502         if (unlikely(pmd_trans_huge(*pmd)))
3503                 return 0;
3504         /*
3505          * A regular pmd is established and it can't morph into a huge pmd
3506          * from under us anymore at this point because we hold the mmap_sem
3507          * read mode and khugepaged takes it in write mode. So now it's
3508          * safe to run pte_offset_map().
3509          */
3510         pte = pte_offset_map(pmd, address);
3511
3512         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3513 }
3514
3515 #ifndef __PAGETABLE_PUD_FOLDED
3516 /*
3517  * Allocate page upper directory.
3518  * We've already handled the fast-path in-line.
3519  */
3520 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3521 {
3522         pud_t *new = pud_alloc_one(mm, address);
3523         if (!new)
3524                 return -ENOMEM;
3525
3526         smp_wmb(); /* See comment in __pte_alloc */
3527
3528         spin_lock(&mm->page_table_lock);
3529         if (pgd_present(*pgd))          /* Another has populated it */
3530                 pud_free(mm, new);
3531         else
3532                 pgd_populate(mm, pgd, new);
3533         spin_unlock(&mm->page_table_lock);
3534         return 0;
3535 }
3536 #endif /* __PAGETABLE_PUD_FOLDED */
3537
3538 #ifndef __PAGETABLE_PMD_FOLDED
3539 /*
3540  * Allocate page middle directory.
3541  * We've already handled the fast-path in-line.
3542  */
3543 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3544 {
3545         pmd_t *new = pmd_alloc_one(mm, address);
3546         if (!new)
3547                 return -ENOMEM;
3548
3549         smp_wmb(); /* See comment in __pte_alloc */
3550
3551         spin_lock(&mm->page_table_lock);
3552 #ifndef __ARCH_HAS_4LEVEL_HACK
3553         if (pud_present(*pud))          /* Another has populated it */
3554                 pmd_free(mm, new);
3555         else
3556                 pud_populate(mm, pud, new);
3557 #else
3558         if (pgd_present(*pud))          /* Another has populated it */
3559                 pmd_free(mm, new);
3560         else
3561                 pgd_populate(mm, pud, new);
3562 #endif /* __ARCH_HAS_4LEVEL_HACK */
3563         spin_unlock(&mm->page_table_lock);
3564         return 0;
3565 }
3566 #endif /* __PAGETABLE_PMD_FOLDED */
3567
3568 int make_pages_present(unsigned long addr, unsigned long end)
3569 {
3570         int ret, len, write;
3571         struct vm_area_struct * vma;
3572
3573         vma = find_vma(current->mm, addr);
3574         if (!vma)
3575                 return -ENOMEM;
3576         /*
3577          * We want to touch writable mappings with a write fault in order
3578          * to break COW, except for shared mappings because these don't COW
3579          * and we would not want to dirty them for nothing.
3580          */
3581         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3582         BUG_ON(addr >= end);
3583         BUG_ON(end > vma->vm_end);
3584         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3585         ret = get_user_pages(current, current->mm, addr,
3586                         len, write, 0, NULL, NULL);
3587         if (ret < 0)
3588                 return ret;
3589         return ret == len ? 0 : -EFAULT;
3590 }
3591
3592 #if !defined(__HAVE_ARCH_GATE_AREA)
3593
3594 #if defined(AT_SYSINFO_EHDR)
3595 static struct vm_area_struct gate_vma;
3596
3597 static int __init gate_vma_init(void)
3598 {
3599         gate_vma.vm_mm = NULL;
3600         gate_vma.vm_start = FIXADDR_USER_START;
3601         gate_vma.vm_end = FIXADDR_USER_END;
3602         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3603         gate_vma.vm_page_prot = __P101;
3604         /*
3605          * Make sure the vDSO gets into every core dump.
3606          * Dumping its contents makes post-mortem fully interpretable later
3607          * without matching up the same kernel and hardware config to see
3608          * what PC values meant.
3609          */
3610         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3611         return 0;
3612 }
3613 __initcall(gate_vma_init);
3614 #endif
3615
3616 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3617 {
3618 #ifdef AT_SYSINFO_EHDR
3619         return &gate_vma;
3620 #else
3621         return NULL;
3622 #endif
3623 }
3624
3625 int in_gate_area_no_mm(unsigned long addr)
3626 {
3627 #ifdef AT_SYSINFO_EHDR
3628         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3629                 return 1;
3630 #endif
3631         return 0;
3632 }
3633
3634 #endif  /* __HAVE_ARCH_GATE_AREA */
3635
3636 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3637                 pte_t **ptepp, spinlock_t **ptlp)
3638 {
3639         pgd_t *pgd;
3640         pud_t *pud;
3641         pmd_t *pmd;
3642         pte_t *ptep;
3643
3644         pgd = pgd_offset(mm, address);
3645         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3646                 goto out;
3647
3648         pud = pud_offset(pgd, address);
3649         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3650                 goto out;
3651
3652         pmd = pmd_offset(pud, address);
3653         VM_BUG_ON(pmd_trans_huge(*pmd));
3654         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3655                 goto out;
3656
3657         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3658         if (pmd_huge(*pmd))
3659                 goto out;
3660
3661         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3662         if (!ptep)
3663                 goto out;
3664         if (!pte_present(*ptep))
3665                 goto unlock;
3666         *ptepp = ptep;
3667         return 0;
3668 unlock:
3669         pte_unmap_unlock(ptep, *ptlp);
3670 out:
3671         return -EINVAL;
3672 }
3673
3674 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3675                              pte_t **ptepp, spinlock_t **ptlp)
3676 {
3677         int res;
3678
3679         /* (void) is needed to make gcc happy */
3680         (void) __cond_lock(*ptlp,
3681                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3682         return res;
3683 }
3684
3685 /**
3686  * follow_pfn - look up PFN at a user virtual address
3687  * @vma: memory mapping
3688  * @address: user virtual address
3689  * @pfn: location to store found PFN
3690  *
3691  * Only IO mappings and raw PFN mappings are allowed.
3692  *
3693  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3694  */
3695 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3696         unsigned long *pfn)
3697 {
3698         int ret = -EINVAL;
3699         spinlock_t *ptl;
3700         pte_t *ptep;
3701
3702         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3703                 return ret;
3704
3705         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3706         if (ret)
3707                 return ret;
3708         *pfn = pte_pfn(*ptep);
3709         pte_unmap_unlock(ptep, ptl);
3710         return 0;
3711 }
3712 EXPORT_SYMBOL(follow_pfn);
3713
3714 #ifdef CONFIG_HAVE_IOREMAP_PROT
3715 int follow_phys(struct vm_area_struct *vma,
3716                 unsigned long address, unsigned int flags,
3717                 unsigned long *prot, resource_size_t *phys)
3718 {
3719         int ret = -EINVAL;
3720         pte_t *ptep, pte;
3721         spinlock_t *ptl;
3722
3723         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3724                 goto out;
3725
3726         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3727                 goto out;
3728         pte = *ptep;
3729
3730         if ((flags & FOLL_WRITE) && !pte_write(pte))
3731                 goto unlock;
3732
3733         *prot = pgprot_val(pte_pgprot(pte));
3734         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3735
3736         ret = 0;
3737 unlock:
3738         pte_unmap_unlock(ptep, ptl);
3739 out:
3740         return ret;
3741 }
3742
3743 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3744                         void *buf, int len, int write)
3745 {
3746         resource_size_t phys_addr;
3747         unsigned long prot = 0;
3748         void __iomem *maddr;
3749         int offset = addr & (PAGE_SIZE-1);
3750
3751         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3752                 return -EINVAL;
3753
3754         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3755         if (write)
3756                 memcpy_toio(maddr + offset, buf, len);
3757         else
3758                 memcpy_fromio(buf, maddr + offset, len);
3759         iounmap(maddr);
3760
3761         return len;
3762 }
3763 #endif
3764
3765 /*
3766  * Access another process' address space as given in mm.  If non-NULL, use the
3767  * given task for page fault accounting.
3768  */
3769 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3770                 unsigned long addr, void *buf, int len, int write)
3771 {
3772         struct vm_area_struct *vma;
3773         void *old_buf = buf;
3774
3775         down_read(&mm->mmap_sem);
3776         /* ignore errors, just check how much was successfully transferred */
3777         while (len) {
3778                 int bytes, ret, offset;
3779                 void *maddr;
3780                 struct page *page = NULL;
3781
3782                 ret = get_user_pages(tsk, mm, addr, 1,
3783                                 write, 1, &page, &vma);
3784                 if (ret <= 0) {
3785                         /*
3786                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3787                          * we can access using slightly different code.
3788                          */
3789 #ifdef CONFIG_HAVE_IOREMAP_PROT
3790                         vma = find_vma(mm, addr);
3791                         if (!vma || vma->vm_start > addr)
3792                                 break;
3793                         if (vma->vm_ops && vma->vm_ops->access)
3794                                 ret = vma->vm_ops->access(vma, addr, buf,
3795                                                           len, write);
3796                         if (ret <= 0)
3797 #endif
3798                                 break;
3799                         bytes = ret;
3800                 } else {
3801                         bytes = len;
3802                         offset = addr & (PAGE_SIZE-1);
3803                         if (bytes > PAGE_SIZE-offset)
3804                                 bytes = PAGE_SIZE-offset;
3805
3806                         maddr = kmap(page);
3807                         if (write) {
3808                                 copy_to_user_page(vma, page, addr,
3809                                                   maddr + offset, buf, bytes);
3810                                 set_page_dirty_lock(page);
3811                         } else {
3812                                 copy_from_user_page(vma, page, addr,
3813                                                     buf, maddr + offset, bytes);
3814                         }
3815                         kunmap(page);
3816                         page_cache_release(page);
3817                 }
3818                 len -= bytes;
3819                 buf += bytes;
3820                 addr += bytes;
3821         }
3822         up_read(&mm->mmap_sem);
3823
3824         return buf - old_buf;
3825 }
3826
3827 /**
3828  * access_remote_vm - access another process' address space
3829  * @mm:         the mm_struct of the target address space
3830  * @addr:       start address to access
3831  * @buf:        source or destination buffer
3832  * @len:        number of bytes to transfer
3833  * @write:      whether the access is a write
3834  *
3835  * The caller must hold a reference on @mm.
