8d8f52569f328ab35d6ac9e68abe94c348434062
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
54 #include <asm/tlb.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
57
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
65
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
69
70 unsigned long num_physpages;
71 /*
72  * A number of key systems in x86 including ioremap() rely on the assumption
73  * that high_memory defines the upper bound on direct map memory, then end
74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76  * and ZONE_HIGHMEM.
77  */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
80
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
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 0;
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_page_state(nr_page_table_pages);
130         tlb->mm->nr_ptes--;
131 }
132
133 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
134                                 unsigned long addr, unsigned long end,
135                                 unsigned long floor, unsigned long ceiling)
136 {
137         pmd_t *pmd;
138         unsigned long next;
139         unsigned long start;
140
141         start = addr;
142         pmd = pmd_offset(pud, addr);
143         do {
144                 next = pmd_addr_end(addr, end);
145                 if (pmd_none_or_clear_bad(pmd))
146                         continue;
147                 free_pte_range(tlb, pmd);
148         } while (pmd++, addr = next, addr != end);
149
150         start &= PUD_MASK;
151         if (start < floor)
152                 return;
153         if (ceiling) {
154                 ceiling &= PUD_MASK;
155                 if (!ceiling)
156                         return;
157         }
158         if (end - 1 > ceiling - 1)
159                 return;
160
161         pmd = pmd_offset(pud, start);
162         pud_clear(pud);
163         pmd_free_tlb(tlb, pmd);
164 }
165
166 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
167                                 unsigned long addr, unsigned long end,
168                                 unsigned long floor, unsigned long ceiling)
169 {
170         pud_t *pud;
171         unsigned long next;
172         unsigned long start;
173
174         start = addr;
175         pud = pud_offset(pgd, addr);
176         do {
177                 next = pud_addr_end(addr, end);
178                 if (pud_none_or_clear_bad(pud))
179                         continue;
180                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
181         } while (pud++, addr = next, addr != end);
182
183         start &= PGDIR_MASK;
184         if (start < floor)
185                 return;
186         if (ceiling) {
187                 ceiling &= PGDIR_MASK;
188                 if (!ceiling)
189                         return;
190         }
191         if (end - 1 > ceiling - 1)
192                 return;
193
194         pud = pud_offset(pgd, start);
195         pgd_clear(pgd);
196         pud_free_tlb(tlb, pud);
197 }
198
199 /*
200  * This function frees user-level page tables of a process.
201  *
202  * Must be called with pagetable lock held.
203  */
204 void free_pgd_range(struct mmu_gather **tlb,
205                         unsigned long addr, unsigned long end,
206                         unsigned long floor, unsigned long ceiling)
207 {
208         pgd_t *pgd;
209         unsigned long next;
210         unsigned long start;
211
212         /*
213          * The next few lines have given us lots of grief...
214          *
215          * Why are we testing PMD* at this top level?  Because often
216          * there will be no work to do at all, and we'd prefer not to
217          * go all the way down to the bottom just to discover that.
218          *
219          * Why all these "- 1"s?  Because 0 represents both the bottom
220          * of the address space and the top of it (using -1 for the
221          * top wouldn't help much: the masks would do the wrong thing).
222          * The rule is that addr 0 and floor 0 refer to the bottom of
223          * the address space, but end 0 and ceiling 0 refer to the top
224          * Comparisons need to use "end - 1" and "ceiling - 1" (though
225          * that end 0 case should be mythical).
226          *
227          * Wherever addr is brought up or ceiling brought down, we must
228          * be careful to reject "the opposite 0" before it confuses the
229          * subsequent tests.  But what about where end is brought down
230          * by PMD_SIZE below? no, end can't go down to 0 there.
231          *
232          * Whereas we round start (addr) and ceiling down, by different
233          * masks at different levels, in order to test whether a table
234          * now has no other vmas using it, so can be freed, we don't
235          * bother to round floor or end up - the tests don't need that.
236          */
237
238         addr &= PMD_MASK;
239         if (addr < floor) {
240                 addr += PMD_SIZE;
241                 if (!addr)
242                         return;
243         }
244         if (ceiling) {
245                 ceiling &= PMD_MASK;
246                 if (!ceiling)
247                         return;
248         }
249         if (end - 1 > ceiling - 1)
250                 end -= PMD_SIZE;
251         if (addr > end - 1)
252                 return;
253
254         start = addr;
255         pgd = pgd_offset((*tlb)->mm, addr);
256         do {
257                 next = pgd_addr_end(addr, end);
258                 if (pgd_none_or_clear_bad(pgd))
259                         continue;
260                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
261         } while (pgd++, addr = next, addr != end);
262
263         if (!(*tlb)->fullmm)
264                 flush_tlb_pgtables((*tlb)->mm, start, end);
265 }
266
267 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
268                 unsigned long floor, unsigned long ceiling)
269 {
270         while (vma) {
271                 struct vm_area_struct *next = vma->vm_next;
272                 unsigned long addr = vma->vm_start;
273
274                 /*
275                  * Hide vma from rmap and vmtruncate before freeing pgtables
276                  */
277                 anon_vma_unlink(vma);
278                 unlink_file_vma(vma);
279
280                 if (is_vm_hugetlb_page(vma)) {
281                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
282                                 floor, next? next->vm_start: ceiling);
283                 } else {
284                         /*
285                          * Optimization: gather nearby vmas into one call down
286                          */
287                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
288                                && !is_vm_hugetlb_page(next)) {
289                                 vma = next;
290                                 next = vma->vm_next;
291                                 anon_vma_unlink(vma);
292                                 unlink_file_vma(vma);
293                         }
294                         free_pgd_range(tlb, addr, vma->vm_end,
295                                 floor, next? next->vm_start: ceiling);
296                 }
297                 vma = next;
298         }
299 }
300
301 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
302 {
303         struct page *new = pte_alloc_one(mm, address);
304         if (!new)
305                 return -ENOMEM;
306
307         pte_lock_init(new);
308         spin_lock(&mm->page_table_lock);
309         if (pmd_present(*pmd)) {        /* Another has populated it */
310                 pte_lock_deinit(new);
311                 pte_free(new);
312         } else {
313                 mm->nr_ptes++;
314                 inc_page_state(nr_page_table_pages);
315                 pmd_populate(mm, pmd, new);
316         }
317         spin_unlock(&mm->page_table_lock);
318         return 0;
319 }
320
321 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
322 {
323         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
324         if (!new)
325                 return -ENOMEM;
326
327         spin_lock(&init_mm.page_table_lock);
328         if (pmd_present(*pmd))          /* Another has populated it */
329                 pte_free_kernel(new);
330         else
331                 pmd_populate_kernel(&init_mm, pmd, new);
332         spin_unlock(&init_mm.page_table_lock);
333         return 0;
334 }
335
336 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
337 {
338         if (file_rss)
339                 add_mm_counter(mm, file_rss, file_rss);
340         if (anon_rss)
341                 add_mm_counter(mm, anon_rss, anon_rss);
342 }
343
344 /*
345  * This function is called to print an error when a bad pte
346  * is found. For example, we might have a PFN-mapped pte in
347  * a region that doesn't allow it.
348  *
349  * The calling function must still handle the error.
350  */
351 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
352 {
353         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
354                         "vm_flags = %lx, vaddr = %lx\n",
355                 (long long)pte_val(pte),
356                 (vma->vm_mm == current->mm ? current->comm : "???"),
357                 vma->vm_flags, vaddr);
358         dump_stack();
359 }
360
361 static inline int is_cow_mapping(unsigned int flags)
362 {
363         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
364 }
365
366 /*
367  * This function gets the "struct page" associated with a pte.
368  *
369  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
370  * will have each page table entry just pointing to a raw page frame
371  * number, and as far as the VM layer is concerned, those do not have
372  * pages associated with them - even if the PFN might point to memory
373  * that otherwise is perfectly fine and has a "struct page".
374  *
375  * The way we recognize those mappings is through the rules set up
376  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
377  * and the vm_pgoff will point to the first PFN mapped: thus every
378  * page that is a raw mapping will always honor the rule
379  *
380  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
381  *
382  * and if that isn't true, the page has been COW'ed (in which case it
383  * _does_ have a "struct page" associated with it even if it is in a
384  * VM_PFNMAP range).
385  */
386 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
387 {
388         unsigned long pfn = pte_pfn(pte);
389
390         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
391                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
392                 if (pfn == vma->vm_pgoff + off)
393                         return NULL;
394                 if (!is_cow_mapping(vma->vm_flags))
395                         return NULL;
396         }
397
398         /*
399          * Add some anal sanity checks for now. Eventually,
400          * we should just do "return pfn_to_page(pfn)", but
401          * in the meantime we check that we get a valid pfn,
402          * and that the resulting page looks ok.
403          */
404         if (unlikely(!pfn_valid(pfn))) {
405                 print_bad_pte(vma, pte, addr);
406                 return NULL;
407         }
408
409         /*
410          * NOTE! We still have PageReserved() pages in the page 
411          * tables. 
412          *
413          * The PAGE_ZERO() pages and various VDSO mappings can
414          * cause them to exist.
415          */
416         return pfn_to_page(pfn);
417 }
418
419 /*
420  * copy one vm_area from one task to the other. Assumes the page tables
421  * already present in the new task to be cleared in the whole range
422  * covered by this vma.
423  */
424
425 static inline void
426 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
427                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
428                 unsigned long addr, int *rss)
429 {
430         unsigned long vm_flags = vma->vm_flags;
431         pte_t pte = *src_pte;
432         struct page *page;
433
434         /* pte contains position in swap or file, so copy. */
435         if (unlikely(!pte_present(pte))) {
436                 if (!pte_file(pte)) {
437                         swap_duplicate(pte_to_swp_entry(pte));
438                         /* make sure dst_mm is on swapoff's mmlist. */
439                         if (unlikely(list_empty(&dst_mm->mmlist))) {
440                                 spin_lock(&mmlist_lock);
441                                 if (list_empty(&dst_mm->mmlist))
442                                         list_add(&dst_mm->mmlist,
443                                                  &src_mm->mmlist);
444                                 spin_unlock(&mmlist_lock);
445                         }
446                 }
447                 goto out_set_pte;
448         }
449
450         /*
451          * If it's a COW mapping, write protect it both
452          * in the parent and the child
453          */
454         if (is_cow_mapping(vm_flags)) {
455                 ptep_set_wrprotect(src_mm, addr, src_pte);
456                 pte = *src_pte;
457         }
458
459         /*
460          * If it's a shared mapping, mark it clean in
461          * the child
462          */
463         if (vm_flags & VM_SHARED)
464                 pte = pte_mkclean(pte);
465         pte = pte_mkold(pte);
466
467         page = vm_normal_page(vma, addr, pte);
468         if (page) {
469                 get_page(page);
470                 page_dup_rmap(page);
471                 rss[!!PageAnon(page)]++;
472         }
473
474 out_set_pte:
475         set_pte_at(dst_mm, addr, dst_pte, pte);
476 }
477
478 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
479                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
480                 unsigned long addr, unsigned long end)
481 {
482         pte_t *src_pte, *dst_pte;
483         spinlock_t *src_ptl, *dst_ptl;
484         int progress = 0;
485         int rss[2];
486
487 again:
488         rss[1] = rss[0] = 0;
489         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
490         if (!dst_pte)
491                 return -ENOMEM;
492         src_pte = pte_offset_map_nested(src_pmd, addr);
493         src_ptl = pte_lockptr(src_mm, src_pmd);
494         spin_lock(src_ptl);
495
496         do {
497                 /*
498                  * We are holding two locks at this point - either of them
499                  * could generate latencies in another task on another CPU.
500                  */
501                 if (progress >= 32) {
502                         progress = 0;
503                         if (need_resched() ||
504                             need_lockbreak(src_ptl) ||
505                             need_lockbreak(dst_ptl))
506                                 break;
507                 }
508                 if (pte_none(*src_pte)) {
509                         progress++;
510                         continue;
511                 }
512                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
513                 progress += 8;
514         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
515
516         spin_unlock(src_ptl);
517         pte_unmap_nested(src_pte - 1);
518         add_mm_rss(dst_mm, rss[0], rss[1]);
519         pte_unmap_unlock(dst_pte - 1, dst_ptl);
520         cond_resched();
521         if (addr != end)
522                 goto again;
523         return 0;
524 }
525
526 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
527                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
528                 unsigned long addr, unsigned long end)
529 {
530         pmd_t *src_pmd, *dst_pmd;
531         unsigned long next;
532
533         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
534         if (!dst_pmd)
535                 return -ENOMEM;
536         src_pmd = pmd_offset(src_pud, addr);
537         do {
538                 next = pmd_addr_end(addr, end);
539                 if (pmd_none_or_clear_bad(src_pmd))
540                         continue;
541                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
542                                                 vma, addr, next))
543                         return -ENOMEM;
544         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
545         return 0;
546 }
547
548 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
549                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
550                 unsigned long addr, unsigned long end)
551 {
552         pud_t *src_pud, *dst_pud;
553         unsigned long next;
554
555         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
556         if (!dst_pud)
557                 return -ENOMEM;
558         src_pud = pud_offset(src_pgd, addr);
559         do {
560                 next = pud_addr_end(addr, end);
561                 if (pud_none_or_clear_bad(src_pud))
562                         continue;
563                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
564                                                 vma, addr, next))
565                         return -ENOMEM;
566         } while (dst_pud++, src_pud++, addr = next, addr != end);
567         return 0;
568 }
569
570 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
571                 struct vm_area_struct *vma)
572 {
573         pgd_t *src_pgd, *dst_pgd;
574         unsigned long next;
575         unsigned long addr = vma->vm_start;
576         unsigned long end = vma->vm_end;
577
578         /*
579          * Don't copy ptes where a page fault will fill them correctly.
580          * Fork becomes much lighter when there are big shared or private
581          * readonly mappings. The tradeoff is that copy_page_range is more
582          * efficient than faulting.
583          */
584         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
585                 if (!vma->anon_vma)
586                         return 0;
587         }
588
589         if (is_vm_hugetlb_page(vma))
590                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
591
592         dst_pgd = pgd_offset(dst_mm, addr);
593         src_pgd = pgd_offset(src_mm, addr);
594         do {
595                 next = pgd_addr_end(addr, end);
596                 if (pgd_none_or_clear_bad(src_pgd))
597                         continue;
598                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
599                                                 vma, addr, next))
600                         return -ENOMEM;
601         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
602         return 0;
603 }
604
605 static unsigned long zap_pte_range(struct mmu_gather *tlb,
606                                 struct vm_area_struct *vma, pmd_t *pmd,
607                                 unsigned long addr, unsigned long end,
608                                 long *zap_work, struct zap_details *details)
609 {
610         struct mm_struct *mm = tlb->mm;
611         pte_t *pte;
612         spinlock_t *ptl;
613         int file_rss = 0;
614         int anon_rss = 0;
615
616         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
617         do {
618                 pte_t ptent = *pte;
619                 if (pte_none(ptent)) {
620                         (*zap_work)--;
621                         continue;
622                 }
623
624                 (*zap_work) -= PAGE_SIZE;
625
626                 if (pte_present(ptent)) {
627                         struct page *page;
628
629                         page = vm_normal_page(vma, addr, ptent);
630                         if (unlikely(details) && page) {
631                                 /*
632                                  * unmap_shared_mapping_pages() wants to
633                                  * invalidate cache without truncating:
634                                  * unmap shared but keep private pages.
635                                  */
636                                 if (details->check_mapping &&
637                                     details->check_mapping != page->mapping)
638                                         continue;
639                                 /*
640                                  * Each page->index must be checked when
641                                  * invalidating or truncating nonlinear.
642                                  */
643                                 if (details->nonlinear_vma &&
644                                     (page->index < details->first_index ||
645                                      page->index > details->last_index))
646                                         continue;
647                         }
648                         ptent = ptep_get_and_clear_full(mm, addr, pte,
649                                                         tlb->fullmm);
650                         tlb_remove_tlb_entry(tlb, pte, addr);
651                         if (unlikely(!page))
652                                 continue;
653                         if (unlikely(details) && details->nonlinear_vma
654                             && linear_page_index(details->nonlinear_vma,
655                                                 addr) != page->index)
656                                 set_pte_at(mm, addr, pte,
657                                            pgoff_to_pte(page->index));
658                         if (PageAnon(page))
659                                 anon_rss--;
660                         else {
661                                 if (pte_dirty(ptent))
662                                         set_page_dirty(page);
663                                 if (pte_young(ptent))
664                                         mark_page_accessed(page);
665                                 file_rss--;
666                         }
667                         page_remove_rmap(page);
668                         tlb_remove_page(tlb, page);
669                         continue;
670                 }
671                 /*
672                  * If details->check_mapping, we leave swap entries;
673                  * if details->nonlinear_vma, we leave file entries.
674                  */
675                 if (unlikely(details))
676                         continue;
677                 if (!pte_file(ptent))
678                         free_swap_and_cache(pte_to_swp_entry(ptent));
679                 pte_clear_full(mm, addr, pte, tlb->fullmm);
680         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
681
682         add_mm_rss(mm, file_rss, anon_rss);
683         pte_unmap_unlock(pte - 1, ptl);
684
685         return addr;
686 }
687
688 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
689                                 struct vm_area_struct *vma, pud_t *pud,
690                                 unsigned long addr, unsigned long end,
691                                 long *zap_work, struct zap_details *details)
692 {
693         pmd_t *pmd;
694         unsigned long next;
695
696         pmd = pmd_offset(pud, addr);
697         do {
698                 next = pmd_addr_end(addr, end);
699                 if (pmd_none_or_clear_bad(pmd)) {
700                         (*zap_work)--;
701                         continue;
702                 }
703                 next = zap_pte_range(tlb, vma, pmd, addr, next,
704                                                 zap_work, details);
705         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
706
707         return addr;
708 }
709
710 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
711                                 struct vm_area_struct *vma, pgd_t *pgd,
712                                 unsigned long addr, unsigned long end,
713                                 long *zap_work, struct zap_details *details)
714 {
715         pud_t *pud;
716         unsigned long next;
717
718         pud = pud_offset(pgd, addr);
719         do {
720                 next = pud_addr_end(addr, end);
721                 if (pud_none_or_clear_bad(pud)) {
722                         (*zap_work)--;
723                         continue;
724                 }
725                 next = zap_pmd_range(tlb, vma, pud, addr, next,
726                                                 zap_work, details);
727         } while (pud++, addr = next, (addr != end && *zap_work > 0));
728
729         return addr;
730 }
731
732 static unsigned long unmap_page_range(struct mmu_gather *tlb,
733                                 struct vm_area_struct *vma,
734                                 unsigned long addr, unsigned long end,
735                                 long *zap_work, struct zap_details *details)
736 {
737         pgd_t *pgd;
738         unsigned long next;
739
740         if (details && !details->check_mapping && !details->nonlinear_vma)
741                 details = NULL;
742
743         BUG_ON(addr >= end);
744         tlb_start_vma(tlb, vma);
745         pgd = pgd_offset(vma->vm_mm, addr);
746         do {
747                 next = pgd_addr_end(addr, end);
748                 if (pgd_none_or_clear_bad(pgd)) {
749                         (*zap_work)--;
750                         continue;
751                 }
752                 next = zap_pud_range(tlb, vma, pgd, addr, next,
753                                                 zap_work, details);
754         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
755         tlb_end_vma(tlb, vma);
756
757         return addr;
758 }
759
760 #ifdef CONFIG_PREEMPT
761 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
762 #else
763 /* No preempt: go for improved straight-line efficiency */
764 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
765 #endif
766
767 /**
768  * unmap_vmas - unmap a range of memory covered by a list of vma's
769  * @tlbp: address of the caller's struct mmu_gather
770  * @vma: the starting vma
771  * @start_addr: virtual address at which to start unmapping
772  * @end_addr: virtual address at which to end unmapping
773  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
774  * @details: details of nonlinear truncation or shared cache invalidation
775  *
776  * Returns the end address of the unmapping (restart addr if interrupted).
777  *
778  * Unmap all pages in the vma list.
779  *
780  * We aim to not hold locks for too long (for scheduling latency reasons).
781  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
782  * return the ending mmu_gather to the caller.
783  *
784  * Only addresses between `start' and `end' will be unmapped.
785  *
786  * The VMA list must be sorted in ascending virtual address order.
787  *
788  * unmap_vmas() assumes that the caller will flush the whole unmapped address
789  * range after unmap_vmas() returns.  So the only responsibility here is to
790  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
791  * drops the lock and schedules.
792  */
793 unsigned long unmap_vmas(struct mmu_gather **tlbp,
794                 struct vm_area_struct *vma, unsigned long start_addr,
795                 unsigned long end_addr, unsigned long *nr_accounted,
796                 struct zap_details *details)
797 {
798         long zap_work = ZAP_BLOCK_SIZE;
799         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
800         int tlb_start_valid = 0;
801         unsigned long start = start_addr;
802         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
803         int fullmm = (*tlbp)->fullmm;
804
805         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
806                 unsigned long end;
807
808                 start = max(vma->vm_start, start_addr);
809                 if (start >= vma->vm_end)
810                         continue;
811                 end = min(vma->vm_end, end_addr);
812                 if (end <= vma->vm_start)
813                         continue;
814
815                 if (vma->vm_flags & VM_ACCOUNT)
816                         *nr_accounted += (end - start) >> PAGE_SHIFT;
817
818                 while (start != end) {
819                         if (!tlb_start_valid) {
820                                 tlb_start = start;
821                                 tlb_start_valid = 1;
822                         }
823
824                         if (unlikely(is_vm_hugetlb_page(vma))) {
825                                 unmap_hugepage_range(vma, start, end);
826                                 zap_work -= (end - start) /
827                                                 (HPAGE_SIZE / PAGE_SIZE);
828                                 start = end;
829                         } else
830                                 start = unmap_page_range(*tlbp, vma,
831                                                 start, end, &zap_work, details);
832
833                         if (zap_work > 0) {
834                                 BUG_ON(start != end);
835                                 break;
836                         }
837
838                         tlb_finish_mmu(*tlbp, tlb_start, start);
839
840                         if (need_resched() ||
841                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
842                                 if (i_mmap_lock) {
843                                         *tlbp = NULL;
844                                         goto out;
845                                 }
846                                 cond_resched();
847                         }
848
849                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
850                         tlb_start_valid = 0;
851                         zap_work = ZAP_BLOCK_SIZE;
852                 }
853         }
854 out:
855         return start;   /* which is now the end (or restart) address */
856 }
857
858 /**
859  * zap_page_range - remove user pages in a given range
860  * @vma: vm_area_struct holding the applicable pages
861  * @address: starting address of pages to zap
862  * @size: number of bytes to zap
863  * @details: details of nonlinear truncation or shared cache invalidation
864  */
865 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
866                 unsigned long size, struct zap_details *details)
867 {
868         struct mm_struct *mm = vma->vm_mm;
869         struct mmu_gather *tlb;
870         unsigned long end = address + size;
871         unsigned long nr_accounted = 0;
872
873         lru_add_drain();
874         tlb = tlb_gather_mmu(mm, 0);
875         update_hiwater_rss(mm);
876         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
877         if (tlb)
878                 tlb_finish_mmu(tlb, address, end);
879         return end;
880 }
881
882 /*
883  * Do a quick page-table lookup for a single page.
884  */
885 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
886                         unsigned int flags)
887 {
888         pgd_t *pgd;
889         pud_t *pud;
890         pmd_t *pmd;
891         pte_t *ptep, pte;
892         spinlock_t *ptl;
893         struct page *page;
894         struct mm_struct *mm = vma->vm_mm;
895
896         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
897         if (!IS_ERR(page)) {
898                 BUG_ON(flags & FOLL_GET);
899                 goto out;
900         }
901
902         page = NULL;
903         pgd = pgd_offset(mm, address);
904         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
905                 goto no_page_table;
906
907         pud = pud_offset(pgd, address);
908         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
909                 goto no_page_table;
910         
911         pmd = pmd_offset(pud, address);
912         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
913                 goto no_page_table;
914
915         if (pmd_huge(*pmd)) {
916                 BUG_ON(flags & FOLL_GET);
917                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
918                 goto out;
919         }
920
921         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
922         if (!ptep)
923                 goto out;
924
925         pte = *ptep;
926         if (!pte_present(pte))
927                 goto unlock;
928         if ((flags & FOLL_WRITE) && !pte_write(pte))
929                 goto unlock;
930         page = vm_normal_page(vma, address, pte);
931         if (unlikely(!page))
932                 goto unlock;
933
934         if (flags & FOLL_GET)
935                 get_page(page);
936         if (flags & FOLL_TOUCH) {
937                 if ((flags & FOLL_WRITE) &&
938                     !pte_dirty(pte) && !PageDirty(page))
939                         set_page_dirty(page);
940                 mark_page_accessed(page);
941         }
942 unlock:
943         pte_unmap_unlock(ptep, ptl);
944 out:
945         return page;
946
947 no_page_table:
948         /*
949          * When core dumping an enormous anonymous area that nobody
950          * has touched so far, we don't want to allocate page tables.
951          */
952         if (flags & FOLL_ANON) {
953                 page = ZERO_PAGE(address);
954                 if (flags & FOLL_GET)
955                         get_page(page);
956                 BUG_ON(flags & FOLL_WRITE);
957         }
958         return page;
959 }
960
961 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
962                 unsigned long start, int len, int write, int force,
963                 struct page **pages, struct vm_area_struct **vmas)
964 {
965         int i;
966         unsigned int vm_flags;
967
968         /* 
969          * Require read or write permissions.
970          * If 'force' is set, we only require the "MAY" flags.
971          */
972         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
973         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
974         i = 0;
975
976         do {
977                 struct vm_area_struct *vma;
978                 unsigned int foll_flags;
979
980                 vma = find_extend_vma(mm, start);
981                 if (!vma && in_gate_area(tsk, start)) {
982                         unsigned long pg = start & PAGE_MASK;
983                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
984                         pgd_t *pgd;
985                         pud_t *pud;
986                         pmd_t *pmd;
987                         pte_t *pte;
988                         if (write) /* user gate pages are read-only */
989                                 return i ? : -EFAULT;
990                         if (pg > TASK_SIZE)
991                                 pgd = pgd_offset_k(pg);
992                         else
993                                 pgd = pgd_offset_gate(mm, pg);
994                         BUG_ON(pgd_none(*pgd));
995                         pud = pud_offset(pgd, pg);
996                         BUG_ON(pud_none(*pud));
997                         pmd = pmd_offset(pud, pg);
998                         if (pmd_none(*pmd))
999                                 return i ? : -EFAULT;
1000                         pte = pte_offset_map(pmd, pg);
1001                         if (pte_none(*pte)) {
1002                                 pte_unmap(pte);
1003                                 return i ? : -EFAULT;
1004                         }
1005                         if (pages) {
1006                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1007                                 pages[i] = page;
1008                                 if (page)
1009                                         get_page(page);
1010                         }
1011                         pte_unmap(pte);
1012                         if (vmas)
1013                                 vmas[i] = gate_vma;
1014                         i++;
1015                         start += PAGE_SIZE;
1016                         len--;
1017                         continue;
1018                 }
1019
1020                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1021                                 || !(vm_flags & vma->vm_flags))
1022                         return i ? : -EFAULT;
1023
1024                 if (is_vm_hugetlb_page(vma)) {
1025                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1026                                                 &start, &len, i);
1027                         continue;
1028                 }
1029
1030                 foll_flags = FOLL_TOUCH;
1031                 if (pages)
1032                         foll_flags |= FOLL_GET;
1033                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1034                     (!vma->vm_ops || !vma->vm_ops->nopage))
1035                         foll_flags |= FOLL_ANON;
1036
1037                 do {
1038                         struct page *page;
1039
1040                         if (write)
1041                                 foll_flags |= FOLL_WRITE;
1042
1043                         cond_resched();
1044                         while (!(page = follow_page(vma, start, foll_flags))) {
1045                                 int ret;
1046                                 ret = __handle_mm_fault(mm, vma, start,
1047                                                 foll_flags & FOLL_WRITE);
1048                                 /*
1049                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1050                                  * broken COW when necessary, even if maybe_mkwrite
1051                                  * decided not to set pte_write. We can thus safely do
1052                                  * subsequent page lookups as if they were reads.
1053                                  */
1054                                 if (ret & VM_FAULT_WRITE)
1055                                         foll_flags &= ~FOLL_WRITE;
1056                                 
1057                                 switch (ret & ~VM_FAULT_WRITE) {
1058                                 case VM_FAULT_MINOR:
1059                                         tsk->min_flt++;
1060                                         break;
1061                                 case VM_FAULT_MAJOR:
1062                                         tsk->maj_flt++;
1063                                         break;
1064                                 case VM_FAULT_SIGBUS:
1065                                         return i ? i : -EFAULT;
1066                                 case VM_FAULT_OOM:
1067                                         return i ? i : -ENOMEM;
1068                                 default:
1069                                         BUG();
1070                                 }
1071                         }
1072                         if (pages) {
1073                                 pages[i] = page;
1074
1075                                 flush_anon_page(page, start);
1076                                 flush_dcache_page(page);
1077                         }
1078                         if (vmas)
1079                                 vmas[i] = vma;
1080                         i++;
1081                         start += PAGE_SIZE;
1082                         len--;
1083                 } while (len && start < vma->vm_end);
1084         } while (len);
1085         return i;
1086 }
1087 EXPORT_SYMBOL(get_user_pages);
1088
1089 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1090                         unsigned long addr, unsigned long end, pgprot_t prot)
1091 {
1092         pte_t *pte;
1093         spinlock_t *ptl;
1094
1095         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1096         if (!pte)
1097                 return -ENOMEM;
1098         do {
1099                 struct page *page = ZERO_PAGE(addr);
1100                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1101                 page_cache_get(page);
1102                 page_add_file_rmap(page);
1103                 inc_mm_counter(mm, file_rss);
1104                 BUG_ON(!pte_none(*pte));
1105                 set_pte_at(mm, addr, pte, zero_pte);
1106         } while (pte++, addr += PAGE_SIZE, addr != end);
1107         pte_unmap_unlock(pte - 1, ptl);
1108         return 0;
1109 }
1110
1111 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1112                         unsigned long addr, unsigned long end, pgprot_t prot)
1113 {
1114         pmd_t *pmd;
1115         unsigned long next;
1116
1117         pmd = pmd_alloc(mm, pud, addr);
1118         if (!pmd)
1119                 return -ENOMEM;
1120         do {
1121                 next = pmd_addr_end(addr, end);
1122                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1123                         return -ENOMEM;
1124         } while (pmd++, addr = next, addr != end);
1125         return 0;
1126 }
1127
1128 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1129                         unsigned long addr, unsigned long end, pgprot_t prot)
1130 {
1131         pud_t *pud;
1132         unsigned long next;
1133
1134         pud = pud_alloc(mm, pgd, addr);
1135         if (!pud)
1136                 return -ENOMEM;
1137         do {
1138                 next = pud_addr_end(addr, end);
1139                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1140                         return -ENOMEM;
1141         } while (pud++, addr = next, addr != end);
1142         return 0;
1143 }
1144
1145 int zeromap_page_range(struct vm_area_struct *vma,
1146                         unsigned long addr, unsigned long size, pgprot_t prot)
1147 {
1148         pgd_t *pgd;
1149         unsigned long next;
1150         unsigned long end = addr + size;
1151         struct mm_struct *mm = vma->vm_mm;
1152         int err;
1153
1154         BUG_ON(addr >= end);
1155         pgd = pgd_offset(mm, addr);
1156         flush_cache_range(vma, addr, end);
1157         do {
1158                 next = pgd_addr_end(addr, end);
1159                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1160                 if (err)
1161                         break;
1162         } while (pgd++, addr = next, addr != end);
1163         return err;
1164 }
1165
1166 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1167 {
1168         pgd_t * pgd = pgd_offset(mm, addr);
1169         pud_t * pud = pud_alloc(mm, pgd, addr);
1170         if (pud) {
1171                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1172                 if (pmd)
1173                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1174         }
1175         return NULL;
1176 }
1177
1178 /*
1179  * This is the old fallback for page remapping.
1180  *
1181  * For historical reasons, it only allows reserved pages. Only
1182  * old drivers should use this, and they needed to mark their
1183  * pages reserved for the old functions anyway.
1184  */
1185 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1186 {
1187         int retval;
1188         pte_t *pte;
1189         spinlock_t *ptl;  
1190
1191         retval = -EINVAL;
1192         if (PageAnon(page))
1193                 goto out;
1194         retval = -ENOMEM;
1195         flush_dcache_page(page);
1196         pte = get_locked_pte(mm, addr, &ptl);
1197         if (!pte)
1198                 goto out;
1199         retval = -EBUSY;
1200         if (!pte_none(*pte))
1201                 goto out_unlock;
1202
1203         /* Ok, finally just insert the thing.. */
1204         get_page(page);
1205         inc_mm_counter(mm, file_rss);
1206         page_add_file_rmap(page);
1207         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1208
1209         retval = 0;
1210 out_unlock:
1211         pte_unmap_unlock(pte, ptl);
1212 out:
1213         return retval;
1214 }
1215
1216 /*
1217  * This allows drivers to insert individual pages they've allocated
1218  * into a user vma.
1219  *
1220  * The page has to be a nice clean _individual_ kernel allocation.
1221  * If you allocate a compound page, you need to have marked it as
1222  * such (__GFP_COMP), or manually just split the page up yourself
1223  * (see split_page()).
1224  *
1225  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1226  * took an arbitrary page protection parameter. This doesn't allow
1227  * that. Your vma protection will have to be set up correctly, which
1228  * means that if you want a shared writable mapping, you'd better
1229  * ask for a shared writable mapping!
1230  *
1231  * The page does not need to be reserved.
1232  */
1233 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1234 {
1235         if (addr < vma->vm_start || addr >= vma->vm_end)
1236                 return -EFAULT;
1237         if (!page_count(page))
1238                 return -EINVAL;
1239         vma->vm_flags |= VM_INSERTPAGE;
1240         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1241 }
1242 EXPORT_SYMBOL(vm_insert_page);
1243
1244 /*
1245  * maps a range of physical memory into the requested pages. the old
1246  * mappings are removed. any references to nonexistent pages results
1247  * in null mappings (currently treated as "copy-on-access")
1248  */
1249 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1250                         unsigned long addr, unsigned long end,
1251                         unsigned long pfn, pgprot_t prot)
1252 {
1253         pte_t *pte;
1254         spinlock_t *ptl;
1255
1256         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1257         if (!pte)
1258                 return -ENOMEM;
1259         do {
1260                 BUG_ON(!pte_none(*pte));
1261                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1262                 pfn++;
1263         } while (pte++, addr += PAGE_SIZE, addr != end);
1264         pte_unmap_unlock(pte - 1, ptl);
1265         return 0;
1266 }
1267
1268 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1269                         unsigned long addr, unsigned long end,
1270                         unsigned long pfn, pgprot_t prot)
1271 {
1272         pmd_t *pmd;
1273         unsigned long next;
1274
1275         pfn -= addr >> PAGE_SHIFT;
1276         pmd = pmd_alloc(mm, pud, addr);
1277         if (!pmd)
1278                 return -ENOMEM;
1279         do {
1280                 next = pmd_addr_end(addr, end);
1281                 if (remap_pte_range(mm, pmd, addr, next,
1282                                 pfn + (addr >> PAGE_SHIFT), prot))
1283                         return -ENOMEM;
1284         } while (pmd++, addr = next, addr != end);
1285         return 0;
1286 }
1287
1288 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1289                         unsigned long addr, unsigned long end,
1290                         unsigned long pfn, pgprot_t prot)
1291 {
1292         pud_t *pud;
1293         unsigned long next;
1294
1295         pfn -= addr >> PAGE_SHIFT;
1296         pud = pud_alloc(mm, pgd, addr);
1297         if (!pud)
1298                 return -ENOMEM;
1299         do {
1300                 next = pud_addr_end(addr, end);
1301                 if (remap_pmd_range(mm, pud, addr, next,
1302                                 pfn + (addr >> PAGE_SHIFT), prot))
1303                         return -ENOMEM;
1304         } while (pud++, addr = next, addr != end);
1305         return 0;
1306 }
1307
1308 /*  Note: this is only safe if the mm semaphore is held when called. */
1309 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1310                     unsigned long pfn, unsigned long size, pgprot_t prot)
1311 {
1312         pgd_t *pgd;
1313         unsigned long next;
1314         unsigned long end = addr + PAGE_ALIGN(size);
1315         struct mm_struct *mm = vma->vm_mm;
1316         int err;
1317
1318         /*
1319          * Physically remapped pages are special. Tell the
1320          * rest of the world about it:
1321          *   VM_IO tells people not to look at these pages
1322          *      (accesses can have side effects).
1323          *   VM_RESERVED is specified all over the place, because
1324          *      in 2.4 it kept swapout's vma scan off this vma; but
1325          *      in 2.6 the LRU scan won't even find its pages, so this
1326          *      flag means no more than count its pages in reserved_vm,
1327          *      and omit it from core dump, even when VM_IO turned off.
1328          *   VM_PFNMAP tells the core MM that the base pages are just
1329          *      raw PFN mappings, and do not have a "struct page" associated
1330          *      with them.
1331          *
1332          * There's a horrible special case to handle copy-on-write
1333          * behaviour that some programs depend on. We mark the "original"
1334          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1335          */
1336         if (is_cow_mapping(vma->vm_flags)) {
1337                 if (addr != vma->vm_start || end != vma->vm_end)
1338                         return -EINVAL;
1339                 vma->vm_pgoff = pfn;
1340         }
1341
1342         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1343
1344         BUG_ON(addr >= end);
1345         pfn -= addr >> PAGE_SHIFT;
1346         pgd = pgd_offset(mm, addr);
1347         flush_cache_range(vma, addr, end);
1348         do {
1349                 next = pgd_addr_end(addr, end);
1350                 err = remap_pud_range(mm, pgd, addr, next,
1351                                 pfn + (addr >> PAGE_SHIFT), prot);
1352                 if (err)
1353                         break;
1354         } while (pgd++, addr = next, addr != end);
1355         return err;
1356 }
1357 EXPORT_SYMBOL(remap_pfn_range);
1358
1359 /*
1360  * handle_pte_fault chooses page fault handler according to an entry
1361  * which was read non-atomically.  Before making any commitment, on
1362  * those architectures or configurations (e.g. i386 with PAE) which
1363  * might give a mix of unmatched parts, do_swap_page and do_file_page
1364  * must check under lock before unmapping the pte and proceeding
1365  * (but do_wp_page is only called after already making such a check;
1366  * and do_anonymous_page and do_no_page can safely check later on).
1367  */
1368 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1369                                 pte_t *page_table, pte_t orig_pte)
1370 {
1371         int same = 1;
1372 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1373         if (sizeof(pte_t) > sizeof(unsigned long)) {
1374                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1375                 spin_lock(ptl);
1376                 same = pte_same(*page_table, orig_pte);
1377                 spin_unlock(ptl);
1378         }
1379 #endif
1380         pte_unmap(page_table);
1381         return same;
1382 }
1383
1384 /*
1385  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1386  * servicing faults for write access.  In the normal case, do always want
1387  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1388  * that do not have writing enabled, when used by access_process_vm.
1389  */
1390 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1391 {
1392         if (likely(vma->vm_flags & VM_WRITE))
1393                 pte = pte_mkwrite(pte);
1394         return pte;
1395 }
1396
1397 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1398 {
1399         /*
1400          * If the source page was a PFN mapping, we don't have
1401          * a "struct page" for it. We do a best-effort copy by
1402          * just copying from the original user address. If that
1403          * fails, we just zero-fill it. Live with it.
1404          */
1405         if (unlikely(!src)) {
1406                 void *kaddr = kmap_atomic(dst, KM_USER0);
1407                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1408
1409                 /*
1410                  * This really shouldn't fail, because the page is there
1411                  * in the page tables. But it might just be unreadable,
1412                  * in which case we just give up and fill the result with
1413                  * zeroes.
1414                  */
1415                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1416                         memset(kaddr, 0, PAGE_SIZE);
1417                 kunmap_atomic(kaddr, KM_USER0);
1418                 return;
1419                 
1420         }
1421         copy_user_highpage(dst, src, va);
1422 }
1423
1424 /*
1425  * This routine handles present pages, when users try to write
1426  * to a shared page. It is done by copying the page to a new address
1427  * and decrementing the shared-page counter for the old page.
1428  *
1429  * Note that this routine assumes that the protection checks have been
1430  * done by the caller (the low-level page fault routine in most cases).
1431  * Thus we can safely just mark it writable once we've done any necessary
1432  * COW.
1433  *
1434  * We also mark the page dirty at this point even though the page will
1435  * change only once the write actually happens. This avoids a few races,
1436  * and potentially makes it more efficient.
1437  *
1438  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1439  * but allow concurrent faults), with pte both mapped and locked.
1440  * We return with mmap_sem still held, but pte unmapped and unlocked.
1441  */
1442 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1443                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1444                 spinlock_t *ptl, pte_t orig_pte)
1445 {
1446         struct page *old_page, *new_page;
1447         pte_t entry;
1448         int ret = VM_FAULT_MINOR;
1449
1450         old_page = vm_normal_page(vma, address, orig_pte);
1451         if (!old_page)
1452                 goto gotten;
1453
1454         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1455                 int reuse = can_share_swap_page(old_page);
1456                 unlock_page(old_page);
1457                 if (reuse) {
1458                         flush_cache_page(vma, address, pte_pfn(orig_pte));
1459                         entry = pte_mkyoung(orig_pte);
1460                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1461                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1462                         update_mmu_cache(vma, address, entry);
1463                         lazy_mmu_prot_update(entry);
1464                         ret |= VM_FAULT_WRITE;
1465                         goto unlock;
1466                 }
1467         }
1468
1469         /*
1470          * Ok, we need to copy. Oh, well..
1471          */
1472         page_cache_get(old_page);
1473 gotten:
1474         pte_unmap_unlock(page_table, ptl);
1475
1476         if (unlikely(anon_vma_prepare(vma)))
1477                 goto oom;
1478         if (old_page == ZERO_PAGE(address)) {
1479                 new_page = alloc_zeroed_user_highpage(vma, address);
1480                 if (!new_page)
1481                         goto oom;
1482         } else {
1483                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1484                 if (!new_page)
1485                         goto oom;
1486                 cow_user_page(new_page, old_page, address);
1487         }
1488
1489         /*
1490          * Re-check the pte - we dropped the lock
1491          */
1492         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1493         if (likely(pte_same(*page_table, orig_pte))) {
1494                 if (old_page) {
1495                         page_remove_rmap(old_page);
1496                         if (!PageAnon(old_page)) {
1497                                 dec_mm_counter(mm, file_rss);
1498                                 inc_mm_counter(mm, anon_rss);
1499                         }
1500                 } else
1501                         inc_mm_counter(mm, anon_rss);
1502                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1503                 entry = mk_pte(new_page, vma->vm_page_prot);
1504                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1505                 ptep_establish(vma, address, page_table, entry);
1506                 update_mmu_cache(vma, address, entry);
1507                 lazy_mmu_prot_update(entry);
1508                 lru_cache_add_active(new_page);
1509                 page_add_new_anon_rmap(new_page, vma, address);
1510
1511                 /* Free the old page.. */
1512                 new_page = old_page;
1513                 ret |= VM_FAULT_WRITE;
1514         }
1515         if (new_page)
1516                 page_cache_release(new_page);
1517         if (old_page)
1518                 page_cache_release(old_page);
1519 unlock:
1520         pte_unmap_unlock(page_table, ptl);
1521         return ret;
1522 oom:
1523         if (old_page)
1524                 page_cache_release(old_page);
1525         return VM_FAULT_OOM;
1526 }
1527
1528 /*
1529  * Helper functions for unmap_mapping_range().
1530  *
1531  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1532  *
1533  * We have to restart searching the prio_tree whenever we drop the lock,
1534  * since the iterator is only valid while the lock is held, and anyway
1535  * a later vma might be split and reinserted earlier while lock dropped.
1536  *
1537  * The list of nonlinear vmas could be handled more efficiently, using
1538  * a placeholder, but handle it in the same way until a need is shown.
1539  * It is important to search the prio_tree before nonlinear list: a vma
1540  * may become nonlinear and be shifted from prio_tree to nonlinear list
1541  * while the lock is dropped; but never shifted from list to prio_tree.
1542  *
1543  * In order to make forward progress despite restarting the search,
1544  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1545  * quickly skip it next time around.  Since the prio_tree search only
1546  * shows us those vmas affected by unmapping the range in question, we
1547  * can't efficiently keep all vmas in step with mapping->truncate_count:
1548  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1549  * mapping->truncate_count and vma->vm_truncate_count are protected by
1550  * i_mmap_lock.
1551  *
1552  * In order to make forward progress despite repeatedly restarting some
1553  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1554  * and restart from that address when we reach that vma again.  It might
1555  * have been split or merged, shrunk or extended, but never shifted: so
1556  * restart_addr remains valid so long as it remains in the vma's range.
1557  * unmap_mapping_range forces truncate_count to leap over page-aligned
1558  * values so we can save vma's restart_addr in its truncate_count field.
1559  */
1560 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1561
1562 static void reset_vma_truncate_counts(struct address_space *mapping)
1563 {
1564         struct vm_area_struct *vma;
1565         struct prio_tree_iter iter;
1566
1567         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1568                 vma->vm_truncate_count = 0;
1569         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1570                 vma->vm_truncate_count = 0;
1571 }
1572
1573 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1574                 unsigned long start_addr, unsigned long end_addr,
1575                 struct zap_details *details)
1576 {
1577         unsigned long restart_addr;
1578         int need_break;
1579
1580 again:
1581         restart_addr = vma->vm_truncate_count;
1582         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1583                 start_addr = restart_addr;
1584                 if (start_addr >= end_addr) {
1585                         /* Top of vma has been split off since last time */
1586                         vma->vm_truncate_count = details->truncate_count;
1587                         return 0;
1588                 }
1589         }
1590
1591         restart_addr = zap_page_range(vma, start_addr,
1592                                         end_addr - start_addr, details);
1593         need_break = need_resched() ||
1594                         need_lockbreak(details->i_mmap_lock);
1595
1596         if (restart_addr >= end_addr) {
1597                 /* We have now completed this vma: mark it so */
1598                 vma->vm_truncate_count = details->truncate_count;
1599                 if (!need_break)
1600                         return 0;
1601         } else {
1602                 /* Note restart_addr in vma's truncate_count field */
1603                 vma->vm_truncate_count = restart_addr;
1604                 if (!need_break)
1605                         goto again;
1606         }
1607
1608         spin_unlock(details->i_mmap_lock);
1609         cond_resched();
1610         spin_lock(details->i_mmap_lock);
1611         return -EINTR;
1612 }
1613
1614 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1615                                             struct zap_details *details)
1616 {
1617         struct vm_area_struct *vma;
1618         struct prio_tree_iter iter;
1619         pgoff_t vba, vea, zba, zea;
1620
1621 restart:
1622         vma_prio_tree_foreach(vma, &iter, root,
1623                         details->first_index, details->last_index) {
1624                 /* Skip quickly over those we have already dealt with */
1625                 if (vma->vm_truncate_count == details->truncate_count)
1626                         continue;
1627
1628                 vba = vma->vm_pgoff;
1629                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1630                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1631                 zba = details->first_index;
1632                 if (zba < vba)
1633                         zba = vba;
1634                 zea = details->last_index;
1635                 if (zea > vea)
1636                         zea = vea;
1637
1638                 if (unmap_mapping_range_vma(vma,
1639                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1640                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1641                                 details) < 0)
1642                         goto restart;
1643         }
1644 }
1645
1646 static inline void unmap_mapping_range_list(struct list_head *head,
1647                                             struct zap_details *details)
1648 {
1649         struct vm_area_struct *vma;
1650
1651         /*
1652          * In nonlinear VMAs there is no correspondence between virtual address
1653          * offset and file offset.  So we must perform an exhaustive search
1654          * across *all* the pages in each nonlinear VMA, not just the pages
1655          * whose virtual address lies outside the file truncation point.
1656          */
1657 restart:
1658         list_for_each_entry(vma, head, shared.vm_set.list) {
1659                 /* Skip quickly over those we have already dealt with */
1660                 if (vma->vm_truncate_count == details->truncate_count)
1661                         continue;
1662                 details->nonlinear_vma = vma;
1663                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1664                                         vma->vm_end, details) < 0)
1665                         goto restart;
1666         }
1667 }
1668
1669 /**
1670  * unmap_mapping_range - unmap the portion of all mmaps
1671  * in the specified address_space corresponding to the specified
1672  * page range in the underlying file.
1673  * @mapping: the address space containing mmaps to be unmapped.
1674  * @holebegin: byte in first page to unmap, relative to the start of
1675  * the underlying file.  This will be rounded down to a PAGE_SIZE
1676  * boundary.  Note that this is different from vmtruncate(), which
1677  * must keep the partial page.  In contrast, we must get rid of
1678  * partial pages.
1679  * @holelen: size of prospective hole in bytes.  This will be rounded
1680  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1681  * end of the file.
1682  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1683  * but 0 when invalidating pagecache, don't throw away private data.
1684  */
1685 void unmap_mapping_range(struct address_space *mapping,
1686                 loff_t const holebegin, loff_t const holelen, int even_cows)
1687 {
1688         struct zap_details details;
1689         pgoff_t hba = holebegin >> PAGE_SHIFT;
1690         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1691
1692         /* Check for overflow. */
1693         if (sizeof(holelen) > sizeof(hlen)) {
1694                 long long holeend =
1695                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1696                 if (holeend & ~(long long)ULONG_MAX)
1697                         hlen = ULONG_MAX - hba + 1;
1698         }
1699
1700         details.check_mapping = even_cows? NULL: mapping;
1701         details.nonlinear_vma = NULL;
1702         details.first_index = hba;
1703         details.last_index = hba + hlen - 1;
1704         if (details.last_index < details.first_index)
1705                 details.last_index = ULONG_MAX;
1706         details.i_mmap_lock = &mapping->i_mmap_lock;
1707
1708         spin_lock(&mapping->i_mmap_lock);
1709
1710         /* serialize i_size write against truncate_count write */
1711         smp_wmb();
1712         /* Protect against page faults, and endless unmapping loops */
1713         mapping->truncate_count++;
1714         /*
1715          * For archs where spin_lock has inclusive semantics like ia64
1716          * this smp_mb() will prevent to read pagetable contents
1717          * before the truncate_count increment is visible to
1718          * other cpus.
1719          */
1720         smp_mb();
1721         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1722                 if (mapping->truncate_count == 0)
1723                         reset_vma_truncate_counts(mapping);
1724                 mapping->truncate_count++;
1725         }
1726         details.truncate_count = mapping->truncate_count;
1727
1728         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1729                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1730         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1731                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1732         spin_unlock(&mapping->i_mmap_lock);
1733 }
1734 EXPORT_SYMBOL(unmap_mapping_range);
1735
1736 /*
1737  * Handle all mappings that got truncated by a "truncate()"
1738  * system call.
1739  *
1740  * NOTE! We have to be ready to update the memory sharing
1741  * between the file and the memory map for a potential last
1742  * incomplete page.  Ugly, but necessary.
1743  */
1744 int vmtruncate(struct inode * inode, loff_t offset)
1745 {
1746         struct address_space *mapping = inode->i_mapping;
1747         unsigned long limit;
1748
1749         if (inode->i_size < offset)
1750                 goto do_expand;
1751         /*
1752          * truncation of in-use swapfiles is disallowed - it would cause
1753          * subsequent swapout to scribble on the now-freed blocks.
1754          */
1755         if (IS_SWAPFILE(inode))
1756                 goto out_busy;
1757         i_size_write(inode, offset);
1758         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1759         truncate_inode_pages(mapping, offset);
1760         goto out_truncate;
1761
1762 do_expand:
1763         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1764         if (limit != RLIM_INFINITY && offset > limit)
1765                 goto out_sig;
1766         if (offset > inode->i_sb->s_maxbytes)
1767                 goto out_big;
1768         i_size_write(inode, offset);
1769
1770 out_truncate:
1771         if (inode->i_op && inode->i_op->truncate)
1772                 inode->i_op->truncate(inode);
1773         return 0;
1774 out_sig:
1775         send_sig(SIGXFSZ, current, 0);
1776 out_big:
1777         return -EFBIG;
1778 out_busy:
1779         return -ETXTBSY;
1780 }
1781 EXPORT_SYMBOL(vmtruncate);
1782
1783 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1784 {
1785         struct address_space *mapping = inode->i_mapping;
1786
1787         /*
1788          * If the underlying filesystem is not going to provide
1789          * a way to truncate a range of blocks (punch a hole) -
1790          * we should return failure right now.
1791          */
1792         if (!inode->i_op || !inode->i_op->truncate_range)
1793                 return -ENOSYS;
1794
1795         mutex_lock(&inode->i_mutex);
1796         down_write(&inode->i_alloc_sem);
1797         unmap_mapping_range(mapping, offset, (end - offset), 1);
1798         truncate_inode_pages_range(mapping, offset, end);
1799         inode->i_op->truncate_range(inode, offset, end);
1800         up_write(&inode->i_alloc_sem);
1801         mutex_unlock(&inode->i_mutex);
1802
1803         return 0;
1804 }
1805 EXPORT_SYMBOL(vmtruncate_range);
1806
1807 /* 
1808  * Primitive swap readahead code. We simply read an aligned block of
1809  * (1 << page_cluster) entries in the swap area. This method is chosen
1810  * because it doesn't cost us any seek time.  We also make sure to queue
1811  * the 'original' request together with the readahead ones...  
1812  *
1813  * This has been extended to use the NUMA policies from the mm triggering
1814  * the readahead.
1815  *
1816  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1817  */
1818 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1819 {
1820 #ifdef CONFIG_NUMA
1821         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1822 #endif
1823         int i, num;
1824         struct page *new_page;
1825         unsigned long offset;
1826
1827         /*
1828          * Get the number of handles we should do readahead io to.
1829          */
1830         num = valid_swaphandles(entry, &offset);
1831         for (i = 0; i < num; offset++, i++) {
1832                 /* Ok, do the async read-ahead now */
1833                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1834                                                            offset), vma, addr);
1835                 if (!new_page)
1836                         break;
1837                 page_cache_release(new_page);
1838 #ifdef CONFIG_NUMA
1839                 /*
1840                  * Find the next applicable VMA for the NUMA policy.
1841                  */
1842                 addr += PAGE_SIZE;
1843                 if (addr == 0)
1844                         vma = NULL;
1845                 if (vma) {
1846                         if (addr >= vma->vm_end) {
1847                                 vma = next_vma;
1848                                 next_vma = vma ? vma->vm_next : NULL;
1849                         }
1850                         if (vma && addr < vma->vm_start)
1851                                 vma = NULL;
1852                 } else {
1853                         if (next_vma && addr >= next_vma->vm_start) {
1854                                 vma = next_vma;
1855                                 next_vma = vma->vm_next;
1856                         }
1857                 }
1858 #endif
1859         }
1860         lru_add_drain();        /* Push any new pages onto the LRU now */
1861 }
1862
1863 /*
1864  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1865  * but allow concurrent faults), and pte mapped but not yet locked.
1866  * We return with mmap_sem still held, but pte unmapped and unlocked.
1867  */
1868 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1869                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1870                 int write_access, pte_t orig_pte)
1871 {
1872         spinlock_t *ptl;
1873         struct page *page;
1874         swp_entry_t entry;
1875         pte_t pte;
1876         int ret = VM_FAULT_MINOR;
1877
1878         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1879                 goto out;
1880
1881         entry = pte_to_swp_entry(orig_pte);
1882 again:
1883         page = lookup_swap_cache(entry);
1884         if (!page) {
1885                 swapin_readahead(entry, address, vma);
1886                 page = read_swap_cache_async(entry, vma, address);
1887                 if (!page) {
1888                         /*
1889                          * Back out if somebody else faulted in this pte
1890                          * while we released the pte lock.
1891                          */
1892                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1893                         if (likely(pte_same(*page_table, orig_pte)))
1894                                 ret = VM_FAULT_OOM;
1895                         goto unlock;
1896                 }
1897
1898                 /* Had to read the page from swap area: Major fault */
1899                 ret = VM_FAULT_MAJOR;
1900                 inc_page_state(pgmajfault);
1901                 grab_swap_token();
1902         }
1903
1904         mark_page_accessed(page);
1905         lock_page(page);
1906         if (!PageSwapCache(page)) {
1907                 /* Page migration has occured */
1908                 unlock_page(page);
1909                 page_cache_release(page);
1910                 goto again;
1911         }
1912
1913         /*
1914          * Back out if somebody else already faulted in this pte.
1915          */
1916         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1917         if (unlikely(!pte_same(*page_table, orig_pte)))
1918                 goto out_nomap;
1919
1920         if (unlikely(!PageUptodate(page))) {
1921                 ret = VM_FAULT_SIGBUS;
1922                 goto out_nomap;
1923         }
1924
1925         /* The page isn't present yet, go ahead with the fault. */
1926
1927         inc_mm_counter(mm, anon_rss);
1928         pte = mk_pte(page, vma->vm_page_prot);
1929         if (write_access && can_share_swap_page(page)) {
1930                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1931                 write_access = 0;
1932         }
1933
1934         flush_icache_page(vma, page);
1935         set_pte_at(mm, address, page_table, pte);
1936         page_add_anon_rmap(page, vma, address);
1937
1938         swap_free(entry);
1939         if (vm_swap_full())
1940                 remove_exclusive_swap_page(page);
1941         unlock_page(page);
1942
1943         if (write_access) {
1944                 if (do_wp_page(mm, vma, address,
1945                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1946                         ret = VM_FAULT_OOM;
1947                 goto out;
1948         }
1949
1950         /* No need to invalidate - it was non-present before */
1951         update_mmu_cache(vma, address, pte);
1952         lazy_mmu_prot_update(pte);
1953 unlock:
1954         pte_unmap_unlock(page_table, ptl);
1955 out:
1956         return ret;
1957 out_nomap:
1958         pte_unmap_unlock(page_table, ptl);
1959         unlock_page(page);
1960         page_cache_release(page);
1961         return ret;
1962 }
1963
1964 /*
1965  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1966  * but allow concurrent faults), and pte mapped but not yet locked.
1967  * We return with mmap_sem still held, but pte unmapped and unlocked.
1968  */
1969 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1970                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1971                 int write_access)
1972 {
1973         struct page *page;
1974         spinlock_t *ptl;
1975         pte_t entry;
1976
1977         if (write_access) {
1978                 /* Allocate our own private page. */
1979                 pte_unmap(page_table);
1980
1981                 if (unlikely(anon_vma_prepare(vma)))
1982                         goto oom;
1983                 page = alloc_zeroed_user_highpage(vma, address);
1984                 if (!page)
1985                         goto oom;
1986
1987                 entry = mk_pte(page, vma->vm_page_prot);
1988                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1989
1990                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1991                 if (!pte_none(*page_table))
1992                         goto release;
1993                 inc_mm_counter(mm, anon_rss);
1994                 lru_cache_add_active(page);
1995                 page_add_new_anon_rmap(page, vma, address);
1996         } else {
1997                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1998                 page = ZERO_PAGE(address);
1999                 page_cache_get(page);
2000                 entry = mk_pte(page, vma->vm_page_prot);
2001
2002                 ptl = pte_lockptr(mm, pmd);
2003                 spin_lock(ptl);
2004                 if (!pte_none(*page_table))
2005                         goto release;
2006                 inc_mm_counter(mm, file_rss);
2007                 page_add_file_rmap(page);
2008         }
2009
2010         set_pte_at(mm, address, page_table, entry);
2011
2012         /* No need to invalidate - it was non-present before */
2013         update_mmu_cache(vma, address, entry);
2014         lazy_mmu_prot_update(entry);
2015 unlock:
2016         pte_unmap_unlock(page_table, ptl);
2017         return VM_FAULT_MINOR;
2018 release:
2019         page_cache_release(page);
2020         goto unlock;
2021 oom:
2022         return VM_FAULT_OOM;
2023 }
2024
2025 /*
2026  * do_no_page() tries to create a new page mapping. It aggressively
2027  * tries to share with existing pages, but makes a separate copy if
2028  * the "write_access" parameter is true in order to avoid the next
2029  * page fault.
2030  *
2031  * As this is called only for pages that do not currently exist, we
2032  * do not need to flush old virtual caches or the TLB.
2033  *
2034  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2035  * but allow concurrent faults), and pte mapped but not yet locked.
2036  * We return with mmap_sem still held, but pte unmapped and unlocked.
2037  */
2038 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2039                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2040                 int write_access)
2041 {
2042         spinlock_t *ptl;
2043         struct page *new_page;
2044         struct address_space *mapping = NULL;
2045         pte_t entry;
2046         unsigned int sequence = 0;
2047         int ret = VM_FAULT_MINOR;
2048         int anon = 0;
2049
2050         pte_unmap(page_table);
2051         BUG_ON(vma->vm_flags & VM_PFNMAP);
2052
2053         if (vma->vm_file) {
2054                 mapping = vma->vm_file->f_mapping;
2055                 sequence = mapping->truncate_count;
2056                 smp_rmb(); /* serializes i_size against truncate_count */
2057         }
2058 retry:
2059         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2060         /*
2061          * No smp_rmb is needed here as long as there's a full
2062          * spin_lock/unlock sequence inside the ->nopage callback
2063          * (for the pagecache lookup) that acts as an implicit
2064          * smp_mb() and prevents the i_size read to happen
2065          * after the next truncate_count read.
2066          */
2067
2068         /* no page was available -- either SIGBUS or OOM */
2069         if (new_page == NOPAGE_SIGBUS)
2070                 return VM_FAULT_SIGBUS;
2071         if (new_page == NOPAGE_OOM)
2072                 return VM_FAULT_OOM;
2073
2074         /*
2075          * Should we do an early C-O-W break?
2076          */
2077         if (write_access && !(vma->vm_flags & VM_SHARED)) {
2078                 struct page *page;
2079
2080                 if (unlikely(anon_vma_prepare(vma)))
2081                         goto oom;
2082                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2083                 if (!page)
2084                         goto oom;
2085                 copy_user_highpage(page, new_page, address);
2086                 page_cache_release(new_page);
2087                 new_page = page;
2088                 anon = 1;
2089         }
2090
2091         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2092         /*
2093          * For a file-backed vma, someone could have truncated or otherwise
2094          * invalidated this page.  If unmap_mapping_range got called,
2095          * retry getting the page.
2096          */
2097         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2098                 pte_unmap_unlock(page_table, ptl);
2099                 page_cache_release(new_page);
2100                 cond_resched();
2101                 sequence = mapping->truncate_count;
2102                 smp_rmb();
2103                 goto retry;
2104         }
2105
2106         /*
2107          * This silly early PAGE_DIRTY setting removes a race
2108          * due to the bad i386 page protection. But it's valid
2109          * for other architectures too.
2110          *
2111          * Note that if write_access is true, we either now have
2112          * an exclusive copy of the page, or this is a shared mapping,
2113          * so we can make it writable and dirty to avoid having to
2114          * handle that later.
2115          */
2116         /* Only go through if we didn't race with anybody else... */
2117         if (pte_none(*page_table)) {
2118                 flush_icache_page(vma, new_page);
2119                 entry = mk_pte(new_page, vma->vm_page_prot);
2120                 if (write_access)
2121                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2122                 set_pte_at(mm, address, page_table, entry);
2123                 if (anon) {
2124                         inc_mm_counter(mm, anon_rss);
2125                         lru_cache_add_active(new_page);
2126                         page_add_new_anon_rmap(new_page, vma, address);
2127                 } else {
2128                         inc_mm_counter(mm, file_rss);
2129                         page_add_file_rmap(new_page);
2130                 }
2131         } else {
2132                 /* One of our sibling threads was faster, back out. */
2133                 page_cache_release(new_page);
2134                 goto unlock;
2135         }
2136
2137         /* no need to invalidate: a not-present page shouldn't be cached */
2138         update_mmu_cache(vma, address, entry);
2139         lazy_mmu_prot_update(entry);
2140 unlock:
2141         pte_unmap_unlock(page_table, ptl);
2142         return ret;
2143 oom:
2144         page_cache_release(new_page);
2145         return VM_FAULT_OOM;
2146 }
2147
2148 /*
2149  * Fault of a previously existing named mapping. Repopulate the pte
2150  * from the encoded file_pte if possible. This enables swappable
2151  * nonlinear vmas.
2152  *
2153  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2154  * but allow concurrent faults), and pte mapped but not yet locked.
2155  * We return with mmap_sem still held, but pte unmapped and unlocked.
2156  */
2157 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2158                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2159                 int write_access, pte_t orig_pte)
2160 {
2161         pgoff_t pgoff;
2162         int err;
2163
2164         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2165                 return VM_FAULT_MINOR;
2166
2167         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2168                 /*
2169                  * Page table corrupted: show pte and kill process.
2170                  */
2171                 print_bad_pte(vma, orig_pte, address);
2172                 return VM_FAULT_OOM;
2173         }
2174         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2175
2176         pgoff = pte_to_pgoff(orig_pte);
2177         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2178                                         vma->vm_page_prot, pgoff, 0);
2179         if (err == -ENOMEM)
2180                 return VM_FAULT_OOM;
2181         if (err)
2182                 return VM_FAULT_SIGBUS;
2183         return VM_FAULT_MAJOR;
2184 }
2185
2186 /*
2187  * These routines also need to handle stuff like marking pages dirty
2188  * and/or accessed for architectures that don't do it in hardware (most
2189  * RISC architectures).  The early dirtying is also good on the i386.
2190  *
2191  * There is also a hook called "update_mmu_cache()" that architectures
2192  * with external mmu caches can use to update those (ie the Sparc or
2193  * PowerPC hashed page tables that act as extended TLBs).
2194  *
2195  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2196  * but allow concurrent faults), and pte mapped but not yet locked.
2197  * We return with mmap_sem still held, but pte unmapped and unlocked.
2198  */
2199 static inline int handle_pte_fault(struct mm_struct *mm,
2200                 struct vm_area_struct *vma, unsigned long address,
2201                 pte_t *pte, pmd_t *pmd, int write_access)
2202 {
2203         pte_t entry;
2204         pte_t old_entry;
2205         spinlock_t *ptl;
2206
2207         old_entry = entry = *pte;
2208         if (!pte_present(entry)) {
2209                 if (pte_none(entry)) {
2210                         if (!vma->vm_ops || !vma->vm_ops->nopage)
2211                                 return do_anonymous_page(mm, vma, address,
2212                                         pte, pmd, write_access);
2213                         return do_no_page(mm, vma, address,
2214                                         pte, pmd, write_access);
2215                 }
2216                 if (pte_file(entry))
2217                         return do_file_page(mm, vma, address,
2218                                         pte, pmd, write_access, entry);
2219                 return do_swap_page(mm, vma, address,
2220                                         pte, pmd, write_access, entry);
2221         }
2222
2223         ptl = pte_lockptr(mm, pmd);
2224         spin_lock(ptl);
2225         if (unlikely(!pte_same(*pte, entry)))
2226                 goto unlock;
2227         if (write_access) {
2228                 if (!pte_write(entry))
2229                         return do_wp_page(mm, vma, address,
2230                                         pte, pmd, ptl, entry);
2231                 entry = pte_mkdirty(entry);
2232         }
2233         entry = pte_mkyoung(entry);
2234         if (!pte_same(old_entry, entry)) {
2235                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2236                 update_mmu_cache(vma, address, entry);
2237                 lazy_mmu_prot_update(entry);
2238         } else {
2239                 /*
2240                  * This is needed only for protection faults but the arch code
2241                  * is not yet telling us if this is a protection fault or not.
2242                  * This still avoids useless tlb flushes for .text page faults
2243                  * with threads.
2244                  */
2245                 if (write_access)
2246                         flush_tlb_page(vma, address);
2247         }
2248 unlock:
2249         pte_unmap_unlock(pte, ptl);
2250         return VM_FAULT_MINOR;
2251 }
2252
2253 /*
2254  * By the time we get here, we already hold the mm semaphore
2255  */
2256 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2257                 unsigned long address, int write_access)
2258 {
2259         pgd_t *pgd;
2260         pud_t *pud;
2261         pmd_t *pmd;
2262         pte_t *pte;
2263
2264         __set_current_state(TASK_RUNNING);
2265
2266         inc_page_state(pgfault);
2267
2268         if (unlikely(is_vm_hugetlb_page(vma)))
2269                 return hugetlb_fault(mm, vma, address, write_access);
2270
2271         pgd = pgd_offset(mm, address);
2272         pud = pud_alloc(mm, pgd, address);
2273         if (!pud)
2274                 return VM_FAULT_OOM;
2275         pmd = pmd_alloc(mm, pud, address);
2276         if (!pmd)
2277                 return VM_FAULT_OOM;
2278         pte = pte_alloc_map(mm, pmd, address);
2279         if (!pte)
2280                 return VM_FAULT_OOM;
2281
2282         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2283 }
2284
2285 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2286
2287 #ifndef __PAGETABLE_PUD_FOLDED
2288 /*
2289  * Allocate page upper directory.
2290  * We've already handled the fast-path in-line.
2291  */
2292 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2293 {
2294         pud_t *new = pud_alloc_one(mm, address);
2295         if (!new)
2296                 return -ENOMEM;
2297
2298         spin_lock(&mm->page_table_lock);
2299         if (pgd_present(*pgd))          /* Another has populated it */
2300                 pud_free(new);
2301         else
2302                 pgd_populate(mm, pgd, new);
2303         spin_unlock(&mm->page_table_lock);
2304         return 0;
2305 }
2306 #else
2307 /* Workaround for gcc 2.96 */
2308 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2309 {
2310         return 0;
2311 }
2312 #endif /* __PAGETABLE_PUD_FOLDED */
2313
2314 #ifndef __PAGETABLE_PMD_FOLDED
2315 /*
2316  * Allocate page middle directory.
2317  * We've already handled the fast-path in-line.
2318  */
2319 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2320 {
2321         pmd_t *new = pmd_alloc_one(mm, address);
2322         if (!new)
2323                 return -ENOMEM;
2324
2325         spin_lock(&mm->page_table_lock);
2326 #ifndef __ARCH_HAS_4LEVEL_HACK
2327         if (pud_present(*pud))          /* Another has populated it */
2328                 pmd_free(new);
2329         else
2330                 pud_populate(mm, pud, new);
2331 #else
2332         if (pgd_present(*pud))          /* Another has populated it */
2333                 pmd_free(new);
2334         else
2335                 pgd_populate(mm, pud, new);
2336 #endif /* __ARCH_HAS_4LEVEL_HACK */
2337         spin_unlock(&mm->page_table_lock);
2338         return 0;
2339 }
2340 #else
2341 /* Workaround for gcc 2.96 */
2342 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2343 {
2344         return 0;
2345 }
2346 #endif /* __PAGETABLE_PMD_FOLDED */
2347
2348 int make_pages_present(unsigned long addr, unsigned long end)
2349 {
2350         int ret, len, write;
2351         struct vm_area_struct * vma;
2352
2353         vma = find_vma(current->mm, addr);
2354         if (!vma)
2355                 return -1;
2356         write = (vma->vm_flags & VM_WRITE) != 0;
2357         BUG_ON(addr >= end);
2358         BUG_ON(end > vma->vm_end);
2359         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2360         ret = get_user_pages(current, current->mm, addr,
2361                         len, write, 0, NULL, NULL);
2362         if (ret < 0)
2363                 return ret;
2364         return ret == len ? 0 : -1;
2365 }
2366
2367 /* 
2368  * Map a vmalloc()-space virtual address to the physical page.
2369  */
2370 struct page * vmalloc_to_page(void * vmalloc_addr)
2371 {
2372         unsigned long addr = (unsigned long) vmalloc_addr;
2373         struct page *page = NULL;
2374         pgd_t *pgd = pgd_offset_k(addr);
2375         pud_t *pud;
2376         pmd_t *pmd;
2377         pte_t *ptep, pte;
2378   
2379         if (!pgd_none(*pgd)) {
2380                 pud = pud_offset(pgd, addr);
2381                 if (!pud_none(*pud)) {
2382                         pmd = pmd_offset(pud, addr);
2383                         if (!pmd_none(*pmd)) {
2384                                 ptep = pte_offset_map(pmd, addr);
2385                                 pte = *ptep;
2386                                 if (pte_present(pte))
2387                                         page = pte_page(pte);
2388                                 pte_unmap(ptep);
2389                         }
2390                 }
2391         }
2392         return page;
2393 }
2394
2395 EXPORT_SYMBOL(vmalloc_to_page);
2396
2397 /*
2398  * Map a vmalloc()-space virtual address to the physical page frame number.
2399  */
2400 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2401 {
2402         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2403 }
2404
2405 EXPORT_SYMBOL(vmalloc_to_pfn);
2406
2407 #if !defined(__HAVE_ARCH_GATE_AREA)
2408
2409 #if defined(AT_SYSINFO_EHDR)
2410 static struct vm_area_struct gate_vma;
2411
2412 static int __init gate_vma_init(void)
2413 {
2414         gate_vma.vm_mm = NULL;
2415         gate_vma.vm_start = FIXADDR_USER_START;
2416         gate_vma.vm_end = FIXADDR_USER_END;
2417         gate_vma.vm_page_prot = PAGE_READONLY;
2418         gate_vma.vm_flags = 0;
2419         return 0;
2420 }
2421 __initcall(gate_vma_init);
2422 #endif
2423
2424 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2425 {
2426 #ifdef AT_SYSINFO_EHDR
2427         return &gate_vma;
2428 #else
2429         return NULL;
2430 #endif
2431 }
2432
2433 int in_gate_area_no_task(unsigned long addr)
2434 {
2435 #ifdef AT_SYSINFO_EHDR
2436         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2437                 return 1;
2438 #endif
2439         return 0;
2440 }
2441
2442 #endif  /* __HAVE_ARCH_GATE_AREA */