19b2d44de9f0b39a97b4a45fc72ac65cfdd7167b
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60
61 #include <asm/io.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
64 #include <asm/tlb.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
67
68 #include "internal.h"
69
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
73 struct page *mem_map;
74
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
77 #endif
78
79 unsigned long num_physpages;
80 /*
81  * A number of key systems in x86 including ioremap() rely on the assumption
82  * that high_memory defines the upper bound on direct map memory, then end
83  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
84  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85  * and ZONE_HIGHMEM.
86  */
87 void * high_memory;
88
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
91
92 /*
93  * Randomize the address space (stacks, mmaps, brk, etc.).
94  *
95  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96  *   as ancient (libc5 based) binaries can segfault. )
97  */
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
100                                         1;
101 #else
102                                         2;
103 #endif
104
105 static int __init disable_randmaps(char *s)
106 {
107         randomize_va_space = 0;
108         return 1;
109 }
110 __setup("norandmaps", disable_randmaps);
111
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
114
115 /*
116  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117  */
118 static int __init init_zero_pfn(void)
119 {
120         zero_pfn = page_to_pfn(ZERO_PAGE(0));
121         return 0;
122 }
123 core_initcall(init_zero_pfn);
124
125
126 #if defined(SPLIT_RSS_COUNTING)
127
128 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
129 {
130         int i;
131
132         for (i = 0; i < NR_MM_COUNTERS; i++) {
133                 if (task->rss_stat.count[i]) {
134                         add_mm_counter(mm, i, task->rss_stat.count[i]);
135                         task->rss_stat.count[i] = 0;
136                 }
137         }
138         task->rss_stat.events = 0;
139 }
140
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142 {
143         struct task_struct *task = current;
144
145         if (likely(task->mm == mm))
146                 task->rss_stat.count[member] += val;
147         else
148                 add_mm_counter(mm, member, val);
149 }
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH  (64)
155 static void check_sync_rss_stat(struct task_struct *task)
156 {
157         if (unlikely(task != current))
158                 return;
159         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160                 __sync_task_rss_stat(task, task->mm);
161 }
162
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
164 {
165         long val = 0;
166
167         /*
168          * Don't use task->mm here...for avoiding to use task_get_mm()..
169          * The caller must guarantee task->mm is not invalid.
170          */
171         val = atomic_long_read(&mm->rss_stat.count[member]);
172         /*
173          * counter is updated in asynchronous manner and may go to minus.
174          * But it's never be expected number for users.
175          */
176         if (val < 0)
177                 return 0;
178         return (unsigned long)val;
179 }
180
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182 {
183         __sync_task_rss_stat(task, mm);
184 }
185 #else
186
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
189
190 static void check_sync_rss_stat(struct task_struct *task)
191 {
192 }
193
194 #endif
195
196 /*
197  * If a p?d_bad entry is found while walking page tables, report
198  * the error, before resetting entry to p?d_none.  Usually (but
199  * very seldom) called out from the p?d_none_or_clear_bad macros.
200  */
201
202 void pgd_clear_bad(pgd_t *pgd)
203 {
204         pgd_ERROR(*pgd);
205         pgd_clear(pgd);
206 }
207
208 void pud_clear_bad(pud_t *pud)
209 {
210         pud_ERROR(*pud);
211         pud_clear(pud);
212 }
213
214 void pmd_clear_bad(pmd_t *pmd)
215 {
216         pmd_ERROR(*pmd);
217         pmd_clear(pmd);
218 }
219
220 /*
221  * Note: this doesn't free the actual pages themselves. That
222  * has been handled earlier when unmapping all the memory regions.
223  */
224 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
225                            unsigned long addr)
226 {
227         pgtable_t token = pmd_pgtable(*pmd);
228         pmd_clear(pmd);
229         pte_free_tlb(tlb, token, addr);
230         tlb->mm->nr_ptes--;
231 }
232
233 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
234                                 unsigned long addr, unsigned long end,
235                                 unsigned long floor, unsigned long ceiling)
236 {
237         pmd_t *pmd;
238         unsigned long next;
239         unsigned long start;
240
241         start = addr;
242         pmd = pmd_offset(pud, addr);
243         do {
244                 next = pmd_addr_end(addr, end);
245                 if (pmd_none_or_clear_bad(pmd))
246                         continue;
247                 free_pte_range(tlb, pmd, addr);
248         } while (pmd++, addr = next, addr != end);
249
250         start &= PUD_MASK;
251         if (start < floor)
252                 return;
253         if (ceiling) {
254                 ceiling &= PUD_MASK;
255                 if (!ceiling)
256                         return;
257         }
258         if (end - 1 > ceiling - 1)
259                 return;
260
261         pmd = pmd_offset(pud, start);
262         pud_clear(pud);
263         pmd_free_tlb(tlb, pmd, start);
264 }
265
266 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
267                                 unsigned long addr, unsigned long end,
268                                 unsigned long floor, unsigned long ceiling)
269 {
270         pud_t *pud;
271         unsigned long next;
272         unsigned long start;
273
274         start = addr;
275         pud = pud_offset(pgd, addr);
276         do {
277                 next = pud_addr_end(addr, end);
278                 if (pud_none_or_clear_bad(pud))
279                         continue;
280                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
281         } while (pud++, addr = next, addr != end);
282
283         start &= PGDIR_MASK;
284         if (start < floor)
285                 return;
286         if (ceiling) {
287                 ceiling &= PGDIR_MASK;
288                 if (!ceiling)
289                         return;
290         }
291         if (end - 1 > ceiling - 1)
292                 return;
293
294         pud = pud_offset(pgd, start);
295         pgd_clear(pgd);
296         pud_free_tlb(tlb, pud, start);
297 }
298
299 /*
300  * This function frees user-level page tables of a process.
301  *
302  * Must be called with pagetable lock held.
303  */
304 void free_pgd_range(struct mmu_gather *tlb,
305                         unsigned long addr, unsigned long end,
306                         unsigned long floor, unsigned long ceiling)
307 {
308         pgd_t *pgd;
309         unsigned long next;
310
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         pgd = pgd_offset(tlb->mm, addr);
354         do {
355                 next = pgd_addr_end(addr, end);
356                 if (pgd_none_or_clear_bad(pgd))
357                         continue;
358                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
359         } while (pgd++, addr = next, addr != end);
360 }
361
362 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
363                 unsigned long floor, unsigned long ceiling)
364 {
365         while (vma) {
366                 struct vm_area_struct *next = vma->vm_next;
367                 unsigned long addr = vma->vm_start;
368
369                 /*
370                  * Hide vma from rmap and truncate_pagecache before freeing
371                  * pgtables
372                  */
373                 unlink_anon_vmas(vma);
374                 unlink_file_vma(vma);
375
376                 if (is_vm_hugetlb_page(vma)) {
377                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
378                                 floor, next? next->vm_start: ceiling);
379                 } else {
380                         /*
381                          * Optimization: gather nearby vmas into one call down
382                          */
383                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
384                                && !is_vm_hugetlb_page(next)) {
385                                 vma = next;
386                                 next = vma->vm_next;
387                                 unlink_anon_vmas(vma);
388                                 unlink_file_vma(vma);
389                         }
390                         free_pgd_range(tlb, addr, vma->vm_end,
391                                 floor, next? next->vm_start: ceiling);
392                 }
393                 vma = next;
394         }
395 }
396
397 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
398                 pmd_t *pmd, unsigned long address)
399 {
400         pgtable_t new = pte_alloc_one(mm, address);
401         int wait_split_huge_page;
402         if (!new)
403                 return -ENOMEM;
404
405         /*
406          * Ensure all pte setup (eg. pte page lock and page clearing) are
407          * visible before the pte is made visible to other CPUs by being
408          * put into page tables.
409          *
410          * The other side of the story is the pointer chasing in the page
411          * table walking code (when walking the page table without locking;
412          * ie. most of the time). Fortunately, these data accesses consist
413          * of a chain of data-dependent loads, meaning most CPUs (alpha
414          * being the notable exception) will already guarantee loads are
415          * seen in-order. See the alpha page table accessors for the
416          * smp_read_barrier_depends() barriers in page table walking code.
417          */
418         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
419
420         spin_lock(&mm->page_table_lock);
421         wait_split_huge_page = 0;
422         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
423                 mm->nr_ptes++;
424                 pmd_populate(mm, pmd, new);
425                 new = NULL;
426         } else if (unlikely(pmd_trans_splitting(*pmd)))
427                 wait_split_huge_page = 1;
428         spin_unlock(&mm->page_table_lock);
429         if (new)
430                 pte_free(mm, new);
431         if (wait_split_huge_page)
432                 wait_split_huge_page(vma->anon_vma, pmd);
433         return 0;
434 }
435
436 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
437 {
438         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
439         if (!new)
440                 return -ENOMEM;
441
442         smp_wmb(); /* See comment in __pte_alloc */
443
444         spin_lock(&init_mm.page_table_lock);
445         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
446                 pmd_populate_kernel(&init_mm, pmd, new);
447                 new = NULL;
448         } else
449                 VM_BUG_ON(pmd_trans_splitting(*pmd));
450         spin_unlock(&init_mm.page_table_lock);
451         if (new)
452                 pte_free_kernel(&init_mm, new);
453         return 0;
454 }
455
456 static inline void init_rss_vec(int *rss)
457 {
458         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
459 }
460
461 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
462 {
463         int i;
464
465         if (current->mm == mm)
466                 sync_mm_rss(current, mm);
467         for (i = 0; i < NR_MM_COUNTERS; i++)
468                 if (rss[i])
469                         add_mm_counter(mm, i, rss[i]);
470 }
471
472 /*
473  * This function is called to print an error when a bad pte
474  * is found. For example, we might have a PFN-mapped pte in
475  * a region that doesn't allow it.
476  *
477  * The calling function must still handle the error.
478  */
479 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
480                           pte_t pte, struct page *page)
481 {
482         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
483         pud_t *pud = pud_offset(pgd, addr);
484         pmd_t *pmd = pmd_offset(pud, addr);
485         struct address_space *mapping;
486         pgoff_t index;
487         static unsigned long resume;
488         static unsigned long nr_shown;
489         static unsigned long nr_unshown;
490
491         /*
492          * Allow a burst of 60 reports, then keep quiet for that minute;
493          * or allow a steady drip of one report per second.
494          */
495         if (nr_shown == 60) {
496                 if (time_before(jiffies, resume)) {
497                         nr_unshown++;
498                         return;
499                 }
500                 if (nr_unshown) {
501                         printk(KERN_ALERT
502                                 "BUG: Bad page map: %lu messages suppressed\n",
503                                 nr_unshown);
504                         nr_unshown = 0;
505                 }
506                 nr_shown = 0;
507         }
508         if (nr_shown++ == 0)
509                 resume = jiffies + 60 * HZ;
510
511         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
512         index = linear_page_index(vma, addr);
513
514         printk(KERN_ALERT
515                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
516                 current->comm,
517                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
518         if (page)
519                 dump_page(page);
520         printk(KERN_ALERT
521                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
522                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
523         /*
524          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
525          */
526         if (vma->vm_ops)
527                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
528                                 (unsigned long)vma->vm_ops->fault);
529         if (vma->vm_file && vma->vm_file->f_op)
530                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
531                                 (unsigned long)vma->vm_file->f_op->mmap);
532         dump_stack();
533         add_taint(TAINT_BAD_PAGE);
534 }
535
536 static inline int is_cow_mapping(unsigned int flags)
537 {
538         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
539 }
540
541 #ifndef is_zero_pfn
542 static inline int is_zero_pfn(unsigned long pfn)
543 {
544         return pfn == zero_pfn;
545 }
546 #endif
547
548 #ifndef my_zero_pfn
549 static inline unsigned long my_zero_pfn(unsigned long addr)
550 {
551         return zero_pfn;
552 }
553 #endif
554
555 /*
556  * vm_normal_page -- This function gets the "struct page" associated with a pte.
557  *
558  * "Special" mappings do not wish to be associated with a "struct page" (either
559  * it doesn't exist, or it exists but they don't want to touch it). In this
560  * case, NULL is returned here. "Normal" mappings do have a struct page.
561  *
562  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
563  * pte bit, in which case this function is trivial. Secondly, an architecture
564  * may not have a spare pte bit, which requires a more complicated scheme,
565  * described below.
566  *
567  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
568  * special mapping (even if there are underlying and valid "struct pages").
569  * COWed pages of a VM_PFNMAP are always normal.
570  *
571  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
572  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
573  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
574  * mapping will always honor the rule
575  *
576  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
577  *
578  * And for normal mappings this is false.
579  *
580  * This restricts such mappings to be a linear translation from virtual address
581  * to pfn. To get around this restriction, we allow arbitrary mappings so long
582  * as the vma is not a COW mapping; in that case, we know that all ptes are
583  * special (because none can have been COWed).
584  *
585  *
586  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
587  *
588  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
589  * page" backing, however the difference is that _all_ pages with a struct
590  * page (that is, those where pfn_valid is true) are refcounted and considered
591  * normal pages by the VM. The disadvantage is that pages are refcounted
592  * (which can be slower and simply not an option for some PFNMAP users). The
593  * advantage is that we don't have to follow the strict linearity rule of
594  * PFNMAP mappings in order to support COWable mappings.
595  *
596  */
597 #ifdef __HAVE_ARCH_PTE_SPECIAL
598 # define HAVE_PTE_SPECIAL 1
599 #else
600 # define HAVE_PTE_SPECIAL 0
601 #endif
602 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
603                                 pte_t pte)
604 {
605         unsigned long pfn = pte_pfn(pte);
606
607         if (HAVE_PTE_SPECIAL) {
608                 if (likely(!pte_special(pte)))
609                         goto check_pfn;
610                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
611                         return NULL;
612                 if (!is_zero_pfn(pfn))
613                         print_bad_pte(vma, addr, pte, NULL);
614                 return NULL;
615         }
616
617         /* !HAVE_PTE_SPECIAL case follows: */
618
619         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
620                 if (vma->vm_flags & VM_MIXEDMAP) {
621                         if (!pfn_valid(pfn))
622                                 return NULL;
623                         goto out;
624                 } else {
625                         unsigned long off;
626                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
627                         if (pfn == vma->vm_pgoff + off)
628                                 return NULL;
629                         if (!is_cow_mapping(vma->vm_flags))
630                                 return NULL;
631                 }
632         }
633
634         if (is_zero_pfn(pfn))
635                 return NULL;
636 check_pfn:
637         if (unlikely(pfn > highest_memmap_pfn)) {
638                 print_bad_pte(vma, addr, pte, NULL);
639                 return NULL;
640         }
641
642         /*
643          * NOTE! We still have PageReserved() pages in the page tables.
644          * eg. VDSO mappings can cause them to exist.
645          */
646 out:
647         return pfn_to_page(pfn);
648 }
649
650 /*
651  * copy one vm_area from one task to the other. Assumes the page tables
652  * already present in the new task to be cleared in the whole range
653  * covered by this vma.
