]> git.karo-electronics.de Git - mv-sheeva.git/blob - mm/memory.c
a4597614f18d8034cd274b91a7be91ef5d134580
[mv-sheeva.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                 anon_vma_unlink(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                                 anon_vma_unlink(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 (is_write_migration_entry(entry) &&
683                                         is_cow_mapping(vm_flags)) {
684                                 /*
685                                  * COW mappings require pages in both parent
686                                  * and child to be set to read.
687                                  */
688                                 make_migration_entry_read(&entry);
689                                 pte = swp_entry_to_pte(entry);
690                                 set_pte_at(src_mm, addr, src_pte, pte);
691                         }
692                 }
693                 goto out_set_pte;
694         }
695
696         /*
697          * If it's a COW mapping, write protect it both
698          * in the parent and the child
699          */
700         if (is_cow_mapping(vm_flags)) {
701                 ptep_set_wrprotect(src_mm, addr, src_pte);
702                 pte = pte_wrprotect(pte);
703         }
704
705         /*
706          * If it's a shared mapping, mark it clean in
707          * the child
708          */
709         if (vm_flags & VM_SHARED)
710                 pte = pte_mkclean(pte);
711         pte = pte_mkold(pte);
712
713         page = vm_normal_page(vma, addr, pte);
714         if (page) {
715                 get_page(page);
716                 page_dup_rmap(page);
717                 if (PageAnon(page))
718                         rss[MM_ANONPAGES]++;
719                 else
720                         rss[MM_FILEPAGES]++;
721         }
722
723 out_set_pte:
724         set_pte_at(dst_mm, addr, dst_pte, pte);
725         return 0;
726 }
727
728 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
729                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
730                 unsigned long addr, unsigned long end)
731 {
732         pte_t *orig_src_pte, *orig_dst_pte;
733         pte_t *src_pte, *dst_pte;
734         spinlock_t *src_ptl, *dst_ptl;
735         int progress = 0;
736         int rss[NR_MM_COUNTERS];
737         swp_entry_t entry = (swp_entry_t){0};
738
739 again:
740         init_rss_vec(rss);
741
742         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
743         if (!dst_pte)
744                 return -ENOMEM;
745         src_pte = pte_offset_map_nested(src_pmd, addr);
746         src_ptl = pte_lockptr(src_mm, src_pmd);
747         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
748         orig_src_pte = src_pte;
749         orig_dst_pte = dst_pte;
750         arch_enter_lazy_mmu_mode();
751
752         do {
753                 /*
754                  * We are holding two locks at this point - either of them
755                  * could generate latencies in another task on another CPU.
756                  */
757                 if (progress >= 32) {
758                         progress = 0;
759                         if (need_resched() ||
760                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
761                                 break;
762                 }
763                 if (pte_none(*src_pte)) {
764                         progress++;
765                         continue;
766                 }
767                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
768                                                         vma, addr, rss);
769                 if (entry.val)
770                         break;
771                 progress += 8;
772         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
773
774         arch_leave_lazy_mmu_mode();
775         spin_unlock(src_ptl);
776         pte_unmap_nested(orig_src_pte);
777         add_mm_rss_vec(dst_mm, rss);
778         pte_unmap_unlock(orig_dst_pte, dst_ptl);
779         cond_resched();
780
781         if (entry.val) {
782                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
783                         return -ENOMEM;
784                 progress = 0;
785         }
786         if (addr != end)
787                 goto again;
788         return 0;
789 }
790
791 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
792                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
793                 unsigned long addr, unsigned long end)
794 {
795         pmd_t *src_pmd, *dst_pmd;
796         unsigned long next;
797
798         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
799         if (!dst_pmd)
800                 return -ENOMEM;
801         src_pmd = pmd_offset(src_pud, addr);
802         do {
803                 next = pmd_addr_end(addr, end);
804                 if (pmd_none_or_clear_bad(src_pmd))
805                         continue;
806                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
807                                                 vma, addr, next))
808                         return -ENOMEM;
809         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
810         return 0;
811 }
812
813 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
814                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
815                 unsigned long addr, unsigned long end)
816 {
817         pud_t *src_pud, *dst_pud;
818         unsigned long next;
819
820         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
821         if (!dst_pud)
822                 return -ENOMEM;
823         src_pud = pud_offset(src_pgd, addr);
824         do {
825                 next = pud_addr_end(addr, end);
826                 if (pud_none_or_clear_bad(src_pud))
827                         continue;
828                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
829                                                 vma, addr, next))
830                         return -ENOMEM;
831         } while (dst_pud++, src_pud++, addr = next, addr != end);
832         return 0;
833 }
834
835 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
836                 struct vm_area_struct *vma)
837 {
838         pgd_t *src_pgd, *dst_pgd;
839         unsigned long next;
840         unsigned long addr = vma->vm_start;
841         unsigned long end = vma->vm_end;
842         int ret;
843
844         /*
845          * Don't copy ptes where a page fault will fill them correctly.
846          * Fork becomes much lighter when there are big shared or private
847          * readonly mappings. The tradeoff is that copy_page_range is more
848          * efficient than faulting.
849          */
850         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
851                 if (!vma->anon_vma)
852                         return 0;
853         }
854
855         if (is_vm_hugetlb_page(vma))
856                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
857
858         if (unlikely(is_pfn_mapping(vma))) {
859                 /*
860                  * We do not free on error cases below as remove_vma
861                  * gets called on error from higher level routine
862                  */
863                 ret = track_pfn_vma_copy(vma);
864                 if (ret)
865                         return ret;
866         }
867
868         /*
869          * We need to invalidate the secondary MMU mappings only when
870          * there could be a permission downgrade on the ptes of the
871          * parent mm. And a permission downgrade will only happen if
872          * is_cow_mapping() returns true.
873          */
874         if (is_cow_mapping(vma->vm_flags))
875                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
876
877         ret = 0;
878         dst_pgd = pgd_offset(dst_mm, addr);
879         src_pgd = pgd_offset(src_mm, addr);
880         do {
881                 next = pgd_addr_end(addr, end);
882                 if (pgd_none_or_clear_bad(src_pgd))
883                         continue;
884                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
885                                             vma, addr, next))) {
886                         ret = -ENOMEM;
887                         break;
888                 }
889         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
890
891         if (is_cow_mapping(vma->vm_flags))
892                 mmu_notifier_invalidate_range_end(src_mm,
893                                                   vma->vm_start, end);
894         return ret;
895 }
896
897 static unsigned long zap_pte_range(struct mmu_gather *tlb,
898                                 struct vm_area_struct *vma, pmd_t *pmd,
899                                 unsigned long addr, unsigned long end,
900                                 long *zap_work, struct zap_details *details)
901 {
902         struct mm_struct *mm = tlb->mm;
903         pte_t *pte;
904         spinlock_t *ptl;
905         int rss[NR_MM_COUNTERS];
906
907         init_rss_vec(rss);
908
909         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
910         arch_enter_lazy_mmu_mode();
911         do {
912                 pte_t ptent = *pte;
913                 if (pte_none(ptent)) {
914                         (*zap_work)--;
915                         continue;
916                 }
917
918                 (*zap_work) -= PAGE_SIZE;
919
920                 if (pte_present(ptent)) {
921                         struct page *page;
922
923                         page = vm_normal_page(vma, addr, ptent);
924                         if (unlikely(details) && page) {
925                                 /*
926                                  * unmap_shared_mapping_pages() wants to
927                                  * invalidate cache without truncating:
928                                  * unmap shared but keep private pages.
929                                  */
930                                 if (details->check_mapping &&
931                                     details->check_mapping != page->mapping)
932                                         continue;
933                                 /*
934                                  * Each page->index must be checked when
935                                  * invalidating or truncating nonlinear.
936                                  */
937                                 if (details->nonlinear_vma &&
938                                     (page->index < details->first_index ||
939                                      page->index > details->last_index))
940                                         continue;
941                         }
942                         ptent = ptep_get_and_clear_full(mm, addr, pte,
943                                                         tlb->fullmm);
944                         tlb_remove_tlb_entry(tlb, pte, addr);
945                         if (unlikely(!page))
946                                 continue;
947                         if (unlikely(details) && details->nonlinear_vma
948                             && linear_page_index(details->nonlinear_vma,
949                                                 addr) != page->index)
950                                 set_pte_at(mm, addr, pte,
951                                            pgoff_to_pte(page->index));
952                         if (PageAnon(page))
953                                 rss[MM_ANONPAGES]--;
954                         else {
955                                 if (pte_dirty(ptent))
956                                         set_page_dirty(page);
957                                 if (pte_young(ptent) &&
958                                     likely(!VM_SequentialReadHint(vma)))
959                                         mark_page_accessed(page);
960                                 rss[MM_FILEPAGES]--;
961                         }
962                         page_remove_rmap(page);
963                         if (unlikely(page_mapcount(page) < 0))
964                                 print_bad_pte(vma, addr, ptent, page);
965                         tlb_remove_page(tlb, page);
966                         continue;
967                 }
968                 /*
969                  * If details->check_mapping, we leave swap entries;
970                  * if details->nonlinear_vma, we leave file entries.
971                  */
972                 if (unlikely(details))
973                         continue;
974                 if (pte_file(ptent)) {
975                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
976                                 print_bad_pte(vma, addr, ptent, NULL);
977                 } else if
978                   (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
979                         print_bad_pte(vma, addr, ptent, NULL);
980                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
981         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
982
983         add_mm_rss_vec(mm, rss);
984         arch_leave_lazy_mmu_mode();
985         pte_unmap_unlock(pte - 1, ptl);
986
987         return addr;
988 }
989
990 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
991                                 struct vm_area_struct *vma, pud_t *pud,
992                                 unsigned long addr, unsigned long end,
993                                 long *zap_work, struct zap_details *details)
994 {
995         pmd_t *pmd;
996         unsigned long next;
997
998         pmd = pmd_offset(pud, addr);
999         do {
1000                 next = pmd_addr_end(addr, end);
1001                 if (pmd_none_or_clear_bad(pmd)) {
1002                         (*zap_work)--;
1003                         continue;
1004                 }
1005                 next = zap_pte_range(tlb, vma, pmd, addr, next,
1006                                                 zap_work, details);
1007         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
1008
1009         return addr;
1010 }
1011
1012 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1013                                 struct vm_area_struct *vma, pgd_t *pgd,
1014                                 unsigned long addr, unsigned long end,
1015                                 long *zap_work, struct zap_details *details)
1016 {
1017         pud_t *pud;
1018         unsigned long next;
1019
1020         pud = pud_offset(pgd, addr);
1021         do {
1022                 next = pud_addr_end(addr, end);
1023                 if (pud_none_or_clear_bad(pud)) {
1024                         (*zap_work)--;
1025                         continue;
1026                 }
1027                 next = zap_pmd_range(tlb, vma, pud, addr, next,
1028                                                 zap_work, details);
1029         } while (pud++, addr = next, (addr != end && *zap_work > 0));
1030
1031         return addr;
1032 }
1033
1034 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1035                                 struct vm_area_struct *vma,
1036                                 unsigned long addr, unsigned long end,
1037                                 long *zap_work, struct zap_details *details)
1038 {
1039         pgd_t *pgd;
1040         unsigned long next;
1041
1042         if (details && !details->check_mapping && !details->nonlinear_vma)
1043                 details = NULL;
1044
1045         BUG_ON(addr >= end);
1046         mem_cgroup_uncharge_start();
1047         tlb_start_vma(tlb, vma);
1048         pgd = pgd_offset(vma->vm_mm, addr);
1049         do {
1050                 next = pgd_addr_end(addr, end);
1051                 if (pgd_none_or_clear_bad(pgd)) {
1052                         (*zap_work)--;
1053                         continue;
1054                 }
1055                 next = zap_pud_range(tlb, vma, pgd, addr, next,
1056                                                 zap_work, details);
1057         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
1058         tlb_end_vma(tlb, vma);
1059         mem_cgroup_uncharge_end();
1060
1061         return addr;
1062 }
1063
1064 #ifdef CONFIG_PREEMPT
1065 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1066 #else
1067 /* No preempt: go for improved straight-line efficiency */
1068 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1069 #endif
1070
1071 /**
1072  * unmap_vmas - unmap a range of memory covered by a list of vma's
1073  * @tlbp: address of the caller's struct mmu_gather
1074  * @vma: the starting vma
1075  * @start_addr: virtual address at which to start unmapping
1076  * @end_addr: virtual address at which to end unmapping
1077  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1078  * @details: details of nonlinear truncation or shared cache invalidation
1079  *
1080  * Returns the end address of the unmapping (restart addr if interrupted).
1081  *
1082  * Unmap all pages in the vma list.
1083  *
1084  * We aim to not hold locks for too long (for scheduling latency reasons).
1085  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
1086  * return the ending mmu_gather to the caller.
1087  *
1088  * Only addresses between `start' and `end' will be unmapped.
1089  *
1090  * The VMA list must be sorted in ascending virtual address order.
1091  *
1092  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1093  * range after unmap_vmas() returns.  So the only responsibility here is to
1094  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1095  * drops the lock and schedules.
1096  */
1097 unsigned long unmap_vmas(struct mmu_gather **tlbp,
1098                 struct vm_area_struct *vma, unsigned long start_addr,
1099                 unsigned long end_addr, unsigned long *nr_accounted,
1100                 struct zap_details *details)
1101 {
1102         long zap_work = ZAP_BLOCK_SIZE;
1103         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
1104         int tlb_start_valid = 0;
1105         unsigned long start = start_addr;
1106         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1107         int fullmm = (*tlbp)->fullmm;
1108         struct mm_struct *mm = vma->vm_mm;
1109
1110         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1111         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1112                 unsigned long end;
1113
1114                 start = max(vma->vm_start, start_addr);
1115                 if (start >= vma->vm_end)
1116                         continue;
1117                 end = min(vma->vm_end, end_addr);
1118                 if (end <= vma->vm_start)
1119                         continue;
1120
1121                 if (vma->vm_flags & VM_ACCOUNT)
1122                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1123
1124                 if (unlikely(is_pfn_mapping(vma)))
1125                         untrack_pfn_vma(vma, 0, 0);
1126
1127                 while (start != end) {
1128                         if (!tlb_start_valid) {
1129                                 tlb_start = start;
1130                                 tlb_start_valid = 1;
1131                         }
1132
1133                         if (unlikely(is_vm_hugetlb_page(vma))) {
1134                                 /*
1135                                  * It is undesirable to test vma->vm_file as it
1136                                  * should be non-null for valid hugetlb area.
1137                                  * However, vm_file will be NULL in the error
1138                                  * cleanup path of do_mmap_pgoff. When
1139                                  * hugetlbfs ->mmap method fails,
1140                                  * do_mmap_pgoff() nullifies vma->vm_file
1141                                  * before calling this function to clean up.
1142                                  * Since no pte has actually been setup, it is
1143                                  * safe to do nothing in this case.
1144                                  */
1145                                 if (vma->vm_file) {
1146                                         unmap_hugepage_range(vma, start, end, NULL);
1147                                         zap_work -= (end - start) /
1148                                         pages_per_huge_page(hstate_vma(vma));
1149                                 }
1150
1151                                 start = end;
1152                         } else
1153                                 start = unmap_page_range(*tlbp, vma,
1154                                                 start, end, &zap_work, details);
1155
1156                         if (zap_work > 0) {
1157                                 BUG_ON(start != end);
1158                                 break;
1159                         }
1160
1161                         tlb_finish_mmu(*tlbp, tlb_start, start);
1162
1163                         if (need_resched() ||
1164                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1165                                 if (i_mmap_lock) {
1166                                         *tlbp = NULL;
1167                                         goto out;
1168                                 }
1169                                 cond_resched();
1170                         }
1171
1172                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1173                         tlb_start_valid = 0;
1174                         zap_work = ZAP_BLOCK_SIZE;
1175                 }
1176         }
1177 out:
1178         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1179         return start;   /* which is now the end (or restart) address */
1180 }
1181
1182 /**
1183  * zap_page_range - remove user pages in a given range
1184  * @vma: vm_area_struct holding the applicable pages
1185  * @address: starting address of pages to zap
1186  * @size: number of bytes to zap
1187  * @details: details of nonlinear truncation or shared cache invalidation
1188  */
1189 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1190                 unsigned long size, struct zap_details *details)
1191 {
1192         struct mm_struct *mm = vma->vm_mm;
1193         struct mmu_gather *tlb;
1194         unsigned long end = address + size;
1195         unsigned long nr_accounted = 0;
1196
1197         lru_add_drain();
1198         tlb = tlb_gather_mmu(mm, 0);
1199         update_hiwater_rss(mm);
1200         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1201         if (tlb)
1202                 tlb_finish_mmu(tlb, address, end);
1203         return end;
1204 }
1205
1206 /**
1207  * zap_vma_ptes - remove ptes mapping the vma
1208  * @vma: vm_area_struct holding ptes to be zapped
1209  * @address: starting address of pages to zap
1210  * @size: number of bytes to zap
1211  *
1212  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1213  *
1214  * The entire address range must be fully contained within the vma.
1215  *
1216  * Returns 0 if successful.
1217  */
1218 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1219                 unsigned long size)
1220 {
1221         if (address < vma->vm_start || address + size > vma->vm_end ||
1222                         !(vma->vm_flags & VM_PFNMAP))
1223                 return -1;
1224         zap_page_range(vma, address, size, NULL);
1225         return 0;
1226 }
1227 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1228
1229 /*
1230  * Do a quick page-table lookup for a single page.
1231  */
1232 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1233                         unsigned int flags)
1234 {
1235         pgd_t *pgd;
1236         pud_t *pud;
1237         pmd_t *pmd;
1238         pte_t *ptep, pte;
1239         spinlock_t *ptl;
1240         struct page *page;
1241         struct mm_struct *mm = vma->vm_mm;
1242
1243         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1244         if (!IS_ERR(page)) {
1245                 BUG_ON(flags & FOLL_GET);
1246                 goto out;
1247         }
1248
1249         page = NULL;
1250         pgd = pgd_offset(mm, address);
1251         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1252                 goto no_page_table;
1253
1254         pud = pud_offset(pgd, address);
1255         if (pud_none(*pud))
1256                 goto no_page_table;
1257         if (pud_huge(*pud)) {
1258                 BUG_ON(flags & FOLL_GET);
1259                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1260                 goto out;
1261         }
1262         if (unlikely(pud_bad(*pud)))
1263                 goto no_page_table;
1264
1265         pmd = pmd_offset(pud, address);
1266         if (pmd_none(*pmd))
1267                 goto no_page_table;
1268         if (pmd_huge(*pmd)) {
1269                 BUG_ON(flags & FOLL_GET);
1270                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1271                 goto out;
1272         }
1273         if (unlikely(pmd_bad(*pmd)))
1274                 goto no_page_table;
1275
1276         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1277
1278         pte = *ptep;
1279         if (!pte_present(pte))
1280                 goto no_page;
1281         if ((flags & FOLL_WRITE) && !pte_write(pte))
1282                 goto unlock;
1283
1284         page = vm_normal_page(vma, address, pte);
1285         if (unlikely(!page)) {
1286                 if ((flags & FOLL_DUMP) ||
1287                     !is_zero_pfn(pte_pfn(pte)))
1288                         goto bad_page;
1289                 page = pte_page(pte);
1290         }
1291
1292         if (flags & FOLL_GET)
1293                 get_page(page);
1294         if (flags & FOLL_TOUCH) {
1295                 if ((flags & FOLL_WRITE) &&
1296                     !pte_dirty(pte) && !PageDirty(page))
1297                         set_page_dirty(page);
1298                 /*
1299                  * pte_mkyoung() would be more correct here, but atomic care
1300                  * is needed to avoid losing the dirty bit: it is easier to use
1301                  * mark_page_accessed().
1302                  */
1303                 mark_page_accessed(page);
1304         }
1305 unlock:
1306         pte_unmap_unlock(ptep, ptl);
1307 out:
1308         return page;
1309
1310 bad_page:
1311         pte_unmap_unlock(ptep, ptl);
1312         return ERR_PTR(-EFAULT);
1313
1314 no_page:
1315         pte_unmap_unlock(ptep, ptl);
1316         if (!pte_none(pte))
1317                 return page;
1318
1319 no_page_table:
1320         /*
1321          * When core dumping an enormous anonymous area that nobody
1322          * has touched so far, we don't want to allocate unnecessary pages or
1323          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1324          * then get_dump_page() will return NULL to leave a hole in the dump.
1325          * But we can only make this optimization where a hole would surely
1326          * be zero-filled if handle_mm_fault() actually did handle it.
1327          */
1328         if ((flags & FOLL_DUMP) &&
1329             (!vma->vm_ops || !vma->vm_ops->fault))
1330                 return ERR_PTR(-EFAULT);
1331         return page;
1332 }
1333
1334 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1335                      unsigned long start, int nr_pages, unsigned int gup_flags,
1336                      struct page **pages, struct vm_area_struct **vmas)
1337 {
1338         int i;
1339         unsigned long vm_flags;
1340
1341         if (nr_pages <= 0)
1342                 return 0;
1343
1344         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1345
1346         /* 
1347          * Require read or write permissions.
1348          * If FOLL_FORCE is set, we only require the "MAY" flags.
1349          */
1350         vm_flags  = (gup_flags & FOLL_WRITE) ?
1351                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1352         vm_flags &= (gup_flags & FOLL_FORCE) ?
1353                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1354         i = 0;
1355
1356         do {
1357                 struct vm_area_struct *vma;
1358
1359                 vma = find_extend_vma(mm, start);
1360                 if (!vma && in_gate_area(tsk, start)) {
1361                         unsigned long pg = start & PAGE_MASK;
1362                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1363                         pgd_t *pgd;
1364                         pud_t *pud;
1365                         pmd_t *pmd;
1366                         pte_t *pte;
1367
1368                         /* user gate pages are read-only */
1369                         if (gup_flags & FOLL_WRITE)
1370                                 return i ? : -EFAULT;
1371                         if (pg > TASK_SIZE)
1372                                 pgd = pgd_offset_k(pg);
1373                         else
1374                                 pgd = pgd_offset_gate(mm, pg);
1375                         BUG_ON(pgd_none(*pgd));
1376                         pud = pud_offset(pgd, pg);
1377                         BUG_ON(pud_none(*pud));
1378                         pmd = pmd_offset(pud, pg);
1379                         if (pmd_none(*pmd))
1380                                 return i ? : -EFAULT;
1381                         pte = pte_offset_map(pmd, pg);
1382                         if (pte_none(*pte)) {
1383                                 pte_unmap(pte);
1384                                 return i ? : -EFAULT;
1385                         }
1386                         if (pages) {
1387                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1388                                 pages[i] = page;
1389                                 if (page)
1390                                         get_page(page);
1391                         }
1392                         pte_unmap(pte);
1393                         if (vmas)
1394                                 vmas[i] = gate_vma;
1395                         i++;
1396                         start += PAGE_SIZE;
1397                         nr_pages--;
1398                         continue;
1399                 }
1400
1401                 if (!vma ||
1402                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1403                     !(vm_flags & vma->vm_flags))
1404                         return i ? : -EFAULT;
1405
1406                 if (is_vm_hugetlb_page(vma)) {
1407                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1408                                         &start, &nr_pages, i, gup_flags);
1409                         continue;
1410                 }
1411
1412                 do {
1413                         struct page *page;
1414                         unsigned int foll_flags = gup_flags;
1415
1416                         /*
1417                          * If we have a pending SIGKILL, don't keep faulting
1418                          * pages and potentially allocating memory.
1419                          */
1420                         if (unlikely(fatal_signal_pending(current)))
1421                                 return i ? i : -ERESTARTSYS;
1422
1423                         cond_resched();
1424                         while (!(page = follow_page(vma, start, foll_flags))) {
1425                                 int ret;
1426
1427                                 ret = handle_mm_fault(mm, vma, start,
1428                                         (foll_flags & FOLL_WRITE) ?
1429                                         FAULT_FLAG_WRITE : 0);
1430
1431                                 if (ret & VM_FAULT_ERROR) {
1432                                         if (ret & VM_FAULT_OOM)
1433                                                 return i ? i : -ENOMEM;
1434                                         if (ret &
1435                                             (VM_FAULT_HWPOISON|VM_FAULT_SIGBUS))
1436                                                 return i ? i : -EFAULT;
1437                                         BUG();
1438                                 }
1439                                 if (ret & VM_FAULT_MAJOR)
1440                                         tsk->maj_flt++;
1441                                 else
1442                                         tsk->min_flt++;
1443
1444                                 /*
1445                                  * The VM_FAULT_WRITE bit tells us that
1446                                  * do_wp_page has broken COW when necessary,
1447                                  * even if maybe_mkwrite decided not to set
1448                                  * pte_write. We can thus safely do subsequent
1449                                  * page lookups as if they were reads. But only
1450                                  * do so when looping for pte_write is futile:
1451                                  * in some cases userspace may also be wanting
1452                                  * to write to the gotten user page, which a
1453                                  * read fault here might prevent (a readonly
1454                                  * page might get reCOWed by userspace write).
1455                                  */
1456                                 if ((ret & VM_FAULT_WRITE) &&
1457                                     !(vma->vm_flags & VM_WRITE))
1458                                         foll_flags &= ~FOLL_WRITE;
1459
1460                                 cond_resched();
1461                         }
1462                         if (IS_ERR(page))
1463                                 return i ? i : PTR_ERR(page);
1464                         if (pages) {
1465                                 pages[i] = page;
1466
1467                                 flush_anon_page(vma, page, start);
1468                                 flush_dcache_page(page);
1469                         }
1470                         if (vmas)
1471                                 vmas[i] = vma;
1472                         i++;
1473                         start += PAGE_SIZE;
1474                         nr_pages--;
1475                 } while (nr_pages && start < vma->vm_end);
1476         } while (nr_pages);
1477         return i;
1478 }
1479
1480 /**
1481  * get_user_pages() - pin user pages in memory
1482  * @tsk:        task_struct of target task
1483  * @mm:         mm_struct of target mm
1484  * @start:      starting user address
1485  * @nr_pages:   number of pages from start to pin
1486  * @write:      whether pages will be written to by the caller
1487  * @force:      whether to force write access even if user mapping is
1488  *              readonly. This will result in the page being COWed even
1489  *              in MAP_SHARED mappings. You do not want this.
1490  * @pages:      array that receives pointers to the pages pinned.
1491  *              Should be at least nr_pages long. Or NULL, if caller
1492  *              only intends to ensure the pages are faulted in.
1493  * @vmas:       array of pointers to vmas corresponding to each page.
1494  *              Or NULL if the caller does not require them.
1495  *
1496  * Returns number of pages pinned. This may be fewer than the number
1497  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1498  * were pinned, returns -errno. Each page returned must be released
1499  * with a put_page() call when it is finished with. vmas will only
1500  * remain valid while mmap_sem is held.
1501  *
1502  * Must be called with mmap_sem held for read or write.
1503  *
1504  * get_user_pages walks a process's page tables and takes a reference to
1505  * each struct page that each user address corresponds to at a given
1506  * instant. That is, it takes the page that would be accessed if a user
1507  * thread accesses the given user virtual address at that instant.
1508  *
1509  * This does not guarantee that the page exists in the user mappings when
1510  * get_user_pages returns, and there may even be a completely different
1511  * page there in some cases (eg. if mmapped pagecache has been invalidated
1512  * and subsequently re faulted). However it does guarantee that the page
1513  * won't be freed completely. And mostly callers simply care that the page
1514  * contains data that was valid *at some point in time*. Typically, an IO
1515  * or similar operation cannot guarantee anything stronger anyway because
1516  * locks can't be held over the syscall boundary.
1517  *
1518  * If write=0, the page must not be written to. If the page is written to,
1519  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1520  * after the page is finished with, and before put_page is called.
1521  *
1522  * get_user_pages is typically used for fewer-copy IO operations, to get a
1523  * handle on the memory by some means other than accesses via the user virtual
1524  * addresses. The pages may be submitted for DMA to devices or accessed via
1525  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1526  * use the correct cache flushing APIs.
1527  *
1528  * See also get_user_pages_fast, for performance critical applications.
1529  */
1530 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1531                 unsigned long start, int nr_pages, int write, int force,
1532                 struct page **pages, struct vm_area_struct **vmas)
1533 {
1534         int flags = FOLL_TOUCH;
1535
1536         if (pages)
1537                 flags |= FOLL_GET;
1538         if (write)
1539                 flags |= FOLL_WRITE;
1540         if (force)
1541                 flags |= FOLL_FORCE;
1542
1543         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
1544 }
1545 EXPORT_SYMBOL(get_user_pages);
1546
1547 /**
1548  * get_dump_page() - pin user page in memory while writing it to core dump
1549  * @addr: user address
1550  *
1551  * Returns struct page pointer of user page pinned for dump,
1552  * to be freed afterwards by page_cache_release() or put_page().
1553  *
1554  * Returns NULL on any kind of failure - a hole must then be inserted into
1555  * the corefile, to preserve alignment with its headers; and also returns
1556  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1557  * allowing a hole to be left in the corefile to save diskspace.
1558  *
1559  * Called without mmap_sem, but after all other threads have been killed.
1560  */
1561 #ifdef CONFIG_ELF_CORE
1562 struct page *get_dump_page(unsigned long addr)
1563 {
1564         struct vm_area_struct *vma;
1565         struct page *page;
1566
1567         if (__get_user_pages(current, current->mm, addr, 1,
1568                         FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma) < 1)
1569                 return NULL;
1570         flush_cache_page(vma, addr, page_to_pfn(page));
1571         return page;
1572 }
1573 #endif /* CONFIG_ELF_CORE */
1574
1575 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1576                         spinlock_t **ptl)
1577 {
1578         pgd_t * pgd = pgd_offset(mm, addr);
1579         pud_t * pud = pud_alloc(mm, pgd, addr);
1580         if (pud) {
1581                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1582                 if (pmd)
1583                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1584         }
1585         return NULL;
1586 }
1587
1588 /*
1589  * This is the old fallback for page remapping.
1590  *
1591  * For historical reasons, it only allows reserved pages. Only
1592  * old drivers should use this, and they needed to mark their
1593  * pages reserved for the old functions anyway.
1594  */
1595 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1596                         struct page *page, pgprot_t prot)
1597 {
1598         struct mm_struct *mm = vma->vm_mm;
1599         int retval;
1600         pte_t *pte;
1601         spinlock_t *ptl;
1602
1603         retval = -EINVAL;
1604         if (PageAnon(page))
1605                 goto out;
1606         retval = -ENOMEM;
1607         flush_dcache_page(page);
1608         pte = get_locked_pte(mm, addr, &ptl);
1609         if (!pte)
1610                 goto out;
1611         retval = -EBUSY;
1612         if (!pte_none(*pte))
1613                 goto out_unlock;
1614
1615         /* Ok, finally just insert the thing.. */
1616         get_page(page);
1617         inc_mm_counter_fast(mm, MM_FILEPAGES);
1618         page_add_file_rmap(page);
1619         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1620
1621         retval = 0;
1622         pte_unmap_unlock(pte, ptl);
1623         return retval;
1624 out_unlock:
1625         pte_unmap_unlock(pte, ptl);
1626 out:
1627         return retval;
1628 }
1629
1630 /**
1631  * vm_insert_page - insert single page into user vma
1632  * @vma: user vma to map to
1633  * @addr: target user address of this page
1634  * @page: source kernel page
1635  *
1636  * This allows drivers to insert individual pages they've allocated
1637  * into a user vma.
1638  *
1639  * The page has to be a nice clean _individual_ kernel allocation.
1640  * If you allocate a compound page, you need to have marked it as
1641  * such (__GFP_COMP), or manually just split the page up yourself
1642  * (see split_page()).
1643  *
1644  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1645  * took an arbitrary page protection parameter. This doesn't allow
1646  * that. Your vma protection will have to be set up correctly, which
1647  * means that if you want a shared writable mapping, you'd better
1648  * ask for a shared writable mapping!
1649  *
1650  * The page does not need to be reserved.
1651  */
1652 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1653                         struct page *page)
1654 {
1655         if (addr < vma->vm_start || addr >= vma->vm_end)
1656                 return -EFAULT;
1657         if (!page_count(page))
1658                 return -EINVAL;
1659         vma->vm_flags |= VM_INSERTPAGE;
1660         return insert_page(vma, addr, page, vma->vm_page_prot);
1661 }
1662 EXPORT_SYMBOL(vm_insert_page);
1663
1664 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1665                         unsigned long pfn, pgprot_t prot)
1666 {
1667         struct mm_struct *mm = vma->vm_mm;
1668         int retval;
1669         pte_t *pte, entry;
1670         spinlock_t *ptl;
1671
1672         retval = -ENOMEM;
1673         pte = get_locked_pte(mm, addr, &ptl);
1674         if (!pte)
1675                 goto out;
1676         retval = -EBUSY;
1677         if (!pte_none(*pte))
1678                 goto out_unlock;
1679
1680         /* Ok, finally just insert the thing.. */
1681         entry = pte_mkspecial(pfn_pte(pfn, prot));
1682         set_pte_at(mm, addr, pte, entry);
1683         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1684
1685         retval = 0;
1686 out_unlock:
1687         pte_unmap_unlock(pte, ptl);
1688 out:
1689         return retval;
1690 }
1691
1692 /**
1693  * vm_insert_pfn - insert single pfn into user vma
1694  * @vma: user vma to map to
1695  * @addr: target user address of this page
1696  * @pfn: source kernel pfn
1697  *
1698  * Similar to vm_inert_page, this allows drivers to insert individual pages
1699  * they've allocated into a user vma. Same comments apply.
1700  *
1701  * This function should only be called from a vm_ops->fault handler, and
1702  * in that case the handler should return NULL.
1703  *
1704  * vma cannot be a COW mapping.
1705  *
1706  * As this is called only for pages that do not currently exist, we
1707  * do not need to flush old virtual caches or the TLB.
1708  */
1709 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1710                         unsigned long pfn)
1711 {
1712         int ret;
1713         pgprot_t pgprot = vma->vm_page_prot;
1714         /*
1715          * Technically, architectures with pte_special can avoid all these
1716          * restrictions (same for remap_pfn_range).  However we would like
1717          * consistency in testing and feature parity among all, so we should
1718          * try to keep these invariants in place for everybody.
1719          */
1720         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1721         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1722                                                 (VM_PFNMAP|VM_MIXEDMAP));
1723         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1724         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1725
1726         if (addr < vma->vm_start || addr >= vma->vm_end)
1727                 return -EFAULT;
1728         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1729                 return -EINVAL;
1730
1731         ret = insert_pfn(vma, addr, pfn, pgprot);
1732
1733         if (ret)
1734                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1735
1736         return ret;
1737 }
1738 EXPORT_SYMBOL(vm_insert_pfn);
1739
1740 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1741                         unsigned long pfn)
1742 {
1743         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1744
1745         if (addr < vma->vm_start || addr >= vma->vm_end)
1746                 return -EFAULT;
1747
1748         /*
1749          * If we don't have pte special, then we have to use the pfn_valid()
1750          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1751          * refcount the page if pfn_valid is true (hence insert_page rather
1752          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1753          * without pte special, it would there be refcounted as a normal page.
1754          */
1755         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1756                 struct page *page;
1757
1758                 page = pfn_to_page(pfn);
1759                 return insert_page(vma, addr, page, vma->vm_page_prot);
1760         }
1761         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1762 }
1763 EXPORT_SYMBOL(vm_insert_mixed);
1764
1765 /*
1766  * maps a range of physical memory into the requested pages. the old
1767  * mappings are removed. any references to nonexistent pages results
1768  * in null mappings (currently treated as "copy-on-access")
1769  */
1770 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1771                         unsigned long addr, unsigned long end,
1772                         unsigned long pfn, pgprot_t prot)
1773 {
1774         pte_t *pte;
1775         spinlock_t *ptl;
1776
1777         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1778         if (!pte)
1779                 return -ENOMEM;
1780         arch_enter_lazy_mmu_mode();
1781         do {
1782                 BUG_ON(!pte_none(*pte));
1783                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1784                 pfn++;
1785         } while (pte++, addr += PAGE_SIZE, addr != end);
1786         arch_leave_lazy_mmu_mode();
1787         pte_unmap_unlock(pte - 1, ptl);
1788         return 0;
1789 }
1790
1791 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1792                         unsigned long addr, unsigned long end,
1793                         unsigned long pfn, pgprot_t prot)
1794 {
1795         pmd_t *pmd;
1796         unsigned long next;
1797
1798         pfn -= addr >> PAGE_SHIFT;
1799         pmd = pmd_alloc(mm, pud, addr);
1800         if (!pmd)
1801                 return -ENOMEM;
1802         do {
1803                 next = pmd_addr_end(addr, end);
1804                 if (remap_pte_range(mm, pmd, addr, next,
1805                                 pfn + (addr >> PAGE_SHIFT), prot))
1806                         return -ENOMEM;
1807         } while (pmd++, addr = next, addr != end);
1808         return 0;
1809 }
1810
1811 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1812                         unsigned long addr, unsigned long end,
1813                         unsigned long pfn, pgprot_t prot)
1814 {
1815         pud_t *pud;
1816         unsigned long next;
1817
1818         pfn -= addr >> PAGE_SHIFT;
1819         pud = pud_alloc(mm, pgd, addr);
1820         if (!pud)
1821                 return -ENOMEM;
1822         do {
1823                 next = pud_addr_end(addr, end);
1824                 if (remap_pmd_range(mm, pud, addr, next,
1825                                 pfn + (addr >> PAGE_SHIFT), prot))
1826                         return -ENOMEM;
1827         } while (pud++, addr = next, addr != end);
1828         return 0;
1829 }
1830
1831 /**
1832  * remap_pfn_range - remap kernel memory to userspace
1833  * @vma: user vma to map to
1834  * @addr: target user address to start at
1835  * @pfn: physical address of kernel memory
1836  * @size: size of map area
1837  * @prot: page protection flags for this mapping
1838  *
1839  *  Note: this is only safe if the mm semaphore is held when called.
1840  */
1841 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1842                     unsigned long pfn, unsigned long size, pgprot_t prot)
1843 {
1844         pgd_t *pgd;
1845         unsigned long next;
1846         unsigned long end = addr + PAGE_ALIGN(size);
1847         struct mm_struct *mm = vma->vm_mm;
1848         int err;
1849
1850         /*
1851          * Physically remapped pages are special. Tell the
1852          * rest of the world about it:
1853          *   VM_IO tells people not to look at these pages
1854          *      (accesses can have side effects).
1855          *   VM_RESERVED is specified all over the place, because
1856          *      in 2.4 it kept swapout's vma scan off this vma; but
1857          *      in 2.6 the LRU scan won't even find its pages, so this
1858          *      flag means no more than count its pages in reserved_vm,
1859          *      and omit it from core dump, even when VM_IO turned off.
1860          *   VM_PFNMAP tells the core MM that the base pages are just
1861          *      raw PFN mappings, and do not have a "struct page" associated
1862          *      with them.
1863          *
1864          * There's a horrible special case to handle copy-on-write
1865          * behaviour that some programs depend on. We mark the "original"
1866          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1867          */
1868         if (addr == vma->vm_start && end == vma->vm_end) {
1869                 vma->vm_pgoff = pfn;
1870                 vma->vm_flags |= VM_PFN_AT_MMAP;
1871         } else if (is_cow_mapping(vma->vm_flags))
1872                 return -EINVAL;
1873
1874         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1875
1876         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1877         if (err) {
1878                 /*
1879                  * To indicate that track_pfn related cleanup is not
1880                  * needed from higher level routine calling unmap_vmas
1881                  */
1882                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1883                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1884                 return -EINVAL;
1885         }
1886
1887         BUG_ON(addr >= end);
1888         pfn -= addr >> PAGE_SHIFT;
1889         pgd = pgd_offset(mm, addr);
1890         flush_cache_range(vma, addr, end);
1891         do {
1892                 next = pgd_addr_end(addr, end);
1893                 err = remap_pud_range(mm, pgd, addr, next,
1894                                 pfn + (addr >> PAGE_SHIFT), prot);
1895                 if (err)
1896                         break;
1897         } while (pgd++, addr = next, addr != end);
1898
1899         if (err)
1900                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1901
1902         return err;
1903 }
1904 EXPORT_SYMBOL(remap_pfn_range);
1905
1906 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1907                                      unsigned long addr, unsigned long end,
1908                                      pte_fn_t fn, void *data)
1909 {
1910         pte_t *pte;
1911         int err;
1912         pgtable_t token;
1913         spinlock_t *uninitialized_var(ptl);
1914
1915         pte = (mm == &init_mm) ?
1916                 pte_alloc_kernel(pmd, addr) :
1917                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1918         if (!pte)
1919                 return -ENOMEM;
1920
1921         BUG_ON(pmd_huge(*pmd));
1922
1923         arch_enter_lazy_mmu_mode();
1924
1925         token = pmd_pgtable(*pmd);
1926
1927         do {
1928                 err = fn(pte++, token, addr, data);
1929                 if (err)
1930                         break;
1931         } while (addr += PAGE_SIZE, addr != end);
1932
1933         arch_leave_lazy_mmu_mode();
1934
1935         if (mm != &init_mm)
1936                 pte_unmap_unlock(pte-1, ptl);
1937         return err;
1938 }
1939
1940 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1941                                      unsigned long addr, unsigned long end,
1942                                      pte_fn_t fn, void *data)
1943 {
1944         pmd_t *pmd;
1945         unsigned long next;
1946         int err;
1947
1948         BUG_ON(pud_huge(*pud));
1949
1950         pmd = pmd_alloc(mm, pud, addr);
1951         if (!pmd)
1952                 return -ENOMEM;
1953         do {
1954                 next = pmd_addr_end(addr, end);
1955                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1956                 if (err)
1957                         break;
1958         } while (pmd++, addr = next, addr != end);
1959         return err;
1960 }
1961
1962 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1963                                      unsigned long addr, unsigned long end,
1964                                      pte_fn_t fn, void *data)
1965 {
1966         pud_t *pud;
1967         unsigned long next;
1968         int err;
1969
1970         pud = pud_alloc(mm, pgd, addr);
1971         if (!pud)
1972                 return -ENOMEM;
1973         do {
1974                 next = pud_addr_end(addr, end);
1975                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1976                 if (err)
1977                         break;
1978         } while (pud++, addr = next, addr != end);
1979         return err;
1980 }
1981
1982 /*
1983  * Scan a region of virtual memory, filling in page tables as necessary
1984  * and calling a provided function on each leaf page table.
1985  */
1986 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1987                         unsigned long size, pte_fn_t fn, void *data)
1988 {
1989         pgd_t *pgd;
1990         unsigned long next;
1991         unsigned long start = addr, end = addr + size;
1992         int err;
1993
1994         BUG_ON(addr >= end);
1995         mmu_notifier_invalidate_range_start(mm, start, end);
1996         pgd = pgd_offset(mm, addr);
1997         do {
1998                 next = pgd_addr_end(addr, end);
1999                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2000                 if (err)
2001                         break;
2002         } while (pgd++, addr = next, addr != end);
2003         mmu_notifier_invalidate_range_end(mm, start, end);
2004         return err;
2005 }
2006 EXPORT_SYMBOL_GPL(apply_to_page_range);
2007
2008 /*
2009  * handle_pte_fault chooses page fault handler according to an entry
2010  * which was read non-atomically.  Before making any commitment, on
2011  * those architectures or configurations (e.g. i386 with PAE) which
2012  * might give a mix of unmatched parts, do_swap_page and do_file_page
2013  * must check under lock before unmapping the pte and proceeding
2014  * (but do_wp_page is only called after already making such a check;
2015  * and do_anonymous_page and do_no_page can safely check later on).
2016  */
2017 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2018                                 pte_t *page_table, pte_t orig_pte)
2019 {
2020         int same = 1;
2021 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2022         if (sizeof(pte_t) > sizeof(unsigned long)) {
2023                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2024                 spin_lock(ptl);
2025                 same = pte_same(*page_table, orig_pte);
2026                 spin_unlock(ptl);
2027         }
2028 #endif
2029         pte_unmap(page_table);
2030         return same;
2031 }
2032
2033 /*
2034  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
2035  * servicing faults for write access.  In the normal case, do always want
2036  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
2037  * that do not have writing enabled, when used by access_process_vm.
2038  */
2039 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
2040 {
2041         if (likely(vma->vm_flags & VM_WRITE))
2042                 pte = pte_mkwrite(pte);
2043         return pte;
2044 }
2045
2046 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2047 {
2048         /*
2049          * If the source page was a PFN mapping, we don't have
2050          * a "struct page" for it. We do a best-effort copy by
2051          * just copying from the original user address. If that
2052          * fails, we just zero-fill it. Live with it.
2053          */
2054         if (unlikely(!src)) {
2055                 void *kaddr = kmap_atomic(dst, KM_USER0);
2056                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2057
2058                 /*
2059                  * This really shouldn't fail, because the page is there
2060                  * in the page tables. But it might just be unreadable,
2061                  * in which case we just give up and fill the result with
2062                  * zeroes.
2063                  */
2064                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2065                         memset(kaddr, 0, PAGE_SIZE);
2066                 kunmap_atomic(kaddr, KM_USER0);
2067                 flush_dcache_page(dst);
2068         } else
2069                 copy_user_highpage(dst, src, va, vma);
2070 }
2071
2072 /*
2073  * This routine handles present pages, when users try to write
2074  * to a shared page. It is done by copying the page to a new address
2075  * and decrementing the shared-page counter for the old page.
2076  *
2077  * Note that this routine assumes that the protection checks have been
2078  * done by the caller (the low-level page fault routine in most cases).
2079  * Thus we can safely just mark it writable once we've done any necessary
2080  * COW.
2081  *
2082  * We also mark the page dirty at this point even though the page will
2083  * change only once the write actually happens. This avoids a few races,
2084  * and potentially makes it more efficient.
2085  *
2086  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2087  * but allow concurrent faults), with pte both mapped and locked.
2088  * We return with mmap_sem still held, but pte unmapped and unlocked.
2089  */
2090 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2091                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2092                 spinlock_t *ptl, pte_t orig_pte)
2093 {
2094         struct page *old_page, *new_page;
2095         pte_t entry;
2096         int reuse = 0, ret = 0;
2097         int page_mkwrite = 0;
2098         struct page *dirty_page = NULL;
2099
2100         old_page = vm_normal_page(vma, address, orig_pte);
2101         if (!old_page) {
2102                 /*
2103                  * VM_MIXEDMAP !pfn_valid() case
2104                  *
2105                  * We should not cow pages in a shared writeable mapping.
2106                  * Just mark the pages writable as we can't do any dirty
2107                  * accounting on raw pfn maps.
2108                  */
2109                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2110                                      (VM_WRITE|VM_SHARED))
2111                         goto reuse;
2112                 goto gotten;
2113         }
2114
2115         /*
2116          * Take out anonymous pages first, anonymous shared vmas are
2117          * not dirty accountable.
2118          */
2119         if (PageAnon(old_page) && !PageKsm(old_page)) {
2120                 if (!trylock_page(old_page)) {
2121                         page_cache_get(old_page);
2122                         pte_unmap_unlock(page_table, ptl);
2123                         lock_page(old_page);
2124                         page_table = pte_offset_map_lock(mm, pmd, address,
2125                                                          &ptl);
2126                         if (!pte_same(*page_table, orig_pte)) {
2127                                 unlock_page(old_page);
2128                                 page_cache_release(old_page);
2129                                 goto unlock;
2130                         }
2131                         page_cache_release(old_page);
2132                 }
2133                 reuse = reuse_swap_page(old_page);
2134                 unlock_page(old_page);
2135         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2136                                         (VM_WRITE|VM_SHARED))) {
2137                 /*
2138                  * Only catch write-faults on shared writable pages,
2139                  * read-only shared pages can get COWed by
2140                  * get_user_pages(.write=1, .force=1).
2141                  */
2142                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2143                         struct vm_fault vmf;
2144                         int tmp;
2145
2146                         vmf.virtual_address = (void __user *)(address &
2147                                                                 PAGE_MASK);
2148                         vmf.pgoff = old_page->index;
2149                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2150                         vmf.page = old_page;
2151
2152                         /*
2153                          * Notify the address space that the page is about to
2154                          * become writable so that it can prohibit this or wait
2155                          * for the page to get into an appropriate state.
2156                          *
2157                          * We do this without the lock held, so that it can
2158                          * sleep if it needs to.
2159                          */
2160                         page_cache_get(old_page);
2161                         pte_unmap_unlock(page_table, ptl);
2162
2163                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2164                         if (unlikely(tmp &
2165                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2166                                 ret = tmp;
2167                                 goto unwritable_page;
2168                         }
2169                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2170                                 lock_page(old_page);
2171                                 if (!old_page->mapping) {
2172                                         ret = 0; /* retry the fault */
2173                                         unlock_page(old_page);
2174                                         goto unwritable_page;
2175                                 }
2176                         } else
2177                                 VM_BUG_ON(!PageLocked(old_page));
2178
2179                         /*
2180                          * Since we dropped the lock we need to revalidate
2181                          * the PTE as someone else may have changed it.  If
2182                          * they did, we just return, as we can count on the
2183                          * MMU to tell us if they didn't also make it writable.
2184                          */
2185                         page_table = pte_offset_map_lock(mm, pmd, address,
2186                                                          &ptl);
2187                         if (!pte_same(*page_table, orig_pte)) {
2188                                 unlock_page(old_page);
2189                                 page_cache_release(old_page);
2190                                 goto unlock;
2191                         }
2192
2193                         page_mkwrite = 1;
2194                 }
2195                 dirty_page = old_page;
2196                 get_page(dirty_page);
2197                 reuse = 1;
2198         }
2199
2200         if (reuse) {
2201 reuse:
2202                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2203                 entry = pte_mkyoung(orig_pte);
2204                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2205                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2206                         update_mmu_cache(vma, address, page_table);
2207                 ret |= VM_FAULT_WRITE;
2208                 goto unlock;
2209         }
2210
2211         /*
2212          * Ok, we need to copy. Oh, well..
2213          */
2214         page_cache_get(old_page);
2215 gotten:
2216         pte_unmap_unlock(page_table, ptl);
2217
2218         if (unlikely(anon_vma_prepare(vma)))
2219                 goto oom;
2220
2221         if (is_zero_pfn(pte_pfn(orig_pte))) {
2222                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2223                 if (!new_page)
2224                         goto oom;
2225         } else {
2226                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2227                 if (!new_page)
2228                         goto oom;
2229                 cow_user_page(new_page, old_page, address, vma);
2230         }
2231         __SetPageUptodate(new_page);
2232
2233         /*
2234          * Don't let another task, with possibly unlocked vma,
2235          * keep the mlocked page.
2236          */
2237         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2238                 lock_page(old_page);    /* for LRU manipulation */
2239                 clear_page_mlock(old_page);
2240                 unlock_page(old_page);
2241         }
2242
2243         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2244                 goto oom_free_new;
2245
2246         /*
2247          * Re-check the pte - we dropped the lock
2248          */
2249         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2250         if (likely(pte_same(*page_table, orig_pte))) {
2251                 if (old_page) {
2252                         if (!PageAnon(old_page)) {
2253                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2254                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2255                         }
2256                 } else
2257                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2258                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2259                 entry = mk_pte(new_page, vma->vm_page_prot);
2260                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2261                 /*
2262                  * Clear the pte entry and flush it first, before updating the
2263                  * pte with the new entry. This will avoid a race condition
2264                  * seen in the presence of one thread doing SMC and another
2265                  * thread doing COW.
2266                  */
2267                 ptep_clear_flush(vma, address, page_table);
2268                 page_add_new_anon_rmap(new_page, vma, address);
2269                 /*
2270                  * We call the notify macro here because, when using secondary
2271                  * mmu page tables (such as kvm shadow page tables), we want the
2272                  * new page to be mapped directly into the secondary page table.
2273                  */
2274                 set_pte_at_notify(mm, address, page_table, entry);
2275                 update_mmu_cache(vma, address, page_table);
2276                 if (old_page) {
2277                         /*
2278                          * Only after switching the pte to the new page may
2279                          * we remove the mapcount here. Otherwise another
2280                          * process may come and find the rmap count decremented
2281                          * before the pte is switched to the new page, and
2282                          * "reuse" the old page writing into it while our pte
2283                          * here still points into it and can be read by other
2284                          * threads.
2285                          *
2286                          * The critical issue is to order this
2287                          * page_remove_rmap with the ptp_clear_flush above.
2288                          * Those stores are ordered by (if nothing else,)
2289                          * the barrier present in the atomic_add_negative
2290                          * in page_remove_rmap.
2291                          *
2292                          * Then the TLB flush in ptep_clear_flush ensures that
2293                          * no process can access the old page before the
2294                          * decremented mapcount is visible. And the old page
2295                          * cannot be reused until after the decremented
2296                          * mapcount is visible. So transitively, TLBs to
2297                          * old page will be flushed before it can be reused.
2298                          */
2299                         page_remove_rmap(old_page);
2300                 }
2301
2302                 /* Free the old page.. */
2303                 new_page = old_page;
2304                 ret |= VM_FAULT_WRITE;
2305         } else
2306                 mem_cgroup_uncharge_page(new_page);
2307
2308         if (new_page)
2309                 page_cache_release(new_page);
2310         if (old_page)
2311                 page_cache_release(old_page);
2312 unlock:
2313         pte_unmap_unlock(page_table, ptl);
2314         if (dirty_page) {
2315                 /*
2316                  * Yes, Virginia, this is actually required to prevent a race
2317                  * with clear_page_dirty_for_io() from clearing the page dirty
2318                  * bit after it clear all dirty ptes, but before a racing
2319                  * do_wp_page installs a dirty pte.
2320                  *
2321                  * do_no_page is protected similarly.
2322                  */
2323                 if (!page_mkwrite) {
2324                         wait_on_page_locked(dirty_page);
2325                         set_page_dirty_balance(dirty_page, page_mkwrite);
2326                 }
2327                 put_page(dirty_page);
2328                 if (page_mkwrite) {
2329                         struct address_space *mapping = dirty_page->mapping;
2330
2331                         set_page_dirty(dirty_page);
2332                         unlock_page(dirty_page);
2333                         page_cache_release(dirty_page);
2334                         if (mapping)    {
2335                                 /*
2336                                  * Some device drivers do not set page.mapping
2337                                  * but still dirty their pages
2338                                  */
2339                                 balance_dirty_pages_ratelimited(mapping);
2340                         }
2341                 }
2342
2343                 /* file_update_time outside page_lock */
2344                 if (vma->vm_file)
2345                         file_update_time(vma->vm_file);
2346         }
2347         return ret;
2348 oom_free_new:
2349         page_cache_release(new_page);
2350 oom:
2351         if (old_page) {
2352                 if (page_mkwrite) {
2353                         unlock_page(old_page);
2354                         page_cache_release(old_page);
2355                 }
2356                 page_cache_release(old_page);
2357         }
2358         return VM_FAULT_OOM;
2359
2360 unwritable_page:
2361         page_cache_release(old_page);
2362         return ret;
2363 }
2364
2365 /*
2366  * Helper functions for unmap_mapping_range().
2367  *
2368  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2369  *
2370  * We have to restart searching the prio_tree whenever we drop the lock,
2371  * since the iterator is only valid while the lock is held, and anyway
2372  * a later vma might be split and reinserted earlier while lock dropped.
2373  *
2374  * The list of nonlinear vmas could be handled more efficiently, using
2375  * a placeholder, but handle it in the same way until a need is shown.
2376  * It is important to search the prio_tree before nonlinear list: a vma
2377  * may become nonlinear and be shifted from prio_tree to nonlinear list
2378  * while the lock is dropped; but never shifted from list to prio_tree.
2379  *
2380  * In order to make forward progress despite restarting the search,
2381  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2382  * quickly skip it next time around.  Since the prio_tree search only
2383  * shows us those vmas affected by unmapping the range in question, we
2384  * can't efficiently keep all vmas in step with mapping->truncate_count:
2385  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2386  * mapping->truncate_count and vma->vm_truncate_count are protected by
2387  * i_mmap_lock.
2388  *
2389  * In order to make forward progress despite repeatedly restarting some
2390  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2391  * and restart from that address when we reach that vma again.  It might
2392  * have been split or merged, shrunk or extended, but never shifted: so
2393  * restart_addr remains valid so long as it remains in the vma's range.
2394  * unmap_mapping_range forces truncate_count to leap over page-aligned
2395  * values so we can save vma's restart_addr in its truncate_count field.
2396  */
2397 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2398
2399 static void reset_vma_truncate_counts(struct address_space *mapping)
2400 {
2401         struct vm_area_struct *vma;
2402         struct prio_tree_iter iter;
2403
2404         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2405                 vma->vm_truncate_count = 0;
2406         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2407                 vma->vm_truncate_count = 0;
2408 }
2409
2410 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2411                 unsigned long start_addr, unsigned long end_addr,
2412                 struct zap_details *details)
2413 {
2414         unsigned long restart_addr;
2415         int need_break;
2416
2417         /*
2418          * files that support invalidating or truncating portions of the
2419          * file from under mmaped areas must have their ->fault function
2420          * return a locked page (and set VM_FAULT_LOCKED in the return).
2421          * This provides synchronisation against concurrent unmapping here.
2422          */
2423
2424 again:
2425         restart_addr = vma->vm_truncate_count;
2426         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2427                 start_addr = restart_addr;
2428                 if (start_addr >= end_addr) {
2429                         /* Top of vma has been split off since last time */
2430                         vma->vm_truncate_count = details->truncate_count;
2431                         return 0;
2432                 }
2433         }
2434
2435         restart_addr = zap_page_range(vma, start_addr,
2436                                         end_addr - start_addr, details);
2437         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2438
2439         if (restart_addr >= end_addr) {
2440                 /* We have now completed this vma: mark it so */
2441                 vma->vm_truncate_count = details->truncate_count;
2442                 if (!need_break)
2443                         return 0;
2444         } else {
2445                 /* Note restart_addr in vma's truncate_count field */
2446                 vma->vm_truncate_count = restart_addr;
2447                 if (!need_break)
2448                         goto again;
2449         }
2450
2451         spin_unlock(details->i_mmap_lock);
2452         cond_resched();
2453         spin_lock(details->i_mmap_lock);
2454         return -EINTR;
2455 }
2456
2457 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2458                                             struct zap_details *details)
2459 {
2460         struct vm_area_struct *vma;
2461         struct prio_tree_iter iter;
2462         pgoff_t vba, vea, zba, zea;
2463
2464 restart:
2465         vma_prio_tree_foreach(vma, &iter, root,
2466                         details->first_index, details->last_index) {
2467                 /* Skip quickly over those we have already dealt with */
2468                 if (vma->vm_truncate_count == details->truncate_count)
2469                         continue;
2470
2471                 vba = vma->vm_pgoff;
2472                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2473                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2474                 zba = details->first_index;
2475                 if (zba < vba)
2476                         zba = vba;
2477                 zea = details->last_index;
2478                 if (zea > vea)
2479                         zea = vea;
2480
2481                 if (unmap_mapping_range_vma(vma,
2482                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2483                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2484                                 details) < 0)
2485                         goto restart;
2486         }
2487 }
2488
2489 static inline void unmap_mapping_range_list(struct list_head *head,
2490                                             struct zap_details *details)
2491 {
2492         struct vm_area_struct *vma;
2493
2494         /*
2495          * In nonlinear VMAs there is no correspondence between virtual address
2496          * offset and file offset.  So we must perform an exhaustive search
2497          * across *all* the pages in each nonlinear VMA, not just the pages
2498          * whose virtual address lies outside the file truncation point.
2499          */
2500 restart:
2501         list_for_each_entry(vma, head, shared.vm_set.list) {
2502                 /* Skip quickly over those we have already dealt with */
2503                 if (vma->vm_truncate_count == details->truncate_count)
2504                         continue;
2505                 details->nonlinear_vma = vma;
2506                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2507                                         vma->vm_end, details) < 0)
2508                         goto restart;
2509         }
2510 }
2511
2512 /**
2513  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2514  * @mapping: the address space containing mmaps to be unmapped.
2515  * @holebegin: byte in first page to unmap, relative to the start of
2516  * the underlying file.  This will be rounded down to a PAGE_SIZE
2517  * boundary.  Note that this is different from truncate_pagecache(), which
2518  * must keep the partial page.  In contrast, we must get rid of
2519  * partial pages.
2520  * @holelen: size of prospective hole in bytes.  This will be rounded
2521  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2522  * end of the file.
2523  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2524  * but 0 when invalidating pagecache, don't throw away private data.
2525  */
2526 void unmap_mapping_range(struct address_space *mapping,
2527                 loff_t const holebegin, loff_t const holelen, int even_cows)
2528 {
2529         struct zap_details details;
2530         pgoff_t hba = holebegin >> PAGE_SHIFT;
2531         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2532
2533         /* Check for overflow. */
2534         if (sizeof(holelen) > sizeof(hlen)) {
2535                 long long holeend =
2536                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2537                 if (holeend & ~(long long)ULONG_MAX)
2538                         hlen = ULONG_MAX - hba + 1;
2539         }
2540
2541         details.check_mapping = even_cows? NULL: mapping;
2542         details.nonlinear_vma = NULL;
2543         details.first_index = hba;
2544         details.last_index = hba + hlen - 1;
2545         if (details.last_index < details.first_index)
2546                 details.last_index = ULONG_MAX;
2547         details.i_mmap_lock = &mapping->i_mmap_lock;
2548
2549         spin_lock(&mapping->i_mmap_lock);
2550
2551         /* Protect against endless unmapping loops */
2552         mapping->truncate_count++;
2553         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2554                 if (mapping->truncate_count == 0)
2555                         reset_vma_truncate_counts(mapping);
2556                 mapping->truncate_count++;
2557         }
2558         details.truncate_count = mapping->truncate_count;
2559
2560         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2561                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2562         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2563                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2564         spin_unlock(&mapping->i_mmap_lock);
2565 }
2566 EXPORT_SYMBOL(unmap_mapping_range);
2567
2568 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2569 {
2570         struct address_space *mapping = inode->i_mapping;
2571
2572         /*
2573          * If the underlying filesystem is not going to provide
2574          * a way to truncate a range of blocks (punch a hole) -
2575          * we should return failure right now.
2576          */
2577         if (!inode->i_op->truncate_range)
2578                 return -ENOSYS;
2579
2580         mutex_lock(&inode->i_mutex);
2581         down_write(&inode->i_alloc_sem);
2582         unmap_mapping_range(mapping, offset, (end - offset), 1);
2583         truncate_inode_pages_range(mapping, offset, end);
2584         unmap_mapping_range(mapping, offset, (end - offset), 1);
2585         inode->i_op->truncate_range(inode, offset, end);
2586         up_write(&inode->i_alloc_sem);
2587         mutex_unlock(&inode->i_mutex);
2588
2589         return 0;
2590 }
2591
2592 /*
2593  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2594  * but allow concurrent faults), and pte mapped but not yet locked.
2595  * We return with mmap_sem still held, but pte unmapped and unlocked.
2596  */
2597 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2598                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2599                 unsigned int flags, pte_t orig_pte)
2600 {
2601         spinlock_t *ptl;
2602         struct page *page;
2603         swp_entry_t entry;
2604         pte_t pte;
2605         struct mem_cgroup *ptr = NULL;
2606         int ret = 0;
2607
2608         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2609                 goto out;
2610
2611         entry = pte_to_swp_entry(orig_pte);
2612         if (unlikely(non_swap_entry(entry))) {
2613                 if (is_migration_entry(entry)) {
2614                         migration_entry_wait(mm, pmd, address);
2615                 } else if (is_hwpoison_entry(entry)) {
2616                         ret = VM_FAULT_HWPOISON;
2617                 } else {
2618                         print_bad_pte(vma, address, orig_pte, NULL);
2619                         ret = VM_FAULT_SIGBUS;
2620                 }
2621                 goto out;
2622         }
2623         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2624         page = lookup_swap_cache(entry);
2625         if (!page) {
2626                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2627                 page = swapin_readahead(entry,
2628                                         GFP_HIGHUSER_MOVABLE, vma, address);
2629                 if (!page) {
2630                         /*
2631                          * Back out if somebody else faulted in this pte
2632                          * while we released the pte lock.
2633                          */
2634                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2635                         if (likely(pte_same(*page_table, orig_pte)))
2636                                 ret = VM_FAULT_OOM;
2637                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2638                         goto unlock;
2639                 }
2640
2641                 /* Had to read the page from swap area: Major fault */
2642                 ret = VM_FAULT_MAJOR;
2643                 count_vm_event(PGMAJFAULT);
2644         } else if (PageHWPoison(page)) {
2645                 /*
2646                  * hwpoisoned dirty swapcache pages are kept for killing
2647                  * owner processes (which may be unknown at hwpoison time)
2648                  */
2649                 ret = VM_FAULT_HWPOISON;
2650                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2651                 goto out_release;
2652         }
2653
2654         lock_page(page);
2655         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2656
2657         page = ksm_might_need_to_copy(page, vma, address);
2658         if (!page) {
2659                 ret = VM_FAULT_OOM;
2660                 goto out;
2661         }
2662
2663         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2664                 ret = VM_FAULT_OOM;
2665                 goto out_page;
2666         }
2667
2668         /*
2669          * Back out if somebody else already faulted in this pte.
2670          */
2671         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2672         if (unlikely(!pte_same(*page_table, orig_pte)))
2673                 goto out_nomap;
2674
2675         if (unlikely(!PageUptodate(page))) {
2676                 ret = VM_FAULT_SIGBUS;
2677                 goto out_nomap;
2678         }
2679
2680         /*
2681          * The page isn't present yet, go ahead with the fault.
2682          *
2683          * Be careful about the sequence of operations here.
2684          * To get its accounting right, reuse_swap_page() must be called
2685          * while the page is counted on swap but not yet in mapcount i.e.
2686          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2687          * must be called after the swap_free(), or it will never succeed.
2688          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2689          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2690          * in page->private. In this case, a record in swap_cgroup  is silently
2691          * discarded at swap_free().
2692          */
2693
2694         inc_mm_counter_fast(mm, MM_ANONPAGES);
2695         pte = mk_pte(page, vma->vm_page_prot);
2696         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2697                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2698                 flags &= ~FAULT_FLAG_WRITE;
2699         }
2700         flush_icache_page(vma, page);
2701         set_pte_at(mm, address, page_table, pte);
2702         page_add_anon_rmap(page, vma, address);
2703         /* It's better to call commit-charge after rmap is established */
2704         mem_cgroup_commit_charge_swapin(page, ptr);
2705
2706         swap_free(entry);
2707         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2708                 try_to_free_swap(page);
2709         unlock_page(page);
2710
2711         if (flags & FAULT_FLAG_WRITE) {
2712                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2713                 if (ret & VM_FAULT_ERROR)
2714                         ret &= VM_FAULT_ERROR;
2715                 goto out;
2716         }
2717
2718         /* No need to invalidate - it was non-present before */
2719         update_mmu_cache(vma, address, page_table);
2720 unlock:
2721         pte_unmap_unlock(page_table, ptl);
2722 out:
2723         return ret;
2724 out_nomap:
2725         mem_cgroup_cancel_charge_swapin(ptr);
2726         pte_unmap_unlock(page_table, ptl);
2727 out_page:
2728         unlock_page(page);
2729 out_release:
2730         page_cache_release(page);
2731         return ret;
2732 }
2733
2734 /*
2735  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2736  * but allow concurrent faults), and pte mapped but not yet locked.
2737  * We return with mmap_sem still held, but pte unmapped and unlocked.
2738  */
2739 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2740                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2741                 unsigned int flags)
2742 {
2743         struct page *page;
2744         spinlock_t *ptl;
2745         pte_t entry;
2746
2747         if (!(flags & FAULT_FLAG_WRITE)) {
2748                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2749                                                 vma->vm_page_prot));
2750                 ptl = pte_lockptr(mm, pmd);
2751                 spin_lock(ptl);
2752                 if (!pte_none(*page_table))
2753                         goto unlock;
2754                 goto setpte;
2755         }
2756
2757         /* Allocate our own private page. */
2758         pte_unmap(page_table);
2759
2760         if (unlikely(anon_vma_prepare(vma)))
2761                 goto oom;
2762         page = alloc_zeroed_user_highpage_movable(vma, address);
2763         if (!page)
2764                 goto oom;
2765         __SetPageUptodate(page);
2766
2767         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2768                 goto oom_free_page;
2769
2770         entry = mk_pte(page, vma->vm_page_prot);
2771         if (vma->vm_flags & VM_WRITE)
2772                 entry = pte_mkwrite(pte_mkdirty(entry));
2773
2774         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2775         if (!pte_none(*page_table))
2776                 goto release;
2777
2778         inc_mm_counter_fast(mm, MM_ANONPAGES);
2779         page_add_new_anon_rmap(page, vma, address);
2780 setpte:
2781         set_pte_at(mm, address, page_table, entry);
2782
2783         /* No need to invalidate - it was non-present before */
2784         update_mmu_cache(vma, address, page_table);
2785 unlock:
2786         pte_unmap_unlock(page_table, ptl);
2787         return 0;
2788 release:
2789         mem_cgroup_uncharge_page(page);
2790         page_cache_release(page);
2791         goto unlock;
2792 oom_free_page:
2793         page_cache_release(page);
2794 oom:
2795         return VM_FAULT_OOM;
2796 }
2797
2798 /*
2799  * __do_fault() tries to create a new page mapping. It aggressively
2800  * tries to share with existing pages, but makes a separate copy if
2801  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2802  * the next page fault.
2803  *
2804  * As this is called only for pages that do not currently exist, we
2805  * do not need to flush old virtual caches or the TLB.
2806  *
2807  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2808  * but allow concurrent faults), and pte neither mapped nor locked.
2809  * We return with mmap_sem still held, but pte unmapped and unlocked.
2810  */
2811 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2812                 unsigned long address, pmd_t *pmd,
2813                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2814 {
2815         pte_t *page_table;
2816         spinlock_t *ptl;
2817         struct page *page;
2818         pte_t entry;
2819         int anon = 0;
2820         int charged = 0;
2821         struct page *dirty_page = NULL;
2822         struct vm_fault vmf;
2823         int ret;
2824         int page_mkwrite = 0;
2825
2826         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2827         vmf.pgoff = pgoff;
2828         vmf.flags = flags;
2829         vmf.page = NULL;
2830
2831         ret = vma->vm_ops->fault(vma, &vmf);
2832         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2833                 return ret;
2834
2835         if (unlikely(PageHWPoison(vmf.page))) {
2836                 if (ret & VM_FAULT_LOCKED)
2837                         unlock_page(vmf.page);
2838                 return VM_FAULT_HWPOISON;
2839         }
2840
2841         /*
2842          * For consistency in subsequent calls, make the faulted page always
2843          * locked.
2844          */
2845         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2846                 lock_page(vmf.page);
2847         else
2848                 VM_BUG_ON(!PageLocked(vmf.page));
2849
2850         /*
2851          * Should we do an early C-O-W break?
2852          */
2853         page = vmf.page;
2854         if (flags & FAULT_FLAG_WRITE) {
2855                 if (!(vma->vm_flags & VM_SHARED)) {
2856                         anon = 1;
2857                         if (unlikely(anon_vma_prepare(vma))) {
2858                                 ret = VM_FAULT_OOM;
2859                                 goto out;
2860                         }
2861                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2862                                                 vma, address);
2863                         if (!page) {
2864                                 ret = VM_FAULT_OOM;
2865                                 goto out;
2866                         }
2867                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2868                                 ret = VM_FAULT_OOM;
2869                                 page_cache_release(page);
2870                                 goto out;
2871                         }
2872                         charged = 1;
2873                         /*
2874                          * Don't let another task, with possibly unlocked vma,
2875                          * keep the mlocked page.
2876                          */
2877                         if (vma->vm_flags & VM_LOCKED)
2878                                 clear_page_mlock(vmf.page);
2879                         copy_user_highpage(page, vmf.page, address, vma);
2880                         __SetPageUptodate(page);
2881                 } else {
2882                         /*
2883                          * If the page will be shareable, see if the backing
2884                          * address space wants to know that the page is about
2885                          * to become writable
2886                          */
2887                         if (vma->vm_ops->page_mkwrite) {
2888                                 int tmp;
2889
2890                                 unlock_page(page);
2891                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2892                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2893                                 if (unlikely(tmp &
2894                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2895                                         ret = tmp;
2896                                         goto unwritable_page;
2897                                 }
2898                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2899                                         lock_page(page);
2900                                         if (!page->mapping) {
2901                                                 ret = 0; /* retry the fault */
2902                                                 unlock_page(page);
2903                                                 goto unwritable_page;
2904                                         }
2905                                 } else
2906                                         VM_BUG_ON(!PageLocked(page));
2907                                 page_mkwrite = 1;
2908                         }
2909                 }
2910
2911         }
2912
2913         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2914
2915         /*
2916          * This silly early PAGE_DIRTY setting removes a race
2917          * due to the bad i386 page protection. But it's valid
2918          * for other architectures too.
2919          *
2920          * Note that if FAULT_FLAG_WRITE is set, we either now have
2921          * an exclusive copy of the page, or this is a shared mapping,
2922          * so we can make it writable and dirty to avoid having to
2923          * handle that later.
2924          */
2925         /* Only go through if we didn't race with anybody else... */
2926         if (likely(pte_same(*page_table, orig_pte))) {
2927                 flush_icache_page(vma, page);
2928                 entry = mk_pte(page, vma->vm_page_prot);
2929                 if (flags & FAULT_FLAG_WRITE)
2930                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2931                 if (anon) {
2932                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2933                         page_add_new_anon_rmap(page, vma, address);
2934                 } else {
2935                         inc_mm_counter_fast(mm, MM_FILEPAGES);
2936                         page_add_file_rmap(page);
2937                         if (flags & FAULT_FLAG_WRITE) {
2938                                 dirty_page = page;
2939                                 get_page(dirty_page);
2940                         }
2941                 }
2942                 set_pte_at(mm, address, page_table, entry);
2943
2944                 /* no need to invalidate: a not-present page won't be cached */
2945                 update_mmu_cache(vma, address, page_table);
2946         } else {
2947                 if (charged)
2948                         mem_cgroup_uncharge_page(page);
2949                 if (anon)
2950                         page_cache_release(page);
2951                 else
2952                         anon = 1; /* no anon but release faulted_page */
2953         }
2954
2955         pte_unmap_unlock(page_table, ptl);
2956
2957 out:
2958         if (dirty_page) {
2959                 struct address_space *mapping = page->mapping;
2960
2961                 if (set_page_dirty(dirty_page))
2962                         page_mkwrite = 1;
2963                 unlock_page(dirty_page);
2964                 put_page(dirty_page);
2965                 if (page_mkwrite && mapping) {
2966                         /*
2967                          * Some device drivers do not set page.mapping but still
2968                          * dirty their pages
2969                          */
2970                         balance_dirty_pages_ratelimited(mapping);
2971                 }
2972
2973                 /* file_update_time outside page_lock */
2974                 if (vma->vm_file)
2975                         file_update_time(vma->vm_file);
2976         } else {
2977                 unlock_page(vmf.page);
2978                 if (anon)
2979                         page_cache_release(vmf.page);
2980         }
2981
2982         return ret;
2983
2984 unwritable_page:
2985         page_cache_release(page);
2986         return ret;
2987 }
2988
2989 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2990                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2991                 unsigned int flags, pte_t orig_pte)
2992 {
2993         pgoff_t pgoff = (((address & PAGE_MASK)
2994                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2995
2996         pte_unmap(page_table);
2997         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2998 }
2999
3000 /*
3001  * Fault of a previously existing named mapping. Repopulate the pte
3002  * from the encoded file_pte if possible. This enables swappable
3003  * nonlinear vmas.
3004  *
3005  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3006  * but allow concurrent faults), and pte mapped but not yet locked.
3007  * We return with mmap_sem still held, but pte unmapped and unlocked.
3008  */
3009 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3010                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3011                 unsigned int flags, pte_t orig_pte)
3012 {
3013         pgoff_t pgoff;
3014
3015         flags |= FAULT_FLAG_NONLINEAR;
3016
3017         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3018                 return 0;
3019
3020         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3021                 /*
3022                  * Page table corrupted: show pte and kill process.
3023                  */
3024                 print_bad_pte(vma, address, orig_pte, NULL);
3025                 return VM_FAULT_SIGBUS;
3026         }
3027
3028         pgoff = pte_to_pgoff(orig_pte);
3029         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3030 }
3031
3032 /*
3033  * These routines also need to handle stuff like marking pages dirty
3034  * and/or accessed for architectures that don't do it in hardware (most
3035  * RISC architectures).  The early dirtying is also good on the i386.
3036  *
3037  * There is also a hook called "update_mmu_cache()" that architectures
3038  * with external mmu caches can use to update those (ie the Sparc or
3039  * PowerPC hashed page tables that act as extended TLBs).
3040  *
3041  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3042  * but allow concurrent faults), and pte mapped but not yet locked.
3043  * We return with mmap_sem still held, but pte unmapped and unlocked.
3044  */
3045 static inline int handle_pte_fault(struct mm_struct *mm,
3046                 struct vm_area_struct *vma, unsigned long address,
3047                 pte_t *pte, pmd_t *pmd, unsigned int flags)
3048 {
3049         pte_t entry;
3050         spinlock_t *ptl;
3051
3052         entry = *pte;
3053         if (!pte_present(entry)) {
3054                 if (pte_none(entry)) {
3055                         if (vma->vm_ops) {
3056                                 if (likely(vma->vm_ops->fault))
3057                                         return do_linear_fault(mm, vma, address,
3058                                                 pte, pmd, flags, entry);
3059                         }
3060                         return do_anonymous_page(mm, vma, address,
3061                                                  pte, pmd, flags);
3062                 }
3063                 if (pte_file(entry))
3064                         return do_nonlinear_fault(mm, vma, address,
3065                                         pte, pmd, flags, entry);
3066                 return do_swap_page(mm, vma, address,
3067                                         pte, pmd, flags, entry);
3068         }
3069
3070         ptl = pte_lockptr(mm, pmd);
3071         spin_lock(ptl);
3072         if (unlikely(!pte_same(*pte, entry)))
3073                 goto unlock;
3074         if (flags & FAULT_FLAG_WRITE) {
3075                 if (!pte_write(entry))
3076                         return do_wp_page(mm, vma, address,
3077                                         pte, pmd, ptl, entry);
3078                 entry = pte_mkdirty(entry);
3079         }
3080         entry = pte_mkyoung(entry);
3081         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3082                 update_mmu_cache(vma, address, pte);
3083         } else {
3084                 /*
3085                  * This is needed only for protection faults but the arch code
3086                  * is not yet telling us if this is a protection fault or not.
3087                  * This still avoids useless tlb flushes for .text page faults
3088                  * with threads.
3089                  */
3090                 if (flags & FAULT_FLAG_WRITE)
3091                         flush_tlb_page(vma, address);
3092         }
3093 unlock:
3094         pte_unmap_unlock(pte, ptl);
3095         return 0;
3096 }
3097
3098 /*
3099  * By the time we get here, we already hold the mm semaphore
3100  */
3101 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3102                 unsigned long address, unsigned int flags)
3103 {
3104         pgd_t *pgd;
3105         pud_t *pud;
3106         pmd_t *pmd;
3107         pte_t *pte;
3108
3109         __set_current_state(TASK_RUNNING);
3110
3111         count_vm_event(PGFAULT);
3112
3113         /* do counter updates before entering really critical section. */
3114         check_sync_rss_stat(current);
3115
3116         if (unlikely(is_vm_hugetlb_page(vma)))
3117                 return hugetlb_fault(mm, vma, address, flags);
3118
3119         pgd = pgd_offset(mm, address);
3120         pud = pud_alloc(mm, pgd, address);
3121         if (!pud)
3122                 return VM_FAULT_OOM;
3123         pmd = pmd_alloc(mm, pud, address);
3124         if (!pmd)
3125                 return VM_FAULT_OOM;
3126         pte = pte_alloc_map(mm, pmd, address);
3127         if (!pte)
3128                 return VM_FAULT_OOM;
3129
3130         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3131 }
3132
3133 #ifndef __PAGETABLE_PUD_FOLDED
3134 /*
3135  * Allocate page upper directory.
3136  * We've already handled the fast-path in-line.
3137  */
3138 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3139 {
3140         pud_t *new = pud_alloc_one(mm, address);
3141         if (!new)
3142                 return -ENOMEM;
3143
3144         smp_wmb(); /* See comment in __pte_alloc */
3145
3146         spin_lock(&mm->page_table_lock);
3147         if (pgd_present(*pgd))          /* Another has populated it */
3148                 pud_free(mm, new);
3149         else
3150                 pgd_populate(mm, pgd, new);
3151         spin_unlock(&mm->page_table_lock);
3152         return 0;
3153 }
3154 #endif /* __PAGETABLE_PUD_FOLDED */
3155
3156 #ifndef __PAGETABLE_PMD_FOLDED
3157 /*
3158  * Allocate page middle directory.
3159  * We've already handled the fast-path in-line.
3160  */
3161 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3162 {
3163         pmd_t *new = pmd_alloc_one(mm, address);
3164         if (!new)
3165                 return -ENOMEM;
3166
3167         smp_wmb(); /* See comment in __pte_alloc */
3168
3169         spin_lock(&mm->page_table_lock);
3170 #ifndef __ARCH_HAS_4LEVEL_HACK
3171         if (pud_present(*pud))          /* Another has populated it */
3172                 pmd_free(mm, new);
3173         else
3174                 pud_populate(mm, pud, new);
3175 #else
3176         if (pgd_present(*pud))          /* Another has populated it */
3177                 pmd_free(mm, new);
3178         else
3179                 pgd_populate(mm, pud, new);
3180 #endif /* __ARCH_HAS_4LEVEL_HACK */
3181         spin_unlock(&mm->page_table_lock);
3182         return 0;
3183 }
3184 #endif /* __PAGETABLE_PMD_FOLDED */
3185
3186 int make_pages_present(unsigned long addr, unsigned long end)
3187 {
3188         int ret, len, write;
3189         struct vm_area_struct * vma;
3190
3191         vma = find_vma(current->mm, addr);
3192         if (!vma)
3193                 return -ENOMEM;
3194         write = (vma->vm_flags & VM_WRITE) != 0;
3195         BUG_ON(addr >= end);
3196         BUG_ON(end > vma->vm_end);
3197         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3198         ret = get_user_pages(current, current->mm, addr,
3199                         len, write, 0, NULL, NULL);
3200         if (ret < 0)
3201                 return ret;
3202         return ret == len ? 0 : -EFAULT;
3203 }
3204
3205 #if !defined(__HAVE_ARCH_GATE_AREA)
3206
3207 #if defined(AT_SYSINFO_EHDR)
3208 static struct vm_area_struct gate_vma;
3209
3210 static int __init gate_vma_init(void)
3211 {
3212         gate_vma.vm_mm = NULL;
3213         gate_vma.vm_start = FIXADDR_USER_START;
3214         gate_vma.vm_end = FIXADDR_USER_END;
3215         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3216         gate_vma.vm_page_prot = __P101;
3217         /*
3218          * Make sure the vDSO gets into every core dump.
3219          * Dumping its contents makes post-mortem fully interpretable later
3220          * without matching up the same kernel and hardware config to see
3221          * what PC values meant.
3222          */
3223         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3224         return 0;
3225 }
3226 __initcall(gate_vma_init);
3227 #endif
3228
3229 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3230 {
3231 #ifdef AT_SYSINFO_EHDR
3232         return &gate_vma;
3233 #else
3234         return NULL;
3235 #endif
3236 }
3237
3238 int in_gate_area_no_task(unsigned long addr)
3239 {
3240 #ifdef AT_SYSINFO_EHDR
3241         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3242                 return 1;
3243 #endif
3244         return 0;
3245 }
3246
3247 #endif  /* __HAVE_ARCH_GATE_AREA */
3248
3249 static int follow_pte(struct mm_struct *mm, unsigned long address,
3250                 pte_t **ptepp, spinlock_t **ptlp)
3251 {
3252         pgd_t *pgd;
3253         pud_t *pud;
3254         pmd_t *pmd;
3255         pte_t *ptep;
3256
3257         pgd = pgd_offset(mm, address);
3258         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3259                 goto out;
3260
3261         pud = pud_offset(pgd, address);
3262         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3263                 goto out;
3264
3265         pmd = pmd_offset(pud, address);
3266         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3267                 goto out;
3268
3269         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3270         if (pmd_huge(*pmd))
3271                 goto out;
3272
3273         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3274         if (!ptep)
3275                 goto out;
3276         if (!pte_present(*ptep))
3277                 goto unlock;
3278         *ptepp = ptep;
3279         return 0;
3280 unlock:
3281         pte_unmap_unlock(ptep, *ptlp);
3282 out:
3283         return -EINVAL;
3284 }
3285
3286 /**
3287  * follow_pfn - look up PFN at a user virtual address
3288  * @vma: memory mapping
3289  * @address: user virtual address
3290  * @pfn: location to store found PFN
3291  *
3292  * Only IO mappings and raw PFN mappings are allowed.
3293  *
3294  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3295  */
3296 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3297         unsigned long *pfn)
3298 {
3299         int ret = -EINVAL;
3300         spinlock_t *ptl;
3301         pte_t *ptep;
3302
3303         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3304                 return ret;
3305
3306         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3307         if (ret)
3308                 return ret;
3309         *pfn = pte_pfn(*ptep);
3310         pte_unmap_unlock(ptep, ptl);
3311         return 0;
3312 }
3313 EXPORT_SYMBOL(follow_pfn);
3314
3315 #ifdef CONFIG_HAVE_IOREMAP_PROT
3316 int follow_phys(struct vm_area_struct *vma,
3317                 unsigned long address, unsigned int flags,
3318                 unsigned long *prot, resource_size_t *phys)
3319 {
3320         int ret = -EINVAL;
3321         pte_t *ptep, pte;
3322         spinlock_t *ptl;
3323
3324         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3325                 goto out;
3326
3327         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3328                 goto out;
3329         pte = *ptep;
3330
3331         if ((flags & FOLL_WRITE) && !pte_write(pte))
3332                 goto unlock;
3333
3334         *prot = pgprot_val(pte_pgprot(pte));
3335         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3336
3337         ret = 0;
3338 unlock:
3339         pte_unmap_unlock(ptep, ptl);
3340 out:
3341         return ret;
3342 }
3343
3344 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3345                         void *buf, int len, int write)
3346 {
3347         resource_size_t phys_addr;
3348         unsigned long prot = 0;
3349         void __iomem *maddr;
3350         int offset = addr & (PAGE_SIZE-1);
3351
3352         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3353                 return -EINVAL;
3354
3355         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3356         if (write)
3357                 memcpy_toio(maddr + offset, buf, len);
3358         else
3359                 memcpy_fromio(buf, maddr + offset, len);
3360         iounmap(maddr);
3361
3362         return len;
3363 }
3364 #endif
3365
3366 /*
3367  * Access another process' address space.
3368  * Source/target buffer must be kernel space,
3369  * Do not walk the page table directly, use get_user_pages
3370  */
3371 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3372 {
3373         struct mm_struct *mm;
3374         struct vm_area_struct *vma;
3375         void *old_buf = buf;
3376
3377         mm = get_task_mm(tsk);
3378         if (!mm)
3379                 return 0;
3380
3381         down_read(&mm->mmap_sem);
3382         /* ignore errors, just check how much was successfully transferred */
3383         while (len) {
3384                 int bytes, ret, offset;
3385                 void *maddr;
3386                 struct page *page = NULL;
3387
3388                 ret = get_user_pages(tsk, mm, addr, 1,
3389                                 write, 1, &page, &vma);
3390                 if (ret <= 0) {
3391                         /*
3392                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3393                          * we can access using slightly different code.
3394                          */
3395 #ifdef CONFIG_HAVE_IOREMAP_PROT
3396                         vma = find_vma(mm, addr);
3397                         if (!vma)
3398                                 break;
3399                         if (vma->vm_ops && vma->vm_ops->access)
3400                                 ret = vma->vm_ops->access(vma, addr, buf,
3401                                                           len, write);
3402                         if (ret <= 0)
3403 #endif
3404                                 break;
3405                         bytes = ret;
3406                 } else {
3407                         bytes = len;
3408                         offset = addr & (PAGE_SIZE-1);
3409                         if (bytes > PAGE_SIZE-offset)
3410                                 bytes = PAGE_SIZE-offset;
3411
3412                         maddr = kmap(page);
3413                         if (write) {
3414                                 copy_to_user_page(vma, page, addr,
3415                                                   maddr + offset, buf, bytes);
3416                                 set_page_dirty_lock(page);
3417                         } else {
3418                                 copy_from_user_page(vma, page, addr,
3419                                                     buf, maddr + offset, bytes);
3420                         }
3421                         kunmap(page);
3422                         page_cache_release(page);
3423                 }
3424                 len -= bytes;
3425                 buf += bytes;
3426                 addr += bytes;
3427         }
3428         up_read(&mm->mmap_sem);
3429         mmput(mm);
3430
3431         return buf - old_buf;
3432 }
3433
3434 /*
3435  * Print the name of a VMA.
3436  */
3437 void print_vma_addr(char *prefix, unsigned long ip)
3438 {
3439         struct mm_struct *mm = current->mm;
3440         struct vm_area_struct *vma;
3441
3442         /*
3443          * Do not print if we are in atomic
3444          * contexts (in exception stacks, etc.):
3445          */
3446         if (preempt_count())
3447                 return;
3448
3449         down_read(&mm->mmap_sem);
3450         vma = find_vma(mm, ip);
3451         if (vma && vma->vm_file) {
3452                 struct file *f = vma->vm_file;
3453                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3454                 if (buf) {
3455                         char *p, *s;
3456
3457                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3458                         if (IS_ERR(p))
3459                                 p = "?";
3460                         s = strrchr(p, '/');
3461                         if (s)
3462                                 p = s+1;
3463                         printk("%s%s[%lx+%lx]", prefix, p,
3464                                         vma->vm_start,
3465                                         vma->vm_end - vma->vm_start);
3466                         free_page((unsigned long)buf);
3467                 }
3468         }
3469         up_read(&current->mm->mmap_sem);
3470 }
3471
3472 #ifdef CONFIG_PROVE_LOCKING
3473 void might_fault(void)
3474 {
3475         /*
3476          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3477          * holding the mmap_sem, this is safe because kernel memory doesn't
3478          * get paged out, therefore we'll never actually fault, and the
3479          * below annotations will generate false positives.
3480          */
3481         if (segment_eq(get_fs(), KERNEL_DS))
3482                 return;
3483
3484         might_sleep();
3485         /*
3486          * it would be nicer only to annotate paths which are not under
3487          * pagefault_disable, however that requires a larger audit and
3488          * providing helpers like get_user_atomic.
3489          */
3490         if (!in_atomic() && current->mm)
3491                 might_lock_read(&current->mm->mmap_sem);
3492 }
3493 EXPORT_SYMBOL(might_fault);
3494 #endif