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