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