[PATCH] mm: copy_one_pte inc rss
[linux-2.6] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/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 /*
86  * If a p?d_bad entry is found while walking page tables, report
87  * the error, before resetting entry to p?d_none.  Usually (but
88  * very seldom) called out from the p?d_none_or_clear_bad macros.
89  */
90
91 void pgd_clear_bad(pgd_t *pgd)
92 {
93         pgd_ERROR(*pgd);
94         pgd_clear(pgd);
95 }
96
97 void pud_clear_bad(pud_t *pud)
98 {
99         pud_ERROR(*pud);
100         pud_clear(pud);
101 }
102
103 void pmd_clear_bad(pmd_t *pmd)
104 {
105         pmd_ERROR(*pmd);
106         pmd_clear(pmd);
107 }
108
109 /*
110  * Note: this doesn't free the actual pages themselves. That
111  * has been handled earlier when unmapping all the memory regions.
112  */
113 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
114 {
115         struct page *page = pmd_page(*pmd);
116         pmd_clear(pmd);
117         pte_free_tlb(tlb, page);
118         dec_page_state(nr_page_table_pages);
119         tlb->mm->nr_ptes--;
120 }
121
122 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
123                                 unsigned long addr, unsigned long end,
124                                 unsigned long floor, unsigned long ceiling)
125 {
126         pmd_t *pmd;
127         unsigned long next;
128         unsigned long start;
129
130         start = addr;
131         pmd = pmd_offset(pud, addr);
132         do {
133                 next = pmd_addr_end(addr, end);
134                 if (pmd_none_or_clear_bad(pmd))
135                         continue;
136                 free_pte_range(tlb, pmd);
137         } while (pmd++, addr = next, addr != end);
138
139         start &= PUD_MASK;
140         if (start < floor)
141                 return;
142         if (ceiling) {
143                 ceiling &= PUD_MASK;
144                 if (!ceiling)
145                         return;
146         }
147         if (end - 1 > ceiling - 1)
148                 return;
149
150         pmd = pmd_offset(pud, start);
151         pud_clear(pud);
152         pmd_free_tlb(tlb, pmd);
153 }
154
155 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
156                                 unsigned long addr, unsigned long end,
157                                 unsigned long floor, unsigned long ceiling)
158 {
159         pud_t *pud;
160         unsigned long next;
161         unsigned long start;
162
163         start = addr;
164         pud = pud_offset(pgd, addr);
165         do {
166                 next = pud_addr_end(addr, end);
167                 if (pud_none_or_clear_bad(pud))
168                         continue;
169                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
170         } while (pud++, addr = next, addr != end);
171
172         start &= PGDIR_MASK;
173         if (start < floor)
174                 return;
175         if (ceiling) {
176                 ceiling &= PGDIR_MASK;
177                 if (!ceiling)
178                         return;
179         }
180         if (end - 1 > ceiling - 1)
181                 return;
182
183         pud = pud_offset(pgd, start);
184         pgd_clear(pgd);
185         pud_free_tlb(tlb, pud);
186 }
187
188 /*
189  * This function frees user-level page tables of a process.
190  *
191  * Must be called with pagetable lock held.
192  */
193 void free_pgd_range(struct mmu_gather **tlb,
194                         unsigned long addr, unsigned long end,
195                         unsigned long floor, unsigned long ceiling)
196 {
197         pgd_t *pgd;
198         unsigned long next;
199         unsigned long start;
200
201         /*
202          * The next few lines have given us lots of grief...
203          *
204          * Why are we testing PMD* at this top level?  Because often
205          * there will be no work to do at all, and we'd prefer not to
206          * go all the way down to the bottom just to discover that.
207          *
208          * Why all these "- 1"s?  Because 0 represents both the bottom
209          * of the address space and the top of it (using -1 for the
210          * top wouldn't help much: the masks would do the wrong thing).
211          * The rule is that addr 0 and floor 0 refer to the bottom of
212          * the address space, but end 0 and ceiling 0 refer to the top
213          * Comparisons need to use "end - 1" and "ceiling - 1" (though
214          * that end 0 case should be mythical).
215          *
216          * Wherever addr is brought up or ceiling brought down, we must
217          * be careful to reject "the opposite 0" before it confuses the
218          * subsequent tests.  But what about where end is brought down
219          * by PMD_SIZE below? no, end can't go down to 0 there.
220          *
221          * Whereas we round start (addr) and ceiling down, by different
222          * masks at different levels, in order to test whether a table
223          * now has no other vmas using it, so can be freed, we don't
224          * bother to round floor or end up - the tests don't need that.
225          */
226
227         addr &= PMD_MASK;
228         if (addr < floor) {
229                 addr += PMD_SIZE;
230                 if (!addr)
231                         return;
232         }
233         if (ceiling) {
234                 ceiling &= PMD_MASK;
235                 if (!ceiling)
236                         return;
237         }
238         if (end - 1 > ceiling - 1)
239                 end -= PMD_SIZE;
240         if (addr > end - 1)
241                 return;
242
243         start = addr;
244         pgd = pgd_offset((*tlb)->mm, addr);
245         do {
246                 next = pgd_addr_end(addr, end);
247                 if (pgd_none_or_clear_bad(pgd))
248                         continue;
249                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
250         } while (pgd++, addr = next, addr != end);
251
252         if (!(*tlb)->fullmm)
253                 flush_tlb_pgtables((*tlb)->mm, start, end);
254 }
255
256 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
257                 unsigned long floor, unsigned long ceiling)
258 {
259         while (vma) {
260                 struct vm_area_struct *next = vma->vm_next;
261                 unsigned long addr = vma->vm_start;
262
263                 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
264                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
265                                 floor, next? next->vm_start: ceiling);
266                 } else {
267                         /*
268                          * Optimization: gather nearby vmas into one call down
269                          */
270                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
271                           && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
272                                                         HPAGE_SIZE)) {
273                                 vma = next;
274                                 next = vma->vm_next;
275                         }
276                         free_pgd_range(tlb, addr, vma->vm_end,
277                                 floor, next? next->vm_start: ceiling);
278                 }
279                 vma = next;
280         }
281 }
282
283 pte_t fastcall *pte_alloc_map(struct mm_struct *mm, pmd_t *pmd,
284                                 unsigned long address)
285 {
286         if (!pmd_present(*pmd)) {
287                 struct page *new;
288
289                 spin_unlock(&mm->page_table_lock);
290                 new = pte_alloc_one(mm, address);
291                 spin_lock(&mm->page_table_lock);
292                 if (!new)
293                         return NULL;
294                 /*
295                  * Because we dropped the lock, we should re-check the
296                  * entry, as somebody else could have populated it..
297                  */
298                 if (pmd_present(*pmd)) {
299                         pte_free(new);
300                         goto out;
301                 }
302                 mm->nr_ptes++;
303                 inc_page_state(nr_page_table_pages);
304                 pmd_populate(mm, pmd, new);
305         }
306 out:
307         return pte_offset_map(pmd, address);
308 }
309
310 pte_t fastcall * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
311 {
312         if (!pmd_present(*pmd)) {
313                 pte_t *new;
314
315                 spin_unlock(&mm->page_table_lock);
316                 new = pte_alloc_one_kernel(mm, address);
317                 spin_lock(&mm->page_table_lock);
318                 if (!new)
319                         return NULL;
320
321                 /*
322                  * Because we dropped the lock, we should re-check the
323                  * entry, as somebody else could have populated it..
324                  */
325                 if (pmd_present(*pmd)) {
326                         pte_free_kernel(new);
327                         goto out;
328                 }
329                 pmd_populate_kernel(mm, pmd, new);
330         }
331 out:
332         return pte_offset_kernel(pmd, address);
333 }
334
335 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
336 {
337         if (file_rss)
338                 add_mm_counter(mm, file_rss, file_rss);
339         if (anon_rss)
340                 add_mm_counter(mm, anon_rss, anon_rss);
341 }
342
343 /*
344  * This function is called to print an error when a pte in a
345  * !VM_RESERVED region is found pointing to an invalid pfn (which
346  * is an error.
347  *
348  * The calling function must still handle the error.
349  */
350 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
351 {
352         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
353                         "vm_flags = %lx, vaddr = %lx\n",
354                 (long long)pte_val(pte),
355                 (vma->vm_mm == current->mm ? current->comm : "???"),
356                 vma->vm_flags, vaddr);
357         dump_stack();
358 }
359
360 /*
361  * copy one vm_area from one task to the other. Assumes the page tables
362  * already present in the new task to be cleared in the whole range
363  * covered by this vma.
364  *
365  * dst->page_table_lock is held on entry and exit,
366  * but may be dropped within p[mg]d_alloc() and pte_alloc_map().
367  */
368
369 static inline void
370 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
371                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
372                 unsigned long addr, int *rss)
373 {
374         unsigned long vm_flags = vma->vm_flags;
375         pte_t pte = *src_pte;
376         struct page *page;
377         unsigned long pfn;
378
379         /* pte contains position in swap or file, so copy. */
380         if (unlikely(!pte_present(pte))) {
381                 if (!pte_file(pte)) {
382                         swap_duplicate(pte_to_swp_entry(pte));
383                         /* make sure dst_mm is on swapoff's mmlist. */
384                         if (unlikely(list_empty(&dst_mm->mmlist))) {
385                                 spin_lock(&mmlist_lock);
386                                 list_add(&dst_mm->mmlist, &src_mm->mmlist);
387                                 spin_unlock(&mmlist_lock);
388                         }
389                 }
390                 goto out_set_pte;
391         }
392
393         /* If the region is VM_RESERVED, the mapping is not
394          * mapped via rmap - duplicate the pte as is.
395          */
396         if (vm_flags & VM_RESERVED)
397                 goto out_set_pte;
398
399         pfn = pte_pfn(pte);
400         /* If the pte points outside of valid memory but
401          * the region is not VM_RESERVED, we have a problem.
402          */
403         if (unlikely(!pfn_valid(pfn))) {
404                 print_bad_pte(vma, pte, addr);
405                 goto out_set_pte; /* try to do something sane */
406         }
407
408         page = pfn_to_page(pfn);
409
410         /*
411          * If it's a COW mapping, write protect it both
412          * in the parent and the child
413          */
414         if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
415                 ptep_set_wrprotect(src_mm, addr, src_pte);
416                 pte = *src_pte;
417         }
418
419         /*
420          * If it's a shared mapping, mark it clean in
421          * the child
422          */
423         if (vm_flags & VM_SHARED)
424                 pte = pte_mkclean(pte);
425         pte = pte_mkold(pte);
426         get_page(page);
427         page_dup_rmap(page);
428         rss[!!PageAnon(page)]++;
429
430 out_set_pte:
431         set_pte_at(dst_mm, addr, dst_pte, pte);
432 }
433
434 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
435                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
436                 unsigned long addr, unsigned long end)
437 {
438         pte_t *src_pte, *dst_pte;
439         int progress = 0;
440         int rss[2];
441
442 again:
443         rss[1] = rss[0] = 0;
444         dst_pte = pte_alloc_map(dst_mm, dst_pmd, addr);
445         if (!dst_pte)
446                 return -ENOMEM;
447         src_pte = pte_offset_map_nested(src_pmd, addr);
448
449         spin_lock(&src_mm->page_table_lock);
450         do {
451                 /*
452                  * We are holding two locks at this point - either of them
453                  * could generate latencies in another task on another CPU.
454                  */
455                 if (progress >= 32) {
456                         progress = 0;
457                         if (need_resched() ||
458                             need_lockbreak(&src_mm->page_table_lock) ||
459                             need_lockbreak(&dst_mm->page_table_lock))
460                                 break;
461                 }
462                 if (pte_none(*src_pte)) {
463                         progress++;
464                         continue;
465                 }
466                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
467                 progress += 8;
468         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
469         spin_unlock(&src_mm->page_table_lock);
470
471         pte_unmap_nested(src_pte - 1);
472         pte_unmap(dst_pte - 1);
473         add_mm_rss(dst_mm, rss[0], rss[1]);
474         cond_resched_lock(&dst_mm->page_table_lock);
475         if (addr != end)
476                 goto again;
477         return 0;
478 }
479
480 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
481                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
482                 unsigned long addr, unsigned long end)
483 {
484         pmd_t *src_pmd, *dst_pmd;
485         unsigned long next;
486
487         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
488         if (!dst_pmd)
489                 return -ENOMEM;
490         src_pmd = pmd_offset(src_pud, addr);
491         do {
492                 next = pmd_addr_end(addr, end);
493                 if (pmd_none_or_clear_bad(src_pmd))
494                         continue;
495                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
496                                                 vma, addr, next))
497                         return -ENOMEM;
498         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
499         return 0;
500 }
501
502 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
503                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
504                 unsigned long addr, unsigned long end)
505 {
506         pud_t *src_pud, *dst_pud;
507         unsigned long next;
508
509         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
510         if (!dst_pud)
511                 return -ENOMEM;
512         src_pud = pud_offset(src_pgd, addr);
513         do {
514                 next = pud_addr_end(addr, end);
515                 if (pud_none_or_clear_bad(src_pud))
516                         continue;
517                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
518                                                 vma, addr, next))
519                         return -ENOMEM;
520         } while (dst_pud++, src_pud++, addr = next, addr != end);
521         return 0;
522 }
523
524 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
525                 struct vm_area_struct *vma)
526 {
527         pgd_t *src_pgd, *dst_pgd;
528         unsigned long next;
529         unsigned long addr = vma->vm_start;
530         unsigned long end = vma->vm_end;
531
532         /*
533          * Don't copy ptes where a page fault will fill them correctly.
534          * Fork becomes much lighter when there are big shared or private
535          * readonly mappings. The tradeoff is that copy_page_range is more
536          * efficient than faulting.
537          */
538         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_RESERVED))) {
539                 if (!vma->anon_vma)
540                         return 0;
541         }
542
543         if (is_vm_hugetlb_page(vma))
544                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
545
546         dst_pgd = pgd_offset(dst_mm, addr);
547         src_pgd = pgd_offset(src_mm, addr);
548         do {
549                 next = pgd_addr_end(addr, end);
550                 if (pgd_none_or_clear_bad(src_pgd))
551                         continue;
552                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
553                                                 vma, addr, next))
554                         return -ENOMEM;
555         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
556         return 0;
557 }
558
559 static void zap_pte_range(struct mmu_gather *tlb,
560                                 struct vm_area_struct *vma, pmd_t *pmd,
561                                 unsigned long addr, unsigned long end,
562                                 struct zap_details *details)
563 {
564         struct mm_struct *mm = tlb->mm;
565         pte_t *pte;
566         int file_rss = 0;
567         int anon_rss = 0;
568
569         pte = pte_offset_map(pmd, addr);
570         do {
571                 pte_t ptent = *pte;
572                 if (pte_none(ptent))
573                         continue;
574                 if (pte_present(ptent)) {
575                         struct page *page = NULL;
576                         if (!(vma->vm_flags & VM_RESERVED)) {
577                                 unsigned long pfn = pte_pfn(ptent);
578                                 if (unlikely(!pfn_valid(pfn)))
579                                         print_bad_pte(vma, ptent, addr);
580                                 else
581                                         page = pfn_to_page(pfn);
582                         }
583                         if (unlikely(details) && page) {
584                                 /*
585                                  * unmap_shared_mapping_pages() wants to
586                                  * invalidate cache without truncating:
587                                  * unmap shared but keep private pages.
588                                  */
589                                 if (details->check_mapping &&
590                                     details->check_mapping != page->mapping)
591                                         continue;
592                                 /*
593                                  * Each page->index must be checked when
594                                  * invalidating or truncating nonlinear.
595                                  */
596                                 if (details->nonlinear_vma &&
597                                     (page->index < details->first_index ||
598                                      page->index > details->last_index))
599                                         continue;
600                         }
601                         ptent = ptep_get_and_clear_full(mm, addr, pte,
602                                                         tlb->fullmm);
603                         tlb_remove_tlb_entry(tlb, pte, addr);
604                         if (unlikely(!page))
605                                 continue;
606                         if (unlikely(details) && details->nonlinear_vma
607                             && linear_page_index(details->nonlinear_vma,
608                                                 addr) != page->index)
609                                 set_pte_at(mm, addr, pte,
610                                            pgoff_to_pte(page->index));
611                         if (PageAnon(page))
612                                 anon_rss++;
613                         else {
614                                 if (pte_dirty(ptent))
615                                         set_page_dirty(page);
616                                 if (pte_young(ptent))
617                                         mark_page_accessed(page);
618                                 file_rss++;
619                         }
620                         page_remove_rmap(page);
621                         tlb_remove_page(tlb, page);
622                         continue;
623                 }
624                 /*
625                  * If details->check_mapping, we leave swap entries;
626                  * if details->nonlinear_vma, we leave file entries.
627                  */
628                 if (unlikely(details))
629                         continue;
630                 if (!pte_file(ptent))
631                         free_swap_and_cache(pte_to_swp_entry(ptent));
632                 pte_clear_full(mm, addr, pte, tlb->fullmm);
633         } while (pte++, addr += PAGE_SIZE, addr != end);
634
635         add_mm_rss(mm, -file_rss, -anon_rss);
636         pte_unmap(pte - 1);
637 }
638
639 static inline void zap_pmd_range(struct mmu_gather *tlb,
640                                 struct vm_area_struct *vma, pud_t *pud,
641                                 unsigned long addr, unsigned long end,
642                                 struct zap_details *details)
643 {
644         pmd_t *pmd;
645         unsigned long next;
646
647         pmd = pmd_offset(pud, addr);
648         do {
649                 next = pmd_addr_end(addr, end);
650                 if (pmd_none_or_clear_bad(pmd))
651                         continue;
652                 zap_pte_range(tlb, vma, pmd, addr, next, details);
653         } while (pmd++, addr = next, addr != end);
654 }
655
656 static inline void zap_pud_range(struct mmu_gather *tlb,
657                                 struct vm_area_struct *vma, pgd_t *pgd,
658                                 unsigned long addr, unsigned long end,
659                                 struct zap_details *details)
660 {
661         pud_t *pud;
662         unsigned long next;
663
664         pud = pud_offset(pgd, addr);
665         do {
666                 next = pud_addr_end(addr, end);
667                 if (pud_none_or_clear_bad(pud))
668                         continue;
669                 zap_pmd_range(tlb, vma, pud, addr, next, details);
670         } while (pud++, addr = next, addr != end);
671 }
672
673 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
674                                 unsigned long addr, unsigned long end,
675                                 struct zap_details *details)
676 {
677         pgd_t *pgd;
678         unsigned long next;
679
680         if (details && !details->check_mapping && !details->nonlinear_vma)
681                 details = NULL;
682
683         BUG_ON(addr >= end);
684         tlb_start_vma(tlb, vma);
685         pgd = pgd_offset(vma->vm_mm, addr);
686         do {
687                 next = pgd_addr_end(addr, end);
688                 if (pgd_none_or_clear_bad(pgd))
689                         continue;
690                 zap_pud_range(tlb, vma, pgd, addr, next, details);
691         } while (pgd++, addr = next, addr != end);
692         tlb_end_vma(tlb, vma);
693 }
694
695 #ifdef CONFIG_PREEMPT
696 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
697 #else
698 /* No preempt: go for improved straight-line efficiency */
699 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
700 #endif
701
702 /**
703  * unmap_vmas - unmap a range of memory covered by a list of vma's
704  * @tlbp: address of the caller's struct mmu_gather
705  * @mm: the controlling mm_struct
706  * @vma: the starting vma
707  * @start_addr: virtual address at which to start unmapping
708  * @end_addr: virtual address at which to end unmapping
709  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
710  * @details: details of nonlinear truncation or shared cache invalidation
711  *
712  * Returns the end address of the unmapping (restart addr if interrupted).
713  *
714  * Unmap all pages in the vma list.  Called under page_table_lock.
715  *
716  * We aim to not hold page_table_lock for too long (for scheduling latency
717  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
718  * return the ending mmu_gather to the caller.
719  *
720  * Only addresses between `start' and `end' will be unmapped.
721  *
722  * The VMA list must be sorted in ascending virtual address order.
723  *
724  * unmap_vmas() assumes that the caller will flush the whole unmapped address
725  * range after unmap_vmas() returns.  So the only responsibility here is to
726  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
727  * drops the lock and schedules.
728  */
729 unsigned long unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
730                 struct vm_area_struct *vma, unsigned long start_addr,
731                 unsigned long end_addr, unsigned long *nr_accounted,
732                 struct zap_details *details)
733 {
734         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
735         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
736         int tlb_start_valid = 0;
737         unsigned long start = start_addr;
738         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
739         int fullmm = (*tlbp)->fullmm;
740
741         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
742                 unsigned long end;
743
744                 start = max(vma->vm_start, start_addr);
745                 if (start >= vma->vm_end)
746                         continue;
747                 end = min(vma->vm_end, end_addr);
748                 if (end <= vma->vm_start)
749                         continue;
750
751                 if (vma->vm_flags & VM_ACCOUNT)
752                         *nr_accounted += (end - start) >> PAGE_SHIFT;
753
754                 while (start != end) {
755                         unsigned long block;
756
757                         if (!tlb_start_valid) {
758                                 tlb_start = start;
759                                 tlb_start_valid = 1;
760                         }
761
762                         if (is_vm_hugetlb_page(vma)) {
763                                 block = end - start;
764                                 unmap_hugepage_range(vma, start, end);
765                         } else {
766                                 block = min(zap_bytes, end - start);
767                                 unmap_page_range(*tlbp, vma, start,
768                                                 start + block, details);
769                         }
770
771                         start += block;
772                         zap_bytes -= block;
773                         if ((long)zap_bytes > 0)
774                                 continue;
775
776                         tlb_finish_mmu(*tlbp, tlb_start, start);
777
778                         if (need_resched() ||
779                                 need_lockbreak(&mm->page_table_lock) ||
780                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
781                                 if (i_mmap_lock) {
782                                         /* must reset count of rss freed */
783                                         *tlbp = tlb_gather_mmu(mm, fullmm);
784                                         goto out;
785                                 }
786                                 spin_unlock(&mm->page_table_lock);
787                                 cond_resched();
788                                 spin_lock(&mm->page_table_lock);
789                         }
790
791                         *tlbp = tlb_gather_mmu(mm, fullmm);
792                         tlb_start_valid = 0;
793                         zap_bytes = ZAP_BLOCK_SIZE;
794                 }
795         }
796 out:
797         return start;   /* which is now the end (or restart) address */
798 }
799
800 /**
801  * zap_page_range - remove user pages in a given range
802  * @vma: vm_area_struct holding the applicable pages
803  * @address: starting address of pages to zap
804  * @size: number of bytes to zap
805  * @details: details of nonlinear truncation or shared cache invalidation
806  */
807 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
808                 unsigned long size, struct zap_details *details)
809 {
810         struct mm_struct *mm = vma->vm_mm;
811         struct mmu_gather *tlb;
812         unsigned long end = address + size;
813         unsigned long nr_accounted = 0;
814
815         if (is_vm_hugetlb_page(vma)) {
816                 zap_hugepage_range(vma, address, size);
817                 return end;
818         }
819
820         lru_add_drain();
821         spin_lock(&mm->page_table_lock);
822         tlb = tlb_gather_mmu(mm, 0);
823         end = unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
824         tlb_finish_mmu(tlb, address, end);
825         spin_unlock(&mm->page_table_lock);
826         return end;
827 }
828
829 /*
830  * Do a quick page-table lookup for a single page.
831  * mm->page_table_lock must be held.
832  */
833 static struct page *__follow_page(struct mm_struct *mm, unsigned long address,
834                         int read, int write, int accessed)
835 {
836         pgd_t *pgd;
837         pud_t *pud;
838         pmd_t *pmd;
839         pte_t *ptep, pte;
840         unsigned long pfn;
841         struct page *page;
842
843         page = follow_huge_addr(mm, address, write);
844         if (! IS_ERR(page))
845                 return page;
846
847         pgd = pgd_offset(mm, address);
848         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
849                 goto out;
850
851         pud = pud_offset(pgd, address);
852         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
853                 goto out;
854         
855         pmd = pmd_offset(pud, address);
856         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
857                 goto out;
858         if (pmd_huge(*pmd))
859                 return follow_huge_pmd(mm, address, pmd, write);
860
861         ptep = pte_offset_map(pmd, address);
862         if (!ptep)
863                 goto out;
864
865         pte = *ptep;
866         pte_unmap(ptep);
867         if (pte_present(pte)) {
868                 if (write && !pte_write(pte))
869                         goto out;
870                 if (read && !pte_read(pte))
871                         goto out;
872                 pfn = pte_pfn(pte);
873                 if (pfn_valid(pfn)) {
874                         page = pfn_to_page(pfn);
875                         if (accessed) {
876                                 if (write && !pte_dirty(pte) &&!PageDirty(page))
877                                         set_page_dirty(page);
878                                 mark_page_accessed(page);
879                         }
880                         return page;
881                 }
882         }
883
884 out:
885         return NULL;
886 }
887
888 inline struct page *
889 follow_page(struct mm_struct *mm, unsigned long address, int write)
890 {
891         return __follow_page(mm, address, 0, write, 1);
892 }
893
894 /*
895  * check_user_page_readable() can be called frm niterrupt context by oprofile,
896  * so we need to avoid taking any non-irq-safe locks
897  */
898 int check_user_page_readable(struct mm_struct *mm, unsigned long address)
899 {
900         return __follow_page(mm, address, 1, 0, 0) != NULL;
901 }
902 EXPORT_SYMBOL(check_user_page_readable);
903
904 static inline int
905 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
906                          unsigned long address)
907 {
908         pgd_t *pgd;
909         pud_t *pud;
910         pmd_t *pmd;
911
912         /* Check if the vma is for an anonymous mapping. */
913         if (vma->vm_ops && vma->vm_ops->nopage)
914                 return 0;
915
916         /* Check if page directory entry exists. */
917         pgd = pgd_offset(mm, address);
918         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
919                 return 1;
920
921         pud = pud_offset(pgd, address);
922         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
923                 return 1;
924
925         /* Check if page middle directory entry exists. */
926         pmd = pmd_offset(pud, address);
927         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
928                 return 1;
929
930         /* There is a pte slot for 'address' in 'mm'. */
931         return 0;
932 }
933
934 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
935                 unsigned long start, int len, int write, int force,
936                 struct page **pages, struct vm_area_struct **vmas)
937 {
938         int i;
939         unsigned int flags;
940
941         /* 
942          * Require read or write permissions.
943          * If 'force' is set, we only require the "MAY" flags.
944          */
945         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
946         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
947         i = 0;
948
949         do {
950                 struct vm_area_struct * vma;
951
952                 vma = find_extend_vma(mm, start);
953                 if (!vma && in_gate_area(tsk, start)) {
954                         unsigned long pg = start & PAGE_MASK;
955                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
956                         pgd_t *pgd;
957                         pud_t *pud;
958                         pmd_t *pmd;
959                         pte_t *pte;
960                         if (write) /* user gate pages are read-only */
961                                 return i ? : -EFAULT;
962                         if (pg > TASK_SIZE)
963                                 pgd = pgd_offset_k(pg);
964                         else
965                                 pgd = pgd_offset_gate(mm, pg);
966                         BUG_ON(pgd_none(*pgd));
967                         pud = pud_offset(pgd, pg);
968                         BUG_ON(pud_none(*pud));
969                         pmd = pmd_offset(pud, pg);
970                         if (pmd_none(*pmd))
971                                 return i ? : -EFAULT;
972                         pte = pte_offset_map(pmd, pg);
973                         if (pte_none(*pte)) {
974                                 pte_unmap(pte);
975                                 return i ? : -EFAULT;
976                         }
977                         if (pages) {
978                                 pages[i] = pte_page(*pte);
979                                 get_page(pages[i]);
980                         }
981                         pte_unmap(pte);
982                         if (vmas)
983                                 vmas[i] = gate_vma;
984                         i++;
985                         start += PAGE_SIZE;
986                         len--;
987                         continue;
988                 }
989
990                 if (!vma || (vma->vm_flags & (VM_IO | VM_RESERVED))
991                                 || !(flags & vma->vm_flags))
992                         return i ? : -EFAULT;
993
994                 if (is_vm_hugetlb_page(vma)) {
995                         i = follow_hugetlb_page(mm, vma, pages, vmas,
996                                                 &start, &len, i);
997                         continue;
998                 }
999                 spin_lock(&mm->page_table_lock);
1000                 do {
1001                         int write_access = write;
1002                         struct page *page;
1003
1004                         cond_resched_lock(&mm->page_table_lock);
1005                         while (!(page = follow_page(mm, start, write_access))) {
1006                                 int ret;
1007
1008                                 /*
1009                                  * Shortcut for anonymous pages. We don't want
1010                                  * to force the creation of pages tables for
1011                                  * insanely big anonymously mapped areas that
1012                                  * nobody touched so far. This is important
1013                                  * for doing a core dump for these mappings.
1014                                  */
1015                                 if (!write && untouched_anonymous_page(mm,vma,start)) {
1016                                         page = ZERO_PAGE(start);
1017                                         break;
1018                                 }
1019                                 spin_unlock(&mm->page_table_lock);
1020                                 ret = __handle_mm_fault(mm, vma, start, write_access);
1021
1022                                 /*
1023                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1024                                  * broken COW when necessary, even if maybe_mkwrite
1025                                  * decided not to set pte_write. We can thus safely do
1026                                  * subsequent page lookups as if they were reads.
1027                                  */
1028                                 if (ret & VM_FAULT_WRITE)
1029                                         write_access = 0;
1030                                 
1031                                 switch (ret & ~VM_FAULT_WRITE) {
1032                                 case VM_FAULT_MINOR:
1033                                         tsk->min_flt++;
1034                                         break;
1035                                 case VM_FAULT_MAJOR:
1036                                         tsk->maj_flt++;
1037                                         break;
1038                                 case VM_FAULT_SIGBUS:
1039                                         return i ? i : -EFAULT;
1040                                 case VM_FAULT_OOM:
1041                                         return i ? i : -ENOMEM;
1042                                 default:
1043                                         BUG();
1044                                 }
1045                                 spin_lock(&mm->page_table_lock);
1046                         }
1047                         if (pages) {
1048                                 pages[i] = page;
1049                                 flush_dcache_page(page);
1050                                 page_cache_get(page);
1051                         }
1052                         if (vmas)
1053                                 vmas[i] = vma;
1054                         i++;
1055                         start += PAGE_SIZE;
1056                         len--;
1057                 } while (len && start < vma->vm_end);
1058                 spin_unlock(&mm->page_table_lock);
1059         } while (len);
1060         return i;
1061 }
1062 EXPORT_SYMBOL(get_user_pages);
1063
1064 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1065                         unsigned long addr, unsigned long end, pgprot_t prot)
1066 {
1067         pte_t *pte;
1068
1069         pte = pte_alloc_map(mm, pmd, addr);
1070         if (!pte)
1071                 return -ENOMEM;
1072         do {
1073                 struct page *page = ZERO_PAGE(addr);
1074                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1075                 page_cache_get(page);
1076                 page_add_file_rmap(page);
1077                 inc_mm_counter(mm, file_rss);
1078                 BUG_ON(!pte_none(*pte));
1079                 set_pte_at(mm, addr, pte, zero_pte);
1080         } while (pte++, addr += PAGE_SIZE, addr != end);
1081         pte_unmap(pte - 1);
1082         return 0;
1083 }
1084
1085 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1086                         unsigned long addr, unsigned long end, pgprot_t prot)
1087 {
1088         pmd_t *pmd;
1089         unsigned long next;
1090
1091         pmd = pmd_alloc(mm, pud, addr);
1092         if (!pmd)
1093                 return -ENOMEM;
1094         do {
1095                 next = pmd_addr_end(addr, end);
1096                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1097                         return -ENOMEM;
1098         } while (pmd++, addr = next, addr != end);
1099         return 0;
1100 }
1101
1102 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1103                         unsigned long addr, unsigned long end, pgprot_t prot)
1104 {
1105         pud_t *pud;
1106         unsigned long next;
1107
1108         pud = pud_alloc(mm, pgd, addr);
1109         if (!pud)
1110                 return -ENOMEM;
1111         do {
1112                 next = pud_addr_end(addr, end);
1113                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1114                         return -ENOMEM;
1115         } while (pud++, addr = next, addr != end);
1116         return 0;
1117 }
1118
1119 int zeromap_page_range(struct vm_area_struct *vma,
1120                         unsigned long addr, unsigned long size, pgprot_t prot)
1121 {
1122         pgd_t *pgd;
1123         unsigned long next;
1124         unsigned long end = addr + size;
1125         struct mm_struct *mm = vma->vm_mm;
1126         int err;
1127
1128         BUG_ON(addr >= end);
1129         pgd = pgd_offset(mm, addr);
1130         flush_cache_range(vma, addr, end);
1131         spin_lock(&mm->page_table_lock);
1132         do {
1133                 next = pgd_addr_end(addr, end);
1134                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1135                 if (err)
1136                         break;
1137         } while (pgd++, addr = next, addr != end);
1138         spin_unlock(&mm->page_table_lock);
1139         return err;
1140 }
1141
1142 /*
1143  * maps a range of physical memory into the requested pages. the old
1144  * mappings are removed. any references to nonexistent pages results
1145  * in null mappings (currently treated as "copy-on-access")
1146  */
1147 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1148                         unsigned long addr, unsigned long end,
1149                         unsigned long pfn, pgprot_t prot)
1150 {
1151         pte_t *pte;
1152
1153         pte = pte_alloc_map(mm, pmd, addr);
1154         if (!pte)
1155                 return -ENOMEM;
1156         do {
1157                 BUG_ON(!pte_none(*pte));
1158                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1159                 pfn++;
1160         } while (pte++, addr += PAGE_SIZE, addr != end);
1161         pte_unmap(pte - 1);
1162         return 0;
1163 }
1164
1165 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1166                         unsigned long addr, unsigned long end,
1167                         unsigned long pfn, pgprot_t prot)
1168 {
1169         pmd_t *pmd;
1170         unsigned long next;
1171
1172         pfn -= addr >> PAGE_SHIFT;
1173         pmd = pmd_alloc(mm, pud, addr);
1174         if (!pmd)
1175                 return -ENOMEM;
1176         do {
1177                 next = pmd_addr_end(addr, end);
1178                 if (remap_pte_range(mm, pmd, addr, next,
1179                                 pfn + (addr >> PAGE_SHIFT), prot))
1180                         return -ENOMEM;
1181         } while (pmd++, addr = next, addr != end);
1182         return 0;
1183 }
1184
1185 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1186                         unsigned long addr, unsigned long end,
1187                         unsigned long pfn, pgprot_t prot)
1188 {
1189         pud_t *pud;
1190         unsigned long next;
1191
1192         pfn -= addr >> PAGE_SHIFT;
1193         pud = pud_alloc(mm, pgd, addr);
1194         if (!pud)
1195                 return -ENOMEM;
1196         do {
1197                 next = pud_addr_end(addr, end);
1198                 if (remap_pmd_range(mm, pud, addr, next,
1199                                 pfn + (addr >> PAGE_SHIFT), prot))
1200                         return -ENOMEM;
1201         } while (pud++, addr = next, addr != end);
1202         return 0;
1203 }
1204
1205 /*  Note: this is only safe if the mm semaphore is held when called. */
1206 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1207                     unsigned long pfn, unsigned long size, pgprot_t prot)
1208 {
1209         pgd_t *pgd;
1210         unsigned long next;
1211         unsigned long end = addr + PAGE_ALIGN(size);
1212         struct mm_struct *mm = vma->vm_mm;
1213         int err;
1214
1215         /*
1216          * Physically remapped pages are special. Tell the
1217          * rest of the world about it:
1218          *   VM_IO tells people not to look at these pages
1219          *      (accesses can have side effects).
1220          *   VM_RESERVED tells the core MM not to "manage" these pages
1221          *      (e.g. refcount, mapcount, try to swap them out).
1222          */
1223         vma->vm_flags |= VM_IO | VM_RESERVED;
1224
1225         BUG_ON(addr >= end);
1226         pfn -= addr >> PAGE_SHIFT;
1227         pgd = pgd_offset(mm, addr);
1228         flush_cache_range(vma, addr, end);
1229         spin_lock(&mm->page_table_lock);
1230         do {
1231                 next = pgd_addr_end(addr, end);
1232                 err = remap_pud_range(mm, pgd, addr, next,
1233                                 pfn + (addr >> PAGE_SHIFT), prot);
1234                 if (err)
1235                         break;
1236         } while (pgd++, addr = next, addr != end);
1237         spin_unlock(&mm->page_table_lock);
1238         return err;
1239 }
1240 EXPORT_SYMBOL(remap_pfn_range);
1241
1242 /*
1243  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1244  * servicing faults for write access.  In the normal case, do always want
1245  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1246  * that do not have writing enabled, when used by access_process_vm.
1247  */
1248 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1249 {
1250         if (likely(vma->vm_flags & VM_WRITE))
1251                 pte = pte_mkwrite(pte);
1252         return pte;
1253 }
1254
1255 /*
1256  * This routine handles present pages, when users try to write
1257  * to a shared page. It is done by copying the page to a new address
1258  * and decrementing the shared-page counter for the old page.
1259  *
1260  * Note that this routine assumes that the protection checks have been
1261  * done by the caller (the low-level page fault routine in most cases).
1262  * Thus we can safely just mark it writable once we've done any necessary
1263  * COW.
1264  *
1265  * We also mark the page dirty at this point even though the page will
1266  * change only once the write actually happens. This avoids a few races,
1267  * and potentially makes it more efficient.
1268  *
1269  * We hold the mm semaphore and the page_table_lock on entry and exit
1270  * with the page_table_lock released.
1271  */
1272 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1273                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1274                 pte_t orig_pte)
1275 {
1276         struct page *old_page, *new_page;
1277         unsigned long pfn = pte_pfn(orig_pte);
1278         pte_t entry;
1279         int ret = VM_FAULT_MINOR;
1280
1281         BUG_ON(vma->vm_flags & VM_RESERVED);
1282
1283         if (unlikely(!pfn_valid(pfn))) {
1284                 /*
1285                  * Page table corrupted: show pte and kill process.
1286                  */
1287                 print_bad_pte(vma, orig_pte, address);
1288                 ret = VM_FAULT_OOM;
1289                 goto unlock;
1290         }
1291         old_page = pfn_to_page(pfn);
1292
1293         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1294                 int reuse = can_share_swap_page(old_page);
1295                 unlock_page(old_page);
1296                 if (reuse) {
1297                         flush_cache_page(vma, address, pfn);
1298                         entry = pte_mkyoung(orig_pte);
1299                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1300                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1301                         update_mmu_cache(vma, address, entry);
1302                         lazy_mmu_prot_update(entry);
1303                         ret |= VM_FAULT_WRITE;
1304                         goto unlock;
1305                 }
1306         }
1307
1308         /*
1309          * Ok, we need to copy. Oh, well..
1310          */
1311         page_cache_get(old_page);
1312         pte_unmap(page_table);
1313         spin_unlock(&mm->page_table_lock);
1314
1315         if (unlikely(anon_vma_prepare(vma)))
1316                 goto oom;
1317         if (old_page == ZERO_PAGE(address)) {
1318                 new_page = alloc_zeroed_user_highpage(vma, address);
1319                 if (!new_page)
1320                         goto oom;
1321         } else {
1322                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1323                 if (!new_page)
1324                         goto oom;
1325                 copy_user_highpage(new_page, old_page, address);
1326         }
1327
1328         /*
1329          * Re-check the pte - we dropped the lock
1330          */
1331         spin_lock(&mm->page_table_lock);
1332         page_table = pte_offset_map(pmd, address);
1333         if (likely(pte_same(*page_table, orig_pte))) {
1334                 page_remove_rmap(old_page);
1335                 if (!PageAnon(old_page)) {
1336                         inc_mm_counter(mm, anon_rss);
1337                         dec_mm_counter(mm, file_rss);
1338                 }
1339                 flush_cache_page(vma, address, pfn);
1340                 entry = mk_pte(new_page, vma->vm_page_prot);
1341                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1342                 ptep_establish(vma, address, page_table, entry);
1343                 update_mmu_cache(vma, address, entry);
1344                 lazy_mmu_prot_update(entry);
1345
1346                 lru_cache_add_active(new_page);
1347                 page_add_anon_rmap(new_page, vma, address);
1348
1349                 /* Free the old page.. */
1350                 new_page = old_page;
1351                 ret |= VM_FAULT_WRITE;
1352         }
1353         page_cache_release(new_page);
1354         page_cache_release(old_page);
1355 unlock:
1356         pte_unmap(page_table);
1357         spin_unlock(&mm->page_table_lock);
1358         return ret;
1359 oom:
1360         page_cache_release(old_page);
1361         return VM_FAULT_OOM;
1362 }
1363
1364 /*
1365  * Helper functions for unmap_mapping_range().
1366  *
1367  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1368  *
1369  * We have to restart searching the prio_tree whenever we drop the lock,
1370  * since the iterator is only valid while the lock is held, and anyway
1371  * a later vma might be split and reinserted earlier while lock dropped.
1372  *
1373  * The list of nonlinear vmas could be handled more efficiently, using
1374  * a placeholder, but handle it in the same way until a need is shown.
1375  * It is important to search the prio_tree before nonlinear list: a vma
1376  * may become nonlinear and be shifted from prio_tree to nonlinear list
1377  * while the lock is dropped; but never shifted from list to prio_tree.
1378  *
1379  * In order to make forward progress despite restarting the search,
1380  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1381  * quickly skip it next time around.  Since the prio_tree search only
1382  * shows us those vmas affected by unmapping the range in question, we
1383  * can't efficiently keep all vmas in step with mapping->truncate_count:
1384  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1385  * mapping->truncate_count and vma->vm_truncate_count are protected by
1386  * i_mmap_lock.
1387  *
1388  * In order to make forward progress despite repeatedly restarting some
1389  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1390  * and restart from that address when we reach that vma again.  It might
1391  * have been split or merged, shrunk or extended, but never shifted: so
1392  * restart_addr remains valid so long as it remains in the vma's range.
1393  * unmap_mapping_range forces truncate_count to leap over page-aligned
1394  * values so we can save vma's restart_addr in its truncate_count field.
1395  */
1396 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1397
1398 static void reset_vma_truncate_counts(struct address_space *mapping)
1399 {
1400         struct vm_area_struct *vma;
1401         struct prio_tree_iter iter;
1402
1403         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1404                 vma->vm_truncate_count = 0;
1405         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1406                 vma->vm_truncate_count = 0;
1407 }
1408
1409 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1410                 unsigned long start_addr, unsigned long end_addr,
1411                 struct zap_details *details)
1412 {
1413         unsigned long restart_addr;
1414         int need_break;
1415
1416 again:
1417         restart_addr = vma->vm_truncate_count;
1418         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1419                 start_addr = restart_addr;
1420                 if (start_addr >= end_addr) {
1421                         /* Top of vma has been split off since last time */
1422                         vma->vm_truncate_count = details->truncate_count;
1423                         return 0;
1424                 }
1425         }
1426
1427         restart_addr = zap_page_range(vma, start_addr,
1428                                         end_addr - start_addr, details);
1429
1430         /*
1431          * We cannot rely on the break test in unmap_vmas:
1432          * on the one hand, we don't want to restart our loop
1433          * just because that broke out for the page_table_lock;
1434          * on the other hand, it does no test when vma is small.
1435          */
1436         need_break = need_resched() ||
1437                         need_lockbreak(details->i_mmap_lock);
1438
1439         if (restart_addr >= end_addr) {
1440                 /* We have now completed this vma: mark it so */
1441                 vma->vm_truncate_count = details->truncate_count;
1442                 if (!need_break)
1443                         return 0;
1444         } else {
1445                 /* Note restart_addr in vma's truncate_count field */
1446                 vma->vm_truncate_count = restart_addr;
1447                 if (!need_break)
1448                         goto again;
1449         }
1450
1451         spin_unlock(details->i_mmap_lock);
1452         cond_resched();
1453         spin_lock(details->i_mmap_lock);
1454         return -EINTR;
1455 }
1456
1457 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1458                                             struct zap_details *details)
1459 {
1460         struct vm_area_struct *vma;
1461         struct prio_tree_iter iter;
1462         pgoff_t vba, vea, zba, zea;
1463
1464 restart:
1465         vma_prio_tree_foreach(vma, &iter, root,
1466                         details->first_index, details->last_index) {
1467                 /* Skip quickly over those we have already dealt with */
1468                 if (vma->vm_truncate_count == details->truncate_count)
1469                         continue;
1470
1471                 vba = vma->vm_pgoff;
1472                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1473                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1474                 zba = details->first_index;
1475                 if (zba < vba)
1476                         zba = vba;
1477                 zea = details->last_index;
1478                 if (zea > vea)
1479                         zea = vea;
1480
1481                 if (unmap_mapping_range_vma(vma,
1482                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1483                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1484                                 details) < 0)
1485                         goto restart;
1486         }
1487 }
1488
1489 static inline void unmap_mapping_range_list(struct list_head *head,
1490                                             struct zap_details *details)
1491 {
1492         struct vm_area_struct *vma;
1493
1494         /*
1495          * In nonlinear VMAs there is no correspondence between virtual address
1496          * offset and file offset.  So we must perform an exhaustive search
1497          * across *all* the pages in each nonlinear VMA, not just the pages
1498          * whose virtual address lies outside the file truncation point.
1499          */
1500 restart:
1501         list_for_each_entry(vma, head, shared.vm_set.list) {
1502                 /* Skip quickly over those we have already dealt with */
1503                 if (vma->vm_truncate_count == details->truncate_count)
1504                         continue;
1505                 details->nonlinear_vma = vma;
1506                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1507                                         vma->vm_end, details) < 0)
1508                         goto restart;
1509         }
1510 }
1511
1512 /**
1513  * unmap_mapping_range - unmap the portion of all mmaps
1514  * in the specified address_space corresponding to the specified
1515  * page range in the underlying file.
1516  * @mapping: the address space containing mmaps to be unmapped.
1517  * @holebegin: byte in first page to unmap, relative to the start of
1518  * the underlying file.  This will be rounded down to a PAGE_SIZE
1519  * boundary.  Note that this is different from vmtruncate(), which
1520  * must keep the partial page.  In contrast, we must get rid of
1521  * partial pages.
1522  * @holelen: size of prospective hole in bytes.  This will be rounded
1523  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1524  * end of the file.
1525  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1526  * but 0 when invalidating pagecache, don't throw away private data.
1527  */
1528 void unmap_mapping_range(struct address_space *mapping,
1529                 loff_t const holebegin, loff_t const holelen, int even_cows)
1530 {
1531         struct zap_details details;
1532         pgoff_t hba = holebegin >> PAGE_SHIFT;
1533         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1534
1535         /* Check for overflow. */
1536         if (sizeof(holelen) > sizeof(hlen)) {
1537                 long long holeend =
1538                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1539                 if (holeend & ~(long long)ULONG_MAX)
1540                         hlen = ULONG_MAX - hba + 1;
1541         }
1542
1543         details.check_mapping = even_cows? NULL: mapping;
1544         details.nonlinear_vma = NULL;
1545         details.first_index = hba;
1546         details.last_index = hba + hlen - 1;
1547         if (details.last_index < details.first_index)
1548                 details.last_index = ULONG_MAX;
1549         details.i_mmap_lock = &mapping->i_mmap_lock;
1550
1551         spin_lock(&mapping->i_mmap_lock);
1552
1553         /* serialize i_size write against truncate_count write */
1554         smp_wmb();
1555         /* Protect against page faults, and endless unmapping loops */
1556         mapping->truncate_count++;
1557         /*
1558          * For archs where spin_lock has inclusive semantics like ia64
1559          * this smp_mb() will prevent to read pagetable contents
1560          * before the truncate_count increment is visible to
1561          * other cpus.
1562          */
1563         smp_mb();
1564         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1565                 if (mapping->truncate_count == 0)
1566                         reset_vma_truncate_counts(mapping);
1567                 mapping->truncate_count++;
1568         }
1569         details.truncate_count = mapping->truncate_count;
1570
1571         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1572                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1573         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1574                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1575         spin_unlock(&mapping->i_mmap_lock);
1576 }
1577 EXPORT_SYMBOL(unmap_mapping_range);
1578
1579 /*
1580  * Handle all mappings that got truncated by a "truncate()"
1581  * system call.
1582  *
1583  * NOTE! We have to be ready to update the memory sharing
1584  * between the file and the memory map for a potential last
1585  * incomplete page.  Ugly, but necessary.
1586  */
1587 int vmtruncate(struct inode * inode, loff_t offset)
1588 {
1589         struct address_space *mapping = inode->i_mapping;
1590         unsigned long limit;
1591
1592         if (inode->i_size < offset)
1593                 goto do_expand;
1594         /*
1595          * truncation of in-use swapfiles is disallowed - it would cause
1596          * subsequent swapout to scribble on the now-freed blocks.
1597          */
1598         if (IS_SWAPFILE(inode))
1599                 goto out_busy;
1600         i_size_write(inode, offset);
1601         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1602         truncate_inode_pages(mapping, offset);
1603         goto out_truncate;
1604
1605 do_expand:
1606         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1607         if (limit != RLIM_INFINITY && offset > limit)
1608                 goto out_sig;
1609         if (offset > inode->i_sb->s_maxbytes)
1610                 goto out_big;
1611         i_size_write(inode, offset);
1612
1613 out_truncate:
1614         if (inode->i_op && inode->i_op->truncate)
1615                 inode->i_op->truncate(inode);
1616         return 0;
1617 out_sig:
1618         send_sig(SIGXFSZ, current, 0);
1619 out_big:
1620         return -EFBIG;
1621 out_busy:
1622         return -ETXTBSY;
1623 }
1624
1625 EXPORT_SYMBOL(vmtruncate);
1626
1627 /* 
1628  * Primitive swap readahead code. We simply read an aligned block of
1629  * (1 << page_cluster) entries in the swap area. This method is chosen
1630  * because it doesn't cost us any seek time.  We also make sure to queue
1631  * the 'original' request together with the readahead ones...  
1632  *
1633  * This has been extended to use the NUMA policies from the mm triggering
1634  * the readahead.
1635  *
1636  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1637  */
1638 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1639 {
1640 #ifdef CONFIG_NUMA
1641         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1642 #endif
1643         int i, num;
1644         struct page *new_page;
1645         unsigned long offset;
1646
1647         /*
1648          * Get the number of handles we should do readahead io to.
1649          */
1650         num = valid_swaphandles(entry, &offset);
1651         for (i = 0; i < num; offset++, i++) {
1652                 /* Ok, do the async read-ahead now */
1653                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1654                                                            offset), vma, addr);
1655                 if (!new_page)
1656                         break;
1657                 page_cache_release(new_page);
1658 #ifdef CONFIG_NUMA
1659                 /*
1660                  * Find the next applicable VMA for the NUMA policy.
1661                  */
1662                 addr += PAGE_SIZE;
1663                 if (addr == 0)
1664                         vma = NULL;
1665                 if (vma) {
1666                         if (addr >= vma->vm_end) {
1667                                 vma = next_vma;
1668                                 next_vma = vma ? vma->vm_next : NULL;
1669                         }
1670                         if (vma && addr < vma->vm_start)
1671                                 vma = NULL;
1672                 } else {
1673                         if (next_vma && addr >= next_vma->vm_start) {
1674                                 vma = next_vma;
1675                                 next_vma = vma->vm_next;
1676                         }
1677                 }
1678 #endif
1679         }
1680         lru_add_drain();        /* Push any new pages onto the LRU now */
1681 }
1682
1683 /*
1684  * We hold the mm semaphore and the page_table_lock on entry and
1685  * should release the pagetable lock on exit..
1686  */
1687 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1688                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1689                 int write_access, pte_t orig_pte)
1690 {
1691         struct page *page;
1692         swp_entry_t entry;
1693         pte_t pte;
1694         int ret = VM_FAULT_MINOR;
1695
1696         pte_unmap(page_table);
1697         spin_unlock(&mm->page_table_lock);
1698
1699         entry = pte_to_swp_entry(orig_pte);
1700         page = lookup_swap_cache(entry);
1701         if (!page) {
1702                 swapin_readahead(entry, address, vma);
1703                 page = read_swap_cache_async(entry, vma, address);
1704                 if (!page) {
1705                         /*
1706                          * Back out if somebody else faulted in this pte while
1707                          * we released the page table lock.
1708                          */
1709                         spin_lock(&mm->page_table_lock);
1710                         page_table = pte_offset_map(pmd, address);
1711                         if (likely(pte_same(*page_table, orig_pte)))
1712                                 ret = VM_FAULT_OOM;
1713                         goto unlock;
1714                 }
1715
1716                 /* Had to read the page from swap area: Major fault */
1717                 ret = VM_FAULT_MAJOR;
1718                 inc_page_state(pgmajfault);
1719                 grab_swap_token();
1720         }
1721
1722         mark_page_accessed(page);
1723         lock_page(page);
1724
1725         /*
1726          * Back out if somebody else faulted in this pte while we
1727          * released the page table lock.
1728          */
1729         spin_lock(&mm->page_table_lock);
1730         page_table = pte_offset_map(pmd, address);
1731         if (unlikely(!pte_same(*page_table, orig_pte))) {
1732                 ret = VM_FAULT_MINOR;
1733                 goto out_nomap;
1734         }
1735
1736         if (unlikely(!PageUptodate(page))) {
1737                 ret = VM_FAULT_SIGBUS;
1738                 goto out_nomap;
1739         }
1740
1741         /* The page isn't present yet, go ahead with the fault. */
1742
1743         inc_mm_counter(mm, anon_rss);
1744         pte = mk_pte(page, vma->vm_page_prot);
1745         if (write_access && can_share_swap_page(page)) {
1746                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1747                 write_access = 0;
1748         }
1749
1750         flush_icache_page(vma, page);
1751         set_pte_at(mm, address, page_table, pte);
1752         page_add_anon_rmap(page, vma, address);
1753
1754         swap_free(entry);
1755         if (vm_swap_full())
1756                 remove_exclusive_swap_page(page);
1757         unlock_page(page);
1758
1759         if (write_access) {
1760                 if (do_wp_page(mm, vma, address,
1761                                 page_table, pmd, pte) == VM_FAULT_OOM)
1762                         ret = VM_FAULT_OOM;
1763                 goto out;
1764         }
1765
1766         /* No need to invalidate - it was non-present before */
1767         update_mmu_cache(vma, address, pte);
1768         lazy_mmu_prot_update(pte);
1769 unlock:
1770         pte_unmap(page_table);
1771         spin_unlock(&mm->page_table_lock);
1772 out:
1773         return ret;
1774 out_nomap:
1775         pte_unmap(page_table);
1776         spin_unlock(&mm->page_table_lock);
1777         unlock_page(page);
1778         page_cache_release(page);
1779         return ret;
1780 }
1781
1782 /*
1783  * We are called with the MM semaphore and page_table_lock
1784  * spinlock held to protect against concurrent faults in
1785  * multithreaded programs. 
1786  */
1787 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1788                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1789                 int write_access)
1790 {
1791         struct page *page = ZERO_PAGE(addr);
1792         pte_t entry;
1793
1794         /* Mapping of ZERO_PAGE - vm_page_prot is readonly */
1795         entry = mk_pte(page, vma->vm_page_prot);
1796
1797         if (write_access) {
1798                 /* Allocate our own private page. */
1799                 pte_unmap(page_table);
1800                 spin_unlock(&mm->page_table_lock);
1801
1802                 if (unlikely(anon_vma_prepare(vma)))
1803                         goto oom;
1804                 page = alloc_zeroed_user_highpage(vma, address);
1805                 if (!page)
1806                         goto oom;
1807
1808                 spin_lock(&mm->page_table_lock);
1809                 page_table = pte_offset_map(pmd, address);
1810
1811                 if (!pte_none(*page_table)) {
1812                         page_cache_release(page);
1813                         goto unlock;
1814                 }
1815                 inc_mm_counter(mm, anon_rss);
1816                 entry = mk_pte(page, vma->vm_page_prot);
1817                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1818                 lru_cache_add_active(page);
1819                 SetPageReferenced(page);
1820                 page_add_anon_rmap(page, vma, address);
1821         } else {
1822                 inc_mm_counter(mm, file_rss);
1823                 page_add_file_rmap(page);
1824                 page_cache_get(page);
1825         }
1826
1827         set_pte_at(mm, address, page_table, entry);
1828
1829         /* No need to invalidate - it was non-present before */
1830         update_mmu_cache(vma, address, entry);
1831         lazy_mmu_prot_update(entry);
1832 unlock:
1833         pte_unmap(page_table);
1834         spin_unlock(&mm->page_table_lock);
1835         return VM_FAULT_MINOR;
1836 oom:
1837         return VM_FAULT_OOM;
1838 }
1839
1840 /*
1841  * do_no_page() tries to create a new page mapping. It aggressively
1842  * tries to share with existing pages, but makes a separate copy if
1843  * the "write_access" parameter is true in order to avoid the next
1844  * page fault.
1845  *
1846  * As this is called only for pages that do not currently exist, we
1847  * do not need to flush old virtual caches or the TLB.
1848  *
1849  * This is called with the MM semaphore held and the page table
1850  * spinlock held. Exit with the spinlock released.
1851  */
1852 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1853                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1854                 int write_access)
1855 {
1856         struct page *new_page;
1857         struct address_space *mapping = NULL;
1858         pte_t entry;
1859         unsigned int sequence = 0;
1860         int ret = VM_FAULT_MINOR;
1861         int anon = 0;
1862
1863         pte_unmap(page_table);
1864         spin_unlock(&mm->page_table_lock);
1865
1866         if (vma->vm_file) {
1867                 mapping = vma->vm_file->f_mapping;
1868                 sequence = mapping->truncate_count;
1869                 smp_rmb(); /* serializes i_size against truncate_count */
1870         }
1871 retry:
1872         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1873         /*
1874          * No smp_rmb is needed here as long as there's a full
1875          * spin_lock/unlock sequence inside the ->nopage callback
1876          * (for the pagecache lookup) that acts as an implicit
1877          * smp_mb() and prevents the i_size read to happen
1878          * after the next truncate_count read.
1879          */
1880
1881         /* no page was available -- either SIGBUS or OOM */
1882         if (new_page == NOPAGE_SIGBUS)
1883                 return VM_FAULT_SIGBUS;
1884         if (new_page == NOPAGE_OOM)
1885                 return VM_FAULT_OOM;
1886
1887         /*
1888          * Should we do an early C-O-W break?
1889          */
1890         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1891                 struct page *page;
1892
1893                 if (unlikely(anon_vma_prepare(vma)))
1894                         goto oom;
1895                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1896                 if (!page)
1897                         goto oom;
1898                 copy_user_highpage(page, new_page, address);
1899                 page_cache_release(new_page);
1900                 new_page = page;
1901                 anon = 1;
1902         }
1903
1904         spin_lock(&mm->page_table_lock);
1905         /*
1906          * For a file-backed vma, someone could have truncated or otherwise
1907          * invalidated this page.  If unmap_mapping_range got called,
1908          * retry getting the page.
1909          */
1910         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1911                 spin_unlock(&mm->page_table_lock);
1912                 page_cache_release(new_page);
1913                 cond_resched();
1914                 sequence = mapping->truncate_count;
1915                 smp_rmb();
1916                 goto retry;
1917         }
1918         page_table = pte_offset_map(pmd, address);
1919
1920         /*
1921          * This silly early PAGE_DIRTY setting removes a race
1922          * due to the bad i386 page protection. But it's valid
1923          * for other architectures too.
1924          *
1925          * Note that if write_access is true, we either now have
1926          * an exclusive copy of the page, or this is a shared mapping,
1927          * so we can make it writable and dirty to avoid having to
1928          * handle that later.
1929          */
1930         /* Only go through if we didn't race with anybody else... */
1931         if (pte_none(*page_table)) {
1932                 flush_icache_page(vma, new_page);
1933                 entry = mk_pte(new_page, vma->vm_page_prot);
1934                 if (write_access)
1935                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1936                 set_pte_at(mm, address, page_table, entry);
1937                 if (anon) {
1938                         inc_mm_counter(mm, anon_rss);
1939                         lru_cache_add_active(new_page);
1940                         page_add_anon_rmap(new_page, vma, address);
1941                 } else if (!(vma->vm_flags & VM_RESERVED)) {
1942                         inc_mm_counter(mm, file_rss);
1943                         page_add_file_rmap(new_page);
1944                 }
1945         } else {
1946                 /* One of our sibling threads was faster, back out. */
1947                 page_cache_release(new_page);
1948                 goto unlock;
1949         }
1950
1951         /* no need to invalidate: a not-present page shouldn't be cached */
1952         update_mmu_cache(vma, address, entry);
1953         lazy_mmu_prot_update(entry);
1954 unlock:
1955         pte_unmap(page_table);
1956         spin_unlock(&mm->page_table_lock);
1957         return ret;
1958 oom:
1959         page_cache_release(new_page);
1960         return VM_FAULT_OOM;
1961 }
1962
1963 /*
1964  * Fault of a previously existing named mapping. Repopulate the pte
1965  * from the encoded file_pte if possible. This enables swappable
1966  * nonlinear vmas.
1967  */
1968 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
1969                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1970                 int write_access, pte_t orig_pte)
1971 {
1972         pgoff_t pgoff;
1973         int err;
1974
1975         pte_unmap(page_table);
1976         spin_unlock(&mm->page_table_lock);
1977
1978         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
1979                 /*
1980                  * Page table corrupted: show pte and kill process.
1981                  */
1982                 print_bad_pte(vma, orig_pte, address);
1983                 return VM_FAULT_OOM;
1984         }
1985         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
1986
1987         pgoff = pte_to_pgoff(orig_pte);
1988         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
1989                                         vma->vm_page_prot, pgoff, 0);
1990         if (err == -ENOMEM)
1991                 return VM_FAULT_OOM;
1992         if (err)
1993                 return VM_FAULT_SIGBUS;
1994         return VM_FAULT_MAJOR;
1995 }
1996
1997 /*
1998  * These routines also need to handle stuff like marking pages dirty
1999  * and/or accessed for architectures that don't do it in hardware (most
2000  * RISC architectures).  The early dirtying is also good on the i386.
2001  *
2002  * There is also a hook called "update_mmu_cache()" that architectures
2003  * with external mmu caches can use to update those (ie the Sparc or
2004  * PowerPC hashed page tables that act as extended TLBs).
2005  *
2006  * Note the "page_table_lock". It is to protect against kswapd removing
2007  * pages from under us. Note that kswapd only ever _removes_ pages, never
2008  * adds them. As such, once we have noticed that the page is not present,
2009  * we can drop the lock early.
2010  *
2011  * The adding of pages is protected by the MM semaphore (which we hold),
2012  * so we don't need to worry about a page being suddenly been added into
2013  * our VM.
2014  *
2015  * We enter with the pagetable spinlock held, we are supposed to
2016  * release it when done.
2017  */
2018 static inline int handle_pte_fault(struct mm_struct *mm,
2019                 struct vm_area_struct *vma, unsigned long address,
2020                 pte_t *pte, pmd_t *pmd, int write_access)
2021 {
2022         pte_t entry;
2023
2024         entry = *pte;
2025         if (!pte_present(entry)) {
2026                 if (pte_none(entry)) {
2027                         if (!vma->vm_ops || !vma->vm_ops->nopage)
2028                                 return do_anonymous_page(mm, vma, address,
2029                                         pte, pmd, write_access);
2030                         return do_no_page(mm, vma, address,
2031                                         pte, pmd, write_access);
2032                 }
2033                 if (pte_file(entry))
2034                         return do_file_page(mm, vma, address,
2035                                         pte, pmd, write_access, entry);
2036                 return do_swap_page(mm, vma, address,
2037                                         pte, pmd, write_access, entry);
2038         }
2039
2040         if (write_access) {
2041                 if (!pte_write(entry))
2042                         return do_wp_page(mm, vma, address, pte, pmd, entry);
2043                 entry = pte_mkdirty(entry);
2044         }
2045         entry = pte_mkyoung(entry);
2046         ptep_set_access_flags(vma, address, pte, entry, write_access);
2047         update_mmu_cache(vma, address, entry);
2048         lazy_mmu_prot_update(entry);
2049         pte_unmap(pte);
2050         spin_unlock(&mm->page_table_lock);
2051         return VM_FAULT_MINOR;
2052 }
2053
2054 /*
2055  * By the time we get here, we already hold the mm semaphore
2056  */
2057 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2058                 unsigned long address, int write_access)
2059 {
2060         pgd_t *pgd;
2061         pud_t *pud;
2062         pmd_t *pmd;
2063         pte_t *pte;
2064
2065         __set_current_state(TASK_RUNNING);
2066
2067         inc_page_state(pgfault);
2068
2069         if (unlikely(is_vm_hugetlb_page(vma)))
2070                 return hugetlb_fault(mm, vma, address, write_access);
2071
2072         /*
2073          * We need the page table lock to synchronize with kswapd
2074          * and the SMP-safe atomic PTE updates.
2075          */
2076         pgd = pgd_offset(mm, address);
2077         spin_lock(&mm->page_table_lock);
2078
2079         pud = pud_alloc(mm, pgd, address);
2080         if (!pud)
2081                 goto oom;
2082
2083         pmd = pmd_alloc(mm, pud, address);
2084         if (!pmd)
2085                 goto oom;
2086
2087         pte = pte_alloc_map(mm, pmd, address);
2088         if (!pte)
2089                 goto oom;
2090         
2091         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2092
2093  oom:
2094         spin_unlock(&mm->page_table_lock);
2095         return VM_FAULT_OOM;
2096 }
2097
2098 #ifndef __PAGETABLE_PUD_FOLDED
2099 /*
2100  * Allocate page upper directory.
2101  *
2102  * We've already handled the fast-path in-line, and we own the
2103  * page table lock.
2104  */
2105 pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2106 {
2107         pud_t *new;
2108
2109         spin_unlock(&mm->page_table_lock);
2110         new = pud_alloc_one(mm, address);
2111         spin_lock(&mm->page_table_lock);
2112         if (!new)
2113                 return NULL;
2114
2115         /*
2116          * Because we dropped the lock, we should re-check the
2117          * entry, as somebody else could have populated it..
2118          */
2119         if (pgd_present(*pgd)) {
2120                 pud_free(new);
2121                 goto out;
2122         }
2123         pgd_populate(mm, pgd, new);
2124  out:
2125         return pud_offset(pgd, address);
2126 }
2127 #endif /* __PAGETABLE_PUD_FOLDED */
2128
2129 #ifndef __PAGETABLE_PMD_FOLDED
2130 /*
2131  * Allocate page middle directory.
2132  *
2133  * We've already handled the fast-path in-line, and we own the
2134  * page table lock.
2135  */
2136 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2137 {
2138         pmd_t *new;
2139
2140         spin_unlock(&mm->page_table_lock);
2141         new = pmd_alloc_one(mm, address);
2142         spin_lock(&mm->page_table_lock);
2143         if (!new)
2144                 return NULL;
2145
2146         /*
2147          * Because we dropped the lock, we should re-check the
2148          * entry, as somebody else could have populated it..
2149          */
2150 #ifndef __ARCH_HAS_4LEVEL_HACK
2151         if (pud_present(*pud)) {
2152                 pmd_free(new);
2153                 goto out;
2154         }
2155         pud_populate(mm, pud, new);
2156 #else
2157         if (pgd_present(*pud)) {
2158                 pmd_free(new);
2159                 goto out;
2160         }
2161         pgd_populate(mm, pud, new);
2162 #endif /* __ARCH_HAS_4LEVEL_HACK */
2163
2164  out:
2165         return pmd_offset(pud, address);
2166 }
2167 #endif /* __PAGETABLE_PMD_FOLDED */
2168
2169 int make_pages_present(unsigned long addr, unsigned long end)
2170 {
2171         int ret, len, write;
2172         struct vm_area_struct * vma;
2173
2174         vma = find_vma(current->mm, addr);
2175         if (!vma)
2176                 return -1;
2177         write = (vma->vm_flags & VM_WRITE) != 0;
2178         if (addr >= end)
2179                 BUG();
2180         if (end > vma->vm_end)
2181                 BUG();
2182         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2183         ret = get_user_pages(current, current->mm, addr,
2184                         len, write, 0, NULL, NULL);
2185         if (ret < 0)
2186                 return ret;
2187         return ret == len ? 0 : -1;
2188 }
2189
2190 /* 
2191  * Map a vmalloc()-space virtual address to the physical page.
2192  */
2193 struct page * vmalloc_to_page(void * vmalloc_addr)
2194 {
2195         unsigned long addr = (unsigned long) vmalloc_addr;
2196         struct page *page = NULL;
2197         pgd_t *pgd = pgd_offset_k(addr);
2198         pud_t *pud;
2199         pmd_t *pmd;
2200         pte_t *ptep, pte;
2201   
2202         if (!pgd_none(*pgd)) {
2203                 pud = pud_offset(pgd, addr);
2204                 if (!pud_none(*pud)) {
2205                         pmd = pmd_offset(pud, addr);
2206                         if (!pmd_none(*pmd)) {
2207                                 ptep = pte_offset_map(pmd, addr);
2208                                 pte = *ptep;
2209                                 if (pte_present(pte))
2210                                         page = pte_page(pte);
2211                                 pte_unmap(ptep);
2212                         }
2213                 }
2214         }
2215         return page;
2216 }
2217
2218 EXPORT_SYMBOL(vmalloc_to_page);
2219
2220 /*
2221  * Map a vmalloc()-space virtual address to the physical page frame number.
2222  */
2223 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2224 {
2225         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2226 }
2227
2228 EXPORT_SYMBOL(vmalloc_to_pfn);
2229
2230 /*
2231  * update_mem_hiwater
2232  *      - update per process rss and vm high water data
2233  */
2234 void update_mem_hiwater(struct task_struct *tsk)
2235 {
2236         if (tsk->mm) {
2237                 unsigned long rss = get_mm_rss(tsk->mm);
2238
2239                 if (tsk->mm->hiwater_rss < rss)
2240                         tsk->mm->hiwater_rss = rss;
2241                 if (tsk->mm->hiwater_vm < tsk->mm->total_vm)
2242                         tsk->mm->hiwater_vm = tsk->mm->total_vm;
2243         }
2244 }
2245
2246 #if !defined(__HAVE_ARCH_GATE_AREA)
2247
2248 #if defined(AT_SYSINFO_EHDR)
2249 static struct vm_area_struct gate_vma;
2250
2251 static int __init gate_vma_init(void)
2252 {
2253         gate_vma.vm_mm = NULL;
2254         gate_vma.vm_start = FIXADDR_USER_START;
2255         gate_vma.vm_end = FIXADDR_USER_END;
2256         gate_vma.vm_page_prot = PAGE_READONLY;
2257         gate_vma.vm_flags = VM_RESERVED;
2258         return 0;
2259 }
2260 __initcall(gate_vma_init);
2261 #endif
2262
2263 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2264 {
2265 #ifdef AT_SYSINFO_EHDR
2266         return &gate_vma;
2267 #else
2268         return NULL;
2269 #endif
2270 }
2271
2272 int in_gate_area_no_task(unsigned long addr)
2273 {
2274 #ifdef AT_SYSINFO_EHDR
2275         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2276                 return 1;
2277 #endif
2278         return 0;
2279 }
2280
2281 #endif  /* __HAVE_ARCH_GATE_AREA */