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