[PATCH] check_user_page_readable() deadlock fix
[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_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 *__follow_page(struct mm_struct *mm, unsigned long address,
780                         int read, int write, int accessed)
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 (accessed) {
822                                 if (write && !pte_dirty(pte) &&!PageDirty(page))
823                                         set_page_dirty(page);
824                                 mark_page_accessed(page);
825                         }
826                         return page;
827                 }
828         }
829
830 out:
831         return NULL;
832 }
833
834 inline struct page *
835 follow_page(struct mm_struct *mm, unsigned long address, int write)
836 {
837         return __follow_page(mm, address, 0, write, 1);
838 }
839
840 /*
841  * check_user_page_readable() can be called frm niterrupt context by oprofile,
842  * so we need to avoid taking any non-irq-safe locks
843  */
844 int check_user_page_readable(struct mm_struct *mm, unsigned long address)
845 {
846         return __follow_page(mm, address, 1, 0, 0) != NULL;
847 }
848 EXPORT_SYMBOL(check_user_page_readable);
849
850 static inline int
851 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
852                          unsigned long address)
853 {
854         pgd_t *pgd;
855         pud_t *pud;
856         pmd_t *pmd;
857
858         /* Check if the vma is for an anonymous mapping. */
859         if (vma->vm_ops && vma->vm_ops->nopage)
860                 return 0;
861
862         /* Check if page directory entry exists. */
863         pgd = pgd_offset(mm, address);
864         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
865                 return 1;
866
867         pud = pud_offset(pgd, address);
868         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
869                 return 1;
870
871         /* Check if page middle directory entry exists. */
872         pmd = pmd_offset(pud, address);
873         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
874                 return 1;
875
876         /* There is a pte slot for 'address' in 'mm'. */
877         return 0;
878 }
879
880 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
881                 unsigned long start, int len, int write, int force,
882                 struct page **pages, struct vm_area_struct **vmas)
883 {
884         int i;
885         unsigned int flags;
886
887         /* 
888          * Require read or write permissions.
889          * If 'force' is set, we only require the "MAY" flags.
890          */
891         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
892         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
893         i = 0;
894
895         do {
896                 struct vm_area_struct * vma;
897
898                 vma = find_extend_vma(mm, start);
899                 if (!vma && in_gate_area(tsk, start)) {
900                         unsigned long pg = start & PAGE_MASK;
901                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
902                         pgd_t *pgd;
903                         pud_t *pud;
904                         pmd_t *pmd;
905                         pte_t *pte;
906                         if (write) /* user gate pages are read-only */
907                                 return i ? : -EFAULT;
908                         if (pg > TASK_SIZE)
909                                 pgd = pgd_offset_k(pg);
910                         else
911                                 pgd = pgd_offset_gate(mm, pg);
912                         BUG_ON(pgd_none(*pgd));
913                         pud = pud_offset(pgd, pg);
914                         BUG_ON(pud_none(*pud));
915                         pmd = pmd_offset(pud, pg);
916                         BUG_ON(pmd_none(*pmd));
917                         pte = pte_offset_map(pmd, pg);
918                         BUG_ON(pte_none(*pte));
919                         if (pages) {
920                                 pages[i] = pte_page(*pte);
921                                 get_page(pages[i]);
922                         }
923                         pte_unmap(pte);
924                         if (vmas)
925                                 vmas[i] = gate_vma;
926                         i++;
927                         start += PAGE_SIZE;
928                         len--;
929                         continue;
930                 }
931
932                 if (!vma || (vma->vm_flags & VM_IO)
933                                 || !(flags & vma->vm_flags))
934                         return i ? : -EFAULT;
935
936                 if (is_vm_hugetlb_page(vma)) {
937                         i = follow_hugetlb_page(mm, vma, pages, vmas,
938                                                 &start, &len, i);
939                         continue;
940                 }
941                 spin_lock(&mm->page_table_lock);
942                 do {
943                         struct page *page;
944                         int lookup_write = write;
945
946                         cond_resched_lock(&mm->page_table_lock);
947                         while (!(page = follow_page(mm, start, lookup_write))) {
948                                 /*
949                                  * Shortcut for anonymous pages. We don't want
950                                  * to force the creation of pages tables for
951                                  * insanely big anonymously mapped areas that
952                                  * nobody touched so far. This is important
953                                  * for doing a core dump for these mappings.
954                                  */
955                                 if (!lookup_write &&
956                                     untouched_anonymous_page(mm,vma,start)) {
957                                         page = ZERO_PAGE(start);
958                                         break;
959                                 }
960                                 spin_unlock(&mm->page_table_lock);
961                                 switch (handle_mm_fault(mm,vma,start,write)) {
962                                 case VM_FAULT_MINOR:
963                                         tsk->min_flt++;
964                                         break;
965                                 case VM_FAULT_MAJOR:
966                                         tsk->maj_flt++;
967                                         break;
968                                 case VM_FAULT_SIGBUS:
969                                         return i ? i : -EFAULT;
970                                 case VM_FAULT_OOM:
971                                         return i ? i : -ENOMEM;
972                                 default:
973                                         BUG();
974                                 }
975                                 /*
976                                  * Now that we have performed a write fault
977                                  * and surely no longer have a shared page we
978                                  * shouldn't write, we shouldn't ignore an
979                                  * unwritable page in the page table if
980                                  * we are forcing write access.
981                                  */
982                                 lookup_write = write && !force;
983                                 spin_lock(&mm->page_table_lock);
984                         }
985                         if (pages) {
986                                 pages[i] = page;
987                                 flush_dcache_page(page);
988                                 if (!PageReserved(page))
989                                         page_cache_get(page);
990                         }
991                         if (vmas)
992                                 vmas[i] = vma;
993                         i++;
994                         start += PAGE_SIZE;
995                         len--;
996                 } while (len && start < vma->vm_end);
997                 spin_unlock(&mm->page_table_lock);
998         } while (len);
999         return i;
1000 }
1001 EXPORT_SYMBOL(get_user_pages);
1002
1003 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1004                         unsigned long addr, unsigned long end, pgprot_t prot)
1005 {
1006         pte_t *pte;
1007
1008         pte = pte_alloc_map(mm, pmd, addr);
1009         if (!pte)
1010                 return -ENOMEM;
1011         do {
1012                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(addr), prot));
1013                 BUG_ON(!pte_none(*pte));
1014                 set_pte_at(mm, addr, pte, zero_pte);
1015         } while (pte++, addr += PAGE_SIZE, addr != end);
1016         pte_unmap(pte - 1);
1017         return 0;
1018 }
1019
1020 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1021                         unsigned long addr, unsigned long end, pgprot_t prot)
1022 {
1023         pmd_t *pmd;
1024         unsigned long next;
1025
1026         pmd = pmd_alloc(mm, pud, addr);
1027         if (!pmd)
1028                 return -ENOMEM;
1029         do {
1030                 next = pmd_addr_end(addr, end);
1031                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1032                         return -ENOMEM;
1033         } while (pmd++, addr = next, addr != end);
1034         return 0;
1035 }
1036
1037 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1038                         unsigned long addr, unsigned long end, pgprot_t prot)
1039 {
1040         pud_t *pud;
1041         unsigned long next;
1042
1043         pud = pud_alloc(mm, pgd, addr);
1044         if (!pud)
1045                 return -ENOMEM;
1046         do {
1047                 next = pud_addr_end(addr, end);
1048                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1049                         return -ENOMEM;
1050         } while (pud++, addr = next, addr != end);
1051         return 0;
1052 }
1053
1054 int zeromap_page_range(struct vm_area_struct *vma,
1055                         unsigned long addr, unsigned long size, pgprot_t prot)
1056 {
1057         pgd_t *pgd;
1058         unsigned long next;
1059         unsigned long end = addr + size;
1060         struct mm_struct *mm = vma->vm_mm;
1061         int err;
1062
1063         BUG_ON(addr >= end);
1064         pgd = pgd_offset(mm, addr);
1065         flush_cache_range(vma, addr, end);
1066         spin_lock(&mm->page_table_lock);
1067         do {
1068                 next = pgd_addr_end(addr, end);
1069                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1070                 if (err)
1071                         break;
1072         } while (pgd++, addr = next, addr != end);
1073         spin_unlock(&mm->page_table_lock);
1074         return err;
1075 }
1076
1077 /*
1078  * maps a range of physical memory into the requested pages. the old
1079  * mappings are removed. any references to nonexistent pages results
1080  * in null mappings (currently treated as "copy-on-access")
1081  */
1082 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1083                         unsigned long addr, unsigned long end,
1084                         unsigned long pfn, pgprot_t prot)
1085 {
1086         pte_t *pte;
1087
1088         pte = pte_alloc_map(mm, pmd, addr);
1089         if (!pte)
1090                 return -ENOMEM;
1091         do {
1092                 BUG_ON(!pte_none(*pte));
1093                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
1094                         set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1095                 pfn++;
1096         } while (pte++, addr += PAGE_SIZE, addr != end);
1097         pte_unmap(pte - 1);
1098         return 0;
1099 }
1100
1101 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1102                         unsigned long addr, unsigned long end,
1103                         unsigned long pfn, pgprot_t prot)
1104 {
1105         pmd_t *pmd;
1106         unsigned long next;
1107
1108         pfn -= addr >> PAGE_SHIFT;
1109         pmd = pmd_alloc(mm, pud, addr);
1110         if (!pmd)
1111                 return -ENOMEM;
1112         do {
1113                 next = pmd_addr_end(addr, end);
1114                 if (remap_pte_range(mm, pmd, addr, next,
1115                                 pfn + (addr >> PAGE_SHIFT), prot))
1116                         return -ENOMEM;
1117         } while (pmd++, addr = next, addr != end);
1118         return 0;
1119 }
1120
1121 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1122                         unsigned long addr, unsigned long end,
1123                         unsigned long pfn, pgprot_t prot)
1124 {
1125         pud_t *pud;
1126         unsigned long next;
1127
1128         pfn -= addr >> PAGE_SHIFT;
1129         pud = pud_alloc(mm, pgd, addr);
1130         if (!pud)
1131                 return -ENOMEM;
1132         do {
1133                 next = pud_addr_end(addr, end);
1134                 if (remap_pmd_range(mm, pud, addr, next,
1135                                 pfn + (addr >> PAGE_SHIFT), prot))
1136                         return -ENOMEM;
1137         } while (pud++, addr = next, addr != end);
1138         return 0;
1139 }
1140
1141 /*  Note: this is only safe if the mm semaphore is held when called. */
1142 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1143                     unsigned long pfn, unsigned long size, pgprot_t prot)
1144 {
1145         pgd_t *pgd;
1146         unsigned long next;
1147         unsigned long end = addr + PAGE_ALIGN(size);
1148         struct mm_struct *mm = vma->vm_mm;
1149         int err;
1150
1151         /*
1152          * Physically remapped pages are special. Tell the
1153          * rest of the world about it:
1154          *   VM_IO tells people not to look at these pages
1155          *      (accesses can have side effects).
1156          *   VM_RESERVED tells swapout not to try to touch
1157          *      this region.
1158          */
1159         vma->vm_flags |= VM_IO | VM_RESERVED;
1160
1161         BUG_ON(addr >= end);
1162         pfn -= addr >> PAGE_SHIFT;
1163         pgd = pgd_offset(mm, addr);
1164         flush_cache_range(vma, addr, end);
1165         spin_lock(&mm->page_table_lock);
1166         do {
1167                 next = pgd_addr_end(addr, end);
1168                 err = remap_pud_range(mm, pgd, addr, next,
1169                                 pfn + (addr >> PAGE_SHIFT), prot);
1170                 if (err)
1171                         break;
1172         } while (pgd++, addr = next, addr != end);
1173         spin_unlock(&mm->page_table_lock);
1174         return err;
1175 }
1176 EXPORT_SYMBOL(remap_pfn_range);
1177
1178 /*
1179  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1180  * servicing faults for write access.  In the normal case, do always want
1181  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1182  * that do not have writing enabled, when used by access_process_vm.
1183  */
1184 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1185 {
1186         if (likely(vma->vm_flags & VM_WRITE))
1187                 pte = pte_mkwrite(pte);
1188         return pte;
1189 }
1190
1191 /*
1192  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1193  */
1194 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address, 
1195                 pte_t *page_table)
1196 {
1197         pte_t entry;
1198
1199         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1200                               vma);
1201         ptep_establish(vma, address, page_table, entry);
1202         update_mmu_cache(vma, address, entry);
1203         lazy_mmu_prot_update(entry);
1204 }
1205
1206 /*
1207  * This routine handles present pages, when users try to write
1208  * to a shared page. It is done by copying the page to a new address
1209  * and decrementing the shared-page counter for the old page.
1210  *
1211  * Goto-purists beware: the only reason for goto's here is that it results
1212  * in better assembly code.. The "default" path will see no jumps at all.
1213  *
1214  * Note that this routine assumes that the protection checks have been
1215  * done by the caller (the low-level page fault routine in most cases).
1216  * Thus we can safely just mark it writable once we've done any necessary
1217  * COW.
1218  *
1219  * We also mark the page dirty at this point even though the page will
1220  * change only once the write actually happens. This avoids a few races,
1221  * and potentially makes it more efficient.
1222  *
1223  * We hold the mm semaphore and the page_table_lock on entry and exit
1224  * with the page_table_lock released.
1225  */
1226 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1227         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1228 {
1229         struct page *old_page, *new_page;
1230         unsigned long pfn = pte_pfn(pte);
1231         pte_t entry;
1232
1233         if (unlikely(!pfn_valid(pfn))) {
1234                 /*
1235                  * This should really halt the system so it can be debugged or
1236                  * at least the kernel stops what it's doing before it corrupts
1237                  * data, but for the moment just pretend this is OOM.
1238                  */
1239                 pte_unmap(page_table);
1240                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1241                                 address);
1242                 spin_unlock(&mm->page_table_lock);
1243                 return VM_FAULT_OOM;
1244         }
1245         old_page = pfn_to_page(pfn);
1246
1247         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1248                 int reuse = can_share_swap_page(old_page);
1249                 unlock_page(old_page);
1250                 if (reuse) {
1251                         flush_cache_page(vma, address, pfn);
1252                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1253                                               vma);
1254                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1255                         update_mmu_cache(vma, address, entry);
1256                         lazy_mmu_prot_update(entry);
1257                         pte_unmap(page_table);
1258                         spin_unlock(&mm->page_table_lock);
1259                         return VM_FAULT_MINOR;
1260                 }
1261         }
1262         pte_unmap(page_table);
1263
1264         /*
1265          * Ok, we need to copy. Oh, well..
1266          */
1267         if (!PageReserved(old_page))
1268                 page_cache_get(old_page);
1269         spin_unlock(&mm->page_table_lock);
1270
1271         if (unlikely(anon_vma_prepare(vma)))
1272                 goto no_new_page;
1273         if (old_page == ZERO_PAGE(address)) {
1274                 new_page = alloc_zeroed_user_highpage(vma, address);
1275                 if (!new_page)
1276                         goto no_new_page;
1277         } else {
1278                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1279                 if (!new_page)
1280                         goto no_new_page;
1281                 copy_user_highpage(new_page, old_page, address);
1282         }
1283         /*
1284          * Re-check the pte - we dropped the lock
1285          */
1286         spin_lock(&mm->page_table_lock);
1287         page_table = pte_offset_map(pmd, address);
1288         if (likely(pte_same(*page_table, pte))) {
1289                 if (PageAnon(old_page))
1290                         dec_mm_counter(mm, anon_rss);
1291                 if (PageReserved(old_page))
1292                         inc_mm_counter(mm, rss);
1293                 else
1294                         page_remove_rmap(old_page);
1295                 flush_cache_page(vma, address, pfn);
1296                 break_cow(vma, new_page, address, page_table);
1297                 lru_cache_add_active(new_page);
1298                 page_add_anon_rmap(new_page, vma, address);
1299
1300                 /* Free the old page.. */
1301                 new_page = old_page;
1302         }
1303         pte_unmap(page_table);
1304         page_cache_release(new_page);
1305         page_cache_release(old_page);
1306         spin_unlock(&mm->page_table_lock);
1307         return VM_FAULT_MINOR;
1308
1309 no_new_page:
1310         page_cache_release(old_page);
1311         return VM_FAULT_OOM;
1312 }
1313
1314 /*
1315  * Helper functions for unmap_mapping_range().
1316  *
1317  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1318  *
1319  * We have to restart searching the prio_tree whenever we drop the lock,
1320  * since the iterator is only valid while the lock is held, and anyway
1321  * a later vma might be split and reinserted earlier while lock dropped.
1322  *
1323  * The list of nonlinear vmas could be handled more efficiently, using
1324  * a placeholder, but handle it in the same way until a need is shown.
1325  * It is important to search the prio_tree before nonlinear list: a vma
1326  * may become nonlinear and be shifted from prio_tree to nonlinear list
1327  * while the lock is dropped; but never shifted from list to prio_tree.
1328  *
1329  * In order to make forward progress despite restarting the search,
1330  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1331  * quickly skip it next time around.  Since the prio_tree search only
1332  * shows us those vmas affected by unmapping the range in question, we
1333  * can't efficiently keep all vmas in step with mapping->truncate_count:
1334  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1335  * mapping->truncate_count and vma->vm_truncate_count are protected by
1336  * i_mmap_lock.
1337  *
1338  * In order to make forward progress despite repeatedly restarting some
1339  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1340  * and restart from that address when we reach that vma again.  It might
1341  * have been split or merged, shrunk or extended, but never shifted: so
1342  * restart_addr remains valid so long as it remains in the vma's range.
1343  * unmap_mapping_range forces truncate_count to leap over page-aligned
1344  * values so we can save vma's restart_addr in its truncate_count field.
1345  */
1346 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1347
1348 static void reset_vma_truncate_counts(struct address_space *mapping)
1349 {
1350         struct vm_area_struct *vma;
1351         struct prio_tree_iter iter;
1352
1353         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1354                 vma->vm_truncate_count = 0;
1355         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1356                 vma->vm_truncate_count = 0;
1357 }
1358
1359 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1360                 unsigned long start_addr, unsigned long end_addr,
1361                 struct zap_details *details)
1362 {
1363         unsigned long restart_addr;
1364         int need_break;
1365
1366 again:
1367         restart_addr = vma->vm_truncate_count;
1368         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1369                 start_addr = restart_addr;
1370                 if (start_addr >= end_addr) {
1371                         /* Top of vma has been split off since last time */
1372                         vma->vm_truncate_count = details->truncate_count;
1373                         return 0;
1374                 }
1375         }
1376
1377         restart_addr = zap_page_range(vma, start_addr,
1378                                         end_addr - start_addr, details);
1379
1380         /*
1381          * We cannot rely on the break test in unmap_vmas:
1382          * on the one hand, we don't want to restart our loop
1383          * just because that broke out for the page_table_lock;
1384          * on the other hand, it does no test when vma is small.
1385          */
1386         need_break = need_resched() ||
1387                         need_lockbreak(details->i_mmap_lock);
1388
1389         if (restart_addr >= end_addr) {
1390                 /* We have now completed this vma: mark it so */
1391                 vma->vm_truncate_count = details->truncate_count;
1392                 if (!need_break)
1393                         return 0;
1394         } else {
1395                 /* Note restart_addr in vma's truncate_count field */
1396                 vma->vm_truncate_count = restart_addr;
1397                 if (!need_break)
1398                         goto again;
1399         }
1400
1401         spin_unlock(details->i_mmap_lock);
1402         cond_resched();
1403         spin_lock(details->i_mmap_lock);
1404         return -EINTR;
1405 }
1406
1407 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1408                                             struct zap_details *details)
1409 {
1410         struct vm_area_struct *vma;
1411         struct prio_tree_iter iter;
1412         pgoff_t vba, vea, zba, zea;
1413
1414 restart:
1415         vma_prio_tree_foreach(vma, &iter, root,
1416                         details->first_index, details->last_index) {
1417                 /* Skip quickly over those we have already dealt with */
1418                 if (vma->vm_truncate_count == details->truncate_count)
1419                         continue;
1420
1421                 vba = vma->vm_pgoff;
1422                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1423                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1424                 zba = details->first_index;
1425                 if (zba < vba)
1426                         zba = vba;
1427                 zea = details->last_index;
1428                 if (zea > vea)
1429                         zea = vea;
1430
1431                 if (unmap_mapping_range_vma(vma,
1432                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1433                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1434                                 details) < 0)
1435                         goto restart;
1436         }
1437 }
1438
1439 static inline void unmap_mapping_range_list(struct list_head *head,
1440                                             struct zap_details *details)
1441 {
1442         struct vm_area_struct *vma;
1443
1444         /*
1445          * In nonlinear VMAs there is no correspondence between virtual address
1446          * offset and file offset.  So we must perform an exhaustive search
1447          * across *all* the pages in each nonlinear VMA, not just the pages
1448          * whose virtual address lies outside the file truncation point.
1449          */
1450 restart:
1451         list_for_each_entry(vma, head, shared.vm_set.list) {
1452                 /* Skip quickly over those we have already dealt with */
1453                 if (vma->vm_truncate_count == details->truncate_count)
1454                         continue;
1455                 details->nonlinear_vma = vma;
1456                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1457                                         vma->vm_end, details) < 0)
1458                         goto restart;
1459         }
1460 }
1461
1462 /**
1463  * unmap_mapping_range - unmap the portion of all mmaps
1464  * in the specified address_space corresponding to the specified
1465  * page range in the underlying file.
1466  * @mapping: the address space containing mmaps to be unmapped.
1467  * @holebegin: byte in first page to unmap, relative to the start of
1468  * the underlying file.  This will be rounded down to a PAGE_SIZE
1469  * boundary.  Note that this is different from vmtruncate(), which
1470  * must keep the partial page.  In contrast, we must get rid of
1471  * partial pages.
1472  * @holelen: size of prospective hole in bytes.  This will be rounded
1473  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1474  * end of the file.
1475  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1476  * but 0 when invalidating pagecache, don't throw away private data.
1477  */
1478 void unmap_mapping_range(struct address_space *mapping,
1479                 loff_t const holebegin, loff_t const holelen, int even_cows)
1480 {
1481         struct zap_details details;
1482         pgoff_t hba = holebegin >> PAGE_SHIFT;
1483         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1484
1485         /* Check for overflow. */
1486         if (sizeof(holelen) > sizeof(hlen)) {
1487                 long long holeend =
1488                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1489                 if (holeend & ~(long long)ULONG_MAX)
1490                         hlen = ULONG_MAX - hba + 1;
1491         }
1492
1493         details.check_mapping = even_cows? NULL: mapping;
1494         details.nonlinear_vma = NULL;
1495         details.first_index = hba;
1496         details.last_index = hba + hlen - 1;
1497         if (details.last_index < details.first_index)
1498                 details.last_index = ULONG_MAX;
1499         details.i_mmap_lock = &mapping->i_mmap_lock;
1500
1501         spin_lock(&mapping->i_mmap_lock);
1502
1503         /* serialize i_size write against truncate_count write */
1504         smp_wmb();
1505         /* Protect against page faults, and endless unmapping loops */
1506         mapping->truncate_count++;
1507         /*
1508          * For archs where spin_lock has inclusive semantics like ia64
1509          * this smp_mb() will prevent to read pagetable contents
1510          * before the truncate_count increment is visible to
1511          * other cpus.
1512          */
1513         smp_mb();
1514         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1515                 if (mapping->truncate_count == 0)
1516                         reset_vma_truncate_counts(mapping);
1517                 mapping->truncate_count++;
1518         }
1519         details.truncate_count = mapping->truncate_count;
1520
1521         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1522                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1523         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1524                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1525         spin_unlock(&mapping->i_mmap_lock);
1526 }
1527 EXPORT_SYMBOL(unmap_mapping_range);
1528
1529 /*
1530  * Handle all mappings that got truncated by a "truncate()"
1531  * system call.
1532  *
1533  * NOTE! We have to be ready to update the memory sharing
1534  * between the file and the memory map for a potential last
1535  * incomplete page.  Ugly, but necessary.
1536  */
1537 int vmtruncate(struct inode * inode, loff_t offset)
1538 {
1539         struct address_space *mapping = inode->i_mapping;
1540         unsigned long limit;
1541
1542         if (inode->i_size < offset)
1543                 goto do_expand;
1544         /*
1545          * truncation of in-use swapfiles is disallowed - it would cause
1546          * subsequent swapout to scribble on the now-freed blocks.
1547          */
1548         if (IS_SWAPFILE(inode))
1549                 goto out_busy;
1550         i_size_write(inode, offset);
1551         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1552         truncate_inode_pages(mapping, offset);
1553         goto out_truncate;
1554
1555 do_expand:
1556         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1557         if (limit != RLIM_INFINITY && offset > limit)
1558                 goto out_sig;
1559         if (offset > inode->i_sb->s_maxbytes)
1560                 goto out_big;
1561         i_size_write(inode, offset);
1562
1563 out_truncate:
1564         if (inode->i_op && inode->i_op->truncate)
1565                 inode->i_op->truncate(inode);
1566         return 0;
1567 out_sig:
1568         send_sig(SIGXFSZ, current, 0);
1569 out_big:
1570         return -EFBIG;
1571 out_busy:
1572         return -ETXTBSY;
1573 }
1574
1575 EXPORT_SYMBOL(vmtruncate);
1576
1577 /* 
1578  * Primitive swap readahead code. We simply read an aligned block of
1579  * (1 << page_cluster) entries in the swap area. This method is chosen
1580  * because it doesn't cost us any seek time.  We also make sure to queue
1581  * the 'original' request together with the readahead ones...  
1582  *
1583  * This has been extended to use the NUMA policies from the mm triggering
1584  * the readahead.
1585  *
1586  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1587  */
1588 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1589 {
1590 #ifdef CONFIG_NUMA
1591         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1592 #endif
1593         int i, num;
1594         struct page *new_page;
1595         unsigned long offset;
1596
1597         /*
1598          * Get the number of handles we should do readahead io to.
1599          */
1600         num = valid_swaphandles(entry, &offset);
1601         for (i = 0; i < num; offset++, i++) {
1602                 /* Ok, do the async read-ahead now */
1603                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1604                                                            offset), vma, addr);
1605                 if (!new_page)
1606                         break;
1607                 page_cache_release(new_page);
1608 #ifdef CONFIG_NUMA
1609                 /*
1610                  * Find the next applicable VMA for the NUMA policy.
1611                  */
1612                 addr += PAGE_SIZE;
1613                 if (addr == 0)
1614                         vma = NULL;
1615                 if (vma) {
1616                         if (addr >= vma->vm_end) {
1617                                 vma = next_vma;
1618                                 next_vma = vma ? vma->vm_next : NULL;
1619                         }
1620                         if (vma && addr < vma->vm_start)
1621                                 vma = NULL;
1622                 } else {
1623                         if (next_vma && addr >= next_vma->vm_start) {
1624                                 vma = next_vma;
1625                                 next_vma = vma->vm_next;
1626                         }
1627                 }
1628 #endif
1629         }
1630         lru_add_drain();        /* Push any new pages onto the LRU now */
1631 }
1632
1633 /*
1634  * We hold the mm semaphore and the page_table_lock on entry and
1635  * should release the pagetable lock on exit..
1636  */
1637 static int do_swap_page(struct mm_struct * mm,
1638         struct vm_area_struct * vma, unsigned long address,
1639         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1640 {
1641         struct page *page;
1642         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1643         pte_t pte;
1644         int ret = VM_FAULT_MINOR;
1645
1646         pte_unmap(page_table);
1647         spin_unlock(&mm->page_table_lock);
1648         page = lookup_swap_cache(entry);
1649         if (!page) {
1650                 swapin_readahead(entry, address, vma);
1651                 page = read_swap_cache_async(entry, vma, address);
1652                 if (!page) {
1653                         /*
1654                          * Back out if somebody else faulted in this pte while
1655                          * we released the page table lock.
1656                          */
1657                         spin_lock(&mm->page_table_lock);
1658                         page_table = pte_offset_map(pmd, address);
1659                         if (likely(pte_same(*page_table, orig_pte)))
1660                                 ret = VM_FAULT_OOM;
1661                         else
1662                                 ret = VM_FAULT_MINOR;
1663                         pte_unmap(page_table);
1664                         spin_unlock(&mm->page_table_lock);
1665                         goto out;
1666                 }
1667
1668                 /* Had to read the page from swap area: Major fault */
1669                 ret = VM_FAULT_MAJOR;
1670                 inc_page_state(pgmajfault);
1671                 grab_swap_token();
1672         }
1673
1674         mark_page_accessed(page);
1675         lock_page(page);
1676
1677         /*
1678          * Back out if somebody else faulted in this pte while we
1679          * released the page table lock.
1680          */
1681         spin_lock(&mm->page_table_lock);
1682         page_table = pte_offset_map(pmd, address);
1683         if (unlikely(!pte_same(*page_table, orig_pte))) {
1684                 ret = VM_FAULT_MINOR;
1685                 goto out_nomap;
1686         }
1687
1688         if (unlikely(!PageUptodate(page))) {
1689                 ret = VM_FAULT_SIGBUS;
1690                 goto out_nomap;
1691         }
1692
1693         /* The page isn't present yet, go ahead with the fault. */
1694
1695         inc_mm_counter(mm, rss);
1696         pte = mk_pte(page, vma->vm_page_prot);
1697         if (write_access && can_share_swap_page(page)) {
1698                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1699                 write_access = 0;
1700         }
1701
1702         flush_icache_page(vma, page);
1703         set_pte_at(mm, address, page_table, pte);
1704         page_add_anon_rmap(page, vma, address);
1705
1706         swap_free(entry);
1707         if (vm_swap_full())
1708                 remove_exclusive_swap_page(page);
1709         unlock_page(page);
1710
1711         if (write_access) {
1712                 if (do_wp_page(mm, vma, address,
1713                                 page_table, pmd, pte) == VM_FAULT_OOM)
1714                         ret = VM_FAULT_OOM;
1715                 goto out;
1716         }
1717
1718         /* No need to invalidate - it was non-present before */
1719         update_mmu_cache(vma, address, pte);
1720         lazy_mmu_prot_update(pte);
1721         pte_unmap(page_table);
1722         spin_unlock(&mm->page_table_lock);
1723 out:
1724         return ret;
1725 out_nomap:
1726         pte_unmap(page_table);
1727         spin_unlock(&mm->page_table_lock);
1728         unlock_page(page);
1729         page_cache_release(page);
1730         goto out;
1731 }
1732
1733 /*
1734  * We are called with the MM semaphore and page_table_lock
1735  * spinlock held to protect against concurrent faults in
1736  * multithreaded programs. 
1737  */
1738 static int
1739 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1740                 pte_t *page_table, pmd_t *pmd, int write_access,
1741                 unsigned long addr)
1742 {
1743         pte_t entry;
1744         struct page * page = ZERO_PAGE(addr);
1745
1746         /* Read-only mapping of ZERO_PAGE. */
1747         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1748
1749         /* ..except if it's a write access */
1750         if (write_access) {
1751                 /* Allocate our own private page. */
1752                 pte_unmap(page_table);
1753                 spin_unlock(&mm->page_table_lock);
1754
1755                 if (unlikely(anon_vma_prepare(vma)))
1756                         goto no_mem;
1757                 page = alloc_zeroed_user_highpage(vma, addr);
1758                 if (!page)
1759                         goto no_mem;
1760
1761                 spin_lock(&mm->page_table_lock);
1762                 page_table = pte_offset_map(pmd, addr);
1763
1764                 if (!pte_none(*page_table)) {
1765                         pte_unmap(page_table);
1766                         page_cache_release(page);
1767                         spin_unlock(&mm->page_table_lock);
1768                         goto out;
1769                 }
1770                 inc_mm_counter(mm, rss);
1771                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1772                                                          vma->vm_page_prot)),
1773                                       vma);
1774                 lru_cache_add_active(page);
1775                 SetPageReferenced(page);
1776                 page_add_anon_rmap(page, vma, addr);
1777         }
1778
1779         set_pte_at(mm, addr, page_table, entry);
1780         pte_unmap(page_table);
1781
1782         /* No need to invalidate - it was non-present before */
1783         update_mmu_cache(vma, addr, entry);
1784         lazy_mmu_prot_update(entry);
1785         spin_unlock(&mm->page_table_lock);
1786 out:
1787         return VM_FAULT_MINOR;
1788 no_mem:
1789         return VM_FAULT_OOM;
1790 }
1791
1792 /*
1793  * do_no_page() tries to create a new page mapping. It aggressively
1794  * tries to share with existing pages, but makes a separate copy if
1795  * the "write_access" parameter is true in order to avoid the next
1796  * page fault.
1797  *
1798  * As this is called only for pages that do not currently exist, we
1799  * do not need to flush old virtual caches or the TLB.
1800  *
1801  * This is called with the MM semaphore held and the page table
1802  * spinlock held. Exit with the spinlock released.
1803  */
1804 static int
1805 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1806         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1807 {
1808         struct page * new_page;
1809         struct address_space *mapping = NULL;
1810         pte_t entry;
1811         unsigned int sequence = 0;
1812         int ret = VM_FAULT_MINOR;
1813         int anon = 0;
1814
1815         if (!vma->vm_ops || !vma->vm_ops->nopage)
1816                 return do_anonymous_page(mm, vma, page_table,
1817                                         pmd, write_access, address);
1818         pte_unmap(page_table);
1819         spin_unlock(&mm->page_table_lock);
1820
1821         if (vma->vm_file) {
1822                 mapping = vma->vm_file->f_mapping;
1823                 sequence = mapping->truncate_count;
1824                 smp_rmb(); /* serializes i_size against truncate_count */
1825         }
1826 retry:
1827         cond_resched();
1828         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1829         /*
1830          * No smp_rmb is needed here as long as there's a full
1831          * spin_lock/unlock sequence inside the ->nopage callback
1832          * (for the pagecache lookup) that acts as an implicit
1833          * smp_mb() and prevents the i_size read to happen
1834          * after the next truncate_count read.
1835          */
1836
1837         /* no page was available -- either SIGBUS or OOM */
1838         if (new_page == NOPAGE_SIGBUS)
1839                 return VM_FAULT_SIGBUS;
1840         if (new_page == NOPAGE_OOM)
1841                 return VM_FAULT_OOM;
1842
1843         /*
1844          * Should we do an early C-O-W break?
1845          */
1846         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1847                 struct page *page;
1848
1849                 if (unlikely(anon_vma_prepare(vma)))
1850                         goto oom;
1851                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1852                 if (!page)
1853                         goto oom;
1854                 copy_user_highpage(page, new_page, address);
1855                 page_cache_release(new_page);
1856                 new_page = page;
1857                 anon = 1;
1858         }
1859
1860         spin_lock(&mm->page_table_lock);
1861         /*
1862          * For a file-backed vma, someone could have truncated or otherwise
1863          * invalidated this page.  If unmap_mapping_range got called,
1864          * retry getting the page.
1865          */
1866         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1867                 sequence = mapping->truncate_count;
1868                 spin_unlock(&mm->page_table_lock);
1869                 page_cache_release(new_page);
1870                 goto retry;
1871         }
1872         page_table = pte_offset_map(pmd, address);
1873
1874         /*
1875          * This silly early PAGE_DIRTY setting removes a race
1876          * due to the bad i386 page protection. But it's valid
1877          * for other architectures too.
1878          *
1879          * Note that if write_access is true, we either now have
1880          * an exclusive copy of the page, or this is a shared mapping,
1881          * so we can make it writable and dirty to avoid having to
1882          * handle that later.
1883          */
1884         /* Only go through if we didn't race with anybody else... */
1885         if (pte_none(*page_table)) {
1886                 if (!PageReserved(new_page))
1887                         inc_mm_counter(mm, rss);
1888
1889                 flush_icache_page(vma, new_page);
1890                 entry = mk_pte(new_page, vma->vm_page_prot);
1891                 if (write_access)
1892                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1893                 set_pte_at(mm, address, page_table, entry);
1894                 if (anon) {
1895                         lru_cache_add_active(new_page);
1896                         page_add_anon_rmap(new_page, vma, address);
1897                 } else
1898                         page_add_file_rmap(new_page);
1899                 pte_unmap(page_table);
1900         } else {
1901                 /* One of our sibling threads was faster, back out. */
1902                 pte_unmap(page_table);
1903                 page_cache_release(new_page);
1904                 spin_unlock(&mm->page_table_lock);
1905                 goto out;
1906         }
1907
1908         /* no need to invalidate: a not-present page shouldn't be cached */
1909         update_mmu_cache(vma, address, entry);
1910         lazy_mmu_prot_update(entry);
1911         spin_unlock(&mm->page_table_lock);
1912 out:
1913         return ret;
1914 oom:
1915         page_cache_release(new_page);
1916         ret = VM_FAULT_OOM;
1917         goto out;
1918 }
1919
1920 /*
1921  * Fault of a previously existing named mapping. Repopulate the pte
1922  * from the encoded file_pte if possible. This enables swappable
1923  * nonlinear vmas.
1924  */
1925 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1926         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1927 {
1928         unsigned long pgoff;
1929         int err;
1930
1931         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1932         /*
1933          * Fall back to the linear mapping if the fs does not support
1934          * ->populate:
1935          */
1936         if (!vma->vm_ops || !vma->vm_ops->populate || 
1937                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1938                 pte_clear(mm, address, pte);
1939                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1940         }
1941
1942         pgoff = pte_to_pgoff(*pte);
1943
1944         pte_unmap(pte);
1945         spin_unlock(&mm->page_table_lock);
1946
1947         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1948         if (err == -ENOMEM)
1949                 return VM_FAULT_OOM;
1950         if (err)
1951                 return VM_FAULT_SIGBUS;
1952         return VM_FAULT_MAJOR;
1953 }
1954
1955 /*
1956  * These routines also need to handle stuff like marking pages dirty
1957  * and/or accessed for architectures that don't do it in hardware (most
1958  * RISC architectures).  The early dirtying is also good on the i386.
1959  *
1960  * There is also a hook called "update_mmu_cache()" that architectures
1961  * with external mmu caches can use to update those (ie the Sparc or
1962  * PowerPC hashed page tables that act as extended TLBs).
1963  *
1964  * Note the "page_table_lock". It is to protect against kswapd removing
1965  * pages from under us. Note that kswapd only ever _removes_ pages, never
1966  * adds them. As such, once we have noticed that the page is not present,
1967  * we can drop the lock early.
1968  *
1969  * The adding of pages is protected by the MM semaphore (which we hold),
1970  * so we don't need to worry about a page being suddenly been added into
1971  * our VM.
1972  *
1973  * We enter with the pagetable spinlock held, we are supposed to
1974  * release it when done.
1975  */
1976 static inline int handle_pte_fault(struct mm_struct *mm,
1977         struct vm_area_struct * vma, unsigned long address,
1978         int write_access, pte_t *pte, pmd_t *pmd)
1979 {
1980         pte_t entry;
1981
1982         entry = *pte;
1983         if (!pte_present(entry)) {
1984                 /*
1985                  * If it truly wasn't present, we know that kswapd
1986                  * and the PTE updates will not touch it later. So
1987                  * drop the lock.
1988                  */
1989                 if (pte_none(entry))
1990                         return do_no_page(mm, vma, address, write_access, pte, pmd);
1991                 if (pte_file(entry))
1992                         return do_file_page(mm, vma, address, write_access, pte, pmd);
1993                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1994         }
1995
1996         if (write_access) {
1997                 if (!pte_write(entry))
1998                         return do_wp_page(mm, vma, address, pte, pmd, entry);
1999
2000                 entry = pte_mkdirty(entry);
2001         }
2002         entry = pte_mkyoung(entry);
2003         ptep_set_access_flags(vma, address, pte, entry, write_access);
2004         update_mmu_cache(vma, address, entry);
2005         lazy_mmu_prot_update(entry);
2006         pte_unmap(pte);
2007         spin_unlock(&mm->page_table_lock);
2008         return VM_FAULT_MINOR;
2009 }
2010
2011 /*
2012  * By the time we get here, we already hold the mm semaphore
2013  */
2014 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
2015                 unsigned long address, int write_access)
2016 {
2017         pgd_t *pgd;
2018         pud_t *pud;
2019         pmd_t *pmd;
2020         pte_t *pte;
2021
2022         __set_current_state(TASK_RUNNING);
2023
2024         inc_page_state(pgfault);
2025
2026         if (is_vm_hugetlb_page(vma))
2027                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
2028
2029         /*
2030          * We need the page table lock to synchronize with kswapd
2031          * and the SMP-safe atomic PTE updates.
2032          */
2033         pgd = pgd_offset(mm, address);
2034         spin_lock(&mm->page_table_lock);
2035
2036         pud = pud_alloc(mm, pgd, address);
2037         if (!pud)
2038                 goto oom;
2039
2040         pmd = pmd_alloc(mm, pud, address);
2041         if (!pmd)
2042                 goto oom;
2043
2044         pte = pte_alloc_map(mm, pmd, address);
2045         if (!pte)
2046                 goto oom;
2047         
2048         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
2049
2050  oom:
2051         spin_unlock(&mm->page_table_lock);
2052         return VM_FAULT_OOM;
2053 }
2054
2055 #ifndef __PAGETABLE_PUD_FOLDED
2056 /*
2057  * Allocate page upper directory.
2058  *
2059  * We've already handled the fast-path in-line, and we own the
2060  * page table lock.
2061  */
2062 pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2063 {
2064         pud_t *new;
2065
2066         spin_unlock(&mm->page_table_lock);
2067         new = pud_alloc_one(mm, address);
2068         spin_lock(&mm->page_table_lock);
2069         if (!new)
2070                 return NULL;
2071
2072         /*
2073          * Because we dropped the lock, we should re-check the
2074          * entry, as somebody else could have populated it..
2075          */
2076         if (pgd_present(*pgd)) {
2077                 pud_free(new);
2078                 goto out;
2079         }
2080         pgd_populate(mm, pgd, new);
2081  out:
2082         return pud_offset(pgd, address);
2083 }
2084 #endif /* __PAGETABLE_PUD_FOLDED */
2085
2086 #ifndef __PAGETABLE_PMD_FOLDED
2087 /*
2088  * Allocate page middle directory.
2089  *
2090  * We've already handled the fast-path in-line, and we own the
2091  * page table lock.
2092  */
2093 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2094 {
2095         pmd_t *new;
2096
2097         spin_unlock(&mm->page_table_lock);
2098         new = pmd_alloc_one(mm, address);
2099         spin_lock(&mm->page_table_lock);
2100         if (!new)
2101                 return NULL;
2102
2103         /*
2104          * Because we dropped the lock, we should re-check the
2105          * entry, as somebody else could have populated it..
2106          */
2107 #ifndef __ARCH_HAS_4LEVEL_HACK
2108         if (pud_present(*pud)) {
2109                 pmd_free(new);
2110                 goto out;
2111         }
2112         pud_populate(mm, pud, new);
2113 #else
2114         if (pgd_present(*pud)) {
2115                 pmd_free(new);
2116                 goto out;
2117         }
2118         pgd_populate(mm, pud, new);
2119 #endif /* __ARCH_HAS_4LEVEL_HACK */
2120
2121  out:
2122         return pmd_offset(pud, address);
2123 }
2124 #endif /* __PAGETABLE_PMD_FOLDED */
2125
2126 int make_pages_present(unsigned long addr, unsigned long end)
2127 {
2128         int ret, len, write;
2129         struct vm_area_struct * vma;
2130
2131         vma = find_vma(current->mm, addr);
2132         if (!vma)
2133                 return -1;
2134         write = (vma->vm_flags & VM_WRITE) != 0;
2135         if (addr >= end)
2136                 BUG();
2137         if (end > vma->vm_end)
2138                 BUG();
2139         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2140         ret = get_user_pages(current, current->mm, addr,
2141                         len, write, 0, NULL, NULL);
2142         if (ret < 0)
2143                 return ret;
2144         return ret == len ? 0 : -1;
2145 }
2146
2147 /* 
2148  * Map a vmalloc()-space virtual address to the physical page.
2149  */
2150 struct page * vmalloc_to_page(void * vmalloc_addr)
2151 {
2152         unsigned long addr = (unsigned long) vmalloc_addr;
2153         struct page *page = NULL;
2154         pgd_t *pgd = pgd_offset_k(addr);
2155         pud_t *pud;
2156         pmd_t *pmd;
2157         pte_t *ptep, pte;
2158   
2159         if (!pgd_none(*pgd)) {
2160                 pud = pud_offset(pgd, addr);
2161                 if (!pud_none(*pud)) {
2162                         pmd = pmd_offset(pud, addr);
2163                         if (!pmd_none(*pmd)) {
2164                                 ptep = pte_offset_map(pmd, addr);
2165                                 pte = *ptep;
2166                                 if (pte_present(pte))
2167                                         page = pte_page(pte);
2168                                 pte_unmap(ptep);
2169                         }
2170                 }
2171         }
2172         return page;
2173 }
2174
2175 EXPORT_SYMBOL(vmalloc_to_page);
2176
2177 /*
2178  * Map a vmalloc()-space virtual address to the physical page frame number.
2179  */
2180 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2181 {
2182         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2183 }
2184
2185 EXPORT_SYMBOL(vmalloc_to_pfn);
2186
2187 /*
2188  * update_mem_hiwater
2189  *      - update per process rss and vm high water data
2190  */
2191 void update_mem_hiwater(struct task_struct *tsk)
2192 {
2193         if (tsk->mm) {
2194                 unsigned long rss = get_mm_counter(tsk->mm, rss);
2195
2196                 if (tsk->mm->hiwater_rss < rss)
2197                         tsk->mm->hiwater_rss = rss;
2198                 if (tsk->mm->hiwater_vm < tsk->mm->total_vm)
2199                         tsk->mm->hiwater_vm = tsk->mm->total_vm;
2200         }
2201 }
2202
2203 #if !defined(__HAVE_ARCH_GATE_AREA)
2204
2205 #if defined(AT_SYSINFO_EHDR)
2206 struct vm_area_struct gate_vma;
2207
2208 static int __init gate_vma_init(void)
2209 {
2210         gate_vma.vm_mm = NULL;
2211         gate_vma.vm_start = FIXADDR_USER_START;
2212         gate_vma.vm_end = FIXADDR_USER_END;
2213         gate_vma.vm_page_prot = PAGE_READONLY;
2214         gate_vma.vm_flags = 0;
2215         return 0;
2216 }
2217 __initcall(gate_vma_init);
2218 #endif
2219
2220 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2221 {
2222 #ifdef AT_SYSINFO_EHDR
2223         return &gate_vma;
2224 #else
2225         return NULL;
2226 #endif
2227 }
2228
2229 int in_gate_area_no_task(unsigned long addr)
2230 {
2231 #ifdef AT_SYSINFO_EHDR
2232         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2233                 return 1;
2234 #endif
2235         return 0;
2236 }
2237
2238 #endif  /* __HAVE_ARCH_GATE_AREA */