654  */
655
656 static inline unsigned long
657 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
658                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
659                 unsigned long addr, int *rss)
660 {
661         unsigned long vm_flags = vma->vm_flags;
662         pte_t pte = *src_pte;
663         struct page *page;
664
665         /* pte contains position in swap or file, so copy. */
666         if (unlikely(!pte_present(pte))) {
667                 if (!pte_file(pte)) {
668                         swp_entry_t entry = pte_to_swp_entry(pte);
669
670                         if (swap_duplicate(entry) < 0)
671                                 return entry.val;
672
673                         /* make sure dst_mm is on swapoff's mmlist. */
674                         if (unlikely(list_empty(&dst_mm->mmlist))) {
675                                 spin_lock(&mmlist_lock);
676                                 if (list_empty(&dst_mm->mmlist))
677                                         list_add(&dst_mm->mmlist,
678                                                  &src_mm->mmlist);
679                                 spin_unlock(&mmlist_lock);
680                         }
681                         if (likely(!non_swap_entry(entry)))
682                                 rss[MM_SWAPENTS]++;
683                         else if (is_write_migration_entry(entry) &&
684                                         is_cow_mapping(vm_flags)) {
685                                 /*
686                                  * COW mappings require pages in both parent
687                                  * and child to be set to read.
688                                  */
689                                 make_migration_entry_read(&entry);
690                                 pte = swp_entry_to_pte(entry);
691                                 set_pte_at(src_mm, addr, src_pte, pte);
692                         }
693                 }
694                 goto out_set_pte;
695         }
696
697         /*
698          * If it's a COW mapping, write protect it both
699          * in the parent and the child
700          */
701         if (is_cow_mapping(vm_flags)) {
702                 ptep_set_wrprotect(src_mm, addr, src_pte);
703                 pte = pte_wrprotect(pte);
704         }
705
706         /*
707          * If it's a shared mapping, mark it clean in
708          * the child
709          */
710         if (vm_flags & VM_SHARED)
711                 pte = pte_mkclean(pte);
712         pte = pte_mkold(pte);
713
714         page = vm_normal_page(vma, addr, pte);
715         if (page) {
716                 get_page(page);
717                 page_dup_rmap(page);
718                 if (PageAnon(page))
719                         rss[MM_ANONPAGES]++;
720                 else
721                         rss[MM_FILEPAGES]++;
722         }
723
724 out_set_pte:
725         set_pte_at(dst_mm, addr, dst_pte, pte);
726         return 0;
727 }
728
729 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
730                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
731                    unsigned long addr, unsigned long end)
732 {
733         pte_t *orig_src_pte, *orig_dst_pte;
734         pte_t *src_pte, *dst_pte;
735         spinlock_t *src_ptl, *dst_ptl;
736         int progress = 0;
737         int rss[NR_MM_COUNTERS];
738         swp_entry_t entry = (swp_entry_t){0};
739
740 again:
741         init_rss_vec(rss);
742
743         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
744         if (!dst_pte)
745                 return -ENOMEM;
746         src_pte = pte_offset_map(src_pmd, addr);
747         src_ptl = pte_lockptr(src_mm, src_pmd);
748         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
749         orig_src_pte = src_pte;
750         orig_dst_pte = dst_pte;
751         arch_enter_lazy_mmu_mode();
752
753         do {
754                 /*
755                  * We are holding two locks at this point - either of them
756                  * could generate latencies in another task on another CPU.
757                  */
758                 if (progress >= 32) {
759                         progress = 0;
760                         if (need_resched() ||
761                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
762                                 break;
763                 }
764                 if (pte_none(*src_pte)) {
765                         progress++;
766                         continue;
767                 }
768                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
769                                                         vma, addr, rss);
770                 if (entry.val)
771                         break;
772                 progress += 8;
773         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
774
775         arch_leave_lazy_mmu_mode();
776         spin_unlock(src_ptl);
777         pte_unmap(orig_src_pte);
778         add_mm_rss_vec(dst_mm, rss);
779         pte_unmap_unlock(orig_dst_pte, dst_ptl);
780         cond_resched();
781
782         if (entry.val) {
783                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
784                         return -ENOMEM;
785                 progress = 0;
786         }
787         if (addr != end)
788                 goto again;
789         return 0;
790 }
791
792 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
793                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
794                 unsigned long addr, unsigned long end)
795 {
796         pmd_t *src_pmd, *dst_pmd;
797         unsigned long next;
798
799         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
800         if (!dst_pmd)
801                 return -ENOMEM;
802         src_pmd = pmd_offset(src_pud, addr);
803         do {
804                 next = pmd_addr_end(addr, end);
805                 if (pmd_trans_huge(*src_pmd)) {
806                         int err;
807                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
808                         err = copy_huge_pmd(dst_mm, src_mm,
809                                             dst_pmd, src_pmd, addr, vma);
810                         if (err == -ENOMEM)
811                                 return -ENOMEM;
812                         if (!err)
813                                 continue;
814                         /* fall through */
815                 }
816                 if (pmd_none_or_clear_bad(src_pmd))
817                         continue;
818                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
819                                                 vma, addr, next))
820                         return -ENOMEM;
821         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
822         return 0;
823 }
824
825 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
826                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
827                 unsigned long addr, unsigned long end)
828 {
829         pud_t *src_pud, *dst_pud;
830         unsigned long next;
831
832         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
833         if (!dst_pud)
834                 return -ENOMEM;
835         src_pud = pud_offset(src_pgd, addr);
836         do {
837                 next = pud_addr_end(addr, end);
838                 if (pud_none_or_clear_bad(src_pud))
839                         continue;
840                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
841                                                 vma, addr, next))
842                         return -ENOMEM;
843         } while (dst_pud++, src_pud++, addr = next, addr != end);
844         return 0;
845 }
846
847 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
848                 struct vm_area_struct *vma)
849 {
850         pgd_t *src_pgd, *dst_pgd;
851         unsigned long next;
852         unsigned long addr = vma->vm_start;
853         unsigned long end = vma->vm_end;
854         int ret;
855
856         /*
857          * Don't copy ptes where a page fault will fill them correctly.
858          * Fork becomes much lighter when there are big shared or private
859          * readonly mappings. The tradeoff is that copy_page_range is more
860          * efficient than faulting.
861          */
862         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
863                 if (!vma->anon_vma)
864                         return 0;
865         }
866
867         if (is_vm_hugetlb_page(vma))
868                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
869
870         if (unlikely(is_pfn_mapping(vma))) {
871                 /*
872                  * We do not free on error cases below as remove_vma
873                  * gets called on error from higher level routine
874                  */
875                 ret = track_pfn_vma_copy(vma);
876                 if (ret)
877                         return ret;
878         }
879
880         /*
881          * We need to invalidate the secondary MMU mappings only when
882          * there could be a permission downgrade on the ptes of the
883          * parent mm. And a permission downgrade will only happen if
884          * is_cow_mapping() returns true.
885          */
886         if (is_cow_mapping(vma->vm_flags))
887                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
888
889         ret = 0;
890         dst_pgd = pgd_offset(dst_mm, addr);
891         src_pgd = pgd_offset(src_mm, addr);
892         do {
893                 next = pgd_addr_end(addr, end);
894                 if (pgd_none_or_clear_bad(src_pgd))
895                         continue;
896                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
897                                             vma, addr, next))) {
898                         ret = -ENOMEM;
899                         break;
900                 }
901         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
902
903         if (is_cow_mapping(vma->vm_flags))
904                 mmu_notifier_invalidate_range_end(src_mm,
905                                                   vma->vm_start, end);
906         return ret;
907 }
908
909 static unsigned long zap_pte_range(struct mmu_gather *tlb,
910                                 struct vm_area_struct *vma, pmd_t *pmd,
911                                 unsigned long addr, unsigned long end,
912                                 long *zap_work, struct zap_details *details)
913 {
914         struct mm_struct *mm = tlb->mm;
915         int force_flush = 0;
916         pte_t *pte;
917         spinlock_t *ptl;
918         int rss[NR_MM_COUNTERS];
919
920         init_rss_vec(rss);
921 again:
922         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
923         arch_enter_lazy_mmu_mode();
924         do {
925                 pte_t ptent = *pte;
926                 if (pte_none(ptent)) {
927                         (*zap_work)--;
928                         continue;
929                 }
930
931                 (*zap_work) -= PAGE_SIZE;
932
933                 if (pte_present(ptent)) {
934                         struct page *page;
935
936                         page = vm_normal_page(vma, addr, ptent);
937                         if (unlikely(details) && page) {
938                                 /*
939                                  * unmap_shared_mapping_pages() wants to
940                                  * invalidate cache without truncating:
941                                  * unmap shared but keep private pages.
942                                  */
943                                 if (details->check_mapping &&
944                                     details->check_mapping != page->mapping)
945                                         continue;
946                                 /*
947                                  * Each page->index must be checked when
948                                  * invalidating or truncating nonlinear.
949                                  */
950                                 if (details->nonlinear_vma &&
951                                     (page->index < details->first_index ||
952                                      page->index > details->last_index))
953                                         continue;
954                         }
955                         ptent = ptep_get_and_clear_full(mm, addr, pte,
956                                                         tlb->fullmm);
957                         tlb_remove_tlb_entry(tlb, pte, addr);
958                         if (unlikely(!page))
959                                 continue;
960                         if (unlikely(details) && details->nonlinear_vma
961                             && linear_page_index(details->nonlinear_vma,
962                                                 addr) != page->index)
963                                 set_pte_at(mm, addr, pte,
964                                            pgoff_to_pte(page->index));
965                         if (PageAnon(page))
966                                 rss[MM_ANONPAGES]--;
967                         else {
968                                 if (pte_dirty(ptent))
969                                         set_page_dirty(page);
970                                 if (pte_young(ptent) &&
971                                     likely(!VM_SequentialReadHint(vma)))
972                                         mark_page_accessed(page);
973                                 rss[MM_FILEPAGES]--;
974                         }
975                         page_remove_rmap(page);
976                         if (unlikely(page_mapcount(page) < 0))
977                                 print_bad_pte(vma, addr, ptent, page);
978                         force_flush = !__tlb_remove_page(tlb, page);
979                         if (force_flush)
980                                 break;
981                         continue;
982                 }
983                 /*
984                  * If details->check_mapping, we leave swap entries;
985                  * if details->nonlinear_vma, we leave file entries.
986                  */
987                 if (unlikely(details))
988                         continue;
989                 if (pte_file(ptent)) {
990                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
991                                 print_bad_pte(vma, addr, ptent, NULL);
992                 } else {
993                         swp_entry_t entry = pte_to_swp_entry(ptent);
994
995                         if (!non_swap_entry(entry))
996                                 rss[MM_SWAPENTS]--;
997                         if (unlikely(!free_swap_and_cache(entry)))
998                                 print_bad_pte(vma, addr, ptent, NULL);
999                 }
1000                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1001         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
1002
1003         add_mm_rss_vec(mm, rss);
1004         arch_leave_lazy_mmu_mode();
1005         pte_unmap_unlock(pte - 1, ptl);
1006
1007         /*
1008          * mmu_gather ran out of room to batch pages, we break out of
1009          * the PTE lock to avoid doing the potential expensive TLB invalidate
1010          * and page-free while holding it.
1011          */
1012         if (force_flush) {
1013                 force_flush = 0;
1014                 tlb_flush_mmu(tlb);
1015                 if (addr != end)
1016                         goto again;
1017         }
1018
1019         return addr;
1020 }
1021
1022 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1023                                 struct vm_area_struct *vma, pud_t *pud,
1024                                 unsigned long addr, unsigned long end,
1025                                 long *zap_work, struct zap_details *details)
1026 {
1027         pmd_t *pmd;
1028         unsigned long next;
1029
1030         pmd = pmd_offset(pud, addr);
1031         do {
1032                 next = pmd_addr_end(addr, end);
1033                 if (pmd_trans_huge(*pmd)) {
1034                         if (next-addr != HPAGE_PMD_SIZE) {
1035                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1036                                 split_huge_page_pmd(vma->vm_mm, pmd);
1037                         } else if (zap_huge_pmd(tlb, vma, pmd)) {
1038                                 (*zap_work)--;
1039                                 continue;
1040                         }
1041                         /* fall through */
1042                 }
1043                 if (pmd_none_or_clear_bad(pmd)) {
1044                         (*zap_work)--;
1045                         continue;
1046                 }
1047                 next = zap_pte_range(tlb, vma, pmd, addr, next,
1048                                                 zap_work, details);
1049         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
1050
1051         return addr;
1052 }
1053
1054 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1055                                 struct vm_area_struct *vma, pgd_t *pgd,
1056                                 unsigned long addr, unsigned long end,
1057                                 long *zap_work, struct zap_details *details)
1058 {
1059         pud_t *pud;
1060         unsigned long next;
1061
1062         pud = pud_offset(pgd, addr);
1063         do {
1064                 next = pud_addr_end(addr, end);
1065                 if (pud_none_or_clear_bad(pud)) {
1066                         (*zap_work)--;
1067                         continue;
1068                 }
1069                 next = zap_pmd_range(tlb, vma, pud, addr, next,
1070                                                 zap_work, details);
1071         } while (pud++, addr = next, (addr != end && *zap_work > 0));
1072
1073         return addr;
1074 }
1075
1076 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1077                                 struct vm_area_struct *vma,
1078                                 unsigned long addr, unsigned long end,
1079                                 long *zap_work, struct zap_details *details)
1080 {
1081         pgd_t *pgd;
1082         unsigned long next;
1083
1084         if (details && !details->check_mapping && !details->nonlinear_vma)
1085                 details = NULL;
1086
1087         BUG_ON(addr >= end);
1088         mem_cgroup_uncharge_start();
1089         tlb_start_vma(tlb, vma);
1090         pgd = pgd_offset(vma->vm_mm, addr);
1091         do {
1092                 next = pgd_addr_end(addr, end);
1093                 if (pgd_none_or_clear_bad(pgd)) {
1094                         (*zap_work)--;
1095                         continue;
1096                 }
1097                 next = zap_pud_range(tlb, vma, pgd, addr, next,
1098                                                 zap_work, details);
1099         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
1100         tlb_end_vma(tlb, vma);
1101         mem_cgroup_uncharge_end();
1102
1103         return addr;
1104 }
1105
1106 #ifdef CONFIG_PREEMPT
1107 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1108 #else
1109 /* No preempt: go for improved straight-line efficiency */
1110 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1111 #endif
1112
1113 /**
1114  * unmap_vmas - unmap a range of memory covered by a list of vma's
1115  * @tlbp: address of the caller's struct mmu_gather
1116  * @vma: the starting vma
1117  * @start_addr: virtual address at which to start unmapping
1118  * @end_addr: virtual address at which to end unmapping
1119  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1120  * @details: details of nonlinear truncation or shared cache invalidation
1121  *
1122  * Returns the end address of the unmapping (restart addr if interrupted).
1123  *
1124  * Unmap all pages in the vma list.
1125  *
1126  * We aim to not hold locks for too long (for scheduling latency reasons).
1127  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
1128  * return the ending mmu_gather to the caller.
1129  *
1130  * Only addresses between `start' and `end' will be unmapped.
1131  *
1132  * The VMA list must be sorted in ascending virtual address order.
1133  *
1134  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1135  * range after unmap_vmas() returns.  So the only responsibility here is to
1136  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1137  * drops the lock and schedules.
1138  */
1139 unsigned long unmap_vmas(struct mmu_gather *tlb,
1140                 struct vm_area_struct *vma, unsigned long start_addr,
1141                 unsigned long end_addr, unsigned long *nr_accounted,
1142                 struct zap_details *details)
1143 {
1144         long zap_work = ZAP_BLOCK_SIZE;
1145         unsigned long start = start_addr;
1146         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1147         struct mm_struct *mm = vma->vm_mm;
1148
1149         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1150         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1151                 unsigned long end;
1152
1153                 start = max(vma->vm_start, start_addr);
1154                 if (start >= vma->vm_end)
1155                         continue;
1156                 end = min(vma->vm_end, end_addr);
1157                 if (end <= vma->vm_start)
1158                         continue;
1159
1160                 if (vma->vm_flags & VM_ACCOUNT)
1161                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1162
1163                 if (unlikely(is_pfn_mapping(vma)))
1164                         untrack_pfn_vma(vma, 0, 0);
1165
1166                 while (start != end) {
1167                         if (unlikely(is_vm_hugetlb_page(vma))) {
1168                                 /*
1169                                  * It is undesirable to test vma->vm_file as it
1170                                  * should be non-null for valid hugetlb area.
1171                                  * However, vm_file will be NULL in the error
1172                                  * cleanup path of do_mmap_pgoff. When
1173                                  * hugetlbfs ->mmap method fails,
1174                                  * do_mmap_pgoff() nullifies vma->vm_file
1175                                  * before calling this function to clean up.
1176                                  * Since no pte has actually been setup, it is
1177                                  * safe to do nothing in this case.
1178                                  */
1179                                 if (vma->vm_file) {
1180                                         unmap_hugepage_range(vma, start, end, NULL);
1181                                         zap_work -= (end - start) /
1182                                         pages_per_huge_page(hstate_vma(vma));
1183                                 }
1184
1185                                 start = end;
1186                         } else
1187                                 start = unmap_page_range(tlb, vma,
1188                                                 start, end, &zap_work, details);
1189
1190                         if (zap_work > 0) {
1191                                 BUG_ON(start != end);
1192                                 break;
1193                         }
1194
1195                         if (need_resched() ||
1196                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1197                                 if (i_mmap_lock)
1198                                         goto out;
1199                                 cond_resched();
1200                         }
1201
1202                         zap_work = ZAP_BLOCK_SIZE;
1203                 }
1204         }
1205 out:
1206         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1207         return start;   /* which is now the end (or restart) address */
1208 }
1209
1210 /**
1211  * zap_page_range - remove user pages in a given range
1212  * @vma: vm_area_struct holding the applicable pages
1213  * @address: starting address of pages to zap
1214  * @size: number of bytes to zap
1215  * @details: details of nonlinear truncation or shared cache invalidation
1216  */
1217 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1218                 unsigned long size, struct zap_details *details)
1219 {
1220         struct mm_struct *mm = vma->vm_mm;
1221         struct mmu_gather tlb;
1222         unsigned long end = address + size;
1223         unsigned long nr_accounted = 0;
1224
1225         lru_add_drain();
1226         tlb_gather_mmu(&tlb, mm, 0);
1227         update_hiwater_rss(mm);
1228         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1229         tlb_finish_mmu(&tlb, address, end);
1230         return end;
1231 }
1232
1233 /**
1234  * zap_vma_ptes - remove ptes mapping the vma
1235  * @vma: vm_area_struct holding ptes to be zapped
1236  * @address: starting address of pages to zap
1237  * @size: number of bytes to zap
1238  *
1239  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1240  *
1241  * The entire address range must be fully contained within the vma.
1242  *
1243  * Returns 0 if successful.
1244  */
1245 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1246                 unsigned long size)
1247 {
1248         if (address < vma->vm_start || address + size > vma->vm_end ||
1249                         !(vma->vm_flags & VM_PFNMAP))
1250                 return -1;
1251         zap_page_range(vma, address, size, NULL);
1252         return 0;
1253 }
1254 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1255
1256 /**
1257  * follow_page - look up a page descriptor from a user-virtual address
1258  * @vma: vm_area_struct mapping @address
1259  * @address: virtual address to look up
1260  * @flags: flags modifying lookup behaviour
1261  *
1262  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1263  *
1264  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1265  * an error pointer if there is a mapping to something not represented
1266  * by a page descriptor (see also vm_normal_page()).
1267  */
1268 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1269                         unsigned int flags)
1270 {
1271         pgd_t *pgd;
1272         pud_t *pud;
1273         pmd_t *pmd;
1274         pte_t *ptep, pte;
1275         spinlock_t *ptl;
1276         struct page *page;
1277         struct mm_struct *mm = vma->vm_mm;
1278
1279         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1280         if (!IS_ERR(page)) {
1281                 BUG_ON(flags & FOLL_GET);
1282                 goto out;
1283         }
1284
1285         page = NULL;
1286         pgd = pgd_offset(mm, address);
1287         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1288                 goto no_page_table;
1289
1290         pud = pud_offset(pgd, address);
1291         if (pud_none(*pud))
1292                 goto no_page_table;
1293         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1294                 BUG_ON(flags & FOLL_GET);
1295                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1296                 goto out;
1297         }
1298         if (unlikely(pud_bad(*pud)))
1299                 goto no_page_table;
1300
1301         pmd = pmd_offset(pud, address);
1302         if (pmd_none(*pmd))
1303                 goto no_page_table;
1304         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1305                 BUG_ON(flags & FOLL_GET);
1306                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1307                 goto out;
1308         }
1309         if (pmd_trans_huge(*pmd)) {
1310                 if (flags & FOLL_SPLIT) {
1311                         split_huge_page_pmd(mm, pmd);
1312                         goto split_fallthrough;
1313                 }
1314                 spin_lock(&mm->page_table_lock);
1315                 if (likely(pmd_trans_huge(*pmd))) {
1316                         if (unlikely(pmd_trans_splitting(*pmd))) {
1317                                 spin_unlock(&mm->page_table_lock);
1318                                 wait_split_huge_page(vma->anon_vma, pmd);
1319                         } else {
1320                                 page = follow_trans_huge_pmd(mm, address,
1321                                                              pmd, flags);
1322                                 spin_unlock(&mm->page_table_lock);
1323                                 goto out;
1324                         }
1325                 } else
1326                         spin_unlock(&mm->page_table_lock);
1327                 /* fall through */
1328         }
1329 split_fallthrough:
1330         if (unlikely(pmd_bad(*pmd)))
1331                 goto no_page_table;
1332
1333         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1334
1335         pte = *ptep;
1336         if (!pte_present(pte))
1337                 goto no_page;
1338         if ((flags & FOLL_WRITE) && !pte_write(pte))
1339                 goto unlock;
1340
1341         page = vm_normal_page(vma, address, pte);
1342         if (unlikely(!page)) {
1343                 if ((flags & FOLL_DUMP) ||
1344                     !is_zero_pfn(pte_pfn(pte)))
1345                         goto bad_page;
1346                 page = pte_page(pte);
1347         }
1348
1349         if (flags & FOLL_GET)
1350                 get_page(page);
1351         if (flags & FOLL_TOUCH) {
1352                 if ((flags & FOLL_WRITE) &&
1353                     !pte_dirty(pte) && !PageDirty(page))
1354                         set_page_dirty(page);
1355                 /*
1356                  * pte_mkyoung() would be more correct here, but atomic care
1357                  * is needed to avoid losing the dirty bit: it is easier to use
1358                  * mark_page_accessed().
1359                  */
1360                 mark_page_accessed(page);
1361         }
1362         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1363                 /*
1364                  * The preliminary mapping check is mainly to avoid the
1365                  * pointless overhead of lock_page on the ZERO_PAGE
1366                  * which might bounce very badly if there is contention.
1367                  *
1368                  * If the page is already locked, we don't need to
1369                  * handle it now - vmscan will handle it later if and
1370                  * when it attempts to reclaim the page.
1371                  */
1372                 if (page->mapping && trylock_page(page)) {
1373                         lru_add_drain();  /* push cached pages to LRU */
1374                         /*
1375                          * Because we lock page here and migration is
1376                          * blocked by the pte's page reference, we need
1377                          * only check for file-cache page truncation.
1378                          */
1379                         if (page->mapping)
1380                                 mlock_vma_page(page);
1381                         unlock_page(page);
1382                 }
1383         }
1384 unlock:
1385         pte_unmap_unlock(ptep, ptl);
1386 out:
1387         return page;
1388
1389 bad_page:
1390         pte_unmap_unlock(ptep, ptl);
1391         return ERR_PTR(-EFAULT);
1392
1393 no_page:
1394         pte_unmap_unlock(ptep, ptl);
1395         if (!pte_none(pte))
1396                 return page;
1397
1398 no_page_table:
1399         /*
1400          * When core dumping an enormous anonymous area that nobody
1401          * has touched so far, we don't want to allocate unnecessary pages or
1402          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1403          * then get_dump_page() will return NULL to leave a hole in the dump.
1404          * But we can only make this optimization where a hole would surely
1405          * be zero-filled if handle_mm_fault() actually did handle it.
1406          */
1407         if ((flags & FOLL_DUMP) &&
1408             (!vma->vm_ops || !vma->vm_ops->fault))
1409                 return ERR_PTR(-EFAULT);
1410         return page;
1411 }
1412
1413 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1414 {
1415         return stack_guard_page_start(vma, addr) ||
1416                stack_guard_page_end(vma, addr+PAGE_SIZE);
1417 }
1418
1419 /**
1420  * __get_user_pages() - pin user pages in memory
1421  * @tsk:        task_struct of target task
1422  * @mm:         mm_struct of target mm
1423  * @start:      starting user address
1424  * @nr_pages:   number of pages from start to pin
1425  * @gup_flags:  flags modifying pin behaviour
1426  * @pages:      array that receives pointers to the pages pinned.
1427  *              Should be at least nr_pages long. Or NULL, if caller
1428  *              only intends to ensure the pages are faulted in.
1429  * @vmas:       array of pointers to vmas corresponding to each page.
1430  *              Or NULL if the caller does not require them.
1431  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1432  *
1433  * Returns number of pages pinned. This may be fewer than the number
1434  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1435  * were pinned, returns -errno. Each page returned must be released
1436  * with a put_page() call when it is finished with. vmas will only
1437  * remain valid while mmap_sem is held.
1438  *
1439  * Must be called with mmap_sem held for read or write.
1440  *
1441  * __get_user_pages walks a process's page tables and takes a reference to
1442  * each struct page that each user address corresponds to at a given
1443  * instant. That is, it takes the page that would be accessed if a user
1444  * thread accesses the given user virtual address at that instant.
1445  *
1446  * This does not guarantee that the page exists in the user mappings when
1447  * __get_user_pages returns, and there may even be a completely different
1448  * page there in some cases (eg. if mmapped pagecache has been invalidated
1449  * and subsequently re faulted). However it does guarantee that the page
1450  * won't be freed completely. And mostly callers simply care that the page
1451  * contains data that was valid *at some point in time*. Typically, an IO
1452  * or similar operation cannot guarantee anything stronger anyway because
1453  * locks can't be held over the syscall boundary.
1454  *
1455  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1456  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1457  * appropriate) must be called after the page is finished with, and
1458  * before put_page is called.
1459  *
1460  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1461  * or mmap_sem contention, and if waiting is needed to pin all pages,
1462  * *@nonblocking will be set to 0.
1463  *
1464  * In most cases, get_user_pages or get_user_pages_fast should be used
1465  * instead of __get_user_pages. __get_user_pages should be used only if
1466  * you need some special @gup_flags.
1467  */
1468 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1469                      unsigned long start, int nr_pages, unsigned int gup_flags,
1470                      struct page **pages, struct vm_area_struct **vmas,
1471                      int *nonblocking)
1472 {
1473         int i;
1474         unsigned long vm_flags;
1475
1476         if (nr_pages <= 0)
1477                 return 0;
1478
1479         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1480
1481         /* 
1482          * Require read or write permissions.
1483          * If FOLL_FORCE is set, we only require the "MAY" flags.
1484          */
1485         vm_flags  = (gup_flags & FOLL_WRITE) ?
1486                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1487         vm_flags &= (gup_flags & FOLL_FORCE) ?
1488                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1489         i = 0;
1490
1491         do {
1492                 struct vm_area_struct *vma;
1493
1494                 vma = find_extend_vma(mm, start);
1495                 if (!vma && in_gate_area(mm, start)) {
1496                         unsigned long pg = start & PAGE_MASK;
1497                         pgd_t *pgd;
1498                         pud_t *pud;
1499                         pmd_t *pmd;
1500                         pte_t *pte;
1501
1502                         /* user gate pages are read-only */
1503                         if (gup_flags & FOLL_WRITE)
1504                                 return i ? : -EFAULT;
1505                         if (pg > TASK_SIZE)
1506                                 pgd = pgd_offset_k(pg);
1507                         else
1508                                 pgd = pgd_offset_gate(mm, pg);
1509                         BUG_ON(pgd_none(*pgd));
1510                         pud = pud_offset(pgd, pg);
1511                         BUG_ON(pud_none(*pud));
1512                         pmd = pmd_offset(pud, pg);
1513                         if (pmd_none(*pmd))
1514                                 return i ? : -EFAULT;
1515                         VM_BUG_ON(pmd_trans_huge(*pmd));
1516                         pte = pte_offset_map(pmd, pg);
1517                         if (pte_none(*pte)) {
1518                                 pte_unmap(pte);
1519                                 return i ? : -EFAULT;
1520                         }
1521                         vma = get_gate_vma(mm);
1522                         if (pages) {
1523                                 struct page *page;
1524
1525                                 page = vm_normal_page(vma, start, *pte);
1526                                 if (!page) {
1527                                         if (!(gup_flags & FOLL_DUMP) &&
1528                                              is_zero_pfn(pte_pfn(*pte)))
1529                                                 page = pte_page(*pte);
1530                                         else {
1531                                                 pte_unmap(pte);
1532                                                 return i ? : -EFAULT;
1533                                         }
1534                                 }
1535                                 pages[i] = page;
1536                                 get_page(page);
1537                         }
1538                         pte_unmap(pte);
1539                         goto next_page;
1540                 }
1541
1542                 if (!vma ||
1543                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1544                     !(vm_flags & vma->vm_flags))
1545                         return i ? : -EFAULT;
1546
1547                 if (is_vm_hugetlb_page(vma)) {
1548                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1549                                         &start, &nr_pages, i, gup_flags);
1550                         continue;
1551                 }
1552
1553                 do {
1554                         struct page *page;
1555                         unsigned int foll_flags = gup_flags;
1556
1557                         /*
1558                          * If we have a pending SIGKILL, don't keep faulting
1559                          * pages and potentially allocating memory.
1560                          */
1561                         if (unlikely(fatal_signal_pending(current)))
1562                                 return i ? i : -ERESTARTSYS;
1563
1564                         cond_resched();
1565                         while (!(page = follow_page(vma, start, foll_flags))) {
1566                                 int ret;
1567                                 unsigned int fault_flags = 0;
1568
1569                                 /* For mlock, just skip the stack guard page. */
1570                                 if (foll_flags & FOLL_MLOCK) {
1571                                         if (stack_guard_page(vma, start))
1572                                                 goto next_page;
1573                                 }
1574                                 if (foll_flags & FOLL_WRITE)
1575                                         fault_flags |= FAULT_FLAG_WRITE;
1576                                 if (nonblocking)
1577                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1578                                 if (foll_flags & FOLL_NOWAIT)
1579                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1580
1581                                 ret = handle_mm_fault(mm, vma, start,
1582                                                         fault_flags);
1583
1584                                 if (ret & VM_FAULT_ERROR) {
1585                                         if (ret & VM_FAULT_OOM)
1586                                                 return i ? i : -ENOMEM;
1587                                         if (ret & (VM_FAULT_HWPOISON |
1588                                                    VM_FAULT_HWPOISON_LARGE)) {
1589                                                 if (i)
1590                                                         return i;
1591                                                 else if (gup_flags & FOLL_HWPOISON)
1592                                                         return -EHWPOISON;
1593                                                 else
1594                                                         return -EFAULT;
1595                                         }
1596                                         if (ret & VM_FAULT_SIGBUS)
1597                                                 return i ? i : -EFAULT;
1598                                         BUG();
1599                                 }
1600
1601                                 if (tsk) {
1602                                         if (ret & VM_FAULT_MAJOR)
1603                                                 tsk->maj_flt++;
1604                                         else
1605                                                 tsk->min_flt++;
1606                                 }
1607
1608                                 if (ret & VM_FAULT_RETRY) {
1609                                         if (nonblocking)
1610                                                 *nonblocking = 0;
1611                                         return i;
1612                                 }
1613
1614                                 /*
1615                                  * The VM_FAULT_WRITE bit tells us that
1616                                  * do_wp_page has broken COW when necessary,
1617                                  * even if maybe_mkwrite decided not to set
1618                                  * pte_write. We can thus safely do subsequent
1619                                  * page lookups as if they were reads. But only
1620                                  * do so when looping for pte_write is futile:
1621                                  * in some cases userspace may also be wanting
1622                                  * to write to the gotten user page, which a
1623                                  * read fault here might prevent (a readonly
1624                                  * page might get reCOWed by userspace write).
1625                                  */
1626                                 if ((ret & VM_FAULT_WRITE) &&
1627                                     !(vma->vm_flags & VM_WRITE))
1628                                         foll_flags &= ~FOLL_WRITE;
1629
1630                                 cond_resched();
1631                         }
1632                         if (IS_ERR(page))
1633                                 return i ? i : PTR_ERR(page);
1634                         if (pages) {
1635                                 pages[i] = page;
1636
1637                                 flush_anon_page(vma, page, start);
1638                                 flush_dcache_page(page);
1639                         }
1640 next_page:
1641                         if (vmas)
1642                                 vmas[i] = vma;
1643                         i++;
1644                         start += PAGE_SIZE;
1645                         nr_pages--;
1646                 } while (nr_pages && start < vma->vm_end);
1647         } while (nr_pages);
1648         return i;
1649 }
1650 EXPORT_SYMBOL(__get_user_pages);
1651
1652 /**
1653  * get_user_pages() - pin user pages in memory
1654  * @tsk:        the task_struct to use for page fault accounting, or
1655  *              NULL if faults are not to be recorded.
1656  * @mm:         mm_struct of target mm
1657  * @start:      starting user address
1658  * @nr_pages:   number of pages from start to pin
1659  * @write:      whether pages will be written to by the caller
1660  * @force:      whether to force write access even if user mapping is
1661  *              readonly. This will result in the page being COWed even
1662  *              in MAP_SHARED mappings. You do not want this.
1663  * @pages:      array that receives pointers to the pages pinned.
1664  *              Should be at least nr_pages long. Or NULL, if caller
1665  *              only intends to ensure the pages are faulted in.
1666  * @vmas:       array of pointers to vmas corresponding to each page.
1667  *              Or NULL if the caller does not require them.
1668  *
1669  * Returns number of pages pinned. This may be fewer than the number
1670  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1671  * were pinned, returns -errno. Each page returned must be released
1672  * with a put_page() call when it is finished with. vmas will only
1673  * remain valid while mmap_sem is held.
1674  *
1675  * Must be called with mmap_sem held for read or write.
1676  *
1677  * get_user_pages walks a process's page tables and takes a reference to
1678  * each struct page that each user address corresponds to at a given
1679  * instant. That is, it takes the page that would be accessed if a user
1680  * thread accesses the given user virtual address at that instant.
1681  *
1682  * This does not guarantee that the page exists in the user mappings when
1683  * get_user_pages returns, and there may even be a completely different
1684  * page there in some cases (eg. if mmapped pagecache has been invalidated
1685  * and subsequently re faulted). However it does guarantee that the page
1686  * won't be freed completely. And mostly callers simply care that the page
1687  * contains data that was valid *at some point in time*. Typically, an IO
1688  * or similar operation cannot guarantee anything stronger anyway because
1689  * locks can't be held over the syscall boundary.
1690  *
1691  * If write=0, the page must not be written to. If the page is written to,
1692  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1693  * after the page is finished with, and before put_page is called.
1694  *
1695  * get_user_pages is typically used for fewer-copy IO operations, to get a
1696  * handle on the memory by some means other than accesses via the user virtual
1697  * addresses. The pages may be submitted for DMA to devices or accessed via
1698  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1699  * use the correct cache flushing APIs.
1700  *
1701  * See also get_user_pages_fast, for performance critical applications.
1702  */
1703 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1704                 unsigned long start, int nr_pages, int write, int force,
1705                 struct page **pages, struct vm_area_struct **vmas)
1706 {
1707         int flags = FOLL_TOUCH;
1708
1709         if (pages)
1710                 flags |= FOLL_GET;
1711         if (write)
1712                 flags |= FOLL_WRITE;
1713         if (force)
1714                 flags |= FOLL_FORCE;
1715
1716         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1717                                 NULL);
1718 }
1719 EXPORT_SYMBOL(get_user_pages);
1720
1721 /**
1722  * get_dump_page() - pin user page in memory while writing it to core dump
1723  * @addr: user address
1724  *
1725  * Returns struct page pointer of user page pinned for dump,
1726  * to be freed afterwards by page_cache_release() or put_page().
1727  *
1728  * Returns NULL on any kind of failure - a hole must then be inserted into
1729  * the corefile, to preserve alignment with its headers; and also returns
1730  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1731  * allowing a hole to be left in the corefile to save diskspace.
1732  *
1733  * Called without mmap_sem, but after all other threads have been killed.
1734  */
1735 #ifdef CONFIG_ELF_CORE
1736 struct page *get_dump_page(unsigned long addr)
1737 {
1738         struct vm_area_struct *vma;
1739         struct page *page;
1740
1741         if (__get_user_pages(current, current->mm, addr, 1,
1742                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1743                              NULL) < 1)
1744                 return NULL;
1745         flush_cache_page(vma, addr, page_to_pfn(page));
1746         return page;
1747 }
1748 #endif /* CONFIG_ELF_CORE */
1749
1750 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1751                         spinlock_t **ptl)
1752 {
1753         pgd_t * pgd = pgd_offset(mm, addr);
1754         pud_t * pud = pud_alloc(mm, pgd, addr);
1755         if (pud) {
1756                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1757                 if (pmd) {
1758                         VM_BUG_ON(pmd_trans_huge(*pmd));
1759                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1760                 }
1761         }
1762         return NULL;
1763 }
1764
1765 /*
1766  * This is the old fallback for page remapping.
1767  *
1768  * For historical reasons, it only allows reserved pages. Only
1769  * old drivers should use this, and they needed to mark their
1770  * pages reserved for the old functions anyway.
1771  */
1772 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1773                         struct page *page, pgprot_t prot)
1774 {
1775         struct mm_struct *mm = vma->vm_mm;
1776         int retval;
1777         pte_t *pte;
1778         spinlock_t *ptl;
1779
1780         retval = -EINVAL;
1781         if (PageAnon(page))
1782                 goto out;
1783         retval = -ENOMEM;
1784         flush_dcache_page(page);
1785         pte = get_locked_pte(mm, addr, &ptl);
1786         if (!pte)
1787                 goto out;
1788         retval = -EBUSY;
1789         if (!pte_none(*pte))
1790                 goto out_unlock;
1791
1792         /* Ok, finally just insert the thing.. */
1793         get_page(page);
1794         inc_mm_counter_fast(mm, MM_FILEPAGES);
1795         page_add_file_rmap(page);
1796         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1797
1798         retval = 0;
1799         pte_unmap_unlock(pte, ptl);
1800         return retval;
1801 out_unlock:
1802         pte_unmap_unlock(pte, ptl);
1803 out:
1804         return retval;
1805 }
1806
1807 /**
1808  * vm_insert_page - insert single page into user vma
1809  * @vma: user vma to map to
1810  * @addr: target user address of this page
1811  * @page: source kernel page
1812  *
1813  * This allows drivers to insert individual pages they've allocated
1814  * into a user vma.
1815  *
1816  * The page has to be a nice clean _individual_ kernel allocation.
1817  * If you allocate a compound page, you need to have marked it as
1818  * such (__GFP_COMP), or manually just split the page up yourself
1819  * (see split_page()).
1820  *
1821  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1822  * took an arbitrary page protection parameter. This doesn't allow
1823  * that. Your vma protection will have to be set up correctly, which
1824  * means that if you want a shared writable mapping, you'd better
1825  * ask for a shared writable mapping!
1826  *
1827  * The page does not need to be reserved.
1828  */
1829 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1830                         struct page *page)
1831 {
1832         if (addr < vma->vm_start || addr >= vma->vm_end)
1833                 return -EFAULT;
1834         if (!page_count(page))
1835                 return -EINVAL;
1836         vma->vm_flags |= VM_INSERTPAGE;
1837         return insert_page(vma, addr, page, vma->vm_page_prot);
1838 }
1839 EXPORT_SYMBOL(vm_insert_page);
1840
1841 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1842                         unsigned long pfn, pgprot_t prot)
1843 {
1844         struct mm_struct *mm = vma->vm_mm;
1845         int retval;
1846         pte_t *pte, entry;
1847         spinlock_t *ptl;
1848
1849         retval = -ENOMEM;
1850         pte = get_locked_pte(mm, addr, &ptl);
1851         if (!pte)
1852                 goto out;
1853         retval = -EBUSY;
1854         if (!pte_none(*pte))
1855                 goto out_unlock;
1856
1857         /* Ok, finally just insert the thing.. */
1858         entry = pte_mkspecial(pfn_pte(pfn, prot));
1859         set_pte_at(mm, addr, pte, entry);
1860         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1861
1862         retval = 0;
1863 out_unlock:
1864         pte_unmap_unlock(pte, ptl);
1865 out:
1866         return retval;
1867 }
1868
1869 /**
1870  * vm_insert_pfn - insert single pfn into user vma
1871  * @vma: user vma to map to
1872  * @addr: target user address of this page
1873  * @pfn: source kernel pfn
1874  *
1875  * Similar to vm_inert_page, this allows drivers to insert individual pages
1876  * they've allocated into a user vma. Same comments apply.
1877  *
1878  * This function should only be called from a vm_ops->fault handler, and
1879  * in that case the handler should return NULL.
1880  *
1881  * vma cannot be a COW mapping.
1882  *
1883  * As this is called only for pages that do not currently exist, we
1884  * do not need to flush old virtual caches or the TLB.
1885  */
1886 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1887                         unsigned long pfn)
1888 {
1889         int ret;
1890         pgprot_t pgprot = vma->vm_page_prot;
1891         /*
1892          * Technically, architectures with pte_special can avoid all these
1893          * restrictions (same for remap_pfn_range).  However we would like
1894          * consistency in testing and feature parity among all, so we should
1895          * try to keep these invariants in place for everybody.
1896          */
1897         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1898         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1899                                                 (VM_PFNMAP|VM_MIXEDMAP));
1900         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1901         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1902
1903         if (addr < vma->vm_start || addr >= vma->vm_end)
1904                 return -EFAULT;
1905         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1906                 return -EINVAL;
1907
1908         ret = insert_pfn(vma, addr, pfn, pgprot);
1909
1910         if (ret)
1911                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1912
1913         return ret;
1914 }
1915 EXPORT_SYMBOL(vm_insert_pfn);
1916
1917 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1918                         unsigned long pfn)
1919 {
1920         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1921
1922         if (addr < vma->vm_start || addr >= vma->vm_end)
1923                 return -EFAULT;
1924
1925         /*
1926          * If we don't have pte special, then we have to use the pfn_valid()
1927          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1928          * refcount the page if pfn_valid is true (hence insert_page rather
1929          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1930          * without pte special, it would there be refcounted as a normal page.
1931          */
1932         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1933                 struct page *page;
1934
1935                 page = pfn_to_page(pfn);
1936                 return insert_page(vma, addr, page, vma->vm_page_prot);
1937         }
1938         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1939 }
1940 EXPORT_SYMBOL(vm_insert_mixed);
1941
1942 /*
1943  * maps a range of physical memory into the requested pages. the old
1944  * mappings are removed. any references to nonexistent pages results
1945  * in null mappings (currently treated as "copy-on-access")
1946  */
1947 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1948                         unsigned long addr, unsigned long end,
1949                         unsigned long pfn, pgprot_t prot)
1950 {
1951         pte_t *pte;
1952         spinlock_t *ptl;
1953
1954         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1955         if (!pte)
1956                 return -ENOMEM;
1957         arch_enter_lazy_mmu_mode();
1958         do {
1959                 BUG_ON(!pte_none(*pte));
1960                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1961                 pfn++;
1962         } while (pte++, addr += PAGE_SIZE, addr != end);
1963         arch_leave_lazy_mmu_mode();
1964         pte_unmap_unlock(pte - 1, ptl);
1965         return 0;
1966 }
1967
1968 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1969                         unsigned long addr, unsigned long end,
1970                         unsigned long pfn, pgprot_t prot)
1971 {
1972         pmd_t *pmd;
1973         unsigned long next;
1974
1975         pfn -= addr >> PAGE_SHIFT;
1976         pmd = pmd_alloc(mm, pud, addr);
1977         if (!pmd)
1978                 return -ENOMEM;
1979         VM_BUG_ON(pmd_trans_huge(*pmd));
1980         do {
1981                 next = pmd_addr_end(addr, end);
1982                 if (remap_pte_range(mm, pmd, addr, next,
1983                                 pfn + (addr >> PAGE_SHIFT), prot))
1984                         return -ENOMEM;
1985         } while (pmd++, addr = next, addr != end);
1986         return 0;
1987 }
1988
1989 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1990                         unsigned long addr, unsigned long end,
1991                         unsigned long pfn, pgprot_t prot)
1992 {
1993         pud_t *pud;
1994         unsigned long next;
1995
1996         pfn -= addr >> PAGE_SHIFT;
1997         pud = pud_alloc(mm, pgd, addr);
1998         if (!pud)
1999                 return -ENOMEM;
2000         do {
2001                 next = pud_addr_end(addr, end);
2002                 if (remap_pmd_range(mm, pud, addr, next,
2003                                 pfn + (addr >> PAGE_SHIFT), prot))
2004                         return -ENOMEM;
2005         } while (pud++, addr = next, addr != end);
2006         return 0;
2007 }
2008
2009 /**
2010  * remap_pfn_range - remap kernel memory to userspace
2011  * @vma: user vma to map to
2012  * @addr: target user address to start at
2013  * @pfn: physical address of kernel memory
2014  * @size: size of map area
2015  * @prot: page protection flags for this mapping
2016  *
2017  *  Note: this is only safe if the mm semaphore is held when called.
2018  */
2019 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2020                     unsigned long pfn, unsigned long size, pgprot_t prot)
2021 {
2022         pgd_t *pgd;
2023         unsigned long next;
2024         unsigned long end = addr + PAGE_ALIGN(size);
2025         struct mm_struct *mm = vma->vm_mm;
2026         int err;
2027
2028         /*
2029          * Physically remapped pages are special. Tell the
2030          * rest of the world about it:
2031          *   VM_IO tells people not to look at these pages
2032          *      (accesses can have side effects).
2033          *   VM_RESERVED is specified all over the place, because
2034          *      in 2.4 it kept swapout's vma scan off this vma; but
2035          *      in 2.6 the LRU scan won't even find its pages, so this
2036          *      flag means no more than count its pages in reserved_vm,
2037          *      and omit it from core dump, even when VM_IO turned off.
2038          *   VM_PFNMAP tells the core MM that the base pages are just
2039          *      raw PFN mappings, and do not have a "struct page" associated
2040          *      with them.
2041          *
2042          * There's a horrible special case to handle copy-on-write
2043          * behaviour that some programs depend on. We mark the "original"
2044          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2045          */
2046         if (addr == vma->vm_start && end == vma->vm_end) {
2047                 vma->vm_pgoff = pfn;
2048                 vma->vm_flags |= VM_PFN_AT_MMAP;
2049         } else if (is_cow_mapping(vma->vm_flags))
2050                 return -EINVAL;
2051
2052         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2053
2054         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2055         if (err) {
2056                 /*
2057                  * To indicate that track_pfn related cleanup is not
2058                  * needed from higher level routine calling unmap_vmas
2059                  */
2060                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2061                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2062                 return -EINVAL;
2063         }
2064
2065         BUG_ON(addr >= end);
2066         pfn -= addr >> PAGE_SHIFT;
2067         pgd = pgd_offset(mm, addr);
2068         flush_cache_range(vma, addr, end);
2069         do {
2070                 next = pgd_addr_end(addr, end);
2071                 err = remap_pud_range(mm, pgd, addr, next,
2072                                 pfn + (addr >> PAGE_SHIFT), prot);
2073                 if (err)
2074                         break;
2075         } while (pgd++, addr = next, addr != end);
2076
2077         if (err)
2078                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2079
2080         return err;
2081 }
2082 EXPORT_SYMBOL(remap_pfn_range);
2083
2084 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2085                                      unsigned long addr, unsigned long end,
2086                                      pte_fn_t fn, void *data)
2087 {
2088         pte_t *pte;
2089         int err;
2090         pgtable_t token;
2091         spinlock_t *uninitialized_var(ptl);
2092
2093         pte = (mm == &init_mm) ?
2094                 pte_alloc_kernel(pmd, addr) :
2095                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2096         if (!pte)
2097                 return -ENOMEM;
2098
2099         BUG_ON(pmd_huge(*pmd));
2100
2101         arch_enter_lazy_mmu_mode();
2102
2103         token = pmd_pgtable(*pmd);
2104
2105         do {
2106                 err = fn(pte++, token, addr, data);
2107                 if (err)
2108                         break;
2109         } while (addr += PAGE_SIZE, addr != end);
2110
2111         arch_leave_lazy_mmu_mode();
2112
2113         if (mm != &init_mm)
2114                 pte_unmap_unlock(pte-1, ptl);
2115         return err;
2116 }
2117
2118 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2119                                      unsigned long addr, unsigned long end,
2120                                      pte_fn_t fn, void *data)
2121 {
2122         pmd_t *pmd;
2123         unsigned long next;
2124         int err;
2125
2126         BUG_ON(pud_huge(*pud));
2127
2128         pmd = pmd_alloc(mm, pud, addr);
2129         if (!pmd)
2130                 return -ENOMEM;
2131         do {
2132                 next = pmd_addr_end(addr, end);
2133                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2134                 if (err)
2135                         break;
2136         } while (pmd++, addr = next, addr != end);
2137         return err;
2138 }
2139
2140 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2141                                      unsigned long addr, unsigned long end,
2142                                      pte_fn_t fn, void *data)
2143 {
2144         pud_t *pud;
2145         unsigned long next;
2146         int err;
2147
2148         pud = pud_alloc(mm, pgd, addr);
2149         if (!pud)
2150                 return -ENOMEM;
2151         do {
2152                 next = pud_addr_end(addr, end);
2153                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2154                 if (err)
2155                         break;
2156         } while (pud++, addr = next, addr != end);
2157         return err;
2158 }
2159
2160 /*
2161  * Scan a region of virtual memory, filling in page tables as necessary
2162  * and calling a provided function on each leaf page table.
2163  */
2164 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2165                         unsigned long size, pte_fn_t fn, void *data)
2166 {
2167         pgd_t *pgd;
2168         unsigned long next;
2169         unsigned long end = addr + size;
2170         int err;
2171
2172         BUG_ON(addr >= end);
2173         pgd = pgd_offset(mm, addr);
2174         do {
2175                 next = pgd_addr_end(addr, end);
2176                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2177                 if (err)
2178                         break;
2179         } while (pgd++, addr = next, addr != end);
2180
2181         return err;
2182 }
2183 EXPORT_SYMBOL_GPL(apply_to_page_range);
2184
2185 /*
2186  * handle_pte_fault chooses page fault handler according to an entry
2187  * which was read non-atomically.  Before making any commitment, on
2188  * those architectures or configurations (e.g. i386 with PAE) which
2189  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2190  * must check under lock before unmapping the pte and proceeding
2191  * (but do_wp_page is only called after already making such a check;
2192  * and do_anonymous_page can safely check later on).
2193  */
2194 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2195                                 pte_t *page_table, pte_t orig_pte)
2196 {
2197         int same = 1;
2198 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2199         if (sizeof(pte_t) > sizeof(unsigned long)) {
2200                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2201                 spin_lock(ptl);
2202                 same = pte_same(*page_table, orig_pte);
2203                 spin_unlock(ptl);
2204         }
2205 #endif
2206         pte_unmap(page_table);
2207         return same;
2208 }
2209
2210 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2211 {
2212         /*
2213          * If the source page was a PFN mapping, we don't have
2214          * a "struct page" for it. We do a best-effort copy by
2215          * just copying from the original user address. If that
2216          * fails, we just zero-fill it. Live with it.
2217          */
2218         if (unlikely(!src)) {
2219                 void *kaddr = kmap_atomic(dst, KM_USER0);
2220                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2221
2222                 /*
2223                  * This really shouldn't fail, because the page is there
2224                  * in the page tables. But it might just be unreadable,
2225                  * in which case we just give up and fill the result with
2226                  * zeroes.
2227                  */
2228                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2229                         clear_page(kaddr);
2230                 kunmap_atomic(kaddr, KM_USER0);
2231                 flush_dcache_page(dst);
2232         } else
2233                 copy_user_highpage(dst, src, va, vma);
2234 }
2235
2236 /*
2237  * This routine handles present pages, when users try to write
2238  * to a shared page. It is done by copying the page to a new address
2239  * and decrementing the shared-page counter for the old page.
2240  *
2241  * Note that this routine assumes that the protection checks have been
2242  * done by the caller (the low-level page fault routine in most cases).
2243  * Thus we can safely just mark it writable once we've done any necessary
2244  * COW.
2245  *
2246  * We also mark the page dirty at this point even though the page will
2247  * change only once the write actually happens. This avoids a few races,
2248  * and potentially makes it more efficient.
2249  *
2250  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2251  * but allow concurrent faults), with pte both mapped and locked.
2252  * We return with mmap_sem still held, but pte unmapped and unlocked.
2253  */
2254 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2255                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2256                 spinlock_t *ptl, pte_t orig_pte)
2257         __releases(ptl)
2258 {
2259         struct page *old_page, *new_page;
2260         pte_t entry;
2261         int ret = 0;
2262         int page_mkwrite = 0;
2263         struct page *dirty_page = NULL;
2264
2265         old_page = vm_normal_page(vma, address, orig_pte);
2266         if (!old_page) {
2267                 /*
2268                  * VM_MIXEDMAP !pfn_valid() case
2269                  *
2270                  * We should not cow pages in a shared writeable mapping.
2271                  * Just mark the pages writable as we can't do any dirty
2272                  * accounting on raw pfn maps.
2273                  */
2274                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2275                                      (VM_WRITE|VM_SHARED))
2276                         goto reuse;
2277                 goto gotten;
2278         }
2279
2280         /*
2281          * Take out anonymous pages first, anonymous shared vmas are
2282          * not dirty accountable.
2283          */
2284         if (PageAnon(old_page) && !PageKsm(old_page)) {
2285                 if (!trylock_page(old_page)) {
2286                         page_cache_get(old_page);
2287                         pte_unmap_unlock(page_table, ptl);
2288                         lock_page(old_page);
2289                         page_table = pte_offset_map_lock(mm, pmd, address,
2290                                                          &ptl);
2291                         if (!pte_same(*page_table, orig_pte)) {
2292                                 unlock_page(old_page);
2293                                 goto unlock;
2294                         }
2295                         page_cache_release(old_page);
2296                 }
2297                 if (reuse_swap_page(old_page)) {
2298                         /*
2299                          * The page is all ours.  Move it to our anon_vma so
2300                          * the rmap code will not search our parent or siblings.
2301                          * Protected against the rmap code by the page lock.
2302                          */
2303                         page_move_anon_rmap(old_page, vma, address);
2304                         unlock_page(old_page);
2305                         goto reuse;
2306                 }
2307                 unlock_page(old_page);
2308         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2309                                         (VM_WRITE|VM_SHARED))) {
2310                 /*
2311                  * Only catch write-faults on shared writable pages,
2312                  * read-only shared pages can get COWed by
2313                  * get_user_pages(.write=1, .force=1).
2314                  */
2315                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2316                         struct vm_fault vmf;
2317                         int tmp;
2318
2319                         vmf.virtual_address = (void __user *)(address &
2320                                                                 PAGE_MASK);
2321                         vmf.pgoff = old_page->index;
2322                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2323                         vmf.page = old_page;
2324
2325                         /*
2326                          * Notify the address space that the page is about to
2327                          * become writable so that it can prohibit this or wait
2328                          * for the page to get into an appropriate state.
2329                          *
2330                          * We do this without the lock held, so that it can
2331                          * sleep if it needs to.
2332                          */
2333                         page_cache_get(old_page);
2334                         pte_unmap_unlock(page_table, ptl);
2335
2336                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2337                         if (unlikely(tmp &
2338                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2339                                 ret = tmp;
2340                                 goto unwritable_page;
2341                         }
2342                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2343                                 lock_page(old_page);
2344                                 if (!old_page->mapping) {
2345                                         ret = 0; /* retry the fault */
2346                                         unlock_page(old_page);
2347                                         goto unwritable_page;
2348                                 }
2349                         } else
2350                                 VM_BUG_ON(!PageLocked(old_page));
2351
2352                         /*
2353                          * Since we dropped the lock we need to revalidate
2354                          * the PTE as someone else may have changed it.  If
2355                          * they did, we just return, as we can count on the
2356                          * MMU to tell us if they didn't also make it writable.
2357                          */
2358                         page_table = pte_offset_map_lock(mm, pmd, address,
2359                                                          &ptl);
2360                         if (!pte_same(*page_table, orig_pte)) {
2361                                 unlock_page(old_page);
2362                                 goto unlock;
2363                         }
2364
2365                         page_mkwrite = 1;
2366                 }
2367                 dirty_page = old_page;
2368                 get_page(dirty_page);
2369
2370 reuse:
2371                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2372                 entry = pte_mkyoung(orig_pte);
2373                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2374                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2375                         update_mmu_cache(vma, address, page_table);
2376                 pte_unmap_unlock(page_table, ptl);
2377                 ret |= VM_FAULT_WRITE;
2378
2379                 if (!dirty_page)
2380                         return ret;
2381
2382                 /*
2383                  * Yes, Virginia, this is actually required to prevent a race
2384                  * with clear_page_dirty_for_io() from clearing the page dirty
2385                  * bit after it clear all dirty ptes, but before a racing
2386                  * do_wp_page installs a dirty pte.
2387                  *
2388                  * __do_fault is protected similarly.
2389                  */
2390                 if (!page_mkwrite) {
2391                         wait_on_page_locked(dirty_page);
2392                         set_page_dirty_balance(dirty_page, page_mkwrite);
2393                 }
2394                 put_page(dirty_page);
2395                 if (page_mkwrite) {
2396                         struct address_space *mapping = dirty_page->mapping;
2397
2398                         set_page_dirty(dirty_page);
2399                         unlock_page(dirty_page);
2400                         page_cache_release(dirty_page);
2401                         if (mapping)    {
2402                                 /*
2403                                  * Some device drivers do not set page.mapping
2404                                  * but still dirty their pages
2405                                  */
2406                                 balance_dirty_pages_ratelimited(mapping);
2407                         }
2408                 }
2409
2410                 /* file_update_time outside page_lock */
2411                 if (vma->vm_file)
2412                         file_update_time(vma->vm_file);
2413
2414                 return ret;
2415         }
2416
2417         /*
2418          * Ok, we need to copy. Oh, well..
2419          */
2420         page_cache_get(old_page);
2421 gotten:
2422         pte_unmap_unlock(page_table, ptl);
2423
2424         if (unlikely(anon_vma_prepare(vma)))
2425                 goto oom;
2426
2427         if (is_zero_pfn(pte_pfn(orig_pte))) {
2428                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2429                 if (!new_page)
2430                         goto oom;
2431         } else {
2432                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2433                 if (!new_page)
2434                         goto oom;
2435                 cow_user_page(new_page, old_page, address, vma);
2436         }
2437         __SetPageUptodate(new_page);
2438
2439         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2440                 goto oom_free_new;
2441
2442         /*
2443          * Re-check the pte - we dropped the lock
2444          */
2445         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2446         if (likely(pte_same(*page_table, orig_pte))) {
2447                 if (old_page) {
2448                         if (!PageAnon(old_page)) {
2449                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2450                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2451                         }
2452                 } else
2453                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2454                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2455                 entry = mk_pte(new_page, vma->vm_page_prot);
2456                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2457                 /*
2458                  * Clear the pte entry and flush it first, before updating the
2459                  * pte with the new entry. This will avoid a race condition
2460                  * seen in the presence of one thread doing SMC and another
2461                  * thread doing COW.
2462                  */
2463                 ptep_clear_flush(vma, address, page_table);
2464                 page_add_new_anon_rmap(new_page, vma, address);
2465                 /*
2466                  * We call the notify macro here because, when using secondary
2467                  * mmu page tables (such as kvm shadow page tables), we want the
2468                  * new page to be mapped directly into the secondary page table.
2469                  */
2470                 set_pte_at_notify(mm, address, page_table, entry);
2471                 update_mmu_cache(vma, address, page_table);
2472                 if (old_page) {
2473                         /*
2474                          * Only after switching the pte to the new page may
2475                          * we remove the mapcount here. Otherwise another
2476                          * process may come and find the rmap count decremented
2477                          * before the pte is switched to the new page, and
2478                          * "reuse" the old page writing into it while our pte
2479                          * here still points into it and can be read by other
2480                          * threads.
2481                          *
2482                          * The critical issue is to order this
2483                          * page_remove_rmap with the ptp_clear_flush above.
2484                          * Those stores are ordered by (if nothing else,)
2485                          * the barrier present in the atomic_add_negative
2486                          * in page_remove_rmap.
2487                          *
2488                          * Then the TLB flush in ptep_clear_flush ensures that
2489                          * no process can access the old page before the
2490                          * decremented mapcount is visible. And the old page
2491                          * cannot be reused until after the decremented
2492                          * mapcount is visible. So transitively, TLBs to
2493                          * old page will be flushed before it can be reused.
2494                          */
2495                         page_remove_rmap(old_page);
2496                 }
2497
2498                 /* Free the old page.. */
2499                 new_page = old_page;
2500                 ret |= VM_FAULT_WRITE;
2501         } else
2502                 mem_cgroup_uncharge_page(new_page);
2503
2504         if (new_page)
2505                 page_cache_release(new_page);
2506 unlock:
2507         pte_unmap_unlock(page_table, ptl);
2508         if (old_page) {
2509                 /*
2510                  * Don't let another task, with possibly unlocked vma,
2511                  * keep the mlocked page.
2512                  */
2513                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2514                         lock_page(old_page);    /* LRU manipulation */
2515                         munlock_vma_page(old_page);
2516                         unlock_page(old_page);
2517                 }
2518                 page_cache_release(old_page);
2519         }
2520         return ret;
2521 oom_free_new:
2522         page_cache_release(new_page);
2523 oom:
2524         if (old_page) {
2525                 if (page_mkwrite) {
2526                         unlock_page(old_page);
2527                         page_cache_release(old_page);
2528                 }
2529                 page_cache_release(old_page);
2530         }
2531         return VM_FAULT_OOM;
2532
2533 unwritable_page:
2534         page_cache_release(old_page);
2535         return ret;
2536 }
2537
2538 /*
2539  * Helper functions for unmap_mapping_range().
2540  *
2541  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2542  *
2543  * We have to restart searching the prio_tree whenever we drop the lock,
2544  * since the iterator is only valid while the lock is held, and anyway
2545  * a later vma might be split and reinserted earlier while lock dropped.
2546  *
2547  * The list of nonlinear vmas could be handled more efficiently, using
2548  * a placeholder, but handle it in the same way until a need is shown.
2549  * It is important to search the prio_tree before nonlinear list: a vma
2550  * may become nonlinear and be shifted from prio_tree to nonlinear list
2551  * while the lock is dropped; but never shifted from list to prio_tree.
2552  *
2553  * In order to make forward progress despite restarting the search,
2554  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2555  * quickly skip it next time around.  Since the prio_tree search only
2556  * shows us those vmas affected by unmapping the range in question, we
2557  * can't efficiently keep all vmas in step with mapping->truncate_count:
2558  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2559  * mapping->truncate_count and vma->vm_truncate_count are protected by
2560  * i_mmap_lock.
2561  *
2562  * In order to make forward progress despite repeatedly restarting some
2563  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2564  * and restart from that address when we reach that vma again.  It might
2565  * have been split or merged, shrunk or extended, but never shifted: so
2566  * restart_addr remains valid so long as it remains in the vma's range.
2567  * unmap_mapping_range forces truncate_count to leap over page-aligned
2568  * values so we can save vma's restart_addr in its truncate_count field.
2569  */
2570 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2571
2572 static void reset_vma_truncate_counts(struct address_space *mapping)
2573 {
2574         struct vm_area_struct *vma;
2575         struct prio_tree_iter iter;
2576
2577         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2578                 vma->vm_truncate_count = 0;
2579         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2580                 vma->vm_truncate_count = 0;
2581 }
2582
2583 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2584                 unsigned long start_addr, unsigned long end_addr,
2585                 struct zap_details *details)
2586 {
2587         unsigned long restart_addr;
2588         int need_break;
2589
2590         /*
2591          * files that support invalidating or truncating portions of the
2592          * file from under mmaped areas must have their ->fault function
2593          * return a locked page (and set VM_FAULT_LOCKED in the return).
2594          * This provides synchronisation against concurrent unmapping here.
2595          */
2596
2597 again:
2598         restart_addr = vma->vm_truncate_count;
2599         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2600                 start_addr = restart_addr;
2601                 if (start_addr >= end_addr) {
2602                         /* Top of vma has been split off since last time */
2603                         vma->vm_truncate_count = details->truncate_count;
2604                         return 0;
2605                 }
2606         }
2607
2608         restart_addr = zap_page_range(vma, start_addr,
2609                                         end_addr - start_addr, details);
2610         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2611
2612         if (restart_addr >= end_addr) {
2613                 /* We have now completed this vma: mark it so */
2614                 vma->vm_truncate_count = details->truncate_count;
2615                 if (!need_break)
2616                         return 0;
2617         } else {
2618                 /* Note restart_addr in vma's truncate_count field */
2619                 vma->vm_truncate_count = restart_addr;
2620                 if (!need_break)
2621                         goto again;
2622         }
2623
2624         spin_unlock(details->i_mmap_lock);
2625         cond_resched();
2626         spin_lock(details->i_mmap_lock);
2627         return -EINTR;
2628 }
2629
2630 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2631                                             struct zap_details *details)
2632 {
2633         struct vm_area_struct *vma;
2634         struct prio_tree_iter iter;
2635         pgoff_t vba, vea, zba, zea;
2636
2637 restart:
2638         vma_prio_tree_foreach(vma, &iter, root,
2639                         details->first_index, details->last_index) {
2640                 /* Skip quickly over those we have already dealt with */
2641                 if (vma->vm_truncate_count == details->truncate_count)
2642                         continue;
2643
2644                 vba = vma->vm_pgoff;
2645                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2646                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2647                 zba = details->first_index;
2648                 if (zba < vba)
2649                         zba = vba;
2650                 zea = details->last_index;
2651                 if (zea > vea)
2652                         zea = vea;
2653
2654                 if (unmap_mapping_range_vma(vma,
2655                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2656                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2657                                 details) < 0)
2658                         goto restart;
2659         }
2660 }
2661
2662 static inline void unmap_mapping_range_list(struct list_head *head,
2663                                             struct zap_details *details)
2664 {
2665         struct vm_area_struct *vma;
2666
2667         /*
2668          * In nonlinear VMAs there is no correspondence between virtual address
2669          * offset and file offset.  So we must perform an exhaustive search
2670          * across *all* the pages in each nonlinear VMA, not just the pages
2671          * whose virtual address lies outside the file truncation point.
2672          */
2673 restart:
2674         list_for_each_entry(vma, head, shared.vm_set.list) {
2675                 /* Skip quickly over those we have already dealt with */
2676                 if (vma->vm_truncate_count == details->truncate_count)
2677                         continue;
2678                 details->nonlinear_vma = vma;
2679                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2680                                         vma->vm_end, details) < 0)
2681                         goto restart;
2682         }
2683 }
2684
2685 /**
2686  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2687  * @mapping: the address space containing mmaps to be unmapped.
2688  * @holebegin: byte in first page to unmap, relative to the start of
2689  * the underlying file.  This will be rounded down to a PAGE_SIZE
2690  * boundary.  Note that this is different from truncate_pagecache(), which
2691  * must keep the partial page.  In contrast, we must get rid of
2692  * partial pages.
2693  * @holelen: size of prospective hole in bytes.  This will be rounded
2694  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2695  * end of the file.
2696  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2697  * but 0 when invalidating pagecache, don't throw away private data.
2698  */
2699 void unmap_mapping_range(struct address_space *mapping,
2700                 loff_t const holebegin, loff_t const holelen, int even_cows)
2701 {
2702         struct zap_details details;
2703         pgoff_t hba = holebegin >> PAGE_SHIFT;
2704         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2705
2706         /* Check for overflow. */
2707         if (sizeof(holelen) > sizeof(hlen)) {
2708                 long long holeend =
2709                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2710                 if (holeend & ~(long long)ULONG_MAX)
2711                         hlen = ULONG_MAX - hba + 1;
2712         }
2713
2714         details.check_mapping = even_cows? NULL: mapping;
2715         details.nonlinear_vma = NULL;
2716         details.first_index = hba;
2717         details.last_index = hba + hlen - 1;
2718         if (details.last_index < details.first_index)
2719                 details.last_index = ULONG_MAX;
2720         details.i_mmap_lock = &mapping->i_mmap_lock;
2721
2722         mutex_lock(&mapping->unmap_mutex);
2723         spin_lock(&mapping->i_mmap_lock);
2724
2725         /* Protect against endless unmapping loops */
2726         mapping->truncate_count++;
2727         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2728                 if (mapping->truncate_count == 0)
2729                         reset_vma_truncate_counts(mapping);
2730                 mapping->truncate_count++;
2731         }
2732         details.truncate_count = mapping->truncate_count;
2733
2734         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2735                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2736         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2737                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2738         spin_unlock(&mapping->i_mmap_lock);
2739         mutex_unlock(&mapping->unmap_mutex);
2740 }
2741 EXPORT_SYMBOL(unmap_mapping_range);
2742
2743 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2744 {
2745         struct address_space *mapping = inode->i_mapping;
2746
2747         /*
2748          * If the underlying filesystem is not going to provide
2749          * a way to truncate a range of blocks (punch a hole) -
2750          * we should return failure right now.
2751          */
2752         if (!inode->i_op->truncate_range)
2753                 return -ENOSYS;
2754
2755         mutex_lock(&inode->i_mutex);
2756         down_write(&inode->i_alloc_sem);
2757         unmap_mapping_range(mapping, offset, (end - offset), 1);
2758         truncate_inode_pages_range(mapping, offset, end);
2759         unmap_mapping_range(mapping, offset, (end - offset), 1);
2760         inode->i_op->truncate_range(inode, offset, end);
2761         up_write(&inode->i_alloc_sem);
2762         mutex_unlock(&inode->i_mutex);
2763
2764         return 0;
2765 }
2766
2767 /*
2768  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2769  * but allow concurrent faults), and pte mapped but not yet locked.
2770  * We return with mmap_sem still held, but pte unmapped and unlocked.
2771  */
2772 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2773                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2774                 unsigned int flags, pte_t orig_pte)
2775 {
2776         spinlock_t *ptl;
2777         struct page *page, *swapcache = NULL;
2778         swp_entry_t entry;
2779         pte_t pte;
2780         int locked;
2781         struct mem_cgroup *ptr;
2782         int exclusive = 0;
2783         int ret = 0;
2784
2785         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2786                 goto out;
2787
2788         entry = pte_to_swp_entry(orig_pte);
2789         if (unlikely(non_swap_entry(entry))) {
2790                 if (is_migration_entry(entry)) {
2791                         migration_entry_wait(mm, pmd, address);
2792                 } else if (is_hwpoison_entry(entry)) {
2793                         ret = VM_FAULT_HWPOISON;
2794                 } else {
2795                         print_bad_pte(vma, address, orig_pte, NULL);
2796                         ret = VM_FAULT_SIGBUS;
2797                 }
2798                 goto out;
2799         }
2800         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2801         page = lookup_swap_cache(entry);
2802         if (!page) {
2803                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2804                 page = swapin_readahead(entry,
2805                                         GFP_HIGHUSER_MOVABLE, vma, address);
2806                 if (!page) {
2807                         /*
2808                          * Back out if somebody else faulted in this pte
2809                          * while we released the pte lock.
2810                          */
2811                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2812                         if (likely(pte_same(*page_table, orig_pte)))
2813                                 ret = VM_FAULT_OOM;
2814                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2815                         goto unlock;
2816                 }
2817
2818                 /* Had to read the page from swap area: Major fault */
2819                 ret = VM_FAULT_MAJOR;
2820                 count_vm_event(PGMAJFAULT);
2821         } else if (PageHWPoison(page)) {
2822                 /*
2823                  * hwpoisoned dirty swapcache pages are kept for killing
2824                  * owner processes (which may be unknown at hwpoison time)
2825                  */
2826                 ret = VM_FAULT_HWPOISON;
2827                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2828                 goto out_release;
2829         }
2830
2831         locked = lock_page_or_retry(page, mm, flags);
2832         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2833         if (!locked) {
2834                 ret |= VM_FAULT_RETRY;
2835                 goto out_release;
2836         }
2837
2838         /*
2839          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2840          * release the swapcache from under us.  The page pin, and pte_same
2841          * test below, are not enough to exclude that.  Even if it is still
2842          * swapcache, we need to check that the page's swap has not changed.
2843          */
2844         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2845                 goto out_page;
2846
2847         if (ksm_might_need_to_copy(page, vma, address)) {
2848                 swapcache = page;
2849                 page = ksm_does_need_to_copy(page, vma, address);
2850
2851                 if (unlikely(!page)) {
2852                         ret = VM_FAULT_OOM;
2853                         page = swapcache;
2854                         swapcache = NULL;
2855                         goto out_page;
2856                 }
2857         }
2858
2859         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2860                 ret = VM_FAULT_OOM;
2861                 goto out_page;
2862         }
2863
2864         /*
2865          * Back out if somebody else already faulted in this pte.
2866          */
2867         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2868         if (unlikely(!pte_same(*page_table, orig_pte)))
2869                 goto out_nomap;
2870
2871         if (unlikely(!PageUptodate(page))) {
2872                 ret = VM_FAULT_SIGBUS;
2873                 goto out_nomap;
2874         }
2875
2876         /*
2877          * The page isn't present yet, go ahead with the fault.
2878          *
2879          * Be careful about the sequence of operations here.
2880          * To get its accounting right, reuse_swap_page() must be called
2881          * while the page is counted on swap but not yet in mapcount i.e.
2882          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2883          * must be called after the swap_free(), or it will never succeed.
2884          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2885          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2886          * in page->private. In this case, a record in swap_cgroup  is silently
2887          * discarded at swap_free().
2888          */
2889
2890         inc_mm_counter_fast(mm, MM_ANONPAGES);
2891         dec_mm_counter_fast(mm, MM_SWAPENTS);
2892         pte = mk_pte(page, vma->vm_page_prot);
2893         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2894                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2895                 flags &= ~FAULT_FLAG_WRITE;
2896                 ret |= VM_FAULT_WRITE;
2897                 exclusive = 1;
2898         }
2899         flush_icache_page(vma, page);
2900         set_pte_at(mm, address, page_table, pte);
2901         do_page_add_anon_rmap(page, vma, address, exclusive);
2902         /* It's better to call commit-charge after rmap is established */
2903         mem_cgroup_commit_charge_swapin(page, ptr);
2904
2905         swap_free(entry);
2906         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2907                 try_to_free_swap(page);
2908         unlock_page(page);
2909         if (swapcache) {
2910                 /*
2911                  * Hold the lock to avoid the swap entry to be reused
2912                  * until we take the PT lock for the pte_same() check
2913                  * (to avoid false positives from pte_same). For
2914                  * further safety release the lock after the swap_free
2915                  * so that the swap count won't change under a
2916                  * parallel locked swapcache.
2917                  */
2918                 unlock_page(swapcache);
2919                 page_cache_release(swapcache);
2920         }
2921
2922         if (flags & FAULT_FLAG_WRITE) {
2923                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2924                 if (ret & VM_FAULT_ERROR)
2925                         ret &= VM_FAULT_ERROR;
2926                 goto out;
2927         }
2928
2929         /* No need to invalidate - it was non-present before */
2930         update_mmu_cache(vma, address, page_table);
2931 unlock:
2932         pte_unmap_unlock(page_table, ptl);
2933 out:
2934         return ret;
2935 out_nomap:
2936         mem_cgroup_cancel_charge_swapin(ptr);
2937         pte_unmap_unlock(page_table, ptl);
2938 out_page:
2939         unlock_page(page);
2940 out_release:
2941         page_cache_release(page);
2942         if (swapcache) {
2943                 unlock_page(swapcache);
2944                 page_cache_release(swapcache);
2945         }
2946         return ret;
2947 }
2948
2949 /*
2950  * This is like a special single-page "expand_{down|up}wards()",
2951  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2952  * doesn't hit another vma.
2953  */
2954 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2955 {
2956         address &= PAGE_MASK;
2957         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2958                 struct vm_area_struct *prev = vma->vm_prev;
2959
2960                 /*
2961                  * Is there a mapping abutting this one below?
2962                  *
2963                  * That's only ok if it's the same stack mapping
2964                  * that has gotten split..
2965                  */
2966                 if (prev && prev->vm_end == address)
2967                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2968
2969                 expand_downwards(vma, address - PAGE_SIZE);
2970         }
2971         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2972                 struct vm_area_struct *next = vma->vm_next;
2973
2974                 /* As VM_GROWSDOWN but s/below/above/ */
2975                 if (next && next->vm_start == address + PAGE_SIZE)
2976                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2977
2978                 expand_upwards(vma, address + PAGE_SIZE);
2979         }
2980         return 0;
2981 }
2982
2983 /*
2984  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2985  * but allow concurrent faults), and pte mapped but not yet locked.
2986  * We return with mmap_sem still held, but pte unmapped and unlocked.
2987  */
2988 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2989                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2990                 unsigned int flags)
2991 {
2992         struct page *page;
2993         spinlock_t *ptl;
2994         pte_t entry;
2995
2996         pte_unmap(page_table);
2997
2998         /* Check if we need to add a guard page to the stack */
2999         if (check_stack_guard_page(vma, address) < 0)
3000                 return VM_FAULT_SIGBUS;
3001
3002         /* Use the zero-page for reads */
3003         if (!(flags & FAULT_FLAG_WRITE)) {
3004                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3005                                                 vma->vm_page_prot));
3006                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3007                 if (!pte_none(*page_table))
3008                         goto unlock;
3009                 goto setpte;
3010         }
3011
3012         /* Allocate our own private page. */
3013         if (unlikely(anon_vma_prepare(vma)))
3014                 goto oom;
3015         page = alloc_zeroed_user_highpage_movable(vma, address);
3016         if (!page)
3017                 goto oom;
3018         __SetPageUptodate(page);
3019
3020         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3021                 goto oom_free_page;
3022
3023         entry = mk_pte(page, vma->vm_page_prot);
3024         if (vma->vm_flags & VM_WRITE)
3025                 entry = pte_mkwrite(pte_mkdirty(entry));
3026
3027         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3028         if (!pte_none(*page_table))
3029                 goto release;
3030
3031         inc_mm_counter_fast(mm, MM_ANONPAGES);
3032         page_add_new_anon_rmap(page, vma, address);
3033 setpte:
3034         set_pte_at(mm, address, page_table, entry);
3035
3036         /* No need to invalidate - it was non-present before */
3037         update_mmu_cache(vma, address, page_table);
3038 unlock:
3039         pte_unmap_unlock(page_table, ptl);
3040         return 0;
3041 release:
3042         mem_cgroup_uncharge_page(page);
3043         page_cache_release(page);
3044         goto unlock;
3045 oom_free_page:
3046         page_cache_release(page);
3047 oom:
3048         return VM_FAULT_OOM;
3049 }
3050
3051 /*
3052  * __do_fault() tries to create a new page mapping. It aggressively
3053  * tries to share with existing pages, but makes a separate copy if
3054  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3055  * the next page fault.
3056  *
3057  * As this is called only for pages that do not currently exist, we
3058  * do not need to flush old virtual caches or the TLB.
3059  *
3060  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3061  * but allow concurrent faults), and pte neither mapped nor locked.
3062  * We return with mmap_sem still held, but pte unmapped and unlocked.
3063  */
3064 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3065                 unsigned long address, pmd_t *pmd,
3066                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3067 {
3068         pte_t *page_table;
3069         spinlock_t *ptl;
3070         struct page *page;
3071         pte_t entry;
3072         int anon = 0;
3073         int charged = 0;
3074         struct page *dirty_page = NULL;
3075         struct vm_fault vmf;
3076         int ret;
3077         int page_mkwrite = 0;
3078
3079         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3080         vmf.pgoff = pgoff;
3081         vmf.flags = flags;
3082         vmf.page = NULL;
3083
3084         ret = vma->vm_ops->fault(vma, &vmf);
3085         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3086                             VM_FAULT_RETRY)))
3087                 return ret;
3088
3089         if (unlikely(PageHWPoison(vmf.page))) {
3090                 if (ret & VM_FAULT_LOCKED)
3091                         unlock_page(vmf.page);
3092                 return VM_FAULT_HWPOISON;
3093         }
3094
3095         /*
3096          * For consistency in subsequent calls, make the faulted page always
3097          * locked.
3098          */
3099         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3100                 lock_page(vmf.page);
3101         else
3102                 VM_BUG_ON(!PageLocked(vmf.page));
3103
3104         /*
3105          * Should we do an early C-O-W break?
3106          */
3107         page = vmf.page;
3108         if (flags & FAULT_FLAG_WRITE) {
3109                 if (!(vma->vm_flags & VM_SHARED)) {
3110                         anon = 1;
3111                         if (unlikely(anon_vma_prepare(vma))) {
3112                                 ret = VM_FAULT_OOM;
3113                                 goto out;
3114                         }
3115                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3116                                                 vma, address);
3117                         if (!page) {
3118                                 ret = VM_FAULT_OOM;
3119                                 goto out;
3120                         }
3121                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3122                                 ret = VM_FAULT_OOM;
3123                                 page_cache_release(page);
3124                                 goto out;
3125                         }
3126                         charged = 1;
3127                         copy_user_highpage(page, vmf.page, address, vma);
3128                         __SetPageUptodate(page);
3129                 } else {
3130                         /*
3131                          * If the page will be shareable, see if the backing
3132                          * address space wants to know that the page is about
3133                          * to become writable
3134                          */
3135                         if (vma->vm_ops->page_mkwrite) {
3136                                 int tmp;
3137
3138                                 unlock_page(page);
3139                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3140                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3141                                 if (unlikely(tmp &
3142                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3143                                         ret = tmp;
3144                                         goto unwritable_page;
3145                                 }
3146                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3147                                         lock_page(page);
3148                                         if (!page->mapping) {
3149                                                 ret = 0; /* retry the fault */
3150                                                 unlock_page(page);
3151                                                 goto unwritable_page;
3152                                         }
3153                                 } else
3154                                         VM_BUG_ON(!PageLocked(page));
3155                                 page_mkwrite = 1;
3156                         }
3157                 }
3158
3159         }
3160
3161         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3162
3163         /*
3164          * This silly early PAGE_DIRTY setting removes a race
3165          * due to the bad i386 page protection. But it's valid
3166          * for other architectures too.
3167          *
3168          * Note that if FAULT_FLAG_WRITE is set, we either now have
3169          * an exclusive copy of the page, or this is a shared mapping,
3170          * so we can make it writable and dirty to avoid having to
3171          * handle that later.
3172          */
3173         /* Only go through if we didn't race with anybody else... */
3174         if (likely(pte_same(*page_table, orig_pte))) {
3175                 flush_icache_page(vma, page);
3176                 entry = mk_pte(page, vma->vm_page_prot);
3177                 if (flags & FAULT_FLAG_WRITE)
3178                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3179                 if (anon) {
3180                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3181                         page_add_new_anon_rmap(page, vma, address);
3182                 } else {
3183                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3184                         page_add_file_rmap(page);
3185                         if (flags & FAULT_FLAG_WRITE) {
3186                                 dirty_page = page;
3187                                 get_page(dirty_page);
3188                         }
3189                 }
3190                 set_pte_at(mm, address, page_table, entry);
3191
3192                 /* no need to invalidate: a not-present page won't be cached */
3193                 update_mmu_cache(vma, address, page_table);
3194         } else {
3195                 if (charged)
3196                         mem_cgroup_uncharge_page(page);
3197                 if (anon)
3198                         page_cache_release(page);
3199                 else
3200                         anon = 1; /* no anon but release faulted_page */
3201         }
3202
3203         pte_unmap_unlock(page_table, ptl);
3204
3205 out:
3206         if (dirty_page) {
3207                 struct address_space *mapping = page->mapping;
3208
3209                 if (set_page_dirty(dirty_page))
3210                         page_mkwrite = 1;
3211                 unlock_page(dirty_page);
3212                 put_page(dirty_page);
3213                 if (page_mkwrite && mapping) {
3214                         /*
3215                          * Some device drivers do not set page.mapping but still
3216                          * dirty their pages
3217                          */
3218                         balance_dirty_pages_ratelimited(mapping);
3219                 }
3220
3221                 /* file_update_time outside page_lock */
3222                 if (vma->vm_file)
3223                         file_update_time(vma->vm_file);
3224         } else {
3225                 unlock_page(vmf.page);
3226                 if (anon)
3227                         page_cache_release(vmf.page);
3228         }
3229
3230         return ret;
3231
3232 unwritable_page:
3233         page_cache_release(page);
3234         return ret;
3235 }
3236
3237 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3238                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3239                 unsigned int flags, pte_t orig_pte)
3240 {
3241         pgoff_t pgoff = (((address & PAGE_MASK)
3242                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3243
3244         pte_unmap(page_table);
3245         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3246 }
3247
3248 /*
3249  * Fault of a previously existing named mapping. Repopulate the pte
3250  * from the encoded file_pte if possible. This enables swappable
3251  * nonlinear vmas.
3252  *
3253  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3254  * but allow concurrent faults), and pte mapped but not yet locked.
3255  * We return with mmap_sem still held, but pte unmapped and unlocked.
3256  */
3257 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3258                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3259                 unsigned int flags, pte_t orig_pte)
3260 {
3261         pgoff_t pgoff;
3262
3263         flags |= FAULT_FLAG_NONLINEAR;
3264
3265         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3266                 return 0;
3267
3268         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3269                 /*
3270                  * Page table corrupted: show pte and kill process.
3271                  */
3272                 print_bad_pte(vma, address, orig_pte, NULL);
3273                 return VM_FAULT_SIGBUS;
3274         }
3275
3276         pgoff = pte_to_pgoff(orig_pte);
3277         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3278 }
3279
3280 /*
3281  * These routines also need to handle stuff like marking pages dirty
3282  * and/or accessed for architectures that don't do it in hardware (most
3283  * RISC architectures).  The early dirtying is also good on the i386.
3284  *
3285  * There is also a hook called "update_mmu_cache()" that architectures
3286  * with external mmu caches can use to update those (ie the Sparc or
3287  * PowerPC hashed page tables that act as extended TLBs).
3288  *
3289  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3290  * but allow concurrent faults), and pte mapped but not yet locked.
3291  * We return with mmap_sem still held, but pte unmapped and unlocked.
3292  */
3293 int handle_pte_fault(struct mm_struct *mm,
3294                      struct vm_area_struct *vma, unsigned long address,
3295                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3296 {
3297         pte_t entry;
3298         spinlock_t *ptl;
3299
3300         entry = *pte;
3301         if (!pte_present(entry)) {
3302                 if (pte_none(entry)) {
3303                         if (vma->vm_ops) {
3304                                 if (likely(vma->vm_ops->fault))
3305                                         return do_linear_fault(mm, vma, address,
3306                                                 pte, pmd, flags, entry);
3307                         }
3308                         return do_anonymous_page(mm, vma, address,
3309                                                  pte, pmd, flags);
3310                 }
3311                 if (pte_file(entry))
3312                         return do_nonlinear_fault(mm, vma, address,
3313                                         pte, pmd, flags, entry);
3314                 return do_swap_page(mm, vma, address,
3315                                         pte, pmd, flags, entry);
3316         }
3317
3318         ptl = pte_lockptr(mm, pmd);
3319         spin_lock(ptl);
3320         if (unlikely(!pte_same(*pte, entry)))
3321                 goto unlock;
3322         if (flags & FAULT_FLAG_WRITE) {
3323                 if (!pte_write(entry))
3324                         return do_wp_page(mm, vma, address,
3325                                         pte, pmd, ptl, entry);
3326                 entry = pte_mkdirty(entry);
3327         }
3328         entry = pte_mkyoung(entry);
3329         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3330                 update_mmu_cache(vma, address, pte);
3331         } else {
3332                 /*
3333                  * This is needed only for protection faults but the arch code
3334                  * is not yet telling us if this is a protection fault or not.
3335                  * This still avoids useless tlb flushes for .text page faults
3336                  * with threads.
3337                  */
3338                 if (flags & FAULT_FLAG_WRITE)
3339                         flush_tlb_fix_spurious_fault(vma, address);
3340         }
3341 unlock:
3342         pte_unmap_unlock(pte, ptl);
3343         return 0;
3344 }
3345
3346 /*
3347  * By the time we get here, we already hold the mm semaphore
3348  */
3349 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3350                 unsigned long address, unsigned int flags)
3351 {
3352         pgd_t *pgd;
3353         pud_t *pud;
3354         pmd_t *pmd;
3355         pte_t *pte;
3356
3357         __set_current_state(TASK_RUNNING);
3358
3359         count_vm_event(PGFAULT);
3360
3361         /* do counter updates before entering really critical section. */
3362         check_sync_rss_stat(current);
3363
3364         if (unlikely(is_vm_hugetlb_page(vma)))
3365                 return hugetlb_fault(mm, vma, address, flags);
3366
3367         pgd = pgd_offset(mm, address);
3368         pud = pud_alloc(mm, pgd, address);
3369         if (!pud)
3370                 return VM_FAULT_OOM;
3371         pmd = pmd_alloc(mm, pud, address);
3372         if (!pmd)
3373                 return VM_FAULT_OOM;
3374         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3375                 if (!vma->vm_ops)
3376                         return do_huge_pmd_anonymous_page(mm, vma, address,
3377                                                           pmd, flags);
3378         } else {
3379                 pmd_t orig_pmd = *pmd;
3380                 barrier();
3381                 if (pmd_trans_huge(orig_pmd)) {
3382                         if (flags & FAULT_FLAG_WRITE &&
3383                             !pmd_write(orig_pmd) &&
3384                             !pmd_trans_splitting(orig_pmd))
3385                                 return do_huge_pmd_wp_page(mm, vma, address,
3386                                                            pmd, orig_pmd);
3387                         return 0;
3388                 }
3389         }
3390
3391         /*
3392          * Use __pte_alloc instead of pte_alloc_map, because we can't
3393          * run pte_offset_map on the pmd, if an huge pmd could
3394          * materialize from under us from a different thread.
3395          */
3396         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3397                 return VM_FAULT_OOM;
3398         /* if an huge pmd materialized from under us just retry later */
3399         if (unlikely(pmd_trans_huge(*pmd)))
3400                 return 0;
3401         /*
3402          * A regular pmd is established and it can't morph into a huge pmd
3403          * from under us anymore at this point because we hold the mmap_sem
3404          * read mode and khugepaged takes it in write mode. So now it's
3405          * safe to run pte_offset_map().
3406          */
3407         pte = pte_offset_map(pmd, address);
3408
3409         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3410 }
3411
3412 #ifndef __PAGETABLE_PUD_FOLDED
3413 /*
3414  * Allocate page upper directory.
3415  * We've already handled the fast-path in-line.
3416  */
3417 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3418 {
3419         pud_t *new = pud_alloc_one(mm, address);
3420         if (!new)
3421                 return -ENOMEM;
3422
3423         smp_wmb(); /* See comment in __pte_alloc */
3424
3425         spin_lock(&mm->page_table_lock);
3426         if (pgd_present(*pgd))          /* Another has populated it */
3427                 pud_free(mm, new);
3428         else
3429                 pgd_populate(mm, pgd, new);
3430         spin_unlock(&mm->page_table_lock);
3431         return 0;
3432 }
3433 #endif /* __PAGETABLE_PUD_FOLDED */
3434
3435 #ifndef __PAGETABLE_PMD_FOLDED
3436 /*
3437  * Allocate page middle directory.
3438  * We've already handled the fast-path in-line.
3439  */
3440 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3441 {
3442         pmd_t *new = pmd_alloc_one(mm, address);
3443         if (!new)
3444                 return -ENOMEM;
3445
3446         smp_wmb(); /* See comment in __pte_alloc */
3447
3448         spin_lock(&mm->page_table_lock);
3449 #ifndef __ARCH_HAS_4LEVEL_HACK
3450         if (pud_present(*pud))          /* Another has populated it */
3451                 pmd_free(mm, new);
3452         else
3453                 pud_populate(mm, pud, new);
3454 #else
3455         if (pgd_present(*pud))          /* Another has populated it */
3456                 pmd_free(mm, new);
3457         else
3458                 pgd_populate(mm, pud, new);
3459 #endif /* __ARCH_HAS_4LEVEL_HACK */
3460         spin_unlock(&mm->page_table_lock);
3461         return 0;
3462 }
3463 #endif /* __PAGETABLE_PMD_FOLDED */
3464
3465 int make_pages_present(unsigned long addr, unsigned long end)
3466 {
3467         int ret, len, write;
3468         struct vm_area_struct * vma;
3469
3470         vma = find_vma(current->mm, addr);
3471         if (!vma)
3472                 return -ENOMEM;
3473         /*
3474          * We want to touch writable mappings with a write fault in order
3475          * to break COW, except for shared mappings because these don't COW
3476          * and we would not want to dirty them for nothing.
3477          */
3478         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3479         BUG_ON(addr >= end);
3480         BUG_ON(end > vma->vm_end);
3481         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3482         ret = get_user_pages(current, current->mm, addr,
3483                         len, write, 0, NULL, NULL);
3484         if (ret < 0)
3485                 return ret;
3486         return ret == len ? 0 : -EFAULT;
3487 }
3488
3489 #if !defined(__HAVE_ARCH_GATE_AREA)
3490
3491 #if defined(AT_SYSINFO_EHDR)
3492 static struct vm_area_struct gate_vma;
3493
3494 static int __init gate_vma_init(void)
3495 {
3496         gate_vma.vm_mm = NULL;
3497         gate_vma.vm_start = FIXADDR_USER_START;
3498         gate_vma.vm_end = FIXADDR_USER_END;
3499         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3500         gate_vma.vm_page_prot = __P101;
3501         /*
3502          * Make sure the vDSO gets into every core dump.
3503          * Dumping its contents makes post-mortem fully interpretable later
3504          * without matching up the same kernel and hardware config to see
3505          * what PC values meant.
3506          */
3507         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3508         return 0;
3509 }
3510 __initcall(gate_vma_init);
3511 #endif
3512
3513 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3514 {
3515 #ifdef AT_SYSINFO_EHDR
3516         return &gate_vma;
3517 #else
3518         return NULL;
3519 #endif
3520 }
3521
3522 int in_gate_area_no_mm(unsigned long addr)
3523 {
3524 #ifdef AT_SYSINFO_EHDR
3525         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3526                 return 1;
3527 #endif
3528         return 0;
3529 }
3530
3531 #endif  /* __HAVE_ARCH_GATE_AREA */
3532
3533 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3534                 pte_t **ptepp, spinlock_t **ptlp)
3535 {
3536         pgd_t *pgd;
3537         pud_t *pud;
3538         pmd_t *pmd;
3539         pte_t *ptep;
3540
3541         pgd = pgd_offset(mm, address);
3542         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3543                 goto out;
3544
3545         pud = pud_offset(pgd, address);
3546         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3547                 goto out;
3548
3549         pmd = pmd_offset(pud, address);
3550         VM_BUG_ON(pmd_trans_huge(*pmd));
3551         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3552                 goto out;
3553
3554         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3555         if (pmd_huge(*pmd))
3556                 goto out;
3557
3558         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3559         if (!ptep)
3560                 goto out;
3561         if (!pte_present(*ptep))
3562                 goto unlock;
3563         *ptepp = ptep;
3564         return 0;
3565 unlock:
3566         pte_unmap_unlock(ptep, *ptlp);
3567 out:
3568         return -EINVAL;
3569 }
3570
3571 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3572                              pte_t **ptepp, spinlock_t **ptlp)
3573 {
3574         int res;
3575
3576         /* (void) is needed to make gcc happy */
3577         (void) __cond_lock(*ptlp,
3578                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3579         return res;
3580 }
3581
3582 /**
3583  * follow_pfn - look up PFN at a user virtual address
3584  * @vma: memory mapping
3585  * @address: user virtual address
3586  * @pfn: location to store found PFN
3587  *
3588  * Only IO mappings and raw PFN mappings are allowed.
3589  *
3590  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3591  */
3592 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3593         unsigned long *pfn)
3594 {
3595         int ret = -EINVAL;
3596         spinlock_t *ptl;
3597         pte_t *ptep;
3598
3599         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3600                 return ret;
3601
3602         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3603         if (ret)
3604                 return ret;
3605         *pfn = pte_pfn(*ptep);
3606         pte_unmap_unlock(ptep, ptl);
3607         return 0;
3608 }
3609 EXPORT_SYMBOL(follow_pfn);
3610
3611 #ifdef CONFIG_HAVE_IOREMAP_PROT
3612 int follow_phys(struct vm_area_struct *vma,
3613                 unsigned long address, unsigned int flags,
3614                 unsigned long *prot, resource_size_t *phys)
3615 {
3616         int ret = -EINVAL;
3617         pte_t *ptep, pte;
3618         spinlock_t *ptl;
3619
3620         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3621                 goto out;
3622
3623         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3624                 goto out;
3625         pte = *ptep;
3626
3627         if ((flags & FOLL_WRITE) && !pte_write(pte))
3628                 goto unlock;
3629
3630         *prot = pgprot_val(pte_pgprot(pte));
3631         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3632
3633         ret = 0;
3634 unlock:
3635         pte_unmap_unlock(ptep, ptl);
3636 out:
3637         return ret;
3638 }
3639
3640 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3641                         void *buf, int len, int write)
3642 {
3643         resource_size_t phys_addr;
3644         unsigned long prot = 0;
3645         void __iomem *maddr;
3646         int offset = addr & (PAGE_SIZE-1);
3647
3648         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3649                 return -EINVAL;
3650
3651         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3652         if (write)
3653                 memcpy_toio(maddr + offset, buf, len);
3654         else
3655                 memcpy_fromio(buf, maddr + offset, len);
3656         iounmap(maddr);
3657
3658         return len;
3659 }
3660 #endif
3661
3662 /*
3663  * Access another process' address space as given in mm.  If non-NULL, use the
3664  * given task for page fault accounting.
3665  */
3666 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3667                 unsigned long addr, void *buf, int len, int write)
3668 {
3669         struct vm_area_struct *vma;
3670         void *old_buf = buf;
3671
3672         down_read(&mm->mmap_sem);
3673         /* ignore errors, just check how much was successfully transferred */
3674         while (len) {
3675                 int bytes, ret, offset;
3676                 void *maddr;
3677                 struct page *page = NULL;
3678
3679                 ret = get_user_pages(tsk, mm, addr, 1,
3680                                 write, 1, &page, &vma);
3681                 if (ret <= 0) {
3682                         /*
3683                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3684                          * we can access using slightly different code.
3685                          */
3686 #ifdef CONFIG_HAVE_IOREMAP_PROT
3687                         vma = find_vma(mm, addr);
3688                         if (!vma || vma->vm_start > addr)
3689                                 break;
3690                         if (vma->vm_ops && vma->vm_ops->access)
3691                                 ret = vma->vm_ops->access(vma, addr, buf,
3692                                                           len, write);
3693                         if (ret <= 0)
3694 #endif
3695                                 break;
3696                         bytes = ret;
3697                 } else {
3698                         bytes = len;
3699                         offset = addr & (PAGE_SIZE-1);
3700                         if (bytes > PAGE_SIZE-offset)
3701                                 bytes = PAGE_SIZE-offset;
3702
3703                         maddr = kmap(page);
3704                         if (write) {
3705                                 copy_to_user_page(vma, page, addr,
3706                                                   maddr + offset, buf, bytes);
3707                                 set_page_dirty_lock(page);
3708                         } else {
3709                                 copy_from_user_page(vma, page, addr,
3710                                                     buf, maddr + offset, bytes);
3711                         }
3712                         kunmap(page);
3713                         page_cache_release(page);
3714                 }
3715                 len -= bytes;
3716                 buf += bytes;
3717                 addr += bytes;
3718         }
3719         up_read(&mm->mmap_sem);
3720
3721         return buf - old_buf;
3722 }
3723
3724 /**
3725  * access_remote_vm - access another process' address space
3726  * @mm:         the mm_struct of the target address space
3727  * @addr:       start address to access
3728  * @buf:        source or destination buffer
3729  * @len:        number of bytes to transfer
3730  * @write:      whether the access is a write
3731  *
3732  * The caller must hold a reference on @mm.
3733  */
3734 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3735                 void *buf, int len, int write)
3736 {
3737         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3738 }
3739
3740 /*
3741  * Access another process' address space.
3742  * Source/target buffer must be kernel space,
3743  * Do not walk the page table directly, use get_user_pages
3744  */
3745 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3746                 void *buf, int len, int write)
3747 {
3748         struct mm_struct *mm;
3749         int ret;
3750
3751         mm = get_task_mm(tsk);
3752         if (!mm)
3753                 return 0;
3754
3755         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3756         mmput(mm);
3757
3758         return ret;
3759 }
3760
3761 /*
3762  * Print the name of a VMA.
3763  */
3764 void print_vma_addr(char *prefix, unsigned long ip)
3765 {
3766         struct mm_struct *mm = current->mm;
3767         struct vm_area_struct *vma;
3768
3769         /*
3770          * Do not print if we are in atomic
3771          * contexts (in exception stacks, etc.):
3772          */
3773         if (preempt_count())
3774                 return;
3775
3776         down_read(&mm->mmap_sem);
3777         vma = find_vma(mm, ip);
3778         if (vma && vma->vm_file) {
3779                 struct file *f = vma->vm_file;
3780                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3781                 if (buf) {
3782                         char *p, *s;
3783
3784                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3785                         if (IS_ERR(p))
3786                                 p = "?";
3787                         s = strrchr(p, '/');
3788                         if (s)
3789                                 p = s+1;
3790                         printk("%s%s[%lx+%lx]", prefix, p,
3791                                         vma->vm_start,
3792                                         vma->vm_end - vma->vm_start);
3793                         free_page((unsigned long)buf);
3794                 }
3795         }
3796         up_read(&current->mm->mmap_sem);
3797 }
3798
3799 #ifdef CONFIG_PROVE_LOCKING
3800 void might_fault(void)
3801 {
3802         /*
3803          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3804          * holding the mmap_sem, this is safe because kernel memory doesn't
3805          * get paged out, therefore we'll never actually fault, and the
3806          * below annotations will generate false positives.
3807          */
3808         if (segment_eq(get_fs(), KERNEL_DS))
3809                 return;
3810
3811         might_sleep();
3812         /*
3813          * it would be nicer only to annotate paths which are not under
3814          * pagefault_disable, however that requires a larger audit and
3815          * providing helpers like get_user_atomic.
3816          */
3817         if (!in_atomic() && current->mm)
3818                 might_lock_read(&current->mm->mmap_sem);
3819 }
3820 EXPORT_SYMBOL(might_fault);
3821 #endif
3822
3823 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3824 static void clear_gigantic_page(struct page *page,
3825                                 unsigned long addr,
3826                                 unsigned int pages_per_huge_page)
3827 {
3828         int i;
3829         struct page *p = page;
3830
3831         might_sleep();
3832         for (i = 0; i < pages_per_huge_page;
3833              i++, p = mem_map_next(p, page, i)) {
3834                 cond_resched();
3835                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3836         }
3837 }
3838 void clear_huge_page(struct page *page,
3839                      unsigned long addr, unsigned int pages_per_huge_page)
3840 {
3841         int i;
3842
3843         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3844                 clear_gigantic_page(page, addr, pages_per_huge_page);
3845                 return;
3846         }
3847
3848         might_sleep();
3849         for (i = 0; i < pages_per_huge_page; i++) {
3850                 cond_resched();
3851                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3852         }
3853 }
3854
3855 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3856                                     unsigned long addr,
3857                                     struct vm_area_struct *vma,
3858                                     unsigned int pages_per_huge_page)
3859 {
3860         int i;
3861         struct page *dst_base = dst;
3862         struct page *src_base = src;
3863
3864         for (i = 0; i < pages_per_huge_page; ) {
3865                 cond_resched();
3866                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3867
3868                 i++;
3869                 dst = mem_map_next(dst, dst_base, i);
3870                 src = mem_map_next(src, src_base, i);
3871         }
3872 }
3873
3874 void copy_user_huge_page(struct page *dst, struct page *src,
3875                          unsigned long addr, struct vm_area_struct *vma,
3876                          unsigned int pages_per_huge_page)
3877 {
3878         int i;
3879
3880         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3881                 copy_user_gigantic_page(dst, src, addr, vma,
3882                                         pages_per_huge_page);
3883                 return;
3884         }
3885
3886         might_sleep();
3887         for (i = 0; i < pages_per_huge_page; i++) {
3888                 cond_resched();
3889                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3890         }
3891 }
3892 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */