thp: mprotect: transparent huge page support
[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                         err = copy_huge_pmd(dst_mm, src_mm,
808                                             dst_pmd, src_pmd, addr, vma);
809                         if (err == -ENOMEM)
810                                 return -ENOMEM;
811                         if (!err)
812                                 continue;
813                         /* fall through */
814                 }
815                 if (pmd_none_or_clear_bad(src_pmd))
816                         continue;
817                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
818                                                 vma, addr, next))
819                         return -ENOMEM;
820         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
821         return 0;
822 }
823
824 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
825                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
826                 unsigned long addr, unsigned long end)
827 {
828         pud_t *src_pud, *dst_pud;
829         unsigned long next;
830
831         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
832         if (!dst_pud)
833                 return -ENOMEM;
834         src_pud = pud_offset(src_pgd, addr);
835         do {
836                 next = pud_addr_end(addr, end);
837                 if (pud_none_or_clear_bad(src_pud))
838                         continue;
839                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
840                                                 vma, addr, next))
841                         return -ENOMEM;
842         } while (dst_pud++, src_pud++, addr = next, addr != end);
843         return 0;
844 }
845
846 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
847                 struct vm_area_struct *vma)
848 {
849         pgd_t *src_pgd, *dst_pgd;
850         unsigned long next;
851         unsigned long addr = vma->vm_start;
852         unsigned long end = vma->vm_end;
853         int ret;
854
855         /*
856          * Don't copy ptes where a page fault will fill them correctly.
857          * Fork becomes much lighter when there are big shared or private
858          * readonly mappings. The tradeoff is that copy_page_range is more
859          * efficient than faulting.
860          */
861         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
862                 if (!vma->anon_vma)
863                         return 0;
864         }
865
866         if (is_vm_hugetlb_page(vma))
867                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
868
869         if (unlikely(is_pfn_mapping(vma))) {
870                 /*
871                  * We do not free on error cases below as remove_vma
872                  * gets called on error from higher level routine
873                  */
874                 ret = track_pfn_vma_copy(vma);
875                 if (ret)
876                         return ret;
877         }
878
879         /*
880          * We need to invalidate the secondary MMU mappings only when
881          * there could be a permission downgrade on the ptes of the
882          * parent mm. And a permission downgrade will only happen if
883          * is_cow_mapping() returns true.
884          */
885         if (is_cow_mapping(vma->vm_flags))
886                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
887
888         ret = 0;
889         dst_pgd = pgd_offset(dst_mm, addr);
890         src_pgd = pgd_offset(src_mm, addr);
891         do {
892                 next = pgd_addr_end(addr, end);
893                 if (pgd_none_or_clear_bad(src_pgd))
894                         continue;
895                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
896                                             vma, addr, next))) {
897                         ret = -ENOMEM;
898                         break;
899                 }
900         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
901
902         if (is_cow_mapping(vma->vm_flags))
903                 mmu_notifier_invalidate_range_end(src_mm,
904                                                   vma->vm_start, end);
905         return ret;
906 }
907
908 static unsigned long zap_pte_range(struct mmu_gather *tlb,
909                                 struct vm_area_struct *vma, pmd_t *pmd,
910                                 unsigned long addr, unsigned long end,
911                                 long *zap_work, struct zap_details *details)
912 {
913         struct mm_struct *mm = tlb->mm;
914         pte_t *pte;
915         spinlock_t *ptl;
916         int rss[NR_MM_COUNTERS];
917
918         init_rss_vec(rss);
919
920         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
921         arch_enter_lazy_mmu_mode();
922         do {
923                 pte_t ptent = *pte;
924                 if (pte_none(ptent)) {
925                         (*zap_work)--;
926                         continue;
927                 }
928
929                 (*zap_work) -= PAGE_SIZE;
930
931                 if (pte_present(ptent)) {
932                         struct page *page;
933
934                         page = vm_normal_page(vma, addr, ptent);
935                         if (unlikely(details) && page) {
936                                 /*
937                                  * unmap_shared_mapping_pages() wants to
938                                  * invalidate cache without truncating:
939                                  * unmap shared but keep private pages.
940                                  */
941                                 if (details->check_mapping &&
942                                     details->check_mapping != page->mapping)
943                                         continue;
944                                 /*
945                                  * Each page->index must be checked when
946                                  * invalidating or truncating nonlinear.
947                                  */
948                                 if (details->nonlinear_vma &&
949                                     (page->index < details->first_index ||
950                                      page->index > details->last_index))
951                                         continue;
952                         }
953                         ptent = ptep_get_and_clear_full(mm, addr, pte,
954                                                         tlb->fullmm);
955                         tlb_remove_tlb_entry(tlb, pte, addr);
956                         if (unlikely(!page))
957                                 continue;
958                         if (unlikely(details) && details->nonlinear_vma
959                             && linear_page_index(details->nonlinear_vma,
960                                                 addr) != page->index)
961                                 set_pte_at(mm, addr, pte,
962                                            pgoff_to_pte(page->index));
963                         if (PageAnon(page))
964                                 rss[MM_ANONPAGES]--;
965                         else {
966                                 if (pte_dirty(ptent))
967                                         set_page_dirty(page);
968                                 if (pte_young(ptent) &&
969                                     likely(!VM_SequentialReadHint(vma)))
970                                         mark_page_accessed(page);
971                                 rss[MM_FILEPAGES]--;
972                         }
973                         page_remove_rmap(page);
974                         if (unlikely(page_mapcount(page) < 0))
975                                 print_bad_pte(vma, addr, ptent, page);
976                         tlb_remove_page(tlb, page);
977                         continue;
978                 }
979                 /*
980                  * If details->check_mapping, we leave swap entries;
981                  * if details->nonlinear_vma, we leave file entries.
982                  */
983                 if (unlikely(details))
984                         continue;
985                 if (pte_file(ptent)) {
986                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
987                                 print_bad_pte(vma, addr, ptent, NULL);
988                 } else {
989                         swp_entry_t entry = pte_to_swp_entry(ptent);
990
991                         if (!non_swap_entry(entry))
992                                 rss[MM_SWAPENTS]--;
993                         if (unlikely(!free_swap_and_cache(entry)))
994                                 print_bad_pte(vma, addr, ptent, NULL);
995                 }
996                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
997         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
998
999         add_mm_rss_vec(mm, rss);
1000         arch_leave_lazy_mmu_mode();
1001         pte_unmap_unlock(pte - 1, ptl);
1002
1003         return addr;
1004 }
1005
1006 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1007                                 struct vm_area_struct *vma, pud_t *pud,
1008                                 unsigned long addr, unsigned long end,
1009                                 long *zap_work, struct zap_details *details)
1010 {
1011         pmd_t *pmd;
1012         unsigned long next;
1013
1014         pmd = pmd_offset(pud, addr);
1015         do {
1016                 next = pmd_addr_end(addr, end);
1017                 if (pmd_trans_huge(*pmd)) {
1018                         if (next-addr != HPAGE_PMD_SIZE)
1019                                 split_huge_page_pmd(vma->vm_mm, pmd);
1020                         else if (zap_huge_pmd(tlb, vma, pmd)) {
1021                                 (*zap_work)--;
1022                                 continue;
1023                         }
1024                         /* fall through */
1025                 }
1026                 if (pmd_none_or_clear_bad(pmd)) {
1027                         (*zap_work)--;
1028                         continue;
1029                 }
1030                 next = zap_pte_range(tlb, vma, pmd, addr, next,
1031                                                 zap_work, details);
1032         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
1033
1034         return addr;
1035 }
1036
1037 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1038                                 struct vm_area_struct *vma, pgd_t *pgd,
1039                                 unsigned long addr, unsigned long end,
1040                                 long *zap_work, struct zap_details *details)
1041 {
1042         pud_t *pud;
1043         unsigned long next;
1044
1045         pud = pud_offset(pgd, addr);
1046         do {
1047                 next = pud_addr_end(addr, end);
1048                 if (pud_none_or_clear_bad(pud)) {
1049                         (*zap_work)--;
1050                         continue;
1051                 }
1052                 next = zap_pmd_range(tlb, vma, pud, addr, next,
1053                                                 zap_work, details);
1054         } while (pud++, addr = next, (addr != end && *zap_work > 0));
1055
1056         return addr;
1057 }
1058
1059 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1060                                 struct vm_area_struct *vma,
1061                                 unsigned long addr, unsigned long end,
1062                                 long *zap_work, struct zap_details *details)
1063 {
1064         pgd_t *pgd;
1065         unsigned long next;
1066
1067         if (details && !details->check_mapping && !details->nonlinear_vma)
1068                 details = NULL;
1069
1070         BUG_ON(addr >= end);
1071         mem_cgroup_uncharge_start();
1072         tlb_start_vma(tlb, vma);
1073         pgd = pgd_offset(vma->vm_mm, addr);
1074         do {
1075                 next = pgd_addr_end(addr, end);
1076                 if (pgd_none_or_clear_bad(pgd)) {
1077                         (*zap_work)--;
1078                         continue;
1079                 }
1080                 next = zap_pud_range(tlb, vma, pgd, addr, next,
1081                                                 zap_work, details);
1082         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
1083         tlb_end_vma(tlb, vma);
1084         mem_cgroup_uncharge_end();
1085
1086         return addr;
1087 }
1088
1089 #ifdef CONFIG_PREEMPT
1090 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1091 #else
1092 /* No preempt: go for improved straight-line efficiency */
1093 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1094 #endif
1095
1096 /**
1097  * unmap_vmas - unmap a range of memory covered by a list of vma's
1098  * @tlbp: address of the caller's struct mmu_gather
1099  * @vma: the starting vma
1100  * @start_addr: virtual address at which to start unmapping
1101  * @end_addr: virtual address at which to end unmapping
1102  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1103  * @details: details of nonlinear truncation or shared cache invalidation
1104  *
1105  * Returns the end address of the unmapping (restart addr if interrupted).
1106  *
1107  * Unmap all pages in the vma list.
1108  *
1109  * We aim to not hold locks for too long (for scheduling latency reasons).
1110  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
1111  * return the ending mmu_gather to the caller.
1112  *
1113  * Only addresses between `start' and `end' will be unmapped.
1114  *
1115  * The VMA list must be sorted in ascending virtual address order.
1116  *
1117  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1118  * range after unmap_vmas() returns.  So the only responsibility here is to
1119  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1120  * drops the lock and schedules.
1121  */
1122 unsigned long unmap_vmas(struct mmu_gather **tlbp,
1123                 struct vm_area_struct *vma, unsigned long start_addr,
1124                 unsigned long end_addr, unsigned long *nr_accounted,
1125                 struct zap_details *details)
1126 {
1127         long zap_work = ZAP_BLOCK_SIZE;
1128         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
1129         int tlb_start_valid = 0;
1130         unsigned long start = start_addr;
1131         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1132         int fullmm = (*tlbp)->fullmm;
1133         struct mm_struct *mm = vma->vm_mm;
1134
1135         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1136         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1137                 unsigned long end;
1138
1139                 start = max(vma->vm_start, start_addr);
1140                 if (start >= vma->vm_end)
1141                         continue;
1142                 end = min(vma->vm_end, end_addr);
1143                 if (end <= vma->vm_start)
1144                         continue;
1145
1146                 if (vma->vm_flags & VM_ACCOUNT)
1147                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1148
1149                 if (unlikely(is_pfn_mapping(vma)))
1150                         untrack_pfn_vma(vma, 0, 0);
1151
1152                 while (start != end) {
1153                         if (!tlb_start_valid) {
1154                                 tlb_start = start;
1155                                 tlb_start_valid = 1;
1156                         }
1157
1158                         if (unlikely(is_vm_hugetlb_page(vma))) {
1159                                 /*
1160                                  * It is undesirable to test vma->vm_file as it
1161                                  * should be non-null for valid hugetlb area.
1162                                  * However, vm_file will be NULL in the error
1163                                  * cleanup path of do_mmap_pgoff. When
1164                                  * hugetlbfs ->mmap method fails,
1165                                  * do_mmap_pgoff() nullifies vma->vm_file
1166                                  * before calling this function to clean up.
1167                                  * Since no pte has actually been setup, it is
1168                                  * safe to do nothing in this case.
