Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/trivial
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60
61 #include <asm/io.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
64 #include <asm/tlb.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
67
68 #include "internal.h"
69
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
73 struct page *mem_map;
74
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
77 #endif
78
79 unsigned long num_physpages;
80 /*
81  * A number of key systems in x86 including ioremap() rely on the assumption
82  * that high_memory defines the upper bound on direct map memory, then end
83  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
84  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85  * and ZONE_HIGHMEM.
86  */
87 void * high_memory;
88
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
91
92 /*
93  * Randomize the address space (stacks, mmaps, brk, etc.).
94  *
95  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96  *   as ancient (libc5 based) binaries can segfault. )
97  */
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
100                                         1;
101 #else
102                                         2;
103 #endif
104
105 static int __init disable_randmaps(char *s)
106 {
107         randomize_va_space = 0;
108         return 1;
109 }
110 __setup("norandmaps", disable_randmaps);
111
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
114
115 /*
116  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117  */
118 static int __init init_zero_pfn(void)
119 {
120         zero_pfn = page_to_pfn(ZERO_PAGE(0));
121         return 0;
122 }
123 core_initcall(init_zero_pfn);
124
125
126 #if defined(SPLIT_RSS_COUNTING)
127
128 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
129 {
130         int i;
131
132         for (i = 0; i < NR_MM_COUNTERS; i++) {
133                 if (task->rss_stat.count[i]) {
134                         add_mm_counter(mm, i, task->rss_stat.count[i]);
135                         task->rss_stat.count[i] = 0;
136                 }
137         }
138         task->rss_stat.events = 0;
139 }
140
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142 {
143         struct task_struct *task = current;
144
145         if (likely(task->mm == mm))
146                 task->rss_stat.count[member] += val;
147         else
148                 add_mm_counter(mm, member, val);
149 }
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH  (64)
155 static void check_sync_rss_stat(struct task_struct *task)
156 {
157         if (unlikely(task != current))
158                 return;
159         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160                 __sync_task_rss_stat(task, task->mm);
161 }
162
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
164 {
165         long val = 0;
166
167         /*
168          * Don't use task->mm here...for avoiding to use task_get_mm()..
169          * The caller must guarantee task->mm is not invalid.
170          */
171         val = atomic_long_read(&mm->rss_stat.count[member]);
172         /*
173          * counter is updated in asynchronous manner and may go to minus.
174          * But it's never be expected number for users.
175          */
176         if (val < 0)
177                 return 0;
178         return (unsigned long)val;
179 }
180
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182 {
183         __sync_task_rss_stat(task, mm);
184 }
185 #else
186
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
189
190 static void check_sync_rss_stat(struct task_struct *task)
191 {
192 }
193
194 #endif
195
196 /*
197  * If a p?d_bad entry is found while walking page tables, report
198  * the error, before resetting entry to p?d_none.  Usually (but
199  * very seldom) called out from the p?d_none_or_clear_bad macros.
200  */
201
202 void pgd_clear_bad(pgd_t *pgd)
203 {
204         pgd_ERROR(*pgd);
205         pgd_clear(pgd);
206 }
207
208 void pud_clear_bad(pud_t *pud)
209 {
210         pud_ERROR(*pud);
211         pud_clear(pud);
212 }
213
214 void pmd_clear_bad(pmd_t *pmd)
215 {
216         pmd_ERROR(*pmd);
217         pmd_clear(pmd);
218 }
219
220 /*
221  * Note: this doesn't free the actual pages themselves. That
222  * has been handled earlier when unmapping all the memory regions.
223  */
224 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
225                            unsigned long addr)
226 {
227         pgtable_t token = pmd_pgtable(*pmd);
228         pmd_clear(pmd);
229         pte_free_tlb(tlb, token, addr);
230         tlb->mm->nr_ptes--;
231 }
232
233 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
234                                 unsigned long addr, unsigned long end,
235                                 unsigned long floor, unsigned long ceiling)
236 {
237         pmd_t *pmd;
238         unsigned long next;
239         unsigned long start;
240
241         start = addr;
242         pmd = pmd_offset(pud, addr);
243         do {
244                 next = pmd_addr_end(addr, end);
245                 if (pmd_none_or_clear_bad(pmd))
246                         continue;
247                 free_pte_range(tlb, pmd, addr);
248         } while (pmd++, addr = next, addr != end);
249
250         start &= PUD_MASK;
251         if (start < floor)
252                 return;
253         if (ceiling) {
254                 ceiling &= PUD_MASK;
255                 if (!ceiling)
256                         return;
257         }
258         if (end - 1 > ceiling - 1)
259                 return;
260
261         pmd = pmd_offset(pud, start);
262         pud_clear(pud);
263         pmd_free_tlb(tlb, pmd, start);
264 }
265
266 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
267                                 unsigned long addr, unsigned long end,
268                                 unsigned long floor, unsigned long ceiling)
269 {
270         pud_t *pud;
271         unsigned long next;
272         unsigned long start;
273
274         start = addr;
275         pud = pud_offset(pgd, addr);
276         do {
277                 next = pud_addr_end(addr, end);
278                 if (pud_none_or_clear_bad(pud))
279                         continue;
280                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
281         } while (pud++, addr = next, addr != end);
282
283         start &= PGDIR_MASK;
284         if (start < floor)
285                 return;
286         if (ceiling) {
287                 ceiling &= PGDIR_MASK;
288                 if (!ceiling)
289                         return;
290         }
291         if (end - 1 > ceiling - 1)
292                 return;
293
294         pud = pud_offset(pgd, start);
295         pgd_clear(pgd);
296         pud_free_tlb(tlb, pud, start);
297 }
298
299 /*
300  * This function frees user-level page tables of a process.
301  *
302  * Must be called with pagetable lock held.
303  */
304 void free_pgd_range(struct mmu_gather *tlb,
305                         unsigned long addr, unsigned long end,
306                         unsigned long floor, unsigned long ceiling)
307 {
308         pgd_t *pgd;
309         unsigned long next;
310
311         /*
312          * The next few lines have given us lots of grief...
313          *
314          * Why are we testing PMD* at this top level?  Because often
315          * there will be no work to do at all, and we'd prefer not to
316          * go all the way down to the bottom just to discover that.
317          *
318          * Why all these "- 1"s?  Because 0 represents both the bottom
319          * of the address space and the top of it (using -1 for the
320          * top wouldn't help much: the masks would do the wrong thing).
321          * The rule is that addr 0 and floor 0 refer to the bottom of
322          * the address space, but end 0 and ceiling 0 refer to the top
323          * Comparisons need to use "end - 1" and "ceiling - 1" (though
324          * that end 0 case should be mythical).
325          *
326          * Wherever addr is brought up or ceiling brought down, we must
327          * be careful to reject "the opposite 0" before it confuses the
328          * subsequent tests.  But what about where end is brought down
329          * by PMD_SIZE below? no, end can't go down to 0 there.
330          *
331          * Whereas we round start (addr) and ceiling down, by different
332          * masks at different levels, in order to test whether a table
333          * now has no other vmas using it, so can be freed, we don't
334          * bother to round floor or end up - the tests don't need that.
335          */
336
337         addr &= PMD_MASK;
338         if (addr < floor) {
339                 addr += PMD_SIZE;
340                 if (!addr)
341                         return;
342         }
343         if (ceiling) {
344                 ceiling &= PMD_MASK;
345                 if (!ceiling)
346                         return;
347         }
348         if (end - 1 > ceiling - 1)
349                 end -= PMD_SIZE;
350         if (addr > end - 1)
351                 return;
352
353         pgd = pgd_offset(tlb->mm, addr);
354         do {
355                 next = pgd_addr_end(addr, end);
356                 if (pgd_none_or_clear_bad(pgd))
357                         continue;
358                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
359         } while (pgd++, addr = next, addr != end);
360 }
361
362 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
363                 unsigned long floor, unsigned long ceiling)
364 {
365         while (vma) {
366                 struct vm_area_struct *next = vma->vm_next;
367                 unsigned long addr = vma->vm_start;
368
369                 /*
370                  * Hide vma from rmap and truncate_pagecache before freeing
371                  * pgtables
372                  */
373                 unlink_anon_vmas(vma);
374                 unlink_file_vma(vma);
375
376                 if (is_vm_hugetlb_page(vma)) {
377                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
378                                 floor, next? next->vm_start: ceiling);
379                 } else {
380                         /*
381                          * Optimization: gather nearby vmas into one call down
382                          */
383                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
384                                && !is_vm_hugetlb_page(next)) {
385                                 vma = next;
386                                 next = vma->vm_next;
387                                 unlink_anon_vmas(vma);
388                                 unlink_file_vma(vma);
389                         }
390                         free_pgd_range(tlb, addr, vma->vm_end,
391                                 floor, next? next->vm_start: ceiling);
392                 }
393                 vma = next;
394         }
395 }
396
397 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
398                 pmd_t *pmd, unsigned long address)
399 {
400         pgtable_t new = pte_alloc_one(mm, address);
401         int wait_split_huge_page;
402         if (!new)
403                 return -ENOMEM;
404
405         /*
406          * Ensure all pte setup (eg. pte page lock and page clearing) are
407          * visible before the pte is made visible to other CPUs by being
408          * put into page tables.
409          *
410          * The other side of the story is the pointer chasing in the page
411          * table walking code (when walking the page table without locking;
412          * ie. most of the time). Fortunately, these data accesses consist
413          * of a chain of data-dependent loads, meaning most CPUs (alpha
414          * being the notable exception) will already guarantee loads are
415          * seen in-order. See the alpha page table accessors for the
416          * smp_read_barrier_depends() barriers in page table walking code.
417          */
418         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
419
420         spin_lock(&mm->page_table_lock);
421         wait_split_huge_page = 0;
422         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
423                 mm->nr_ptes++;
424                 pmd_populate(mm, pmd, new);
425                 new = NULL;
426         } else if (unlikely(pmd_trans_splitting(*pmd)))
427                 wait_split_huge_page = 1;
428         spin_unlock(&mm->page_table_lock);
429         if (new)
430                 pte_free(mm, new);
431         if (wait_split_huge_page)
432                 wait_split_huge_page(vma->anon_vma, pmd);
433         return 0;
434 }
435
436 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
437 {
438         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
439         if (!new)
440                 return -ENOMEM;
441
442         smp_wmb(); /* See comment in __pte_alloc */
443
444         spin_lock(&init_mm.page_table_lock);
445         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
446                 pmd_populate_kernel(&init_mm, pmd, new);
447                 new = NULL;
448         } else
449                 VM_BUG_ON(pmd_trans_splitting(*pmd));
450         spin_unlock(&init_mm.page_table_lock);
451         if (new)
452                 pte_free_kernel(&init_mm, new);
453         return 0;
454 }
455
456 static inline void init_rss_vec(int *rss)
457 {
458         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
459 }
460
461 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
462 {
463         int i;
464
465         if (current->mm == mm)
466                 sync_mm_rss(current, mm);
467         for (i = 0; i < NR_MM_COUNTERS; i++)
468                 if (rss[i])
469                         add_mm_counter(mm, i, rss[i]);
470 }
471
472 /*
473  * This function is called to print an error when a bad pte
474  * is found. For example, we might have a PFN-mapped pte in
475  * a region that doesn't allow it.
476  *
477  * The calling function must still handle the error.
478  */
479 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
480                           pte_t pte, struct page *page)
481 {
482         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
483         pud_t *pud = pud_offset(pgd, addr);
484         pmd_t *pmd = pmd_offset(pud, addr);
485         struct address_space *mapping;
486         pgoff_t index;
487         static unsigned long resume;
488         static unsigned long nr_shown;
489         static unsigned long nr_unshown;
490
491         /*
492          * Allow a burst of 60 reports, then keep quiet for that minute;
493          * or allow a steady drip of one report per second.
494          */
495         if (nr_shown == 60) {
496                 if (time_before(jiffies, resume)) {
497                         nr_unshown++;
498                         return;
499                 }
500                 if (nr_unshown) {
501                         printk(KERN_ALERT
502                                 "BUG: Bad page map: %lu messages suppressed\n",
503                                 nr_unshown);
504                         nr_unshown = 0;
505                 }
506                 nr_shown = 0;
507         }
508         if (nr_shown++ == 0)
509                 resume = jiffies + 60 * HZ;
510
511         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
512         index = linear_page_index(vma, addr);
513
514         printk(KERN_ALERT
515                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
516                 current->comm,
517                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
518         if (page)
519                 dump_page(page);
520         printk(KERN_ALERT
521                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
522                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
523         /*
524          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
525          */
526         if (vma->vm_ops)
527                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
528                                 (unsigned long)vma->vm_ops->fault);
529         if (vma->vm_file && vma->vm_file->f_op)
530                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
531                                 (unsigned long)vma->vm_file->f_op->mmap);
532         dump_stack();
533         add_taint(TAINT_BAD_PAGE);
534 }
535
536 static inline int is_cow_mapping(unsigned int flags)
537 {
538         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
539 }
540
541 #ifndef is_zero_pfn
542 static inline int is_zero_pfn(unsigned long pfn)
543 {
544         return pfn == zero_pfn;
545 }
546 #endif
547
548 #ifndef my_zero_pfn
549 static inline unsigned long my_zero_pfn(unsigned long addr)
550 {
551         return zero_pfn;
552 }
553 #endif
554
555 /*
556  * vm_normal_page -- This function gets the "struct page" associated with a pte.
557  *
558  * "Special" mappings do not wish to be associated with a "struct page" (either
559  * it doesn't exist, or it exists but they don't want to touch it). In this
560  * case, NULL is returned here. "Normal" mappings do have a struct page.
561  *
562  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
563  * pte bit, in which case this function is trivial. Secondly, an architecture
564  * may not have a spare pte bit, which requires a more complicated scheme,
565  * described below.
566  *
567  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
568  * special mapping (even if there are underlying and valid "struct pages").
569  * COWed pages of a VM_PFNMAP are always normal.
570  *
571  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
572  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
573  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
574  * mapping will always honor the rule
575  *
576  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
577  *
578  * And for normal mappings this is false.
579  *
580  * This restricts such mappings to be a linear translation from virtual address
581  * to pfn. To get around this restriction, we allow arbitrary mappings so long
582  * as the vma is not a COW mapping; in that case, we know that all ptes are
583  * special (because none can have been COWed).
584  *
585  *
586  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
587  *
588  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
589  * page" backing, however the difference is that _all_ pages with a struct
590  * page (that is, those where pfn_valid is true) are refcounted and considered
591  * normal pages by the VM. The disadvantage is that pages are refcounted
592  * (which can be slower and simply not an option for some PFNMAP users). The
593  * advantage is that we don't have to follow the strict linearity rule of
594  * PFNMAP mappings in order to support COWable mappings.
595  *
596  */
597 #ifdef __HAVE_ARCH_PTE_SPECIAL
598 # define HAVE_PTE_SPECIAL 1
599 #else
600 # define HAVE_PTE_SPECIAL 0
601 #endif
602 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
603                                 pte_t pte)
604 {
605         unsigned long pfn = pte_pfn(pte);
606
607         if (HAVE_PTE_SPECIAL) {
608                 if (likely(!pte_special(pte)))
609                         goto check_pfn;
610                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
611                         return NULL;
612                 if (!is_zero_pfn(pfn))
613                         print_bad_pte(vma, addr, pte, NULL);
614                 return NULL;
615         }
616
617         /* !HAVE_PTE_SPECIAL case follows: */
618
619         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
620                 if (vma->vm_flags & VM_MIXEDMAP) {
621                         if (!pfn_valid(pfn))
622                                 return NULL;
623                         goto out;
624                 } else {
625                         unsigned long off;
626                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
627                         if (pfn == vma->vm_pgoff + off)
628                                 return NULL;
629                         if (!is_cow_mapping(vma->vm_flags))
630                                 return NULL;
631                 }
632         }
633
634         if (is_zero_pfn(pfn))
635                 return NULL;
636 check_pfn:
637         if (unlikely(pfn > highest_memmap_pfn)) {
638                 print_bad_pte(vma, addr, pte, NULL);
639                 return NULL;
640         }
641
642         /*
643          * NOTE! We still have PageReserved() pages in the page tables.
644          * eg. VDSO mappings can cause them to exist.
645          */
646 out:
647         return pfn_to_page(pfn);
648 }
649
650 /*
651  * copy one vm_area from one task to the other. Assumes the page tables
652  * already present in the new task to be cleared in the whole range
653  * covered by this vma.
654  */
655
656 static inline unsigned long
657 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
658                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
659                 unsigned long addr, int *rss)
660 {
661         unsigned long vm_flags = vma->vm_flags;
662         pte_t pte = *src_pte;
663         struct page *page;
664
665         /* pte contains position in swap or file, so copy. */
666         if (unlikely(!pte_present(pte))) {
667                 if (!pte_file(pte)) {
668                         swp_entry_t entry = pte_to_swp_entry(pte);
669
670                         if (swap_duplicate(entry) < 0)
671                                 return entry.val;
672
673                         /* make sure dst_mm is on swapoff's mmlist. */
674                         if (unlikely(list_empty(&dst_mm->mmlist))) {
675                                 spin_lock(&mmlist_lock);
676                                 if (list_empty(&dst_mm->mmlist))
677                                         list_add(&dst_mm->mmlist,
678                                                  &src_mm->mmlist);
679                                 spin_unlock(&mmlist_lock);
680                         }
681                         if (likely(!non_swap_entry(entry)))
682                                 rss[MM_SWAPENTS]++;
683                         else if (is_write_migration_entry(entry) &&
684                                         is_cow_mapping(vm_flags)) {
685                                 /*
686                                  * COW mappings require pages in both parent
687                                  * and child to be set to read.
688                                  */
689                                 make_migration_entry_read(&entry);
690                                 pte = swp_entry_to_pte(entry);
691                                 set_pte_at(src_mm, addr, src_pte, pte);
692                         }
693                 }
694                 goto out_set_pte;
695         }
696
697         /*
698          * If it's a COW mapping, write protect it both
699          * in the parent and the child
700          */
701         if (is_cow_mapping(vm_flags)) {
702                 ptep_set_wrprotect(src_mm, addr, src_pte);
703                 pte = pte_wrprotect(pte);
704         }
705
706         /*
707          * If it's a shared mapping, mark it clean in
708          * the child
709          */
710         if (vm_flags & VM_SHARED)
711                 pte = pte_mkclean(pte);
712         pte = pte_mkold(pte);
713
714         page = vm_normal_page(vma, addr, pte);
715         if (page) {
716                 get_page(page);
717                 page_dup_rmap(page);
718                 if (PageAnon(page))
719                         rss[MM_ANONPAGES]++;
720                 else
721                         rss[MM_FILEPAGES]++;
722         }
723
724 out_set_pte:
725         set_pte_at(dst_mm, addr, dst_pte, pte);
726         return 0;
727 }
728
729 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
730                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
731                    unsigned long addr, unsigned long end)
732 {
733         pte_t *orig_src_pte, *orig_dst_pte;
734         pte_t *src_pte, *dst_pte;
735         spinlock_t *src_ptl, *dst_ptl;
736         int progress = 0;
737         int rss[NR_MM_COUNTERS];
738         swp_entry_t entry = (swp_entry_t){0};
739
740 again:
741         init_rss_vec(rss);
742
743         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
744         if (!dst_pte)
745                 return -ENOMEM;
746         src_pte = pte_offset_map(src_pmd, addr);
747         src_ptl = pte_lockptr(src_mm, src_pmd);
748         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
749         orig_src_pte = src_pte;
750         orig_dst_pte = dst_pte;
751         arch_enter_lazy_mmu_mode();
752
753         do {
754                 /*
755                  * We are holding two locks at this point - either of them
756                  * could generate latencies in another task on another CPU.
757                  */
758                 if (progress >= 32) {
759                         progress = 0;
760                         if (need_resched() ||
761                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
762                                 break;
763                 }
764                 if (pte_none(*src_pte)) {
765                         progress++;
766                         continue;
767                 }
768                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
769                                                         vma, addr, rss);
770                 if (entry.val)
771                         break;
772                 progress += 8;
773         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
774
775         arch_leave_lazy_mmu_mode();
776         spin_unlock(src_ptl);
777         pte_unmap(orig_src_pte);
778         add_mm_rss_vec(dst_mm, rss);
779         pte_unmap_unlock(orig_dst_pte, dst_ptl);
780         cond_resched();
781
782         if (entry.val) {
783                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
784                         return -ENOMEM;
785                 progress = 0;
786         }
787         if (addr != end)
788                 goto again;
789         return 0;
790 }
791
792 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
793                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
794                 unsigned long addr, unsigned long end)
795 {
796         pmd_t *src_pmd, *dst_pmd;
797         unsigned long next;
798
799         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
800         if (!dst_pmd)
801                 return -ENOMEM;
802         src_pmd = pmd_offset(src_pud, addr);
803         do {
804                 next = pmd_addr_end(addr, end);
805                 if (pmd_trans_huge(*src_pmd)) {
806                         int err;
807                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
808                         err = copy_huge_pmd(dst_mm, src_mm,
809                                             dst_pmd, src_pmd, addr, vma);
810                         if (err == -ENOMEM)
811                                 return -ENOMEM;
812                         if (!err)
813                                 continue;
814                         /* fall through */
815                 }
816                 if (pmd_none_or_clear_bad(src_pmd))
817                         continue;
818                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
819                                                 vma, addr, next))
820                         return -ENOMEM;
821         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
822         return 0;
823 }
824
825 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
826                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
827                 unsigned long addr, unsigned long end)
828 {
829         pud_t *src_pud, *dst_pud;
830         unsigned long next;
831
832         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
833         if (!dst_pud)
834                 return -ENOMEM;
835         src_pud = pud_offset(src_pgd, addr);
836         do {
837                 next = pud_addr_end(addr, end);
838                 if (pud_none_or_clear_bad(src_pud))
839                         continue;
840                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
841                                                 vma, addr, next))
842                         return -ENOMEM;
843         } while (dst_pud++, src_pud++, addr = next, addr != end);
844         return 0;
845 }
846
847 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
848                 struct vm_area_struct *vma)
849 {
850         pgd_t *src_pgd, *dst_pgd;
851         unsigned long next;
852         unsigned long addr = vma->vm_start;
853         unsigned long end = vma->vm_end;
854         int ret;
855
856         /*
857          * Don't copy ptes where a page fault will fill them correctly.
858          * Fork becomes much lighter when there are big shared or private
859          * readonly mappings. The tradeoff is that copy_page_range is more
860          * efficient than faulting.
861          */
862         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
863                 if (!vma->anon_vma)
864                         return 0;
865         }
866
867         if (is_vm_hugetlb_page(vma))
868                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
869
870         if (unlikely(is_pfn_mapping(vma))) {
871                 /*
872                  * We do not free on error cases below as remove_vma
873                  * gets called on error from higher level routine
874                  */
875                 ret = track_pfn_vma_copy(vma);
876                 if (ret)
877                         return ret;
878         }
879
880         /*
881          * We need to invalidate the secondary MMU mappings only when
882          * there could be a permission downgrade on the ptes of the
883          * parent mm. And a permission downgrade will only happen if
884          * is_cow_mapping() returns true.
885          */
886         if (is_cow_mapping(vma->vm_flags))
887                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
888
889         ret = 0;
890         dst_pgd = pgd_offset(dst_mm, addr);
891         src_pgd = pgd_offset(src_mm, addr);
892         do {
893                 next = pgd_addr_end(addr, end);
894                 if (pgd_none_or_clear_bad(src_pgd))
895                         continue;
896                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
897                                             vma, addr, next))) {
898                         ret = -ENOMEM;
899                         break;
900                 }
901         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
902
903         if (is_cow_mapping(vma->vm_flags))
904                 mmu_notifier_invalidate_range_end(src_mm,
905                                                   vma->vm_start, end);
906         return ret;
907 }
908
909 static unsigned long zap_pte_range(struct mmu_gather *tlb,
910                                 struct vm_area_struct *vma, pmd_t *pmd,
911                                 unsigned long addr, unsigned long end,
912                                 long *zap_work, struct zap_details *details)
913 {
914         struct mm_struct *mm = tlb->mm;
915         pte_t *pte;
916         spinlock_t *ptl;
917         int rss[NR_MM_COUNTERS];
918
919         init_rss_vec(rss);
920
921         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
922         arch_enter_lazy_mmu_mode();
923         do {
924                 pte_t ptent = *pte;
925                 if (pte_none(ptent)) {
926                         (*zap_work)--;
927                         continue;
928                 }
929
930                 (*zap_work) -= PAGE_SIZE;
931
932                 if (pte_present(ptent)) {
933                         struct page *page;
934
935                         page = vm_normal_page(vma, addr, ptent);
936                         if (unlikely(details) && page) {
937                                 /*
938                                  * unmap_shared_mapping_pages() wants to
939                                  * invalidate cache without truncating:
940                                  * unmap shared but keep private pages.
941                                  */
942                                 if (details->check_mapping &&
943                                     details->check_mapping != page->mapping)
944                                         continue;
945                                 /*
946                                  * Each page->index must be checked when
947                                  * invalidating or truncating nonlinear.
948                                  */
949                                 if (details->nonlinear_vma &&
950                                     (page->index < details->first_index ||
951                                      page->index > details->last_index))
952                                         continue;
953                         }
954                         ptent = ptep_get_and_clear_full(mm, addr, pte,
955                                                         tlb->fullmm);
956                         tlb_remove_tlb_entry(tlb, pte, addr);
957                         if (unlikely(!page))
958                                 continue;
959                         if (unlikely(details) && details->nonlinear_vma
960                             && linear_page_index(details->nonlinear_vma,
961                                                 addr) != page->index)
962                                 set_pte_at(mm, addr, pte,
963                                            pgoff_to_pte(page->index));
964                         if (PageAnon(page))
965                                 rss[MM_ANONPAGES]--;
966                         else {
967                                 if (pte_dirty(ptent))
968                                         set_page_dirty(page);
969                                 if (pte_young(ptent) &&
970                                     likely(!VM_SequentialReadHint(vma)))
971                                         mark_page_accessed(page);
972                                 rss[MM_FILEPAGES]--;
973                         }
974                         page_remove_rmap(page);
975                         if (unlikely(page_mapcount(page) < 0))
976                                 print_bad_pte(vma, addr, ptent, page);
977                         tlb_remove_page(tlb, page);
978                         continue;
979                 }
980                 /*
981                  * If details->check_mapping, we leave swap entries;
982                  * if details->nonlinear_vma, we leave file entries.
983                  */
984                 if (unlikely(details))
985                         continue;
986                 if (pte_file(ptent)) {
987                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
988                                 print_bad_pte(vma, addr, ptent, NULL);
989                 } else {
990                         swp_entry_t entry = pte_to_swp_entry(ptent);
991
992                         if (!non_swap_entry(entry))
993                                 rss[MM_SWAPENTS]--;
994                         if (unlikely(!free_swap_and_cache(entry)))
995                                 print_bad_pte(vma, addr, ptent, NULL);
996                 }
997                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
998         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
999
1000         add_mm_rss_vec(mm, rss);
1001         arch_leave_lazy_mmu_mode();
1002         pte_unmap_unlock(pte - 1, ptl);
1003
1004         return addr;
1005 }
1006
1007 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1008                                 struct vm_area_struct *vma, pud_t *pud,
1009                                 unsigned long addr, unsigned long end,
1010                                 long *zap_work, struct zap_details *details)
1011 {
1012         pmd_t *pmd;
1013         unsigned long next;
1014
1015         pmd = pmd_offset(pud, addr);
1016         do {
1017                 next = pmd_addr_end(addr, end);
1018                 if (pmd_trans_huge(*pmd)) {
1019                         if (next-addr != HPAGE_PMD_SIZE) {
1020                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1021                                 split_huge_page_pmd(vma->vm_mm, pmd);
1022                         } else if (zap_huge_pmd(tlb, vma, pmd)) {
1023                                 (*zap_work)--;
1024                                 continue;
1025                         }
1026                         /* fall through */
1027                 }
1028                 if (pmd_none_or_clear_bad(pmd)) {
1029                         (*zap_work)--;
1030                         continue;
1031                 }
1032                 next = zap_pte_range(tlb, vma, pmd, addr, next,
1033                                                 zap_work, details);
1034         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
1035
1036         return addr;
1037 }
1038
1039 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1040                                 struct vm_area_struct *vma, pgd_t *pgd,
1041                                 unsigned long addr, unsigned long end,
1042                                 long *zap_work, struct zap_details *details)
1043 {
1044         pud_t *pud;
1045         unsigned long next;
1046
1047         pud = pud_offset(pgd, addr);
1048         do {
1049                 next = pud_addr_end(addr, end);
1050                 if (pud_none_or_clear_bad(pud)) {
1051                         (*zap_work)--;
1052                         continue;
1053                 }
1054                 next = zap_pmd_range(tlb, vma, pud, addr, next,
1055                                                 zap_work, details);
1056         } while (pud++, addr = next, (addr != end && *zap_work > 0));
1057
1058         return addr;
1059 }
1060
1061 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1062                                 struct vm_area_struct *vma,
1063                                 unsigned long addr, unsigned long end,
1064                                 long *zap_work, struct zap_details *details)
1065 {
1066         pgd_t *pgd;
1067         unsigned long next;
1068
1069         if (details && !details->check_mapping && !details->nonlinear_vma)
1070                 details = NULL;
1071
1072         BUG_ON(addr >= end);
1073         mem_cgroup_uncharge_start();
1074         tlb_start_vma(tlb, vma);
1075         pgd = pgd_offset(vma->vm_mm, addr);
1076         do {
1077                 next = pgd_addr_end(addr, end);
1078                 if (pgd_none_or_clear_bad(pgd)) {
1079                         (*zap_work)--;
1080                         continue;
1081                 }
1082                 next = zap_pud_range(tlb, vma, pgd, addr, next,
1083                                                 zap_work, details);
1084         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
1085         tlb_end_vma(tlb, vma);
1086         mem_cgroup_uncharge_end();
1087
1088         return addr;
1089 }
1090
1091 #ifdef CONFIG_PREEMPT
1092 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1093 #else
1094 /* No preempt: go for improved straight-line efficiency */
1095 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1096 #endif
1097
1098 /**
1099  * unmap_vmas - unmap a range of memory covered by a list of vma's
1100  * @tlbp: address of the caller's struct mmu_gather
1101  * @vma: the starting vma
1102  * @start_addr: virtual address at which to start unmapping
1103  * @end_addr: virtual address at which to end unmapping
1104  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1105  * @details: details of nonlinear truncation or shared cache invalidation
1106  *
1107  * Returns the end address of the unmapping (restart addr if interrupted).
1108  *
1109  * Unmap all pages in the vma list.
1110  *
1111  * We aim to not hold locks for too long (for scheduling latency reasons).
1112  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
1113  * return the ending mmu_gather to the caller.
1114  *
1115  * Only addresses between `start' and `end' will be unmapped.
1116  *
1117  * The VMA list must be sorted in ascending virtual address order.
1118  *
1119  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1120  * range after unmap_vmas() returns.  So the only responsibility here is to
1121  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1122  * drops the lock and schedules.
1123  */
1124 unsigned long unmap_vmas(struct mmu_gather **tlbp,
1125                 struct vm_area_struct *vma, unsigned long start_addr,
1126                 unsigned long end_addr, unsigned long *nr_accounted,
1127                 struct zap_details *details)
1128 {
1129         long zap_work = ZAP_BLOCK_SIZE;
1130         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
1131         int tlb_start_valid = 0;
1132         unsigned long start = start_addr;
1133         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1134         int fullmm = (*tlbp)->fullmm;
1135         struct mm_struct *mm = vma->vm_mm;
1136
1137         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1138         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1139                 unsigned long end;
1140
1141                 start = max(vma->vm_start, start_addr);
1142                 if (start >= vma->vm_end)
1143                         continue;
1144                 end = min(vma->vm_end, end_addr);
1145                 if (end <= vma->vm_start)
1146                         continue;
1147
1148                 if (vma->vm_flags & VM_ACCOUNT)
1149                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1150
1151                 if (unlikely(is_pfn_mapping(vma)))
1152                         untrack_pfn_vma(vma, 0, 0);
1153
1154                 while (start != end) {
1155                         if (!tlb_start_valid) {
1156                                 tlb_start = start;
1157                                 tlb_start_valid = 1;
1158                         }
1159
1160                         if (unlikely(is_vm_hugetlb_page(vma))) {
1161                                 /*
1162                                  * It is undesirable to test vma->vm_file as it
1163                                  * should be non-null for valid hugetlb area.
1164                                  * However, vm_file will be NULL in the error
1165                                  * cleanup path of do_mmap_pgoff. When
1166                                  * hugetlbfs ->mmap method fails,
1167                                  * do_mmap_pgoff() nullifies vma->vm_file
1168                                  * before calling this function to clean up.
1169                                  * Since no pte has actually been setup, it is
1170                                  * safe to do nothing in this case.
1171                                  */
1172                                 if (vma->vm_file) {
1173                                         unmap_hugepage_range(vma, start, end, NULL);
1174                                         zap_work -= (end - start) /
1175                                         pages_per_huge_page(hstate_vma(vma));
1176                                 }
1177
1178                                 start = end;
1179                         } else
1180                                 start = unmap_page_range(*tlbp, vma,
1181                                                 start, end, &zap_work, details);
1182
1183                         if (zap_work > 0) {
1184                                 BUG_ON(start != end);
1185                                 break;
1186                         }
1187
1188                         tlb_finish_mmu(*tlbp, tlb_start, start);
1189
1190                         if (need_resched() ||
1191                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1192                                 if (i_mmap_lock) {
1193                                         *tlbp = NULL;
1194                                         goto out;
1195                                 }
1196                                 cond_resched();
1197                         }
1198
1199                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1200                         tlb_start_valid = 0;
1201                         zap_work = ZAP_BLOCK_SIZE;
1202                 }
1203         }
1204 out:
1205         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1206         return start;   /* which is now the end (or restart) address */
1207 }
1208
1209 /**
1210  * zap_page_range - remove user pages in a given range
1211  * @vma: vm_area_struct holding the applicable pages
1212  * @address: starting address of pages to zap
1213  * @size: number of bytes to zap
1214  * @details: details of nonlinear truncation or shared cache invalidation
1215  */
1216 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1217                 unsigned long size, struct zap_details *details)
1218 {
1219         struct mm_struct *mm = vma->vm_mm;
1220         struct mmu_gather *tlb;
1221         unsigned long end = address + size;
1222         unsigned long nr_accounted = 0;
1223
1224         lru_add_drain();
1225         tlb = tlb_gather_mmu(mm, 0);
1226         update_hiwater_rss(mm);
1227         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1228         if (tlb)
1229                 tlb_finish_mmu(tlb, address, end);
1230         return end;
1231 }
1232
1233 /**
1234  * zap_vma_ptes - remove ptes mapping the vma
1235  * @vma: vm_area_struct holding ptes to be zapped
1236  * @address: starting address of pages to zap
1237  * @size: number of bytes to zap
1238  *
1239  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1240  *
1241  * The entire address range must be fully contained within the vma.
1242  *
1243  * Returns 0 if successful.
1244  */
1245 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1246                 unsigned long size)
1247 {
1248         if (address < vma->vm_start || address + size > vma->vm_end ||
1249                         !(vma->vm_flags & VM_PFNMAP))
1250                 return -1;
1251         zap_page_range(vma, address, size, NULL);
1252         return 0;
1253 }
1254 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1255
1256 /**
1257  * follow_page - look up a page descriptor from a user-virtual address
1258  * @vma: vm_area_struct mapping @address
1259  * @address: virtual address to look up
1260  * @flags: flags modifying lookup behaviour
1261  *
1262  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1263  *
1264  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1265  * an error pointer if there is a mapping to something not represented
1266  * by a page descriptor (see also vm_normal_page()).
1267  */
1268 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1269                         unsigned int flags)
1270 {
1271         pgd_t *pgd;
1272         pud_t *pud;
1273         pmd_t *pmd;
1274         pte_t *ptep, pte;
1275         spinlock_t *ptl;
1276         struct page *page;
1277         struct mm_struct *mm = vma->vm_mm;
1278
1279         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1280         if (!IS_ERR(page)) {
1281                 BUG_ON(flags & FOLL_GET);
1282                 goto out;
1283         }
1284
1285         page = NULL;
1286         pgd = pgd_offset(mm, address);
1287         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1288                 goto no_page_table;
1289
1290         pud = pud_offset(pgd, address);
1291         if (pud_none(*pud))
1292                 goto no_page_table;
1293         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1294                 BUG_ON(flags & FOLL_GET);
1295                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1296                 goto out;
1297         }
1298         if (unlikely(pud_bad(*pud)))
1299                 goto no_page_table;
1300
1301         pmd = pmd_offset(pud, address);
1302         if (pmd_none(*pmd))
1303                 goto no_page_table;
1304         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1305                 BUG_ON(flags & FOLL_GET);
1306                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1307                 goto out;
1308         }
1309         if (pmd_trans_huge(*pmd)) {
1310                 if (flags & FOLL_SPLIT) {
1311                         split_huge_page_pmd(mm, pmd);
1312                         goto split_fallthrough;
1313                 }
1314                 spin_lock(&mm->page_table_lock);
1315                 if (likely(pmd_trans_huge(*pmd))) {
1316                         if (unlikely(pmd_trans_splitting(*pmd))) {
1317                                 spin_unlock(&mm->page_table_lock);
1318                                 wait_split_huge_page(vma->anon_vma, pmd);
1319                         } else {
1320                                 page = follow_trans_huge_pmd(mm, address,
1321                                                              pmd, flags);
1322                                 spin_unlock(&mm->page_table_lock);
1323                                 goto out;
1324                         }
1325                 } else
1326                         spin_unlock(&mm->page_table_lock);
1327                 /* fall through */
1328         }
1329 split_fallthrough:
1330         if (unlikely(pmd_bad(*pmd)))
1331                 goto no_page_table;
1332
1333         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1334
1335         pte = *ptep;
1336         if (!pte_present(pte))
1337                 goto no_page;
1338         if ((flags & FOLL_WRITE) && !pte_write(pte))
1339                 goto unlock;
1340
1341         page = vm_normal_page(vma, address, pte);
1342         if (unlikely(!page)) {
1343                 if ((flags & FOLL_DUMP) ||
1344                     !is_zero_pfn(pte_pfn(pte)))
1345                         goto bad_page;
1346                 page = pte_page(pte);
1347         }
1348
1349         if (flags & FOLL_GET)
1350                 get_page(page);
1351         if (flags & FOLL_TOUCH) {
1352                 if ((flags & FOLL_WRITE) &&
1353                     !pte_dirty(pte) && !PageDirty(page))
1354                         set_page_dirty(page);
1355                 /*
1356                  * pte_mkyoung() would be more correct here, but atomic care
1357                  * is needed to avoid losing the dirty bit: it is easier to use
1358                  * mark_page_accessed().
1359                  */
1360                 mark_page_accessed(page);
1361         }
1362         if (flags & FOLL_MLOCK) {
1363                 /*
1364                  * The preliminary mapping check is mainly to avoid the
1365                  * pointless overhead of lock_page on the ZERO_PAGE
1366                  * which might bounce very badly if there is contention.
1367                  *
1368                  * If the page is already locked, we don't need to
1369                  * handle it now - vmscan will handle it later if and
1370                  * when it attempts to reclaim the page.
1371                  */
1372                 if (page->mapping && trylock_page(page)) {
1373                         lru_add_drain();  /* push cached pages to LRU */
1374                         /*
1375                          * Because we lock page here and migration is
1376                          * blocked by the pte's page reference, we need
1377                          * only check for file-cache page truncation.
1378                          */
1379                         if (page->mapping)
1380                                 mlock_vma_page(page);
1381                         unlock_page(page);
1382                 }
1383         }
1384 unlock:
1385         pte_unmap_unlock(ptep, ptl);
1386 out:
1387         return page;
1388
1389 bad_page:
1390         pte_unmap_unlock(ptep, ptl);
1391         return ERR_PTR(-EFAULT);
1392
1393 no_page:
1394         pte_unmap_unlock(ptep, ptl);
1395         if (!pte_none(pte))
1396                 return page;
1397
1398 no_page_table:
1399         /*
1400          * When core dumping an enormous anonymous area that nobody
1401          * has touched so far, we don't want to allocate unnecessary pages or
1402          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1403          * then get_dump_page() will return NULL to leave a hole in the dump.
1404          * But we can only make this optimization where a hole would surely
1405          * be zero-filled if handle_mm_fault() actually did handle it.
1406          */
1407         if ((flags & FOLL_DUMP) &&
1408             (!vma->vm_ops || !vma->vm_ops->fault))
1409                 return ERR_PTR(-EFAULT);
1410         return page;
1411 }
1412
1413 /**
1414  * __get_user_pages() - pin user pages in memory
1415  * @tsk:        task_struct of target task
1416  * @mm:         mm_struct of target mm
1417  * @start:      starting user address
1418  * @nr_pages:   number of pages from start to pin
1419  * @gup_flags:  flags modifying pin behaviour
1420  * @pages:      array that receives pointers to the pages pinned.
1421  *              Should be at least nr_pages long. Or NULL, if caller
1422  *              only intends to ensure the pages are faulted in.
1423  * @vmas:       array of pointers to vmas corresponding to each page.
1424  *              Or NULL if the caller does not require them.
1425  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1426  *
1427  * Returns number of pages pinned. This may be fewer than the number
1428  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1429  * were pinned, returns -errno. Each page returned must be released
1430  * with a put_page() call when it is finished with. vmas will only
1431  * remain valid while mmap_sem is held.
1432  *
1433  * Must be called with mmap_sem held for read or write.
1434  *
1435  * __get_user_pages walks a process's page tables and takes a reference to
1436  * each struct page that each user address corresponds to at a given
1437  * instant. That is, it takes the page that would be accessed if a user
1438  * thread accesses the given user virtual address at that instant.
1439  *
1440  * This does not guarantee that the page exists in the user mappings when
1441  * __get_user_pages returns, and there may even be a completely different
1442  * page there in some cases (eg. if mmapped pagecache has been invalidated
1443  * and subsequently re faulted). However it does guarantee that the page
1444  * won't be freed completely. And mostly callers simply care that the page
1445  * contains data that was valid *at some point in time*. Typically, an IO
1446  * or similar operation cannot guarantee anything stronger anyway because
1447  * locks can't be held over the syscall boundary.
1448  *
1449  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1450  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1451  * appropriate) must be called after the page is finished with, and
1452  * before put_page is called.
1453  *
1454  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1455  * or mmap_sem contention, and if waiting is needed to pin all pages,
1456  * *@nonblocking will be set to 0.
1457  *
1458  * In most cases, get_user_pages or get_user_pages_fast should be used
1459  * instead of __get_user_pages. __get_user_pages should be used only if
1460  * you need some special @gup_flags.
1461  */
1462 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1463                      unsigned long start, int nr_pages, unsigned int gup_flags,
1464                      struct page **pages, struct vm_area_struct **vmas,
1465                      int *nonblocking)
1466 {
1467         int i;
1468         unsigned long vm_flags;
1469
1470         if (nr_pages <= 0)
1471                 return 0;
1472
1473         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1474
1475         /* 
1476          * Require read or write permissions.
1477          * If FOLL_FORCE is set, we only require the "MAY" flags.
1478          */
1479         vm_flags  = (gup_flags & FOLL_WRITE) ?
1480                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1481         vm_flags &= (gup_flags & FOLL_FORCE) ?
1482                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1483         i = 0;
1484
1485         do {
1486                 struct vm_area_struct *vma;
1487
1488                 vma = find_extend_vma(mm, start);
1489                 if (!vma && in_gate_area(tsk, start)) {
1490                         unsigned long pg = start & PAGE_MASK;
1491                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1492                         pgd_t *pgd;
1493                         pud_t *pud;
1494                         pmd_t *pmd;
1495                         pte_t *pte;
1496
1497                         /* user gate pages are read-only */
1498                         if (gup_flags & FOLL_WRITE)
1499                                 return i ? : -EFAULT;
1500                         if (pg > TASK_SIZE)
1501                                 pgd = pgd_offset_k(pg);
1502                         else
1503                                 pgd = pgd_offset_gate(mm, pg);
1504                         BUG_ON(pgd_none(*pgd));
1505                         pud = pud_offset(pgd, pg);
1506                         BUG_ON(pud_none(*pud));
1507                         pmd = pmd_offset(pud, pg);
1508                         if (pmd_none(*pmd))
1509                                 return i ? : -EFAULT;
1510                         VM_BUG_ON(pmd_trans_huge(*pmd));
1511                         pte = pte_offset_map(pmd, pg);
1512                         if (pte_none(*pte)) {
1513                                 pte_unmap(pte);
1514                                 return i ? : -EFAULT;
1515                         }
1516                         if (pages) {
1517                                 struct page *page;
1518
1519                                 page = vm_normal_page(gate_vma, start, *pte);
1520                                 if (!page) {
1521                                         if (!(gup_flags & FOLL_DUMP) &&
1522                                              is_zero_pfn(pte_pfn(*pte)))
1523                                                 page = pte_page(*pte);
1524                                         else {
1525                                                 pte_unmap(pte);
1526                                                 return i ? : -EFAULT;
1527                                         }
1528                                 }
1529                                 pages[i] = page;
1530                                 get_page(page);
1531                         }
1532                         pte_unmap(pte);
1533                         if (vmas)
1534                                 vmas[i] = gate_vma;
1535                         i++;
1536                         start += PAGE_SIZE;
1537                         nr_pages--;
1538                         continue;
1539                 }
1540
1541                 if (!vma ||
1542                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1543                     !(vm_flags & vma->vm_flags))
1544                         return i ? : -EFAULT;
1545
1546                 if (is_vm_hugetlb_page(vma)) {
1547                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1548                                         &start, &nr_pages, i, gup_flags);
1549                         continue;
1550                 }
1551
1552                 do {
1553                         struct page *page;
1554                         unsigned int foll_flags = gup_flags;
1555
1556                         /*
1557                          * If we have a pending SIGKILL, don't keep faulting
1558                          * pages and potentially allocating memory.
1559                          */
1560                         if (unlikely(fatal_signal_pending(current)))
1561                                 return i ? i : -ERESTARTSYS;
1562
1563                         cond_resched();
1564                         while (!(page = follow_page(vma, start, foll_flags))) {
1565                                 int ret;
1566                                 unsigned int fault_flags = 0;
1567
1568                                 if (foll_flags & FOLL_WRITE)
1569                                         fault_flags |= FAULT_FLAG_WRITE;
1570                                 if (nonblocking)
1571                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1572
1573                                 ret = handle_mm_fault(mm, vma, start,
1574                                                         fault_flags);
1575
1576                                 if (ret & VM_FAULT_ERROR) {
1577                                         if (ret & VM_FAULT_OOM)
1578                                                 return i ? i : -ENOMEM;
1579                                         if (ret & (VM_FAULT_HWPOISON |
1580                                                    VM_FAULT_HWPOISON_LARGE)) {
1581                                                 if (i)
1582                                                         return i;
1583                                                 else if (gup_flags & FOLL_HWPOISON)
1584                                                         return -EHWPOISON;
1585                                                 else
1586                                                         return -EFAULT;
1587                                         }
1588                                         if (ret & VM_FAULT_SIGBUS)
1589                                                 return i ? i : -EFAULT;
1590                                         BUG();
1591                                 }
1592                                 if (ret & VM_FAULT_MAJOR)
1593                                         tsk->maj_flt++;
1594                                 else
1595                                         tsk->min_flt++;
1596
1597                                 if (ret & VM_FAULT_RETRY) {
1598                                         *nonblocking = 0;
1599                                         return i;
1600                                 }
1601
1602                                 /*
1603                                  * The VM_FAULT_WRITE bit tells us that
1604                                  * do_wp_page has broken COW when necessary,
1605                                  * even if maybe_mkwrite decided not to set
1606                                  * pte_write. We can thus safely do subsequent
1607                                  * page lookups as if they were reads. But only
1608                                  * do so when looping for pte_write is futile:
1609                                  * in some cases userspace may also be wanting
1610                                  * to write to the gotten user page, which a
1611                                  * read fault here might prevent (a readonly
1612                                  * page might get reCOWed by userspace write).
1613                                  */
1614                                 if ((ret & VM_FAULT_WRITE) &&
1615                                     !(vma->vm_flags & VM_WRITE))
1616                                         foll_flags &= ~FOLL_WRITE;
1617
1618                                 cond_resched();
1619                         }
1620                         if (IS_ERR(page))
1621                                 return i ? i : PTR_ERR(page);
1622                         if (pages) {
1623                                 pages[i] = page;
1624
1625                                 flush_anon_page(vma, page, start);
1626                                 flush_dcache_page(page);
1627                         }
1628                         if (vmas)
1629                                 vmas[i] = vma;
1630                         i++;
1631                         start += PAGE_SIZE;
1632                         nr_pages--;
1633                 } while (nr_pages && start < vma->vm_end);
1634         } while (nr_pages);
1635         return i;
1636 }
1637 EXPORT_SYMBOL(__get_user_pages);
1638
1639 /**
1640  * get_user_pages() - pin user pages in memory
1641  * @tsk:        task_struct of target task
1642  * @mm:         mm_struct of target mm
1643  * @start:      starting user address
1644  * @nr_pages:   number of pages from start to pin
1645  * @write:      whether pages will be written to by the caller
1646  * @force:      whether to force write access even if user mapping is
1647  *              readonly. This will result in the page being COWed even
1648  *              in MAP_SHARED mappings. You do not want this.
1649  * @pages:      array that receives pointers to the pages pinned.
1650  *              Should be at least nr_pages long. Or NULL, if caller
1651  *              only intends to ensure the pages are faulted in.
1652  * @vmas:       array of pointers to vmas corresponding to each page.
1653  *              Or NULL if the caller does not require them.
1654  *
1655  * Returns number of pages pinned. This may be fewer than the number
1656  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1657  * were pinned, returns -errno. Each page returned must be released
1658  * with a put_page() call when it is finished with. vmas will only
1659  * remain valid while mmap_sem is held.
1660  *
1661  * Must be called with mmap_sem held for read or write.
1662  *
1663  * get_user_pages walks a process's page tables and takes a reference to
1664  * each struct page that each user address corresponds to at a given
1665  * instant. That is, it takes the page that would be accessed if a user
1666  * thread accesses the given user virtual address at that instant.
1667  *
1668  * This does not guarantee that the page exists in the user mappings when
1669  * get_user_pages returns, and there may even be a completely different
1670  * page there in some cases (eg. if mmapped pagecache has been invalidated
1671  * and subsequently re faulted). However it does guarantee that the page
1672  * won't be freed completely. And mostly callers simply care that the page
1673  * contains data that was valid *at some point in time*. Typically, an IO
1674  * or similar operation cannot guarantee anything stronger anyway because
1675  * locks can't be held over the syscall boundary.
1676  *
1677  * If write=0, the page must not be written to. If the page is written to,
1678  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1679  * after the page is finished with, and before put_page is called.
1680  *
1681  * get_user_pages is typically used for fewer-copy IO operations, to get a
1682  * handle on the memory by some means other than accesses via the user virtual
1683  * addresses. The pages may be submitted for DMA to devices or accessed via
1684  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1685  * use the correct cache flushing APIs.
1686  *
1687  * See also get_user_pages_fast, for performance critical applications.
1688  */
1689 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1690                 unsigned long start, int nr_pages, int write, int force,
1691                 struct page **pages, struct vm_area_struct **vmas)
1692 {
1693         int flags = FOLL_TOUCH;
1694
1695         if (pages)
1696                 flags |= FOLL_GET;
1697         if (write)
1698                 flags |= FOLL_WRITE;
1699         if (force)
1700                 flags |= FOLL_FORCE;
1701
1702         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1703                                 NULL);
1704 }
1705 EXPORT_SYMBOL(get_user_pages);
1706
1707 /**
1708  * get_dump_page() - pin user page in memory while writing it to core dump
1709  * @addr: user address
1710  *
1711  * Returns struct page pointer of user page pinned for dump,
1712  * to be freed afterwards by page_cache_release() or put_page().
1713  *
1714  * Returns NULL on any kind of failure - a hole must then be inserted into
1715  * the corefile, to preserve alignment with its headers; and also returns
1716  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1717  * allowing a hole to be left in the corefile to save diskspace.
1718  *
1719  * Called without mmap_sem, but after all other threads have been killed.
1720  */
1721 #ifdef CONFIG_ELF_CORE
1722 struct page *get_dump_page(unsigned long addr)
1723 {
1724         struct vm_area_struct *vma;
1725         struct page *page;
1726
1727         if (__get_user_pages(current, current->mm, addr, 1,
1728                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1729                              NULL) < 1)
1730                 return NULL;
1731         flush_cache_page(vma, addr, page_to_pfn(page));
1732         return page;
1733 }
1734 #endif /* CONFIG_ELF_CORE */
1735
1736 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1737                         spinlock_t **ptl)
1738 {
1739         pgd_t * pgd = pgd_offset(mm, addr);
1740         pud_t * pud = pud_alloc(mm, pgd, addr);
1741         if (pud) {
1742                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1743                 if (pmd) {
1744                         VM_BUG_ON(pmd_trans_huge(*pmd));
1745                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1746                 }
1747         }
1748         return NULL;
1749 }
1750
1751 /*
1752  * This is the old fallback for page remapping.
1753  *
1754  * For historical reasons, it only allows reserved pages. Only
1755  * old drivers should use this, and they needed to mark their
1756  * pages reserved for the old functions anyway.
1757  */
1758 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1759                         struct page *page, pgprot_t prot)
1760 {
1761         struct mm_struct *mm = vma->vm_mm;
1762         int retval;
1763         pte_t *pte;
1764         spinlock_t *ptl;
1765
1766         retval = -EINVAL;
1767         if (PageAnon(page))
1768                 goto out;
1769         retval = -ENOMEM;
1770         flush_dcache_page(page);
1771         pte = get_locked_pte(mm, addr, &ptl);
1772         if (!pte)
1773                 goto out;
1774         retval = -EBUSY;
1775         if (!pte_none(*pte))
1776                 goto out_unlock;
1777
1778         /* Ok, finally just insert the thing.. */
1779         get_page(page);
1780         inc_mm_counter_fast(mm, MM_FILEPAGES);
1781         page_add_file_rmap(page);
1782         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1783
1784         retval = 0;
1785         pte_unmap_unlock(pte, ptl);
1786         return retval;
1787 out_unlock:
1788         pte_unmap_unlock(pte, ptl);
1789 out:
1790         return retval;
1791 }
1792
1793 /**
1794  * vm_insert_page - insert single page into user vma
1795  * @vma: user vma to map to
1796  * @addr: target user address of this page
1797  * @page: source kernel page
1798  *
1799  * This allows drivers to insert individual pages they've allocated
1800  * into a user vma.
1801  *
1802  * The page has to be a nice clean _individual_ kernel allocation.
1803  * If you allocate a compound page, you need to have marked it as
1804  * such (__GFP_COMP), or manually just split the page up yourself
1805  * (see split_page()).
1806  *
1807  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1808  * took an arbitrary page protection parameter. This doesn't allow
1809  * that. Your vma protection will have to be set up correctly, which
1810  * means that if you want a shared writable mapping, you'd better
1811  * ask for a shared writable mapping!
1812  *
1813  * The page does not need to be reserved.
1814  */
1815 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1816                         struct page *page)
1817 {
1818         if (addr < vma->vm_start || addr >= vma->vm_end)
1819                 return -EFAULT;
1820         if (!page_count(page))
1821                 return -EINVAL;
1822         vma->vm_flags |= VM_INSERTPAGE;
1823         return insert_page(vma, addr, page, vma->vm_page_prot);
1824 }
1825 EXPORT_SYMBOL(vm_insert_page);
1826
1827 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1828                         unsigned long pfn, pgprot_t prot)
1829 {
1830         struct mm_struct *mm = vma->vm_mm;
1831         int retval;
1832         pte_t *pte, entry;
1833         spinlock_t *ptl;
1834
1835         retval = -ENOMEM;
1836         pte = get_locked_pte(mm, addr, &ptl);
1837         if (!pte)
1838                 goto out;
1839         retval = -EBUSY;
1840         if (!pte_none(*pte))
1841                 goto out_unlock;
1842
1843         /* Ok, finally just insert the thing.. */
1844         entry = pte_mkspecial(pfn_pte(pfn, prot));
1845         set_pte_at(mm, addr, pte, entry);
1846         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1847
1848         retval = 0;
1849 out_unlock:
1850         pte_unmap_unlock(pte, ptl);
1851 out:
1852         return retval;
1853 }
1854
1855 /**
1856  * vm_insert_pfn - insert single pfn into user vma
1857  * @vma: user vma to map to
1858  * @addr: target user address of this page
1859  * @pfn: source kernel pfn
1860  *
1861  * Similar to vm_inert_page, this allows drivers to insert individual pages
1862  * they've allocated into a user vma. Same comments apply.
1863  *
1864  * This function should only be called from a vm_ops->fault handler, and
1865  * in that case the handler should return NULL.
1866  *
1867  * vma cannot be a COW mapping.
1868  *
1869  * As this is called only for pages that do not currently exist, we
1870  * do not need to flush old virtual caches or the TLB.
1871  */
1872 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1873                         unsigned long pfn)
1874 {
1875         int ret;
1876         pgprot_t pgprot = vma->vm_page_prot;
1877         /*
1878          * Technically, architectures with pte_special can avoid all these
1879          * restrictions (same for remap_pfn_range).  However we would like
1880          * consistency in testing and feature parity among all, so we should
1881          * try to keep these invariants in place for everybody.
1882          */
1883         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1884         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1885                                                 (VM_PFNMAP|VM_MIXEDMAP));
1886         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1887         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1888
1889         if (addr < vma->vm_start || addr >= vma->vm_end)
1890                 return -EFAULT;
1891         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1892                 return -EINVAL;
1893
1894         ret = insert_pfn(vma, addr, pfn, pgprot);
1895
1896         if (ret)
1897                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1898
1899         return ret;
1900 }
1901 EXPORT_SYMBOL(vm_insert_pfn);
1902
1903 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1904                         unsigned long pfn)
1905 {
1906         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1907
1908         if (addr < vma->vm_start || addr >= vma->vm_end)
1909                 return -EFAULT;
1910
1911         /*
1912          * If we don't have pte special, then we have to use the pfn_valid()
1913          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1914          * refcount the page if pfn_valid is true (hence insert_page rather
1915          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1916          * without pte special, it would there be refcounted as a normal page.
1917          */
1918         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1919                 struct page *page;
1920
1921                 page = pfn_to_page(pfn);
1922                 return insert_page(vma, addr, page, vma->vm_page_prot);
1923         }
1924         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1925 }
1926 EXPORT_SYMBOL(vm_insert_mixed);
1927
1928 /*
1929  * maps a range of physical memory into the requested pages. the old
1930  * mappings are removed. any references to nonexistent pages results
1931  * in null mappings (currently treated as "copy-on-access")
1932  */
1933 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1934                         unsigned long addr, unsigned long end,
1935                         unsigned long pfn, pgprot_t prot)
1936 {
1937         pte_t *pte;
1938         spinlock_t *ptl;
1939
1940         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1941         if (!pte)
1942                 return -ENOMEM;
1943         arch_enter_lazy_mmu_mode();
1944         do {
1945                 BUG_ON(!pte_none(*pte));
1946                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1947                 pfn++;
1948         } while (pte++, addr += PAGE_SIZE, addr != end);
1949         arch_leave_lazy_mmu_mode();
1950         pte_unmap_unlock(pte - 1, ptl);
1951         return 0;
1952 }
1953
1954 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1955                         unsigned long addr, unsigned long end,
1956                         unsigned long pfn, pgprot_t prot)
1957 {
1958         pmd_t *pmd;
1959         unsigned long next;
1960
1961         pfn -= addr >> PAGE_SHIFT;
1962         pmd = pmd_alloc(mm, pud, addr);
1963         if (!pmd)
1964                 return -ENOMEM;
1965         VM_BUG_ON(pmd_trans_huge(*pmd));
1966         do {
1967                 next = pmd_addr_end(addr, end);
1968                 if (remap_pte_range(mm, pmd, addr, next,
1969                                 pfn + (addr >> PAGE_SHIFT), prot))
1970                         return -ENOMEM;
1971         } while (pmd++, addr = next, addr != end);
1972         return 0;
1973 }
1974
1975 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1976                         unsigned long addr, unsigned long end,
1977                         unsigned long pfn, pgprot_t prot)
1978 {
1979         pud_t *pud;
1980         unsigned long next;
1981
1982         pfn -= addr >> PAGE_SHIFT;
1983         pud = pud_alloc(mm, pgd, addr);
1984         if (!pud)
1985                 return -ENOMEM;
1986         do {
1987                 next = pud_addr_end(addr, end);
1988                 if (remap_pmd_range(mm, pud, addr, next,
1989                                 pfn + (addr >> PAGE_SHIFT), prot))
1990                         return -ENOMEM;
1991         } while (pud++, addr = next, addr != end);
1992         return 0;
1993 }
1994
1995 /**
1996  * remap_pfn_range - remap kernel memory to userspace
1997  * @vma: user vma to map to
1998  * @addr: target user address to start at
1999  * @pfn: physical address of kernel memory
2000  * @size: size of map area
2001  * @prot: page protection flags for this mapping
2002  *
2003  *  Note: this is only safe if the mm semaphore is held when called.
2004  */
2005 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2006                     unsigned long pfn, unsigned long size, pgprot_t prot)
2007 {
2008         pgd_t *pgd;
2009         unsigned long next;
2010         unsigned long end = addr + PAGE_ALIGN(size);
2011         struct mm_struct *mm = vma->vm_mm;
2012         int err;
2013
2014         /*
2015          * Physically remapped pages are special. Tell the
2016          * rest of the world about it:
2017          *   VM_IO tells people not to look at these pages
2018          *      (accesses can have side effects).
2019          *   VM_RESERVED is specified all over the place, because
2020          *      in 2.4 it kept swapout's vma scan off this vma; but
2021          *      in 2.6 the LRU scan won't even find its pages, so this
2022          *      flag means no more than count its pages in reserved_vm,
2023          *      and omit it from core dump, even when VM_IO turned off.
2024          *   VM_PFNMAP tells the core MM that the base pages are just
2025          *      raw PFN mappings, and do not have a "struct page" associated
2026          *      with them.
2027          *
2028          * There's a horrible special case to handle copy-on-write
2029          * behaviour that some programs depend on. We mark the "original"
2030          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2031          */
2032         if (addr == vma->vm_start && end == vma->vm_end) {
2033                 vma->vm_pgoff = pfn;
2034                 vma->vm_flags |= VM_PFN_AT_MMAP;
2035         } else if (is_cow_mapping(vma->vm_flags))
2036                 return -EINVAL;
2037
2038         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2039
2040         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2041         if (err) {
2042                 /*
2043                  * To indicate that track_pfn related cleanup is not
2044                  * needed from higher level routine calling unmap_vmas
2045                  */
2046                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2047                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2048                 return -EINVAL;
2049         }
2050
2051         BUG_ON(addr >= end);
2052         pfn -= addr >> PAGE_SHIFT;
2053         pgd = pgd_offset(mm, addr);
2054         flush_cache_range(vma, addr, end);
2055         do {
2056                 next = pgd_addr_end(addr, end);
2057                 err = remap_pud_range(mm, pgd, addr, next,
2058                                 pfn + (addr >> PAGE_SHIFT), prot);
2059                 if (err)
2060                         break;
2061         } while (pgd++, addr = next, addr != end);
2062
2063         if (err)
2064                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2065
2066         return err;
2067 }
2068 EXPORT_SYMBOL(remap_pfn_range);
2069
2070 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2071                                      unsigned long addr, unsigned long end,
2072                                      pte_fn_t fn, void *data)
2073 {
2074         pte_t *pte;
2075         int err;
2076         pgtable_t token;
2077         spinlock_t *uninitialized_var(ptl);
2078
2079         pte = (mm == &init_mm) ?
2080                 pte_alloc_kernel(pmd, addr) :
2081                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2082         if (!pte)
2083                 return -ENOMEM;
2084
2085         BUG_ON(pmd_huge(*pmd));
2086
2087         arch_enter_lazy_mmu_mode();
2088
2089         token = pmd_pgtable(*pmd);
2090
2091         do {
2092                 err = fn(pte++, token, addr, data);
2093                 if (err)
2094                         break;
2095         } while (addr += PAGE_SIZE, addr != end);
2096
2097         arch_leave_lazy_mmu_mode();
2098
2099         if (mm != &init_mm)
2100                 pte_unmap_unlock(pte-1, ptl);
2101         return err;
2102 }
2103
2104 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2105                                      unsigned long addr, unsigned long end,
2106                                      pte_fn_t fn, void *data)
2107 {
2108         pmd_t *pmd;
2109         unsigned long next;
2110         int err;
2111
2112         BUG_ON(pud_huge(*pud));
2113
2114         pmd = pmd_alloc(mm, pud, addr);
2115         if (!pmd)
2116                 return -ENOMEM;
2117         do {
2118                 next = pmd_addr_end(addr, end);
2119                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2120                 if (err)
2121                         break;
2122         } while (pmd++, addr = next, addr != end);
2123         return err;
2124 }
2125
2126 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2127                                      unsigned long addr, unsigned long end,
2128                                      pte_fn_t fn, void *data)
2129 {
2130         pud_t *pud;
2131         unsigned long next;
2132         int err;
2133
2134         pud = pud_alloc(mm, pgd, addr);
2135         if (!pud)
2136                 return -ENOMEM;
2137         do {
2138                 next = pud_addr_end(addr, end);
2139                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2140                 if (err)
2141                         break;
2142         } while (pud++, addr = next, addr != end);
2143         return err;
2144 }
2145
2146 /*
2147  * Scan a region of virtual memory, filling in page tables as necessary
2148  * and calling a provided function on each leaf page table.
2149  */
2150 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2151                         unsigned long size, pte_fn_t fn, void *data)
2152 {
2153         pgd_t *pgd;
2154         unsigned long next;
2155         unsigned long end = addr + size;
2156         int err;
2157
2158         BUG_ON(addr >= end);
2159         pgd = pgd_offset(mm, addr);
2160         do {
2161                 next = pgd_addr_end(addr, end);
2162                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2163                 if (err)
2164                         break;
2165         } while (pgd++, addr = next, addr != end);
2166
2167         return err;
2168 }
2169 EXPORT_SYMBOL_GPL(apply_to_page_range);
2170
2171 /*
2172  * handle_pte_fault chooses page fault handler according to an entry
2173  * which was read non-atomically.  Before making any commitment, on
2174  * those architectures or configurations (e.g. i386 with PAE) which
2175  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2176  * must check under lock before unmapping the pte and proceeding
2177  * (but do_wp_page is only called after already making such a check;
2178  * and do_anonymous_page can safely check later on).
2179  */
2180 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2181                                 pte_t *page_table, pte_t orig_pte)
2182 {
2183         int same = 1;
2184 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2185         if (sizeof(pte_t) > sizeof(unsigned long)) {
2186                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2187                 spin_lock(ptl);
2188                 same = pte_same(*page_table, orig_pte);
2189                 spin_unlock(ptl);
2190         }
2191 #endif
2192         pte_unmap(page_table);
2193         return same;
2194 }
2195
2196 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2197 {
2198         /*
2199          * If the source page was a PFN mapping, we don't have
2200          * a "struct page" for it. We do a best-effort copy by
2201          * just copying from the original user address. If that
2202          * fails, we just zero-fill it. Live with it.
2203          */
2204         if (unlikely(!src)) {
2205                 void *kaddr = kmap_atomic(dst, KM_USER0);
2206                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2207
2208                 /*
2209                  * This really shouldn't fail, because the page is there
2210                  * in the page tables. But it might just be unreadable,
2211                  * in which case we just give up and fill the result with
2212                  * zeroes.
2213                  */
2214                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2215                         clear_page(kaddr);
2216                 kunmap_atomic(kaddr, KM_USER0);
2217                 flush_dcache_page(dst);
2218         } else
2219                 copy_user_highpage(dst, src, va, vma);
2220 }
2221
2222 /*
2223  * This routine handles present pages, when users try to write
2224  * to a shared page. It is done by copying the page to a new address
2225  * and decrementing the shared-page counter for the old page.
2226  *
2227  * Note that this routine assumes that the protection checks have been
2228  * done by the caller (the low-level page fault routine in most cases).
2229  * Thus we can safely just mark it writable once we've done any necessary
2230  * COW.
2231  *
2232  * We also mark the page dirty at this point even though the page will
2233  * change only once the write actually happens. This avoids a few races,
2234  * and potentially makes it more efficient.
2235  *
2236  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2237  * but allow concurrent faults), with pte both mapped and locked.
2238  * We return with mmap_sem still held, but pte unmapped and unlocked.
2239  */
2240 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2241                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2242                 spinlock_t *ptl, pte_t orig_pte)
2243         __releases(ptl)
2244 {
2245         struct page *old_page, *new_page;
2246         pte_t entry;
2247         int ret = 0;
2248         int page_mkwrite = 0;
2249         struct page *dirty_page = NULL;
2250
2251         old_page = vm_normal_page(vma, address, orig_pte);
2252         if (!old_page) {
2253                 /*
2254                  * VM_MIXEDMAP !pfn_valid() case
2255                  *
2256                  * We should not cow pages in a shared writeable mapping.
2257                  * Just mark the pages writable as we can't do any dirty
2258                  * accounting on raw pfn maps.
2259                  */
2260                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2261                                      (VM_WRITE|VM_SHARED))
2262                         goto reuse;
2263                 goto gotten;
2264         }
2265
2266         /*
2267          * Take out anonymous pages first, anonymous shared vmas are
2268          * not dirty accountable.
2269          */
2270         if (PageAnon(old_page) && !PageKsm(old_page)) {
2271                 if (!trylock_page(old_page)) {
2272                         page_cache_get(old_page);
2273                         pte_unmap_unlock(page_table, ptl);
2274                         lock_page(old_page);
2275                         page_table = pte_offset_map_lock(mm, pmd, address,
2276                                                          &ptl);
2277                         if (!pte_same(*page_table, orig_pte)) {
2278                                 unlock_page(old_page);
2279                                 goto unlock;
2280                         }
2281                         page_cache_release(old_page);
2282                 }
2283                 if (reuse_swap_page(old_page)) {
2284                         /*
2285                          * The page is all ours.  Move it to our anon_vma so
2286                          * the rmap code will not search our parent or siblings.
2287                          * Protected against the rmap code by the page lock.
2288                          */
2289                         page_move_anon_rmap(old_page, vma, address);
2290                         unlock_page(old_page);
2291                         goto reuse;
2292                 }
2293                 unlock_page(old_page);
2294         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2295                                         (VM_WRITE|VM_SHARED))) {
2296                 /*
2297                  * Only catch write-faults on shared writable pages,
2298                  * read-only shared pages can get COWed by
2299                  * get_user_pages(.write=1, .force=1).
2300                  */
2301                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2302                         struct vm_fault vmf;
2303                         int tmp;
2304
2305                         vmf.virtual_address = (void __user *)(address &
2306                                                                 PAGE_MASK);
2307                         vmf.pgoff = old_page->index;
2308                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2309                         vmf.page = old_page;
2310
2311                         /*
2312                          * Notify the address space that the page is about to
2313                          * become writable so that it can prohibit this or wait
2314                          * for the page to get into an appropriate state.
2315                          *
2316                          * We do this without the lock held, so that it can
2317                          * sleep if it needs to.
2318                          */
2319                         page_cache_get(old_page);
2320                         pte_unmap_unlock(page_table, ptl);
2321
2322                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2323                         if (unlikely(tmp &
2324                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2325                                 ret = tmp;
2326                                 goto unwritable_page;
2327                         }
2328                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2329                                 lock_page(old_page);
2330                                 if (!old_page->mapping) {
2331                                         ret = 0; /* retry the fault */
2332                                         unlock_page(old_page);
2333                                         goto unwritable_page;
2334                                 }
2335                         } else
2336                                 VM_BUG_ON(!PageLocked(old_page));
2337
2338                         /*
2339                          * Since we dropped the lock we need to revalidate
2340                          * the PTE as someone else may have changed it.  If
2341                          * they did, we just return, as we can count on the
2342                          * MMU to tell us if they didn't also make it writable.
2343                          */
2344                         page_table = pte_offset_map_lock(mm, pmd, address,
2345                                                          &ptl);
2346                         if (!pte_same(*page_table, orig_pte)) {
2347                                 unlock_page(old_page);
2348                                 goto unlock;
2349                         }
2350
2351                         page_mkwrite = 1;
2352                 }
2353                 dirty_page = old_page;
2354                 get_page(dirty_page);
2355
2356 reuse:
2357                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2358                 entry = pte_mkyoung(orig_pte);
2359                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2360                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2361                         update_mmu_cache(vma, address, page_table);
2362                 pte_unmap_unlock(page_table, ptl);
2363                 ret |= VM_FAULT_WRITE;
2364
2365                 if (!dirty_page)
2366                         return ret;
2367
2368                 /*
2369                  * Yes, Virginia, this is actually required to prevent a race
2370                  * with clear_page_dirty_for_io() from clearing the page dirty
2371                  * bit after it clear all dirty ptes, but before a racing
2372                  * do_wp_page installs a dirty pte.
2373                  *
2374                  * __do_fault is protected similarly.
2375                  */
2376                 if (!page_mkwrite) {
2377                         wait_on_page_locked(dirty_page);
2378                         set_page_dirty_balance(dirty_page, page_mkwrite);
2379                 }
2380                 put_page(dirty_page);
2381                 if (page_mkwrite) {
2382                         struct address_space *mapping = dirty_page->mapping;
2383
2384                         set_page_dirty(dirty_page);
2385                         unlock_page(dirty_page);
2386                         page_cache_release(dirty_page);
2387                         if (mapping)    {
2388                                 /*
2389                                  * Some device drivers do not set page.mapping
2390                                  * but still dirty their pages
2391                                  */
2392                                 balance_dirty_pages_ratelimited(mapping);
2393                         }
2394                 }
2395
2396                 /* file_update_time outside page_lock */
2397                 if (vma->vm_file)
2398                         file_update_time(vma->vm_file);
2399
2400                 return ret;
2401         }
2402
2403         /*
2404          * Ok, we need to copy. Oh, well..
2405          */
2406         page_cache_get(old_page);
2407 gotten:
2408         pte_unmap_unlock(page_table, ptl);
2409
2410         if (unlikely(anon_vma_prepare(vma)))
2411                 goto oom;
2412
2413         if (is_zero_pfn(pte_pfn(orig_pte))) {
2414                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2415                 if (!new_page)
2416                         goto oom;
2417         } else {
2418                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2419                 if (!new_page)
2420                         goto oom;
2421                 cow_user_page(new_page, old_page, address, vma);
2422         }
2423         __SetPageUptodate(new_page);
2424
2425         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2426                 goto oom_free_new;
2427
2428         /*
2429          * Re-check the pte - we dropped the lock
2430          */
2431         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2432         if (likely(pte_same(*page_table, orig_pte))) {
2433                 if (old_page) {
2434                         if (!PageAnon(old_page)) {
2435                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2436                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2437                         }
2438                 } else
2439                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2440                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2441                 entry = mk_pte(new_page, vma->vm_page_prot);
2442                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2443                 /*
2444                  * Clear the pte entry and flush it first, before updating the
2445                  * pte with the new entry. This will avoid a race condition
2446                  * seen in the presence of one thread doing SMC and another
2447                  * thread doing COW.
2448                  */
2449                 ptep_clear_flush(vma, address, page_table);
2450                 page_add_new_anon_rmap(new_page, vma, address);
2451                 /*
2452                  * We call the notify macro here because, when using secondary
2453                  * mmu page tables (such as kvm shadow page tables), we want the
2454                  * new page to be mapped directly into the secondary page table.
2455                  */
2456                 set_pte_at_notify(mm, address, page_table, entry);
2457                 update_mmu_cache(vma, address, page_table);
2458                 if (old_page) {
2459                         /*
2460                          * Only after switching the pte to the new page may
2461                          * we remove the mapcount here. Otherwise another
2462                          * process may come and find the rmap count decremented
2463                          * before the pte is switched to the new page, and
2464                          * "reuse" the old page writing into it while our pte
2465                          * here still points into it and can be read by other
2466                          * threads.
2467                          *
2468                          * The critical issue is to order this
2469                          * page_remove_rmap with the ptp_clear_flush above.
2470                          * Those stores are ordered by (if nothing else,)
2471                          * the barrier present in the atomic_add_negative
2472                          * in page_remove_rmap.
2473                          *
2474                          * Then the TLB flush in ptep_clear_flush ensures that
2475                          * no process can access the old page before the
2476                          * decremented mapcount is visible. And the old page
2477                          * cannot be reused until after the decremented
2478                          * mapcount is visible. So transitively, TLBs to
2479                          * old page will be flushed before it can be reused.
2480                          */
2481                         page_remove_rmap(old_page);
2482                 }
2483
2484                 /* Free the old page.. */
2485                 new_page = old_page;
2486                 ret |= VM_FAULT_WRITE;
2487         } else
2488                 mem_cgroup_uncharge_page(new_page);
2489
2490         if (new_page)
2491                 page_cache_release(new_page);
2492 unlock:
2493         pte_unmap_unlock(page_table, ptl);
2494         if (old_page) {
2495                 /*
2496                  * Don't let another task, with possibly unlocked vma,
2497                  * keep the mlocked page.
2498                  */
2499                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2500                         lock_page(old_page);    /* LRU manipulation */
2501                         munlock_vma_page(old_page);
2502                         unlock_page(old_page);
2503                 }
2504                 page_cache_release(old_page);
2505         }
2506         return ret;
2507 oom_free_new:
2508         page_cache_release(new_page);
2509 oom:
2510         if (old_page) {
2511                 if (page_mkwrite) {
2512                         unlock_page(old_page);
2513                         page_cache_release(old_page);
2514                 }
2515                 page_cache_release(old_page);
2516         }
2517         return VM_FAULT_OOM;
2518
2519 unwritable_page:
2520         page_cache_release(old_page);
2521         return ret;
2522 }
2523
2524 /*
2525  * Helper functions for unmap_mapping_range().
2526  *
2527  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2528  *
2529  * We have to restart searching the prio_tree whenever we drop the lock,
2530  * since the iterator is only valid while the lock is held, and anyway
2531  * a later vma might be split and reinserted earlier while lock dropped.
2532  *
2533  * The list of nonlinear vmas could be handled more efficiently, using
2534  * a placeholder, but handle it in the same way until a need is shown.
2535  * It is important to search the prio_tree before nonlinear list: a vma
2536  * may become nonlinear and be shifted from prio_tree to nonlinear list
2537  * while the lock is dropped; but never shifted from list to prio_tree.
2538  *
2539  * In order to make forward progress despite restarting the search,
2540  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2541  * quickly skip it next time around.  Since the prio_tree search only
2542  * shows us those vmas affected by unmapping the range in question, we
2543  * can't efficiently keep all vmas in step with mapping->truncate_count:
2544  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2545  * mapping->truncate_count and vma->vm_truncate_count are protected by
2546  * i_mmap_lock.
2547  *
2548  * In order to make forward progress despite repeatedly restarting some
2549  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2550  * and restart from that address when we reach that vma again.  It might
2551  * have been split or merged, shrunk or extended, but never shifted: so
2552  * restart_addr remains valid so long as it remains in the vma's range.
2553  * unmap_mapping_range forces truncate_count to leap over page-aligned
2554  * values so we can save vma's restart_addr in its truncate_count field.
2555  */
2556 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2557
2558 static void reset_vma_truncate_counts(struct address_space *mapping)
2559 {
2560         struct vm_area_struct *vma;
2561         struct prio_tree_iter iter;
2562
2563         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2564                 vma->vm_truncate_count = 0;
2565         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2566                 vma->vm_truncate_count = 0;
2567 }
2568
2569 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2570                 unsigned long start_addr, unsigned long end_addr,
2571                 struct zap_details *details)
2572 {
2573         unsigned long restart_addr;
2574         int need_break;
2575
2576         /*
2577          * files that support invalidating or truncating portions of the
2578          * file from under mmaped areas must have their ->fault function
2579          * return a locked page (and set VM_FAULT_LOCKED in the return).
2580          * This provides synchronisation against concurrent unmapping here.
2581          */
2582
2583 again:
2584         restart_addr = vma->vm_truncate_count;
2585         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2586                 start_addr = restart_addr;
2587                 if (start_addr >= end_addr) {
2588                         /* Top of vma has been split off since last time */
2589                         vma->vm_truncate_count = details->truncate_count;
2590                         return 0;
2591                 }
2592         }
2593
2594         restart_addr = zap_page_range(vma, start_addr,
2595                                         end_addr - start_addr, details);
2596         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2597
2598         if (restart_addr >= end_addr) {
2599                 /* We have now completed this vma: mark it so */
2600                 vma->vm_truncate_count = details->truncate_count;
2601                 if (!need_break)
2602                         return 0;
2603         } else {
2604                 /* Note restart_addr in vma's truncate_count field */
2605                 vma->vm_truncate_count = restart_addr;
2606                 if (!need_break)
2607                         goto again;
2608         }
2609
2610         spin_unlock(details->i_mmap_lock);
2611         cond_resched();
2612         spin_lock(details->i_mmap_lock);
2613         return -EINTR;
2614 }
2615
2616 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2617                                             struct zap_details *details)
2618 {
2619         struct vm_area_struct *vma;
2620         struct prio_tree_iter iter;
2621         pgoff_t vba, vea, zba, zea;
2622
2623 restart:
2624         vma_prio_tree_foreach(vma, &iter, root,
2625                         details->first_index, details->last_index) {
2626                 /* Skip quickly over those we have already dealt with */
2627                 if (vma->vm_truncate_count == details->truncate_count)
2628                         continue;
2629
2630                 vba = vma->vm_pgoff;
2631                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2632                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2633                 zba = details->first_index;
2634                 if (zba < vba)
2635                         zba = vba;
2636                 zea = details->last_index;
2637                 if (zea > vea)
2638                         zea = vea;
2639
2640                 if (unmap_mapping_range_vma(vma,
2641                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2642                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2643                                 details) < 0)
2644                         goto restart;
2645         }
2646 }
2647
2648 static inline void unmap_mapping_range_list(struct list_head *head,
2649                                             struct zap_details *details)
2650 {
2651         struct vm_area_struct *vma;
2652
2653         /*
2654          * In nonlinear VMAs there is no correspondence between virtual address
2655          * offset and file offset.  So we must perform an exhaustive search
2656          * across *all* the pages in each nonlinear VMA, not just the pages
2657          * whose virtual address lies outside the file truncation point.
2658          */
2659 restart:
2660         list_for_each_entry(vma, head, shared.vm_set.list) {
2661                 /* Skip quickly over those we have already dealt with */
2662                 if (vma->vm_truncate_count == details->truncate_count)
2663                         continue;
2664                 details->nonlinear_vma = vma;
2665                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2666                                         vma->vm_end, details) < 0)
2667                         goto restart;
2668         }
2669 }
2670
2671 /**
2672  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2673  * @mapping: the address space containing mmaps to be unmapped.
2674  * @holebegin: byte in first page to unmap, relative to the start of
2675  * the underlying file.  This will be rounded down to a PAGE_SIZE
2676  * boundary.  Note that this is different from truncate_pagecache(), which
2677  * must keep the partial page.  In contrast, we must get rid of
2678  * partial pages.
2679  * @holelen: size of prospective hole in bytes.  This will be rounded
2680  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2681  * end of the file.
2682  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2683  * but 0 when invalidating pagecache, don't throw away private data.
2684  */
2685 void unmap_mapping_range(struct address_space *mapping,
2686                 loff_t const holebegin, loff_t const holelen, int even_cows)
2687 {
2688         struct zap_details details;
2689         pgoff_t hba = holebegin >> PAGE_SHIFT;
2690         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2691
2692         /* Check for overflow. */
2693         if (sizeof(holelen) > sizeof(hlen)) {
2694                 long long holeend =
2695                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2696                 if (holeend & ~(long long)ULONG_MAX)
2697                         hlen = ULONG_MAX - hba + 1;
2698         }
2699
2700         details.check_mapping = even_cows? NULL: mapping;
2701         details.nonlinear_vma = NULL;
2702         details.first_index = hba;
2703         details.last_index = hba + hlen - 1;
2704         if (details.last_index < details.first_index)
2705                 details.last_index = ULONG_MAX;
2706         details.i_mmap_lock = &mapping->i_mmap_lock;
2707
2708         mutex_lock(&mapping->unmap_mutex);
2709         spin_lock(&mapping->i_mmap_lock);
2710
2711         /* Protect against endless unmapping loops */
2712         mapping->truncate_count++;
2713         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2714                 if (mapping->truncate_count == 0)
2715                         reset_vma_truncate_counts(mapping);
2716                 mapping->truncate_count++;
2717         }
2718         details.truncate_count = mapping->truncate_count;
2719
2720         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2721                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2722         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2723                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2724         spin_unlock(&mapping->i_mmap_lock);
2725         mutex_unlock(&mapping->unmap_mutex);
2726 }
2727 EXPORT_SYMBOL(unmap_mapping_range);
2728
2729 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2730 {
2731         struct address_space *mapping = inode->i_mapping;
2732
2733         /*
2734          * If the underlying filesystem is not going to provide
2735          * a way to truncate a range of blocks (punch a hole) -
2736          * we should return failure right now.
2737          */
2738         if (!inode->i_op->truncate_range)
2739                 return -ENOSYS;
2740
2741         mutex_lock(&inode->i_mutex);
2742         down_write(&inode->i_alloc_sem);
2743         unmap_mapping_range(mapping, offset, (end - offset), 1);
2744         truncate_inode_pages_range(mapping, offset, end);
2745         unmap_mapping_range(mapping, offset, (end - offset), 1);
2746         inode->i_op->truncate_range(inode, offset, end);
2747         up_write(&inode->i_alloc_sem);
2748         mutex_unlock(&inode->i_mutex);
2749
2750         return 0;
2751 }
2752
2753 /*
2754  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2755  * but allow concurrent faults), and pte mapped but not yet locked.
2756  * We return with mmap_sem still held, but pte unmapped and unlocked.
2757  */
2758 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2759                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2760                 unsigned int flags, pte_t orig_pte)
2761 {
2762         spinlock_t *ptl;
2763         struct page *page, *swapcache = NULL;
2764         swp_entry_t entry;
2765         pte_t pte;
2766         int locked;
2767         struct mem_cgroup *ptr = NULL;
2768         int exclusive = 0;
2769         int ret = 0;
2770
2771         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2772                 goto out;
2773
2774         entry = pte_to_swp_entry(orig_pte);
2775         if (unlikely(non_swap_entry(entry))) {
2776                 if (is_migration_entry(entry)) {
2777                         migration_entry_wait(mm, pmd, address);
2778                 } else if (is_hwpoison_entry(entry)) {
2779                         ret = VM_FAULT_HWPOISON;
2780                 } else {
2781                         print_bad_pte(vma, address, orig_pte, NULL);
2782                         ret = VM_FAULT_SIGBUS;
2783                 }
2784                 goto out;
2785         }
2786         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2787         page = lookup_swap_cache(entry);
2788         if (!page) {
2789                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2790                 page = swapin_readahead(entry,
2791                                         GFP_HIGHUSER_MOVABLE, vma, address);
2792                 if (!page) {
2793                         /*
2794                          * Back out if somebody else faulted in this pte
2795                          * while we released the pte lock.
2796                          */
2797                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2798                         if (likely(pte_same(*page_table, orig_pte)))
2799                                 ret = VM_FAULT_OOM;
2800                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2801                         goto unlock;
2802                 }
2803
2804                 /* Had to read the page from swap area: Major fault */
2805                 ret = VM_FAULT_MAJOR;
2806                 count_vm_event(PGMAJFAULT);
2807         } else if (PageHWPoison(page)) {
2808                 /*
2809                  * hwpoisoned dirty swapcache pages are kept for killing
2810                  * owner processes (which may be unknown at hwpoison time)
2811                  */
2812                 ret = VM_FAULT_HWPOISON;
2813                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2814                 goto out_release;
2815         }
2816
2817         locked = lock_page_or_retry(page, mm, flags);
2818         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2819         if (!locked) {
2820                 ret |= VM_FAULT_RETRY;
2821                 goto out_release;
2822         }
2823
2824         /*
2825          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2826          * release the swapcache from under us.  The page pin, and pte_same
2827          * test below, are not enough to exclude that.  Even if it is still
2828          * swapcache, we need to check that the page's swap has not changed.
2829          */
2830         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2831                 goto out_page;
2832
2833         if (ksm_might_need_to_copy(page, vma, address)) {
2834                 swapcache = page;
2835                 page = ksm_does_need_to_copy(page, vma, address);
2836
2837                 if (unlikely(!page)) {
2838                         ret = VM_FAULT_OOM;
2839                         page = swapcache;
2840                         swapcache = NULL;
2841                         goto out_page;
2842                 }
2843         }
2844
2845         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2846                 ret = VM_FAULT_OOM;
2847                 goto out_page;
2848         }
2849
2850         /*
2851          * Back out if somebody else already faulted in this pte.
2852          */
2853         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2854         if (unlikely(!pte_same(*page_table, orig_pte)))
2855                 goto out_nomap;
2856
2857         if (unlikely(!PageUptodate(page))) {
2858                 ret = VM_FAULT_SIGBUS;
2859                 goto out_nomap;
2860         }
2861
2862         /*
2863          * The page isn't present yet, go ahead with the fault.
2864          *
2865          * Be careful about the sequence of operations here.
2866          * To get its accounting right, reuse_swap_page() must be called
2867          * while the page is counted on swap but not yet in mapcount i.e.
2868          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2869          * must be called after the swap_free(), or it will never succeed.
2870          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2871          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2872          * in page->private. In this case, a record in swap_cgroup  is silently
2873          * discarded at swap_free().
2874          */
2875
2876         inc_mm_counter_fast(mm, MM_ANONPAGES);
2877         dec_mm_counter_fast(mm, MM_SWAPENTS);
2878         pte = mk_pte(page, vma->vm_page_prot);
2879         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2880                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2881                 flags &= ~FAULT_FLAG_WRITE;
2882                 ret |= VM_FAULT_WRITE;
2883                 exclusive = 1;
2884         }
2885         flush_icache_page(vma, page);
2886         set_pte_at(mm, address, page_table, pte);
2887         do_page_add_anon_rmap(page, vma, address, exclusive);
2888         /* It's better to call commit-charge after rmap is established */
2889         mem_cgroup_commit_charge_swapin(page, ptr);
2890
2891         swap_free(entry);
2892         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2893                 try_to_free_swap(page);
2894         unlock_page(page);
2895         if (swapcache) {
2896                 /*
2897                  * Hold the lock to avoid the swap entry to be reused
2898                  * until we take the PT lock for the pte_same() check
2899                  * (to avoid false positives from pte_same). For
2900                  * further safety release the lock after the swap_free
2901                  * so that the swap count won't change under a
2902                  * parallel locked swapcache.
2903                  */
2904                 unlock_page(swapcache);
2905                 page_cache_release(swapcache);
2906         }
2907
2908         if (flags & FAULT_FLAG_WRITE) {
2909                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2910                 if (ret & VM_FAULT_ERROR)
2911                         ret &= VM_FAULT_ERROR;
2912                 goto out;
2913         }
2914
2915         /* No need to invalidate - it was non-present before */
2916         update_mmu_cache(vma, address, page_table);
2917 unlock:
2918         pte_unmap_unlock(page_table, ptl);
2919 out:
2920         return ret;
2921 out_nomap:
2922         mem_cgroup_cancel_charge_swapin(ptr);
2923         pte_unmap_unlock(page_table, ptl);
2924 out_page:
2925         unlock_page(page);
2926 out_release:
2927         page_cache_release(page);
2928         if (swapcache) {
2929                 unlock_page(swapcache);
2930                 page_cache_release(swapcache);
2931         }
2932         return ret;
2933 }
2934
2935 /*
2936  * This is like a special single-page "expand_{down|up}wards()",
2937  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2938  * doesn't hit another vma.
2939  */
2940 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2941 {
2942         address &= PAGE_MASK;
2943         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2944                 struct vm_area_struct *prev = vma->vm_prev;
2945
2946                 /*
2947                  * Is there a mapping abutting this one below?
2948                  *
2949                  * That's only ok if it's the same stack mapping
2950                  * that has gotten split..
2951                  */
2952                 if (prev && prev->vm_end == address)
2953                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2954
2955                 expand_stack(vma, address - PAGE_SIZE);
2956         }
2957         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2958                 struct vm_area_struct *next = vma->vm_next;
2959
2960                 /* As VM_GROWSDOWN but s/below/above/ */
2961                 if (next && next->vm_start == address + PAGE_SIZE)
2962                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2963
2964                 expand_upwards(vma, address + PAGE_SIZE);
2965         }
2966         return 0;
2967 }
2968
2969 /*
2970  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2971  * but allow concurrent faults), and pte mapped but not yet locked.
2972  * We return with mmap_sem still held, but pte unmapped and unlocked.
2973  */
2974 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2975                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2976                 unsigned int flags)
2977 {
2978         struct page *page;
2979         spinlock_t *ptl;
2980         pte_t entry;
2981
2982         pte_unmap(page_table);
2983
2984         /* Check if we need to add a guard page to the stack */
2985         if (check_stack_guard_page(vma, address) < 0)
2986                 return VM_FAULT_SIGBUS;
2987
2988         /* Use the zero-page for reads */
2989         if (!(flags & FAULT_FLAG_WRITE)) {
2990                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2991                                                 vma->vm_page_prot));
2992                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2993                 if (!pte_none(*page_table))
2994                         goto unlock;
2995                 goto setpte;
2996         }
2997
2998         /* Allocate our own private page. */
2999         if (unlikely(anon_vma_prepare(vma)))
3000                 goto oom;
3001         page = alloc_zeroed_user_highpage_movable(vma, address);
3002         if (!page)
3003                 goto oom;
3004         __SetPageUptodate(page);
3005
3006         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3007                 goto oom_free_page;
3008
3009         entry = mk_pte(page, vma->vm_page_prot);
3010         if (vma->vm_flags & VM_WRITE)
3011                 entry = pte_mkwrite(pte_mkdirty(entry));
3012
3013         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3014         if (!pte_none(*page_table))
3015                 goto release;
3016
3017         inc_mm_counter_fast(mm, MM_ANONPAGES);
3018         page_add_new_anon_rmap(page, vma, address);
3019 setpte:
3020         set_pte_at(mm, address, page_table, entry);
3021
3022         /* No need to invalidate - it was non-present before */
3023         update_mmu_cache(vma, address, page_table);
3024 unlock:
3025         pte_unmap_unlock(page_table, ptl);
3026         return 0;
3027 release:
3028         mem_cgroup_uncharge_page(page);
3029         page_cache_release(page);
3030         goto unlock;
3031 oom_free_page:
3032         page_cache_release(page);
3033 oom:
3034         return VM_FAULT_OOM;
3035 }
3036
3037 /*
3038  * __do_fault() tries to create a new page mapping. It aggressively
3039  * tries to share with existing pages, but makes a separate copy if
3040  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3041  * the next page fault.
3042  *
3043  * As this is called only for pages that do not currently exist, we
3044  * do not need to flush old virtual caches or the TLB.
3045  *
3046  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3047  * but allow concurrent faults), and pte neither mapped nor locked.
3048  * We return with mmap_sem still held, but pte unmapped and unlocked.
3049  */
3050 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3051                 unsigned long address, pmd_t *pmd,
3052                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3053 {
3054         pte_t *page_table;
3055         spinlock_t *ptl;
3056         struct page *page;
3057         pte_t entry;
3058         int anon = 0;
3059         int charged = 0;
3060         struct page *dirty_page = NULL;
3061         struct vm_fault vmf;
3062         int ret;
3063         int page_mkwrite = 0;
3064
3065         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3066         vmf.pgoff = pgoff;
3067         vmf.flags = flags;
3068         vmf.page = NULL;
3069
3070         ret = vma->vm_ops->fault(vma, &vmf);
3071         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3072                             VM_FAULT_RETRY)))
3073                 return ret;
3074
3075         if (unlikely(PageHWPoison(vmf.page))) {
3076                 if (ret & VM_FAULT_LOCKED)
3077                         unlock_page(vmf.page);
3078                 return VM_FAULT_HWPOISON;
3079         }
3080
3081         /*
3082          * For consistency in subsequent calls, make the faulted page always
3083          * locked.
3084          */
3085         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3086                 lock_page(vmf.page);
3087         else
3088                 VM_BUG_ON(!PageLocked(vmf.page));
3089
3090         /*
3091          * Should we do an early C-O-W break?
3092          */
3093         page = vmf.page;
3094         if (flags & FAULT_FLAG_WRITE) {
3095                 if (!(vma->vm_flags & VM_SHARED)) {
3096                         anon = 1;
3097                         if (unlikely(anon_vma_prepare(vma))) {
3098                                 ret = VM_FAULT_OOM;
3099                                 goto out;
3100                         }
3101                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3102                                                 vma, address);
3103                         if (!page) {
3104                                 ret = VM_FAULT_OOM;
3105                                 goto out;
3106                         }
3107                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3108                                 ret = VM_FAULT_OOM;
3109                                 page_cache_release(page);
3110                                 goto out;
3111                         }
3112                         charged = 1;
3113                         copy_user_highpage(page, vmf.page, address, vma);
3114                         __SetPageUptodate(page);
3115                 } else {
3116                         /*
3117                          * If the page will be shareable, see if the backing
3118                          * address space wants to know that the page is about
3119                          * to become writable
3120                          */
3121                         if (vma->vm_ops->page_mkwrite) {
3122                                 int tmp;
3123
3124                                 unlock_page(page);
3125                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3126                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3127                                 if (unlikely(tmp &
3128                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3129                                         ret = tmp;
3130                                         goto unwritable_page;
3131                                 }
3132                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3133                                         lock_page(page);
3134                                         if (!page->mapping) {
3135                                                 ret = 0; /* retry the fault */
3136                                                 unlock_page(page);
3137                                                 goto unwritable_page;
3138                                         }
3139                                 } else
3140                                         VM_BUG_ON(!PageLocked(page));
3141                                 page_mkwrite = 1;
3142                         }
3143                 }
3144
3145         }
3146
3147         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3148
3149         /*
3150          * This silly early PAGE_DIRTY setting removes a race
3151          * due to the bad i386 page protection. But it's valid
3152          * for other architectures too.
3153          *
3154          * Note that if FAULT_FLAG_WRITE is set, we either now have
3155          * an exclusive copy of the page, or this is a shared mapping,
3156          * so we can make it writable and dirty to avoid having to
3157          * handle that later.
3158          */
3159         /* Only go through if we didn't race with anybody else... */
3160         if (likely(pte_same(*page_table, orig_pte))) {
3161                 flush_icache_page(vma, page);
3162                 entry = mk_pte(page, vma->vm_page_prot);
3163                 if (flags & FAULT_FLAG_WRITE)
3164                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3165                 if (anon) {
3166                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3167                         page_add_new_anon_rmap(page, vma, address);
3168                 } else {
3169                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3170                         page_add_file_rmap(page);
3171                         if (flags & FAULT_FLAG_WRITE) {
3172                                 dirty_page = page;
3173                                 get_page(dirty_page);
3174                         }
3175                 }
3176                 set_pte_at(mm, address, page_table, entry);
3177
3178                 /* no need to invalidate: a not-present page won't be cached */
3179                 update_mmu_cache(vma, address, page_table);
3180         } else {
3181                 if (charged)
3182                         mem_cgroup_uncharge_page(page);
3183                 if (anon)
3184                         page_cache_release(page);
3185                 else
3186                         anon = 1; /* no anon but release faulted_page */
3187         }
3188
3189         pte_unmap_unlock(page_table, ptl);
3190
3191 out:
3192         if (dirty_page) {
3193                 struct address_space *mapping = page->mapping;
3194
3195                 if (set_page_dirty(dirty_page))
3196                         page_mkwrite = 1;
3197                 unlock_page(dirty_page);
3198                 put_page(dirty_page);
3199                 if (page_mkwrite && mapping) {
3200                         /*
3201                          * Some device drivers do not set page.mapping but still
3202                          * dirty their pages
3203                          */
3204                         balance_dirty_pages_ratelimited(mapping);
3205                 }
3206
3207                 /* file_update_time outside page_lock */
3208                 if (vma->vm_file)
3209                         file_update_time(vma->vm_file);
3210         } else {
3211                 unlock_page(vmf.page);
3212                 if (anon)
3213                         page_cache_release(vmf.page);
3214         }
3215
3216         return ret;
3217
3218 unwritable_page:
3219         page_cache_release(page);
3220         return ret;
3221 }
3222
3223 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3224                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3225                 unsigned int flags, pte_t orig_pte)
3226 {
3227         pgoff_t pgoff = (((address & PAGE_MASK)
3228                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3229
3230         pte_unmap(page_table);
3231         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3232 }
3233
3234 /*
3235  * Fault of a previously existing named mapping. Repopulate the pte
3236  * from the encoded file_pte if possible. This enables swappable
3237  * nonlinear vmas.
3238  *
3239  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3240  * but allow concurrent faults), and pte mapped but not yet locked.
3241  * We return with mmap_sem still held, but pte unmapped and unlocked.
3242  */
3243 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3244                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3245                 unsigned int flags, pte_t orig_pte)
3246 {
3247         pgoff_t pgoff;
3248
3249         flags |= FAULT_FLAG_NONLINEAR;
3250
3251         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3252                 return 0;
3253
3254         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3255                 /*
3256                  * Page table corrupted: show pte and kill process.
3257                  */
3258                 print_bad_pte(vma, address, orig_pte, NULL);
3259                 return VM_FAULT_SIGBUS;
3260         }
3261
3262         pgoff = pte_to_pgoff(orig_pte);
3263         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3264 }
3265
3266 /*
3267  * These routines also need to handle stuff like marking pages dirty
3268  * and/or accessed for architectures that don't do it in hardware (most
3269  * RISC architectures).  The early dirtying is also good on the i386.
3270  *
3271  * There is also a hook called "update_mmu_cache()" that architectures
3272  * with external mmu caches can use to update those (ie the Sparc or
3273  * PowerPC hashed page tables that act as extended TLBs).
3274  *
3275  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3276  * but allow concurrent faults), and pte mapped but not yet locked.
3277  * We return with mmap_sem still held, but pte unmapped and unlocked.
3278  */
3279 int handle_pte_fault(struct mm_struct *mm,
3280                      struct vm_area_struct *vma, unsigned long address,
3281                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3282 {
3283         pte_t entry;
3284         spinlock_t *ptl;
3285
3286         entry = *pte;
3287         if (!pte_present(entry)) {
3288                 if (pte_none(entry)) {
3289                         if (vma->vm_ops) {
3290                                 if (likely(vma->vm_ops->fault))
3291                                         return do_linear_fault(mm, vma, address,
3292                                                 pte, pmd, flags, entry);
3293                         }
3294                         return do_anonymous_page(mm, vma, address,
3295                                                  pte, pmd, flags);
3296                 }
3297                 if (pte_file(entry))
3298                         return do_nonlinear_fault(mm, vma, address,
3299                                         pte, pmd, flags, entry);
3300                 return do_swap_page(mm, vma, address,
3301                                         pte, pmd, flags, entry);
3302         }
3303
3304         ptl = pte_lockptr(mm, pmd);
3305         spin_lock(ptl);
3306         if (unlikely(!pte_same(*pte, entry)))
3307                 goto unlock;
3308         if (flags & FAULT_FLAG_WRITE) {
3309                 if (!pte_write(entry))
3310                         return do_wp_page(mm, vma, address,
3311                                         pte, pmd, ptl, entry);
3312                 entry = pte_mkdirty(entry);
3313         }
3314         entry = pte_mkyoung(entry);
3315         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3316                 update_mmu_cache(vma, address, pte);
3317         } else {
3318                 /*
3319                  * This is needed only for protection faults but the arch code
3320                  * is not yet telling us if this is a protection fault or not.
3321                  * This still avoids useless tlb flushes for .text page faults
3322                  * with threads.
3323                  */
3324                 if (flags & FAULT_FLAG_WRITE)
3325                         flush_tlb_fix_spurious_fault(vma, address);
3326         }
3327 unlock:
3328         pte_unmap_unlock(pte, ptl);
3329         return 0;
3330 }
3331
3332 /*
3333  * By the time we get here, we already hold the mm semaphore
3334  */
3335 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3336                 unsigned long address, unsigned int flags)
3337 {
3338         pgd_t *pgd;
3339         pud_t *pud;
3340         pmd_t *pmd;
3341         pte_t *pte;
3342
3343         __set_current_state(TASK_RUNNING);
3344
3345         count_vm_event(PGFAULT);
3346
3347         /* do counter updates before entering really critical section. */
3348         check_sync_rss_stat(current);
3349
3350         if (unlikely(is_vm_hugetlb_page(vma)))
3351                 return hugetlb_fault(mm, vma, address, flags);
3352
3353         pgd = pgd_offset(mm, address);
3354         pud = pud_alloc(mm, pgd, address);
3355         if (!pud)
3356                 return VM_FAULT_OOM;
3357         pmd = pmd_alloc(mm, pud, address);
3358         if (!pmd)
3359                 return VM_FAULT_OOM;
3360         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3361                 if (!vma->vm_ops)
3362                         return do_huge_pmd_anonymous_page(mm, vma, address,
3363                                                           pmd, flags);
3364         } else {
3365                 pmd_t orig_pmd = *pmd;
3366                 barrier();
3367                 if (pmd_trans_huge(orig_pmd)) {
3368                         if (flags & FAULT_FLAG_WRITE &&
3369                             !pmd_write(orig_pmd) &&
3370                             !pmd_trans_splitting(orig_pmd))
3371                                 return do_huge_pmd_wp_page(mm, vma, address,
3372                                                            pmd, orig_pmd);
3373                         return 0;
3374                 }
3375         }
3376
3377         /*
3378          * Use __pte_alloc instead of pte_alloc_map, because we can't
3379          * run pte_offset_map on the pmd, if an huge pmd could
3380          * materialize from under us from a different thread.
3381          */
3382         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
3383                 return VM_FAULT_OOM;
3384         /* if an huge pmd materialized from under us just retry later */
3385         if (unlikely(pmd_trans_huge(*pmd)))
3386                 return 0;
3387         /*
3388          * A regular pmd is established and it can't morph into a huge pmd
3389          * from under us anymore at this point because we hold the mmap_sem
3390          * read mode and khugepaged takes it in write mode. So now it's
3391          * safe to run pte_offset_map().
3392          */
3393         pte = pte_offset_map(pmd, address);
3394
3395         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3396 }
3397
3398 #ifndef __PAGETABLE_PUD_FOLDED
3399 /*
3400  * Allocate page upper directory.
3401  * We've already handled the fast-path in-line.
3402  */
3403 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3404 {
3405         pud_t *new = pud_alloc_one(mm, address);
3406         if (!new)
3407                 return -ENOMEM;
3408
3409         smp_wmb(); /* See comment in __pte_alloc */
3410
3411         spin_lock(&mm->page_table_lock);
3412         if (pgd_present(*pgd))          /* Another has populated it */
3413                 pud_free(mm, new);
3414         else
3415                 pgd_populate(mm, pgd, new);
3416         spin_unlock(&mm->page_table_lock);
3417         return 0;
3418 }
3419 #endif /* __PAGETABLE_PUD_FOLDED */
3420
3421 #ifndef __PAGETABLE_PMD_FOLDED
3422 /*
3423  * Allocate page middle directory.
3424  * We've already handled the fast-path in-line.
3425  */
3426 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3427 {
3428         pmd_t *new = pmd_alloc_one(mm, address);
3429         if (!new)
3430                 return -ENOMEM;
3431
3432         smp_wmb(); /* See comment in __pte_alloc */
3433
3434         spin_lock(&mm->page_table_lock);
3435 #ifndef __ARCH_HAS_4LEVEL_HACK
3436         if (pud_present(*pud))          /* Another has populated it */
3437                 pmd_free(mm, new);
3438         else
3439                 pud_populate(mm, pud, new);
3440 #else
3441         if (pgd_present(*pud))          /* Another has populated it */
3442                 pmd_free(mm, new);
3443         else
3444                 pgd_populate(mm, pud, new);
3445 #endif /* __ARCH_HAS_4LEVEL_HACK */
3446         spin_unlock(&mm->page_table_lock);
3447         return 0;
3448 }
3449 #endif /* __PAGETABLE_PMD_FOLDED */
3450
3451 int make_pages_present(unsigned long addr, unsigned long end)
3452 {
3453         int ret, len, write;
3454         struct vm_area_struct * vma;
3455
3456         vma = find_vma(current->mm, addr);
3457         if (!vma)
3458                 return -ENOMEM;
3459         /*
3460          * We want to touch writable mappings with a write fault in order
3461          * to break COW, except for shared mappings because these don't COW
3462          * and we would not want to dirty them for nothing.
3463          */
3464         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3465         BUG_ON(addr >= end);
3466         BUG_ON(end > vma->vm_end);
3467         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3468         ret = get_user_pages(current, current->mm, addr,
3469                         len, write, 0, NULL, NULL);
3470         if (ret < 0)
3471                 return ret;
3472         return ret == len ? 0 : -EFAULT;
3473 }
3474
3475 #if !defined(__HAVE_ARCH_GATE_AREA)
3476
3477 #if defined(AT_SYSINFO_EHDR)
3478 static struct vm_area_struct gate_vma;
3479
3480 static int __init gate_vma_init(void)
3481 {
3482         gate_vma.vm_mm = NULL;
3483         gate_vma.vm_start = FIXADDR_USER_START;
3484         gate_vma.vm_end = FIXADDR_USER_END;
3485         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3486         gate_vma.vm_page_prot = __P101;
3487         /*
3488          * Make sure the vDSO gets into every core dump.
3489          * Dumping its contents makes post-mortem fully interpretable later
3490          * without matching up the same kernel and hardware config to see
3491          * what PC values meant.
3492          */
3493         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3494         return 0;
3495 }
3496 __initcall(gate_vma_init);
3497 #endif
3498
3499 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3500 {
3501 #ifdef AT_SYSINFO_EHDR
3502         return &gate_vma;
3503 #else
3504         return NULL;
3505 #endif
3506 }
3507
3508 int in_gate_area_no_task(unsigned long addr)
3509 {
3510 #ifdef AT_SYSINFO_EHDR
3511         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3512                 return 1;
3513 #endif
3514         return 0;
3515 }
3516
3517 #endif  /* __HAVE_ARCH_GATE_AREA */
3518
3519 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3520                 pte_t **ptepp, spinlock_t **ptlp)
3521 {
3522         pgd_t *pgd;
3523         pud_t *pud;
3524         pmd_t *pmd;
3525         pte_t *ptep;
3526
3527         pgd = pgd_offset(mm, address);
3528         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3529                 goto out;
3530
3531         pud = pud_offset(pgd, address);
3532         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3533                 goto out;
3534
3535         pmd = pmd_offset(pud, address);
3536         VM_BUG_ON(pmd_trans_huge(*pmd));
3537         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3538                 goto out;
3539
3540         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3541         if (pmd_huge(*pmd))
3542                 goto out;
3543
3544         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3545         if (!ptep)
3546                 goto out;
3547         if (!pte_present(*ptep))
3548                 goto unlock;
3549         *ptepp = ptep;
3550         return 0;
3551 unlock:
3552         pte_unmap_unlock(ptep, *ptlp);
3553 out:
3554         return -EINVAL;
3555 }
3556
3557 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3558                              pte_t **ptepp, spinlock_t **ptlp)
3559 {
3560         int res;
3561
3562         /* (void) is needed to make gcc happy */
3563         (void) __cond_lock(*ptlp,
3564                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3565         return res;
3566 }
3567
3568 /**
3569  * follow_pfn - look up PFN at a user virtual address
3570  * @vma: memory mapping
3571  * @address: user virtual address
3572  * @pfn: location to store found PFN
3573  *
3574  * Only IO mappings and raw PFN mappings are allowed.
3575  *
3576  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3577  */
3578 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3579         unsigned long *pfn)
3580 {
3581         int ret = -EINVAL;
3582         spinlock_t *ptl;
3583         pte_t *ptep;
3584
3585         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3586                 return ret;
3587
3588         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3589         if (ret)
3590                 return ret;
3591         *pfn = pte_pfn(*ptep);
3592         pte_unmap_unlock(ptep, ptl);
3593         return 0;
3594 }
3595 EXPORT_SYMBOL(follow_pfn);
3596
3597 #ifdef CONFIG_HAVE_IOREMAP_PROT
3598 int follow_phys(struct vm_area_struct *vma,
3599                 unsigned long address, unsigned int flags,
3600                 unsigned long *prot, resource_size_t *phys)
3601 {
3602         int ret = -EINVAL;
3603         pte_t *ptep, pte;
3604         spinlock_t *ptl;
3605
3606         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3607                 goto out;
3608
3609         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3610                 goto out;
3611         pte = *ptep;
3612
3613         if ((flags & FOLL_WRITE) && !pte_write(pte))
3614                 goto unlock;
3615
3616         *prot = pgprot_val(pte_pgprot(pte));
3617         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3618
3619         ret = 0;
3620 unlock:
3621         pte_unmap_unlock(ptep, ptl);
3622 out:
3623         return ret;
3624 }
3625
3626 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3627                         void *buf, int len, int write)
3628 {
3629         resource_size_t phys_addr;
3630         unsigned long prot = 0;
3631         void __iomem *maddr;
3632         int offset = addr & (PAGE_SIZE-1);
3633
3634         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3635                 return -EINVAL;
3636
3637         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3638         if (write)
3639                 memcpy_toio(maddr + offset, buf, len);
3640         else
3641                 memcpy_fromio(buf, maddr + offset, len);
3642         iounmap(maddr);
3643
3644         return len;
3645 }
3646 #endif
3647
3648 /*
3649  * Access another process' address space.
3650  * Source/target buffer must be kernel space,
3651  * Do not walk the page table directly, use get_user_pages
3652  */
3653 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3654 {
3655         struct mm_struct *mm;
3656         struct vm_area_struct *vma;
3657         void *old_buf = buf;
3658
3659         mm = get_task_mm(tsk);
3660         if (!mm)
3661                 return 0;
3662
3663         down_read(&mm->mmap_sem);
3664         /* ignore errors, just check how much was successfully transferred */
3665         while (len) {
3666                 int bytes, ret, offset;
3667                 void *maddr;
3668                 struct page *page = NULL;
3669
3670                 ret = get_user_pages(tsk, mm, addr, 1,
3671                                 write, 1, &page, &vma);
3672                 if (ret <= 0) {
3673                         /*
3674                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3675                          * we can access using slightly different code.
3676                          */
3677 #ifdef CONFIG_HAVE_IOREMAP_PROT
3678                         vma = find_vma(mm, addr);
3679                         if (!vma)
3680                                 break;
3681                         if (vma->vm_ops && vma->vm_ops->access)
3682                                 ret = vma->vm_ops->access(vma, addr, buf,
3683                                                           len, write);
3684                         if (ret <= 0)
3685 #endif
3686                                 break;
3687                         bytes = ret;
3688                 } else {
3689                         bytes = len;
3690                         offset = addr & (PAGE_SIZE-1);
3691                         if (bytes > PAGE_SIZE-offset)
3692                                 bytes = PAGE_SIZE-offset;
3693
3694                         maddr = kmap(page);
3695                         if (write) {
3696                                 copy_to_user_page(vma, page, addr,
3697                                                   maddr + offset, buf, bytes);
3698                                 set_page_dirty_lock(page);
3699                         } else {
3700                                 copy_from_user_page(vma, page, addr,
3701                                                     buf, maddr + offset, bytes);
3702                         }
3703                         kunmap(page);
3704                         page_cache_release(page);
3705                 }
3706                 len -= bytes;
3707                 buf += bytes;
3708                 addr += bytes;
3709         }
3710         up_read(&mm->mmap_sem);
3711         mmput(mm);
3712
3713         return buf - old_buf;
3714 }
3715
3716 /*
3717  * Print the name of a VMA.
3718  */
3719 void print_vma_addr(char *prefix, unsigned long ip)
3720 {
3721         struct mm_struct *mm = current->mm;
3722         struct vm_area_struct *vma;
3723
3724         /*
3725          * Do not print if we are in atomic
3726          * contexts (in exception stacks, etc.):
3727          */
3728         if (preempt_count())
3729                 return;
3730
3731         down_read(&mm->mmap_sem);
3732         vma = find_vma(mm, ip);
3733         if (vma && vma->vm_file) {
3734                 struct file *f = vma->vm_file;
3735                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3736                 if (buf) {
3737                         char *p, *s;
3738
3739                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3740                         if (IS_ERR(p))
3741                                 p = "?";
3742                         s = strrchr(p, '/');
3743                         if (s)
3744                                 p = s+1;
3745                         printk("%s%s[%lx+%lx]", prefix, p,
3746                                         vma->vm_start,
3747                                         vma->vm_end - vma->vm_start);
3748                         free_page((unsigned long)buf);
3749                 }
3750         }
3751         up_read(&current->mm->mmap_sem);
3752 }
3753
3754 #ifdef CONFIG_PROVE_LOCKING
3755 void might_fault(void)
3756 {
3757         /*
3758          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3759          * holding the mmap_sem, this is safe because kernel memory doesn't
3760          * get paged out, therefore we'll never actually fault, and the
3761          * below annotations will generate false positives.
3762          */
3763         if (segment_eq(get_fs(), KERNEL_DS))
3764                 return;
3765
3766         might_sleep();
3767         /*
3768          * it would be nicer only to annotate paths which are not under
3769          * pagefault_disable, however that requires a larger audit and
3770          * providing helpers like get_user_atomic.
3771          */
3772         if (!in_atomic() && current->mm)
3773                 might_lock_read(&current->mm->mmap_sem);
3774 }
3775 EXPORT_SYMBOL(might_fault);
3776 #endif
3777
3778 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3779 static void clear_gigantic_page(struct page *page,
3780                                 unsigned long addr,
3781                                 unsigned int pages_per_huge_page)
3782 {
3783         int i;
3784         struct page *p = page;
3785
3786         might_sleep();
3787         for (i = 0; i < pages_per_huge_page;
3788              i++, p = mem_map_next(p, page, i)) {
3789                 cond_resched();
3790                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3791         }
3792 }
3793 void clear_huge_page(struct page *page,
3794                      unsigned long addr, unsigned int pages_per_huge_page)
3795 {
3796         int i;
3797
3798         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3799                 clear_gigantic_page(page, addr, pages_per_huge_page);
3800                 return;
3801         }
3802
3803         might_sleep();
3804         for (i = 0; i < pages_per_huge_page; i++) {
3805                 cond_resched();
3806                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3807         }
3808 }
3809
3810 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3811                                     unsigned long addr,
3812                                     struct vm_area_struct *vma,
3813                                     unsigned int pages_per_huge_page)
3814 {
3815         int i;
3816         struct page *dst_base = dst;
3817         struct page *src_base = src;
3818
3819         for (i = 0; i < pages_per_huge_page; ) {
3820                 cond_resched();
3821                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3822
3823                 i++;
3824                 dst = mem_map_next(dst, dst_base, i);
3825                 src = mem_map_next(src, src_base, i);
3826         }
3827 }
3828
3829 void copy_user_huge_page(struct page *dst, struct page *src,
3830                          unsigned long addr, struct vm_area_struct *vma,
3831                          unsigned int pages_per_huge_page)
3832 {
3833         int i;
3834
3835         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3836                 copy_user_gigantic_page(dst, src, addr, vma,
3837                                         pages_per_huge_page);
3838                 return;
3839         }
3840
3841         might_sleep();
3842         for (i = 0; i < pages_per_huge_page; i++) {
3843                 cond_resched();
3844                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3845         }
3846 }
3847 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */