VM: skip the stack guard page lookup in get_user_pages only for mlock
[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         pte_t *pte;
916         spinlock_t *ptl;
917         int rss[NR_MM_COUNTERS];
918
919         init_rss_vec(rss);
920
921         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
922         arch_enter_lazy_mmu_mode();
923         do {
924                 pte_t ptent = *pte;
925                 if (pte_none(ptent)) {
926                         (*zap_work)--;
927                         continue;
928                 }
929
930                 (*zap_work) -= PAGE_SIZE;
931
932                 if (pte_present(ptent)) {
933                         struct page *page;
934
935                         page = vm_normal_page(vma, addr, ptent);
936                         if (unlikely(details) && page) {
937                                 /*
938                                  * unmap_shared_mapping_pages() wants to
939                                  * invalidate cache without truncating:
940                                  * unmap shared but keep private pages.
941                                  */
942                                 if (details->check_mapping &&
943                                     details->check_mapping != page->mapping)
944                                         continue;
945                                 /*
946                                  * Each page->index must be checked when
947                                  * invalidating or truncating nonlinear.
948                                  */
949                                 if (details->nonlinear_vma &&
950                                     (page->index < details->first_index ||
951                                      page->index > details->last_index))
952                                         continue;
953                         }
954                         ptent = ptep_get_and_clear_full(mm, addr, pte,
955                                                         tlb->fullmm);
956                         tlb_remove_tlb_entry(tlb, pte, addr);
957                         if (unlikely(!page))
958                                 continue;
959                         if (unlikely(details) && details->nonlinear_vma
960                             && linear_page_index(details->nonlinear_vma,
961                                                 addr) != page->index)
962                                 set_pte_at(mm, addr, pte,
963                                            pgoff_to_pte(page->index));
964                         if (PageAnon(page))
965                                 rss[MM_ANONPAGES]--;
966                         else {
967                                 if (pte_dirty(ptent))
968                                         set_page_dirty(page);
969                                 if (pte_young(ptent) &&
970                                     likely(!VM_SequentialReadHint(vma)))
971                                         mark_page_accessed(page);
972                                 rss[MM_FILEPAGES]--;
973                         }
974                         page_remove_rmap(page);
975                         if (unlikely(page_mapcount(page) < 0))
976                                 print_bad_pte(vma, addr, ptent, page);
977                         tlb_remove_page(tlb, page);
978                         continue;
979                 }
980                 /*
981                  * If details->check_mapping, we leave swap entries;
982                  * if details->nonlinear_vma, we leave file entries.
983                  */
984                 if (unlikely(details))
985                         continue;
986                 if (pte_file(ptent)) {
987                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
988                                 print_bad_pte(vma, addr, ptent, NULL);
989                 } else {
990                         swp_entry_t entry = pte_to_swp_entry(ptent);
991
992                         if (!non_swap_entry(entry))
993                                 rss[MM_SWAPENTS]--;
994                         if (unlikely(!free_swap_and_cache(entry)))
995                                 print_bad_pte(vma, addr, ptent, NULL);
996                 }
997                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
998         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
999
1000         add_mm_rss_vec(mm, rss);
1001         arch_leave_lazy_mmu_mode();
1002         pte_unmap_unlock(pte - 1, ptl);
1003
1004         return addr;
1005 }
1006
1007 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1008                                 struct vm_area_struct *vma, pud_t *pud,
1009                                 unsigned long addr, unsigned long end,
1010                                 long *zap_work, struct zap_details *details)
1011 {
1012         pmd_t *pmd;
1013         unsigned long next;
1014
1015         pmd = pmd_offset(pud, addr);
1016         do {
1017                 next = pmd_addr_end(addr, end);
1018                 if (pmd_trans_huge(*pmd)) {
1019                         if (next-addr != HPAGE_PMD_SIZE) {
1020                                 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1021                                 split_huge_page_pmd(vma->vm_mm, pmd);
1022                         } else if (zap_huge_pmd(tlb, vma, pmd)) {
1023                                 (*zap_work)--;
1024                                 continue;
1025                         }
1026                         /* fall through */
1027                 }
1028                 if (pmd_none_or_clear_bad(pmd)) {
1029                         (*zap_work)--;
1030                         continue;
1031                 }
1032                 next = zap_pte_range(tlb, vma, pmd, addr, next,
1033                                                 zap_work, details);
1034         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
1035
1036         return addr;
1037 }
1038
1039 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1040                                 struct vm_area_struct *vma, pgd_t *pgd,
1041                                 unsigned long addr, unsigned long end,
1042                                 long *zap_work, struct zap_details *details)
1043 {
1044         pud_t *pud;
1045         unsigned long next;
1046
1047         pud = pud_offset(pgd, addr);
1048         do {
1049                 next = pud_addr_end(addr, end);
1050                 if (pud_none_or_clear_bad(pud)) {
1051                         (*zap_work)--;
1052                         continue;
1053                 }
1054                 next = zap_pmd_range(tlb, vma, pud, addr, next,
1055                                                 zap_work, details);
1056         } while (pud++, addr = next, (addr != end && *zap_work > 0));
1057
1058         return addr;
1059 }
1060
1061 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1062                                 struct vm_area_struct *vma,
1063                                 unsigned long addr, unsigned long end,
1064                                 long *zap_work, struct zap_details *details)
1065 {
1066         pgd_t *pgd;
1067         unsigned long next;
1068
1069         if (details && !details->check_mapping && !details->nonlinear_vma)
1070                 details = NULL;
1071
1072         BUG_ON(addr >= end);
1073         mem_cgroup_uncharge_start();
1074         tlb_start_vma(tlb, vma);
1075         pgd = pgd_offset(vma->vm_mm, addr);
1076         do {
1077                 next = pgd_addr_end(addr, end);
1078                 if (pgd_none_or_clear_bad(pgd)) {
1079                         (*zap_work)--;
1080                         continue;
1081                 }
1082                 next = zap_pud_range(tlb, vma, pgd, addr, next,
1083                                                 zap_work, details);
1084         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
1085         tlb_end_vma(tlb, vma);
1086         mem_cgroup_uncharge_end();
1087
1088         return addr;
1089 }
1090
1091 #ifdef CONFIG_PREEMPT
1092 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1093 #else
1094 /* No preempt: go for improved straight-line efficiency */
1095 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1096 #endif
1097
1098 /**
1099  * unmap_vmas - unmap a range of memory covered by a list of vma's
1100  * @tlbp: address of the caller's struct mmu_gather
1101  * @vma: the starting vma
1102  * @start_addr: virtual address at which to start unmapping
1103  * @end_addr: virtual address at which to end unmapping
1104  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1105  * @details: details of nonlinear truncation or shared cache invalidation
1106  *
1107  * Returns the end address of the unmapping (restart addr if interrupted).
1108  *
1109  * Unmap all pages in the vma list.
1110  *
1111  * We aim to not hold locks for too long (for scheduling latency reasons).
1112  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
1113  * return the ending mmu_gather to the caller.
1114  *
1115  * Only addresses between `start' and `end' will be unmapped.
1116  *
1117  * The VMA list must be sorted in ascending virtual address order.
1118  *
1119  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1120  * range after unmap_vmas() returns.  So the only responsibility here is to
1121  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1122  * drops the lock and schedules.
1123  */
1124 unsigned long unmap_vmas(struct mmu_gather **tlbp,
1125                 struct vm_area_struct *vma, unsigned long start_addr,
1126                 unsigned long end_addr, unsigned long *nr_accounted,
1127                 struct zap_details *details)
1128 {
1129         long zap_work = ZAP_BLOCK_SIZE;
1130         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
1131         int tlb_start_valid = 0;
1132         unsigned long start = start_addr;
1133         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1134         int fullmm = (*tlbp)->fullmm;
1135         struct mm_struct *mm = vma->vm_mm;
1136
1137         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1138         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1139                 unsigned long end;
1140
1141                 start = max(vma->vm_start, start_addr);
1142                 if (start >= vma->vm_end)
1143                         continue;
1144                 end = min(vma->vm_end, end_addr);
1145                 if (end <= vma->vm_start)
1146                         continue;
1147
1148                 if (vma->vm_flags & VM_ACCOUNT)
1149                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1150
1151                 if (unlikely(is_pfn_mapping(vma)))
1152                         untrack_pfn_vma(vma, 0, 0);
1153
1154                 while (start != end) {
1155                         if (!tlb_start_valid) {
1156                                 tlb_start = start;
1157                                 tlb_start_valid = 1;
1158                         }
1159
1160                         if (unlikely(is_vm_hugetlb_page(vma))) {
1161                                 /*
1162                                  * It is undesirable to test vma->vm_file as it
1163                                  * should be non-null for valid hugetlb area.
1164                                  * However, vm_file will be NULL in the error
1165                                  * cleanup path of do_mmap_pgoff. When
1166                                  * hugetlbfs ->mmap method fails,
1167                                  * do_mmap_pgoff() nullifies vma->vm_file
1168                                  * before calling this function to clean up.
1169                                  * Since no pte has actually been setup, it is
1170                                  * safe to do nothing in this case.
1171                                  */
1172                                 if (vma->vm_file) {
1173                                         unmap_hugepage_range(vma, start, end, NULL);
1174                                         zap_work -= (end - start) /
1175                                         pages_per_huge_page(hstate_vma(vma));
1176                                 }
1177
1178                                 start = end;
1179                         } else
1180                                 start = unmap_page_range(*tlbp, vma,
1181                                                 start, end, &zap_work, details);
1182
1183                         if (zap_work > 0) {
1184                                 BUG_ON(start != end);
1185                                 break;
1186                         }
1187
1188                         tlb_finish_mmu(*tlbp, tlb_start, start);
1189
1190                         if (need_resched() ||
1191                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1192                                 if (i_mmap_lock) {
1193                                         *tlbp = NULL;
1194                                         goto out;
1195                                 }
1196                                 cond_resched();
1197                         }
1198
1199                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1200                         tlb_start_valid = 0;
1201                         zap_work = ZAP_BLOCK_SIZE;
1202                 }
1203         }
1204 out:
1205         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1206         return start;   /* which is now the end (or restart) address */
1207 }
1208
1209 /**
1210  * zap_page_range - remove user pages in a given range
1211  * @vma: vm_area_struct holding the applicable pages
1212  * @address: starting address of pages to zap
1213  * @size: number of bytes to zap
1214  * @details: details of nonlinear truncation or shared cache invalidation
1215  */
1216 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1217                 unsigned long size, struct zap_details *details)
1218 {
1219         struct mm_struct *mm = vma->vm_mm;
1220         struct mmu_gather *tlb;
1221         unsigned long end = address + size;
1222         unsigned long nr_accounted = 0;
1223
1224         lru_add_drain();
1225         tlb = tlb_gather_mmu(mm, 0);
1226         update_hiwater_rss(mm);
1227         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1228         if (tlb)
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 (vma->vm_flags & VM_GROWSDOWN) &&
1416                 (vma->vm_start == addr) &&
1417                 !vma_stack_continue(vma->vm_prev, addr);
1418 }
1419
1420 /**
1421  * __get_user_pages() - pin user pages in memory
1422  * @tsk:        task_struct of target task
1423  * @mm:         mm_struct of target mm
1424  * @start:      starting user address
1425  * @nr_pages:   number of pages from start to pin
1426  * @gup_flags:  flags modifying pin behaviour
1427  * @pages:      array that receives pointers to the pages pinned.
1428  *              Should be at least nr_pages long. Or NULL, if caller
1429  *              only intends to ensure the pages are faulted in.
1430  * @vmas:       array of pointers to vmas corresponding to each page.
1431  *              Or NULL if the caller does not require them.
1432  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1433  *
1434  * Returns number of pages pinned. This may be fewer than the number
1435  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1436  * were pinned, returns -errno. Each page returned must be released
1437  * with a put_page() call when it is finished with. vmas will only
1438  * remain valid while mmap_sem is held.
1439  *
1440  * Must be called with mmap_sem held for read or write.
1441  *
1442  * __get_user_pages walks a process's page tables and takes a reference to
1443  * each struct page that each user address corresponds to at a given
1444  * instant. That is, it takes the page that would be accessed if a user
1445  * thread accesses the given user virtual address at that instant.
1446  *
1447  * This does not guarantee that the page exists in the user mappings when
1448  * __get_user_pages returns, and there may even be a completely different
1449  * page there in some cases (eg. if mmapped pagecache has been invalidated
1450  * and subsequently re faulted). However it does guarantee that the page
1451  * won't be freed completely. And mostly callers simply care that the page
1452  * contains data that was valid *at some point in time*. Typically, an IO
1453  * or similar operation cannot guarantee anything stronger anyway because
1454  * locks can't be held over the syscall boundary.
1455  *
1456  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1457  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1458  * appropriate) must be called after the page is finished with, and
1459  * before put_page is called.
1460  *
1461  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1462  * or mmap_sem contention, and if waiting is needed to pin all pages,
1463  * *@nonblocking will be set to 0.
1464  *
1465  * In most cases, get_user_pages or get_user_pages_fast should be used
1466  * instead of __get_user_pages. __get_user_pages should be used only if
1467  * you need some special @gup_flags.
1468  */
1469 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1470                      unsigned long start, int nr_pages, unsigned int gup_flags,
1471                      struct page **pages, struct vm_area_struct **vmas,
1472                      int *nonblocking)
1473 {
1474         int i;
1475         unsigned long vm_flags;
1476
1477         if (nr_pages <= 0)
1478                 return 0;
1479
1480         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1481
1482         /* 
1483          * Require read or write permissions.
1484          * If FOLL_FORCE is set, we only require the "MAY" flags.
1485          */
1486         vm_flags  = (gup_flags & FOLL_WRITE) ?
1487                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1488         vm_flags &= (gup_flags & FOLL_FORCE) ?
1489                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1490         i = 0;
1491
1492         do {
1493                 struct vm_area_struct *vma;
1494
1495                 vma = find_extend_vma(mm, start);
1496                 if (!vma && in_gate_area(mm, start)) {
1497                         unsigned long pg = start & PAGE_MASK;
1498                         pgd_t *pgd;
1499                         pud_t *pud;
1500                         pmd_t *pmd;
1501                         pte_t *pte;
1502
1503                         /* user gate pages are read-only */
1504                         if (gup_flags & FOLL_WRITE)
1505                                 return i ? : -EFAULT;
1506                         if (pg > TASK_SIZE)
1507                                 pgd = pgd_offset_k(pg);
1508                         else
1509                                 pgd = pgd_offset_gate(mm, pg);
1510                         BUG_ON(pgd_none(*pgd));
1511                         pud = pud_offset(pgd, pg);
1512                         BUG_ON(pud_none(*pud));
1513                         pmd = pmd_offset(pud, pg);
1514                         if (pmd_none(*pmd))
1515                                 return i ? : -EFAULT;
1516                         VM_BUG_ON(pmd_trans_huge(*pmd));
1517                         pte = pte_offset_map(pmd, pg);
1518                         if (pte_none(*pte)) {
1519                                 pte_unmap(pte);
1520                                 return i ? : -EFAULT;
1521                         }
1522                         vma = get_gate_vma(mm);
1523                         if (pages) {
1524                                 struct page *page;
1525
1526                                 page = vm_normal_page(vma, start, *pte);
1527                                 if (!page) {
1528                                         if (!(gup_flags & FOLL_DUMP) &&
1529                                              is_zero_pfn(pte_pfn(*pte)))
1530                                                 page = pte_page(*pte);
1531                                         else {
1532                                                 pte_unmap(pte);
1533                                                 return i ? : -EFAULT;
1534                                         }
1535                                 }
1536                                 pages[i] = page;
1537                                 get_page(page);
1538                         }
1539                         pte_unmap(pte);
1540                         goto next_page;
1541                 }
1542
1543                 if (!vma ||
1544                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1545                     !(vm_flags & vma->vm_flags))
1546                         return i ? : -EFAULT;
1547
1548                 if (is_vm_hugetlb_page(vma)) {
1549                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1550                                         &start, &nr_pages, i, gup_flags);
1551                         continue;
1552                 }
1553
1554                 /*
1555                  * For mlock, just skip the stack guard page.
1556                  */
1557                 if ((gup_flags & FOLL_MLOCK) && stack_guard_page(vma, start))
1558                         goto next_page;
1559
1560                 do {
1561                         struct page *page;
1562                         unsigned int foll_flags = gup_flags;
1563
1564                         /*
1565                          * If we have a pending SIGKILL, don't keep faulting
1566                          * pages and potentially allocating memory.
1567                          */
1568                         if (unlikely(fatal_signal_pending(current)))
1569                                 return i ? i : -ERESTARTSYS;
1570
1571                         cond_resched();
1572                         while (!(page = follow_page(vma, start, foll_flags))) {
1573                                 int ret;
1574                                 unsigned int fault_flags = 0;
1575
1576                                 if (foll_flags & FOLL_WRITE)
1577                                         fault_flags |= FAULT_FLAG_WRITE;
1578                                 if (nonblocking)
1579                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1580                                 if (foll_flags & FOLL_NOWAIT)
1581                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1582
1583                                 ret = handle_mm_fault(mm, vma, start,
1584                                                         fault_flags);
1585
1586                                 if (ret & VM_FAULT_ERROR) {
1587                                         if (ret & VM_FAULT_OOM)
1588                                                 return i ? i : -ENOMEM;
1589                                         if (ret & (VM_FAULT_HWPOISON |
1590                                                    VM_FAULT_HWPOISON_LARGE)) {
1591                                                 if (i)
1592                                                         return i;
1593                                                 else if (gup_flags & FOLL_HWPOISON)
1594                                                         return -EHWPOISON;
1595                                                 else
1596                                                         return -EFAULT;
1597                                         }
1598                                         if (ret & VM_FAULT_SIGBUS)
1599                                                 return i ? i : -EFAULT;
1600                                         BUG();
1601                                 }
1602
1603                                 if (tsk) {
1604                                         if (ret & VM_FAULT_MAJOR)
1605                                                 tsk->maj_flt++;
1606                                         else
1607                                                 tsk->min_flt++;
1608                                 }
1609
1610                                 if (ret & VM_FAULT_RETRY) {
1611                                         if (nonblocking)
1612                                                 *nonblocking = 0;
1613                                         return i;
1614                                 }
1615
1616                                 /*
1617                                  * The VM_FAULT_WRITE bit tells us that
1618                                  * do_wp_page has broken COW when necessary,
1619                                  * even if maybe_mkwrite decided not to set
1620                                  * pte_write. We can thus safely do subsequent
1621                                  * page lookups as if they were reads. But only
1622                                  * do so when looping for pte_write is futile:
1623                                  * in some cases userspace may also be wanting
1624                                  * to write to the gotten user page, which a
1625                                  * read fault here might prevent (a readonly
1626                                  * page might get reCOWed by userspace write).
1627                                  */
1628                                 if ((ret & VM_FAULT_WRITE) &&
1629                                     !(vma->vm_flags & VM_WRITE))
1630                                         foll_flags &= ~FOLL_WRITE;
1631
1632                                 cond_resched();
1633                         }
1634                         if (IS_ERR(page))
1635                                 return i ? i : PTR_ERR(page);
1636                         if (pages) {
1637                                 pages[i] = page;
1638
1639                                 flush_anon_page(vma, page, start);
1640                                 flush_dcache_page(page);
1641                         }
1642 next_page:
1643                         if (vmas)
1644                                 vmas[i] = vma;
1645                         i++;
1646                         start += PAGE_SIZE;
1647                         nr_pages--;
1648                 } while (nr_pages && start < vma->vm_end);
1649         } while (nr_pages);
1650         return i;
1651 }
1652 EXPORT_SYMBOL(__get_user_pages);
1653
1654 /**
1655  * get_user_pages() - pin user pages in memory
1656  * @tsk:        the task_struct to use for page fault accounting, or
1657  *              NULL if faults are not to be recorded.
1658  * @mm:         mm_struct of target mm
1659  * @start:      starting user address
1660  * @nr_pages:   number of pages from start to pin
1661  * @write:      whether pages will be written to by the caller
1662  * @force:      whether to force write access even if user mapping is
1663  *              readonly. This will result in the page being COWed even
1664  *              in MAP_SHARED mappings. You do not want this.
1665  * @pages:      array that receives pointers to the pages pinned.
1666  *              Should be at least nr_pages long. Or NULL, if caller
1667  *              only intends to ensure the pages are faulted in.
1668  * @vmas:       array of pointers to vmas corresponding to each page.
1669  *              Or NULL if the caller does not require them.
1670  *
1671  * Returns number of pages pinned. This may be fewer than the number
1672  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1673  * were pinned, returns -errno. Each page returned must be released
1674  * with a put_page() call when it is finished with. vmas will only
1675  * remain valid while mmap_sem is held.
1676  *
1677  * Must be called with mmap_sem held for read or write.
1678  *
1679  * get_user_pages walks a process's page tables and takes a reference to
1680  * each struct page that each user address corresponds to at a given
1681  * instant. That is, it takes the page that would be accessed if a user
1682  * thread accesses the given user virtual address at that instant.
1683  *
1684  * This does not guarantee that the page exists in the user mappings when
1685  * get_user_pages returns, and there may even be a completely different
1686  * page there in some cases (eg. if mmapped pagecache has been invalidated
1687  * and subsequently re faulted). However it does guarantee that the page
1688  * won't be freed completely. And mostly callers simply care that the page
1689  * contains data that was valid *at some point in time*. Typically, an IO
1690  * or similar operation cannot guarantee anything stronger anyway because
1691  * locks can't be held over the syscall boundary.
1692  *
1693  * If write=0, the page must not be written to. If the page is written to,
1694  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1695  * after the page is finished with, and before put_page is called.
1696  *
1697  * get_user_pages is typically used for fewer-copy IO operations, to get a
1698  * handle on the memory by some means other than accesses via the user virtual
1699  * addresses. The pages may be submitted for DMA to devices or accessed via
1700  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1701  * use the correct cache flushing APIs.
1702  *
1703  * See also get_user_pages_fast, for performance critical applications.
1704  */
1705 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1706                 unsigned long start, int nr_pages, int write, int force,
1707                 struct page **pages, struct vm_area_struct **vmas)
1708 {
1709         int flags = FOLL_TOUCH;
1710
1711         if (pages)
1712                 flags |= FOLL_GET;
1713         if (write)
1714                 flags |= FOLL_WRITE;
1715         if (force)
1716                 flags |= FOLL_FORCE;
1717
1718         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1719                                 NULL);
1720 }
1721 EXPORT_SYMBOL(get_user_pages);
1722
1723 /**
1724  * get_dump_page() - pin user page in memory while writing it to core dump
1725  * @addr: user address
1726  *
1727  * Returns struct page pointer of user page pinned for dump,
1728  * to be freed afterwards by page_cache_release() or put_page().
1729  *
1730  * Returns NULL on any kind of failure - a hole must then be inserted into
1731  * the corefile, to preserve alignment with its headers; and also returns
1732  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1733  * allowing a hole to be left in the corefile to save diskspace.
1734  *
1735  * Called without mmap_sem, but after all other threads have been killed.
1736  */
1737 #ifdef CONFIG_ELF_CORE
1738 struct page *get_dump_page(unsigned long addr)
1739 {
1740         struct vm_area_struct *vma;
1741         struct page *page;
1742
1743         if (__get_user_pages(current, current->mm, addr, 1,
1744                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1745                              NULL) < 1)
1746                 return NULL;
1747         flush_cache_page(vma, addr, page_to_pfn(page));
1748         return page;
1749 }
1750 #endif /* CONFIG_ELF_CORE */
1751
1752 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1753                         spinlock_t **ptl)
1754 {
1755         pgd_t * pgd = pgd_offset(mm, addr);
1756         pud_t * pud = pud_alloc(mm, pgd, addr);
1757         if (pud) {
1758                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1759                 if (pmd) {
1760                         VM_BUG_ON(pmd_trans_huge(*pmd));
1761                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1762                 }
1763         }
1764         return NULL;
1765 }
1766
1767 /*
1768  * This is the old fallback for page remapping.
1769  *
1770  * For historical reasons, it only allows reserved pages. Only
1771  * old drivers should use this, and they needed to mark their
1772  * pages reserved for the old functions anyway.
1773  */
1774 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1775                         struct page *page, pgprot_t prot)
1776 {
1777         struct mm_struct *mm = vma->vm_mm;
1778         int retval;
1779         pte_t *pte;
1780         spinlock_t *ptl;
1781
1782         retval = -EINVAL;
1783         if (PageAnon(page))
1784                 goto out;
1785         retval = -ENOMEM;
1786         flush_dcache_page(page);
1787         pte = get_locked_pte(mm, addr, &ptl);
1788         if (!pte)
1789                 goto out;
1790         retval = -EBUSY;
1791         if (!pte_none(*pte))
1792                 goto out_unlock;
1793
1794         /* Ok, finally just insert the thing.. */
1795         get_page(page);
1796         inc_mm_counter_fast(mm, MM_FILEPAGES);
1797         page_add_file_rmap(page);
1798         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1799
1800         retval = 0;
1801         pte_unmap_unlock(pte, ptl);
1802         return retval;
1803 out_unlock:
1804         pte_unmap_unlock(pte, ptl);
1805 out:
1806         return retval;
1807 }
1808
1809 /**
1810  * vm_insert_page - insert single page into user vma
1811  * @vma: user vma to map to
1812  * @addr: target user address of this page
1813  * @page: source kernel page
1814  *
1815  * This allows drivers to insert individual pages they've allocated
1816  * into a user vma.
1817  *
1818  * The page has to be a nice clean _individual_ kernel allocation.
1819  * If you allocate a compound page, you need to have marked it as
1820  * such (__GFP_COMP), or manually just split the page up yourself
1821  * (see split_page()).
1822  *
1823  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1824  * took an arbitrary page protection parameter. This doesn't allow
1825  * that. Your vma protection will have to be set up correctly, which
1826  * means that if you want a shared writable mapping, you'd better
1827  * ask for a shared writable mapping!
1828  *
1829  * The page does not need to be reserved.
1830  */
1831 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1832                         struct page *page)
1833 {
1834         if (addr < vma->vm_start || addr >= vma->vm_end)
1835                 return -EFAULT;
1836         if (!page_count(page))
1837                 return -EINVAL;
1838         vma->vm_flags |= VM_INSERTPAGE;
1839         return insert_page(vma, addr, page, vma->vm_page_prot);
1840 }
1841 EXPORT_SYMBOL(vm_insert_page);
1842
1843 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1844                         unsigned long pfn, pgprot_t prot)
1845 {
1846         struct mm_struct *mm = vma->vm_mm;
1847         int retval;
1848         pte_t *pte, entry;
1849         spinlock_t *ptl;
1850
1851         retval = -ENOMEM;
1852         pte = get_locked_pte(mm, addr, &ptl);
1853         if (!pte)
1854                 goto out;
1855         retval = -EBUSY;
1856         if (!pte_none(*pte))
1857                 goto out_unlock;
1858
1859         /* Ok, finally just insert the thing.. */
1860         entry = pte_mkspecial(pfn_pte(pfn, prot));
1861         set_pte_at(mm, addr, pte, entry);
1862         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1863
1864         retval = 0;
1865 out_unlock:
1866         pte_unmap_unlock(pte, ptl);
1867 out:
1868         return retval;
1869 }
1870
1871 /**
1872  * vm_insert_pfn - insert single pfn into user vma
1873  * @vma: user vma to map to
1874  * @addr: target user address of this page
1875  * @pfn: source kernel pfn
1876  *
1877  * Similar to vm_inert_page, this allows drivers to insert individual pages
1878  * they've allocated into a user vma. Same comments apply.
1879  *
1880  * This function should only be called from a vm_ops->fault handler, and
1881  * in that case the handler should return NULL.
1882  *
1883  * vma cannot be a COW mapping.
1884  *
1885  * As this is called only for pages that do not currently exist, we
1886  * do not need to flush old virtual caches or the TLB.
1887  */
1888 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1889                         unsigned long pfn)
1890 {
1891         int ret;
1892         pgprot_t pgprot = vma->vm_page_prot;
1893         /*
1894          * Technically, architectures with pte_special can avoid all these
1895          * restrictions (same for remap_pfn_range).  However we would like
1896          * consistency in testing and feature parity among all, so we should
1897          * try to keep these invariants in place for everybody.
1898          */
1899         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1900         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1901                                                 (VM_PFNMAP|VM_MIXEDMAP));
1902         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1903         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1904
1905         if (addr < vma->vm_start || addr >= vma->vm_end)
1906                 return -EFAULT;
1907         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1908                 return -EINVAL;
1909
1910         ret = insert_pfn(vma, addr, pfn, pgprot);
1911
1912         if (ret)
1913                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1914
1915         return ret;
1916 }
1917 EXPORT_SYMBOL(vm_insert_pfn);
1918
1919 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1920                         unsigned long pfn)
1921 {
1922         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1923
1924         if (addr < vma->vm_start || addr >= vma->vm_end)
1925                 return -EFAULT;
1926
1927         /*
1928          * If we don't have pte special, then we have to use the pfn_valid()
1929          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1930          * refcount the page if pfn_valid is true (hence insert_page rather
1931          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1932          * without pte special, it would there be refcounted as a normal page.
1933          */
1934         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1935                 struct page *page;
1936
1937                 page = pfn_to_page(pfn);
1938                 return insert_page(vma, addr, page, vma->vm_page_prot);
1939         }
1940         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1941 }
1942 EXPORT_SYMBOL(vm_insert_mixed);
1943
1944 /*
1945  * maps a range of physical memory into the requested pages. the old
1946  * mappings are removed. any references to nonexistent pages results
1947  * in null mappings (currently treated as "copy-on-access")
1948  */
1949 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1950                         unsigned long addr, unsigned long end,
1951                         unsigned long pfn, pgprot_t prot)
1952 {
1953         pte_t *pte;
1954         spinlock_t *ptl;
1955
1956         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1957         if (!pte)
1958                 return -ENOMEM;
1959         arch_enter_lazy_mmu_mode();
1960         do {
1961                 BUG_ON(!pte_none(*pte));
1962                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1963                 pfn++;
1964         } while (pte++, addr += PAGE_SIZE, addr != end);
1965         arch_leave_lazy_mmu_mode();
1966         pte_unmap_unlock(pte - 1, ptl);
1967         return 0;
1968 }
1969
1970 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1971                         unsigned long addr, unsigned long end,
1972                         unsigned long pfn, pgprot_t prot)
1973 {
1974         pmd_t *pmd;
1975         unsigned long next;
1976
1977         pfn -= addr >> PAGE_SHIFT;
1978         pmd = pmd_alloc(mm, pud, addr);
1979         if (!pmd)
1980                 return -ENOMEM;
1981         VM_BUG_ON(pmd_trans_huge(*pmd));
1982         do {
1983                 next = pmd_addr_end(addr, end);
1984                 if (remap_pte_range(mm, pmd, addr, next,
1985                                 pfn + (addr >> PAGE_SHIFT), prot))
1986                         return -ENOMEM;
1987         } while (pmd++, addr = next, addr != end);
1988         return 0;
1989 }
1990
1991 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1992                         unsigned long addr, unsigned long end,
1993                         unsigned long pfn, pgprot_t prot)
1994 {
1995         pud_t *pud;
1996         unsigned long next;
1997
1998         pfn -= addr >> PAGE_SHIFT;
1999         pud = pud_alloc(mm, pgd, addr);
2000         if (!pud)
2001                 return -ENOMEM;
2002         do {
2003                 next = pud_addr_end(addr, end);
2004                 if (remap_pmd_range(mm, pud, addr, next,
2005                                 pfn + (addr >> PAGE_SHIFT), prot))
2006                         return -ENOMEM;
2007         } while (pud++, addr = next, addr != end);
2008         return 0;
2009 }
2010
2011 /**
2012  * remap_pfn_range - remap kernel memory to userspace
2013  * @vma: user vma to map to
2014  * @addr: target user address to start at
2015  * @pfn: physical address of kernel memory
2016  * @size: size of map area
2017  * @prot: page protection flags for this mapping
2018  *
2019  *  Note: this is only safe if the mm semaphore is held when called.
2020  */
2021 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2022                     unsigned long pfn, unsigned long size, pgprot_t prot)
2023 {
2024         pgd_t *pgd;
2025         unsigned long next;
2026         unsigned long end = addr + PAGE_ALIGN(size);
2027         struct mm_struct *mm = vma->vm_mm;
2028         int err;
2029
2030         /*
2031          * Physically remapped pages are special. Tell the
2032          * rest of the world about it:
2033          *   VM_IO tells people not to look at these pages
2034          *      (accesses can have side effects).
2035          *   VM_RESERVED is specified all over the place, because
2036          *      in 2.4 it kept swapout's vma scan off this vma; but
2037          *      in 2.6 the LRU scan won't even find its pages, so this
2038          *      flag means no more than count its pages in reserved_vm,
2039          *      and omit it from core dump, even when VM_IO turned off.
2040          *   VM_PFNMAP tells the core MM that the base pages are just
2041          *      raw PFN mappings, and do not have a "struct page" associated
2042          *      with them.
2043          *
2044          * There's a horrible special case to handle copy-on-write
2045          * behaviour that some programs depend on. We mark the "original"
2046          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2047          */
2048         if (addr == vma->vm_start && end == vma->vm_end) {
2049                 vma->vm_pgoff = pfn;
2050                 vma->vm_flags |= VM_PFN_AT_MMAP;
2051         } else if (is_cow_mapping(vma->vm_flags))
2052                 return -EINVAL;
2053
2054         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2055
2056         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2057         if (err) {
2058                 /*
2059                  * To indicate that track_pfn related cleanup is not
2060                  * needed from higher level routine calling unmap_vmas
2061                  */
2062                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2063                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2064                 return -EINVAL;
2065         }
2066
2067         BUG_ON(addr >= end);
2068         pfn -= addr >> PAGE_SHIFT;
2069         pgd = pgd_offset(mm, addr);
2070         flush_cache_range(vma, addr, end);
2071         do {
2072                 next = pgd_addr_end(addr, end);
2073                 err = remap_pud_range(mm, pgd, addr, next,
2074                                 pfn + (addr >> PAGE_SHIFT), prot);
2075                 if (err)
2076                         break;
2077         } while (pgd++, addr = next, addr != end);
2078
2079         if (err)
2080                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2081
2082         return err;
2083 }
2084 EXPORT_SYMBOL(remap_pfn_range);
2085
2086 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2087                                      unsigned long addr, unsigned long end,
2088                                      pte_fn_t fn, void *data)
2089 {
2090         pte_t *pte;
2091         int err;
2092         pgtable_t token;
2093         spinlock_t *uninitialized_var(ptl);
2094
2095         pte = (mm == &init_mm) ?
2096                 pte_alloc_kernel(pmd, addr) :
2097                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2098         if (!pte)
2099                 return -ENOMEM;
2100
2101         BUG_ON(pmd_huge(*pmd));
2102
2103         arch_enter_lazy_mmu_mode();
2104
2105         token = pmd_pgtable(*pmd);
2106
2107         do {
2108                 err = fn(pte++, token, addr, data);
2109                 if (err)
2110                         break;
2111         } while (addr += PAGE_SIZE, addr != end);
2112
2113         arch_leave_lazy_mmu_mode();
2114
2115         if (mm != &init_mm)
2116                 pte_unmap_unlock(pte-1, ptl);
2117         return err;
2118 }
2119
2120 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2121                                      unsigned long addr, unsigned long end,
2122                                      pte_fn_t fn, void *data)
2123 {
2124         pmd_t *pmd;
2125         unsigned long next;
2126         int err;
2127
2128         BUG_ON(pud_huge(*pud));
2129
2130         pmd = pmd_alloc(mm, pud, addr);
2131         if (!pmd)
2132                 return -ENOMEM;
2133         do {
2134                 next = pmd_addr_end(addr, end);
2135                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2136                 if (err)
2137                         break;
2138         } while (pmd++, addr = next, addr != end);
2139         return err;
2140 }
2141
2142 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2143                                      unsigned long addr, unsigned long end,
2144                                      pte_fn_t fn, void *data)
2145 {
2146         pud_t *pud;
2147         unsigned long next;
2148         int err;
2149
2150         pud = pud_alloc(mm, pgd, addr);
2151         if (!pud)
2152                 return -ENOMEM;
2153         do {
2154                 next = pud_addr_end(addr, end);
2155                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2156                 if (err)
2157                         break;
2158         } while (pud++, addr = next, addr != end);
2159         return err;
2160 }
2161
2162 /*
2163  * Scan a region of virtual memory, filling in page tables as necessary
2164  * and calling a provided function on each leaf page table.
2165  */
2166 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2167                         unsigned long size, pte_fn_t fn, void *data)
2168 {
2169         pgd_t *pgd;
2170         unsigned long next;
2171         unsigned long end = addr + size;
2172         int err;
2173
2174         BUG_ON(addr >= end);
2175         pgd = pgd_offset(mm, addr);
2176         do {
2177                 next = pgd_addr_end(addr, end);
2178                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2179                 if (err)
2180                         break;
2181         } while (pgd++, addr = next, addr != end);
2182
2183         return err;
2184 }
2185 EXPORT_SYMBOL_GPL(apply_to_page_range);
2186
2187 /*
2188  * handle_pte_fault chooses page fault handler according to an entry
2189  * which was read non-atomically.  Before making any commitment, on
2190  * those architectures or configurations (e.g. i386 with PAE) which
2191  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2192  * must check under lock before unmapping the pte and proceeding
2193  * (but do_wp_page is only called after already making such a check;
2194  * and do_anonymous_page can safely check later on).
2195  */
2196 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2197                                 pte_t *page_table, pte_t orig_pte)
2198 {
2199         int same = 1;
2200 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2201         if (sizeof(pte_t) > sizeof(unsigned long)) {
2202                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2203                 spin_lock(ptl);
2204                 same = pte_same(*page_table, orig_pte);
2205                 spin_unlock(ptl);
2206         }
2207 #endif
2208         pte_unmap(page_table);
2209         return same;
2210 }
2211
2212 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2213 {
2214         /*
2215          * If the source page was a PFN mapping, we don't have
2216          * a "struct page" for it. We do a best-effort copy by
2217          * just copying from the original user address. If that
2218          * fails, we just zero-fill it. Live with it.
2219          */
2220         if (unlikely(!src)) {
2221                 void *kaddr = kmap_atomic(dst, KM_USER0);
2222                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2223
2224                 /*
2225                  * This really shouldn't fail, because the page is there
2226                  * in the page tables. But it might just be unreadable,
2227                  * in which case we just give up and fill the result with
2228                  * zeroes.
2229                  */
2230                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2231                         clear_page(kaddr);
2232                 kunmap_atomic(kaddr, KM_USER0);
2233                 flush_dcache_page(dst);
2234         } else
2235                 copy_user_highpage(dst, src, va, vma);
2236 }
2237
2238 /*
2239  * This routine handles present pages, when users try to write
2240  * to a shared page. It is done by copying the page to a new address
2241  * and decrementing the shared-page counter for the old page.
2242  *
2243  * Note that this routine assumes that the protection checks have been
2244  * done by the caller (the low-level page fault routine in most cases).
2245  * Thus we can safely just mark it writable once we've done any necessary
2246  * COW.
2247  *
2248  * We also mark the page dirty at this point even though the page will
2249  * change only once the write actually happens. This avoids a few races,
2250  * and potentially makes it more efficient.
2251  *
2252  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2253  * but allow concurrent faults), with pte both mapped and locked.
2254  * We return with mmap_sem still held, but pte unmapped and unlocked.
2255  */
2256 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2257                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2258                 spinlock_t *ptl, pte_t orig_pte)
2259         __releases(ptl)
2260 {
2261         struct page *old_page, *new_page;
2262         pte_t entry;
2263         int ret = 0;
2264         int page_mkwrite = 0;
2265         struct page *dirty_page = NULL;
2266
2267         old_page = vm_normal_page(vma, address, orig_pte);
2268         if (!old_page) {
2269                 /*
2270                  * VM_MIXEDMAP !pfn_valid() case
2271                  *
2272                  * We should not cow pages in a shared writeable mapping.
2273                  * Just mark the pages writable as we can't do any dirty
2274                  * accounting on raw pfn maps.
2275                  */
2276                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2277                                      (VM_WRITE|VM_SHARED))
2278                         goto reuse;
2279                 goto gotten;
2280         }
2281
2282         /*
2283          * Take out anonymous pages first, anonymous shared vmas are
2284          * not dirty accountable.
2285          */
2286         if (PageAnon(old_page) && !PageKsm(old_page)) {
2287                 if (!trylock_page(old_page)) {
2288                         page_cache_get(old_page);
2289                         pte_unmap_unlock(page_table, ptl);
2290                         lock_page(old_page);
2291                         page_table = pte_offset_map_lock(mm, pmd, address,
2292                                                          &ptl);
2293                         if (!pte_same(*page_table, orig_pte)) {
2294                                 unlock_page(old_page);
2295                                 goto unlock;
2296                         }
2297                         page_cache_release(old_page);
2298                 }
2299                 if (reuse_swap_page(old_page)) {
2300                         /*
2301                          * The page is all ours.  Move it to our anon_vma so
2302                          * the rmap code will not search our parent or siblings.
2303                          * Protected against the rmap code by the page lock.
2304                          */
2305                         page_move_anon_rmap(old_page, vma, address);
2306                         unlock_page(old_page);
2307                         goto reuse;
2308                 }
2309                 unlock_page(old_page);
2310         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2311                                         (VM_WRITE|VM_SHARED))) {
2312                 /*
2313                  * Only catch write-faults on shared writable pages,
2314                  * read-only shared pages can get COWed by
2315                  * get_user_pages(.write=1, .force=1).
2316                  */
2317                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2318                         struct vm_fault vmf;
2319                         int tmp;
2320
2321                         vmf.virtual_address = (void __user *)(address &
2322                                                                 PAGE_MASK);
2323                         vmf.pgoff = old_page->index;
2324                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2325                         vmf.page = old_page;
2326
2327                         /*
2328                          * Notify the address space that the page is about to
2329                          * become writable so that it can prohibit this or wait
2330                          * for the page to get into an appropriate state.
2331                          *
2332                          * We do this without the lock held, so that it can
2333                          * sleep if it needs to.
2334                          */
2335                         page_cache_get(old_page);
2336                         pte_unmap_unlock(page_table, ptl);
2337
2338                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2339                         if (unlikely(tmp &
2340                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2341                                 ret = tmp;
2342                                 goto unwritable_page;
2343                         }
2344                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2345                                 lock_page(old_page);
2346                                 if (!old_page->mapping) {
2347                                         ret = 0; /* retry the fault */
2348                                         unlock_page(old_page);
2349                                         goto unwritable_page;
2350                                 }
2351                         } else
2352                                 VM_BUG_ON(!PageLocked(old_page));
2353
2354                         /*
2355                          * Since we dropped the lock we need to revalidate
2356                          * the PTE as someone else may have changed it.  If
2357                          * they did, we just return, as we can count on the
2358                          * MMU to tell us if they didn't also make it writable.
2359                          */
2360                         page_table = pte_offset_map_lock(mm, pmd, address,
2361                                                          &ptl);
2362                         if (!pte_same(*page_table, orig_pte)) {
2363                                 unlock_page(old_page);
2364                                 goto unlock;
2365                         }
2366
2367                         page_mkwrite = 1;
2368                 }
2369                 dirty_page = old_page;
2370                 get_page(dirty_page);
2371
2372 reuse:
2373                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2374                 entry = pte_mkyoung(orig_pte);
2375                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2376                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2377                         update_mmu_cache(vma, address, page_table);
2378                 pte_unmap_unlock(page_table, ptl);
2379                 ret |= VM_FAULT_WRITE;
2380
2381                 if (!dirty_page)
2382                         return ret;
2383
2384                 /*
2385                  * Yes, Virginia, this is actually required to prevent a race
2386                  * with clear_page_dirty_for_io() from clearing the page dirty
2387                  * bit after it clear all dirty ptes, but before a racing
2388                  * do_wp_page installs a dirty pte.
2389                  *
2390                  * __do_fault is protected similarly.
2391                  */
2392                 if (!page_mkwrite) {
2393                         wait_on_page_locked(dirty_page);
2394                         set_page_dirty_balance(dirty_page, page_mkwrite);
2395                 }
2396                 put_page(dirty_page);
2397                 if (page_mkwrite) {
2398                         struct address_space *mapping = dirty_page->mapping;
2399
2400                         set_page_dirty(dirty_page);
2401                         unlock_page(dirty_page);
2402                         page_cache_release(dirty_page);
2403                         if (mapping)    {
2404                                 /*
2405                                  * Some device drivers do not set page.mapping
2406                                  * but still dirty their pages
2407                                  */
2408                                 balance_dirty_pages_ratelimited(mapping);
2409                         }
2410                 }
2411
2412                 /* file_update_time outside page_lock */
2413                 if (vma->vm_file)
2414                         file_update_time(vma->vm_file);
2415
2416                 return ret;
2417         }
2418
2419         /*
2420          * Ok, we need to copy. Oh, well..
2421          */
2422         page_cache_get(old_page);
2423 gotten:
2424         pte_unmap_unlock(page_table, ptl);
2425
2426         if (unlikely(anon_vma_prepare(vma)))
2427                 goto oom;
2428
2429         if (is_zero_pfn(pte_pfn(orig_pte))) {
2430                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2431                 if (!new_page)
2432                         goto oom;
2433         } else {
2434                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2435                 if (!new_page)
2436                         goto oom;
2437                 cow_user_page(new_page, old_page, address, vma);
2438         }
2439         __SetPageUptodate(new_page);
2440
2441         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2442                 goto oom_free_new;
2443
2444         /*
2445          * Re-check the pte - we dropped the lock
2446          */
2447         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2448         if (likely(pte_same(*page_table, orig_pte))) {
2449                 if (old_page) {
2450                         if (!PageAnon(old_page)) {
2451                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2452                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2453                         }
2454                 } else
2455                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2456                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2457                 entry = mk_pte(new_page, vma->vm_page_prot);
2458                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2459                 /*
2460                  * Clear the pte entry and flush it first, before updating the
2461                  * pte with the new entry. This will avoid a race condition
2462                  * seen in the presence of one thread doing SMC and another
2463                  * thread doing COW.
2464                  */
2465                 ptep_clear_flush(vma, address, page_table);
2466                 page_add_new_anon_rmap(new_page, vma, address);
2467                 /*
2468                  * We call the notify macro here because, when using secondary
2469                  * mmu page tables (such as kvm shadow page tables), we want the
2470                  * new page to be mapped directly into the secondary page table.
2471                  */
2472                 set_pte_at_notify(mm, address, page_table, entry);
2473                 update_mmu_cache(vma, address, page_table);
2474                 if (old_page) {
2475                         /*
2476                          * Only after switching the pte to the new page may
2477                          * we remove the mapcount here. Otherwise another
2478                          * process may come and find the rmap count decremented
2479                          * before the pte is switched to the new page, and
2480                          * "reuse" the old page writing into it while our pte
2481                          * here still points into it and can be read by other
2482                          * threads.
2483                          *
2484                          * The critical issue is to order this
2485                          * page_remove_rmap with the ptp_clear_flush above.
2486                          * Those stores are ordered by (if nothing else,)
2487                          * the barrier present in the atomic_add_negative
2488                          * in page_remove_rmap.
2489                          *
2490                          * Then the TLB flush in ptep_clear_flush ensures that
2491                          * no process can access the old page before the
2492                          * decremented mapcount is visible. And the old page
2493                          * cannot be reused until after the decremented
2494                          * mapcount is visible. So transitively, TLBs to
2495                          * old page will be flushed before it can be reused.
2496                          */
2497                         page_remove_rmap(old_page);
2498                 }
2499
2500                 /* Free the old page.. */
2501                 new_page = old_page;
2502                 ret |= VM_FAULT_WRITE;
2503         } else
2504                 mem_cgroup_uncharge_page(new_page);
2505
2506         if (new_page)
2507                 page_cache_release(new_page);
2508 unlock:
2509         pte_unmap_unlock(page_table, ptl);
2510         if (old_page) {
2511                 /*
2512                  * Don't let another task, with possibly unlocked vma,
2513                  * keep the mlocked page.
2514                  */
2515                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2516                         lock_page(old_page);    /* LRU manipulation */
2517                         munlock_vma_page(old_page);
2518                         unlock_page(old_page);
2519                 }
2520                 page_cache_release(old_page);
2521         }
2522         return ret;
2523 oom_free_new:
2524         page_cache_release(new_page);
2525 oom:
2526         if (old_page) {
2527                 if (page_mkwrite) {
2528                         unlock_page(old_page);
2529                         page_cache_release(old_page);
2530                 }
2531                 page_cache_release(old_page);
2532         }
2533         return VM_FAULT_OOM;
2534
2535 unwritable_page:
2536         page_cache_release(old_page);
2537         return ret;
2538 }
2539
2540 /*
2541  * Helper functions for unmap_mapping_range().
2542  *
2543  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2544  *
2545  * We have to restart searching the prio_tree whenever we drop the lock,
2546  * since the iterator is only valid while the lock is held, and anyway
2547  * a later vma might be split and reinserted earlier while lock dropped.
2548  *
2549  * The list of nonlinear vmas could be handled more efficiently, using
2550  * a placeholder, but handle it in the same way until a need is shown.
2551  * It is important to search the prio_tree before nonlinear list: a vma
2552  * may become nonlinear and be shifted from prio_tree to nonlinear list
2553  * while the lock is dropped; but never shifted from list to prio_tree.
2554  *
2555  * In order to make forward progress despite restarting the search,
2556  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2557  * quickly skip it next time around.  Since the prio_tree search only
2558  * shows us those vmas affected by unmapping the range in question, we
2559  * can't efficiently keep all vmas in step with mapping->truncate_count:
2560  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2561  * mapping->truncate_count and vma->vm_truncate_count are protected by
2562  * i_mmap_lock.
2563  *
2564  * In order to make forward progress despite repeatedly restarting some
2565  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2566  * and restart from that address when we reach that vma again.  It might
2567  * have been split or merged, shrunk or extended, but never shifted: so
2568  * restart_addr remains valid so long as it remains in the vma's range.
2569  * unmap_mapping_range forces truncate_count to leap over page-aligned
2570  * values so we can save vma's restart_addr in its truncate_count field.
2571  */
2572 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2573
2574 static void reset_vma_truncate_counts(struct address_space *mapping)
2575 {
2576         struct vm_area_struct *vma;
2577         struct prio_tree_iter iter;
2578
2579         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2580                 vma->vm_truncate_count = 0;
2581         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2582                 vma->vm_truncate_count = 0;
2583 }
2584
2585 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2586                 unsigned long start_addr, unsigned long end_addr,
2587                 struct zap_details *details)
2588 {
2589         unsigned long restart_addr;
2590         int need_break;
2591
2592         /*
2593          * files that support invalidating or truncating portions of the
2594          * file from under mmaped areas must have their ->fault function
2595          * return a locked page (and set VM_FAULT_LOCKED in the return).
2596          * This provides synchronisation against concurrent unmapping here.
2597          */
2598
2599 again:
2600         restart_addr = vma->vm_truncate_count;
2601         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2602                 start_addr = restart_addr;
2603                 if (start_addr >= end_addr) {
2604                         /* Top of vma has been split off since last time */
2605                         vma->vm_truncate_count = details->truncate_count;
2606                         return 0;
2607                 }
2608         }
2609
2610         restart_addr = zap_page_range(vma, start_addr,
2611                                         end_addr - start_addr, details);
2612         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2613
2614         if (restart_addr >= end_addr) {
2615                 /* We have now completed this vma: mark it so */
2616                 vma->vm_truncate_count = details->truncate_count;
2617                 if (!need_break)
2618                         return 0;
2619         } else {
2620                 /* Note restart_addr in vma's truncate_count field */
2621                 vma->vm_truncate_count = restart_addr;
2622                 if (!need_break)
2623                         goto again;
2624         }
2625
2626         spin_unlock(details->i_mmap_lock);
2627         cond_resched();
2628         spin_lock(details->i_mmap_lock);
2629         return -EINTR;
2630 }
2631
2632 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2633                                             struct zap_details *details)
2634 {
2635         struct vm_area_struct *vma;
2636         struct prio_tree_iter iter;
2637         pgoff_t vba, vea, zba, zea;
2638
2639 restart:
2640         vma_prio_tree_foreach(vma, &iter, root,
2641                         details->first_index, details->last_index) {
2642                 /* Skip quickly over those we have already dealt with */
2643                 if (vma->vm_truncate_count == details->truncate_count)
2644                         continue;
2645
2646                 vba = vma->vm_pgoff;
2647                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2648                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2649                 zba = details->first_index;
2650                 if (zba < vba)
2651                         zba = vba;
2652                 zea = details->last_index;
2653                 if (zea > vea)
2654                         zea = vea;
2655
2656                 if (unmap_mapping_range_vma(vma,
2657                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2658                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2659                                 details) < 0)
2660                         goto restart;
2661         }
2662 }
2663
2664 static inline void unmap_mapping_range_list(struct list_head *head,
2665                                             struct zap_details *details)
2666 {
2667         struct vm_area_struct *vma;
2668
2669         /*
2670          * In nonlinear VMAs there is no correspondence between virtual address
2671          * offset and file offset.  So we must perform an exhaustive search
2672          * across *all* the pages in each nonlinear VMA, not just the pages
2673          * whose virtual address lies outside the file truncation point.
2674          */
2675 restart:
2676         list_for_each_entry(vma, head, shared.vm_set.list) {
2677                 /* Skip quickly over those we have already dealt with */
2678                 if (vma->vm_truncate_count == details->truncate_count)
2679                         continue;
2680                 details->nonlinear_vma = vma;
2681                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2682                                         vma->vm_end, details) < 0)
2683                         goto restart;
2684         }
2685 }
2686
2687 /**
2688  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2689  * @mapping: the address space containing mmaps to be unmapped.
2690  * @holebegin: byte in first page to unmap, relative to the start of
2691  * the underlying file.  This will be rounded down to a PAGE_SIZE
2692  * boundary.  Note that this is different from truncate_pagecache(), which
2693  * must keep the partial page.  In contrast, we must get rid of
2694  * partial pages.
2695  * @holelen: size of prospective hole in bytes.  This will be rounded
2696  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2697  * end of the file.
2698  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2699  * but 0 when invalidating pagecache, don't throw away private data.
2700  */
2701 void unmap_mapping_range(struct address_space *mapping,
2702                 loff_t const holebegin, loff_t const holelen, int even_cows)
2703 {
2704         struct zap_details details;
2705         pgoff_t hba = holebegin >> PAGE_SHIFT;
2706         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2707
2708         /* Check for overflow. */
2709         if (sizeof(holelen) > sizeof(hlen)) {
2710                 long long holeend =
2711                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2712                 if (holeend & ~(long long)ULONG_MAX)
2713                         hlen = ULONG_MAX - hba + 1;
2714         }
2715
2716         details.check_mapping = even_cows? NULL: mapping;
2717         details.nonlinear_vma = NULL;
2718         details.first_index = hba;
2719         details.last_index = hba + hlen - 1;
2720         if (details.last_index < details.first_index)
2721                 details.last_index = ULONG_MAX;
2722         details.i_mmap_lock = &mapping->i_mmap_lock;
2723
2724         mutex_lock(&mapping->unmap_mutex);
2725         spin_lock(&mapping->i_mmap_lock);
2726
2727         /* Protect against endless unmapping loops */
2728         mapping->truncate_count++;
2729         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2730                 if (mapping->truncate_count == 0)
2731                         reset_vma_truncate_counts(mapping);
2732                 mapping->truncate_count++;
2733         }
2734         details.truncate_count = mapping->truncate_count;
2735
2736         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2737                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2738         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2739                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2740         spin_unlock(&mapping->i_mmap_lock);
2741         mutex_unlock(&mapping->unmap_mutex);
2742 }
2743 EXPORT_SYMBOL(unmap_mapping_range);
2744
2745 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2746 {
2747         struct address_space *mapping = inode->i_mapping;
2748
2749         /*
2750          * If the underlying filesystem is not going to provide
2751          * a way to truncate a range of blocks (punch a hole) -
2752          * we should return failure right now.
2753          */
2754         if (!inode->i_op->truncate_range)
2755                 return -ENOSYS;
2756
2757         mutex_lock(&inode->i_mutex);
2758         down_write(&inode->i_alloc_sem);
2759         unmap_mapping_range(mapping, offset, (end - offset), 1);
2760         truncate_inode_pages_range(mapping, offset, end);
2761         unmap_mapping_range(mapping, offset, (end - offset), 1);
2762         inode->i_op->truncate_range(inode, offset, end);
2763         up_write(&inode->i_alloc_sem);
2764         mutex_unlock(&inode->i_mutex);
2765
2766         return 0;
2767 }
2768
2769 /*
2770  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2771  * but allow concurrent faults), and pte mapped but not yet locked.
2772  * We return with mmap_sem still held, but pte unmapped and unlocked.
