Merge /spare/repo/linux-2.6/
[linux-2.6] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
54 #include <asm/tlb.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
57
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
65
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
69
70 unsigned long num_physpages;
71 /*
72  * A number of key systems in x86 including ioremap() rely on the assumption
73  * that high_memory defines the upper bound on direct map memory, then end
74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76  * and ZONE_HIGHMEM.
77  */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
80
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85 /*
86  * If a p?d_bad entry is found while walking page tables, report
87  * the error, before resetting entry to p?d_none.  Usually (but
88  * very seldom) called out from the p?d_none_or_clear_bad macros.
89  */
90
91 void pgd_clear_bad(pgd_t *pgd)
92 {
93         pgd_ERROR(*pgd);
94         pgd_clear(pgd);
95 }
96
97 void pud_clear_bad(pud_t *pud)
98 {
99         pud_ERROR(*pud);
100         pud_clear(pud);
101 }
102
103 void pmd_clear_bad(pmd_t *pmd)
104 {
105         pmd_ERROR(*pmd);
106         pmd_clear(pmd);
107 }
108
109 /*
110  * Note: this doesn't free the actual pages themselves. That
111  * has been handled earlier when unmapping all the memory regions.
112  */
113 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
114 {
115         struct page *page = pmd_page(*pmd);
116         pmd_clear(pmd);
117         pte_free_tlb(tlb, page);
118         dec_page_state(nr_page_table_pages);
119         tlb->mm->nr_ptes--;
120 }
121
122 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
123                                 unsigned long addr, unsigned long end,
124                                 unsigned long floor, unsigned long ceiling)
125 {
126         pmd_t *pmd;
127         unsigned long next;
128         unsigned long start;
129
130         start = addr;
131         pmd = pmd_offset(pud, addr);
132         do {
133                 next = pmd_addr_end(addr, end);
134                 if (pmd_none_or_clear_bad(pmd))
135                         continue;
136                 free_pte_range(tlb, pmd);
137         } while (pmd++, addr = next, addr != end);
138
139         start &= PUD_MASK;
140         if (start < floor)
141                 return;
142         if (ceiling) {
143                 ceiling &= PUD_MASK;
144                 if (!ceiling)
145                         return;
146         }
147         if (end - 1 > ceiling - 1)
148                 return;
149
150         pmd = pmd_offset(pud, start);
151         pud_clear(pud);
152         pmd_free_tlb(tlb, pmd);
153 }
154
155 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
156                                 unsigned long addr, unsigned long end,
157                                 unsigned long floor, unsigned long ceiling)
158 {
159         pud_t *pud;
160         unsigned long next;
161         unsigned long start;
162
163         start = addr;
164         pud = pud_offset(pgd, addr);
165         do {
166                 next = pud_addr_end(addr, end);
167                 if (pud_none_or_clear_bad(pud))
168                         continue;
169                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
170         } while (pud++, addr = next, addr != end);
171
172         start &= PGDIR_MASK;
173         if (start < floor)
174                 return;
175         if (ceiling) {
176                 ceiling &= PGDIR_MASK;
177                 if (!ceiling)
178                         return;
179         }
180         if (end - 1 > ceiling - 1)
181                 return;
182
183         pud = pud_offset(pgd, start);
184         pgd_clear(pgd);
185         pud_free_tlb(tlb, pud);
186 }
187
188 /*
189  * This function frees user-level page tables of a process.
190  *
191  * Must be called with pagetable lock held.
192  */
193 void free_pgd_range(struct mmu_gather **tlb,
194                         unsigned long addr, unsigned long end,
195                         unsigned long floor, unsigned long ceiling)
196 {
197         pgd_t *pgd;
198         unsigned long next;
199         unsigned long start;
200
201         /*
202          * The next few lines have given us lots of grief...
203          *
204          * Why are we testing PMD* at this top level?  Because often
205          * there will be no work to do at all, and we'd prefer not to
206          * go all the way down to the bottom just to discover that.
207          *
208          * Why all these "- 1"s?  Because 0 represents both the bottom
209          * of the address space and the top of it (using -1 for the
210          * top wouldn't help much: the masks would do the wrong thing).
211          * The rule is that addr 0 and floor 0 refer to the bottom of
212          * the address space, but end 0 and ceiling 0 refer to the top
213          * Comparisons need to use "end - 1" and "ceiling - 1" (though
214          * that end 0 case should be mythical).
215          *
216          * Wherever addr is brought up or ceiling brought down, we must
217          * be careful to reject "the opposite 0" before it confuses the
218          * subsequent tests.  But what about where end is brought down
219          * by PMD_SIZE below? no, end can't go down to 0 there.
220          *
221          * Whereas we round start (addr) and ceiling down, by different
222          * masks at different levels, in order to test whether a table
223          * now has no other vmas using it, so can be freed, we don't
224          * bother to round floor or end up - the tests don't need that.
225          */
226
227         addr &= PMD_MASK;
228         if (addr < floor) {
229                 addr += PMD_SIZE;
230                 if (!addr)
231                         return;
232         }
233         if (ceiling) {
234                 ceiling &= PMD_MASK;
235                 if (!ceiling)
236                         return;
237         }
238         if (end - 1 > ceiling - 1)
239                 end -= PMD_SIZE;
240         if (addr > end - 1)
241                 return;
242
243         start = addr;
244         pgd = pgd_offset((*tlb)->mm, addr);
245         do {
246                 next = pgd_addr_end(addr, end);
247                 if (pgd_none_or_clear_bad(pgd))
248                         continue;
249                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
250         } while (pgd++, addr = next, addr != end);
251
252         if (!tlb_is_full_mm(*tlb))
253                 flush_tlb_pgtables((*tlb)->mm, start, end);
254 }
255
256 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
257                 unsigned long floor, unsigned long ceiling)
258 {
259         while (vma) {
260                 struct vm_area_struct *next = vma->vm_next;
261                 unsigned long addr = vma->vm_start;
262
263                 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
264                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
265                                 floor, next? next->vm_start: ceiling);
266                 } else {
267                         /*
268                          * Optimization: gather nearby vmas into one call down
269                          */
270                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
271                           && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
272                                                         HPAGE_SIZE)) {
273                                 vma = next;
274                                 next = vma->vm_next;
275                         }
276                         free_pgd_range(tlb, addr, vma->vm_end,
277                                 floor, next? next->vm_start: ceiling);
278                 }
279                 vma = next;
280         }
281 }
282
283 pte_t fastcall *pte_alloc_map(struct mm_struct *mm, pmd_t *pmd,
284                                 unsigned long address)
285 {
286         if (!pmd_present(*pmd)) {
287                 struct page *new;
288
289                 spin_unlock(&mm->page_table_lock);
290                 new = pte_alloc_one(mm, address);
291                 spin_lock(&mm->page_table_lock);
292                 if (!new)
293                         return NULL;
294                 /*
295                  * Because we dropped the lock, we should re-check the
296                  * entry, as somebody else could have populated it..
297                  */
298                 if (pmd_present(*pmd)) {
299                         pte_free(new);
300                         goto out;
301                 }
302                 mm->nr_ptes++;
303                 inc_page_state(nr_page_table_pages);
304                 pmd_populate(mm, pmd, new);
305         }
306 out:
307         return pte_offset_map(pmd, address);
308 }
309
310 pte_t fastcall * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
311 {
312         if (!pmd_present(*pmd)) {
313                 pte_t *new;
314
315                 spin_unlock(&mm->page_table_lock);
316                 new = pte_alloc_one_kernel(mm, address);
317                 spin_lock(&mm->page_table_lock);
318                 if (!new)
319                         return NULL;
320
321                 /*
322                  * Because we dropped the lock, we should re-check the
323                  * entry, as somebody else could have populated it..
324                  */
325                 if (pmd_present(*pmd)) {
326                         pte_free_kernel(new);
327                         goto out;
328                 }
329                 pmd_populate_kernel(mm, pmd, new);
330         }
331 out:
332         return pte_offset_kernel(pmd, address);
333 }
334
335 /*
336  * copy one vm_area from one task to the other. Assumes the page tables
337  * already present in the new task to be cleared in the whole range
338  * covered by this vma.
339  *
340  * dst->page_table_lock is held on entry and exit,
341  * but may be dropped within p[mg]d_alloc() and pte_alloc_map().
342  */
343
344 static inline void
345 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
346                 pte_t *dst_pte, pte_t *src_pte, unsigned long vm_flags,
347                 unsigned long addr)
348 {
349         pte_t pte = *src_pte;
350         struct page *page;
351         unsigned long pfn;
352
353         /* pte contains position in swap or file, so copy. */
354         if (unlikely(!pte_present(pte))) {
355                 if (!pte_file(pte)) {
356                         swap_duplicate(pte_to_swp_entry(pte));
357                         /* make sure dst_mm is on swapoff's mmlist. */
358                         if (unlikely(list_empty(&dst_mm->mmlist))) {
359                                 spin_lock(&mmlist_lock);
360                                 list_add(&dst_mm->mmlist, &src_mm->mmlist);
361                                 spin_unlock(&mmlist_lock);
362                         }
363                 }
364                 set_pte_at(dst_mm, addr, dst_pte, pte);
365                 return;
366         }
367
368         pfn = pte_pfn(pte);
369         /* the pte points outside of valid memory, the
370          * mapping is assumed to be good, meaningful
371          * and not mapped via rmap - duplicate the
372          * mapping as is.
373          */
374         page = NULL;
375         if (pfn_valid(pfn))
376                 page = pfn_to_page(pfn);
377
378         if (!page || PageReserved(page)) {
379                 set_pte_at(dst_mm, addr, dst_pte, pte);
380                 return;
381         }
382
383         /*
384          * If it's a COW mapping, write protect it both
385          * in the parent and the child
386          */
387         if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
388                 ptep_set_wrprotect(src_mm, addr, src_pte);
389                 pte = *src_pte;
390         }
391
392         /*
393          * If it's a shared mapping, mark it clean in
394          * the child
395          */
396         if (vm_flags & VM_SHARED)
397                 pte = pte_mkclean(pte);
398         pte = pte_mkold(pte);
399         get_page(page);
400         inc_mm_counter(dst_mm, rss);
401         if (PageAnon(page))
402                 inc_mm_counter(dst_mm, anon_rss);
403         set_pte_at(dst_mm, addr, dst_pte, pte);
404         page_dup_rmap(page);
405 }
406
407 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
408                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
409                 unsigned long addr, unsigned long end)
410 {
411         pte_t *src_pte, *dst_pte;
412         unsigned long vm_flags = vma->vm_flags;
413         int progress;
414
415 again:
416         dst_pte = pte_alloc_map(dst_mm, dst_pmd, addr);
417         if (!dst_pte)
418                 return -ENOMEM;
419         src_pte = pte_offset_map_nested(src_pmd, addr);
420
421         progress = 0;
422         spin_lock(&src_mm->page_table_lock);
423         do {
424                 /*
425                  * We are holding two locks at this point - either of them
426                  * could generate latencies in another task on another CPU.
427                  */
428                 if (progress >= 32 && (need_resched() ||
429                     need_lockbreak(&src_mm->page_table_lock) ||
430                     need_lockbreak(&dst_mm->page_table_lock)))
431                         break;
432                 if (pte_none(*src_pte)) {
433                         progress++;
434                         continue;
435                 }
436                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vm_flags, addr);
437                 progress += 8;
438         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
439         spin_unlock(&src_mm->page_table_lock);
440
441         pte_unmap_nested(src_pte - 1);
442         pte_unmap(dst_pte - 1);
443         cond_resched_lock(&dst_mm->page_table_lock);
444         if (addr != end)
445                 goto again;
446         return 0;
447 }
448
449 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
450                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
451                 unsigned long addr, unsigned long end)
452 {
453         pmd_t *src_pmd, *dst_pmd;
454         unsigned long next;
455
456         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
457         if (!dst_pmd)
458                 return -ENOMEM;
459         src_pmd = pmd_offset(src_pud, addr);
460         do {
461                 next = pmd_addr_end(addr, end);
462                 if (pmd_none_or_clear_bad(src_pmd))
463                         continue;
464                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
465                                                 vma, addr, next))
466                         return -ENOMEM;
467         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
468         return 0;
469 }
470
471 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
472                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
473                 unsigned long addr, unsigned long end)
474 {
475         pud_t *src_pud, *dst_pud;
476         unsigned long next;
477
478         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
479         if (!dst_pud)
480                 return -ENOMEM;
481         src_pud = pud_offset(src_pgd, addr);
482         do {
483                 next = pud_addr_end(addr, end);
484                 if (pud_none_or_clear_bad(src_pud))
485                         continue;
486                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
487                                                 vma, addr, next))
488                         return -ENOMEM;
489         } while (dst_pud++, src_pud++, addr = next, addr != end);
490         return 0;
491 }
492
493 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
494                 struct vm_area_struct *vma)
495 {
496         pgd_t *src_pgd, *dst_pgd;
497         unsigned long next;
498         unsigned long addr = vma->vm_start;
499         unsigned long end = vma->vm_end;
500
501         if (is_vm_hugetlb_page(vma))
502                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
503
504         dst_pgd = pgd_offset(dst_mm, addr);
505         src_pgd = pgd_offset(src_mm, addr);
506         do {
507                 next = pgd_addr_end(addr, end);
508                 if (pgd_none_or_clear_bad(src_pgd))
509                         continue;
510                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
511                                                 vma, addr, next))
512                         return -ENOMEM;
513         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
514         return 0;
515 }
516
517 static void zap_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
518                                 unsigned long addr, unsigned long end,
519                                 struct zap_details *details)
520 {
521         pte_t *pte;
522
523         pte = pte_offset_map(pmd, addr);
524         do {
525                 pte_t ptent = *pte;
526                 if (pte_none(ptent))
527                         continue;
528                 if (pte_present(ptent)) {
529                         struct page *page = NULL;
530                         unsigned long pfn = pte_pfn(ptent);
531                         if (pfn_valid(pfn)) {
532                                 page = pfn_to_page(pfn);
533                                 if (PageReserved(page))
534                                         page = NULL;
535                         }
536                         if (unlikely(details) && page) {
537                                 /*
538                                  * unmap_shared_mapping_pages() wants to
539                                  * invalidate cache without truncating:
540                                  * unmap shared but keep private pages.
541                                  */
542                                 if (details->check_mapping &&
543                                     details->check_mapping != page->mapping)
544                                         continue;
545                                 /*
546                                  * Each page->index must be checked when
547                                  * invalidating or truncating nonlinear.
548                                  */
549                                 if (details->nonlinear_vma &&
550                                     (page->index < details->first_index ||
551                                      page->index > details->last_index))
552                                         continue;
553                         }
554                         ptent = ptep_get_and_clear(tlb->mm, addr, pte);
555                         tlb_remove_tlb_entry(tlb, pte, addr);
556                         if (unlikely(!page))
557                                 continue;
558                         if (unlikely(details) && details->nonlinear_vma
559                             && linear_page_index(details->nonlinear_vma,
560                                                 addr) != page->index)
561                                 set_pte_at(tlb->mm, addr, pte,
562                                            pgoff_to_pte(page->index));
563                         if (pte_dirty(ptent))
564                                 set_page_dirty(page);
565                         if (PageAnon(page))
566                                 dec_mm_counter(tlb->mm, anon_rss);
567                         else if (pte_young(ptent))
568                                 mark_page_accessed(page);
569                         tlb->freed++;
570                         page_remove_rmap(page);
571                         tlb_remove_page(tlb, page);
572                         continue;
573                 }
574                 /*
575                  * If details->check_mapping, we leave swap entries;
576                  * if details->nonlinear_vma, we leave file entries.
577                  */
578                 if (unlikely(details))
579                         continue;
580                 if (!pte_file(ptent))
581                         free_swap_and_cache(pte_to_swp_entry(ptent));
582                 pte_clear(tlb->mm, addr, pte);
583         } while (pte++, addr += PAGE_SIZE, addr != end);
584         pte_unmap(pte - 1);
585 }
586
587 static inline void zap_pmd_range(struct mmu_gather *tlb, pud_t *pud,
588                                 unsigned long addr, unsigned long end,
589                                 struct zap_details *details)
590 {
591         pmd_t *pmd;
592         unsigned long next;
593
594         pmd = pmd_offset(pud, addr);
595         do {
596                 next = pmd_addr_end(addr, end);
597                 if (pmd_none_or_clear_bad(pmd))
598                         continue;
599                 zap_pte_range(tlb, pmd, addr, next, details);
600         } while (pmd++, addr = next, addr != end);
601 }
602
603 static inline void zap_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
604                                 unsigned long addr, unsigned long end,
605                                 struct zap_details *details)
606 {
607         pud_t *pud;
608         unsigned long next;
609
610         pud = pud_offset(pgd, addr);
611         do {
612                 next = pud_addr_end(addr, end);
613                 if (pud_none_or_clear_bad(pud))
614                         continue;
615                 zap_pmd_range(tlb, pud, addr, next, details);
616         } while (pud++, addr = next, addr != end);
617 }
618
619 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
620                                 unsigned long addr, unsigned long end,
621                                 struct zap_details *details)
622 {
623         pgd_t *pgd;
624         unsigned long next;
625
626         if (details && !details->check_mapping && !details->nonlinear_vma)
627                 details = NULL;
628
629         BUG_ON(addr >= end);
630         tlb_start_vma(tlb, vma);
631         pgd = pgd_offset(vma->vm_mm, addr);
632         do {
633                 next = pgd_addr_end(addr, end);
634                 if (pgd_none_or_clear_bad(pgd))
635                         continue;
636                 zap_pud_range(tlb, pgd, addr, next, details);
637         } while (pgd++, addr = next, addr != end);
638         tlb_end_vma(tlb, vma);
639 }
640
641 #ifdef CONFIG_PREEMPT
642 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
643 #else
644 /* No preempt: go for improved straight-line efficiency */
645 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
646 #endif
647
648 /**
649  * unmap_vmas - unmap a range of memory covered by a list of vma's
650  * @tlbp: address of the caller's struct mmu_gather
651  * @mm: the controlling mm_struct
652  * @vma: the starting vma
653  * @start_addr: virtual address at which to start unmapping
654  * @end_addr: virtual address at which to end unmapping
655  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
656  * @details: details of nonlinear truncation or shared cache invalidation
657  *
658  * Returns the end address of the unmapping (restart addr if interrupted).
659  *
660  * Unmap all pages in the vma list.  Called under page_table_lock.
661  *
662  * We aim to not hold page_table_lock for too long (for scheduling latency
663  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
664  * return the ending mmu_gather to the caller.
665  *
666  * Only addresses between `start' and `end' will be unmapped.
667  *
668  * The VMA list must be sorted in ascending virtual address order.
669  *
670  * unmap_vmas() assumes that the caller will flush the whole unmapped address
671  * range after unmap_vmas() returns.  So the only responsibility here is to
672  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
673  * drops the lock and schedules.
674  */
675 unsigned long unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
676                 struct vm_area_struct *vma, unsigned long start_addr,
677                 unsigned long end_addr, unsigned long *nr_accounted,
678                 struct zap_details *details)
679 {
680         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
681         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
682         int tlb_start_valid = 0;
683         unsigned long start = start_addr;
684         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
685         int fullmm = tlb_is_full_mm(*tlbp);
686
687         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
688                 unsigned long end;
689
690                 start = max(vma->vm_start, start_addr);
691                 if (start >= vma->vm_end)
692                         continue;
693                 end = min(vma->vm_end, end_addr);
694                 if (end <= vma->vm_start)
695                         continue;
696
697                 if (vma->vm_flags & VM_ACCOUNT)
698                         *nr_accounted += (end - start) >> PAGE_SHIFT;
699
700                 while (start != end) {
701                         unsigned long block;
702
703                         if (!tlb_start_valid) {
704                                 tlb_start = start;
705                                 tlb_start_valid = 1;
706                         }
707
708                         if (is_vm_hugetlb_page(vma)) {
709                                 block = end - start;
710                                 unmap_hugepage_range(vma, start, end);
711                         } else {
712                                 block = min(zap_bytes, end - start);
713                                 unmap_page_range(*tlbp, vma, start,
714                                                 start + block, details);
715                         }
716
717                         start += block;
718                         zap_bytes -= block;
719                         if ((long)zap_bytes > 0)
720                                 continue;
721
722                         tlb_finish_mmu(*tlbp, tlb_start, start);
723
724                         if (need_resched() ||
725                                 need_lockbreak(&mm->page_table_lock) ||
726                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
727                                 if (i_mmap_lock) {
728                                         /* must reset count of rss freed */
729                                         *tlbp = tlb_gather_mmu(mm, fullmm);
730                                         goto out;
731                                 }
732                                 spin_unlock(&mm->page_table_lock);
733                                 cond_resched();
734                                 spin_lock(&mm->page_table_lock);
735                         }
736
737                         *tlbp = tlb_gather_mmu(mm, fullmm);
738                         tlb_start_valid = 0;
739                         zap_bytes = ZAP_BLOCK_SIZE;
740                 }
741         }
742 out:
743         return start;   /* which is now the end (or restart) address */
744 }
745
746 /**
747  * zap_page_range - remove user pages in a given range
748  * @vma: vm_area_struct holding the applicable pages
749  * @address: starting address of pages to zap
750  * @size: number of bytes to zap
751  * @details: details of nonlinear truncation or shared cache invalidation
752  */
753 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
754                 unsigned long size, struct zap_details *details)
755 {
756         struct mm_struct *mm = vma->vm_mm;
757         struct mmu_gather *tlb;
758         unsigned long end = address + size;
759         unsigned long nr_accounted = 0;
760
761         if (is_vm_hugetlb_page(vma)) {
762                 zap_hugepage_range(vma, address, size);
763                 return end;
764         }
765
766         lru_add_drain();
767         spin_lock(&mm->page_table_lock);
768         tlb = tlb_gather_mmu(mm, 0);
769         end = unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
770         tlb_finish_mmu(tlb, address, end);
771         spin_unlock(&mm->page_table_lock);
772         return end;
773 }
774
775 /*
776  * Do a quick page-table lookup for a single page.
777  * mm->page_table_lock must be held.
778  */
779 static struct page *__follow_page(struct mm_struct *mm, unsigned long address,
780                         int read, int write, int accessed)
781 {
782         pgd_t *pgd;
783         pud_t *pud;
784         pmd_t *pmd;
785         pte_t *ptep, pte;
786         unsigned long pfn;
787         struct page *page;
788
789         page = follow_huge_addr(mm, address, write);
790         if (! IS_ERR(page))
791                 return page;
792
793         pgd = pgd_offset(mm, address);
794         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
795                 goto out;
796
797         pud = pud_offset(pgd, address);
798         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
799                 goto out;
800         
801         pmd = pmd_offset(pud, address);
802         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
803                 goto out;
804         if (pmd_huge(*pmd))
805                 return follow_huge_pmd(mm, address, pmd, write);
806
807         ptep = pte_offset_map(pmd, address);
808         if (!ptep)
809                 goto out;
810
811         pte = *ptep;
812         pte_unmap(ptep);
813         if (pte_present(pte)) {
814                 if (write && !pte_write(pte))
815                         goto out;
816                 if (read && !pte_read(pte))
817                         goto out;
818                 pfn = pte_pfn(pte);
819                 if (pfn_valid(pfn)) {
820                         page = pfn_to_page(pfn);
821                         if (accessed) {
822                                 if (write && !pte_dirty(pte) &&!PageDirty(page))
823                                         set_page_dirty(page);
824                                 mark_page_accessed(page);
825                         }
826                         return page;
827                 }
828         }
829
830 out:
831         return NULL;
832 }
833
834 inline struct page *
835 follow_page(struct mm_struct *mm, unsigned long address, int write)
836 {
837         return __follow_page(mm, address, 0, write, 1);
838 }
839
840 /*
841  * check_user_page_readable() can be called frm niterrupt context by oprofile,
842  * so we need to avoid taking any non-irq-safe locks
843  */
844 int check_user_page_readable(struct mm_struct *mm, unsigned long address)
845 {
846         return __follow_page(mm, address, 1, 0, 0) != NULL;
847 }
848 EXPORT_SYMBOL(check_user_page_readable);
849
850 static inline int
851 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
852                          unsigned long address)
853 {
854         pgd_t *pgd;
855         pud_t *pud;
856         pmd_t *pmd;
857
858         /* Check if the vma is for an anonymous mapping. */
859         if (vma->vm_ops && vma->vm_ops->nopage)
860                 return 0;
861
862         /* Check if page directory entry exists. */
863         pgd = pgd_offset(mm, address);
864         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
865                 return 1;
866
867         pud = pud_offset(pgd, address);
868         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
869                 return 1;
870
871         /* Check if page middle directory entry exists. */
872         pmd = pmd_offset(pud, address);
873         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
874                 return 1;
875
876         /* There is a pte slot for 'address' in 'mm'. */
877         return 0;
878 }
879
880 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
881                 unsigned long start, int len, int write, int force,
882                 struct page **pages, struct vm_area_struct **vmas)
883 {
884         int i;
885         unsigned int flags;
886
887         /* 
888          * Require read or write permissions.
889          * If 'force' is set, we only require the "MAY" flags.
890          */
891         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
892         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
893         i = 0;
894
895         do {
896                 struct vm_area_struct * vma;
897
898                 vma = find_extend_vma(mm, start);
899                 if (!vma && in_gate_area(tsk, start)) {
900                         unsigned long pg = start & PAGE_MASK;
901                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
902                         pgd_t *pgd;
903                         pud_t *pud;
904                         pmd_t *pmd;
905                         pte_t *pte;
906                         if (write) /* user gate pages are read-only */
907                                 return i ? : -EFAULT;
908                         if (pg > TASK_SIZE)
909                                 pgd = pgd_offset_k(pg);
910                         else
911                                 pgd = pgd_offset_gate(mm, pg);
912                         BUG_ON(pgd_none(*pgd));
913                         pud = pud_offset(pgd, pg);
914                         BUG_ON(pud_none(*pud));
915                         pmd = pmd_offset(pud, pg);
916                         if (pmd_none(*pmd))
917                                 return i ? : -EFAULT;
918                         pte = pte_offset_map(pmd, pg);
919                         if (pte_none(*pte)) {
920                                 pte_unmap(pte);
921                                 return i ? : -EFAULT;
922                         }
923                         if (pages) {
924                                 pages[i] = pte_page(*pte);
925                                 get_page(pages[i]);
926                         }
927                         pte_unmap(pte);
928                         if (vmas)
929                                 vmas[i] = gate_vma;
930                         i++;
931                         start += PAGE_SIZE;
932                         len--;
933                         continue;
934                 }
935
936                 if (!vma || (vma->vm_flags & VM_IO)
937                                 || !(flags & vma->vm_flags))
938                         return i ? : -EFAULT;
939
940                 if (is_vm_hugetlb_page(vma)) {
941                         i = follow_hugetlb_page(mm, vma, pages, vmas,
942                                                 &start, &len, i);
943                         continue;
944                 }
945                 spin_lock(&mm->page_table_lock);
946                 do {
947                         int write_access = write;
948                         struct page *page;
949
950                         cond_resched_lock(&mm->page_table_lock);
951                         while (!(page = follow_page(mm, start, write_access))) {
952                                 int ret;
953
954                                 /*
955                                  * Shortcut for anonymous pages. We don't want
956                                  * to force the creation of pages tables for
957                                  * insanely big anonymously mapped areas that
958                                  * nobody touched so far. This is important
959                                  * for doing a core dump for these mappings.
960                                  */
961                                 if (!write && untouched_anonymous_page(mm,vma,start)) {
962                                         page = ZERO_PAGE(start);
963                                         break;
964                                 }
965                                 spin_unlock(&mm->page_table_lock);
966                                 ret = __handle_mm_fault(mm, vma, start, write_access);
967
968                                 /*
969                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
970                                  * broken COW when necessary, even if maybe_mkwrite
971                                  * decided not to set pte_write. We can thus safely do
972                                  * subsequent page lookups as if they were reads.
973                                  */
974                                 if (ret & VM_FAULT_WRITE)
975                                         write_access = 0;
976                                 
977                                 switch (ret & ~VM_FAULT_WRITE) {
978                                 case VM_FAULT_MINOR:
979                                         tsk->min_flt++;
980                                         break;
981                                 case VM_FAULT_MAJOR:
982                                         tsk->maj_flt++;
983                                         break;
984                                 case VM_FAULT_SIGBUS:
985                                         return i ? i : -EFAULT;
986                                 case VM_FAULT_OOM:
987                                         return i ? i : -ENOMEM;
988                                 default:
989                                         BUG();
990                                 }
991                                 spin_lock(&mm->page_table_lock);
992                         }
993                         if (pages) {
994                                 pages[i] = page;
995                                 flush_dcache_page(page);
996                                 if (!PageReserved(page))
997                                         page_cache_get(page);
998                         }
999                         if (vmas)
1000                                 vmas[i] = vma;
1001                         i++;
1002                         start += PAGE_SIZE;
1003                         len--;
1004                 } while (len && start < vma->vm_end);
1005                 spin_unlock(&mm->page_table_lock);
1006         } while (len);
1007         return i;
1008 }
1009 EXPORT_SYMBOL(get_user_pages);
1010
1011 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1012                         unsigned long addr, unsigned long end, pgprot_t prot)
1013 {
1014         pte_t *pte;
1015
1016         pte = pte_alloc_map(mm, pmd, addr);
1017         if (!pte)
1018                 return -ENOMEM;
1019         do {
1020                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(addr), prot));
1021                 BUG_ON(!pte_none(*pte));
1022                 set_pte_at(mm, addr, pte, zero_pte);
1023         } while (pte++, addr += PAGE_SIZE, addr != end);
1024         pte_unmap(pte - 1);
1025         return 0;
1026 }
1027
1028 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1029                         unsigned long addr, unsigned long end, pgprot_t prot)
1030 {
1031         pmd_t *pmd;
1032         unsigned long next;
1033
1034         pmd = pmd_alloc(mm, pud, addr);
1035         if (!pmd)
1036                 return -ENOMEM;
1037         do {
1038                 next = pmd_addr_end(addr, end);
1039                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1040                         return -ENOMEM;
1041         } while (pmd++, addr = next, addr != end);
1042         return 0;
1043 }
1044
1045 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1046                         unsigned long addr, unsigned long end, pgprot_t prot)
1047 {
1048         pud_t *pud;
1049         unsigned long next;
1050
1051         pud = pud_alloc(mm, pgd, addr);
1052         if (!pud)
1053                 return -ENOMEM;
1054         do {
1055                 next = pud_addr_end(addr, end);
1056                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1057                         return -ENOMEM;
1058         } while (pud++, addr = next, addr != end);
1059         return 0;
1060 }
1061
1062 int zeromap_page_range(struct vm_area_struct *vma,
1063                         unsigned long addr, unsigned long size, pgprot_t prot)
1064 {
1065         pgd_t *pgd;
1066         unsigned long next;
1067         unsigned long end = addr + size;
1068         struct mm_struct *mm = vma->vm_mm;
1069         int err;
1070
1071         BUG_ON(addr >= end);
1072         pgd = pgd_offset(mm, addr);
1073         flush_cache_range(vma, addr, end);
1074         spin_lock(&mm->page_table_lock);
1075         do {
1076                 next = pgd_addr_end(addr, end);
1077                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1078                 if (err)
1079                         break;
1080         } while (pgd++, addr = next, addr != end);
1081         spin_unlock(&mm->page_table_lock);
1082         return err;
1083 }
1084
1085 /*
1086  * maps a range of physical memory into the requested pages. the old
1087  * mappings are removed. any references to nonexistent pages results
1088  * in null mappings (currently treated as "copy-on-access")
1089  */
1090 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1091                         unsigned long addr, unsigned long end,
1092                         unsigned long pfn, pgprot_t prot)
1093 {
1094         pte_t *pte;
1095
1096         pte = pte_alloc_map(mm, pmd, addr);
1097         if (!pte)
1098                 return -ENOMEM;
1099         do {
1100                 BUG_ON(!pte_none(*pte));
1101                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
1102                         set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1103                 pfn++;
1104         } while (pte++, addr += PAGE_SIZE, addr != end);
1105         pte_unmap(pte - 1);
1106         return 0;
1107 }
1108
1109 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1110                         unsigned long addr, unsigned long end,
1111                         unsigned long pfn, pgprot_t prot)
1112 {
1113         pmd_t *pmd;
1114         unsigned long next;
1115
1116         pfn -= addr >> PAGE_SHIFT;
1117         pmd = pmd_alloc(mm, pud, addr);
1118         if (!pmd)
1119                 return -ENOMEM;
1120         do {
1121                 next = pmd_addr_end(addr, end);
1122                 if (remap_pte_range(mm, pmd, addr, next,
1123                                 pfn + (addr >> PAGE_SHIFT), prot))
1124                         return -ENOMEM;
1125         } while (pmd++, addr = next, addr != end);
1126         return 0;
1127 }
1128
1129 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1130                         unsigned long addr, unsigned long end,
1131                         unsigned long pfn, pgprot_t prot)
1132 {
1133         pud_t *pud;
1134         unsigned long next;
1135
1136         pfn -= addr >> PAGE_SHIFT;
1137         pud = pud_alloc(mm, pgd, addr);
1138         if (!pud)
1139                 return -ENOMEM;
1140         do {
1141                 next = pud_addr_end(addr, end);
1142                 if (remap_pmd_range(mm, pud, addr, next,
1143                                 pfn + (addr >> PAGE_SHIFT), prot))
1144                         return -ENOMEM;
1145         } while (pud++, addr = next, addr != end);
1146         return 0;
1147 }
1148
1149 /*  Note: this is only safe if the mm semaphore is held when called. */
1150 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1151                     unsigned long pfn, unsigned long size, pgprot_t prot)
1152 {
1153         pgd_t *pgd;
1154         unsigned long next;
1155         unsigned long end = addr + PAGE_ALIGN(size);
1156         struct mm_struct *mm = vma->vm_mm;
1157         int err;
1158
1159         /*
1160          * Physically remapped pages are special. Tell the
1161          * rest of the world about it:
1162          *   VM_IO tells people not to look at these pages
1163          *      (accesses can have side effects).
1164          *   VM_RESERVED tells swapout not to try to touch
1165          *      this region.
1166          */
1167         vma->vm_flags |= VM_IO | VM_RESERVED;
1168
1169         BUG_ON(addr >= end);
1170         pfn -= addr >> PAGE_SHIFT;
1171         pgd = pgd_offset(mm, addr);
1172         flush_cache_range(vma, addr, end);
1173         spin_lock(&mm->page_table_lock);
1174         do {
1175                 next = pgd_addr_end(addr, end);
1176                 err = remap_pud_range(mm, pgd, addr, next,
1177                                 pfn + (addr >> PAGE_SHIFT), prot);
1178                 if (err)
1179                         break;
1180         } while (pgd++, addr = next, addr != end);
1181         spin_unlock(&mm->page_table_lock);
1182         return err;
1183 }
1184 EXPORT_SYMBOL(remap_pfn_range);
1185
1186 /*
1187  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1188  * servicing faults for write access.  In the normal case, do always want
1189  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1190  * that do not have writing enabled, when used by access_process_vm.
1191  */
1192 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1193 {
1194         if (likely(vma->vm_flags & VM_WRITE))
1195                 pte = pte_mkwrite(pte);
1196         return pte;
1197 }
1198
1199 /*
1200  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1201  */
1202 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address, 
1203                 pte_t *page_table)
1204 {
1205         pte_t entry;
1206
1207         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1208                               vma);
1209         ptep_establish(vma, address, page_table, entry);
1210         update_mmu_cache(vma, address, entry);
1211         lazy_mmu_prot_update(entry);
1212 }
1213
1214 /*
1215  * This routine handles present pages, when users try to write
1216  * to a shared page. It is done by copying the page to a new address
1217  * and decrementing the shared-page counter for the old page.
1218  *
1219  * Goto-purists beware: the only reason for goto's here is that it results
1220  * in better assembly code.. The "default" path will see no jumps at all.
1221  *
1222  * Note that this routine assumes that the protection checks have been
1223  * done by the caller (the low-level page fault routine in most cases).
1224  * Thus we can safely just mark it writable once we've done any necessary
1225  * COW.
1226  *
1227  * We also mark the page dirty at this point even though the page will
1228  * change only once the write actually happens. This avoids a few races,
1229  * and potentially makes it more efficient.
1230  *
1231  * We hold the mm semaphore and the page_table_lock on entry and exit
1232  * with the page_table_lock released.
1233  */
1234 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1235         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1236 {
1237         struct page *old_page, *new_page;
1238         unsigned long pfn = pte_pfn(pte);
1239         pte_t entry;
1240         int ret;
1241
1242         if (unlikely(!pfn_valid(pfn))) {
1243                 /*
1244                  * This should really halt the system so it can be debugged or
1245                  * at least the kernel stops what it's doing before it corrupts
1246                  * data, but for the moment just pretend this is OOM.
1247                  */
1248                 pte_unmap(page_table);
1249                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1250                                 address);
1251                 spin_unlock(&mm->page_table_lock);
1252                 return VM_FAULT_OOM;
1253         }
1254         old_page = pfn_to_page(pfn);
1255
1256         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1257                 int reuse = can_share_swap_page(old_page);
1258                 unlock_page(old_page);
1259                 if (reuse) {
1260                         flush_cache_page(vma, address, pfn);
1261                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1262                                               vma);
1263                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1264                         update_mmu_cache(vma, address, entry);
1265                         lazy_mmu_prot_update(entry);
1266                         pte_unmap(page_table);
1267                         spin_unlock(&mm->page_table_lock);
1268                         return VM_FAULT_MINOR|VM_FAULT_WRITE;
1269                 }
1270         }
1271         pte_unmap(page_table);
1272
1273         /*
1274          * Ok, we need to copy. Oh, well..
1275          */
1276         if (!PageReserved(old_page))
1277                 page_cache_get(old_page);
1278         spin_unlock(&mm->page_table_lock);
1279
1280         if (unlikely(anon_vma_prepare(vma)))
1281                 goto no_new_page;
1282         if (old_page == ZERO_PAGE(address)) {
1283                 new_page = alloc_zeroed_user_highpage(vma, address);
1284                 if (!new_page)
1285                         goto no_new_page;
1286         } else {
1287                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1288                 if (!new_page)
1289                         goto no_new_page;
1290                 copy_user_highpage(new_page, old_page, address);
1291         }
1292         /*
1293          * Re-check the pte - we dropped the lock
1294          */
1295         ret = VM_FAULT_MINOR;
1296         spin_lock(&mm->page_table_lock);
1297         page_table = pte_offset_map(pmd, address);
1298         if (likely(pte_same(*page_table, pte))) {
1299                 if (PageAnon(old_page))
1300                         dec_mm_counter(mm, anon_rss);
1301                 if (PageReserved(old_page))
1302                         inc_mm_counter(mm, rss);
1303                 else
1304                         page_remove_rmap(old_page);
1305                 flush_cache_page(vma, address, pfn);
1306                 break_cow(vma, new_page, address, page_table);
1307                 lru_cache_add_active(new_page);
1308                 page_add_anon_rmap(new_page, vma, address);
1309
1310                 /* Free the old page.. */
1311                 new_page = old_page;
1312                 ret |= VM_FAULT_WRITE;
1313         }
1314         pte_unmap(page_table);
1315         page_cache_release(new_page);
1316         page_cache_release(old_page);
1317         spin_unlock(&mm->page_table_lock);
1318         return ret;
1319
1320 no_new_page:
1321         page_cache_release(old_page);
1322         return VM_FAULT_OOM;
1323 }
1324
1325 /*
1326  * Helper functions for unmap_mapping_range().
1327  *
1328  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1329  *
1330  * We have to restart searching the prio_tree whenever we drop the lock,
1331  * since the iterator is only valid while the lock is held, and anyway
1332  * a later vma might be split and reinserted earlier while lock dropped.
1333  *
1334  * The list of nonlinear vmas could be handled more efficiently, using
1335  * a placeholder, but handle it in the same way until a need is shown.
1336  * It is important to search the prio_tree before nonlinear list: a vma
1337  * may become nonlinear and be shifted from prio_tree to nonlinear list
1338  * while the lock is dropped; but never shifted from list to prio_tree.
1339  *
1340  * In order to make forward progress despite restarting the search,
1341  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1342  * quickly skip it next time around.  Since the prio_tree search only
1343  * shows us those vmas affected by unmapping the range in question, we
1344  * can't efficiently keep all vmas in step with mapping->truncate_count:
1345  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1346  * mapping->truncate_count and vma->vm_truncate_count are protected by
1347  * i_mmap_lock.
1348  *
1349  * In order to make forward progress despite repeatedly restarting some
1350  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1351  * and restart from that address when we reach that vma again.  It might
1352  * have been split or merged, shrunk or extended, but never shifted: so
1353  * restart_addr remains valid so long as it remains in the vma's range.
1354  * unmap_mapping_range forces truncate_count to leap over page-aligned
1355  * values so we can save vma's restart_addr in its truncate_count field.
1356  */
1357 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1358
1359 static void reset_vma_truncate_counts(struct address_space *mapping)
1360 {
1361         struct vm_area_struct *vma;
1362         struct prio_tree_iter iter;
1363
1364         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1365                 vma->vm_truncate_count = 0;
1366         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1367                 vma->vm_truncate_count = 0;
1368 }
1369
1370 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1371                 unsigned long start_addr, unsigned long end_addr,
1372                 struct zap_details *details)
1373 {
1374         unsigned long restart_addr;
1375         int need_break;
1376
1377 again:
1378         restart_addr = vma->vm_truncate_count;
1379         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1380                 start_addr = restart_addr;
1381                 if (start_addr >= end_addr) {
1382                         /* Top of vma has been split off since last time */
1383                         vma->vm_truncate_count = details->truncate_count;
1384                         return 0;
1385                 }
1386         }
1387
1388         restart_addr = zap_page_range(vma, start_addr,
1389                                         end_addr - start_addr, details);
1390
1391         /*
1392          * We cannot rely on the break test in unmap_vmas:
1393          * on the one hand, we don't want to restart our loop
1394          * just because that broke out for the page_table_lock;
1395          * on the other hand, it does no test when vma is small.
1396          */
1397         need_break = need_resched() ||
1398                         need_lockbreak(details->i_mmap_lock);
1399
1400         if (restart_addr >= end_addr) {
1401                 /* We have now completed this vma: mark it so */
1402                 vma->vm_truncate_count = details->truncate_count;
1403                 if (!need_break)
1404                         return 0;
1405         } else {
1406                 /* Note restart_addr in vma's truncate_count field */
1407                 vma->vm_truncate_count = restart_addr;
1408                 if (!need_break)
1409                         goto again;
1410         }
1411
1412         spin_unlock(details->i_mmap_lock);
1413         cond_resched();
1414         spin_lock(details->i_mmap_lock);
1415         return -EINTR;
1416 }
1417
1418 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1419                                             struct zap_details *details)
1420 {
1421         struct vm_area_struct *vma;
1422         struct prio_tree_iter iter;
1423         pgoff_t vba, vea, zba, zea;
1424
1425 restart:
1426         vma_prio_tree_foreach(vma, &iter, root,
1427                         details->first_index, details->last_index) {
1428                 /* Skip quickly over those we have already dealt with */
1429                 if (vma->vm_truncate_count == details->truncate_count)
1430                         continue;
1431
1432                 vba = vma->vm_pgoff;
1433                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1434                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1435                 zba = details->first_index;
1436                 if (zba < vba)
1437                         zba = vba;
1438                 zea = details->last_index;
1439                 if (zea > vea)
1440                         zea = vea;
1441
1442                 if (unmap_mapping_range_vma(vma,
1443                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1444                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1445                                 details) < 0)
1446                         goto restart;
1447         }
1448 }
1449
1450 static inline void unmap_mapping_range_list(struct list_head *head,
1451                                             struct zap_details *details)
1452 {
1453         struct vm_area_struct *vma;
1454
1455         /*
1456          * In nonlinear VMAs there is no correspondence between virtual address
1457          * offset and file offset.  So we must perform an exhaustive search
1458          * across *all* the pages in each nonlinear VMA, not just the pages
1459          * whose virtual address lies outside the file truncation point.
1460          */
1461 restart:
1462         list_for_each_entry(vma, head, shared.vm_set.list) {
1463                 /* Skip quickly over those we have already dealt with */
1464                 if (vma->vm_truncate_count == details->truncate_count)
1465                         continue;
1466                 details->nonlinear_vma = vma;
1467                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1468                                         vma->vm_end, details) < 0)
1469                         goto restart;
1470         }
1471 }
1472
1473 /**
1474  * unmap_mapping_range - unmap the portion of all mmaps
1475  * in the specified address_space corresponding to the specified
1476  * page range in the underlying file.
1477  * @mapping: the address space containing mmaps to be unmapped.
1478  * @holebegin: byte in first page to unmap, relative to the start of
1479  * the underlying file.  This will be rounded down to a PAGE_SIZE
1480  * boundary.  Note that this is different from vmtruncate(), which
1481  * must keep the partial page.  In contrast, we must get rid of
1482  * partial pages.
1483  * @holelen: size of prospective hole in bytes.  This will be rounded
1484  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1485  * end of the file.
1486  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1487  * but 0 when invalidating pagecache, don't throw away private data.
1488  */
1489 void unmap_mapping_range(struct address_space *mapping,
1490                 loff_t const holebegin, loff_t const holelen, int even_cows)
1491 {
1492         struct zap_details details;
1493         pgoff_t hba = holebegin >> PAGE_SHIFT;
1494         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1495
1496         /* Check for overflow. */
1497         if (sizeof(holelen) > sizeof(hlen)) {
1498                 long long holeend =
1499                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1500                 if (holeend & ~(long long)ULONG_MAX)
1501                         hlen = ULONG_MAX - hba + 1;
1502         }
1503
1504         details.check_mapping = even_cows? NULL: mapping;
1505         details.nonlinear_vma = NULL;
1506         details.first_index = hba;
1507         details.last_index = hba + hlen - 1;
1508         if (details.last_index < details.first_index)
1509                 details.last_index = ULONG_MAX;
1510         details.i_mmap_lock = &mapping->i_mmap_lock;
1511
1512         spin_lock(&mapping->i_mmap_lock);
1513
1514         /* serialize i_size write against truncate_count write */
1515         smp_wmb();
1516         /* Protect against page faults, and endless unmapping loops */
1517         mapping->truncate_count++;
1518         /*
1519          * For archs where spin_lock has inclusive semantics like ia64
1520          * this smp_mb() will prevent to read pagetable contents
1521          * before the truncate_count increment is visible to
1522          * other cpus.
1523          */
1524         smp_mb();
1525         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1526                 if (mapping->truncate_count == 0)
1527                         reset_vma_truncate_counts(mapping);
1528                 mapping->truncate_count++;
1529         }
1530         details.truncate_count = mapping->truncate_count;
1531
1532         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1533                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1534         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1535                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1536         spin_unlock(&mapping->i_mmap_lock);
1537 }
1538 EXPORT_SYMBOL(unmap_mapping_range);
1539
1540 /*
1541  * Handle all mappings that got truncated by a "truncate()"
1542  * system call.
1543  *
1544  * NOTE! We have to be ready to update the memory sharing
1545  * between the file and the memory map for a potential last
1546  * incomplete page.  Ugly, but necessary.
1547  */
1548 int vmtruncate(struct inode * inode, loff_t offset)
1549 {
1550         struct address_space *mapping = inode->i_mapping;
1551         unsigned long limit;
1552
1553         if (inode->i_size < offset)
1554                 goto do_expand;
1555         /*
1556          * truncation of in-use swapfiles is disallowed - it would cause
1557          * subsequent swapout to scribble on the now-freed blocks.
1558          */
1559         if (IS_SWAPFILE(inode))
1560                 goto out_busy;
1561         i_size_write(inode, offset);
1562         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1563         truncate_inode_pages(mapping, offset);
1564         goto out_truncate;
1565
1566 do_expand:
1567         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1568         if (limit != RLIM_INFINITY && offset > limit)
1569                 goto out_sig;
1570         if (offset > inode->i_sb->s_maxbytes)
1571                 goto out_big;
1572         i_size_write(inode, offset);
1573
1574 out_truncate:
1575         if (inode->i_op && inode->i_op->truncate)
1576                 inode->i_op->truncate(inode);
1577         return 0;
1578 out_sig:
1579         send_sig(SIGXFSZ, current, 0);
1580 out_big:
1581         return -EFBIG;
1582 out_busy:
1583         return -ETXTBSY;
1584 }
1585
1586 EXPORT_SYMBOL(vmtruncate);
1587
1588 /* 
1589  * Primitive swap readahead code. We simply read an aligned block of
1590  * (1 << page_cluster) entries in the swap area. This method is chosen
1591  * because it doesn't cost us any seek time.  We also make sure to queue
1592  * the 'original' request together with the readahead ones...  
1593  *
1594  * This has been extended to use the NUMA policies from the mm triggering
1595  * the readahead.
1596  *
1597  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1598  */
1599 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1600 {
1601 #ifdef CONFIG_NUMA
1602         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1603 #endif
1604         int i, num;
1605         struct page *new_page;
1606         unsigned long offset;
1607
1608         /*
1609          * Get the number of handles we should do readahead io to.
1610          */
1611         num = valid_swaphandles(entry, &offset);
1612         for (i = 0; i < num; offset++, i++) {
1613                 /* Ok, do the async read-ahead now */
1614                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1615                                                            offset), vma, addr);
1616                 if (!new_page)
1617                         break;
1618                 page_cache_release(new_page);
1619 #ifdef CONFIG_NUMA
1620                 /*
1621                  * Find the next applicable VMA for the NUMA policy.
1622                  */
1623                 addr += PAGE_SIZE;
1624                 if (addr == 0)
1625                         vma = NULL;
1626                 if (vma) {
1627                         if (addr >= vma->vm_end) {
1628                                 vma = next_vma;
1629                                 next_vma = vma ? vma->vm_next : NULL;
1630                         }
1631                         if (vma && addr < vma->vm_start)
1632                                 vma = NULL;
1633                 } else {
1634                         if (next_vma && addr >= next_vma->vm_start) {
1635                                 vma = next_vma;
1636                                 next_vma = vma->vm_next;
1637                         }
1638                 }
1639 #endif
1640         }
1641         lru_add_drain();        /* Push any new pages onto the LRU now */
1642 }
1643
1644 /*
1645  * We hold the mm semaphore and the page_table_lock on entry and
1646  * should release the pagetable lock on exit..
1647  */
1648 static int do_swap_page(struct mm_struct * mm,
1649         struct vm_area_struct * vma, unsigned long address,
1650         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1651 {
1652         struct page *page;
1653         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1654         pte_t pte;
1655         int ret = VM_FAULT_MINOR;
1656
1657         pte_unmap(page_table);
1658         spin_unlock(&mm->page_table_lock);
1659         page = lookup_swap_cache(entry);
1660         if (!page) {
1661                 swapin_readahead(entry, address, vma);
1662                 page = read_swap_cache_async(entry, vma, address);
1663                 if (!page) {
1664                         /*
1665                          * Back out if somebody else faulted in this pte while
1666                          * we released the page table lock.
1667                          */
1668                         spin_lock(&mm->page_table_lock);
1669                         page_table = pte_offset_map(pmd, address);
1670                         if (likely(pte_same(*page_table, orig_pte)))
1671                                 ret = VM_FAULT_OOM;
1672                         else
1673                                 ret = VM_FAULT_MINOR;
1674                         pte_unmap(page_table);
1675                         spin_unlock(&mm->page_table_lock);
1676                         goto out;
1677                 }
1678
1679                 /* Had to read the page from swap area: Major fault */
1680                 ret = VM_FAULT_MAJOR;
1681                 inc_page_state(pgmajfault);
1682                 grab_swap_token();
1683         }
1684
1685         mark_page_accessed(page);
1686         lock_page(page);
1687
1688         /*
1689          * Back out if somebody else faulted in this pte while we
1690          * released the page table lock.
1691          */
1692         spin_lock(&mm->page_table_lock);
1693         page_table = pte_offset_map(pmd, address);
1694         if (unlikely(!pte_same(*page_table, orig_pte))) {
1695                 ret = VM_FAULT_MINOR;
1696                 goto out_nomap;
1697         }
1698
1699         if (unlikely(!PageUptodate(page))) {
1700                 ret = VM_FAULT_SIGBUS;
1701                 goto out_nomap;
1702         }
1703
1704         /* The page isn't present yet, go ahead with the fault. */
1705
1706         inc_mm_counter(mm, rss);
1707         pte = mk_pte(page, vma->vm_page_prot);
1708         if (write_access && can_share_swap_page(page)) {
1709                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1710                 write_access = 0;
1711         }
1712
1713         flush_icache_page(vma, page);
1714         set_pte_at(mm, address, page_table, pte);
1715         page_add_anon_rmap(page, vma, address);
1716
1717         swap_free(entry);
1718         if (vm_swap_full())
1719                 remove_exclusive_swap_page(page);
1720         unlock_page(page);
1721
1722         if (write_access) {
1723                 if (do_wp_page(mm, vma, address,
1724                                 page_table, pmd, pte) == VM_FAULT_OOM)
1725                         ret = VM_FAULT_OOM;
1726                 goto out;
1727         }
1728
1729         /* No need to invalidate - it was non-present before */
1730         update_mmu_cache(vma, address, pte);
1731         lazy_mmu_prot_update(pte);
1732         pte_unmap(page_table);
1733         spin_unlock(&mm->page_table_lock);
1734 out:
1735         return ret;
1736 out_nomap:
1737         pte_unmap(page_table);
1738         spin_unlock(&mm->page_table_lock);
1739         unlock_page(page);
1740         page_cache_release(page);
1741         goto out;
1742 }
1743
1744 /*
1745  * We are called with the MM semaphore and page_table_lock
1746  * spinlock held to protect against concurrent faults in
1747  * multithreaded programs. 
1748  */
1749 static int
1750 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1751                 pte_t *page_table, pmd_t *pmd, int write_access,
1752                 unsigned long addr)
1753 {
1754         pte_t entry;
1755         struct page * page = ZERO_PAGE(addr);
1756
1757         /* Read-only mapping of ZERO_PAGE. */
1758         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1759
1760         /* ..except if it's a write access */
1761         if (write_access) {
1762                 /* Allocate our own private page. */
1763                 pte_unmap(page_table);
1764                 spin_unlock(&mm->page_table_lock);
1765
1766                 if (unlikely(anon_vma_prepare(vma)))
1767                         goto no_mem;
1768                 page = alloc_zeroed_user_highpage(vma, addr);
1769                 if (!page)
1770                         goto no_mem;
1771
1772                 spin_lock(&mm->page_table_lock);
1773                 page_table = pte_offset_map(pmd, addr);
1774
1775                 if (!pte_none(*page_table)) {
1776                         pte_unmap(page_table);
1777                         page_cache_release(page);
1778                         spin_unlock(&mm->page_table_lock);
1779                         goto out;
1780                 }
1781                 inc_mm_counter(mm, rss);
1782                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1783                                                          vma->vm_page_prot)),
1784                                       vma);
1785                 lru_cache_add_active(page);
1786                 SetPageReferenced(page);
1787                 page_add_anon_rmap(page, vma, addr);
1788         }
1789
1790         set_pte_at(mm, addr, page_table, entry);
1791         pte_unmap(page_table);
1792
1793         /* No need to invalidate - it was non-present before */
1794         update_mmu_cache(vma, addr, entry);
1795         lazy_mmu_prot_update(entry);
1796         spin_unlock(&mm->page_table_lock);
1797 out:
1798         return VM_FAULT_MINOR;
1799 no_mem:
1800         return VM_FAULT_OOM;
1801 }
1802
1803 /*
1804  * do_no_page() tries to create a new page mapping. It aggressively
1805  * tries to share with existing pages, but makes a separate copy if
1806  * the "write_access" parameter is true in order to avoid the next
1807  * page fault.
1808  *
1809  * As this is called only for pages that do not currently exist, we
1810  * do not need to flush old virtual caches or the TLB.
1811  *
1812  * This is called with the MM semaphore held and the page table
1813  * spinlock held. Exit with the spinlock released.
1814  */
1815 static int
1816 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1817         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1818 {
1819         struct page * new_page;
1820         struct address_space *mapping = NULL;
1821         pte_t entry;
1822         unsigned int sequence = 0;
1823         int ret = VM_FAULT_MINOR;
1824         int anon = 0;
1825
1826         if (!vma->vm_ops || !vma->vm_ops->nopage)
1827                 return do_anonymous_page(mm, vma, page_table,
1828                                         pmd, write_access, address);
1829         pte_unmap(page_table);
1830         spin_unlock(&mm->page_table_lock);
1831
1832         if (vma->vm_file) {
1833                 mapping = vma->vm_file->f_mapping;
1834                 sequence = mapping->truncate_count;
1835                 smp_rmb(); /* serializes i_size against truncate_count */
1836         }
1837 retry:
1838         cond_resched();
1839         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1840         /*
1841          * No smp_rmb is needed here as long as there's a full
1842          * spin_lock/unlock sequence inside the ->nopage callback
1843          * (for the pagecache lookup) that acts as an implicit
1844          * smp_mb() and prevents the i_size read to happen
1845          * after the next truncate_count read.
1846          */
1847
1848         /* no page was available -- either SIGBUS or OOM */
1849         if (new_page == NOPAGE_SIGBUS)
1850                 return VM_FAULT_SIGBUS;
1851         if (new_page == NOPAGE_OOM)
1852                 return VM_FAULT_OOM;
1853
1854         /*
1855          * Should we do an early C-O-W break?
1856          */
1857         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1858                 struct page *page;
1859
1860                 if (unlikely(anon_vma_prepare(vma)))
1861                         goto oom;
1862                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1863                 if (!page)
1864                         goto oom;
1865                 copy_user_highpage(page, new_page, address);
1866                 page_cache_release(new_page);
1867                 new_page = page;
1868                 anon = 1;
1869         }
1870
1871         spin_lock(&mm->page_table_lock);
1872         /*
1873          * For a file-backed vma, someone could have truncated or otherwise
1874          * invalidated this page.  If unmap_mapping_range got called,
1875          * retry getting the page.
1876          */
1877         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1878                 sequence = mapping->truncate_count;
1879                 spin_unlock(&mm->page_table_lock);
1880                 page_cache_release(new_page);
1881                 goto retry;
1882         }
1883         page_table = pte_offset_map(pmd, address);
1884
1885         /*
1886          * This silly early PAGE_DIRTY setting removes a race
1887          * due to the bad i386 page protection. But it's valid
1888          * for other architectures too.
1889          *
1890          * Note that if write_access is true, we either now have
1891          * an exclusive copy of the page, or this is a shared mapping,
1892          * so we can make it writable and dirty to avoid having to
1893          * handle that later.
1894          */
1895         /* Only go through if we didn't race with anybody else... */
1896         if (pte_none(*page_table)) {
1897                 if (!PageReserved(new_page))
1898                         inc_mm_counter(mm, rss);
1899
1900                 flush_icache_page(vma, new_page);
1901                 entry = mk_pte(new_page, vma->vm_page_prot);
1902                 if (write_access)
1903                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1904                 set_pte_at(mm, address, page_table, entry);
1905                 if (anon) {
1906                         lru_cache_add_active(new_page);
1907                         page_add_anon_rmap(new_page, vma, address);
1908                 } else
1909                         page_add_file_rmap(new_page);
1910                 pte_unmap(page_table);
1911         } else {
1912                 /* One of our sibling threads was faster, back out. */
1913                 pte_unmap(page_table);
1914                 page_cache_release(new_page);
1915                 spin_unlock(&mm->page_table_lock);
1916                 goto out;
1917         }
1918
1919         /* no need to invalidate: a not-present page shouldn't be cached */
1920         update_mmu_cache(vma, address, entry);
1921         lazy_mmu_prot_update(entry);
1922         spin_unlock(&mm->page_table_lock);
1923 out:
1924         return ret;
1925 oom:
1926         page_cache_release(new_page);
1927         ret = VM_FAULT_OOM;
1928         goto out;
1929 }
1930
1931 /*
1932  * Fault of a previously existing named mapping. Repopulate the pte
1933  * from the encoded file_pte if possible. This enables swappable
1934  * nonlinear vmas.
1935  */
1936 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1937         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1938 {
1939         unsigned long pgoff;
1940         int err;
1941
1942         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1943         /*
1944          * Fall back to the linear mapping if the fs does not support
1945          * ->populate:
1946          */
1947         if (!vma->vm_ops || !vma->vm_ops->populate || 
1948                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1949                 pte_clear(mm, address, pte);
1950                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1951         }
1952
1953         pgoff = pte_to_pgoff(*pte);
1954
1955         pte_unmap(pte);
1956         spin_unlock(&mm->page_table_lock);
1957
1958         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1959         if (err == -ENOMEM)
1960                 return VM_FAULT_OOM;
1961         if (err)
1962                 return VM_FAULT_SIGBUS;
1963         return VM_FAULT_MAJOR;
1964 }
1965
1966 /*
1967  * These routines also need to handle stuff like marking pages dirty
1968  * and/or accessed for architectures that don't do it in hardware (most
1969  * RISC architectures).  The early dirtying is also good on the i386.
1970  *
1971  * There is also a hook called "update_mmu_cache()" that architectures
1972  * with external mmu caches can use to update those (ie the Sparc or
1973  * PowerPC hashed page tables that act as extended TLBs).
1974  *
1975  * Note the "page_table_lock". It is to protect against kswapd removing
1976  * pages from under us. Note that kswapd only ever _removes_ pages, never
1977  * adds them. As such, once we have noticed that the page is not present,
1978  * we can drop the lock early.
1979  *
1980  * The adding of pages is protected by the MM semaphore (which we hold),
1981  * so we don't need to worry about a page being suddenly been added into
1982  * our VM.
1983  *
1984  * We enter with the pagetable spinlock held, we are supposed to
1985  * release it when done.
1986  */
1987 static inline int handle_pte_fault(struct mm_struct *mm,
1988         struct vm_area_struct * vma, unsigned long address,
1989         int write_access, pte_t *pte, pmd_t *pmd)
1990 {
1991         pte_t entry;
1992
1993         entry = *pte;
1994         if (!pte_present(entry)) {
1995                 /*
1996                  * If it truly wasn't present, we know that kswapd
1997                  * and the PTE updates will not touch it later. So
1998                  * drop the lock.
1999                  */
2000                 if (pte_none(entry))
2001                         return do_no_page(mm, vma, address, write_access, pte, pmd);
2002                 if (pte_file(entry))
2003                         return do_file_page(mm, vma, address, write_access, pte, pmd);
2004                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
2005         }
2006
2007         if (write_access) {
2008                 if (!pte_write(entry))
2009                         return do_wp_page(mm, vma, address, pte, pmd, entry);
2010                 entry = pte_mkdirty(entry);
2011         }
2012         entry = pte_mkyoung(entry);
2013         ptep_set_access_flags(vma, address, pte, entry, write_access);
2014         update_mmu_cache(vma, address, entry);
2015         lazy_mmu_prot_update(entry);
2016         pte_unmap(pte);
2017         spin_unlock(&mm->page_table_lock);
2018         return VM_FAULT_MINOR;
2019 }
2020
2021 /*
2022  * By the time we get here, we already hold the mm semaphore
2023  */
2024 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
2025                 unsigned long address, int write_access)
2026 {
2027         pgd_t *pgd;
2028         pud_t *pud;
2029         pmd_t *pmd;
2030         pte_t *pte;
2031
2032         __set_current_state(TASK_RUNNING);
2033
2034         inc_page_state(pgfault);
2035
2036         if (is_vm_hugetlb_page(vma))
2037                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
2038
2039         /*
2040          * We need the page table lock to synchronize with kswapd
2041          * and the SMP-safe atomic PTE updates.
2042          */
2043         pgd = pgd_offset(mm, address);
2044         spin_lock(&mm->page_table_lock);
2045
2046         pud = pud_alloc(mm, pgd, address);
2047         if (!pud)
2048                 goto oom;
2049
2050         pmd = pmd_alloc(mm, pud, address);
2051         if (!pmd)
2052                 goto oom;
2053
2054         pte = pte_alloc_map(mm, pmd, address);
2055         if (!pte)
2056                 goto oom;
2057         
2058         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
2059
2060  oom:
2061         spin_unlock(&mm->page_table_lock);
2062         return VM_FAULT_OOM;
2063 }
2064
2065 #ifndef __PAGETABLE_PUD_FOLDED
2066 /*
2067  * Allocate page upper directory.
2068  *
2069  * We've already handled the fast-path in-line, and we own the
2070  * page table lock.
2071  */
2072 pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2073 {
2074         pud_t *new;
2075
2076         spin_unlock(&mm->page_table_lock);
2077         new = pud_alloc_one(mm, address);
2078         spin_lock(&mm->page_table_lock);
2079         if (!new)
2080                 return NULL;
2081
2082         /*
2083          * Because we dropped the lock, we should re-check the
2084          * entry, as somebody else could have populated it..
2085          */
2086         if (pgd_present(*pgd)) {
2087                 pud_free(new);
2088                 goto out;
2089         }
2090         pgd_populate(mm, pgd, new);
2091  out:
2092         return pud_offset(pgd, address);
2093 }
2094 #endif /* __PAGETABLE_PUD_FOLDED */
2095
2096 #ifndef __PAGETABLE_PMD_FOLDED
2097 /*
2098  * Allocate page middle directory.
2099  *
2100  * We've already handled the fast-path in-line, and we own the
2101  * page table lock.
2102  */
2103 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2104 {
2105         pmd_t *new;
2106
2107         spin_unlock(&mm->page_table_lock);
2108         new = pmd_alloc_one(mm, address);
2109         spin_lock(&mm->page_table_lock);
2110         if (!new)
2111                 return NULL;
2112
2113         /*
2114          * Because we dropped the lock, we should re-check the
2115          * entry, as somebody else could have populated it..
2116          */
2117 #ifndef __ARCH_HAS_4LEVEL_HACK
2118         if (pud_present(*pud)) {
2119                 pmd_free(new);
2120                 goto out;
2121         }
2122         pud_populate(mm, pud, new);
2123 #else
2124         if (pgd_present(*pud)) {
2125                 pmd_free(new);
2126                 goto out;
2127         }
2128         pgd_populate(mm, pud, new);
2129 #endif /* __ARCH_HAS_4LEVEL_HACK */
2130
2131  out:
2132         return pmd_offset(pud, address);
2133 }
2134 #endif /* __PAGETABLE_PMD_FOLDED */
2135
2136 int make_pages_present(unsigned long addr, unsigned long end)
2137 {
2138         int ret, len, write;
2139         struct vm_area_struct * vma;
2140
2141         vma = find_vma(current->mm, addr);
2142         if (!vma)
2143                 return -1;
2144         write = (vma->vm_flags & VM_WRITE) != 0;
2145         if (addr >= end)
2146                 BUG();
2147         if (end > vma->vm_end)
2148                 BUG();
2149         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2150         ret = get_user_pages(current, current->mm, addr,
2151                         len, write, 0, NULL, NULL);
2152         if (ret < 0)
2153                 return ret;
2154         return ret == len ? 0 : -1;
2155 }
2156
2157 /* 
2158  * Map a vmalloc()-space virtual address to the physical page.
2159  */
2160 struct page * vmalloc_to_page(void * vmalloc_addr)
2161 {
2162         unsigned long addr = (unsigned long) vmalloc_addr;
2163         struct page *page = NULL;
2164         pgd_t *pgd = pgd_offset_k(addr);
2165         pud_t *pud;
2166         pmd_t *pmd;
2167         pte_t *ptep, pte;
2168   
2169         if (!pgd_none(*pgd)) {
2170                 pud = pud_offset(pgd, addr);
2171                 if (!pud_none(*pud)) {
2172                         pmd = pmd_offset(pud, addr);
2173                         if (!pmd_none(*pmd)) {
2174                                 ptep = pte_offset_map(pmd, addr);
2175                                 pte = *ptep;
2176                                 if (pte_present(pte))
2177                                         page = pte_page(pte);
2178                                 pte_unmap(ptep);
2179                         }
2180                 }
2181         }
2182         return page;
2183 }
2184
2185 EXPORT_SYMBOL(vmalloc_to_page);
2186
2187 /*
2188  * Map a vmalloc()-space virtual address to the physical page frame number.
2189  */
2190 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2191 {
2192         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2193 }
2194
2195 EXPORT_SYMBOL(vmalloc_to_pfn);
2196
2197 /*
2198  * update_mem_hiwater
2199  *      - update per process rss and vm high water data
2200  */
2201 void update_mem_hiwater(struct task_struct *tsk)
2202 {
2203         if (tsk->mm) {
2204                 unsigned long rss = get_mm_counter(tsk->mm, rss);
2205
2206                 if (tsk->mm->hiwater_rss < rss)
2207                         tsk->mm->hiwater_rss = rss;
2208                 if (tsk->mm->hiwater_vm < tsk->mm->total_vm)
2209                         tsk->mm->hiwater_vm = tsk->mm->total_vm;
2210         }
2211 }
2212
2213 #if !defined(__HAVE_ARCH_GATE_AREA)
2214
2215 #if defined(AT_SYSINFO_EHDR)
2216 struct vm_area_struct gate_vma;
2217
2218 static int __init gate_vma_init(void)
2219 {
2220         gate_vma.vm_mm = NULL;
2221         gate_vma.vm_start = FIXADDR_USER_START;
2222         gate_vma.vm_end = FIXADDR_USER_END;
2223         gate_vma.vm_page_prot = PAGE_READONLY;
2224         gate_vma.vm_flags = 0;
2225         return 0;
2226 }
2227 __initcall(gate_vma_init);
2228 #endif
2229
2230 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2231 {
2232 #ifdef AT_SYSINFO_EHDR
2233         return &gate_vma;
2234 #else
2235         return NULL;
2236 #endif
2237 }
2238
2239 int in_gate_area_no_task(unsigned long addr)
2240 {
2241 #ifdef AT_SYSINFO_EHDR
2242         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2243                 return 1;
2244 #endif
2245         return 0;
2246 }
2247
2248 #endif  /* __HAVE_ARCH_GATE_AREA */