mm: fix ia64 crash when gcore reads gate area
[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;
1398
1399                                 page = vm_normal_page(gate_vma, start, *pte);
1400                                 if (!page) {
1401                                         if (!(gup_flags & FOLL_DUMP) &&
1402                                              is_zero_pfn(pte_pfn(*pte)))
1403                                                 page = pte_page(*pte);
1404                                         else {
1405                                                 pte_unmap(pte);
1406                                                 return i ? : -EFAULT;
1407                                         }
1408                                 }
1409                                 pages[i] = page;
1410                                 get_page(page);
1411                         }
1412                         pte_unmap(pte);
1413                         if (vmas)
1414                                 vmas[i] = gate_vma;
1415                         i++;
1416                         start += PAGE_SIZE;
1417                         nr_pages--;
1418                         continue;
1419                 }
1420
1421                 if (!vma ||
1422                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1423                     !(vm_flags & vma->vm_flags))
1424                         return i ? : -EFAULT;
1425
1426                 if (is_vm_hugetlb_page(vma)) {
1427                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1428                                         &start, &nr_pages, i, gup_flags);
1429                         continue;
1430                 }
1431
1432                 do {
1433                         struct page *page;
1434                         unsigned int foll_flags = gup_flags;
1435
1436                         /*
1437                          * If we have a pending SIGKILL, don't keep faulting
1438                          * pages and potentially allocating memory.
1439                          */
1440                         if (unlikely(fatal_signal_pending(current)))
1441                                 return i ? i : -ERESTARTSYS;
1442
1443                         cond_resched();
1444                         while (!(page = follow_page(vma, start, foll_flags))) {
1445                                 int ret;
1446
1447                                 ret = handle_mm_fault(mm, vma, start,
1448                                         (foll_flags & FOLL_WRITE) ?
1449                                         FAULT_FLAG_WRITE : 0);
1450
1451                                 if (ret & VM_FAULT_ERROR) {
1452                                         if (ret & VM_FAULT_OOM)
1453                                                 return i ? i : -ENOMEM;
1454                                         if (ret &
1455                                             (VM_FAULT_HWPOISON|VM_FAULT_SIGBUS))
1456                                                 return i ? i : -EFAULT;
1457                                         BUG();
1458                                 }
1459                                 if (ret & VM_FAULT_MAJOR)
1460                                         tsk->maj_flt++;
1461                                 else
1462                                         tsk->min_flt++;
1463
1464                                 /*
1465                                  * The VM_FAULT_WRITE bit tells us that
1466                                  * do_wp_page has broken COW when necessary,
1467                                  * even if maybe_mkwrite decided not to set
1468                                  * pte_write. We can thus safely do subsequent
1469                                  * page lookups as if they were reads. But only
1470                                  * do so when looping for pte_write is futile:
1471                                  * in some cases userspace may also be wanting
1472                                  * to write to the gotten user page, which a
1473                                  * read fault here might prevent (a readonly
1474                                  * page might get reCOWed by userspace write).
1475                                  */
1476                                 if ((ret & VM_FAULT_WRITE) &&
1477                                     !(vma->vm_flags & VM_WRITE))
1478                                         foll_flags &= ~FOLL_WRITE;
1479
1480                                 cond_resched();
1481                         }
1482                         if (IS_ERR(page))
1483                                 return i ? i : PTR_ERR(page);
1484                         if (pages) {
1485                                 pages[i] = page;
1486
1487                                 flush_anon_page(vma, page, start);
1488                                 flush_dcache_page(page);
1489                         }
1490                         if (vmas)
1491                                 vmas[i] = vma;
1492                         i++;
1493                         start += PAGE_SIZE;
1494                         nr_pages--;
1495                 } while (nr_pages && start < vma->vm_end);
1496         } while (nr_pages);
1497         return i;
1498 }
1499
1500 /**
1501  * get_user_pages() - pin user pages in memory
1502  * @tsk:        task_struct of target task
1503  * @mm:         mm_struct of target mm
1504  * @start:      starting user address
1505  * @nr_pages:   number of pages from start to pin
1506  * @write:      whether pages will be written to by the caller
1507  * @force:      whether to force write access even if user mapping is
1508  *              readonly. This will result in the page being COWed even
1509  *              in MAP_SHARED mappings. You do not want this.
1510  * @pages:      array that receives pointers to the pages pinned.
1511  *              Should be at least nr_pages long. Or NULL, if caller
1512  *              only intends to ensure the pages are faulted in.
1513  * @vmas:       array of pointers to vmas corresponding to each page.
1514  *              Or NULL if the caller does not require them.
1515  *
1516  * Returns number of pages pinned. This may be fewer than the number
1517  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1518  * were pinned, returns -errno. Each page returned must be released
1519  * with a put_page() call when it is finished with. vmas will only
1520  * remain valid while mmap_sem is held.
1521  *
1522  * Must be called with mmap_sem held for read or write.
1523  *
1524  * get_user_pages walks a process's page tables and takes a reference to
1525  * each struct page that each user address corresponds to at a given
1526  * instant. That is, it takes the page that would be accessed if a user
1527  * thread accesses the given user virtual address at that instant.
1528  *
1529  * This does not guarantee that the page exists in the user mappings when
1530  * get_user_pages returns, and there may even be a completely different
1531  * page there in some cases (eg. if mmapped pagecache has been invalidated
1532  * and subsequently re faulted). However it does guarantee that the page
1533  * won't be freed completely. And mostly callers simply care that the page
1534  * contains data that was valid *at some point in time*. Typically, an IO
1535  * or similar operation cannot guarantee anything stronger anyway because
1536  * locks can't be held over the syscall boundary.
1537  *
1538  * If write=0, the page must not be written to. If the page is written to,
1539  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1540  * after the page is finished with, and before put_page is called.
1541  *
1542  * get_user_pages is typically used for fewer-copy IO operations, to get a
1543  * handle on the memory by some means other than accesses via the user virtual
1544  * addresses. The pages may be submitted for DMA to devices or accessed via
1545  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1546  * use the correct cache flushing APIs.
1547  *
1548  * See also get_user_pages_fast, for performance critical applications.
1549  */
1550 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1551                 unsigned long start, int nr_pages, int write, int force,
1552                 struct page **pages, struct vm_area_struct **vmas)
1553 {
1554         int flags = FOLL_TOUCH;
1555
1556         if (pages)
1557                 flags |= FOLL_GET;
1558         if (write)
1559                 flags |= FOLL_WRITE;
1560         if (force)
1561                 flags |= FOLL_FORCE;
1562
1563         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
1564 }
1565 EXPORT_SYMBOL(get_user_pages);
1566
1567 /**
1568  * get_dump_page() - pin user page in memory while writing it to core dump
1569  * @addr: user address
1570  *
1571  * Returns struct page pointer of user page pinned for dump,
1572  * to be freed afterwards by page_cache_release() or put_page().
1573  *
1574  * Returns NULL on any kind of failure - a hole must then be inserted into
1575  * the corefile, to preserve alignment with its headers; and also returns
1576  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1577  * allowing a hole to be left in the corefile to save diskspace.
1578  *
1579  * Called without mmap_sem, but after all other threads have been killed.
1580  */
1581 #ifdef CONFIG_ELF_CORE
1582 struct page *get_dump_page(unsigned long addr)
1583 {
1584         struct vm_area_struct *vma;
1585         struct page *page;
1586
1587         if (__get_user_pages(current, current->mm, addr, 1,
1588                         FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma) < 1)
1589                 return NULL;
1590         flush_cache_page(vma, addr, page_to_pfn(page));
1591         return page;
1592 }
1593 #endif /* CONFIG_ELF_CORE */
1594
1595 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1596                         spinlock_t **ptl)
1597 {
1598         pgd_t * pgd = pgd_offset(mm, addr);
1599         pud_t * pud = pud_alloc(mm, pgd, addr);
1600         if (pud) {
1601                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1602                 if (pmd)
1603                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1604         }
1605         return NULL;
1606 }
1607
1608 /*
1609  * This is the old fallback for page remapping.
1610  *
1611  * For historical reasons, it only allows reserved pages. Only
1612  * old drivers should use this, and they needed to mark their
1613  * pages reserved for the old functions anyway.
1614  */
1615 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1616                         struct page *page, pgprot_t prot)
1617 {
1618         struct mm_struct *mm = vma->vm_mm;
1619         int retval;
1620         pte_t *pte;
1621         spinlock_t *ptl;
1622
1623         retval = -EINVAL;
1624         if (PageAnon(page))
1625                 goto out;
1626         retval = -ENOMEM;
1627         flush_dcache_page(page);
1628         pte = get_locked_pte(mm, addr, &ptl);
1629         if (!pte)
1630                 goto out;
1631         retval = -EBUSY;
1632         if (!pte_none(*pte))
1633                 goto out_unlock;
1634
1635         /* Ok, finally just insert the thing.. */
1636         get_page(page);
1637         inc_mm_counter_fast(mm, MM_FILEPAGES);
1638         page_add_file_rmap(page);
1639         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1640
1641         retval = 0;
1642         pte_unmap_unlock(pte, ptl);
1643         return retval;
1644 out_unlock:
1645         pte_unmap_unlock(pte, ptl);
1646 out:
1647         return retval;
1648 }
1649
1650 /**
1651  * vm_insert_page - insert single page into user vma
1652  * @vma: user vma to map to
1653  * @addr: target user address of this page
1654  * @page: source kernel page
1655  *
1656  * This allows drivers to insert individual pages they've allocated
1657  * into a user vma.
1658  *
1659  * The page has to be a nice clean _individual_ kernel allocation.
1660  * If you allocate a compound page, you need to have marked it as
1661  * such (__GFP_COMP), or manually just split the page up yourself
1662  * (see split_page()).
1663  *
1664  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1665  * took an arbitrary page protection parameter. This doesn't allow
1666  * that. Your vma protection will have to be set up correctly, which
1667  * means that if you want a shared writable mapping, you'd better
1668  * ask for a shared writable mapping!
1669  *
1670  * The page does not need to be reserved.
1671  */
1672 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1673                         struct page *page)
1674 {
1675         if (addr < vma->vm_start || addr >= vma->vm_end)
1676                 return -EFAULT;
1677         if (!page_count(page))
1678                 return -EINVAL;
1679         vma->vm_flags |= VM_INSERTPAGE;
1680         return insert_page(vma, addr, page, vma->vm_page_prot);
1681 }
1682 EXPORT_SYMBOL(vm_insert_page);
1683
1684 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1685                         unsigned long pfn, pgprot_t prot)
1686 {
1687         struct mm_struct *mm = vma->vm_mm;
1688         int retval;
1689         pte_t *pte, entry;
1690         spinlock_t *ptl;
1691
1692         retval = -ENOMEM;
1693         pte = get_locked_pte(mm, addr, &ptl);
1694         if (!pte)
1695                 goto out;
1696         retval = -EBUSY;
1697         if (!pte_none(*pte))
1698                 goto out_unlock;
1699
1700         /* Ok, finally just insert the thing.. */
1701         entry = pte_mkspecial(pfn_pte(pfn, prot));
1702         set_pte_at(mm, addr, pte, entry);
1703         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1704
1705         retval = 0;
1706 out_unlock:
1707         pte_unmap_unlock(pte, ptl);
1708 out:
1709         return retval;
1710 }
1711
1712 /**
1713  * vm_insert_pfn - insert single pfn into user vma
1714  * @vma: user vma to map to
1715  * @addr: target user address of this page
1716  * @pfn: source kernel pfn
1717  *
1718  * Similar to vm_inert_page, this allows drivers to insert individual pages
1719  * they've allocated into a user vma. Same comments apply.
1720  *
1721  * This function should only be called from a vm_ops->fault handler, and
1722  * in that case the handler should return NULL.
1723  *
1724  * vma cannot be a COW mapping.
1725  *
1726  * As this is called only for pages that do not currently exist, we
1727  * do not need to flush old virtual caches or the TLB.
1728  */
1729 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1730                         unsigned long pfn)
1731 {
1732         int ret;
1733         pgprot_t pgprot = vma->vm_page_prot;
1734         /*
1735          * Technically, architectures with pte_special can avoid all these
1736          * restrictions (same for remap_pfn_range).  However we would like
1737          * consistency in testing and feature parity among all, so we should
1738          * try to keep these invariants in place for everybody.
1739          */
1740         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1741         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1742                                                 (VM_PFNMAP|VM_MIXEDMAP));
1743         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1744         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1745
1746         if (addr < vma->vm_start || addr >= vma->vm_end)
1747                 return -EFAULT;
1748         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1749                 return -EINVAL;
1750
1751         ret = insert_pfn(vma, addr, pfn, pgprot);
1752
1753         if (ret)
1754                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1755
1756         return ret;
1757 }
1758 EXPORT_SYMBOL(vm_insert_pfn);
1759
1760 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1761                         unsigned long pfn)
1762 {
1763         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1764
1765         if (addr < vma->vm_start || addr >= vma->vm_end)
1766                 return -EFAULT;
1767
1768         /*
1769          * If we don't have pte special, then we have to use the pfn_valid()
1770          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1771          * refcount the page if pfn_valid is true (hence insert_page rather
1772          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1773          * without pte special, it would there be refcounted as a normal page.
1774          */
1775         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1776                 struct page *page;
1777
1778                 page = pfn_to_page(pfn);
1779                 return insert_page(vma, addr, page, vma->vm_page_prot);
1780         }
1781         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1782 }
1783 EXPORT_SYMBOL(vm_insert_mixed);
1784
1785 /*
1786  * maps a range of physical memory into the requested pages. the old
1787  * mappings are removed. any references to nonexistent pages results
1788  * in null mappings (currently treated as "copy-on-access")
1789  */
1790 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1791                         unsigned long addr, unsigned long end,
1792                         unsigned long pfn, pgprot_t prot)
1793 {
1794         pte_t *pte;
1795         spinlock_t *ptl;
1796
1797         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1798         if (!pte)
1799                 return -ENOMEM;
1800         arch_enter_lazy_mmu_mode();
1801         do {
1802                 BUG_ON(!pte_none(*pte));
1803                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1804                 pfn++;
1805         } while (pte++, addr += PAGE_SIZE, addr != end);
1806         arch_leave_lazy_mmu_mode();
1807         pte_unmap_unlock(pte - 1, ptl);
1808         return 0;
1809 }
1810
1811 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1812                         unsigned long addr, unsigned long end,
1813                         unsigned long pfn, pgprot_t prot)
1814 {
1815         pmd_t *pmd;
1816         unsigned long next;
1817
1818         pfn -= addr >> PAGE_SHIFT;
1819         pmd = pmd_alloc(mm, pud, addr);
1820         if (!pmd)
1821                 return -ENOMEM;
1822         do {
1823                 next = pmd_addr_end(addr, end);
1824                 if (remap_pte_range(mm, pmd, addr, next,
1825                                 pfn + (addr >> PAGE_SHIFT), prot))
1826                         return -ENOMEM;
1827         } while (pmd++, addr = next, addr != end);
1828         return 0;
1829 }
1830
1831 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1832                         unsigned long addr, unsigned long end,
1833                         unsigned long pfn, pgprot_t prot)
1834 {
1835         pud_t *pud;
1836         unsigned long next;
1837
1838         pfn -= addr >> PAGE_SHIFT;
1839         pud = pud_alloc(mm, pgd, addr);
1840         if (!pud)
1841                 return -ENOMEM;
1842         do {
1843                 next = pud_addr_end(addr, end);
1844                 if (remap_pmd_range(mm, pud, addr, next,
1845                                 pfn + (addr >> PAGE_SHIFT), prot))
1846                         return -ENOMEM;
1847         } while (pud++, addr = next, addr != end);
1848         return 0;
1849 }
1850
1851 /**
1852  * remap_pfn_range - remap kernel memory to userspace
1853  * @vma: user vma to map to
1854  * @addr: target user address to start at
1855  * @pfn: physical address of kernel memory
1856  * @size: size of map area
1857  * @prot: page protection flags for this mapping
1858  *
1859  *  Note: this is only safe if the mm semaphore is held when called.
1860  */
1861 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1862                     unsigned long pfn, unsigned long size, pgprot_t prot)
1863 {
1864         pgd_t *pgd;
1865         unsigned long next;
1866         unsigned long end = addr + PAGE_ALIGN(size);
1867         struct mm_struct *mm = vma->vm_mm;
1868         int err;
1869
1870         /*
1871          * Physically remapped pages are special. Tell the
1872          * rest of the world about it:
1873          *   VM_IO tells people not to look at these pages
1874          *      (accesses can have side effects).
1875          *   VM_RESERVED is specified all over the place, because
1876          *      in 2.4 it kept swapout's vma scan off this vma; but
1877          *      in 2.6 the LRU scan won't even find its pages, so this
1878          *      flag means no more than count its pages in reserved_vm,
1879          *      and omit it from core dump, even when VM_IO turned off.
1880          *   VM_PFNMAP tells the core MM that the base pages are just
1881          *      raw PFN mappings, and do not have a "struct page" associated
1882          *      with them.
1883          *
1884          * There's a horrible special case to handle copy-on-write
1885          * behaviour that some programs depend on. We mark the "original"
1886          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1887          */
1888         if (addr == vma->vm_start && end == vma->vm_end) {
1889                 vma->vm_pgoff = pfn;
1890                 vma->vm_flags |= VM_PFN_AT_MMAP;
1891         } else if (is_cow_mapping(vma->vm_flags))
1892                 return -EINVAL;
1893
1894         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1895
1896         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1897         if (err) {
1898                 /*
1899                  * To indicate that track_pfn related cleanup is not
1900                  * needed from higher level routine calling unmap_vmas
1901                  */
1902                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1903                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1904                 return -EINVAL;
1905         }
1906
1907         BUG_ON(addr >= end);
1908         pfn -= addr >> PAGE_SHIFT;
1909         pgd = pgd_offset(mm, addr);
1910         flush_cache_range(vma, addr, end);
1911         do {
1912                 next = pgd_addr_end(addr, end);
1913                 err = remap_pud_range(mm, pgd, addr, next,
1914                                 pfn + (addr >> PAGE_SHIFT), prot);
1915                 if (err)
1916                         break;
1917         } while (pgd++, addr = next, addr != end);
1918
1919         if (err)
1920                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1921
1922         return err;
1923 }
1924 EXPORT_SYMBOL(remap_pfn_range);
1925
1926 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1927                                      unsigned long addr, unsigned long end,
1928                                      pte_fn_t fn, void *data)
1929 {
1930         pte_t *pte;
1931         int err;
1932         pgtable_t token;
1933         spinlock_t *uninitialized_var(ptl);
1934
1935         pte = (mm == &init_mm) ?
1936                 pte_alloc_kernel(pmd, addr) :
1937                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1938         if (!pte)
1939                 return -ENOMEM;
1940
1941         BUG_ON(pmd_huge(*pmd));
1942
1943         arch_enter_lazy_mmu_mode();
1944
1945         token = pmd_pgtable(*pmd);
1946
1947         do {
1948                 err = fn(pte++, token, addr, data);
1949                 if (err)
1950                         break;
1951         } while (addr += PAGE_SIZE, addr != end);
1952
1953         arch_leave_lazy_mmu_mode();
1954
1955         if (mm != &init_mm)
1956                 pte_unmap_unlock(pte-1, ptl);
1957         return err;
1958 }
1959
1960 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1961                                      unsigned long addr, unsigned long end,
1962                                      pte_fn_t fn, void *data)
1963 {
1964         pmd_t *pmd;
1965         unsigned long next;
1966         int err;
1967
1968         BUG_ON(pud_huge(*pud));
1969
1970         pmd = pmd_alloc(mm, pud, addr);
1971         if (!pmd)
1972                 return -ENOMEM;
1973         do {
1974                 next = pmd_addr_end(addr, end);
1975                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1976                 if (err)
1977                         break;
1978         } while (pmd++, addr = next, addr != end);
1979         return err;
1980 }
1981
1982 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1983                                      unsigned long addr, unsigned long end,
1984                                      pte_fn_t fn, void *data)
1985 {
1986         pud_t *pud;
1987         unsigned long next;
1988         int err;
1989
1990         pud = pud_alloc(mm, pgd, addr);
1991         if (!pud)
1992                 return -ENOMEM;
1993         do {
1994                 next = pud_addr_end(addr, end);
1995                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1996                 if (err)
1997                         break;
1998         } while (pud++, addr = next, addr != end);
1999         return err;
2000 }
2001
2002 /*
2003  * Scan a region of virtual memory, filling in page tables as necessary
2004  * and calling a provided function on each leaf page table.
2005  */
2006 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2007                         unsigned long size, pte_fn_t fn, void *data)
2008 {
2009         pgd_t *pgd;
2010         unsigned long next;
2011         unsigned long start = addr, end = addr + size;
2012         int err;
2013
2014         BUG_ON(addr >= end);
2015         mmu_notifier_invalidate_range_start(mm, start, end);
2016         pgd = pgd_offset(mm, addr);
2017         do {
2018                 next = pgd_addr_end(addr, end);
2019                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2020                 if (err)
2021                         break;
2022         } while (pgd++, addr = next, addr != end);
2023         mmu_notifier_invalidate_range_end(mm, start, end);
2024         return err;
2025 }
2026 EXPORT_SYMBOL_GPL(apply_to_page_range);
2027
2028 /*
2029  * handle_pte_fault chooses page fault handler according to an entry
2030  * which was read non-atomically.  Before making any commitment, on
2031  * those architectures or configurations (e.g. i386 with PAE) which
2032  * might give a mix of unmatched parts, do_swap_page and do_file_page
2033  * must check under lock before unmapping the pte and proceeding
2034  * (but do_wp_page is only called after already making such a check;
2035  * and do_anonymous_page and do_no_page can safely check later on).
2036  */
2037 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2038                                 pte_t *page_table, pte_t orig_pte)
2039 {
2040         int same = 1;
2041 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2042         if (sizeof(pte_t) > sizeof(unsigned long)) {
2043                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2044                 spin_lock(ptl);
2045                 same = pte_same(*page_table, orig_pte);
2046                 spin_unlock(ptl);
2047         }
2048 #endif
2049         pte_unmap(page_table);
2050         return same;
2051 }
2052
2053 /*
2054  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
2055  * servicing faults for write access.  In the normal case, do always want
2056  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
2057  * that do not have writing enabled, when used by access_process_vm.
2058  */
2059 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
2060 {
2061         if (likely(vma->vm_flags & VM_WRITE))
2062                 pte = pte_mkwrite(pte);
2063         return pte;
2064 }
2065
2066 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2067 {
2068         /*
2069          * If the source page was a PFN mapping, we don't have
2070          * a "struct page" for it. We do a best-effort copy by
2071          * just copying from the original user address. If that
2072          * fails, we just zero-fill it. Live with it.
2073          */
2074         if (unlikely(!src)) {
2075                 void *kaddr = kmap_atomic(dst, KM_USER0);
2076                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2077
2078                 /*
2079                  * This really shouldn't fail, because the page is there
2080                  * in the page tables. But it might just be unreadable,
2081                  * in which case we just give up and fill the result with
2082                  * zeroes.
2083                  */
2084                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2085                         memset(kaddr, 0, PAGE_SIZE);
2086                 kunmap_atomic(kaddr, KM_USER0);
2087                 flush_dcache_page(dst);
2088         } else
2089                 copy_user_highpage(dst, src, va, vma);
2090 }
2091
2092 /*
2093  * This routine handles present pages, when users try to write
2094  * to a shared page. It is done by copying the page to a new address
2095  * and decrementing the shared-page counter for the old page.
2096  *
2097  * Note that this routine assumes that the protection checks have been
2098  * done by the caller (the low-level page fault routine in most cases).
2099  * Thus we can safely just mark it writable once we've done any necessary
2100  * COW.
2101  *
2102  * We also mark the page dirty at this point even though the page will
2103  * change only once the write actually happens. This avoids a few races,
2104  * and potentially makes it more efficient.
2105  *
2106  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2107  * but allow concurrent faults), with pte both mapped and locked.
2108  * We return with mmap_sem still held, but pte unmapped and unlocked.
2109  */
2110 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2111                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2112                 spinlock_t *ptl, pte_t orig_pte)
2113 {
2114         struct page *old_page, *new_page;
2115         pte_t entry;
2116         int reuse = 0, ret = 0;
2117         int page_mkwrite = 0;
2118         struct page *dirty_page = NULL;
2119
2120         old_page = vm_normal_page(vma, address, orig_pte);
2121         if (!old_page) {
2122                 /*
2123                  * VM_MIXEDMAP !pfn_valid() case
2124                  *
2125                  * We should not cow pages in a shared writeable mapping.
2126                  * Just mark the pages writable as we can't do any dirty
2127                  * accounting on raw pfn maps.
2128                  */
2129                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2130                                      (VM_WRITE|VM_SHARED))
2131                         goto reuse;
2132                 goto gotten;
2133         }
2134
2135         /*
2136          * Take out anonymous pages first, anonymous shared vmas are
2137          * not dirty accountable.
2138          */
2139         if (PageAnon(old_page) && !PageKsm(old_page)) {
2140                 if (!trylock_page(old_page)) {
2141                         page_cache_get(old_page);
2142                         pte_unmap_unlock(page_table, ptl);
2143                         lock_page(old_page);
2144                         page_table = pte_offset_map_lock(mm, pmd, address,
2145                                                          &ptl);
2146                         if (!pte_same(*page_table, orig_pte)) {
2147                                 unlock_page(old_page);
2148                                 page_cache_release(old_page);
2149                                 goto unlock;
2150                         }
2151                         page_cache_release(old_page);
2152                 }
2153                 reuse = reuse_swap_page(old_page);
2154                 if (reuse)
2155                         /*
2156                          * The page is all ours.  Move it to our anon_vma so
2157                          * the rmap code will not search our parent or siblings.
2158                          * Protected against the rmap code by the page lock.
2159                          */
2160                         page_move_anon_rmap(old_page, vma, address);
2161                 unlock_page(old_page);
2162         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2163                                         (VM_WRITE|VM_SHARED))) {
2164                 /*
2165                  * Only catch write-faults on shared writable pages,
2166                  * read-only shared pages can get COWed by
2167                  * get_user_pages(.write=1, .force=1).
2168                  */
2169                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2170                         struct vm_fault vmf;
2171                         int tmp;
2172
2173                         vmf.virtual_address = (void __user *)(address &
2174                                                                 PAGE_MASK);
2175                         vmf.pgoff = old_page->index;
2176                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2177                         vmf.page = old_page;
2178
2179                         /*
2180                          * Notify the address space that the page is about to
2181                          * become writable so that it can prohibit this or wait
2182                          * for the page to get into an appropriate state.
2183                          *
2184                          * We do this without the lock held, so that it can
2185                          * sleep if it needs to.
2186                          */
2187                         page_cache_get(old_page);
2188                         pte_unmap_unlock(page_table, ptl);
2189
2190                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2191                         if (unlikely(tmp &
2192                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2193                                 ret = tmp;
2194                                 goto unwritable_page;
2195                         }
2196                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2197                                 lock_page(old_page);
2198                                 if (!old_page->mapping) {
2199                                         ret = 0; /* retry the fault */
2200                                         unlock_page(old_page);
2201                                         goto unwritable_page;
2202                                 }
2203                         } else
2204                                 VM_BUG_ON(!PageLocked(old_page));
2205
2206                         /*
2207                          * Since we dropped the lock we need to revalidate
2208                          * the PTE as someone else may have changed it.  If
2209                          * they did, we just return, as we can count on the
2210                          * MMU to tell us if they didn't also make it writable.
2211                          */
2212                         page_table = pte_offset_map_lock(mm, pmd, address,
2213                                                          &ptl);
2214                         if (!pte_same(*page_table, orig_pte)) {
2215                                 unlock_page(old_page);
2216                                 page_cache_release(old_page);
2217                                 goto unlock;
2218                         }
2219
2220                         page_mkwrite = 1;
2221                 }
2222                 dirty_page = old_page;
2223                 get_page(dirty_page);
2224                 reuse = 1;
2225         }
2226
2227         if (reuse) {
2228 reuse:
2229                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2230                 entry = pte_mkyoung(orig_pte);
2231                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2232                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2233                         update_mmu_cache(vma, address, page_table);
2234                 ret |= VM_FAULT_WRITE;
2235                 goto unlock;
2236         }
2237
2238         /*
2239          * Ok, we need to copy. Oh, well..
2240          */
2241         page_cache_get(old_page);
2242 gotten:
2243         pte_unmap_unlock(page_table, ptl);
2244
2245         if (unlikely(anon_vma_prepare(vma)))
2246                 goto oom;
2247
2248         if (is_zero_pfn(pte_pfn(orig_pte))) {
2249                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2250                 if (!new_page)
2251                         goto oom;
2252         } else {
2253                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2254                 if (!new_page)
2255                         goto oom;
2256                 cow_user_page(new_page, old_page, address, vma);
2257         }
2258         __SetPageUptodate(new_page);
2259
2260         /*
2261          * Don't let another task, with possibly unlocked vma,
2262          * keep the mlocked page.
2263          */
2264         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2265                 lock_page(old_page);    /* for LRU manipulation */
2266                 clear_page_mlock(old_page);
2267                 unlock_page(old_page);
2268         }
2269
2270         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2271                 goto oom_free_new;
2272
2273         /*
2274          * Re-check the pte - we dropped the lock
2275          */
2276         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2277         if (likely(pte_same(*page_table, orig_pte))) {
2278                 if (old_page) {
2279                         if (!PageAnon(old_page)) {
2280                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2281                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2282                         }
2283                 } else
2284                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2285                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2286                 entry = mk_pte(new_page, vma->vm_page_prot);
2287                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2288                 /*
2289                  * Clear the pte entry and flush it first, before updating the
2290                  * pte with the new entry. This will avoid a race condition
2291                  * seen in the presence of one thread doing SMC and another
2292                  * thread doing COW.
2293                  */
2294                 ptep_clear_flush(vma, address, page_table);
2295                 page_add_new_anon_rmap(new_page, vma, address);
2296                 /*
2297                  * We call the notify macro here because, when using secondary
2298                  * mmu page tables (such as kvm shadow page tables), we want the
2299                  * new page to be mapped directly into the secondary page table.
2300                  */
2301                 set_pte_at_notify(mm, address, page_table, entry);
2302                 update_mmu_cache(vma, address, page_table);
2303                 if (old_page) {
2304                         /*
2305                          * Only after switching the pte to the new page may
2306                          * we remove the mapcount here. Otherwise another
2307                          * process may come and find the rmap count decremented
2308                          * before the pte is switched to the new page, and
2309                          * "reuse" the old page writing into it while our pte
2310                          * here still points into it and can be read by other
2311                          * threads.
2312                          *
2313                          * The critical issue is to order this
2314                          * page_remove_rmap with the ptp_clear_flush above.
2315                          * Those stores are ordered by (if nothing else,)
2316                          * the barrier present in the atomic_add_negative
2317                          * in page_remove_rmap.
2318                          *
2319                          * Then the TLB flush in ptep_clear_flush ensures that
2320                          * no process can access the old page before the
2321                          * decremented mapcount is visible. And the old page
2322                          * cannot be reused until after the decremented
2323                          * mapcount is visible. So transitively, TLBs to
2324                          * old page will be flushed before it can be reused.
2325                          */
2326                         page_remove_rmap(old_page);
2327                 }
2328
2329                 /* Free the old page.. */
2330                 new_page = old_page;
2331                 ret |= VM_FAULT_WRITE;
2332         } else
2333                 mem_cgroup_uncharge_page(new_page);
2334
2335         if (new_page)
2336                 page_cache_release(new_page);
2337         if (old_page)
2338                 page_cache_release(old_page);
2339 unlock:
2340         pte_unmap_unlock(page_table, ptl);
2341         if (dirty_page) {
2342                 /*
2343                  * Yes, Virginia, this is actually required to prevent a race
2344                  * with clear_page_dirty_for_io() from clearing the page dirty
2345                  * bit after it clear all dirty ptes, but before a racing
2346                  * do_wp_page installs a dirty pte.
2347                  *
2348                  * do_no_page is protected similarly.
2349                  */
2350                 if (!page_mkwrite) {
2351                         wait_on_page_locked(dirty_page);
2352                         set_page_dirty_balance(dirty_page, page_mkwrite);
2353                 }
2354                 put_page(dirty_page);
2355                 if (page_mkwrite) {
2356                         struct address_space *mapping = dirty_page->mapping;
2357
2358                         set_page_dirty(dirty_page);
2359                         unlock_page(dirty_page);
2360                         page_cache_release(dirty_page);
2361                         if (mapping)    {
2362                                 /*
2363                                  * Some device drivers do not set page.mapping
2364                                  * but still dirty their pages
2365                                  */
2366                                 balance_dirty_pages_ratelimited(mapping);
2367                         }
2368                 }
2369
2370                 /* file_update_time outside page_lock */
2371                 if (vma->vm_file)
2372                         file_update_time(vma->vm_file);
2373         }
2374         return ret;
2375 oom_free_new:
2376         page_cache_release(new_page);
2377 oom:
2378         if (old_page) {
2379                 if (page_mkwrite) {
2380                         unlock_page(old_page);
2381                         page_cache_release(old_page);
2382                 }
2383                 page_cache_release(old_page);
2384         }
2385         return VM_FAULT_OOM;
2386
2387 unwritable_page:
2388         page_cache_release(old_page);
2389         return ret;
2390 }
2391
2392 /*
2393  * Helper functions for unmap_mapping_range().
2394  *
2395  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2396  *
2397  * We have to restart searching the prio_tree whenever we drop the lock,
2398  * since the iterator is only valid while the lock is held, and anyway
2399  * a later vma might be split and reinserted earlier while lock dropped.
2400  *
2401  * The list of nonlinear vmas could be handled more efficiently, using
2402  * a placeholder, but handle it in the same way until a need is shown.
2403  * It is important to search the prio_tree before nonlinear list: a vma
2404  * may become nonlinear and be shifted from prio_tree to nonlinear list
2405  * while the lock is dropped; but never shifted from list to prio_tree.
2406  *
2407  * In order to make forward progress despite restarting the search,
2408  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2409  * quickly skip it next time around.  Since the prio_tree search only
2410  * shows us those vmas affected by unmapping the range in question, we
2411  * can't efficiently keep all vmas in step with mapping->truncate_count:
2412  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2413  * mapping->truncate_count and vma->vm_truncate_count are protected by
2414  * i_mmap_lock.
2415  *
2416  * In order to make forward progress despite repeatedly restarting some
2417  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2418  * and restart from that address when we reach that vma again.  It might
2419  * have been split or merged, shrunk or extended, but never shifted: so
2420  * restart_addr remains valid so long as it remains in the vma's range.
2421  * unmap_mapping_range forces truncate_count to leap over page-aligned
2422  * values so we can save vma's restart_addr in its truncate_count field.
2423  */
2424 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2425
2426 static void reset_vma_truncate_counts(struct address_space *mapping)
2427 {
2428         struct vm_area_struct *vma;
2429         struct prio_tree_iter iter;
2430
2431         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2432                 vma->vm_truncate_count = 0;
2433         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2434                 vma->vm_truncate_count = 0;
2435 }
2436
2437 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2438                 unsigned long start_addr, unsigned long end_addr,
2439                 struct zap_details *details)
2440 {
2441         unsigned long restart_addr;
2442         int need_break;
2443
2444         /*
2445          * files that support invalidating or truncating portions of the
2446          * file from under mmaped areas must have their ->fault function
2447          * return a locked page (and set VM_FAULT_LOCKED in the return).
2448          * This provides synchronisation against concurrent unmapping here.
2449          */
2450
2451 again:
2452         restart_addr = vma->vm_truncate_count;
2453         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2454                 start_addr = restart_addr;
2455                 if (start_addr >= end_addr) {
2456                         /* Top of vma has been split off since last time */
2457                         vma->vm_truncate_count = details->truncate_count;
2458                         return 0;
2459                 }
2460         }
2461
2462         restart_addr = zap_page_range(vma, start_addr,
2463                                         end_addr - start_addr, details);
2464         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2465
2466         if (restart_addr >= end_addr) {
2467                 /* We have now completed this vma: mark it so */
2468                 vma->vm_truncate_count = details->truncate_count;
2469                 if (!need_break)
2470                         return 0;
2471         } else {
2472                 /* Note restart_addr in vma's truncate_count field */
2473                 vma->vm_truncate_count = restart_addr;
2474                 if (!need_break)
2475                         goto again;
2476         }
2477
2478         spin_unlock(details->i_mmap_lock);
2479         cond_resched();
2480         spin_lock(details->i_mmap_lock);
2481         return -EINTR;
2482 }
2483
2484 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2485                                             struct zap_details *details)
2486 {
2487         struct vm_area_struct *vma;
2488         struct prio_tree_iter iter;
2489         pgoff_t vba, vea, zba, zea;
2490
2491 restart:
2492         vma_prio_tree_foreach(vma, &iter, root,
2493                         details->first_index, details->last_index) {
2494                 /* Skip quickly over those we have already dealt with */
2495                 if (vma->vm_truncate_count == details->truncate_count)
2496                         continue;
2497
2498                 vba = vma->vm_pgoff;
2499                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2500                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2501                 zba = details->first_index;
2502                 if (zba < vba)
2503                         zba = vba;
2504                 zea = details->last_index;
2505                 if (zea > vea)
2506                         zea = vea;
2507
2508                 if (unmap_mapping_range_vma(vma,
2509                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2510                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2511                                 details) < 0)
2512                         goto restart;
2513         }
2514 }
2515
2516 static inline void unmap_mapping_range_list(struct list_head *head,
2517                                             struct zap_details *details)
2518 {
2519         struct vm_area_struct *vma;
2520
2521         /*
2522          * In nonlinear VMAs there is no correspondence between virtual address
2523          * offset and file offset.  So we must perform an exhaustive search
2524          * across *all* the pages in each nonlinear VMA, not just the pages
2525          * whose virtual address lies outside the file truncation point.
2526          */
2527 restart:
2528         list_for_each_entry(vma, head, shared.vm_set.list) {
2529                 /* Skip quickly over those we have already dealt with */
2530                 if (vma->vm_truncate_count == details->truncate_count)
2531                         continue;
2532                 details->nonlinear_vma = vma;
2533                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2534                                         vma->vm_end, details) < 0)
2535                         goto restart;
2536         }
2537 }
2538
2539 /**
2540  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2541  * @mapping: the address space containing mmaps to be unmapped.
2542  * @holebegin: byte in first page to unmap, relative to the start of
2543  * the underlying file.  This will be rounded down to a PAGE_SIZE
2544  * boundary.  Note that this is different from truncate_pagecache(), which
2545  * must keep the partial page.  In contrast, we must get rid of
2546  * partial pages.
2547  * @holelen: size of prospective hole in bytes.  This will be rounded
2548  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2549  * end of the file.
2550  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2551  * but 0 when invalidating pagecache, don't throw away private data.
2552  */
2553 void unmap_mapping_range(struct address_space *mapping,
2554                 loff_t const holebegin, loff_t const holelen, int even_cows)
2555 {
2556         struct zap_details details;
2557         pgoff_t hba = holebegin >> PAGE_SHIFT;
2558         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2559
2560         /* Check for overflow. */
2561         if (sizeof(holelen) > sizeof(hlen)) {
2562                 long long holeend =
2563                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2564                 if (holeend & ~(long long)ULONG_MAX)
2565                         hlen = ULONG_MAX - hba + 1;
2566         }
2567
2568         details.check_mapping = even_cows? NULL: mapping;
2569         details.nonlinear_vma = NULL;
2570         details.first_index = hba;
2571         details.last_index = hba + hlen - 1;
2572         if (details.last_index < details.first_index)
2573                 details.last_index = ULONG_MAX;
2574         details.i_mmap_lock = &mapping->i_mmap_lock;
2575
2576         spin_lock(&mapping->i_mmap_lock);
2577
2578         /* Protect against endless unmapping loops */
2579         mapping->truncate_count++;
2580         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2581                 if (mapping->truncate_count == 0)
2582                         reset_vma_truncate_counts(mapping);
2583                 mapping->truncate_count++;
2584         }
2585         details.truncate_count = mapping->truncate_count;
2586
2587         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2588                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2589         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2590                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2591         spin_unlock(&mapping->i_mmap_lock);
2592 }
2593 EXPORT_SYMBOL(unmap_mapping_range);
2594
2595 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2596 {
2597         struct address_space *mapping = inode->i_mapping;
2598
2599         /*
2600          * If the underlying filesystem is not going to provide
2601          * a way to truncate a range of blocks (punch a hole) -
2602          * we should return failure right now.
2603          */
2604         if (!inode->i_op->truncate_range)
2605                 return -ENOSYS;
2606
2607         mutex_lock(&inode->i_mutex);
2608         down_write(&inode->i_alloc_sem);
2609         unmap_mapping_range(mapping, offset, (end - offset), 1);
2610         truncate_inode_pages_range(mapping, offset, end);
2611         unmap_mapping_range(mapping, offset, (end - offset), 1);
2612         inode->i_op->truncate_range(inode, offset, end);
2613         up_write(&inode->i_alloc_sem);
2614         mutex_unlock(&inode->i_mutex);
2615
2616         return 0;
2617 }
2618
2619 /*
2620  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2621  * but allow concurrent faults), and pte mapped but not yet locked.
2622  * We return with mmap_sem still held, but pte unmapped and unlocked.
2623  */
2624 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2625                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2626                 unsigned int flags, pte_t orig_pte)
2627 {
2628         spinlock_t *ptl;
2629         struct page *page;
2630         swp_entry_t entry;
2631         pte_t pte;
2632         struct mem_cgroup *ptr = NULL;
2633         int ret = 0;
2634
2635         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2636                 goto out;
2637
2638         entry = pte_to_swp_entry(orig_pte);
2639         if (unlikely(non_swap_entry(entry))) {
2640                 if (is_migration_entry(entry)) {
2641                         migration_entry_wait(mm, pmd, address);
2642                 } else if (is_hwpoison_entry(entry)) {
2643                         ret = VM_FAULT_HWPOISON;
2644                 } else {
2645                         print_bad_pte(vma, address, orig_pte, NULL);
2646                         ret = VM_FAULT_SIGBUS;
2647                 }
2648                 goto out;
2649         }
2650         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2651         page = lookup_swap_cache(entry);
2652         if (!page) {
2653                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2654                 page = swapin_readahead(entry,
2655                                         GFP_HIGHUSER_MOVABLE, vma, address);
2656                 if (!page) {
2657                         /*
2658                          * Back out if somebody else faulted in this pte
2659                          * while we released the pte lock.
2660                          */
2661                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2662                         if (likely(pte_same(*page_table, orig_pte)))
2663                                 ret = VM_FAULT_OOM;
2664                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2665                         goto unlock;
2666                 }
2667
2668                 /* Had to read the page from swap area: Major fault */
2669                 ret = VM_FAULT_MAJOR;
2670                 count_vm_event(PGMAJFAULT);
2671         } else if (PageHWPoison(page)) {
2672                 /*
2673                  * hwpoisoned dirty swapcache pages are kept for killing
2674                  * owner processes (which may be unknown at hwpoison time)
2675                  */
2676                 ret = VM_FAULT_HWPOISON;
2677                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2678                 goto out_release;
2679         }
2680
2681         lock_page(page);
2682         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2683
2684         page = ksm_might_need_to_copy(page, vma, address);
2685         if (!page) {
2686                 ret = VM_FAULT_OOM;
2687                 goto out;
2688         }
2689
2690         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2691                 ret = VM_FAULT_OOM;
2692                 goto out_page;
2693         }
2694
2695         /*
2696          * Back out if somebody else already faulted in this pte.
2697          */
2698         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2699         if (unlikely(!pte_same(*page_table, orig_pte)))
2700                 goto out_nomap;
2701
2702         if (unlikely(!PageUptodate(page))) {
2703                 ret = VM_FAULT_SIGBUS;
2704                 goto out_nomap;
2705         }
2706
2707         /*
2708          * The page isn't present yet, go ahead with the fault.
2709          *
2710          * Be careful about the sequence of operations here.
2711          * To get its accounting right, reuse_swap_page() must be called
2712          * while the page is counted on swap but not yet in mapcount i.e.
2713          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2714          * must be called after the swap_free(), or it will never succeed.
2715          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2716          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2717          * in page->private. In this case, a record in swap_cgroup  is silently
2718          * discarded at swap_free().
2719          */
2720
2721         inc_mm_counter_fast(mm, MM_ANONPAGES);
2722         dec_mm_counter_fast(mm, MM_SWAPENTS);
2723         pte = mk_pte(page, vma->vm_page_prot);
2724         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2725                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2726                 flags &= ~FAULT_FLAG_WRITE;
2727         }
2728         flush_icache_page(vma, page);
2729         set_pte_at(mm, address, page_table, pte);
2730         page_add_anon_rmap(page, vma, address);
2731         /* It's better to call commit-charge after rmap is established */
2732         mem_cgroup_commit_charge_swapin(page, ptr);
2733
2734         swap_free(entry);
2735         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2736                 try_to_free_swap(page);
2737         unlock_page(page);
2738
2739         if (flags & FAULT_FLAG_WRITE) {
2740                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2741                 if (ret & VM_FAULT_ERROR)
2742                         ret &= VM_FAULT_ERROR;
2743                 goto out;
2744         }
2745
2746         /* No need to invalidate - it was non-present before */
2747         update_mmu_cache(vma, address, page_table);
2748 unlock:
2749         pte_unmap_unlock(page_table, ptl);
2750 out:
2751         return ret;
2752 out_nomap:
2753         mem_cgroup_cancel_charge_swapin(ptr);
2754         pte_unmap_unlock(page_table, ptl);
2755 out_page:
2756         unlock_page(page);
2757 out_release:
2758         page_cache_release(page);
2759         return ret;
2760 }
2761
2762 /*
2763  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2764  * but allow concurrent faults), and pte mapped but not yet locked.
2765  * We return with mmap_sem still held, but pte unmapped and unlocked.
2766  */
2767 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2768                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2769                 unsigned int flags)
2770 {
2771         struct page *page;
2772         spinlock_t *ptl;
2773         pte_t entry;
2774
2775         if (!(flags & FAULT_FLAG_WRITE)) {
2776                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2777                                                 vma->vm_page_prot));
2778                 ptl = pte_lockptr(mm, pmd);
2779                 spin_lock(ptl);
2780                 if (!pte_none(*page_table))
2781                         goto unlock;
2782                 goto setpte;
2783         }
2784
2785         /* Allocate our own private page. */
2786         pte_unmap(page_table);
2787
2788         if (unlikely(anon_vma_prepare(vma)))
2789                 goto oom;
2790         page = alloc_zeroed_user_highpage_movable(vma, address);
2791         if (!page)
2792                 goto oom;
2793         __SetPageUptodate(page);
2794
2795         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2796                 goto oom_free_page;
2797
2798         entry = mk_pte(page, vma->vm_page_prot);
2799         if (vma->vm_flags & VM_WRITE)
2800                 entry = pte_mkwrite(pte_mkdirty(entry));
2801
2802         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2803         if (!pte_none(*page_table))
2804                 goto release;
2805
2806         inc_mm_counter_fast(mm, MM_ANONPAGES);
2807         page_add_new_anon_rmap(page, vma, address);
2808 setpte:
2809         set_pte_at(mm, address, page_table, entry);
2810
2811         /* No need to invalidate - it was non-present before */
2812         update_mmu_cache(vma, address, page_table);
2813 unlock:
2814         pte_unmap_unlock(page_table, ptl);
2815         return 0;
2816 release:
2817         mem_cgroup_uncharge_page(page);
2818         page_cache_release(page);
2819         goto unlock;
2820 oom_free_page:
2821         page_cache_release(page);
2822 oom:
2823         return VM_FAULT_OOM;
2824 }
2825
2826 /*
2827  * __do_fault() tries to create a new page mapping. It aggressively
2828  * tries to share with existing pages, but makes a separate copy if
2829  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2830  * the next page fault.
2831  *
2832  * As this is called only for pages that do not currently exist, we
2833  * do not need to flush old virtual caches or the TLB.
2834  *
2835  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2836  * but allow concurrent faults), and pte neither mapped nor locked.
2837  * We return with mmap_sem still held, but pte unmapped and unlocked.
2838  */
2839 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2840                 unsigned long address, pmd_t *pmd,
2841                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2842 {
2843         pte_t *page_table;
2844         spinlock_t *ptl;
2845         struct page *page;
2846         pte_t entry;
2847         int anon = 0;
2848         int charged = 0;
2849         struct page *dirty_page = NULL;
2850         struct vm_fault vmf;
2851         int ret;
2852         int page_mkwrite = 0;
2853
2854         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2855         vmf.pgoff = pgoff;
2856         vmf.flags = flags;
2857         vmf.page = NULL;
2858
2859         ret = vma->vm_ops->fault(vma, &vmf);
2860         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2861                 return ret;
2862
2863         if (unlikely(PageHWPoison(vmf.page))) {
2864                 if (ret & VM_FAULT_LOCKED)
2865                         unlock_page(vmf.page);
2866                 return VM_FAULT_HWPOISON;
2867         }
2868
2869         /*
2870          * For consistency in subsequent calls, make the faulted page always
2871          * locked.
2872          */
2873         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2874                 lock_page(vmf.page);
2875         else
2876                 VM_BUG_ON(!PageLocked(vmf.page));
2877
2878         /*
2879          * Should we do an early C-O-W break?
2880          */
2881         page = vmf.page;
2882         if (flags & FAULT_FLAG_WRITE) {
2883                 if (!(vma->vm_flags & VM_SHARED)) {
2884                         anon = 1;
2885                         if (unlikely(anon_vma_prepare(vma))) {
2886                                 ret = VM_FAULT_OOM;
2887                                 goto out;
2888                         }
2889                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2890                                                 vma, address);
2891                         if (!page) {
2892                                 ret = VM_FAULT_OOM;
2893                                 goto out;
2894                         }
2895                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2896                                 ret = VM_FAULT_OOM;
2897                                 page_cache_release(page);
2898                                 goto out;
2899                         }
2900                         charged = 1;
2901                         /*
2902                          * Don't let another task, with possibly unlocked vma,
2903                          * keep the mlocked page.
2904                          */
2905                         if (vma->vm_flags & VM_LOCKED)
2906                                 clear_page_mlock(vmf.page);
2907                         copy_user_highpage(page, vmf.page, address, vma);
2908                         __SetPageUptodate(page);
2909                 } else {
2910                         /*
2911                          * If the page will be shareable, see if the backing
2912                          * address space wants to know that the page is about
2913                          * to become writable
2914                          */
2915                         if (vma->vm_ops->page_mkwrite) {
2916                                 int tmp;
2917
2918                                 unlock_page(page);
2919                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2920                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2921                                 if (unlikely(tmp &
2922                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2923                                         ret = tmp;
2924                                         goto unwritable_page;
2925                                 }
2926                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2927                                         lock_page(page);
2928                                         if (!page->mapping) {
2929                                                 ret = 0; /* retry the fault */
2930                                                 unlock_page(page);
2931                                                 goto unwritable_page;
2932                                         }
2933                                 } else
2934                                         VM_BUG_ON(!PageLocked(page));
2935                                 page_mkwrite = 1;
2936                         }
2937                 }
2938
2939         }
2940
2941         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2942
2943         /*
2944          * This silly early PAGE_DIRTY setting removes a race
2945          * due to the bad i386 page protection. But it's valid
2946          * for other architectures too.
2947          *
2948          * Note that if FAULT_FLAG_WRITE is set, we either now have
2949          * an exclusive copy of the page, or this is a shared mapping,
2950          * so we can make it writable and dirty to avoid having to
2951          * handle that later.
2952          */
2953         /* Only go through if we didn't race with anybody else... */
2954         if (likely(pte_same(*page_table, orig_pte))) {
2955                 flush_icache_page(vma, page);
2956                 entry = mk_pte(page, vma->vm_page_prot);
2957                 if (flags & FAULT_FLAG_WRITE)
2958                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2959                 if (anon) {
2960                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2961                         page_add_new_anon_rmap(page, vma, address);
2962                 } else {
2963                         inc_mm_counter_fast(mm, MM_FILEPAGES);
2964                         page_add_file_rmap(page);
2965                         if (flags & FAULT_FLAG_WRITE) {
2966                                 dirty_page = page;
2967                                 get_page(dirty_page);
2968                         }
2969                 }
2970                 set_pte_at(mm, address, page_table, entry);
2971
2972                 /* no need to invalidate: a not-present page won't be cached */
2973                 update_mmu_cache(vma, address, page_table);
2974         } else {
2975                 if (charged)
2976                         mem_cgroup_uncharge_page(page);
2977                 if (anon)
2978                         page_cache_release(page);
2979                 else
2980                         anon = 1; /* no anon but release faulted_page */
2981         }
2982
2983         pte_unmap_unlock(page_table, ptl);
2984
2985 out:
2986         if (dirty_page) {
2987                 struct address_space *mapping = page->mapping;
2988
2989                 if (set_page_dirty(dirty_page))
2990                         page_mkwrite = 1;
2991                 unlock_page(dirty_page);
2992                 put_page(dirty_page);
2993                 if (page_mkwrite && mapping) {
2994                         /*
2995                          * Some device drivers do not set page.mapping but still
2996                          * dirty their pages
2997                          */
2998                         balance_dirty_pages_ratelimited(mapping);
2999                 }
3000
3001                 /* file_update_time outside page_lock */
3002                 if (vma->vm_file)
3003                         file_update_time(vma->vm_file);
3004         } else {
3005                 unlock_page(vmf.page);
3006                 if (anon)
3007                         page_cache_release(vmf.page);
3008         }
3009
3010         return ret;
3011
3012 unwritable_page:
3013         page_cache_release(page);
3014         return ret;
3015 }
3016
3017 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3018                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3019                 unsigned int flags, pte_t orig_pte)
3020 {
3021         pgoff_t pgoff = (((address & PAGE_MASK)
3022                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3023
3024         pte_unmap(page_table);
3025         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3026 }
3027
3028 /*
3029  * Fault of a previously existing named mapping. Repopulate the pte
3030  * from the encoded file_pte if possible. This enables swappable
3031  * nonlinear vmas.
3032  *
3033  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3034  * but allow concurrent faults), and pte mapped but not yet locked.
3035  * We return with mmap_sem still held, but pte unmapped and unlocked.
3036  */
3037 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3038                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3039                 unsigned int flags, pte_t orig_pte)
3040 {
3041         pgoff_t pgoff;
3042
3043         flags |= FAULT_FLAG_NONLINEAR;
3044
3045         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3046                 return 0;
3047
3048         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3049                 /*
3050                  * Page table corrupted: show pte and kill process.
3051                  */
3052                 print_bad_pte(vma, address, orig_pte, NULL);
3053                 return VM_FAULT_SIGBUS;
3054         }
3055
3056         pgoff = pte_to_pgoff(orig_pte);
3057         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3058 }
3059
3060 /*
3061  * These routines also need to handle stuff like marking pages dirty
3062  * and/or accessed for architectures that don't do it in hardware (most
3063  * RISC architectures).  The early dirtying is also good on the i386.
3064  *
3065  * There is also a hook called "update_mmu_cache()" that architectures
3066  * with external mmu caches can use to update those (ie the Sparc or
3067  * PowerPC hashed page tables that act as extended TLBs).
3068  *
3069  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3070  * but allow concurrent faults), and pte mapped but not yet locked.
3071  * We return with mmap_sem still held, but pte unmapped and unlocked.
3072  */
3073 static inline int handle_pte_fault(struct mm_struct *mm,
3074                 struct vm_area_struct *vma, unsigned long address,
3075                 pte_t *pte, pmd_t *pmd, unsigned int flags)
3076 {
3077         pte_t entry;
3078         spinlock_t *ptl;
3079
3080         entry = *pte;
3081         if (!pte_present(entry)) {
3082                 if (pte_none(entry)) {
3083                         if (vma->vm_ops) {
3084                                 if (likely(vma->vm_ops->fault))
3085                                         return do_linear_fault(mm, vma, address,
3086                                                 pte, pmd, flags, entry);
3087                         }
3088                         return do_anonymous_page(mm, vma, address,
3089                                                  pte, pmd, flags);
3090                 }
3091                 if (pte_file(entry))
3092                         return do_nonlinear_fault(mm, vma, address,
3093                                         pte, pmd, flags, entry);
3094                 return do_swap_page(mm, vma, address,
3095                                         pte, pmd, flags, entry);
3096         }
3097
3098         ptl = pte_lockptr(mm, pmd);
3099         spin_lock(ptl);
3100         if (unlikely(!pte_same(*pte, entry)))
3101                 goto unlock;
3102         if (flags & FAULT_FLAG_WRITE) {
3103                 if (!pte_write(entry))
3104                         return do_wp_page(mm, vma, address,
3105                                         pte, pmd, ptl, entry);
3106                 entry = pte_mkdirty(entry);
3107         }
3108         entry = pte_mkyoung(entry);
3109         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3110                 update_mmu_cache(vma, address, pte);
3111         } else {
3112                 /*
3113                  * This is needed only for protection faults but the arch code
3114                  * is not yet telling us if this is a protection fault or not.
3115                  * This still avoids useless tlb flushes for .text page faults
3116                  * with threads.
3117                  */
3118                 if (flags & FAULT_FLAG_WRITE)
3119                         flush_tlb_page(vma, address);
3120         }
3121 unlock:
3122         pte_unmap_unlock(pte, ptl);
3123         return 0;
3124 }
3125
3126 /*
3127  * By the time we get here, we already hold the mm semaphore
3128  */
3129 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3130                 unsigned long address, unsigned int flags)
3131 {
3132         pgd_t *pgd;
3133         pud_t *pud;
3134         pmd_t *pmd;
3135         pte_t *pte;
3136
3137         __set_current_state(TASK_RUNNING);
3138
3139         count_vm_event(PGFAULT);
3140
3141         /* do counter updates before entering really critical section. */
3142         check_sync_rss_stat(current);
3143
3144         if (unlikely(is_vm_hugetlb_page(vma)))
3145                 return hugetlb_fault(mm, vma, address, flags);
3146
3147         pgd = pgd_offset(mm, address);
3148         pud = pud_alloc(mm, pgd, address);
3149         if (!pud)
3150                 return VM_FAULT_OOM;
3151         pmd = pmd_alloc(mm, pud, address);
3152         if (!pmd)
3153                 return VM_FAULT_OOM;
3154         pte = pte_alloc_map(mm, pmd, address);
3155         if (!pte)
3156                 return VM_FAULT_OOM;
3157
3158         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3159 }
3160
3161 #ifndef __PAGETABLE_PUD_FOLDED
3162 /*
3163  * Allocate page upper directory.
3164  * We've already handled the fast-path in-line.
3165  */
3166 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3167 {
3168         pud_t *new = pud_alloc_one(mm, address);
3169         if (!new)
3170                 return -ENOMEM;
3171
3172         smp_wmb(); /* See comment in __pte_alloc */
3173
3174         spin_lock(&mm->page_table_lock);
3175         if (pgd_present(*pgd))          /* Another has populated it */
3176                 pud_free(mm, new);
3177         else
3178                 pgd_populate(mm, pgd, new);
3179         spin_unlock(&mm->page_table_lock);
3180         return 0;
3181 }
3182 #endif /* __PAGETABLE_PUD_FOLDED */
3183
3184 #ifndef __PAGETABLE_PMD_FOLDED
3185 /*
3186  * Allocate page middle directory.
3187  * We've already handled the fast-path in-line.
3188  */
3189 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3190 {
3191         pmd_t *new = pmd_alloc_one(mm, address);
3192         if (!new)
3193                 return -ENOMEM;
3194
3195         smp_wmb(); /* See comment in __pte_alloc */
3196
3197         spin_lock(&mm->page_table_lock);
3198 #ifndef __ARCH_HAS_4LEVEL_HACK
3199         if (pud_present(*pud))          /* Another has populated it */
3200                 pmd_free(mm, new);
3201         else
3202                 pud_populate(mm, pud, new);
3203 #else
3204         if (pgd_present(*pud))          /* Another has populated it */
3205                 pmd_free(mm, new);
3206         else
3207                 pgd_populate(mm, pud, new);
3208 #endif /* __ARCH_HAS_4LEVEL_HACK */
3209         spin_unlock(&mm->page_table_lock);
3210         return 0;
3211 }
3212 #endif /* __PAGETABLE_PMD_FOLDED */
3213
3214 int make_pages_present(unsigned long addr, unsigned long end)
3215 {
3216         int ret, len, write;
3217         struct vm_area_struct * vma;
3218
3219         vma = find_vma(current->mm, addr);
3220         if (!vma)
3221                 return -ENOMEM;
3222         write = (vma->vm_flags & VM_WRITE) != 0;
3223         BUG_ON(addr >= end);
3224         BUG_ON(end > vma->vm_end);
3225         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3226         ret = get_user_pages(current, current->mm, addr,
3227                         len, write, 0, NULL, NULL);
3228         if (ret < 0)
3229                 return ret;
3230         return ret == len ? 0 : -EFAULT;
3231 }
3232
3233 #if !defined(__HAVE_ARCH_GATE_AREA)
3234
3235 #if defined(AT_SYSINFO_EHDR)
3236 static struct vm_area_struct gate_vma;
3237
3238 static int __init gate_vma_init(void)
3239 {
3240         gate_vma.vm_mm = NULL;
3241         gate_vma.vm_start = FIXADDR_USER_START;
3242         gate_vma.vm_end = FIXADDR_USER_END;
3243         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3244         gate_vma.vm_page_prot = __P101;
3245         /*
3246          * Make sure the vDSO gets into every core dump.
3247          * Dumping its contents makes post-mortem fully interpretable later
3248          * without matching up the same kernel and hardware config to see
3249          * what PC values meant.
3250          */
3251         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3252         return 0;
3253 }
3254 __initcall(gate_vma_init);
3255 #endif
3256
3257 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3258 {
3259 #ifdef AT_SYSINFO_EHDR
3260         return &gate_vma;
3261 #else
3262         return NULL;
3263 #endif
3264 }
3265
3266 int in_gate_area_no_task(unsigned long addr)
3267 {
3268 #ifdef AT_SYSINFO_EHDR
3269         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3270                 return 1;
3271 #endif
3272         return 0;
3273 }
3274
3275 #endif  /* __HAVE_ARCH_GATE_AREA */
3276
3277 static int follow_pte(struct mm_struct *mm, unsigned long address,
3278                 pte_t **ptepp, spinlock_t **ptlp)
3279 {
3280         pgd_t *pgd;
3281         pud_t *pud;
3282         pmd_t *pmd;
3283         pte_t *ptep;
3284
3285         pgd = pgd_offset(mm, address);
3286         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3287                 goto out;
3288
3289         pud = pud_offset(pgd, address);
3290         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3291                 goto out;
3292
3293         pmd = pmd_offset(pud, address);
3294         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3295                 goto out;
3296
3297         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3298         if (pmd_huge(*pmd))
3299                 goto out;
3300
3301         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3302         if (!ptep)
3303                 goto out;
3304         if (!pte_present(*ptep))
3305                 goto unlock;
3306         *ptepp = ptep;
3307         return 0;
3308 unlock:
3309         pte_unmap_unlock(ptep, *ptlp);
3310 out:
3311         return -EINVAL;
3312 }
3313
3314 /**
3315  * follow_pfn - look up PFN at a user virtual address
3316  * @vma: memory mapping
3317  * @address: user virtual address
3318  * @pfn: location to store found PFN
3319  *
3320  * Only IO mappings and raw PFN mappings are allowed.
3321  *
3322  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3323  */
3324 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3325         unsigned long *pfn)
3326 {
3327         int ret = -EINVAL;
3328         spinlock_t *ptl;
3329         pte_t *ptep;
3330
3331         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3332                 return ret;
3333
3334         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3335         if (ret)
3336                 return ret;
3337         *pfn = pte_pfn(*ptep);
3338         pte_unmap_unlock(ptep, ptl);
3339         return 0;
3340 }
3341 EXPORT_SYMBOL(follow_pfn);
3342
3343 #ifdef CONFIG_HAVE_IOREMAP_PROT
3344 int follow_phys(struct vm_area_struct *vma,
3345                 unsigned long address, unsigned int flags,
3346                 unsigned long *prot, resource_size_t *phys)
3347 {
3348         int ret = -EINVAL;
3349         pte_t *ptep, pte;
3350         spinlock_t *ptl;
3351
3352         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3353                 goto out;
3354
3355         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3356                 goto out;
3357         pte = *ptep;
3358
3359         if ((flags & FOLL_WRITE) && !pte_write(pte))
3360                 goto unlock;
3361
3362         *prot = pgprot_val(pte_pgprot(pte));
3363         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3364
3365         ret = 0;
3366 unlock:
3367         pte_unmap_unlock(ptep, ptl);
3368 out:
3369         return ret;
3370 }
3371
3372 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3373                         void *buf, int len, int write)
3374 {
3375         resource_size_t phys_addr;
3376         unsigned long prot = 0;
3377         void __iomem *maddr;
3378         int offset = addr & (PAGE_SIZE-1);
3379
3380         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3381                 return -EINVAL;
3382
3383         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3384         if (write)
3385                 memcpy_toio(maddr + offset, buf, len);
3386         else
3387                 memcpy_fromio(buf, maddr + offset, len);
3388         iounmap(maddr);
3389
3390         return len;
3391 }
3392 #endif
3393
3394 /*
3395  * Access another process' address space.
3396  * Source/target buffer must be kernel space,
3397  * Do not walk the page table directly, use get_user_pages
3398  */
3399 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3400 {
3401         struct mm_struct *mm;
3402         struct vm_area_struct *vma;
3403         void *old_buf = buf;
3404
3405         mm = get_task_mm(tsk);
3406         if (!mm)
3407                 return 0;
3408
3409         down_read(&mm->mmap_sem);
3410         /* ignore errors, just check how much was successfully transferred */
3411         while (len) {
3412                 int bytes, ret, offset;
3413                 void *maddr;
3414                 struct page *page = NULL;
3415
3416                 ret = get_user_pages(tsk, mm, addr, 1,
3417                                 write, 1, &page, &vma);
3418                 if (ret <= 0) {
3419                         /*
3420                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3421                          * we can access using slightly different code.
3422                          */
3423 #ifdef CONFIG_HAVE_IOREMAP_PROT
3424                         vma = find_vma(mm, addr);
3425                         if (!vma)
3426                                 break;
3427                         if (vma->vm_ops && vma->vm_ops->access)
3428                                 ret = vma->vm_ops->access(vma, addr, buf,
3429                                                           len, write);
3430                         if (ret <= 0)
3431 #endif
3432                                 break;
3433                         bytes = ret;
3434                 } else {
3435                         bytes = len;
3436                         offset = addr & (PAGE_SIZE-1);
3437                         if (bytes > PAGE_SIZE-offset)
3438                                 bytes = PAGE_SIZE-offset;
3439
3440                         maddr = kmap(page);
3441                         if (write) {
3442                                 copy_to_user_page(vma, page, addr,
3443                                                   maddr + offset, buf, bytes);
3444                                 set_page_dirty_lock(page);
3445                         } else {
3446                                 copy_from_user_page(vma, page, addr,
3447                                                     buf, maddr + offset, bytes);
3448                         }
3449                         kunmap(page);
3450                         page_cache_release(page);
3451                 }
3452                 len -= bytes;
3453                 buf += bytes;
3454                 addr += bytes;
3455         }
3456         up_read(&mm->mmap_sem);
3457         mmput(mm);
3458
3459         return buf - old_buf;
3460 }
3461
3462 /*
3463  * Print the name of a VMA.
3464  */
3465 void print_vma_addr(char *prefix, unsigned long ip)
3466 {
3467         struct mm_struct *mm = current->mm;
3468         struct vm_area_struct *vma;
3469
3470         /*
3471          * Do not print if we are in atomic
3472          * contexts (in exception stacks, etc.):
3473          */
3474         if (preempt_count())
3475                 return;
3476
3477         down_read(&mm->mmap_sem);
3478         vma = find_vma(mm, ip);
3479         if (vma && vma->vm_file) {
3480                 struct file *f = vma->vm_file;
3481                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3482                 if (buf) {
3483                         char *p, *s;
3484
3485                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3486                         if (IS_ERR(p))
3487                                 p = "?";
3488                         s = strrchr(p, '/');
3489                         if (s)
3490                                 p = s+1;
3491                         printk("%s%s[%lx+%lx]", prefix, p,
3492                                         vma->vm_start,
3493                                         vma->vm_end - vma->vm_start);
3494                         free_page((unsigned long)buf);
3495                 }
3496         }
3497         up_read(&current->mm->mmap_sem);
3498 }
3499
3500 #ifdef CONFIG_PROVE_LOCKING
3501 void might_fault(void)
3502 {
3503         /*
3504          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3505          * holding the mmap_sem, this is safe because kernel memory doesn't
3506          * get paged out, therefore we'll never actually fault, and the
3507          * below annotations will generate false positives.
3508          */
3509         if (segment_eq(get_fs(), KERNEL_DS))
3510                 return;
3511
3512         might_sleep();
3513         /*
3514          * it would be nicer only to annotate paths which are not under
3515          * pagefault_disable, however that requires a larger audit and
3516          * providing helpers like get_user_atomic.
3517          */
3518         if (!in_atomic() && current->mm)
3519                 might_lock_read(&current->mm->mmap_sem);
3520 }
3521 EXPORT_SYMBOL(might_fault);
3522 #endif