2773  */
2774 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2775                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2776                 unsigned int flags, pte_t orig_pte)
2777 {
2778         spinlock_t *ptl;
2779         struct page *page, *swapcache = NULL;
2780         swp_entry_t entry;
2781         pte_t pte;
2782         int locked;
2783         struct mem_cgroup *ptr;
2784         int exclusive = 0;
2785         int ret = 0;
2786
2787         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2788                 goto out;
2789
2790         entry = pte_to_swp_entry(orig_pte);
2791         if (unlikely(non_swap_entry(entry))) {
2792                 if (is_migration_entry(entry)) {
2793                         migration_entry_wait(mm, pmd, address);
2794                 } else if (is_hwpoison_entry(entry)) {
2795                         ret = VM_FAULT_HWPOISON;
2796                 } else {
2797                         print_bad_pte(vma, address, orig_pte, NULL);
2798                         ret = VM_FAULT_SIGBUS;
2799                 }
2800                 goto out;
2801         }
2802         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2803         page = lookup_swap_cache(entry);
2804         if (!page) {
2805                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2806                 page = swapin_readahead(entry,
2807                                         GFP_HIGHUSER_MOVABLE, vma, address);
2808                 if (!page) {
2809                         /*
2810                          * Back out if somebody else faulted in this pte
2811                          * while we released the pte lock.
2812                          */
2813                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2814                         if (likely(pte_same(*page_table, orig_pte)))
2815                                 ret = VM_FAULT_OOM;
2816                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2817                         goto unlock;
2818                 }
2819
2820                 /* Had to read the page from swap area: Major fault */
2821                 ret = VM_FAULT_MAJOR;
2822                 count_vm_event(PGMAJFAULT);
2823         } else if (PageHWPoison(page)) {
2824                 /*
2825                  * hwpoisoned dirty swapcache pages are kept for killing
2826                  * owner processes (which may be unknown at hwpoison time)
2827                  */
2828                 ret = VM_FAULT_HWPOISON;
2829                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2830                 goto out_release;
2831         }
2832
2833         locked = lock_page_or_retry(page, mm, flags);
2834         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2835         if (!locked) {
2836                 ret |= VM_FAULT_RETRY;
2837                 goto out_release;
2838         }
2839
2840         /*
2841          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2842          * release the swapcache from under us.  The page pin, and pte_same
2843          * test below, are not enough to exclude that.  Even if it is still
2844          * swapcache, we need to check that the page's swap has not changed.
2845          */
2846         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2847                 goto out_page;
2848
2849         if (ksm_might_need_to_copy(page, vma, address)) {
2850                 swapcache = page;
2851                 page = ksm_does_need_to_copy(page, vma, address);
2852
2853                 if (unlikely(!page)) {
2854                         ret = VM_FAULT_OOM;
2855                         page = swapcache;
2856                         swapcache = NULL;
2857                         goto out_page;
2858                 }
2859         }
2860
2861         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2862                 ret = VM_FAULT_OOM;
2863                 goto out_page;
2864         }
2865
2866         /*
2867          * Back out if somebody else already faulted in this pte.
2868          */
2869         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2870         if (unlikely(!pte_same(*page_table, orig_pte)))
2871                 goto out_nomap;
2872
2873         if (unlikely(!PageUptodate(page))) {
2874                 ret = VM_FAULT_SIGBUS;
2875                 goto out_nomap;
2876         }
2877
2878         /*
2879          * The page isn't present yet, go ahead with the fault.
2880          *
2881          * Be careful about the sequence of operations here.
2882          * To get its accounting right, reuse_swap_page() must be called
2883          * while the page is counted on swap but not yet in mapcount i.e.
2884          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2885          * must be called after the swap_free(), or it will never succeed.
2886          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2887          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2888          * in page->private. In this case, a record in swap_cgroup  is silently
2889          * discarded at swap_free().
2890          */
2891
2892         inc_mm_counter_fast(mm, MM_ANONPAGES);
2893         dec_mm_counter_fast(mm, MM_SWAPENTS);
2894         pte = mk_pte(page, vma->vm_page_prot);
2895         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2896                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2897                 flags &= ~FAULT_FLAG_WRITE;
2898                 ret |= VM_FAULT_WRITE;
2899                 exclusive = 1;
2900         }
2901         flush_icache_page(vma, page);
2902         set_pte_at(mm, address, page_table, pte);
2903         do_page_add_anon_rmap(page, vma, address, exclusive);
2904         /* It's better to call commit-charge after rmap is established */
2905         mem_cgroup_commit_charge_swapin(page, ptr);
2906
2907         swap_free(entry);
2908         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2909                 try_to_free_swap(page);
2910         unlock_page(page);
2911         if (swapcache) {
2912                 /*
2913                  * Hold the lock to avoid the swap entry to be reused
2914                  * until we take the PT lock for the pte_same() check
2915                  * (to avoid false positives from pte_same). For
2916                  * further safety release the lock after the swap_free
2917                  * so that the swap count won't change under a
2918                  * parallel locked swapcache.
2919                  */
2920                 unlock_page(swapcache);
2921                 page_cache_release(swapcache);
2922         }
2923
2924         if (flags & FAULT_FLAG_WRITE) {
2925                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2926                 if (ret & VM_FAULT_ERROR)
2927                         ret &= VM_FAULT_ERROR;
2928                 goto out;
2929         }
2930
2931         /* No need to invalidate - it was non-present before */
2932         update_mmu_cache(vma, address, page_table);
2933 unlock:
2934         pte_unmap_unlock(page_table, ptl);
2935 out:
2936         return ret;
2937 out_nomap:
2938         mem_cgroup_cancel_charge_swapin(ptr);
2939         pte_unmap_unlock(page_table, ptl);
2940 out_page:
2941         unlock_page(page);
2942 out_release:
2943         page_cache_release(page);
2944         if (swapcache) {
2945                 unlock_page(swapcache);
2946                 page_cache_release(swapcache);
2947         }
2948         return ret;
2949 }
2950
2951 /*
2952  * This is like a special single-page "expand_{down|up}wards()",
2953  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2954  * doesn't hit another vma.
2955  */
2956 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2957 {
2958         address &= PAGE_MASK;
2959         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2960                 struct vm_area_struct *prev = vma->vm_prev;
2961
2962                 /*
2963                  * Is there a mapping abutting this one below?
2964                  *
2965                  * That's only ok if it's the same stack mapping
2966                  * that has gotten split..
2967                  */
2968                 if (prev && prev->vm_end == address)
2969                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2970
2971                 expand_stack(vma, address - PAGE_SIZE);
2972         }
2973         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2974                 struct vm_area_struct *next = vma->vm_next;
2975
2976                 /* As VM_GROWSDOWN but s/below/above/ */
2977                 if (next && next->vm_start == address + PAGE_SIZE)
2978                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2979
2980                 expand_upwards(vma, address + PAGE_SIZE);
2981         }
2982         return 0;
2983 }
2984
2985 /*
2986  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2987  * but allow concurrent faults), and pte mapped but not yet locked.
2988  * We return with mmap_sem still held, but pte unmapped and unlocked.
2989  */
2990 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2991                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2992                 unsigned int flags)
2993 {
2994         struct page *page;
2995         spinlock_t *ptl;
2996         pte_t entry;
2997
2998         pte_unmap(page_table);
2999
3000         /* Check if we need to add a guard page to the stack */
3001         if (check_stack_guard_page(vma, address) < 0)
3002                 return VM_FAULT_SIGBUS;
3003
3004         /* Use the zero-page for reads */
3005         if (!(flags & FAULT_FLAG_WRITE)) {
3006                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3007                                                 vma->vm_page_prot));
3008                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3009                 if (!pte_none(*page_table))
3010                         goto unlock;
3011                 goto setpte;
3012         }
3013
3014         /* Allocate our own private page. */
3015         if (unlikely(anon_vma_prepare(vma)))
3016                 goto oom;
3017         page = alloc_zeroed_user_highpage_movable(vma, address);
3018         if (!page)
3019                 goto oom;
3020         __SetPageUptodate(page);
3021
3022         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3023                 goto oom_free_page;
3024
3025         entry = mk_pte(page, vma->vm_page_prot);
3026         if (vma->vm_flags & VM_WRITE)
3027                 entry = pte_mkwrite(pte_mkdirty(entry));
3028
3029         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3030         if (!pte_none(*page_table))
3031                 goto release;
3032
3033         inc_mm_counter_fast(mm, MM_ANONPAGES);
3034         page_add_new_anon_rmap(page, vma, address);
3035 setpte:
3036         set_pte_at(mm, address, page_table, entry);
3037
3038         /* No need to invalidate - it was non-present before */
3039         update_mmu_cache(vma, address, page_table);
3040 unlock:
3041         pte_unmap_unlock(page_table, ptl);
3042         return 0;
3043 release:
3044         mem_cgroup_uncharge_page(page);
3045         page_cache_release(page);
3046         goto unlock;
3047 oom_free_page:
3048         page_cache_release(page);
3049 oom:
3050         return VM_FAULT_OOM;
3051 }
3052
3053 /*
3054  * __do_fault() tries to create a new page mapping. It aggressively
3055  * tries to share with existing pages, but makes a separate copy if
3056  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3057  * the next page fault.
3058  *
3059  * As this is called only for pages that do not currently exist, we
3060  * do not need to flush old virtual caches or the TLB.
3061  *
3062  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3063  * but allow concurrent faults), and pte neither mapped nor locked.
3064  * We return with mmap_sem still held, but pte unmapped and unlocked.
3065  */
3066 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3067                 unsigned long address, pmd_t *pmd,
3068                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3069 {
3070         pte_t *page_table;
3071         spinlock_t *ptl;
3072         struct page *page;
3073         pte_t entry;
3074         int anon = 0;
3075         int charged = 0;
3076         struct page *dirty_page = NULL;
3077         struct vm_fault vmf;
3078         int ret;
3079         int page_mkwrite = 0;
3080
3081         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3082         vmf.pgoff = pgoff;
3083         vmf.flags = flags;
3084         vmf.page = NULL;
3085
3086         ret = vma->vm_ops->fault(vma, &vmf);
3087         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3088                             VM_FAULT_RETRY)))
3089                 return ret;
3090
3091         if (unlikely(PageHWPoison(vmf.page))) {
3092                 if (ret & VM_FAULT_LOCKED)
3093                         unlock_page(vmf.page);
3094                 return VM_FAULT_HWPOISON;
3095         }
3096
3097         /*
3098          * For consistency in subsequent calls, make the faulted page always
3099          * locked.
3100          */
3101         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3102                 lock_page(vmf.page);
3103         else
3104                 VM_BUG_ON(!PageLocked(vmf.page));
3105
3106         /*
3107          * Should we do an early C-O-W break?
3108          */
3109         page = vmf.page;
3110         if (flags & FAULT_FLAG_WRITE) {
3111                 if (!(vma->vm_flags & VM_SHARED)) {
3112                         anon = 1;
3113                         if (unlikely(anon_vma_prepare(vma))) {
3114                                 ret = VM_FAULT_OOM;
3115                                 goto out;
3116                         }
3117                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3118                                                 vma, address);
3119                         if (!page) {
3120                                 ret = VM_FAULT_OOM;
3121                                 goto out;
3122                         }
3123                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3124                                 ret = VM_FAULT_OOM;
3125                                 page_cache_release(page);
3126                                 goto out;
3127                         }
3128                         charged = 1;
3129                         copy_user_highpage(page, vmf.page, address, vma);
3130                         __SetPageUptodate(page);
3131                 } else {
3132                         /*
3133                          * If the page will be shareable, see if the backing
3134                          * address space wants to know that the page is about
3135                          * to become writable
3136                          */
3137                         if (vma->vm_ops->page_mkwrite) {
3138                                 int tmp;
3139
3140                                 unlock_page(page);
3141                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3142                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3143                                 if (unlikely(tmp &
3144                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3145                                         ret = tmp;
3146                                         goto unwritable_page;
3147                                 }
3148                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3149                                         lock_page(page);
3150                                         if (!page->mapping) {
3151                                                 ret = 0; /* retry the fault */
3152                                                 unlock_page(page);
3153                                                 goto unwritable_page;
3154                                         }
3155                                 } else
3156                                         VM_BUG_ON(!PageLocked(page));
3157                                 page_mkwrite = 1;
3158                         }
3159                 }
3160
3161         }
3162
3163         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3164
3165         /*
3166          * This silly early PAGE_DIRTY setting removes a race
3167          * due to the bad i386 page protection. But it's valid
3168          * for other architectures too.
3169          *
3170          * Note that if FAULT_FLAG_WRITE is set, we either now have
3171          * an exclusive copy of the page, or this is a shared mapping,
3172          * so we can make it writable and dirty to avoid having to
3173          * handle that later.
3174          */
3175         /* Only go through if we didn't race with anybody else... */
3176         if (likely(pte_same(*page_table, orig_pte))) {
3177                 flush_icache_page(vma, page);
3178                 entry = mk_pte(page, vma->vm_page_prot);
3179                 if (flags & FAULT_FLAG_WRITE)
3180                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3181                 if (anon) {
3182                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3183                         page_add_new_anon_rmap(page, vma, address);
3184                 } else {
3185                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3186                         page_add_file_rmap(page);
3187                         if (flags & FAULT_FLAG_WRITE) {
3188                                 dirty_page = page;
3189                                 get_page(dirty_page);
3190                         }
3191                 }
3192                 set_pte_at(mm, address, page_table, entry);
3193
3194                 /* no need to invalidate: a not-present page won't be cached */
3195                 update_mmu_cache(vma, address, page_table);
3196         } else {
3197                 if (charged)
3198                         mem_cgroup_uncharge_page(page);
3199                 if (anon)
3200                         page_cache_release(page);
3201                 else
3202                         anon = 1; /* no anon but release faulted_page */
3203         }
3204
3205         pte_unmap_unlock(page_table, ptl);
3206
3207 out:
3208         if (dirty_page) {
3209                 struct address_space *mapping = page->mapping;
3210
3211                 if (set_page_dirty(dirty_page))
3212                         page_mkwrite = 1;
3213                 unlock_page(dirty_page);
3214                 put_page(dirty_page);
3215                 if (page_mkwrite && mapping) {
3216                         /*
3217                          * Some device drivers do not set page.mapping but still
3218                          * dirty their pages
3219                          */
3220                         balance_dirty_pages_ratelimited(mapping);
3221                 }
3222
3223                 /* file_update_time outside page_lock */
3224                 if (vma->vm_file)
3225                         file_update_time(vma->vm_file);
3226         } else {
3227                 unlock_page(vmf.page);
3228                 if (anon)
3229                         page_cache_release(vmf.page);
3230         }
3231
3232         return ret;
3233
3234 unwritable_page:
3235         page_cache_release(page);
3236         return ret;
3237 }
3238
3239 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3240                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3241                 unsigned int flags, pte_t orig_pte)
3242 {
3243         pgoff_t pgoff = (((address & PAGE_MASK)
3244                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3245
3246         pte_unmap(page_table);
3247         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3248 }
3249
3250 /*
3251  * Fault of a previously existing named mapping. Repopulate the pte
3252  * from the encoded file_pte if possible. This enables swappable
3253  * nonlinear vmas.
3254  *
3255  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3256  * but allow concurrent faults), and pte mapped but not yet locked.
3257  * We return with mmap_sem still held, but pte unmapped and unlocked.
3258  */
3259 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3260                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3261                 unsigned int flags, pte_t orig_pte)
3262 {
3263         pgoff_t pgoff;
3264
3265         flags |= FAULT_FLAG_NONLINEAR;
3266
3267         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3268                 return 0;
3269
3270         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3271                 /*
3272                  * Page table corrupted: show pte and kill process.
3273                  */
3274                 print_bad_pte(vma, address, orig_pte, NULL);
3275                 return VM_FAULT_SIGBUS;
3276         }
3277
3278         pgoff = pte_to_pgoff(orig_pte);
3279         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3280 }
3281
3282 /*
3283  * These routines also need to handle stuff like marking pages dirty
3284  * and/or accessed for architectures that don't do it in hardware (most
3285  * RISC architectures).  The early dirtying is also good on the i386.
3286  *
3287  * There is also a hook called "update_mmu_cache()" that architectures
3288  * with external mmu caches can use to update those (ie the Sparc or
3289  * PowerPC hashed page tables that act as extended TLBs).
3290  *
3291  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3292  * but allow concurrent faults), and pte mapped but not yet locked.
3293  * We return with mmap_sem still held, but pte unmapped and unlocked.
3294  */
3295 int handle_pte_fault(struct mm_struct *mm,
3296                      struct vm_area_struct *vma, unsigned long address,
3297                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3298 {
3299         pte_t entry;
3300         spinlock_t *ptl;
3301
3302         entry = *pte;
3303         if (!pte_present(entry)) {
3304                 if (pte_none(entry)) {
3305                         if (vma->vm_ops) {
3306                                 if (likely(vma->vm_ops->fault))
3307                                         return do_linear_fault(mm, vma, address,
3308                                                 pte, pmd, flags, entry);
3309                         }
3310                         return do_anonymous_page(mm, vma, address,
3311                                                  pte, pmd, flags);
3312                 }
3313                 if (pte_file(entry))
3314                         return do_nonlinear_fault(mm, vma, address,
3315                                         pte, pmd, flags, entry);
3316                 return do_swap_page(mm, vma, address,
3317                                         pte, pmd, flags, entry);
3318         }
3319
3320         ptl = pte_lockptr(mm, pmd);
3321         spin_lock(ptl);
3322         if (unlikely(!pte_same(*pte, entry)))
3323                 goto unlock;
3324         if (flags & FAULT_FLAG_WRITE) {
3325                 if (!pte_write(entry))
3326                         return do_wp_page(mm, vma, address,
3327                                         pte, pmd, ptl, entry);
3328                 entry = pte_mkdirty(entry);
3329         }
3330         entry = pte_mkyoung(entry);
3331         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3332                 update_mmu_cache(vma, address, pte);
3333         } else {
3334                 /*
3335                  * This is needed only for protection faults but the arch code
3336                  * is not yet telling us if this is a protection fault or not.
3337                  * This still avoids useless tlb flushes for .text page faults
3338                  * with threads.
3339                  */
3340                 if (flags & FAULT_FLAG_WRITE)
3341                         flush_tlb_fix_spurious_fault(vma, address);
3342         }
3343 unlock:
3344         pte_unmap_unlock(pte, ptl);
3345         return 0;
3346 }
3347
3348 /*
3349  * By the time we get here, we already hold the mm semaphore
3350  */
3351 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3352                 unsigned long address, unsigned int flags)
3353 {
3354         pgd_t *pgd;
3355         pud_t *pud;
3356         pmd_t *pmd;
3357         pte_t *pte;
3358
3359         __set_current_state(TASK_RUNNING);
3360
3361         count_vm_event(PGFAULT);
3362
3363         /* do counter updates before entering really critical section. */
3364         check_sync_rss_stat(current);
3365
3366         if (unlikely(is_vm_hugetlb_page(vma)))
3367                 return hugetlb_fault(mm, vma, address, flags);
3368
3369         pgd = pgd_offset(mm, address);
3370         pud = pud_alloc(mm, pgd, address);
3371         if (!pud)
3372                 return VM_FAULT_OOM;
3373         pmd = pmd_alloc(mm, pud, address);
3374         if (!pmd)
3375                 return VM_FAULT_OOM;
3376         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3377                 if (!vma->vm_ops)
3378                         return do_huge_pmd_anonymous_page(mm, vma, address,
3379                                                           pmd, flags);
3380         } else {
3381                 pmd_t orig_pmd = *pmd;
3382                 barrier();
3383                 if (pmd_trans_huge(orig_pmd)) {
3384                         if (flags & FAULT_FLAG_WRITE &&
3385                             !pmd_write(orig_pmd) &&
3386                             !pmd_trans_splitting(orig_pmd))
3387                                 return do_huge_pmd_wp_page(mm, vma, address,
3388                                                            pmd, orig_pmd);
3389                         return 0;
3390                 }
3391         }
3392
3393         /*
3394          * Use __pte_alloc instead of pte_alloc_map, because we can't
3395          * run pte_offset_map on the pmd, if an huge pmd could
3396          * materialize from under us from a different thread.
3397          */
3398         if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3399                 return VM_FAULT_OOM;
3400         /* if an huge pmd materialized from under us just retry later */
3401         if (unlikely(pmd_trans_huge(*pmd)))
3402                 return 0;
3403         /*
3404          * A regular pmd is established and it can't morph into a huge pmd
3405          * from under us anymore at this point because we hold the mmap_sem
3406          * read mode and khugepaged takes it in write mode. So now it's
3407          * safe to run pte_offset_map().
3408          */
3409         pte = pte_offset_map(pmd, address);
3410
3411         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3412 }
3413
3414 #ifndef __PAGETABLE_PUD_FOLDED
3415 /*
3416  * Allocate page upper directory.
3417  * We've already handled the fast-path in-line.
3418  */
3419 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3420 {
3421         pud_t *new = pud_alloc_one(mm, address);
3422         if (!new)
3423                 return -ENOMEM;
3424
3425         smp_wmb(); /* See comment in __pte_alloc */
3426
3427         spin_lock(&mm->page_table_lock);
3428         if (pgd_present(*pgd))          /* Another has populated it */
3429                 pud_free(mm, new);
3430         else
3431                 pgd_populate(mm, pgd, new);
3432         spin_unlock(&mm->page_table_lock);
3433         return 0;
3434 }
3435 #endif /* __PAGETABLE_PUD_FOLDED */
3436
3437 #ifndef __PAGETABLE_PMD_FOLDED
3438 /*
3439  * Allocate page middle directory.
3440  * We've already handled the fast-path in-line.
3441  */
3442 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3443 {
3444         pmd_t *new = pmd_alloc_one(mm, address);
3445         if (!new)
3446                 return -ENOMEM;
3447
3448         smp_wmb(); /* See comment in __pte_alloc */
3449
3450         spin_lock(&mm->page_table_lock);
3451 #ifndef __ARCH_HAS_4LEVEL_HACK
3452         if (pud_present(*pud))          /* Another has populated it */
3453                 pmd_free(mm, new);
3454         else
3455                 pud_populate(mm, pud, new);
3456 #else
3457         if (pgd_present(*pud))          /* Another has populated it */
3458                 pmd_free(mm, new);
3459         else
3460                 pgd_populate(mm, pud, new);
3461 #endif /* __ARCH_HAS_4LEVEL_HACK */
3462         spin_unlock(&mm->page_table_lock);
3463         return 0;
3464 }
3465 #endif /* __PAGETABLE_PMD_FOLDED */
3466
3467 int make_pages_present(unsigned long addr, unsigned long end)
3468 {
3469         int ret, len, write;
3470         struct vm_area_struct * vma;
3471
3472         vma = find_vma(current->mm, addr);
3473         if (!vma)
3474                 return -ENOMEM;
3475         /*
3476          * We want to touch writable mappings with a write fault in order
3477          * to break COW, except for shared mappings because these don't COW
3478          * and we would not want to dirty them for nothing.
3479          */
3480         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3481         BUG_ON(addr >= end);
3482         BUG_ON(end > vma->vm_end);
3483         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3484         ret = get_user_pages(current, current->mm, addr,
3485                         len, write, 0, NULL, NULL);
3486         if (ret < 0)
3487                 return ret;
3488         return ret == len ? 0 : -EFAULT;
3489 }
3490
3491 #if !defined(__HAVE_ARCH_GATE_AREA)
3492
3493 #if defined(AT_SYSINFO_EHDR)
3494 static struct vm_area_struct gate_vma;
3495
3496 static int __init gate_vma_init(void)
3497 {
3498         gate_vma.vm_mm = NULL;
3499         gate_vma.vm_start = FIXADDR_USER_START;
3500         gate_vma.vm_end = FIXADDR_USER_END;
3501         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3502         gate_vma.vm_page_prot = __P101;
3503         /*
3504          * Make sure the vDSO gets into every core dump.
3505          * Dumping its contents makes post-mortem fully interpretable later
3506          * without matching up the same kernel and hardware config to see
3507          * what PC values meant.
3508          */
3509         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3510         return 0;
3511 }
3512 __initcall(gate_vma_init);
3513 #endif
3514
3515 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3516 {
3517 #ifdef AT_SYSINFO_EHDR
3518         return &gate_vma;
3519 #else
3520         return NULL;
3521 #endif
3522 }
3523
3524 int in_gate_area_no_mm(unsigned long addr)
3525 {
3526 #ifdef AT_SYSINFO_EHDR
3527         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3528                 return 1;
3529 #endif
3530         return 0;
3531 }
3532
3533 #endif  /* __HAVE_ARCH_GATE_AREA */
3534
3535 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3536                 pte_t **ptepp, spinlock_t **ptlp)
3537 {
3538         pgd_t *pgd;
3539         pud_t *pud;
3540         pmd_t *pmd;
3541         pte_t *ptep;
3542
3543         pgd = pgd_offset(mm, address);
3544         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3545                 goto out;
3546
3547         pud = pud_offset(pgd, address);
3548         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3549                 goto out;
3550
3551         pmd = pmd_offset(pud, address);
3552         VM_BUG_ON(pmd_trans_huge(*pmd));
3553         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3554                 goto out;
3555
3556         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3557         if (pmd_huge(*pmd))
3558                 goto out;
3559
3560         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3561         if (!ptep)
3562                 goto out;
3563         if (!pte_present(*ptep))
3564                 goto unlock;
3565         *ptepp = ptep;
3566         return 0;
3567 unlock:
3568         pte_unmap_unlock(ptep, *ptlp);
3569 out:
3570         return -EINVAL;
3571 }
3572
3573 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3574                              pte_t **ptepp, spinlock_t **ptlp)
3575 {
3576         int res;
3577
3578         /* (void) is needed to make gcc happy */
3579         (void) __cond_lock(*ptlp,
3580                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3581         return res;
3582 }
3583
3584 /**
3585  * follow_pfn - look up PFN at a user virtual address
3586  * @vma: memory mapping
3587  * @address: user virtual address
3588  * @pfn: location to store found PFN
3589  *
3590  * Only IO mappings and raw PFN mappings are allowed.
3591  *
3592  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3593  */
3594 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3595         unsigned long *pfn)
3596 {
3597         int ret = -EINVAL;
3598         spinlock_t *ptl;
3599         pte_t *ptep;
3600
3601         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3602                 return ret;
3603
3604         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3605         if (ret)
3606                 return ret;
3607         *pfn = pte_pfn(*ptep);
3608         pte_unmap_unlock(ptep, ptl);
3609         return 0;
3610 }
3611 EXPORT_SYMBOL(follow_pfn);
3612
3613 #ifdef CONFIG_HAVE_IOREMAP_PROT
3614 int follow_phys(struct vm_area_struct *vma,
3615                 unsigned long address, unsigned int flags,
3616                 unsigned long *prot, resource_size_t *phys)
3617 {
3618         int ret = -EINVAL;
3619         pte_t *ptep, pte;
3620         spinlock_t *ptl;
3621
3622         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3623                 goto out;
3624
3625         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3626                 goto out;
3627         pte = *ptep;
3628
3629         if ((flags & FOLL_WRITE) && !pte_write(pte))
3630                 goto unlock;
3631
3632         *prot = pgprot_val(pte_pgprot(pte));
3633         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3634
3635         ret = 0;
3636 unlock:
3637         pte_unmap_unlock(ptep, ptl);
3638 out:
3639         return ret;
3640 }
3641
3642 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3643                         void *buf, int len, int write)
3644 {
3645         resource_size_t phys_addr;
3646         unsigned long prot = 0;
3647         void __iomem *maddr;
3648         int offset = addr & (PAGE_SIZE-1);
3649
3650         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3651                 return -EINVAL;
3652
3653         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3654         if (write)
3655                 memcpy_toio(maddr + offset, buf, len);
3656         else
3657                 memcpy_fromio(buf, maddr + offset, len);
3658         iounmap(maddr);
3659
3660         return len;
3661 }
3662 #endif
3663
3664 /*
3665  * Access another process' address space as given in mm.  If non-NULL, use the
3666  * given task for page fault accounting.
3667  */
3668 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3669                 unsigned long addr, void *buf, int len, int write)
3670 {
3671         struct vm_area_struct *vma;
3672         void *old_buf = buf;
3673
3674         down_read(&mm->mmap_sem);
3675         /* ignore errors, just check how much was successfully transferred */
3676         while (len) {
3677                 int bytes, ret, offset;
3678                 void *maddr;
3679                 struct page *page = NULL;
3680
3681                 ret = get_user_pages(tsk, mm, addr, 1,
3682                                 write, 1, &page, &vma);
3683                 if (ret <= 0) {
3684                         /*
3685                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3686                          * we can access using slightly different code.
3687                          */
3688 #ifdef CONFIG_HAVE_IOREMAP_PROT
3689                         vma = find_vma(mm, addr);
3690                         if (!vma || vma->vm_start > addr)
3691                                 break;
3692                         if (vma->vm_ops && vma->vm_ops->access)
3693                                 ret = vma->vm_ops->access(vma, addr, buf,
3694                                                           len, write);
3695                         if (ret <= 0)
3696 #endif
3697                                 break;
3698                         bytes = ret;
3699                 } else {
3700                         bytes = len;
3701                         offset = addr & (PAGE_SIZE-1);
3702                         if (bytes > PAGE_SIZE-offset)
3703                                 bytes = PAGE_SIZE-offset;
3704
3705                         maddr = kmap(page);
3706                         if (write) {
3707                                 copy_to_user_page(vma, page, addr,
3708                                                   maddr + offset, buf, bytes);
3709                                 set_page_dirty_lock(page);
3710                         } else {
3711                                 copy_from_user_page(vma, page, addr,
3712                                                     buf, maddr + offset, bytes);
3713                         }
3714                         kunmap(page);
3715                         page_cache_release(page);
3716                 }
3717                 len -= bytes;
3718                 buf += bytes;
3719                 addr += bytes;
3720         }
3721         up_read(&mm->mmap_sem);
3722
3723         return buf - old_buf;
3724 }
3725
3726 /**
3727  * access_remote_vm - access another process' address space
3728  * @mm:         the mm_struct of the target address space
3729  * @addr:       start address to access
3730  * @buf:        source or destination buffer
3731  * @len:        number of bytes to transfer
3732  * @write:      whether the access is a write
3733  *
3734  * The caller must hold a reference on @mm.
3735  */
3736 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3737                 void *buf, int len, int write)
3738 {
3739         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3740 }
3741
3742 /*
3743  * Access another process' address space.
3744  * Source/target buffer must be kernel space,
3745  * Do not walk the page table directly, use get_user_pages
3746  */
3747 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3748                 void *buf, int len, int write)
3749 {
3750         struct mm_struct *mm;
3751         int ret;
3752
3753         mm = get_task_mm(tsk);
3754         if (!mm)
3755                 return 0;
3756
3757         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3758         mmput(mm);
3759
3760         return ret;
3761 }
3762
3763 /*
3764  * Print the name of a VMA.
3765  */
3766 void print_vma_addr(char *prefix, unsigned long ip)
3767 {
3768         struct mm_struct *mm = current->mm;
3769         struct vm_area_struct *vma;
3770
3771         /*
3772          * Do not print if we are in atomic
3773          * contexts (in exception stacks, etc.):
3774          */
3775         if (preempt_count())
3776                 return;
3777
3778         down_read(&mm->mmap_sem);
3779         vma = find_vma(mm, ip);
3780         if (vma && vma->vm_file) {
3781                 struct file *f = vma->vm_file;
3782                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3783                 if (buf) {
3784                         char *p, *s;
3785
3786                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3787                         if (IS_ERR(p))
3788                                 p = "?";
3789                         s = strrchr(p, '/');
3790                         if (s)
3791                                 p = s+1;
3792                         printk("%s%s[%lx+%lx]", prefix, p,
3793                                         vma->vm_start,
3794                                         vma->vm_end - vma->vm_start);
3795                         free_page((unsigned long)buf);
3796                 }
3797         }
3798         up_read(&current->mm->mmap_sem);
3799 }
3800
3801 #ifdef CONFIG_PROVE_LOCKING
3802 void might_fault(void)
3803 {
3804         /*
3805          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3806          * holding the mmap_sem, this is safe because kernel memory doesn't
3807          * get paged out, therefore we'll never actually fault, and the
3808          * below annotations will generate false positives.
3809          */
3810         if (segment_eq(get_fs(), KERNEL_DS))
3811                 return;
3812
3813         might_sleep();
3814         /*
3815          * it would be nicer only to annotate paths which are not under
3816          * pagefault_disable, however that requires a larger audit and
3817          * providing helpers like get_user_atomic.
3818          */
3819         if (!in_atomic() && current->mm)
3820                 might_lock_read(&current->mm->mmap_sem);
3821 }
3822 EXPORT_SYMBOL(might_fault);
3823 #endif
3824
3825 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3826 static void clear_gigantic_page(struct page *page,
3827                                 unsigned long addr,
3828                                 unsigned int pages_per_huge_page)
3829 {
3830         int i;
3831         struct page *p = page;
3832
3833         might_sleep();
3834         for (i = 0; i < pages_per_huge_page;
3835              i++, p = mem_map_next(p, page, i)) {
3836                 cond_resched();
3837                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3838         }
3839 }
3840 void clear_huge_page(struct page *page,
3841                      unsigned long addr, unsigned int pages_per_huge_page)
3842 {
3843         int i;
3844
3845         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3846                 clear_gigantic_page(page, addr, pages_per_huge_page);
3847                 return;
3848         }
3849
3850         might_sleep();
3851         for (i = 0; i < pages_per_huge_page; i++) {
3852                 cond_resched();
3853                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3854         }
3855 }
3856
3857 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3858                                     unsigned long addr,
3859                                     struct vm_area_struct *vma,
3860                                     unsigned int pages_per_huge_page)
3861 {
3862         int i;
3863         struct page *dst_base = dst;
3864         struct page *src_base = src;
3865
3866         for (i = 0; i < pages_per_huge_page; ) {
3867                 cond_resched();
3868                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3869
3870                 i++;
3871                 dst = mem_map_next(dst, dst_base, i);
3872                 src = mem_map_next(src, src_base, i);
3873         }
3874 }
3875
3876 void copy_user_huge_page(struct page *dst, struct page *src,
3877                          unsigned long addr, struct vm_area_struct *vma,
3878                          unsigned int pages_per_huge_page)
3879 {
3880         int i;
3881
3882         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3883                 copy_user_gigantic_page(dst, src, addr, vma,
3884                                         pages_per_huge_page);
3885                 return;
3886         }
3887
3888         might_sleep();
3889         for (i = 0; i < pages_per_huge_page; i++) {
3890                 cond_resched();
3891                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3892         }
3893 }
3894 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */