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