mm: keep a guard page below a grow-down stack segment
[0xlab-kernel:kernel.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
60 #include <asm/io.h>
61 #include <asm/pgalloc.h>
62 #include <asm/uaccess.h>
63 #include <asm/tlb.h>
64 #include <asm/tlbflush.h>
65 #include <asm/pgtable.h>
66
67 #include "internal.h"
68
69 #ifndef CONFIG_NEED_MULTIPLE_NODES
70 /* use the per-pgdat data instead for discontigmem - mbligh */
71 unsigned long max_mapnr;
72 struct page *mem_map;
73
74 EXPORT_SYMBOL(max_mapnr);
75 EXPORT_SYMBOL(mem_map);
76 #endif
77
78 unsigned long num_physpages;
79 /*
80  * A number of key systems in x86 including ioremap() rely on the assumption
81  * that high_memory defines the upper bound on direct map memory, then end
82  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
83  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
84  * and ZONE_HIGHMEM.
85  */
86 void * high_memory;
87
88 EXPORT_SYMBOL(num_physpages);
89 EXPORT_SYMBOL(high_memory);
90
91 /*
92  * Randomize the address space (stacks, mmaps, brk, etc.).
93  *
94  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
95  *   as ancient (libc5 based) binaries can segfault. )
96  */
97 int randomize_va_space __read_mostly =
98 #ifdef CONFIG_COMPAT_BRK
99                                         1;
100 #else
101                                         2;
102 #endif
103
104 static int __init disable_randmaps(char *s)
105 {
106         randomize_va_space = 0;
107         return 1;
108 }
109 __setup("norandmaps", disable_randmaps);
110
111 unsigned long zero_pfn __read_mostly;
112 unsigned long highest_memmap_pfn __read_mostly;
113
114 /*
115  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
116  */
117 static int __init init_zero_pfn(void)
118 {
119         zero_pfn = page_to_pfn(ZERO_PAGE(0));
120         return 0;
121 }
122 core_initcall(init_zero_pfn);
123
124 /*
125  * If a p?d_bad entry is found while walking page tables, report
126  * the error, before resetting entry to p?d_none.  Usually (but
127  * very seldom) called out from the p?d_none_or_clear_bad macros.
128  */
129
130 void pgd_clear_bad(pgd_t *pgd)
131 {
132         pgd_ERROR(*pgd);
133         pgd_clear(pgd);
134 }
135
136 void pud_clear_bad(pud_t *pud)
137 {
138         pud_ERROR(*pud);
139         pud_clear(pud);
140 }
141
142 void pmd_clear_bad(pmd_t *pmd)
143 {
144         pmd_ERROR(*pmd);
145         pmd_clear(pmd);
146 }
147
148 /*
149  * Note: this doesn't free the actual pages themselves. That
150  * has been handled earlier when unmapping all the memory regions.
151  */
152 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
153                            unsigned long addr)
154 {
155         pgtable_t token = pmd_pgtable(*pmd);
156         pmd_clear(pmd);
157         pte_free_tlb(tlb, token, addr);
158         tlb->mm->nr_ptes--;
159 }
160
161 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
162                                 unsigned long addr, unsigned long end,
163                                 unsigned long floor, unsigned long ceiling)
164 {
165         pmd_t *pmd;
166         unsigned long next;
167         unsigned long start;
168
169         start = addr;
170         pmd = pmd_offset(pud, addr);
171         do {
172                 next = pmd_addr_end(addr, end);
173                 if (pmd_none_or_clear_bad(pmd))
174                         continue;
175                 free_pte_range(tlb, pmd, addr);
176         } while (pmd++, addr = next, addr != end);
177
178         start &= PUD_MASK;
179         if (start < floor)
180                 return;
181         if (ceiling) {
182                 ceiling &= PUD_MASK;
183                 if (!ceiling)
184                         return;
185         }
186         if (end - 1 > ceiling - 1)
187                 return;
188
189         pmd = pmd_offset(pud, start);
190         pud_clear(pud);
191         pmd_free_tlb(tlb, pmd, start);
192 }
193
194 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
195                                 unsigned long addr, unsigned long end,
196                                 unsigned long floor, unsigned long ceiling)
197 {
198         pud_t *pud;
199         unsigned long next;
200         unsigned long start;
201
202         start = addr;
203         pud = pud_offset(pgd, addr);
204         do {
205                 next = pud_addr_end(addr, end);
206                 if (pud_none_or_clear_bad(pud))
207                         continue;
208                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
209         } while (pud++, addr = next, addr != end);
210
211         start &= PGDIR_MASK;
212         if (start < floor)
213                 return;
214         if (ceiling) {
215                 ceiling &= PGDIR_MASK;
216                 if (!ceiling)
217                         return;
218         }
219         if (end - 1 > ceiling - 1)
220                 return;
221
222         pud = pud_offset(pgd, start);
223         pgd_clear(pgd);
224         pud_free_tlb(tlb, pud, start);
225 }
226
227 /*
228  * This function frees user-level page tables of a process.
229  *
230  * Must be called with pagetable lock held.
231  */
232 void free_pgd_range(struct mmu_gather *tlb,
233                         unsigned long addr, unsigned long end,
234                         unsigned long floor, unsigned long ceiling)
235 {
236         pgd_t *pgd;
237         unsigned long next;
238         unsigned long start;
239
240         /*
241          * The next few lines have given us lots of grief...
242          *
243          * Why are we testing PMD* at this top level?  Because often
244          * there will be no work to do at all, and we'd prefer not to
245          * go all the way down to the bottom just to discover that.
246          *
247          * Why all these "- 1"s?  Because 0 represents both the bottom
248          * of the address space and the top of it (using -1 for the
249          * top wouldn't help much: the masks would do the wrong thing).
250          * The rule is that addr 0 and floor 0 refer to the bottom of
251          * the address space, but end 0 and ceiling 0 refer to the top
252          * Comparisons need to use "end - 1" and "ceiling - 1" (though
253          * that end 0 case should be mythical).
254          *
255          * Wherever addr is brought up or ceiling brought down, we must
256          * be careful to reject "the opposite 0" before it confuses the
257          * subsequent tests.  But what about where end is brought down
258          * by PMD_SIZE below? no, end can't go down to 0 there.
259          *
260          * Whereas we round start (addr) and ceiling down, by different
261          * masks at different levels, in order to test whether a table
262          * now has no other vmas using it, so can be freed, we don't
263          * bother to round floor or end up - the tests don't need that.
264          */
265
266         addr &= PMD_MASK;
267         if (addr < floor) {
268                 addr += PMD_SIZE;
269                 if (!addr)
270                         return;
271         }
272         if (ceiling) {
273                 ceiling &= PMD_MASK;
274                 if (!ceiling)
275                         return;
276         }
277         if (end - 1 > ceiling - 1)
278                 end -= PMD_SIZE;
279         if (addr > end - 1)
280                 return;
281
282         start = addr;
283         pgd = pgd_offset(tlb->mm, addr);
284         do {
285                 next = pgd_addr_end(addr, end);
286                 if (pgd_none_or_clear_bad(pgd))
287                         continue;
288                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
289         } while (pgd++, addr = next, addr != end);
290 }
291
292 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
293                 unsigned long floor, unsigned long ceiling)
294 {
295         while (vma) {
296                 struct vm_area_struct *next = vma->vm_next;
297                 unsigned long addr = vma->vm_start;
298
299                 /*
300                  * Hide vma from rmap and truncate_pagecache before freeing
301                  * pgtables
302                  */
303                 anon_vma_unlink(vma);
304                 unlink_file_vma(vma);
305
306                 if (is_vm_hugetlb_page(vma)) {
307                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
308                                 floor, next? next->vm_start: ceiling);
309                 } else {
310                         /*
311                          * Optimization: gather nearby vmas into one call down
312                          */
313                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
314                                && !is_vm_hugetlb_page(next)) {
315                                 vma = next;
316                                 next = vma->vm_next;
317                                 anon_vma_unlink(vma);
318                                 unlink_file_vma(vma);
319                         }
320                         free_pgd_range(tlb, addr, vma->vm_end,
321                                 floor, next? next->vm_start: ceiling);
322                 }
323                 vma = next;
324         }
325 }
326
327 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
328 {
329         pgtable_t new = pte_alloc_one(mm, address);
330         if (!new)
331                 return -ENOMEM;
332
333         /*
334          * Ensure all pte setup (eg. pte page lock and page clearing) are
335          * visible before the pte is made visible to other CPUs by being
336          * put into page tables.
337          *
338          * The other side of the story is the pointer chasing in the page
339          * table walking code (when walking the page table without locking;
340          * ie. most of the time). Fortunately, these data accesses consist
341          * of a chain of data-dependent loads, meaning most CPUs (alpha
342          * being the notable exception) will already guarantee loads are
343          * seen in-order. See the alpha page table accessors for the
344          * smp_read_barrier_depends() barriers in page table walking code.
345          */
346         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
347
348         spin_lock(&mm->page_table_lock);
349         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
350                 mm->nr_ptes++;
351                 pmd_populate(mm, pmd, new);
352                 new = NULL;
353         }
354         spin_unlock(&mm->page_table_lock);
355         if (new)
356                 pte_free(mm, new);
357         return 0;
358 }
359
360 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
361 {
362         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
363         if (!new)
364                 return -ENOMEM;
365
366         smp_wmb(); /* See comment in __pte_alloc */
367
368         spin_lock(&init_mm.page_table_lock);
369         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
370                 pmd_populate_kernel(&init_mm, pmd, new);
371                 new = NULL;
372         }
373         spin_unlock(&init_mm.page_table_lock);
374         if (new)
375                 pte_free_kernel(&init_mm, new);
376         return 0;
377 }
378
379 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
380 {
381         if (file_rss)
382                 add_mm_counter(mm, file_rss, file_rss);
383         if (anon_rss)
384                 add_mm_counter(mm, anon_rss, anon_rss);
385 }
386
387 /*
388  * This function is called to print an error when a bad pte
389  * is found. For example, we might have a PFN-mapped pte in
390  * a region that doesn't allow it.
391  *
392  * The calling function must still handle the error.
393  */
394 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
395                           pte_t pte, struct page *page)
396 {
397         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
398         pud_t *pud = pud_offset(pgd, addr);
399         pmd_t *pmd = pmd_offset(pud, addr);
400         struct address_space *mapping;
401         pgoff_t index;
402         static unsigned long resume;
403         static unsigned long nr_shown;
404         static unsigned long nr_unshown;
405
406         /*
407          * Allow a burst of 60 reports, then keep quiet for that minute;
408          * or allow a steady drip of one report per second.
409          */
410         if (nr_shown == 60) {
411                 if (time_before(jiffies, resume)) {
412                         nr_unshown++;
413                         return;
414                 }
415                 if (nr_unshown) {
416                         printk(KERN_ALERT
417                                 "BUG: Bad page map: %lu messages suppressed\n",
418                                 nr_unshown);
419                         nr_unshown = 0;
420                 }
421                 nr_shown = 0;
422         }
423         if (nr_shown++ == 0)
424                 resume = jiffies + 60 * HZ;
425
426         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
427         index = linear_page_index(vma, addr);
428
429         printk(KERN_ALERT
430                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
431                 current->comm,
432                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
433         if (page) {
434                 printk(KERN_ALERT
435                 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
436                 page, (void *)page->flags, page_count(page),
437                 page_mapcount(page), page->mapping, page->index);
438         }
439         printk(KERN_ALERT
440                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
441                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
442         /*
443          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
444          */
445         if (vma->vm_ops)
446                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
447                                 (unsigned long)vma->vm_ops->fault);
448         if (vma->vm_file && vma->vm_file->f_op)
449                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
450                                 (unsigned long)vma->vm_file->f_op->mmap);
451         dump_stack();
452         add_taint(TAINT_BAD_PAGE);
453 }
454
455 static inline int is_cow_mapping(unsigned int flags)
456 {
457         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
458 }
459
460 #ifndef is_zero_pfn
461 static inline int is_zero_pfn(unsigned long pfn)
462 {
463         return pfn == zero_pfn;
464 }
465 #endif
466
467 #ifndef my_zero_pfn
468 static inline unsigned long my_zero_pfn(unsigned long addr)
469 {
470         return zero_pfn;
471 }
472 #endif
473
474 /*
475  * vm_normal_page -- This function gets the "struct page" associated with a pte.
476  *
477  * "Special" mappings do not wish to be associated with a "struct page" (either
478  * it doesn't exist, or it exists but they don't want to touch it). In this
479  * case, NULL is returned here. "Normal" mappings do have a struct page.
480  *
481  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
482  * pte bit, in which case this function is trivial. Secondly, an architecture
483  * may not have a spare pte bit, which requires a more complicated scheme,
484  * described below.
485  *
486  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
487  * special mapping (even if there are underlying and valid "struct pages").
488  * COWed pages of a VM_PFNMAP are always normal.
489  *
490  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
491  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
492  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
493  * mapping will always honor the rule
494  *
495  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
496  *
497  * And for normal mappings this is false.
498  *
499  * This restricts such mappings to be a linear translation from virtual address
500  * to pfn. To get around this restriction, we allow arbitrary mappings so long
501  * as the vma is not a COW mapping; in that case, we know that all ptes are
502  * special (because none can have been COWed).
503  *
504  *
505  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
506  *
507  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
508  * page" backing, however the difference is that _all_ pages with a struct
509  * page (that is, those where pfn_valid is true) are refcounted and considered
510  * normal pages by the VM. The disadvantage is that pages are refcounted
511  * (which can be slower and simply not an option for some PFNMAP users). The
512  * advantage is that we don't have to follow the strict linearity rule of
513  * PFNMAP mappings in order to support COWable mappings.
514  *
515  */
516 #ifdef __HAVE_ARCH_PTE_SPECIAL
517 # define HAVE_PTE_SPECIAL 1
518 #else
519 # define HAVE_PTE_SPECIAL 0
520 #endif
521 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
522                                 pte_t pte)
523 {
524         unsigned long pfn = pte_pfn(pte);
525
526         if (HAVE_PTE_SPECIAL) {
527                 if (likely(!pte_special(pte)))
528                         goto check_pfn;
529                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
530                         return NULL;
531                 if (!is_zero_pfn(pfn))
532                         print_bad_pte(vma, addr, pte, NULL);
533                 return NULL;
534         }
535
536         /* !HAVE_PTE_SPECIAL case follows: */
537
538         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
539                 if (vma->vm_flags & VM_MIXEDMAP) {
540                         if (!pfn_valid(pfn))
541                                 return NULL;
542                         goto out;
543                 } else {
544                         unsigned long off;
545                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
546                         if (pfn == vma->vm_pgoff + off)
547                                 return NULL;
548                         if (!is_cow_mapping(vma->vm_flags))
549                                 return NULL;
550                 }
551         }
552
553         if (is_zero_pfn(pfn))
554                 return NULL;
555 check_pfn:
556         if (unlikely(pfn > highest_memmap_pfn)) {
557                 print_bad_pte(vma, addr, pte, NULL);
558                 return NULL;
559         }
560
561         /*
562          * NOTE! We still have PageReserved() pages in the page tables.
563          * eg. VDSO mappings can cause them to exist.
564          */
565 out:
566         return pfn_to_page(pfn);
567 }
568
569 /*
570  * copy one vm_area from one task to the other. Assumes the page tables
571  * already present in the new task to be cleared in the whole range
572  * covered by this vma.
573  */
574
575 static inline void
576 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
577                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
578                 unsigned long addr, int *rss)
579 {
580         unsigned long vm_flags = vma->vm_flags;
581         pte_t pte = *src_pte;
582         struct page *page;
583
584         /* pte contains position in swap or file, so copy. */
585         if (unlikely(!pte_present(pte))) {
586                 if (!pte_file(pte)) {
587                         swp_entry_t entry = pte_to_swp_entry(pte);
588
589                         swap_duplicate(entry);
590                         /* make sure dst_mm is on swapoff's mmlist. */
591                         if (unlikely(list_empty(&dst_mm->mmlist))) {
592                                 spin_lock(&mmlist_lock);
593                                 if (list_empty(&dst_mm->mmlist))
594                                         list_add(&dst_mm->mmlist,
595                                                  &src_mm->mmlist);
596                                 spin_unlock(&mmlist_lock);
597                         }
598                         if (is_write_migration_entry(entry) &&
599                                         is_cow_mapping(vm_flags)) {
600                                 /*
601                                  * COW mappings require pages in both parent
602                                  * and child to be set to read.
603                                  */
604                                 make_migration_entry_read(&entry);
605                                 pte = swp_entry_to_pte(entry);
606                                 set_pte_at(src_mm, addr, src_pte, pte);
607                         }
608                 }
609                 goto out_set_pte;
610         }
611
612         /*
613          * If it's a COW mapping, write protect it both
614          * in the parent and the child
615          */
616         if (is_cow_mapping(vm_flags)) {
617                 ptep_set_wrprotect(src_mm, addr, src_pte);
618                 pte = pte_wrprotect(pte);
619         }
620
621         /*
622          * If it's a shared mapping, mark it clean in
623          * the child
624          */
625         if (vm_flags & VM_SHARED)
626                 pte = pte_mkclean(pte);
627         pte = pte_mkold(pte);
628
629         page = vm_normal_page(vma, addr, pte);
630         if (page) {
631                 get_page(page);
632                 page_dup_rmap(page);
633                 rss[PageAnon(page)]++;
634         }
635
636 out_set_pte:
637         set_pte_at(dst_mm, addr, dst_pte, pte);
638 }
639
640 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
641                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
642                 unsigned long addr, unsigned long end)
643 {
644         pte_t *orig_src_pte, *orig_dst_pte;
645         pte_t *src_pte, *dst_pte;
646         spinlock_t *src_ptl, *dst_ptl;
647         int progress = 0;
648         int rss[2];
649
650 again:
651         rss[1] = rss[0] = 0;
652         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
653         if (!dst_pte)
654                 return -ENOMEM;
655         src_pte = pte_offset_map_nested(src_pmd, addr);
656         src_ptl = pte_lockptr(src_mm, src_pmd);
657         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
658         orig_src_pte = src_pte;
659         orig_dst_pte = dst_pte;
660         arch_enter_lazy_mmu_mode();
661
662         do {
663                 /*
664                  * We are holding two locks at this point - either of them
665                  * could generate latencies in another task on another CPU.
666                  */
667                 if (progress >= 32) {
668                         progress = 0;
669                         if (need_resched() ||
670                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
671                                 break;
672                 }
673                 if (pte_none(*src_pte)) {
674                         progress++;
675                         continue;
676                 }
677                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
678                 progress += 8;
679         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
680
681         arch_leave_lazy_mmu_mode();
682         spin_unlock(src_ptl);
683         pte_unmap_nested(orig_src_pte);
684         add_mm_rss(dst_mm, rss[0], rss[1]);
685         pte_unmap_unlock(orig_dst_pte, dst_ptl);
686         cond_resched();
687         if (addr != end)
688                 goto again;
689         return 0;
690 }
691
692 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
693                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
694                 unsigned long addr, unsigned long end)
695 {
696         pmd_t *src_pmd, *dst_pmd;
697         unsigned long next;
698
699         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
700         if (!dst_pmd)
701                 return -ENOMEM;
702         src_pmd = pmd_offset(src_pud, addr);
703         do {
704                 next = pmd_addr_end(addr, end);
705                 if (pmd_none_or_clear_bad(src_pmd))
706                         continue;
707                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
708                                                 vma, addr, next))
709                         return -ENOMEM;
710         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
711         return 0;
712 }
713
714 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
715                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
716                 unsigned long addr, unsigned long end)
717 {
718         pud_t *src_pud, *dst_pud;
719         unsigned long next;
720
721         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
722         if (!dst_pud)
723                 return -ENOMEM;
724         src_pud = pud_offset(src_pgd, addr);
725         do {
726                 next = pud_addr_end(addr, end);
727                 if (pud_none_or_clear_bad(src_pud))
728                         continue;
729                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
730                                                 vma, addr, next))
731                         return -ENOMEM;
732         } while (dst_pud++, src_pud++, addr = next, addr != end);
733         return 0;
734 }
735
736 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
737                 struct vm_area_struct *vma)
738 {
739         pgd_t *src_pgd, *dst_pgd;
740         unsigned long next;
741         unsigned long addr = vma->vm_start;
742         unsigned long end = vma->vm_end;
743         int ret;
744
745         /*
746          * Don't copy ptes where a page fault will fill them correctly.
747          * Fork becomes much lighter when there are big shared or private
748          * readonly mappings. The tradeoff is that copy_page_range is more
749          * efficient than faulting.
750          */
751         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
752                 if (!vma->anon_vma)
753                         return 0;
754         }
755
756         if (is_vm_hugetlb_page(vma))
757                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
758
759         if (unlikely(is_pfn_mapping(vma))) {
760                 /*
761                  * We do not free on error cases below as remove_vma
762                  * gets called on error from higher level routine
763                  */
764                 ret = track_pfn_vma_copy(vma);
765                 if (ret)
766                         return ret;
767         }
768
769         /*
770          * We need to invalidate the secondary MMU mappings only when
771          * there could be a permission downgrade on the ptes of the
772          * parent mm. And a permission downgrade will only happen if
773          * is_cow_mapping() returns true.
774          */
775         if (is_cow_mapping(vma->vm_flags))
776                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
777
778         ret = 0;
779         dst_pgd = pgd_offset(dst_mm, addr);
780         src_pgd = pgd_offset(src_mm, addr);
781         do {
782                 next = pgd_addr_end(addr, end);
783                 if (pgd_none_or_clear_bad(src_pgd))
784                         continue;
785                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
786                                             vma, addr, next))) {
787                         ret = -ENOMEM;
788                         break;
789                 }
790         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
791
792         if (is_cow_mapping(vma->vm_flags))
793                 mmu_notifier_invalidate_range_end(src_mm,
794                                                   vma->vm_start, end);
795         return ret;
796 }
797
798 static unsigned long zap_pte_range(struct mmu_gather *tlb,
799                                 struct vm_area_struct *vma, pmd_t *pmd,
800                                 unsigned long addr, unsigned long end,
801                                 long *zap_work, struct zap_details *details)
802 {
803         struct mm_struct *mm = tlb->mm;
804         pte_t *pte;
805         spinlock_t *ptl;
806         int file_rss = 0;
807         int anon_rss = 0;
808
809         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
810         arch_enter_lazy_mmu_mode();
811         do {
812                 pte_t ptent = *pte;
813                 if (pte_none(ptent)) {
814                         (*zap_work)--;
815                         continue;
816                 }
817
818                 (*zap_work) -= PAGE_SIZE;
819
820                 if (pte_present(ptent)) {
821                         struct page *page;
822
823                         page = vm_normal_page(vma, addr, ptent);
824                         if (unlikely(details) && page) {
825                                 /*
826                                  * unmap_shared_mapping_pages() wants to
827                                  * invalidate cache without truncating:
828                                  * unmap shared but keep private pages.
829                                  */
830                                 if (details->check_mapping &&
831                                     details->check_mapping != page->mapping)
832                                         continue;
833                                 /*
834                                  * Each page->index must be checked when
835                                  * invalidating or truncating nonlinear.
836                                  */
837                                 if (details->nonlinear_vma &&
838                                     (page->index < details->first_index ||
839                                      page->index > details->last_index))
840                                         continue;
841                         }
842                         ptent = ptep_get_and_clear_full(mm, addr, pte,
843                                                         tlb->fullmm);
844                         tlb_remove_tlb_entry(tlb, pte, addr);
845                         if (unlikely(!page))
846                                 continue;
847                         if (unlikely(details) && details->nonlinear_vma
848                             && linear_page_index(details->nonlinear_vma,
849                                                 addr) != page->index)
850                                 set_pte_at(mm, addr, pte,
851                                            pgoff_to_pte(page->index));
852                         if (PageAnon(page))
853                                 anon_rss--;
854                         else {
855                                 if (pte_dirty(ptent))
856                                         set_page_dirty(page);
857                                 if (pte_young(ptent) &&
858                                     likely(!VM_SequentialReadHint(vma)))
859                                         mark_page_accessed(page);
860                                 file_rss--;
861                         }
862                         page_remove_rmap(page);
863                         if (unlikely(page_mapcount(page) < 0))
864                                 print_bad_pte(vma, addr, ptent, page);
865                         tlb_remove_page(tlb, page);
866                         continue;
867                 }
868                 /*
869                  * If details->check_mapping, we leave swap entries;
870                  * if details->nonlinear_vma, we leave file entries.
871                  */
872                 if (unlikely(details))
873                         continue;
874                 if (pte_file(ptent)) {
875                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
876                                 print_bad_pte(vma, addr, ptent, NULL);
877                 } else if
878                   (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
879                         print_bad_pte(vma, addr, ptent, NULL);
880                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
881         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
882
883         add_mm_rss(mm, file_rss, anon_rss);
884         arch_leave_lazy_mmu_mode();
885         pte_unmap_unlock(pte - 1, ptl);
886
887         return addr;
888 }
889
890 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
891                                 struct vm_area_struct *vma, pud_t *pud,
892                                 unsigned long addr, unsigned long end,
893                                 long *zap_work, struct zap_details *details)
894 {
895         pmd_t *pmd;
896         unsigned long next;
897
898         pmd = pmd_offset(pud, addr);
899         do {
900                 next = pmd_addr_end(addr, end);
901                 if (pmd_none_or_clear_bad(pmd)) {
902                         (*zap_work)--;
903                         continue;
904                 }
905                 next = zap_pte_range(tlb, vma, pmd, addr, next,
906                                                 zap_work, details);
907         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
908
909         return addr;
910 }
911
912 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
913                                 struct vm_area_struct *vma, pgd_t *pgd,
914                                 unsigned long addr, unsigned long end,
915                                 long *zap_work, struct zap_details *details)
916 {
917         pud_t *pud;
918         unsigned long next;
919
920         pud = pud_offset(pgd, addr);
921         do {
922                 next = pud_addr_end(addr, end);
923                 if (pud_none_or_clear_bad(pud)) {
924                         (*zap_work)--;
925                         continue;
926                 }
927                 next = zap_pmd_range(tlb, vma, pud, addr, next,
928                                                 zap_work, details);
929         } while (pud++, addr = next, (addr != end && *zap_work > 0));
930
931         return addr;
932 }
933
934 static unsigned long unmap_page_range(struct mmu_gather *tlb,
935                                 struct vm_area_struct *vma,
936                                 unsigned long addr, unsigned long end,
937                                 long *zap_work, struct zap_details *details)
938 {
939         pgd_t *pgd;
940         unsigned long next;
941
942         if (details && !details->check_mapping && !details->nonlinear_vma)
943                 details = NULL;
944
945         BUG_ON(addr >= end);
946         tlb_start_vma(tlb, vma);
947         pgd = pgd_offset(vma->vm_mm, addr);
948         do {
949                 next = pgd_addr_end(addr, end);
950                 if (pgd_none_or_clear_bad(pgd)) {
951                         (*zap_work)--;
952                         continue;
953                 }
954                 next = zap_pud_range(tlb, vma, pgd, addr, next,
955                                                 zap_work, details);
956         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
957         tlb_end_vma(tlb, vma);
958
959         return addr;
960 }
961
962 #ifdef CONFIG_PREEMPT
963 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
964 #else
965 /* No preempt: go for improved straight-line efficiency */
966 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
967 #endif
968
969 /**
970  * unmap_vmas - unmap a range of memory covered by a list of vma's
971  * @tlbp: address of the caller's struct mmu_gather
972  * @vma: the starting vma
973  * @start_addr: virtual address at which to start unmapping
974  * @end_addr: virtual address at which to end unmapping
975  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
976  * @details: details of nonlinear truncation or shared cache invalidation
977  *
978  * Returns the end address of the unmapping (restart addr if interrupted).
979  *
980  * Unmap all pages in the vma list.
981  *
982  * We aim to not hold locks for too long (for scheduling latency reasons).
983  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
984  * return the ending mmu_gather to the caller.
985  *
986  * Only addresses between `start' and `end' will be unmapped.
987  *
988  * The VMA list must be sorted in ascending virtual address order.
989  *
990  * unmap_vmas() assumes that the caller will flush the whole unmapped address
991  * range after unmap_vmas() returns.  So the only responsibility here is to
992  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
993  * drops the lock and schedules.
994  */
995 unsigned long unmap_vmas(struct mmu_gather **tlbp,
996                 struct vm_area_struct *vma, unsigned long start_addr,
997                 unsigned long end_addr, unsigned long *nr_accounted,
998                 struct zap_details *details)
999 {
1000         long zap_work = ZAP_BLOCK_SIZE;
1001         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
1002         int tlb_start_valid = 0;
1003         unsigned long start = start_addr;
1004         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1005         int fullmm = (*tlbp)->fullmm;
1006         struct mm_struct *mm = vma->vm_mm;
1007
1008         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1009         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1010                 unsigned long end;
1011
1012                 start = max(vma->vm_start, start_addr);
1013                 if (start >= vma->vm_end)
1014                         continue;
1015                 end = min(vma->vm_end, end_addr);
1016                 if (end <= vma->vm_start)
1017                         continue;
1018
1019                 if (vma->vm_flags & VM_ACCOUNT)
1020                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1021
1022                 if (unlikely(is_pfn_mapping(vma)))
1023                         untrack_pfn_vma(vma, 0, 0);
1024
1025                 while (start != end) {
1026                         if (!tlb_start_valid) {
1027                                 tlb_start = start;
1028                                 tlb_start_valid = 1;
1029                         }
1030
1031                         if (unlikely(is_vm_hugetlb_page(vma))) {
1032                                 /*
1033                                  * It is undesirable to test vma->vm_file as it
1034                                  * should be non-null for valid hugetlb area.
1035                                  * However, vm_file will be NULL in the error
1036                                  * cleanup path of do_mmap_pgoff. When
1037                                  * hugetlbfs ->mmap method fails,
1038                                  * do_mmap_pgoff() nullifies vma->vm_file
1039                                  * before calling this function to clean up.
1040                                  * Since no pte has actually been setup, it is
1041                                  * safe to do nothing in this case.
1042                                  */
1043                                 if (vma->vm_file) {
1044                                         unmap_hugepage_range(vma, start, end, NULL);
1045                                         zap_work -= (end - start) /
1046                                         pages_per_huge_page(hstate_vma(vma));
1047                                 }
1048
1049                                 start = end;
1050                         } else
1051                                 start = unmap_page_range(*tlbp, vma,
1052                                                 start, end, &zap_work, details);
1053
1054                         if (zap_work > 0) {
1055                                 BUG_ON(start != end);
1056                                 break;
1057                         }
1058
1059                         tlb_finish_mmu(*tlbp, tlb_start, start);
1060
1061                         if (need_resched() ||
1062                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1063                                 if (i_mmap_lock) {
1064                                         *tlbp = NULL;
1065                                         goto out;
1066                                 }
1067                                 cond_resched();
1068                         }
1069
1070                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1071                         tlb_start_valid = 0;
1072                         zap_work = ZAP_BLOCK_SIZE;
1073                 }
1074         }
1075 out:
1076         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1077         return start;   /* which is now the end (or restart) address */
1078 }
1079
1080 /**
1081  * zap_page_range - remove user pages in a given range
1082  * @vma: vm_area_struct holding the applicable pages
1083  * @address: starting address of pages to zap
1084  * @size: number of bytes to zap
1085  * @details: details of nonlinear truncation or shared cache invalidation
1086  */
1087 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1088                 unsigned long size, struct zap_details *details)
1089 {
1090         struct mm_struct *mm = vma->vm_mm;
1091         struct mmu_gather *tlb;
1092         unsigned long end = address + size;
1093         unsigned long nr_accounted = 0;
1094
1095         lru_add_drain();
1096         tlb = tlb_gather_mmu(mm, 0);
1097         update_hiwater_rss(mm);
1098         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1099         if (tlb)
1100                 tlb_finish_mmu(tlb, address, end);
1101         return end;
1102 }
1103
1104 /**
1105  * zap_vma_ptes - remove ptes mapping the vma
1106  * @vma: vm_area_struct holding ptes to be zapped
1107  * @address: starting address of pages to zap
1108  * @size: number of bytes to zap
1109  *
1110  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1111  *
1112  * The entire address range must be fully contained within the vma.
1113  *
1114  * Returns 0 if successful.
1115  */
1116 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1117                 unsigned long size)
1118 {
1119         if (address < vma->vm_start || address + size > vma->vm_end ||
1120                         !(vma->vm_flags & VM_PFNMAP))
1121                 return -1;
1122         zap_page_range(vma, address, size, NULL);
1123         return 0;
1124 }
1125 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1126
1127 /*
1128  * Do a quick page-table lookup for a single page.
1129  */
1130 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1131                         unsigned int flags)
1132 {
1133         pgd_t *pgd;
1134         pud_t *pud;
1135         pmd_t *pmd;
1136         pte_t *ptep, pte;
1137         spinlock_t *ptl;
1138         struct page *page;
1139         struct mm_struct *mm = vma->vm_mm;
1140
1141         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1142         if (!IS_ERR(page)) {
1143                 BUG_ON(flags & FOLL_GET);
1144                 goto out;
1145         }
1146
1147         page = NULL;
1148         pgd = pgd_offset(mm, address);
1149         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1150                 goto no_page_table;
1151
1152         pud = pud_offset(pgd, address);
1153         if (pud_none(*pud))
1154                 goto no_page_table;
1155         if (pud_huge(*pud)) {
1156                 BUG_ON(flags & FOLL_GET);
1157                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1158                 goto out;
1159         }
1160         if (unlikely(pud_bad(*pud)))
1161                 goto no_page_table;
1162
1163         pmd = pmd_offset(pud, address);
1164         if (pmd_none(*pmd))
1165                 goto no_page_table;
1166         if (pmd_huge(*pmd)) {
1167                 BUG_ON(flags & FOLL_GET);
1168                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1169                 goto out;
1170         }
1171         if (unlikely(pmd_bad(*pmd)))
1172                 goto no_page_table;
1173
1174         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1175
1176         pte = *ptep;
1177         if (!pte_present(pte))
1178                 goto no_page;
1179         if ((flags & FOLL_WRITE) && !pte_write(pte))
1180                 goto unlock;
1181
1182         page = vm_normal_page(vma, address, pte);
1183         if (unlikely(!page)) {
1184                 if ((flags & FOLL_DUMP) ||
1185                     !is_zero_pfn(pte_pfn(pte)))
1186                         goto bad_page;
1187                 page = pte_page(pte);
1188         }
1189
1190         if (flags & FOLL_GET)
1191                 get_page(page);
1192         if (flags & FOLL_TOUCH) {
1193                 if ((flags & FOLL_WRITE) &&
1194                     !pte_dirty(pte) && !PageDirty(page))
1195                         set_page_dirty(page);
1196                 /*
1197                  * pte_mkyoung() would be more correct here, but atomic care
1198                  * is needed to avoid losing the dirty bit: it is easier to use
1199                  * mark_page_accessed().
1200                  */
1201                 mark_page_accessed(page);
1202         }
1203 unlock:
1204         pte_unmap_unlock(ptep, ptl);
1205 out:
1206         return page;
1207
1208 bad_page:
1209         pte_unmap_unlock(ptep, ptl);
1210         return ERR_PTR(-EFAULT);
1211
1212 no_page:
1213         pte_unmap_unlock(ptep, ptl);
1214         if (!pte_none(pte))
1215                 return page;
1216
1217 no_page_table:
1218         /*
1219          * When core dumping an enormous anonymous area that nobody
1220          * has touched so far, we don't want to allocate unnecessary pages or
1221          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1222          * then get_dump_page() will return NULL to leave a hole in the dump.
1223          * But we can only make this optimization where a hole would surely
1224          * be zero-filled if handle_mm_fault() actually did handle it.
1225          */
1226         if ((flags & FOLL_DUMP) &&
1227             (!vma->vm_ops || !vma->vm_ops->fault))
1228                 return ERR_PTR(-EFAULT);
1229         return page;
1230 }
1231 EXPORT_SYMBOL_GPL(follow_page);
1232
1233 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1234                      unsigned long start, int nr_pages, unsigned int gup_flags,
1235                      struct page **pages, struct vm_area_struct **vmas)
1236 {
1237         int i;
1238         unsigned long vm_flags;
1239
1240         if (nr_pages <= 0)
1241                 return 0;
1242
1243         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1244
1245         /* 
1246          * Require read or write permissions.
1247          * If FOLL_FORCE is set, we only require the "MAY" flags.
1248          */
1249         vm_flags  = (gup_flags & FOLL_WRITE) ?
1250                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1251         vm_flags &= (gup_flags & FOLL_FORCE) ?
1252                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1253         i = 0;
1254
1255         do {
1256                 struct vm_area_struct *vma;
1257
1258                 vma = find_extend_vma(mm, start);
1259                 if (!vma && in_gate_area(tsk, start)) {
1260                         unsigned long pg = start & PAGE_MASK;
1261                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1262                         pgd_t *pgd;
1263                         pud_t *pud;
1264                         pmd_t *pmd;
1265                         pte_t *pte;
1266
1267                         /* user gate pages are read-only */
1268                         if (gup_flags & FOLL_WRITE)
1269                                 return i ? : -EFAULT;
1270                         if (pg > TASK_SIZE)
1271                                 pgd = pgd_offset_k(pg);
1272                         else
1273                                 pgd = pgd_offset_gate(mm, pg);
1274                         BUG_ON(pgd_none(*pgd));
1275                         pud = pud_offset(pgd, pg);
1276                         BUG_ON(pud_none(*pud));
1277                         pmd = pmd_offset(pud, pg);
1278                         if (pmd_none(*pmd))
1279                                 return i ? : -EFAULT;
1280                         pte = pte_offset_map(pmd, pg);
1281                         if (pte_none(*pte)) {
1282                                 pte_unmap(pte);
1283                                 return i ? : -EFAULT;
1284                         }
1285                         if (pages) {
1286                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1287                                 pages[i] = page;
1288                                 if (page)
1289                                         get_page(page);
1290                         }
1291                         pte_unmap(pte);
1292                         if (vmas)
1293                                 vmas[i] = gate_vma;
1294                         i++;
1295                         start += PAGE_SIZE;
1296                         nr_pages--;
1297                         continue;
1298                 }
1299
1300                 if (!vma ||
1301                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1302                     !(vm_flags & vma->vm_flags))
1303                         return i ? : -EFAULT;
1304
1305                 if (is_vm_hugetlb_page(vma)) {
1306                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1307                                         &start, &nr_pages, i, gup_flags);
1308                         continue;
1309                 }
1310
1311                 do {
1312                         struct page *page;
1313                         unsigned int foll_flags = gup_flags;
1314
1315                         /*
1316                          * If we have a pending SIGKILL, don't keep faulting
1317                          * pages and potentially allocating memory.
1318                          */
1319                         if (unlikely(fatal_signal_pending(current)))
1320                                 return i ? i : -ERESTARTSYS;
1321
1322                         cond_resched();
1323                         while (!(page = follow_page(vma, start, foll_flags))) {
1324                                 int ret;
1325
1326                                 ret = handle_mm_fault(mm, vma, start,
1327                                         (foll_flags & FOLL_WRITE) ?
1328                                         FAULT_FLAG_WRITE : 0);
1329
1330                                 if (ret & VM_FAULT_ERROR) {
1331                                         if (ret & VM_FAULT_OOM)
1332                                                 return i ? i : -ENOMEM;
1333                                         if (ret &
1334                                             (VM_FAULT_HWPOISON|VM_FAULT_SIGBUS))
1335                                                 return i ? i : -EFAULT;
1336                                         BUG();
1337                                 }
1338                                 if (ret & VM_FAULT_MAJOR)
1339                                         tsk->maj_flt++;
1340                                 else
1341                                         tsk->min_flt++;
1342
1343                                 /*
1344                                  * The VM_FAULT_WRITE bit tells us that
1345                                  * do_wp_page has broken COW when necessary,
1346                                  * even if maybe_mkwrite decided not to set
1347                                  * pte_write. We can thus safely do subsequent
1348                                  * page lookups as if they were reads. But only
1349                                  * do so when looping for pte_write is futile:
1350                                  * in some cases userspace may also be wanting
1351                                  * to write to the gotten user page, which a
1352                                  * read fault here might prevent (a readonly
1353                                  * page might get reCOWed by userspace write).
1354                                  */
1355                                 if ((ret & VM_FAULT_WRITE) &&
1356                                     !(vma->vm_flags & VM_WRITE))
1357                                         foll_flags &= ~FOLL_WRITE;
1358
1359                                 cond_resched();
1360                         }
1361                         if (IS_ERR(page))
1362                                 return i ? i : PTR_ERR(page);
1363                         if (pages) {
1364                                 pages[i] = page;
1365
1366                                 flush_anon_page(vma, page, start);
1367                                 flush_dcache_page(page);
1368                         }
1369                         if (vmas)
1370                                 vmas[i] = vma;
1371                         i++;
1372                         start += PAGE_SIZE;
1373                         nr_pages--;
1374                 } while (nr_pages && start < vma->vm_end);
1375         } while (nr_pages);
1376         return i;
1377 }
1378
1379 /**
1380  * get_user_pages() - pin user pages in memory
1381  * @tsk:        task_struct of target task
1382  * @mm:         mm_struct of target mm
1383  * @start:      starting user address
1384  * @nr_pages:   number of pages from start to pin
1385  * @write:      whether pages will be written to by the caller
1386  * @force:      whether to force write access even if user mapping is
1387  *              readonly. This will result in the page being COWed even
1388  *              in MAP_SHARED mappings. You do not want this.
1389  * @pages:      array that receives pointers to the pages pinned.
1390  *              Should be at least nr_pages long. Or NULL, if caller
1391  *              only intends to ensure the pages are faulted in.
1392  * @vmas:       array of pointers to vmas corresponding to each page.
1393  *              Or NULL if the caller does not require them.
1394  *
1395  * Returns number of pages pinned. This may be fewer than the number
1396  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1397  * were pinned, returns -errno. Each page returned must be released
1398  * with a put_page() call when it is finished with. vmas will only
1399  * remain valid while mmap_sem is held.
1400  *
1401  * Must be called with mmap_sem held for read or write.
1402  *
1403  * get_user_pages walks a process's page tables and takes a reference to
1404  * each struct page that each user address corresponds to at a given
1405  * instant. That is, it takes the page that would be accessed if a user
1406  * thread accesses the given user virtual address at that instant.
1407  *
1408  * This does not guarantee that the page exists in the user mappings when
1409  * get_user_pages returns, and there may even be a completely different
1410  * page there in some cases (eg. if mmapped pagecache has been invalidated
1411  * and subsequently re faulted). However it does guarantee that the page
1412  * won't be freed completely. And mostly callers simply care that the page
1413  * contains data that was valid *at some point in time*. Typically, an IO
1414  * or similar operation cannot guarantee anything stronger anyway because
1415  * locks can't be held over the syscall boundary.
1416  *
1417  * If write=0, the page must not be written to. If the page is written to,
1418  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1419  * after the page is finished with, and before put_page is called.
1420  *
1421  * get_user_pages is typically used for fewer-copy IO operations, to get a
1422  * handle on the memory by some means other than accesses via the user virtual
1423  * addresses. The pages may be submitted for DMA to devices or accessed via
1424  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1425  * use the correct cache flushing APIs.
1426  *
1427  * See also get_user_pages_fast, for performance critical applications.
1428  */
1429 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1430                 unsigned long start, int nr_pages, int write, int force,
1431                 struct page **pages, struct vm_area_struct **vmas)
1432 {
1433         int flags = FOLL_TOUCH;
1434
1435         if (pages)
1436                 flags |= FOLL_GET;
1437         if (write)
1438                 flags |= FOLL_WRITE;
1439         if (force)
1440                 flags |= FOLL_FORCE;
1441
1442         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
1443 }
1444 EXPORT_SYMBOL(get_user_pages);
1445
1446 /**
1447  * get_dump_page() - pin user page in memory while writing it to core dump
1448  * @addr: user address
1449  *
1450  * Returns struct page pointer of user page pinned for dump,
1451  * to be freed afterwards by page_cache_release() or put_page().
1452  *
1453  * Returns NULL on any kind of failure - a hole must then be inserted into
1454  * the corefile, to preserve alignment with its headers; and also returns
1455  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1456  * allowing a hole to be left in the corefile to save diskspace.
1457  *
1458  * Called without mmap_sem, but after all other threads have been killed.
1459  */
1460 #ifdef CONFIG_ELF_CORE
1461 struct page *get_dump_page(unsigned long addr)
1462 {
1463         struct vm_area_struct *vma;
1464         struct page *page;
1465
1466         if (__get_user_pages(current, current->mm, addr, 1,
1467                         FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma) < 1)
1468                 return NULL;
1469         flush_cache_page(vma, addr, page_to_pfn(page));
1470         return page;
1471 }
1472 #endif /* CONFIG_ELF_CORE */
1473
1474 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1475                         spinlock_t **ptl)
1476 {
1477         pgd_t * pgd = pgd_offset(mm, addr);
1478         pud_t * pud = pud_alloc(mm, pgd, addr);
1479         if (pud) {
1480                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1481                 if (pmd)
1482                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1483         }
1484         return NULL;
1485 }
1486
1487 /*
1488  * This is the old fallback for page remapping.
1489  *
1490  * For historical reasons, it only allows reserved pages. Only
1491  * old drivers should use this, and they needed to mark their
1492  * pages reserved for the old functions anyway.
1493  */
1494 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1495                         struct page *page, pgprot_t prot)
1496 {
1497         struct mm_struct *mm = vma->vm_mm;
1498         int retval;
1499         pte_t *pte;
1500         spinlock_t *ptl;
1501
1502         retval = -EINVAL;
1503         if (PageAnon(page))
1504                 goto out;
1505         retval = -ENOMEM;
1506         flush_dcache_page(page);
1507         pte = get_locked_pte(mm, addr, &ptl);
1508         if (!pte)
1509                 goto out;
1510         retval = -EBUSY;
1511         if (!pte_none(*pte))
1512                 goto out_unlock;
1513
1514         /* Ok, finally just insert the thing.. */
1515         get_page(page);
1516         inc_mm_counter(mm, file_rss);
1517         page_add_file_rmap(page);
1518         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1519
1520         retval = 0;
1521         pte_unmap_unlock(pte, ptl);
1522         return retval;
1523 out_unlock:
1524         pte_unmap_unlock(pte, ptl);
1525 out:
1526         return retval;
1527 }
1528
1529 /**
1530  * vm_insert_page - insert single page into user vma
1531  * @vma: user vma to map to
1532  * @addr: target user address of this page
1533  * @page: source kernel page
1534  *
1535  * This allows drivers to insert individual pages they've allocated
1536  * into a user vma.
1537  *
1538  * The page has to be a nice clean _individual_ kernel allocation.
1539  * If you allocate a compound page, you need to have marked it as
1540  * such (__GFP_COMP), or manually just split the page up yourself
1541  * (see split_page()).
1542  *
1543  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1544  * took an arbitrary page protection parameter. This doesn't allow
1545  * that. Your vma protection will have to be set up correctly, which
1546  * means that if you want a shared writable mapping, you'd better
1547  * ask for a shared writable mapping!
1548  *
1549  * The page does not need to be reserved.
1550  */
1551 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1552                         struct page *page)
1553 {
1554         if (addr < vma->vm_start || addr >= vma->vm_end)
1555                 return -EFAULT;
1556         if (!page_count(page))
1557                 return -EINVAL;
1558         vma->vm_flags |= VM_INSERTPAGE;
1559         return insert_page(vma, addr, page, vma->vm_page_prot);
1560 }
1561 EXPORT_SYMBOL(vm_insert_page);
1562
1563 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1564                         unsigned long pfn, pgprot_t prot)
1565 {
1566         struct mm_struct *mm = vma->vm_mm;
1567         int retval;
1568         pte_t *pte, entry;
1569         spinlock_t *ptl;
1570
1571         retval = -ENOMEM;
1572         pte = get_locked_pte(mm, addr, &ptl);
1573         if (!pte)
1574                 goto out;
1575         retval = -EBUSY;
1576         if (!pte_none(*pte))
1577                 goto out_unlock;
1578
1579         /* Ok, finally just insert the thing.. */
1580         entry = pte_mkspecial(pfn_pte(pfn, prot));
1581         set_pte_at(mm, addr, pte, entry);
1582         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1583
1584         retval = 0;
1585 out_unlock:
1586         pte_unmap_unlock(pte, ptl);
1587 out:
1588         return retval;
1589 }
1590
1591 /**
1592  * vm_insert_pfn - insert single pfn into user vma
1593  * @vma: user vma to map to
1594  * @addr: target user address of this page
1595  * @pfn: source kernel pfn
1596  *
1597  * Similar to vm_inert_page, this allows drivers to insert individual pages
1598  * they've allocated into a user vma. Same comments apply.
1599  *
1600  * This function should only be called from a vm_ops->fault handler, and
1601  * in that case the handler should return NULL.
1602  *
1603  * vma cannot be a COW mapping.
1604  *
1605  * As this is called only for pages that do not currently exist, we
1606  * do not need to flush old virtual caches or the TLB.
1607  */
1608 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1609                         unsigned long pfn)
1610 {
1611         int ret;
1612         pgprot_t pgprot = vma->vm_page_prot;
1613         /*
1614          * Technically, architectures with pte_special can avoid all these
1615          * restrictions (same for remap_pfn_range).  However we would like
1616          * consistency in testing and feature parity among all, so we should
1617          * try to keep these invariants in place for everybody.
1618          */
1619         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1620         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1621                                                 (VM_PFNMAP|VM_MIXEDMAP));
1622         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1623         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1624
1625         if (addr < vma->vm_start || addr >= vma->vm_end)
1626                 return -EFAULT;
1627         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1628                 return -EINVAL;
1629
1630         ret = insert_pfn(vma, addr, pfn, pgprot);
1631
1632         if (ret)
1633                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1634
1635         return ret;
1636 }
1637 EXPORT_SYMBOL(vm_insert_pfn);
1638
1639 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1640                         unsigned long pfn)
1641 {
1642         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1643
1644         if (addr < vma->vm_start || addr >= vma->vm_end)
1645                 return -EFAULT;
1646
1647         /*
1648          * If we don't have pte special, then we have to use the pfn_valid()
1649          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1650          * refcount the page if pfn_valid is true (hence insert_page rather
1651          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1652          * without pte special, it would there be refcounted as a normal page.
1653          */
1654         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1655                 struct page *page;
1656
1657                 page = pfn_to_page(pfn);
1658                 return insert_page(vma, addr, page, vma->vm_page_prot);
1659         }
1660         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1661 }
1662 EXPORT_SYMBOL(vm_insert_mixed);
1663
1664 /*
1665  * maps a range of physical memory into the requested pages. the old
1666  * mappings are removed. any references to nonexistent pages results
1667  * in null mappings (currently treated as "copy-on-access")
1668  */
1669 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1670                         unsigned long addr, unsigned long end,
1671                         unsigned long pfn, pgprot_t prot)
1672 {
1673         pte_t *pte;
1674         spinlock_t *ptl;
1675
1676         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1677         if (!pte)
1678                 return -ENOMEM;
1679         arch_enter_lazy_mmu_mode();
1680         do {
1681                 BUG_ON(!pte_none(*pte));
1682                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1683                 pfn++;
1684         } while (pte++, addr += PAGE_SIZE, addr != end);
1685         arch_leave_lazy_mmu_mode();
1686         pte_unmap_unlock(pte - 1, ptl);
1687         return 0;
1688 }
1689
1690 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1691                         unsigned long addr, unsigned long end,
1692                         unsigned long pfn, pgprot_t prot)
1693 {
1694         pmd_t *pmd;
1695         unsigned long next;
1696
1697         pfn -= addr >> PAGE_SHIFT;
1698         pmd = pmd_alloc(mm, pud, addr);
1699         if (!pmd)
1700                 return -ENOMEM;
1701         do {
1702                 next = pmd_addr_end(addr, end);
1703                 if (remap_pte_range(mm, pmd, addr, next,
1704                                 pfn + (addr >> PAGE_SHIFT), prot))
1705                         return -ENOMEM;
1706         } while (pmd++, addr = next, addr != end);
1707         return 0;
1708 }
1709
1710 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1711                         unsigned long addr, unsigned long end,
1712                         unsigned long pfn, pgprot_t prot)
1713 {
1714         pud_t *pud;
1715         unsigned long next;
1716
1717         pfn -= addr >> PAGE_SHIFT;
1718         pud = pud_alloc(mm, pgd, addr);
1719         if (!pud)
1720                 return -ENOMEM;
1721         do {
1722                 next = pud_addr_end(addr, end);
1723                 if (remap_pmd_range(mm, pud, addr, next,
1724                                 pfn + (addr >> PAGE_SHIFT), prot))
1725                         return -ENOMEM;
1726         } while (pud++, addr = next, addr != end);
1727         return 0;
1728 }
1729
1730 /**
1731  * remap_pfn_range - remap kernel memory to userspace
1732  * @vma: user vma to map to
1733  * @addr: target user address to start at
1734  * @pfn: physical address of kernel memory
1735  * @size: size of map area
1736  * @prot: page protection flags for this mapping
1737  *
1738  *  Note: this is only safe if the mm semaphore is held when called.
1739  */
1740 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1741                     unsigned long pfn, unsigned long size, pgprot_t prot)
1742 {
1743         pgd_t *pgd;
1744         unsigned long next;
1745         unsigned long end = addr + PAGE_ALIGN(size);
1746         struct mm_struct *mm = vma->vm_mm;
1747         int err;
1748
1749         /*
1750          * Physically remapped pages are special. Tell the
1751          * rest of the world about it:
1752          *   VM_IO tells people not to look at these pages
1753          *      (accesses can have side effects).
1754          *   VM_RESERVED is specified all over the place, because
1755          *      in 2.4 it kept swapout's vma scan off this vma; but
1756          *      in 2.6 the LRU scan won't even find its pages, so this
1757          *      flag means no more than count its pages in reserved_vm,
1758          *      and omit it from core dump, even when VM_IO turned off.
1759          *   VM_PFNMAP tells the core MM that the base pages are just
1760          *      raw PFN mappings, and do not have a "struct page" associated
1761          *      with them.
1762          *
1763          * There's a horrible special case to handle copy-on-write
1764          * behaviour that some programs depend on. We mark the "original"
1765          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1766          */
1767         if (addr == vma->vm_start && end == vma->vm_end) {
1768                 vma->vm_pgoff = pfn;
1769                 vma->vm_flags |= VM_PFN_AT_MMAP;
1770         } else if (is_cow_mapping(vma->vm_flags))
1771                 return -EINVAL;
1772
1773         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1774
1775         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1776         if (err) {
1777                 /*
1778                  * To indicate that track_pfn related cleanup is not
1779                  * needed from higher level routine calling unmap_vmas
1780                  */
1781                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1782                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1783                 return -EINVAL;
1784         }
1785
1786         BUG_ON(addr >= end);
1787         pfn -= addr >> PAGE_SHIFT;
1788         pgd = pgd_offset(mm, addr);
1789         flush_cache_range(vma, addr, end);
1790         do {
1791                 next = pgd_addr_end(addr, end);
1792                 err = remap_pud_range(mm, pgd, addr, next,
1793                                 pfn + (addr >> PAGE_SHIFT), prot);
1794                 if (err)
1795                         break;
1796         } while (pgd++, addr = next, addr != end);
1797
1798         if (err)
1799                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1800
1801         return err;
1802 }
1803 EXPORT_SYMBOL(remap_pfn_range);
1804
1805 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1806                                      unsigned long addr, unsigned long end,
1807                                      pte_fn_t fn, void *data)
1808 {
1809         pte_t *pte;
1810         int err;
1811         pgtable_t token;
1812         spinlock_t *uninitialized_var(ptl);
1813
1814         pte = (mm == &init_mm) ?
1815                 pte_alloc_kernel(pmd, addr) :
1816                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1817         if (!pte)
1818                 return -ENOMEM;
1819
1820         BUG_ON(pmd_huge(*pmd));
1821
1822         arch_enter_lazy_mmu_mode();
1823
1824         token = pmd_pgtable(*pmd);
1825
1826         do {
1827                 err = fn(pte++, token, addr, data);
1828                 if (err)
1829                         break;
1830         } while (addr += PAGE_SIZE, addr != end);
1831
1832         arch_leave_lazy_mmu_mode();
1833
1834         if (mm != &init_mm)
1835                 pte_unmap_unlock(pte-1, ptl);
1836         return err;
1837 }
1838
1839 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1840                                      unsigned long addr, unsigned long end,
1841                                      pte_fn_t fn, void *data)
1842 {
1843         pmd_t *pmd;
1844         unsigned long next;
1845         int err;
1846
1847         BUG_ON(pud_huge(*pud));
1848
1849         pmd = pmd_alloc(mm, pud, addr);
1850         if (!pmd)
1851                 return -ENOMEM;
1852         do {
1853                 next = pmd_addr_end(addr, end);
1854                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1855                 if (err)
1856                         break;
1857         } while (pmd++, addr = next, addr != end);
1858         return err;
1859 }
1860
1861 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1862                                      unsigned long addr, unsigned long end,
1863                                      pte_fn_t fn, void *data)
1864 {
1865         pud_t *pud;
1866         unsigned long next;
1867         int err;
1868
1869         pud = pud_alloc(mm, pgd, addr);
1870         if (!pud)
1871                 return -ENOMEM;
1872         do {
1873                 next = pud_addr_end(addr, end);
1874                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1875                 if (err)
1876                         break;
1877         } while (pud++, addr = next, addr != end);
1878         return err;
1879 }
1880
1881 /*
1882  * Scan a region of virtual memory, filling in page tables as necessary
1883  * and calling a provided function on each leaf page table.
1884  */
1885 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1886                         unsigned long size, pte_fn_t fn, void *data)
1887 {
1888         pgd_t *pgd;
1889         unsigned long next;
1890         unsigned long start = addr, end = addr + size;
1891         int err;
1892
1893         BUG_ON(addr >= end);
1894         mmu_notifier_invalidate_range_start(mm, start, end);
1895         pgd = pgd_offset(mm, addr);
1896         do {
1897                 next = pgd_addr_end(addr, end);
1898                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1899                 if (err)
1900                         break;
1901         } while (pgd++, addr = next, addr != end);
1902         mmu_notifier_invalidate_range_end(mm, start, end);
1903         return err;
1904 }
1905 EXPORT_SYMBOL_GPL(apply_to_page_range);
1906
1907 /*
1908  * handle_pte_fault chooses page fault handler according to an entry
1909  * which was read non-atomically.  Before making any commitment, on
1910  * those architectures or configurations (e.g. i386 with PAE) which
1911  * might give a mix of unmatched parts, do_swap_page and do_file_page
1912  * must check under lock before unmapping the pte and proceeding
1913  * (but do_wp_page is only called after already making such a check;
1914  * and do_anonymous_page and do_no_page can safely check later on).
1915  */
1916 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1917                                 pte_t *page_table, pte_t orig_pte)
1918 {
1919         int same = 1;
1920 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1921         if (sizeof(pte_t) > sizeof(unsigned long)) {
1922                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1923                 spin_lock(ptl);
1924                 same = pte_same(*page_table, orig_pte);
1925                 spin_unlock(ptl);
1926         }
1927 #endif
1928         pte_unmap(page_table);
1929         return same;
1930 }
1931
1932 /*
1933  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1934  * servicing faults for write access.  In the normal case, do always want
1935  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1936  * that do not have writing enabled, when used by access_process_vm.
1937  */
1938 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1939 {
1940         if (likely(vma->vm_flags & VM_WRITE))
1941                 pte = pte_mkwrite(pte);
1942         return pte;
1943 }
1944
1945 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1946 {
1947         /*
1948          * If the source page was a PFN mapping, we don't have
1949          * a "struct page" for it. We do a best-effort copy by
1950          * just copying from the original user address. If that
1951          * fails, we just zero-fill it. Live with it.
1952          */
1953         if (unlikely(!src)) {
1954                 void *kaddr = kmap_atomic(dst, KM_USER0);
1955                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1956
1957                 /*
1958                  * This really shouldn't fail, because the page is there
1959                  * in the page tables. But it might just be unreadable,
1960                  * in which case we just give up and fill the result with
1961                  * zeroes.
1962                  */
1963                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1964                         memset(kaddr, 0, PAGE_SIZE);
1965                 kunmap_atomic(kaddr, KM_USER0);
1966                 flush_dcache_page(dst);
1967         } else
1968                 copy_user_highpage(dst, src, va, vma);
1969 }
1970
1971 /*
1972  * This routine handles present pages, when users try to write
1973  * to a shared page. It is done by copying the page to a new address
1974  * and decrementing the shared-page counter for the old page.
1975  *
1976  * Note that this routine assumes that the protection checks have been
1977  * done by the caller (the low-level page fault routine in most cases).
1978  * Thus we can safely just mark it writable once we've done any necessary
1979  * COW.
1980  *
1981  * We also mark the page dirty at this point even though the page will
1982  * change only once the write actually happens. This avoids a few races,
1983  * and potentially makes it more efficient.
1984  *
1985  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1986  * but allow concurrent faults), with pte both mapped and locked.
1987  * We return with mmap_sem still held, but pte unmapped and unlocked.
1988  */
1989 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1990                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1991                 spinlock_t *ptl, pte_t orig_pte)
1992 {
1993         struct page *old_page, *new_page;
1994         pte_t entry;
1995         int reuse = 0, ret = 0;
1996         int page_mkwrite = 0;
1997         struct page *dirty_page = NULL;
1998
1999         old_page = vm_normal_page(vma, address, orig_pte);
2000         if (!old_page) {
2001                 /*
2002                  * VM_MIXEDMAP !pfn_valid() case
2003                  *
2004                  * We should not cow pages in a shared writeable mapping.
2005                  * Just mark the pages writable as we can't do any dirty
2006                  * accounting on raw pfn maps.
2007                  */
2008                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2009                                      (VM_WRITE|VM_SHARED))
2010                         goto reuse;
2011                 goto gotten;
2012         }
2013
2014         /*
2015          * Take out anonymous pages first, anonymous shared vmas are
2016          * not dirty accountable.
2017          */
2018         if (PageAnon(old_page) && !PageKsm(old_page)) {
2019                 if (!trylock_page(old_page)) {
2020                         page_cache_get(old_page);
2021                         pte_unmap_unlock(page_table, ptl);
2022                         lock_page(old_page);
2023                         page_table = pte_offset_map_lock(mm, pmd, address,
2024                                                          &ptl);
2025                         if (!pte_same(*page_table, orig_pte)) {
2026                                 unlock_page(old_page);
2027                                 page_cache_release(old_page);
2028                                 goto unlock;
2029                         }
2030                         page_cache_release(old_page);
2031                 }
2032                 reuse = reuse_swap_page(old_page);
2033                 unlock_page(old_page);
2034         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2035                                         (VM_WRITE|VM_SHARED))) {
2036                 /*
2037                  * Only catch write-faults on shared writable pages,
2038                  * read-only shared pages can get COWed by
2039                  * get_user_pages(.write=1, .force=1).
2040                  */
2041                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2042                         struct vm_fault vmf;
2043                         int tmp;
2044
2045                         vmf.virtual_address = (void __user *)(address &
2046                                                                 PAGE_MASK);
2047                         vmf.pgoff = old_page->index;
2048                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2049                         vmf.page = old_page;
2050
2051                         /*
2052                          * Notify the address space that the page is about to
2053                          * become writable so that it can prohibit this or wait
2054                          * for the page to get into an appropriate state.
2055                          *
2056                          * We do this without the lock held, so that it can
2057                          * sleep if it needs to.
2058                          */
2059                         page_cache_get(old_page);
2060                         pte_unmap_unlock(page_table, ptl);
2061
2062                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2063                         if (unlikely(tmp &
2064                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2065                                 ret = tmp;
2066                                 goto unwritable_page;
2067                         }
2068                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2069                                 lock_page(old_page);
2070                                 if (!old_page->mapping) {
2071                                         ret = 0; /* retry the fault */
2072                                         unlock_page(old_page);
2073                                         goto unwritable_page;
2074                                 }
2075                         } else
2076                                 VM_BUG_ON(!PageLocked(old_page));
2077
2078                         /*
2079                          * Since we dropped the lock we need to revalidate
2080                          * the PTE as someone else may have changed it.  If
2081                          * they did, we just return, as we can count on the
2082                          * MMU to tell us if they didn't also make it writable.
2083                          */
2084                         page_table = pte_offset_map_lock(mm, pmd, address,
2085                                                          &ptl);
2086                         if (!pte_same(*page_table, orig_pte)) {
2087                                 unlock_page(old_page);
2088                                 page_cache_release(old_page);
2089                                 goto unlock;
2090                         }
2091
2092                         page_mkwrite = 1;
2093                 }
2094                 dirty_page = old_page;
2095                 get_page(dirty_page);
2096                 reuse = 1;
2097         }
2098
2099         if (reuse) {
2100 reuse:
2101                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2102                 entry = pte_mkyoung(orig_pte);
2103                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2104                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2105                         update_mmu_cache(vma, address, entry);
2106                 ret |= VM_FAULT_WRITE;
2107                 goto unlock;
2108         }
2109
2110         /*
2111          * Ok, we need to copy. Oh, well..
2112          */
2113         page_cache_get(old_page);
2114 gotten:
2115         pte_unmap_unlock(page_table, ptl);
2116
2117         if (unlikely(anon_vma_prepare(vma)))
2118                 goto oom;
2119
2120         if (is_zero_pfn(pte_pfn(orig_pte))) {
2121                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2122                 if (!new_page)
2123                         goto oom;
2124         } else {
2125                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2126                 if (!new_page)
2127                         goto oom;
2128                 cow_user_page(new_page, old_page, address, vma);
2129         }
2130         __SetPageUptodate(new_page);
2131
2132         /*
2133          * Don't let another task, with possibly unlocked vma,
2134          * keep the mlocked page.
2135          */
2136         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2137                 lock_page(old_page);    /* for LRU manipulation */
2138                 clear_page_mlock(old_page);
2139                 unlock_page(old_page);
2140         }
2141
2142         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2143                 goto oom_free_new;
2144
2145         /*
2146          * Re-check the pte - we dropped the lock
2147          */
2148         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2149         if (likely(pte_same(*page_table, orig_pte))) {
2150                 if (old_page) {
2151                         if (!PageAnon(old_page)) {
2152                                 dec_mm_counter(mm, file_rss);
2153                                 inc_mm_counter(mm, anon_rss);
2154                         }
2155                 } else
2156                         inc_mm_counter(mm, anon_rss);
2157                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2158                 entry = mk_pte(new_page, vma->vm_page_prot);
2159                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2160                 /*
2161                  * Clear the pte entry and flush it first, before updating the
2162                  * pte with the new entry. This will avoid a race condition
2163                  * seen in the presence of one thread doing SMC and another
2164                  * thread doing COW.
2165                  */
2166                 ptep_clear_flush(vma, address, page_table);
2167                 page_add_new_anon_rmap(new_page, vma, address);
2168                 /*
2169                  * We call the notify macro here because, when using secondary
2170                  * mmu page tables (such as kvm shadow page tables), we want the
2171                  * new page to be mapped directly into the secondary page table.
2172                  */
2173                 set_pte_at_notify(mm, address, page_table, entry);
2174                 update_mmu_cache(vma, address, entry);
2175                 if (old_page) {
2176                         /*
2177                          * Only after switching the pte to the new page may
2178                          * we remove the mapcount here. Otherwise another
2179                          * process may come and find the rmap count decremented
2180                          * before the pte is switched to the new page, and
2181                          * "reuse" the old page writing into it while our pte
2182                          * here still points into it and can be read by other
2183                          * threads.
2184                          *
2185                          * The critical issue is to order this
2186                          * page_remove_rmap with the ptp_clear_flush above.
2187                          * Those stores are ordered by (if nothing else,)
2188                          * the barrier present in the atomic_add_negative
2189                          * in page_remove_rmap.
2190                          *
2191                          * Then the TLB flush in ptep_clear_flush ensures that
2192                          * no process can access the old page before the
2193                          * decremented mapcount is visible. And the old page
2194                          * cannot be reused until after the decremented
2195                          * mapcount is visible. So transitively, TLBs to
2196                          * old page will be flushed before it can be reused.
2197                          */
2198                         page_remove_rmap(old_page);
2199                 }
2200
2201                 /* Free the old page.. */
2202                 new_page = old_page;
2203                 ret |= VM_FAULT_WRITE;
2204         } else
2205                 mem_cgroup_uncharge_page(new_page);
2206
2207         if (new_page)
2208                 page_cache_release(new_page);
2209         if (old_page)
2210                 page_cache_release(old_page);
2211 unlock:
2212         pte_unmap_unlock(page_table, ptl);
2213         if (dirty_page) {
2214                 /*
2215                  * Yes, Virginia, this is actually required to prevent a race
2216                  * with clear_page_dirty_for_io() from clearing the page dirty
2217                  * bit after it clear all dirty ptes, but before a racing
2218                  * do_wp_page installs a dirty pte.
2219                  *
2220                  * do_no_page is protected similarly.
2221                  */
2222                 if (!page_mkwrite) {
2223                         wait_on_page_locked(dirty_page);
2224                         set_page_dirty_balance(dirty_page, page_mkwrite);
2225                 }
2226                 put_page(dirty_page);
2227                 if (page_mkwrite) {
2228                         struct address_space *mapping = dirty_page->mapping;
2229
2230                         set_page_dirty(dirty_page);
2231                         unlock_page(dirty_page);
2232                         page_cache_release(dirty_page);
2233                         if (mapping)    {
2234                                 /*
2235                                  * Some device drivers do not set page.mapping
2236                                  * but still dirty their pages
2237                                  */
2238                                 balance_dirty_pages_ratelimited(mapping);
2239                         }
2240                 }
2241
2242                 /* file_update_time outside page_lock */
2243                 if (vma->vm_file)
2244                         file_update_time(vma->vm_file);
2245         }
2246         return ret;
2247 oom_free_new:
2248         page_cache_release(new_page);
2249 oom:
2250         if (old_page) {
2251                 if (page_mkwrite) {
2252                         unlock_page(old_page);
2253                         page_cache_release(old_page);
2254                 }
2255                 page_cache_release(old_page);
2256         }
2257         return VM_FAULT_OOM;
2258
2259 unwritable_page:
2260         page_cache_release(old_page);
2261         return ret;
2262 }
2263
2264 /*
2265  * Helper functions for unmap_mapping_range().
2266  *
2267  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2268  *
2269  * We have to restart searching the prio_tree whenever we drop the lock,
2270  * since the iterator is only valid while the lock is held, and anyway
2271  * a later vma might be split and reinserted earlier while lock dropped.
2272  *
2273  * The list of nonlinear vmas could be handled more efficiently, using
2274  * a placeholder, but handle it in the same way until a need is shown.
2275  * It is important to search the prio_tree before nonlinear list: a vma
2276  * may become nonlinear and be shifted from prio_tree to nonlinear list
2277  * while the lock is dropped; but never shifted from list to prio_tree.
2278  *
2279  * In order to make forward progress despite restarting the search,
2280  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2281  * quickly skip it next time around.  Since the prio_tree search only
2282  * shows us those vmas affected by unmapping the range in question, we
2283  * can't efficiently keep all vmas in step with mapping->truncate_count:
2284  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2285  * mapping->truncate_count and vma->vm_truncate_count are protected by
2286  * i_mmap_lock.
2287  *
2288  * In order to make forward progress despite repeatedly restarting some
2289  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2290  * and restart from that address when we reach that vma again.  It might
2291  * have been split or merged, shrunk or extended, but never shifted: so
2292  * restart_addr remains valid so long as it remains in the vma's range.
2293  * unmap_mapping_range forces truncate_count to leap over page-aligned
2294  * values so we can save vma's restart_addr in its truncate_count field.
2295  */
2296 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2297
2298 static void reset_vma_truncate_counts(struct address_space *mapping)
2299 {
2300         struct vm_area_struct *vma;
2301         struct prio_tree_iter iter;
2302
2303         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2304                 vma->vm_truncate_count = 0;
2305         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2306                 vma->vm_truncate_count = 0;
2307 }
2308
2309 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2310                 unsigned long start_addr, unsigned long end_addr,
2311                 struct zap_details *details)
2312 {
2313         unsigned long restart_addr;
2314         int need_break;
2315
2316         /*
2317          * files that support invalidating or truncating portions of the
2318          * file from under mmaped areas must have their ->fault function
2319          * return a locked page (and set VM_FAULT_LOCKED in the return).
2320          * This provides synchronisation against concurrent unmapping here.
2321          */
2322
2323 again:
2324         restart_addr = vma->vm_truncate_count;
2325         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2326                 start_addr = restart_addr;
2327                 if (start_addr >= end_addr) {
2328                         /* Top of vma has been split off since last time */
2329                         vma->vm_truncate_count = details->truncate_count;
2330                         return 0;
2331                 }
2332         }
2333
2334         restart_addr = zap_page_range(vma, start_addr,
2335                                         end_addr - start_addr, details);
2336         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2337
2338         if (restart_addr >= end_addr) {
2339                 /* We have now completed this vma: mark it so */
2340                 vma->vm_truncate_count = details->truncate_count;
2341                 if (!need_break)
2342                         return 0;
2343         } else {
2344                 /* Note restart_addr in vma's truncate_count field */
2345                 vma->vm_truncate_count = restart_addr;
2346                 if (!need_break)
2347                         goto again;
2348         }
2349
2350         spin_unlock(details->i_mmap_lock);
2351         cond_resched();
2352         spin_lock(details->i_mmap_lock);
2353         return -EINTR;
2354 }
2355
2356 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2357                                             struct zap_details *details)
2358 {
2359         struct vm_area_struct *vma;
2360         struct prio_tree_iter iter;
2361         pgoff_t vba, vea, zba, zea;
2362
2363 restart:
2364         vma_prio_tree_foreach(vma, &iter, root,
2365                         details->first_index, details->last_index) {
2366                 /* Skip quickly over those we have already dealt with */
2367                 if (vma->vm_truncate_count == details->truncate_count)
2368                         continue;
2369
2370                 vba = vma->vm_pgoff;
2371                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2372                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2373                 zba = details->first_index;
2374                 if (zba < vba)
2375                         zba = vba;
2376                 zea = details->last_index;
2377                 if (zea > vea)
2378                         zea = vea;
2379
2380                 if (unmap_mapping_range_vma(vma,
2381                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2382                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2383                                 details) < 0)
2384                         goto restart;
2385         }
2386 }
2387
2388 static inline void unmap_mapping_range_list(struct list_head *head,
2389                                             struct zap_details *details)
2390 {
2391         struct vm_area_struct *vma;
2392
2393         /*
2394          * In nonlinear VMAs there is no correspondence between virtual address
2395          * offset and file offset.  So we must perform an exhaustive search
2396          * across *all* the pages in each nonlinear VMA, not just the pages
2397          * whose virtual address lies outside the file truncation point.
2398          */
2399 restart:
2400         list_for_each_entry(vma, head, shared.vm_set.list) {
2401                 /* Skip quickly over those we have already dealt with */
2402                 if (vma->vm_truncate_count == details->truncate_count)
2403                         continue;
2404                 details->nonlinear_vma = vma;
2405                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2406                                         vma->vm_end, details) < 0)
2407                         goto restart;
2408         }
2409 }
2410
2411 /**
2412  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2413  * @mapping: the address space containing mmaps to be unmapped.
2414  * @holebegin: byte in first page to unmap, relative to the start of
2415  * the underlying file.  This will be rounded down to a PAGE_SIZE
2416  * boundary.  Note that this is different from truncate_pagecache(), which
2417  * must keep the partial page.  In contrast, we must get rid of
2418  * partial pages.
2419  * @holelen: size of prospective hole in bytes.  This will be rounded
2420  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2421  * end of the file.
2422  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2423  * but 0 when invalidating pagecache, don't throw away private data.
2424  */
2425 void unmap_mapping_range(struct address_space *mapping,
2426                 loff_t const holebegin, loff_t const holelen, int even_cows)
2427 {
2428         struct zap_details details;
2429         pgoff_t hba = holebegin >> PAGE_SHIFT;
2430         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2431
2432         /* Check for overflow. */
2433         if (sizeof(holelen) > sizeof(hlen)) {
2434                 long long holeend =
2435                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2436                 if (holeend & ~(long long)ULONG_MAX)
2437                         hlen = ULONG_MAX - hba + 1;
2438         }
2439
2440         details.check_mapping = even_cows? NULL: mapping;
2441         details.nonlinear_vma = NULL;
2442         details.first_index = hba;
2443         details.last_index = hba + hlen - 1;
2444         if (details.last_index < details.first_index)
2445                 details.last_index = ULONG_MAX;
2446         details.i_mmap_lock = &mapping->i_mmap_lock;
2447
2448         spin_lock(&mapping->i_mmap_lock);
2449
2450         /* Protect against endless unmapping loops */
2451         mapping->truncate_count++;
2452         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2453                 if (mapping->truncate_count == 0)
2454                         reset_vma_truncate_counts(mapping);
2455                 mapping->truncate_count++;
2456         }
2457         details.truncate_count = mapping->truncate_count;
2458
2459         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2460                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2461         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2462                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2463         spin_unlock(&mapping->i_mmap_lock);
2464 }
2465 EXPORT_SYMBOL(unmap_mapping_range);
2466
2467 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2468 {
2469         struct address_space *mapping = inode->i_mapping;
2470
2471         /*
2472          * If the underlying filesystem is not going to provide
2473          * a way to truncate a range of blocks (punch a hole) -
2474          * we should return failure right now.
2475          */
2476         if (!inode->i_op->truncate_range)
2477                 return -ENOSYS;
2478
2479         mutex_lock(&inode->i_mutex);
2480         down_write(&inode->i_alloc_sem);
2481         unmap_mapping_range(mapping, offset, (end - offset), 1);
2482         truncate_inode_pages_range(mapping, offset, end);
2483         unmap_mapping_range(mapping, offset, (end - offset), 1);
2484         inode->i_op->truncate_range(inode, offset, end);
2485         up_write(&inode->i_alloc_sem);
2486         mutex_unlock(&inode->i_mutex);
2487
2488         return 0;
2489 }
2490
2491 /*
2492  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2493  * but allow concurrent faults), and pte mapped but not yet locked.
2494  * We return with mmap_sem still held, but pte unmapped and unlocked.
2495  */
2496 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2497                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2498                 unsigned int flags, pte_t orig_pte)
2499 {
2500         spinlock_t *ptl;
2501         struct page *page;
2502         swp_entry_t entry;
2503         pte_t pte;
2504         struct mem_cgroup *ptr = NULL;
2505         int ret = 0;
2506
2507         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2508                 goto out;
2509
2510         entry = pte_to_swp_entry(orig_pte);
2511         if (unlikely(non_swap_entry(entry))) {
2512                 if (is_migration_entry(entry)) {
2513                         migration_entry_wait(mm, pmd, address);
2514                 } else if (is_hwpoison_entry(entry)) {
2515                         ret = VM_FAULT_HWPOISON;
2516                 } else {
2517                         print_bad_pte(vma, address, orig_pte, NULL);
2518                         ret = VM_FAULT_SIGBUS;
2519                 }
2520                 goto out;
2521         }
2522         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2523         page = lookup_swap_cache(entry);
2524         if (!page) {
2525                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2526                 page = swapin_readahead(entry,
2527                                         GFP_HIGHUSER_MOVABLE, vma, address);
2528                 if (!page) {
2529                         /*
2530                          * Back out if somebody else faulted in this pte
2531                          * while we released the pte lock.
2532                          */
2533                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2534                         if (likely(pte_same(*page_table, orig_pte)))
2535                                 ret = VM_FAULT_OOM;
2536                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2537                         goto unlock;
2538                 }
2539
2540                 /* Had to read the page from swap area: Major fault */
2541                 ret = VM_FAULT_MAJOR;
2542                 count_vm_event(PGMAJFAULT);
2543         } else if (PageHWPoison(page)) {
2544                 ret = VM_FAULT_HWPOISON;
2545                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2546                 goto out_release;
2547         }
2548
2549         lock_page(page);
2550         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2551
2552         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2553                 ret = VM_FAULT_OOM;
2554                 goto out_page;
2555         }
2556
2557         /*
2558          * Back out if somebody else already faulted in this pte.
2559          */
2560         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2561         if (unlikely(!pte_same(*page_table, orig_pte)))
2562                 goto out_nomap;
2563
2564         if (unlikely(!PageUptodate(page))) {
2565                 ret = VM_FAULT_SIGBUS;
2566                 goto out_nomap;
2567         }
2568
2569         /*
2570          * The page isn't present yet, go ahead with the fault.
2571          *
2572          * Be careful about the sequence of operations here.
2573          * To get its accounting right, reuse_swap_page() must be called
2574          * while the page is counted on swap but not yet in mapcount i.e.
2575          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2576          * must be called after the swap_free(), or it will never succeed.
2577          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2578          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2579          * in page->private. In this case, a record in swap_cgroup  is silently
2580          * discarded at swap_free().
2581          */
2582
2583         inc_mm_counter(mm, anon_rss);
2584         pte = mk_pte(page, vma->vm_page_prot);
2585         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2586                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2587                 flags &= ~FAULT_FLAG_WRITE;
2588         }
2589         flush_icache_page(vma, page);
2590         set_pte_at(mm, address, page_table, pte);
2591         page_add_anon_rmap(page, vma, address);
2592         /* It's better to call commit-charge after rmap is established */
2593         mem_cgroup_commit_charge_swapin(page, ptr);
2594
2595         swap_free(entry);
2596         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2597                 try_to_free_swap(page);
2598         unlock_page(page);
2599
2600         if (flags & FAULT_FLAG_WRITE) {
2601                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2602                 if (ret & VM_FAULT_ERROR)
2603                         ret &= VM_FAULT_ERROR;
2604                 goto out;
2605         }
2606
2607         /* No need to invalidate - it was non-present before */
2608         update_mmu_cache(vma, address, pte);
2609 unlock:
2610         pte_unmap_unlock(page_table, ptl);
2611 out:
2612         return ret;
2613 out_nomap:
2614         mem_cgroup_cancel_charge_swapin(ptr);
2615         pte_unmap_unlock(page_table, ptl);
2616 out_page:
2617         unlock_page(page);
2618 out_release:
2619         page_cache_release(page);
2620         return ret;
2621 }
2622
2623 /*
2624  * This is like a special single-page "expand_downwards()",
2625  * except we must first make sure that 'address-PAGE_SIZE'
2626  * doesn't hit another vma.
2627  *
2628  * The "find_vma()" will do the right thing even if we wrap
2629  */
2630 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2631 {
2632         address &= PAGE_MASK;
2633         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2634                 address -= PAGE_SIZE;
2635                 if (find_vma(vma->vm_mm, address) != vma)
2636                         return -ENOMEM;
2637
2638                 expand_stack(vma, address);
2639         }
2640         return 0;
2641 }
2642
2643 /*
2644  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2645  * but allow concurrent faults), and pte mapped but not yet locked.
2646  * We return with mmap_sem still held, but pte unmapped and unlocked.
2647  */
2648 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2649                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2650                 unsigned int flags)
2651 {
2652         struct page *page;
2653         spinlock_t *ptl;
2654         pte_t entry;
2655
2656         if (check_stack_guard_page(vma, address) < 0)
2657                 return VM_FAULT_SIGBUS;
2658
2659         if (!(flags & FAULT_FLAG_WRITE)) {
2660                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2661                                                 vma->vm_page_prot));
2662                 ptl = pte_lockptr(mm, pmd);
2663                 spin_lock(ptl);
2664                 if (!pte_none(*page_table))
2665                         goto unlock;
2666                 goto setpte;
2667         }
2668
2669         /* Allocate our own private page. */
2670         pte_unmap(page_table);
2671
2672         if (unlikely(anon_vma_prepare(vma)))
2673                 goto oom;
2674         page = alloc_zeroed_user_highpage_movable(vma, address);
2675         if (!page)
2676                 goto oom;
2677         __SetPageUptodate(page);
2678
2679         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2680                 goto oom_free_page;
2681
2682         entry = mk_pte(page, vma->vm_page_prot);
2683         if (vma->vm_flags & VM_WRITE)
2684                 entry = pte_mkwrite(pte_mkdirty(entry));
2685
2686         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2687         if (!pte_none(*page_table))
2688                 goto release;
2689
2690         inc_mm_counter(mm, anon_rss);
2691         page_add_new_anon_rmap(page, vma, address);
2692 setpte:
2693         set_pte_at(mm, address, page_table, entry);
2694
2695         /* No need to invalidate - it was non-present before */
2696         update_mmu_cache(vma, address, entry);
2697 unlock:
2698         pte_unmap_unlock(page_table, ptl);
2699         return 0;
2700 release:
2701         mem_cgroup_uncharge_page(page);
2702         page_cache_release(page);
2703         goto unlock;
2704 oom_free_page:
2705         page_cache_release(page);
2706 oom:
2707         return VM_FAULT_OOM;
2708 }
2709
2710 /*
2711  * __do_fault() tries to create a new page mapping. It aggressively
2712  * tries to share with existing pages, but makes a separate copy if
2713  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2714  * the next page fault.
2715  *
2716  * As this is called only for pages that do not currently exist, we
2717  * do not need to flush old virtual caches or the TLB.
2718  *
2719  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2720  * but allow concurrent faults), and pte neither mapped nor locked.
2721  * We return with mmap_sem still held, but pte unmapped and unlocked.
2722  */
2723 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2724                 unsigned long address, pmd_t *pmd,
2725                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2726 {