vm: fix mlock() on stack guard page
[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) {
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                  * If we don't actually want the page itself,
1556                  * and it's the stack guard page, just skip it.
1557                  */
1558                 if (!pages && stack_guard_page(vma, start))
1559                         goto next_page;
1560
1561                 do {
1562                         struct page *page;
1563                         unsigned int foll_flags = gup_flags;
1564
1565                         /*
1566                          * If we have a pending SIGKILL, don't keep faulting
1567                          * pages and potentially allocating memory.
1568                          */
1569                         if (unlikely(fatal_signal_pending(current)))
1570                                 return i ? i : -ERESTARTSYS;
1571
1572                         cond_resched();
1573                         while (!(page = follow_page(vma, start, foll_flags))) {
1574                                 int ret;
1575                                 unsigned int fault_flags = 0;
1576
1577                                 if (foll_flags & FOLL_WRITE)
1578                                         fault_flags |= FAULT_FLAG_WRITE;
1579                                 if (nonblocking)
1580                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1581                                 if (foll_flags & FOLL_NOWAIT)
1582                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1583
1584                                 ret = handle_mm_fault(mm, vma, start,
1585                                                         fault_flags);
1586
1587                                 if (ret & VM_FAULT_ERROR) {
1588                                         if (ret & VM_FAULT_OOM)
1589                                                 return i ? i : -ENOMEM;
1590                                         if (ret & (VM_FAULT_HWPOISON |
1591                                                    VM_FAULT_HWPOISON_LARGE)) {
1592                                                 if (i)
1593                                                         return i;
1594                                                 else if (gup_flags & FOLL_HWPOISON)
1595                                                         return -EHWPOISON;
1596                                                 else
1597                                                         return -EFAULT;
1598                                         }
1599                                         if (ret & VM_FAULT_SIGBUS)
1600                                                 return i ? i : -EFAULT;
1601                                         BUG();
1602                                 }
1603
1604                                 if (tsk) {
1605                                         if (ret & VM_FAULT_MAJOR)
1606                                                 tsk->maj_flt++;
1607                                         else
1608                                                 tsk->min_flt++;
1609                                 }
1610
1611                                 if (ret & VM_FAULT_RETRY) {
1612                                         if (nonblocking)
1613                                                 *nonblocking = 0;
1614                                         return i;
1615                                 }
1616
1617                                 /*
1618                                  * The VM_FAULT_WRITE bit tells us that
1619                                  * do_wp_page has broken COW when necessary,
1620                                  * even if maybe_mkwrite decided not to set
1621                                  * pte_write. We can thus safely do subsequent
1622                                  * page lookups as if they were reads. But only
1623                                  * do so when looping for pte_write is futile:
1624                                  * in some cases userspace may also be wanting
1625                                  * to write to the gotten user page, which a
1626                                  * read fault here might prevent (a readonly
1627                                  * page might get reCOWed by userspace write).
1628                                  */
1629                                 if ((ret & VM_FAULT_WRITE) &&
1630                                     !(vma->vm_flags & VM_WRITE))
1631                                         foll_flags &= ~FOLL_WRITE;
1632
1633                                 cond_resched();
1634                         }
1635                         if (IS_ERR(page))
1636                                 return i ? i : PTR_ERR(page);
1637                         if (pages) {
1638                                 pages[i] = page;
1639
1640                                 flush_anon_page(vma, page, start);
1641                                 flush_dcache_page(page);
1642                         }
1643 next_page:
1644                         if (vmas)
1645                                 vmas[i] = vma;
1646                         i++;
1647                         start += PAGE_SIZE;
1648                         nr_pages--;
1649                 } while (nr_pages && start < vma->vm_end);
1650         } while (nr_pages);
1651         return i;
1652 }
1653 EXPORT_SYMBOL(__get_user_pages);
1654
1655 /**
1656  * get_user_pages() - pin user pages in memory
1657  * @tsk:        the task_struct to use for page fault accounting, or
1658  *              NULL if faults are not to be recorded.
1659  * @mm:         mm_struct of target mm
1660  * @start:      starting user address
1661  * @nr_pages:   number of pages from start to pin
1662  * @write:      whether pages will be written to by the caller
1663  * @force:      whether to force write access even if user mapping is
1664  *              readonly. This will result in the page being COWed even
1665  *              in MAP_SHARED mappings. You do not want this.
1666  * @pages:      array that receives pointers to the pages pinned.
1667  *              Should be at least nr_pages long. Or NULL, if caller
1668  *              only intends to ensure the pages are faulted in.
1669  * @vmas:       array of pointers to vmas corresponding to each page.
1670  *              Or NULL if the caller does not require them.
1671  *
1672  * Returns number of pages pinned. This may be fewer than the number
1673  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1674  * were pinned, returns -errno. Each page returned must be released
1675  * with a put_page() call when it is finished with. vmas will only
1676  * remain valid while mmap_sem is held.
1677  *
1678  * Must be called with mmap_sem held for read or write.
1679  *
1680  * get_user_pages walks a process's page tables and takes a reference to
1681  * each struct page that each user address corresponds to at a given
1682  * instant. That is, it takes the page that would be accessed if a user
1683  * thread accesses the given user virtual address at that instant.
1684  *
1685  * This does not guarantee that the page exists in the user mappings when
1686  * get_user_pages returns, and there may even be a completely different
1687  * page there in some cases (eg. if mmapped pagecache has been invalidated
1688  * and subsequently re faulted). However it does guarantee that the page
1689  * won't be freed completely. And mostly callers simply care that the page
1690  * contains data that was valid *at some point in time*. Typically, an IO
1691  * or similar operation cannot guarantee anything stronger anyway because
1692  * locks can't be held over the syscall boundary.
1693  *
1694  * If write=0, the page must not be written to. If the page is written to,
1695  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1696  * after the page is finished with, and before put_page is called.
1697  *
1698  * get_user_pages is typically used for fewer-copy IO operations, to get a
1699  * handle on the memory by some means other than accesses via the user virtual
1700  * addresses. The pages may be submitted for DMA to devices or accessed via
1701  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1702  * use the correct cache flushing APIs.
1703  *
1704  * See also get_user_pages_fast, for performance critical applications.
1705  */
1706 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1707                 unsigned long start, int nr_pages, int write, int force,
1708                 struct page **pages, struct vm_area_struct **vmas)
1709 {
1710         int flags = FOLL_TOUCH;
1711
1712         if (pages)
1713                 flags |= FOLL_GET;
1714         if (write)
1715                 flags |= FOLL_WRITE;
1716         if (force)
1717                 flags |= FOLL_FORCE;
1718
1719         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1720                                 NULL);
1721 }
1722 EXPORT_SYMBOL(get_user_pages);
1723
1724 /**
1725  * get_dump_page() - pin user page in memory while writing it to core dump
1726  * @addr: user address
1727  *
1728  * Returns struct page pointer of user page pinned for dump,
1729  * to be freed afterwards by page_cache_release() or put_page().
1730  *
1731  * Returns NULL on any kind of failure - a hole must then be inserted into
1732  * the corefile, to preserve alignment with its headers; and also returns
1733  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1734  * allowing a hole to be left in the corefile to save diskspace.
1735  *
1736  * Called without mmap_sem, but after all other threads have been killed.
1737  */
1738 #ifdef CONFIG_ELF_CORE
1739 struct page *get_dump_page(unsigned long addr)
1740 {
1741         struct vm_area_struct *vma;
1742         struct page *page;
1743
1744         if (__get_user_pages(current, current->mm, addr, 1,
1745                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1746                              NULL) < 1)
1747                 return NULL;
1748         flush_cache_page(vma, addr, page_to_pfn(page));
1749         return page;
1750 }
1751 #endif /* CONFIG_ELF_CORE */
1752
1753 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1754                         spinlock_t **ptl)
1755 {
1756         pgd_t * pgd = pgd_offset(mm, addr);
1757         pud_t * pud = pud_alloc(mm, pgd, addr);
1758         if (pud) {
1759                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1760                 if (pmd) {
1761                         VM_BUG_ON(pmd_trans_huge(*pmd));
1762                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1763                 }
1764         }
1765         return NULL;
1766 }
1767
1768 /*
1769  * This is the old fallback for page remapping.
1770  *
1771  * For historical reasons, it only allows reserved pages. Only
1772  * old drivers should use this, and they needed to mark their
1773  * pages reserved for the old functions anyway.
1774  */
1775 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1776                         struct page *page, pgprot_t prot)
1777 {
1778         struct mm_struct *mm = vma->vm_mm;
1779         int retval;
1780         pte_t *pte;
1781         spinlock_t *ptl;
1782
1783         retval = -EINVAL;
1784         if (PageAnon(page))
1785                 goto out;
1786         retval = -ENOMEM;
1787         flush_dcache_page(page);
1788         pte = get_locked_pte(mm, addr, &ptl);
1789         if (!pte)
1790                 goto out;
1791         retval = -EBUSY;
1792         if (!pte_none(*pte))
1793                 goto out_unlock;
1794
1795         /* Ok, finally just insert the thing.. */
1796         get_page(page);
1797         inc_mm_counter_fast(mm, MM_FILEPAGES);
1798         page_add_file_rmap(page);
1799         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1800
1801         retval = 0;
1802         pte_unmap_unlock(pte, ptl);
1803         return retval;
1804 out_unlock:
1805         pte_unmap_unlock(pte, ptl);
1806 out:
1807         return retval;
1808 }
1809
1810 /**
1811  * vm_insert_page - insert single page into user vma
1812  * @vma: user vma to map to
1813  * @addr: target user address of this page
1814  * @page: source kernel page
1815  *
1816  * This allows drivers to insert individual pages they've allocated
1817  * into a user vma.
