/spare/repo/netdev-2.6 branch 'ieee80211'
[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         /*
502          * Don't copy ptes where a page fault will fill them correctly.
503          * Fork becomes much lighter when there are big shared or private
504          * readonly mappings. The tradeoff is that copy_page_range is more
505          * efficient than faulting.
506          */
507         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_RESERVED))) {
508                 if (!vma->anon_vma)
509                         return 0;
510         }
511
512         if (is_vm_hugetlb_page(vma))
513                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
514
515         dst_pgd = pgd_offset(dst_mm, addr);
516         src_pgd = pgd_offset(src_mm, addr);
517         do {
518                 next = pgd_addr_end(addr, end);
519                 if (pgd_none_or_clear_bad(src_pgd))
520                         continue;
521                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
522                                                 vma, addr, next))
523                         return -ENOMEM;
524         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
525         return 0;
526 }
527
528 static void zap_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
529                                 unsigned long addr, unsigned long end,
530                                 struct zap_details *details)
531 {
532         pte_t *pte;
533
534         pte = pte_offset_map(pmd, addr);
535         do {
536                 pte_t ptent = *pte;
537                 if (pte_none(ptent))
538                         continue;
539                 if (pte_present(ptent)) {
540                         struct page *page = NULL;
541                         unsigned long pfn = pte_pfn(ptent);
542                         if (pfn_valid(pfn)) {
543                                 page = pfn_to_page(pfn);
544                                 if (PageReserved(page))
545                                         page = NULL;
546                         }
547                         if (unlikely(details) && page) {
548                                 /*
549                                  * unmap_shared_mapping_pages() wants to
550                                  * invalidate cache without truncating:
551                                  * unmap shared but keep private pages.
552                                  */
553                                 if (details->check_mapping &&
554                                     details->check_mapping != page->mapping)
555                                         continue;
556                                 /*
557                                  * Each page->index must be checked when
558                                  * invalidating or truncating nonlinear.
559                                  */
560                                 if (details->nonlinear_vma &&
561                                     (page->index < details->first_index ||
562                                      page->index > details->last_index))
563                                         continue;
564                         }
565                         ptent = ptep_get_and_clear(tlb->mm, addr, pte);
566                         tlb_remove_tlb_entry(tlb, pte, addr);
567                         if (unlikely(!page))
568                                 continue;
569                         if (unlikely(details) && details->nonlinear_vma
570                             && linear_page_index(details->nonlinear_vma,
571                                                 addr) != page->index)
572                                 set_pte_at(tlb->mm, addr, pte,
573                                            pgoff_to_pte(page->index));
574                         if (pte_dirty(ptent))
575                                 set_page_dirty(page);
576                         if (PageAnon(page))
577                                 dec_mm_counter(tlb->mm, anon_rss);
578                         else if (pte_young(ptent))
579                                 mark_page_accessed(page);
580                         tlb->freed++;
581                         page_remove_rmap(page);
582                         tlb_remove_page(tlb, page);
583                         continue;
584                 }
585                 /*
586                  * If details->check_mapping, we leave swap entries;
587                  * if details->nonlinear_vma, we leave file entries.
588                  */
589                 if (unlikely(details))
590                         continue;
591                 if (!pte_file(ptent))
592                         free_swap_and_cache(pte_to_swp_entry(ptent));
593                 pte_clear(tlb->mm, addr, pte);
594         } while (pte++, addr += PAGE_SIZE, addr != end);
595         pte_unmap(pte - 1);
596 }
597
598 static inline void zap_pmd_range(struct mmu_gather *tlb, pud_t *pud,
599                                 unsigned long addr, unsigned long end,
600                                 struct zap_details *details)
601 {
602         pmd_t *pmd;
603         unsigned long next;
604
605         pmd = pmd_offset(pud, addr);
606         do {
607                 next = pmd_addr_end(addr, end);
608                 if (pmd_none_or_clear_bad(pmd))
609                         continue;
610                 zap_pte_range(tlb, pmd, addr, next, details);
611         } while (pmd++, addr = next, addr != end);
612 }
613
614 static inline void zap_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
615                                 unsigned long addr, unsigned long end,
616                                 struct zap_details *details)
617 {
618         pud_t *pud;
619         unsigned long next;
620
621         pud = pud_offset(pgd, addr);
622         do {
623                 next = pud_addr_end(addr, end);
624                 if (pud_none_or_clear_bad(pud))
625                         continue;
626                 zap_pmd_range(tlb, pud, addr, next, details);
627         } while (pud++, addr = next, addr != end);
628 }
629
630 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
631                                 unsigned long addr, unsigned long end,
632                                 struct zap_details *details)
633 {
634         pgd_t *pgd;
635         unsigned long next;
636
637         if (details && !details->check_mapping && !details->nonlinear_vma)
638                 details = NULL;
639
640         BUG_ON(addr >= end);
641         tlb_start_vma(tlb, vma);
642         pgd = pgd_offset(vma->vm_mm, addr);
643         do {
644                 next = pgd_addr_end(addr, end);
645                 if (pgd_none_or_clear_bad(pgd))
646                         continue;
647                 zap_pud_range(tlb, pgd, addr, next, details);
648         } while (pgd++, addr = next, addr != end);
649         tlb_end_vma(tlb, vma);
650 }
651
652 #ifdef CONFIG_PREEMPT
653 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
654 #else
655 /* No preempt: go for improved straight-line efficiency */
656 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
657 #endif
658
659 /**
660  * unmap_vmas - unmap a range of memory covered by a list of vma's
661  * @tlbp: address of the caller's struct mmu_gather
662  * @mm: the controlling mm_struct
663  * @vma: the starting vma
664  * @start_addr: virtual address at which to start unmapping
665  * @end_addr: virtual address at which to end unmapping
666  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
667  * @details: details of nonlinear truncation or shared cache invalidation
668  *
669  * Returns the end address of the unmapping (restart addr if interrupted).
670  *
671  * Unmap all pages in the vma list.  Called under page_table_lock.
672  *
673  * We aim to not hold page_table_lock for too long (for scheduling latency
674  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
675  * return the ending mmu_gather to the caller.
676  *
677  * Only addresses between `start' and `end' will be unmapped.
678  *
679  * The VMA list must be sorted in ascending virtual address order.
680  *
681  * unmap_vmas() assumes that the caller will flush the whole unmapped address
682  * range after unmap_vmas() returns.  So the only responsibility here is to
683  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
684  * drops the lock and schedules.
685  */
686 unsigned long unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
687                 struct vm_area_struct *vma, unsigned long start_addr,
688                 unsigned long end_addr, unsigned long *nr_accounted,
689                 struct zap_details *details)
690 {
691         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
692         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
693         int tlb_start_valid = 0;
694         unsigned long start = start_addr;
695         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
696         int fullmm = tlb_is_full_mm(*tlbp);
697
698         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
699                 unsigned long end;
700
701                 start = max(vma->vm_start, start_addr);
702                 if (start >= vma->vm_end)
703                         continue;
704                 end = min(vma->vm_end, end_addr);
705                 if (end <= vma->vm_start)
706                         continue;
707
708                 if (vma->vm_flags & VM_ACCOUNT)
709                         *nr_accounted += (end - start) >> PAGE_SHIFT;
710
711                 while (start != end) {
712                         unsigned long block;
713
714                         if (!tlb_start_valid) {
715                                 tlb_start = start;
716                                 tlb_start_valid = 1;
717                         }
718
719                         if (is_vm_hugetlb_page(vma)) {
720                                 block = end - start;
721                                 unmap_hugepage_range(vma, start, end);
722                         } else {
723                                 block = min(zap_bytes, end - start);
724                                 unmap_page_range(*tlbp, vma, start,
725                                                 start + block, details);
726                         }
727
728                         start += block;
729                         zap_bytes -= block;
730                         if ((long)zap_bytes > 0)
731                                 continue;
732
733                         tlb_finish_mmu(*tlbp, tlb_start, start);
734
735                         if (need_resched() ||
736                                 need_lockbreak(&mm->page_table_lock) ||
737                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
738                                 if (i_mmap_lock) {
739                                         /* must reset count of rss freed */
740                                         *tlbp = tlb_gather_mmu(mm, fullmm);
741                                         goto out;
742                                 }
743                                 spin_unlock(&mm->page_table_lock);
744                                 cond_resched();
745                                 spin_lock(&mm->page_table_lock);
746                         }
747
748                         *tlbp = tlb_gather_mmu(mm, fullmm);
749                         tlb_start_valid = 0;
750                         zap_bytes = ZAP_BLOCK_SIZE;
751                 }
752         }
753 out:
754         return start;   /* which is now the end (or restart) address */
755 }
756
757 /**
758  * zap_page_range - remove user pages in a given range
759  * @vma: vm_area_struct holding the applicable pages
760  * @address: starting address of pages to zap
761  * @size: number of bytes to zap
762  * @details: details of nonlinear truncation or shared cache invalidation
763  */
764 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
765                 unsigned long size, struct zap_details *details)
766 {
767         struct mm_struct *mm = vma->vm_mm;
768         struct mmu_gather *tlb;
769         unsigned long end = address + size;
770         unsigned long nr_accounted = 0;
771
772         if (is_vm_hugetlb_page(vma)) {
773                 zap_hugepage_range(vma, address, size);
774                 return end;
775         }
776
777         lru_add_drain();
778         spin_lock(&mm->page_table_lock);
779         tlb = tlb_gather_mmu(mm, 0);
780         end = unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
781         tlb_finish_mmu(tlb, address, end);
782         spin_unlock(&mm->page_table_lock);
783         return end;
784 }
785
786 /*
787  * Do a quick page-table lookup for a single page.
788  * mm->page_table_lock must be held.
789  */
790 static struct page *__follow_page(struct mm_struct *mm, unsigned long address,
791                         int read, int write, int accessed)
792 {
793         pgd_t *pgd;
794         pud_t *pud;
795         pmd_t *pmd;
796         pte_t *ptep, pte;
797         unsigned long pfn;
798         struct page *page;
799
800         page = follow_huge_addr(mm, address, write);
801         if (! IS_ERR(page))
802                 return page;
803
804         pgd = pgd_offset(mm, address);
805         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
806                 goto out;
807
808         pud = pud_offset(pgd, address);
809         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
810                 goto out;
811         
812         pmd = pmd_offset(pud, address);
813         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
814                 goto out;
815         if (pmd_huge(*pmd))
816                 return follow_huge_pmd(mm, address, pmd, write);
817
818         ptep = pte_offset_map(pmd, address);
819         if (!ptep)
820                 goto out;
821
822         pte = *ptep;
823         pte_unmap(ptep);
824         if (pte_present(pte)) {
825                 if (write && !pte_write(pte))
826                         goto out;
827                 if (read && !pte_read(pte))
828                         goto out;
829                 pfn = pte_pfn(pte);
830                 if (pfn_valid(pfn)) {
831                         page = pfn_to_page(pfn);
832                         if (accessed) {
833                                 if (write && !pte_dirty(pte) &&!PageDirty(page))
834                                         set_page_dirty(page);
835                                 mark_page_accessed(page);
836                         }
837                         return page;
838                 }
839         }
840
841 out:
842         return NULL;
843 }
844
845 inline struct page *
846 follow_page(struct mm_struct *mm, unsigned long address, int write)
847 {
848         return __follow_page(mm, address, 0, write, 1);
849 }
850
851 /*
852  * check_user_page_readable() can be called frm niterrupt context by oprofile,
853  * so we need to avoid taking any non-irq-safe locks
854  */
855 int check_user_page_readable(struct mm_struct *mm, unsigned long address)
856 {
857         return __follow_page(mm, address, 1, 0, 0) != NULL;
858 }
859 EXPORT_SYMBOL(check_user_page_readable);
860
861 static inline int
862 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
863                          unsigned long address)
864 {
865         pgd_t *pgd;
866         pud_t *pud;
867         pmd_t *pmd;
868
869         /* Check if the vma is for an anonymous mapping. */
870         if (vma->vm_ops && vma->vm_ops->nopage)
871                 return 0;
872
873         /* Check if page directory entry exists. */
874         pgd = pgd_offset(mm, address);
875         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
876                 return 1;
877
878         pud = pud_offset(pgd, address);
879         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
880                 return 1;
881
882         /* Check if page middle directory entry exists. */
883         pmd = pmd_offset(pud, address);
884         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
885                 return 1;
886
887         /* There is a pte slot for 'address' in 'mm'. */
888         return 0;
889 }
890
891 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
892                 unsigned long start, int len, int write, int force,
893                 struct page **pages, struct vm_area_struct **vmas)
894 {
895         int i;
896         unsigned int flags;
897
898         /* 
899          * Require read or write permissions.
900          * If 'force' is set, we only require the "MAY" flags.
901          */
902         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
903         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
904         i = 0;
905
906         do {
907                 struct vm_area_struct * vma;
908
909                 vma = find_extend_vma(mm, start);
910                 if (!vma && in_gate_area(tsk, start)) {
911                         unsigned long pg = start & PAGE_MASK;
912                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
913                         pgd_t *pgd;
914                         pud_t *pud;
915                         pmd_t *pmd;
916                         pte_t *pte;
917                         if (write) /* user gate pages are read-only */
918                                 return i ? : -EFAULT;
919                         if (pg > TASK_SIZE)
920                                 pgd = pgd_offset_k(pg);
921                         else
922                                 pgd = pgd_offset_gate(mm, pg);
923                         BUG_ON(pgd_none(*pgd));
924                         pud = pud_offset(pgd, pg);
925                         BUG_ON(pud_none(*pud));
926                         pmd = pmd_offset(pud, pg);
927                         if (pmd_none(*pmd))
928                                 return i ? : -EFAULT;
929                         pte = pte_offset_map(pmd, pg);
930                         if (pte_none(*pte)) {
931                                 pte_unmap(pte);
932                                 return i ? : -EFAULT;
933                         }
934                         if (pages) {
935                                 pages[i] = pte_page(*pte);
936                                 get_page(pages[i]);
937                         }
938                         pte_unmap(pte);
939                         if (vmas)
940                                 vmas[i] = gate_vma;
941                         i++;
942                         start += PAGE_SIZE;
943                         len--;
944                         continue;
945                 }
946
947                 if (!vma || (vma->vm_flags & VM_IO)
948                                 || !(flags & vma->vm_flags))
949                         return i ? : -EFAULT;
950
951                 if (is_vm_hugetlb_page(vma)) {
952                         i = follow_hugetlb_page(mm, vma, pages, vmas,
953                                                 &start, &len, i);
954                         continue;
955                 }
956                 spin_lock(&mm->page_table_lock);
957                 do {
958                         int write_access = write;
959                         struct page *page;
960
961                         cond_resched_lock(&mm->page_table_lock);
962                         while (!(page = follow_page(mm, start, write_access))) {
963                                 int ret;
964
965                                 /*
966                                  * Shortcut for anonymous pages. We don't want
967                                  * to force the creation of pages tables for
968                                  * insanely big anonymously mapped areas that
969                                  * nobody touched so far. This is important
970                                  * for doing a core dump for these mappings.
971                                  */
972                                 if (!write && untouched_anonymous_page(mm,vma,start)) {
973                                         page = ZERO_PAGE(start);
974                                         break;
975                                 }
976                                 spin_unlock(&mm->page_table_lock);
977                                 ret = __handle_mm_fault(mm, vma, start, write_access);
978
979                                 /*
980                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
981                                  * broken COW when necessary, even if maybe_mkwrite
982                                  * decided not to set pte_write. We can thus safely do
983                                  * subsequent page lookups as if they were reads.
984                                  */
985                                 if (ret & VM_FAULT_WRITE)
986                                         write_access = 0;
987                                 
988                                 switch (ret & ~VM_FAULT_WRITE) {
989                                 case VM_FAULT_MINOR:
990                                         tsk->min_flt++;
991                                         break;
992                                 case VM_FAULT_MAJOR:
993                                         tsk->maj_flt++;
994                                         break;
995                                 case VM_FAULT_SIGBUS:
996                                         return i ? i : -EFAULT;
997                                 case VM_FAULT_OOM:
998                                         return i ? i : -ENOMEM;
999                                 default:
1000                                         BUG();
1001                                 }
1002                                 spin_lock(&mm->page_table_lock);
1003                         }
1004                         if (pages) {
1005                                 pages[i] = page;
1006                                 flush_dcache_page(page);
1007                                 if (!PageReserved(page))
1008                                         page_cache_get(page);
1009                         }
1010                         if (vmas)
1011                                 vmas[i] = vma;
1012                         i++;
1013                         start += PAGE_SIZE;
1014                         len--;
1015                 } while (len && start < vma->vm_end);
1016                 spin_unlock(&mm->page_table_lock);
1017         } while (len);
1018         return i;
1019 }
1020 EXPORT_SYMBOL(get_user_pages);
1021
1022 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1023                         unsigned long addr, unsigned long end, pgprot_t prot)
1024 {
1025         pte_t *pte;
1026
1027         pte = pte_alloc_map(mm, pmd, addr);
1028         if (!pte)
1029                 return -ENOMEM;
1030         do {
1031                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(addr), prot));
1032                 BUG_ON(!pte_none(*pte));
1033                 set_pte_at(mm, addr, pte, zero_pte);
1034         } while (pte++, addr += PAGE_SIZE, addr != end);
1035         pte_unmap(pte - 1);
1036         return 0;
1037 }
1038
1039 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1040                         unsigned long addr, unsigned long end, pgprot_t prot)
1041 {
1042         pmd_t *pmd;
1043         unsigned long next;
1044
1045         pmd = pmd_alloc(mm, pud, addr);
1046         if (!pmd)
1047                 return -ENOMEM;
1048         do {
1049                 next = pmd_addr_end(addr, end);
1050                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1051                         return -ENOMEM;
1052         } while (pmd++, addr = next, addr != end);
1053         return 0;
1054 }
1055
1056 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1057                         unsigned long addr, unsigned long end, pgprot_t prot)
1058 {
1059         pud_t *pud;
1060         unsigned long next;
1061
1062         pud = pud_alloc(mm, pgd, addr);
1063         if (!pud)
1064                 return -ENOMEM;
1065         do {
1066                 next = pud_addr_end(addr, end);
1067                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1068                         return -ENOMEM;
1069         } while (pud++, addr = next, addr != end);
1070         return 0;
1071 }
1072
1073 int zeromap_page_range(struct vm_area_struct *vma,
1074                         unsigned long addr, unsigned long size, pgprot_t prot)
1075 {
1076         pgd_t *pgd;
1077         unsigned long next;
1078         unsigned long end = addr + size;
1079         struct mm_struct *mm = vma->vm_mm;
1080         int err;
1081
1082         BUG_ON(addr >= end);
1083         pgd = pgd_offset(mm, addr);
1084         flush_cache_range(vma, addr, end);
1085         spin_lock(&mm->page_table_lock);
1086         do {
1087                 next = pgd_addr_end(addr, end);
1088                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1089                 if (err)
1090                         break;
1091         } while (pgd++, addr = next, addr != end);
1092         spin_unlock(&mm->page_table_lock);
1093         return err;
1094 }
1095
1096 /*
1097  * maps a range of physical memory into the requested pages. the old
1098  * mappings are removed. any references to nonexistent pages results
1099  * in null mappings (currently treated as "copy-on-access")
1100  */
1101 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1102                         unsigned long addr, unsigned long end,
1103                         unsigned long pfn, pgprot_t prot)
1104 {
1105         pte_t *pte;
1106
1107         pte = pte_alloc_map(mm, pmd, addr);
1108         if (!pte)
1109                 return -ENOMEM;
1110         do {
1111                 BUG_ON(!pte_none(*pte));
1112                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
1113                         set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1114                 pfn++;
1115         } while (pte++, addr += PAGE_SIZE, addr != end);
1116         pte_unmap(pte - 1);
1117         return 0;
1118 }
1119
1120 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1121                         unsigned long addr, unsigned long end,
1122                         unsigned long pfn, pgprot_t prot)
1123 {
1124         pmd_t *pmd;
1125         unsigned long next;
1126
1127         pfn -= addr >> PAGE_SHIFT;
1128         pmd = pmd_alloc(mm, pud, addr);
1129         if (!pmd)
1130                 return -ENOMEM;
1131         do {
1132                 next = pmd_addr_end(addr, end);
1133                 if (remap_pte_range(mm, pmd, addr, next,
1134                                 pfn + (addr >> PAGE_SHIFT), prot))
1135                         return -ENOMEM;
1136         } while (pmd++, addr = next, addr != end);
1137         return 0;
1138 }
1139
1140 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1141                         unsigned long addr, unsigned long end,
1142                         unsigned long pfn, pgprot_t prot)
1143 {
1144         pud_t *pud;
1145         unsigned long next;
1146
1147         pfn -= addr >> PAGE_SHIFT;
1148         pud = pud_alloc(mm, pgd, addr);
1149         if (!pud)
1150                 return -ENOMEM;
1151         do {
1152                 next = pud_addr_end(addr, end);
1153                 if (remap_pmd_range(mm, pud, addr, next,
1154                                 pfn + (addr >> PAGE_SHIFT), prot))
1155                         return -ENOMEM;
1156         } while (pud++, addr = next, addr != end);
1157         return 0;
1158 }
1159
1160 /*  Note: this is only safe if the mm semaphore is held when called. */
1161 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1162                     unsigned long pfn, unsigned long size, pgprot_t prot)
1163 {
1164         pgd_t *pgd;
1165         unsigned long next;
1166         unsigned long end = addr + PAGE_ALIGN(size);
1167         struct mm_struct *mm = vma->vm_mm;
1168         int err;
1169
1170         /*
1171          * Physically remapped pages are special. Tell the
1172          * rest of the world about it:
1173          *   VM_IO tells people not to look at these pages
1174          *      (accesses can have side effects).
1175          *   VM_RESERVED tells swapout not to try to touch
1176          *      this region.
1177          */
1178         vma->vm_flags |= VM_IO | VM_RESERVED;
1179
1180         BUG_ON(addr >= end);
1181         pfn -= addr >> PAGE_SHIFT;
1182         pgd = pgd_offset(mm, addr);
1183         flush_cache_range(vma, addr, end);
1184         spin_lock(&mm->page_table_lock);
1185         do {
1186                 next = pgd_addr_end(addr, end);
1187                 err = remap_pud_range(mm, pgd, addr, next,
1188                                 pfn + (addr >> PAGE_SHIFT), prot);
1189                 if (err)
1190                         break;
1191         } while (pgd++, addr = next, addr != end);
1192         spin_unlock(&mm->page_table_lock);
1193         return err;
1194 }
1195 EXPORT_SYMBOL(remap_pfn_range);
1196
1197 /*
1198  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1199  * servicing faults for write access.  In the normal case, do always want
1200  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1201  * that do not have writing enabled, when used by access_process_vm.
1202  */
1203 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1204 {
1205         if (likely(vma->vm_flags & VM_WRITE))
1206                 pte = pte_mkwrite(pte);
1207         return pte;
1208 }
1209
1210 /*
1211  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1212  */
1213 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address, 
1214                 pte_t *page_table)
1215 {
1216         pte_t entry;
1217
1218         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1219                               vma);
1220         ptep_establish(vma, address, page_table, entry);
1221         update_mmu_cache(vma, address, entry);
1222         lazy_mmu_prot_update(entry);
1223 }
1224
1225 /*
1226  * This routine handles present pages, when users try to write
1227  * to a shared page. It is done by copying the page to a new address
1228  * and decrementing the shared-page counter for the old page.
1229  *
1230  * Goto-purists beware: the only reason for goto's here is that it results
1231  * in better assembly code.. The "default" path will see no jumps at all.
1232  *
1233  * Note that this routine assumes that the protection checks have been
1234  * done by the caller (the low-level page fault routine in most cases).
1235  * Thus we can safely just mark it writable once we've done any necessary
1236  * COW.
1237  *
1238  * We also mark the page dirty at this point even though the page will
1239  * change only once the write actually happens. This avoids a few races,
1240  * and potentially makes it more efficient.
1241  *
1242  * We hold the mm semaphore and the page_table_lock on entry and exit
1243  * with the page_table_lock released.
1244  */
1245 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1246         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1247 {
1248         struct page *old_page, *new_page;
1249         unsigned long pfn = pte_pfn(pte);
1250         pte_t entry;
1251         int ret;
1252
1253         if (unlikely(!pfn_valid(pfn))) {
1254                 /*
1255                  * This should really halt the system so it can be debugged or
1256                  * at least the kernel stops what it's doing before it corrupts
1257                  * data, but for the moment just pretend this is OOM.
1258                  */
1259                 pte_unmap(page_table);
1260                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1261                                 address);
1262                 spin_unlock(&mm->page_table_lock);
1263                 return VM_FAULT_OOM;
1264         }
1265         old_page = pfn_to_page(pfn);
1266
1267         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1268                 int reuse = can_share_swap_page(old_page);
1269                 unlock_page(old_page);
1270                 if (reuse) {
1271                         flush_cache_page(vma, address, pfn);
1272                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1273                                               vma);
1274                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1275                         update_mmu_cache(vma, address, entry);
1276                         lazy_mmu_prot_update(entry);
1277                         pte_unmap(page_table);
1278                         spin_unlock(&mm->page_table_lock);
1279                         return VM_FAULT_MINOR|VM_FAULT_WRITE;
1280                 }
1281         }
1282         pte_unmap(page_table);
1283
1284         /*
1285          * Ok, we need to copy. Oh, well..
1286          */
1287         if (!PageReserved(old_page))
1288                 page_cache_get(old_page);
1289         spin_unlock(&mm->page_table_lock);
1290
1291         if (unlikely(anon_vma_prepare(vma)))
1292                 goto no_new_page;
1293         if (old_page == ZERO_PAGE(address)) {
1294                 new_page = alloc_zeroed_user_highpage(vma, address);
1295                 if (!new_page)
1296                         goto no_new_page;
1297         } else {
1298                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1299                 if (!new_page)
1300                         goto no_new_page;
1301                 copy_user_highpage(new_page, old_page, address);
1302         }
1303         /*
1304          * Re-check the pte - we dropped the lock
1305          */
1306         ret = VM_FAULT_MINOR;
1307         spin_lock(&mm->page_table_lock);
1308         page_table = pte_offset_map(pmd, address);
1309         if (likely(pte_same(*page_table, pte))) {
1310                 if (PageAnon(old_page))
1311                         dec_mm_counter(mm, anon_rss);
1312                 if (PageReserved(old_page))
1313                         inc_mm_counter(mm, rss);
1314                 else
1315                         page_remove_rmap(old_page);
1316                 flush_cache_page(vma, address, pfn);
1317                 break_cow(vma, new_page, address, page_table);
1318                 lru_cache_add_active(new_page);
1319                 page_add_anon_rmap(new_page, vma, address);
1320
1321                 /* Free the old page.. */
1322                 new_page = old_page;
1323                 ret |= VM_FAULT_WRITE;
1324         }
1325         pte_unmap(page_table);
1326         page_cache_release(new_page);
1327         page_cache_release(old_page);
1328         spin_unlock(&mm->page_table_lock);
1329         return ret;
1330
1331 no_new_page:
1332         page_cache_release(old_page);
1333         return VM_FAULT_OOM;
1334 }
1335
1336 /*
1337  * Helper functions for unmap_mapping_range().
1338  *
1339  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1340  *
1341  * We have to restart searching the prio_tree whenever we drop the lock,
1342  * since the iterator is only valid while the lock is held, and anyway
1343  * a later vma might be split and reinserted earlier while lock dropped.
1344  *
1345  * The list of nonlinear vmas could be handled more efficiently, using
1346  * a placeholder, but handle it in the same way until a need is shown.
1347  * It is important to search the prio_tree before nonlinear list: a vma
1348  * may become nonlinear and be shifted from prio_tree to nonlinear list
1349  * while the lock is dropped; but never shifted from list to prio_tree.
1350  *
1351  * In order to make forward progress despite restarting the search,
1352  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1353  * quickly skip it next time around.  Since the prio_tree search only
1354  * shows us those vmas affected by unmapping the range in question, we
1355  * can't efficiently keep all vmas in step with mapping->truncate_count:
1356  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1357  * mapping->truncate_count and vma->vm_truncate_count are protected by
1358  * i_mmap_lock.
1359  *
1360  * In order to make forward progress despite repeatedly restarting some
1361  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1362  * and restart from that address when we reach that vma again.  It might
1363  * have been split or merged, shrunk or extended, but never shifted: so
1364  * restart_addr remains valid so long as it remains in the vma's range.
1365  * unmap_mapping_range forces truncate_count to leap over page-aligned
1366  * values so we can save vma's restart_addr in its truncate_count field.
1367  */
1368 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1369
1370 static void reset_vma_truncate_counts(struct address_space *mapping)
1371 {
1372         struct vm_area_struct *vma;
1373         struct prio_tree_iter iter;
1374
1375         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1376                 vma->vm_truncate_count = 0;
1377         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1378                 vma->vm_truncate_count = 0;
1379 }
1380
1381 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1382                 unsigned long start_addr, unsigned long end_addr,
1383                 struct zap_details *details)
1384 {
1385         unsigned long restart_addr;
1386         int need_break;
1387
1388 again:
1389         restart_addr = vma->vm_truncate_count;
1390         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1391                 start_addr = restart_addr;
1392                 if (start_addr >= end_addr) {
1393                         /* Top of vma has been split off since last time */
1394                         vma->vm_truncate_count = details->truncate_count;
1395                         return 0;
1396                 }
1397         }
1398
1399         restart_addr = zap_page_range(vma, start_addr,
1400                                         end_addr - start_addr, details);
1401
1402         /*
1403          * We cannot rely on the break test in unmap_vmas:
1404          * on the one hand, we don't want to restart our loop
1405          * just because that broke out for the page_table_lock;
1406          * on the other hand, it does no test when vma is small.
1407          */
1408         need_break = need_resched() ||
1409                         need_lockbreak(details->i_mmap_lock);
1410
1411         if (restart_addr >= end_addr) {
1412                 /* We have now completed this vma: mark it so */
1413                 vma->vm_truncate_count = details->truncate_count;
1414                 if (!need_break)
1415                         return 0;
1416         } else {
1417                 /* Note restart_addr in vma's truncate_count field */
1418                 vma->vm_truncate_count = restart_addr;
1419                 if (!need_break)
1420                         goto again;
1421         }
1422
1423         spin_unlock(details->i_mmap_lock);
1424         cond_resched();
1425         spin_lock(details->i_mmap_lock);
1426         return -EINTR;
1427 }
1428
1429 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1430                                             struct zap_details *details)
1431 {
1432         struct vm_area_struct *vma;
1433         struct prio_tree_iter iter;
1434         pgoff_t vba, vea, zba, zea;
1435
1436 restart:
1437         vma_prio_tree_foreach(vma, &iter, root,
1438                         details->first_index, details->last_index) {
1439                 /* Skip quickly over those we have already dealt with */
1440                 if (vma->vm_truncate_count == details->truncate_count)
1441                         continue;
1442
1443                 vba = vma->vm_pgoff;
1444                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1445                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1446                 zba = details->first_index;
1447                 if (zba < vba)
1448                         zba = vba;
1449                 zea = details->last_index;
1450                 if (zea > vea)
1451                         zea = vea;
1452
1453                 if (unmap_mapping_range_vma(vma,
1454                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1455                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1456                                 details) < 0)
1457                         goto restart;
1458         }
1459 }
1460
1461 static inline void unmap_mapping_range_list(struct list_head *head,
1462                                             struct zap_details *details)
1463 {
1464         struct vm_area_struct *vma;
1465
1466         /*
1467          * In nonlinear VMAs there is no correspondence between virtual address
1468          * offset and file offset.  So we must perform an exhaustive search
1469          * across *all* the pages in each nonlinear VMA, not just the pages
1470          * whose virtual address lies outside the file truncation point.
1471          */
1472 restart:
1473         list_for_each_entry(vma, head, shared.vm_set.list) {
1474                 /* Skip quickly over those we have already dealt with */
1475                 if (vma->vm_truncate_count == details->truncate_count)
1476                         continue;
1477                 details->nonlinear_vma = vma;
1478                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1479                                         vma->vm_end, details) < 0)
1480                         goto restart;
1481         }
1482 }
1483
1484 /**
1485  * unmap_mapping_range - unmap the portion of all mmaps
1486  * in the specified address_space corresponding to the specified
1487  * page range in the underlying file.
1488  * @mapping: the address space containing mmaps to be unmapped.
1489  * @holebegin: byte in first page to unmap, relative to the start of
1490  * the underlying file.  This will be rounded down to a PAGE_SIZE
1491  * boundary.  Note that this is different from vmtruncate(), which
1492  * must keep the partial page.  In contrast, we must get rid of
1493  * partial pages.
1494  * @holelen: size of prospective hole in bytes.  This will be rounded
1495  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1496  * end of the file.
1497  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1498  * but 0 when invalidating pagecache, don't throw away private data.
1499  */
1500 void unmap_mapping_range(struct address_space *mapping,
1501                 loff_t const holebegin, loff_t const holelen, int even_cows)
1502 {
1503         struct zap_details details;
1504         pgoff_t hba = holebegin >> PAGE_SHIFT;
1505         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1506
1507         /* Check for overflow. */
1508         if (sizeof(holelen) > sizeof(hlen)) {
1509                 long long holeend =
1510                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1511                 if (holeend & ~(long long)ULONG_MAX)
1512                         hlen = ULONG_MAX - hba + 1;
1513         }
1514
1515         details.check_mapping = even_cows? NULL: mapping;
1516         details.nonlinear_vma = NULL;
1517         details.first_index = hba;
1518         details.last_index = hba + hlen - 1;
1519         if (details.last_index < details.first_index)
1520                 details.last_index = ULONG_MAX;
1521         details.i_mmap_lock = &mapping->i_mmap_lock;
1522
1523         spin_lock(&mapping->i_mmap_lock);
1524
1525         /* serialize i_size write against truncate_count write */
1526         smp_wmb();
1527         /* Protect against page faults, and endless unmapping loops */
1528         mapping->truncate_count++;
1529         /*
1530          * For archs where spin_lock has inclusive semantics like ia64
1531          * this smp_mb() will prevent to read pagetable contents
1532          * before the truncate_count increment is visible to
1533          * other cpus.
1534          */
1535         smp_mb();
1536         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1537                 if (mapping->truncate_count == 0)
1538                         reset_vma_truncate_counts(mapping);
1539                 mapping->truncate_count++;
1540         }
1541         details.truncate_count = mapping->truncate_count;
1542
1543         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1544                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1545         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1546                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1547         spin_unlock(&mapping->i_mmap_lock);
1548 }
1549 EXPORT_SYMBOL(unmap_mapping_range);
1550
1551 /*
1552  * Handle all mappings that got truncated by a "truncate()"
1553  * system call.
1554  *
1555  * NOTE! We have to be ready to update the memory sharing
1556  * between the file and the memory map for a potential last
1557  * incomplete page.  Ugly, but necessary.
1558  */
1559 int vmtruncate(struct inode * inode, loff_t offset)
1560 {
1561         struct address_space *mapping = inode->i_mapping;
1562         unsigned long limit;
1563
1564         if (inode->i_size < offset)
1565                 goto do_expand;
1566         /*
1567          * truncation of in-use swapfiles is disallowed - it would cause
1568          * subsequent swapout to scribble on the now-freed blocks.
1569          */
1570         if (IS_SWAPFILE(inode))
1571                 goto out_busy;
1572         i_size_write(inode, offset);
1573         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1574         truncate_inode_pages(mapping, offset);
1575         goto out_truncate;
1576
1577 do_expand:
1578         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1579         if (limit != RLIM_INFINITY && offset > limit)
1580                 goto out_sig;
1581         if (offset > inode->i_sb->s_maxbytes)
1582                 goto out_big;
1583         i_size_write(inode, offset);
1584
1585 out_truncate:
1586         if (inode->i_op && inode->i_op->truncate)
1587                 inode->i_op->truncate(inode);
1588         return 0;
1589 out_sig:
1590         send_sig(SIGXFSZ, current, 0);
1591 out_big:
1592         return -EFBIG;
1593 out_busy:
1594         return -ETXTBSY;
1595 }
1596
1597 EXPORT_SYMBOL(vmtruncate);
1598
1599 /* 
1600  * Primitive swap readahead code. We simply read an aligned block of
1601  * (1 << page_cluster) entries in the swap area. This method is chosen
1602  * because it doesn't cost us any seek time.  We also make sure to queue
1603  * the 'original' request together with the readahead ones...  
1604  *
1605  * This has been extended to use the NUMA policies from the mm triggering
1606  * the readahead.
1607  *
1608  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1609  */
1610 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1611 {
1612 #ifdef CONFIG_NUMA
1613         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1614 #endif
1615         int i, num;
1616         struct page *new_page;
1617         unsigned long offset;
1618
1619         /*
1620          * Get the number of handles we should do readahead io to.
1621          */
1622         num = valid_swaphandles(entry, &offset);
1623         for (i = 0; i < num; offset++, i++) {
1624                 /* Ok, do the async read-ahead now */
1625                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1626                                                            offset), vma, addr);
1627                 if (!new_page)
1628                         break;
1629                 page_cache_release(new_page);
1630 #ifdef CONFIG_NUMA
1631                 /*
1632                  * Find the next applicable VMA for the NUMA policy.
1633                  */
1634                 addr += PAGE_SIZE;
1635                 if (addr == 0)
1636                         vma = NULL;
1637                 if (vma) {
1638                         if (addr >= vma->vm_end) {
1639                                 vma = next_vma;
1640                                 next_vma = vma ? vma->vm_next : NULL;
1641                         }
1642                         if (vma && addr < vma->vm_start)
1643                                 vma = NULL;
1644                 } else {
1645                         if (next_vma && addr >= next_vma->vm_start) {
1646                                 vma = next_vma;
1647                                 next_vma = vma->vm_next;
1648                         }
1649                 }
1650 #endif
1651         }
1652         lru_add_drain();        /* Push any new pages onto the LRU now */
1653 }
1654
1655 /*
1656  * We hold the mm semaphore and the page_table_lock on entry and
1657  * should release the pagetable lock on exit..
1658  */
1659 static int do_swap_page(struct mm_struct * mm,
1660         struct vm_area_struct * vma, unsigned long address,
1661         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1662 {
1663         struct page *page;
1664         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1665         pte_t pte;
1666         int ret = VM_FAULT_MINOR;
1667
1668         pte_unmap(page_table);
1669         spin_unlock(&mm->page_table_lock);
1670         page = lookup_swap_cache(entry);
1671         if (!page) {
1672                 swapin_readahead(entry, address, vma);
1673                 page = read_swap_cache_async(entry, vma, address);
1674                 if (!page) {
1675                         /*
1676                          * Back out if somebody else faulted in this pte while
1677                          * we released the page table lock.
1678                          */
1679                         spin_lock(&mm->page_table_lock);
1680                         page_table = pte_offset_map(pmd, address);
1681                         if (likely(pte_same(*page_table, orig_pte)))
1682                                 ret = VM_FAULT_OOM;
1683                         else
1684                                 ret = VM_FAULT_MINOR;
1685                         pte_unmap(page_table);
1686                         spin_unlock(&mm->page_table_lock);
1687                         goto out;
1688                 }
1689
1690                 /* Had to read the page from swap area: Major fault */
1691                 ret = VM_FAULT_MAJOR;
1692                 inc_page_state(pgmajfault);
1693                 grab_swap_token();
1694         }
1695
1696         mark_page_accessed(page);
1697         lock_page(page);
1698
1699         /*
1700          * Back out if somebody else faulted in this pte while we
1701          * released the page table lock.
1702          */
1703         spin_lock(&mm->page_table_lock);
1704         page_table = pte_offset_map(pmd, address);
1705         if (unlikely(!pte_same(*page_table, orig_pte))) {
1706                 ret = VM_FAULT_MINOR;
1707                 goto out_nomap;
1708         }
1709
1710         if (unlikely(!PageUptodate(page))) {
1711                 ret = VM_FAULT_SIGBUS;
1712                 goto out_nomap;
1713         }
1714
1715         /* The page isn't present yet, go ahead with the fault. */
1716
1717         inc_mm_counter(mm, rss);
1718         pte = mk_pte(page, vma->vm_page_prot);
1719         if (write_access && can_share_swap_page(page)) {
1720                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1721                 write_access = 0;
1722         }
1723
1724         flush_icache_page(vma, page);
1725         set_pte_at(mm, address, page_table, pte);
1726         page_add_anon_rmap(page, vma, address);
1727
1728         swap_free(entry);
1729         if (vm_swap_full())
1730                 remove_exclusive_swap_page(page);
1731         unlock_page(page);
1732
1733         if (write_access) {
1734                 if (do_wp_page(mm, vma, address,
1735                                 page_table, pmd, pte) == VM_FAULT_OOM)
1736                         ret = VM_FAULT_OOM;
1737                 goto out;
1738         }
1739
1740         /* No need to invalidate - it was non-present before */
1741         update_mmu_cache(vma, address, pte);
1742         lazy_mmu_prot_update(pte);
1743         pte_unmap(page_table);
1744         spin_unlock(&mm->page_table_lock);
1745 out:
1746         return ret;
1747 out_nomap:
1748         pte_unmap(page_table);
1749         spin_unlock(&mm->page_table_lock);
1750         unlock_page(page);
1751         page_cache_release(page);
1752         goto out;
1753 }
1754
1755 /*
1756  * We are called with the MM semaphore and page_table_lock
1757  * spinlock held to protect against concurrent faults in
1758  * multithreaded programs. 
1759  */
1760 static int
1761 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1762                 pte_t *page_table, pmd_t *pmd, int write_access,
1763                 unsigned long addr)
1764 {
1765         pte_t entry;
1766         struct page * page = ZERO_PAGE(addr);
1767
1768         /* Read-only mapping of ZERO_PAGE. */
1769         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1770
1771         /* ..except if it's a write access */
1772         if (write_access) {
1773                 /* Allocate our own private page. */
1774                 pte_unmap(page_table);
1775                 spin_unlock(&mm->page_table_lock);
1776
1777                 if (unlikely(anon_vma_prepare(vma)))
1778                         goto no_mem;
1779                 page = alloc_zeroed_user_highpage(vma, addr);
1780                 if (!page)
1781                         goto no_mem;
1782
1783                 spin_lock(&mm->page_table_lock);
1784                 page_table = pte_offset_map(pmd, addr);
1785
1786                 if (!pte_none(*page_table)) {
1787                         pte_unmap(page_table);
1788                         page_cache_release(page);
1789                         spin_unlock(&mm->page_table_lock);
1790                         goto out;
1791                 }
1792                 inc_mm_counter(mm, rss);
1793                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1794                                                          vma->vm_page_prot)),
1795                                       vma);
1796                 lru_cache_add_active(page);
1797                 SetPageReferenced(page);
1798                 page_add_anon_rmap(page, vma, addr);
1799         }
1800
1801         set_pte_at(mm, addr, page_table, entry);
1802         pte_unmap(page_table);
1803
1804         /* No need to invalidate - it was non-present before */
1805         update_mmu_cache(vma, addr, entry);
1806         lazy_mmu_prot_update(entry);
1807         spin_unlock(&mm->page_table_lock);
1808 out:
1809         return VM_FAULT_MINOR;
1810 no_mem:
1811         return VM_FAULT_OOM;
1812 }
1813
1814 /*
1815  * do_no_page() tries to create a new page mapping. It aggressively
1816  * tries to share with existing pages, but makes a separate copy if
1817  * the "write_access" parameter is true in order to avoid the next
1818  * page fault.
1819  *
1820  * As this is called only for pages that do not currently exist, we
1821  * do not need to flush old virtual caches or the TLB.
1822  *
1823  * This is called with the MM semaphore held and the page table
1824  * spinlock held. Exit with the spinlock released.
1825  */
1826 static int
1827 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1828         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1829 {
1830         struct page * new_page;
1831         struct address_space *mapping = NULL;
1832         pte_t entry;
1833         unsigned int sequence = 0;
1834         int ret = VM_FAULT_MINOR;
1835         int anon = 0;
1836
1837         if (!vma->vm_ops || !vma->vm_ops->nopage)
1838                 return do_anonymous_page(mm, vma, page_table,
1839                                         pmd, write_access, address);
1840         pte_unmap(page_table);
1841         spin_unlock(&mm->page_table_lock);
1842
1843         if (vma->vm_file) {
1844                 mapping = vma->vm_file->f_mapping;
1845                 sequence = mapping->truncate_count;
1846                 smp_rmb(); /* serializes i_size against truncate_count */
1847         }
1848 retry:
1849         cond_resched();
1850         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1851         /*
1852          * No smp_rmb is needed here as long as there's a full
1853          * spin_lock/unlock sequence inside the ->nopage callback
1854          * (for the pagecache lookup) that acts as an implicit
1855          * smp_mb() and prevents the i_size read to happen
1856          * after the next truncate_count read.
1857          */
1858
1859         /* no page was available -- either SIGBUS or OOM */
1860         if (new_page == NOPAGE_SIGBUS)
1861                 return VM_FAULT_SIGBUS;
1862         if (new_page == NOPAGE_OOM)
1863                 return VM_FAULT_OOM;
1864
1865         /*
1866          * Should we do an early C-O-W break?
1867          */
1868         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1869                 struct page *page;
1870
1871                 if (unlikely(anon_vma_prepare(vma)))
1872                         goto oom;
1873                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1874                 if (!page)
1875                         goto oom;
1876                 copy_user_highpage(page, new_page, address);
1877                 page_cache_release(new_page);
1878                 new_page = page;
1879                 anon = 1;
1880         }
1881
1882         spin_lock(&mm->page_table_lock);
1883         /*
1884          * For a file-backed vma, someone could have truncated or otherwise
1885          * invalidated this page.  If unmap_mapping_range got called,
1886          * retry getting the page.
1887          */
1888         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1889                 sequence = mapping->truncate_count;
1890                 spin_unlock(&mm->page_table_lock);
1891                 page_cache_release(new_page);
1892                 goto retry;
1893         }
1894         page_table = pte_offset_map(pmd, address);
1895
1896         /*
1897          * This silly early PAGE_DIRTY setting removes a race
1898          * due to the bad i386 page protection. But it's valid
1899          * for other architectures too.
1900          *
1901          * Note that if write_access is true, we either now have
1902          * an exclusive copy of the page, or this is a shared mapping,
1903          * so we can make it writable and dirty to avoid having to
1904          * handle that later.
1905          */
1906         /* Only go through if we didn't race with anybody else... */
1907         if (pte_none(*page_table)) {
1908                 if (!PageReserved(new_page))
1909                         inc_mm_counter(mm, rss);
1910
1911                 flush_icache_page(vma, new_page);
1912                 entry = mk_pte(new_page, vma->vm_page_prot);
1913                 if (write_access)
1914                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1915                 set_pte_at(mm, address, page_table, entry);
1916                 if (anon) {
1917                         lru_cache_add_active(new_page);
1918                         page_add_anon_rmap(new_page, vma, address);
1919                 } else
1920                         page_add_file_rmap(new_page);
1921                 pte_unmap(page_table);
1922         } else {
1923                 /* One of our sibling threads was faster, back out. */
1924                 pte_unmap(page_table);
1925                 page_cache_release(new_page);
1926                 spin_unlock(&mm->page_table_lock);
1927                 goto out;
1928         }
1929
1930         /* no need to invalidate: a not-present page shouldn't be cached */
1931         update_mmu_cache(vma, address, entry);
1932         lazy_mmu_prot_update(entry);
1933         spin_unlock(&mm->page_table_lock);
1934 out:
1935         return ret;
1936 oom:
1937         page_cache_release(new_page);
1938         ret = VM_FAULT_OOM;
1939         goto out;
1940 }
1941
1942 /*
1943  * Fault of a previously existing named mapping. Repopulate the pte
1944  * from the encoded file_pte if possible. This enables swappable
1945  * nonlinear vmas.
1946  */
1947 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1948         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1949 {
1950         unsigned long pgoff;
1951         int err;
1952
1953         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1954         /*
1955          * Fall back to the linear mapping if the fs does not support
1956          * ->populate:
1957          */
1958         if (!vma->vm_ops || !vma->vm_ops->populate || 
1959                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1960                 pte_clear(mm, address, pte);
1961                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1962         }
1963
1964         pgoff = pte_to_pgoff(*pte);
1965
1966         pte_unmap(pte);
1967         spin_unlock(&mm->page_table_lock);
1968
1969         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1970         if (err == -ENOMEM)
1971                 return VM_FAULT_OOM;
1972         if (err)
1973                 return VM_FAULT_SIGBUS;
1974         return VM_FAULT_MAJOR;
1975 }
1976
1977 /*
1978  * These routines also need to handle stuff like marking pages dirty
1979  * and/or accessed for architectures that don't do it in hardware (most
1980  * RISC architectures).  The early dirtying is also good on the i386.
1981  *
1982  * There is also a hook called "update_mmu_cache()" that architectures
1983  * with external mmu caches can use to update those (ie the Sparc or
1984  * PowerPC hashed page tables that act as extended TLBs).
1985  *
1986  * Note the "page_table_lock". It is to protect against kswapd removing
1987  * pages from under us. Note that kswapd only ever _removes_ pages, never
1988  * adds them. As such, once we have noticed that the page is not present,
1989  * we can drop the lock early.
1990  *
1991  * The adding of pages is protected by the MM semaphore (which we hold),
1992  * so we don't need to worry about a page being suddenly been added into
1993  * our VM.
1994  *
1995  * We enter with the pagetable spinlock held, we are supposed to
1996  * release it when done.
1997  */
1998 static inline int handle_pte_fault(struct mm_struct *mm,
1999         struct vm_area_struct * vma, unsigned long address,
2000         int write_access, pte_t *pte, pmd_t *pmd)
2001 {
2002         pte_t entry;
2003
2004         entry = *pte;
2005         if (!pte_present(entry)) {
2006                 /*
2007                  * If it truly wasn't present, we know that kswapd
2008                  * and the PTE updates will not touch it later. So
2009                  * drop the lock.
2010                  */
2011                 if (pte_none(entry))
2012                         return do_no_page(mm, vma, address, write_access, pte, pmd);
2013                 if (pte_file(entry))
2014                         return do_file_page(mm, vma, address, write_access, pte, pmd);
2015                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
2016         }
2017
2018         if (write_access) {
2019                 if (!pte_write(entry))
2020                         return do_wp_page(mm, vma, address, pte, pmd, entry);
2021                 entry = pte_mkdirty(entry);
2022         }
2023         entry = pte_mkyoung(entry);
2024         ptep_set_access_flags(vma, address, pte, entry, write_access);
2025         update_mmu_cache(vma, address, entry);
2026         lazy_mmu_prot_update(entry);
2027         pte_unmap(pte);
2028         spin_unlock(&mm->page_table_lock);
2029         return VM_FAULT_MINOR;
2030 }
2031
2032 /*
2033  * By the time we get here, we already hold the mm semaphore
2034  */
2035 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
2036                 unsigned long address, int write_access)
2037 {
2038         pgd_t *pgd;
2039         pud_t *pud;
2040         pmd_t *pmd;
2041         pte_t *pte;
2042
2043         __set_current_state(TASK_RUNNING);
2044
2045         inc_page_state(pgfault);
2046
2047         if (is_vm_hugetlb_page(vma))
2048                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
2049
2050         /*
2051          * We need the page table lock to synchronize with kswapd
2052          * and the SMP-safe atomic PTE updates.
2053          */
2054         pgd = pgd_offset(mm, address);
2055         spin_lock(&mm->page_table_lock);
2056
2057         pud = pud_alloc(mm, pgd, address);
2058         if (!pud)
2059                 goto oom;
2060
2061         pmd = pmd_alloc(mm, pud, address);
2062         if (!pmd)
2063                 goto oom;
2064
2065         pte = pte_alloc_map(mm, pmd, address);
2066         if (!pte)
2067                 goto oom;
2068         
2069         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
2070
2071  oom:
2072         spin_unlock(&mm->page_table_lock);
2073         return VM_FAULT_OOM;
2074 }
2075
2076 #ifndef __PAGETABLE_PUD_FOLDED
2077 /*
2078  * Allocate page upper directory.
2079  *
2080  * We've already handled the fast-path in-line, and we own the
2081  * page table lock.
2082  */
2083 pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2084 {
2085         pud_t *new;
2086
2087         spin_unlock(&mm->page_table_lock);
2088         new = pud_alloc_one(mm, address);
2089         spin_lock(&mm->page_table_lock);
2090         if (!new)
2091                 return NULL;
2092
2093         /*
2094          * Because we dropped the lock, we should re-check the
2095          * entry, as somebody else could have populated it..
2096          */
2097         if (pgd_present(*pgd)) {
2098                 pud_free(new);
2099                 goto out;
2100         }
2101         pgd_populate(mm, pgd, new);
2102  out:
2103         return pud_offset(pgd, address);
2104 }
2105 #endif /* __PAGETABLE_PUD_FOLDED */
2106
2107 #ifndef __PAGETABLE_PMD_FOLDED
2108 /*
2109  * Allocate page middle directory.
2110  *
2111  * We've already handled the fast-path in-line, and we own the
2112  * page table lock.
2113  */
2114 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2115 {
2116         pmd_t *new;
2117
2118         spin_unlock(&mm->page_table_lock);
2119         new = pmd_alloc_one(mm, address);
2120         spin_lock(&mm->page_table_lock);
2121         if (!new)
2122                 return NULL;
2123
2124         /*
2125          * Because we dropped the lock, we should re-check the
2126          * entry, as somebody else could have populated it..
2127          */
2128 #ifndef __ARCH_HAS_4LEVEL_HACK
2129         if (pud_present(*pud)) {
2130                 pmd_free(new);
2131                 goto out;
2132         }
2133         pud_populate(mm, pud, new);
2134 #else
2135         if (pgd_present(*pud)) {
2136                 pmd_free(new);
2137                 goto out;
2138         }
2139         pgd_populate(mm, pud, new);
2140 #endif /* __ARCH_HAS_4LEVEL_HACK */
2141
2142  out:
2143         return pmd_offset(pud, address);
2144 }
2145 #endif /* __PAGETABLE_PMD_FOLDED */
2146
2147 int make_pages_present(unsigned long addr, unsigned long end)
2148 {
2149         int ret, len, write;
2150         struct vm_area_struct * vma;
2151
2152         vma = find_vma(current->mm, addr);
2153         if (!vma)
2154                 return -1;
2155         write = (vma->vm_flags & VM_WRITE) != 0;
2156         if (addr >= end)
2157                 BUG();
2158         if (end > vma->vm_end)
2159                 BUG();
2160         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2161         ret = get_user_pages(current, current->mm, addr,
2162                         len, write, 0, NULL, NULL);
2163         if (ret < 0)
2164                 return ret;
2165         return ret == len ? 0 : -1;
2166 }
2167
2168 /* 
2169  * Map a vmalloc()-space virtual address to the physical page.
2170  */
2171 struct page * vmalloc_to_page(void * vmalloc_addr)
2172 {
2173         unsigned long addr = (unsigned long) vmalloc_addr;
2174         struct page *page = NULL;
2175         pgd_t *pgd = pgd_offset_k(addr);
2176         pud_t *pud;
2177         pmd_t *pmd;
2178         pte_t *ptep, pte;
2179   
2180         if (!pgd_none(*pgd)) {
2181                 pud = pud_offset(pgd, addr);
2182                 if (!pud_none(*pud)) {
2183                         pmd = pmd_offset(pud, addr);
2184                         if (!pmd_none(*pmd)) {
2185                                 ptep = pte_offset_map(pmd, addr);
2186                                 pte = *ptep;
2187                                 if (pte_present(pte))
2188                                         page = pte_page(pte);
2189                                 pte_unmap(ptep);
2190                         }
2191                 }
2192         }
2193         return page;
2194 }
2195
2196 EXPORT_SYMBOL(vmalloc_to_page);
2197
2198 /*
2199  * Map a vmalloc()-space virtual address to the physical page frame number.
2200  */
2201 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2202 {
2203         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2204 }
2205
2206 EXPORT_SYMBOL(vmalloc_to_pfn);
2207
2208 /*
2209  * update_mem_hiwater
2210  *      - update per process rss and vm high water data
2211  */
2212 void update_mem_hiwater(struct task_struct *tsk)
2213 {
2214         if (tsk->mm) {
2215                 unsigned long rss = get_mm_counter(tsk->mm, rss);
2216
2217                 if (tsk->mm->hiwater_rss < rss)
2218                         tsk->mm->hiwater_rss = rss;
2219                 if (tsk->mm->hiwater_vm < tsk->mm->total_vm)
2220                         tsk->mm->hiwater_vm = tsk->mm->total_vm;
2221         }
2222 }
2223
2224 #if !defined(__HAVE_ARCH_GATE_AREA)
2225
2226 #if defined(AT_SYSINFO_EHDR)
2227 struct vm_area_struct gate_vma;
2228
2229 static int __init gate_vma_init(void)
2230 {
2231         gate_vma.vm_mm = NULL;
2232         gate_vma.vm_start = FIXADDR_USER_START;
2233         gate_vma.vm_end = FIXADDR_USER_END;
2234         gate_vma.vm_page_prot = PAGE_READONLY;
2235         gate_vma.vm_flags = 0;
2236         return 0;
2237 }
2238 __initcall(gate_vma_init);
2239 #endif
2240
2241 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2242 {
2243 #ifdef AT_SYSINFO_EHDR
2244         return &gate_vma;
2245 #else
2246         return NULL;
2247 #endif
2248 }
2249
2250 int in_gate_area_no_task(unsigned long addr)
2251 {
2252 #ifdef AT_SYSINFO_EHDR
2253         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2254                 return 1;
2255 #endif
2256         return 0;
2257 }
2258
2259 #endif  /* __HAVE_ARCH_GATE_AREA */