3836  */
3837 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3838                 void *buf, int len, int write)
3839 {
3840         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3841 }
3842
3843 /*
3844  * Access another process' address space.
3845  * Source/target buffer must be kernel space,
3846  * Do not walk the page table directly, use get_user_pages
3847  */
3848 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3849                 void *buf, int len, int write)
3850 {
3851         struct mm_struct *mm;
3852         int ret;
3853
3854         mm = get_task_mm(tsk);
3855         if (!mm)
3856                 return 0;
3857
3858         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3859         mmput(mm);
3860
3861         return ret;
3862 }
3863
3864 /*
3865  * Print the name of a VMA.
3866  */
3867 void print_vma_addr(char *prefix, unsigned long ip)
3868 {
3869         struct mm_struct *mm = current->mm;
3870         struct vm_area_struct *vma;
3871
3872         /*
3873          * Do not print if we are in atomic
3874          * contexts (in exception stacks, etc.):
3875          */
3876         if (preempt_count())
3877                 return;
3878
3879         down_read(&mm->mmap_sem);
3880         vma = find_vma(mm, ip);
3881         if (vma && vma->vm_file) {
3882                 struct file *f = vma->vm_file;
3883                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3884                 if (buf) {
3885                         char *p, *s;
3886
3887                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3888                         if (IS_ERR(p))
3889                                 p = "?";
3890                         s = strrchr(p, '/');
3891                         if (s)
3892                                 p = s+1;
3893                         printk("%s%s[%lx+%lx]", prefix, p,
3894                                         vma->vm_start,
3895                                         vma->vm_end - vma->vm_start);
3896                         free_page((unsigned long)buf);
3897                 }
3898         }
3899         up_read(&current->mm->mmap_sem);
3900 }
3901
3902 #ifdef CONFIG_PROVE_LOCKING
3903 void might_fault(void)
3904 {
3905         /*
3906          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3907          * holding the mmap_sem, this is safe because kernel memory doesn't
3908          * get paged out, therefore we'll never actually fault, and the
3909          * below annotations will generate false positives.
3910          */
3911         if (segment_eq(get_fs(), KERNEL_DS))
3912                 return;
3913
3914         might_sleep();
3915         /*
3916          * it would be nicer only to annotate paths which are not under
3917          * pagefault_disable, however that requires a larger audit and
3918          * providing helpers like get_user_atomic.
3919          */
3920         if (!in_atomic() && current->mm)
3921                 might_lock_read(&current->mm->mmap_sem);
3922 }
3923 EXPORT_SYMBOL(might_fault);
3924 #endif
3925
3926 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3927 static void clear_gigantic_page(struct page *page,
3928                                 unsigned long addr,
3929                                 unsigned int pages_per_huge_page)
3930 {
3931         int i;
3932         struct page *p = page;
3933
3934         might_sleep();
3935         for (i = 0; i < pages_per_huge_page;
3936              i++, p = mem_map_next(p, page, i)) {
3937                 cond_resched();
3938                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3939         }
3940 }
3941 void clear_huge_page(struct page *page,
3942                      unsigned long addr, unsigned int pages_per_huge_page)
3943 {
3944         int i;
3945
3946         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3947                 clear_gigantic_page(page, addr, pages_per_huge_page);
3948                 return;
3949         }
3950
3951         might_sleep();
3952         for (i = 0; i < pages_per_huge_page; i++) {
3953                 cond_resched();
3954                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3955         }
3956 }
3957
3958 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3959                                     unsigned long addr,
3960                                     struct vm_area_struct *vma,
3961                                     unsigned int pages_per_huge_page)
3962 {
3963         int i;
3964         struct page *dst_base = dst;
3965         struct page *src_base = src;
3966
3967         for (i = 0; i < pages_per_huge_page; ) {
3968                 cond_resched();
3969                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3970
3971                 i++;
3972                 dst = mem_map_next(dst, dst_base, i);
3973                 src = mem_map_next(src, src_base, i);
3974         }
3975 }
3976
3977 void copy_user_huge_page(struct page *dst, struct page *src,
3978                          unsigned long addr, struct vm_area_struct *vma,
3979                          unsigned int pages_per_huge_page)
3980 {
3981         int i;
3982
3983         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3984                 copy_user_gigantic_page(dst, src, addr, vma,
3985                                         pages_per_huge_page);
3986                 return;
3987         }
3988
3989         might_sleep();
3990         for (i = 0; i < pages_per_huge_page; i++) {
3991                 cond_resched();
3992                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3993         }
3994 }
3995 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */