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