sys_swapon: call swap_cgroup_swapon() earlier
[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                                 if (foll_flags & FOLL_NOWAIT)
1573                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1574
1575                                 ret = handle_mm_fault(mm, vma, start,
1576                                                         fault_flags);
1577
1578                                 if (ret & VM_FAULT_ERROR) {
1579                                         if (ret & VM_FAULT_OOM)
1580                                                 return i ? i : -ENOMEM;
1581                                         if (ret & (VM_FAULT_HWPOISON |
1582                                                    VM_FAULT_HWPOISON_LARGE)) {
1583                                                 if (i)
1584                                                         return i;
1585                                                 else if (gup_flags & FOLL_HWPOISON)
1586                                                         return -EHWPOISON;
1587                                                 else
1588                                                         return -EFAULT;
1589                                         }
1590                                         if (ret & VM_FAULT_SIGBUS)
1591                                                 return i ? i : -EFAULT;
1592                                         BUG();
1593                                 }
1594                                 if (ret & VM_FAULT_MAJOR)
1595                                         tsk->maj_flt++;
1596                                 else
1597                                         tsk->min_flt++;
1598
1599                                 if (ret & VM_FAULT_RETRY) {
1600                                         if (nonblocking)
1601                                                 *nonblocking = 0;
1602                                         return i;
1603                                 }
1604
1605                                 /*
1606                                  * The VM_FAULT_WRITE bit tells us that
1607                                  * do_wp_page has broken COW when necessary,
1608                                  * even if maybe_mkwrite decided not to set
1609                                  * pte_write. We can thus safely do subsequent
1610                                  * page lookups as if they were reads. But only
1611                                  * do so when looping for pte_write is futile:
1612                                  * in some cases userspace may also be wanting
1613                                  * to write to the gotten user page, which a
1614                                  * read fault here might prevent (a readonly
1615                                  * page might get reCOWed by userspace write).
1616                                  */
1617                                 if ((ret & VM_FAULT_WRITE) &&
1618                                     !(vma->vm_flags & VM_WRITE))
1619                                         foll_flags &= ~FOLL_WRITE;
1620
1621                                 cond_resched();
1622                         }
1623                         if (IS_ERR(page))
1624                                 return i ? i : PTR_ERR(page);
1625                         if (pages) {
1626                                 pages[i] = page;
1627
1628                                 flush_anon_page(vma, page, start);
1629                                 flush_dcache_page(page);
1630                         }
1631                         if (vmas)
1632                                 vmas[i] = vma;
1633                         i++;
1634                         start += PAGE_SIZE;
1635                         nr_pages--;
1636                 } while (nr_pages && start < vma->vm_end);
1637         } while (nr_pages);
1638         return i;
1639 }
1640 EXPORT_SYMBOL(__get_user_pages);
1641
1642 /**
1643  * get_user_pages() - pin user pages in memory
1644  * @tsk:        task_struct of target task
1645  * @mm:         mm_struct of target mm
1646  * @start:      starting user address
1647  * @nr_pages:   number of pages from start to pin
1648  * @write:      whether pages will be written to by the caller
1649  * @force:      whether to force write access even if user mapping is
1650  *              readonly. This will result in the page being COWed even
1651  *              in MAP_SHARED mappings. You do not want this.
1652  * @pages:      array that receives pointers to the pages pinned.
1653  *              Should be at least nr_pages long. Or NULL, if caller
1654  *              only intends to ensure the pages are faulted in.
1655  * @vmas:       array of pointers to vmas corresponding to each page.
1656  *              Or NULL if the caller does not require them.
1657  *
1658  * Returns number of pages pinned. This may be fewer than the number
1659  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1660  * were pinned, returns -errno. Each page returned must be released
1661  * with a put_page() call when it is finished with. vmas will only
1662  * remain valid while mmap_sem is held.
1663  *
1664  * Must be called with mmap_sem held for read or write.
1665  *
1666  * get_user_pages walks a process's page tables and takes a reference to
1667  * each struct page that each user address corresponds to at a given
1668  * instant. That is, it takes the page that would be accessed if a user
1669  * thread accesses the given user virtual address at that instant.
1670  *
1671  * This does not guarantee that the page exists in the user mappings when
1672  * get_user_pages returns, and there may even be a completely different
1673  * page there in some cases (eg. if mmapped pagecache has been invalidated
1674  * and subsequently re faulted). However it does guarantee that the page
1675  * won't be freed completely. And mostly callers simply care that the page
1676  * contains data that was valid *at some point in time*. Typically, an IO
1677  * or similar operation cannot guarantee anything stronger anyway because
1678  * locks can't be held over the syscall boundary.
1679  *
1680  * If write=0, the page must not be written to. If the page is written to,
1681  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1682  * after the page is finished with, and before put_page is called.
1683  *
1684  * get_user_pages is typically used for fewer-copy IO operations, to get a
1685  * handle on the memory by some means other than accesses via the user virtual
1686  * addresses. The pages may be submitted for DMA to devices or accessed via
1687  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1688  * use the correct cache flushing APIs.
1689  *
1690  * See also get_user_pages_fast, for performance critical applications.
1691  */
1692 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1693                 unsigned long start, int nr_pages, int write, int force,
1694                 struct page **pages, struct vm_area_struct **vmas)
1695 {
1696         int flags = FOLL_TOUCH;
1697
1698         if (pages)
1699                 flags |= FOLL_GET;
1700         if (write)
1701                 flags |= FOLL_WRITE;
1702         if (force)
1703                 flags |= FOLL_FORCE;
1704
1705         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1706                                 NULL);
1707 }
1708 EXPORT_SYMBOL(get_user_pages);
1709
1710 /**
1711  * get_dump_page() - pin user page in memory while writing it to core dump
1712  * @addr: user address
1713  *
1714  * Returns struct page pointer of user page pinned for dump,
1715  * to be freed afterwards by page_cache_release() or put_page().
1716  *
1717  * Returns NULL on any kind of failure - a hole must then be inserted into
1718  * the corefile, to preserve alignment with its headers; and also returns
1719  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1720  * allowing a hole to be left in the corefile to save diskspace.
1721  *
1722  * Called without mmap_sem, but after all other threads have been killed.
1723  */
1724 #ifdef CONFIG_ELF_CORE
1725 struct page *get_dump_page(unsigned long addr)
1726 {
1727         struct vm_area_struct *vma;
1728         struct page *page;
1729
1730         if (__get_user_pages(current, current->mm, addr, 1,
1731                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1732                              NULL) < 1)
1733                 return NULL;
1734         flush_cache_page(vma, addr, page_to_pfn(page));
1735         return page;
1736 }
1737 #endif /* CONFIG_ELF_CORE */
1738
1739 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1740                         spinlock_t **ptl)
1741 {
1742         pgd_t * pgd = pgd_offset(mm, addr);
1743         pud_t * pud = pud_alloc(mm, pgd, addr);
1744         if (pud) {
1745                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1746                 if (pmd) {
1747                         VM_BUG_ON(pmd_trans_huge(*pmd));
1748                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1749                 }
1750         }
1751         return NULL;
1752 }
1753
1754 /*
1755  * This is the old fallback for page remapping.
1756  *
1757  * For historical reasons, it only allows reserved pages. Only
1758  * old drivers should use this, and they needed to mark their
1759  * pages reserved for the old functions anyway.
1760  */
1761 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1762                         struct page *page, pgprot_t prot)
1763 {
1764         struct mm_struct *mm = vma->vm_mm;
1765         int retval;
1766         pte_t *pte;
1767         spinlock_t *ptl;
1768
1769         retval = -EINVAL;
1770         if (PageAnon(page))
1771                 goto out;
1772         retval = -ENOMEM;
1773         flush_dcache_page(page);
1774         pte = get_locked_pte(mm, addr, &ptl);
1775         if (!pte)
1776                 goto out;
1777         retval = -EBUSY;
1778         if (!pte_none(*pte))
1779                 goto out_unlock;
1780
1781         /* Ok, finally just insert the thing.. */
1782         get_page(page);
1783         inc_mm_counter_fast(mm, MM_FILEPAGES);
1784         page_add_file_rmap(page);
1785         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1786
1787         retval = 0;
1788         pte_unmap_unlock(pte, ptl);
1789         return retval;
1790 out_unlock:
1791         pte_unmap_unlock(pte, ptl);
1792 out:
1793         return retval;
1794 }
1795
1796 /**
1797  * vm_insert_page - insert single page into user vma
1798  * @vma: user vma to map to
1799  * @addr: target user address of this page
1800  * @page: source kernel page
1801  *
1802  * This allows drivers to insert individual pages they've allocated
1803  * into a user vma.
1804  *
1805  * The page has to be a nice clean _individual_ kernel allocation.
1806  * If you allocate a compound page, you need to have marked it as
1807  * such (__GFP_COMP), or manually just split the page up yourself
1808  * (see split_page()).
1809  *
1810  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1811  * took an arbitrary page protection parameter. This doesn't allow
1812  * that. Your vma protection will have to be set up correctly, which
1813  * means that if you want a shared writable mapping, you'd better
1814  * ask for a shared writable mapping!
1815  *
1816  * The page does not need to be reserved.
1817  */
1818 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1819                         struct page *page)
1820 {
1821         if (addr < vma->vm_start || addr >= vma->vm_end)
1822                 return -EFAULT;
1823         if (!page_count(page))
1824                 return -EINVAL;
1825         vma->vm_flags |= VM_INSERTPAGE;
1826         return insert_page(vma, addr, page, vma->vm_page_prot);
1827 }
1828 EXPORT_SYMBOL(vm_insert_page);
1829
1830 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1831                         unsigned long pfn, pgprot_t prot)
1832 {
1833         struct mm_struct *mm = vma->vm_mm;
1834         int retval;
1835         pte_t *pte, entry;
1836         spinlock_t *ptl;
1837
1838         retval = -ENOMEM;
1839         pte = get_locked_pte(mm, addr, &ptl);
1840         if (!pte)
1841                 goto out;
1842         retval = -EBUSY;
1843         if (!pte_none(*pte))
1844                 goto out_unlock;
1845
1846         /* Ok, finally just insert the thing.. */
1847         entry = pte_mkspecial(pfn_pte(pfn, prot));
1848         set_pte_at(mm, addr, pte, entry);
1849         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1850
1851         retval = 0;
1852 out_unlock:
1853         pte_unmap_unlock(pte, ptl);
1854 out:
1855         return retval;
1856 }
1857
1858 /**
1859  * vm_insert_pfn - insert single pfn into user vma
1860  * @vma: user vma to map to
1861  * @addr: target user address of this page
1862  * @pfn: source kernel pfn
1863  *
1864  * Similar to vm_inert_page, this allows drivers to insert individual pages
1865  * they've allocated into a user vma. Same comments apply.
1866  *
1867  * This function should only be called from a vm_ops->fault handler, and
1868  * in that case the handler should return NULL.
1869  *
1870  * vma cannot be a COW mapping.
1871  *
1872  * As this is called only for pages that do not currently exist, we
1873  * do not need to flush old virtual caches or the TLB.
1874  */
1875 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1876                         unsigned long pfn)
1877 {
1878         int ret;
1879         pgprot_t pgprot = vma->vm_page_prot;
1880         /*
1881          * Technically, architectures with pte_special can avoid all these
1882          * restrictions (same for remap_pfn_range).  However we would like
1883          * consistency in testing and feature parity among all, so we should
1884          * try to keep these invariants in place for everybody.
1885          */
1886         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1887         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1888                                                 (VM_PFNMAP|VM_MIXEDMAP));
1889         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1890         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1891
1892         if (addr < vma->vm_start || addr >= vma->vm_end)
1893                 return -EFAULT;
1894         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1895                 return -EINVAL;
1896
1897         ret = insert_pfn(vma, addr, pfn, pgprot);
1898
1899         if (ret)
1900                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1901
1902         return ret;
1903 }
1904 EXPORT_SYMBOL(vm_insert_pfn);
1905
1906 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1907                         unsigned long pfn)
1908 {
1909         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1910
1911         if (addr < vma->vm_start || addr >= vma->vm_end)
1912                 return -EFAULT;
1913
1914         /*
1915          * If we don't have pte special, then we have to use the pfn_valid()
1916          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1917          * refcount the page if pfn_valid is true (hence insert_page rather
1918          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1919          * without pte special, it would there be refcounted as a normal page.
1920          */
1921         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1922                 struct page *page;
1923
1924                 page = pfn_to_page(pfn);
1925                 return insert_page(vma, addr, page, vma->vm_page_prot);
1926         }
1927         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1928 }
1929 EXPORT_SYMBOL(vm_insert_mixed);
1930
1931 /*
1932  * maps a range of physical memory into the requested pages. the old
1933  * mappings are removed. any references to nonexistent pages results
1934  * in null mappings (currently treated as "copy-on-access")
1935  */
1936 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1937                         unsigned long addr, unsigned long end,
1938                         unsigned long pfn, pgprot_t prot)
1939 {
1940         pte_t *pte;
1941         spinlock_t *ptl;
1942
1943         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1944         if (!pte)
1945                 return -ENOMEM;
1946         arch_enter_lazy_mmu_mode();
1947         do {
1948                 BUG_ON(!pte_none(*pte));
1949                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1950                 pfn++;
1951         } while (pte++, addr += PAGE_SIZE, addr != end);
1952         arch_leave_lazy_mmu_mode();
1953         pte_unmap_unlock(pte - 1, ptl);
1954         return 0;
1955 }
1956
1957 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1958                         unsigned long addr, unsigned long end,
1959                         unsigned long pfn, pgprot_t prot)
1960 {
1961         pmd_t *pmd;
1962         unsigned long next;
1963
1964         pfn -= addr >> PAGE_SHIFT;
1965         pmd = pmd_alloc(mm, pud, addr);
1966         if (!pmd)
1967                 return -ENOMEM;
1968         VM_BUG_ON(pmd_trans_huge(*pmd));
1969         do {
1970                 next = pmd_addr_end(addr, end);
1971                 if (remap_pte_range(mm, pmd, addr, next,
1972                                 pfn + (addr >> PAGE_SHIFT), prot))
1973                         return -ENOMEM;
1974         } while (pmd++, addr = next, addr != end);
1975         return 0;
1976 }
1977
1978 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1979                         unsigned long addr, unsigned long end,
1980                         unsigned long pfn, pgprot_t prot)
1981 {
1982         pud_t *pud;
1983         unsigned long next;
1984
1985         pfn -= addr >> PAGE_SHIFT;
1986         pud = pud_alloc(mm, pgd, addr);
1987         if (!pud)
1988                 return -ENOMEM;
1989         do {
1990                 next = pud_addr_end(addr, end);
1991                 if (remap_pmd_range(mm, pud, addr, next,
1992                                 pfn + (addr >> PAGE_SHIFT), prot))
1993                         return -ENOMEM;
1994         } while (pud++, addr = next, addr != end);
1995         return 0;
1996 }
1997
1998 /**
1999  * remap_pfn_range - remap kernel memory to userspace
2000  * @vma: user vma to map to
2001  * @addr: target user address to start at
2002  * @pfn: physical address of kernel memory
2003  * @size: size of map area
2004  * @prot: page protection flags for this mapping
2005  *
2006  *  Note: this is only safe if the mm semaphore is held when called.
2007  */
2008 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2009                     unsigned long pfn, unsigned long size, pgprot_t prot)
2010 {
2011         pgd_t *pgd;
2012         unsigned long next;
2013         unsigned long end = addr + PAGE_ALIGN(size);
2014         struct mm_struct *mm = vma->vm_mm;
2015         int err;
2016
2017         /*
2018          * Physically remapped pages are special. Tell the
2019          * rest of the world about it:
2020          *   VM_IO tells people not to look at these pages
2021          *      (accesses can have side effects).
2022          *   VM_RESERVED is specified all over the place, because
2023          *      in 2.4 it kept swapout's vma scan off this vma; but
2024          *      in 2.6 the LRU scan won't even find its pages, so this
2025          *      flag means no more than count its pages in reserved_vm,
2026          *      and omit it from core dump, even when VM_IO turned off.
2027          *   VM_PFNMAP tells the core MM that the base pages are just
2028          *      raw PFN mappings, and do not have a "struct page" associated
2029          *      with them.
2030          *
2031          * There's a horrible special case to handle copy-on-write
2032          * behaviour that some programs depend on. We mark the "original"
2033          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2034          */
2035         if (addr == vma->vm_start && end == vma->vm_end) {
2036                 vma->vm_pgoff = pfn;
2037                 vma->vm_flags |= VM_PFN_AT_MMAP;
2038         } else if (is_cow_mapping(vma->vm_flags))
2039                 return -EINVAL;
2040
2041         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2042
2043         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2044         if (err) {
2045                 /*
2046                  * To indicate that track_pfn related cleanup is not
2047                  * needed from higher level routine calling unmap_vmas
2048                  */
2049                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2050                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2051                 return -EINVAL;
2052         }
2053
2054         BUG_ON(addr >= end);
2055         pfn -= addr >> PAGE_SHIFT;
2056         pgd = pgd_offset(mm, addr);
2057         flush_cache_range(vma, addr, end);
2058         do {
2059                 next = pgd_addr_end(addr, end);
2060                 err = remap_pud_range(mm, pgd, addr, next,
2061                                 pfn + (addr >> PAGE_SHIFT), prot);
2062                 if (err)
2063                         break;
2064         } while (pgd++, addr = next, addr != end);
2065
2066         if (err)
2067                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2068
2069         return err;
2070 }
2071 EXPORT_SYMBOL(remap_pfn_range);
2072
2073 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2074                                      unsigned long addr, unsigned long end,
2075                                      pte_fn_t fn, void *data)
2076 {
2077         pte_t *pte;
2078         int err;
2079         pgtable_t token;
2080         spinlock_t *uninitialized_var(ptl);
2081
2082         pte = (mm == &init_mm) ?
2083                 pte_alloc_kernel(pmd, addr) :
2084                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2085         if (!pte)
2086                 return -ENOMEM;
2087
2088         BUG_ON(pmd_huge(*pmd));
2089
2090         arch_enter_lazy_mmu_mode();
2091
2092         token = pmd_pgtable(*pmd);
2093
2094         do {
2095                 err = fn(pte++, token, addr, data);
2096                 if (err)
2097                         break;
2098         } while (addr += PAGE_SIZE, addr != end);
2099
2100         arch_leave_lazy_mmu_mode();
2101
2102         if (mm != &init_mm)
2103                 pte_unmap_unlock(pte-1, ptl);
2104         return err;
2105 }
2106
2107 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2108                                      unsigned long addr, unsigned long end,
2109                                      pte_fn_t fn, void *data)
2110 {
2111         pmd_t *pmd;
2112         unsigned long next;
2113         int err;
2114
2115         BUG_ON(pud_huge(*pud));
2116
2117         pmd = pmd_alloc(mm, pud, addr);
2118         if (!pmd)
2119                 return -ENOMEM;
2120         do {
2121                 next = pmd_addr_end(addr, end);
2122                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2123                 if (err)
2124                         break;
2125         } while (pmd++, addr = next, addr != end);
2126         return err;
2127 }
2128
2129 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2130                                      unsigned long addr, unsigned long end,
2131                                      pte_fn_t fn, void *data)
2132 {
2133         pud_t *pud;
2134         unsigned long next;
2135         int err;
2136
2137         pud = pud_alloc(mm, pgd, addr);
2138         if (!pud)
2139                 return -ENOMEM;
2140         do {
2141                 next = pud_addr_end(addr, end);
2142                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2143                 if (err)
2144                         break;
2145         } while (pud++, addr = next, addr != end);
2146         return err;
2147 }
2148
2149 /*
2150  * Scan a region of virtual memory, filling in page tables as necessary
2151  * and calling a provided function on each leaf page table.
2152  */
2153 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2154                         unsigned long size, pte_fn_t fn, void *data)
2155 {
2156         pgd_t *pgd;
2157         unsigned long next;
2158         unsigned long end = addr + size;
2159         int err;
2160
2161         BUG_ON(addr >= end);
2162         pgd = pgd_offset(mm, addr);
2163         do {
2164                 next = pgd_addr_end(addr, end);
2165                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2166                 if (err)
2167                         break;
2168         } while (pgd++, addr = next, addr != end);
2169
2170         return err;
2171 }
2172 EXPORT_SYMBOL_GPL(apply_to_page_range);
2173
2174 /*
2175  * handle_pte_fault chooses page fault handler according to an entry
2176  * which was read non-atomically.  Before making any commitment, on
2177  * those architectures or configurations (e.g. i386 with PAE) which
2178  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2179  * must check under lock before unmapping the pte and proceeding
2180  * (but do_wp_page is only called after already making such a check;
2181  * and do_anonymous_page can safely check later on).
2182  */
2183 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2184                                 pte_t *page_table, pte_t orig_pte)
2185 {
2186         int same = 1;
2187 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2188         if (sizeof(pte_t) > sizeof(unsigned long)) {
2189                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2190                 spin_lock(ptl);
2191                 same = pte_same(*page_table, orig_pte);
2192                 spin_unlock(ptl);
2193         }
2194 #endif
2195         pte_unmap(page_table);
2196         return same;
2197 }
2198
2199 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2200 {
2201         /*
2202          * If the source page was a PFN mapping, we don't have
2203          * a "struct page" for it. We do a best-effort copy by
2204          * just copying from the original user address. If that
2205          * fails, we just zero-fill it. Live with it.
2206          */
2207         if (unlikely(!src)) {
2208                 void *kaddr = kmap_atomic(dst, KM_USER0);
2209                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2210
2211                 /*
2212                  * This really shouldn't fail, because the page is there
2213                  * in the page tables. But it might just be unreadable,
2214                  * in which case we just give up and fill the result with
2215                  * zeroes.
2216                  */
2217                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2218                         clear_page(kaddr);
2219                 kunmap_atomic(kaddr, KM_USER0);
2220                 flush_dcache_page(dst);
2221         } else
2222                 copy_user_highpage(dst, src, va, vma);
2223 }
2224
2225 /*
2226  * This routine handles present pages, when users try to write
2227  * to a shared page. It is done by copying the page to a new address
2228  * and decrementing the shared-page counter for the old page.
2229  *
2230  * Note that this routine assumes that the protection checks have been
2231  * done by the caller (the low-level page fault routine in most cases).
2232  * Thus we can safely just mark it writable once we've done any necessary
2233  * COW.
2234  *
2235  * We also mark the page dirty at this point even though the page will
2236  * change only once the write actually happens. This avoids a few races,
2237  * and potentially makes it more efficient.
2238  *
2239  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2240  * but allow concurrent faults), with pte both mapped and locked.
2241  * We return with mmap_sem still held, but pte unmapped and unlocked.
2242  */
2243 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2244                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2245                 spinlock_t *ptl, pte_t orig_pte)
2246         __releases(ptl)
2247 {
2248         struct page *old_page, *new_page;
2249         pte_t entry;
2250         int ret = 0;
2251         int page_mkwrite = 0;
2252         struct page *dirty_page = NULL;
2253
2254         old_page = vm_normal_page(vma, address, orig_pte);
2255         if (!old_page) {
2256                 /*
2257                  * VM_MIXEDMAP !pfn_valid() case
2258                  *
2259                  * We should not cow pages in a shared writeable mapping.
2260                  * Just mark the pages writable as we can't do any dirty
2261                  * accounting on raw pfn maps.
2262                  */
2263                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2264                                      (VM_WRITE|VM_SHARED))
2265                         goto reuse;
2266                 goto gotten;
2267         }
2268
2269         /*
2270          * Take out anonymous pages first, anonymous shared vmas are
2271          * not dirty accountable.
2272          */
2273         if (PageAnon(old_page) && !PageKsm(old_page)) {
2274                 if (!trylock_page(old_page)) {
2275                         page_cache_get(old_page);
2276                         pte_unmap_unlock(page_table, ptl);
2277                         lock_page(old_page);
2278                         page_table = pte_offset_map_lock(mm, pmd, address,
2279                                                          &ptl);
2280                         if (!pte_same(*page_table, orig_pte)) {
2281                                 unlock_page(old_page);
2282                                 goto unlock;
2283                         }
2284                         page_cache_release(old_page);
2285                 }
2286                 if (reuse_swap_page(old_page)) {
2287                         /*
2288                          * The page is all ours.  Move it to our anon_vma so
2289                          * the rmap code will not search our parent or siblings.
2290                          * Protected against the rmap code by the page lock.
2291                          */
2292                         page_move_anon_rmap(old_page, vma, address);
2293                         unlock_page(old_page);
2294                         goto reuse;
2295                 }
2296                 unlock_page(old_page);
2297         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2298                                         (VM_WRITE|VM_SHARED))) {
2299                 /*
2300                  * Only catch write-faults on shared writable pages,
2301                  * read-only shared pages can get COWed by
2302                  * get_user_pages(.write=1, .force=1).
2303                  */
2304                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2305                         struct vm_fault vmf;
2306                         int tmp;
2307
2308                         vmf.virtual_address = (void __user *)(address &
2309                                                                 PAGE_MASK);
2310                         vmf.pgoff = old_page->index;
2311                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2312                         vmf.page = old_page;
2313
2314                         /*
2315                          * Notify the address space that the page is about to
2316                          * become writable so that it can prohibit this or wait
2317                          * for the page to get into an appropriate state.
2318                          *
2319                          * We do this without the lock held, so that it can
2320                          * sleep if it needs to.
2321                          */
2322                         page_cache_get(old_page);
2323                         pte_unmap_unlock(page_table, ptl);
2324
2325                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2326                         if (unlikely(tmp &
2327                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2328                                 ret = tmp;
2329                                 goto unwritable_page;
2330                         }
2331                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2332                                 lock_page(old_page);
2333                                 if (!old_page->mapping) {
2334                                         ret = 0; /* retry the fault */
2335                                         unlock_page(old_page);
2336                                         goto unwritable_page;
2337                                 }
2338                         } else
2339                                 VM_BUG_ON(!PageLocked(old_page));
2340
2341                         /*
2342                          * Since we dropped the lock we need to revalidate
2343                          * the PTE as someone else may have changed it.  If
2344                          * they did, we just return, as we can count on the
2345                          * MMU to tell us if they didn't also make it writable.
2346                          */
2347                         page_table = pte_offset_map_lock(mm, pmd, address,
2348                                                          &ptl);
2349                         if (!pte_same(*page_table, orig_pte)) {
2350                                 unlock_page(old_page);
2351                                 goto unlock;
2352                         }
2353
2354                         page_mkwrite = 1;
2355                 }
2356                 dirty_page = old_page;
2357                 get_page(dirty_page);
2358
2359 reuse:
2360                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2361                 entry = pte_mkyoung(orig_pte);
2362                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2363                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2364                         update_mmu_cache(vma, address, page_table);
2365                 pte_unmap_unlock(page_table, ptl);
2366                 ret |= VM_FAULT_WRITE;
2367
2368                 if (!dirty_page)
2369                         return ret;
2370
2371                 /*
2372                  * Yes, Virginia, this is actually required to prevent a race
2373                  * with clear_page_dirty_for_io() from clearing the page dirty
2374                  * bit after it clear all dirty ptes, but before a racing
2375                  * do_wp_page installs a dirty pte.
2376                  *
2377                  * __do_fault is protected similarly.
2378                  */
2379                 if (!page_mkwrite) {
2380                         wait_on_page_locked(dirty_page);
2381                         set_page_dirty_balance(dirty_page, page_mkwrite);
2382                 }
2383                 put_page(dirty_page);
2384                 if (page_mkwrite) {
2385                         struct address_space *mapping = dirty_page->mapping;
2386
2387                         set_page_dirty(dirty_page);
2388                         unlock_page(dirty_page);
2389                         page_cache_release(dirty_page);
2390                         if (mapping)    {
2391                                 /*
2392                                  * Some device drivers do not set page.mapping
2393                                  * but still dirty their pages
2394                                  */
2395                                 balance_dirty_pages_ratelimited(mapping);
2396                         }
2397                 }
2398
2399                 /* file_update_time outside page_lock */
2400                 if (vma->vm_file)
2401                         file_update_time(vma->vm_file);
2402
2403                 return ret;
2404         }
2405
2406         /*
2407          * Ok, we need to copy. Oh, well..
2408          */
2409         page_cache_get(old_page);
2410 gotten:
2411         pte_unmap_unlock(page_table, ptl);
2412
2413         if (unlikely(anon_vma_prepare(vma)))
2414                 goto oom;
2415
2416         if (is_zero_pfn(pte_pfn(orig_pte))) {
2417                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2418                 if (!new_page)
2419                         goto oom;
2420         } else {
2421                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2422                 if (!new_page)
2423                         goto oom;
2424                 cow_user_page(new_page, old_page, address, vma);
2425         }
2426         __SetPageUptodate(new_page);
2427
2428         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2429                 goto oom_free_new;
2430
2431         /*
2432          * Re-check the pte - we dropped the lock
2433          */
2434         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2435         if (likely(pte_same(*page_table, orig_pte))) {
2436                 if (old_page) {
2437                         if (!PageAnon(old_page)) {
2438                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2439                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2440                         }
2441                 } else
2442                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2443                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2444                 entry = mk_pte(new_page, vma->vm_page_prot);
2445                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2446                 /*
2447                  * Clear the pte entry and flush it first, before updating the
2448                  * pte with the new entry. This will avoid a race condition
2449                  * seen in the presence of one thread doing SMC and another
2450                  * thread doing COW.
2451                  */
2452                 ptep_clear_flush(vma, address, page_table);
2453                 page_add_new_anon_rmap(new_page, vma, address);
2454                 /*
2455                  * We call the notify macro here because, when using secondary
2456                  * mmu page tables (such as kvm shadow page tables), we want the
2457                  * new page to be mapped directly into the secondary page table.
2458                  */
2459                 set_pte_at_notify(mm, address, page_table, entry);
2460                 update_mmu_cache(vma, address, page_table);
2461                 if (old_page) {
2462                         /*
2463                          * Only after switching the pte to the new page may
2464                          * we remove the mapcount here. Otherwise another
2465                          * process may come and find the rmap count decremented
2466                          * before the pte is switched to the new page, and
2467                          * "reuse" the old page writing into it while our pte
2468                          * here still points into it and can be read by other
2469                          * threads.
2470                          *
2471                          * The critical issue is to order this
2472                          * page_remove_rmap with the ptp_clear_flush above.
2473                          * Those stores are ordered by (if nothing else,)
2474                          * the barrier present in the atomic_add_negative
2475                          * in page_remove_rmap.
2476                          *
2477                          * Then the TLB flush in ptep_clear_flush ensures that
2478                          * no process can access the old page before the
2479                          * decremented mapcount is visible. And the old page
2480                          * cannot be reused until after the decremented
2481                          * mapcount is visible. So transitively, TLBs to
2482                          * old page will be flushed before it can be reused.
2483                          */
2484                         page_remove_rmap(old_page);
2485                 }
2486
2487                 /* Free the old page.. */
2488                 new_page = old_page;
2489                 ret |= VM_FAULT_WRITE;
2490         } else
2491                 mem_cgroup_uncharge_page(new_page);
2492
2493         if (new_page)
2494                 page_cache_release(new_page);
2495 unlock:
2496         pte_unmap_unlock(page_table, ptl);
2497         if (old_page) {
2498                 /*
2499                  * Don't let another task, with possibly unlocked vma,
2500                  * keep the mlocked page.
2501                  */
2502                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2503                         lock_page(old_page);    /* LRU manipulation */
2504                         munlock_vma_page(old_page);
2505                         unlock_page(old_page);
2506                 }
2507                 page_cache_release(old_page);
2508         }
2509         return ret;
2510 oom_free_new:
2511         page_cache_release(new_page);
2512 oom:
2513         if (old_page) {
2514                 if (page_mkwrite) {
2515                         unlock_page(old_page);
2516                         page_cache_release(old_page);
2517                 }
2518                 page_cache_release(old_page);
2519         }
2520         return VM_FAULT_OOM;
2521
2522 unwritable_page:
2523         page_cache_release(old_page);
2524         return ret;
2525 }
2526
2527 /*
2528  * Helper functions for unmap_mapping_range().
2529  *
2530  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2531  *
2532  * We have to restart searching the prio_tree whenever we drop the lock,
2533  * since the iterator is only valid while the lock is held, and anyway
2534  * a later vma might be split and reinserted earlier while lock dropped.
2535  *
2536  * The list of nonlinear vmas could be handled more efficiently, using
2537  * a placeholder, but handle it in the same way until a need is shown.
2538  * It is important to search the prio_tree before nonlinear list: a vma
2539  * may become nonlinear and be shifted from prio_tree to nonlinear list
2540  * while the lock is dropped; but never shifted from list to prio_tree.
2541  *
2542  * In order to make forward progress despite restarting the search,
2543  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2544  * quickly skip it next time around.  Since the prio_tree search only
2545  * shows us those vmas affected by unmapping the range in question, we
2546  * can't efficiently keep all vmas in step with mapping->truncate_count:
2547  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2548  * mapping->truncate_count and vma->vm_truncate_count are protected by
2549  * i_mmap_lock.
2550  *
2551  * In order to make forward progress despite repeatedly restarting some
2552  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2553  * and restart from that address when we reach that vma again.  It might
2554  * have been split or merged, shrunk or extended, but never shifted: so
2555  * restart_addr remains valid so long as it remains in the vma's range.
2556  * unmap_mapping_range forces truncate_count to leap over page-aligned
2557  * values so we can save vma's restart_addr in its truncate_count field.
2558  */
2559 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2560
2561 static void reset_vma_truncate_counts(struct address_space *mapping)
2562 {
2563         struct vm_area_struct *vma;
2564         struct prio_tree_iter iter;
2565
2566         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2567                 vma->vm_truncate_count = 0;
2568         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2569                 vma->vm_truncate_count = 0;
2570 }
2571
2572 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2573                 unsigned long start_addr, unsigned long end_addr,
2574                 struct zap_details *details)
2575 {
2576         unsigned long restart_addr;
2577         int need_break;
2578
2579         /*
2580          * files that support invalidating or truncating portions of the
2581          * file from under mmaped areas must have their ->fault function
2582          * return a locked page (and set VM_FAULT_LOCKED in the return).
2583          * This provides synchronisation against concurrent unmapping here.
2584          */
2585
2586 again:
2587         restart_addr = vma->vm_truncate_count;
2588         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2589                 start_addr = restart_addr;
2590                 if (start_addr >= end_addr) {
2591                         /* Top of vma has been split off since last time */
2592                         vma->vm_truncate_count = details->truncate_count;
2593                         return 0;
2594                 }
2595         }
2596
2597         restart_addr = zap_page_range(vma, start_addr,
2598                                         end_addr - start_addr, details);
2599         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2600
2601         if (restart_addr >= end_addr) {
2602                 /* We have now completed this vma: mark it so */
2603                 vma->vm_truncate_count = details->truncate_count;
2604                 if (!need_break)
2605                         return 0;
2606         } else {
2607                 /* Note restart_addr in vma's truncate_count field */
2608                 vma->vm_truncate_count = restart_addr;
2609                 if (!need_break)
2610                         goto again;
2611         }
2612
2613         spin_unlock(details->i_mmap_lock);
2614         cond_resched();
2615         spin_lock(details->i_mmap_lock);
2616         return -EINTR;
2617 }
2618
2619 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2620                                             struct zap_details *details)
2621 {
2622         struct vm_area_struct *vma;
2623         struct prio_tree_iter iter;
2624         pgoff_t vba, vea, zba, zea;
2625
2626 restart:
2627         vma_prio_tree_foreach(vma, &iter, root,
2628                         details->first_index, details->last_index) {
2629                 /* Skip quickly over those we have already dealt with */
2630                 if (vma->vm_truncate_count == details->truncate_count)
2631                         continue;
2632
2633                 vba = vma->vm_pgoff;
2634                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2635                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2636                 zba = details->first_index;
2637                 if (zba < vba)
2638                         zba = vba;
2639                 zea = details->last_index;
2640                 if (zea > vea)
2641                         zea = vea;
2642
2643                 if (unmap_mapping_range_vma(vma,
2644                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2645                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2646                                 details) < 0)
2647                         goto restart;
2648         }
2649 }
2650
2651 static inline void unmap_mapping_range_list(struct list_head *head,
2652                                             struct zap_details *details)
2653 {
2654         struct vm_area_struct *vma;
2655
2656         /*
2657          * In nonlinear VMAs there is no correspondence between virtual address
2658          * offset and file offset.  So we must perform an exhaustive search
2659          * across *all* the pages in each nonlinear VMA, not just the pages
2660          * whose virtual address lies outside the file truncation point.
2661          */
2662 restart:
2663         list_for_each_entry(vma, head, shared.vm_set.list) {
2664                 /* Skip quickly over those we have already dealt with */
2665                 if (vma->vm_truncate_count == details->truncate_count)
2666                         continue;
2667                 details->nonlinear_vma = vma;
2668                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2669                                         vma->vm_end, details) < 0)
2670                         goto restart;
2671         }
2672 }
2673
2674 /**
2675  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2676  * @mapping: the address space containing mmaps to be unmapped.
2677  * @holebegin: byte in first page to unmap, relative to the start of
2678  * the underlying file.  This will be rounded down to a PAGE_SIZE
2679  * boundary.  Note that this is different from truncate_pagecache(), which
2680  * must keep the partial page.  In contrast, we must get rid of
2681  * partial pages.
2682  * @holelen: size of prospective hole in bytes.  This will be rounded
2683  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2684  * end of the file.
2685  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2686  * but 0 when invalidating pagecache, don't throw away private data.
2687  */
2688 void unmap_mapping_range(struct address_space *mapping,
2689                 loff_t const holebegin, loff_t const holelen, int even_cows)
2690 {
2691         struct zap_details details;
2692         pgoff_t hba = holebegin >> PAGE_SHIFT;
2693         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2694
2695         /* Check for overflow. */
2696         if (sizeof(holelen) > sizeof(hlen)) {
2697                 long long holeend =
2698                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2699                 if (holeend & ~(long long)ULONG_MAX)
2700                         hlen = ULONG_MAX - hba + 1;
2701         }
2702
2703         details.check_mapping = even_cows? NULL: mapping;
2704         details.nonlinear_vma = NULL;
2705         details.first_index = hba;
2706         details.last_index = hba + hlen - 1;
2707         if (details.last_index < details.first_index)
2708                 details.last_index = ULONG_MAX;
2709         details.i_mmap_lock = &mapping->i_mmap_lock;
2710
2711         mutex_lock(&mapping->unmap_mutex);
2712         spin_lock(&mapping->i_mmap_lock);
2713
2714         /* Protect against endless unmapping loops */
2715         mapping->truncate_count++;
2716         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2717                 if (mapping->truncate_count == 0)
2718                         reset_vma_truncate_counts(mapping);
2719                 mapping->truncate_count++;
2720         }
2721         details.truncate_count = mapping->truncate_count;
2722
2723         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2724                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2725         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2726                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2727         spin_unlock(&mapping->i_mmap_lock);
2728         mutex_unlock(&mapping->unmap_mutex);
2729 }
2730 EXPORT_SYMBOL(unmap_mapping_range);
2731
2732 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2733 {
2734         struct address_space *mapping = inode->i_mapping;
2735
2736         /*
2737          * If the underlying filesystem is not going to provide
2738          * a way to truncate a range of blocks (punch a hole) -
2739          * we should return failure right now.
2740          */
2741         if (!inode->i_op->truncate_range)
2742                 return -ENOSYS;
2743
2744         mutex_lock(&inode->i_mutex);
2745         down_write(&inode->i_alloc_sem);
2746         unmap_mapping_range(mapping, offset, (end - offset), 1);
2747         truncate_inode_pages_range(mapping, offset, end);
2748         unmap_mapping_range(mapping, offset, (end - offset), 1);
2749         inode->i_op->truncate_range(inode, offset, end);
2750         up_write(&inode->i_alloc_sem);
2751         mutex_unlock(&inode->i_mutex);
2752
2753         return 0;
2754 }
2755
2756 /*
2757  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2758  * but allow concurrent faults), and pte mapped but not yet locked.
2759  * We return with mmap_sem still held, but pte unmapped and unlocked.
2760  */
2761 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2762                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2763                 unsigned int flags, pte_t orig_pte)
2764 {
2765         spinlock_t *ptl;
2766         struct page *page, *swapcache = NULL;
2767         swp_entry_t entry;
2768         pte_t pte;
2769         int locked;
2770         struct mem_cgroup *ptr = NULL;
2771         int exclusive = 0;
2772         int ret = 0;
2773
2774         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2775                 goto out;
2776
2777         entry = pte_to_swp_entry(orig_pte);
2778         if (unlikely(non_swap_entry(entry))) {
2779                 if (is_migration_entry(entry)) {
2780                         migration_entry_wait(mm, pmd, address);
2781                 } else if (is_hwpoison_entry(entry)) {
2782                         ret = VM_FAULT_HWPOISON;
2783                 } else {
2784                         print_bad_pte(vma, address, orig_pte, NULL);
2785                         ret = VM_FAULT_SIGBUS;
2786                 }
2787                 goto out;
2788         }
2789         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2790         page = lookup_swap_cache(entry);
2791         if (!page) {
2792                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2793                 page = swapin_readahead(entry,
2794                                         GFP_HIGHUSER_MOVABLE, vma, address);
2795                 if (!page) {
2796                         /*
2797                          * Back out if somebody else faulted in this pte
2798                          * while we released the pte lock.
2799                          */
2800                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2801                         if (likely(pte_same(*page_table, orig_pte)))
2802                                 ret = VM_FAULT_OOM;
2803                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2804                         goto unlock;
2805                 }
2806
2807                 /* Had to read the page from swap area: Major fault */
2808                 ret = VM_FAULT_MAJOR;
2809                 count_vm_event(PGMAJFAULT);
2810         } else if (PageHWPoison(page)) {
2811                 /*
2812                  * hwpoisoned dirty swapcache pages are kept for killing
2813                  * owner processes (which may be unknown at hwpoison time)
2814                  */
2815                 ret = VM_FAULT_HWPOISON;
2816                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2817                 goto out_release;
2818         }
2819
2820         locked = lock_page_or_retry(page, mm, flags);
2821         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2822         if (!locked) {
2823                 ret |= VM_FAULT_RETRY;
2824                 goto out_release;
2825         }
2826
2827         /*
2828          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2829          * release the swapcache from under us.  The page pin, and pte_same
2830          * test below, are not enough to exclude that.  Even if it is still
2831          * swapcache, we need to check that the page's swap has not changed.
2832          */
2833         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2834                 goto out_page;
2835
2836         if (ksm_might_need_to_copy(page, vma, address)) {
2837                 swapcache = page;
2838                 page = ksm_does_need_to_copy(page, vma, address);
2839
2840                 if (unlikely(!page)) {
2841                         ret = VM_FAULT_OOM;
2842                         page = swapcache;
2843                         swapcache = NULL;
2844                         goto out_page;
2845                 }
2846         }
2847
2848         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2849                 ret = VM_FAULT_OOM;
2850                 goto out_page;
2851         }
2852
2853         /*
2854          * Back out if somebody else already faulted in this pte.
2855          */
2856         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2857         if (unlikely(!pte_same(*page_table, orig_pte)))
2858                 goto out_nomap;
2859
2860         if (unlikely(!PageUptodate(page))) {
2861                 ret = VM_FAULT_SIGBUS;
2862                 goto out_nomap;
2863         }
2864
2865         /*
2866          * The page isn't present yet, go ahead with the fault.
2867          *
2868          * Be careful about the sequence of operations here.
2869          * To get its accounting right, reuse_swap_page() must be called
2870          * while the page is counted on swap but not yet in mapcount i.e.
2871          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2872          * must be called after the swap_free(), or it will never succeed.
2873          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2874          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2875          * in page->private. In this case, a record in swap_cgroup  is silently
2876          * discarded at swap_free().
2877          */
2878
2879         inc_mm_counter_fast(mm, MM_ANONPAGES);
2880         dec_mm_counter_fast(mm, MM_SWAPENTS);
2881         pte = mk_pte(page, vma->vm_page_prot);
2882         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2883                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2884                 flags &= ~FAULT_FLAG_WRITE;
2885                 ret |= VM_FAULT_WRITE;
2886                 exclusive = 1;
2887         }
2888         flush_icache_page(vma, page);
2889         set_pte_at(mm, address, page_table, pte);
2890         do_page_add_anon_rmap(page, vma, address, exclusive);
2891         /* It's better to call commit-charge after rmap is established */
2892         mem_cgroup_commit_charge_swapin(page, ptr);
2893
2894         swap_free(entry);
2895         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2896                 try_to_free_swap(page);
2897         unlock_page(page);
2898         if (swapcache) {
2899                 /*
2900                  * Hold the lock to avoid the swap entry to be reused
2901                  * until we take the PT lock for the pte_same() check
2902                  * (to avoid false positives from pte_same). For
2903                  * further safety release the lock after the swap_free
2904                  * so that the swap count won't change under a
2905                  * parallel locked swapcache.
2906                  */
2907                 unlock_page(swapcache);
2908                 page_cache_release(swapcache);
2909         }
2910
2911         if (flags & FAULT_FLAG_WRITE) {
2912                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2913                 if (ret & VM_FAULT_ERROR)
2914                         ret &= VM_FAULT_ERROR;
2915                 goto out;
2916         }
2917
2918         /* No need to invalidate - it was non-present before */
2919         update_mmu_cache(vma, address, page_table);
2920 unlock:
2921         pte_unmap_unlock(page_table, ptl);
2922 out:
2923         return ret;
2924 out_nomap:
2925         mem_cgroup_cancel_charge_swapin(ptr);
2926         pte_unmap_unlock(page_table, ptl);
2927 out_page:
2928         unlock_page(page);
2929 out_release:
2930         page_cache_release(page);
2931         if (swapcache) {
2932                 unlock_page(swapcache);
2933                 page_cache_release(swapcache);
2934         }
2935         return ret;
2936 }
2937
2938 /*
2939  * This is like a special single-page "expand_{down|up}wards()",
2940  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2941  * doesn't hit another vma.
2942  */
2943 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2944 {
2945         address &= PAGE_MASK;
2946         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2947                 struct vm_area_struct *prev = vma->vm_prev;
2948
2949                 /*
2950                  * Is there a mapping abutting this one below?
2951                  *
2952                  * That's only ok if it's the same stack mapping
2953                  * that has gotten split..
2954                  */
2955                 if (prev && prev->vm_end == address)
2956                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2957
2958                 expand_stack(vma, address - PAGE_SIZE);
2959         }
2960         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2961                 struct vm_area_struct *next = vma->vm_next;
2962
2963                 /* As VM_GROWSDOWN but s/below/above/ */
2964                 if (next && next->vm_start == address + PAGE_SIZE)
2965                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2966
2967                 expand_upwards(vma, address + PAGE_SIZE);
2968         }
2969         return 0;
2970 }
2971
2972 /*
2973  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2974  * but allow concurrent faults), and pte mapped but not yet locked.
2975  * We return with mmap_sem still held, but pte unmapped and unlocked.
2976  */
2977 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2978                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2979                 unsigned int flags)
2980 {
2981         struct page *page;
2982         spinlock_t *ptl;
2983         pte_t entry;
2984
2985         pte_unmap(page_table);
2986
2987         /* Check if we need to add a guard page to the stack */
2988         if (check_stack_guard_page(vma, address) < 0)
2989                 return VM_FAULT_SIGBUS;
2990
2991         /* Use the zero-page for reads */
2992         if (!(flags & FAULT_FLAG_WRITE)) {
2993                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2994                                                 vma->vm_page_prot));
2995                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2996                 if (!pte_none(*page_table))
2997                         goto unlock;
2998                 goto setpte;
2999         }
3000
3001         /* Allocate our own private page. */
3002         if (unlikely(anon_vma_prepare(vma)))
3003                 goto oom;
3004         page = alloc_zeroed_user_highpage_movable(vma, address);
3005         if (!page)
3006                 goto oom;
3007         __SetPageUptodate(page);
3008
3009         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3010                 goto oom_free_page;
3011
3012         entry = mk_pte(page, vma->vm_page_prot);
3013         if (vma->vm_flags & VM_WRITE)
3014                 entry = pte_mkwrite(pte_mkdirty(entry));
3015
3016         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3017         if (!pte_none(*page_table))
3018                 goto release;
3019
3020         inc_mm_counter_fast(mm, MM_ANONPAGES);
3021         page_add_new_anon_rmap(page, vma, address);
3022 setpte:
3023         set_pte_at(mm, address, page_table, entry);
3024
3025         /* No need to invalidate - it was non-present before */
3026         update_mmu_cache(vma, address, page_table);
3027 unlock:
3028         pte_unmap_unlock(page_table, ptl);
3029         return 0;
3030 release:
3031         mem_cgroup_uncharge_page(page);
3032         page_cache_release(page);
3033         goto unlock;
3034 oom_free_page:
3035         page_cache_release(page);
3036 oom:
3037         return VM_FAULT_OOM;
3038 }
3039
3040 /*
3041  * __do_fault() tries to create a new page mapping. It aggressively
3042  * tries to share with existing pages, but makes a separate copy if
3043  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3044  * the next page fault.
3045  *
3046  * As this is called only for pages that do not currently exist, we
3047  * do not need to flush old virtual caches or the TLB.
3048  *
3049  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3050  * but allow concurrent faults), and pte neither mapped nor locked.
3051  * We return with mmap_sem still held, but pte unmapped and unlocked.
3052  */
3053 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3054                 unsigned long address, pmd_t *pmd,
3055                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3056 {
3057         pte_t *page_table;
3058         spinlock_t *ptl;
3059         struct page *page;
3060         pte_t entry;
3061         int anon = 0;
3062         int charged = 0;
3063         struct page *dirty_page = NULL;
3064         struct vm_fault vmf;
3065         int ret;
3066         int page_mkwrite = 0;
3067
3068         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3069         vmf.pgoff = pgoff;
3070         vmf.flags = flags;
3071         vmf.page = NULL;
3072
3073         ret = vma->vm_ops->fault(vma, &vmf);
3074         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3075                             VM_FAULT_RETRY)))
3076                 return ret;
3077
3078         if (unlikely(PageHWPoison(vmf.page))) {
3079                 if (ret & VM_FAULT_LOCKED)
3080                         unlock_page(vmf.page);
3081                 return VM_FAULT_HWPOISON;
3082         }
3083
3084         /*
3085          * For consistency in subsequent calls, make the faulted page always
3086          * locked.
3087          */
3088         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3089                 lock_page(vmf.page);
3090         else
3091                 VM_BUG_ON(!PageLocked(vmf.page));
3092
3093         /*
3094          * Should we do an early C-O-W break?
3095          */
3096         page = vmf.page;
3097         if (flags & FAULT_FLAG_WRITE) {
3098                 if (!(vma->vm_flags & VM_SHARED)) {
3099                         anon = 1;
3100                         if (unlikely(anon_vma_prepare(vma))) {
3101                                 ret = VM_FAULT_OOM;
3102                                 goto out;
3103                         }
3104                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3105                                                 vma, address);
3106                         if (!page) {
3107                                 ret = VM_FAULT_OOM;
3108                                 goto out;
3109                         }
3110                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3111                                 ret = VM_FAULT_OOM;
3112                                 page_cache_release(page);
3113                                 goto out;
3114                         }
3115                         charged = 1;
3116                         copy_user_highpage(page, vmf.page, address, vma);
3117                         __SetPageUptodate(page);
3118                 } else {
3119                         /*
3120                          * If the page will be shareable, see if the backing
3121                          * address space wants to know that the page is about
3122                          * to become writable
3123                          */
3124                         if (vma->vm_ops->page_mkwrite) {
3125                                 int tmp;
3126
3127                                 unlock_page(page);
3128                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3129                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3130                                 if (unlikely(tmp &
3131                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3132                                         ret = tmp;
3133                                         goto unwritable_page;
3134                                 }
3135                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3136                                         lock_page(page);
3137                                         if (!page->mapping) {
3138                                                 ret = 0; /* retry the fault */
3139                                                 unlock_page(page);
3140                                                 goto unwritable_page;
3141                                         }
3142                                 } else
3143                                         VM_BUG_ON(!PageLocked(page));
3144                                 page_mkwrite = 1;
3145                         }
3146                 }
3147
3148         }
3149
3150         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3151
3152         /*
3153          * This silly early PAGE_DIRTY setting removes a race
3154          * due to the bad i386 page protection. But it's valid
3155          * for other architectures too.
3156          *
3157          * Note that if FAULT_FLAG_WRITE is set, we either now have
3158          * an exclusive copy of the page, or this is a shared mapping,
3159          * so we can make it writable and dirty to avoid having to
3160          * handle that later.
3161          */
3162         /* Only go through if we didn't race with anybody else... */
3163         if (likely(pte_same(*page_table, orig_pte))) {
3164                 flush_icache_page(vma, page);
3165                 entry = mk_pte(page, vma->vm_page_prot);
3166                 if (flags & FAULT_FLAG_WRITE)
3167                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3168                 if (anon) {
3169                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3170                         page_add_new_anon_rmap(page, vma, address);
3171                 } else {
3172                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3173                         page_add_file_rmap(page);
3174                         if (flags & FAULT_FLAG_WRITE) {
3175                                 dirty_page = page;
3176                                 get_page(dirty_page);
3177                         }
3178                 }
3179                 set_pte_at(mm, address, page_table, entry);
3180
3181                 /* no need to invalidate: a not-present page won't be cached */
3182                 update_mmu_cache(vma, address, page_table);
3183         } else {
3184                 if (charged)
3185                         mem_cgroup_uncharge_page(page);
3186                 if (anon)
3187                         page_cache_release(page);
3188                 else
3189                         anon = 1; /* no anon but release faulted_page */
3190         }
3191
3192         pte_unmap_unlock(page_table, ptl);
3193
3194 out:
3195         if (dirty_page) {
3196                 struct address_space *mapping = page->mapping;
3197
3198                 if (set_page_dirty(dirty_page))
3199                         page_mkwrite = 1;
3200                 unlock_page(dirty_page);
3201                 put_page(dirty_page);
3202                 if (page_mkwrite && mapping) {
3203                         /*
3204                          * Some device drivers do not set page.mapping but still
3205                          * dirty their pages
3206                          */
3207                         balance_dirty_pages_ratelimited(mapping);
3208                 }
3209
3210                 /* file_update_time outside page_lock */
3211                 if (vma->vm_file)
3212                         file_update_time(vma->vm_file);
3213         } else {
3214                 unlock_page(vmf.page);
3215                 if (anon)
3216                         page_cache_release(vmf.page);
3217         }
3218
3219         return ret;
3220
3221 unwritable_page:
3222         page_cache_release(page);
3223         return ret;
3224 }
3225
3226 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3227                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3228                 unsigned int flags, pte_t orig_pte)
3229 {
3230         pgoff_t pgoff = (((address & PAGE_MASK)
3231                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3232
3233         pte_unmap(page_table);
3234         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3235 }
3236
3237 /*
3238  * Fault of a previously existing named mapping. Repopulate the pte
3239  * from the encoded file_pte if possible. This enables swappable
3240  * nonlinear vmas.
3241  *
3242  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3243  * but allow concurrent faults), and pte mapped but not yet locked.
3244  * We return with mmap_sem still held, but pte unmapped and unlocked.
3245  */
3246 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3247                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3248                 unsigned int flags, pte_t orig_pte)
3249 {
3250         pgoff_t pgoff;
3251
3252         flags |= FAULT_FLAG_NONLINEAR;
3253
3254         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3255                 return 0;
3256
3257         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3258                 /*
3259                  * Page table corrupted: show pte and kill process.
3260                  */
3261                 print_bad_pte(vma, address, orig_pte, NULL);
3262                 return VM_FAULT_SIGBUS;
3263         }
3264
3265         pgoff = pte_to_pgoff(orig_pte);
3266         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3267 }
3268
3269 /*
3270  * These routines also need to handle stuff like marking pages dirty
3271  * and/or accessed for architectures that don't do it in hardware (most
3272  * RISC architectures).  The early dirtying is also good on the i386.
3273  *
3274  * There is also a hook called "update_mmu_cache()" that architectures
3275  * with external mmu caches can use to update those (ie the Sparc or
3276  * PowerPC hashed page tables that act as extended TLBs).
3277  *
3278  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3279  * but allow concurrent faults), and pte mapped but not yet locked.
3280  * We return with mmap_sem still held, but pte unmapped and unlocked.
3281  */
3282 int handle_pte_fault(struct mm_struct *mm,
3283                      struct vm_area_struct *vma, unsigned long address,
3284                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3285 {
3286         pte_t entry;
3287         spinlock_t *ptl;
3288
3289         entry = *pte;
3290         if (!pte_present(entry)) {
3291                 if (pte_none(entry)) {
3292                         if (vma->vm_ops) {
3293                                 if (likely(vma->vm_ops->fault))
3294                                         return do_linear_fault(mm, vma, address,
3295                                                 pte, pmd, flags, entry);
3296                         }
3297                         return do_anonymous_page(mm, vma, address,
3298                                                  pte, pmd, flags);
3299                 }
3300                 if (pte_file(entry))
3301                         return do_nonlinear_fault(mm, vma, address,
3302                                         pte, pmd, flags, entry);
3303                 return do_swap_page(mm, vma, address,
3304                                         pte, pmd, flags, entry);
3305         }
3306
3307         ptl = pte_lockptr(mm, pmd);
3308         spin_lock(ptl);
3309         if (unlikely(!pte_same(*pte, entry)))
3310                 goto unlock;
3311         if (flags & FAULT_FLAG_WRITE) {
3312                 if (!pte_write(entry))
3313                         return do_wp_page(mm, vma, address,
3314                                         pte, pmd, ptl, entry);
3315                 entry = pte_mkdirty(entry);
3316         }
3317         entry = pte_mkyoung(entry);
3318         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3319                 update_mmu_cache(vma, address, pte);
3320         } else {
3321                 /*
3322                  * This is needed only for protection faults but the arch code
3323                  * is not yet telling us if this is a protection fault or not.
3324                  * This still avoids useless tlb flushes for .text page faults
3325                  * with threads.
3326                  */
3327                 if (flags & FAULT_FLAG_WRITE)
3328                         flush_tlb_fix_spurious_fault(vma, address);
3329         }
3330 unlock:
3331         pte_unmap_unlock(pte, ptl);
3332         return 0;
3333 }
3334
3335 /*
3336  * By the time we get here, we already hold the mm semaphore
3337  */
3338 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3339                 unsigned long address, unsigned int flags)
3340 {
3341         pgd_t *pgd;
3342         pud_t *pud;
3343         pmd_t *pmd;
3344         pte_t *pte;
3345
3346         __set_current_state(TASK_RUNNING);
3347
3348         count_vm_event(PGFAULT);
3349
3350         /* do counter updates before entering really critical section. */
3351         check_sync_rss_stat(current);
3352
3353         if (unlikely(is_vm_hugetlb_page(vma)))
3354                 return hugetlb_fault(mm, vma, address, flags);
3355
3356         pgd = pgd_offset(mm, address);
3357         pud = pud_alloc(mm, pgd, address);
3358         if (!pud)
3359                 return VM_FAULT_OOM;
3360         pmd = pmd_alloc(mm, pud, address);
3361         if (!pmd)
3362                 return VM_FAULT_OOM;
3363         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3364                 if (!vma->vm_ops)
3365                         return do_huge_pmd_anonymous_page(mm, vma, address,
3366                                                           pmd, flags);
3367         } else {
3368                 pmd_t orig_pmd = *pmd;
3369                 barrier();
3370                 if (pmd_trans_huge(orig_pmd)) {
3371                         if (flags & FAULT_FLAG_WRITE &&
3372                             !pmd_write(orig_pmd) &&
3373                             !pmd_trans_splitting(orig_pmd))
3374                                 return do_huge_pmd_wp_page(mm, vma, address,
3375                                                            pmd, orig_pmd);
3376                         return 0;
3377                 }
3378         }
3379
3380         /*
3381          * Use __pte_alloc instead of pte_alloc_map, because we can't
3382          * run pte_offset_map on the pmd, if an huge pmd could
3383          * materialize from under us from a different thread.
3384          */
3385         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
3386                 return VM_FAULT_OOM;
3387         /* if an huge pmd materialized from under us just retry later */
3388         if (unlikely(pmd_trans_huge(*pmd)))
3389                 return 0;
3390         /*
3391          * A regular pmd is established and it can't morph into a huge pmd
3392          * from under us anymore at this point because we hold the mmap_sem
3393          * read mode and khugepaged takes it in write mode. So now it's
3394          * safe to run pte_offset_map().
3395          */
3396         pte = pte_offset_map(pmd, address);
3397
3398         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3399 }
3400
3401 #ifndef __PAGETABLE_PUD_FOLDED
3402 /*
3403  * Allocate page upper directory.
3404  * We've already handled the fast-path in-line.
3405  */
3406 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3407 {
3408         pud_t *new = pud_alloc_one(mm, address);
3409         if (!new)
3410                 return -ENOMEM;
3411
3412         smp_wmb(); /* See comment in __pte_alloc */
3413
3414         spin_lock(&mm->page_table_lock);
3415         if (pgd_present(*pgd))          /* Another has populated it */
3416                 pud_free(mm, new);
3417         else
3418                 pgd_populate(mm, pgd, new);
3419         spin_unlock(&mm->page_table_lock);
3420         return 0;
3421 }
3422 #endif /* __PAGETABLE_PUD_FOLDED */
3423
3424 #ifndef __PAGETABLE_PMD_FOLDED
3425 /*
3426  * Allocate page middle directory.
3427  * We've already handled the fast-path in-line.
3428  */
3429 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3430 {
3431         pmd_t *new = pmd_alloc_one(mm, address);
3432         if (!new)
3433                 return -ENOMEM;
3434
3435         smp_wmb(); /* See comment in __pte_alloc */
3436
3437         spin_lock(&mm->page_table_lock);
3438 #ifndef __ARCH_HAS_4LEVEL_HACK
3439         if (pud_present(*pud))          /* Another has populated it */
3440                 pmd_free(mm, new);
3441         else
3442                 pud_populate(mm, pud, new);
3443 #else
3444         if (pgd_present(*pud))          /* Another has populated it */
3445                 pmd_free(mm, new);
3446         else
3447                 pgd_populate(mm, pud, new);
3448 #endif /* __ARCH_HAS_4LEVEL_HACK */
3449         spin_unlock(&mm->page_table_lock);
3450         return 0;
3451 }
3452 #endif /* __PAGETABLE_PMD_FOLDED */
3453
3454 int make_pages_present(unsigned long addr, unsigned long end)
3455 {
3456         int ret, len, write;
3457         struct vm_area_struct * vma;
3458
3459         vma = find_vma(current->mm, addr);
3460         if (!vma)
3461                 return -ENOMEM;
3462         /*
3463          * We want to touch writable mappings with a write fault in order
3464          * to break COW, except for shared mappings because these don't COW
3465          * and we would not want to dirty them for nothing.
3466          */
3467         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3468         BUG_ON(addr >= end);
3469         BUG_ON(end > vma->vm_end);
3470         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3471         ret = get_user_pages(current, current->mm, addr,
3472                         len, write, 0, NULL, NULL);
3473         if (ret < 0)
3474                 return ret;
3475         return ret == len ? 0 : -EFAULT;
3476 }
3477
3478 #if !defined(__HAVE_ARCH_GATE_AREA)
3479
3480 #if defined(AT_SYSINFO_EHDR)
3481 static struct vm_area_struct gate_vma;
3482
3483 static int __init gate_vma_init(void)
3484 {
3485         gate_vma.vm_mm = NULL;
3486         gate_vma.vm_start = FIXADDR_USER_START;
3487         gate_vma.vm_end = FIXADDR_USER_END;
3488         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3489         gate_vma.vm_page_prot = __P101;
3490         /*
3491          * Make sure the vDSO gets into every core dump.
3492          * Dumping its contents makes post-mortem fully interpretable later
3493          * without matching up the same kernel and hardware config to see
3494          * what PC values meant.
3495          */
3496         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3497         return 0;
3498 }
3499 __initcall(gate_vma_init);
3500 #endif
3501
3502 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3503 {
3504 #ifdef AT_SYSINFO_EHDR
3505         return &gate_vma;
3506 #else
3507         return NULL;
3508 #endif
3509 }
3510
3511 int in_gate_area_no_task(unsigned long addr)
3512 {
3513 #ifdef AT_SYSINFO_EHDR
3514         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3515                 return 1;
3516 #endif
3517         return 0;
3518 }
3519
3520 #endif  /* __HAVE_ARCH_GATE_AREA */
3521
3522 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3523                 pte_t **ptepp, spinlock_t **ptlp)
3524 {
3525         pgd_t *pgd;
3526         pud_t *pud;
3527         pmd_t *pmd;
3528         pte_t *ptep;
3529
3530         pgd = pgd_offset(mm, address);
3531         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3532                 goto out;
3533
3534         pud = pud_offset(pgd, address);
3535         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3536                 goto out;
3537
3538         pmd = pmd_offset(pud, address);
3539         VM_BUG_ON(pmd_trans_huge(*pmd));
3540         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3541                 goto out;
3542
3543         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3544         if (pmd_huge(*pmd))
3545                 goto out;
3546
3547         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3548         if (!ptep)
3549                 goto out;
3550         if (!pte_present(*ptep))
3551                 goto unlock;
3552         *ptepp = ptep;
3553         return 0;
3554 unlock:
3555         pte_unmap_unlock(ptep, *ptlp);
3556 out:
3557         return -EINVAL;
3558 }
3559
3560 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3561                              pte_t **ptepp, spinlock_t **ptlp)
3562 {
3563         int res;
3564
3565         /* (void) is needed to make gcc happy */
3566         (void) __cond_lock(*ptlp,
3567                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3568         return res;
3569 }
3570
3571 /**
3572  * follow_pfn - look up PFN at a user virtual address
3573  * @vma: memory mapping
3574  * @address: user virtual address
3575  * @pfn: location to store found PFN
3576  *
3577  * Only IO mappings and raw PFN mappings are allowed.
3578  *
3579  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3580  */
3581 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3582         unsigned long *pfn)
3583 {
3584         int ret = -EINVAL;
3585         spinlock_t *ptl;
3586         pte_t *ptep;
3587
3588         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3589                 return ret;
3590
3591         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3592         if (ret)
3593                 return ret;
3594         *pfn = pte_pfn(*ptep);
3595         pte_unmap_unlock(ptep, ptl);
3596         return 0;
3597 }
3598 EXPORT_SYMBOL(follow_pfn);
3599
3600 #ifdef CONFIG_HAVE_IOREMAP_PROT
3601 int follow_phys(struct vm_area_struct *vma,
3602                 unsigned long address, unsigned int flags,
3603                 unsigned long *prot, resource_size_t *phys)
3604 {
3605         int ret = -EINVAL;
3606         pte_t *ptep, pte;
3607         spinlock_t *ptl;
3608
3609         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3610                 goto out;
3611
3612         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3613                 goto out;
3614         pte = *ptep;
3615
3616         if ((flags & FOLL_WRITE) && !pte_write(pte))
3617                 goto unlock;
3618
3619         *prot = pgprot_val(pte_pgprot(pte));
3620         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3621
3622         ret = 0;
3623 unlock:
3624         pte_unmap_unlock(ptep, ptl);
3625 out:
3626         return ret;
3627 }
3628
3629 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3630                         void *buf, int len, int write)
3631 {
3632         resource_size_t phys_addr;
3633         unsigned long prot = 0;
3634         void __iomem *maddr;
3635         int offset = addr & (PAGE_SIZE-1);
3636
3637         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3638                 return -EINVAL;
3639
3640         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3641         if (write)
3642                 memcpy_toio(maddr + offset, buf, len);
3643         else
3644                 memcpy_fromio(buf, maddr + offset, len);
3645         iounmap(maddr);
3646
3647         return len;
3648 }
3649 #endif
3650
3651 /*
3652  * Access another process' address space.
3653  * Source/target buffer must be kernel space,
3654  * Do not walk the page table directly, use get_user_pages
3655  */
3656 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3657 {
3658         struct mm_struct *mm;
3659         struct vm_area_struct *vma;
3660         void *old_buf = buf;
3661
3662         mm = get_task_mm(tsk);
3663         if (!mm)
3664                 return 0;
3665
3666         down_read(&mm->mmap_sem);
3667         /* ignore errors, just check how much was successfully transferred */
3668         while (len) {
3669                 int bytes, ret, offset;
3670                 void *maddr;
3671                 struct page *page = NULL;
3672
3673                 ret = get_user_pages(tsk, mm, addr, 1,
3674                                 write, 1, &page, &vma);
3675                 if (ret <= 0) {
3676                         /*
3677                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3678                          * we can access using slightly different code.
3679                          */
3680 #ifdef CONFIG_HAVE_IOREMAP_PROT
3681                         vma = find_vma(mm, addr);
3682                         if (!vma)
3683                                 break;
3684                         if (vma->vm_ops && vma->vm_ops->access)
3685                                 ret = vma->vm_ops->access(vma, addr, buf,
3686                                                           len, write);
3687                         if (ret <= 0)
3688 #endif
3689                                 break;
3690                         bytes = ret;
3691                 } else {
3692                         bytes = len;
3693                         offset = addr & (PAGE_SIZE-1);
3694                         if (bytes > PAGE_SIZE-offset)
3695                                 bytes = PAGE_SIZE-offset;
3696
3697                         maddr = kmap(page);
3698                         if (write) {
3699                                 copy_to_user_page(vma, page, addr,
3700                                                   maddr + offset, buf, bytes);
3701                                 set_page_dirty_lock(page);
3702                         } else {
3703                                 copy_from_user_page(vma, page, addr,
3704                                                     buf, maddr + offset, bytes);
3705                         }
3706                         kunmap(page);
3707                         page_cache_release(page);
3708                 }
3709                 len -= bytes;
3710                 buf += bytes;
3711                 addr += bytes;
3712         }
3713         up_read(&mm->mmap_sem);
3714         mmput(mm);
3715
3716         return buf - old_buf;
3717 }
3718
3719 /*
3720  * Print the name of a VMA.
3721  */
3722 void print_vma_addr(char *prefix, unsigned long ip)
3723 {
3724         struct mm_struct *mm = current->mm;
3725         struct vm_area_struct *vma;
3726
3727         /*
3728          * Do not print if we are in atomic
3729          * contexts (in exception stacks, etc.):
3730          */
3731         if (preempt_count())
3732                 return;
3733
3734         down_read(&mm->mmap_sem);
3735         vma = find_vma(mm, ip);
3736         if (vma && vma->vm_file) {
3737                 struct file *f = vma->vm_file;
3738                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3739                 if (buf) {
3740                         char *p, *s;
3741
3742                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3743                         if (IS_ERR(p))
3744                                 p = "?";
3745                         s = strrchr(p, '/');
3746                         if (s)
3747                                 p = s+1;
3748                         printk("%s%s[%lx+%lx]", prefix, p,
3749                                         vma->vm_start,
3750                                         vma->vm_end - vma->vm_start);
3751                         free_page((unsigned long)buf);
3752                 }
3753         }
3754         up_read(&current->mm->mmap_sem);
3755 }
3756
3757 #ifdef CONFIG_PROVE_LOCKING
3758 void might_fault(void)
3759 {
3760         /*
3761          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3762          * holding the mmap_sem, this is safe because kernel memory doesn't
3763          * get paged out, therefore we'll never actually fault, and the
3764          * below annotations will generate false positives.
3765          */
3766         if (segment_eq(get_fs(), KERNEL_DS))
3767                 return;
3768
3769         might_sleep();
3770         /*
3771          * it would be nicer only to annotate paths which are not under
3772          * pagefault_disable, however that requires a larger audit and
3773          * providing helpers like get_user_atomic.
3774          */
3775         if (!in_atomic() && current->mm)
3776                 might_lock_read(&current->mm->mmap_sem);
3777 }
3778 EXPORT_SYMBOL(might_fault);
3779 #endif
3780
3781 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3782 static void clear_gigantic_page(struct page *page,
3783                                 unsigned long addr,
3784                                 unsigned int pages_per_huge_page)
3785 {
3786         int i;
3787         struct page *p = page;
3788
3789         might_sleep();
3790         for (i = 0; i < pages_per_huge_page;
3791              i++, p = mem_map_next(p, page, i)) {
3792                 cond_resched();
3793                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3794         }
3795 }
3796 void clear_huge_page(struct page *page,
3797                      unsigned long addr, unsigned int pages_per_huge_page)
3798 {
3799         int i;
3800
3801         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3802                 clear_gigantic_page(page, addr, pages_per_huge_page);
3803                 return;
3804         }
3805
3806         might_sleep();
3807         for (i = 0; i < pages_per_huge_page; i++) {
3808                 cond_resched();
3809                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3810         }
3811 }
3812
3813 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3814                                     unsigned long addr,
3815                                     struct vm_area_struct *vma,
3816                                     unsigned int pages_per_huge_page)
3817 {
3818         int i;
3819         struct page *dst_base = dst;
3820         struct page *src_base = src;
3821
3822         for (i = 0; i < pages_per_huge_page; ) {
3823                 cond_resched();
3824                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3825
3826                 i++;
3827                 dst = mem_map_next(dst, dst_base, i);
3828                 src = mem_map_next(src, src_base, i);
3829         }
3830 }
3831
3832 void copy_user_huge_page(struct page *dst, struct page *src,
3833                          unsigned long addr, struct vm_area_struct *vma,
3834                          unsigned int pages_per_huge_page)
3835 {
3836         int i;
3837
3838         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3839                 copy_user_gigantic_page(dst, src, addr, vma,
3840                                         pages_per_huge_page);
3841                 return;
3842         }
3843
3844         might_sleep();
3845         for (i = 0; i < pages_per_huge_page; i++) {
3846                 cond_resched();
3847                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3848         }
3849 }
3850 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */