f82b359b27452dd1f7e682197479fa1391d19a46
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/delayacct.h>
51 #include <linux/init.h>
52 #include <linux/writeback.h>
53
54 #include <asm/pgalloc.h>
55 #include <asm/uaccess.h>
56 #include <asm/tlb.h>
57 #include <asm/tlbflush.h>
58 #include <asm/pgtable.h>
59
60 #include <linux/swapops.h>
61 #include <linux/elf.h>
62
63 #ifndef CONFIG_NEED_MULTIPLE_NODES
64 /* use the per-pgdat data instead for discontigmem - mbligh */
65 unsigned long max_mapnr;
66 struct page *mem_map;
67
68 EXPORT_SYMBOL(max_mapnr);
69 EXPORT_SYMBOL(mem_map);
70 #endif
71
72 unsigned long num_physpages;
73 /*
74  * A number of key systems in x86 including ioremap() rely on the assumption
75  * that high_memory defines the upper bound on direct map memory, then end
76  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
77  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
78  * and ZONE_HIGHMEM.
79  */
80 void * high_memory;
81
82 EXPORT_SYMBOL(num_physpages);
83 EXPORT_SYMBOL(high_memory);
84
85 int randomize_va_space __read_mostly = 1;
86
87 static int __init disable_randmaps(char *s)
88 {
89         randomize_va_space = 0;
90         return 1;
91 }
92 __setup("norandmaps", disable_randmaps);
93
94
95 /*
96  * If a p?d_bad entry is found while walking page tables, report
97  * the error, before resetting entry to p?d_none.  Usually (but
98  * very seldom) called out from the p?d_none_or_clear_bad macros.
99  */
100
101 void pgd_clear_bad(pgd_t *pgd)
102 {
103         pgd_ERROR(*pgd);
104         pgd_clear(pgd);
105 }
106
107 void pud_clear_bad(pud_t *pud)
108 {
109         pud_ERROR(*pud);
110         pud_clear(pud);
111 }
112
113 void pmd_clear_bad(pmd_t *pmd)
114 {
115         pmd_ERROR(*pmd);
116         pmd_clear(pmd);
117 }
118
119 /*
120  * Note: this doesn't free the actual pages themselves. That
121  * has been handled earlier when unmapping all the memory regions.
122  */
123 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
124 {
125         struct page *page = pmd_page(*pmd);
126         pmd_clear(pmd);
127         pte_lock_deinit(page);
128         pte_free_tlb(tlb, page);
129         dec_zone_page_state(page, NR_PAGETABLE);
130         tlb->mm->nr_ptes--;
131 }
132
133 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
134                                 unsigned long addr, unsigned long end,
135                                 unsigned long floor, unsigned long ceiling)
136 {
137         pmd_t *pmd;
138         unsigned long next;
139         unsigned long start;
140
141         start = addr;
142         pmd = pmd_offset(pud, addr);
143         do {
144                 next = pmd_addr_end(addr, end);
145                 if (pmd_none_or_clear_bad(pmd))
146                         continue;
147                 free_pte_range(tlb, pmd);
148         } while (pmd++, addr = next, addr != end);
149
150         start &= PUD_MASK;
151         if (start < floor)
152                 return;
153         if (ceiling) {
154                 ceiling &= PUD_MASK;
155                 if (!ceiling)
156                         return;
157         }
158         if (end - 1 > ceiling - 1)
159                 return;
160
161         pmd = pmd_offset(pud, start);
162         pud_clear(pud);
163         pmd_free_tlb(tlb, pmd);
164 }
165
166 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
167                                 unsigned long addr, unsigned long end,
168                                 unsigned long floor, unsigned long ceiling)
169 {
170         pud_t *pud;
171         unsigned long next;
172         unsigned long start;
173
174         start = addr;
175         pud = pud_offset(pgd, addr);
176         do {
177                 next = pud_addr_end(addr, end);
178                 if (pud_none_or_clear_bad(pud))
179                         continue;
180                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
181         } while (pud++, addr = next, addr != end);
182
183         start &= PGDIR_MASK;
184         if (start < floor)
185                 return;
186         if (ceiling) {
187                 ceiling &= PGDIR_MASK;
188                 if (!ceiling)
189                         return;
190         }
191         if (end - 1 > ceiling - 1)
192                 return;
193
194         pud = pud_offset(pgd, start);
195         pgd_clear(pgd);
196         pud_free_tlb(tlb, pud);
197 }
198
199 /*
200  * This function frees user-level page tables of a process.
201  *
202  * Must be called with pagetable lock held.
203  */
204 void free_pgd_range(struct mmu_gather **tlb,
205                         unsigned long addr, unsigned long end,
206                         unsigned long floor, unsigned long ceiling)
207 {
208         pgd_t *pgd;
209         unsigned long next;
210         unsigned long start;
211
212         /*
213          * The next few lines have given us lots of grief...
214          *
215          * Why are we testing PMD* at this top level?  Because often
216          * there will be no work to do at all, and we'd prefer not to
217          * go all the way down to the bottom just to discover that.
218          *
219          * Why all these "- 1"s?  Because 0 represents both the bottom
220          * of the address space and the top of it (using -1 for the
221          * top wouldn't help much: the masks would do the wrong thing).
222          * The rule is that addr 0 and floor 0 refer to the bottom of
223          * the address space, but end 0 and ceiling 0 refer to the top
224          * Comparisons need to use "end - 1" and "ceiling - 1" (though
225          * that end 0 case should be mythical).
226          *
227          * Wherever addr is brought up or ceiling brought down, we must
228          * be careful to reject "the opposite 0" before it confuses the
229          * subsequent tests.  But what about where end is brought down
230          * by PMD_SIZE below? no, end can't go down to 0 there.
231          *
232          * Whereas we round start (addr) and ceiling down, by different
233          * masks at different levels, in order to test whether a table
234          * now has no other vmas using it, so can be freed, we don't
235          * bother to round floor or end up - the tests don't need that.
236          */
237
238         addr &= PMD_MASK;
239         if (addr < floor) {
240                 addr += PMD_SIZE;
241                 if (!addr)
242                         return;
243         }
244         if (ceiling) {
245                 ceiling &= PMD_MASK;
246                 if (!ceiling)
247                         return;
248         }
249         if (end - 1 > ceiling - 1)
250                 end -= PMD_SIZE;
251         if (addr > end - 1)
252                 return;
253
254         start = addr;
255         pgd = pgd_offset((*tlb)->mm, addr);
256         do {
257                 next = pgd_addr_end(addr, end);
258                 if (pgd_none_or_clear_bad(pgd))
259                         continue;
260                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
261         } while (pgd++, addr = next, addr != end);
262
263         if (!(*tlb)->fullmm)
264                 flush_tlb_pgtables((*tlb)->mm, start, end);
265 }
266
267 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
268                 unsigned long floor, unsigned long ceiling)
269 {
270         while (vma) {
271                 struct vm_area_struct *next = vma->vm_next;
272                 unsigned long addr = vma->vm_start;
273
274                 /*
275                  * Hide vma from rmap and vmtruncate before freeing pgtables
276                  */
277                 anon_vma_unlink(vma);
278                 unlink_file_vma(vma);
279
280                 if (is_vm_hugetlb_page(vma)) {
281                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
282                                 floor, next? next->vm_start: ceiling);
283                 } else {
284                         /*
285                          * Optimization: gather nearby vmas into one call down
286                          */
287                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
288                                && !is_vm_hugetlb_page(next)) {
289                                 vma = next;
290                                 next = vma->vm_next;
291                                 anon_vma_unlink(vma);
292                                 unlink_file_vma(vma);
293                         }
294                         free_pgd_range(tlb, addr, vma->vm_end,
295                                 floor, next? next->vm_start: ceiling);
296                 }
297                 vma = next;
298         }
299 }
300
301 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
302 {
303         struct page *new = pte_alloc_one(mm, address);
304         if (!new)
305                 return -ENOMEM;
306
307         pte_lock_init(new);
308         spin_lock(&mm->page_table_lock);
309         if (pmd_present(*pmd)) {        /* Another has populated it */
310                 pte_lock_deinit(new);
311                 pte_free(new);
312         } else {
313                 mm->nr_ptes++;
314                 inc_zone_page_state(new, NR_PAGETABLE);
315                 pmd_populate(mm, pmd, new);
316         }
317         spin_unlock(&mm->page_table_lock);
318         return 0;
319 }
320
321 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
322 {
323         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
324         if (!new)
325                 return -ENOMEM;
326
327         spin_lock(&init_mm.page_table_lock);
328         if (pmd_present(*pmd))          /* Another has populated it */
329                 pte_free_kernel(new);
330         else
331                 pmd_populate_kernel(&init_mm, pmd, new);
332         spin_unlock(&init_mm.page_table_lock);
333         return 0;
334 }
335
336 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
337 {
338         if (file_rss)
339                 add_mm_counter(mm, file_rss, file_rss);
340         if (anon_rss)
341                 add_mm_counter(mm, anon_rss, anon_rss);
342 }
343
344 /*
345  * This function is called to print an error when a bad pte
346  * is found. For example, we might have a PFN-mapped pte in
347  * a region that doesn't allow it.
348  *
349  * The calling function must still handle the error.
350  */
351 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
352 {
353         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
354                         "vm_flags = %lx, vaddr = %lx\n",
355                 (long long)pte_val(pte),
356                 (vma->vm_mm == current->mm ? current->comm : "???"),
357                 vma->vm_flags, vaddr);
358         dump_stack();
359 }
360
361 static inline int is_cow_mapping(unsigned int flags)
362 {
363         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
364 }
365
366 /*
367  * This function gets the "struct page" associated with a pte.
368  *
369  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
370  * will have each page table entry just pointing to a raw page frame
371  * number, and as far as the VM layer is concerned, those do not have
372  * pages associated with them - even if the PFN might point to memory
373  * that otherwise is perfectly fine and has a "struct page".
374  *
375  * The way we recognize those mappings is through the rules set up
376  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
377  * and the vm_pgoff will point to the first PFN mapped: thus every
378  * page that is a raw mapping will always honor the rule
379  *
380  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
381  *
382  * and if that isn't true, the page has been COW'ed (in which case it
383  * _does_ have a "struct page" associated with it even if it is in a
384  * VM_PFNMAP range).
385  */
386 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
387 {
388         unsigned long pfn = pte_pfn(pte);
389
390         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
391                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
392                 if (pfn == vma->vm_pgoff + off)
393                         return NULL;
394                 if (!is_cow_mapping(vma->vm_flags))
395                         return NULL;
396         }
397
398         /*
399          * Add some anal sanity checks for now. Eventually,
400          * we should just do "return pfn_to_page(pfn)", but
401          * in the meantime we check that we get a valid pfn,
402          * and that the resulting page looks ok.
403          */
404         if (unlikely(!pfn_valid(pfn))) {
405                 print_bad_pte(vma, pte, addr);
406                 return NULL;
407         }
408
409         /*
410          * NOTE! We still have PageReserved() pages in the page 
411          * tables. 
412          *
413          * The PAGE_ZERO() pages and various VDSO mappings can
414          * cause them to exist.
415          */
416         return pfn_to_page(pfn);
417 }
418
419 /*
420  * copy one vm_area from one task to the other. Assumes the page tables
421  * already present in the new task to be cleared in the whole range
422  * covered by this vma.
423  */
424
425 static inline void
426 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
427                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
428                 unsigned long addr, int *rss)
429 {
430         unsigned long vm_flags = vma->vm_flags;
431         pte_t pte = *src_pte;
432         struct page *page;
433
434         /* pte contains position in swap or file, so copy. */
435         if (unlikely(!pte_present(pte))) {
436                 if (!pte_file(pte)) {
437                         swp_entry_t entry = pte_to_swp_entry(pte);
438
439                         swap_duplicate(entry);
440                         /* make sure dst_mm is on swapoff's mmlist. */
441                         if (unlikely(list_empty(&dst_mm->mmlist))) {
442                                 spin_lock(&mmlist_lock);
443                                 if (list_empty(&dst_mm->mmlist))
444                                         list_add(&dst_mm->mmlist,
445                                                  &src_mm->mmlist);
446                                 spin_unlock(&mmlist_lock);
447                         }
448                         if (is_write_migration_entry(entry) &&
449                                         is_cow_mapping(vm_flags)) {
450                                 /*
451                                  * COW mappings require pages in both parent
452                                  * and child to be set to read.
453                                  */
454                                 make_migration_entry_read(&entry);
455                                 pte = swp_entry_to_pte(entry);
456                                 set_pte_at(src_mm, addr, src_pte, pte);
457                         }
458                 }
459                 goto out_set_pte;
460         }
461
462         /*
463          * If it's a COW mapping, write protect it both
464          * in the parent and the child
465          */
466         if (is_cow_mapping(vm_flags)) {
467                 ptep_set_wrprotect(src_mm, addr, src_pte);
468                 pte = pte_wrprotect(pte);
469         }
470
471         /*
472          * If it's a shared mapping, mark it clean in
473          * the child
474          */
475         if (vm_flags & VM_SHARED)
476                 pte = pte_mkclean(pte);
477         pte = pte_mkold(pte);
478
479         page = vm_normal_page(vma, addr, pte);
480         if (page) {
481                 get_page(page);
482                 page_dup_rmap(page, vma, addr);
483                 rss[!!PageAnon(page)]++;
484         }
485
486 out_set_pte:
487         set_pte_at(dst_mm, addr, dst_pte, pte);
488 }
489
490 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
491                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
492                 unsigned long addr, unsigned long end)
493 {
494         pte_t *src_pte, *dst_pte;
495         spinlock_t *src_ptl, *dst_ptl;
496         int progress = 0;
497         int rss[2];
498
499 again:
500         rss[1] = rss[0] = 0;
501         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
502         if (!dst_pte)
503                 return -ENOMEM;
504         src_pte = pte_offset_map_nested(src_pmd, addr);
505         src_ptl = pte_lockptr(src_mm, src_pmd);
506         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
507         arch_enter_lazy_mmu_mode();
508
509         do {
510                 /*
511                  * We are holding two locks at this point - either of them
512                  * could generate latencies in another task on another CPU.
513                  */
514                 if (progress >= 32) {
515                         progress = 0;
516                         if (need_resched() ||
517                             need_lockbreak(src_ptl) ||
518                             need_lockbreak(dst_ptl))
519                                 break;
520                 }
521                 if (pte_none(*src_pte)) {
522                         progress++;
523                         continue;
524                 }
525                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
526                 progress += 8;
527         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
528
529         arch_leave_lazy_mmu_mode();
530         spin_unlock(src_ptl);
531         pte_unmap_nested(src_pte - 1);
532         add_mm_rss(dst_mm, rss[0], rss[1]);
533         pte_unmap_unlock(dst_pte - 1, dst_ptl);
534         cond_resched();
535         if (addr != end)
536                 goto again;
537         return 0;
538 }
539
540 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
541                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
542                 unsigned long addr, unsigned long end)
543 {
544         pmd_t *src_pmd, *dst_pmd;
545         unsigned long next;
546
547         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
548         if (!dst_pmd)
549                 return -ENOMEM;
550         src_pmd = pmd_offset(src_pud, addr);
551         do {
552                 next = pmd_addr_end(addr, end);
553                 if (pmd_none_or_clear_bad(src_pmd))
554                         continue;
555                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
556                                                 vma, addr, next))
557                         return -ENOMEM;
558         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
559         return 0;
560 }
561
562 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
563                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
564                 unsigned long addr, unsigned long end)
565 {
566         pud_t *src_pud, *dst_pud;
567         unsigned long next;
568
569         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
570         if (!dst_pud)
571                 return -ENOMEM;
572         src_pud = pud_offset(src_pgd, addr);
573         do {
574                 next = pud_addr_end(addr, end);
575                 if (pud_none_or_clear_bad(src_pud))
576                         continue;
577                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
578                                                 vma, addr, next))
579                         return -ENOMEM;
580         } while (dst_pud++, src_pud++, addr = next, addr != end);
581         return 0;
582 }
583
584 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
585                 struct vm_area_struct *vma)
586 {
587         pgd_t *src_pgd, *dst_pgd;
588         unsigned long next;
589         unsigned long addr = vma->vm_start;
590         unsigned long end = vma->vm_end;
591
592         /*
593          * Don't copy ptes where a page fault will fill them correctly.
594          * Fork becomes much lighter when there are big shared or private
595          * readonly mappings. The tradeoff is that copy_page_range is more
596          * efficient than faulting.
597          */
598         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
599                 if (!vma->anon_vma)
600                         return 0;
601         }
602
603         if (is_vm_hugetlb_page(vma))
604                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
605
606         dst_pgd = pgd_offset(dst_mm, addr);
607         src_pgd = pgd_offset(src_mm, addr);
608         do {
609                 next = pgd_addr_end(addr, end);
610                 if (pgd_none_or_clear_bad(src_pgd))
611                         continue;
612                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
613                                                 vma, addr, next))
614                         return -ENOMEM;
615         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
616         return 0;
617 }
618
619 static unsigned long zap_pte_range(struct mmu_gather *tlb,
620                                 struct vm_area_struct *vma, pmd_t *pmd,
621                                 unsigned long addr, unsigned long end,
622                                 long *zap_work, struct zap_details *details)
623 {
624         struct mm_struct *mm = tlb->mm;
625         pte_t *pte;
626         spinlock_t *ptl;
627         int file_rss = 0;
628         int anon_rss = 0;
629
630         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
631         arch_enter_lazy_mmu_mode();
632         do {
633                 pte_t ptent = *pte;
634                 if (pte_none(ptent)) {
635                         (*zap_work)--;
636                         continue;
637                 }
638
639                 (*zap_work) -= PAGE_SIZE;
640
641                 if (pte_present(ptent)) {
642                         struct page *page;
643
644                         page = vm_normal_page(vma, addr, ptent);
645                         if (unlikely(details) && page) {
646                                 /*
647                                  * unmap_shared_mapping_pages() wants to
648                                  * invalidate cache without truncating:
649                                  * unmap shared but keep private pages.
650                                  */
651                                 if (details->check_mapping &&
652                                     details->check_mapping != page->mapping)
653                                         continue;
654                                 /*
655                                  * Each page->index must be checked when
656                                  * invalidating or truncating nonlinear.
657                                  */
658                                 if (details->nonlinear_vma &&
659                                     (page->index < details->first_index ||
660                                      page->index > details->last_index))
661                                         continue;
662                         }
663                         ptent = ptep_get_and_clear_full(mm, addr, pte,
664                                                         tlb->fullmm);
665                         tlb_remove_tlb_entry(tlb, pte, addr);
666                         if (unlikely(!page))
667                                 continue;
668                         if (unlikely(details) && details->nonlinear_vma
669                             && linear_page_index(details->nonlinear_vma,
670                                                 addr) != page->index)
671                                 set_pte_at(mm, addr, pte,
672                                            pgoff_to_pte(page->index));
673                         if (PageAnon(page))
674                                 anon_rss--;
675                         else {
676                                 if (pte_dirty(ptent))
677                                         set_page_dirty(page);
678                                 if (pte_young(ptent))
679                                         SetPageReferenced(page);
680                                 file_rss--;
681                         }
682                         page_remove_rmap(page, vma);
683                         tlb_remove_page(tlb, page);
684                         continue;
685                 }
686                 /*
687                  * If details->check_mapping, we leave swap entries;
688                  * if details->nonlinear_vma, we leave file entries.
689                  */
690                 if (unlikely(details))
691                         continue;
692                 if (!pte_file(ptent))
693                         free_swap_and_cache(pte_to_swp_entry(ptent));
694                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
695         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
696
697         add_mm_rss(mm, file_rss, anon_rss);
698         arch_leave_lazy_mmu_mode();
699         pte_unmap_unlock(pte - 1, ptl);
700
701         return addr;
702 }
703
704 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
705                                 struct vm_area_struct *vma, pud_t *pud,
706                                 unsigned long addr, unsigned long end,
707                                 long *zap_work, struct zap_details *details)
708 {
709         pmd_t *pmd;
710         unsigned long next;
711
712         pmd = pmd_offset(pud, addr);
713         do {
714                 next = pmd_addr_end(addr, end);
715                 if (pmd_none_or_clear_bad(pmd)) {
716                         (*zap_work)--;
717                         continue;
718                 }
719                 next = zap_pte_range(tlb, vma, pmd, addr, next,
720                                                 zap_work, details);
721         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
722
723         return addr;
724 }
725
726 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
727                                 struct vm_area_struct *vma, pgd_t *pgd,
728                                 unsigned long addr, unsigned long end,
729                                 long *zap_work, struct zap_details *details)
730 {
731         pud_t *pud;
732         unsigned long next;
733
734         pud = pud_offset(pgd, addr);
735         do {
736                 next = pud_addr_end(addr, end);
737                 if (pud_none_or_clear_bad(pud)) {
738                         (*zap_work)--;
739                         continue;
740                 }
741                 next = zap_pmd_range(tlb, vma, pud, addr, next,
742                                                 zap_work, details);
743         } while (pud++, addr = next, (addr != end && *zap_work > 0));
744
745         return addr;
746 }
747
748 static unsigned long unmap_page_range(struct mmu_gather *tlb,
749                                 struct vm_area_struct *vma,
750                                 unsigned long addr, unsigned long end,
751                                 long *zap_work, struct zap_details *details)
752 {
753         pgd_t *pgd;
754         unsigned long next;
755
756         if (details && !details->check_mapping && !details->nonlinear_vma)
757                 details = NULL;
758
759         BUG_ON(addr >= end);
760         tlb_start_vma(tlb, vma);
761         pgd = pgd_offset(vma->vm_mm, addr);
762         do {
763                 next = pgd_addr_end(addr, end);
764                 if (pgd_none_or_clear_bad(pgd)) {
765                         (*zap_work)--;
766                         continue;
767                 }
768                 next = zap_pud_range(tlb, vma, pgd, addr, next,
769                                                 zap_work, details);
770         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
771         tlb_end_vma(tlb, vma);
772
773         return addr;
774 }
775
776 #ifdef CONFIG_PREEMPT
777 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
778 #else
779 /* No preempt: go for improved straight-line efficiency */
780 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
781 #endif
782
783 /**
784  * unmap_vmas - unmap a range of memory covered by a list of vma's
785  * @tlbp: address of the caller's struct mmu_gather
786  * @vma: the starting vma
787  * @start_addr: virtual address at which to start unmapping
788  * @end_addr: virtual address at which to end unmapping
789  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
790  * @details: details of nonlinear truncation or shared cache invalidation
791  *
792  * Returns the end address of the unmapping (restart addr if interrupted).
793  *
794  * Unmap all pages in the vma list.
795  *
796  * We aim to not hold locks for too long (for scheduling latency reasons).
797  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
798  * return the ending mmu_gather to the caller.
799  *
800  * Only addresses between `start' and `end' will be unmapped.
801  *
802  * The VMA list must be sorted in ascending virtual address order.
803  *
804  * unmap_vmas() assumes that the caller will flush the whole unmapped address
805  * range after unmap_vmas() returns.  So the only responsibility here is to
806  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
807  * drops the lock and schedules.
808  */
809 unsigned long unmap_vmas(struct mmu_gather **tlbp,
810                 struct vm_area_struct *vma, unsigned long start_addr,
811                 unsigned long end_addr, unsigned long *nr_accounted,
812                 struct zap_details *details)
813 {
814         long zap_work = ZAP_BLOCK_SIZE;
815         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
816         int tlb_start_valid = 0;
817         unsigned long start = start_addr;
818         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
819         int fullmm = (*tlbp)->fullmm;
820
821         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
822                 unsigned long end;
823
824                 start = max(vma->vm_start, start_addr);
825                 if (start >= vma->vm_end)
826                         continue;
827                 end = min(vma->vm_end, end_addr);
828                 if (end <= vma->vm_start)
829                         continue;
830
831                 if (vma->vm_flags & VM_ACCOUNT)
832                         *nr_accounted += (end - start) >> PAGE_SHIFT;
833
834                 while (start != end) {
835                         if (!tlb_start_valid) {
836                                 tlb_start = start;
837                                 tlb_start_valid = 1;
838                         }
839
840                         if (unlikely(is_vm_hugetlb_page(vma))) {
841                                 unmap_hugepage_range(vma, start, end);
842                                 zap_work -= (end - start) /
843                                                 (HPAGE_SIZE / PAGE_SIZE);
844                                 start = end;
845                         } else
846                                 start = unmap_page_range(*tlbp, vma,
847                                                 start, end, &zap_work, details);
848
849                         if (zap_work > 0) {
850                                 BUG_ON(start != end);
851                                 break;
852                         }
853
854                         tlb_finish_mmu(*tlbp, tlb_start, start);
855
856                         if (need_resched() ||
857                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
858                                 if (i_mmap_lock) {
859                                         *tlbp = NULL;
860                                         goto out;
861                                 }
862                                 cond_resched();
863                         }
864
865                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
866                         tlb_start_valid = 0;
867                         zap_work = ZAP_BLOCK_SIZE;
868                 }
869         }
870 out:
871         return start;   /* which is now the end (or restart) address */
872 }
873
874 /**
875  * zap_page_range - remove user pages in a given range
876  * @vma: vm_area_struct holding the applicable pages
877  * @address: starting address of pages to zap
878  * @size: number of bytes to zap
879  * @details: details of nonlinear truncation or shared cache invalidation
880  */
881 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
882                 unsigned long size, struct zap_details *details)
883 {
884         struct mm_struct *mm = vma->vm_mm;
885         struct mmu_gather *tlb;
886         unsigned long end = address + size;
887         unsigned long nr_accounted = 0;
888
889         lru_add_drain();
890         tlb = tlb_gather_mmu(mm, 0);
891         update_hiwater_rss(mm);
892         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
893         if (tlb)
894                 tlb_finish_mmu(tlb, address, end);
895         return end;
896 }
897
898 /*
899  * Do a quick page-table lookup for a single page.
900  */
901 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
902                         unsigned int flags)
903 {
904         pgd_t *pgd;
905         pud_t *pud;
906         pmd_t *pmd;
907         pte_t *ptep, pte;
908         spinlock_t *ptl;
909         struct page *page;
910         struct mm_struct *mm = vma->vm_mm;
911
912         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
913         if (!IS_ERR(page)) {
914                 BUG_ON(flags & FOLL_GET);
915                 goto out;
916         }
917
918         page = NULL;
919         pgd = pgd_offset(mm, address);
920         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
921                 goto no_page_table;
922
923         pud = pud_offset(pgd, address);
924         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
925                 goto no_page_table;
926         
927         pmd = pmd_offset(pud, address);
928         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
929                 goto no_page_table;
930
931         if (pmd_huge(*pmd)) {
932                 BUG_ON(flags & FOLL_GET);
933                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
934                 goto out;
935         }
936
937         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
938         if (!ptep)
939                 goto out;
940
941         pte = *ptep;
942         if (!pte_present(pte))
943                 goto unlock;
944         if ((flags & FOLL_WRITE) && !pte_write(pte))
945                 goto unlock;
946         page = vm_normal_page(vma, address, pte);
947         if (unlikely(!page))
948                 goto unlock;
949
950         if (flags & FOLL_GET)
951                 get_page(page);
952         if (flags & FOLL_TOUCH) {
953                 if ((flags & FOLL_WRITE) &&
954                     !pte_dirty(pte) && !PageDirty(page))
955                         set_page_dirty(page);
956                 mark_page_accessed(page);
957         }
958 unlock:
959         pte_unmap_unlock(ptep, ptl);
960 out:
961         return page;
962
963 no_page_table:
964         /*
965          * When core dumping an enormous anonymous area that nobody
966          * has touched so far, we don't want to allocate page tables.
967          */
968         if (flags & FOLL_ANON) {
969                 page = ZERO_PAGE(address);
970                 if (flags & FOLL_GET)
971                         get_page(page);
972                 BUG_ON(flags & FOLL_WRITE);
973         }
974         return page;
975 }
976
977 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
978                 unsigned long start, int len, int write, int force,
979                 struct page **pages, struct vm_area_struct **vmas)
980 {
981         int i;
982         unsigned int vm_flags;
983
984         /* 
985          * Require read or write permissions.
986          * If 'force' is set, we only require the "MAY" flags.
987          */
988         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
989         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
990         i = 0;
991
992         do {
993                 struct vm_area_struct *vma;
994                 unsigned int foll_flags;
995
996                 vma = find_extend_vma(mm, start);
997                 if (!vma && in_gate_area(tsk, start)) {
998                         unsigned long pg = start & PAGE_MASK;
999                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1000                         pgd_t *pgd;
1001                         pud_t *pud;
1002                         pmd_t *pmd;
1003                         pte_t *pte;
1004                         if (write) /* user gate pages are read-only */
1005                                 return i ? : -EFAULT;
1006                         if (pg > TASK_SIZE)
1007                                 pgd = pgd_offset_k(pg);
1008                         else
1009                                 pgd = pgd_offset_gate(mm, pg);
1010                         BUG_ON(pgd_none(*pgd));
1011                         pud = pud_offset(pgd, pg);
1012                         BUG_ON(pud_none(*pud));
1013                         pmd = pmd_offset(pud, pg);
1014                         if (pmd_none(*pmd))
1015                                 return i ? : -EFAULT;
1016                         pte = pte_offset_map(pmd, pg);
1017                         if (pte_none(*pte)) {
1018                                 pte_unmap(pte);
1019                                 return i ? : -EFAULT;
1020                         }
1021                         if (pages) {
1022                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1023                                 pages[i] = page;
1024                                 if (page)
1025                                         get_page(page);
1026                         }
1027                         pte_unmap(pte);
1028                         if (vmas)
1029                                 vmas[i] = gate_vma;
1030                         i++;
1031                         start += PAGE_SIZE;
1032                         len--;
1033                         continue;
1034                 }
1035
1036                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1037                                 || !(vm_flags & vma->vm_flags))
1038                         return i ? : -EFAULT;
1039
1040                 if (is_vm_hugetlb_page(vma)) {
1041                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1042                                                 &start, &len, i);
1043                         continue;
1044                 }
1045
1046                 foll_flags = FOLL_TOUCH;
1047                 if (pages)
1048                         foll_flags |= FOLL_GET;
1049                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1050                     (!vma->vm_ops || (!vma->vm_ops->nopage &&
1051                                         !vma->vm_ops->fault)))
1052                         foll_flags |= FOLL_ANON;
1053
1054                 do {
1055                         struct page *page;
1056
1057                         /*
1058                          * If tsk is ooming, cut off its access to large memory
1059                          * allocations. It has a pending SIGKILL, but it can't
1060                          * be processed until returning to user space.
1061                          */
1062                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1063                                 return -ENOMEM;
1064
1065                         if (write)
1066                                 foll_flags |= FOLL_WRITE;
1067
1068                         cond_resched();
1069                         while (!(page = follow_page(vma, start, foll_flags))) {
1070                                 int ret;
1071                                 ret = handle_mm_fault(mm, vma, start,
1072                                                 foll_flags & FOLL_WRITE);
1073                                 if (ret & VM_FAULT_ERROR) {
1074                                         if (ret & VM_FAULT_OOM)
1075                                                 return i ? i : -ENOMEM;
1076                                         else if (ret & VM_FAULT_SIGBUS)
1077                                                 return i ? i : -EFAULT;
1078                                         BUG();
1079                                 }
1080                                 if (ret & VM_FAULT_MAJOR)
1081                                         tsk->maj_flt++;
1082                                 else
1083                                         tsk->min_flt++;
1084
1085                                 /*
1086                                  * The VM_FAULT_WRITE bit tells us that
1087                                  * do_wp_page has broken COW when necessary,
1088                                  * even if maybe_mkwrite decided not to set
1089                                  * pte_write. We can thus safely do subsequent
1090                                  * page lookups as if they were reads.
1091                                  */
1092                                 if (ret & VM_FAULT_WRITE)
1093                                         foll_flags &= ~FOLL_WRITE;
1094
1095                                 cond_resched();
1096                         }
1097                         if (pages) {
1098                                 pages[i] = page;
1099
1100                                 flush_anon_page(vma, page, start);
1101                                 flush_dcache_page(page);
1102                         }
1103                         if (vmas)
1104                                 vmas[i] = vma;
1105                         i++;
1106                         start += PAGE_SIZE;
1107                         len--;
1108                 } while (len && start < vma->vm_end);
1109         } while (len);
1110         return i;
1111 }
1112 EXPORT_SYMBOL(get_user_pages);
1113
1114 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1115                         unsigned long addr, unsigned long end, pgprot_t prot)
1116 {
1117         pte_t *pte;
1118         spinlock_t *ptl;
1119         int err = 0;
1120
1121         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1122         if (!pte)
1123                 return -EAGAIN;
1124         arch_enter_lazy_mmu_mode();
1125         do {
1126                 struct page *page = ZERO_PAGE(addr);
1127                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1128
1129                 if (unlikely(!pte_none(*pte))) {
1130                         err = -EEXIST;
1131                         pte++;
1132                         break;
1133                 }
1134                 page_cache_get(page);
1135                 page_add_file_rmap(page);
1136                 inc_mm_counter(mm, file_rss);
1137                 set_pte_at(mm, addr, pte, zero_pte);
1138         } while (pte++, addr += PAGE_SIZE, addr != end);
1139         arch_leave_lazy_mmu_mode();
1140         pte_unmap_unlock(pte - 1, ptl);
1141         return err;
1142 }
1143
1144 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1145                         unsigned long addr, unsigned long end, pgprot_t prot)
1146 {
1147         pmd_t *pmd;
1148         unsigned long next;
1149         int err;
1150
1151         pmd = pmd_alloc(mm, pud, addr);
1152         if (!pmd)
1153                 return -EAGAIN;
1154         do {
1155                 next = pmd_addr_end(addr, end);
1156                 err = zeromap_pte_range(mm, pmd, addr, next, prot);
1157                 if (err)
1158                         break;
1159         } while (pmd++, addr = next, addr != end);
1160         return err;
1161 }
1162
1163 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1164                         unsigned long addr, unsigned long end, pgprot_t prot)
1165 {
1166         pud_t *pud;
1167         unsigned long next;
1168         int err;
1169
1170         pud = pud_alloc(mm, pgd, addr);
1171         if (!pud)
1172                 return -EAGAIN;
1173         do {
1174                 next = pud_addr_end(addr, end);
1175                 err = zeromap_pmd_range(mm, pud, addr, next, prot);
1176                 if (err)
1177                         break;
1178         } while (pud++, addr = next, addr != end);
1179         return err;
1180 }
1181
1182 int zeromap_page_range(struct vm_area_struct *vma,
1183                         unsigned long addr, unsigned long size, pgprot_t prot)
1184 {
1185         pgd_t *pgd;
1186         unsigned long next;
1187         unsigned long end = addr + size;
1188         struct mm_struct *mm = vma->vm_mm;
1189         int err;
1190
1191         BUG_ON(addr >= end);
1192         pgd = pgd_offset(mm, addr);
1193         flush_cache_range(vma, addr, end);
1194         do {
1195                 next = pgd_addr_end(addr, end);
1196                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1197                 if (err)
1198                         break;
1199         } while (pgd++, addr = next, addr != end);
1200         return err;
1201 }
1202
1203 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1204 {
1205         pgd_t * pgd = pgd_offset(mm, addr);
1206         pud_t * pud = pud_alloc(mm, pgd, addr);
1207         if (pud) {
1208                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1209                 if (pmd)
1210                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1211         }
1212         return NULL;
1213 }
1214
1215 /*
1216  * This is the old fallback for page remapping.
1217  *
1218  * For historical reasons, it only allows reserved pages. Only
1219  * old drivers should use this, and they needed to mark their
1220  * pages reserved for the old functions anyway.
1221  */
1222 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1223 {
1224         int retval;
1225         pte_t *pte;
1226         spinlock_t *ptl;  
1227
1228         retval = -EINVAL;
1229         if (PageAnon(page))
1230                 goto out;
1231         retval = -ENOMEM;
1232         flush_dcache_page(page);
1233         pte = get_locked_pte(mm, addr, &ptl);
1234         if (!pte)
1235                 goto out;
1236         retval = -EBUSY;
1237         if (!pte_none(*pte))
1238                 goto out_unlock;
1239
1240         /* Ok, finally just insert the thing.. */
1241         get_page(page);
1242         inc_mm_counter(mm, file_rss);
1243         page_add_file_rmap(page);
1244         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1245
1246         retval = 0;
1247 out_unlock:
1248         pte_unmap_unlock(pte, ptl);
1249 out:
1250         return retval;
1251 }
1252
1253 /**
1254  * vm_insert_page - insert single page into user vma
1255  * @vma: user vma to map to
1256  * @addr: target user address of this page
1257  * @page: source kernel page
1258  *
1259  * This allows drivers to insert individual pages they've allocated
1260  * into a user vma.
1261  *
1262  * The page has to be a nice clean _individual_ kernel allocation.
1263  * If you allocate a compound page, you need to have marked it as
1264  * such (__GFP_COMP), or manually just split the page up yourself
1265  * (see split_page()).
1266  *
1267  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1268  * took an arbitrary page protection parameter. This doesn't allow
1269  * that. Your vma protection will have to be set up correctly, which
1270  * means that if you want a shared writable mapping, you'd better
1271  * ask for a shared writable mapping!
1272  *
1273  * The page does not need to be reserved.
1274  */
1275 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1276 {
1277         if (addr < vma->vm_start || addr >= vma->vm_end)
1278                 return -EFAULT;
1279         if (!page_count(page))
1280                 return -EINVAL;
1281         vma->vm_flags |= VM_INSERTPAGE;
1282         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1283 }
1284 EXPORT_SYMBOL(vm_insert_page);
1285
1286 /**
1287  * vm_insert_pfn - insert single pfn into user vma
1288  * @vma: user vma to map to
1289  * @addr: target user address of this page
1290  * @pfn: source kernel pfn
1291  *
1292  * Similar to vm_inert_page, this allows drivers to insert individual pages
1293  * they've allocated into a user vma. Same comments apply.
1294  *
1295  * This function should only be called from a vm_ops->fault handler, and
1296  * in that case the handler should return NULL.
1297  */
1298 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1299                 unsigned long pfn)
1300 {
1301         struct mm_struct *mm = vma->vm_mm;
1302         int retval;
1303         pte_t *pte, entry;
1304         spinlock_t *ptl;
1305
1306         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
1307         BUG_ON(is_cow_mapping(vma->vm_flags));
1308
1309         retval = -ENOMEM;
1310         pte = get_locked_pte(mm, addr, &ptl);
1311         if (!pte)
1312                 goto out;
1313         retval = -EBUSY;
1314         if (!pte_none(*pte))
1315                 goto out_unlock;
1316
1317         /* Ok, finally just insert the thing.. */
1318         entry = pfn_pte(pfn, vma->vm_page_prot);
1319         set_pte_at(mm, addr, pte, entry);
1320         update_mmu_cache(vma, addr, entry);
1321
1322         retval = 0;
1323 out_unlock:
1324         pte_unmap_unlock(pte, ptl);
1325
1326 out:
1327         return retval;
1328 }
1329 EXPORT_SYMBOL(vm_insert_pfn);
1330
1331 /*
1332  * maps a range of physical memory into the requested pages. the old
1333  * mappings are removed. any references to nonexistent pages results
1334  * in null mappings (currently treated as "copy-on-access")
1335  */
1336 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1337                         unsigned long addr, unsigned long end,
1338                         unsigned long pfn, pgprot_t prot)
1339 {
1340         pte_t *pte;
1341         spinlock_t *ptl;
1342
1343         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1344         if (!pte)
1345                 return -ENOMEM;
1346         arch_enter_lazy_mmu_mode();
1347         do {
1348                 BUG_ON(!pte_none(*pte));
1349                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1350                 pfn++;
1351         } while (pte++, addr += PAGE_SIZE, addr != end);
1352         arch_leave_lazy_mmu_mode();
1353         pte_unmap_unlock(pte - 1, ptl);
1354         return 0;
1355 }
1356
1357 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1358                         unsigned long addr, unsigned long end,
1359                         unsigned long pfn, pgprot_t prot)
1360 {
1361         pmd_t *pmd;
1362         unsigned long next;
1363
1364         pfn -= addr >> PAGE_SHIFT;
1365         pmd = pmd_alloc(mm, pud, addr);
1366         if (!pmd)
1367                 return -ENOMEM;
1368         do {
1369                 next = pmd_addr_end(addr, end);
1370                 if (remap_pte_range(mm, pmd, addr, next,
1371                                 pfn + (addr >> PAGE_SHIFT), prot))
1372                         return -ENOMEM;
1373         } while (pmd++, addr = next, addr != end);
1374         return 0;
1375 }
1376
1377 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1378                         unsigned long addr, unsigned long end,
1379                         unsigned long pfn, pgprot_t prot)
1380 {
1381         pud_t *pud;
1382         unsigned long next;
1383
1384         pfn -= addr >> PAGE_SHIFT;
1385         pud = pud_alloc(mm, pgd, addr);
1386         if (!pud)
1387                 return -ENOMEM;
1388         do {
1389                 next = pud_addr_end(addr, end);
1390                 if (remap_pmd_range(mm, pud, addr, next,
1391                                 pfn + (addr >> PAGE_SHIFT), prot))
1392                         return -ENOMEM;
1393         } while (pud++, addr = next, addr != end);
1394         return 0;
1395 }
1396
1397 /**
1398  * remap_pfn_range - remap kernel memory to userspace
1399  * @vma: user vma to map to
1400  * @addr: target user address to start at
1401  * @pfn: physical address of kernel memory
1402  * @size: size of map area
1403  * @prot: page protection flags for this mapping
1404  *
1405  *  Note: this is only safe if the mm semaphore is held when called.
1406  */
1407 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1408                     unsigned long pfn, unsigned long size, pgprot_t prot)
1409 {
1410         pgd_t *pgd;
1411         unsigned long next;
1412         unsigned long end = addr + PAGE_ALIGN(size);
1413         struct mm_struct *mm = vma->vm_mm;
1414         int err;
1415
1416         /*
1417          * Physically remapped pages are special. Tell the
1418          * rest of the world about it:
1419          *   VM_IO tells people not to look at these pages
1420          *      (accesses can have side effects).
1421          *   VM_RESERVED is specified all over the place, because
1422          *      in 2.4 it kept swapout's vma scan off this vma; but
1423          *      in 2.6 the LRU scan won't even find its pages, so this
1424          *      flag means no more than count its pages in reserved_vm,
1425          *      and omit it from core dump, even when VM_IO turned off.
1426          *   VM_PFNMAP tells the core MM that the base pages are just
1427          *      raw PFN mappings, and do not have a "struct page" associated
1428          *      with them.
1429          *
1430          * There's a horrible special case to handle copy-on-write
1431          * behaviour that some programs depend on. We mark the "original"
1432          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1433          */
1434         if (is_cow_mapping(vma->vm_flags)) {
1435                 if (addr != vma->vm_start || end != vma->vm_end)
1436                         return -EINVAL;
1437                 vma->vm_pgoff = pfn;
1438         }
1439
1440         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1441
1442         BUG_ON(addr >= end);
1443         pfn -= addr >> PAGE_SHIFT;
1444         pgd = pgd_offset(mm, addr);
1445         flush_cache_range(vma, addr, end);
1446         do {
1447                 next = pgd_addr_end(addr, end);
1448                 err = remap_pud_range(mm, pgd, addr, next,
1449                                 pfn + (addr >> PAGE_SHIFT), prot);
1450                 if (err)
1451                         break;
1452         } while (pgd++, addr = next, addr != end);
1453         return err;
1454 }
1455 EXPORT_SYMBOL(remap_pfn_range);
1456
1457 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1458                                      unsigned long addr, unsigned long end,
1459                                      pte_fn_t fn, void *data)
1460 {
1461         pte_t *pte;
1462         int err;
1463         struct page *pmd_page;
1464         spinlock_t *uninitialized_var(ptl);
1465
1466         pte = (mm == &init_mm) ?
1467                 pte_alloc_kernel(pmd, addr) :
1468                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1469         if (!pte)
1470                 return -ENOMEM;
1471
1472         BUG_ON(pmd_huge(*pmd));
1473
1474         pmd_page = pmd_page(*pmd);
1475
1476         do {
1477                 err = fn(pte, pmd_page, addr, data);
1478                 if (err)
1479                         break;
1480         } while (pte++, addr += PAGE_SIZE, addr != end);
1481
1482         if (mm != &init_mm)
1483                 pte_unmap_unlock(pte-1, ptl);
1484         return err;
1485 }
1486
1487 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1488                                      unsigned long addr, unsigned long end,
1489                                      pte_fn_t fn, void *data)
1490 {
1491         pmd_t *pmd;
1492         unsigned long next;
1493         int err;
1494
1495         pmd = pmd_alloc(mm, pud, addr);
1496         if (!pmd)
1497                 return -ENOMEM;
1498         do {
1499                 next = pmd_addr_end(addr, end);
1500                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1501                 if (err)
1502                         break;
1503         } while (pmd++, addr = next, addr != end);
1504         return err;
1505 }
1506
1507 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1508                                      unsigned long addr, unsigned long end,
1509                                      pte_fn_t fn, void *data)
1510 {
1511         pud_t *pud;
1512         unsigned long next;
1513         int err;
1514
1515         pud = pud_alloc(mm, pgd, addr);
1516         if (!pud)
1517                 return -ENOMEM;
1518         do {
1519                 next = pud_addr_end(addr, end);
1520                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1521                 if (err)
1522                         break;
1523         } while (pud++, addr = next, addr != end);
1524         return err;
1525 }
1526
1527 /*
1528  * Scan a region of virtual memory, filling in page tables as necessary
1529  * and calling a provided function on each leaf page table.
1530  */
1531 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1532                         unsigned long size, pte_fn_t fn, void *data)
1533 {
1534         pgd_t *pgd;
1535         unsigned long next;
1536         unsigned long end = addr + size;
1537         int err;
1538
1539         BUG_ON(addr >= end);
1540         pgd = pgd_offset(mm, addr);
1541         do {
1542                 next = pgd_addr_end(addr, end);
1543                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1544                 if (err)
1545                         break;
1546         } while (pgd++, addr = next, addr != end);
1547         return err;
1548 }
1549 EXPORT_SYMBOL_GPL(apply_to_page_range);
1550
1551 /*
1552  * handle_pte_fault chooses page fault handler according to an entry
1553  * which was read non-atomically.  Before making any commitment, on
1554  * those architectures or configurations (e.g. i386 with PAE) which
1555  * might give a mix of unmatched parts, do_swap_page and do_file_page
1556  * must check under lock before unmapping the pte and proceeding
1557  * (but do_wp_page is only called after already making such a check;
1558  * and do_anonymous_page and do_no_page can safely check later on).
1559  */
1560 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1561                                 pte_t *page_table, pte_t orig_pte)
1562 {
1563         int same = 1;
1564 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1565         if (sizeof(pte_t) > sizeof(unsigned long)) {
1566                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1567                 spin_lock(ptl);
1568                 same = pte_same(*page_table, orig_pte);
1569                 spin_unlock(ptl);
1570         }
1571 #endif
1572         pte_unmap(page_table);
1573         return same;
1574 }
1575
1576 /*
1577  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1578  * servicing faults for write access.  In the normal case, do always want
1579  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1580  * that do not have writing enabled, when used by access_process_vm.
1581  */
1582 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1583 {
1584         if (likely(vma->vm_flags & VM_WRITE))
1585                 pte = pte_mkwrite(pte);
1586         return pte;
1587 }
1588
1589 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1590 {
1591         /*
1592          * If the source page was a PFN mapping, we don't have
1593          * a "struct page" for it. We do a best-effort copy by
1594          * just copying from the original user address. If that
1595          * fails, we just zero-fill it. Live with it.
1596          */
1597         if (unlikely(!src)) {
1598                 void *kaddr = kmap_atomic(dst, KM_USER0);
1599                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1600
1601                 /*
1602                  * This really shouldn't fail, because the page is there
1603                  * in the page tables. But it might just be unreadable,
1604                  * in which case we just give up and fill the result with
1605                  * zeroes.
1606                  */
1607                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1608                         memset(kaddr, 0, PAGE_SIZE);
1609                 kunmap_atomic(kaddr, KM_USER0);
1610                 flush_dcache_page(dst);
1611                 return;
1612
1613         }
1614         copy_user_highpage(dst, src, va, vma);
1615 }
1616
1617 /*
1618  * This routine handles present pages, when users try to write
1619  * to a shared page. It is done by copying the page to a new address
1620  * and decrementing the shared-page counter for the old page.
1621  *
1622  * Note that this routine assumes that the protection checks have been
1623  * done by the caller (the low-level page fault routine in most cases).
1624  * Thus we can safely just mark it writable once we've done any necessary
1625  * COW.
1626  *
1627  * We also mark the page dirty at this point even though the page will
1628  * change only once the write actually happens. This avoids a few races,
1629  * and potentially makes it more efficient.
1630  *
1631  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1632  * but allow concurrent faults), with pte both mapped and locked.
1633  * We return with mmap_sem still held, but pte unmapped and unlocked.
1634  */
1635 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1636                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1637                 spinlock_t *ptl, pte_t orig_pte)
1638 {
1639         struct page *old_page, *new_page;
1640         pte_t entry;
1641         int reuse = 0, ret = 0;
1642         int page_mkwrite = 0;
1643         struct page *dirty_page = NULL;
1644
1645         old_page = vm_normal_page(vma, address, orig_pte);
1646         if (!old_page)
1647                 goto gotten;
1648
1649         /*
1650          * Take out anonymous pages first, anonymous shared vmas are
1651          * not dirty accountable.
1652          */
1653         if (PageAnon(old_page)) {
1654                 if (!TestSetPageLocked(old_page)) {
1655                         reuse = can_share_swap_page(old_page);
1656                         unlock_page(old_page);
1657                 }
1658         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1659                                         (VM_WRITE|VM_SHARED))) {
1660                 /*
1661                  * Only catch write-faults on shared writable pages,
1662                  * read-only shared pages can get COWed by
1663                  * get_user_pages(.write=1, .force=1).
1664                  */
1665                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1666                         /*
1667                          * Notify the address space that the page is about to
1668                          * become writable so that it can prohibit this or wait
1669                          * for the page to get into an appropriate state.
1670                          *
1671                          * We do this without the lock held, so that it can
1672                          * sleep if it needs to.
1673                          */
1674                         page_cache_get(old_page);
1675                         pte_unmap_unlock(page_table, ptl);
1676
1677                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1678                                 goto unwritable_page;
1679
1680                         /*
1681                          * Since we dropped the lock we need to revalidate
1682                          * the PTE as someone else may have changed it.  If
1683                          * they did, we just return, as we can count on the
1684                          * MMU to tell us if they didn't also make it writable.
1685                          */
1686                         page_table = pte_offset_map_lock(mm, pmd, address,
1687                                                          &ptl);
1688                         page_cache_release(old_page);
1689                         if (!pte_same(*page_table, orig_pte))
1690                                 goto unlock;
1691
1692                         page_mkwrite = 1;
1693                 }
1694                 dirty_page = old_page;
1695                 get_page(dirty_page);
1696                 reuse = 1;
1697         }
1698
1699         if (reuse) {
1700                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1701                 entry = pte_mkyoung(orig_pte);
1702                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1703                 if (ptep_set_access_flags(vma, address, page_table, entry,1)) {
1704                         update_mmu_cache(vma, address, entry);
1705                         lazy_mmu_prot_update(entry);
1706                 }
1707                 ret |= VM_FAULT_WRITE;
1708                 goto unlock;
1709         }
1710
1711         /*
1712          * Ok, we need to copy. Oh, well..
1713          */
1714         page_cache_get(old_page);
1715 gotten:
1716         pte_unmap_unlock(page_table, ptl);
1717
1718         if (unlikely(anon_vma_prepare(vma)))
1719                 goto oom;
1720         if (old_page == ZERO_PAGE(address)) {
1721                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
1722                 if (!new_page)
1723                         goto oom;
1724         } else {
1725                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1726                 if (!new_page)
1727                         goto oom;
1728                 cow_user_page(new_page, old_page, address, vma);
1729         }
1730
1731         /*
1732          * Re-check the pte - we dropped the lock
1733          */
1734         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1735         if (likely(pte_same(*page_table, orig_pte))) {
1736                 if (old_page) {
1737                         page_remove_rmap(old_page, vma);
1738                         if (!PageAnon(old_page)) {
1739                                 dec_mm_counter(mm, file_rss);
1740                                 inc_mm_counter(mm, anon_rss);
1741                         }
1742                 } else
1743                         inc_mm_counter(mm, anon_rss);
1744                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1745                 entry = mk_pte(new_page, vma->vm_page_prot);
1746                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1747                 lazy_mmu_prot_update(entry);
1748                 /*
1749                  * Clear the pte entry and flush it first, before updating the
1750                  * pte with the new entry. This will avoid a race condition
1751                  * seen in the presence of one thread doing SMC and another
1752                  * thread doing COW.
1753                  */
1754                 ptep_clear_flush(vma, address, page_table);
1755                 set_pte_at(mm, address, page_table, entry);
1756                 update_mmu_cache(vma, address, entry);
1757                 lru_cache_add_active(new_page);
1758                 page_add_new_anon_rmap(new_page, vma, address);
1759
1760                 /* Free the old page.. */
1761                 new_page = old_page;
1762                 ret |= VM_FAULT_WRITE;
1763         }
1764         if (new_page)
1765                 page_cache_release(new_page);
1766         if (old_page)
1767                 page_cache_release(old_page);
1768 unlock:
1769         pte_unmap_unlock(page_table, ptl);
1770         if (dirty_page) {
1771                 /*
1772                  * Yes, Virginia, this is actually required to prevent a race
1773                  * with clear_page_dirty_for_io() from clearing the page dirty
1774                  * bit after it clear all dirty ptes, but before a racing
1775                  * do_wp_page installs a dirty pte.
1776                  *
1777                  * do_no_page is protected similarly.
1778                  */
1779                 wait_on_page_locked(dirty_page);
1780                 set_page_dirty_balance(dirty_page, page_mkwrite);
1781                 put_page(dirty_page);
1782         }
1783         return ret;
1784 oom:
1785         if (old_page)
1786                 page_cache_release(old_page);
1787         return VM_FAULT_OOM;
1788
1789 unwritable_page:
1790         page_cache_release(old_page);
1791         return VM_FAULT_SIGBUS;
1792 }
1793
1794 /*
1795  * Helper functions for unmap_mapping_range().
1796  *
1797  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1798  *
1799  * We have to restart searching the prio_tree whenever we drop the lock,
1800  * since the iterator is only valid while the lock is held, and anyway
1801  * a later vma might be split and reinserted earlier while lock dropped.
1802  *
1803  * The list of nonlinear vmas could be handled more efficiently, using
1804  * a placeholder, but handle it in the same way until a need is shown.
1805  * It is important to search the prio_tree before nonlinear list: a vma
1806  * may become nonlinear and be shifted from prio_tree to nonlinear list
1807  * while the lock is dropped; but never shifted from list to prio_tree.
1808  *
1809  * In order to make forward progress despite restarting the search,
1810  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1811  * quickly skip it next time around.  Since the prio_tree search only
1812  * shows us those vmas affected by unmapping the range in question, we
1813  * can't efficiently keep all vmas in step with mapping->truncate_count:
1814  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1815  * mapping->truncate_count and vma->vm_truncate_count are protected by
1816  * i_mmap_lock.
1817  *
1818  * In order to make forward progress despite repeatedly restarting some
1819  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1820  * and restart from that address when we reach that vma again.  It might
1821  * have been split or merged, shrunk or extended, but never shifted: so
1822  * restart_addr remains valid so long as it remains in the vma's range.
1823  * unmap_mapping_range forces truncate_count to leap over page-aligned
1824  * values so we can save vma's restart_addr in its truncate_count field.
1825  */
1826 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1827
1828 static void reset_vma_truncate_counts(struct address_space *mapping)
1829 {
1830         struct vm_area_struct *vma;
1831         struct prio_tree_iter iter;
1832
1833         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1834                 vma->vm_truncate_count = 0;
1835         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1836                 vma->vm_truncate_count = 0;
1837 }
1838
1839 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1840                 unsigned long start_addr, unsigned long end_addr,
1841                 struct zap_details *details)
1842 {
1843         unsigned long restart_addr;
1844         int need_break;
1845
1846         /*
1847          * files that support invalidating or truncating portions of the
1848          * file from under mmaped areas must have their ->fault function
1849          * return a locked page (and set VM_FAULT_LOCKED in the return).
1850          * This provides synchronisation against concurrent unmapping here.
1851          */
1852
1853 again:
1854         restart_addr = vma->vm_truncate_count;
1855         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1856                 start_addr = restart_addr;
1857                 if (start_addr >= end_addr) {
1858                         /* Top of vma has been split off since last time */
1859                         vma->vm_truncate_count = details->truncate_count;
1860                         return 0;
1861                 }
1862         }
1863
1864         restart_addr = zap_page_range(vma, start_addr,
1865                                         end_addr - start_addr, details);
1866         need_break = need_resched() ||
1867                         need_lockbreak(details->i_mmap_lock);
1868
1869         if (restart_addr >= end_addr) {
1870                 /* We have now completed this vma: mark it so */
1871                 vma->vm_truncate_count = details->truncate_count;
1872                 if (!need_break)
1873                         return 0;
1874         } else {
1875                 /* Note restart_addr in vma's truncate_count field */
1876                 vma->vm_truncate_count = restart_addr;
1877                 if (!need_break)
1878                         goto again;
1879         }
1880
1881         spin_unlock(details->i_mmap_lock);
1882         cond_resched();
1883         spin_lock(details->i_mmap_lock);
1884         return -EINTR;
1885 }
1886
1887 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1888                                             struct zap_details *details)
1889 {
1890         struct vm_area_struct *vma;
1891         struct prio_tree_iter iter;
1892         pgoff_t vba, vea, zba, zea;
1893
1894 restart:
1895         vma_prio_tree_foreach(vma, &iter, root,
1896                         details->first_index, details->last_index) {
1897                 /* Skip quickly over those we have already dealt with */
1898                 if (vma->vm_truncate_count == details->truncate_count)
1899                         continue;
1900
1901                 vba = vma->vm_pgoff;
1902                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1903                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1904                 zba = details->first_index;
1905                 if (zba < vba)
1906                         zba = vba;
1907                 zea = details->last_index;
1908                 if (zea > vea)
1909                         zea = vea;
1910
1911                 if (unmap_mapping_range_vma(vma,
1912                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1913                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1914                                 details) < 0)
1915                         goto restart;
1916         }
1917 }
1918
1919 static inline void unmap_mapping_range_list(struct list_head *head,
1920                                             struct zap_details *details)
1921 {
1922         struct vm_area_struct *vma;
1923
1924         /*
1925          * In nonlinear VMAs there is no correspondence between virtual address
1926          * offset and file offset.  So we must perform an exhaustive search
1927          * across *all* the pages in each nonlinear VMA, not just the pages
1928          * whose virtual address lies outside the file truncation point.
1929          */
1930 restart:
1931         list_for_each_entry(vma, head, shared.vm_set.list) {
1932                 /* Skip quickly over those we have already dealt with */
1933                 if (vma->vm_truncate_count == details->truncate_count)
1934                         continue;
1935                 details->nonlinear_vma = vma;
1936                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1937                                         vma->vm_end, details) < 0)
1938                         goto restart;
1939         }
1940 }
1941
1942 /**
1943  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
1944  * @mapping: the address space containing mmaps to be unmapped.
1945  * @holebegin: byte in first page to unmap, relative to the start of
1946  * the underlying file.  This will be rounded down to a PAGE_SIZE
1947  * boundary.  Note that this is different from vmtruncate(), which
1948  * must keep the partial page.  In contrast, we must get rid of
1949  * partial pages.
1950  * @holelen: size of prospective hole in bytes.  This will be rounded
1951  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1952  * end of the file.
1953  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1954  * but 0 when invalidating pagecache, don't throw away private data.
1955  */
1956 void unmap_mapping_range(struct address_space *mapping,
1957                 loff_t const holebegin, loff_t const holelen, int even_cows)
1958 {
1959         struct zap_details details;
1960         pgoff_t hba = holebegin >> PAGE_SHIFT;
1961         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1962
1963         /* Check for overflow. */
1964         if (sizeof(holelen) > sizeof(hlen)) {
1965                 long long holeend =
1966                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1967                 if (holeend & ~(long long)ULONG_MAX)
1968                         hlen = ULONG_MAX - hba + 1;
1969         }
1970
1971         details.check_mapping = even_cows? NULL: mapping;
1972         details.nonlinear_vma = NULL;
1973         details.first_index = hba;
1974         details.last_index = hba + hlen - 1;
1975         if (details.last_index < details.first_index)
1976                 details.last_index = ULONG_MAX;
1977         details.i_mmap_lock = &mapping->i_mmap_lock;
1978
1979         spin_lock(&mapping->i_mmap_lock);
1980
1981         /* Protect against endless unmapping loops */
1982         mapping->truncate_count++;
1983         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1984                 if (mapping->truncate_count == 0)
1985                         reset_vma_truncate_counts(mapping);
1986                 mapping->truncate_count++;
1987         }
1988         details.truncate_count = mapping->truncate_count;
1989
1990         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1991                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1992         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1993                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1994         spin_unlock(&mapping->i_mmap_lock);
1995 }
1996 EXPORT_SYMBOL(unmap_mapping_range);
1997
1998 /**
1999  * vmtruncate - unmap mappings "freed" by truncate() syscall
2000  * @inode: inode of the file used
2001  * @offset: file offset to start truncating
2002  *
2003  * NOTE! We have to be ready to update the memory sharing
2004  * between the file and the memory map for a potential last
2005  * incomplete page.  Ugly, but necessary.
2006  */
2007 int vmtruncate(struct inode * inode, loff_t offset)
2008 {
2009         struct address_space *mapping = inode->i_mapping;
2010         unsigned long limit;
2011
2012         if (inode->i_size < offset)
2013                 goto do_expand;
2014         /*
2015          * truncation of in-use swapfiles is disallowed - it would cause
2016          * subsequent swapout to scribble on the now-freed blocks.
2017          */
2018         if (IS_SWAPFILE(inode))
2019                 goto out_busy;
2020         i_size_write(inode, offset);
2021
2022         /*
2023          * unmap_mapping_range is called twice, first simply for efficiency
2024          * so that truncate_inode_pages does fewer single-page unmaps. However
2025          * after this first call, and before truncate_inode_pages finishes,
2026          * it is possible for private pages to be COWed, which remain after
2027          * truncate_inode_pages finishes, hence the second unmap_mapping_range
2028          * call must be made for correctness.
2029          */
2030         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2031         truncate_inode_pages(mapping, offset);
2032         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2033         goto out_truncate;
2034
2035 do_expand:
2036         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2037         if (limit != RLIM_INFINITY && offset > limit)
2038                 goto out_sig;
2039         if (offset > inode->i_sb->s_maxbytes)
2040                 goto out_big;
2041         i_size_write(inode, offset);
2042
2043 out_truncate:
2044         if (inode->i_op && inode->i_op->truncate)
2045                 inode->i_op->truncate(inode);
2046         return 0;
2047 out_sig:
2048         send_sig(SIGXFSZ, current, 0);
2049 out_big:
2050         return -EFBIG;
2051 out_busy:
2052         return -ETXTBSY;
2053 }
2054 EXPORT_SYMBOL(vmtruncate);
2055
2056 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2057 {
2058         struct address_space *mapping = inode->i_mapping;
2059
2060         /*
2061          * If the underlying filesystem is not going to provide
2062          * a way to truncate a range of blocks (punch a hole) -
2063          * we should return failure right now.
2064          */
2065         if (!inode->i_op || !inode->i_op->truncate_range)
2066                 return -ENOSYS;
2067
2068         mutex_lock(&inode->i_mutex);
2069         down_write(&inode->i_alloc_sem);
2070         unmap_mapping_range(mapping, offset, (end - offset), 1);
2071         truncate_inode_pages_range(mapping, offset, end);
2072         unmap_mapping_range(mapping, offset, (end - offset), 1);
2073         inode->i_op->truncate_range(inode, offset, end);
2074         up_write(&inode->i_alloc_sem);
2075         mutex_unlock(&inode->i_mutex);
2076
2077         return 0;
2078 }
2079
2080 /**
2081  * swapin_readahead - swap in pages in hope we need them soon
2082  * @entry: swap entry of this memory
2083  * @addr: address to start
2084  * @vma: user vma this addresses belong to
2085  *
2086  * Primitive swap readahead code. We simply read an aligned block of
2087  * (1 << page_cluster) entries in the swap area. This method is chosen
2088  * because it doesn't cost us any seek time.  We also make sure to queue
2089  * the 'original' request together with the readahead ones...
2090  *
2091  * This has been extended to use the NUMA policies from the mm triggering
2092  * the readahead.
2093  *
2094  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
2095  */
2096 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
2097 {
2098 #ifdef CONFIG_NUMA
2099         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
2100 #endif
2101         int i, num;
2102         struct page *new_page;
2103         unsigned long offset;
2104
2105         /*
2106          * Get the number of handles we should do readahead io to.
2107          */
2108         num = valid_swaphandles(entry, &offset);
2109         for (i = 0; i < num; offset++, i++) {
2110                 /* Ok, do the async read-ahead now */
2111                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
2112                                                            offset), vma, addr);
2113                 if (!new_page)
2114                         break;
2115                 page_cache_release(new_page);
2116 #ifdef CONFIG_NUMA
2117                 /*
2118                  * Find the next applicable VMA for the NUMA policy.
2119                  */
2120                 addr += PAGE_SIZE;
2121                 if (addr == 0)
2122                         vma = NULL;
2123                 if (vma) {
2124                         if (addr >= vma->vm_end) {
2125                                 vma = next_vma;
2126                                 next_vma = vma ? vma->vm_next : NULL;
2127                         }
2128                         if (vma && addr < vma->vm_start)
2129                                 vma = NULL;
2130                 } else {
2131                         if (next_vma && addr >= next_vma->vm_start) {
2132                                 vma = next_vma;
2133                                 next_vma = vma->vm_next;
2134                         }
2135                 }
2136 #endif
2137         }
2138         lru_add_drain();        /* Push any new pages onto the LRU now */
2139 }
2140
2141 /*
2142  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2143  * but allow concurrent faults), and pte mapped but not yet locked.
2144  * We return with mmap_sem still held, but pte unmapped and unlocked.
2145  */
2146 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2147                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2148                 int write_access, pte_t orig_pte)
2149 {
2150         spinlock_t *ptl;
2151         struct page *page;
2152         swp_entry_t entry;
2153         pte_t pte;
2154         int ret = 0;
2155
2156         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2157                 goto out;
2158
2159         entry = pte_to_swp_entry(orig_pte);
2160         if (is_migration_entry(entry)) {
2161                 migration_entry_wait(mm, pmd, address);
2162                 goto out;
2163         }
2164         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2165         page = lookup_swap_cache(entry);
2166         if (!page) {
2167                 grab_swap_token(); /* Contend for token _before_ read-in */
2168                 swapin_readahead(entry, address, vma);
2169                 page = read_swap_cache_async(entry, vma, address);
2170                 if (!page) {
2171                         /*
2172                          * Back out if somebody else faulted in this pte
2173                          * while we released the pte lock.
2174                          */
2175                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2176                         if (likely(pte_same(*page_table, orig_pte)))
2177                                 ret = VM_FAULT_OOM;
2178                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2179                         goto unlock;
2180                 }
2181
2182                 /* Had to read the page from swap area: Major fault */
2183                 ret = VM_FAULT_MAJOR;
2184                 count_vm_event(PGMAJFAULT);
2185         }
2186
2187         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2188         mark_page_accessed(page);
2189         lock_page(page);
2190
2191         /*
2192          * Back out if somebody else already faulted in this pte.
2193          */
2194         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2195         if (unlikely(!pte_same(*page_table, orig_pte)))
2196                 goto out_nomap;
2197
2198         if (unlikely(!PageUptodate(page))) {
2199                 ret = VM_FAULT_SIGBUS;
2200                 goto out_nomap;
2201         }
2202
2203         /* The page isn't present yet, go ahead with the fault. */
2204
2205         inc_mm_counter(mm, anon_rss);
2206         pte = mk_pte(page, vma->vm_page_prot);
2207         if (write_access && can_share_swap_page(page)) {
2208                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2209                 write_access = 0;
2210         }
2211
2212         flush_icache_page(vma, page);
2213         set_pte_at(mm, address, page_table, pte);
2214         page_add_anon_rmap(page, vma, address);
2215
2216         swap_free(entry);
2217         if (vm_swap_full())
2218                 remove_exclusive_swap_page(page);
2219         unlock_page(page);
2220
2221         if (write_access) {
2222                 /* XXX: We could OR the do_wp_page code with this one? */
2223                 if (do_wp_page(mm, vma, address,
2224                                 page_table, pmd, ptl, pte) & VM_FAULT_OOM)
2225                         ret = VM_FAULT_OOM;
2226                 goto out;
2227         }
2228
2229         /* No need to invalidate - it was non-present before */
2230         update_mmu_cache(vma, address, pte);
2231 unlock:
2232         pte_unmap_unlock(page_table, ptl);
2233 out:
2234         return ret;
2235 out_nomap:
2236         pte_unmap_unlock(page_table, ptl);
2237         unlock_page(page);
2238         page_cache_release(page);
2239         return ret;
2240 }
2241
2242 /*
2243  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2244  * but allow concurrent faults), and pte mapped but not yet locked.
2245  * We return with mmap_sem still held, but pte unmapped and unlocked.
2246  */
2247 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2248                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2249                 int write_access)
2250 {
2251         struct page *page;
2252         spinlock_t *ptl;
2253         pte_t entry;
2254
2255         if (write_access) {
2256                 /* Allocate our own private page. */
2257                 pte_unmap(page_table);
2258
2259                 if (unlikely(anon_vma_prepare(vma)))
2260                         goto oom;
2261                 page = alloc_zeroed_user_highpage_movable(vma, address);
2262                 if (!page)
2263                         goto oom;
2264
2265                 entry = mk_pte(page, vma->vm_page_prot);
2266                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2267
2268                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2269                 if (!pte_none(*page_table))
2270                         goto release;
2271                 inc_mm_counter(mm, anon_rss);
2272                 lru_cache_add_active(page);
2273                 page_add_new_anon_rmap(page, vma, address);
2274         } else {
2275                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2276                 page = ZERO_PAGE(address);
2277                 page_cache_get(page);
2278                 entry = mk_pte(page, vma->vm_page_prot);
2279
2280                 ptl = pte_lockptr(mm, pmd);
2281                 spin_lock(ptl);
2282                 if (!pte_none(*page_table))
2283                         goto release;
2284                 inc_mm_counter(mm, file_rss);
2285                 page_add_file_rmap(page);
2286         }
2287
2288         set_pte_at(mm, address, page_table, entry);
2289
2290         /* No need to invalidate - it was non-present before */
2291         update_mmu_cache(vma, address, entry);
2292         lazy_mmu_prot_update(entry);
2293 unlock:
2294         pte_unmap_unlock(page_table, ptl);
2295         return 0;
2296 release:
2297         page_cache_release(page);
2298         goto unlock;
2299 oom:
2300         return VM_FAULT_OOM;
2301 }
2302
2303 /*
2304  * __do_fault() tries to create a new page mapping. It aggressively
2305  * tries to share with existing pages, but makes a separate copy if
2306  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2307  * the next page fault.
2308  *
2309  * As this is called only for pages that do not currently exist, we
2310  * do not need to flush old virtual caches or the TLB.
2311  *
2312  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2313  * but allow concurrent faults), and pte neither mapped nor locked.
2314  * We return with mmap_sem still held, but pte unmapped and unlocked.
2315  */
2316 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2317                 unsigned long address, pmd_t *pmd,
2318                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2319 {
2320         pte_t *page_table;
2321         spinlock_t *ptl;
2322         struct page *page;
2323         pte_t entry;
2324         int anon = 0;
2325         struct page *dirty_page = NULL;
2326         struct vm_fault vmf;
2327         int ret;
2328         int page_mkwrite = 0;
2329
2330         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2331         vmf.pgoff = pgoff;
2332         vmf.flags = flags;
2333         vmf.page = NULL;
2334
2335         BUG_ON(vma->vm_flags & VM_PFNMAP);
2336
2337         if (likely(vma->vm_ops->fault)) {
2338                 ret = vma->vm_ops->fault(vma, &vmf);
2339                 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2340                         return ret;
2341         } else {
2342                 /* Legacy ->nopage path */
2343                 ret = 0;
2344                 vmf.page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2345                 /* no page was available -- either SIGBUS or OOM */
2346                 if (unlikely(vmf.page == NOPAGE_SIGBUS))
2347                         return VM_FAULT_SIGBUS;
2348                 else if (unlikely(vmf.page == NOPAGE_OOM))
2349                         return VM_FAULT_OOM;
2350         }
2351
2352         /*
2353          * For consistency in subsequent calls, make the faulted page always
2354          * locked.
2355          */
2356         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2357                 lock_page(vmf.page);
2358         else
2359                 VM_BUG_ON(!PageLocked(vmf.page));
2360
2361         /*
2362          * Should we do an early C-O-W break?
2363          */
2364         page = vmf.page;
2365         if (flags & FAULT_FLAG_WRITE) {
2366                 if (!(vma->vm_flags & VM_SHARED)) {
2367                         anon = 1;
2368                         if (unlikely(anon_vma_prepare(vma))) {
2369                                 ret = VM_FAULT_OOM;
2370                                 goto out;
2371                         }
2372                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2373                                                 vma, address);
2374                         if (!page) {
2375                                 ret = VM_FAULT_OOM;
2376                                 goto out;
2377                         }
2378                         copy_user_highpage(page, vmf.page, address, vma);
2379                 } else {
2380                         /*
2381                          * If the page will be shareable, see if the backing
2382                          * address space wants to know that the page is about
2383                          * to become writable
2384                          */
2385                         if (vma->vm_ops->page_mkwrite) {
2386                                 unlock_page(page);
2387                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2388                                         ret = VM_FAULT_SIGBUS;
2389                                         anon = 1; /* no anon but release vmf.page */
2390                                         goto out_unlocked;
2391                                 }
2392                                 lock_page(page);
2393                                 /*
2394                                  * XXX: this is not quite right (racy vs
2395                                  * invalidate) to unlock and relock the page
2396                                  * like this, however a better fix requires
2397                                  * reworking page_mkwrite locking API, which
2398                                  * is better done later.
2399                                  */
2400                                 if (!page->mapping) {
2401                                         ret = 0;
2402                                         anon = 1; /* no anon but release vmf.page */
2403                                         goto out;
2404                                 }
2405                                 page_mkwrite = 1;
2406                         }
2407                 }
2408
2409         }
2410
2411         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2412
2413         /*
2414          * This silly early PAGE_DIRTY setting removes a race
2415          * due to the bad i386 page protection. But it's valid
2416          * for other architectures too.
2417          *
2418          * Note that if write_access is true, we either now have
2419          * an exclusive copy of the page, or this is a shared mapping,
2420          * so we can make it writable and dirty to avoid having to
2421          * handle that later.
2422          */
2423         /* Only go through if we didn't race with anybody else... */
2424         if (likely(pte_same(*page_table, orig_pte))) {
2425                 flush_icache_page(vma, page);
2426                 entry = mk_pte(page, vma->vm_page_prot);
2427                 if (flags & FAULT_FLAG_WRITE)
2428                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2429                 set_pte_at(mm, address, page_table, entry);
2430                 if (anon) {
2431                         inc_mm_counter(mm, anon_rss);
2432                         lru_cache_add_active(page);
2433                         page_add_new_anon_rmap(page, vma, address);
2434                 } else {
2435                         inc_mm_counter(mm, file_rss);
2436                         page_add_file_rmap(page);
2437                         if (flags & FAULT_FLAG_WRITE) {
2438                                 dirty_page = page;
2439                                 get_page(dirty_page);
2440                         }
2441                 }
2442
2443                 /* no need to invalidate: a not-present page won't be cached */
2444                 update_mmu_cache(vma, address, entry);
2445                 lazy_mmu_prot_update(entry);
2446         } else {
2447                 if (anon)
2448                         page_cache_release(page);
2449                 else
2450                         anon = 1; /* no anon but release faulted_page */
2451         }
2452
2453         pte_unmap_unlock(page_table, ptl);
2454
2455 out:
2456         unlock_page(vmf.page);
2457 out_unlocked:
2458         if (anon)
2459                 page_cache_release(vmf.page);
2460         else if (dirty_page) {
2461                 set_page_dirty_balance(dirty_page, page_mkwrite);
2462                 put_page(dirty_page);
2463         }
2464
2465         return ret;
2466 }
2467
2468 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2469                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2470                 int write_access, pte_t orig_pte)
2471 {
2472         pgoff_t pgoff = (((address & PAGE_MASK)
2473                         - vma->vm_start) >> PAGE_CACHE_SHIFT) + vma->vm_pgoff;
2474         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2475
2476         pte_unmap(page_table);
2477         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2478 }
2479
2480
2481 /*
2482  * do_no_pfn() tries to create a new page mapping for a page without
2483  * a struct_page backing it
2484  *
2485  * As this is called only for pages that do not currently exist, we
2486  * do not need to flush old virtual caches or the TLB.
2487  *
2488  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2489  * but allow concurrent faults), and pte mapped but not yet locked.
2490  * We return with mmap_sem still held, but pte unmapped and unlocked.
2491  *
2492  * It is expected that the ->nopfn handler always returns the same pfn
2493  * for a given virtual mapping.
2494  *
2495  * Mark this `noinline' to prevent it from bloating the main pagefault code.
2496  */
2497 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2498                      unsigned long address, pte_t *page_table, pmd_t *pmd,
2499                      int write_access)
2500 {
2501         spinlock_t *ptl;
2502         pte_t entry;
2503         unsigned long pfn;
2504
2505         pte_unmap(page_table);
2506         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
2507         BUG_ON(is_cow_mapping(vma->vm_flags));
2508
2509         pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2510         if (unlikely(pfn == NOPFN_OOM))
2511                 return VM_FAULT_OOM;
2512         else if (unlikely(pfn == NOPFN_SIGBUS))
2513                 return VM_FAULT_SIGBUS;
2514         else if (unlikely(pfn == NOPFN_REFAULT))
2515                 return 0;
2516
2517         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2518
2519         /* Only go through if we didn't race with anybody else... */
2520         if (pte_none(*page_table)) {
2521                 entry = pfn_pte(pfn, vma->vm_page_prot);
2522                 if (write_access)
2523                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2524                 set_pte_at(mm, address, page_table, entry);
2525         }
2526         pte_unmap_unlock(page_table, ptl);
2527         return 0;
2528 }
2529
2530 /*
2531  * Fault of a previously existing named mapping. Repopulate the pte
2532  * from the encoded file_pte if possible. This enables swappable
2533  * nonlinear vmas.
2534  *
2535  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2536  * but allow concurrent faults), and pte mapped but not yet locked.
2537  * We return with mmap_sem still held, but pte unmapped and unlocked.
2538  */
2539 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2540                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2541                 int write_access, pte_t orig_pte)
2542 {
2543         unsigned int flags = FAULT_FLAG_NONLINEAR |
2544                                 (write_access ? FAULT_FLAG_WRITE : 0);
2545         pgoff_t pgoff;
2546
2547         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2548                 return 0;
2549
2550         if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2551                         !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2552                 /*
2553                  * Page table corrupted: show pte and kill process.
2554                  */
2555                 print_bad_pte(vma, orig_pte, address);
2556                 return VM_FAULT_OOM;
2557         }
2558
2559         pgoff = pte_to_pgoff(orig_pte);
2560         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2561 }
2562
2563 /*
2564  * These routines also need to handle stuff like marking pages dirty
2565  * and/or accessed for architectures that don't do it in hardware (most
2566  * RISC architectures).  The early dirtying is also good on the i386.
2567  *
2568  * There is also a hook called "update_mmu_cache()" that architectures
2569  * with external mmu caches can use to update those (ie the Sparc or
2570  * PowerPC hashed page tables that act as extended TLBs).
2571  *
2572  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2573  * but allow concurrent faults), and pte mapped but not yet locked.
2574  * We return with mmap_sem still held, but pte unmapped and unlocked.
2575  */
2576 static inline int handle_pte_fault(struct mm_struct *mm,
2577                 struct vm_area_struct *vma, unsigned long address,
2578                 pte_t *pte, pmd_t *pmd, int write_access)
2579 {
2580         pte_t entry;
2581         spinlock_t *ptl;
2582
2583         entry = *pte;
2584         if (!pte_present(entry)) {
2585                 if (pte_none(entry)) {
2586                         if (vma->vm_ops) {
2587                                 if (vma->vm_ops->fault || vma->vm_ops->nopage)
2588                                         return do_linear_fault(mm, vma, address,
2589                                                 pte, pmd, write_access, entry);
2590                                 if (unlikely(vma->vm_ops->nopfn))
2591                                         return do_no_pfn(mm, vma, address, pte,
2592                                                          pmd, write_access);
2593                         }
2594                         return do_anonymous_page(mm, vma, address,
2595                                                  pte, pmd, write_access);
2596                 }
2597                 if (pte_file(entry))
2598                         return do_nonlinear_fault(mm, vma, address,
2599                                         pte, pmd, write_access, entry);
2600                 return do_swap_page(mm, vma, address,
2601                                         pte, pmd, write_access, entry);
2602         }
2603
2604         ptl = pte_lockptr(mm, pmd);
2605         spin_lock(ptl);
2606         if (unlikely(!pte_same(*pte, entry)))
2607                 goto unlock;
2608         if (write_access) {
2609                 if (!pte_write(entry))
2610                         return do_wp_page(mm, vma, address,
2611                                         pte, pmd, ptl, entry);
2612                 entry = pte_mkdirty(entry);
2613         }
2614         entry = pte_mkyoung(entry);
2615         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2616                 update_mmu_cache(vma, address, entry);
2617                 lazy_mmu_prot_update(entry);
2618         } else {
2619                 /*
2620                  * This is needed only for protection faults but the arch code
2621                  * is not yet telling us if this is a protection fault or not.
2622                  * This still avoids useless tlb flushes for .text page faults
2623                  * with threads.
2624                  */
2625                 if (write_access)
2626                         flush_tlb_page(vma, address);
2627         }
2628 unlock:
2629         pte_unmap_unlock(pte, ptl);
2630         return 0;
2631 }
2632
2633 /*
2634  * By the time we get here, we already hold the mm semaphore
2635  */
2636 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2637                 unsigned long address, int write_access)
2638 {
2639         pgd_t *pgd;
2640         pud_t *pud;
2641         pmd_t *pmd;
2642         pte_t *pte;
2643
2644         __set_current_state(TASK_RUNNING);
2645
2646         count_vm_event(PGFAULT);
2647
2648         if (unlikely(is_vm_hugetlb_page(vma)))
2649                 return hugetlb_fault(mm, vma, address, write_access);
2650
2651         pgd = pgd_offset(mm, address);
2652         pud = pud_alloc(mm, pgd, address);
2653         if (!pud)
2654                 return VM_FAULT_OOM;
2655         pmd = pmd_alloc(mm, pud, address);
2656         if (!pmd)
2657                 return VM_FAULT_OOM;
2658         pte = pte_alloc_map(mm, pmd, address);
2659         if (!pte)
2660                 return VM_FAULT_OOM;
2661
2662         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2663 }
2664
2665 #ifndef __PAGETABLE_PUD_FOLDED
2666 /*
2667  * Allocate page upper directory.
2668  * We've already handled the fast-path in-line.
2669  */
2670 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2671 {
2672         pud_t *new = pud_alloc_one(mm, address);
2673         if (!new)
2674                 return -ENOMEM;
2675
2676         spin_lock(&mm->page_table_lock);
2677         if (pgd_present(*pgd))          /* Another has populated it */
2678                 pud_free(new);
2679         else
2680                 pgd_populate(mm, pgd, new);
2681         spin_unlock(&mm->page_table_lock);
2682         return 0;
2683 }
2684 #endif /* __PAGETABLE_PUD_FOLDED */
2685
2686 #ifndef __PAGETABLE_PMD_FOLDED
2687 /*
2688  * Allocate page middle directory.
2689  * We've already handled the fast-path in-line.
2690  */
2691 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2692 {
2693         pmd_t *new = pmd_alloc_one(mm, address);
2694         if (!new)
2695                 return -ENOMEM;
2696
2697         spin_lock(&mm->page_table_lock);
2698 #ifndef __ARCH_HAS_4LEVEL_HACK
2699         if (pud_present(*pud))          /* Another has populated it */
2700                 pmd_free(new);
2701         else
2702                 pud_populate(mm, pud, new);
2703 #else
2704         if (pgd_present(*pud))          /* Another has populated it */
2705                 pmd_free(new);
2706         else
2707                 pgd_populate(mm, pud, new);
2708 #endif /* __ARCH_HAS_4LEVEL_HACK */
2709         spin_unlock(&mm->page_table_lock);
2710         return 0;
2711 }
2712 #endif /* __PAGETABLE_PMD_FOLDED */
2713
2714 int make_pages_present(unsigned long addr, unsigned long end)
2715 {
2716         int ret, len, write;
2717         struct vm_area_struct * vma;
2718
2719         vma = find_vma(current->mm, addr);
2720         if (!vma)
2721                 return -1;
2722         write = (vma->vm_flags & VM_WRITE) != 0;
2723         BUG_ON(addr >= end);
2724         BUG_ON(end > vma->vm_end);
2725         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2726         ret = get_user_pages(current, current->mm, addr,
2727                         len, write, 0, NULL, NULL);
2728         if (ret < 0)
2729                 return ret;
2730         return ret == len ? 0 : -1;
2731 }
2732
2733 /* 
2734  * Map a vmalloc()-space virtual address to the physical page.
2735  */
2736 struct page * vmalloc_to_page(void * vmalloc_addr)
2737 {
2738         unsigned long addr = (unsigned long) vmalloc_addr;
2739         struct page *page = NULL;
2740         pgd_t *pgd = pgd_offset_k(addr);
2741         pud_t *pud;
2742         pmd_t *pmd;
2743         pte_t *ptep, pte;
2744   
2745         if (!pgd_none(*pgd)) {
2746                 pud = pud_offset(pgd, addr);
2747                 if (!pud_none(*pud)) {
2748                         pmd = pmd_offset(pud, addr);
2749                         if (!pmd_none(*pmd)) {
2750                                 ptep = pte_offset_map(pmd, addr);
2751                                 pte = *ptep;
2752                                 if (pte_present(pte))
2753                                         page = pte_page(pte);
2754                                 pte_unmap(ptep);
2755                         }
2756                 }
2757         }
2758         return page;
2759 }
2760
2761 EXPORT_SYMBOL(vmalloc_to_page);
2762
2763 /*
2764  * Map a vmalloc()-space virtual address to the physical page frame number.
2765  */
2766 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2767 {
2768         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2769 }
2770
2771 EXPORT_SYMBOL(vmalloc_to_pfn);
2772
2773 #if !defined(__HAVE_ARCH_GATE_AREA)
2774
2775 #if defined(AT_SYSINFO_EHDR)
2776 static struct vm_area_struct gate_vma;
2777
2778 static int __init gate_vma_init(void)
2779 {
2780         gate_vma.vm_mm = NULL;
2781         gate_vma.vm_start = FIXADDR_USER_START;
2782         gate_vma.vm_end = FIXADDR_USER_END;
2783         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2784         gate_vma.vm_page_prot = __P101;
2785         /*
2786          * Make sure the vDSO gets into every core dump.
2787          * Dumping its contents makes post-mortem fully interpretable later
2788          * without matching up the same kernel and hardware config to see
2789          * what PC values meant.
2790          */
2791         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2792         return 0;
2793 }
2794 __initcall(gate_vma_init);
2795 #endif
2796
2797 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2798 {
2799 #ifdef AT_SYSINFO_EHDR
2800         return &gate_vma;
2801 #else
2802         return NULL;
2803 #endif
2804 }
2805
2806 int in_gate_area_no_task(unsigned long addr)
2807 {
2808 #ifdef AT_SYSINFO_EHDR
2809         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2810                 return 1;
2811 #endif
2812         return 0;
2813 }
2814
2815 #endif  /* __HAVE_ARCH_GATE_AREA */
2816
2817 /*
2818  * Access another process' address space.
2819  * Source/target buffer must be kernel space,
2820  * Do not walk the page table directly, use get_user_pages
2821  */
2822 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2823 {
2824         struct mm_struct *mm;
2825         struct vm_area_struct *vma;
2826         struct page *page;
2827         void *old_buf = buf;
2828
2829         mm = get_task_mm(tsk);
2830         if (!mm)
2831                 return 0;
2832
2833         down_read(&mm->mmap_sem);
2834         /* ignore errors, just check how much was sucessfully transfered */
2835         while (len) {
2836                 int bytes, ret, offset;
2837                 void *maddr;
2838
2839                 ret = get_user_pages(tsk, mm, addr, 1,
2840                                 write, 1, &page, &vma);
2841                 if (ret <= 0)
2842                         break;
2843
2844                 bytes = len;
2845                 offset = addr & (PAGE_SIZE-1);
2846                 if (bytes > PAGE_SIZE-offset)
2847                         bytes = PAGE_SIZE-offset;
2848
2849                 maddr = kmap(page);
2850                 if (write) {
2851                         copy_to_user_page(vma, page, addr,
2852                                           maddr + offset, buf, bytes);
2853                         set_page_dirty_lock(page);
2854                 } else {
2855                         copy_from_user_page(vma, page, addr,
2856                                             buf, maddr + offset, bytes);
2857                 }
2858                 kunmap(page);
2859                 page_cache_release(page);
2860                 len -= bytes;
2861                 buf += bytes;
2862                 addr += bytes;
2863         }
2864         up_read(&mm->mmap_sem);
2865         mmput(mm);
2866
2867         return buf - old_buf;
2868 }
2869 EXPORT_SYMBOL_GPL(access_process_vm);