1818  *
1819  * The page has to be a nice clean _individual_ kernel allocation.
1820  * If you allocate a compound page, you need to have marked it as
1821  * such (__GFP_COMP), or manually just split the page up yourself
1822  * (see split_page()).
1823  *
1824  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1825  * took an arbitrary page protection parameter. This doesn't allow
1826  * that. Your vma protection will have to be set up correctly, which
1827  * means that if you want a shared writable mapping, you'd better
1828  * ask for a shared writable mapping!
1829  *
1830  * The page does not need to be reserved.
1831  */
1832 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1833                         struct page *page)
1834 {
1835         if (addr < vma->vm_start || addr >= vma->vm_end)
1836                 return -EFAULT;
1837         if (!page_count(page))
1838                 return -EINVAL;
1839         vma->vm_flags |= VM_INSERTPAGE;
1840         return insert_page(vma, addr, page, vma->vm_page_prot);
1841 }
1842 EXPORT_SYMBOL(vm_insert_page);
1843
1844 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1845                         unsigned long pfn, pgprot_t prot)
1846 {
1847         struct mm_struct *mm = vma->vm_mm;
1848         int retval;
1849         pte_t *pte, entry;
1850         spinlock_t *ptl;
1851
1852         retval = -ENOMEM;
1853         pte = get_locked_pte(mm, addr, &ptl);
1854         if (!pte)
1855                 goto out;
1856         retval = -EBUSY;
1857         if (!pte_none(*pte))
1858                 goto out_unlock;
1859
1860         /* Ok, finally just insert the thing.. */
1861         entry = pte_mkspecial(pfn_pte(pfn, prot));
1862         set_pte_at(mm, addr, pte, entry);
1863         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1864
1865         retval = 0;
1866 out_unlock:
1867         pte_unmap_unlock(pte, ptl);
1868 out:
1869         return retval;
1870 }
1871
1872 /**
1873  * vm_insert_pfn - insert single pfn into user vma
1874  * @vma: user vma to map to
1875  * @addr: target user address of this page
1876  * @pfn: source kernel pfn
1877  *
1878  * Similar to vm_inert_page, this allows drivers to insert individual pages
1879  * they've allocated into a user vma. Same comments apply.
1880  *
1881  * This function should only be called from a vm_ops->fault handler, and
1882  * in that case the handler should return NULL.
1883  *
1884  * vma cannot be a COW mapping.
1885  *
1886  * As this is called only for pages that do not currently exist, we
1887  * do not need to flush old virtual caches or the TLB.
1888  */
1889 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1890                         unsigned long pfn)
1891 {
1892         int ret;
1893         pgprot_t pgprot = vma->vm_page_prot;
1894         /*
1895          * Technically, architectures with pte_special can avoid all these
1896          * restrictions (same for remap_pfn_range).  However we would like
1897          * consistency in testing and feature parity among all, so we should
1898          * try to keep these invariants in place for everybody.
1899          */
1900         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1901         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1902                                                 (VM_PFNMAP|VM_MIXEDMAP));
1903         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1904         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1905
1906         if (addr < vma->vm_start || addr >= vma->vm_end)
1907                 return -EFAULT;
1908         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1909                 return -EINVAL;
1910
1911         ret = insert_pfn(vma, addr, pfn, pgprot);
1912
1913         if (ret)
1914                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1915
1916         return ret;
1917 }
1918 EXPORT_SYMBOL(vm_insert_pfn);
1919
1920 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1921                         unsigned long pfn)
1922 {
1923         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1924
1925         if (addr < vma->vm_start || addr >= vma->vm_end)
1926                 return -EFAULT;
1927
1928         /*
1929          * If we don't have pte special, then we have to use the pfn_valid()
1930          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1931          * refcount the page if pfn_valid is true (hence insert_page rather
1932          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1933          * without pte special, it would there be refcounted as a normal page.
1934          */
1935         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1936                 struct page *page;
1937
1938                 page = pfn_to_page(pfn);
1939                 return insert_page(vma, addr, page, vma->vm_page_prot);
1940         }
1941         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1942 }
1943 EXPORT_SYMBOL(vm_insert_mixed);
1944
1945 /*
1946  * maps a range of physical memory into the requested pages. the old
1947  * mappings are removed. any references to nonexistent pages results
1948  * in null mappings (currently treated as "copy-on-access")
1949  */
1950 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1951                         unsigned long addr, unsigned long end,
1952                         unsigned long pfn, pgprot_t prot)
1953 {
1954         pte_t *pte;
1955         spinlock_t *ptl;
1956
1957         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1958         if (!pte)
1959                 return -ENOMEM;
1960         arch_enter_lazy_mmu_mode();
1961         do {
1962                 BUG_ON(!pte_none(*pte));
1963                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1964                 pfn++;
1965         } while (pte++, addr += PAGE_SIZE, addr != end);
1966         arch_leave_lazy_mmu_mode();
1967         pte_unmap_unlock(pte - 1, ptl);
1968         return 0;
1969 }
1970
1971 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1972                         unsigned long addr, unsigned long end,
1973                         unsigned long pfn, pgprot_t prot)
1974 {
1975         pmd_t *pmd;
1976         unsigned long next;
1977
1978         pfn -= addr >> PAGE_SHIFT;
1979         pmd = pmd_alloc(mm, pud, addr);
1980         if (!pmd)
1981                 return -ENOMEM;
1982         VM_BUG_ON(pmd_trans_huge(*pmd));
1983         do {
1984                 next = pmd_addr_end(addr, end);
1985                 if (remap_pte_range(mm, pmd, addr, next,
1986                                 pfn + (addr >> PAGE_SHIFT), prot))
1987                         return -ENOMEM;
1988         } while (pmd++, addr = next, addr != end);
1989         return 0;
1990 }
1991
1992 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1993                         unsigned long addr, unsigned long end,
1994                         unsigned long pfn, pgprot_t prot)
1995 {
1996         pud_t *pud;
1997         unsigned long next;
1998
1999         pfn -= addr >> PAGE_SHIFT;
2000         pud = pud_alloc(mm, pgd, addr);
2001         if (!pud)
2002                 return -ENOMEM;
2003         do {
2004                 next = pud_addr_end(addr, end);
2005                 if (remap_pmd_range(mm, pud, addr, next,
2006                                 pfn + (addr >> PAGE_SHIFT), prot))
2007                         return -ENOMEM;
2008         } while (pud++, addr = next, addr != end);
2009         return 0;
2010 }
2011
2012 /**
2013  * remap_pfn_range - remap kernel memory to userspace
2014  * @vma: user vma to map to
2015  * @addr: target user address to start at
2016  * @pfn: physical address of kernel memory
2017  * @size: size of map area
2018  * @prot: page protection flags for this mapping
2019  *
2020  *  Note: this is only safe if the mm semaphore is held when called.
2021  */
2022 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2023                     unsigned long pfn, unsigned long size, pgprot_t prot)
2024 {
2025         pgd_t *pgd;
2026         unsigned long next;
2027         unsigned long end = addr + PAGE_ALIGN(size);
2028         struct mm_struct *mm = vma->vm_mm;
2029         int err;
2030
2031         /*
2032          * Physically remapped pages are special. Tell the
2033          * rest of the world about it:
2034          *   VM_IO tells people not to look at these pages
2035          *      (accesses can have side effects).
2036          *   VM_RESERVED is specified all over the place, because
2037          *      in 2.4 it kept swapout's vma scan off this vma; but
2038          *      in 2.6 the LRU scan won't even find its pages, so this
2039          *      flag means no more than count its pages in reserved_vm,
2040          *      and omit it from core dump, even when VM_IO turned off.
2041          *   VM_PFNMAP tells the core MM that the base pages are just
2042          *      raw PFN mappings, and do not have a "struct page" associated
2043          *      with them.
2044          *
2045          * There's a horrible special case to handle copy-on-write
2046          * behaviour that some programs depend on. We mark the "original"
2047          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2048          */
2049         if (addr == vma->vm_start && end == vma->vm_end) {
2050                 vma->vm_pgoff = pfn;
2051                 vma->vm_flags |= VM_PFN_AT_MMAP;
2052         } else if (is_cow_mapping(vma->vm_flags))
2053                 return -EINVAL;
2054
2055         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2056
2057         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2058         if (err) {
2059                 /*
2060                  * To indicate that track_pfn related cleanup is not
2061                  * needed from higher level routine calling unmap_vmas
2062                  */
2063                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2064                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2065                 return -EINVAL;
2066         }
2067
2068         BUG_ON(addr >= end);
2069         pfn -= addr >> PAGE_SHIFT;
2070         pgd = pgd_offset(mm, addr);
2071         flush_cache_range(vma, addr, end);
2072         do {
2073                 next = pgd_addr_end(addr, end);
2074                 err = remap_pud_range(mm, pgd, addr, next,
2075                                 pfn + (addr >> PAGE_SHIFT), prot);
2076                 if (err)
2077                         break;
2078         } while (pgd++, addr = next, addr != end);
2079
2080         if (err)
2081                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2082
2083         return err;
2084 }
2085 EXPORT_SYMBOL(remap_pfn_range);
2086
2087 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2088                                      unsigned long addr, unsigned long end,
2089                                      pte_fn_t fn, void *data)
2090 {
2091         pte_t *pte;
2092         int err;
2093         pgtable_t token;
2094         spinlock_t *uninitialized_var(ptl);
2095
2096         pte = (mm == &init_mm) ?
2097                 pte_alloc_kernel(pmd, addr) :
2098                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2099         if (!pte)
2100                 return -ENOMEM;
2101
2102         BUG_ON(pmd_huge(*pmd));
2103
2104         arch_enter_lazy_mmu_mode();
2105
2106         token = pmd_pgtable(*pmd);
2107
2108         do {
2109                 err = fn(pte++, token, addr, data);
2110                 if (err)
2111                         break;
2112         } while (addr += PAGE_SIZE, addr != end);
2113
2114         arch_leave_lazy_mmu_mode();
2115
2116         if (mm != &init_mm)
2117                 pte_unmap_unlock(pte-1, ptl);
2118         return err;
2119 }
2120
2121 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2122                                      unsigned long addr, unsigned long end,
2123                                      pte_fn_t fn, void *data)
2124 {
2125         pmd_t *pmd;
2126         unsigned long next;
2127         int err;
2128
2129         BUG_ON(pud_huge(*pud));
2130
2131         pmd = pmd_alloc(mm, pud, addr);
2132         if (!pmd)
2133                 return -ENOMEM;
2134         do {
2135                 next = pmd_addr_end(addr, end);
2136                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2137                 if (err)
2138                         break;
2139         } while (pmd++, addr = next, addr != end);
2140         return err;
2141 }
2142
2143 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2144                                      unsigned long addr, unsigned long end,
2145                                      pte_fn_t fn, void *data)
2146 {
2147         pud_t *pud;
2148         unsigned long next;
2149         int err;
2150
2151         pud = pud_alloc(mm, pgd, addr);
2152         if (!pud)
2153                 return -ENOMEM;
2154         do {
2155                 next = pud_addr_end(addr, end);
2156                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2157                 if (err)
2158                         break;
2159         } while (pud++, addr = next, addr != end);
2160         return err;
2161 }
2162
2163 /*
2164  * Scan a region of virtual memory, filling in page tables as necessary
2165  * and calling a provided function on each leaf page table.
2166  */
2167 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2168                         unsigned long size, pte_fn_t fn, void *data)
2169 {
2170         pgd_t *pgd;
2171         unsigned long next;
2172         unsigned long end = addr + size;
2173         int err;
2174
2175         BUG_ON(addr >= end);
2176         pgd = pgd_offset(mm, addr);
2177         do {
2178                 next = pgd_addr_end(addr, end);
2179                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2180                 if (err)
2181                         break;
2182         } while (pgd++, addr = next, addr != end);
2183
2184         return err;
2185 }
2186 EXPORT_SYMBOL_GPL(apply_to_page_range);
2187
2188 /*
2189  * handle_pte_fault chooses page fault handler according to an entry
2190  * which was read non-atomically.  Before making any commitment, on
2191  * those architectures or configurations (e.g. i386 with PAE) which
2192  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2193  * must check under lock before unmapping the pte and proceeding
2194  * (but do_wp_page is only called after already making such a check;
2195  * and do_anonymous_page can safely check later on).
2196  */
2197 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2198                                 pte_t *page_table, pte_t orig_pte)
2199 {
2200         int same = 1;
2201 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2202         if (sizeof(pte_t) > sizeof(unsigned long)) {
2203                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2204                 spin_lock(ptl);
2205                 same = pte_same(*page_table, orig_pte);
2206                 spin_unlock(ptl);
2207         }
2208 #endif
2209         pte_unmap(page_table);
2210         return same;
2211 }
2212
2213 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2214 {
2215         /*
2216          * If the source page was a PFN mapping, we don't have
2217          * a "struct page" for it. We do a best-effort copy by
2218          * just copying from the original user address. If that
2219          * fails, we just zero-fill it. Live with it.
2220          */
2221         if (unlikely(!src)) {
2222                 void *kaddr = kmap_atomic(dst, KM_USER0);
2223                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2224
2225                 /*
2226                  * This really shouldn't fail, because the page is there
2227                  * in the page tables. But it might just be unreadable,
2228                  * in which case we just give up and fill the result with
2229                  * zeroes.
2230                  */
2231                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2232                         clear_page(kaddr);
2233                 kunmap_atomic(kaddr, KM_USER0);
2234                 flush_dcache_page(dst);
2235         } else
2236                 copy_user_highpage(dst, src, va, vma);
2237 }
2238
2239 /*
2240  * This routine handles present pages, when users try to write
2241  * to a shared page. It is done by copying the page to a new address
2242  * and decrementing the shared-page counter for the old page.
2243  *
2244  * Note that this routine assumes that the protection checks have been
2245  * done by the caller (the low-level page fault routine in most cases).
2246  * Thus we can safely just mark it writable once we've done any necessary
2247  * COW.
2248  *
2249  * We also mark the page dirty at this point even though the page will
2250  * change only once the write actually happens. This avoids a few races,
2251  * and potentially makes it more efficient.
2252  *
2253  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2254  * but allow concurrent faults), with pte both mapped and locked.
2255  * We return with mmap_sem still held, but pte unmapped and unlocked.
2256  */
2257 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2258                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2259                 spinlock_t *ptl, pte_t orig_pte)
2260         __releases(ptl)
2261 {
2262         struct page *old_page, *new_page;
2263         pte_t entry;
2264         int ret = 0;
2265         int page_mkwrite = 0;
2266         struct page *dirty_page = NULL;
2267
2268         old_page = vm_normal_page(vma, address, orig_pte);
2269         if (!old_page) {
2270                 /*
2271                  * VM_MIXEDMAP !pfn_valid() case
2272                  *
2273                  * We should not cow pages in a shared writeable mapping.
2274                  * Just mark the pages writable as we can't do any dirty
2275                  * accounting on raw pfn maps.
2276                  */
2277                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2278                                      (VM_WRITE|VM_SHARED))
2279                         goto reuse;
2280                 goto gotten;
2281         }
2282
2283         /*
2284          * Take out anonymous pages first, anonymous shared vmas are
2285          * not dirty accountable.
2286          */
2287         if (PageAnon(old_page) && !PageKsm(old_page)) {
2288                 if (!trylock_page(old_page)) {
2289                         page_cache_get(old_page);
2290                         pte_unmap_unlock(page_table, ptl);
2291                         lock_page(old_page);
2292                         page_table = pte_offset_map_lock(mm, pmd, address,
2293                                                          &ptl);
2294                         if (!pte_same(*page_table, orig_pte)) {
2295                                 unlock_page(old_page);
2296                                 goto unlock;
2297                         }
2298                         page_cache_release(old_page);
2299                 }
2300                 if (reuse_swap_page(old_page)) {
2301                         /*
2302                          * The page is all ours.  Move it to our anon_vma so
2303                          * the rmap code will not search our parent or siblings.
2304                          * Protected against the rmap code by the page lock.
2305                          */
2306                         page_move_anon_rmap(old_page, vma, address);
2307                         unlock_page(old_page);
2308                         goto reuse;
2309                 }
2310                 unlock_page(old_page);
2311         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2312                                         (VM_WRITE|VM_SHARED))) {
2313                 /*
2314                  * Only catch write-faults on shared writable pages,
2315                  * read-only shared pages can get COWed by
2316                  * get_user_pages(.write=1, .force=1).
2317                  */
2318                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2319                         struct vm_fault vmf;
2320                         int tmp;
2321
2322                         vmf.virtual_address = (void __user *)(address &
2323                                                                 PAGE_MASK);
2324                         vmf.pgoff = old_page->index;
2325                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2326                         vmf.page = old_page;
2327
2328                         /*
2329                          * Notify the address space that the page is about to
2330                          * become writable so that it can prohibit this or wait
2331                          * for the page to get into an appropriate state.
2332                          *
2333                          * We do this without the lock held, so that it can
2334                          * sleep if it needs to.
2335                          */
2336                         page_cache_get(old_page);
2337                         pte_unmap_unlock(page_table, ptl);
2338
2339                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2340                         if (unlikely(tmp &
2341                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2342                                 ret = tmp;
2343                                 goto unwritable_page;
2344                         }
2345                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2346                                 lock_page(old_page);
2347                                 if (!old_page->mapping) {
2348                                         ret = 0; /* retry the fault */
2349                                         unlock_page(old_page);
2350                                         goto unwritable_page;
2351                                 }
2352                         } else
2353                                 VM_BUG_ON(!PageLocked(old_page));
2354
2355                         /*
2356                          * Since we dropped the lock we need to revalidate
2357                          * the PTE as someone else may have changed it.  If
2358                          * they did, we just return, as we can count on the
2359                          * MMU to tell us if they didn't also make it writable.
2360                          */
2361                         page_table = pte_offset_map_lock(mm, pmd, address,
2362                                                          &ptl);
2363                         if (!pte_same(*page_table, orig_pte)) {
2364                                 unlock_page(old_page);
2365                                 goto unlock;
2366                         }
2367
2368                         page_mkwrite = 1;
2369                 }
2370                 dirty_page = old_page;
2371                 get_page(dirty_page);
2372
2373 reuse:
2374                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2375                 entry = pte_mkyoung(orig_pte);
2376                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2377                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2378                         update_mmu_cache(vma, address, page_table);
2379                 pte_unmap_unlock(page_table, ptl);
2380                 ret |= VM_FAULT_WRITE;
2381
2382                 if (!dirty_page)
2383                         return ret;
2384
2385                 /*
2386                  * Yes, Virginia, this is actually required to prevent a race
2387                  * with clear_page_dirty_for_io() from clearing the page dirty
2388                  * bit after it clear all dirty ptes, but before a racing
2389                  * do_wp_page installs a dirty pte.
2390                  *
2391                  * __do_fault is protected similarly.
2392                  */
2393                 if (!page_mkwrite) {
2394                         wait_on_page_locked(dirty_page);
2395                         set_page_dirty_balance(dirty_page, page_mkwrite);
2396                 }
2397                 put_page(dirty_page);
2398                 if (page_mkwrite) {
2399                         struct address_space *mapping = dirty_page->mapping;
2400
2401                         set_page_dirty(dirty_page);
2402                         unlock_page(dirty_page);
2403                         page_cache_release(dirty_page);
2404                         if (mapping)    {
2405                                 /*
2406                                  * Some device drivers do not set page.mapping
2407                                  * but still dirty their pages
2408                                  */
2409                                 balance_dirty_pages_ratelimited(mapping);
2410                         }
2411                 }
2412
2413                 /* file_update_time outside page_lock */
2414                 if (vma->vm_file)
2415                         file_update_time(vma->vm_file);
2416
2417                 return ret;
2418         }
2419
2420         /*
2421          * Ok, we need to copy. Oh, well..
2422          */
2423         page_cache_get(old_page);
2424 gotten:
2425         pte_unmap_unlock(page_table, ptl);
2426
2427         if (unlikely(anon_vma_prepare(vma)))
2428                 goto oom;
2429
2430         if (is_zero_pfn(pte_pfn(orig_pte))) {
2431                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2432                 if (!new_page)
2433                         goto oom;
2434         } else {
2435                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2436                 if (!new_page)
2437                         goto oom;
2438                 cow_user_page(new_page, old_page, address, vma);
2439         }
2440         __SetPageUptodate(new_page);
2441
2442         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2443                 goto oom_free_new;
2444
2445         /*
2446          * Re-check the pte - we dropped the lock
2447          */
2448         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2449         if (likely(pte_same(*page_table, orig_pte))) {
2450                 if (old_page) {
2451                         if (!PageAnon(old_page)) {
2452                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2453                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2454                         }
2455                 } else
2456                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2457                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2458                 entry = mk_pte(new_page, vma->vm_page_prot);
2459                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2460                 /*
2461                  * Clear the pte entry and flush it first, before updating the
2462                  * pte with the new entry. This will avoid a race condition
2463                  * seen in the presence of one thread doing SMC and another
2464                  * thread doing COW.
2465                  */
2466                 ptep_clear_flush(vma, address, page_table);
2467                 page_add_new_anon_rmap(new_page, vma, address);
2468                 /*
2469                  * We call the notify macro here because, when using secondary
2470                  * mmu page tables (such as kvm shadow page tables), we want the
2471                  * new page to be mapped directly into the secondary page table.
2472                  */
2473                 set_pte_at_notify(mm, address, page_table, entry);
2474                 update_mmu_cache(vma, address, page_table);
2475                 if (old_page) {
2476                         /*
2477                          * Only after switching the pte to the new page may
2478                          * we remove the mapcount here. Otherwise another
2479                          * process may come and find the rmap count decremented
2480                          * before the pte is switched to the new page, and
2481                          * "reuse" the old page writing into it while our pte
2482                          * here still points into it and can be read by other
2483                          * threads.
2484                          *
2485                          * The critical issue is to order this
2486                          * page_remove_rmap with the ptp_clear_flush above.
2487                          * Those stores are ordered by (if nothing else,)
2488                          * the barrier present in the atomic_add_negative
2489                          * in page_remove_rmap.
2490                          *
2491                          * Then the TLB flush in ptep_clear_flush ensures that
2492                          * no process can access the old page before the
2493                          * decremented mapcount is visible. And the old page
2494                          * cannot be reused until after the decremented
2495                          * mapcount is visible. So transitively, TLBs to
2496                          * old page will be flushed before it can be reused.
2497                          */
2498                         page_remove_rmap(old_page);
2499                 }
2500
2501                 /* Free the old page.. */
2502                 new_page = old_page;
2503                 ret |= VM_FAULT_WRITE;
2504         } else
2505                 mem_cgroup_uncharge_page(new_page);
2506
2507         if (new_page)
2508                 page_cache_release(new_page);
2509 unlock:
2510         pte_unmap_unlock(page_table, ptl);
2511         if (old_page) {
2512                 /*
2513                  * Don't let another task, with possibly unlocked vma,
2514                  * keep the mlocked page.
2515                  */
2516                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2517                         lock_page(old_page);    /* LRU manipulation */
2518                         munlock_vma_page(old_page);
2519                         unlock_page(old_page);
2520                 }
2521                 page_cache_release(old_page);
2522         }
2523         return ret;
2524 oom_free_new:
2525         page_cache_release(new_page);
2526 oom:
2527         if (old_page) {
2528                 if (page_mkwrite) {
2529                         unlock_page(old_page);
2530                         page_cache_release(old_page);
2531                 }
2532                 page_cache_release(old_page);
2533         }
2534         return VM_FAULT_OOM;
2535
2536 unwritable_page:
2537         page_cache_release(old_page);
2538         return ret;
2539 }
2540
2541 /*
2542  * Helper functions for unmap_mapping_range().
2543  *
2544  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2545  *
2546  * We have to restart searching the prio_tree whenever we drop the lock,
2547  * since the iterator is only valid while the lock is held, and anyway
2548  * a later vma might be split and reinserted earlier while lock dropped.
2549  *
2550  * The list of nonlinear vmas could be handled more efficiently, using
2551  * a placeholder, but handle it in the same way until a need is shown.
2552  * It is important to search the prio_tree before nonlinear list: a vma
2553  * may become nonlinear and be shifted from prio_tree to nonlinear list
2554  * while the lock is dropped; but never shifted from list to prio_tree.
2555  *
2556  * In order to make forward progress despite restarting the search,
2557  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2558  * quickly skip it next time around.  Since the prio_tree search only
2559  * shows us those vmas affected by unmapping the range in question, we
2560  * can't efficiently keep all vmas in step with mapping->truncate_count:
2561  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2562  * mapping->truncate_count and vma->vm_truncate_count are protected by
2563  * i_mmap_lock.
2564  *
2565  * In order to make forward progress despite repeatedly restarting some
2566  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2567  * and restart from that address when we reach that vma again.  It might
2568  * have been split or merged, shrunk or extended, but never shifted: so
2569  * restart_addr remains valid so long as it remains in the vma's range.
2570  * unmap_mapping_range forces truncate_count to leap over page-aligned
2571  * values so we can save vma's restart_addr in its truncate_count field.
2572  */
2573 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2574
2575 static void reset_vma_truncate_counts(struct address_space *mapping)
2576 {
2577         struct vm_area_struct *vma;
2578         struct prio_tree_iter iter;
2579
2580         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2581                 vma->vm_truncate_count = 0;
2582         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2583                 vma->vm_truncate_count = 0;
2584 }
2585
2586 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2587                 unsigned long start_addr, unsigned long end_addr,
2588                 struct zap_details *details)
2589 {
2590         unsigned long restart_addr;
2591         int need_break;
2592
2593         /*
2594          * files that support invalidating or truncating portions of the
2595          * file from under mmaped areas must have their ->fault function
2596          * return a locked page (and set VM_FAULT_LOCKED in the return).
2597          * This provides synchronisation against concurrent unmapping here.
2598          */
2599
2600 again:
2601         restart_addr = vma->vm_truncate_count;
2602         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2603                 start_addr = restart_addr;
2604                 if (start_addr >= end_addr) {
2605                         /* Top of vma has been split off since last time */
2606                         vma->vm_truncate_count = details->truncate_count;
2607                         return 0;
2608                 }
2609         }
2610
2611         restart_addr = zap_page_range(vma, start_addr,
2612                                         end_addr - start_addr, details);
2613         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2614
2615         if (restart_addr >= end_addr) {
2616                 /* We have now completed this vma: mark it so */
2617                 vma->vm_truncate_count = details->truncate_count;
2618                 if (!need_break)
2619                         return 0;
2620         } else {
2621                 /* Note restart_addr in vma's truncate_count field */
2622                 vma->vm_truncate_count = restart_addr;
2623                 if (!need_break)
2624                         goto again;
2625         }
2626
2627         spin_unlock(details->i_mmap_lock);
2628         cond_resched();
2629         spin_lock(details->i_mmap_lock);
2630         return -EINTR;
2631 }
2632
2633 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2634                                             struct zap_details *details)
2635 {
2636         struct vm_area_struct *vma;
2637         struct prio_tree_iter iter;
2638         pgoff_t vba, vea, zba, zea;
2639
2640 restart:
2641         vma_prio_tree_foreach(vma, &iter, root,
2642                         details->first_index, details->last_index) {
2643                 /* Skip quickly over those we have already dealt with */
2644                 if (vma->vm_truncate_count == details->truncate_count)
2645                         continue;
2646
2647                 vba = vma->vm_pgoff;
2648                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2649                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2650                 zba = details->first_index;
2651                 if (zba < vba)
2652                         zba = vba;
2653                 zea = details->last_index;
2654                 if (zea > vea)
2655                         zea = vea;
2656
2657                 if (unmap_mapping_range_vma(vma,
2658                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2659                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2660                                 details) < 0)
2661                         goto restart;
2662         }
2663 }
2664
2665 static inline void unmap_mapping_range_list(struct list_head *head,
2666                                             struct zap_details *details)
2667 {
2668         struct vm_area_struct *vma;
2669
2670         /*
2671          * In nonlinear VMAs there is no correspondence between virtual address
2672          * offset and file offset.  So we must perform an exhaustive search
2673          * across *all* the pages in each nonlinear VMA, not just the pages
2674          * whose virtual address lies outside the file truncation point.
2675          */
2676 restart:
2677         list_for_each_entry(vma, head, shared.vm_set.list) {
2678                 /* Skip quickly over those we have already dealt with */
2679                 if (vma->vm_truncate_count == details->truncate_count)
2680                         continue;
2681                 details->nonlinear_vma = vma;
2682                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2683                                         vma->vm_end, details) < 0)
2684                         goto restart;
2685         }
2686 }
2687
2688 /**
2689  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2690  * @mapping: the address space containing mmaps to be unmapped.
2691  * @holebegin: byte in first page to unmap, relative to the start of
2692  * the underlying file.  This will be rounded down to a PAGE_SIZE
2693  * boundary.  Note that this is different from truncate_pagecache(), which
2694  * must keep the partial page.  In contrast, we must get rid of
2695  * partial pages.
2696  * @holelen: size of prospective hole in bytes.  This will be rounded
2697  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2698  * end of the file.
2699  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2700  * but 0 when invalidating pagecache, don't throw away private data.
2701  */
2702 void unmap_mapping_range(struct address_space *mapping,
2703                 loff_t const holebegin, loff_t const holelen, int even_cows)
2704 {
2705         struct zap_details details;
2706         pgoff_t hba = holebegin >> PAGE_SHIFT;
2707         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2708
2709         /* Check for overflow. */
2710         if (sizeof(holelen) > sizeof(hlen)) {
2711                 long long holeend =
2712                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2713                 if (holeend & ~(long long)ULONG_MAX)
2714                         hlen = ULONG_MAX - hba + 1;
2715         }
2716
2717         details.check_mapping = even_cows? NULL: mapping;
2718         details.nonlinear_vma = NULL;
2719         details.first_index = hba;
2720         details.last_index = hba + hlen - 1;
2721         if (details.last_index < details.first_index)
2722                 details.last_index = ULONG_MAX;
2723         details.i_mmap_lock = &mapping->i_mmap_lock;
2724
2725         mutex_lock(&mapping->unmap_mutex);
2726         spin_lock(&mapping->i_mmap_lock);
2727
2728         /* Protect against endless unmapping loops */
2729         mapping->truncate_count++;
2730         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2731                 if (mapping->truncate_count == 0)
2732                         reset_vma_truncate_counts(mapping);
2733                 mapping->truncate_count++;
2734         }
2735         details.truncate_count = mapping->truncate_count;
2736
2737         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2738                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2739         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2740                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2741         spin_unlock(&mapping->i_mmap_lock);
2742         mutex_unlock(&mapping->unmap_mutex);
2743 }
2744 EXPORT_SYMBOL(unmap_mapping_range);
2745
2746 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2747 {
2748         struct address_space *mapping = inode->i_mapping;
2749
2750         /*
2751          * If the underlying filesystem is not going to provide
2752          * a way to truncate a range of blocks (punch a hole) -
2753          * we should return failure right now.
2754          */
2755         if (!inode->i_op->truncate_range)
2756                 return -ENOSYS;
2757
2758         mutex_lock(&inode->i_mutex);
2759         down_write(&inode->i_alloc_sem);
2760         unmap_mapping_range(mapping, offset, (end - offset), 1);
2761         truncate_inode_pages_range(mapping, offset, end);
2762         unmap_mapping_range(mapping, offset, (end - offset), 1);
2763         inode->i_op->truncate_range(inode, offset, end);
2764         up_write(&inode->i_alloc_sem);
2765         mutex_unlock(&inode->i_mutex);
2766
2767         return 0;
2768 }
2769
2770 /*
2771  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2772  * but allow concurrent faults), and pte mapped but not yet locked.
2773  * We return with mmap_sem still held, but pte unmapped and unlocked.
2774  */
2775 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2776                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2777                 unsigned int flags, pte_t orig_pte)
2778 {
2779         spinlock_t *ptl;
2780         struct page *page, *swapcache = NULL;
2781         swp_entry_t entry;
2782         pte_t pte;
2783         int locked;
2784         struct mem_cgroup *ptr;
2785         int exclusive = 0;
2786         int ret = 0;
2787
2788         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2789                 goto out;
2790
2791         entry = pte_to_swp_entry(orig_pte);
2792         if (unlikely(non_swap_entry(entry))) {
2793                 if (is_migration_entry(entry)) {
2794                         migration_entry_wait(mm, pmd, address);
2795                 } else if (is_hwpoison_entry(entry)) {
2796                         ret = VM_FAULT_HWPOISON;
2797                 } else {
2798                         print_bad_pte(vma, address, orig_pte, NULL);
2799                         ret = VM_FAULT_SIGBUS;
2800                 }
2801                 goto out;
2802         }
2803         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2804         page = lookup_swap_cache(entry);
2805         if (!page) {
2806                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2807                 page = swapin_readahead(entry,
2808                                         GFP_HIGHUSER_MOVABLE, vma, address);
2809                 if (!page) {
2810                         /*
2811                          * Back out if somebody else faulted in this pte
2812                          * while we released the pte lock.
2813                          */
2814                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2815                         if (likely(pte_same(*page_table, orig_pte)))
2816                                 ret = VM_FAULT_OOM;
2817                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2818                         goto unlock;
2819                 }
2820
2821                 /* Had to read the page from swap area: Major fault */
2822                 ret = VM_FAULT_MAJOR;
2823                 count_vm_event(PGMAJFAULT);
2824         } else if (PageHWPoison(page)) {
2825                 /*
2826                  * hwpoisoned dirty swapcache pages are kept for killing
2827                  * owner processes (which may be unknown at hwpoison time)
2828                  */
2829                 ret = VM_FAULT_HWPOISON;
2830                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2831                 goto out_release;
2832         }
2833
2834         locked = lock_page_or_retry(page, mm, flags);
2835         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2836         if (!locked) {
2837                 ret |= VM_FAULT_RETRY;
2838                 goto out_release;
2839         }
2840
2841         /*
2842          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2843          * release the swapcache from under us.  The page pin, and pte_same
2844          * test below, are not enough to exclude that.  Even if it is still
2845          * swapcache, we need to check that the page's swap has not changed.
2846          */
2847         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2848                 goto out_page;
2849
2850         if (ksm_might_need_to_copy(page, vma, address)) {
2851                 swapcache = page;
2852                 page = ksm_does_need_to_copy(page, vma, address);
2853
2854                 if (unlikely(!page)) {
2855                         ret = VM_FAULT_OOM;
2856                         page = swapcache;
2857                         swapcache = NULL;
2858                         goto out_page;
2859                 }
2860         }
2861
2862         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2863                 ret = VM_FAULT_OOM;
2864                 goto out_page;
2865         }
2866
2867         /*
2868          * Back out if somebody else already faulted in this pte.
2869          */
2870         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2871         if (unlikely(!pte_same(*page_table, orig_pte)))
2872                 goto out_nomap;
2873
2874         if (unlikely(!PageUptodate(page))) {
2875                 ret = VM_FAULT_SIGBUS;
2876                 goto out_nomap;
2877         }
2878
2879         /*
2880          * The page isn't present yet, go ahead with the fault.
2881          *
2882          * Be careful about the sequence of operations here.
2883          * To get its accounting right, reuse_swap_page() must be called
2884          * while the page is counted on swap but not yet in mapcount i.e.
2885          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2886          * must be called after the swap_free(), or it will never succeed.
2887          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2888          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2889          * in page->private. In this case, a record in swap_cgroup  is silently
2890          * discarded at swap_free().
2891          */
2892
2893         inc_mm_counter_fast(mm, MM_ANONPAGES);
2894         dec_mm_counter_fast(mm, MM_SWAPENTS);
2895         pte = mk_pte(page, vma->vm_page_prot);
2896         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2897                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2898                 flags &= ~FAULT_FLAG_WRITE;
2899                 ret |= VM_FAULT_WRITE;
2900                 exclusive = 1;
2901         }
2902         flush_icache_page(vma, page);
2903         set_pte_at(mm, address, page_table, pte);
2904         do_page_add_anon_rmap(page, vma, address, exclusive);
2905         /* It's better to call commit-charge after rmap is established */
2906         mem_cgroup_commit_charge_swapin(page, ptr);
2907
2908         swap_free(entry);
2909         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2910                 try_to_free_swap(page);
2911         unlock_page(page);
2912         if (swapcache) {
2913                 /*
2914                  * Hold the lock to avoid the swap entry to be reused
2915                  * until we take the PT lock for the pte_same() check
2916                  * (to avoid false positives from pte_same). For
2917                  * further safety release the lock after the swap_free
2918                  * so that the swap count won't change under a
2919                  * parallel locked swapcache.
2920                  */
2921                 unlock_page(swapcache);
2922                 page_cache_release(swapcache);
2923         }
2924
2925         if (flags & FAULT_FLAG_WRITE) {
2926                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2927                 if (ret & VM_FAULT_ERROR)
2928                         ret &= VM_FAULT_ERROR;
2929                 goto out;
2930         }
2931
2932         /* No need to invalidate - it was non-present before */
2933         update_mmu_cache(vma, address, page_table);
2934 unlock:
2935         pte_unmap_unlock(page_table, ptl);
2936 out:
2937         return ret;
2938 out_nomap:
2939         mem_cgroup_cancel_charge_swapin(ptr);
2940         pte_unmap_unlock(page_table, ptl);
2941 out_page:
2942         unlock_page(page);
2943 out_release:
2944         page_cache_release(page);
2945         if (swapcache) {
2946                 unlock_page(swapcache);
2947                 page_cache_release(swapcache);
2948         }
2949         return ret;
2950 }
2951
2952 /*
2953  * This is like a special single-page "expand_{down|up}wards()",
2954  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2955  * doesn't hit another vma.
2956  */
2957 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2958 {
2959         address &= PAGE_MASK;
2960         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2961                 struct vm_area_struct *prev = vma->vm_prev;
2962
2963                 /*
2964                  * Is there a mapping abutting this one below?
2965                  *
2966                  * That's only ok if it's the same stack mapping
2967                  * that has gotten split..
