de8bc85dc8f3fea54d1898097038f3d57ce3ea33
[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 1;
91 }
92 __setup("norandmaps", disable_randmaps);
93
94
95 /*
96  * If a p?d_bad entry is found while walking page tables, report
97  * the error, before resetting entry to p?d_none.  Usually (but
98  * very seldom) called out from the p?d_none_or_clear_bad macros.
99  */
100
101 void pgd_clear_bad(pgd_t *pgd)
102 {
103         pgd_ERROR(*pgd);
104         pgd_clear(pgd);
105 }
106
107 void pud_clear_bad(pud_t *pud)
108 {
109         pud_ERROR(*pud);
110         pud_clear(pud);
111 }
112
113 void pmd_clear_bad(pmd_t *pmd)
114 {
115         pmd_ERROR(*pmd);
116         pmd_clear(pmd);
117 }
118
119 /*
120  * Note: this doesn't free the actual pages themselves. That
121  * has been handled earlier when unmapping all the memory regions.
122  */
123 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
124 {
125         struct page *page = pmd_page(*pmd);
126         pmd_clear(pmd);
127         pte_lock_deinit(page);
128         pte_free_tlb(tlb, page);
129         dec_zone_page_state(page, NR_PAGETABLE);
130         tlb->mm->nr_ptes--;
131 }
132
133 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
134                                 unsigned long addr, unsigned long end,
135                                 unsigned long floor, unsigned long ceiling)
136 {
137         pmd_t *pmd;
138         unsigned long next;
139         unsigned long start;
140
141         start = addr;
142         pmd = pmd_offset(pud, addr);
143         do {
144                 next = pmd_addr_end(addr, end);
145                 if (pmd_none_or_clear_bad(pmd))
146                         continue;
147                 free_pte_range(tlb, pmd);
148         } while (pmd++, addr = next, addr != end);
149
150         start &= PUD_MASK;
151         if (start < floor)
152                 return;
153         if (ceiling) {
154                 ceiling &= PUD_MASK;
155                 if (!ceiling)
156                         return;
157         }
158         if (end - 1 > ceiling - 1)
159                 return;
160
161         pmd = pmd_offset(pud, start);
162         pud_clear(pud);
163         pmd_free_tlb(tlb, pmd);
164 }
165
166 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
167                                 unsigned long addr, unsigned long end,
168                                 unsigned long floor, unsigned long ceiling)
169 {
170         pud_t *pud;
171         unsigned long next;
172         unsigned long start;
173
174         start = addr;
175         pud = pud_offset(pgd, addr);
176         do {
177                 next = pud_addr_end(addr, end);
178                 if (pud_none_or_clear_bad(pud))
179                         continue;
180                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
181         } while (pud++, addr = next, addr != end);
182
183         start &= PGDIR_MASK;
184         if (start < floor)
185                 return;
186         if (ceiling) {
187                 ceiling &= PGDIR_MASK;
188                 if (!ceiling)
189                         return;
190         }
191         if (end - 1 > ceiling - 1)
192                 return;
193
194         pud = pud_offset(pgd, start);
195         pgd_clear(pgd);
196         pud_free_tlb(tlb, pud);
197 }
198
199 /*
200  * This function frees user-level page tables of a process.
201  *
202  * Must be called with pagetable lock held.
203  */
204 void free_pgd_range(struct mmu_gather **tlb,
205                         unsigned long addr, unsigned long end,
206                         unsigned long floor, unsigned long ceiling)
207 {
208         pgd_t *pgd;
209         unsigned long next;
210         unsigned long start;
211
212         /*
213          * The next few lines have given us lots of grief...
214          *
215          * Why are we testing PMD* at this top level?  Because often
216          * there will be no work to do at all, and we'd prefer not to
217          * go all the way down to the bottom just to discover that.
218          *
219          * Why all these "- 1"s?  Because 0 represents both the bottom
220          * of the address space and the top of it (using -1 for the
221          * top wouldn't help much: the masks would do the wrong thing).
222          * The rule is that addr 0 and floor 0 refer to the bottom of
223          * the address space, but end 0 and ceiling 0 refer to the top
224          * Comparisons need to use "end - 1" and "ceiling - 1" (though
225          * that end 0 case should be mythical).
226          *
227          * Wherever addr is brought up or ceiling brought down, we must
228          * be careful to reject "the opposite 0" before it confuses the
229          * subsequent tests.  But what about where end is brought down
230          * by PMD_SIZE below? no, end can't go down to 0 there.
231          *
232          * Whereas we round start (addr) and ceiling down, by different
233          * masks at different levels, in order to test whether a table
234          * now has no other vmas using it, so can be freed, we don't
235          * bother to round floor or end up - the tests don't need that.
236          */
237
238         addr &= PMD_MASK;
239         if (addr < floor) {
240                 addr += PMD_SIZE;
241                 if (!addr)
242                         return;
243         }
244         if (ceiling) {
245                 ceiling &= PMD_MASK;
246                 if (!ceiling)
247                         return;
248         }
249         if (end - 1 > ceiling - 1)
250                 end -= PMD_SIZE;
251         if (addr > end - 1)
252                 return;
253
254         start = addr;
255         pgd = pgd_offset((*tlb)->mm, addr);
256         do {
257                 next = pgd_addr_end(addr, end);
258                 if (pgd_none_or_clear_bad(pgd))
259                         continue;
260                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
261         } while (pgd++, addr = next, addr != end);
262
263         if (!(*tlb)->fullmm)
264                 flush_tlb_pgtables((*tlb)->mm, start, end);
265 }
266
267 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
268                 unsigned long floor, unsigned long ceiling)
269 {
270         while (vma) {
271                 struct vm_area_struct *next = vma->vm_next;
272                 unsigned long addr = vma->vm_start;
273
274                 /*
275                  * Hide vma from rmap and vmtruncate before freeing pgtables
276                  */
277                 anon_vma_unlink(vma);
278                 unlink_file_vma(vma);
279
280                 if (is_vm_hugetlb_page(vma)) {
281                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
282                                 floor, next? next->vm_start: ceiling);
283                 } else {
284                         /*
285                          * Optimization: gather nearby vmas into one call down
286                          */
287                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
288                                && !is_vm_hugetlb_page(next)) {
289                                 vma = next;
290                                 next = vma->vm_next;
291                                 anon_vma_unlink(vma);
292                                 unlink_file_vma(vma);
293                         }
294                         free_pgd_range(tlb, addr, vma->vm_end,
295                                 floor, next? next->vm_start: ceiling);
296                 }
297                 vma = next;
298         }
299 }
300
301 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
302 {
303         struct page *new = pte_alloc_one(mm, address);
304         if (!new)
305                 return -ENOMEM;
306
307         pte_lock_init(new);
308         spin_lock(&mm->page_table_lock);
309         if (pmd_present(*pmd)) {        /* Another has populated it */
310                 pte_lock_deinit(new);
311                 pte_free(new);
312         } else {
313                 mm->nr_ptes++;
314                 inc_zone_page_state(new, NR_PAGETABLE);
315                 pmd_populate(mm, pmd, new);
316         }
317         spin_unlock(&mm->page_table_lock);
318         return 0;
319 }
320
321 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
322 {
323         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
324         if (!new)
325                 return -ENOMEM;
326
327         spin_lock(&init_mm.page_table_lock);
328         if (pmd_present(*pmd))          /* Another has populated it */
329                 pte_free_kernel(new);
330         else
331                 pmd_populate_kernel(&init_mm, pmd, new);
332         spin_unlock(&init_mm.page_table_lock);
333         return 0;
334 }
335
336 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
337 {
338         if (file_rss)
339                 add_mm_counter(mm, file_rss, file_rss);
340         if (anon_rss)
341                 add_mm_counter(mm, anon_rss, anon_rss);
342 }
343
344 /*
345  * This function is called to print an error when a bad pte
346  * is found. For example, we might have a PFN-mapped pte in
347  * a region that doesn't allow it.
348  *
349  * The calling function must still handle the error.
350  */
351 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
352 {
353         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
354                         "vm_flags = %lx, vaddr = %lx\n",
355                 (long long)pte_val(pte),
356                 (vma->vm_mm == current->mm ? current->comm : "???"),
357                 vma->vm_flags, vaddr);
358         dump_stack();
359 }
360
361 static inline int is_cow_mapping(unsigned int flags)
362 {
363         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
364 }
365
366 /*
367  * This function gets the "struct page" associated with a pte.
368  *
369  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
370  * will have each page table entry just pointing to a raw page frame
371  * number, and as far as the VM layer is concerned, those do not have
372  * pages associated with them - even if the PFN might point to memory
373  * that otherwise is perfectly fine and has a "struct page".
374  *
375  * The way we recognize those mappings is through the rules set up
376  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
377  * and the vm_pgoff will point to the first PFN mapped: thus every
378  * page that is a raw mapping will always honor the rule
379  *
380  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
381  *
382  * and if that isn't true, the page has been COW'ed (in which case it
383  * _does_ have a "struct page" associated with it even if it is in a
384  * VM_PFNMAP range).
385  */
386 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
387 {
388         unsigned long pfn = pte_pfn(pte);
389
390         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
391                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
392                 if (pfn == vma->vm_pgoff + off)
393                         return NULL;
394                 if (!is_cow_mapping(vma->vm_flags))
395                         return NULL;
396         }
397
398         /*
399          * Add some anal sanity checks for now. Eventually,
400          * we should just do "return pfn_to_page(pfn)", but
401          * in the meantime we check that we get a valid pfn,
402          * and that the resulting page looks ok.
403          */
404         if (unlikely(!pfn_valid(pfn))) {
405                 print_bad_pte(vma, pte, addr);
406                 return NULL;
407         }
408
409         /*
410          * NOTE! We still have PageReserved() pages in the page 
411          * tables. 
412          *
413          * The PAGE_ZERO() pages and various VDSO mappings can
414          * cause them to exist.
415          */
416         return pfn_to_page(pfn);
417 }
418
419 /*
420  * copy one vm_area from one task to the other. Assumes the page tables
421  * already present in the new task to be cleared in the whole range
422  * covered by this vma.
423  */
424
425 static inline void
426 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
427                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
428                 unsigned long addr, int *rss)
429 {
430         unsigned long vm_flags = vma->vm_flags;
431         pte_t pte = *src_pte;
432         struct page *page;
433
434         /* pte contains position in swap or file, so copy. */
435         if (unlikely(!pte_present(pte))) {
436                 if (!pte_file(pte)) {
437                         swp_entry_t entry = pte_to_swp_entry(pte);
438
439                         swap_duplicate(entry);
440                         /* make sure dst_mm is on swapoff's mmlist. */
441                         if (unlikely(list_empty(&dst_mm->mmlist))) {
442                                 spin_lock(&mmlist_lock);
443                                 if (list_empty(&dst_mm->mmlist))
444                                         list_add(&dst_mm->mmlist,
445                                                  &src_mm->mmlist);
446                                 spin_unlock(&mmlist_lock);
447                         }
448                         if (is_write_migration_entry(entry) &&
449                                         is_cow_mapping(vm_flags)) {
450                                 /*
451                                  * COW mappings require pages in both parent
452                                  * and child to be set to read.
453                                  */
454                                 make_migration_entry_read(&entry);
455                                 pte = swp_entry_to_pte(entry);
456                                 set_pte_at(src_mm, addr, src_pte, pte);
457                         }
458                 }
459                 goto out_set_pte;
460         }
461
462         /*
463          * If it's a COW mapping, write protect it both
464          * in the parent and the child
465          */
466         if (is_cow_mapping(vm_flags)) {
467                 ptep_set_wrprotect(src_mm, addr, src_pte);
468                 pte = *src_pte;
469         }
470
471         /*
472          * If it's a shared mapping, mark it clean in
473          * the child
474          */
475         if (vm_flags & VM_SHARED)
476                 pte = pte_mkclean(pte);
477         pte = pte_mkold(pte);
478
479         page = vm_normal_page(vma, addr, pte);
480         if (page) {
481                 get_page(page);
482                 page_dup_rmap(page);
483                 rss[!!PageAnon(page)]++;
484         }
485
486 out_set_pte:
487         set_pte_at(dst_mm, addr, dst_pte, pte);
488 }
489
490 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
491                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
492                 unsigned long addr, unsigned long end)
493 {
494         pte_t *src_pte, *dst_pte;
495         spinlock_t *src_ptl, *dst_ptl;
496         int progress = 0;
497         int rss[2];
498
499 again:
500         rss[1] = rss[0] = 0;
501         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
502         if (!dst_pte)
503                 return -ENOMEM;
504         src_pte = pte_offset_map_nested(src_pmd, addr);
505         src_ptl = pte_lockptr(src_mm, src_pmd);
506         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
507
508         do {
509                 /*
510                  * We are holding two locks at this point - either of them
511                  * could generate latencies in another task on another CPU.
512                  */
513                 if (progress >= 32) {
514                         progress = 0;
515                         if (need_resched() ||
516                             need_lockbreak(src_ptl) ||
517                             need_lockbreak(dst_ptl))
518                                 break;
519                 }
520                 if (pte_none(*src_pte)) {
521                         progress++;
522                         continue;
523                 }
524                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
525                 progress += 8;
526         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
527
528         spin_unlock(src_ptl);
529         pte_unmap_nested(src_pte - 1);
530         add_mm_rss(dst_mm, rss[0], rss[1]);
531         pte_unmap_unlock(dst_pte - 1, dst_ptl);
532         cond_resched();
533         if (addr != end)
534                 goto again;
535         return 0;
536 }
537
538 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
539                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
540                 unsigned long addr, unsigned long end)
541 {
542         pmd_t *src_pmd, *dst_pmd;
543         unsigned long next;
544
545         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
546         if (!dst_pmd)
547                 return -ENOMEM;
548         src_pmd = pmd_offset(src_pud, addr);
549         do {
550                 next = pmd_addr_end(addr, end);
551                 if (pmd_none_or_clear_bad(src_pmd))
552                         continue;
553                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
554                                                 vma, addr, next))
555                         return -ENOMEM;
556         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
557         return 0;
558 }
559
560 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
561                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
562                 unsigned long addr, unsigned long end)
563 {
564         pud_t *src_pud, *dst_pud;
565         unsigned long next;
566
567         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
568         if (!dst_pud)
569                 return -ENOMEM;
570         src_pud = pud_offset(src_pgd, addr);
571         do {
572                 next = pud_addr_end(addr, end);
573                 if (pud_none_or_clear_bad(src_pud))
574                         continue;
575                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
576                                                 vma, addr, next))
577                         return -ENOMEM;
578         } while (dst_pud++, src_pud++, addr = next, addr != end);
579         return 0;
580 }
581
582 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
583                 struct vm_area_struct *vma)
584 {
585         pgd_t *src_pgd, *dst_pgd;
586         unsigned long next;
587         unsigned long addr = vma->vm_start;
588         unsigned long end = vma->vm_end;
589
590         /*
591          * Don't copy ptes where a page fault will fill them correctly.
592          * Fork becomes much lighter when there are big shared or private
593          * readonly mappings. The tradeoff is that copy_page_range is more
594          * efficient than faulting.
595          */
596         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
597                 if (!vma->anon_vma)
598                         return 0;
599         }
600
601         if (is_vm_hugetlb_page(vma))
602                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
603
604         dst_pgd = pgd_offset(dst_mm, addr);
605         src_pgd = pgd_offset(src_mm, addr);
606         do {
607                 next = pgd_addr_end(addr, end);
608                 if (pgd_none_or_clear_bad(src_pgd))
609                         continue;
610                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
611                                                 vma, addr, next))
612                         return -ENOMEM;
613         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
614         return 0;
615 }
616
617 static unsigned long zap_pte_range(struct mmu_gather *tlb,
618                                 struct vm_area_struct *vma, pmd_t *pmd,
619                                 unsigned long addr, unsigned long end,
620                                 long *zap_work, struct zap_details *details)
621 {
622         struct mm_struct *mm = tlb->mm;
623         pte_t *pte;
624         spinlock_t *ptl;
625         int file_rss = 0;
626         int anon_rss = 0;
627
628         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
629         do {
630                 pte_t ptent = *pte;
631                 if (pte_none(ptent)) {
632                         (*zap_work)--;
633                         continue;
634                 }
635
636                 (*zap_work) -= PAGE_SIZE;
637
638                 if (pte_present(ptent)) {
639                         struct page *page;
640
641                         page = vm_normal_page(vma, addr, ptent);
642                         if (unlikely(details) && page) {
643                                 /*
644                                  * unmap_shared_mapping_pages() wants to
645                                  * invalidate cache without truncating:
646                                  * unmap shared but keep private pages.
647                                  */
648                                 if (details->check_mapping &&
649                                     details->check_mapping != page->mapping)
650                                         continue;
651                                 /*
652                                  * Each page->index must be checked when
653                                  * invalidating or truncating nonlinear.
654                                  */
655                                 if (details->nonlinear_vma &&
656                                     (page->index < details->first_index ||
657                                      page->index > details->last_index))
658                                         continue;
659                         }
660                         ptent = ptep_get_and_clear_full(mm, addr, pte,
661                                                         tlb->fullmm);
662                         tlb_remove_tlb_entry(tlb, pte, addr);
663                         if (unlikely(!page))
664                                 continue;
665                         if (unlikely(details) && details->nonlinear_vma
666                             && linear_page_index(details->nonlinear_vma,
667                                                 addr) != page->index)
668                                 set_pte_at(mm, addr, pte,
669                                            pgoff_to_pte(page->index));
670                         if (PageAnon(page))
671                                 anon_rss--;
672                         else {
673                                 if (pte_dirty(ptent))
674                                         set_page_dirty(page);
675                                 if (pte_young(ptent))
676                                         mark_page_accessed(page);
677                                 file_rss--;
678                         }
679                         page_remove_rmap(page);
680                         tlb_remove_page(tlb, page);
681                         continue;
682                 }
683                 /*
684                  * If details->check_mapping, we leave swap entries;
685                  * if details->nonlinear_vma, we leave file entries.
686                  */
687                 if (unlikely(details))
688                         continue;
689                 if (!pte_file(ptent))
690                         free_swap_and_cache(pte_to_swp_entry(ptent));
691                 pte_clear_full(mm, addr, pte, tlb->fullmm);
692         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
693
694         add_mm_rss(mm, file_rss, anon_rss);
695         pte_unmap_unlock(pte - 1, ptl);
696
697         return addr;
698 }
699
700 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
701                                 struct vm_area_struct *vma, pud_t *pud,
702                                 unsigned long addr, unsigned long end,
703                                 long *zap_work, struct zap_details *details)
704 {
705         pmd_t *pmd;
706         unsigned long next;
707
708         pmd = pmd_offset(pud, addr);
709         do {
710                 next = pmd_addr_end(addr, end);
711                 if (pmd_none_or_clear_bad(pmd)) {
712                         (*zap_work)--;
713                         continue;
714                 }
715                 next = zap_pte_range(tlb, vma, pmd, addr, next,
716                                                 zap_work, details);
717         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
718
719         return addr;
720 }
721
722 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
723                                 struct vm_area_struct *vma, pgd_t *pgd,
724                                 unsigned long addr, unsigned long end,
725                                 long *zap_work, struct zap_details *details)
726 {
727         pud_t *pud;
728         unsigned long next;
729
730         pud = pud_offset(pgd, addr);
731         do {
732                 next = pud_addr_end(addr, end);
733                 if (pud_none_or_clear_bad(pud)) {
734                         (*zap_work)--;
735                         continue;
736                 }
737                 next = zap_pmd_range(tlb, vma, pud, addr, next,
738                                                 zap_work, details);
739         } while (pud++, addr = next, (addr != end && *zap_work > 0));
740
741         return addr;
742 }
743
744 static unsigned long unmap_page_range(struct mmu_gather *tlb,
745                                 struct vm_area_struct *vma,
746                                 unsigned long addr, unsigned long end,
747                                 long *zap_work, struct zap_details *details)
748 {
749         pgd_t *pgd;
750         unsigned long next;
751
752         if (details && !details->check_mapping && !details->nonlinear_vma)
753                 details = NULL;
754
755         BUG_ON(addr >= end);
756         tlb_start_vma(tlb, vma);
757         pgd = pgd_offset(vma->vm_mm, addr);
758         do {
759                 next = pgd_addr_end(addr, end);
760                 if (pgd_none_or_clear_bad(pgd)) {
761                         (*zap_work)--;
762                         continue;
763                 }
764                 next = zap_pud_range(tlb, vma, pgd, addr, next,
765                                                 zap_work, details);
766         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
767         tlb_end_vma(tlb, vma);
768
769         return addr;
770 }
771
772 #ifdef CONFIG_PREEMPT
773 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
774 #else
775 /* No preempt: go for improved straight-line efficiency */
776 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
777 #endif
778
779 /**
780  * unmap_vmas - unmap a range of memory covered by a list of vma's
781  * @tlbp: address of the caller's struct mmu_gather
782  * @vma: the starting vma
783  * @start_addr: virtual address at which to start unmapping
784  * @end_addr: virtual address at which to end unmapping
785  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
786  * @details: details of nonlinear truncation or shared cache invalidation
787  *
788  * Returns the end address of the unmapping (restart addr if interrupted).
789  *
790  * Unmap all pages in the vma list.
791  *
792  * We aim to not hold locks for too long (for scheduling latency reasons).
793  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
794  * return the ending mmu_gather to the caller.
795  *
796  * Only addresses between `start' and `end' will be unmapped.
797  *
798  * The VMA list must be sorted in ascending virtual address order.
799  *
800  * unmap_vmas() assumes that the caller will flush the whole unmapped address
801  * range after unmap_vmas() returns.  So the only responsibility here is to
802  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
803  * drops the lock and schedules.
804  */
805 unsigned long unmap_vmas(struct mmu_gather **tlbp,
806                 struct vm_area_struct *vma, unsigned long start_addr,
807                 unsigned long end_addr, unsigned long *nr_accounted,
808                 struct zap_details *details)
809 {
810         long zap_work = ZAP_BLOCK_SIZE;
811         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
812         int tlb_start_valid = 0;
813         unsigned long start = start_addr;
814         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
815         int fullmm = (*tlbp)->fullmm;
816
817         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
818                 unsigned long end;
819
820                 start = max(vma->vm_start, start_addr);
821                 if (start >= vma->vm_end)
822                         continue;
823                 end = min(vma->vm_end, end_addr);
824                 if (end <= vma->vm_start)
825                         continue;
826
827                 if (vma->vm_flags & VM_ACCOUNT)
828                         *nr_accounted += (end - start) >> PAGE_SHIFT;
829
830                 while (start != end) {
831                         if (!tlb_start_valid) {
832                                 tlb_start = start;
833                                 tlb_start_valid = 1;
834                         }
835
836                         if (unlikely(is_vm_hugetlb_page(vma))) {
837                                 unmap_hugepage_range(vma, start, end);
838                                 zap_work -= (end - start) /
839                                                 (HPAGE_SIZE / PAGE_SIZE);
840                                 start = end;
841                         } else
842                                 start = unmap_page_range(*tlbp, vma,
843                                                 start, end, &zap_work, details);
844
845                         if (zap_work > 0) {
846                                 BUG_ON(start != end);
847                                 break;
848                         }
849
850                         tlb_finish_mmu(*tlbp, tlb_start, start);
851
852                         if (need_resched() ||
853                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
854                                 if (i_mmap_lock) {
855                                         *tlbp = NULL;
856                                         goto out;
857                                 }
858                                 cond_resched();
859                         }
860
861                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
862                         tlb_start_valid = 0;
863                         zap_work = ZAP_BLOCK_SIZE;
864                 }
865         }
866 out:
867         return start;   /* which is now the end (or restart) address */
868 }
869
870 /**
871  * zap_page_range - remove user pages in a given range
872  * @vma: vm_area_struct holding the applicable pages
873  * @address: starting address of pages to zap
874  * @size: number of bytes to zap
875  * @details: details of nonlinear truncation or shared cache invalidation
876  */
877 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
878                 unsigned long size, struct zap_details *details)
879 {
880         struct mm_struct *mm = vma->vm_mm;
881         struct mmu_gather *tlb;
882         unsigned long end = address + size;
883         unsigned long nr_accounted = 0;
884
885         lru_add_drain();
886         tlb = tlb_gather_mmu(mm, 0);
887         update_hiwater_rss(mm);
888         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
889         if (tlb)
890                 tlb_finish_mmu(tlb, address, end);
891         return end;
892 }
893
894 /*
895  * Do a quick page-table lookup for a single page.
896  */
897 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
898                         unsigned int flags)
899 {
900         pgd_t *pgd;
901         pud_t *pud;
902         pmd_t *pmd;
903         pte_t *ptep, pte;
904         spinlock_t *ptl;
905         struct page *page;
906         struct mm_struct *mm = vma->vm_mm;
907
908         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
909         if (!IS_ERR(page)) {
910                 BUG_ON(flags & FOLL_GET);
911                 goto out;
912         }
913
914         page = NULL;
915         pgd = pgd_offset(mm, address);
916         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
917                 goto no_page_table;
918
919         pud = pud_offset(pgd, address);
920         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
921                 goto no_page_table;
922         
923         pmd = pmd_offset(pud, address);
924         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
925                 goto no_page_table;
926
927         if (pmd_huge(*pmd)) {
928                 BUG_ON(flags & FOLL_GET);
929                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
930                 goto out;
931         }
932
933         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
934         if (!ptep)
935                 goto out;
936
937         pte = *ptep;
938         if (!pte_present(pte))
939                 goto unlock;
940         if ((flags & FOLL_WRITE) && !pte_write(pte))
941                 goto unlock;
942         page = vm_normal_page(vma, address, pte);
943         if (unlikely(!page))
944                 goto unlock;
945
946         if (flags & FOLL_GET)
947                 get_page(page);
948         if (flags & FOLL_TOUCH) {
949                 if ((flags & FOLL_WRITE) &&
950                     !pte_dirty(pte) && !PageDirty(page))
951                         set_page_dirty(page);
952                 mark_page_accessed(page);
953         }
954 unlock:
955         pte_unmap_unlock(ptep, ptl);
956 out:
957         return page;
958
959 no_page_table:
960         /*
961          * When core dumping an enormous anonymous area that nobody
962          * has touched so far, we don't want to allocate page tables.
963          */
964         if (flags & FOLL_ANON) {
965                 page = ZERO_PAGE(address);
966                 if (flags & FOLL_GET)
967                         get_page(page);
968                 BUG_ON(flags & FOLL_WRITE);
969         }
970         return page;
971 }
972
973 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
974                 unsigned long start, int len, int write, int force,
975                 struct page **pages, struct vm_area_struct **vmas)
976 {
977         int i;
978         unsigned int vm_flags;
979
980         /* 
981          * Require read or write permissions.
982          * If 'force' is set, we only require the "MAY" flags.
983          */
984         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
985         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
986         i = 0;
987
988         do {
989                 struct vm_area_struct *vma;
990                 unsigned int foll_flags;
991
992                 vma = find_extend_vma(mm, start);
993                 if (!vma && in_gate_area(tsk, start)) {
994                         unsigned long pg = start & PAGE_MASK;
995                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
996                         pgd_t *pgd;
997                         pud_t *pud;
998                         pmd_t *pmd;
999                         pte_t *pte;
1000                         if (write) /* user gate pages are read-only */
1001                                 return i ? : -EFAULT;
1002                         if (pg > TASK_SIZE)
1003                                 pgd = pgd_offset_k(pg);
1004                         else
1005                                 pgd = pgd_offset_gate(mm, pg);
1006                         BUG_ON(pgd_none(*pgd));
1007                         pud = pud_offset(pgd, pg);
1008                         BUG_ON(pud_none(*pud));
1009                         pmd = pmd_offset(pud, pg);
1010                         if (pmd_none(*pmd))
1011                                 return i ? : -EFAULT;
1012                         pte = pte_offset_map(pmd, pg);
1013                         if (pte_none(*pte)) {
1014                                 pte_unmap(pte);
1015                                 return i ? : -EFAULT;
1016                         }
1017                         if (pages) {
1018                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1019                                 pages[i] = page;
1020                                 if (page)
1021                                         get_page(page);
1022                         }
1023                         pte_unmap(pte);
1024                         if (vmas)
1025                                 vmas[i] = gate_vma;
1026                         i++;
1027                         start += PAGE_SIZE;
1028                         len--;
1029                         continue;
1030                 }
1031
1032                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1033                                 || !(vm_flags & vma->vm_flags))
1034                         return i ? : -EFAULT;
1035
1036                 if (is_vm_hugetlb_page(vma)) {
1037                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1038                                                 &start, &len, i);
1039                         continue;
1040                 }
1041
1042                 foll_flags = FOLL_TOUCH;
1043                 if (pages)
1044                         foll_flags |= FOLL_GET;
1045                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1046                     (!vma->vm_ops || !vma->vm_ops->nopage))
1047                         foll_flags |= FOLL_ANON;
1048
1049                 do {
1050                         struct page *page;
1051
1052                         if (write)
1053                                 foll_flags |= FOLL_WRITE;
1054
1055                         cond_resched();
1056                         while (!(page = follow_page(vma, start, foll_flags))) {
1057                                 int ret;
1058                                 ret = __handle_mm_fault(mm, vma, start,
1059                                                 foll_flags & FOLL_WRITE);
1060                                 /*
1061                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1062                                  * broken COW when necessary, even if maybe_mkwrite
1063                                  * decided not to set pte_write. We can thus safely do
1064                                  * subsequent page lookups as if they were reads.
1065                                  */
1066                                 if (ret & VM_FAULT_WRITE)
1067                                         foll_flags &= ~FOLL_WRITE;
1068                                 
1069                                 switch (ret & ~VM_FAULT_WRITE) {
1070                                 case VM_FAULT_MINOR:
1071                                         tsk->min_flt++;
1072                                         break;
1073                                 case VM_FAULT_MAJOR:
1074                                         tsk->maj_flt++;
1075                                         break;
1076                                 case VM_FAULT_SIGBUS:
1077                                         return i ? i : -EFAULT;
1078                                 case VM_FAULT_OOM:
1079                                         return i ? i : -ENOMEM;
1080                                 default:
1081                                         BUG();
1082                                 }
1083                         }
1084                         if (pages) {
1085                                 pages[i] = page;
1086
1087                                 flush_anon_page(page, start);
1088                                 flush_dcache_page(page);
1089                         }
1090                         if (vmas)
1091                                 vmas[i] = vma;
1092                         i++;
1093                         start += PAGE_SIZE;
1094                         len--;
1095                 } while (len && start < vma->vm_end);
1096         } while (len);
1097         return i;
1098 }
1099 EXPORT_SYMBOL(get_user_pages);
1100
1101 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1102                         unsigned long addr, unsigned long end, pgprot_t prot)
1103 {
1104         pte_t *pte;
1105         spinlock_t *ptl;
1106
1107         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1108         if (!pte)
1109                 return -ENOMEM;
1110         do {
1111                 struct page *page = ZERO_PAGE(addr);
1112                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1113                 page_cache_get(page);
1114                 page_add_file_rmap(page);
1115                 inc_mm_counter(mm, file_rss);
1116                 BUG_ON(!pte_none(*pte));
1117                 set_pte_at(mm, addr, pte, zero_pte);
1118         } while (pte++, addr += PAGE_SIZE, addr != end);
1119         pte_unmap_unlock(pte - 1, ptl);
1120         return 0;
1121 }
1122
1123 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1124                         unsigned long addr, unsigned long end, pgprot_t prot)
1125 {
1126         pmd_t *pmd;
1127         unsigned long next;
1128
1129         pmd = pmd_alloc(mm, pud, addr);
1130         if (!pmd)
1131                 return -ENOMEM;
1132         do {
1133                 next = pmd_addr_end(addr, end);
1134                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1135                         return -ENOMEM;
1136         } while (pmd++, addr = next, addr != end);
1137         return 0;
1138 }
1139
1140 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1141                         unsigned long addr, unsigned long end, pgprot_t prot)
1142 {
1143         pud_t *pud;
1144         unsigned long next;
1145
1146         pud = pud_alloc(mm, pgd, addr);
1147         if (!pud)
1148                 return -ENOMEM;
1149         do {
1150                 next = pud_addr_end(addr, end);
1151                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1152                         return -ENOMEM;
1153         } while (pud++, addr = next, addr != end);
1154         return 0;
1155 }
1156
1157 int zeromap_page_range(struct vm_area_struct *vma,
1158                         unsigned long addr, unsigned long size, pgprot_t prot)
1159 {
1160         pgd_t *pgd;
1161         unsigned long next;
1162         unsigned long end = addr + size;
1163         struct mm_struct *mm = vma->vm_mm;
1164         int err;
1165
1166         BUG_ON(addr >= end);
1167         pgd = pgd_offset(mm, addr);
1168         flush_cache_range(vma, addr, end);
1169         do {
1170                 next = pgd_addr_end(addr, end);
1171                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1172                 if (err)
1173                         break;
1174         } while (pgd++, addr = next, addr != end);
1175         return err;
1176 }
1177
1178 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1179 {
1180         pgd_t * pgd = pgd_offset(mm, addr);
1181         pud_t * pud = pud_alloc(mm, pgd, addr);
1182         if (pud) {
1183                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1184                 if (pmd)
1185                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1186         }
1187         return NULL;
1188 }
1189
1190 /*
1191  * This is the old fallback for page remapping.
1192  *
1193  * For historical reasons, it only allows reserved pages. Only
1194  * old drivers should use this, and they needed to mark their
1195  * pages reserved for the old functions anyway.
1196  */
1197 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1198 {
1199         int retval;
1200         pte_t *pte;
1201         spinlock_t *ptl;  
1202
1203         retval = -EINVAL;
1204         if (PageAnon(page))
1205                 goto out;
1206         retval = -ENOMEM;
1207         flush_dcache_page(page);
1208         pte = get_locked_pte(mm, addr, &ptl);
1209         if (!pte)
1210                 goto out;
1211         retval = -EBUSY;
1212         if (!pte_none(*pte))
1213                 goto out_unlock;
1214
1215         /* Ok, finally just insert the thing.. */
1216         get_page(page);
1217         inc_mm_counter(mm, file_rss);
1218         page_add_file_rmap(page);
1219         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1220
1221         retval = 0;
1222 out_unlock:
1223         pte_unmap_unlock(pte, ptl);
1224 out:
1225         return retval;
1226 }
1227
1228 /*
1229  * This allows drivers to insert individual pages they've allocated
1230  * into a user vma.
1231  *
1232  * The page has to be a nice clean _individual_ kernel allocation.
1233  * If you allocate a compound page, you need to have marked it as
1234  * such (__GFP_COMP), or manually just split the page up yourself
1235  * (see split_page()).
1236  *
1237  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1238  * took an arbitrary page protection parameter. This doesn't allow
1239  * that. Your vma protection will have to be set up correctly, which
1240  * means that if you want a shared writable mapping, you'd better
1241  * ask for a shared writable mapping!
1242  *
1243  * The page does not need to be reserved.
1244  */
1245 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1246 {
1247         if (addr < vma->vm_start || addr >= vma->vm_end)
1248                 return -EFAULT;
1249         if (!page_count(page))
1250                 return -EINVAL;
1251         vma->vm_flags |= VM_INSERTPAGE;
1252         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1253 }
1254 EXPORT_SYMBOL(vm_insert_page);
1255
1256 /*
1257  * maps a range of physical memory into the requested pages. the old
1258  * mappings are removed. any references to nonexistent pages results
1259  * in null mappings (currently treated as "copy-on-access")
1260  */
1261 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1262                         unsigned long addr, unsigned long end,
1263                         unsigned long pfn, pgprot_t prot)
1264 {
1265         pte_t *pte;
1266         spinlock_t *ptl;
1267
1268         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1269         if (!pte)
1270                 return -ENOMEM;
1271         do {
1272                 BUG_ON(!pte_none(*pte));
1273                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1274                 pfn++;
1275         } while (pte++, addr += PAGE_SIZE, addr != end);
1276         pte_unmap_unlock(pte - 1, ptl);
1277         return 0;
1278 }
1279
1280 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1281                         unsigned long addr, unsigned long end,
1282                         unsigned long pfn, pgprot_t prot)
1283 {
1284         pmd_t *pmd;
1285         unsigned long next;
1286
1287         pfn -= addr >> PAGE_SHIFT;
1288         pmd = pmd_alloc(mm, pud, addr);
1289         if (!pmd)
1290                 return -ENOMEM;
1291         do {
1292                 next = pmd_addr_end(addr, end);
1293                 if (remap_pte_range(mm, pmd, addr, next,
1294                                 pfn + (addr >> PAGE_SHIFT), prot))
1295                         return -ENOMEM;
1296         } while (pmd++, addr = next, addr != end);
1297         return 0;
1298 }
1299
1300 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1301                         unsigned long addr, unsigned long end,
1302                         unsigned long pfn, pgprot_t prot)
1303 {
1304         pud_t *pud;
1305         unsigned long next;
1306
1307         pfn -= addr >> PAGE_SHIFT;
1308         pud = pud_alloc(mm, pgd, addr);
1309         if (!pud)
1310                 return -ENOMEM;
1311         do {
1312                 next = pud_addr_end(addr, end);
1313                 if (remap_pmd_range(mm, pud, addr, next,
1314                                 pfn + (addr >> PAGE_SHIFT), prot))
1315                         return -ENOMEM;
1316         } while (pud++, addr = next, addr != end);
1317         return 0;
1318 }
1319
1320 /*  Note: this is only safe if the mm semaphore is held when called. */
1321 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1322                     unsigned long pfn, unsigned long size, pgprot_t prot)
1323 {
1324         pgd_t *pgd;
1325         unsigned long next;
1326         unsigned long end = addr + PAGE_ALIGN(size);
1327         struct mm_struct *mm = vma->vm_mm;
1328         int err;
1329
1330         /*
1331          * Physically remapped pages are special. Tell the
1332          * rest of the world about it:
1333          *   VM_IO tells people not to look at these pages
1334          *      (accesses can have side effects).
1335          *   VM_RESERVED is specified all over the place, because
1336          *      in 2.4 it kept swapout's vma scan off this vma; but
1337          *      in 2.6 the LRU scan won't even find its pages, so this
1338          *      flag means no more than count its pages in reserved_vm,
1339          *      and omit it from core dump, even when VM_IO turned off.
1340          *   VM_PFNMAP tells the core MM that the base pages are just
1341          *      raw PFN mappings, and do not have a "struct page" associated
1342          *      with them.
1343          *
1344          * There's a horrible special case to handle copy-on-write
1345          * behaviour that some programs depend on. We mark the "original"
1346          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1347          */
1348         if (is_cow_mapping(vma->vm_flags)) {
1349                 if (addr != vma->vm_start || end != vma->vm_end)
1350                         return -EINVAL;
1351                 vma->vm_pgoff = pfn;
1352         }
1353
1354         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1355
1356         BUG_ON(addr >= end);
1357         pfn -= addr >> PAGE_SHIFT;
1358         pgd = pgd_offset(mm, addr);
1359         flush_cache_range(vma, addr, end);
1360         do {
1361                 next = pgd_addr_end(addr, end);
1362                 err = remap_pud_range(mm, pgd, addr, next,
1363                                 pfn + (addr >> PAGE_SHIFT), prot);
1364                 if (err)
1365                         break;
1366         } while (pgd++, addr = next, addr != end);
1367         return err;
1368 }
1369 EXPORT_SYMBOL(remap_pfn_range);
1370
1371 /*
1372  * handle_pte_fault chooses page fault handler according to an entry
1373  * which was read non-atomically.  Before making any commitment, on
1374  * those architectures or configurations (e.g. i386 with PAE) which
1375  * might give a mix of unmatched parts, do_swap_page and do_file_page
1376  * must check under lock before unmapping the pte and proceeding
1377  * (but do_wp_page is only called after already making such a check;
1378  * and do_anonymous_page and do_no_page can safely check later on).
1379  */
1380 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1381                                 pte_t *page_table, pte_t orig_pte)
1382 {
1383         int same = 1;
1384 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1385         if (sizeof(pte_t) > sizeof(unsigned long)) {
1386                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1387                 spin_lock(ptl);
1388                 same = pte_same(*page_table, orig_pte);
1389                 spin_unlock(ptl);
1390         }
1391 #endif
1392         pte_unmap(page_table);
1393         return same;
1394 }
1395
1396 /*
1397  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1398  * servicing faults for write access.  In the normal case, do always want
1399  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1400  * that do not have writing enabled, when used by access_process_vm.
1401  */
1402 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1403 {
1404         if (likely(vma->vm_flags & VM_WRITE))
1405                 pte = pte_mkwrite(pte);
1406         return pte;
1407 }
1408
1409 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1410 {
1411         /*
1412          * If the source page was a PFN mapping, we don't have
1413          * a "struct page" for it. We do a best-effort copy by
1414          * just copying from the original user address. If that
1415          * fails, we just zero-fill it. Live with it.
1416          */
1417         if (unlikely(!src)) {
1418                 void *kaddr = kmap_atomic(dst, KM_USER0);
1419                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1420
1421                 /*
1422                  * This really shouldn't fail, because the page is there
1423                  * in the page tables. But it might just be unreadable,
1424                  * in which case we just give up and fill the result with
1425                  * zeroes.
1426                  */
1427                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1428                         memset(kaddr, 0, PAGE_SIZE);
1429                 kunmap_atomic(kaddr, KM_USER0);
1430                 return;
1431                 
1432         }
1433         copy_user_highpage(dst, src, va);
1434 }
1435
1436 /*
1437  * This routine handles present pages, when users try to write
1438  * to a shared page. It is done by copying the page to a new address
1439  * and decrementing the shared-page counter for the old page.
1440  *
1441  * Note that this routine assumes that the protection checks have been
1442  * done by the caller (the low-level page fault routine in most cases).
1443  * Thus we can safely just mark it writable once we've done any necessary
1444  * COW.
1445  *
1446  * We also mark the page dirty at this point even though the page will
1447  * change only once the write actually happens. This avoids a few races,
1448  * and potentially makes it more efficient.
1449  *
1450  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1451  * but allow concurrent faults), with pte both mapped and locked.
1452  * We return with mmap_sem still held, but pte unmapped and unlocked.
1453  */
1454 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1455                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1456                 spinlock_t *ptl, pte_t orig_pte)
1457 {
1458         struct page *old_page, *new_page;
1459         pte_t entry;
1460         int reuse, ret = VM_FAULT_MINOR;
1461
1462         old_page = vm_normal_page(vma, address, orig_pte);
1463         if (!old_page)
1464                 goto gotten;
1465
1466         if (unlikely((vma->vm_flags & (VM_SHARED|VM_WRITE)) ==
1467                                 (VM_SHARED|VM_WRITE))) {
1468                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1469                         /*
1470                          * Notify the address space that the page is about to
1471                          * become writable so that it can prohibit this or wait
1472                          * for the page to get into an appropriate state.
1473                          *
1474                          * We do this without the lock held, so that it can
1475                          * sleep if it needs to.
1476                          */
1477                         page_cache_get(old_page);
1478                         pte_unmap_unlock(page_table, ptl);
1479
1480                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1481                                 goto unwritable_page;
1482
1483                         page_cache_release(old_page);
1484
1485                         /*
1486                          * Since we dropped the lock we need to revalidate
1487                          * the PTE as someone else may have changed it.  If
1488                          * they did, we just return, as we can count on the
1489                          * MMU to tell us if they didn't also make it writable.
1490                          */
1491                         page_table = pte_offset_map_lock(mm, pmd, address,
1492                                                          &ptl);
1493                         if (!pte_same(*page_table, orig_pte))
1494                                 goto unlock;
1495                 }
1496
1497                 reuse = 1;
1498         } else if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1499                 reuse = can_share_swap_page(old_page);
1500                 unlock_page(old_page);
1501         } else {
1502                 reuse = 0;
1503         }
1504
1505         if (reuse) {
1506                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1507                 entry = pte_mkyoung(orig_pte);
1508                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1509                 ptep_set_access_flags(vma, address, page_table, entry, 1);
1510                 update_mmu_cache(vma, address, entry);
1511                 lazy_mmu_prot_update(entry);
1512                 ret |= VM_FAULT_WRITE;
1513                 goto unlock;
1514         }
1515
1516         /*
1517          * Ok, we need to copy. Oh, well..
1518          */
1519         page_cache_get(old_page);
1520 gotten:
1521         pte_unmap_unlock(page_table, ptl);
1522
1523         if (unlikely(anon_vma_prepare(vma)))
1524                 goto oom;
1525         if (old_page == ZERO_PAGE(address)) {
1526                 new_page = alloc_zeroed_user_highpage(vma, address);
1527                 if (!new_page)
1528                         goto oom;
1529         } else {
1530                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1531                 if (!new_page)
1532                         goto oom;
1533                 cow_user_page(new_page, old_page, address);
1534         }
1535
1536         /*
1537          * Re-check the pte - we dropped the lock
1538          */
1539         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1540         if (likely(pte_same(*page_table, orig_pte))) {
1541                 if (old_page) {
1542                         page_remove_rmap(old_page);
1543                         if (!PageAnon(old_page)) {
1544                                 dec_mm_counter(mm, file_rss);
1545                                 inc_mm_counter(mm, anon_rss);
1546                         }
1547                 } else
1548                         inc_mm_counter(mm, anon_rss);
1549                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1550                 entry = mk_pte(new_page, vma->vm_page_prot);
1551                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1552                 lazy_mmu_prot_update(entry);
1553                 ptep_establish(vma, address, page_table, entry);
1554                 update_mmu_cache(vma, address, entry);
1555                 lru_cache_add_active(new_page);
1556                 page_add_new_anon_rmap(new_page, vma, address);
1557
1558                 /* Free the old page.. */
1559                 new_page = old_page;
1560                 ret |= VM_FAULT_WRITE;
1561         }
1562         if (new_page)
1563                 page_cache_release(new_page);
1564         if (old_page)
1565                 page_cache_release(old_page);
1566 unlock:
1567         pte_unmap_unlock(page_table, ptl);
1568         return ret;
1569 oom:
1570         if (old_page)
1571                 page_cache_release(old_page);
1572         return VM_FAULT_OOM;
1573
1574 unwritable_page:
1575         page_cache_release(old_page);
1576         return VM_FAULT_SIGBUS;
1577 }
1578
1579 /*
1580  * Helper functions for unmap_mapping_range().
1581  *
1582  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1583  *
1584  * We have to restart searching the prio_tree whenever we drop the lock,
1585  * since the iterator is only valid while the lock is held, and anyway
1586  * a later vma might be split and reinserted earlier while lock dropped.
1587  *
1588  * The list of nonlinear vmas could be handled more efficiently, using
1589  * a placeholder, but handle it in the same way until a need is shown.
1590  * It is important to search the prio_tree before nonlinear list: a vma
1591  * may become nonlinear and be shifted from prio_tree to nonlinear list
1592  * while the lock is dropped; but never shifted from list to prio_tree.
1593  *
1594  * In order to make forward progress despite restarting the search,
1595  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1596  * quickly skip it next time around.  Since the prio_tree search only
1597  * shows us those vmas affected by unmapping the range in question, we
1598  * can't efficiently keep all vmas in step with mapping->truncate_count:
1599  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1600  * mapping->truncate_count and vma->vm_truncate_count are protected by
1601  * i_mmap_lock.
1602  *
1603  * In order to make forward progress despite repeatedly restarting some
1604  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1605  * and restart from that address when we reach that vma again.  It might
1606  * have been split or merged, shrunk or extended, but never shifted: so
1607  * restart_addr remains valid so long as it remains in the vma's range.
1608  * unmap_mapping_range forces truncate_count to leap over page-aligned
1609  * values so we can save vma's restart_addr in its truncate_count field.
1610  */
1611 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1612
1613 static void reset_vma_truncate_counts(struct address_space *mapping)
1614 {
1615         struct vm_area_struct *vma;
1616         struct prio_tree_iter iter;
1617
1618         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1619                 vma->vm_truncate_count = 0;
1620         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1621                 vma->vm_truncate_count = 0;
1622 }
1623
1624 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1625                 unsigned long start_addr, unsigned long end_addr,
1626                 struct zap_details *details)
1627 {
1628         unsigned long restart_addr;
1629         int need_break;
1630
1631 again:
1632         restart_addr = vma->vm_truncate_count;
1633         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1634                 start_addr = restart_addr;
1635                 if (start_addr >= end_addr) {
1636                         /* Top of vma has been split off since last time */
1637                         vma->vm_truncate_count = details->truncate_count;
1638                         return 0;
1639                 }
1640         }
1641
1642         restart_addr = zap_page_range(vma, start_addr,
1643                                         end_addr - start_addr, details);
1644         need_break = need_resched() ||
1645                         need_lockbreak(details->i_mmap_lock);
1646
1647         if (restart_addr >= end_addr) {
1648                 /* We have now completed this vma: mark it so */
1649                 vma->vm_truncate_count = details->truncate_count;
1650                 if (!need_break)
1651                         return 0;
1652         } else {
1653                 /* Note restart_addr in vma's truncate_count field */
1654                 vma->vm_truncate_count = restart_addr;
1655                 if (!need_break)
1656                         goto again;
1657         }
1658
1659         spin_unlock(details->i_mmap_lock);
1660         cond_resched();
1661         spin_lock(details->i_mmap_lock);
1662         return -EINTR;
1663 }
1664
1665 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1666                                             struct zap_details *details)
1667 {
1668         struct vm_area_struct *vma;
1669         struct prio_tree_iter iter;
1670         pgoff_t vba, vea, zba, zea;
1671
1672 restart:
1673         vma_prio_tree_foreach(vma, &iter, root,
1674                         details->first_index, details->last_index) {
1675                 /* Skip quickly over those we have already dealt with */
1676                 if (vma->vm_truncate_count == details->truncate_count)
1677                         continue;
1678
1679                 vba = vma->vm_pgoff;
1680                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1681                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1682                 zba = details->first_index;
1683                 if (zba < vba)
1684                         zba = vba;
1685                 zea = details->last_index;
1686                 if (zea > vea)
1687                         zea = vea;
1688
1689                 if (unmap_mapping_range_vma(vma,
1690                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1691                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1692                                 details) < 0)
1693                         goto restart;
1694         }
1695 }
1696
1697 static inline void unmap_mapping_range_list(struct list_head *head,
1698                                             struct zap_details *details)
1699 {
1700         struct vm_area_struct *vma;
1701
1702         /*
1703          * In nonlinear VMAs there is no correspondence between virtual address
1704          * offset and file offset.  So we must perform an exhaustive search
1705          * across *all* the pages in each nonlinear VMA, not just the pages
1706          * whose virtual address lies outside the file truncation point.
1707          */
1708 restart:
1709         list_for_each_entry(vma, head, shared.vm_set.list) {
1710                 /* Skip quickly over those we have already dealt with */
1711                 if (vma->vm_truncate_count == details->truncate_count)
1712                         continue;
1713                 details->nonlinear_vma = vma;
1714                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1715                                         vma->vm_end, details) < 0)
1716                         goto restart;
1717         }
1718 }
1719
1720 /**
1721  * unmap_mapping_range - unmap the portion of all mmaps
1722  * in the specified address_space corresponding to the specified
1723  * page range in the underlying file.
1724  * @mapping: the address space containing mmaps to be unmapped.
1725  * @holebegin: byte in first page to unmap, relative to the start of
1726  * the underlying file.  This will be rounded down to a PAGE_SIZE
1727  * boundary.  Note that this is different from vmtruncate(), which
1728  * must keep the partial page.  In contrast, we must get rid of
1729  * partial pages.
1730  * @holelen: size of prospective hole in bytes.  This will be rounded
1731  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1732  * end of the file.
1733  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1734  * but 0 when invalidating pagecache, don't throw away private data.
1735  */
1736 void unmap_mapping_range(struct address_space *mapping,
1737                 loff_t const holebegin, loff_t const holelen, int even_cows)
1738 {
1739         struct zap_details details;
1740         pgoff_t hba = holebegin >> PAGE_SHIFT;
1741         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1742
1743         /* Check for overflow. */
1744         if (sizeof(holelen) > sizeof(hlen)) {
1745                 long long holeend =
1746                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1747                 if (holeend & ~(long long)ULONG_MAX)
1748                         hlen = ULONG_MAX - hba + 1;
1749         }
1750
1751         details.check_mapping = even_cows? NULL: mapping;
1752         details.nonlinear_vma = NULL;
1753         details.first_index = hba;
1754         details.last_index = hba + hlen - 1;
1755         if (details.last_index < details.first_index)
1756                 details.last_index = ULONG_MAX;
1757         details.i_mmap_lock = &mapping->i_mmap_lock;
1758
1759         spin_lock(&mapping->i_mmap_lock);
1760
1761         /* serialize i_size write against truncate_count write */
1762         smp_wmb();
1763         /* Protect against page faults, and endless unmapping loops */
1764         mapping->truncate_count++;
1765         /*
1766          * For archs where spin_lock has inclusive semantics like ia64
1767          * this smp_mb() will prevent to read pagetable contents
1768          * before the truncate_count increment is visible to
1769          * other cpus.
1770          */
1771         smp_mb();
1772         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1773                 if (mapping->truncate_count == 0)
1774                         reset_vma_truncate_counts(mapping);
1775                 mapping->truncate_count++;
1776         }
1777         details.truncate_count = mapping->truncate_count;
1778
1779         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1780                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1781         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1782                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1783         spin_unlock(&mapping->i_mmap_lock);
1784 }
1785 EXPORT_SYMBOL(unmap_mapping_range);
1786
1787 /*
1788  * Handle all mappings that got truncated by a "truncate()"
1789  * system call.
1790  *
1791  * NOTE! We have to be ready to update the memory sharing
1792  * between the file and the memory map for a potential last
1793  * incomplete page.  Ugly, but necessary.
1794  */
1795 int vmtruncate(struct inode * inode, loff_t offset)
1796 {
1797         struct address_space *mapping = inode->i_mapping;
1798         unsigned long limit;
1799
1800         if (inode->i_size < offset)
1801                 goto do_expand;
1802         /*
1803          * truncation of in-use swapfiles is disallowed - it would cause
1804          * subsequent swapout to scribble on the now-freed blocks.
1805          */
1806         if (IS_SWAPFILE(inode))
1807                 goto out_busy;
1808         i_size_write(inode, offset);
1809         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1810         truncate_inode_pages(mapping, offset);
1811         goto out_truncate;
1812
1813 do_expand:
1814         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1815         if (limit != RLIM_INFINITY && offset > limit)
1816                 goto out_sig;
1817         if (offset > inode->i_sb->s_maxbytes)
1818                 goto out_big;
1819         i_size_write(inode, offset);
1820
1821 out_truncate:
1822         if (inode->i_op && inode->i_op->truncate)
1823                 inode->i_op->truncate(inode);
1824         return 0;
1825 out_sig:
1826         send_sig(SIGXFSZ, current, 0);
1827 out_big:
1828         return -EFBIG;
1829 out_busy:
1830         return -ETXTBSY;
1831 }
1832 EXPORT_SYMBOL(vmtruncate);
1833
1834 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1835 {
1836         struct address_space *mapping = inode->i_mapping;
1837
1838         /*
1839          * If the underlying filesystem is not going to provide
1840          * a way to truncate a range of blocks (punch a hole) -
1841          * we should return failure right now.
1842          */
1843         if (!inode->i_op || !inode->i_op->truncate_range)
1844                 return -ENOSYS;
1845
1846         mutex_lock(&inode->i_mutex);
1847         down_write(&inode->i_alloc_sem);
1848         unmap_mapping_range(mapping, offset, (end - offset), 1);
1849         truncate_inode_pages_range(mapping, offset, end);
1850         inode->i_op->truncate_range(inode, offset, end);
1851         up_write(&inode->i_alloc_sem);
1852         mutex_unlock(&inode->i_mutex);
1853
1854         return 0;
1855 }
1856 EXPORT_UNUSED_SYMBOL(vmtruncate_range);  /*  June 2006  */
1857
1858 /* 
1859  * Primitive swap readahead code. We simply read an aligned block of
1860  * (1 << page_cluster) entries in the swap area. This method is chosen
1861  * because it doesn't cost us any seek time.  We also make sure to queue
1862  * the 'original' request together with the readahead ones...  
1863  *
1864  * This has been extended to use the NUMA policies from the mm triggering
1865  * the readahead.
1866  *
1867  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1868  */
1869 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1870 {
1871 #ifdef CONFIG_NUMA
1872         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1873 #endif
1874         int i, num;
1875         struct page *new_page;
1876         unsigned long offset;
1877
1878         /*
1879          * Get the number of handles we should do readahead io to.
1880          */
1881         num = valid_swaphandles(entry, &offset);
1882         for (i = 0; i < num; offset++, i++) {
1883                 /* Ok, do the async read-ahead now */
1884                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1885                                                            offset), vma, addr);
1886                 if (!new_page)
1887                         break;
1888                 page_cache_release(new_page);
1889 #ifdef CONFIG_NUMA
1890                 /*
1891                  * Find the next applicable VMA for the NUMA policy.
1892                  */
1893                 addr += PAGE_SIZE;
1894                 if (addr == 0)
1895                         vma = NULL;
1896                 if (vma) {
1897                         if (addr >= vma->vm_end) {
1898                                 vma = next_vma;
1899                                 next_vma = vma ? vma->vm_next : NULL;
1900                         }
1901                         if (vma && addr < vma->vm_start)
1902                                 vma = NULL;
1903                 } else {
1904                         if (next_vma && addr >= next_vma->vm_start) {
1905                                 vma = next_vma;
1906                                 next_vma = vma->vm_next;
1907                         }
1908                 }
1909 #endif
1910         }
1911         lru_add_drain();        /* Push any new pages onto the LRU now */
1912 }
1913
1914 /*
1915  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1916  * but allow concurrent faults), and pte mapped but not yet locked.
1917  * We return with mmap_sem still held, but pte unmapped and unlocked.
1918  */
1919 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1920                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1921                 int write_access, pte_t orig_pte)
1922 {
1923         spinlock_t *ptl;
1924         struct page *page;
1925         swp_entry_t entry;
1926         pte_t pte;
1927         int ret = VM_FAULT_MINOR;
1928
1929         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1930                 goto out;
1931
1932         entry = pte_to_swp_entry(orig_pte);
1933         if (is_migration_entry(entry)) {
1934                 migration_entry_wait(mm, pmd, address);
1935                 goto out;
1936         }
1937         page = lookup_swap_cache(entry);
1938         if (!page) {
1939                 swapin_readahead(entry, address, vma);
1940                 page = read_swap_cache_async(entry, vma, address);
1941                 if (!page) {
1942                         /*
1943                          * Back out if somebody else faulted in this pte
1944                          * while we released the pte lock.
1945                          */
1946                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1947                         if (likely(pte_same(*page_table, orig_pte)))
1948                                 ret = VM_FAULT_OOM;
1949                         goto unlock;
1950                 }
1951
1952                 /* Had to read the page from swap area: Major fault */
1953                 ret = VM_FAULT_MAJOR;
1954                 count_vm_event(PGMAJFAULT);
1955                 grab_swap_token();
1956         }
1957
1958         mark_page_accessed(page);
1959         lock_page(page);
1960
1961         /*
1962          * Back out if somebody else already faulted in this pte.
1963          */
1964         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1965         if (unlikely(!pte_same(*page_table, orig_pte)))
1966                 goto out_nomap;
1967
1968         if (unlikely(!PageUptodate(page))) {
1969                 ret = VM_FAULT_SIGBUS;
1970                 goto out_nomap;
1971         }
1972
1973         /* The page isn't present yet, go ahead with the fault. */
1974
1975         inc_mm_counter(mm, anon_rss);
1976         pte = mk_pte(page, vma->vm_page_prot);
1977         if (write_access && can_share_swap_page(page)) {
1978                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1979                 write_access = 0;
1980         }
1981
1982         flush_icache_page(vma, page);
1983         set_pte_at(mm, address, page_table, pte);
1984         page_add_anon_rmap(page, vma, address);
1985
1986         swap_free(entry);
1987         if (vm_swap_full())
1988                 remove_exclusive_swap_page(page);
1989         unlock_page(page);
1990
1991         if (write_access) {
1992                 if (do_wp_page(mm, vma, address,
1993                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1994                         ret = VM_FAULT_OOM;
1995                 goto out;
1996         }
1997
1998         /* No need to invalidate - it was non-present before */
1999         update_mmu_cache(vma, address, pte);
2000         lazy_mmu_prot_update(pte);
2001 unlock:
2002         pte_unmap_unlock(page_table, ptl);
2003 out:
2004         return ret;
2005 out_nomap:
2006         pte_unmap_unlock(page_table, ptl);
2007         unlock_page(page);
2008         page_cache_release(page);
2009         return ret;
2010 }
2011
2012 /*
2013  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2014  * but allow concurrent faults), and pte mapped but not yet locked.
2015  * We return with mmap_sem still held, but pte unmapped and unlocked.
2016  */
2017 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2018                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2019                 int write_access)
2020 {
2021         struct page *page;
2022         spinlock_t *ptl;
2023         pte_t entry;
2024
2025         if (write_access) {
2026                 /* Allocate our own private page. */
2027                 pte_unmap(page_table);
2028
2029                 if (unlikely(anon_vma_prepare(vma)))
2030                         goto oom;
2031                 page = alloc_zeroed_user_highpage(vma, address);
2032                 if (!page)
2033                         goto oom;
2034
2035                 entry = mk_pte(page, vma->vm_page_prot);
2036                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2037
2038                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2039                 if (!pte_none(*page_table))
2040                         goto release;
2041                 inc_mm_counter(mm, anon_rss);
2042                 lru_cache_add_active(page);
2043                 page_add_new_anon_rmap(page, vma, address);
2044         } else {
2045                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2046                 page = ZERO_PAGE(address);
2047                 page_cache_get(page);
2048                 entry = mk_pte(page, vma->vm_page_prot);
2049
2050                 ptl = pte_lockptr(mm, pmd);
2051                 spin_lock(ptl);
2052                 if (!pte_none(*page_table))
2053                         goto release;
2054                 inc_mm_counter(mm, file_rss);
2055                 page_add_file_rmap(page);
2056         }
2057
2058         set_pte_at(mm, address, page_table, entry);
2059
2060         /* No need to invalidate - it was non-present before */
2061         update_mmu_cache(vma, address, entry);
2062         lazy_mmu_prot_update(entry);
2063 unlock:
2064         pte_unmap_unlock(page_table, ptl);
2065         return VM_FAULT_MINOR;
2066 release:
2067         page_cache_release(page);
2068         goto unlock;
2069 oom:
2070         return VM_FAULT_OOM;
2071 }
2072
2073 /*
2074  * do_no_page() tries to create a new page mapping. It aggressively
2075  * tries to share with existing pages, but makes a separate copy if
2076  * the "write_access" parameter is true in order to avoid the next
2077  * page fault.
2078  *
2079  * As this is called only for pages that do not currently exist, we
2080  * do not need to flush old virtual caches or the TLB.
2081  *
2082  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2083  * but allow concurrent faults), and pte mapped but not yet locked.
2084  * We return with mmap_sem still held, but pte unmapped and unlocked.
2085  */
2086 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2087                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2088                 int write_access)
2089 {
2090         spinlock_t *ptl;
2091         struct page *new_page;
2092         struct address_space *mapping = NULL;
2093         pte_t entry;
2094         unsigned int sequence = 0;
2095         int ret = VM_FAULT_MINOR;
2096         int anon = 0;
2097
2098         pte_unmap(page_table);
2099         BUG_ON(vma->vm_flags & VM_PFNMAP);
2100
2101         if (vma->vm_file) {
2102                 mapping = vma->vm_file->f_mapping;
2103                 sequence = mapping->truncate_count;
2104                 smp_rmb(); /* serializes i_size against truncate_count */
2105         }
2106 retry:
2107         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2108         /*
2109          * No smp_rmb is needed here as long as there's a full
2110          * spin_lock/unlock sequence inside the ->nopage callback
2111          * (for the pagecache lookup) that acts as an implicit
2112          * smp_mb() and prevents the i_size read to happen
2113          * after the next truncate_count read.
2114          */
2115
2116         /* no page was available -- either SIGBUS or OOM */
2117         if (new_page == NOPAGE_SIGBUS)
2118                 return VM_FAULT_SIGBUS;
2119         if (new_page == NOPAGE_OOM)
2120                 return VM_FAULT_OOM;
2121
2122         /*
2123          * Should we do an early C-O-W break?
2124          */
2125         if (write_access) {
2126                 if (!(vma->vm_flags & VM_SHARED)) {
2127                         struct page *page;
2128
2129                         if (unlikely(anon_vma_prepare(vma)))
2130                                 goto oom;
2131                         page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2132                         if (!page)
2133                                 goto oom;
2134                         copy_user_highpage(page, new_page, address);
2135                         page_cache_release(new_page);
2136                         new_page = page;
2137                         anon = 1;
2138
2139                 } else {
2140                         /* if the page will be shareable, see if the backing
2141                          * address space wants to know that the page is about
2142                          * to become writable */
2143                         if (vma->vm_ops->page_mkwrite &&
2144                             vma->vm_ops->page_mkwrite(vma, new_page) < 0
2145                             ) {
2146                                 page_cache_release(new_page);
2147                                 return VM_FAULT_SIGBUS;
2148                         }
2149                 }
2150         }
2151
2152         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2153         /*
2154          * For a file-backed vma, someone could have truncated or otherwise
2155          * invalidated this page.  If unmap_mapping_range got called,
2156          * retry getting the page.
2157          */
2158         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2159                 pte_unmap_unlock(page_table, ptl);
2160                 page_cache_release(new_page);
2161                 cond_resched();
2162                 sequence = mapping->truncate_count;
2163                 smp_rmb();
2164                 goto retry;
2165         }
2166
2167         /*
2168          * This silly early PAGE_DIRTY setting removes a race
2169          * due to the bad i386 page protection. But it's valid
2170          * for other architectures too.
2171          *
2172          * Note that if write_access is true, we either now have
2173          * an exclusive copy of the page, or this is a shared mapping,
2174          * so we can make it writable and dirty to avoid having to
2175          * handle that later.
2176          */
2177         /* Only go through if we didn't race with anybody else... */
2178         if (pte_none(*page_table)) {
2179                 flush_icache_page(vma, new_page);
2180                 entry = mk_pte(new_page, vma->vm_page_prot);
2181                 if (write_access)
2182                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2183                 set_pte_at(mm, address, page_table, entry);
2184                 if (anon) {
2185                         inc_mm_counter(mm, anon_rss);
2186                         lru_cache_add_active(new_page);
2187                         page_add_new_anon_rmap(new_page, vma, address);
2188                 } else {
2189                         inc_mm_counter(mm, file_rss);
2190                         page_add_file_rmap(new_page);
2191                 }
2192         } else {
2193                 /* One of our sibling threads was faster, back out. */
2194                 page_cache_release(new_page);
2195                 goto unlock;
2196         }
2197
2198         /* no need to invalidate: a not-present page shouldn't be cached */
2199         update_mmu_cache(vma, address, entry);
2200         lazy_mmu_prot_update(entry);
2201 unlock:
2202         pte_unmap_unlock(page_table, ptl);
2203         return ret;
2204 oom:
2205         page_cache_release(new_page);
2206         return VM_FAULT_OOM;
2207 }
2208
2209 /*
2210  * Fault of a previously existing named mapping. Repopulate the pte
2211  * from the encoded file_pte if possible. This enables swappable
2212  * nonlinear vmas.
2213  *
2214  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2215  * but allow concurrent faults), and pte mapped but not yet locked.
2216  * We return with mmap_sem still held, but pte unmapped and unlocked.
2217  */
2218 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2219                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2220                 int write_access, pte_t orig_pte)
2221 {
2222         pgoff_t pgoff;
2223         int err;
2224
2225         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2226                 return VM_FAULT_MINOR;
2227
2228         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2229                 /*
2230                  * Page table corrupted: show pte and kill process.
2231                  */
2232                 print_bad_pte(vma, orig_pte, address);
2233                 return VM_FAULT_OOM;
2234         }
2235         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2236
2237         pgoff = pte_to_pgoff(orig_pte);
2238         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2239                                         vma->vm_page_prot, pgoff, 0);
2240         if (err == -ENOMEM)
2241                 return VM_FAULT_OOM;
2242         if (err)
2243                 return VM_FAULT_SIGBUS;
2244         return VM_FAULT_MAJOR;
2245 }
2246
2247 /*
2248  * These routines also need to handle stuff like marking pages dirty
2249  * and/or accessed for architectures that don't do it in hardware (most
2250  * RISC architectures).  The early dirtying is also good on the i386.
2251  *
2252  * There is also a hook called "update_mmu_cache()" that architectures
2253  * with external mmu caches can use to update those (ie the Sparc or
2254  * PowerPC hashed page tables that act as extended TLBs).
2255  *
2256  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2257  * but allow concurrent faults), and pte mapped but not yet locked.
2258  * We return with mmap_sem still held, but pte unmapped and unlocked.
2259  */
2260 static inline int handle_pte_fault(struct mm_struct *mm,
2261                 struct vm_area_struct *vma, unsigned long address,
2262                 pte_t *pte, pmd_t *pmd, int write_access)
2263 {
2264         pte_t entry;
2265         pte_t old_entry;
2266         spinlock_t *ptl;
2267
2268         old_entry = entry = *pte;
2269         if (!pte_present(entry)) {
2270                 if (pte_none(entry)) {
2271                         if (!vma->vm_ops || !vma->vm_ops->nopage)
2272                                 return do_anonymous_page(mm, vma, address,
2273                                         pte, pmd, write_access);
2274                         return do_no_page(mm, vma, address,
2275                                         pte, pmd, write_access);
2276                 }
2277                 if (pte_file(entry))
2278                         return do_file_page(mm, vma, address,
2279                                         pte, pmd, write_access, entry);
2280                 return do_swap_page(mm, vma, address,
2281                                         pte, pmd, write_access, entry);
2282         }
2283
2284         ptl = pte_lockptr(mm, pmd);
2285         spin_lock(ptl);
2286         if (unlikely(!pte_same(*pte, entry)))
2287                 goto unlock;
2288         if (write_access) {
2289                 if (!pte_write(entry))
2290                         return do_wp_page(mm, vma, address,
2291                                         pte, pmd, ptl, entry);
2292                 entry = pte_mkdirty(entry);
2293         }
2294         entry = pte_mkyoung(entry);
2295         if (!pte_same(old_entry, entry)) {
2296                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2297                 update_mmu_cache(vma, address, entry);
2298                 lazy_mmu_prot_update(entry);
2299         } else {
2300                 /*
2301                  * This is needed only for protection faults but the arch code
2302                  * is not yet telling us if this is a protection fault or not.
2303                  * This still avoids useless tlb flushes for .text page faults
2304                  * with threads.
2305                  */
2306                 if (write_access)
2307                         flush_tlb_page(vma, address);
2308         }
2309 unlock:
2310         pte_unmap_unlock(pte, ptl);
2311         return VM_FAULT_MINOR;
2312 }
2313
2314 /*
2315  * By the time we get here, we already hold the mm semaphore
2316  */
2317 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2318                 unsigned long address, int write_access)
2319 {
2320         pgd_t *pgd;
2321         pud_t *pud;
2322         pmd_t *pmd;
2323         pte_t *pte;
2324
2325         __set_current_state(TASK_RUNNING);
2326
2327         count_vm_event(PGFAULT);
2328
2329         if (unlikely(is_vm_hugetlb_page(vma)))
2330                 return hugetlb_fault(mm, vma, address, write_access);
2331
2332         pgd = pgd_offset(mm, address);
2333         pud = pud_alloc(mm, pgd, address);
2334         if (!pud)
2335                 return VM_FAULT_OOM;
2336         pmd = pmd_alloc(mm, pud, address);
2337         if (!pmd)
2338                 return VM_FAULT_OOM;
2339         pte = pte_alloc_map(mm, pmd, address);
2340         if (!pte)
2341                 return VM_FAULT_OOM;
2342
2343         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2344 }
2345
2346 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2347
2348 #ifndef __PAGETABLE_PUD_FOLDED
2349 /*
2350  * Allocate page upper directory.
2351  * We've already handled the fast-path in-line.
2352  */
2353 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2354 {
2355         pud_t *new = pud_alloc_one(mm, address);
2356         if (!new)
2357                 return -ENOMEM;
2358
2359         spin_lock(&mm->page_table_lock);
2360         if (pgd_present(*pgd))          /* Another has populated it */
2361                 pud_free(new);
2362         else
2363                 pgd_populate(mm, pgd, new);
2364         spin_unlock(&mm->page_table_lock);
2365         return 0;
2366 }
2367 #else
2368 /* Workaround for gcc 2.96 */
2369 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2370 {
2371         return 0;
2372 }
2373 #endif /* __PAGETABLE_PUD_FOLDED */
2374
2375 #ifndef __PAGETABLE_PMD_FOLDED
2376 /*
2377  * Allocate page middle directory.
2378  * We've already handled the fast-path in-line.
2379  */
2380 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2381 {
2382         pmd_t *new = pmd_alloc_one(mm, address);
2383         if (!new)
2384                 return -ENOMEM;
2385
2386         spin_lock(&mm->page_table_lock);
2387 #ifndef __ARCH_HAS_4LEVEL_HACK
2388         if (pud_present(*pud))          /* Another has populated it */
2389                 pmd_free(new);
2390         else
2391                 pud_populate(mm, pud, new);
2392 #else
2393         if (pgd_present(*pud))          /* Another has populated it */
2394                 pmd_free(new);
2395         else
2396                 pgd_populate(mm, pud, new);
2397 #endif /* __ARCH_HAS_4LEVEL_HACK */
2398         spin_unlock(&mm->page_table_lock);
2399         return 0;
2400 }
2401 #else
2402 /* Workaround for gcc 2.96 */
2403 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2404 {
2405         return 0;
2406 }
2407 #endif /* __PAGETABLE_PMD_FOLDED */
2408
2409 int make_pages_present(unsigned long addr, unsigned long end)
2410 {
2411         int ret, len, write;
2412         struct vm_area_struct * vma;
2413
2414         vma = find_vma(current->mm, addr);
2415         if (!vma)
2416                 return -1;
2417         write = (vma->vm_flags & VM_WRITE) != 0;
2418         BUG_ON(addr >= end);
2419         BUG_ON(end > vma->vm_end);
2420         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2421         ret = get_user_pages(current, current->mm, addr,
2422                         len, write, 0, NULL, NULL);
2423         if (ret < 0)
2424                 return ret;
2425         return ret == len ? 0 : -1;
2426 }
2427
2428 /* 
2429  * Map a vmalloc()-space virtual address to the physical page.
2430  */
2431 struct page * vmalloc_to_page(void * vmalloc_addr)
2432 {
2433         unsigned long addr = (unsigned long) vmalloc_addr;
2434         struct page *page = NULL;
2435         pgd_t *pgd = pgd_offset_k(addr);
2436         pud_t *pud;
2437         pmd_t *pmd;
2438         pte_t *ptep, pte;
2439   
2440         if (!pgd_none(*pgd)) {
2441                 pud = pud_offset(pgd, addr);
2442                 if (!pud_none(*pud)) {
2443                         pmd = pmd_offset(pud, addr);
2444                         if (!pmd_none(*pmd)) {
2445                                 ptep = pte_offset_map(pmd, addr);
2446                                 pte = *ptep;
2447                                 if (pte_present(pte))
2448                                         page = pte_page(pte);
2449                                 pte_unmap(ptep);
2450                         }
2451                 }
2452         }
2453         return page;
2454 }
2455
2456 EXPORT_SYMBOL(vmalloc_to_page);
2457
2458 /*
2459  * Map a vmalloc()-space virtual address to the physical page frame number.
2460  */
2461 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2462 {
2463         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2464 }
2465
2466 EXPORT_SYMBOL(vmalloc_to_pfn);
2467
2468 #if !defined(__HAVE_ARCH_GATE_AREA)
2469
2470 #if defined(AT_SYSINFO_EHDR)
2471 static struct vm_area_struct gate_vma;
2472
2473 static int __init gate_vma_init(void)
2474 {
2475         gate_vma.vm_mm = NULL;
2476         gate_vma.vm_start = FIXADDR_USER_START;
2477         gate_vma.vm_end = FIXADDR_USER_END;
2478         gate_vma.vm_page_prot = PAGE_READONLY;
2479         gate_vma.vm_flags = 0;
2480         return 0;
2481 }
2482 __initcall(gate_vma_init);
2483 #endif
2484
2485 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2486 {
2487 #ifdef AT_SYSINFO_EHDR
2488         return &gate_vma;
2489 #else
2490         return NULL;
2491 #endif
2492 }
2493
2494 int in_gate_area_no_task(unsigned long addr)
2495 {
2496 #ifdef AT_SYSINFO_EHDR
2497         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2498                 return 1;
2499 #endif
2500         return 0;
2501 }
2502
2503 #endif  /* __HAVE_ARCH_GATE_AREA */