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