[PATCH] Fix handling spurious page fault for hugetlb region
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
54 #include <asm/tlb.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
57
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
65
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
69
70 unsigned long num_physpages;
71 /*
72  * A number of key systems in x86 including ioremap() rely on the assumption
73  * that high_memory defines the upper bound on direct map memory, then end
74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76  * and ZONE_HIGHMEM.
77  */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
80
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85 /*
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_full(tlb->mm, addr, pte,
566                                                         tlb->fullmm);
567                         tlb_remove_tlb_entry(tlb, pte, addr);
568                         if (unlikely(!page))
569                                 continue;
570                         if (unlikely(details) && details->nonlinear_vma
571                             && linear_page_index(details->nonlinear_vma,
572                                                 addr) != page->index)
573                                 set_pte_at(tlb->mm, addr, pte,
574                                            pgoff_to_pte(page->index));
575                         if (pte_dirty(ptent))
576                                 set_page_dirty(page);
577                         if (PageAnon(page))
578                                 dec_mm_counter(tlb->mm, anon_rss);
579                         else if (pte_young(ptent))
580                                 mark_page_accessed(page);
581                         tlb->freed++;
582                         page_remove_rmap(page);
583                         tlb_remove_page(tlb, page);
584                         continue;
585                 }
586                 /*
587                  * If details->check_mapping, we leave swap entries;
588                  * if details->nonlinear_vma, we leave file entries.
589                  */
590                 if (unlikely(details))
591                         continue;
592                 if (!pte_file(ptent))
593                         free_swap_and_cache(pte_to_swp_entry(ptent));
594                 pte_clear_full(tlb->mm, addr, pte, tlb->fullmm);
595         } while (pte++, addr += PAGE_SIZE, addr != end);
596         pte_unmap(pte - 1);
597 }
598
599 static inline void zap_pmd_range(struct mmu_gather *tlb, pud_t *pud,
600                                 unsigned long addr, unsigned long end,
601                                 struct zap_details *details)
602 {
603         pmd_t *pmd;
604         unsigned long next;
605
606         pmd = pmd_offset(pud, addr);
607         do {
608                 next = pmd_addr_end(addr, end);
609                 if (pmd_none_or_clear_bad(pmd))
610                         continue;
611                 zap_pte_range(tlb, pmd, addr, next, details);
612         } while (pmd++, addr = next, addr != end);
613 }
614
615 static inline void zap_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
616                                 unsigned long addr, unsigned long end,
617                                 struct zap_details *details)
618 {
619         pud_t *pud;
620         unsigned long next;
621
622         pud = pud_offset(pgd, addr);
623         do {
624                 next = pud_addr_end(addr, end);
625                 if (pud_none_or_clear_bad(pud))
626                         continue;
627                 zap_pmd_range(tlb, pud, addr, next, details);
628         } while (pud++, addr = next, addr != end);
629 }
630
631 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
632                                 unsigned long addr, unsigned long end,
633                                 struct zap_details *details)
634 {
635         pgd_t *pgd;
636         unsigned long next;
637
638         if (details && !details->check_mapping && !details->nonlinear_vma)
639                 details = NULL;
640
641         BUG_ON(addr >= end);
642         tlb_start_vma(tlb, vma);
643         pgd = pgd_offset(vma->vm_mm, addr);
644         do {
645                 next = pgd_addr_end(addr, end);
646                 if (pgd_none_or_clear_bad(pgd))
647                         continue;
648                 zap_pud_range(tlb, pgd, addr, next, details);
649         } while (pgd++, addr = next, addr != end);
650         tlb_end_vma(tlb, vma);
651 }
652
653 #ifdef CONFIG_PREEMPT
654 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
655 #else
656 /* No preempt: go for improved straight-line efficiency */
657 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
658 #endif
659
660 /**
661  * unmap_vmas - unmap a range of memory covered by a list of vma's
662  * @tlbp: address of the caller's struct mmu_gather
663  * @mm: the controlling mm_struct
664  * @vma: the starting vma
665  * @start_addr: virtual address at which to start unmapping
666  * @end_addr: virtual address at which to end unmapping
667  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
668  * @details: details of nonlinear truncation or shared cache invalidation
669  *
670  * Returns the end address of the unmapping (restart addr if interrupted).
671  *
672  * Unmap all pages in the vma list.  Called under page_table_lock.
673  *
674  * We aim to not hold page_table_lock for too long (for scheduling latency
675  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
676  * return the ending mmu_gather to the caller.
677  *
678  * Only addresses between `start' and `end' will be unmapped.
679  *
680  * The VMA list must be sorted in ascending virtual address order.
681  *
682  * unmap_vmas() assumes that the caller will flush the whole unmapped address
683  * range after unmap_vmas() returns.  So the only responsibility here is to
684  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
685  * drops the lock and schedules.
686  */
687 unsigned long unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
688                 struct vm_area_struct *vma, unsigned long start_addr,
689                 unsigned long end_addr, unsigned long *nr_accounted,
690                 struct zap_details *details)
691 {
692         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
693         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
694         int tlb_start_valid = 0;
695         unsigned long start = start_addr;
696         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
697         int fullmm = tlb_is_full_mm(*tlbp);
698
699         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
700                 unsigned long end;
701
702                 start = max(vma->vm_start, start_addr);
703                 if (start >= vma->vm_end)
704                         continue;
705                 end = min(vma->vm_end, end_addr);
706                 if (end <= vma->vm_start)
707                         continue;
708
709                 if (vma->vm_flags & VM_ACCOUNT)
710                         *nr_accounted += (end - start) >> PAGE_SHIFT;
711
712                 while (start != end) {
713                         unsigned long block;
714
715                         if (!tlb_start_valid) {
716                                 tlb_start = start;
717                                 tlb_start_valid = 1;
718                         }
719
720                         if (is_vm_hugetlb_page(vma)) {
721                                 block = end - start;
722                                 unmap_hugepage_range(vma, start, end);
723                         } else {
724                                 block = min(zap_bytes, end - start);
725                                 unmap_page_range(*tlbp, vma, start,
726                                                 start + block, details);
727                         }
728
729                         start += block;
730                         zap_bytes -= block;
731                         if ((long)zap_bytes > 0)
732                                 continue;
733
734                         tlb_finish_mmu(*tlbp, tlb_start, start);
735
736                         if (need_resched() ||
737                                 need_lockbreak(&mm->page_table_lock) ||
738                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
739                                 if (i_mmap_lock) {
740                                         /* must reset count of rss freed */
741                                         *tlbp = tlb_gather_mmu(mm, fullmm);
742                                         goto out;
743                                 }
744                                 spin_unlock(&mm->page_table_lock);
745                                 cond_resched();
746                                 spin_lock(&mm->page_table_lock);
747                         }
748
749                         *tlbp = tlb_gather_mmu(mm, fullmm);
750                         tlb_start_valid = 0;
751                         zap_bytes = ZAP_BLOCK_SIZE;
752                 }
753         }
754 out:
755         return start;   /* which is now the end (or restart) address */
756 }
757
758 /**
759  * zap_page_range - remove user pages in a given range
760  * @vma: vm_area_struct holding the applicable pages
761  * @address: starting address of pages to zap
762  * @size: number of bytes to zap
763  * @details: details of nonlinear truncation or shared cache invalidation
764  */
765 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
766                 unsigned long size, struct zap_details *details)
767 {
768         struct mm_struct *mm = vma->vm_mm;
769         struct mmu_gather *tlb;
770         unsigned long end = address + size;
771         unsigned long nr_accounted = 0;
772
773         if (is_vm_hugetlb_page(vma)) {
774                 zap_hugepage_range(vma, address, size);
775                 return end;
776         }
777
778         lru_add_drain();
779         spin_lock(&mm->page_table_lock);
780         tlb = tlb_gather_mmu(mm, 0);
781         end = unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
782         tlb_finish_mmu(tlb, address, end);
783         spin_unlock(&mm->page_table_lock);
784         return end;
785 }
786
787 /*
788  * Do a quick page-table lookup for a single page.
789  * mm->page_table_lock must be held.
790  */
791 static struct page *__follow_page(struct mm_struct *mm, unsigned long address,
792                         int read, int write, int accessed)
793 {
794         pgd_t *pgd;
795         pud_t *pud;
796         pmd_t *pmd;
797         pte_t *ptep, pte;
798         unsigned long pfn;
799         struct page *page;
800
801         page = follow_huge_addr(mm, address, write);
802         if (! IS_ERR(page))
803                 return page;
804
805         pgd = pgd_offset(mm, address);
806         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
807                 goto out;
808
809         pud = pud_offset(pgd, address);
810         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
811                 goto out;
812         
813         pmd = pmd_offset(pud, address);
814         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
815                 goto out;
816         if (pmd_huge(*pmd))
817                 return follow_huge_pmd(mm, address, pmd, write);
818
819         ptep = pte_offset_map(pmd, address);
820         if (!ptep)
821                 goto out;
822
823         pte = *ptep;
824         pte_unmap(ptep);
825         if (pte_present(pte)) {
826                 if (write && !pte_write(pte))
827                         goto out;
828                 if (read && !pte_read(pte))
829                         goto out;
830                 pfn = pte_pfn(pte);
831                 if (pfn_valid(pfn)) {
832                         page = pfn_to_page(pfn);
833                         if (accessed) {
834                                 if (write && !pte_dirty(pte) &&!PageDirty(page))
835                                         set_page_dirty(page);
836                                 mark_page_accessed(page);
837                         }
838                         return page;
839                 }
840         }
841
842 out:
843         return NULL;
844 }
845
846 inline struct page *
847 follow_page(struct mm_struct *mm, unsigned long address, int write)
848 {
849         return __follow_page(mm, address, 0, write, 1);
850 }
851
852 /*
853  * check_user_page_readable() can be called frm niterrupt context by oprofile,
854  * so we need to avoid taking any non-irq-safe locks
855  */
856 int check_user_page_readable(struct mm_struct *mm, unsigned long address)
857 {
858         return __follow_page(mm, address, 1, 0, 0) != NULL;
859 }
860 EXPORT_SYMBOL(check_user_page_readable);
861
862 static inline int
863 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
864                          unsigned long address)
865 {
866         pgd_t *pgd;
867         pud_t *pud;
868         pmd_t *pmd;
869
870         /* Check if the vma is for an anonymous mapping. */
871         if (vma->vm_ops && vma->vm_ops->nopage)
872                 return 0;
873
874         /* Check if page directory entry exists. */
875         pgd = pgd_offset(mm, address);
876         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
877                 return 1;
878
879         pud = pud_offset(pgd, address);
880         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
881                 return 1;
882
883         /* Check if page middle directory entry exists. */
884         pmd = pmd_offset(pud, address);
885         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
886                 return 1;
887
888         /* There is a pte slot for 'address' in 'mm'. */
889         return 0;
890 }
891
892 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
893                 unsigned long start, int len, int write, int force,
894                 struct page **pages, struct vm_area_struct **vmas)
895 {
896         int i;
897         unsigned int flags;
898
899         /* 
900          * Require read or write permissions.
901          * If 'force' is set, we only require the "MAY" flags.
902          */
903         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
904         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
905         i = 0;
906
907         do {
908                 struct vm_area_struct * vma;
909
910                 vma = find_extend_vma(mm, start);
911                 if (!vma && in_gate_area(tsk, start)) {
912                         unsigned long pg = start & PAGE_MASK;
913                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
914                         pgd_t *pgd;
915                         pud_t *pud;
916                         pmd_t *pmd;
917                         pte_t *pte;
918                         if (write) /* user gate pages are read-only */
919                                 return i ? : -EFAULT;
920                         if (pg > TASK_SIZE)
921                                 pgd = pgd_offset_k(pg);
922                         else
923                                 pgd = pgd_offset_gate(mm, pg);
924                         BUG_ON(pgd_none(*pgd));
925                         pud = pud_offset(pgd, pg);
926                         BUG_ON(pud_none(*pud));
927                         pmd = pmd_offset(pud, pg);
928                         if (pmd_none(*pmd))
929                                 return i ? : -EFAULT;
930                         pte = pte_offset_map(pmd, pg);
931                         if (pte_none(*pte)) {
932                                 pte_unmap(pte);
933                                 return i ? : -EFAULT;
934                         }
935                         if (pages) {
936                                 pages[i] = pte_page(*pte);
937                                 get_page(pages[i]);
938                         }
939                         pte_unmap(pte);
940                         if (vmas)
941                                 vmas[i] = gate_vma;
942                         i++;
943                         start += PAGE_SIZE;
944                         len--;
945                         continue;
946                 }
947
948                 if (!vma || (vma->vm_flags & VM_IO)
949                                 || !(flags & vma->vm_flags))
950                         return i ? : -EFAULT;
951
952                 if (is_vm_hugetlb_page(vma)) {
953                         i = follow_hugetlb_page(mm, vma, pages, vmas,
954                                                 &start, &len, i);
955                         continue;
956                 }
957                 spin_lock(&mm->page_table_lock);
958                 do {
959                         int write_access = write;
960                         struct page *page;
961
962                         cond_resched_lock(&mm->page_table_lock);
963                         while (!(page = follow_page(mm, start, write_access))) {
964                                 int ret;
965
966                                 /*
967                                  * Shortcut for anonymous pages. We don't want
968                                  * to force the creation of pages tables for
969                                  * insanely big anonymously mapped areas that
970                                  * nobody touched so far. This is important
971                                  * for doing a core dump for these mappings.
972                                  */
973                                 if (!write && untouched_anonymous_page(mm,vma,start)) {
974                                         page = ZERO_PAGE(start);
975                                         break;
976                                 }
977                                 spin_unlock(&mm->page_table_lock);
978                                 ret = __handle_mm_fault(mm, vma, start, write_access);
979
980                                 /*
981                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
982                                  * broken COW when necessary, even if maybe_mkwrite
983                                  * decided not to set pte_write. We can thus safely do
984                                  * subsequent page lookups as if they were reads.
985                                  */
986                                 if (ret & VM_FAULT_WRITE)
987                                         write_access = 0;
988                                 
989                                 switch (ret & ~VM_FAULT_WRITE) {
990                                 case VM_FAULT_MINOR:
991                                         tsk->min_flt++;
992                                         break;
993                                 case VM_FAULT_MAJOR:
994                                         tsk->maj_flt++;
995                                         break;
996                                 case VM_FAULT_SIGBUS:
997                                         return i ? i : -EFAULT;
998                                 case VM_FAULT_OOM:
999                                         return i ? i : -ENOMEM;
1000                                 default:
1001                                         BUG();
1002                                 }
1003                                 spin_lock(&mm->page_table_lock);
1004                         }
1005                         if (pages) {
1006                                 pages[i] = page;
1007                                 flush_dcache_page(page);
1008                                 if (!PageReserved(page))
1009                                         page_cache_get(page);
1010                         }
1011                         if (vmas)
1012                                 vmas[i] = vma;
1013                         i++;
1014                         start += PAGE_SIZE;
1015                         len--;
1016                 } while (len && start < vma->vm_end);
1017                 spin_unlock(&mm->page_table_lock);
1018         } while (len);
1019         return i;
1020 }
1021 EXPORT_SYMBOL(get_user_pages);
1022
1023 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1024                         unsigned long addr, unsigned long end, pgprot_t prot)
1025 {
1026         pte_t *pte;
1027
1028         pte = pte_alloc_map(mm, pmd, addr);
1029         if (!pte)
1030                 return -ENOMEM;
1031         do {
1032                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(addr), prot));
1033                 BUG_ON(!pte_none(*pte));
1034                 set_pte_at(mm, addr, pte, zero_pte);
1035         } while (pte++, addr += PAGE_SIZE, addr != end);
1036         pte_unmap(pte - 1);
1037         return 0;
1038 }
1039
1040 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1041                         unsigned long addr, unsigned long end, pgprot_t prot)
1042 {
1043         pmd_t *pmd;
1044         unsigned long next;
1045
1046         pmd = pmd_alloc(mm, pud, addr);
1047         if (!pmd)
1048                 return -ENOMEM;
1049         do {
1050                 next = pmd_addr_end(addr, end);
1051                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1052                         return -ENOMEM;
1053         } while (pmd++, addr = next, addr != end);
1054         return 0;
1055 }
1056
1057 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1058                         unsigned long addr, unsigned long end, pgprot_t prot)
1059 {
1060         pud_t *pud;
1061         unsigned long next;
1062
1063         pud = pud_alloc(mm, pgd, addr);
1064         if (!pud)
1065                 return -ENOMEM;
1066         do {
1067                 next = pud_addr_end(addr, end);
1068                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1069                         return -ENOMEM;
1070         } while (pud++, addr = next, addr != end);
1071         return 0;
1072 }
1073
1074 int zeromap_page_range(struct vm_area_struct *vma,
1075                         unsigned long addr, unsigned long size, pgprot_t prot)
1076 {
1077         pgd_t *pgd;
1078         unsigned long next;
1079         unsigned long end = addr + size;
1080         struct mm_struct *mm = vma->vm_mm;
1081         int err;
1082
1083         BUG_ON(addr >= end);
1084         pgd = pgd_offset(mm, addr);
1085         flush_cache_range(vma, addr, end);
1086         spin_lock(&mm->page_table_lock);
1087         do {
1088                 next = pgd_addr_end(addr, end);
1089                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1090                 if (err)
1091                         break;
1092         } while (pgd++, addr = next, addr != end);
1093         spin_unlock(&mm->page_table_lock);
1094         return err;
1095 }
1096
1097 /*
1098  * maps a range of physical memory into the requested pages. the old
1099  * mappings are removed. any references to nonexistent pages results
1100  * in null mappings (currently treated as "copy-on-access")
1101  */
1102 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1103                         unsigned long addr, unsigned long end,
1104                         unsigned long pfn, pgprot_t prot)
1105 {
1106         pte_t *pte;
1107
1108         pte = pte_alloc_map(mm, pmd, addr);
1109         if (!pte)
1110                 return -ENOMEM;
1111         do {
1112                 BUG_ON(!pte_none(*pte));
1113                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
1114                         set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1115                 pfn++;
1116         } while (pte++, addr += PAGE_SIZE, addr != end);
1117         pte_unmap(pte - 1);
1118         return 0;
1119 }
1120
1121 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1122                         unsigned long addr, unsigned long end,
1123                         unsigned long pfn, pgprot_t prot)
1124 {
1125         pmd_t *pmd;
1126         unsigned long next;
1127
1128         pfn -= addr >> PAGE_SHIFT;
1129         pmd = pmd_alloc(mm, pud, addr);
1130         if (!pmd)
1131                 return -ENOMEM;
1132         do {
1133                 next = pmd_addr_end(addr, end);
1134                 if (remap_pte_range(mm, pmd, addr, next,
1135                                 pfn + (addr >> PAGE_SHIFT), prot))
1136                         return -ENOMEM;
1137         } while (pmd++, addr = next, addr != end);
1138         return 0;
1139 }
1140
1141 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1142                         unsigned long addr, unsigned long end,
1143                         unsigned long pfn, pgprot_t prot)
1144 {
1145         pud_t *pud;
1146         unsigned long next;
1147
1148         pfn -= addr >> PAGE_SHIFT;
1149         pud = pud_alloc(mm, pgd, addr);
1150         if (!pud)
1151                 return -ENOMEM;
1152         do {
1153                 next = pud_addr_end(addr, end);
1154                 if (remap_pmd_range(mm, pud, addr, next,
1155                                 pfn + (addr >> PAGE_SHIFT), prot))
1156                         return -ENOMEM;
1157         } while (pud++, addr = next, addr != end);
1158         return 0;
1159 }
1160
1161 /*  Note: this is only safe if the mm semaphore is held when called. */
1162 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1163                     unsigned long pfn, unsigned long size, pgprot_t prot)
1164 {
1165         pgd_t *pgd;
1166         unsigned long next;
1167         unsigned long end = addr + PAGE_ALIGN(size);
1168         struct mm_struct *mm = vma->vm_mm;
1169         int err;
1170
1171         /*
1172          * Physically remapped pages are special. Tell the
1173          * rest of the world about it:
1174          *   VM_IO tells people not to look at these pages
1175          *      (accesses can have side effects).
1176          *   VM_RESERVED tells swapout not to try to touch
1177          *      this region.
1178          */
1179         vma->vm_flags |= VM_IO | VM_RESERVED;
1180
1181         BUG_ON(addr >= end);
1182         pfn -= addr >> PAGE_SHIFT;
1183         pgd = pgd_offset(mm, addr);
1184         flush_cache_range(vma, addr, end);
1185         spin_lock(&mm->page_table_lock);
1186         do {
1187                 next = pgd_addr_end(addr, end);
1188                 err = remap_pud_range(mm, pgd, addr, next,
1189                                 pfn + (addr >> PAGE_SHIFT), prot);
1190                 if (err)
1191                         break;
1192         } while (pgd++, addr = next, addr != end);
1193         spin_unlock(&mm->page_table_lock);
1194         return err;
1195 }
1196 EXPORT_SYMBOL(remap_pfn_range);
1197
1198 /*
1199  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1200  * servicing faults for write access.  In the normal case, do always want
1201  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1202  * that do not have writing enabled, when used by access_process_vm.
1203  */
1204 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1205 {
1206         if (likely(vma->vm_flags & VM_WRITE))
1207                 pte = pte_mkwrite(pte);
1208         return pte;
1209 }
1210
1211 /*
1212  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1213  */
1214 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address, 
1215                 pte_t *page_table)
1216 {
1217         pte_t entry;
1218
1219         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1220                               vma);
1221         ptep_establish(vma, address, page_table, entry);
1222         update_mmu_cache(vma, address, entry);
1223         lazy_mmu_prot_update(entry);
1224 }
1225
1226 /*
1227  * This routine handles present pages, when users try to write
1228  * to a shared page. It is done by copying the page to a new address
1229  * and decrementing the shared-page counter for the old page.
1230  *
1231  * Goto-purists beware: the only reason for goto's here is that it results
1232  * in better assembly code.. The "default" path will see no jumps at all.
1233  *
1234  * Note that this routine assumes that the protection checks have been
1235  * done by the caller (the low-level page fault routine in most cases).
1236  * Thus we can safely just mark it writable once we've done any necessary
1237  * COW.
1238  *
1239  * We also mark the page dirty at this point even though the page will
1240  * change only once the write actually happens. This avoids a few races,
1241  * and potentially makes it more efficient.
1242  *
1243  * We hold the mm semaphore and the page_table_lock on entry and exit
1244  * with the page_table_lock released.
1245  */
1246 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1247         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1248 {
1249         struct page *old_page, *new_page;
1250         unsigned long pfn = pte_pfn(pte);
1251         pte_t entry;
1252         int ret;
1253
1254         if (unlikely(!pfn_valid(pfn))) {
1255                 /*
1256                  * This should really halt the system so it can be debugged or
1257                  * at least the kernel stops what it's doing before it corrupts
1258                  * data, but for the moment just pretend this is OOM.
1259                  */
1260                 pte_unmap(page_table);
1261                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1262                                 address);
1263                 spin_unlock(&mm->page_table_lock);
1264                 return VM_FAULT_OOM;
1265         }
1266         old_page = pfn_to_page(pfn);
1267
1268         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1269                 int reuse = can_share_swap_page(old_page);
1270                 unlock_page(old_page);
1271                 if (reuse) {
1272                         flush_cache_page(vma, address, pfn);
1273                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1274                                               vma);
1275                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1276                         update_mmu_cache(vma, address, entry);
1277                         lazy_mmu_prot_update(entry);
1278                         pte_unmap(page_table);
1279                         spin_unlock(&mm->page_table_lock);
1280                         return VM_FAULT_MINOR|VM_FAULT_WRITE;
1281                 }
1282         }
1283         pte_unmap(page_table);
1284
1285         /*
1286          * Ok, we need to copy. Oh, well..
1287          */
1288         if (!PageReserved(old_page))
1289                 page_cache_get(old_page);
1290         spin_unlock(&mm->page_table_lock);
1291
1292         if (unlikely(anon_vma_prepare(vma)))
1293                 goto no_new_page;
1294         if (old_page == ZERO_PAGE(address)) {
1295                 new_page = alloc_zeroed_user_highpage(vma, address);
1296                 if (!new_page)
1297                         goto no_new_page;
1298         } else {
1299                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1300                 if (!new_page)
1301                         goto no_new_page;
1302                 copy_user_highpage(new_page, old_page, address);
1303         }
1304         /*
1305          * Re-check the pte - we dropped the lock
1306          */
1307         ret = VM_FAULT_MINOR;
1308         spin_lock(&mm->page_table_lock);
1309         page_table = pte_offset_map(pmd, address);
1310         if (likely(pte_same(*page_table, pte))) {
1311                 if (PageAnon(old_page))
1312                         dec_mm_counter(mm, anon_rss);
1313                 if (PageReserved(old_page))
1314                         inc_mm_counter(mm, rss);
1315                 else
1316                         page_remove_rmap(old_page);
1317                 flush_cache_page(vma, address, pfn);
1318                 break_cow(vma, new_page, address, page_table);
1319                 lru_cache_add_active(new_page);
1320                 page_add_anon_rmap(new_page, vma, address);
1321
1322                 /* Free the old page.. */
1323                 new_page = old_page;
1324                 ret |= VM_FAULT_WRITE;
1325         }
1326         pte_unmap(page_table);
1327         page_cache_release(new_page);
1328         page_cache_release(old_page);
1329         spin_unlock(&mm->page_table_lock);
1330         return ret;
1331
1332 no_new_page:
1333         page_cache_release(old_page);
1334         return VM_FAULT_OOM;
1335 }
1336
1337 /*
1338  * Helper functions for unmap_mapping_range().
1339  *
1340  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1341  *
1342  * We have to restart searching the prio_tree whenever we drop the lock,
1343  * since the iterator is only valid while the lock is held, and anyway
1344  * a later vma might be split and reinserted earlier while lock dropped.
1345  *
1346  * The list of nonlinear vmas could be handled more efficiently, using
1347  * a placeholder, but handle it in the same way until a need is shown.
1348  * It is important to search the prio_tree before nonlinear list: a vma
1349  * may become nonlinear and be shifted from prio_tree to nonlinear list
1350  * while the lock is dropped; but never shifted from list to prio_tree.
1351  *
1352  * In order to make forward progress despite restarting the search,
1353  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1354  * quickly skip it next time around.  Since the prio_tree search only
1355  * shows us those vmas affected by unmapping the range in question, we
1356  * can't efficiently keep all vmas in step with mapping->truncate_count:
1357  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1358  * mapping->truncate_count and vma->vm_truncate_count are protected by
1359  * i_mmap_lock.
1360  *
1361  * In order to make forward progress despite repeatedly restarting some
1362  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1363  * and restart from that address when we reach that vma again.  It might
1364  * have been split or merged, shrunk or extended, but never shifted: so
1365  * restart_addr remains valid so long as it remains in the vma's range.
1366  * unmap_mapping_range forces truncate_count to leap over page-aligned
1367  * values so we can save vma's restart_addr in its truncate_count field.
1368  */
1369 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1370
1371 static void reset_vma_truncate_counts(struct address_space *mapping)
1372 {
1373         struct vm_area_struct *vma;
1374         struct prio_tree_iter iter;
1375
1376         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1377                 vma->vm_truncate_count = 0;
1378         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1379                 vma->vm_truncate_count = 0;
1380 }
1381
1382 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1383                 unsigned long start_addr, unsigned long end_addr,
1384                 struct zap_details *details)
1385 {
1386         unsigned long restart_addr;
1387         int need_break;
1388
1389 again:
1390         restart_addr = vma->vm_truncate_count;
1391         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1392                 start_addr = restart_addr;
1393                 if (start_addr >= end_addr) {
1394                         /* Top of vma has been split off since last time */
1395                         vma->vm_truncate_count = details->truncate_count;
1396                         return 0;
1397                 }
1398         }
1399
1400         restart_addr = zap_page_range(vma, start_addr,
1401                                         end_addr - start_addr, details);
1402
1403         /*
1404          * We cannot rely on the break test in unmap_vmas:
1405          * on the one hand, we don't want to restart our loop
1406          * just because that broke out for the page_table_lock;
1407          * on the other hand, it does no test when vma is small.
1408          */
1409         need_break = need_resched() ||
1410                         need_lockbreak(details->i_mmap_lock);
1411
1412         if (restart_addr >= end_addr) {
1413                 /* We have now completed this vma: mark it so */
1414                 vma->vm_truncate_count = details->truncate_count;
1415                 if (!need_break)
1416                         return 0;
1417         } else {
1418                 /* Note restart_addr in vma's truncate_count field */
1419                 vma->vm_truncate_count = restart_addr;
1420                 if (!need_break)
1421                         goto again;
1422         }
1423
1424         spin_unlock(details->i_mmap_lock);
1425         cond_resched();
1426         spin_lock(details->i_mmap_lock);
1427         return -EINTR;
1428 }
1429
1430 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1431                                             struct zap_details *details)
1432 {
1433         struct vm_area_struct *vma;
1434         struct prio_tree_iter iter;
1435         pgoff_t vba, vea, zba, zea;
1436
1437 restart:
1438         vma_prio_tree_foreach(vma, &iter, root,
1439                         details->first_index, details->last_index) {
1440                 /* Skip quickly over those we have already dealt with */
1441                 if (vma->vm_truncate_count == details->truncate_count)
1442                         continue;
1443
1444                 vba = vma->vm_pgoff;
1445                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1446                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1447                 zba = details->first_index;
1448                 if (zba < vba)
1449                         zba = vba;
1450                 zea = details->last_index;
1451                 if (zea > vea)
1452                         zea = vea;
1453
1454                 if (unmap_mapping_range_vma(vma,
1455                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1456                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1457                                 details) < 0)
1458                         goto restart;
1459         }
1460 }
1461
1462 static inline void unmap_mapping_range_list(struct list_head *head,
1463                                             struct zap_details *details)
1464 {
1465         struct vm_area_struct *vma;
1466
1467         /*
1468          * In nonlinear VMAs there is no correspondence between virtual address
1469          * offset and file offset.  So we must perform an exhaustive search
1470          * across *all* the pages in each nonlinear VMA, not just the pages
1471          * whose virtual address lies outside the file truncation point.
1472          */
1473 restart:
1474         list_for_each_entry(vma, head, shared.vm_set.list) {
1475                 /* Skip quickly over those we have already dealt with */
1476                 if (vma->vm_truncate_count == details->truncate_count)
1477                         continue;
1478                 details->nonlinear_vma = vma;
1479                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1480                                         vma->vm_end, details) < 0)
1481                         goto restart;
1482         }
1483 }
1484
1485 /**
1486  * unmap_mapping_range - unmap the portion of all mmaps
1487  * in the specified address_space corresponding to the specified
1488  * page range in the underlying file.
1489  * @mapping: the address space containing mmaps to be unmapped.
1490  * @holebegin: byte in first page to unmap, relative to the start of
1491  * the underlying file.  This will be rounded down to a PAGE_SIZE
1492  * boundary.  Note that this is different from vmtruncate(), which
1493  * must keep the partial page.  In contrast, we must get rid of
1494  * partial pages.
1495  * @holelen: size of prospective hole in bytes.  This will be rounded
1496  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1497  * end of the file.
1498  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1499  * but 0 when invalidating pagecache, don't throw away private data.
1500  */
1501 void unmap_mapping_range(struct address_space *mapping,
1502                 loff_t const holebegin, loff_t const holelen, int even_cows)
1503 {
1504         struct zap_details details;
1505         pgoff_t hba = holebegin >> PAGE_SHIFT;
1506         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1507
1508         /* Check for overflow. */
1509         if (sizeof(holelen) > sizeof(hlen)) {
1510                 long long holeend =
1511                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1512                 if (holeend & ~(long long)ULONG_MAX)
1513                         hlen = ULONG_MAX - hba + 1;
1514         }
1515
1516         details.check_mapping = even_cows? NULL: mapping;
1517         details.nonlinear_vma = NULL;
1518         details.first_index = hba;
1519         details.last_index = hba + hlen - 1;
1520         if (details.last_index < details.first_index)
1521                 details.last_index = ULONG_MAX;
1522         details.i_mmap_lock = &mapping->i_mmap_lock;
1523
1524         spin_lock(&mapping->i_mmap_lock);
1525
1526         /* serialize i_size write against truncate_count write */
1527         smp_wmb();
1528         /* Protect against page faults, and endless unmapping loops */
1529         mapping->truncate_count++;
1530         /*
1531          * For archs where spin_lock has inclusive semantics like ia64
1532          * this smp_mb() will prevent to read pagetable contents
1533          * before the truncate_count increment is visible to
1534          * other cpus.
1535          */
1536         smp_mb();
1537         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1538                 if (mapping->truncate_count == 0)
1539                         reset_vma_truncate_counts(mapping);
1540                 mapping->truncate_count++;
1541         }
1542         details.truncate_count = mapping->truncate_count;
1543
1544         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1545                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1546         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1547                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1548         spin_unlock(&mapping->i_mmap_lock);
1549 }
1550 EXPORT_SYMBOL(unmap_mapping_range);
1551
1552 /*
1553  * Handle all mappings that got truncated by a "truncate()"
1554  * system call.
1555  *
1556  * NOTE! We have to be ready to update the memory sharing
1557  * between the file and the memory map for a potential last
1558  * incomplete page.  Ugly, but necessary.
1559  */
1560 int vmtruncate(struct inode * inode, loff_t offset)
1561 {
1562         struct address_space *mapping = inode->i_mapping;
1563         unsigned long limit;
1564
1565         if (inode->i_size < offset)
1566                 goto do_expand;
1567         /*
1568          * truncation of in-use swapfiles is disallowed - it would cause
1569          * subsequent swapout to scribble on the now-freed blocks.
1570          */
1571         if (IS_SWAPFILE(inode))
1572                 goto out_busy;
1573         i_size_write(inode, offset);
1574         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1575         truncate_inode_pages(mapping, offset);
1576         goto out_truncate;
1577
1578 do_expand:
1579         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1580         if (limit != RLIM_INFINITY && offset > limit)
1581                 goto out_sig;
1582         if (offset > inode->i_sb->s_maxbytes)
1583                 goto out_big;
1584         i_size_write(inode, offset);
1585
1586 out_truncate:
1587         if (inode->i_op && inode->i_op->truncate)
1588                 inode->i_op->truncate(inode);
1589         return 0;
1590 out_sig:
1591         send_sig(SIGXFSZ, current, 0);
1592 out_big:
1593         return -EFBIG;
1594 out_busy:
1595         return -ETXTBSY;
1596 }
1597
1598 EXPORT_SYMBOL(vmtruncate);
1599
1600 /* 
1601  * Primitive swap readahead code. We simply read an aligned block of
1602  * (1 << page_cluster) entries in the swap area. This method is chosen
1603  * because it doesn't cost us any seek time.  We also make sure to queue
1604  * the 'original' request together with the readahead ones...  
1605  *
1606  * This has been extended to use the NUMA policies from the mm triggering
1607  * the readahead.
1608  *
1609  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1610  */
1611 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1612 {
1613 #ifdef CONFIG_NUMA
1614         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1615 #endif
1616         int i, num;
1617         struct page *new_page;
1618         unsigned long offset;
1619
1620         /*
1621          * Get the number of handles we should do readahead io to.
1622          */
1623         num = valid_swaphandles(entry, &offset);
1624         for (i = 0; i < num; offset++, i++) {
1625                 /* Ok, do the async read-ahead now */
1626                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1627                                                            offset), vma, addr);
1628                 if (!new_page)
1629                         break;
1630                 page_cache_release(new_page);
1631 #ifdef CONFIG_NUMA
1632                 /*
1633                  * Find the next applicable VMA for the NUMA policy.
1634                  */
1635                 addr += PAGE_SIZE;
1636                 if (addr == 0)
1637                         vma = NULL;
1638                 if (vma) {
1639                         if (addr >= vma->vm_end) {
1640                                 vma = next_vma;
1641                                 next_vma = vma ? vma->vm_next : NULL;
1642                         }
1643                         if (vma && addr < vma->vm_start)
1644                                 vma = NULL;
1645                 } else {
1646                         if (next_vma && addr >= next_vma->vm_start) {
1647                                 vma = next_vma;
1648                                 next_vma = vma->vm_next;
1649                         }
1650                 }
1651 #endif
1652         }
1653         lru_add_drain();        /* Push any new pages onto the LRU now */
1654 }
1655
1656 /*
1657  * We hold the mm semaphore and the page_table_lock on entry and
1658  * should release the pagetable lock on exit..
1659  */
1660 static int do_swap_page(struct mm_struct * mm,
1661         struct vm_area_struct * vma, unsigned long address,
1662         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1663 {
1664         struct page *page;
1665         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1666         pte_t pte;
1667         int ret = VM_FAULT_MINOR;
1668
1669         pte_unmap(page_table);
1670         spin_unlock(&mm->page_table_lock);
1671         page = lookup_swap_cache(entry);
1672         if (!page) {
1673                 swapin_readahead(entry, address, vma);
1674                 page = read_swap_cache_async(entry, vma, address);
1675                 if (!page) {
1676                         /*
1677                          * Back out if somebody else faulted in this pte while
1678                          * we released the page table lock.
1679                          */
1680                         spin_lock(&mm->page_table_lock);
1681                         page_table = pte_offset_map(pmd, address);
1682                         if (likely(pte_same(*page_table, orig_pte)))
1683                                 ret = VM_FAULT_OOM;
1684                         else
1685                                 ret = VM_FAULT_MINOR;
1686                         pte_unmap(page_table);
1687                         spin_unlock(&mm->page_table_lock);
1688                         goto out;
1689                 }
1690
1691                 /* Had to read the page from swap area: Major fault */
1692                 ret = VM_FAULT_MAJOR;
1693                 inc_page_state(pgmajfault);
1694                 grab_swap_token();
1695         }
1696
1697         mark_page_accessed(page);
1698         lock_page(page);
1699
1700         /*
1701          * Back out if somebody else faulted in this pte while we
1702          * released the page table lock.
1703          */
1704         spin_lock(&mm->page_table_lock);
1705         page_table = pte_offset_map(pmd, address);
1706         if (unlikely(!pte_same(*page_table, orig_pte))) {
1707                 ret = VM_FAULT_MINOR;
1708                 goto out_nomap;
1709         }
1710
1711         if (unlikely(!PageUptodate(page))) {
1712                 ret = VM_FAULT_SIGBUS;
1713                 goto out_nomap;
1714         }
1715
1716         /* The page isn't present yet, go ahead with the fault. */
1717
1718         inc_mm_counter(mm, rss);
1719         pte = mk_pte(page, vma->vm_page_prot);
1720         if (write_access && can_share_swap_page(page)) {
1721                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1722                 write_access = 0;
1723         }
1724
1725         flush_icache_page(vma, page);
1726         set_pte_at(mm, address, page_table, pte);
1727         page_add_anon_rmap(page, vma, address);
1728
1729         swap_free(entry);
1730         if (vm_swap_full())
1731                 remove_exclusive_swap_page(page);
1732         unlock_page(page);
1733
1734         if (write_access) {
1735                 if (do_wp_page(mm, vma, address,
1736                                 page_table, pmd, pte) == VM_FAULT_OOM)
1737                         ret = VM_FAULT_OOM;
1738                 goto out;
1739         }
1740
1741         /* No need to invalidate - it was non-present before */
1742         update_mmu_cache(vma, address, pte);
1743         lazy_mmu_prot_update(pte);
1744         pte_unmap(page_table);
1745         spin_unlock(&mm->page_table_lock);
1746 out:
1747         return ret;
1748 out_nomap:
1749         pte_unmap(page_table);
1750         spin_unlock(&mm->page_table_lock);
1751         unlock_page(page);
1752         page_cache_release(page);
1753         goto out;
1754 }
1755
1756 /*
1757  * We are called with the MM semaphore and page_table_lock
1758  * spinlock held to protect against concurrent faults in
1759  * multithreaded programs. 
1760  */
1761 static int
1762 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1763                 pte_t *page_table, pmd_t *pmd, int write_access,
1764                 unsigned long addr)
1765 {
1766         pte_t entry;
1767         struct page * page = ZERO_PAGE(addr);
1768
1769         /* Read-only mapping of ZERO_PAGE. */
1770         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1771
1772         /* ..except if it's a write access */
1773         if (write_access) {
1774                 /* Allocate our own private page. */
1775                 pte_unmap(page_table);
1776                 spin_unlock(&mm->page_table_lock);
1777
1778                 if (unlikely(anon_vma_prepare(vma)))
1779                         goto no_mem;
1780                 page = alloc_zeroed_user_highpage(vma, addr);
1781                 if (!page)
1782                         goto no_mem;
1783
1784                 spin_lock(&mm->page_table_lock);
1785                 page_table = pte_offset_map(pmd, addr);
1786
1787                 if (!pte_none(*page_table)) {
1788                         pte_unmap(page_table);
1789                         page_cache_release(page);
1790                         spin_unlock(&mm->page_table_lock);
1791                         goto out;
1792                 }
1793                 inc_mm_counter(mm, rss);
1794                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1795                                                          vma->vm_page_prot)),
1796                                       vma);
1797                 lru_cache_add_active(page);
1798                 SetPageReferenced(page);
1799                 page_add_anon_rmap(page, vma, addr);
1800         }
1801
1802         set_pte_at(mm, addr, page_table, entry);
1803         pte_unmap(page_table);
1804
1805         /* No need to invalidate - it was non-present before */
1806         update_mmu_cache(vma, addr, entry);
1807         lazy_mmu_prot_update(entry);
1808         spin_unlock(&mm->page_table_lock);
1809 out:
1810         return VM_FAULT_MINOR;
1811 no_mem:
1812         return VM_FAULT_OOM;
1813 }
1814
1815 /*
1816  * do_no_page() tries to create a new page mapping. It aggressively
1817  * tries to share with existing pages, but makes a separate copy if
1818  * the "write_access" parameter is true in order to avoid the next
1819  * page fault.
1820  *
1821  * As this is called only for pages that do not currently exist, we
1822  * do not need to flush old virtual caches or the TLB.
1823  *
1824  * This is called with the MM semaphore held and the page table
1825  * spinlock held. Exit with the spinlock released.
1826  */
1827 static int
1828 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1829         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1830 {
1831         struct page * new_page;
1832         struct address_space *mapping = NULL;
1833         pte_t entry;
1834         unsigned int sequence = 0;
1835         int ret = VM_FAULT_MINOR;
1836         int anon = 0;
1837
1838         if (!vma->vm_ops || !vma->vm_ops->nopage)
1839                 return do_anonymous_page(mm, vma, page_table,
1840                                         pmd, write_access, address);
1841         pte_unmap(page_table);
1842         spin_unlock(&mm->page_table_lock);
1843
1844         if (vma->vm_file) {
1845                 mapping = vma->vm_file->f_mapping;
1846                 sequence = mapping->truncate_count;
1847                 smp_rmb(); /* serializes i_size against truncate_count */
1848         }
1849 retry:
1850         cond_resched();
1851         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1852         /*
1853          * No smp_rmb is needed here as long as there's a full
1854          * spin_lock/unlock sequence inside the ->nopage callback
1855          * (for the pagecache lookup) that acts as an implicit
1856          * smp_mb() and prevents the i_size read to happen
1857          * after the next truncate_count read.
1858          */
1859
1860         /* no page was available -- either SIGBUS or OOM */
1861         if (new_page == NOPAGE_SIGBUS)
1862                 return VM_FAULT_SIGBUS;
1863         if (new_page == NOPAGE_OOM)
1864                 return VM_FAULT_OOM;
1865
1866         /*
1867          * Should we do an early C-O-W break?
1868          */
1869         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1870                 struct page *page;
1871
1872                 if (unlikely(anon_vma_prepare(vma)))
1873                         goto oom;
1874                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1875                 if (!page)
1876                         goto oom;
1877                 copy_user_highpage(page, new_page, address);
1878                 page_cache_release(new_page);
1879                 new_page = page;
1880                 anon = 1;
1881         }
1882
1883         spin_lock(&mm->page_table_lock);
1884         /*
1885          * For a file-backed vma, someone could have truncated or otherwise
1886          * invalidated this page.  If unmap_mapping_range got called,
1887          * retry getting the page.
1888          */
1889         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1890                 sequence = mapping->truncate_count;
1891                 spin_unlock(&mm->page_table_lock);
1892                 page_cache_release(new_page);
1893                 goto retry;
1894         }
1895         page_table = pte_offset_map(pmd, address);
1896
1897         /*
1898          * This silly early PAGE_DIRTY setting removes a race
1899          * due to the bad i386 page protection. But it's valid
1900          * for other architectures too.
1901          *
1902          * Note that if write_access is true, we either now have
1903          * an exclusive copy of the page, or this is a shared mapping,
1904          * so we can make it writable and dirty to avoid having to
1905          * handle that later.
1906          */
1907         /* Only go through if we didn't race with anybody else... */
1908         if (pte_none(*page_table)) {
1909                 if (!PageReserved(new_page))
1910                         inc_mm_counter(mm, rss);
1911
1912                 flush_icache_page(vma, new_page);
1913                 entry = mk_pte(new_page, vma->vm_page_prot);
1914                 if (write_access)
1915                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1916                 set_pte_at(mm, address, page_table, entry);
1917                 if (anon) {
1918                         lru_cache_add_active(new_page);
1919                         page_add_anon_rmap(new_page, vma, address);
1920                 } else
1921                         page_add_file_rmap(new_page);
1922                 pte_unmap(page_table);
1923         } else {
1924                 /* One of our sibling threads was faster, back out. */
1925                 pte_unmap(page_table);
1926                 page_cache_release(new_page);
1927                 spin_unlock(&mm->page_table_lock);
1928                 goto out;
1929         }
1930
1931         /* no need to invalidate: a not-present page shouldn't be cached */
1932         update_mmu_cache(vma, address, entry);
1933         lazy_mmu_prot_update(entry);
1934         spin_unlock(&mm->page_table_lock);
1935 out:
1936         return ret;
1937 oom:
1938         page_cache_release(new_page);
1939         ret = VM_FAULT_OOM;
1940         goto out;
1941 }
1942
1943 /*
1944  * Fault of a previously existing named mapping. Repopulate the pte
1945  * from the encoded file_pte if possible. This enables swappable
1946  * nonlinear vmas.
1947  */
1948 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1949         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1950 {
1951         unsigned long pgoff;
1952         int err;
1953
1954         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1955         /*
1956          * Fall back to the linear mapping if the fs does not support
1957          * ->populate:
1958          */
1959         if (!vma->vm_ops->populate ||
1960                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1961                 pte_clear(mm, address, pte);
1962                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1963         }
1964
1965         pgoff = pte_to_pgoff(*pte);
1966
1967         pte_unmap(pte);
1968         spin_unlock(&mm->page_table_lock);
1969
1970         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1971         if (err == -ENOMEM)
1972                 return VM_FAULT_OOM;
1973         if (err)
1974                 return VM_FAULT_SIGBUS;
1975         return VM_FAULT_MAJOR;
1976 }
1977
1978 /*
1979  * These routines also need to handle stuff like marking pages dirty
1980  * and/or accessed for architectures that don't do it in hardware (most
1981  * RISC architectures).  The early dirtying is also good on the i386.
1982  *
1983  * There is also a hook called "update_mmu_cache()" that architectures
1984  * with external mmu caches can use to update those (ie the Sparc or
1985  * PowerPC hashed page tables that act as extended TLBs).
1986  *
1987  * Note the "page_table_lock". It is to protect against kswapd removing
1988  * pages from under us. Note that kswapd only ever _removes_ pages, never
1989  * adds them. As such, once we have noticed that the page is not present,
1990  * we can drop the lock early.
1991  *
1992  * The adding of pages is protected by the MM semaphore (which we hold),
1993  * so we don't need to worry about a page being suddenly been added into
1994  * our VM.
1995  *
1996  * We enter with the pagetable spinlock held, we are supposed to
1997  * release it when done.
1998  */
1999 static inline int handle_pte_fault(struct mm_struct *mm,
2000         struct vm_area_struct * vma, unsigned long address,
2001         int write_access, pte_t *pte, pmd_t *pmd)
2002 {
2003         pte_t entry;
2004
2005         entry = *pte;
2006         if (!pte_present(entry)) {
2007                 /*
2008                  * If it truly wasn't present, we know that kswapd
2009                  * and the PTE updates will not touch it later. So
2010                  * drop the lock.
2011                  */
2012                 if (pte_none(entry))
2013                         return do_no_page(mm, vma, address, write_access, pte, pmd);
2014                 if (pte_file(entry))
2015                         return do_file_page(mm, vma, address, write_access, pte, pmd);
2016                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
2017         }
2018
2019         if (write_access) {
2020                 if (!pte_write(entry))
2021                         return do_wp_page(mm, vma, address, pte, pmd, entry);
2022                 entry = pte_mkdirty(entry);
2023         }
2024         entry = pte_mkyoung(entry);
2025         ptep_set_access_flags(vma, address, pte, entry, write_access);
2026         update_mmu_cache(vma, address, entry);
2027         lazy_mmu_prot_update(entry);
2028         pte_unmap(pte);
2029         spin_unlock(&mm->page_table_lock);
2030         return VM_FAULT_MINOR;
2031 }
2032
2033 /*
2034  * By the time we get here, we already hold the mm semaphore
2035  */
2036 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
2037                 unsigned long address, int write_access)
2038 {
2039         pgd_t *pgd;
2040         pud_t *pud;
2041         pmd_t *pmd;
2042         pte_t *pte;
2043
2044         __set_current_state(TASK_RUNNING);
2045
2046         inc_page_state(pgfault);
2047
2048         if (unlikely(is_vm_hugetlb_page(vma)))
2049                 return hugetlb_fault(mm, vma, address, write_access);
2050
2051         /*
2052          * We need the page table lock to synchronize with kswapd
2053          * and the SMP-safe atomic PTE updates.
2054          */
2055         pgd = pgd_offset(mm, address);
2056         spin_lock(&mm->page_table_lock);
2057
2058         pud = pud_alloc(mm, pgd, address);
2059         if (!pud)
2060                 goto oom;
2061
2062         pmd = pmd_alloc(mm, pud, address);
2063         if (!pmd)
2064                 goto oom;
2065
2066         pte = pte_alloc_map(mm, pmd, address);
2067         if (!pte)
2068                 goto oom;
2069         
2070         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
2071
2072  oom:
2073         spin_unlock(&mm->page_table_lock);
2074         return VM_FAULT_OOM;
2075 }
2076
2077 #ifndef __PAGETABLE_PUD_FOLDED
2078 /*
2079  * Allocate page upper directory.
2080  *
2081  * We've already handled the fast-path in-line, and we own the
2082  * page table lock.
2083  */
2084 pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2085 {
2086         pud_t *new;
2087
2088         spin_unlock(&mm->page_table_lock);
2089         new = pud_alloc_one(mm, address);
2090         spin_lock(&mm->page_table_lock);
2091         if (!new)
2092                 return NULL;
2093
2094         /*
2095          * Because we dropped the lock, we should re-check the
2096          * entry, as somebody else could have populated it..
2097          */
2098         if (pgd_present(*pgd)) {
2099                 pud_free(new);
2100                 goto out;
2101         }
2102         pgd_populate(mm, pgd, new);
2103  out:
2104         return pud_offset(pgd, address);
2105 }
2106 #endif /* __PAGETABLE_PUD_FOLDED */
2107
2108 #ifndef __PAGETABLE_PMD_FOLDED
2109 /*
2110  * Allocate page middle directory.
2111  *
2112  * We've already handled the fast-path in-line, and we own the
2113  * page table lock.
2114  */
2115 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2116 {
2117         pmd_t *new;
2118
2119         spin_unlock(&mm->page_table_lock);
2120         new = pmd_alloc_one(mm, address);
2121         spin_lock(&mm->page_table_lock);
2122         if (!new)
2123                 return NULL;
2124
2125         /*
2126          * Because we dropped the lock, we should re-check the
2127          * entry, as somebody else could have populated it..
2128          */
2129 #ifndef __ARCH_HAS_4LEVEL_HACK
2130         if (pud_present(*pud)) {
2131                 pmd_free(new);
2132                 goto out;
2133         }
2134         pud_populate(mm, pud, new);
2135 #else
2136         if (pgd_present(*pud)) {
2137                 pmd_free(new);
2138                 goto out;
2139         }
2140         pgd_populate(mm, pud, new);
2141 #endif /* __ARCH_HAS_4LEVEL_HACK */
2142
2143  out:
2144         return pmd_offset(pud, address);
2145 }
2146 #endif /* __PAGETABLE_PMD_FOLDED */
2147
2148 int make_pages_present(unsigned long addr, unsigned long end)
2149 {
2150         int ret, len, write;
2151         struct vm_area_struct * vma;
2152
2153         vma = find_vma(current->mm, addr);
2154         if (!vma)
2155                 return -1;
2156         write = (vma->vm_flags & VM_WRITE) != 0;
2157         if (addr >= end)
2158                 BUG();
2159         if (end > vma->vm_end)
2160                 BUG();
2161         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2162         ret = get_user_pages(current, current->mm, addr,
2163                         len, write, 0, NULL, NULL);
2164         if (ret < 0)
2165                 return ret;
2166         return ret == len ? 0 : -1;
2167 }
2168
2169 /* 
2170  * Map a vmalloc()-space virtual address to the physical page.
2171  */
2172 struct page * vmalloc_to_page(void * vmalloc_addr)
2173 {
2174         unsigned long addr = (unsigned long) vmalloc_addr;
2175         struct page *page = NULL;
2176         pgd_t *pgd = pgd_offset_k(addr);
2177         pud_t *pud;
2178         pmd_t *pmd;
2179         pte_t *ptep, pte;
2180   
2181         if (!pgd_none(*pgd)) {
2182                 pud = pud_offset(pgd, addr);
2183                 if (!pud_none(*pud)) {
2184                         pmd = pmd_offset(pud, addr);
2185                         if (!pmd_none(*pmd)) {
2186                                 ptep = pte_offset_map(pmd, addr);
2187                                 pte = *ptep;
2188                                 if (pte_present(pte))
2189                                         page = pte_page(pte);
2190                                 pte_unmap(ptep);
2191                         }
2192                 }
2193         }
2194         return page;
2195 }
2196
2197 EXPORT_SYMBOL(vmalloc_to_page);
2198
2199 /*
2200  * Map a vmalloc()-space virtual address to the physical page frame number.
2201  */
2202 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2203 {
2204         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2205 }
2206
2207 EXPORT_SYMBOL(vmalloc_to_pfn);
2208
2209 /*
2210  * update_mem_hiwater
2211  *      - update per process rss and vm high water data
2212  */
2213 void update_mem_hiwater(struct task_struct *tsk)
2214 {
2215         if (tsk->mm) {
2216                 unsigned long rss = get_mm_counter(tsk->mm, rss);
2217
2218                 if (tsk->mm->hiwater_rss < rss)
2219                         tsk->mm->hiwater_rss = rss;
2220                 if (tsk->mm->hiwater_vm < tsk->mm->total_vm)
2221                         tsk->mm->hiwater_vm = tsk->mm->total_vm;
2222         }
2223 }
2224
2225 #if !defined(__HAVE_ARCH_GATE_AREA)
2226
2227 #if defined(AT_SYSINFO_EHDR)
2228 static struct vm_area_struct gate_vma;
2229
2230 static int __init gate_vma_init(void)
2231 {
2232         gate_vma.vm_mm = NULL;
2233         gate_vma.vm_start = FIXADDR_USER_START;
2234         gate_vma.vm_end = FIXADDR_USER_END;
2235         gate_vma.vm_page_prot = PAGE_READONLY;
2236         gate_vma.vm_flags = 0;
2237         return 0;
2238 }
2239 __initcall(gate_vma_init);
2240 #endif
2241
2242 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2243 {
2244 #ifdef AT_SYSINFO_EHDR
2245         return &gate_vma;
2246 #else
2247         return NULL;
2248 #endif
2249 }
2250
2251 int in_gate_area_no_task(unsigned long addr)
2252 {
2253 #ifdef AT_SYSINFO_EHDR
2254         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2255                 return 1;
2256 #endif
2257         return 0;
2258 }
2259
2260 #endif  /* __HAVE_ARCH_GATE_AREA */