Merge tag 'balancenuma-v11' of git://git.kernel.org/pub/scm/linux/kernel/git/mel...
[~shefty/rdma-dev.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61
62 #include <asm/io.h>
63 #include <asm/pgalloc.h>
64 #include <asm/uaccess.h>
65 #include <asm/tlb.h>
66 #include <asm/tlbflush.h>
67 #include <asm/pgtable.h>
68
69 #include "internal.h"
70
71 #ifndef CONFIG_NEED_MULTIPLE_NODES
72 /* use the per-pgdat data instead for discontigmem - mbligh */
73 unsigned long max_mapnr;
74 struct page *mem_map;
75
76 EXPORT_SYMBOL(max_mapnr);
77 EXPORT_SYMBOL(mem_map);
78 #endif
79
80 unsigned long num_physpages;
81 /*
82  * A number of key systems in x86 including ioremap() rely on the assumption
83  * that high_memory defines the upper bound on direct map memory, then end
84  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
85  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
86  * and ZONE_HIGHMEM.
87  */
88 void * high_memory;
89
90 EXPORT_SYMBOL(num_physpages);
91 EXPORT_SYMBOL(high_memory);
92
93 /*
94  * Randomize the address space (stacks, mmaps, brk, etc.).
95  *
96  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
97  *   as ancient (libc5 based) binaries can segfault. )
98  */
99 int randomize_va_space __read_mostly =
100 #ifdef CONFIG_COMPAT_BRK
101                                         1;
102 #else
103                                         2;
104 #endif
105
106 static int __init disable_randmaps(char *s)
107 {
108         randomize_va_space = 0;
109         return 1;
110 }
111 __setup("norandmaps", disable_randmaps);
112
113 unsigned long zero_pfn __read_mostly;
114 unsigned long highest_memmap_pfn __read_mostly;
115
116 /*
117  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
118  */
119 static int __init init_zero_pfn(void)
120 {
121         zero_pfn = page_to_pfn(ZERO_PAGE(0));
122         return 0;
123 }
124 core_initcall(init_zero_pfn);
125
126
127 #if defined(SPLIT_RSS_COUNTING)
128
129 void sync_mm_rss(struct mm_struct *mm)
130 {
131         int i;
132
133         for (i = 0; i < NR_MM_COUNTERS; i++) {
134                 if (current->rss_stat.count[i]) {
135                         add_mm_counter(mm, i, current->rss_stat.count[i]);
136                         current->rss_stat.count[i] = 0;
137                 }
138         }
139         current->rss_stat.events = 0;
140 }
141
142 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
143 {
144         struct task_struct *task = current;
145
146         if (likely(task->mm == mm))
147                 task->rss_stat.count[member] += val;
148         else
149                 add_mm_counter(mm, member, val);
150 }
151 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
152 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
153
154 /* sync counter once per 64 page faults */
155 #define TASK_RSS_EVENTS_THRESH  (64)
156 static void check_sync_rss_stat(struct task_struct *task)
157 {
158         if (unlikely(task != current))
159                 return;
160         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
161                 sync_mm_rss(task->mm);
162 }
163 #else /* SPLIT_RSS_COUNTING */
164
165 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
166 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
167
168 static void check_sync_rss_stat(struct task_struct *task)
169 {
170 }
171
172 #endif /* SPLIT_RSS_COUNTING */
173
174 #ifdef HAVE_GENERIC_MMU_GATHER
175
176 static int tlb_next_batch(struct mmu_gather *tlb)
177 {
178         struct mmu_gather_batch *batch;
179
180         batch = tlb->active;
181         if (batch->next) {
182                 tlb->active = batch->next;
183                 return 1;
184         }
185
186         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
187         if (!batch)
188                 return 0;
189
190         batch->next = NULL;
191         batch->nr   = 0;
192         batch->max  = MAX_GATHER_BATCH;
193
194         tlb->active->next = batch;
195         tlb->active = batch;
196
197         return 1;
198 }
199
200 /* tlb_gather_mmu
201  *      Called to initialize an (on-stack) mmu_gather structure for page-table
202  *      tear-down from @mm. The @fullmm argument is used when @mm is without
203  *      users and we're going to destroy the full address space (exit/execve).
204  */
205 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
206 {
207         tlb->mm = mm;
208
209         tlb->fullmm     = fullmm;
210         tlb->start      = -1UL;
211         tlb->end        = 0;
212         tlb->need_flush = 0;
213         tlb->fast_mode  = (num_possible_cpus() == 1);
214         tlb->local.next = NULL;
215         tlb->local.nr   = 0;
216         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
217         tlb->active     = &tlb->local;
218
219 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
220         tlb->batch = NULL;
221 #endif
222 }
223
224 void tlb_flush_mmu(struct mmu_gather *tlb)
225 {
226         struct mmu_gather_batch *batch;
227
228         if (!tlb->need_flush)
229                 return;
230         tlb->need_flush = 0;
231         tlb_flush(tlb);
232 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
233         tlb_table_flush(tlb);
234 #endif
235
236         if (tlb_fast_mode(tlb))
237                 return;
238
239         for (batch = &tlb->local; batch; batch = batch->next) {
240                 free_pages_and_swap_cache(batch->pages, batch->nr);
241                 batch->nr = 0;
242         }
243         tlb->active = &tlb->local;
244 }
245
246 /* tlb_finish_mmu
247  *      Called at the end of the shootdown operation to free up any resources
248  *      that were required.
249  */
250 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
251 {
252         struct mmu_gather_batch *batch, *next;
253
254         tlb->start = start;
255         tlb->end   = end;
256         tlb_flush_mmu(tlb);
257
258         /* keep the page table cache within bounds */
259         check_pgt_cache();
260
261         for (batch = tlb->local.next; batch; batch = next) {
262                 next = batch->next;
263                 free_pages((unsigned long)batch, 0);
264         }
265         tlb->local.next = NULL;
266 }
267
268 /* __tlb_remove_page
269  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
270  *      handling the additional races in SMP caused by other CPUs caching valid
271  *      mappings in their TLBs. Returns the number of free page slots left.
272  *      When out of page slots we must call tlb_flush_mmu().
273  */
274 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
275 {
276         struct mmu_gather_batch *batch;
277
278         VM_BUG_ON(!tlb->need_flush);
279
280         if (tlb_fast_mode(tlb)) {
281                 free_page_and_swap_cache(page);
282                 return 1; /* avoid calling tlb_flush_mmu() */
283         }
284
285         batch = tlb->active;
286         batch->pages[batch->nr++] = page;
287         if (batch->nr == batch->max) {
288                 if (!tlb_next_batch(tlb))
289                         return 0;
290                 batch = tlb->active;
291         }
292         VM_BUG_ON(batch->nr > batch->max);
293
294         return batch->max - batch->nr;
295 }
296
297 #endif /* HAVE_GENERIC_MMU_GATHER */
298
299 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
300
301 /*
302  * See the comment near struct mmu_table_batch.
303  */
304
305 static void tlb_remove_table_smp_sync(void *arg)
306 {
307         /* Simply deliver the interrupt */
308 }
309
310 static void tlb_remove_table_one(void *table)
311 {
312         /*
313          * This isn't an RCU grace period and hence the page-tables cannot be
314          * assumed to be actually RCU-freed.
315          *
316          * It is however sufficient for software page-table walkers that rely on
317          * IRQ disabling. See the comment near struct mmu_table_batch.
318          */
319         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
320         __tlb_remove_table(table);
321 }
322
323 static void tlb_remove_table_rcu(struct rcu_head *head)
324 {
325         struct mmu_table_batch *batch;
326         int i;
327
328         batch = container_of(head, struct mmu_table_batch, rcu);
329
330         for (i = 0; i < batch->nr; i++)
331                 __tlb_remove_table(batch->tables[i]);
332
333         free_page((unsigned long)batch);
334 }
335
336 void tlb_table_flush(struct mmu_gather *tlb)
337 {
338         struct mmu_table_batch **batch = &tlb->batch;
339
340         if (*batch) {
341                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
342                 *batch = NULL;
343         }
344 }
345
346 void tlb_remove_table(struct mmu_gather *tlb, void *table)
347 {
348         struct mmu_table_batch **batch = &tlb->batch;
349
350         tlb->need_flush = 1;
351
352         /*
353          * When there's less then two users of this mm there cannot be a
354          * concurrent page-table walk.
355          */
356         if (atomic_read(&tlb->mm->mm_users) < 2) {
357                 __tlb_remove_table(table);
358                 return;
359         }
360
361         if (*batch == NULL) {
362                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
363                 if (*batch == NULL) {
364                         tlb_remove_table_one(table);
365                         return;
366                 }
367                 (*batch)->nr = 0;
368         }
369         (*batch)->tables[(*batch)->nr++] = table;
370         if ((*batch)->nr == MAX_TABLE_BATCH)
371                 tlb_table_flush(tlb);
372 }
373
374 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
375
376 /*
377  * If a p?d_bad entry is found while walking page tables, report
378  * the error, before resetting entry to p?d_none.  Usually (but
379  * very seldom) called out from the p?d_none_or_clear_bad macros.
380  */
381
382 void pgd_clear_bad(pgd_t *pgd)
383 {
384         pgd_ERROR(*pgd);
385         pgd_clear(pgd);
386 }
387
388 void pud_clear_bad(pud_t *pud)
389 {
390         pud_ERROR(*pud);
391         pud_clear(pud);
392 }
393
394 void pmd_clear_bad(pmd_t *pmd)
395 {
396         pmd_ERROR(*pmd);
397         pmd_clear(pmd);
398 }
399
400 /*
401  * Note: this doesn't free the actual pages themselves. That
402  * has been handled earlier when unmapping all the memory regions.
403  */
404 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
405                            unsigned long addr)
406 {
407         pgtable_t token = pmd_pgtable(*pmd);
408         pmd_clear(pmd);
409         pte_free_tlb(tlb, token, addr);
410         tlb->mm->nr_ptes--;
411 }
412
413 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
414                                 unsigned long addr, unsigned long end,
415                                 unsigned long floor, unsigned long ceiling)
416 {
417         pmd_t *pmd;
418         unsigned long next;
419         unsigned long start;
420
421         start = addr;
422         pmd = pmd_offset(pud, addr);
423         do {
424                 next = pmd_addr_end(addr, end);
425                 if (pmd_none_or_clear_bad(pmd))
426                         continue;
427                 free_pte_range(tlb, pmd, addr);
428         } while (pmd++, addr = next, addr != end);
429
430         start &= PUD_MASK;
431         if (start < floor)
432                 return;
433         if (ceiling) {
434                 ceiling &= PUD_MASK;
435                 if (!ceiling)
436                         return;
437         }
438         if (end - 1 > ceiling - 1)
439                 return;
440
441         pmd = pmd_offset(pud, start);
442         pud_clear(pud);
443         pmd_free_tlb(tlb, pmd, start);
444 }
445
446 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
447                                 unsigned long addr, unsigned long end,
448                                 unsigned long floor, unsigned long ceiling)
449 {
450         pud_t *pud;
451         unsigned long next;
452         unsigned long start;
453
454         start = addr;
455         pud = pud_offset(pgd, addr);
456         do {
457                 next = pud_addr_end(addr, end);
458                 if (pud_none_or_clear_bad(pud))
459                         continue;
460                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
461         } while (pud++, addr = next, addr != end);
462
463         start &= PGDIR_MASK;
464         if (start < floor)
465                 return;
466         if (ceiling) {
467                 ceiling &= PGDIR_MASK;
468                 if (!ceiling)
469                         return;
470         }
471         if (end - 1 > ceiling - 1)
472                 return;
473
474         pud = pud_offset(pgd, start);
475         pgd_clear(pgd);
476         pud_free_tlb(tlb, pud, start);
477 }
478
479 /*
480  * This function frees user-level page tables of a process.
481  *
482  * Must be called with pagetable lock held.
483  */
484 void free_pgd_range(struct mmu_gather *tlb,
485                         unsigned long addr, unsigned long end,
486                         unsigned long floor, unsigned long ceiling)
487 {
488         pgd_t *pgd;
489         unsigned long next;
490
491         /*
492          * The next few lines have given us lots of grief...
493          *
494          * Why are we testing PMD* at this top level?  Because often
495          * there will be no work to do at all, and we'd prefer not to
496          * go all the way down to the bottom just to discover that.
497          *
498          * Why all these "- 1"s?  Because 0 represents both the bottom
499          * of the address space and the top of it (using -1 for the
500          * top wouldn't help much: the masks would do the wrong thing).
501          * The rule is that addr 0 and floor 0 refer to the bottom of
502          * the address space, but end 0 and ceiling 0 refer to the top
503          * Comparisons need to use "end - 1" and "ceiling - 1" (though
504          * that end 0 case should be mythical).
505          *
506          * Wherever addr is brought up or ceiling brought down, we must
507          * be careful to reject "the opposite 0" before it confuses the
508          * subsequent tests.  But what about where end is brought down
509          * by PMD_SIZE below? no, end can't go down to 0 there.
510          *
511          * Whereas we round start (addr) and ceiling down, by different
512          * masks at different levels, in order to test whether a table
513          * now has no other vmas using it, so can be freed, we don't
514          * bother to round floor or end up - the tests don't need that.
515          */
516
517         addr &= PMD_MASK;
518         if (addr < floor) {
519                 addr += PMD_SIZE;
520                 if (!addr)
521                         return;
522         }
523         if (ceiling) {
524                 ceiling &= PMD_MASK;
525                 if (!ceiling)
526                         return;
527         }
528         if (end - 1 > ceiling - 1)
529                 end -= PMD_SIZE;
530         if (addr > end - 1)
531                 return;
532
533         pgd = pgd_offset(tlb->mm, addr);
534         do {
535                 next = pgd_addr_end(addr, end);
536                 if (pgd_none_or_clear_bad(pgd))
537                         continue;
538                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
539         } while (pgd++, addr = next, addr != end);
540 }
541
542 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
543                 unsigned long floor, unsigned long ceiling)
544 {
545         while (vma) {
546                 struct vm_area_struct *next = vma->vm_next;
547                 unsigned long addr = vma->vm_start;
548
549                 /*
550                  * Hide vma from rmap and truncate_pagecache before freeing
551                  * pgtables
552                  */
553                 unlink_anon_vmas(vma);
554                 unlink_file_vma(vma);
555
556                 if (is_vm_hugetlb_page(vma)) {
557                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
558                                 floor, next? next->vm_start: ceiling);
559                 } else {
560                         /*
561                          * Optimization: gather nearby vmas into one call down
562                          */
563                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
564                                && !is_vm_hugetlb_page(next)) {
565                                 vma = next;
566                                 next = vma->vm_next;
567                                 unlink_anon_vmas(vma);
568                                 unlink_file_vma(vma);
569                         }
570                         free_pgd_range(tlb, addr, vma->vm_end,
571                                 floor, next? next->vm_start: ceiling);
572                 }
573                 vma = next;
574         }
575 }
576
577 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
578                 pmd_t *pmd, unsigned long address)
579 {
580         pgtable_t new = pte_alloc_one(mm, address);
581         int wait_split_huge_page;
582         if (!new)
583                 return -ENOMEM;
584
585         /*
586          * Ensure all pte setup (eg. pte page lock and page clearing) are
587          * visible before the pte is made visible to other CPUs by being
588          * put into page tables.
589          *
590          * The other side of the story is the pointer chasing in the page
591          * table walking code (when walking the page table without locking;
592          * ie. most of the time). Fortunately, these data accesses consist
593          * of a chain of data-dependent loads, meaning most CPUs (alpha
594          * being the notable exception) will already guarantee loads are
595          * seen in-order. See the alpha page table accessors for the
596          * smp_read_barrier_depends() barriers in page table walking code.
597          */
598         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
599
600         spin_lock(&mm->page_table_lock);
601         wait_split_huge_page = 0;
602         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
603                 mm->nr_ptes++;
604                 pmd_populate(mm, pmd, new);
605                 new = NULL;
606         } else if (unlikely(pmd_trans_splitting(*pmd)))
607                 wait_split_huge_page = 1;
608         spin_unlock(&mm->page_table_lock);
609         if (new)
610                 pte_free(mm, new);
611         if (wait_split_huge_page)
612                 wait_split_huge_page(vma->anon_vma, pmd);
613         return 0;
614 }
615
616 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
617 {
618         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
619         if (!new)
620                 return -ENOMEM;
621
622         smp_wmb(); /* See comment in __pte_alloc */
623
624         spin_lock(&init_mm.page_table_lock);
625         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
626                 pmd_populate_kernel(&init_mm, pmd, new);
627                 new = NULL;
628         } else
629                 VM_BUG_ON(pmd_trans_splitting(*pmd));
630         spin_unlock(&init_mm.page_table_lock);
631         if (new)
632                 pte_free_kernel(&init_mm, new);
633         return 0;
634 }
635
636 static inline void init_rss_vec(int *rss)
637 {
638         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
639 }
640
641 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
642 {
643         int i;
644
645         if (current->mm == mm)
646                 sync_mm_rss(mm);
647         for (i = 0; i < NR_MM_COUNTERS; i++)
648                 if (rss[i])
649                         add_mm_counter(mm, i, rss[i]);
650 }
651
652 /*
653  * This function is called to print an error when a bad pte
654  * is found. For example, we might have a PFN-mapped pte in
655  * a region that doesn't allow it.
656  *
657  * The calling function must still handle the error.
658  */
659 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
660                           pte_t pte, struct page *page)
661 {
662         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
663         pud_t *pud = pud_offset(pgd, addr);
664         pmd_t *pmd = pmd_offset(pud, addr);
665         struct address_space *mapping;
666         pgoff_t index;
667         static unsigned long resume;
668         static unsigned long nr_shown;
669         static unsigned long nr_unshown;
670
671         /*
672          * Allow a burst of 60 reports, then keep quiet for that minute;
673          * or allow a steady drip of one report per second.
674          */
675         if (nr_shown == 60) {
676                 if (time_before(jiffies, resume)) {
677                         nr_unshown++;
678                         return;
679                 }
680                 if (nr_unshown) {
681                         printk(KERN_ALERT
682                                 "BUG: Bad page map: %lu messages suppressed\n",
683                                 nr_unshown);
684                         nr_unshown = 0;
685                 }
686                 nr_shown = 0;
687         }
688         if (nr_shown++ == 0)
689                 resume = jiffies + 60 * HZ;
690
691         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
692         index = linear_page_index(vma, addr);
693
694         printk(KERN_ALERT
695                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
696                 current->comm,
697                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
698         if (page)
699                 dump_page(page);
700         printk(KERN_ALERT
701                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
702                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
703         /*
704          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
705          */
706         if (vma->vm_ops)
707                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
708                                 (unsigned long)vma->vm_ops->fault);
709         if (vma->vm_file && vma->vm_file->f_op)
710                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
711                                 (unsigned long)vma->vm_file->f_op->mmap);
712         dump_stack();
713         add_taint(TAINT_BAD_PAGE);
714 }
715
716 static inline bool is_cow_mapping(vm_flags_t flags)
717 {
718         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
719 }
720
721 /*
722  * vm_normal_page -- This function gets the "struct page" associated with a pte.
723  *
724  * "Special" mappings do not wish to be associated with a "struct page" (either
725  * it doesn't exist, or it exists but they don't want to touch it). In this
726  * case, NULL is returned here. "Normal" mappings do have a struct page.
727  *
728  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
729  * pte bit, in which case this function is trivial. Secondly, an architecture
730  * may not have a spare pte bit, which requires a more complicated scheme,
731  * described below.
732  *
733  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
734  * special mapping (even if there are underlying and valid "struct pages").
735  * COWed pages of a VM_PFNMAP are always normal.
736  *
737  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
738  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
739  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
740  * mapping will always honor the rule
741  *
742  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
743  *
744  * And for normal mappings this is false.
745  *
746  * This restricts such mappings to be a linear translation from virtual address
747  * to pfn. To get around this restriction, we allow arbitrary mappings so long
748  * as the vma is not a COW mapping; in that case, we know that all ptes are
749  * special (because none can have been COWed).
750  *
751  *
752  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
753  *
754  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
755  * page" backing, however the difference is that _all_ pages with a struct
756  * page (that is, those where pfn_valid is true) are refcounted and considered
757  * normal pages by the VM. The disadvantage is that pages are refcounted
758  * (which can be slower and simply not an option for some PFNMAP users). The
759  * advantage is that we don't have to follow the strict linearity rule of
760  * PFNMAP mappings in order to support COWable mappings.
761  *
762  */
763 #ifdef __HAVE_ARCH_PTE_SPECIAL
764 # define HAVE_PTE_SPECIAL 1
765 #else
766 # define HAVE_PTE_SPECIAL 0
767 #endif
768 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
769                                 pte_t pte)
770 {
771         unsigned long pfn = pte_pfn(pte);
772
773         if (HAVE_PTE_SPECIAL) {
774                 if (likely(!pte_special(pte)))
775                         goto check_pfn;
776                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
777                         return NULL;
778                 if (!is_zero_pfn(pfn))
779                         print_bad_pte(vma, addr, pte, NULL);
780                 return NULL;
781         }
782
783         /* !HAVE_PTE_SPECIAL case follows: */
784
785         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
786                 if (vma->vm_flags & VM_MIXEDMAP) {
787                         if (!pfn_valid(pfn))
788                                 return NULL;
789                         goto out;
790                 } else {
791                         unsigned long off;
792                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
793                         if (pfn == vma->vm_pgoff + off)
794                                 return NULL;
795                         if (!is_cow_mapping(vma->vm_flags))
796                                 return NULL;
797                 }
798         }
799
800         if (is_zero_pfn(pfn))
801                 return NULL;
802 check_pfn:
803         if (unlikely(pfn > highest_memmap_pfn)) {
804                 print_bad_pte(vma, addr, pte, NULL);
805                 return NULL;
806         }
807
808         /*
809          * NOTE! We still have PageReserved() pages in the page tables.
810          * eg. VDSO mappings can cause them to exist.
811          */
812 out:
813         return pfn_to_page(pfn);
814 }
815
816 /*
817  * copy one vm_area from one task to the other. Assumes the page tables
818  * already present in the new task to be cleared in the whole range
819  * covered by this vma.
820  */
821
822 static inline unsigned long
823 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
824                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
825                 unsigned long addr, int *rss)
826 {
827         unsigned long vm_flags = vma->vm_flags;
828         pte_t pte = *src_pte;
829         struct page *page;
830
831         /* pte contains position in swap or file, so copy. */
832         if (unlikely(!pte_present(pte))) {
833                 if (!pte_file(pte)) {
834                         swp_entry_t entry = pte_to_swp_entry(pte);
835
836                         if (swap_duplicate(entry) < 0)
837                                 return entry.val;
838
839                         /* make sure dst_mm is on swapoff's mmlist. */
840                         if (unlikely(list_empty(&dst_mm->mmlist))) {
841                                 spin_lock(&mmlist_lock);
842                                 if (list_empty(&dst_mm->mmlist))
843                                         list_add(&dst_mm->mmlist,
844                                                  &src_mm->mmlist);
845                                 spin_unlock(&mmlist_lock);
846                         }
847                         if (likely(!non_swap_entry(entry)))
848                                 rss[MM_SWAPENTS]++;
849                         else if (is_migration_entry(entry)) {
850                                 page = migration_entry_to_page(entry);
851
852                                 if (PageAnon(page))
853                                         rss[MM_ANONPAGES]++;
854                                 else
855                                         rss[MM_FILEPAGES]++;
856
857                                 if (is_write_migration_entry(entry) &&
858                                     is_cow_mapping(vm_flags)) {
859                                         /*
860                                          * COW mappings require pages in both
861                                          * parent and child to be set to read.
862                                          */
863                                         make_migration_entry_read(&entry);
864                                         pte = swp_entry_to_pte(entry);
865                                         set_pte_at(src_mm, addr, src_pte, pte);
866                                 }
867                         }
868                 }
869                 goto out_set_pte;
870         }
871
872         /*
873          * If it's a COW mapping, write protect it both
874          * in the parent and the child
875          */
876         if (is_cow_mapping(vm_flags)) {
877                 ptep_set_wrprotect(src_mm, addr, src_pte);
878                 pte = pte_wrprotect(pte);
879         }
880
881         /*
882          * If it's a shared mapping, mark it clean in
883          * the child
884          */
885         if (vm_flags & VM_SHARED)
886                 pte = pte_mkclean(pte);
887         pte = pte_mkold(pte);
888
889         page = vm_normal_page(vma, addr, pte);
890         if (page) {
891                 get_page(page);
892                 page_dup_rmap(page);
893                 if (PageAnon(page))
894                         rss[MM_ANONPAGES]++;
895                 else
896                         rss[MM_FILEPAGES]++;
897         }
898
899 out_set_pte:
900         set_pte_at(dst_mm, addr, dst_pte, pte);
901         return 0;
902 }
903
904 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
905                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
906                    unsigned long addr, unsigned long end)
907 {
908         pte_t *orig_src_pte, *orig_dst_pte;
909         pte_t *src_pte, *dst_pte;
910         spinlock_t *src_ptl, *dst_ptl;
911         int progress = 0;
912         int rss[NR_MM_COUNTERS];
913         swp_entry_t entry = (swp_entry_t){0};
914
915 again:
916         init_rss_vec(rss);
917
918         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
919         if (!dst_pte)
920                 return -ENOMEM;
921         src_pte = pte_offset_map(src_pmd, addr);
922         src_ptl = pte_lockptr(src_mm, src_pmd);
923         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
924         orig_src_pte = src_pte;
925         orig_dst_pte = dst_pte;
926         arch_enter_lazy_mmu_mode();
927
928         do {
929                 /*
930                  * We are holding two locks at this point - either of them
931                  * could generate latencies in another task on another CPU.
932                  */
933                 if (progress >= 32) {
934                         progress = 0;
935                         if (need_resched() ||
936                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
937                                 break;
938                 }
939                 if (pte_none(*src_pte)) {
940                         progress++;
941                         continue;
942                 }
943                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
944                                                         vma, addr, rss);
945                 if (entry.val)
946                         break;
947                 progress += 8;
948         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
949
950         arch_leave_lazy_mmu_mode();
951         spin_unlock(src_ptl);
952         pte_unmap(orig_src_pte);
953         add_mm_rss_vec(dst_mm, rss);
954         pte_unmap_unlock(orig_dst_pte, dst_ptl);
955         cond_resched();
956
957         if (entry.val) {
958                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
959                         return -ENOMEM;
960                 progress = 0;
961         }
962         if (addr != end)
963                 goto again;
964         return 0;
965 }
966
967 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
968                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
969                 unsigned long addr, unsigned long end)
970 {
971         pmd_t *src_pmd, *dst_pmd;
972         unsigned long next;
973
974         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
975         if (!dst_pmd)
976                 return -ENOMEM;
977         src_pmd = pmd_offset(src_pud, addr);
978         do {
979                 next = pmd_addr_end(addr, end);
980                 if (pmd_trans_huge(*src_pmd)) {
981                         int err;
982                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
983                         err = copy_huge_pmd(dst_mm, src_mm,
984                                             dst_pmd, src_pmd, addr, vma);
985                         if (err == -ENOMEM)
986                                 return -ENOMEM;
987                         if (!err)
988                                 continue;
989                         /* fall through */
990                 }
991                 if (pmd_none_or_clear_bad(src_pmd))
992                         continue;
993                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
994                                                 vma, addr, next))
995                         return -ENOMEM;
996         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
997         return 0;
998 }
999
1000 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1001                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1002                 unsigned long addr, unsigned long end)
1003 {
1004         pud_t *src_pud, *dst_pud;
1005         unsigned long next;
1006
1007         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1008         if (!dst_pud)
1009                 return -ENOMEM;
1010         src_pud = pud_offset(src_pgd, addr);
1011         do {
1012                 next = pud_addr_end(addr, end);
1013                 if (pud_none_or_clear_bad(src_pud))
1014                         continue;
1015                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1016                                                 vma, addr, next))
1017                         return -ENOMEM;
1018         } while (dst_pud++, src_pud++, addr = next, addr != end);
1019         return 0;
1020 }
1021
1022 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023                 struct vm_area_struct *vma)
1024 {
1025         pgd_t *src_pgd, *dst_pgd;
1026         unsigned long next;
1027         unsigned long addr = vma->vm_start;
1028         unsigned long end = vma->vm_end;
1029         unsigned long mmun_start;       /* For mmu_notifiers */
1030         unsigned long mmun_end;         /* For mmu_notifiers */
1031         bool is_cow;
1032         int ret;
1033
1034         /*
1035          * Don't copy ptes where a page fault will fill them correctly.
1036          * Fork becomes much lighter when there are big shared or private
1037          * readonly mappings. The tradeoff is that copy_page_range is more
1038          * efficient than faulting.
1039          */
1040         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1041                                VM_PFNMAP | VM_MIXEDMAP))) {
1042                 if (!vma->anon_vma)
1043                         return 0;
1044         }
1045
1046         if (is_vm_hugetlb_page(vma))
1047                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1048
1049         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1050                 /*
1051                  * We do not free on error cases below as remove_vma
1052                  * gets called on error from higher level routine
1053                  */
1054                 ret = track_pfn_copy(vma);
1055                 if (ret)
1056                         return ret;
1057         }
1058
1059         /*
1060          * We need to invalidate the secondary MMU mappings only when
1061          * there could be a permission downgrade on the ptes of the
1062          * parent mm. And a permission downgrade will only happen if
1063          * is_cow_mapping() returns true.
1064          */
1065         is_cow = is_cow_mapping(vma->vm_flags);
1066         mmun_start = addr;
1067         mmun_end   = end;
1068         if (is_cow)
1069                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1070                                                     mmun_end);
1071
1072         ret = 0;
1073         dst_pgd = pgd_offset(dst_mm, addr);
1074         src_pgd = pgd_offset(src_mm, addr);
1075         do {
1076                 next = pgd_addr_end(addr, end);
1077                 if (pgd_none_or_clear_bad(src_pgd))
1078                         continue;
1079                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1080                                             vma, addr, next))) {
1081                         ret = -ENOMEM;
1082                         break;
1083                 }
1084         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1085
1086         if (is_cow)
1087                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1088         return ret;
1089 }
1090
1091 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1092                                 struct vm_area_struct *vma, pmd_t *pmd,
1093                                 unsigned long addr, unsigned long end,
1094                                 struct zap_details *details)
1095 {
1096         struct mm_struct *mm = tlb->mm;
1097         int force_flush = 0;
1098         int rss[NR_MM_COUNTERS];
1099         spinlock_t *ptl;
1100         pte_t *start_pte;
1101         pte_t *pte;
1102
1103 again:
1104         init_rss_vec(rss);
1105         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1106         pte = start_pte;
1107         arch_enter_lazy_mmu_mode();
1108         do {
1109                 pte_t ptent = *pte;
1110                 if (pte_none(ptent)) {
1111                         continue;
1112                 }
1113
1114                 if (pte_present(ptent)) {
1115                         struct page *page;
1116
1117                         page = vm_normal_page(vma, addr, ptent);
1118                         if (unlikely(details) && page) {
1119                                 /*
1120                                  * unmap_shared_mapping_pages() wants to
1121                                  * invalidate cache without truncating:
1122                                  * unmap shared but keep private pages.
1123                                  */
1124                                 if (details->check_mapping &&
1125                                     details->check_mapping != page->mapping)
1126                                         continue;
1127                                 /*
1128                                  * Each page->index must be checked when
1129                                  * invalidating or truncating nonlinear.
1130                                  */
1131                                 if (details->nonlinear_vma &&
1132                                     (page->index < details->first_index ||
1133                                      page->index > details->last_index))
1134                                         continue;
1135                         }
1136                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1137                                                         tlb->fullmm);
1138                         tlb_remove_tlb_entry(tlb, pte, addr);
1139                         if (unlikely(!page))
1140                                 continue;
1141                         if (unlikely(details) && details->nonlinear_vma
1142                             && linear_page_index(details->nonlinear_vma,
1143                                                 addr) != page->index)
1144                                 set_pte_at(mm, addr, pte,
1145                                            pgoff_to_pte(page->index));
1146                         if (PageAnon(page))
1147                                 rss[MM_ANONPAGES]--;
1148                         else {
1149                                 if (pte_dirty(ptent))
1150                                         set_page_dirty(page);
1151                                 if (pte_young(ptent) &&
1152                                     likely(!VM_SequentialReadHint(vma)))
1153                                         mark_page_accessed(page);
1154                                 rss[MM_FILEPAGES]--;
1155                         }
1156                         page_remove_rmap(page);
1157                         if (unlikely(page_mapcount(page) < 0))
1158                                 print_bad_pte(vma, addr, ptent, page);
1159                         force_flush = !__tlb_remove_page(tlb, page);
1160                         if (force_flush)
1161                                 break;
1162                         continue;
1163                 }
1164                 /*
1165                  * If details->check_mapping, we leave swap entries;
1166                  * if details->nonlinear_vma, we leave file entries.
1167                  */
1168                 if (unlikely(details))
1169                         continue;
1170                 if (pte_file(ptent)) {
1171                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1172                                 print_bad_pte(vma, addr, ptent, NULL);
1173                 } else {
1174                         swp_entry_t entry = pte_to_swp_entry(ptent);
1175
1176                         if (!non_swap_entry(entry))
1177                                 rss[MM_SWAPENTS]--;
1178                         else if (is_migration_entry(entry)) {
1179                                 struct page *page;
1180
1181                                 page = migration_entry_to_page(entry);
1182
1183                                 if (PageAnon(page))
1184                                         rss[MM_ANONPAGES]--;
1185                                 else
1186                                         rss[MM_FILEPAGES]--;
1187                         }
1188                         if (unlikely(!free_swap_and_cache(entry)))
1189                                 print_bad_pte(vma, addr, ptent, NULL);
1190                 }
1191                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1192         } while (pte++, addr += PAGE_SIZE, addr != end);
1193
1194         add_mm_rss_vec(mm, rss);
1195         arch_leave_lazy_mmu_mode();
1196         pte_unmap_unlock(start_pte, ptl);
1197
1198         /*
1199          * mmu_gather ran out of room to batch pages, we break out of
1200          * the PTE lock to avoid doing the potential expensive TLB invalidate
1201          * and page-free while holding it.
1202          */
1203         if (force_flush) {
1204                 force_flush = 0;
1205
1206 #ifdef HAVE_GENERIC_MMU_GATHER
1207                 tlb->start = addr;
1208                 tlb->end = end;
1209 #endif
1210                 tlb_flush_mmu(tlb);
1211                 if (addr != end)
1212                         goto again;
1213         }
1214
1215         return addr;
1216 }
1217
1218 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1219                                 struct vm_area_struct *vma, pud_t *pud,
1220                                 unsigned long addr, unsigned long end,
1221                                 struct zap_details *details)
1222 {
1223         pmd_t *pmd;
1224         unsigned long next;
1225
1226         pmd = pmd_offset(pud, addr);
1227         do {
1228                 next = pmd_addr_end(addr, end);
1229                 if (pmd_trans_huge(*pmd)) {
1230                         if (next - addr != HPAGE_PMD_SIZE) {
1231 #ifdef CONFIG_DEBUG_VM
1232                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1233                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1234                                                 __func__, addr, end,
1235                                                 vma->vm_start,
1236                                                 vma->vm_end);
1237                                         BUG();
1238                                 }
1239 #endif
1240                                 split_huge_page_pmd(vma, addr, pmd);
1241                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1242                                 goto next;
1243                         /* fall through */
1244                 }
1245                 /*
1246                  * Here there can be other concurrent MADV_DONTNEED or
1247                  * trans huge page faults running, and if the pmd is
1248                  * none or trans huge it can change under us. This is
1249                  * because MADV_DONTNEED holds the mmap_sem in read
1250                  * mode.
1251                  */
1252                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1253                         goto next;
1254                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1255 next:
1256                 cond_resched();
1257         } while (pmd++, addr = next, addr != end);
1258
1259         return addr;
1260 }
1261
1262 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1263                                 struct vm_area_struct *vma, pgd_t *pgd,
1264                                 unsigned long addr, unsigned long end,
1265                                 struct zap_details *details)
1266 {
1267         pud_t *pud;
1268         unsigned long next;
1269
1270         pud = pud_offset(pgd, addr);
1271         do {
1272                 next = pud_addr_end(addr, end);
1273                 if (pud_none_or_clear_bad(pud))
1274                         continue;
1275                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1276         } while (pud++, addr = next, addr != end);
1277
1278         return addr;
1279 }
1280
1281 static void unmap_page_range(struct mmu_gather *tlb,
1282                              struct vm_area_struct *vma,
1283                              unsigned long addr, unsigned long end,
1284                              struct zap_details *details)
1285 {
1286         pgd_t *pgd;
1287         unsigned long next;
1288
1289         if (details && !details->check_mapping && !details->nonlinear_vma)
1290                 details = NULL;
1291
1292         BUG_ON(addr >= end);
1293         mem_cgroup_uncharge_start();
1294         tlb_start_vma(tlb, vma);
1295         pgd = pgd_offset(vma->vm_mm, addr);
1296         do {
1297                 next = pgd_addr_end(addr, end);
1298                 if (pgd_none_or_clear_bad(pgd))
1299                         continue;
1300                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1301         } while (pgd++, addr = next, addr != end);
1302         tlb_end_vma(tlb, vma);
1303         mem_cgroup_uncharge_end();
1304 }
1305
1306
1307 static void unmap_single_vma(struct mmu_gather *tlb,
1308                 struct vm_area_struct *vma, unsigned long start_addr,
1309                 unsigned long end_addr,
1310                 struct zap_details *details)
1311 {
1312         unsigned long start = max(vma->vm_start, start_addr);
1313         unsigned long end;
1314
1315         if (start >= vma->vm_end)
1316                 return;
1317         end = min(vma->vm_end, end_addr);
1318         if (end <= vma->vm_start)
1319                 return;
1320
1321         if (vma->vm_file)
1322                 uprobe_munmap(vma, start, end);
1323
1324         if (unlikely(vma->vm_flags & VM_PFNMAP))
1325                 untrack_pfn(vma, 0, 0);
1326
1327         if (start != end) {
1328                 if (unlikely(is_vm_hugetlb_page(vma))) {
1329                         /*
1330                          * It is undesirable to test vma->vm_file as it
1331                          * should be non-null for valid hugetlb area.
1332                          * However, vm_file will be NULL in the error
1333                          * cleanup path of do_mmap_pgoff. When
1334                          * hugetlbfs ->mmap method fails,
1335                          * do_mmap_pgoff() nullifies vma->vm_file
1336                          * before calling this function to clean up.
1337                          * Since no pte has actually been setup, it is
1338                          * safe to do nothing in this case.
1339                          */
1340                         if (vma->vm_file) {
1341                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1342                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1343                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1344                         }
1345                 } else
1346                         unmap_page_range(tlb, vma, start, end, details);
1347         }
1348 }
1349
1350 /**
1351  * unmap_vmas - unmap a range of memory covered by a list of vma's
1352  * @tlb: address of the caller's struct mmu_gather
1353  * @vma: the starting vma
1354  * @start_addr: virtual address at which to start unmapping
1355  * @end_addr: virtual address at which to end unmapping
1356  *
1357  * Unmap all pages in the vma list.
1358  *
1359  * Only addresses between `start' and `end' will be unmapped.
1360  *
1361  * The VMA list must be sorted in ascending virtual address order.
1362  *
1363  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1364  * range after unmap_vmas() returns.  So the only responsibility here is to
1365  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1366  * drops the lock and schedules.
1367  */
1368 void unmap_vmas(struct mmu_gather *tlb,
1369                 struct vm_area_struct *vma, unsigned long start_addr,
1370                 unsigned long end_addr)
1371 {
1372         struct mm_struct *mm = vma->vm_mm;
1373
1374         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1375         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1376                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1377         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1378 }
1379
1380 /**
1381  * zap_page_range - remove user pages in a given range
1382  * @vma: vm_area_struct holding the applicable pages
1383  * @start: starting address of pages to zap
1384  * @size: number of bytes to zap
1385  * @details: details of nonlinear truncation or shared cache invalidation
1386  *
1387  * Caller must protect the VMA list
1388  */
1389 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1390                 unsigned long size, struct zap_details *details)
1391 {
1392         struct mm_struct *mm = vma->vm_mm;
1393         struct mmu_gather tlb;
1394         unsigned long end = start + size;
1395
1396         lru_add_drain();
1397         tlb_gather_mmu(&tlb, mm, 0);
1398         update_hiwater_rss(mm);
1399         mmu_notifier_invalidate_range_start(mm, start, end);
1400         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1401                 unmap_single_vma(&tlb, vma, start, end, details);
1402         mmu_notifier_invalidate_range_end(mm, start, end);
1403         tlb_finish_mmu(&tlb, start, end);
1404 }
1405
1406 /**
1407  * zap_page_range_single - remove user pages in a given range
1408  * @vma: vm_area_struct holding the applicable pages
1409  * @address: starting address of pages to zap
1410  * @size: number of bytes to zap
1411  * @details: details of nonlinear truncation or shared cache invalidation
1412  *
1413  * The range must fit into one VMA.
1414  */
1415 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1416                 unsigned long size, struct zap_details *details)
1417 {
1418         struct mm_struct *mm = vma->vm_mm;
1419         struct mmu_gather tlb;
1420         unsigned long end = address + size;
1421
1422         lru_add_drain();
1423         tlb_gather_mmu(&tlb, mm, 0);
1424         update_hiwater_rss(mm);
1425         mmu_notifier_invalidate_range_start(mm, address, end);
1426         unmap_single_vma(&tlb, vma, address, end, details);
1427         mmu_notifier_invalidate_range_end(mm, address, end);
1428         tlb_finish_mmu(&tlb, address, end);
1429 }
1430
1431 /**
1432  * zap_vma_ptes - remove ptes mapping the vma
1433  * @vma: vm_area_struct holding ptes to be zapped
1434  * @address: starting address of pages to zap
1435  * @size: number of bytes to zap
1436  *
1437  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1438  *
1439  * The entire address range must be fully contained within the vma.
1440  *
1441  * Returns 0 if successful.
1442  */
1443 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1444                 unsigned long size)
1445 {
1446         if (address < vma->vm_start || address + size > vma->vm_end ||
1447                         !(vma->vm_flags & VM_PFNMAP))
1448                 return -1;
1449         zap_page_range_single(vma, address, size, NULL);
1450         return 0;
1451 }
1452 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1453
1454 /**
1455  * follow_page - look up a page descriptor from a user-virtual address
1456  * @vma: vm_area_struct mapping @address
1457  * @address: virtual address to look up
1458  * @flags: flags modifying lookup behaviour
1459  *
1460  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1461  *
1462  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1463  * an error pointer if there is a mapping to something not represented
1464  * by a page descriptor (see also vm_normal_page()).
1465  */
1466 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1467                         unsigned int flags)
1468 {
1469         pgd_t *pgd;
1470         pud_t *pud;
1471         pmd_t *pmd;
1472         pte_t *ptep, pte;
1473         spinlock_t *ptl;
1474         struct page *page;
1475         struct mm_struct *mm = vma->vm_mm;
1476
1477         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1478         if (!IS_ERR(page)) {
1479                 BUG_ON(flags & FOLL_GET);
1480                 goto out;
1481         }
1482
1483         page = NULL;
1484         pgd = pgd_offset(mm, address);
1485         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1486                 goto no_page_table;
1487
1488         pud = pud_offset(pgd, address);
1489         if (pud_none(*pud))
1490                 goto no_page_table;
1491         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1492                 BUG_ON(flags & FOLL_GET);
1493                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1494                 goto out;
1495         }
1496         if (unlikely(pud_bad(*pud)))
1497                 goto no_page_table;
1498
1499         pmd = pmd_offset(pud, address);
1500         if (pmd_none(*pmd))
1501                 goto no_page_table;
1502         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1503                 BUG_ON(flags & FOLL_GET);
1504                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1505                 goto out;
1506         }
1507         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1508                 goto no_page_table;
1509         if (pmd_trans_huge(*pmd)) {
1510                 if (flags & FOLL_SPLIT) {
1511                         split_huge_page_pmd(vma, address, pmd);
1512                         goto split_fallthrough;
1513                 }
1514                 spin_lock(&mm->page_table_lock);
1515                 if (likely(pmd_trans_huge(*pmd))) {
1516                         if (unlikely(pmd_trans_splitting(*pmd))) {
1517                                 spin_unlock(&mm->page_table_lock);
1518                                 wait_split_huge_page(vma->anon_vma, pmd);
1519                         } else {
1520                                 page = follow_trans_huge_pmd(vma, address,
1521                                                              pmd, flags);
1522                                 spin_unlock(&mm->page_table_lock);
1523                                 goto out;
1524                         }
1525                 } else
1526                         spin_unlock(&mm->page_table_lock);
1527                 /* fall through */
1528         }
1529 split_fallthrough:
1530         if (unlikely(pmd_bad(*pmd)))
1531                 goto no_page_table;
1532
1533         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1534
1535         pte = *ptep;
1536         if (!pte_present(pte))
1537                 goto no_page;
1538         if ((flags & FOLL_NUMA) && pte_numa(pte))
1539                 goto no_page;
1540         if ((flags & FOLL_WRITE) && !pte_write(pte))
1541                 goto unlock;
1542
1543         page = vm_normal_page(vma, address, pte);
1544         if (unlikely(!page)) {
1545                 if ((flags & FOLL_DUMP) ||
1546                     !is_zero_pfn(pte_pfn(pte)))
1547                         goto bad_page;
1548                 page = pte_page(pte);
1549         }
1550
1551         if (flags & FOLL_GET)
1552                 get_page_foll(page);
1553         if (flags & FOLL_TOUCH) {
1554                 if ((flags & FOLL_WRITE) &&
1555                     !pte_dirty(pte) && !PageDirty(page))
1556                         set_page_dirty(page);
1557                 /*
1558                  * pte_mkyoung() would be more correct here, but atomic care
1559                  * is needed to avoid losing the dirty bit: it is easier to use
1560                  * mark_page_accessed().
1561                  */
1562                 mark_page_accessed(page);
1563         }
1564         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1565                 /*
1566                  * The preliminary mapping check is mainly to avoid the
1567                  * pointless overhead of lock_page on the ZERO_PAGE
1568                  * which might bounce very badly if there is contention.
1569                  *
1570                  * If the page is already locked, we don't need to
1571                  * handle it now - vmscan will handle it later if and
1572                  * when it attempts to reclaim the page.
1573                  */
1574                 if (page->mapping && trylock_page(page)) {
1575                         lru_add_drain();  /* push cached pages to LRU */
1576                         /*
1577                          * Because we lock page here, and migration is
1578                          * blocked by the pte's page reference, and we
1579                          * know the page is still mapped, we don't even
1580                          * need to check for file-cache page truncation.
1581                          */
1582                         mlock_vma_page(page);
1583                         unlock_page(page);
1584                 }
1585         }
1586 unlock:
1587         pte_unmap_unlock(ptep, ptl);
1588 out:
1589         return page;
1590
1591 bad_page:
1592         pte_unmap_unlock(ptep, ptl);
1593         return ERR_PTR(-EFAULT);
1594
1595 no_page:
1596         pte_unmap_unlock(ptep, ptl);
1597         if (!pte_none(pte))
1598                 return page;
1599
1600 no_page_table:
1601         /*
1602          * When core dumping an enormous anonymous area that nobody
1603          * has touched so far, we don't want to allocate unnecessary pages or
1604          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1605          * then get_dump_page() will return NULL to leave a hole in the dump.
1606          * But we can only make this optimization where a hole would surely
1607          * be zero-filled if handle_mm_fault() actually did handle it.
1608          */
1609         if ((flags & FOLL_DUMP) &&
1610             (!vma->vm_ops || !vma->vm_ops->fault))
1611                 return ERR_PTR(-EFAULT);
1612         return page;
1613 }
1614
1615 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1616 {
1617         return stack_guard_page_start(vma, addr) ||
1618                stack_guard_page_end(vma, addr+PAGE_SIZE);
1619 }
1620
1621 /**
1622  * __get_user_pages() - pin user pages in memory
1623  * @tsk:        task_struct of target task
1624  * @mm:         mm_struct of target mm
1625  * @start:      starting user address
1626  * @nr_pages:   number of pages from start to pin
1627  * @gup_flags:  flags modifying pin behaviour
1628  * @pages:      array that receives pointers to the pages pinned.
1629  *              Should be at least nr_pages long. Or NULL, if caller
1630  *              only intends to ensure the pages are faulted in.
1631  * @vmas:       array of pointers to vmas corresponding to each page.
1632  *              Or NULL if the caller does not require them.
1633  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1634  *
1635  * Returns number of pages pinned. This may be fewer than the number
1636  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1637  * were pinned, returns -errno. Each page returned must be released
1638  * with a put_page() call when it is finished with. vmas will only
1639  * remain valid while mmap_sem is held.
1640  *
1641  * Must be called with mmap_sem held for read or write.
1642  *
1643  * __get_user_pages walks a process's page tables and takes a reference to
1644  * each struct page that each user address corresponds to at a given
1645  * instant. That is, it takes the page that would be accessed if a user
1646  * thread accesses the given user virtual address at that instant.
1647  *
1648  * This does not guarantee that the page exists in the user mappings when
1649  * __get_user_pages returns, and there may even be a completely different
1650  * page there in some cases (eg. if mmapped pagecache has been invalidated
1651  * and subsequently re faulted). However it does guarantee that the page
1652  * won't be freed completely. And mostly callers simply care that the page
1653  * contains data that was valid *at some point in time*. Typically, an IO
1654  * or similar operation cannot guarantee anything stronger anyway because
1655  * locks can't be held over the syscall boundary.
1656  *
1657  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1658  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1659  * appropriate) must be called after the page is finished with, and
1660  * before put_page is called.
1661  *
1662  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1663  * or mmap_sem contention, and if waiting is needed to pin all pages,
1664  * *@nonblocking will be set to 0.
1665  *
1666  * In most cases, get_user_pages or get_user_pages_fast should be used
1667  * instead of __get_user_pages. __get_user_pages should be used only if
1668  * you need some special @gup_flags.
1669  */
1670 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1671                      unsigned long start, int nr_pages, unsigned int gup_flags,
1672                      struct page **pages, struct vm_area_struct **vmas,
1673                      int *nonblocking)
1674 {
1675         int i;
1676         unsigned long vm_flags;
1677
1678         if (nr_pages <= 0)
1679                 return 0;
1680
1681         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1682
1683         /* 
1684          * Require read or write permissions.
1685          * If FOLL_FORCE is set, we only require the "MAY" flags.
1686          */
1687         vm_flags  = (gup_flags & FOLL_WRITE) ?
1688                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1689         vm_flags &= (gup_flags & FOLL_FORCE) ?
1690                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1691
1692         /*
1693          * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1694          * would be called on PROT_NONE ranges. We must never invoke
1695          * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1696          * page faults would unprotect the PROT_NONE ranges if
1697          * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1698          * bitflag. So to avoid that, don't set FOLL_NUMA if
1699          * FOLL_FORCE is set.
1700          */
1701         if (!(gup_flags & FOLL_FORCE))
1702                 gup_flags |= FOLL_NUMA;
1703
1704         i = 0;
1705
1706         do {
1707                 struct vm_area_struct *vma;
1708
1709                 vma = find_extend_vma(mm, start);
1710                 if (!vma && in_gate_area(mm, start)) {
1711                         unsigned long pg = start & PAGE_MASK;
1712                         pgd_t *pgd;
1713                         pud_t *pud;
1714                         pmd_t *pmd;
1715                         pte_t *pte;
1716
1717                         /* user gate pages are read-only */
1718                         if (gup_flags & FOLL_WRITE)
1719                                 return i ? : -EFAULT;
1720                         if (pg > TASK_SIZE)
1721                                 pgd = pgd_offset_k(pg);
1722                         else
1723                                 pgd = pgd_offset_gate(mm, pg);
1724                         BUG_ON(pgd_none(*pgd));
1725                         pud = pud_offset(pgd, pg);
1726                         BUG_ON(pud_none(*pud));
1727                         pmd = pmd_offset(pud, pg);
1728                         if (pmd_none(*pmd))
1729                                 return i ? : -EFAULT;
1730                         VM_BUG_ON(pmd_trans_huge(*pmd));
1731                         pte = pte_offset_map(pmd, pg);
1732                         if (pte_none(*pte)) {
1733                                 pte_unmap(pte);
1734                                 return i ? : -EFAULT;
1735                         }
1736                         vma = get_gate_vma(mm);
1737                         if (pages) {
1738                                 struct page *page;
1739
1740                                 page = vm_normal_page(vma, start, *pte);
1741                                 if (!page) {
1742                                         if (!(gup_flags & FOLL_DUMP) &&
1743                                              is_zero_pfn(pte_pfn(*pte)))
1744                                                 page = pte_page(*pte);
1745                                         else {
1746                                                 pte_unmap(pte);
1747                                                 return i ? : -EFAULT;
1748                                         }
1749                                 }
1750                                 pages[i] = page;
1751                                 get_page(page);
1752                         }
1753                         pte_unmap(pte);
1754                         goto next_page;
1755                 }
1756
1757                 if (!vma ||
1758                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1759                     !(vm_flags & vma->vm_flags))
1760                         return i ? : -EFAULT;
1761
1762                 if (is_vm_hugetlb_page(vma)) {
1763                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1764                                         &start, &nr_pages, i, gup_flags);
1765                         continue;
1766                 }
1767
1768                 do {
1769                         struct page *page;
1770                         unsigned int foll_flags = gup_flags;
1771
1772                         /*
1773                          * If we have a pending SIGKILL, don't keep faulting
1774                          * pages and potentially allocating memory.
1775                          */
1776                         if (unlikely(fatal_signal_pending(current)))
1777                                 return i ? i : -ERESTARTSYS;
1778
1779                         cond_resched();
1780                         while (!(page = follow_page(vma, start, foll_flags))) {
1781                                 int ret;
1782                                 unsigned int fault_flags = 0;
1783
1784                                 /* For mlock, just skip the stack guard page. */
1785                                 if (foll_flags & FOLL_MLOCK) {
1786                                         if (stack_guard_page(vma, start))
1787                                                 goto next_page;
1788                                 }
1789                                 if (foll_flags & FOLL_WRITE)
1790                                         fault_flags |= FAULT_FLAG_WRITE;
1791                                 if (nonblocking)
1792                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1793                                 if (foll_flags & FOLL_NOWAIT)
1794                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1795
1796                                 ret = handle_mm_fault(mm, vma, start,
1797                                                         fault_flags);
1798
1799                                 if (ret & VM_FAULT_ERROR) {
1800                                         if (ret & VM_FAULT_OOM)
1801                                                 return i ? i : -ENOMEM;
1802                                         if (ret & (VM_FAULT_HWPOISON |
1803                                                    VM_FAULT_HWPOISON_LARGE)) {
1804                                                 if (i)
1805                                                         return i;
1806                                                 else if (gup_flags & FOLL_HWPOISON)
1807                                                         return -EHWPOISON;
1808                                                 else
1809                                                         return -EFAULT;
1810                                         }
1811                                         if (ret & VM_FAULT_SIGBUS)
1812                                                 return i ? i : -EFAULT;
1813                                         BUG();
1814                                 }
1815
1816                                 if (tsk) {
1817                                         if (ret & VM_FAULT_MAJOR)
1818                                                 tsk->maj_flt++;
1819                                         else
1820                                                 tsk->min_flt++;
1821                                 }
1822
1823                                 if (ret & VM_FAULT_RETRY) {
1824                                         if (nonblocking)
1825                                                 *nonblocking = 0;
1826                                         return i;
1827                                 }
1828
1829                                 /*
1830                                  * The VM_FAULT_WRITE bit tells us that
1831                                  * do_wp_page has broken COW when necessary,
1832                                  * even if maybe_mkwrite decided not to set
1833                                  * pte_write. We can thus safely do subsequent
1834                                  * page lookups as if they were reads. But only
1835                                  * do so when looping for pte_write is futile:
1836                                  * in some cases userspace may also be wanting
1837                                  * to write to the gotten user page, which a
1838                                  * read fault here might prevent (a readonly
1839                                  * page might get reCOWed by userspace write).
1840                                  */
1841                                 if ((ret & VM_FAULT_WRITE) &&
1842                                     !(vma->vm_flags & VM_WRITE))
1843                                         foll_flags &= ~FOLL_WRITE;
1844
1845                                 cond_resched();
1846                         }
1847                         if (IS_ERR(page))
1848                                 return i ? i : PTR_ERR(page);
1849                         if (pages) {
1850                                 pages[i] = page;
1851
1852                                 flush_anon_page(vma, page, start);
1853                                 flush_dcache_page(page);
1854                         }
1855 next_page:
1856                         if (vmas)
1857                                 vmas[i] = vma;
1858                         i++;
1859                         start += PAGE_SIZE;
1860                         nr_pages--;
1861                 } while (nr_pages && start < vma->vm_end);
1862         } while (nr_pages);
1863         return i;
1864 }
1865 EXPORT_SYMBOL(__get_user_pages);
1866
1867 /*
1868  * fixup_user_fault() - manually resolve a user page fault
1869  * @tsk:        the task_struct to use for page fault accounting, or
1870  *              NULL if faults are not to be recorded.
1871  * @mm:         mm_struct of target mm
1872  * @address:    user address
1873  * @fault_flags:flags to pass down to handle_mm_fault()
1874  *
1875  * This is meant to be called in the specific scenario where for locking reasons
1876  * we try to access user memory in atomic context (within a pagefault_disable()
1877  * section), this returns -EFAULT, and we want to resolve the user fault before
1878  * trying again.
1879  *
1880  * Typically this is meant to be used by the futex code.
1881  *
1882  * The main difference with get_user_pages() is that this function will
1883  * unconditionally call handle_mm_fault() which will in turn perform all the
1884  * necessary SW fixup of the dirty and young bits in the PTE, while
1885  * handle_mm_fault() only guarantees to update these in the struct page.
1886  *
1887  * This is important for some architectures where those bits also gate the
1888  * access permission to the page because they are maintained in software.  On
1889  * such architectures, gup() will not be enough to make a subsequent access
1890  * succeed.
1891  *
1892  * This should be called with the mm_sem held for read.
1893  */
1894 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1895                      unsigned long address, unsigned int fault_flags)
1896 {
1897         struct vm_area_struct *vma;
1898         int ret;
1899
1900         vma = find_extend_vma(mm, address);
1901         if (!vma || address < vma->vm_start)
1902                 return -EFAULT;
1903
1904         ret = handle_mm_fault(mm, vma, address, fault_flags);
1905         if (ret & VM_FAULT_ERROR) {
1906                 if (ret & VM_FAULT_OOM)
1907                         return -ENOMEM;
1908                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1909                         return -EHWPOISON;
1910                 if (ret & VM_FAULT_SIGBUS)
1911                         return -EFAULT;
1912                 BUG();
1913         }
1914         if (tsk) {
1915                 if (ret & VM_FAULT_MAJOR)
1916                         tsk->maj_flt++;
1917                 else
1918                         tsk->min_flt++;
1919         }
1920         return 0;
1921 }
1922
1923 /*
1924  * get_user_pages() - pin user pages in memory
1925  * @tsk:        the task_struct to use for page fault accounting, or
1926  *              NULL if faults are not to be recorded.
1927  * @mm:         mm_struct of target mm
1928  * @start:      starting user address
1929  * @nr_pages:   number of pages from start to pin
1930  * @write:      whether pages will be written to by the caller
1931  * @force:      whether to force write access even if user mapping is
1932  *              readonly. This will result in the page being COWed even
1933  *              in MAP_SHARED mappings. You do not want this.
1934  * @pages:      array that receives pointers to the pages pinned.
1935  *              Should be at least nr_pages long. Or NULL, if caller
1936  *              only intends to ensure the pages are faulted in.
1937  * @vmas:       array of pointers to vmas corresponding to each page.
1938  *              Or NULL if the caller does not require them.
1939  *
1940  * Returns number of pages pinned. This may be fewer than the number
1941  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1942  * were pinned, returns -errno. Each page returned must be released
1943  * with a put_page() call when it is finished with. vmas will only
1944  * remain valid while mmap_sem is held.
1945  *
1946  * Must be called with mmap_sem held for read or write.
1947  *
1948  * get_user_pages walks a process's page tables and takes a reference to
1949  * each struct page that each user address corresponds to at a given
1950  * instant. That is, it takes the page that would be accessed if a user
1951  * thread accesses the given user virtual address at that instant.
1952  *
1953  * This does not guarantee that the page exists in the user mappings when
1954  * get_user_pages returns, and there may even be a completely different
1955  * page there in some cases (eg. if mmapped pagecache has been invalidated
1956  * and subsequently re faulted). However it does guarantee that the page
1957  * won't be freed completely. And mostly callers simply care that the page
1958  * contains data that was valid *at some point in time*. Typically, an IO
1959  * or similar operation cannot guarantee anything stronger anyway because
1960  * locks can't be held over the syscall boundary.
1961  *
1962  * If write=0, the page must not be written to. If the page is written to,
1963  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1964  * after the page is finished with, and before put_page is called.
1965  *
1966  * get_user_pages is typically used for fewer-copy IO operations, to get a
1967  * handle on the memory by some means other than accesses via the user virtual
1968  * addresses. The pages may be submitted for DMA to devices or accessed via
1969  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1970  * use the correct cache flushing APIs.
1971  *
1972  * See also get_user_pages_fast, for performance critical applications.
1973  */
1974 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1975                 unsigned long start, int nr_pages, int write, int force,
1976                 struct page **pages, struct vm_area_struct **vmas)
1977 {
1978         int flags = FOLL_TOUCH;
1979
1980         if (pages)
1981                 flags |= FOLL_GET;
1982         if (write)
1983                 flags |= FOLL_WRITE;
1984         if (force)
1985                 flags |= FOLL_FORCE;
1986
1987         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1988                                 NULL);
1989 }
1990 EXPORT_SYMBOL(get_user_pages);
1991
1992 /**
1993  * get_dump_page() - pin user page in memory while writing it to core dump
1994  * @addr: user address
1995  *
1996  * Returns struct page pointer of user page pinned for dump,
1997  * to be freed afterwards by page_cache_release() or put_page().
1998  *
1999  * Returns NULL on any kind of failure - a hole must then be inserted into
2000  * the corefile, to preserve alignment with its headers; and also returns
2001  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2002  * allowing a hole to be left in the corefile to save diskspace.
2003  *
2004  * Called without mmap_sem, but after all other threads have been killed.
2005  */
2006 #ifdef CONFIG_ELF_CORE
2007 struct page *get_dump_page(unsigned long addr)
2008 {
2009         struct vm_area_struct *vma;
2010         struct page *page;
2011
2012         if (__get_user_pages(current, current->mm, addr, 1,
2013                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2014                              NULL) < 1)
2015                 return NULL;
2016         flush_cache_page(vma, addr, page_to_pfn(page));
2017         return page;
2018 }
2019 #endif /* CONFIG_ELF_CORE */
2020
2021 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2022                         spinlock_t **ptl)
2023 {
2024         pgd_t * pgd = pgd_offset(mm, addr);
2025         pud_t * pud = pud_alloc(mm, pgd, addr);
2026         if (pud) {
2027                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2028                 if (pmd) {
2029                         VM_BUG_ON(pmd_trans_huge(*pmd));
2030                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2031                 }
2032         }
2033         return NULL;
2034 }
2035
2036 /*
2037  * This is the old fallback for page remapping.
2038  *
2039  * For historical reasons, it only allows reserved pages. Only
2040  * old drivers should use this, and they needed to mark their
2041  * pages reserved for the old functions anyway.
2042  */
2043 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2044                         struct page *page, pgprot_t prot)
2045 {
2046         struct mm_struct *mm = vma->vm_mm;
2047         int retval;
2048         pte_t *pte;
2049         spinlock_t *ptl;
2050
2051         retval = -EINVAL;
2052         if (PageAnon(page))
2053                 goto out;
2054         retval = -ENOMEM;
2055         flush_dcache_page(page);
2056         pte = get_locked_pte(mm, addr, &ptl);
2057         if (!pte)
2058                 goto out;
2059         retval = -EBUSY;
2060         if (!pte_none(*pte))
2061                 goto out_unlock;
2062
2063         /* Ok, finally just insert the thing.. */
2064         get_page(page);
2065         inc_mm_counter_fast(mm, MM_FILEPAGES);
2066         page_add_file_rmap(page);
2067         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2068
2069         retval = 0;
2070         pte_unmap_unlock(pte, ptl);
2071         return retval;
2072 out_unlock:
2073         pte_unmap_unlock(pte, ptl);
2074 out:
2075         return retval;
2076 }
2077
2078 /**
2079  * vm_insert_page - insert single page into user vma
2080  * @vma: user vma to map to
2081  * @addr: target user address of this page
2082  * @page: source kernel page
2083  *
2084  * This allows drivers to insert individual pages they've allocated
2085  * into a user vma.
2086  *
2087  * The page has to be a nice clean _individual_ kernel allocation.
2088  * If you allocate a compound page, you need to have marked it as
2089  * such (__GFP_COMP), or manually just split the page up yourself
2090  * (see split_page()).
2091  *
2092  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2093  * took an arbitrary page protection parameter. This doesn't allow
2094  * that. Your vma protection will have to be set up correctly, which
2095  * means that if you want a shared writable mapping, you'd better
2096  * ask for a shared writable mapping!
2097  *
2098  * The page does not need to be reserved.
2099  *
2100  * Usually this function is called from f_op->mmap() handler
2101  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2102  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2103  * function from other places, for example from page-fault handler.
2104  */
2105 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2106                         struct page *page)
2107 {
2108         if (addr < vma->vm_start || addr >= vma->vm_end)
2109                 return -EFAULT;
2110         if (!page_count(page))
2111                 return -EINVAL;
2112         if (!(vma->vm_flags & VM_MIXEDMAP)) {
2113                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2114                 BUG_ON(vma->vm_flags & VM_PFNMAP);
2115                 vma->vm_flags |= VM_MIXEDMAP;
2116         }
2117         return insert_page(vma, addr, page, vma->vm_page_prot);
2118 }
2119 EXPORT_SYMBOL(vm_insert_page);
2120
2121 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2122                         unsigned long pfn, pgprot_t prot)
2123 {
2124         struct mm_struct *mm = vma->vm_mm;
2125         int retval;
2126         pte_t *pte, entry;
2127         spinlock_t *ptl;
2128
2129         retval = -ENOMEM;
2130         pte = get_locked_pte(mm, addr, &ptl);
2131         if (!pte)
2132                 goto out;
2133         retval = -EBUSY;
2134         if (!pte_none(*pte))
2135                 goto out_unlock;
2136
2137         /* Ok, finally just insert the thing.. */
2138         entry = pte_mkspecial(pfn_pte(pfn, prot));
2139         set_pte_at(mm, addr, pte, entry);
2140         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2141
2142         retval = 0;
2143 out_unlock:
2144         pte_unmap_unlock(pte, ptl);
2145 out:
2146         return retval;
2147 }
2148
2149 /**
2150  * vm_insert_pfn - insert single pfn into user vma
2151  * @vma: user vma to map to
2152  * @addr: target user address of this page
2153  * @pfn: source kernel pfn
2154  *
2155  * Similar to vm_insert_page, this allows drivers to insert individual pages
2156  * they've allocated into a user vma. Same comments apply.
2157  *
2158  * This function should only be called from a vm_ops->fault handler, and
2159  * in that case the handler should return NULL.
2160  *
2161  * vma cannot be a COW mapping.
2162  *
2163  * As this is called only for pages that do not currently exist, we
2164  * do not need to flush old virtual caches or the TLB.
2165  */
2166 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2167                         unsigned long pfn)
2168 {
2169         int ret;
2170         pgprot_t pgprot = vma->vm_page_prot;
2171         /*
2172          * Technically, architectures with pte_special can avoid all these
2173          * restrictions (same for remap_pfn_range).  However we would like
2174          * consistency in testing and feature parity among all, so we should
2175          * try to keep these invariants in place for everybody.
2176          */
2177         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2178         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2179                                                 (VM_PFNMAP|VM_MIXEDMAP));
2180         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2181         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2182
2183         if (addr < vma->vm_start || addr >= vma->vm_end)
2184                 return -EFAULT;
2185         if (track_pfn_insert(vma, &pgprot, pfn))
2186                 return -EINVAL;
2187
2188         ret = insert_pfn(vma, addr, pfn, pgprot);
2189
2190         return ret;
2191 }
2192 EXPORT_SYMBOL(vm_insert_pfn);
2193
2194 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2195                         unsigned long pfn)
2196 {
2197         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2198
2199         if (addr < vma->vm_start || addr >= vma->vm_end)
2200                 return -EFAULT;
2201
2202         /*
2203          * If we don't have pte special, then we have to use the pfn_valid()
2204          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2205          * refcount the page if pfn_valid is true (hence insert_page rather
2206          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2207          * without pte special, it would there be refcounted as a normal page.
2208          */
2209         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2210                 struct page *page;
2211
2212                 page = pfn_to_page(pfn);
2213                 return insert_page(vma, addr, page, vma->vm_page_prot);
2214         }
2215         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2216 }
2217 EXPORT_SYMBOL(vm_insert_mixed);
2218
2219 /*
2220  * maps a range of physical memory into the requested pages. the old
2221  * mappings are removed. any references to nonexistent pages results
2222  * in null mappings (currently treated as "copy-on-access")
2223  */
2224 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2225                         unsigned long addr, unsigned long end,
2226                         unsigned long pfn, pgprot_t prot)
2227 {
2228         pte_t *pte;
2229         spinlock_t *ptl;
2230
2231         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2232         if (!pte)
2233                 return -ENOMEM;
2234         arch_enter_lazy_mmu_mode();
2235         do {
2236                 BUG_ON(!pte_none(*pte));
2237                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2238                 pfn++;
2239         } while (pte++, addr += PAGE_SIZE, addr != end);
2240         arch_leave_lazy_mmu_mode();
2241         pte_unmap_unlock(pte - 1, ptl);
2242         return 0;
2243 }
2244
2245 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2246                         unsigned long addr, unsigned long end,
2247                         unsigned long pfn, pgprot_t prot)
2248 {
2249         pmd_t *pmd;
2250         unsigned long next;
2251
2252         pfn -= addr >> PAGE_SHIFT;
2253         pmd = pmd_alloc(mm, pud, addr);
2254         if (!pmd)
2255                 return -ENOMEM;
2256         VM_BUG_ON(pmd_trans_huge(*pmd));
2257         do {
2258                 next = pmd_addr_end(addr, end);
2259                 if (remap_pte_range(mm, pmd, addr, next,
2260                                 pfn + (addr >> PAGE_SHIFT), prot))
2261                         return -ENOMEM;
2262         } while (pmd++, addr = next, addr != end);
2263         return 0;
2264 }
2265
2266 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2267                         unsigned long addr, unsigned long end,
2268                         unsigned long pfn, pgprot_t prot)
2269 {
2270         pud_t *pud;
2271         unsigned long next;
2272
2273         pfn -= addr >> PAGE_SHIFT;
2274         pud = pud_alloc(mm, pgd, addr);
2275         if (!pud)
2276                 return -ENOMEM;
2277         do {
2278                 next = pud_addr_end(addr, end);
2279                 if (remap_pmd_range(mm, pud, addr, next,
2280                                 pfn + (addr >> PAGE_SHIFT), prot))
2281                         return -ENOMEM;
2282         } while (pud++, addr = next, addr != end);
2283         return 0;
2284 }
2285
2286 /**
2287  * remap_pfn_range - remap kernel memory to userspace
2288  * @vma: user vma to map to
2289  * @addr: target user address to start at
2290  * @pfn: physical address of kernel memory
2291  * @size: size of map area
2292  * @prot: page protection flags for this mapping
2293  *
2294  *  Note: this is only safe if the mm semaphore is held when called.
2295  */
2296 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2297                     unsigned long pfn, unsigned long size, pgprot_t prot)
2298 {
2299         pgd_t *pgd;
2300         unsigned long next;
2301         unsigned long end = addr + PAGE_ALIGN(size);
2302         struct mm_struct *mm = vma->vm_mm;
2303         int err;
2304
2305         /*
2306          * Physically remapped pages are special. Tell the
2307          * rest of the world about it:
2308          *   VM_IO tells people not to look at these pages
2309          *      (accesses can have side effects).
2310          *   VM_PFNMAP tells the core MM that the base pages are just
2311          *      raw PFN mappings, and do not have a "struct page" associated
2312          *      with them.
2313          *   VM_DONTEXPAND
2314          *      Disable vma merging and expanding with mremap().
2315          *   VM_DONTDUMP
2316          *      Omit vma from core dump, even when VM_IO turned off.
2317          *
2318          * There's a horrible special case to handle copy-on-write
2319          * behaviour that some programs depend on. We mark the "original"
2320          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2321          * See vm_normal_page() for details.
2322          */
2323         if (is_cow_mapping(vma->vm_flags)) {
2324                 if (addr != vma->vm_start || end != vma->vm_end)
2325                         return -EINVAL;
2326                 vma->vm_pgoff = pfn;
2327         }
2328
2329         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2330         if (err)
2331                 return -EINVAL;
2332
2333         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2334
2335         BUG_ON(addr >= end);
2336         pfn -= addr >> PAGE_SHIFT;
2337         pgd = pgd_offset(mm, addr);
2338         flush_cache_range(vma, addr, end);
2339         do {
2340                 next = pgd_addr_end(addr, end);
2341                 err = remap_pud_range(mm, pgd, addr, next,
2342                                 pfn + (addr >> PAGE_SHIFT), prot);
2343                 if (err)
2344                         break;
2345         } while (pgd++, addr = next, addr != end);
2346
2347         if (err)
2348                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2349
2350         return err;
2351 }
2352 EXPORT_SYMBOL(remap_pfn_range);
2353
2354 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2355                                      unsigned long addr, unsigned long end,
2356                                      pte_fn_t fn, void *data)
2357 {
2358         pte_t *pte;
2359         int err;
2360         pgtable_t token;
2361         spinlock_t *uninitialized_var(ptl);
2362
2363         pte = (mm == &init_mm) ?
2364                 pte_alloc_kernel(pmd, addr) :
2365                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2366         if (!pte)
2367                 return -ENOMEM;
2368
2369         BUG_ON(pmd_huge(*pmd));
2370
2371         arch_enter_lazy_mmu_mode();
2372
2373         token = pmd_pgtable(*pmd);
2374
2375         do {
2376                 err = fn(pte++, token, addr, data);
2377                 if (err)
2378                         break;
2379         } while (addr += PAGE_SIZE, addr != end);
2380
2381         arch_leave_lazy_mmu_mode();
2382
2383         if (mm != &init_mm)
2384                 pte_unmap_unlock(pte-1, ptl);
2385         return err;
2386 }
2387
2388 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2389                                      unsigned long addr, unsigned long end,
2390                                      pte_fn_t fn, void *data)
2391 {
2392         pmd_t *pmd;
2393         unsigned long next;
2394         int err;
2395
2396         BUG_ON(pud_huge(*pud));
2397
2398         pmd = pmd_alloc(mm, pud, addr);
2399         if (!pmd)
2400                 return -ENOMEM;
2401         do {
2402                 next = pmd_addr_end(addr, end);
2403                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2404                 if (err)
2405                         break;
2406         } while (pmd++, addr = next, addr != end);
2407         return err;
2408 }
2409
2410 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2411                                      unsigned long addr, unsigned long end,
2412                                      pte_fn_t fn, void *data)
2413 {
2414         pud_t *pud;
2415         unsigned long next;
2416         int err;
2417
2418         pud = pud_alloc(mm, pgd, addr);
2419         if (!pud)
2420                 return -ENOMEM;
2421         do {
2422                 next = pud_addr_end(addr, end);
2423                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2424                 if (err)
2425                         break;
2426         } while (pud++, addr = next, addr != end);
2427         return err;
2428 }
2429
2430 /*
2431  * Scan a region of virtual memory, filling in page tables as necessary
2432  * and calling a provided function on each leaf page table.
2433  */
2434 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2435                         unsigned long size, pte_fn_t fn, void *data)
2436 {
2437         pgd_t *pgd;
2438         unsigned long next;
2439         unsigned long end = addr + size;
2440         int err;
2441
2442         BUG_ON(addr >= end);
2443         pgd = pgd_offset(mm, addr);
2444         do {
2445                 next = pgd_addr_end(addr, end);
2446                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2447                 if (err)
2448                         break;
2449         } while (pgd++, addr = next, addr != end);
2450
2451         return err;
2452 }
2453 EXPORT_SYMBOL_GPL(apply_to_page_range);
2454
2455 /*
2456  * handle_pte_fault chooses page fault handler according to an entry
2457  * which was read non-atomically.  Before making any commitment, on
2458  * those architectures or configurations (e.g. i386 with PAE) which
2459  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2460  * must check under lock before unmapping the pte and proceeding
2461  * (but do_wp_page is only called after already making such a check;
2462  * and do_anonymous_page can safely check later on).
2463  */
2464 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2465                                 pte_t *page_table, pte_t orig_pte)
2466 {
2467         int same = 1;
2468 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2469         if (sizeof(pte_t) > sizeof(unsigned long)) {
2470                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2471                 spin_lock(ptl);
2472                 same = pte_same(*page_table, orig_pte);
2473                 spin_unlock(ptl);
2474         }
2475 #endif
2476         pte_unmap(page_table);
2477         return same;
2478 }
2479
2480 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2481 {
2482         /*
2483          * If the source page was a PFN mapping, we don't have
2484          * a "struct page" for it. We do a best-effort copy by
2485          * just copying from the original user address. If that
2486          * fails, we just zero-fill it. Live with it.
2487          */
2488         if (unlikely(!src)) {
2489                 void *kaddr = kmap_atomic(dst);
2490                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2491
2492                 /*
2493                  * This really shouldn't fail, because the page is there
2494                  * in the page tables. But it might just be unreadable,
2495                  * in which case we just give up and fill the result with
2496                  * zeroes.
2497                  */
2498                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2499                         clear_page(kaddr);
2500                 kunmap_atomic(kaddr);
2501                 flush_dcache_page(dst);
2502         } else
2503                 copy_user_highpage(dst, src, va, vma);
2504 }
2505
2506 /*
2507  * This routine handles present pages, when users try to write
2508  * to a shared page. It is done by copying the page to a new address
2509  * and decrementing the shared-page counter for the old page.
2510  *
2511  * Note that this routine assumes that the protection checks have been
2512  * done by the caller (the low-level page fault routine in most cases).
2513  * Thus we can safely just mark it writable once we've done any necessary
2514  * COW.
2515  *
2516  * We also mark the page dirty at this point even though the page will
2517  * change only once the write actually happens. This avoids a few races,
2518  * and potentially makes it more efficient.
2519  *
2520  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2521  * but allow concurrent faults), with pte both mapped and locked.
2522  * We return with mmap_sem still held, but pte unmapped and unlocked.
2523  */
2524 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2525                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2526                 spinlock_t *ptl, pte_t orig_pte)
2527         __releases(ptl)
2528 {
2529         struct page *old_page, *new_page = NULL;
2530         pte_t entry;
2531         int ret = 0;
2532         int page_mkwrite = 0;
2533         struct page *dirty_page = NULL;
2534         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2535         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2536
2537         old_page = vm_normal_page(vma, address, orig_pte);
2538         if (!old_page) {
2539                 /*
2540                  * VM_MIXEDMAP !pfn_valid() case
2541                  *
2542                  * We should not cow pages in a shared writeable mapping.
2543                  * Just mark the pages writable as we can't do any dirty
2544                  * accounting on raw pfn maps.
2545                  */
2546                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2547                                      (VM_WRITE|VM_SHARED))
2548                         goto reuse;
2549                 goto gotten;
2550         }
2551
2552         /*
2553          * Take out anonymous pages first, anonymous shared vmas are
2554          * not dirty accountable.
2555          */
2556         if (PageAnon(old_page) && !PageKsm(old_page)) {
2557                 if (!trylock_page(old_page)) {
2558                         page_cache_get(old_page);
2559                         pte_unmap_unlock(page_table, ptl);
2560                         lock_page(old_page);
2561                         page_table = pte_offset_map_lock(mm, pmd, address,
2562                                                          &ptl);
2563                         if (!pte_same(*page_table, orig_pte)) {
2564                                 unlock_page(old_page);
2565                                 goto unlock;
2566                         }
2567                         page_cache_release(old_page);
2568                 }
2569                 if (reuse_swap_page(old_page)) {
2570                         /*
2571                          * The page is all ours.  Move it to our anon_vma so
2572                          * the rmap code will not search our parent or siblings.
2573                          * Protected against the rmap code by the page lock.
2574                          */
2575                         page_move_anon_rmap(old_page, vma, address);
2576                         unlock_page(old_page);
2577                         goto reuse;
2578                 }
2579                 unlock_page(old_page);
2580         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2581                                         (VM_WRITE|VM_SHARED))) {
2582                 /*
2583                  * Only catch write-faults on shared writable pages,
2584                  * read-only shared pages can get COWed by
2585                  * get_user_pages(.write=1, .force=1).
2586                  */
2587                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2588                         struct vm_fault vmf;
2589                         int tmp;
2590
2591                         vmf.virtual_address = (void __user *)(address &
2592                                                                 PAGE_MASK);
2593                         vmf.pgoff = old_page->index;
2594                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2595                         vmf.page = old_page;
2596
2597                         /*
2598                          * Notify the address space that the page is about to
2599                          * become writable so that it can prohibit this or wait
2600                          * for the page to get into an appropriate state.
2601                          *
2602                          * We do this without the lock held, so that it can
2603                          * sleep if it needs to.
2604                          */
2605                         page_cache_get(old_page);
2606                         pte_unmap_unlock(page_table, ptl);
2607
2608                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2609                         if (unlikely(tmp &
2610                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2611                                 ret = tmp;
2612                                 goto unwritable_page;
2613                         }
2614                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2615                                 lock_page(old_page);
2616                                 if (!old_page->mapping) {
2617                                         ret = 0; /* retry the fault */
2618                                         unlock_page(old_page);
2619                                         goto unwritable_page;
2620                                 }
2621                         } else
2622                                 VM_BUG_ON(!PageLocked(old_page));
2623
2624                         /*
2625                          * Since we dropped the lock we need to revalidate
2626                          * the PTE as someone else may have changed it.  If
2627                          * they did, we just return, as we can count on the
2628                          * MMU to tell us if they didn't also make it writable.
2629                          */
2630                         page_table = pte_offset_map_lock(mm, pmd, address,
2631                                                          &ptl);
2632                         if (!pte_same(*page_table, orig_pte)) {
2633                                 unlock_page(old_page);
2634                                 goto unlock;
2635                         }
2636
2637                         page_mkwrite = 1;
2638                 }
2639                 dirty_page = old_page;
2640                 get_page(dirty_page);
2641
2642 reuse:
2643                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2644                 entry = pte_mkyoung(orig_pte);
2645                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2646                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2647                         update_mmu_cache(vma, address, page_table);
2648                 pte_unmap_unlock(page_table, ptl);
2649                 ret |= VM_FAULT_WRITE;
2650
2651                 if (!dirty_page)
2652                         return ret;
2653
2654                 /*
2655                  * Yes, Virginia, this is actually required to prevent a race
2656                  * with clear_page_dirty_for_io() from clearing the page dirty
2657                  * bit after it clear all dirty ptes, but before a racing
2658                  * do_wp_page installs a dirty pte.
2659                  *
2660                  * __do_fault is protected similarly.
2661                  */
2662                 if (!page_mkwrite) {
2663                         wait_on_page_locked(dirty_page);
2664                         set_page_dirty_balance(dirty_page, page_mkwrite);
2665                         /* file_update_time outside page_lock */
2666                         if (vma->vm_file)
2667                                 file_update_time(vma->vm_file);
2668                 }
2669                 put_page(dirty_page);
2670                 if (page_mkwrite) {
2671                         struct address_space *mapping = dirty_page->mapping;
2672
2673                         set_page_dirty(dirty_page);
2674                         unlock_page(dirty_page);
2675                         page_cache_release(dirty_page);
2676                         if (mapping)    {
2677                                 /*
2678                                  * Some device drivers do not set page.mapping
2679                                  * but still dirty their pages
2680                                  */
2681                                 balance_dirty_pages_ratelimited(mapping);
2682                         }
2683                 }
2684
2685                 return ret;
2686         }
2687
2688         /*
2689          * Ok, we need to copy. Oh, well..
2690          */
2691         page_cache_get(old_page);
2692 gotten:
2693         pte_unmap_unlock(page_table, ptl);
2694
2695         if (unlikely(anon_vma_prepare(vma)))
2696                 goto oom;
2697
2698         if (is_zero_pfn(pte_pfn(orig_pte))) {
2699                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2700                 if (!new_page)
2701                         goto oom;
2702         } else {
2703                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2704                 if (!new_page)
2705                         goto oom;
2706                 cow_user_page(new_page, old_page, address, vma);
2707         }
2708         __SetPageUptodate(new_page);
2709
2710         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2711                 goto oom_free_new;
2712
2713         mmun_start  = address & PAGE_MASK;
2714         mmun_end    = mmun_start + PAGE_SIZE;
2715         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2716
2717         /*
2718          * Re-check the pte - we dropped the lock
2719          */
2720         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2721         if (likely(pte_same(*page_table, orig_pte))) {
2722                 if (old_page) {
2723                         if (!PageAnon(old_page)) {
2724                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2725                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2726                         }
2727                 } else
2728                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2729                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2730                 entry = mk_pte(new_page, vma->vm_page_prot);
2731                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2732                 /*
2733                  * Clear the pte entry and flush it first, before updating the
2734                  * pte with the new entry. This will avoid a race condition
2735                  * seen in the presence of one thread doing SMC and another
2736                  * thread doing COW.
2737                  */
2738                 ptep_clear_flush(vma, address, page_table);
2739                 page_add_new_anon_rmap(new_page, vma, address);
2740                 /*
2741                  * We call the notify macro here because, when using secondary
2742                  * mmu page tables (such as kvm shadow page tables), we want the
2743                  * new page to be mapped directly into the secondary page table.
2744                  */
2745                 set_pte_at_notify(mm, address, page_table, entry);
2746                 update_mmu_cache(vma, address, page_table);
2747                 if (old_page) {
2748                         /*
2749                          * Only after switching the pte to the new page may
2750                          * we remove the mapcount here. Otherwise another
2751                          * process may come and find the rmap count decremented
2752                          * before the pte is switched to the new page, and
2753                          * "reuse" the old page writing into it while our pte
2754                          * here still points into it and can be read by other
2755                          * threads.
2756                          *
2757                          * The critical issue is to order this
2758                          * page_remove_rmap with the ptp_clear_flush above.
2759                          * Those stores are ordered by (if nothing else,)
2760                          * the barrier present in the atomic_add_negative
2761                          * in page_remove_rmap.
2762                          *
2763                          * Then the TLB flush in ptep_clear_flush ensures that
2764                          * no process can access the old page before the
2765                          * decremented mapcount is visible. And the old page
2766                          * cannot be reused until after the decremented
2767                          * mapcount is visible. So transitively, TLBs to
2768                          * old page will be flushed before it can be reused.
2769                          */
2770                         page_remove_rmap(old_page);
2771                 }
2772
2773                 /* Free the old page.. */
2774                 new_page = old_page;
2775                 ret |= VM_FAULT_WRITE;
2776         } else
2777                 mem_cgroup_uncharge_page(new_page);
2778
2779         if (new_page)
2780                 page_cache_release(new_page);
2781 unlock:
2782         pte_unmap_unlock(page_table, ptl);
2783         if (mmun_end > mmun_start)
2784                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2785         if (old_page) {
2786                 /*
2787                  * Don't let another task, with possibly unlocked vma,
2788                  * keep the mlocked page.
2789                  */
2790                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2791                         lock_page(old_page);    /* LRU manipulation */
2792                         munlock_vma_page(old_page);
2793                         unlock_page(old_page);
2794                 }
2795                 page_cache_release(old_page);
2796         }
2797         return ret;
2798 oom_free_new:
2799         page_cache_release(new_page);
2800 oom:
2801         if (old_page)
2802                 page_cache_release(old_page);
2803         return VM_FAULT_OOM;
2804
2805 unwritable_page:
2806         page_cache_release(old_page);
2807         return ret;
2808 }
2809
2810 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2811                 unsigned long start_addr, unsigned long end_addr,
2812                 struct zap_details *details)
2813 {
2814         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2815 }
2816
2817 static inline void unmap_mapping_range_tree(struct rb_root *root,
2818                                             struct zap_details *details)
2819 {
2820         struct vm_area_struct *vma;
2821         pgoff_t vba, vea, zba, zea;
2822
2823         vma_interval_tree_foreach(vma, root,
2824                         details->first_index, details->last_index) {
2825
2826                 vba = vma->vm_pgoff;
2827                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2828                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2829                 zba = details->first_index;
2830                 if (zba < vba)
2831                         zba = vba;
2832                 zea = details->last_index;
2833                 if (zea > vea)
2834                         zea = vea;
2835
2836                 unmap_mapping_range_vma(vma,
2837                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2838                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2839                                 details);
2840         }
2841 }
2842
2843 static inline void unmap_mapping_range_list(struct list_head *head,
2844                                             struct zap_details *details)
2845 {
2846         struct vm_area_struct *vma;
2847
2848         /*
2849          * In nonlinear VMAs there is no correspondence between virtual address
2850          * offset and file offset.  So we must perform an exhaustive search
2851          * across *all* the pages in each nonlinear VMA, not just the pages
2852          * whose virtual address lies outside the file truncation point.
2853          */
2854         list_for_each_entry(vma, head, shared.nonlinear) {
2855                 details->nonlinear_vma = vma;
2856                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2857         }
2858 }
2859
2860 /**
2861  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2862  * @mapping: the address space containing mmaps to be unmapped.
2863  * @holebegin: byte in first page to unmap, relative to the start of
2864  * the underlying file.  This will be rounded down to a PAGE_SIZE
2865  * boundary.  Note that this is different from truncate_pagecache(), which
2866  * must keep the partial page.  In contrast, we must get rid of
2867  * partial pages.
2868  * @holelen: size of prospective hole in bytes.  This will be rounded
2869  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2870  * end of the file.
2871  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2872  * but 0 when invalidating pagecache, don't throw away private data.
2873  */
2874 void unmap_mapping_range(struct address_space *mapping,
2875                 loff_t const holebegin, loff_t const holelen, int even_cows)
2876 {
2877         struct zap_details details;
2878         pgoff_t hba = holebegin >> PAGE_SHIFT;
2879         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2880
2881         /* Check for overflow. */
2882         if (sizeof(holelen) > sizeof(hlen)) {
2883                 long long holeend =
2884                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2885                 if (holeend & ~(long long)ULONG_MAX)
2886                         hlen = ULONG_MAX - hba + 1;
2887         }
2888
2889         details.check_mapping = even_cows? NULL: mapping;
2890         details.nonlinear_vma = NULL;
2891         details.first_index = hba;
2892         details.last_index = hba + hlen - 1;
2893         if (details.last_index < details.first_index)
2894                 details.last_index = ULONG_MAX;
2895
2896
2897         mutex_lock(&mapping->i_mmap_mutex);
2898         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2899                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2900         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2901                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2902         mutex_unlock(&mapping->i_mmap_mutex);
2903 }
2904 EXPORT_SYMBOL(unmap_mapping_range);
2905
2906 /*
2907  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2908  * but allow concurrent faults), and pte mapped but not yet locked.
2909  * We return with mmap_sem still held, but pte unmapped and unlocked.
2910  */
2911 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2912                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2913                 unsigned int flags, pte_t orig_pte)
2914 {
2915         spinlock_t *ptl;
2916         struct page *page, *swapcache = NULL;
2917         swp_entry_t entry;
2918         pte_t pte;
2919         int locked;
2920         struct mem_cgroup *ptr;
2921         int exclusive = 0;
2922         int ret = 0;
2923
2924         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2925                 goto out;
2926
2927         entry = pte_to_swp_entry(orig_pte);
2928         if (unlikely(non_swap_entry(entry))) {
2929                 if (is_migration_entry(entry)) {
2930                         migration_entry_wait(mm, pmd, address);
2931                 } else if (is_hwpoison_entry(entry)) {
2932                         ret = VM_FAULT_HWPOISON;
2933                 } else {
2934                         print_bad_pte(vma, address, orig_pte, NULL);
2935                         ret = VM_FAULT_SIGBUS;
2936                 }
2937                 goto out;
2938         }
2939         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2940         page = lookup_swap_cache(entry);
2941         if (!page) {
2942                 page = swapin_readahead(entry,
2943                                         GFP_HIGHUSER_MOVABLE, vma, address);
2944                 if (!page) {
2945                         /*
2946                          * Back out if somebody else faulted in this pte
2947                          * while we released the pte lock.
2948                          */
2949                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2950                         if (likely(pte_same(*page_table, orig_pte)))
2951                                 ret = VM_FAULT_OOM;
2952                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2953                         goto unlock;
2954                 }
2955
2956                 /* Had to read the page from swap area: Major fault */
2957                 ret = VM_FAULT_MAJOR;
2958                 count_vm_event(PGMAJFAULT);
2959                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2960         } else if (PageHWPoison(page)) {
2961                 /*
2962                  * hwpoisoned dirty swapcache pages are kept for killing
2963                  * owner processes (which may be unknown at hwpoison time)
2964                  */
2965                 ret = VM_FAULT_HWPOISON;
2966                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2967                 goto out_release;
2968         }
2969
2970         locked = lock_page_or_retry(page, mm, flags);
2971
2972         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2973         if (!locked) {
2974                 ret |= VM_FAULT_RETRY;
2975                 goto out_release;
2976         }
2977
2978         /*
2979          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2980          * release the swapcache from under us.  The page pin, and pte_same
2981          * test below, are not enough to exclude that.  Even if it is still
2982          * swapcache, we need to check that the page's swap has not changed.
2983          */
2984         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2985                 goto out_page;
2986
2987         if (ksm_might_need_to_copy(page, vma, address)) {
2988                 swapcache = page;
2989                 page = ksm_does_need_to_copy(page, vma, address);
2990
2991                 if (unlikely(!page)) {
2992                         ret = VM_FAULT_OOM;
2993                         page = swapcache;
2994                         swapcache = NULL;
2995                         goto out_page;
2996                 }
2997         }
2998
2999         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3000                 ret = VM_FAULT_OOM;
3001                 goto out_page;
3002         }
3003
3004         /*
3005          * Back out if somebody else already faulted in this pte.
3006          */
3007         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3008         if (unlikely(!pte_same(*page_table, orig_pte)))
3009                 goto out_nomap;
3010
3011         if (unlikely(!PageUptodate(page))) {
3012                 ret = VM_FAULT_SIGBUS;
3013                 goto out_nomap;
3014         }
3015
3016         /*
3017          * The page isn't present yet, go ahead with the fault.
3018          *
3019          * Be careful about the sequence of operations here.
3020          * To get its accounting right, reuse_swap_page() must be called
3021          * while the page is counted on swap but not yet in mapcount i.e.
3022          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3023          * must be called after the swap_free(), or it will never succeed.
3024          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3025          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3026          * in page->private. In this case, a record in swap_cgroup  is silently
3027          * discarded at swap_free().
3028          */
3029
3030         inc_mm_counter_fast(mm, MM_ANONPAGES);
3031         dec_mm_counter_fast(mm, MM_SWAPENTS);
3032         pte = mk_pte(page, vma->vm_page_prot);
3033         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3034                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3035                 flags &= ~FAULT_FLAG_WRITE;
3036                 ret |= VM_FAULT_WRITE;
3037                 exclusive = 1;
3038         }
3039         flush_icache_page(vma, page);
3040         set_pte_at(mm, address, page_table, pte);
3041         do_page_add_anon_rmap(page, vma, address, exclusive);
3042         /* It's better to call commit-charge after rmap is established */
3043         mem_cgroup_commit_charge_swapin(page, ptr);
3044
3045         swap_free(entry);
3046         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3047                 try_to_free_swap(page);
3048         unlock_page(page);
3049         if (swapcache) {
3050                 /*
3051                  * Hold the lock to avoid the swap entry to be reused
3052                  * until we take the PT lock for the pte_same() check
3053                  * (to avoid false positives from pte_same). For
3054                  * further safety release the lock after the swap_free
3055                  * so that the swap count won't change under a
3056                  * parallel locked swapcache.
3057                  */
3058                 unlock_page(swapcache);
3059                 page_cache_release(swapcache);
3060         }
3061
3062         if (flags & FAULT_FLAG_WRITE) {
3063                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3064                 if (ret & VM_FAULT_ERROR)
3065                         ret &= VM_FAULT_ERROR;
3066                 goto out;
3067         }
3068
3069         /* No need to invalidate - it was non-present before */
3070         update_mmu_cache(vma, address, page_table);
3071 unlock:
3072         pte_unmap_unlock(page_table, ptl);
3073 out:
3074         return ret;
3075 out_nomap:
3076         mem_cgroup_cancel_charge_swapin(ptr);
3077         pte_unmap_unlock(page_table, ptl);
3078 out_page:
3079         unlock_page(page);
3080 out_release:
3081         page_cache_release(page);
3082         if (swapcache) {
3083                 unlock_page(swapcache);
3084                 page_cache_release(swapcache);
3085         }
3086         return ret;
3087 }
3088
3089 /*
3090  * This is like a special single-page "expand_{down|up}wards()",
3091  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3092  * doesn't hit another vma.
3093  */
3094 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3095 {
3096         address &= PAGE_MASK;
3097         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3098                 struct vm_area_struct *prev = vma->vm_prev;
3099
3100                 /*
3101                  * Is there a mapping abutting this one below?
3102                  *
3103                  * That's only ok if it's the same stack mapping
3104                  * that has gotten split..
3105                  */
3106                 if (prev && prev->vm_end == address)
3107                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3108
3109                 expand_downwards(vma, address - PAGE_SIZE);
3110         }
3111         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3112                 struct vm_area_struct *next = vma->vm_next;
3113
3114                 /* As VM_GROWSDOWN but s/below/above/ */
3115                 if (next && next->vm_start == address + PAGE_SIZE)
3116                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3117
3118                 expand_upwards(vma, address + PAGE_SIZE);
3119         }
3120         return 0;
3121 }
3122
3123 /*
3124  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3125  * but allow concurrent faults), and pte mapped but not yet locked.
3126  * We return with mmap_sem still held, but pte unmapped and unlocked.
3127  */
3128 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3129                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3130                 unsigned int flags)
3131 {
3132         struct page *page;
3133         spinlock_t *ptl;
3134         pte_t entry;
3135
3136         pte_unmap(page_table);
3137
3138         /* Check if we need to add a guard page to the stack */
3139         if (check_stack_guard_page(vma, address) < 0)
3140                 return VM_FAULT_SIGBUS;
3141
3142         /* Use the zero-page for reads */
3143         if (!(flags & FAULT_FLAG_WRITE)) {
3144                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3145                                                 vma->vm_page_prot));
3146                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3147                 if (!pte_none(*page_table))
3148                         goto unlock;
3149                 goto setpte;
3150         }
3151
3152         /* Allocate our own private page. */
3153         if (unlikely(anon_vma_prepare(vma)))
3154                 goto oom;
3155         page = alloc_zeroed_user_highpage_movable(vma, address);
3156         if (!page)
3157                 goto oom;
3158         __SetPageUptodate(page);
3159
3160         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3161                 goto oom_free_page;
3162
3163         entry = mk_pte(page, vma->vm_page_prot);
3164         if (vma->vm_flags & VM_WRITE)
3165                 entry = pte_mkwrite(pte_mkdirty(entry));
3166
3167         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3168         if (!pte_none(*page_table))
3169                 goto release;
3170
3171         inc_mm_counter_fast(mm, MM_ANONPAGES);
3172         page_add_new_anon_rmap(page, vma, address);
3173 setpte:
3174         set_pte_at(mm, address, page_table, entry);
3175
3176         /* No need to invalidate - it was non-present before */
3177         update_mmu_cache(vma, address, page_table);
3178 unlock:
3179         pte_unmap_unlock(page_table, ptl);
3180         return 0;
3181 release:
3182         mem_cgroup_uncharge_page(page);
3183         page_cache_release(page);
3184         goto unlock;
3185 oom_free_page:
3186         page_cache_release(page);
3187 oom:
3188         return VM_FAULT_OOM;
3189 }
3190
3191 /*
3192  * __do_fault() tries to create a new page mapping. It aggressively
3193  * tries to share with existing pages, but makes a separate copy if
3194  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3195  * the next page fault.
3196  *
3197  * As this is called only for pages that do not currently exist, we
3198  * do not need to flush old virtual caches or the TLB.
3199  *
3200  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3201  * but allow concurrent faults), and pte neither mapped nor locked.
3202  * We return with mmap_sem still held, but pte unmapped and unlocked.
3203  */
3204 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3205                 unsigned long address, pmd_t *pmd,
3206                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3207 {
3208         pte_t *page_table;
3209         spinlock_t *ptl;
3210         struct page *page;
3211         struct page *cow_page;
3212         pte_t entry;
3213         int anon = 0;
3214         struct page *dirty_page = NULL;
3215         struct vm_fault vmf;
3216         int ret;
3217         int page_mkwrite = 0;
3218
3219         /*
3220          * If we do COW later, allocate page befor taking lock_page()
3221          * on the file cache page. This will reduce lock holding time.
3222          */
3223         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3224
3225                 if (unlikely(anon_vma_prepare(vma)))
3226                         return VM_FAULT_OOM;
3227
3228                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3229                 if (!cow_page)
3230                         return VM_FAULT_OOM;
3231
3232                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3233                         page_cache_release(cow_page);
3234                         return VM_FAULT_OOM;
3235                 }
3236         } else
3237                 cow_page = NULL;
3238
3239         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3240         vmf.pgoff = pgoff;
3241         vmf.flags = flags;
3242         vmf.page = NULL;
3243
3244         ret = vma->vm_ops->fault(vma, &vmf);
3245         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3246                             VM_FAULT_RETRY)))
3247                 goto uncharge_out;
3248
3249         if (unlikely(PageHWPoison(vmf.page))) {
3250                 if (ret & VM_FAULT_LOCKED)
3251                         unlock_page(vmf.page);
3252                 ret = VM_FAULT_HWPOISON;
3253                 goto uncharge_out;
3254         }
3255
3256         /*
3257          * For consistency in subsequent calls, make the faulted page always
3258          * locked.
3259          */
3260         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3261                 lock_page(vmf.page);
3262         else
3263                 VM_BUG_ON(!PageLocked(vmf.page));
3264
3265         /*
3266          * Should we do an early C-O-W break?
3267          */
3268         page = vmf.page;
3269         if (flags & FAULT_FLAG_WRITE) {
3270                 if (!(vma->vm_flags & VM_SHARED)) {
3271                         page = cow_page;
3272                         anon = 1;
3273                         copy_user_highpage(page, vmf.page, address, vma);
3274                         __SetPageUptodate(page);
3275                 } else {
3276                         /*
3277                          * If the page will be shareable, see if the backing
3278                          * address space wants to know that the page is about
3279                          * to become writable
3280                          */
3281                         if (vma->vm_ops->page_mkwrite) {
3282                                 int tmp;
3283
3284                                 unlock_page(page);
3285                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3286                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3287                                 if (unlikely(tmp &
3288                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3289                                         ret = tmp;
3290                                         goto unwritable_page;
3291                                 }
3292                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3293                                         lock_page(page);
3294                                         if (!page->mapping) {
3295                                                 ret = 0; /* retry the fault */
3296                                                 unlock_page(page);
3297                                                 goto unwritable_page;
3298                                         }
3299                                 } else
3300                                         VM_BUG_ON(!PageLocked(page));
3301                                 page_mkwrite = 1;
3302                         }
3303                 }
3304
3305         }
3306
3307         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3308
3309         /*
3310          * This silly early PAGE_DIRTY setting removes a race
3311          * due to the bad i386 page protection. But it's valid
3312          * for other architectures too.
3313          *
3314          * Note that if FAULT_FLAG_WRITE is set, we either now have
3315          * an exclusive copy of the page, or this is a shared mapping,
3316          * so we can make it writable and dirty to avoid having to
3317          * handle that later.
3318          */
3319         /* Only go through if we didn't race with anybody else... */
3320         if (likely(pte_same(*page_table, orig_pte))) {
3321                 flush_icache_page(vma, page);
3322                 entry = mk_pte(page, vma->vm_page_prot);
3323                 if (flags & FAULT_FLAG_WRITE)
3324                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3325                 if (anon) {
3326                         inc_mm_counter_fast(mm, MM_ANONPAGES);