1169                                  */
1170                                 if (vma->vm_file) {
1171                                         unmap_hugepage_range(vma, start, end, NULL);
1172                                         zap_work -= (end - start) /
1173                                         pages_per_huge_page(hstate_vma(vma));
1174                                 }
1175
1176                                 start = end;
1177                         } else
1178                                 start = unmap_page_range(*tlbp, vma,
1179                                                 start, end, &zap_work, details);
1180
1181                         if (zap_work > 0) {
1182                                 BUG_ON(start != end);
1183                                 break;
1184                         }
1185
1186                         tlb_finish_mmu(*tlbp, tlb_start, start);
1187
1188                         if (need_resched() ||
1189                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1190                                 if (i_mmap_lock) {
1191                                         *tlbp = NULL;
1192                                         goto out;
1193                                 }
1194                                 cond_resched();
1195                         }
1196
1197                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1198                         tlb_start_valid = 0;
1199                         zap_work = ZAP_BLOCK_SIZE;
1200                 }
1201         }
1202 out:
1203         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1204         return start;   /* which is now the end (or restart) address */
1205 }
1206
1207 /**
1208  * zap_page_range - remove user pages in a given range
1209  * @vma: vm_area_struct holding the applicable pages
1210  * @address: starting address of pages to zap
1211  * @size: number of bytes to zap
1212  * @details: details of nonlinear truncation or shared cache invalidation
1213  */
1214 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1215                 unsigned long size, struct zap_details *details)
1216 {
1217         struct mm_struct *mm = vma->vm_mm;
1218         struct mmu_gather *tlb;
1219         unsigned long end = address + size;
1220         unsigned long nr_accounted = 0;
1221
1222         lru_add_drain();
1223         tlb = tlb_gather_mmu(mm, 0);
1224         update_hiwater_rss(mm);
1225         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1226         if (tlb)
1227                 tlb_finish_mmu(tlb, address, end);
1228         return end;
1229 }
1230
1231 /**
1232  * zap_vma_ptes - remove ptes mapping the vma
1233  * @vma: vm_area_struct holding ptes to be zapped
1234  * @address: starting address of pages to zap
1235  * @size: number of bytes to zap
1236  *
1237  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1238  *
1239  * The entire address range must be fully contained within the vma.
1240  *
1241  * Returns 0 if successful.
1242  */
1243 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1244                 unsigned long size)
1245 {
1246         if (address < vma->vm_start || address + size > vma->vm_end ||
1247                         !(vma->vm_flags & VM_PFNMAP))
1248                 return -1;
1249         zap_page_range(vma, address, size, NULL);
1250         return 0;
1251 }
1252 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1253
1254 /**
1255  * follow_page - look up a page descriptor from a user-virtual address
1256  * @vma: vm_area_struct mapping @address
1257  * @address: virtual address to look up
1258  * @flags: flags modifying lookup behaviour
1259  *
1260  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1261  *
1262  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1263  * an error pointer if there is a mapping to something not represented
1264  * by a page descriptor (see also vm_normal_page()).
1265  */
1266 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1267                         unsigned int flags)
1268 {
1269         pgd_t *pgd;
1270         pud_t *pud;
1271         pmd_t *pmd;
1272         pte_t *ptep, pte;
1273         spinlock_t *ptl;
1274         struct page *page;
1275         struct mm_struct *mm = vma->vm_mm;
1276
1277         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1278         if (!IS_ERR(page)) {
1279                 BUG_ON(flags & FOLL_GET);
1280                 goto out;
1281         }
1282
1283         page = NULL;
1284         pgd = pgd_offset(mm, address);
1285         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1286                 goto no_page_table;
1287
1288         pud = pud_offset(pgd, address);
1289         if (pud_none(*pud))
1290                 goto no_page_table;
1291         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1292                 BUG_ON(flags & FOLL_GET);
1293                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1294                 goto out;
1295         }
1296         if (unlikely(pud_bad(*pud)))
1297                 goto no_page_table;
1298
1299         pmd = pmd_offset(pud, address);
1300         if (pmd_none(*pmd))
1301                 goto no_page_table;
1302         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1303                 BUG_ON(flags & FOLL_GET);
1304                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1305                 goto out;
1306         }
1307         if (pmd_trans_huge(*pmd)) {
1308                 if (flags & FOLL_SPLIT) {
1309                         split_huge_page_pmd(mm, pmd);
1310                         goto split_fallthrough;
1311                 }
1312                 spin_lock(&mm->page_table_lock);
1313                 if (likely(pmd_trans_huge(*pmd))) {
1314                         if (unlikely(pmd_trans_splitting(*pmd))) {
1315                                 spin_unlock(&mm->page_table_lock);
1316                                 wait_split_huge_page(vma->anon_vma, pmd);
1317                         } else {
1318                                 page = follow_trans_huge_pmd(mm, address,
1319                                                              pmd, flags);
1320                                 spin_unlock(&mm->page_table_lock);
1321                                 goto out;
1322                         }
1323                 } else
1324                         spin_unlock(&mm->page_table_lock);
1325                 /* fall through */
1326         }
1327 split_fallthrough:
1328         if (unlikely(pmd_bad(*pmd)))
1329                 goto no_page_table;
1330
1331         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1332
1333         pte = *ptep;
1334         if (!pte_present(pte))
1335                 goto no_page;
1336         if ((flags & FOLL_WRITE) && !pte_write(pte))
1337                 goto unlock;
1338
1339         page = vm_normal_page(vma, address, pte);
1340         if (unlikely(!page)) {
1341                 if ((flags & FOLL_DUMP) ||
1342                     !is_zero_pfn(pte_pfn(pte)))
1343                         goto bad_page;
1344                 page = pte_page(pte);
1345         }
1346
1347         if (flags & FOLL_GET)
1348                 get_page(page);
1349         if (flags & FOLL_TOUCH) {
1350                 if ((flags & FOLL_WRITE) &&
1351                     !pte_dirty(pte) && !PageDirty(page))
1352                         set_page_dirty(page);
1353                 /*
1354                  * pte_mkyoung() would be more correct here, but atomic care
1355                  * is needed to avoid losing the dirty bit: it is easier to use
1356                  * mark_page_accessed().
1357                  */
1358                 mark_page_accessed(page);
1359         }
1360         if (flags & FOLL_MLOCK) {
1361                 /*
1362                  * The preliminary mapping check is mainly to avoid the
1363                  * pointless overhead of lock_page on the ZERO_PAGE
1364                  * which might bounce very badly if there is contention.
1365                  *
1366                  * If the page is already locked, we don't need to
1367                  * handle it now - vmscan will handle it later if and
1368                  * when it attempts to reclaim the page.
1369                  */
1370                 if (page->mapping && trylock_page(page)) {
1371                         lru_add_drain();  /* push cached pages to LRU */
1372                         /*
1373                          * Because we lock page here and migration is
1374                          * blocked by the pte's page reference, we need
1375                          * only check for file-cache page truncation.
1376                          */
1377                         if (page->mapping)
1378                                 mlock_vma_page(page);
1379                         unlock_page(page);
1380                 }
1381         }
1382 unlock:
1383         pte_unmap_unlock(ptep, ptl);
1384 out:
1385         return page;
1386
1387 bad_page:
1388         pte_unmap_unlock(ptep, ptl);
1389         return ERR_PTR(-EFAULT);
1390
1391 no_page:
1392         pte_unmap_unlock(ptep, ptl);
1393         if (!pte_none(pte))
1394                 return page;
1395
1396 no_page_table:
1397         /*
1398          * When core dumping an enormous anonymous area that nobody
1399          * has touched so far, we don't want to allocate unnecessary pages or
1400          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1401          * then get_dump_page() will return NULL to leave a hole in the dump.
1402          * But we can only make this optimization where a hole would surely
1403          * be zero-filled if handle_mm_fault() actually did handle it.
1404          */
1405         if ((flags & FOLL_DUMP) &&
1406             (!vma->vm_ops || !vma->vm_ops->fault))
1407                 return ERR_PTR(-EFAULT);
1408         return page;
1409 }
1410
1411 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1412                      unsigned long start, int nr_pages, unsigned int gup_flags,
1413                      struct page **pages, struct vm_area_struct **vmas,
1414                      int *nonblocking)
1415 {
1416         int i;
1417         unsigned long vm_flags;
1418
1419         if (nr_pages <= 0)
1420                 return 0;
1421
1422         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1423
1424         /* 
1425          * Require read or write permissions.
1426          * If FOLL_FORCE is set, we only require the "MAY" flags.
1427          */
1428         vm_flags  = (gup_flags & FOLL_WRITE) ?
1429                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1430         vm_flags &= (gup_flags & FOLL_FORCE) ?
1431                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1432         i = 0;
1433
1434         do {
1435                 struct vm_area_struct *vma;
1436
1437                 vma = find_extend_vma(mm, start);
1438                 if (!vma && in_gate_area(tsk, start)) {
1439                         unsigned long pg = start & PAGE_MASK;
1440                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1441                         pgd_t *pgd;
1442                         pud_t *pud;
1443                         pmd_t *pmd;
1444                         pte_t *pte;
1445
1446                         /* user gate pages are read-only */
1447                         if (gup_flags & FOLL_WRITE)
1448                                 return i ? : -EFAULT;
1449                         if (pg > TASK_SIZE)
1450                                 pgd = pgd_offset_k(pg);
1451                         else
1452                                 pgd = pgd_offset_gate(mm, pg);
1453                         BUG_ON(pgd_none(*pgd));
1454                         pud = pud_offset(pgd, pg);
1455                         BUG_ON(pud_none(*pud));
1456                         pmd = pmd_offset(pud, pg);
1457                         if (pmd_none(*pmd))
1458                                 return i ? : -EFAULT;
1459                         VM_BUG_ON(pmd_trans_huge(*pmd));
1460                         pte = pte_offset_map(pmd, pg);
1461                         if (pte_none(*pte)) {
1462                                 pte_unmap(pte);
1463                                 return i ? : -EFAULT;
1464                         }
1465                         if (pages) {
1466                                 struct page *page;
1467
1468                                 page = vm_normal_page(gate_vma, start, *pte);
1469                                 if (!page) {
1470                                         if (!(gup_flags & FOLL_DUMP) &&
1471                                              is_zero_pfn(pte_pfn(*pte)))
1472                                                 page = pte_page(*pte);
1473                                         else {
1474                                                 pte_unmap(pte);
1475                                                 return i ? : -EFAULT;
1476                                         }
1477                                 }
1478                                 pages[i] = page;
1479                                 get_page(page);
1480                         }
1481                         pte_unmap(pte);
1482                         if (vmas)
1483                                 vmas[i] = gate_vma;
1484                         i++;
1485                         start += PAGE_SIZE;
1486                         nr_pages--;
1487                         continue;
1488                 }
1489
1490                 if (!vma ||
1491                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1492                     !(vm_flags & vma->vm_flags))
1493                         return i ? : -EFAULT;
1494
1495                 if (is_vm_hugetlb_page(vma)) {
1496                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1497                                         &start, &nr_pages, i, gup_flags);
1498                         continue;
1499                 }
1500
1501                 do {
1502                         struct page *page;
1503                         unsigned int foll_flags = gup_flags;
1504
1505                         /*
1506                          * If we have a pending SIGKILL, don't keep faulting
1507                          * pages and potentially allocating memory.
1508                          */
1509                         if (unlikely(fatal_signal_pending(current)))
1510                                 return i ? i : -ERESTARTSYS;
1511
1512                         cond_resched();
1513                         while (!(page = follow_page(vma, start, foll_flags))) {
1514                                 int ret;
1515                                 unsigned int fault_flags = 0;
1516
1517                                 if (foll_flags & FOLL_WRITE)
1518                                         fault_flags |= FAULT_FLAG_WRITE;
1519                                 if (nonblocking)
1520                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1521
1522                                 ret = handle_mm_fault(mm, vma, start,
1523                                                         fault_flags);
1524
1525                                 if (ret & VM_FAULT_ERROR) {
1526                                         if (ret & VM_FAULT_OOM)
1527                                                 return i ? i : -ENOMEM;
1528                                         if (ret &
1529                                             (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE|
1530                                              VM_FAULT_SIGBUS))
1531                                                 return i ? i : -EFAULT;
1532                                         BUG();
1533                                 }
1534                                 if (ret & VM_FAULT_MAJOR)
1535                                         tsk->maj_flt++;
1536                                 else
1537                                         tsk->min_flt++;
1538
1539                                 if (ret & VM_FAULT_RETRY) {
1540                                         *nonblocking = 0;
1541                                         return i;
1542                                 }
1543
1544                                 /*
1545                                  * The VM_FAULT_WRITE bit tells us that
1546                                  * do_wp_page has broken COW when necessary,
1547                                  * even if maybe_mkwrite decided not to set
1548                                  * pte_write. We can thus safely do subsequent
1549                                  * page lookups as if they were reads. But only
1550                                  * do so when looping for pte_write is futile:
1551                                  * in some cases userspace may also be wanting
1552                                  * to write to the gotten user page, which a
1553                                  * read fault here might prevent (a readonly
1554                                  * page might get reCOWed by userspace write).
1555                                  */
1556                                 if ((ret & VM_FAULT_WRITE) &&
1557                                     !(vma->vm_flags & VM_WRITE))
1558                                         foll_flags &= ~FOLL_WRITE;
1559
1560                                 cond_resched();
1561                         }
1562                         if (IS_ERR(page))
1563                                 return i ? i : PTR_ERR(page);
1564                         if (pages) {
1565                                 pages[i] = page;
1566
1567                                 flush_anon_page(vma, page, start);
1568                                 flush_dcache_page(page);
1569                         }
1570                         if (vmas)
1571                                 vmas[i] = vma;
1572                         i++;
1573                         start += PAGE_SIZE;
1574                         nr_pages--;
1575                 } while (nr_pages && start < vma->vm_end);
1576         } while (nr_pages);
1577         return i;
1578 }
1579
1580 /**
1581  * get_user_pages() - pin user pages in memory
1582  * @tsk:        task_struct of target task
1583  * @mm:         mm_struct of target mm
1584  * @start:      starting user address
1585  * @nr_pages:   number of pages from start to pin
1586  * @write:      whether pages will be written to by the caller
1587  * @force:      whether to force write access even if user mapping is
1588  *              readonly. This will result in the page being COWed even
1589  *              in MAP_SHARED mappings. You do not want this.
1590  * @pages:      array that receives pointers to the pages pinned.
1591  *              Should be at least nr_pages long. Or NULL, if caller
1592  *              only intends to ensure the pages are faulted in.
1593  * @vmas:       array of pointers to vmas corresponding to each page.
1594  *              Or NULL if the caller does not require them.
1595  *
1596  * Returns number of pages pinned. This may be fewer than the number
1597  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1598  * were pinned, returns -errno. Each page returned must be released
1599  * with a put_page() call when it is finished with. vmas will only
1600  * remain valid while mmap_sem is held.
1601  *
1602  * Must be called with mmap_sem held for read or write.
1603  *
1604  * get_user_pages walks a process's page tables and takes a reference to
1605  * each struct page that each user address corresponds to at a given
1606  * instant. That is, it takes the page that would be accessed if a user
1607  * thread accesses the given user virtual address at that instant.
1608  *
1609  * This does not guarantee that the page exists in the user mappings when
1610  * get_user_pages returns, and there may even be a completely different
1611  * page there in some cases (eg. if mmapped pagecache has been invalidated
1612  * and subsequently re faulted). However it does guarantee that the page
1613  * won't be freed completely. And mostly callers simply care that the page
1614  * contains data that was valid *at some point in time*. Typically, an IO
1615  * or similar operation cannot guarantee anything stronger anyway because
1616  * locks can't be held over the syscall boundary.
1617  *
1618  * If write=0, the page must not be written to. If the page is written to,
1619  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1620  * after the page is finished with, and before put_page is called.
1621  *
1622  * get_user_pages is typically used for fewer-copy IO operations, to get a
1623  * handle on the memory by some means other than accesses via the user virtual
1624  * addresses. The pages may be submitted for DMA to devices or accessed via
1625  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1626  * use the correct cache flushing APIs.
1627  *
1628  * See also get_user_pages_fast, for performance critical applications.
1629  */
1630 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1631                 unsigned long start, int nr_pages, int write, int force,
1632                 struct page **pages, struct vm_area_struct **vmas)
1633 {
1634         int flags = FOLL_TOUCH;
1635
1636         if (pages)
1637                 flags |= FOLL_GET;
1638         if (write)
1639                 flags |= FOLL_WRITE;
1640         if (force)
1641                 flags |= FOLL_FORCE;
1642
1643         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1644                                 NULL);
1645 }
1646 EXPORT_SYMBOL(get_user_pages);
1647
1648 /**
1649  * get_dump_page() - pin user page in memory while writing it to core dump
1650  * @addr: user address
1651  *
1652  * Returns struct page pointer of user page pinned for dump,
1653  * to be freed afterwards by page_cache_release() or put_page().
1654  *
1655  * Returns NULL on any kind of failure - a hole must then be inserted into
1656  * the corefile, to preserve alignment with its headers; and also returns
1657  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1658  * allowing a hole to be left in the corefile to save diskspace.
1659  *
1660  * Called without mmap_sem, but after all other threads have been killed.
1661  */
1662 #ifdef CONFIG_ELF_CORE
1663 struct page *get_dump_page(unsigned long addr)
1664 {
1665         struct vm_area_struct *vma;
1666         struct page *page;
1667
1668         if (__get_user_pages(current, current->mm, addr, 1,
1669                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1670                              NULL) < 1)
1671                 return NULL;
1672         flush_cache_page(vma, addr, page_to_pfn(page));
1673         return page;
1674 }
1675 #endif /* CONFIG_ELF_CORE */
1676
1677 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1678                         spinlock_t **ptl)
1679 {
1680         pgd_t * pgd = pgd_offset(mm, addr);
1681         pud_t * pud = pud_alloc(mm, pgd, addr);
1682         if (pud) {
1683                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1684                 if (pmd) {
1685                         VM_BUG_ON(pmd_trans_huge(*pmd));
1686                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1687                 }
1688         }
1689         return NULL;
1690 }
1691
1692 /*
1693  * This is the old fallback for page remapping.
1694  *
1695  * For historical reasons, it only allows reserved pages. Only
1696  * old drivers should use this, and they needed to mark their
1697  * pages reserved for the old functions anyway.
1698  */
1699 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1700                         struct page *page, pgprot_t prot)
1701 {
1702         struct mm_struct *mm = vma->vm_mm;
1703         int retval;
1704         pte_t *pte;
1705         spinlock_t *ptl;
1706
1707         retval = -EINVAL;
1708         if (PageAnon(page))
1709                 goto out;
1710         retval = -ENOMEM;
1711         flush_dcache_page(page);
1712         pte = get_locked_pte(mm, addr, &ptl);
1713         if (!pte)
1714                 goto out;
1715         retval = -EBUSY;
1716         if (!pte_none(*pte))
1717                 goto out_unlock;
1718
1719         /* Ok, finally just insert the thing.. */
1720         get_page(page);
1721         inc_mm_counter_fast(mm, MM_FILEPAGES);
1722         page_add_file_rmap(page);
1723         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1724
1725         retval = 0;
1726         pte_unmap_unlock(pte, ptl);
1727         return retval;
1728 out_unlock:
1729         pte_unmap_unlock(pte, ptl);
1730 out:
1731         return retval;
1732 }
1733
1734 /**
1735  * vm_insert_page - insert single page into user vma
1736  * @vma: user vma to map to
1737  * @addr: target user address of this page
1738  * @page: source kernel page
1739  *
1740  * This allows drivers to insert individual pages they've allocated
1741  * into a user vma.
1742  *
1743  * The page has to be a nice clean _individual_ kernel allocation.
1744  * If you allocate a compound page, you need to have marked it as
1745  * such (__GFP_COMP), or manually just split the page up yourself
1746  * (see split_page()).
1747  *
1748  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1749  * took an arbitrary page protection parameter. This doesn't allow
1750  * that. Your vma protection will have to be set up correctly, which
1751  * means that if you want a shared writable mapping, you'd better
1752  * ask for a shared writable mapping!
1753  *
1754  * The page does not need to be reserved.
1755  */
1756 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1757                         struct page *page)
1758 {
1759         if (addr < vma->vm_start || addr >= vma->vm_end)
1760                 return -EFAULT;
1761         if (!page_count(page))
1762                 return -EINVAL;
1763         vma->vm_flags |= VM_INSERTPAGE;
1764         return insert_page(vma, addr, page, vma->vm_page_prot);
1765 }
1766 EXPORT_SYMBOL(vm_insert_page);
1767
1768 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1769                         unsigned long pfn, pgprot_t prot)
1770 {
1771         struct mm_struct *mm = vma->vm_mm;
1772         int retval;
1773         pte_t *pte, entry;
1774         spinlock_t *ptl;
1775
1776         retval = -ENOMEM;
1777         pte = get_locked_pte(mm, addr, &ptl);
1778         if (!pte)
1779                 goto out;
1780         retval = -EBUSY;
1781         if (!pte_none(*pte))
1782                 goto out_unlock;
1783
1784         /* Ok, finally just insert the thing.. */
1785         entry = pte_mkspecial(pfn_pte(pfn, prot));
1786         set_pte_at(mm, addr, pte, entry);
1787         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1788
1789         retval = 0;
1790 out_unlock:
1791         pte_unmap_unlock(pte, ptl);
1792 out:
1793         return retval;
1794 }
1795
1796 /**
1797  * vm_insert_pfn - insert single pfn into user vma
1798  * @vma: user vma to map to
1799  * @addr: target user address of this page
1800  * @pfn: source kernel pfn
1801  *
1802  * Similar to vm_inert_page, this allows drivers to insert individual pages
1803  * they've allocated into a user vma. Same comments apply.
1804  *
1805  * This function should only be called from a vm_ops->fault handler, and
1806  * in that case the handler should return NULL.
1807  *
1808  * vma cannot be a COW mapping.
1809  *
1810  * As this is called only for pages that do not currently exist, we
1811  * do not need to flush old virtual caches or the TLB.
1812  */
1813 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1814                         unsigned long pfn)
1815 {
1816         int ret;
1817         pgprot_t pgprot = vma->vm_page_prot;
1818         /*
1819          * Technically, architectures with pte_special can avoid all these
1820          * restrictions (same for remap_pfn_range).  However we would like
1821          * consistency in testing and feature parity among all, so we should
1822          * try to keep these invariants in place for everybody.
1823          */
1824         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1825         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1826                                                 (VM_PFNMAP|VM_MIXEDMAP));
1827         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1828         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1829
1830         if (addr < vma->vm_start || addr >= vma->vm_end)
1831                 return -EFAULT;
1832         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1833                 return -EINVAL;
1834
1835         ret = insert_pfn(vma, addr, pfn, pgprot);
1836
1837         if (ret)
1838                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1839
1840         return ret;
1841 }
1842 EXPORT_SYMBOL(vm_insert_pfn);
1843
1844 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1845                         unsigned long pfn)
1846 {
1847         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1848
1849         if (addr < vma->vm_start || addr >= vma->vm_end)
1850                 return -EFAULT;
1851
1852         /*
1853          * If we don't have pte special, then we have to use the pfn_valid()
1854          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1855          * refcount the page if pfn_valid is true (hence insert_page rather
1856          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1857          * without pte special, it would there be refcounted as a normal page.
1858          */
1859         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1860                 struct page *page;
1861
1862                 page = pfn_to_page(pfn);
1863                 return insert_page(vma, addr, page, vma->vm_page_prot);
1864         }
1865         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1866 }
1867 EXPORT_SYMBOL(vm_insert_mixed);
1868
1869 /*
1870  * maps a range of physical memory into the requested pages. the old
1871  * mappings are removed. any references to nonexistent pages results
1872  * in null mappings (currently treated as "copy-on-access")
1873  */
1874 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1875                         unsigned long addr, unsigned long end,
1876                         unsigned long pfn, pgprot_t prot)
1877 {
1878         pte_t *pte;
1879         spinlock_t *ptl;
1880
1881         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1882         if (!pte)
1883                 return -ENOMEM;
1884         arch_enter_lazy_mmu_mode();
1885         do {
1886                 BUG_ON(!pte_none(*pte));
1887                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1888                 pfn++;
1889         } while (pte++, addr += PAGE_SIZE, addr != end);
1890         arch_leave_lazy_mmu_mode();
1891         pte_unmap_unlock(pte - 1, ptl);
1892         return 0;
1893 }
1894
1895 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1896                         unsigned long addr, unsigned long end,
1897                         unsigned long pfn, pgprot_t prot)
1898 {
1899         pmd_t *pmd;
1900         unsigned long next;
1901
1902         pfn -= addr >> PAGE_SHIFT;
1903         pmd = pmd_alloc(mm, pud, addr);
1904         if (!pmd)
1905                 return -ENOMEM;
1906         VM_BUG_ON(pmd_trans_huge(*pmd));
1907         do {
1908                 next = pmd_addr_end(addr, end);
1909                 if (remap_pte_range(mm, pmd, addr, next,
1910                                 pfn + (addr >> PAGE_SHIFT), prot))
1911                         return -ENOMEM;
1912         } while (pmd++, addr = next, addr != end);
1913         return 0;
1914 }
1915
1916 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1917                         unsigned long addr, unsigned long end,
1918                         unsigned long pfn, pgprot_t prot)
1919 {
1920         pud_t *pud;
1921         unsigned long next;
1922
1923         pfn -= addr >> PAGE_SHIFT;
1924         pud = pud_alloc(mm, pgd, addr);
1925         if (!pud)
1926                 return -ENOMEM;
1927         do {
1928                 next = pud_addr_end(addr, end);
1929                 if (remap_pmd_range(mm, pud, addr, next,
1930                                 pfn + (addr >> PAGE_SHIFT), prot))
1931                         return -ENOMEM;
1932         } while (pud++, addr = next, addr != end);
1933         return 0;
1934 }
1935
1936 /**
1937  * remap_pfn_range - remap kernel memory to userspace
1938  * @vma: user vma to map to
1939  * @addr: target user address to start at
1940  * @pfn: physical address of kernel memory
1941  * @size: size of map area
1942  * @prot: page protection flags for this mapping
1943  *
1944  *  Note: this is only safe if the mm semaphore is held when called.
1945  */
1946 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1947                     unsigned long pfn, unsigned long size, pgprot_t prot)
1948 {
1949         pgd_t *pgd;
1950         unsigned long next;
1951         unsigned long end = addr + PAGE_ALIGN(size);
1952         struct mm_struct *mm = vma->vm_mm;
1953         int err;
1954
1955         /*
1956          * Physically remapped pages are special. Tell the
1957          * rest of the world about it:
1958          *   VM_IO tells people not to look at these pages
1959          *      (accesses can have side effects).
1960          *   VM_RESERVED is specified all over the place, because
1961          *      in 2.4 it kept swapout's vma scan off this vma; but
1962          *      in 2.6 the LRU scan won't even find its pages, so this
1963          *      flag means no more than count its pages in reserved_vm,
1964          *      and omit it from core dump, even when VM_IO turned off.
1965          *   VM_PFNMAP tells the core MM that the base pages are just
1966          *      raw PFN mappings, and do not have a "struct page" associated
1967          *      with them.
1968          *
1969          * There's a horrible special case to handle copy-on-write
1970          * behaviour that some programs depend on. We mark the "original"
1971          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1972          */
1973         if (addr == vma->vm_start && end == vma->vm_end) {
1974                 vma->vm_pgoff = pfn;
1975                 vma->vm_flags |= VM_PFN_AT_MMAP;
1976         } else if (is_cow_mapping(vma->vm_flags))
1977                 return -EINVAL;
1978
1979         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1980
1981         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1982         if (err) {
1983                 /*
1984                  * To indicate that track_pfn related cleanup is not
1985                  * needed from higher level routine calling unmap_vmas
1986                  */
1987                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1988                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1989                 return -EINVAL;
1990         }
1991
1992         BUG_ON(addr >= end);
1993         pfn -= addr >> PAGE_SHIFT;
1994         pgd = pgd_offset(mm, addr);
1995         flush_cache_range(vma, addr, end);
1996         do {
1997                 next = pgd_addr_end(addr, end);
1998                 err = remap_pud_range(mm, pgd, addr, next,
1999                                 pfn + (addr >> PAGE_SHIFT), prot);
2000                 if (err)
2001                         break;
2002         } while (pgd++, addr = next, addr != end);
2003
2004         if (err)
2005                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2006
2007         return err;
2008 }
2009 EXPORT_SYMBOL(remap_pfn_range);
2010
2011 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2012                                      unsigned long addr, unsigned long end,
2013                                      pte_fn_t fn, void *data)
2014 {
2015         pte_t *pte;
2016         int err;
2017         pgtable_t token;
2018         spinlock_t *uninitialized_var(ptl);
2019
2020         pte = (mm == &init_mm) ?
2021                 pte_alloc_kernel(pmd, addr) :
2022                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2023         if (!pte)
2024                 return -ENOMEM;
2025
2026         BUG_ON(pmd_huge(*pmd));
2027
2028         arch_enter_lazy_mmu_mode();
2029
2030         token = pmd_pgtable(*pmd);
2031
2032         do {
2033                 err = fn(pte++, token, addr, data);
2034                 if (err)
2035                         break;
2036         } while (addr += PAGE_SIZE, addr != end);
2037
2038         arch_leave_lazy_mmu_mode();
2039
2040         if (mm != &init_mm)
2041                 pte_unmap_unlock(pte-1, ptl);
2042         return err;
2043 }
2044
2045 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2046                                      unsigned long addr, unsigned long end,
2047                                      pte_fn_t fn, void *data)
2048 {
2049         pmd_t *pmd;
2050         unsigned long next;
2051         int err;
2052
2053         BUG_ON(pud_huge(*pud));
2054
2055         pmd = pmd_alloc(mm, pud, addr);
2056         if (!pmd)
2057                 return -ENOMEM;
2058         do {
2059                 next = pmd_addr_end(addr, end);
2060                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2061                 if (err)
2062                         break;
2063         } while (pmd++, addr = next, addr != end);
2064         return err;
2065 }
2066
2067 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2068                                      unsigned long addr, unsigned long end,
2069                                      pte_fn_t fn, void *data)
2070 {
2071         pud_t *pud;
2072         unsigned long next;
2073         int err;
2074
2075         pud = pud_alloc(mm, pgd, addr);
2076         if (!pud)
2077                 return -ENOMEM;
2078         do {
2079                 next = pud_addr_end(addr, end);
2080                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2081                 if (err)
2082                         break;
2083         } while (pud++, addr = next, addr != end);
2084         return err;
2085 }
2086
2087 /*
2088  * Scan a region of virtual memory, filling in page tables as necessary
2089  * and calling a provided function on each leaf page table.
2090  */
2091 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2092                         unsigned long size, pte_fn_t fn, void *data)
2093 {
2094         pgd_t *pgd;
2095         unsigned long next;
2096         unsigned long end = addr + size;
2097         int err;
2098
2099         BUG_ON(addr >= end);
2100         pgd = pgd_offset(mm, addr);
2101         do {
2102                 next = pgd_addr_end(addr, end);
2103                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2104                 if (err)
2105                         break;
2106         } while (pgd++, addr = next, addr != end);
2107
2108         return err;
2109 }
2110 EXPORT_SYMBOL_GPL(apply_to_page_range);
2111
2112 /*
2113  * handle_pte_fault chooses page fault handler according to an entry
2114  * which was read non-atomically.  Before making any commitment, on
2115  * those architectures or configurations (e.g. i386 with PAE) which
2116  * might give a mix of unmatched parts, do_swap_page and do_file_page
2117  * must check under lock before unmapping the pte and proceeding
2118  * (but do_wp_page is only called after already making such a check;
2119  * and do_anonymous_page and do_no_page can safely check later on).
2120  */
2121 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2122                                 pte_t *page_table, pte_t orig_pte)
2123 {
2124         int same = 1;
2125 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2126         if (sizeof(pte_t) > sizeof(unsigned long)) {
2127                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2128                 spin_lock(ptl);
2129                 same = pte_same(*page_table, orig_pte);
2130                 spin_unlock(ptl);
2131         }
2132 #endif
2133         pte_unmap(page_table);
2134         return same;
2135 }
2136
2137 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2138 {
2139         /*
2140          * If the source page was a PFN mapping, we don't have
2141          * a "struct page" for it. We do a best-effort copy by
2142          * just copying from the original user address. If that
2143          * fails, we just zero-fill it. Live with it.
2144          */
2145         if (unlikely(!src)) {
2146                 void *kaddr = kmap_atomic(dst, KM_USER0);
2147                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2148
2149                 /*
2150                  * This really shouldn't fail, because the page is there
2151                  * in the page tables. But it might just be unreadable,
2152                  * in which case we just give up and fill the result with
2153                  * zeroes.
2154                  */
2155                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2156                         clear_page(kaddr);
2157                 kunmap_atomic(kaddr, KM_USER0);
2158                 flush_dcache_page(dst);
2159         } else
2160                 copy_user_highpage(dst, src, va, vma);
2161 }
2162
2163 /*
2164  * This routine handles present pages, when users try to write
2165  * to a shared page. It is done by copying the page to a new address
2166  * and decrementing the shared-page counter for the old page.
2167  *
2168  * Note that this routine assumes that the protection checks have been
2169  * done by the caller (the low-level page fault routine in most cases).
2170  * Thus we can safely just mark it writable once we've done any necessary
2171  * COW.
2172  *
2173  * We also mark the page dirty at this point even though the page will
2174  * change only once the write actually happens. This avoids a few races,
2175  * and potentially makes it more efficient.
2176  *
2177  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2178  * but allow concurrent faults), with pte both mapped and locked.
2179  * We return with mmap_sem still held, but pte unmapped and unlocked.
2180  */
2181 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2182                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2183                 spinlock_t *ptl, pte_t orig_pte)
2184         __releases(ptl)
2185 {
2186         struct page *old_page, *new_page;
2187         pte_t entry;
2188         int ret = 0;
2189         int page_mkwrite = 0;
2190         struct page *dirty_page = NULL;
2191
2192         old_page = vm_normal_page(vma, address, orig_pte);
2193         if (!old_page) {
2194                 /*
2195                  * VM_MIXEDMAP !pfn_valid() case
2196                  *
2197                  * We should not cow pages in a shared writeable mapping.
2198                  * Just mark the pages writable as we can't do any dirty
2199                  * accounting on raw pfn maps.
2200                  */
2201                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2202                                      (VM_WRITE|VM_SHARED))
2203                         goto reuse;
2204                 goto gotten;
2205         }
2206
2207         /*
2208          * Take out anonymous pages first, anonymous shared vmas are
2209          * not dirty accountable.
2210          */
2211         if (PageAnon(old_page) && !PageKsm(old_page)) {
2212                 if (!trylock_page(old_page)) {
2213                         page_cache_get(old_page);
2214                         pte_unmap_unlock(page_table, ptl);
2215                         lock_page(old_page);
2216                         page_table = pte_offset_map_lock(mm, pmd, address,
2217                                                          &ptl);
2218                         if (!pte_same(*page_table, orig_pte)) {
2219                                 unlock_page(old_page);
2220                                 page_cache_release(old_page);
2221                                 goto unlock;
2222                         }
2223                         page_cache_release(old_page);
2224                 }
2225                 if (reuse_swap_page(old_page)) {
2226                         /*
2227                          * The page is all ours.  Move it to our anon_vma so
2228                          * the rmap code will not search our parent or siblings.
2229                          * Protected against the rmap code by the page lock.
2230                          */
2231                         page_move_anon_rmap(old_page, vma, address);
2232                         unlock_page(old_page);
2233                         goto reuse;
2234                 }
2235                 unlock_page(old_page);
2236         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2237                                         (VM_WRITE|VM_SHARED))) {
2238                 /*
2239                  * Only catch write-faults on shared writable pages,
2240                  * read-only shared pages can get COWed by
2241                  * get_user_pages(.write=1, .force=1).
2242                  */
2243                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2244                         struct vm_fault vmf;
2245                         int tmp;
2246
2247                         vmf.virtual_address = (void __user *)(address &
2248                                                                 PAGE_MASK);
2249                         vmf.pgoff = old_page->index;
2250                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2251                         vmf.page = old_page;
2252
2253                         /*
2254                          * Notify the address space that the page is about to
2255                          * become writable so that it can prohibit this or wait
2256                          * for the page to get into an appropriate state.
2257                          *
2258                          * We do this without the lock held, so that it can
2259                          * sleep if it needs to.
2260                          */
2261                         page_cache_get(old_page);
2262                         pte_unmap_unlock(page_table, ptl);
2263
2264                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2265                         if (unlikely(tmp &
2266                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2267                                 ret = tmp;
2268                                 goto unwritable_page;
2269                         }
2270                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2271                                 lock_page(old_page);
2272                                 if (!old_page->mapping) {
2273                                         ret = 0; /* retry the fault */
2274                                         unlock_page(old_page);
2275                                         goto unwritable_page;
2276                                 }
2277                         } else
2278                                 VM_BUG_ON(!PageLocked(old_page));
2279
2280                         /*
2281                          * Since we dropped the lock we need to revalidate
2282                          * the PTE as someone else may have changed it.  If
2283                          * they did, we just return, as we can count on the
2284                          * MMU to tell us if they didn't also make it writable.
2285                          */
2286                         page_table = pte_offset_map_lock(mm, pmd, address,
2287                                                          &ptl);
2288                         if (!pte_same(*page_table, orig_pte)) {
2289                                 unlock_page(old_page);
2290                                 page_cache_release(old_page);
2291                                 goto unlock;
2292                         }
2293
2294                         page_mkwrite = 1;
2295                 }
2296                 dirty_page = old_page;
2297                 get_page(dirty_page);
2298
2299 reuse:
2300                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2301                 entry = pte_mkyoung(orig_pte);
2302                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2303                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2304                         update_mmu_cache(vma, address, page_table);
2305                 pte_unmap_unlock(page_table, ptl);
2306                 ret |= VM_FAULT_WRITE;
2307
2308                 if (!dirty_page)
2309                         return ret;
2310
2311                 /*
2312                  * Yes, Virginia, this is actually required to prevent a race
2313                  * with clear_page_dirty_for_io() from clearing the page dirty
2314                  * bit after it clear all dirty ptes, but before a racing
2315                  * do_wp_page installs a dirty pte.
2316                  *
2317                  * do_no_page is protected similarly.
2318                  */
2319                 if (!page_mkwrite) {
2320                         wait_on_page_locked(dirty_page);
2321                         set_page_dirty_balance(dirty_page, page_mkwrite);
2322                 }
2323                 put_page(dirty_page);
2324                 if (page_mkwrite) {
2325                         struct address_space *mapping = dirty_page->mapping;
2326
2327                         set_page_dirty(dirty_page);
2328                         unlock_page(dirty_page);
2329                         page_cache_release(dirty_page);
2330                         if (mapping)    {
2331                                 /*
2332                                  * Some device drivers do not set page.mapping
2333                                  * but still dirty their pages
2334                                  */
2335                                 balance_dirty_pages_ratelimited(mapping);
2336                         }
2337                 }
2338
2339                 /* file_update_time outside page_lock */
2340                 if (vma->vm_file)
2341                         file_update_time(vma->vm_file);
2342
2343                 return ret;
2344         }
2345
2346         /*
2347          * Ok, we need to copy. Oh, well..
2348          */
2349         page_cache_get(old_page);
2350 gotten:
2351         pte_unmap_unlock(page_table, ptl);
2352
2353         if (unlikely(anon_vma_prepare(vma)))
2354                 goto oom;
2355
2356         if (is_zero_pfn(pte_pfn(orig_pte))) {
2357                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2358                 if (!new_page)
2359                         goto oom;
2360         } else {
2361                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2362                 if (!new_page)
2363                         goto oom;
2364                 cow_user_page(new_page, old_page, address, vma);
2365         }
2366         __SetPageUptodate(new_page);
2367
2368         /*
2369          * Don't let another task, with possibly unlocked vma,
2370          * keep the mlocked page.
2371          */
2372         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2373                 lock_page(old_page);    /* for LRU manipulation */
2374                 clear_page_mlock(old_page);
2375                 unlock_page(old_page);
2376         }
2377
2378         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2379                 goto oom_free_new;
2380
2381         /*
2382          * Re-check the pte - we dropped the lock
2383          */
2384         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2385         if (likely(pte_same(*page_table, orig_pte))) {
2386                 if (old_page) {
2387                         if (!PageAnon(old_page)) {
2388                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2389                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2390                         }
2391                 } else
2392                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2393                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2394                 entry = mk_pte(new_page, vma->vm_page_prot);
2395                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2396                 /*
2397                  * Clear the pte entry and flush it first, before updating the
2398                  * pte with the new entry. This will avoid a race condition
2399                  * seen in the presence of one thread doing SMC and another
2400                  * thread doing COW.
2401                  */
2402                 ptep_clear_flush(vma, address, page_table);
2403                 page_add_new_anon_rmap(new_page, vma, address);
2404                 /*
2405                  * We call the notify macro here because, when using secondary
2406                  * mmu page tables (such as kvm shadow page tables), we want the
2407                  * new page to be mapped directly into the secondary page table.
2408                  */
2409                 set_pte_at_notify(mm, address, page_table, entry);
2410                 update_mmu_cache(vma, address, page_table);
2411                 if (old_page) {
2412                         /*
2413                          * Only after switching the pte to the new page may
2414                          * we remove the mapcount here. Otherwise another
2415                          * process may come and find the rmap count decremented
2416                          * before the pte is switched to the new page, and
2417                          * "reuse" the old page writing into it while our pte
2418                          * here still points into it and can be read by other
2419                          * threads.
2420                          *
2421                          * The critical issue is to order this
2422                          * page_remove_rmap with the ptp_clear_flush above.
2423                          * Those stores are ordered by (if nothing else,)
2424                          * the barrier present in the atomic_add_negative
2425                          * in page_remove_rmap.
2426                          *
2427                          * Then the TLB flush in ptep_clear_flush ensures that
2428                          * no process can access the old page before the
2429                          * decremented mapcount is visible. And the old page
2430                          * cannot be reused until after the decremented
2431                          * mapcount is visible. So transitively, TLBs to
2432                          * old page will be flushed before it can be reused.
2433                          */
2434                         page_remove_rmap(old_page);
2435                 }
2436
2437                 /* Free the old page.. */
2438                 new_page = old_page;
2439                 ret |= VM_FAULT_WRITE;
2440         } else
2441                 mem_cgroup_uncharge_page(new_page);
2442
2443         if (new_page)
2444                 page_cache_release(new_page);
2445         if (old_page)
2446                 page_cache_release(old_page);
2447 unlock:
2448         pte_unmap_unlock(page_table, ptl);
2449         return ret;
2450 oom_free_new:
2451         page_cache_release(new_page);
2452 oom:
2453         if (old_page) {
2454                 if (page_mkwrite) {
2455                         unlock_page(old_page);
2456                         page_cache_release(old_page);
2457                 }
2458                 page_cache_release(old_page);
2459         }
2460         return VM_FAULT_OOM;
2461
2462 unwritable_page:
2463         page_cache_release(old_page);
2464         return ret;
2465 }
2466
2467 /*
2468  * Helper functions for unmap_mapping_range().
2469  *
2470  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2471  *
2472  * We have to restart searching the prio_tree whenever we drop the lock,
2473  * since the iterator is only valid while the lock is held, and anyway
2474  * a later vma might be split and reinserted earlier while lock dropped.
2475  *
2476  * The list of nonlinear vmas could be handled more efficiently, using
2477  * a placeholder, but handle it in the same way until a need is shown.
2478  * It is important to search the prio_tree before nonlinear list: a vma
2479  * may become nonlinear and be shifted from prio_tree to nonlinear list
2480  * while the lock is dropped; but never shifted from list to prio_tree.
2481  *
2482  * In order to make forward progress despite restarting the search,
2483  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2484  * quickly skip it next time around.  Since the prio_tree search only
2485  * shows us those vmas affected by unmapping the range in question, we
2486  * can't efficiently keep all vmas in step with mapping->truncate_count:
2487  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2488  * mapping->truncate_count and vma->vm_truncate_count are protected by
2489  * i_mmap_lock.
2490  *
2491  * In order to make forward progress despite repeatedly restarting some
2492  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2493  * and restart from that address when we reach that vma again.  It might
2494  * have been split or merged, shrunk or extended, but never shifted: so
2495  * restart_addr remains valid so long as it remains in the vma's range.
2496  * unmap_mapping_range forces truncate_count to leap over page-aligned
2497  * values so we can save vma's restart_addr in its truncate_count field.
2498  */
2499 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2500
2501 static void reset_vma_truncate_counts(struct address_space *mapping)
2502 {
2503         struct vm_area_struct *vma;
2504         struct prio_tree_iter iter;
2505
2506         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2507                 vma->vm_truncate_count = 0;
2508         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2509                 vma->vm_truncate_count = 0;
2510 }
2511
2512 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2513                 unsigned long start_addr, unsigned long end_addr,
2514                 struct zap_details *details)
2515 {
2516         unsigned long restart_addr;
2517         int need_break;
2518
2519         /*
2520          * files that support invalidating or truncating portions of the
2521          * file from under mmaped areas must have their ->fault function
2522          * return a locked page (and set VM_FAULT_LOCKED in the return).
2523          * This provides synchronisation against concurrent unmapping here.
2524          */
2525
2526 again:
2527         restart_addr = vma->vm_truncate_count;
2528         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2529                 start_addr = restart_addr;
2530                 if (start_addr >= end_addr) {
2531                         /* Top of vma has been split off since last time */
2532                         vma->vm_truncate_count = details->truncate_count;
2533                         return 0;
2534                 }
2535         }
2536
2537         restart_addr = zap_page_range(vma, start_addr,
2538                                         end_addr - start_addr, details);
2539         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2540
2541         if (restart_addr >= end_addr) {
2542                 /* We have now completed this vma: mark it so */
2543                 vma->vm_truncate_count = details->truncate_count;
2544                 if (!need_break)
2545                         return 0;
2546         } else {
2547                 /* Note restart_addr in vma's truncate_count field */
2548                 vma->vm_truncate_count = restart_addr;
2549                 if (!need_break)
2550                         goto again;
2551         }
2552
2553         spin_unlock(details->i_mmap_lock);
2554         cond_resched();
2555         spin_lock(details->i_mmap_lock);
2556         return -EINTR;
2557 }
2558
2559 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2560                                             struct zap_details *details)
2561 {
2562         struct vm_area_struct *vma;
2563         struct prio_tree_iter iter;
2564         pgoff_t vba, vea, zba, zea;
2565
2566 restart:
2567         vma_prio_tree_foreach(vma, &iter, root,
2568                         details->first_index, details->last_index) {
2569                 /* Skip quickly over those we have already dealt with */
2570                 if (vma->vm_truncate_count == details->truncate_count)
2571                         continue;
2572
2573                 vba = vma->vm_pgoff;
2574                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2575                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2576                 zba = details->first_index;
2577                 if (zba < vba)
2578                         zba = vba;
2579                 zea = details->last_index;
2580                 if (zea > vea)
2581                         zea = vea;
2582
2583                 if (unmap_mapping_range_vma(vma,
2584                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2585                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2586                                 details) < 0)
2587                         goto restart;
2588         }
2589 }
2590
2591 static inline void unmap_mapping_range_list(struct list_head *head,
2592                                             struct zap_details *details)
2593 {
2594         struct vm_area_struct *vma;
2595
2596         /*
2597          * In nonlinear VMAs there is no correspondence between virtual address
2598          * offset and file offset.  So we must perform an exhaustive search
2599          * across *all* the pages in each nonlinear VMA, not just the pages
2600          * whose virtual address lies outside the file truncation point.
2601          */
2602 restart:
2603         list_for_each_entry(vma, head, shared.vm_set.list) {
2604                 /* Skip quickly over those we have already dealt with */
2605                 if (vma->vm_truncate_count == details->truncate_count)
2606                         continue;
2607                 details->nonlinear_vma = vma;
2608                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2609                                         vma->vm_end, details) < 0)
2610                         goto restart;
2611         }
2612 }
2613
2614 /**
2615  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2616  * @mapping: the address space containing mmaps to be unmapped.
2617  * @holebegin: byte in first page to unmap, relative to the start of
2618  * the underlying file.  This will be rounded down to a PAGE_SIZE
2619  * boundary.  Note that this is different from truncate_pagecache(), which
2620  * must keep the partial page.  In contrast, we must get rid of
2621  * partial pages.
2622  * @holelen: size of prospective hole in bytes.  This will be rounded
2623  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2624  * end of the file.
2625  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2626  * but 0 when invalidating pagecache, don't throw away private data.
2627  */
2628 void unmap_mapping_range(struct address_space *mapping,
2629                 loff_t const holebegin, loff_t const holelen, int even_cows)
2630 {
2631         struct zap_details details;
2632         pgoff_t hba = holebegin >> PAGE_SHIFT;
2633         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2634
2635         /* Check for overflow. */
2636         if (sizeof(holelen) > sizeof(hlen)) {
2637                 long long holeend =
2638                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2639                 if (holeend & ~(long long)ULONG_MAX)
2640                         hlen = ULONG_MAX - hba + 1;
2641         }
2642
2643         details.check_mapping = even_cows? NULL: mapping;
2644         details.nonlinear_vma = NULL;
2645         details.first_index = hba;
2646         details.last_index = hba + hlen - 1;
2647         if (details.last_index < details.first_index)
2648                 details.last_index = ULONG_MAX;
2649         details.i_mmap_lock = &mapping->i_mmap_lock;
2650
2651         spin_lock(&mapping->i_mmap_lock);
2652
2653         /* Protect against endless unmapping loops */
2654         mapping->truncate_count++;
2655         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2656                 if (mapping->truncate_count == 0)
2657                         reset_vma_truncate_counts(mapping);
2658                 mapping->truncate_count++;
2659         }
2660         details.truncate_count = mapping->truncate_count;
2661
2662         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2663                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2664         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2665                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2666         spin_unlock(&mapping->i_mmap_lock);
2667 }
2668 EXPORT_SYMBOL(unmap_mapping_range);
2669
2670 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2671 {
2672         struct address_space *mapping = inode->i_mapping;
2673
2674         /*
2675          * If the underlying filesystem is not going to provide
2676          * a way to truncate a range of blocks (punch a hole) -
2677          * we should return failure right now.
2678          */
2679         if (!inode->i_op->truncate_range)
2680                 return -ENOSYS;
2681
2682         mutex_lock(&inode->i_mutex);
2683         down_write(&inode->i_alloc_sem);
2684         unmap_mapping_range(mapping, offset, (end - offset), 1);
2685         truncate_inode_pages_range(mapping, offset, end);
2686         unmap_mapping_range(mapping, offset, (end - offset), 1);
2687         inode->i_op->truncate_range(inode, offset, end);
2688         up_write(&inode->i_alloc_sem);
2689         mutex_unlock(&inode->i_mutex);
2690
2691         return 0;
2692 }
2693
2694 /*
2695  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2696  * but allow concurrent faults), and pte mapped but not yet locked.
2697  * We return with mmap_sem still held, but pte unmapped and unlocked.
2698  */
2699 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2700                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2701                 unsigned int flags, pte_t orig_pte)
2702 {
2703         spinlock_t *ptl;
2704         struct page *page, *swapcache = NULL;
2705         swp_entry_t entry;
2706         pte_t pte;
2707         int locked;
2708         struct mem_cgroup *ptr = NULL;
2709         int exclusive = 0;
2710         int ret = 0;
2711
2712         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2713                 goto out;
2714
2715         entry = pte_to_swp_entry(orig_pte);
2716         if (unlikely(non_swap_entry(entry))) {
2717                 if (is_migration_entry(entry)) {
2718                         migration_entry_wait(mm, pmd, address);
2719                 } else if (is_hwpoison_entry(entry)) {
2720                         ret = VM_FAULT_HWPOISON;
2721                 } else {
2722                         print_bad_pte(vma, address, orig_pte, NULL);
2723                         ret = VM_FAULT_SIGBUS;
2724                 }
2725                 goto out;
2726         }
2727         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2728         page = lookup_swap_cache(entry);
2729         if (!page) {
2730                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2731                 page = swapin_readahead(entry,
2732                                         GFP_HIGHUSER_MOVABLE, vma, address);
2733                 if (!page) {
2734                         /*
2735                          * Back out if somebody else faulted in this pte
2736                          * while we released the pte lock.
2737                          */
2738                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2739                         if (likely(pte_same(*page_table, orig_pte)))
2740                                 ret = VM_FAULT_OOM;
2741                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2742                         goto unlock;
2743                 }
2744
2745                 /* Had to read the page from swap area: Major fault */
2746                 ret = VM_FAULT_MAJOR;
2747                 count_vm_event(PGMAJFAULT);
2748         } else if (PageHWPoison(page)) {
2749                 /*
2750                  * hwpoisoned dirty swapcache pages are kept for killing
2751                  * owner processes (which may be unknown at hwpoison time)
2752                  */
2753                 ret = VM_FAULT_HWPOISON;
2754                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2755                 goto out_release;
2756         }
2757
2758         locked = lock_page_or_retry(page, mm, flags);
2759         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2760         if (!locked) {
2761                 ret |= VM_FAULT_RETRY;
2762                 goto out_release;
2763         }
2764
2765         /*
2766          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2767          * release the swapcache from under us.  The page pin, and pte_same
2768          * test below, are not enough to exclude that.  Even if it is still
2769          * swapcache, we need to check that the page's swap has not changed.
2770          */
2771         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2772                 goto out_page;
2773
2774         if (ksm_might_need_to_copy(page, vma, address)) {
2775                 swapcache = page;
2776                 page = ksm_does_need_to_copy(page, vma, address);
2777
2778                 if (unlikely(!page)) {
2779                         ret = VM_FAULT_OOM;
2780                         page = swapcache;
2781                         swapcache = NULL;
2782                         goto out_page;
2783                 }
2784         }
2785
2786         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2787                 ret = VM_FAULT_OOM;
2788                 goto out_page;
2789         }
2790
2791         /*
2792          * Back out if somebody else already faulted in this pte.
2793          */
2794         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2795         if (unlikely(!pte_same(*page_table, orig_pte)))
2796                 goto out_nomap;
2797
2798         if (unlikely(!PageUptodate(page))) {
2799                 ret = VM_FAULT_SIGBUS;
2800                 goto out_nomap;
2801         }
2802
2803         /*
2804          * The page isn't present yet, go ahead with the fault.
2805          *
2806          * Be careful about the sequence of operations here.
2807          * To get its accounting right, reuse_swap_page() must be called
2808          * while the page is counted on swap but not yet in mapcount i.e.
2809          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2810          * must be called after the swap_free(), or it will never succeed.
2811          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2812          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2813          * in page->private. In this case, a record in swap_cgroup  is silently
2814          * discarded at swap_free().
2815          */
2816
2817         inc_mm_counter_fast(mm, MM_ANONPAGES);
2818         dec_mm_counter_fast(mm, MM_SWAPENTS);
2819         pte = mk_pte(page, vma->vm_page_prot);
2820         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2821                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2822                 flags &= ~FAULT_FLAG_WRITE;
2823                 ret |= VM_FAULT_WRITE;
2824                 exclusive = 1;
2825         }
2826         flush_icache_page(vma, page);
2827         set_pte_at(mm, address, page_table, pte);
2828         do_page_add_anon_rmap(page, vma, address, exclusive);
2829         /* It's better to call commit-charge after rmap is established */
2830         mem_cgroup_commit_charge_swapin(page, ptr);
2831
2832         swap_free(entry);
2833         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2834                 try_to_free_swap(page);
2835         unlock_page(page);
2836         if (swapcache) {
2837                 /*
2838                  * Hold the lock to avoid the swap entry to be reused
2839                  * until we take the PT lock for the pte_same() check
2840                  * (to avoid false positives from pte_same). For
2841                  * further safety release the lock after the swap_free
2842                  * so that the swap count won't change under a
2843                  * parallel locked swapcache.
2844                  */
2845                 unlock_page(swapcache);
2846                 page_cache_release(swapcache);
2847         }
2848
2849         if (flags & FAULT_FLAG_WRITE) {
2850                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2851                 if (ret & VM_FAULT_ERROR)
2852                         ret &= VM_FAULT_ERROR;
2853                 goto out;
2854         }
2855
2856         /* No need to invalidate - it was non-present before */
2857         update_mmu_cache(vma, address, page_table);
2858 unlock:
2859         pte_unmap_unlock(page_table, ptl);
2860 out:
2861         return ret;
2862 out_nomap:
2863         mem_cgroup_cancel_charge_swapin(ptr);
2864         pte_unmap_unlock(page_table, ptl);
2865 out_page:
2866         unlock_page(page);
2867 out_release:
2868         page_cache_release(page);
2869         if (swapcache) {
2870                 unlock_page(swapcache);
2871                 page_cache_release(swapcache);
2872         }
2873         return ret;
2874 }
2875
2876 /*
2877  * This is like a special single-page "expand_{down|up}wards()",
2878  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2879  * doesn't hit another vma.
2880  */
2881 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2882 {
2883         address &= PAGE_MASK;
2884         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2885                 struct vm_area_struct *prev = vma->vm_prev;
2886
2887                 /*
2888                  * Is there a mapping abutting this one below?
2889                  *
2890                  * That's only ok if it's the same stack mapping
2891                  * that has gotten split..
2892                  */
2893                 if (prev && prev->vm_end == address)
2894                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2895
2896                 expand_stack(vma, address - PAGE_SIZE);
2897         }
2898         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2899                 struct vm_area_struct *next = vma->vm_next;
2900
2901                 /* As VM_GROWSDOWN but s/below/above/ */
2902                 if (next && next->vm_start == address + PAGE_SIZE)
2903                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2904
2905                 expand_upwards(vma, address + PAGE_SIZE);
2906         }
2907         return 0;
2908 }
2909
2910 /*
2911  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2912  * but allow concurrent faults), and pte mapped but not yet locked.
2913  * We return with mmap_sem still held, but pte unmapped and unlocked.
2914  */
2915 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2916                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2917                 unsigned int flags)
2918 {
2919         struct page *page;
2920         spinlock_t *ptl;
2921         pte_t entry;
2922
2923         pte_unmap(page_table);
2924
2925         /* Check if we need to add a guard page to the stack */
2926         if (check_stack_guard_page(vma, address) < 0)
2927                 return VM_FAULT_SIGBUS;
2928
2929         /* Use the zero-page for reads */
2930         if (!(flags & FAULT_FLAG_WRITE)) {
2931                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2932                                                 vma->vm_page_prot));
2933                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2934                 if (!pte_none(*page_table))
2935                         goto unlock;
2936                 goto setpte;
2937         }
2938
2939         /* Allocate our own private page. */
2940         if (unlikely(anon_vma_prepare(vma)))
2941                 goto oom;
2942         page = alloc_zeroed_user_highpage_movable(vma, address);
2943         if (!page)
2944                 goto oom;
2945         __SetPageUptodate(page);
2946
2947         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2948                 goto oom_free_page;
2949
2950         entry = mk_pte(page, vma->vm_page_prot);
2951         if (vma->vm_flags & VM_WRITE)
2952                 entry = pte_mkwrite(pte_mkdirty(entry));
2953
2954         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2955         if (!pte_none(*page_table))
2956                 goto release;
2957
2958         inc_mm_counter_fast(mm, MM_ANONPAGES);
2959         page_add_new_anon_rmap(page, vma, address);
2960 setpte:
2961         set_pte_at(mm, address, page_table, entry);
2962
2963         /* No need to invalidate - it was non-present before */
2964         update_mmu_cache(vma, address, page_table);
2965 unlock:
2966         pte_unmap_unlock(page_table, ptl);
2967         return 0;
2968 release:
2969         mem_cgroup_uncharge_page(page);
2970         page_cache_release(page);
2971         goto unlock;
2972 oom_free_page:
2973         page_cache_release(page);
2974 oom:
2975         return VM_FAULT_OOM;
2976 }
2977
2978 /*
2979  * __do_fault() tries to create a new page mapping. It aggressively
2980  * tries to share with existing pages, but makes a separate copy if
2981  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2982  * the next page fault.
2983  *
2984  * As this is called only for pages that do not currently exist, we
2985  * do not need to flush old virtual caches or the TLB.
2986  *
2987  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2988  * but allow concurrent faults), and pte neither mapped nor locked.
2989  * We return with mmap_sem still held, but pte unmapped and unlocked.
2990  */
2991 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2992                 unsigned long address, pmd_t *pmd,
2993                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2994 {
2995         pte_t *page_table;
2996         spinlock_t *ptl;
2997         struct page *page;
2998         pte_t entry;
2999         int anon = 0;
3000         int charged = 0;
3001         struct page *dirty_page = NULL;
3002         struct vm_fault vmf;
3003         int ret;
3004         int page_mkwrite = 0;
3005
3006         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3007         vmf.pgoff = pgoff;
3008         vmf.flags = flags;
3009         vmf.page = NULL;
3010
3011         ret = vma->vm_ops->fault(vma, &vmf);
3012         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3013                             VM_FAULT_RETRY)))
3014                 return ret;
3015
3016         if (unlikely(PageHWPoison(vmf.page))) {
3017                 if (ret & VM_FAULT_LOCKED)
3018                         unlock_page(vmf.page);
3019                 return VM_FAULT_HWPOISON;
3020         }
3021
3022         /*
3023          * For consistency in subsequent calls, make the faulted page always
3024          * locked.
3025          */
3026         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3027                 lock_page(vmf.page);
3028         else
3029                 VM_BUG_ON(!PageLocked(vmf.page));
3030
3031         /*
3032          * Should we do an early C-O-W break?
3033          */
3034         page = vmf.page;
3035         if (flags & FAULT_FLAG_WRITE) {
3036                 if (!(vma->vm_flags & VM_SHARED)) {
3037                         anon = 1;
3038                         if (unlikely(anon_vma_prepare(vma))) {
3039                                 ret = VM_FAULT_OOM;
3040                                 goto out;
3041                         }
3042                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3043                                                 vma, address);
3044                         if (!page) {
3045                                 ret = VM_FAULT_OOM;
3046                                 goto out;
3047                         }
3048                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3049                                 ret = VM_FAULT_OOM;
3050                                 page_cache_release(page);
3051                                 goto out;
3052                         }
3053                         charged = 1;
3054                         /*
3055                          * Don't let another task, with possibly unlocked vma,
3056                          * keep the mlocked page.
3057                          */
3058                         if (vma->vm_flags & VM_LOCKED)
3059                                 clear_page_mlock(vmf.page);
3060                         copy_user_highpage(page, vmf.page, address, vma);
3061                         __SetPageUptodate(page);
3062                 } else {
3063                         /*
3064                          * If the page will be shareable, see if the backing
3065                          * address space wants to know that the page is about
3066                          * to become writable
3067                          */
3068                         if (vma->vm_ops->page_mkwrite) {
3069                                 int tmp;
3070
3071                                 unlock_page(page);
3072                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3073                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3074                                 if (unlikely(tmp &
3075                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3076                                         ret = tmp;
3077                                         goto unwritable_page;
3078                                 }
3079                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3080                                         lock_page(page);
3081                                         if (!page->mapping) {
3082                                                 ret = 0; /* retry the fault */
3083                                                 unlock_page(page);
3084                                                 goto unwritable_page;
3085                                         }
3086                                 } else
3087                                         VM_BUG_ON(!PageLocked(page));
3088                                 page_mkwrite = 1;
3089                         }
3090                 }
3091
3092         }
3093
3094         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3095
3096         /*
3097          * This silly early PAGE_DIRTY setting removes a race
3098          * due to the bad i386 page protection. But it's valid
3099          * for other architectures too.
3100          *
3101          * Note that if FAULT_FLAG_WRITE is set, we either now have
3102          * an exclusive copy of the page, or this is a shared mapping,
3103          * so we can make it writable and dirty to avoid having to
3104          * handle that later.
3105          */
3106         /* Only go through if we didn't race with anybody else... */
3107         if (likely(pte_same(*page_table, orig_pte))) {
3108                 flush_icache_page(vma, page);
3109                 entry = mk_pte(page, vma->vm_page_prot);
3110                 if (flags & FAULT_FLAG_WRITE)
3111                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3112                 if (anon) {
3113                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3114                         page_add_new_anon_rmap(page, vma, address);
3115                 } else {
3116                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3117                         page_add_file_rmap(page);
3118                         if (flags & FAULT_FLAG_WRITE) {
3119                                 dirty_page = page;
3120                                 get_page(dirty_page);
3121                         }
3122                 }
3123                 set_pte_at(mm, address, page_table, entry);
3124
3125                 /* no need to invalidate: a not-present page won't be cached */
3126                 update_mmu_cache(vma, address, page_table);
3127         } else {
3128                 if (charged)
3129                         mem_cgroup_uncharge_page(page);
3130                 if (anon)
3131                         page_cache_release(page);
3132                 else
3133                         anon = 1; /* no anon but release faulted_page */
3134         }
3135
3136         pte_unmap_unlock(page_table, ptl);
3137
3138 out:
3139         if (dirty_page) {
3140                 struct address_space *mapping = page->mapping;
3141
3142                 if (set_page_dirty(dirty_page))
3143                         page_mkwrite = 1;
3144                 unlock_page(dirty_page);
3145                 put_page(dirty_page);
3146                 if (page_mkwrite && mapping) {
3147                         /*
3148                          * Some device drivers do not set page.mapping but still
3149                          * dirty their pages
3150                          */
3151                         balance_dirty_pages_ratelimited(mapping);
3152                 }
3153
3154                 /* file_update_time outside page_lock */
3155                 if (vma->vm_file)
3156                         file_update_time(vma->vm_file);
3157         } else {
3158                 unlock_page(vmf.page);
3159                 if (anon)
3160                         page_cache_release(vmf.page);
3161         }
3162
3163         return ret;
3164
3165 unwritable_page:
3166         page_cache_release(page);
3167         return ret;
3168 }
3169
3170 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3171                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3172                 unsigned int flags, pte_t orig_pte)
3173 {
3174         pgoff_t pgoff = (((address & PAGE_MASK)
3175                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3176
3177         pte_unmap(page_table);
3178         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3179 }
3180
3181 /*
3182  * Fault of a previously existing named mapping. Repopulate the pte
3183  * from the encoded file_pte if possible. This enables swappable
3184  * nonlinear vmas.
3185  *
3186  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3187  * but allow concurrent faults), and pte mapped but not yet locked.
3188  * We return with mmap_sem still held, but pte unmapped and unlocked.
3189  */
3190 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3191                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3192                 unsigned int flags, pte_t orig_pte)
3193 {
3194         pgoff_t pgoff;
3195
3196         flags |= FAULT_FLAG_NONLINEAR;
3197
3198         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3199                 return 0;
3200
3201         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3202                 /*
3203                  * Page table corrupted: show pte and kill process.
3204                  */
3205                 print_bad_pte(vma, address, orig_pte, NULL);
3206                 return VM_FAULT_SIGBUS;
3207         }
3208
3209         pgoff = pte_to_pgoff(orig_pte);
3210         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3211 }
3212
3213 /*
3214  * These routines also need to handle stuff like marking pages dirty
3215  * and/or accessed for architectures that don't do it in hardware (most
3216  * RISC architectures).  The early dirtying is also good on the i386.
3217  *
3218  * There is also a hook called "update_mmu_cache()" that architectures
3219  * with external mmu caches can use to update those (ie the Sparc or
3220  * PowerPC hashed page tables that act as extended TLBs).
3221  *
3222  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3223  * but allow concurrent faults), and pte mapped but not yet locked.
3224  * We return with mmap_sem still held, but pte unmapped and unlocked.
3225  */
3226 int handle_pte_fault(struct mm_struct *mm,
3227                      struct vm_area_struct *vma, unsigned long address,
3228                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3229 {
3230         pte_t entry;
3231         spinlock_t *ptl;
3232
3233         entry = *pte;
3234         if (!pte_present(entry)) {
3235                 if (pte_none(entry)) {
3236                         if (vma->vm_ops) {
3237                                 if (likely(vma->vm_ops->fault))
3238                                         return do_linear_fault(mm, vma, address,
3239                                                 pte, pmd, flags, entry);
3240                         }
3241                         return do_anonymous_page(mm, vma, address,
3242                                                  pte, pmd, flags);
3243                 }
3244                 if (pte_file(entry))
3245                         return do_nonlinear_fault(mm, vma, address,
3246                                         pte, pmd, flags, entry);
3247                 return do_swap_page(mm, vma, address,
3248                                         pte, pmd, flags, entry);
3249         }
3250
3251         ptl = pte_lockptr(mm, pmd);
3252         spin_lock(ptl);
3253         if (unlikely(!pte_same(*pte, entry)))
3254                 goto unlock;
3255         if (flags & FAULT_FLAG_WRITE) {
3256                 if (!pte_write(entry))
3257                         return do_wp_page(mm, vma, address,
3258                                         pte, pmd, ptl, entry);
3259                 entry = pte_mkdirty(entry);
3260         }
3261         entry = pte_mkyoung(entry);
3262         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3263                 update_mmu_cache(vma, address, pte);
3264         } else {
3265                 /*
3266                  * This is needed only for protection faults but the arch code
3267                  * is not yet telling us if this is a protection fault or not.
3268                  * This still avoids useless tlb flushes for .text page faults
3269                  * with threads.
3270                  */
3271                 if (flags & FAULT_FLAG_WRITE)
3272                         flush_tlb_fix_spurious_fault(vma, address);
3273         }
3274 unlock:
3275         pte_unmap_unlock(pte, ptl);
3276         return 0;
3277 }
3278
3279 /*
3280  * By the time we get here, we already hold the mm semaphore
3281  */
3282 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3283                 unsigned long address, unsigned int flags)
3284 {
3285         pgd_t *pgd;
3286         pud_t *pud;
3287         pmd_t *pmd;
3288         pte_t *pte;
3289
3290         __set_current_state(TASK_RUNNING);
3291
3292         count_vm_event(PGFAULT);
3293
3294         /* do counter updates before entering really critical section. */
3295         check_sync_rss_stat(current);
3296
3297         if (unlikely(is_vm_hugetlb_page(vma)))
3298                 return hugetlb_fault(mm, vma, address, flags);
3299
3300         pgd = pgd_offset(mm, address);
3301         pud = pud_alloc(mm, pgd, address);
3302         if (!pud)
3303                 return VM_FAULT_OOM;
3304         pmd = pmd_alloc(mm, pud, address);
3305         if (!pmd)
3306                 return VM_FAULT_OOM;
3307         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3308                 if (!vma->vm_ops)
3309                         return do_huge_pmd_anonymous_page(mm, vma, address,
3310                                                           pmd, flags);
3311         } else {
3312                 pmd_t orig_pmd = *pmd;
3313                 barrier();
3314                 if (pmd_trans_huge(orig_pmd)) {
3315                         if (flags & FAULT_FLAG_WRITE &&
3316                             !pmd_write(orig_pmd) &&
3317                             !pmd_trans_splitting(orig_pmd))
3318                                 return do_huge_pmd_wp_page(mm, vma, address,
3319                                                            pmd, orig_pmd);
3320                         return 0;
3321                 }
3322         }
3323
3324         /*
3325          * Use __pte_alloc instead of pte_alloc_map, because we can't
3326          * run pte_offset_map on the pmd, if an huge pmd could
3327          * materialize from under us from a different thread.
3328          */
3329         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
3330                 return VM_FAULT_OOM;
3331         /* if an huge pmd materialized from under us just retry later */
3332         if (unlikely(pmd_trans_huge(*pmd)))
3333                 return 0;
3334         /*
3335          * A regular pmd is established and it can't morph into a huge pmd
3336          * from under us anymore at this point because we hold the mmap_sem
3337          * read mode and khugepaged takes it in write mode. So now it's
3338          * safe to run pte_offset_map().
3339          */
3340         pte = pte_offset_map(pmd, address);
3341
3342         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3343 }
3344
3345 #ifndef __PAGETABLE_PUD_FOLDED
3346 /*
3347  * Allocate page upper directory.
3348  * We've already handled the fast-path in-line.
3349  */
3350 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3351 {
3352         pud_t *new = pud_alloc_one(mm, address);
3353         if (!new)
3354                 return -ENOMEM;
3355
3356         smp_wmb(); /* See comment in __pte_alloc */
3357
3358         spin_lock(&mm->page_table_lock);
3359         if (pgd_present(*pgd))          /* Another has populated it */
3360                 pud_free(mm, new);
3361         else
3362                 pgd_populate(mm, pgd, new);
3363         spin_unlock(&mm->page_table_lock);
3364         return 0;
3365 }
3366 #endif /* __PAGETABLE_PUD_FOLDED */
3367
3368 #ifndef __PAGETABLE_PMD_FOLDED
3369 /*
3370  * Allocate page middle directory.
3371  * We've already handled the fast-path in-line.
3372  */
3373 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3374 {
3375         pmd_t *new = pmd_alloc_one(mm, address);
3376         if (!new)
3377                 return -ENOMEM;
3378
3379         smp_wmb(); /* See comment in __pte_alloc */
3380
3381         spin_lock(&mm->page_table_lock);
3382 #ifndef __ARCH_HAS_4LEVEL_HACK
3383         if (pud_present(*pud))          /* Another has populated it */
3384                 pmd_free(mm, new);
3385         else
3386                 pud_populate(mm, pud, new);
3387 #else
3388         if (pgd_present(*pud))          /* Another has populated it */
3389                 pmd_free(mm, new);
3390         else
3391                 pgd_populate(mm, pud, new);
3392 #endif /* __ARCH_HAS_4LEVEL_HACK */
3393         spin_unlock(&mm->page_table_lock);
3394         return 0;
3395 }
3396 #endif /* __PAGETABLE_PMD_FOLDED */
3397
3398 int make_pages_present(unsigned long addr, unsigned long end)
3399 {
3400         int ret, len, write;
3401         struct vm_area_struct * vma;
3402
3403         vma = find_vma(current->mm, addr);
3404         if (!vma)
3405                 return -ENOMEM;
3406         /*
3407          * We want to touch writable mappings with a write fault in order
3408          * to break COW, except for shared mappings because these don't COW
3409          * and we would not want to dirty them for nothing.
3410          */
3411         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3412         BUG_ON(addr >= end);
3413         BUG_ON(end > vma->vm_end);
3414         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3415         ret = get_user_pages(current, current->mm, addr,
3416                         len, write, 0, NULL, NULL);
3417         if (ret < 0)
3418                 return ret;
3419         return ret == len ? 0 : -EFAULT;
3420 }
3421
3422 #if !defined(__HAVE_ARCH_GATE_AREA)
3423
3424 #if defined(AT_SYSINFO_EHDR)
3425 static struct vm_area_struct gate_vma;
3426
3427 static int __init gate_vma_init(void)
3428 {
3429         gate_vma.vm_mm = NULL;
3430         gate_vma.vm_start = FIXADDR_USER_START;
3431         gate_vma.vm_end = FIXADDR_USER_END;
3432         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3433         gate_vma.vm_page_prot = __P101;
3434         /*
3435          * Make sure the vDSO gets into every core dump.
3436          * Dumping its contents makes post-mortem fully interpretable later
3437          * without matching up the same kernel and hardware config to see
3438          * what PC values meant.
3439          */
3440         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3441         return 0;
3442 }
3443 __initcall(gate_vma_init);
3444 #endif
3445
3446 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3447 {
3448 #ifdef AT_SYSINFO_EHDR
3449         return &gate_vma;
3450 #else
3451         return NULL;
3452 #endif
3453 }
3454
3455 int in_gate_area_no_task(unsigned long addr)
3456 {
3457 #ifdef AT_SYSINFO_EHDR
3458         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3459                 return 1;
3460 #endif
3461         return 0;
3462 }
3463
3464 #endif  /* __HAVE_ARCH_GATE_AREA */
3465
3466 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3467                 pte_t **ptepp, spinlock_t **ptlp)
3468 {
3469         pgd_t *pgd;
3470         pud_t *pud;
3471         pmd_t *pmd;
3472         pte_t *ptep;
3473
3474         pgd = pgd_offset(mm, address);
3475         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3476                 goto out;
3477
3478         pud = pud_offset(pgd, address);
3479         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3480                 goto out;
3481
3482         pmd = pmd_offset(pud, address);
3483         VM_BUG_ON(pmd_trans_huge(*pmd));
3484         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3485                 goto out;
3486
3487         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3488         if (pmd_huge(*pmd))
3489                 goto out;
3490
3491         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3492         if (!ptep)
3493                 goto out;
3494         if (!pte_present(*ptep))
3495                 goto unlock;
3496         *ptepp = ptep;
3497         return 0;
3498 unlock:
3499         pte_unmap_unlock(ptep, *ptlp);
3500 out:
3501         return -EINVAL;
3502 }
3503
3504 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3505                              pte_t **ptepp, spinlock_t **ptlp)
3506 {
3507         int res;
3508
3509         /* (void) is needed to make gcc happy */
3510         (void) __cond_lock(*ptlp,
3511                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3512         return res;
3513 }
3514
3515 /**
3516  * follow_pfn - look up PFN at a user virtual address
3517  * @vma: memory mapping
3518  * @address: user virtual address
3519  * @pfn: location to store found PFN
3520  *
3521  * Only IO mappings and raw PFN mappings are allowed.
3522  *
3523  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3524  */
3525 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3526         unsigned long *pfn)
3527 {
3528         int ret = -EINVAL;
3529         spinlock_t *ptl;
3530         pte_t *ptep;
3531
3532         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3533                 return ret;
3534
3535         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3536         if (ret)
3537                 return ret;
3538         *pfn = pte_pfn(*ptep);
3539         pte_unmap_unlock(ptep, ptl);
3540         return 0;
3541 }
3542 EXPORT_SYMBOL(follow_pfn);
3543
3544 #ifdef CONFIG_HAVE_IOREMAP_PROT
3545 int follow_phys(struct vm_area_struct *vma,
3546                 unsigned long address, unsigned int flags,
3547                 unsigned long *prot, resource_size_t *phys)
3548 {
3549         int ret = -EINVAL;
3550         pte_t *ptep, pte;
3551         spinlock_t *ptl;
3552
3553         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3554                 goto out;
3555
3556         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3557                 goto out;
3558         pte = *ptep;
3559
3560         if ((flags & FOLL_WRITE) && !pte_write(pte))
3561                 goto unlock;
3562
3563         *prot = pgprot_val(pte_pgprot(pte));
3564         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3565
3566         ret = 0;
3567 unlock:
3568         pte_unmap_unlock(ptep, ptl);
3569 out:
3570         return ret;
3571 }
3572
3573 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3574                         void *buf, int len, int write)
3575 {
3576         resource_size_t phys_addr;
3577         unsigned long prot = 0;
3578         void __iomem *maddr;
3579         int offset = addr & (PAGE_SIZE-1);
3580
3581         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3582                 return -EINVAL;
3583
3584         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3585         if (write)
3586                 memcpy_toio(maddr + offset, buf, len);
3587         else
3588                 memcpy_fromio(buf, maddr + offset, len);
3589         iounmap(maddr);
3590
3591         return len;
3592 }
3593 #endif
3594
3595 /*
3596  * Access another process' address space.
3597  * Source/target buffer must be kernel space,
3598  * Do not walk the page table directly, use get_user_pages
3599  */
3600 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3601 {
3602         struct mm_struct *mm;
3603         struct vm_area_struct *vma;
3604         void *old_buf = buf;
3605
3606         mm = get_task_mm(tsk);
3607         if (!mm)
3608                 return 0;
3609
3610         down_read(&mm->mmap_sem);
3611         /* ignore errors, just check how much was successfully transferred */
3612         while (len) {
3613                 int bytes, ret, offset;
3614                 void *maddr;
3615                 struct page *page = NULL;
3616
3617                 ret = get_user_pages(tsk, mm, addr, 1,
3618                                 write, 1, &page, &vma);
3619                 if (ret <= 0) {
3620                         /*
3621                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3622                          * we can access using slightly different code.
3623                          */
3624 #ifdef CONFIG_HAVE_IOREMAP_PROT
3625                         vma = find_vma(mm, addr);
3626                         if (!vma)
3627                                 break;
3628                         if (vma->vm_ops && vma->vm_ops->access)
3629                                 ret = vma->vm_ops->access(vma, addr, buf,
3630                                                           len, write);
3631                         if (ret <= 0)
3632 #endif
3633                                 break;
3634                         bytes = ret;
3635                 } else {
3636                         bytes = len;
3637                         offset = addr & (PAGE_SIZE-1);
3638                         if (bytes > PAGE_SIZE-offset)
3639                                 bytes = PAGE_SIZE-offset;
3640
3641                         maddr = kmap(page);
3642                         if (write) {
3643                                 copy_to_user_page(vma, page, addr,
3644                                                   maddr + offset, buf, bytes);
3645                                 set_page_dirty_lock(page);
3646                         } else {
3647                                 copy_from_user_page(vma, page, addr,
3648                                                     buf, maddr + offset, bytes);
3649                         }
3650                         kunmap(page);
3651                         page_cache_release(page);
3652                 }
3653                 len -= bytes;
3654                 buf += bytes;
3655                 addr += bytes;
3656         }
3657         up_read(&mm->mmap_sem);
3658         mmput(mm);
3659
3660         return buf - old_buf;
3661 }
3662
3663 /*
3664  * Print the name of a VMA.
3665  */
3666 void print_vma_addr(char *prefix, unsigned long ip)
3667 {
3668         struct mm_struct *mm = current->mm;
3669         struct vm_area_struct *vma;
3670
3671         /*
3672          * Do not print if we are in atomic
3673          * contexts (in exception stacks, etc.):
3674          */
3675         if (preempt_count())
3676                 return;
3677
3678         down_read(&mm->mmap_sem);
3679         vma = find_vma(mm, ip);
3680         if (vma && vma->vm_file) {
3681                 struct file *f = vma->vm_file;
3682                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3683                 if (buf) {
3684                         char *p, *s;
3685
3686                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3687                         if (IS_ERR(p))
3688                                 p = "?";
3689                         s = strrchr(p, '/');
3690                         if (s)
3691                                 p = s+1;
3692                         printk("%s%s[%lx+%lx]", prefix, p,
3693                                         vma->vm_start,
3694                                         vma->vm_end - vma->vm_start);
3695                         free_page((unsigned long)buf);
3696                 }
3697         }
3698         up_read(&current->mm->mmap_sem);
3699 }
3700
3701 #ifdef CONFIG_PROVE_LOCKING
3702 void might_fault(void)
3703 {
3704         /*
3705          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3706          * holding the mmap_sem, this is safe because kernel memory doesn't
3707          * get paged out, therefore we'll never actually fault, and the
3708          * below annotations will generate false positives.
3709          */
3710         if (segment_eq(get_fs(), KERNEL_DS))
3711                 return;
3712
3713         might_sleep();
3714         /*
3715          * it would be nicer only to annotate paths which are not under
3716          * pagefault_disable, however that requires a larger audit and
3717          * providing helpers like get_user_atomic.
3718          */
3719         if (!in_atomic() && current->mm)
3720                 might_lock_read(&current->mm->mmap_sem);
3721 }
3722 EXPORT_SYMBOL(might_fault);
3723 #endif
3724
3725 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3726 static void clear_gigantic_page(struct page *page,
3727                                 unsigned long addr,
3728                                 unsigned int pages_per_huge_page)
3729 {
3730         int i;
3731         struct page *p = page;
3732
3733         might_sleep();
3734         for (i = 0; i < pages_per_huge_page;
3735              i++, p = mem_map_next(p, page, i)) {
3736                 cond_resched();
3737                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3738         }
3739 }
3740 void clear_huge_page(struct page *page,
3741                      unsigned long addr, unsigned int pages_per_huge_page)
3742 {
3743         int i;
3744
3745         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3746                 clear_gigantic_page(page, addr, pages_per_huge_page);
3747                 return;
3748         }
3749
3750         might_sleep();
3751         for (i = 0; i < pages_per_huge_page; i++) {
3752                 cond_resched();
3753                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3754         }
3755 }
3756
3757 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3758                                     unsigned long addr,
3759                                     struct vm_area_struct *vma,
3760                                     unsigned int pages_per_huge_page)
3761 {
3762         int i;
3763         struct page *dst_base = dst;
3764         struct page *src_base = src;
3765
3766         for (i = 0; i < pages_per_huge_page; ) {
3767                 cond_resched();
3768                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3769
3770                 i++;
3771                 dst = mem_map_next(dst, dst_base, i);
3772                 src = mem_map_next(src, src_base, i);
3773         }
3774 }
3775
3776 void copy_user_huge_page(struct page *dst, struct page *src,
3777                          unsigned long addr, struct vm_area_struct *vma,
3778                          unsigned int pages_per_huge_page)
3779 {
3780         int i;
3781
3782         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3783                 copy_user_gigantic_page(dst, src, addr, vma,
3784                                         pages_per_huge_page);
3785                 return;
3786         }
3787
3788         might_sleep();
3789         for (i = 0; i < pages_per_huge_page; i++) {
3790                 cond_resched();
3791                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3792         }
3793 }
3794 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */