5b7f2002e54b17e8d9b67e75cff0aaa06c9296b4
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59
60 #include <asm/io.h>
61 #include <asm/pgalloc.h>
62 #include <asm/uaccess.h>
63 #include <asm/tlb.h>
64 #include <asm/tlbflush.h>
65 #include <asm/pgtable.h>
66
67 #include "internal.h"
68
69 #ifndef CONFIG_NEED_MULTIPLE_NODES
70 /* use the per-pgdat data instead for discontigmem - mbligh */
71 unsigned long max_mapnr;
72 struct page *mem_map;
73
74 EXPORT_SYMBOL(max_mapnr);
75 EXPORT_SYMBOL(mem_map);
76 #endif
77
78 unsigned long num_physpages;
79 /*
80  * A number of key systems in x86 including ioremap() rely on the assumption
81  * that high_memory defines the upper bound on direct map memory, then end
82  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
83  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
84  * and ZONE_HIGHMEM.
85  */
86 void * high_memory;
87
88 EXPORT_SYMBOL(num_physpages);
89 EXPORT_SYMBOL(high_memory);
90
91 /*
92  * Randomize the address space (stacks, mmaps, brk, etc.).
93  *
94  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
95  *   as ancient (libc5 based) binaries can segfault. )
96  */
97 int randomize_va_space __read_mostly =
98 #ifdef CONFIG_COMPAT_BRK
99                                         1;
100 #else
101                                         2;
102 #endif
103
104 static int __init disable_randmaps(char *s)
105 {
106         randomize_va_space = 0;
107         return 1;
108 }
109 __setup("norandmaps", disable_randmaps);
110
111 unsigned long zero_pfn __read_mostly;
112 unsigned long highest_memmap_pfn __read_mostly;
113
114 /*
115  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
116  */
117 static int __init init_zero_pfn(void)
118 {
119         zero_pfn = page_to_pfn(ZERO_PAGE(0));
120         return 0;
121 }
122 core_initcall(init_zero_pfn);
123
124
125 #if defined(SPLIT_RSS_COUNTING)
126
127 void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
128 {
129         int i;
130
131         for (i = 0; i < NR_MM_COUNTERS; i++) {
132                 if (task->rss_stat.count[i]) {
133                         add_mm_counter(mm, i, task->rss_stat.count[i]);
134                         task->rss_stat.count[i] = 0;
135                 }
136         }
137         task->rss_stat.events = 0;
138 }
139
140 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
141 {
142         struct task_struct *task = current;
143
144         if (likely(task->mm == mm))
145                 task->rss_stat.count[member] += val;
146         else
147                 add_mm_counter(mm, member, val);
148 }
149 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
150 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
151
152 /* sync counter once per 64 page faults */
153 #define TASK_RSS_EVENTS_THRESH  (64)
154 static void check_sync_rss_stat(struct task_struct *task)
155 {
156         if (unlikely(task != current))
157                 return;
158         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
159                 __sync_task_rss_stat(task, task->mm);
160 }
161
162 unsigned long get_mm_counter(struct mm_struct *mm, int member)
163 {
164         long val = 0;
165
166         /*
167          * Don't use task->mm here...for avoiding to use task_get_mm()..
168          * The caller must guarantee task->mm is not invalid.
169          */
170         val = atomic_long_read(&mm->rss_stat.count[member]);
171         /*
172          * counter is updated in asynchronous manner and may go to minus.
173          * But it's never be expected number for users.
174          */
175         if (val < 0)
176                 return 0;
177         return (unsigned long)val;
178 }
179
180 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
181 {
182         __sync_task_rss_stat(task, mm);
183 }
184 #else
185
186 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
187 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
188
189 static void check_sync_rss_stat(struct task_struct *task)
190 {
191 }
192
193 #endif
194
195 /*
196  * If a p?d_bad entry is found while walking page tables, report
197  * the error, before resetting entry to p?d_none.  Usually (but
198  * very seldom) called out from the p?d_none_or_clear_bad macros.
199  */
200
201 void pgd_clear_bad(pgd_t *pgd)
202 {
203         pgd_ERROR(*pgd);
204         pgd_clear(pgd);
205 }
206
207 void pud_clear_bad(pud_t *pud)
208 {
209         pud_ERROR(*pud);
210         pud_clear(pud);
211 }
212
213 void pmd_clear_bad(pmd_t *pmd)
214 {
215         pmd_ERROR(*pmd);
216         pmd_clear(pmd);
217 }
218
219 /*
220  * Note: this doesn't free the actual pages themselves. That
221  * has been handled earlier when unmapping all the memory regions.
222  */
223 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
224                            unsigned long addr)
225 {
226         pgtable_t token = pmd_pgtable(*pmd);
227         pmd_clear(pmd);
228         pte_free_tlb(tlb, token, addr);
229         tlb->mm->nr_ptes--;
230 }
231
232 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
233                                 unsigned long addr, unsigned long end,
234                                 unsigned long floor, unsigned long ceiling)
235 {
236         pmd_t *pmd;
237         unsigned long next;
238         unsigned long start;
239
240         start = addr;
241         pmd = pmd_offset(pud, addr);
242         do {
243                 next = pmd_addr_end(addr, end);
244                 if (pmd_none_or_clear_bad(pmd))
245                         continue;
246                 free_pte_range(tlb, pmd, addr);
247         } while (pmd++, addr = next, addr != end);
248
249         start &= PUD_MASK;
250         if (start < floor)
251                 return;
252         if (ceiling) {
253                 ceiling &= PUD_MASK;
254                 if (!ceiling)
255                         return;
256         }
257         if (end - 1 > ceiling - 1)
258                 return;
259
260         pmd = pmd_offset(pud, start);
261         pud_clear(pud);
262         pmd_free_tlb(tlb, pmd, start);
263 }
264
265 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
266                                 unsigned long addr, unsigned long end,
267                                 unsigned long floor, unsigned long ceiling)
268 {
269         pud_t *pud;
270         unsigned long next;
271         unsigned long start;
272
273         start = addr;
274         pud = pud_offset(pgd, addr);
275         do {
276                 next = pud_addr_end(addr, end);
277                 if (pud_none_or_clear_bad(pud))
278                         continue;
279                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
280         } while (pud++, addr = next, addr != end);
281
282         start &= PGDIR_MASK;
283         if (start < floor)
284                 return;
285         if (ceiling) {
286                 ceiling &= PGDIR_MASK;
287                 if (!ceiling)
288                         return;
289         }
290         if (end - 1 > ceiling - 1)
291                 return;
292
293         pud = pud_offset(pgd, start);
294         pgd_clear(pgd);
295         pud_free_tlb(tlb, pud, start);
296 }
297
298 /*
299  * This function frees user-level page tables of a process.
300  *
301  * Must be called with pagetable lock held.
302  */
303 void free_pgd_range(struct mmu_gather *tlb,
304                         unsigned long addr, unsigned long end,
305                         unsigned long floor, unsigned long ceiling)
306 {
307         pgd_t *pgd;
308         unsigned long next;
309         unsigned long start;
310
311         /*
312          * The next few lines have given us lots of grief...
313          *
314          * Why are we testing PMD* at this top level?  Because often
315          * there will be no work to do at all, and we'd prefer not to
316          * go all the way down to the bottom just to discover that.
317          *
318          * Why all these "- 1"s?  Because 0 represents both the bottom
319          * of the address space and the top of it (using -1 for the
320          * top wouldn't help much: the masks would do the wrong thing).
321          * The rule is that addr 0 and floor 0 refer to the bottom of
322          * the address space, but end 0 and ceiling 0 refer to the top
323          * Comparisons need to use "end - 1" and "ceiling - 1" (though
324          * that end 0 case should be mythical).
325          *
326          * Wherever addr is brought up or ceiling brought down, we must
327          * be careful to reject "the opposite 0" before it confuses the
328          * subsequent tests.  But what about where end is brought down
329          * by PMD_SIZE below? no, end can't go down to 0 there.
330          *
331          * Whereas we round start (addr) and ceiling down, by different
332          * masks at different levels, in order to test whether a table
333          * now has no other vmas using it, so can be freed, we don't
334          * bother to round floor or end up - the tests don't need that.
335          */
336
337         addr &= PMD_MASK;
338         if (addr < floor) {
339                 addr += PMD_SIZE;
340                 if (!addr)
341                         return;
342         }
343         if (ceiling) {
344                 ceiling &= PMD_MASK;
345                 if (!ceiling)
346                         return;
347         }
348         if (end - 1 > ceiling - 1)
349                 end -= PMD_SIZE;
350         if (addr > end - 1)
351                 return;
352
353         start = addr;
354         pgd = pgd_offset(tlb->mm, addr);
355         do {
356                 next = pgd_addr_end(addr, end);
357                 if (pgd_none_or_clear_bad(pgd))
358                         continue;
359                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
360         } while (pgd++, addr = next, addr != end);
361 }
362
363 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
364                 unsigned long floor, unsigned long ceiling)
365 {
366         while (vma) {
367                 struct vm_area_struct *next = vma->vm_next;
368                 unsigned long addr = vma->vm_start;
369
370                 /*
371                  * Hide vma from rmap and truncate_pagecache before freeing
372                  * pgtables
373                  */
374                 unlink_anon_vmas(vma);
375                 unlink_file_vma(vma);
376
377                 if (is_vm_hugetlb_page(vma)) {
378                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
379                                 floor, next? next->vm_start: ceiling);
380                 } else {
381                         /*
382                          * Optimization: gather nearby vmas into one call down
383                          */
384                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
385                                && !is_vm_hugetlb_page(next)) {
386                                 vma = next;
387                                 next = vma->vm_next;
388                                 unlink_anon_vmas(vma);
389                                 unlink_file_vma(vma);
390                         }
391                         free_pgd_range(tlb, addr, vma->vm_end,
392                                 floor, next? next->vm_start: ceiling);
393                 }
394                 vma = next;
395         }
396 }
397
398 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
399 {
400         pgtable_t new = pte_alloc_one(mm, address);
401         if (!new)
402                 return -ENOMEM;
403
404         /*
405          * Ensure all pte setup (eg. pte page lock and page clearing) are
406          * visible before the pte is made visible to other CPUs by being
407          * put into page tables.
408          *
409          * The other side of the story is the pointer chasing in the page
410          * table walking code (when walking the page table without locking;
411          * ie. most of the time). Fortunately, these data accesses consist
412          * of a chain of data-dependent loads, meaning most CPUs (alpha
413          * being the notable exception) will already guarantee loads are
414          * seen in-order. See the alpha page table accessors for the
415          * smp_read_barrier_depends() barriers in page table walking code.
416          */
417         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
418
419         spin_lock(&mm->page_table_lock);
420         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
421                 mm->nr_ptes++;
422                 pmd_populate(mm, pmd, new);
423                 new = NULL;
424         }
425         spin_unlock(&mm->page_table_lock);
426         if (new)
427                 pte_free(mm, new);
428         return 0;
429 }
430
431 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
432 {
433         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
434         if (!new)
435                 return -ENOMEM;
436
437         smp_wmb(); /* See comment in __pte_alloc */
438
439         spin_lock(&init_mm.page_table_lock);
440         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
441                 pmd_populate_kernel(&init_mm, pmd, new);
442                 new = NULL;
443         }
444         spin_unlock(&init_mm.page_table_lock);
445         if (new)
446                 pte_free_kernel(&init_mm, new);
447         return 0;
448 }
449
450 static inline void init_rss_vec(int *rss)
451 {
452         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
453 }
454
455 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
456 {
457         int i;
458
459         if (current->mm == mm)
460                 sync_mm_rss(current, mm);
461         for (i = 0; i < NR_MM_COUNTERS; i++)
462                 if (rss[i])
463                         add_mm_counter(mm, i, rss[i]);
464 }
465
466 /*
467  * This function is called to print an error when a bad pte
468  * is found. For example, we might have a PFN-mapped pte in
469  * a region that doesn't allow it.
470  *
471  * The calling function must still handle the error.
472  */
473 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
474                           pte_t pte, struct page *page)
475 {
476         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
477         pud_t *pud = pud_offset(pgd, addr);
478         pmd_t *pmd = pmd_offset(pud, addr);
479         struct address_space *mapping;
480         pgoff_t index;
481         static unsigned long resume;
482         static unsigned long nr_shown;
483         static unsigned long nr_unshown;
484
485         /*
486          * Allow a burst of 60 reports, then keep quiet for that minute;
487          * or allow a steady drip of one report per second.
488          */
489         if (nr_shown == 60) {
490                 if (time_before(jiffies, resume)) {
491                         nr_unshown++;
492                         return;
493                 }
494                 if (nr_unshown) {
495                         printk(KERN_ALERT
496                                 "BUG: Bad page map: %lu messages suppressed\n",
497                                 nr_unshown);
498                         nr_unshown = 0;
499                 }
500                 nr_shown = 0;
501         }
502         if (nr_shown++ == 0)
503                 resume = jiffies + 60 * HZ;
504
505         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
506         index = linear_page_index(vma, addr);
507
508         printk(KERN_ALERT
509                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
510                 current->comm,
511                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
512         if (page)
513                 dump_page(page);
514         printk(KERN_ALERT
515                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
516                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
517         /*
518          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
519          */
520         if (vma->vm_ops)
521                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
522                                 (unsigned long)vma->vm_ops->fault);
523         if (vma->vm_file && vma->vm_file->f_op)
524                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
525                                 (unsigned long)vma->vm_file->f_op->mmap);
526         dump_stack();
527         add_taint(TAINT_BAD_PAGE);
528 }
529
530 static inline int is_cow_mapping(unsigned int flags)
531 {
532         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
533 }
534
535 #ifndef is_zero_pfn
536 static inline int is_zero_pfn(unsigned long pfn)
537 {
538         return pfn == zero_pfn;
539 }
540 #endif
541
542 #ifndef my_zero_pfn
543 static inline unsigned long my_zero_pfn(unsigned long addr)
544 {
545         return zero_pfn;
546 }
547 #endif
548
549 /*
550  * vm_normal_page -- This function gets the "struct page" associated with a pte.
551  *
552  * "Special" mappings do not wish to be associated with a "struct page" (either
553  * it doesn't exist, or it exists but they don't want to touch it). In this
554  * case, NULL is returned here. "Normal" mappings do have a struct page.
555  *
556  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
557  * pte bit, in which case this function is trivial. Secondly, an architecture
558  * may not have a spare pte bit, which requires a more complicated scheme,
559  * described below.
560  *
561  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
562  * special mapping (even if there are underlying and valid "struct pages").
563  * COWed pages of a VM_PFNMAP are always normal.
564  *
565  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
566  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
567  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
568  * mapping will always honor the rule
569  *
570  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
571  *
572  * And for normal mappings this is false.
573  *
574  * This restricts such mappings to be a linear translation from virtual address
575  * to pfn. To get around this restriction, we allow arbitrary mappings so long
576  * as the vma is not a COW mapping; in that case, we know that all ptes are
577  * special (because none can have been COWed).
578  *
579  *
580  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
581  *
582  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
583  * page" backing, however the difference is that _all_ pages with a struct
584  * page (that is, those where pfn_valid is true) are refcounted and considered
585  * normal pages by the VM. The disadvantage is that pages are refcounted
586  * (which can be slower and simply not an option for some PFNMAP users). The
587  * advantage is that we don't have to follow the strict linearity rule of
588  * PFNMAP mappings in order to support COWable mappings.
589  *
590  */
591 #ifdef __HAVE_ARCH_PTE_SPECIAL
592 # define HAVE_PTE_SPECIAL 1
593 #else
594 # define HAVE_PTE_SPECIAL 0
595 #endif
596 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
597                                 pte_t pte)
598 {
599         unsigned long pfn = pte_pfn(pte);
600
601         if (HAVE_PTE_SPECIAL) {
602                 if (likely(!pte_special(pte)))
603                         goto check_pfn;
604                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
605                         return NULL;
606                 if (!is_zero_pfn(pfn))
607                         print_bad_pte(vma, addr, pte, NULL);
608                 return NULL;
609         }
610
611         /* !HAVE_PTE_SPECIAL case follows: */
612
613         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
614                 if (vma->vm_flags & VM_MIXEDMAP) {
615                         if (!pfn_valid(pfn))
616                                 return NULL;
617                         goto out;
618                 } else {
619                         unsigned long off;
620                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
621                         if (pfn == vma->vm_pgoff + off)
622                                 return NULL;
623                         if (!is_cow_mapping(vma->vm_flags))
624                                 return NULL;
625                 }
626         }
627
628         if (is_zero_pfn(pfn))
629                 return NULL;
630 check_pfn:
631         if (unlikely(pfn > highest_memmap_pfn)) {
632                 print_bad_pte(vma, addr, pte, NULL);
633                 return NULL;
634         }
635
636         /*
637          * NOTE! We still have PageReserved() pages in the page tables.
638          * eg. VDSO mappings can cause them to exist.
639          */
640 out:
641         return pfn_to_page(pfn);
642 }
643
644 /*
645  * copy one vm_area from one task to the other. Assumes the page tables
646  * already present in the new task to be cleared in the whole range
647  * covered by this vma.
648  */
649
650 static inline unsigned long
651 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
652                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
653                 unsigned long addr, int *rss)
654 {
655         unsigned long vm_flags = vma->vm_flags;
656         pte_t pte = *src_pte;
657         struct page *page;
658
659         /* pte contains position in swap or file, so copy. */
660         if (unlikely(!pte_present(pte))) {
661                 if (!pte_file(pte)) {
662                         swp_entry_t entry = pte_to_swp_entry(pte);
663
664                         if (swap_duplicate(entry) < 0)
665                                 return entry.val;
666
667                         /* make sure dst_mm is on swapoff's mmlist. */
668                         if (unlikely(list_empty(&dst_mm->mmlist))) {
669                                 spin_lock(&mmlist_lock);
670                                 if (list_empty(&dst_mm->mmlist))
671                                         list_add(&dst_mm->mmlist,
672                                                  &src_mm->mmlist);
673                                 spin_unlock(&mmlist_lock);
674                         }
675                         if (likely(!non_swap_entry(entry)))
676                                 rss[MM_SWAPENTS]++;
677                         else if (is_write_migration_entry(entry) &&
678                                         is_cow_mapping(vm_flags)) {
679                                 /*
680                                  * COW mappings require pages in both parent
681                                  * and child to be set to read.
682                                  */
683                                 make_migration_entry_read(&entry);
684                                 pte = swp_entry_to_pte(entry);
685                                 set_pte_at(src_mm, addr, src_pte, pte);
686                         }
687                 }
688                 goto out_set_pte;
689         }
690
691         /*
692          * If it's a COW mapping, write protect it both
693          * in the parent and the child
694          */
695         if (is_cow_mapping(vm_flags)) {
696                 ptep_set_wrprotect(src_mm, addr, src_pte);
697                 pte = pte_wrprotect(pte);
698         }
699
700         /*
701          * If it's a shared mapping, mark it clean in
702          * the child
703          */
704         if (vm_flags & VM_SHARED)
705                 pte = pte_mkclean(pte);
706         pte = pte_mkold(pte);
707
708         page = vm_normal_page(vma, addr, pte);
709         if (page) {
710                 get_page(page);
711                 page_dup_rmap(page);
712                 if (PageAnon(page))
713                         rss[MM_ANONPAGES]++;
714                 else
715                         rss[MM_FILEPAGES]++;
716         }
717
718 out_set_pte:
719         set_pte_at(dst_mm, addr, dst_pte, pte);
720         return 0;
721 }
722
723 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
724                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
725                 unsigned long addr, unsigned long end)
726 {
727         pte_t *orig_src_pte, *orig_dst_pte;
728         pte_t *src_pte, *dst_pte;
729         spinlock_t *src_ptl, *dst_ptl;
730         int progress = 0;
731         int rss[NR_MM_COUNTERS];
732         swp_entry_t entry = (swp_entry_t){0};
733
734 again:
735         init_rss_vec(rss);
736
737         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
738         if (!dst_pte)
739                 return -ENOMEM;
740         src_pte = pte_offset_map_nested(src_pmd, addr);
741         src_ptl = pte_lockptr(src_mm, src_pmd);
742         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
743         orig_src_pte = src_pte;
744         orig_dst_pte = dst_pte;
745         arch_enter_lazy_mmu_mode();
746
747         do {
748                 /*
749                  * We are holding two locks at this point - either of them
750                  * could generate latencies in another task on another CPU.
751                  */
752                 if (progress >= 32) {
753                         progress = 0;
754                         if (need_resched() ||
755                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
756                                 break;
757                 }
758                 if (pte_none(*src_pte)) {
759                         progress++;
760                         continue;
761                 }
762                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
763                                                         vma, addr, rss);
764                 if (entry.val)
765                         break;
766                 progress += 8;
767         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
768
769         arch_leave_lazy_mmu_mode();
770         spin_unlock(src_ptl);
771         pte_unmap_nested(orig_src_pte);
772         add_mm_rss_vec(dst_mm, rss);
773         pte_unmap_unlock(orig_dst_pte, dst_ptl);
774         cond_resched();
775
776         if (entry.val) {
777                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
778                         return -ENOMEM;
779                 progress = 0;
780         }
781         if (addr != end)
782                 goto again;
783         return 0;
784 }
785
786 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
787                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
788                 unsigned long addr, unsigned long end)
789 {
790         pmd_t *src_pmd, *dst_pmd;
791         unsigned long next;
792
793         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
794         if (!dst_pmd)
795                 return -ENOMEM;
796         src_pmd = pmd_offset(src_pud, addr);
797         do {
798                 next = pmd_addr_end(addr, end);
799                 if (pmd_none_or_clear_bad(src_pmd))
800                         continue;
801                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
802                                                 vma, addr, next))
803                         return -ENOMEM;
804         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
805         return 0;
806 }
807
808 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
809                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
810                 unsigned long addr, unsigned long end)
811 {
812         pud_t *src_pud, *dst_pud;
813         unsigned long next;
814
815         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
816         if (!dst_pud)
817                 return -ENOMEM;
818         src_pud = pud_offset(src_pgd, addr);
819         do {
820                 next = pud_addr_end(addr, end);
821                 if (pud_none_or_clear_bad(src_pud))
822                         continue;
823                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
824                                                 vma, addr, next))
825                         return -ENOMEM;
826         } while (dst_pud++, src_pud++, addr = next, addr != end);
827         return 0;
828 }
829
830 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
831                 struct vm_area_struct *vma)
832 {
833         pgd_t *src_pgd, *dst_pgd;
834         unsigned long next;
835         unsigned long addr = vma->vm_start;
836         unsigned long end = vma->vm_end;
837         int ret;
838
839         /*
840          * Don't copy ptes where a page fault will fill them correctly.
841          * Fork becomes much lighter when there are big shared or private
842          * readonly mappings. The tradeoff is that copy_page_range is more
843          * efficient than faulting.
844          */
845         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
846                 if (!vma->anon_vma)
847                         return 0;
848         }
849
850         if (is_vm_hugetlb_page(vma))
851                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
852
853         if (unlikely(is_pfn_mapping(vma))) {
854                 /*
855                  * We do not free on error cases below as remove_vma
856                  * gets called on error from higher level routine
857                  */
858                 ret = track_pfn_vma_copy(vma);
859                 if (ret)
860                         return ret;
861         }
862
863         /*
864          * We need to invalidate the secondary MMU mappings only when
865          * there could be a permission downgrade on the ptes of the
866          * parent mm. And a permission downgrade will only happen if
867          * is_cow_mapping() returns true.
868          */
869         if (is_cow_mapping(vma->vm_flags))
870                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
871
872         ret = 0;
873         dst_pgd = pgd_offset(dst_mm, addr);
874         src_pgd = pgd_offset(src_mm, addr);
875         do {
876                 next = pgd_addr_end(addr, end);
877                 if (pgd_none_or_clear_bad(src_pgd))
878                         continue;
879                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
880                                             vma, addr, next))) {
881                         ret = -ENOMEM;
882                         break;
883                 }
884         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
885
886         if (is_cow_mapping(vma->vm_flags))
887                 mmu_notifier_invalidate_range_end(src_mm,
888                                                   vma->vm_start, end);
889         return ret;
890 }
891
892 static unsigned long zap_pte_range(struct mmu_gather *tlb,
893                                 struct vm_area_struct *vma, pmd_t *pmd,
894                                 unsigned long addr, unsigned long end,
895                                 long *zap_work, struct zap_details *details)
896 {
897         struct mm_struct *mm = tlb->mm;
898         pte_t *pte;
899         spinlock_t *ptl;
900         int rss[NR_MM_COUNTERS];
901
902         init_rss_vec(rss);
903
904         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
905         arch_enter_lazy_mmu_mode();
906         do {
907                 pte_t ptent = *pte;
908                 if (pte_none(ptent)) {
909                         (*zap_work)--;
910                         continue;
911                 }
912
913                 (*zap_work) -= PAGE_SIZE;
914
915                 if (pte_present(ptent)) {
916                         struct page *page;
917
918                         page = vm_normal_page(vma, addr, ptent);
919                         if (unlikely(details) && page) {
920                                 /*
921                                  * unmap_shared_mapping_pages() wants to
922                                  * invalidate cache without truncating:
923                                  * unmap shared but keep private pages.
924                                  */
925                                 if (details->check_mapping &&
926                                     details->check_mapping != page->mapping)
927                                         continue;
928                                 /*
929                                  * Each page->index must be checked when
930                                  * invalidating or truncating nonlinear.
931                                  */
932                                 if (details->nonlinear_vma &&
933                                     (page->index < details->first_index ||
934                                      page->index > details->last_index))
935                                         continue;
936                         }
937                         ptent = ptep_get_and_clear_full(mm, addr, pte,
938                                                         tlb->fullmm);
939                         tlb_remove_tlb_entry(tlb, pte, addr);
940                         if (unlikely(!page))
941                                 continue;
942                         if (unlikely(details) && details->nonlinear_vma
943                             && linear_page_index(details->nonlinear_vma,
944                                                 addr) != page->index)
945                                 set_pte_at(mm, addr, pte,
946                                            pgoff_to_pte(page->index));
947                         if (PageAnon(page))
948                                 rss[MM_ANONPAGES]--;
949                         else {
950                                 if (pte_dirty(ptent))
951                                         set_page_dirty(page);
952                                 if (pte_young(ptent) &&
953                                     likely(!VM_SequentialReadHint(vma)))
954                                         mark_page_accessed(page);
955                                 rss[MM_FILEPAGES]--;
956                         }
957                         page_remove_rmap(page);
958                         if (unlikely(page_mapcount(page) < 0))
959                                 print_bad_pte(vma, addr, ptent, page);
960                         tlb_remove_page(tlb, page);
961                         continue;
962                 }
963                 /*
964                  * If details->check_mapping, we leave swap entries;
965                  * if details->nonlinear_vma, we leave file entries.
966                  */
967                 if (unlikely(details))
968                         continue;
969                 if (pte_file(ptent)) {
970                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
971                                 print_bad_pte(vma, addr, ptent, NULL);
972                 } else {
973                         swp_entry_t entry = pte_to_swp_entry(ptent);
974
975                         if (!non_swap_entry(entry))
976                                 rss[MM_SWAPENTS]--;
977                         if (unlikely(!free_swap_and_cache(entry)))
978                                 print_bad_pte(vma, addr, ptent, NULL);
979                 }
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                 if (reuse)
2135                         /*
2136                          * The page is all ours.  Move it to our anon_vma so
2137                          * the rmap code will not search our parent or siblings.
2138                          * Protected against the rmap code by the page lock.
2139                          */
2140                         page_move_anon_rmap(old_page, vma, address);
2141                 unlock_page(old_page);
2142         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2143                                         (VM_WRITE|VM_SHARED))) {
2144                 /*
2145                  * Only catch write-faults on shared writable pages,
2146                  * read-only shared pages can get COWed by
2147                  * get_user_pages(.write=1, .force=1).
2148                  */
2149                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2150                         struct vm_fault vmf;
2151                         int tmp;
2152
2153                         vmf.virtual_address = (void __user *)(address &
2154                                                                 PAGE_MASK);
2155                         vmf.pgoff = old_page->index;
2156                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2157                         vmf.page = old_page;
2158
2159                         /*
2160                          * Notify the address space that the page is about to
2161                          * become writable so that it can prohibit this or wait
2162                          * for the page to get into an appropriate state.
2163                          *
2164                          * We do this without the lock held, so that it can
2165                          * sleep if it needs to.
2166                          */
2167                         page_cache_get(old_page);
2168                         pte_unmap_unlock(page_table, ptl);
2169
2170                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2171                         if (unlikely(tmp &
2172                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2173                                 ret = tmp;
2174                                 goto unwritable_page;
2175                         }
2176                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2177                                 lock_page(old_page);
2178                                 if (!old_page->mapping) {
2179                                         ret = 0; /* retry the fault */
2180                                         unlock_page(old_page);
2181                                         goto unwritable_page;
2182                                 }
2183                         } else
2184                                 VM_BUG_ON(!PageLocked(old_page));
2185
2186                         /*
2187                          * Since we dropped the lock we need to revalidate
2188                          * the PTE as someone else may have changed it.  If
2189                          * they did, we just return, as we can count on the
2190                          * MMU to tell us if they didn't also make it writable.
2191                          */
2192                         page_table = pte_offset_map_lock(mm, pmd, address,
2193                                                          &ptl);
2194                         if (!pte_same(*page_table, orig_pte)) {
2195                                 unlock_page(old_page);
2196                                 page_cache_release(old_page);
2197                                 goto unlock;
2198                         }
2199
2200                         page_mkwrite = 1;
2201                 }
2202                 dirty_page = old_page;
2203                 get_page(dirty_page);
2204                 reuse = 1;
2205         }
2206
2207         if (reuse) {
2208 reuse:
2209                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2210                 entry = pte_mkyoung(orig_pte);
2211                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2212                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2213                         update_mmu_cache(vma, address, page_table);
2214                 ret |= VM_FAULT_WRITE;
2215                 goto unlock;
2216         }
2217
2218         /*
2219          * Ok, we need to copy. Oh, well..
2220          */
2221         page_cache_get(old_page);
2222 gotten:
2223         pte_unmap_unlock(page_table, ptl);
2224
2225         if (unlikely(anon_vma_prepare(vma)))
2226                 goto oom;
2227
2228         if (is_zero_pfn(pte_pfn(orig_pte))) {
2229                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2230                 if (!new_page)
2231                         goto oom;
2232         } else {
2233                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2234                 if (!new_page)
2235                         goto oom;
2236                 cow_user_page(new_page, old_page, address, vma);
2237         }
2238         __SetPageUptodate(new_page);
2239
2240         /*
2241          * Don't let another task, with possibly unlocked vma,
2242          * keep the mlocked page.
2243          */
2244         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2245                 lock_page(old_page);    /* for LRU manipulation */
2246                 clear_page_mlock(old_page);
2247                 unlock_page(old_page);
2248         }
2249
2250         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2251                 goto oom_free_new;
2252
2253         /*
2254          * Re-check the pte - we dropped the lock
2255          */
2256         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2257         if (likely(pte_same(*page_table, orig_pte))) {
2258                 if (old_page) {
2259                         if (!PageAnon(old_page)) {
2260                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2261                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2262                         }
2263                 } else
2264                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2265                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2266                 entry = mk_pte(new_page, vma->vm_page_prot);
2267                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2268                 /*
2269                  * Clear the pte entry and flush it first, before updating the
2270                  * pte with the new entry. This will avoid a race condition
2271                  * seen in the presence of one thread doing SMC and another
2272                  * thread doing COW.
2273                  */
2274                 ptep_clear_flush(vma, address, page_table);
2275                 page_add_new_anon_rmap(new_page, vma, address);
2276                 /*
2277                  * We call the notify macro here because, when using secondary
2278                  * mmu page tables (such as kvm shadow page tables), we want the
2279                  * new page to be mapped directly into the secondary page table.
2280                  */
2281                 set_pte_at_notify(mm, address, page_table, entry);
2282                 update_mmu_cache(vma, address, page_table);
2283                 if (old_page) {
2284                         /*
2285                          * Only after switching the pte to the new page may
2286                          * we remove the mapcount here. Otherwise another
2287                          * process may come and find the rmap count decremented
2288                          * before the pte is switched to the new page, and
2289                          * "reuse" the old page writing into it while our pte
2290                          * here still points into it and can be read by other
2291                          * threads.
2292                          *
2293                          * The critical issue is to order this
2294                          * page_remove_rmap with the ptp_clear_flush above.
2295                          * Those stores are ordered by (if nothing else,)
2296                          * the barrier present in the atomic_add_negative
2297                          * in page_remove_rmap.
2298                          *
2299                          * Then the TLB flush in ptep_clear_flush ensures that
2300                          * no process can access the old page before the
2301                          * decremented mapcount is visible. And the old page
2302                          * cannot be reused until after the decremented
2303                          * mapcount is visible. So transitively, TLBs to
2304                          * old page will be flushed before it can be reused.
2305                          */
2306                         page_remove_rmap(old_page);
2307                 }
2308
2309                 /* Free the old page.. */
2310                 new_page = old_page;
2311                 ret |= VM_FAULT_WRITE;
2312         } else
2313                 mem_cgroup_uncharge_page(new_page);
2314
2315         if (new_page)
2316                 page_cache_release(new_page);
2317         if (old_page)
2318                 page_cache_release(old_page);
2319 unlock:
2320         pte_unmap_unlock(page_table, ptl);
2321         if (dirty_page) {
2322                 /*
2323                  * Yes, Virginia, this is actually required to prevent a race
2324                  * with clear_page_dirty_for_io() from clearing the page dirty
2325                  * bit after it clear all dirty ptes, but before a racing
2326                  * do_wp_page installs a dirty pte.
2327                  *
2328                  * do_no_page is protected similarly.
2329                  */
2330                 if (!page_mkwrite) {
2331                         wait_on_page_locked(dirty_page);
2332                         set_page_dirty_balance(dirty_page, page_mkwrite);
2333                 }
2334                 put_page(dirty_page);
2335                 if (page_mkwrite) {
2336                         struct address_space *mapping = dirty_page->mapping;
2337
2338                         set_page_dirty(dirty_page);
2339                         unlock_page(dirty_page);
2340                         page_cache_release(dirty_page);
2341                         if (mapping)    {
2342                                 /*
2343                                  * Some device drivers do not set page.mapping
2344                                  * but still dirty their pages
2345                                  */
2346                                 balance_dirty_pages_ratelimited(mapping);
2347                         }
2348                 }
2349
2350                 /* file_update_time outside page_lock */
2351                 if (vma->vm_file)
2352                         file_update_time(vma->vm_file);
2353         }
2354         return ret;
2355 oom_free_new:
2356         page_cache_release(new_page);
2357 oom:
2358         if (old_page) {
2359                 if (page_mkwrite) {
2360                         unlock_page(old_page);
2361                         page_cache_release(old_page);
2362                 }
2363                 page_cache_release(old_page);
2364         }
2365         return VM_FAULT_OOM;
2366
2367 unwritable_page:
2368         page_cache_release(old_page);
2369         return ret;
2370 }
2371
2372 /*
2373  * Helper functions for unmap_mapping_range().
2374  *
2375  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2376  *
2377  * We have to restart searching the prio_tree whenever we drop the lock,
2378  * since the iterator is only valid while the lock is held, and anyway
2379  * a later vma might be split and reinserted earlier while lock dropped.
2380  *
2381  * The list of nonlinear vmas could be handled more efficiently, using
2382  * a placeholder, but handle it in the same way until a need is shown.
2383  * It is important to search the prio_tree before nonlinear list: a vma
2384  * may become nonlinear and be shifted from prio_tree to nonlinear list
2385  * while the lock is dropped; but never shifted from list to prio_tree.
2386  *
2387  * In order to make forward progress despite restarting the search,
2388  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2389  * quickly skip it next time around.  Since the prio_tree search only
2390  * shows us those vmas affected by unmapping the range in question, we
2391  * can't efficiently keep all vmas in step with mapping->truncate_count:
2392  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2393  * mapping->truncate_count and vma->vm_truncate_count are protected by
2394  * i_mmap_lock.
2395  *
2396  * In order to make forward progress despite repeatedly restarting some
2397  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2398  * and restart from that address when we reach that vma again.  It might
2399  * have been split or merged, shrunk or extended, but never shifted: so
2400  * restart_addr remains valid so long as it remains in the vma's range.
2401  * unmap_mapping_range forces truncate_count to leap over page-aligned
2402  * values so we can save vma's restart_addr in its truncate_count field.
2403  */
2404 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2405
2406 static void reset_vma_truncate_counts(struct address_space *mapping)
2407 {
2408         struct vm_area_struct *vma;
2409         struct prio_tree_iter iter;
2410
2411         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2412                 vma->vm_truncate_count = 0;
2413         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2414                 vma->vm_truncate_count = 0;
2415 }
2416
2417 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2418                 unsigned long start_addr, unsigned long end_addr,
2419                 struct zap_details *details)
2420 {
2421         unsigned long restart_addr;
2422         int need_break;
2423
2424         /*
2425          * files that support invalidating or truncating portions of the
2426          * file from under mmaped areas must have their ->fault function
2427          * return a locked page (and set VM_FAULT_LOCKED in the return).
2428          * This provides synchronisation against concurrent unmapping here.
2429          */
2430
2431 again:
2432         restart_addr = vma->vm_truncate_count;
2433         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2434                 start_addr = restart_addr;
2435                 if (start_addr >= end_addr) {
2436                         /* Top of vma has been split off since last time */
2437                         vma->vm_truncate_count = details->truncate_count;
2438                         return 0;
2439                 }
2440         }
2441
2442         restart_addr = zap_page_range(vma, start_addr,
2443                                         end_addr - start_addr, details);
2444         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2445
2446         if (restart_addr >= end_addr) {
2447                 /* We have now completed this vma: mark it so */
2448                 vma->vm_truncate_count = details->truncate_count;
2449                 if (!need_break)
2450                         return 0;
2451         } else {
2452                 /* Note restart_addr in vma's truncate_count field */
2453                 vma->vm_truncate_count = restart_addr;
2454                 if (!need_break)
2455                         goto again;
2456         }
2457
2458         spin_unlock(details->i_mmap_lock);
2459         cond_resched();
2460         spin_lock(details->i_mmap_lock);
2461         return -EINTR;
2462 }
2463
2464 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2465                                             struct zap_details *details)
2466 {
2467         struct vm_area_struct *vma;
2468         struct prio_tree_iter iter;
2469         pgoff_t vba, vea, zba, zea;
2470
2471 restart:
2472         vma_prio_tree_foreach(vma, &iter, root,
2473                         details->first_index, details->last_index) {
2474                 /* Skip quickly over those we have already dealt with */
2475                 if (vma->vm_truncate_count == details->truncate_count)
2476                         continue;
2477
2478                 vba = vma->vm_pgoff;
2479                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2480                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2481                 zba = details->first_index;
2482                 if (zba < vba)
2483                         zba = vba;
2484                 zea = details->last_index;
2485                 if (zea > vea)
2486                         zea = vea;
2487
2488                 if (unmap_mapping_range_vma(vma,
2489                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2490                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2491                                 details) < 0)
2492                         goto restart;
2493         }
2494 }
2495
2496 static inline void unmap_mapping_range_list(struct list_head *head,
2497                                             struct zap_details *details)
2498 {
2499         struct vm_area_struct *vma;
2500
2501         /*
2502          * In nonlinear VMAs there is no correspondence between virtual address
2503          * offset and file offset.  So we must perform an exhaustive search
2504          * across *all* the pages in each nonlinear VMA, not just the pages
2505          * whose virtual address lies outside the file truncation point.
2506          */
2507 restart:
2508         list_for_each_entry(vma, head, shared.vm_set.list) {
2509                 /* Skip quickly over those we have already dealt with */
2510                 if (vma->vm_truncate_count == details->truncate_count)
2511                         continue;
2512                 details->nonlinear_vma = vma;
2513                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2514                                         vma->vm_end, details) < 0)
2515                         goto restart;
2516         }
2517 }
2518
2519 /**
2520  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2521  * @mapping: the address space containing mmaps to be unmapped.
2522  * @holebegin: byte in first page to unmap, relative to the start of
2523  * the underlying file.  This will be rounded down to a PAGE_SIZE
2524  * boundary.  Note that this is different from truncate_pagecache(), which
2525  * must keep the partial page.  In contrast, we must get rid of
2526  * partial pages.
2527  * @holelen: size of prospective hole in bytes.  This will be rounded
2528  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2529  * end of the file.
2530  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2531  * but 0 when invalidating pagecache, don't throw away private data.
2532  */
2533 void unmap_mapping_range(struct address_space *mapping,
2534                 loff_t const holebegin, loff_t const holelen, int even_cows)
2535 {
2536         struct zap_details details;
2537         pgoff_t hba = holebegin >> PAGE_SHIFT;
2538         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2539
2540         /* Check for overflow. */
2541         if (sizeof(holelen) > sizeof(hlen)) {
2542                 long long holeend =
2543                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2544                 if (holeend & ~(long long)ULONG_MAX)
2545                         hlen = ULONG_MAX - hba + 1;
2546         }
2547
2548         details.check_mapping = even_cows? NULL: mapping;
2549         details.nonlinear_vma = NULL;
2550         details.first_index = hba;
2551         details.last_index = hba + hlen - 1;
2552         if (details.last_index < details.first_index)
2553                 details.last_index = ULONG_MAX;
2554         details.i_mmap_lock = &mapping->i_mmap_lock;
2555
2556         spin_lock(&mapping->i_mmap_lock);
2557
2558         /* Protect against endless unmapping loops */
2559         mapping->truncate_count++;
2560         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2561                 if (mapping->truncate_count == 0)
2562                         reset_vma_truncate_counts(mapping);
2563                 mapping->truncate_count++;
2564         }
2565         details.truncate_count = mapping->truncate_count;
2566
2567         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2568                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2569         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2570                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2571         spin_unlock(&mapping->i_mmap_lock);
2572 }
2573 EXPORT_SYMBOL(unmap_mapping_range);
2574
2575 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2576 {
2577         struct address_space *mapping = inode->i_mapping;
2578
2579         /*
2580          * If the underlying filesystem is not going to provide
2581          * a way to truncate a range of blocks (punch a hole) -
2582          * we should return failure right now.
2583          */
2584         if (!inode->i_op->truncate_range)
2585                 return -ENOSYS;
2586
2587         mutex_lock(&inode->i_mutex);
2588         down_write(&inode->i_alloc_sem);
2589         unmap_mapping_range(mapping, offset, (end - offset), 1);
2590         truncate_inode_pages_range(mapping, offset, end);
2591         unmap_mapping_range(mapping, offset, (end - offset), 1);
2592         inode->i_op->truncate_range(inode, offset, end);
2593         up_write(&inode->i_alloc_sem);
2594         mutex_unlock(&inode->i_mutex);
2595
2596         return 0;
2597 }
2598
2599 /*
2600  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2601  * but allow concurrent faults), and pte mapped but not yet locked.
2602  * We return with mmap_sem still held, but pte unmapped and unlocked.
2603  */
2604 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2605                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2606                 unsigned int flags, pte_t orig_pte)
2607 {
2608         spinlock_t *ptl;
2609         struct page *page;
2610         swp_entry_t entry;
2611         pte_t pte;
2612         struct mem_cgroup *ptr = NULL;
2613         int ret = 0;
2614
2615         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2616                 goto out;
2617
2618         entry = pte_to_swp_entry(orig_pte);
2619         if (unlikely(non_swap_entry(entry))) {
2620                 if (is_migration_entry(entry)) {
2621                         migration_entry_wait(mm, pmd, address);
2622                 } else if (is_hwpoison_entry(entry)) {
2623                         ret = VM_FAULT_HWPOISON;
2624                 } else {
2625                         print_bad_pte(vma, address, orig_pte, NULL);
2626                         ret = VM_FAULT_SIGBUS;
2627                 }
2628                 goto out;
2629         }
2630         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2631         page = lookup_swap_cache(entry);
2632         if (!page) {
2633                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2634                 page = swapin_readahead(entry,
2635                                         GFP_HIGHUSER_MOVABLE, vma, address);
2636                 if (!page) {
2637                         /*
2638                          * Back out if somebody else faulted in this pte
2639                          * while we released the pte lock.
2640                          */
2641                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2642                         if (likely(pte_same(*page_table, orig_pte)))
2643                                 ret = VM_FAULT_OOM;
2644                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2645                         goto unlock;
2646                 }
2647
2648                 /* Had to read the page from swap area: Major fault */
2649                 ret = VM_FAULT_MAJOR;
2650                 count_vm_event(PGMAJFAULT);
2651         } else if (PageHWPoison(page)) {
2652                 /*
2653                  * hwpoisoned dirty swapcache pages are kept for killing
2654                  * owner processes (which may be unknown at hwpoison time)
2655                  */
2656                 ret = VM_FAULT_HWPOISON;
2657                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2658                 goto out_release;
2659         }
2660
2661         lock_page(page);
2662         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2663
2664         page = ksm_might_need_to_copy(page, vma, address);
2665         if (!page) {
2666                 ret = VM_FAULT_OOM;
2667                 goto out;
2668         }
2669
2670         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2671                 ret = VM_FAULT_OOM;
2672                 goto out_page;
2673         }
2674
2675         /*
2676          * Back out if somebody else already faulted in this pte.
2677          */
2678         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2679         if (unlikely(!pte_same(*page_table, orig_pte)))
2680                 goto out_nomap;
2681
2682         if (unlikely(!PageUptodate(page))) {
2683                 ret = VM_FAULT_SIGBUS;
2684                 goto out_nomap;
2685         }
2686
2687         /*
2688          * The page isn't present yet, go ahead with the fault.
2689          *
2690          * Be careful about the sequence of operations here.
2691          * To get its accounting right, reuse_swap_page() must be called
2692          * while the page is counted on swap but not yet in mapcount i.e.
2693          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2694          * must be called after the swap_free(), or it will never succeed.
2695          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2696          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2697          * in page->private. In this case, a record in swap_cgroup  is silently
2698          * discarded at swap_free().
2699          */
2700
2701         inc_mm_counter_fast(mm, MM_ANONPAGES);
2702         dec_mm_counter_fast(mm, MM_SWAPENTS);
2703         pte = mk_pte(page, vma->vm_page_prot);
2704         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2705                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2706                 flags &= ~FAULT_FLAG_WRITE;
2707         }
2708         flush_icache_page(vma, page);
2709         set_pte_at(mm, address, page_table, pte);
2710         page_add_anon_rmap(page, vma, address);
2711         /* It's better to call commit-charge after rmap is established */
2712         mem_cgroup_commit_charge_swapin(page, ptr);
2713
2714         swap_free(entry);
2715         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2716                 try_to_free_swap(page);
2717         unlock_page(page);
2718
2719         if (flags & FAULT_FLAG_WRITE) {
2720                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2721                 if (ret & VM_FAULT_ERROR)
2722                         ret &= VM_FAULT_ERROR;
2723                 goto out;
2724         }
2725
2726         /* No need to invalidate - it was non-present before */
2727         update_mmu_cache(vma, address, page_table);
2728 unlock:
2729         pte_unmap_unlock(page_table, ptl);
2730 out:
2731         return ret;
2732 out_nomap:
2733         mem_cgroup_cancel_charge_swapin(ptr);
2734         pte_unmap_unlock(page_table, ptl);
2735 out_page:
2736         unlock_page(page);
2737 out_release:
2738         page_cache_release(page);
2739         return ret;
2740 }
2741
2742 /*
2743  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2744  * but allow concurrent faults), and pte mapped but not yet locked.
2745  * We return with mmap_sem still held, but pte unmapped and unlocked.
2746  */
2747 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2748                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2749                 unsigned int flags)
2750 {
2751         struct page *page;
2752         spinlock_t *ptl;
2753         pte_t entry;
2754
2755         if (!(flags & FAULT_FLAG_WRITE)) {
2756                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2757                                                 vma->vm_page_prot));
2758                 ptl = pte_lockptr(mm, pmd);
2759                 spin_lock(ptl);
2760                 if (!pte_none(*page_table))
2761                         goto unlock;
2762                 goto setpte;
2763         }
2764
2765         /* Allocate our own private page. */
2766         pte_unmap(page_table);
2767
2768         if (unlikely(anon_vma_prepare(vma)))
2769                 goto oom;
2770         page = alloc_zeroed_user_highpage_movable(vma, address);
2771         if (!page)
2772                 goto oom;
2773         __SetPageUptodate(page);
2774
2775         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2776                 goto oom_free_page;
2777
2778         entry = mk_pte(page, vma->vm_page_prot);
2779         if (vma->vm_flags & VM_WRITE)
2780                 entry = pte_mkwrite(pte_mkdirty(entry));
2781
2782         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2783         if (!pte_none(*page_table))
2784                 goto release;
2785
2786         inc_mm_counter_fast(mm, MM_ANONPAGES);
2787         page_add_new_anon_rmap(page, vma, address);
2788 setpte:
2789         set_pte_at(mm, address, page_table, entry);
2790
2791         /* No need to invalidate - it was non-present before */
2792         update_mmu_cache(vma, address, page_table);
2793 unlock:
2794         pte_unmap_unlock(page_table, ptl);
2795         return 0;
2796 release:
2797         mem_cgroup_uncharge_page(page);
2798         page_cache_release(page);
2799         goto unlock;
2800 oom_free_page:
2801         page_cache_release(page);
2802 oom:
2803         return VM_FAULT_OOM;
2804 }
2805
2806 /*
2807  * __do_fault() tries to create a new page mapping. It aggressively
2808  * tries to share with existing pages, but makes a separate copy if
2809  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2810  * the next page fault.
2811  *
2812  * As this is called only for pages that do not currently exist, we
2813  * do not need to flush old virtual caches or the TLB.
2814  *
2815  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2816  * but allow concurrent faults), and pte neither mapped nor locked.
2817  * We return with mmap_sem still held, but pte unmapped and unlocked.
2818  */
2819 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2820                 unsigned long address, pmd_t *pmd,
2821                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2822 {
2823         pte_t *page_table;
2824         spinlock_t *ptl;
2825         struct page *page;
2826         pte_t entry;
2827         int anon = 0;
2828         int charged = 0;
2829         struct page *dirty_page = NULL;
2830         struct vm_fault vmf;
2831         int ret;
2832         int page_mkwrite = 0;
2833
2834         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2835         vmf.pgoff = pgoff;
2836         vmf.flags = flags;
2837         vmf.page = NULL;
2838
2839         ret = vma->vm_ops->fault(vma, &vmf);
2840         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2841                 return ret;
2842
2843         if (unlikely(PageHWPoison(vmf.page))) {
2844                 if (ret & VM_FAULT_LOCKED)
2845                         unlock_page(vmf.page);
2846                 return VM_FAULT_HWPOISON;
2847         }
2848
2849         /*
2850          * For consistency in subsequent calls, make the faulted page always
2851          * locked.
2852          */
2853         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2854                 lock_page(vmf.page);
2855         else
2856                 VM_BUG_ON(!PageLocked(vmf.page));
2857
2858         /*
2859          * Should we do an early C-O-W break?
2860          */
2861         page = vmf.page;
2862         if (flags & FAULT_FLAG_WRITE) {
2863                 if (!(vma->vm_flags & VM_SHARED)) {
2864                         anon = 1;
2865                         if (unlikely(anon_vma_prepare(vma))) {
2866                                 ret = VM_FAULT_OOM;
2867                                 goto out;
2868                         }
2869                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2870                                                 vma, address);
2871                         if (!page) {
2872                                 ret = VM_FAULT_OOM;
2873                                 goto out;
2874                         }
2875                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2876                                 ret = VM_FAULT_OOM;
2877                                 page_cache_release(page);
2878                                 goto out;
2879                         }
2880                         charged = 1;
2881                         /*
2882                          * Don't let another task, with possibly unlocked vma,
2883                          * keep the mlocked page.
2884                          */
2885                         if (vma->vm_flags & VM_LOCKED)
2886                                 clear_page_mlock(vmf.page);
2887                         copy_user_highpage(page, vmf.page, address, vma);
2888                         __SetPageUptodate(page);
2889                 } else {
2890                         /*
2891                          * If the page will be shareable, see if the backing
2892                          * address space wants to know that the page is about
2893                          * to become writable
2894                          */
2895                         if (vma->vm_ops->page_mkwrite) {
2896                                 int tmp;
2897
2898                                 unlock_page(page);
2899                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2900                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2901                                 if (unlikely(tmp &
2902                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2903                                         ret = tmp;
2904                                         goto unwritable_page;
2905                                 }
2906                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2907                                         lock_page(page);
2908                                         if (!page->mapping) {
2909                                                 ret = 0; /* retry the fault */
2910                                                 unlock_page(page);
2911                                                 goto unwritable_page;
2912                                         }
2913                                 } else
2914                                         VM_BUG_ON(!PageLocked(page));
2915                                 page_mkwrite = 1;
2916                         }
2917                 }
2918
2919         }
2920
2921         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2922
2923         /*
2924          * This silly early PAGE_DIRTY setting removes a race
2925          * due to the bad i386 page protection. But it's valid
2926          * for other architectures too.
2927          *
2928          * Note that if FAULT_FLAG_WRITE is set, we either now have
2929          * an exclusive copy of the page, or this is a shared mapping,
2930          * so we can make it writable and dirty to avoid having to
2931          * handle that later.
2932          */
2933         /* Only go through if we didn't race with anybody else... */
2934         if (likely(pte_same(*page_table, orig_pte))) {
2935                 flush_icache_page(vma, page);
2936                 entry = mk_pte(page, vma->vm_page_prot);
2937                 if (flags & FAULT_FLAG_WRITE)
2938                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2939                 if (anon) {
2940                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2941                         page_add_new_anon_rmap(page, vma, address);
2942                 } else {
2943                         inc_mm_counter_fast(mm, MM_FILEPAGES);
2944                         page_add_file_rmap(page);
2945                         if (flags & FAULT_FLAG_WRITE) {
2946                                 dirty_page = page;
2947                                 get_page(dirty_page);
2948                         }
2949                 }
2950                 set_pte_at(mm, address, page_table, entry);
2951
2952                 /* no need to invalidate: a not-present page won't be cached */
2953                 update_mmu_cache(vma, address, page_table);
2954         } else {
2955                 if (charged)
2956                         mem_cgroup_uncharge_page(page);
2957                 if (anon)
2958                         page_cache_release(page);
2959                 else
2960                         anon = 1; /* no anon but release faulted_page */
2961         }
2962
2963         pte_unmap_unlock(page_table, ptl);
2964
2965 out:
2966         if (dirty_page) {
2967                 struct address_space *mapping = page->mapping;
2968
2969                 if (set_page_dirty(dirty_page))
2970                         page_mkwrite = 1;
2971                 unlock_page(dirty_page);
2972                 put_page(dirty_page);
2973                 if (page_mkwrite && mapping) {
2974                         /*
2975                          * Some device drivers do not set page.mapping but still
2976                          * dirty their pages
2977                          */
2978                         balance_dirty_pages_ratelimited(mapping);
2979                 }
2980
2981                 /* file_update_time outside page_lock */
2982                 if (vma->vm_file)
2983                         file_update_time(vma->vm_file);
2984         } else {
2985                 unlock_page(vmf.page);
2986                 if (anon)
2987                         page_cache_release(vmf.page);
2988         }
2989
2990         return ret;
2991
2992 unwritable_page:
2993         page_cache_release(page);
2994         return ret;
2995 }
2996
2997 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2998                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2999                 unsigned int flags, pte_t orig_pte)
3000 {
3001         pgoff_t pgoff = (((address & PAGE_MASK)
3002                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3003
3004         pte_unmap(page_table);
3005         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3006 }
3007
3008 /*
3009  * Fault of a previously existing named mapping. Repopulate the pte
3010  * from the encoded file_pte if possible. This enables swappable
3011  * nonlinear vmas.
3012  *
3013  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3014  * but allow concurrent faults), and pte mapped but not yet locked.
3015  * We return with mmap_sem still held, but pte unmapped and unlocked.
3016  */
3017 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3018                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3019                 unsigned int flags, pte_t orig_pte)
3020 {
3021         pgoff_t pgoff;
3022
3023         flags |= FAULT_FLAG_NONLINEAR;
3024
3025         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3026                 return 0;
3027
3028         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3029                 /*
3030                  * Page table corrupted: show pte and kill process.
3031                  */
3032                 print_bad_pte(vma, address, orig_pte, NULL);
3033                 return VM_FAULT_SIGBUS;
3034         }
3035
3036         pgoff = pte_to_pgoff(orig_pte);
3037         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3038 }
3039
3040 /*
3041  * These routines also need to handle stuff like marking pages dirty
3042  * and/or accessed for architectures that don't do it in hardware (most
3043  * RISC architectures).  The early dirtying is also good on the i386.
3044  *
3045  * There is also a hook called "update_mmu_cache()" that architectures
3046  * with external mmu caches can use to update those (ie the Sparc or
3047  * PowerPC hashed page tables that act as extended TLBs).
3048  *
3049  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3050  * but allow concurrent faults), and pte mapped but not yet locked.
3051  * We return with mmap_sem still held, but pte unmapped and unlocked.
3052  */
3053 static inline int handle_pte_fault(struct mm_struct *mm,
3054                 struct vm_area_struct *vma, unsigned long address,
3055                 pte_t *pte, pmd_t *pmd, unsigned int flags)
3056 {
3057         pte_t entry;
3058         spinlock_t *ptl;
3059
3060         entry = *pte;
3061         if (!pte_present(entry)) {
3062                 if (pte_none(entry)) {
3063                         if (vma->vm_ops) {
3064                                 if (likely(vma->vm_ops->fault))
3065                                         return do_linear_fault(mm, vma, address,
3066                                                 pte, pmd, flags, entry);
3067                         }
3068                         return do_anonymous_page(mm, vma, address,
3069                                                  pte, pmd, flags);
3070                 }
3071                 if (pte_file(entry))
3072                         return do_nonlinear_fault(mm, vma, address,
3073                                         pte, pmd, flags, entry);
3074                 return do_swap_page(mm, vma, address,
3075                                         pte, pmd, flags, entry);
3076         }
3077
3078         ptl = pte_lockptr(mm, pmd);
3079         spin_lock(ptl);
3080         if (unlikely(!pte_same(*pte, entry)))
3081                 goto unlock;
3082         if (flags & FAULT_FLAG_WRITE) {
3083                 if (!pte_write(entry))
3084                         return do_wp_page(mm, vma, address,
3085                                         pte, pmd, ptl, entry);
3086                 entry = pte_mkdirty(entry);
3087         }
3088         entry = pte_mkyoung(entry);
3089         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3090                 update_mmu_cache(vma, address, pte);
3091         } else {
3092                 /*
3093                  * This is needed only for protection faults but the arch code
3094                  * is not yet telling us if this is a protection fault or not.
3095                  * This still avoids useless tlb flushes for .text page faults
3096                  * with threads.
3097                  */
3098                 if (flags & FAULT_FLAG_WRITE)
3099                         flush_tlb_page(vma, address);
3100         }
3101 unlock:
3102         pte_unmap_unlock(pte, ptl);
3103         return 0;
3104 }
3105
3106 /*
3107  * By the time we get here, we already hold the mm semaphore
3108  */
3109 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3110                 unsigned long address, unsigned int flags)
3111 {
3112         pgd_t *pgd;
3113         pud_t *pud;
3114         pmd_t *pmd;
3115         pte_t *pte;
3116
3117         __set_current_state(TASK_RUNNING);
3118
3119         count_vm_event(PGFAULT);
3120
3121         /* do counter updates before entering really critical section. */
3122         check_sync_rss_stat(current);
3123
3124         if (unlikely(is_vm_hugetlb_page(vma)))
3125                 return hugetlb_fault(mm, vma, address, flags);
3126
3127         pgd = pgd_offset(mm, address);
3128         pud = pud_alloc(mm, pgd, address);
3129         if (!pud)
3130                 return VM_FAULT_OOM;
3131         pmd = pmd_alloc(mm, pud, address);
3132         if (!pmd)
3133                 return VM_FAULT_OOM;
3134         pte = pte_alloc_map(mm, pmd, address);
3135         if (!pte)
3136                 return VM_FAULT_OOM;
3137
3138         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3139 }
3140
3141 #ifndef __PAGETABLE_PUD_FOLDED
3142 /*
3143  * Allocate page upper directory.
3144  * We've already handled the fast-path in-line.
3145  */
3146 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3147 {
3148         pud_t *new = pud_alloc_one(mm, address);
3149         if (!new)
3150                 return -ENOMEM;
3151
3152         smp_wmb(); /* See comment in __pte_alloc */
3153
3154         spin_lock(&mm->page_table_lock);
3155         if (pgd_present(*pgd))          /* Another has populated it */
3156                 pud_free(mm, new);
3157         else
3158                 pgd_populate(mm, pgd, new);
3159         spin_unlock(&mm->page_table_lock);
3160         return 0;
3161 }
3162 #endif /* __PAGETABLE_PUD_FOLDED */
3163
3164 #ifndef __PAGETABLE_PMD_FOLDED
3165 /*
3166  * Allocate page middle directory.
3167  * We've already handled the fast-path in-line.
3168  */
3169 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3170 {
3171         pmd_t *new = pmd_alloc_one(mm, address);
3172         if (!new)
3173                 return -ENOMEM;
3174
3175         smp_wmb(); /* See comment in __pte_alloc */
3176
3177         spin_lock(&mm->page_table_lock);
3178 #ifndef __ARCH_HAS_4LEVEL_HACK
3179         if (pud_present(*pud))          /* Another has populated it */
3180                 pmd_free(mm, new);
3181         else
3182                 pud_populate(mm, pud, new);
3183 #else
3184         if (pgd_present(*pud))          /* Another has populated it */
3185                 pmd_free(mm, new);
3186         else
3187                 pgd_populate(mm, pud, new);
3188 #endif /* __ARCH_HAS_4LEVEL_HACK */
3189         spin_unlock(&mm->page_table_lock);
3190         return 0;
3191 }
3192 #endif /* __PAGETABLE_PMD_FOLDED */
3193
3194 int make_pages_present(unsigned long addr, unsigned long end)
3195 {
3196         int ret, len, write;
3197         struct vm_area_struct * vma;
3198
3199         vma = find_vma(current->mm, addr);
3200         if (!vma)
3201                 return -ENOMEM;
3202         write = (vma->vm_flags & VM_WRITE) != 0;
3203         BUG_ON(addr >= end);
3204         BUG_ON(end > vma->vm_end);
3205         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3206         ret = get_user_pages(current, current->mm, addr,
3207                         len, write, 0, NULL, NULL);
3208         if (ret < 0)
3209                 return ret;
3210         return ret == len ? 0 : -EFAULT;
3211 }
3212
3213 #if !defined(__HAVE_ARCH_GATE_AREA)
3214
3215 #if defined(AT_SYSINFO_EHDR)
3216 static struct vm_area_struct gate_vma;
3217
3218 static int __init gate_vma_init(void)
3219 {
3220         gate_vma.vm_mm = NULL;
3221         gate_vma.vm_start = FIXADDR_USER_START;
3222         gate_vma.vm_end = FIXADDR_USER_END;
3223         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3224         gate_vma.vm_page_prot = __P101;
3225         /*
3226          * Make sure the vDSO gets into every core dump.
3227          * Dumping its contents makes post-mortem fully interpretable later
3228          * without matching up the same kernel and hardware config to see
3229          * what PC values meant.
3230          */
3231         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3232         return 0;
3233 }
3234 __initcall(gate_vma_init);
3235 #endif
3236
3237 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3238 {
3239 #ifdef AT_SYSINFO_EHDR
3240         return &gate_vma;
3241 #else
3242         return NULL;
3243 #endif
3244 }
3245
3246 int in_gate_area_no_task(unsigned long addr)
3247 {
3248 #ifdef AT_SYSINFO_EHDR
3249         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3250                 return 1;
3251 #endif
3252         return 0;
3253 }
3254
3255 #endif  /* __HAVE_ARCH_GATE_AREA */
3256
3257 static int follow_pte(struct mm_struct *mm, unsigned long address,
3258                 pte_t **ptepp, spinlock_t **ptlp)
3259 {
3260         pgd_t *pgd;
3261         pud_t *pud;
3262         pmd_t *pmd;
3263         pte_t *ptep;
3264
3265         pgd = pgd_offset(mm, address);
3266         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3267                 goto out;
3268
3269         pud = pud_offset(pgd, address);
3270         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3271                 goto out;
3272
3273         pmd = pmd_offset(pud, address);
3274         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3275                 goto out;
3276
3277         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3278         if (pmd_huge(*pmd))
3279                 goto out;
3280
3281         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3282         if (!ptep)
3283                 goto out;
3284         if (!pte_present(*ptep))
3285                 goto unlock;
3286         *ptepp = ptep;
3287         return 0;
3288 unlock:
3289         pte_unmap_unlock(ptep, *ptlp);
3290 out:
3291         return -EINVAL;
3292 }
3293
3294 /**
3295  * follow_pfn - look up PFN at a user virtual address
3296  * @vma: memory mapping
3297  * @address: user virtual address
3298  * @pfn: location to store found PFN
3299  *
3300  * Only IO mappings and raw PFN mappings are allowed.
3301  *
3302  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3303  */
3304 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3305         unsigned long *pfn)
3306 {
3307         int ret = -EINVAL;
3308         spinlock_t *ptl;
3309         pte_t *ptep;
3310
3311         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3312                 return ret;
3313
3314         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3315         if (ret)
3316                 return ret;
3317         *pfn = pte_pfn(*ptep);
3318         pte_unmap_unlock(ptep, ptl);
3319         return 0;
3320 }
3321 EXPORT_SYMBOL(follow_pfn);
3322
3323 #ifdef CONFIG_HAVE_IOREMAP_PROT
3324 int follow_phys(struct vm_area_struct *vma,
3325                 unsigned long address, unsigned int flags,
3326                 unsigned long *prot, resource_size_t *phys)
3327 {
3328         int ret = -EINVAL;
3329         pte_t *ptep, pte;
3330         spinlock_t *ptl;
3331
3332         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3333                 goto out;
3334
3335         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3336                 goto out;
3337         pte = *ptep;
3338
3339         if ((flags & FOLL_WRITE) && !pte_write(pte))
3340                 goto unlock;
3341
3342         *prot = pgprot_val(pte_pgprot(pte));
3343         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3344
3345         ret = 0;
3346 unlock:
3347         pte_unmap_unlock(ptep, ptl);
3348 out:
3349         return ret;
3350 }
3351
3352 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3353                         void *buf, int len, int write)
3354 {
3355         resource_size_t phys_addr;
3356         unsigned long prot = 0;
3357         void __iomem *maddr;
3358         int offset = addr & (PAGE_SIZE-1);
3359
3360         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3361                 return -EINVAL;
3362
3363         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3364         if (write)
3365                 memcpy_toio(maddr + offset, buf, len);
3366         else
3367                 memcpy_fromio(buf, maddr + offset, len);
3368         iounmap(maddr);
3369
3370         return len;
3371 }
3372 #endif
3373
3374 /*
3375  * Access another process' address space.
3376  * Source/target buffer must be kernel space,
3377  * Do not walk the page table directly, use get_user_pages
3378  */
3379 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3380 {
3381         struct mm_struct *mm;
3382         struct vm_area_struct *vma;
3383         void *old_buf = buf;
3384
3385         mm = get_task_mm(tsk);
3386         if (!mm)
3387                 return 0;
3388
3389         down_read(&mm->mmap_sem);
3390         /* ignore errors, just check how much was successfully transferred */
3391         while (len) {
3392                 int bytes, ret, offset;
3393                 void *maddr;
3394                 struct page *page = NULL;
3395
3396                 ret = get_user_pages(tsk, mm, addr, 1,
3397                                 write, 1, &page, &vma);
3398                 if (ret <= 0) {
3399                         /*
3400                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3401                          * we can access using slightly different code.
3402                          */
3403 #ifdef CONFIG_HAVE_IOREMAP_PROT
3404                         vma = find_vma(mm, addr);
3405                         if (!vma)
3406                                 break;
3407                         if (vma->vm_ops && vma->vm_ops->access)
3408                                 ret = vma->vm_ops->access(vma, addr, buf,
3409                                                           len, write);
3410                         if (ret <= 0)
3411 #endif
3412                                 break;
3413                         bytes = ret;
3414                 } else {
3415                         bytes = len;
3416                         offset = addr & (PAGE_SIZE-1);
3417                         if (bytes > PAGE_SIZE-offset)
3418                                 bytes = PAGE_SIZE-offset;
3419
3420                         maddr = kmap(page);
3421                         if (write) {
3422                                 copy_to_user_page(vma, page, addr,
3423                                                   maddr + offset, buf, bytes);
3424                                 set_page_dirty_lock(page);
3425                         } else {
3426                                 copy_from_user_page(vma, page, addr,
3427                                                     buf, maddr + offset, bytes);
3428                         }
3429                         kunmap(page);
3430                         page_cache_release(page);
3431                 }
3432                 len -= bytes;
3433                 buf += bytes;
3434                 addr += bytes;
3435         }
3436         up_read(&mm->mmap_sem);
3437         mmput(mm);
3438
3439         return buf - old_buf;
3440 }
3441
3442 /*
3443  * Print the name of a VMA.
3444  */
3445 void print_vma_addr(char *prefix, unsigned long ip)
3446 {
3447         struct mm_struct *mm = current->mm;
3448         struct vm_area_struct *vma;
3449
3450         /*
3451          * Do not print if we are in atomic
3452          * contexts (in exception stacks, etc.):
3453          */
3454         if (preempt_count())
3455                 return;
3456
3457         down_read(&mm->mmap_sem);
3458         vma = find_vma(mm, ip);
3459         if (vma && vma->vm_file) {
3460                 struct file *f = vma->vm_file;
3461                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3462                 if (buf) {
3463                         char *p, *s;
3464
3465                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3466                         if (IS_ERR(p))
3467                                 p = "?";
3468                         s = strrchr(p, '/');
3469                         if (s)
3470                                 p = s+1;
3471                         printk("%s%s[%lx+%lx]", prefix, p,
3472                                         vma->vm_start,
3473                                         vma->vm_end - vma->vm_start);
3474                         free_page((unsigned long)buf);
3475                 }
3476         }
3477         up_read(&current->mm->mmap_sem);
3478 }
3479
3480 #ifdef CONFIG_PROVE_LOCKING
3481 void might_fault(void)
3482 {
3483         /*
3484          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3485          * holding the mmap_sem, this is safe because kernel memory doesn't
3486          * get paged out, therefore we'll never actually fault, and the
3487          * below annotations will generate false positives.
3488          */
3489         if (segment_eq(get_fs(), KERNEL_DS))
3490                 return;
3491
3492         might_sleep();
3493         /*
3494          * it would be nicer only to annotate paths which are not under
3495          * pagefault_disable, however that requires a larger audit and
3496          * providing helpers like get_user_atomic.
3497          */
3498         if (!in_atomic() && current->mm)
3499                 might_lock_read(&current->mm->mmap_sem);
3500 }
3501 EXPORT_SYMBOL(might_fault);
3502 #endif