2968                  */
2969                 if (prev && prev->vm_end == address)
2970                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2971
2972                 expand_stack(vma, address - PAGE_SIZE);
2973         }
2974         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2975                 struct vm_area_struct *next = vma->vm_next;
2976
2977                 /* As VM_GROWSDOWN but s/below/above/ */
2978                 if (next && next->vm_start == address + PAGE_SIZE)
2979                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2980
2981                 expand_upwards(vma, address + PAGE_SIZE);
2982         }
2983         return 0;
2984 }
2985
2986 /*
2987  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2988  * but allow concurrent faults), and pte mapped but not yet locked.
2989  * We return with mmap_sem still held, but pte unmapped and unlocked.
2990  */
2991 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2992                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2993                 unsigned int flags)
2994 {
2995         struct page *page;
2996         spinlock_t *ptl;
2997         pte_t entry;
2998
2999         pte_unmap(page_table);
3000
3001         /* Check if we need to add a guard page to the stack */
3002         if (check_stack_guard_page(vma, address) < 0)
3003                 return VM_FAULT_SIGBUS;
3004
3005         /* Use the zero-page for reads */
3006         if (!(flags & FAULT_FLAG_WRITE)) {
3007                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3008                                                 vma->vm_page_prot));
3009                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3010                 if (!pte_none(*page_table))
3011                         goto unlock;
3012                 goto setpte;
3013         }
3014
3015         /* Allocate our own private page. */
3016         if (unlikely(anon_vma_prepare(vma)))
3017                 goto oom;
3018         page = alloc_zeroed_user_highpage_movable(vma, address);
3019         if (!page)
3020                 goto oom;
3021         __SetPageUptodate(page);
3022
3023         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3024                 goto oom_free_page;
3025
3026         entry = mk_pte(page, vma->vm_page_prot);
3027         if (vma->vm_flags & VM_WRITE)
3028                 entry = pte_mkwrite(pte_mkdirty(entry));
3029
3030         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3031         if (!pte_none(*page_table))
3032                 goto release;
3033
3034         inc_mm_counter_fast(mm, MM_ANONPAGES);
3035         page_add_new_anon_rmap(page, vma, address);
3036 setpte:
3037         set_pte_at(mm, address, page_table, entry);
3038
3039         /* No need to invalidate - it was non-present before */
3040         update_mmu_cache(vma, address, page_table);
3041 unlock:
3042         pte_unmap_unlock(page_table, ptl);
3043         return 0;
3044 release:
3045         mem_cgroup_uncharge_page(page);
3046         page_cache_release(page);
3047         goto unlock;
3048 oom_free_page:
3049         page_cache_release(page);
3050 oom:
3051         return VM_FAULT_OOM;
3052 }
3053
3054 /*
3055  * __do_fault() tries to create a new page mapping. It aggressively
3056  * tries to share with existing pages, but makes a separate copy if
3057  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3058  * the next page fault.
3059  *
3060  * As this is called only for pages that do not currently exist, we
3061  * do not need to flush old virtual caches or the TLB.
3062  *
3063  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3064  * but allow concurrent faults), and pte neither mapped nor locked.
3065  * We return with mmap_sem still held, but pte unmapped and unlocked.
3066  */
3067 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3068                 unsigned long address, pmd_t *pmd,
3069                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3070 {
3071         pte_t *page_table;
3072         spinlock_t *ptl;
3073         struct page *page;
3074         pte_t entry;
3075         int anon = 0;
3076         int charged = 0;
3077         struct page *dirty_page = NULL;
3078         struct vm_fault vmf;
3079         int ret;
3080         int page_mkwrite = 0;
3081
3082         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3083         vmf.pgoff = pgoff;
3084         vmf.flags = flags;
3085         vmf.page = NULL;
3086
3087         ret = vma->vm_ops->fault(vma, &vmf);
3088         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3089                             VM_FAULT_RETRY)))
3090                 return ret;
3091
3092         if (unlikely(PageHWPoison(vmf.page))) {
3093                 if (ret & VM_FAULT_LOCKED)
3094                         unlock_page(vmf.page);
3095                 return VM_FAULT_HWPOISON;
3096         }
3097
3098         /*
3099          * For consistency in subsequent calls, make the faulted page always
3100          * locked.
3101          */
3102         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3103                 lock_page(vmf.page);
3104         else
3105                 VM_BUG_ON(!PageLocked(vmf.page));
3106
3107         /*
3108          * Should we do an early C-O-W break?
3109          */
3110         page = vmf.page;
3111         if (flags & FAULT_FLAG_WRITE) {
3112                 if (!(vma->vm_flags & VM_SHARED)) {
3113                         anon = 1;
3114                         if (unlikely(anon_vma_prepare(vma))) {
3115                                 ret = VM_FAULT_OOM;
3116                                 goto out;
3117                         }
3118                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3119                                                 vma, address);
3120                         if (!page) {
3121                                 ret = VM_FAULT_OOM;
3122                                 goto out;
3123                         }
3124                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3125                                 ret = VM_FAULT_OOM;
3126                                 page_cache_release(page);
3127                                 goto out;
3128                         }
3129                         charged = 1;
3130                         copy_user_highpage(page, vmf.page, address, vma);
3131                         __SetPageUptodate(page);
3132                 } else {
3133                         /*
3134                          * If the page will be shareable, see if the backing
3135                          * address space wants to know that the page is about
3136                          * to become writable
3137                          */
3138                         if (vma->vm_ops->page_mkwrite) {
3139                                 int tmp;
3140
3141                                 unlock_page(page);
3142                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3143                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3144                                 if (unlikely(tmp &
3145                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3146                                         ret = tmp;
3147                                         goto unwritable_page;
3148                                 }
3149                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3150                                         lock_page(page);
3151                                         if (!page->mapping) {
3152                                                 ret = 0; /* retry the fault */
3153                                                 unlock_page(page);
3154                                                 goto unwritable_page;
3155                                         }
3156                                 } else
3157                                         VM_BUG_ON(!PageLocked(page));
3158                                 page_mkwrite = 1;
3159                         }
3160                 }
3161
3162         }
3163
3164         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3165
3166         /*
3167          * This silly early PAGE_DIRTY setting removes a race
3168          * due to the bad i386 page protection. But it's valid
3169          * for other architectures too.
3170          *
3171          * Note that if FAULT_FLAG_WRITE is set, we either now have
3172          * an exclusive copy of the page, or this is a shared mapping,
3173          * so we can make it writable and dirty to avoid having to
3174          * handle that later.
3175          */
3176         /* Only go through if we didn't race with anybody else... */
3177         if (likely(pte_same(*page_table, orig_pte))) {
3178                 flush_icache_page(vma, page);
3179                 entry = mk_pte(page, vma->vm_page_prot);
3180                 if (flags & FAULT_FLAG_WRITE)
3181                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3182                 if (anon) {
3183                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3184                         page_add_new_anon_rmap(page, vma, address);
3185                 } else {
3186                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3187                         page_add_file_rmap(page);
3188                         if (flags & FAULT_FLAG_WRITE) {
3189                                 dirty_page = page;
3190                                 get_page(dirty_page);
3191                         }
3192                 }
3193                 set_pte_at(mm, address, page_table, entry);
3194
3195                 /* no need to invalidate: a not-present page won't be cached */
3196                 update_mmu_cache(vma, address, page_table);
3197         } else {
3198                 if (charged)
3199                         mem_cgroup_uncharge_page(page);
3200                 if (anon)
3201                         page_cache_release(page);
3202                 else
3203                         anon = 1; /* no anon but release faulted_page */
3204         }
3205
3206         pte_unmap_unlock(page_table, ptl);
3207
3208 out:
3209         if (dirty_page) {
3210                 struct address_space *mapping = page->mapping;
3211
3212                 if (set_page_dirty(dirty_page))
3213                         page_mkwrite = 1;
3214                 unlock_page(dirty_page);
3215                 put_page(dirty_page);
3216                 if (page_mkwrite && mapping) {
3217                         /*
3218                          * Some device drivers do not set page.mapping but still
3219                          * dirty their pages
3220                          */
3221                         balance_dirty_pages_ratelimited(mapping);
3222                 }
3223
3224                 /* file_update_time outside page_lock */
3225                 if (vma->vm_file)
3226                         file_update_time(vma->vm_file);
3227         } else {
3228                 unlock_page(vmf.page);
3229                 if (anon)
3230                         page_cache_release(vmf.page);
3231         }
3232
3233         return ret;
3234
3235 unwritable_page:
3236         page_cache_release(page);
3237         return ret;
3238 }
3239
3240 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3241                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3242                 unsigned int flags, pte_t orig_pte)
3243 {
3244         pgoff_t pgoff = (((address & PAGE_MASK)
3245                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3246
3247         pte_unmap(page_table);
3248         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3249 }
3250
3251 /*
3252  * Fault of a previously existing named mapping. Repopulate the pte
3253  * from the encoded file_pte if possible. This enables swappable
3254  * nonlinear vmas.
3255  *
3256  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3257  * but allow concurrent faults), and pte mapped but not yet locked.
3258  * We return with mmap_sem still held, but pte unmapped and unlocked.
3259  */
3260 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3261                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3262                 unsigned int flags, pte_t orig_pte)
3263 {
3264         pgoff_t pgoff;
3265
3266         flags |= FAULT_FLAG_NONLINEAR;
3267
3268         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3269                 return 0;
3270
3271         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3272                 /*
3273                  * Page table corrupted: show pte and kill process.
3274                  */
3275                 print_bad_pte(vma, address, orig_pte, NULL);
3276                 return VM_FAULT_SIGBUS;
3277         }
3278
3279         pgoff = pte_to_pgoff(orig_pte);
3280         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3281 }
3282
3283 /*
3284  * These routines also need to handle stuff like marking pages dirty
3285  * and/or accessed for architectures that don't do it in hardware (most
3286  * RISC architectures).  The early dirtying is also good on the i386.
3287  *
3288  * There is also a hook called "update_mmu_cache()" that architectures
3289  * with external mmu caches can use to update those (ie the Sparc or
3290  * PowerPC hashed page tables that act as extended TLBs).
3291  *
3292  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3293  * but allow concurrent faults), and pte mapped but not yet locked.
3294  * We return with mmap_sem still held, but pte unmapped and unlocked.
3295  */
3296 int handle_pte_fault(struct mm_struct *mm,
3297                      struct vm_area_struct *vma, unsigned long address,
3298                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3299 {
3300         pte_t entry;
3301         spinlock_t *ptl;
3302
3303         entry = *pte;
3304         if (!pte_present(entry)) {
3305                 if (pte_none(entry)) {
3306                         if (vma->vm_ops) {
3307                                 if (likely(vma->vm_ops->fault))
3308                                         return do_linear_fault(mm, vma, address,
3309                                                 pte, pmd, flags, entry);
3310                         }
3311                         return do_anonymous_page(mm, vma, address,
3312                                                  pte, pmd, flags);
3313                 }
3314                 if (pte_file(entry))
3315                         return do_nonlinear_fault(mm, vma, address,
3316                                         pte, pmd, flags, entry);
3317                 return do_swap_page(mm, vma, address,
3318                                         pte, pmd, flags, entry);
3319         }
3320
3321         ptl = pte_lockptr(mm, pmd);
3322         spin_lock(ptl);
3323         if (unlikely(!pte_same(*pte, entry)))
3324                 goto unlock;
3325         if (flags & FAULT_FLAG_WRITE) {
3326                 if (!pte_write(entry))
3327                         return do_wp_page(mm, vma, address,
3328                                         pte, pmd, ptl, entry);
3329                 entry = pte_mkdirty(entry);
3330         }
3331         entry = pte_mkyoung(entry);
3332         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3333                 update_mmu_cache(vma, address, pte);
3334         } else {
3335                 /*
3336                  * This is needed only for protection faults but the arch code
3337                  * is not yet telling us if this is a protection fault or not.
3338                  * This still avoids useless tlb flushes for .text page faults
3339                  * with threads.
3340                  */
3341                 if (flags & FAULT_FLAG_WRITE)
3342                         flush_tlb_fix_spurious_fault(vma, address);
3343         }
3344 unlock:
3345         pte_unmap_unlock(pte, ptl);
3346         return 0;
3347 }
3348
3349 /*
3350  * By the time we get here, we already hold the mm semaphore
3351  */
3352 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3353                 unsigned long address, unsigned int flags)
3354 {
3355         pgd_t *pgd;
3356         pud_t *pud;
3357         pmd_t *pmd;
3358         pte_t *pte;
3359
3360         __set_current_state(TASK_RUNNING);
3361
3362         count_vm_event(PGFAULT);
3363
3364         /* do counter updates before entering really critical section. */
3365         check_sync_rss_stat(current);
3366
3367         if (unlikely(is_vm_hugetlb_page(vma)))
3368                 return hugetlb_fault(mm, vma, address, flags);
3369
3370         pgd = pgd_offset(mm, address);
3371         pud = pud_alloc(mm, pgd, address);
3372         if (!pud)
3373                 return VM_FAULT_OOM;
3374         pmd = pmd_alloc(mm, pud, address);
3375         if (!pmd)
3376                 return VM_FAULT_OOM;
3377         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3378                 if (!vma->vm_ops)
3379                         return do_huge_pmd_anonymous_page(mm, vma, address,
3380                                                           pmd, flags);
3381         } else {
3382                 pmd_t orig_pmd = *pmd;
3383                 barrier();
3384                 if (pmd_trans_huge(orig_pmd)) {
3385                         if (flags & FAULT_FLAG_WRITE &&
3386                             !pmd_write(orig_pmd) &&
3387                             !pmd_trans_splitting(orig_pmd))
3388                                 return do_huge_pmd_wp_page(mm, vma, address,
3389                                                            pmd, orig_pmd);
3390                         return 0;
3391                 }
3392         }
3393
3394         /*
3395          * Use __pte_alloc instead of pte_alloc_map, because we can't
3396          * run pte_offset_map on the pmd, if an huge pmd could
3397          * materialize from under us from a different thread.
3398          */
3399         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
3400                 return VM_FAULT_OOM;
3401         /* if an huge pmd materialized from under us just retry later */
3402         if (unlikely(pmd_trans_huge(*pmd)))
3403                 return 0;
3404         /*
3405          * A regular pmd is established and it can't morph into a huge pmd
3406          * from under us anymore at this point because we hold the mmap_sem
3407          * read mode and khugepaged takes it in write mode. So now it's
3408          * safe to run pte_offset_map().
3409          */
3410         pte = pte_offset_map(pmd, address);
3411
3412         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3413 }
3414
3415 #ifndef __PAGETABLE_PUD_FOLDED
3416 /*
3417  * Allocate page upper directory.
3418  * We've already handled the fast-path in-line.
3419  */
3420 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3421 {
3422         pud_t *new = pud_alloc_one(mm, address);
3423         if (!new)
3424                 return -ENOMEM;
3425
3426         smp_wmb(); /* See comment in __pte_alloc */
3427
3428         spin_lock(&mm->page_table_lock);
3429         if (pgd_present(*pgd))          /* Another has populated it */
3430                 pud_free(mm, new);
3431         else
3432                 pgd_populate(mm, pgd, new);
3433         spin_unlock(&mm->page_table_lock);
3434         return 0;
3435 }
3436 #endif /* __PAGETABLE_PUD_FOLDED */
3437
3438 #ifndef __PAGETABLE_PMD_FOLDED
3439 /*
3440  * Allocate page middle directory.
3441  * We've already handled the fast-path in-line.
3442  */
3443 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3444 {
3445         pmd_t *new = pmd_alloc_one(mm, address);
3446         if (!new)
3447                 return -ENOMEM;
3448
3449         smp_wmb(); /* See comment in __pte_alloc */
3450
3451         spin_lock(&mm->page_table_lock);
3452 #ifndef __ARCH_HAS_4LEVEL_HACK
3453         if (pud_present(*pud))          /* Another has populated it */
3454                 pmd_free(mm, new);
3455         else
3456                 pud_populate(mm, pud, new);
3457 #else
3458         if (pgd_present(*pud))          /* Another has populated it */
3459                 pmd_free(mm, new);
3460         else
3461                 pgd_populate(mm, pud, new);
3462 #endif /* __ARCH_HAS_4LEVEL_HACK */
3463         spin_unlock(&mm->page_table_lock);
3464         return 0;
3465 }
3466 #endif /* __PAGETABLE_PMD_FOLDED */
3467
3468 int make_pages_present(unsigned long addr, unsigned long end)
3469 {
3470         int ret, len, write;
3471         struct vm_area_struct * vma;
3472
3473         vma = find_vma(current->mm, addr);
3474         if (!vma)
3475                 return -ENOMEM;
3476         /*
3477          * We want to touch writable mappings with a write fault in order
3478          * to break COW, except for shared mappings because these don't COW
3479          * and we would not want to dirty them for nothing.
3480          */
3481         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3482         BUG_ON(addr >= end);
3483         BUG_ON(end > vma->vm_end);
3484         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3485         ret = get_user_pages(current, current->mm, addr,
3486                         len, write, 0, NULL, NULL);
3487         if (ret < 0)
3488                 return ret;
3489         return ret == len ? 0 : -EFAULT;
3490 }
3491
3492 #if !defined(__HAVE_ARCH_GATE_AREA)
3493
3494 #if defined(AT_SYSINFO_EHDR)
3495 static struct vm_area_struct gate_vma;
3496
3497 static int __init gate_vma_init(void)
3498 {
3499         gate_vma.vm_mm = NULL;
3500         gate_vma.vm_start = FIXADDR_USER_START;
3501         gate_vma.vm_end = FIXADDR_USER_END;
3502         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3503         gate_vma.vm_page_prot = __P101;
3504         /*
3505          * Make sure the vDSO gets into every core dump.
3506          * Dumping its contents makes post-mortem fully interpretable later
3507          * without matching up the same kernel and hardware config to see
3508          * what PC values meant.
3509          */
3510         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3511         return 0;
3512 }
3513 __initcall(gate_vma_init);
3514 #endif
3515
3516 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3517 {
3518 #ifdef AT_SYSINFO_EHDR
3519         return &gate_vma;
3520 #else
3521         return NULL;
3522 #endif
3523 }
3524
3525 int in_gate_area_no_mm(unsigned long addr)
3526 {
3527 #ifdef AT_SYSINFO_EHDR
3528         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3529                 return 1;
3530 #endif
3531         return 0;
3532 }
3533
3534 #endif  /* __HAVE_ARCH_GATE_AREA */
3535
3536 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3537                 pte_t **ptepp, spinlock_t **ptlp)
3538 {
3539         pgd_t *pgd;
3540         pud_t *pud;
3541         pmd_t *pmd;
3542         pte_t *ptep;
3543
3544         pgd = pgd_offset(mm, address);
3545         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3546                 goto out;
3547
3548         pud = pud_offset(pgd, address);
3549         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3550                 goto out;
3551
3552         pmd = pmd_offset(pud, address);
3553         VM_BUG_ON(pmd_trans_huge(*pmd));
3554         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3555                 goto out;
3556
3557         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3558         if (pmd_huge(*pmd))
3559                 goto out;
3560
3561         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3562         if (!ptep)
3563                 goto out;
3564         if (!pte_present(*ptep))
3565                 goto unlock;
3566         *ptepp = ptep;
3567         return 0;
3568 unlock:
3569         pte_unmap_unlock(ptep, *ptlp);
3570 out:
3571         return -EINVAL;
3572 }
3573
3574 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3575                              pte_t **ptepp, spinlock_t **ptlp)
3576 {
3577         int res;
3578
3579         /* (void) is needed to make gcc happy */
3580         (void) __cond_lock(*ptlp,
3581                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3582         return res;
3583 }
3584
3585 /**
3586  * follow_pfn - look up PFN at a user virtual address
3587  * @vma: memory mapping
3588  * @address: user virtual address
3589  * @pfn: location to store found PFN
3590  *
3591  * Only IO mappings and raw PFN mappings are allowed.
3592  *
3593  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3594  */
3595 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3596         unsigned long *pfn)
3597 {
3598         int ret = -EINVAL;
3599         spinlock_t *ptl;
3600         pte_t *ptep;
3601
3602         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3603                 return ret;
3604
3605         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3606         if (ret)
3607                 return ret;
3608         *pfn = pte_pfn(*ptep);
3609         pte_unmap_unlock(ptep, ptl);
3610         return 0;
3611 }
3612 EXPORT_SYMBOL(follow_pfn);
3613
3614 #ifdef CONFIG_HAVE_IOREMAP_PROT
3615 int follow_phys(struct vm_area_struct *vma,
3616                 unsigned long address, unsigned int flags,
3617                 unsigned long *prot, resource_size_t *phys)
3618 {
3619         int ret = -EINVAL;
3620         pte_t *ptep, pte;
3621         spinlock_t *ptl;
3622
3623         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3624                 goto out;
3625
3626         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3627                 goto out;
3628         pte = *ptep;
3629
3630         if ((flags & FOLL_WRITE) && !pte_write(pte))
3631                 goto unlock;
3632
3633         *prot = pgprot_val(pte_pgprot(pte));
3634         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3635
3636         ret = 0;
3637 unlock:
3638         pte_unmap_unlock(ptep, ptl);
3639 out:
3640         return ret;
3641 }
3642
3643 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3644                         void *buf, int len, int write)
3645 {
3646         resource_size_t phys_addr;
3647         unsigned long prot = 0;
3648         void __iomem *maddr;
3649         int offset = addr & (PAGE_SIZE-1);
3650
3651         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3652                 return -EINVAL;
3653
3654         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3655         if (write)
3656                 memcpy_toio(maddr + offset, buf, len);
3657         else
3658                 memcpy_fromio(buf, maddr + offset, len);
3659         iounmap(maddr);
3660
3661         return len;
3662 }
3663 #endif
3664
3665 /*
3666  * Access another process' address space as given in mm.  If non-NULL, use the
3667  * given task for page fault accounting.
3668  */
3669 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3670                 unsigned long addr, void *buf, int len, int write)
3671 {
3672         struct vm_area_struct *vma;
3673         void *old_buf = buf;
3674
3675         down_read(&mm->mmap_sem);
3676         /* ignore errors, just check how much was successfully transferred */
3677         while (len) {
3678                 int bytes, ret, offset;
3679                 void *maddr;
3680                 struct page *page = NULL;
3681
3682                 ret = get_user_pages(tsk, mm, addr, 1,
3683                                 write, 1, &page, &vma);
3684                 if (ret <= 0) {
3685                         /*
3686                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3687                          * we can access using slightly different code.
3688                          */
3689 #ifdef CONFIG_HAVE_IOREMAP_PROT
3690                         vma = find_vma(mm, addr);
3691                         if (!vma)
3692                                 break;
3693                         if (vma->vm_ops && vma->vm_ops->access)
3694                                 ret = vma->vm_ops->access(vma, addr, buf,
3695                                                           len, write);
3696                         if (ret <= 0)
3697 #endif
3698                                 break;
3699                         bytes = ret;
3700                 } else {
3701                         bytes = len;
3702                         offset = addr & (PAGE_SIZE-1);
3703                         if (bytes > PAGE_SIZE-offset)
3704                                 bytes = PAGE_SIZE-offset;
3705
3706                         maddr = kmap(page);
3707                         if (write) {
3708                                 copy_to_user_page(vma, page, addr,
3709                                                   maddr + offset, buf, bytes);
3710                                 set_page_dirty_lock(page);
3711                         } else {
3712                                 copy_from_user_page(vma, page, addr,
3713                                                     buf, maddr + offset, bytes);
3714                         }
3715                         kunmap(page);
3716                         page_cache_release(page);
3717                 }
3718                 len -= bytes;
3719                 buf += bytes;
3720                 addr += bytes;
3721         }
3722         up_read(&mm->mmap_sem);
3723
3724         return buf - old_buf;
3725 }
3726
3727 /**
3728  * access_remote_vm - access another process' address space
3729  * @mm:         the mm_struct of the target address space
3730  * @addr:       start address to access
3731  * @buf:        source or destination buffer
3732  * @len:        number of bytes to transfer
3733  * @write:      whether the access is a write
3734  *
3735  * The caller must hold a reference on @mm.
3736  */
3737 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3738                 void *buf, int len, int write)
3739 {
3740         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3741 }
3742
3743 /*
3744  * Access another process' address space.
3745  * Source/target buffer must be kernel space,
3746  * Do not walk the page table directly, use get_user_pages
3747  */
3748 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3749                 void *buf, int len, int write)
3750 {
3751         struct mm_struct *mm;
3752         int ret;
3753
3754         mm = get_task_mm(tsk);
3755         if (!mm)
3756                 return 0;
3757
3758         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3759         mmput(mm);
3760
3761         return ret;
3762 }
3763
3764 /*
3765  * Print the name of a VMA.
3766  */
3767 void print_vma_addr(char *prefix, unsigned long ip)
3768 {
3769         struct mm_struct *mm = current->mm;
3770         struct vm_area_struct *vma;
3771
3772         /*
3773          * Do not print if we are in atomic
3774          * contexts (in exception stacks, etc.):
3775          */
3776         if (preempt_count())
3777                 return;
3778
3779         down_read(&mm->mmap_sem);
3780         vma = find_vma(mm, ip);
3781         if (vma && vma->vm_file) {
3782                 struct file *f = vma->vm_file;
3783                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3784                 if (buf) {
3785                         char *p, *s;
3786
3787                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3788                         if (IS_ERR(p))
3789                                 p = "?";
3790                         s = strrchr(p, '/');
3791                         if (s)
3792                                 p = s+1;
3793                         printk("%s%s[%lx+%lx]", prefix, p,
3794                                         vma->vm_start,
3795                                         vma->vm_end - vma->vm_start);
3796                         free_page((unsigned long)buf);
3797                 }
3798         }
3799         up_read(&current->mm->mmap_sem);
3800 }
3801
3802 #ifdef CONFIG_PROVE_LOCKING
3803 void might_fault(void)
3804 {
3805         /*
3806          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3807          * holding the mmap_sem, this is safe because kernel memory doesn't
3808          * get paged out, therefore we'll never actually fault, and the
3809          * below annotations will generate false positives.
3810          */
3811         if (segment_eq(get_fs(), KERNEL_DS))
3812                 return;
3813
3814         might_sleep();
3815         /*
3816          * it would be nicer only to annotate paths which are not under
3817          * pagefault_disable, however that requires a larger audit and
3818          * providing helpers like get_user_atomic.
3819          */
3820         if (!in_atomic() && current->mm)
3821                 might_lock_read(&current->mm->mmap_sem);
3822 }
3823 EXPORT_SYMBOL(might_fault);
3824 #endif
3825
3826 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3827 static void clear_gigantic_page(struct page *page,
3828                                 unsigned long addr,
3829                                 unsigned int pages_per_huge_page)
3830 {
3831         int i;
3832         struct page *p = page;
3833
3834         might_sleep();
3835         for (i = 0; i < pages_per_huge_page;
3836              i++, p = mem_map_next(p, page, i)) {
3837                 cond_resched();
3838                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3839         }
3840 }
3841 void clear_huge_page(struct page *page,
3842                      unsigned long addr, unsigned int pages_per_huge_page)
3843 {
3844         int i;
3845
3846         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3847                 clear_gigantic_page(page, addr, pages_per_huge_page);
3848                 return;
3849         }
3850
3851         might_sleep();
3852         for (i = 0; i < pages_per_huge_page; i++) {
3853                 cond_resched();
3854                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3855         }
3856 }
3857
3858 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3859                                     unsigned long addr,
3860                                     struct vm_area_struct *vma,
3861                                     unsigned int pages_per_huge_page)
3862 {
3863         int i;
3864         struct page *dst_base = dst;
3865         struct page *src_base = src;
3866
3867         for (i = 0; i < pages_per_huge_page; ) {
3868                 cond_resched();
3869                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3870
3871                 i++;
3872                 dst = mem_map_next(dst, dst_base, i);
3873                 src = mem_map_next(src, src_base, i);
3874         }
3875 }
3876
3877 void copy_user_huge_page(struct page *dst, struct page *src,
3878                          unsigned long addr, struct vm_area_struct *vma,
3879                          unsigned int pages_per_huge_page)
3880 {
3881         int i;
3882
3883         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3884                 copy_user_gigantic_page(dst, src, addr, vma,
3885                                         pages_per_huge_page);
3886                 return;
3887         }
3888
3889         might_sleep();
3890         for (i = 0; i < pages_per_huge_page; i++) {
3891                 cond_resched();
3892                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3893         }
3894 }
3895 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */