854bd90eeca1928b15dfb61088dd3d13360cb373
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
54 #include <asm/tlb.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
57
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60
61 #ifndef CONFIG_DISCONTIGMEM
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
65
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
69
70 unsigned long num_physpages;
71 /*
72  * A number of key systems in x86 including ioremap() rely on the assumption
73  * that high_memory defines the upper bound on direct map memory, then end
74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76  * and ZONE_HIGHMEM.
77  */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
80
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85 /*
86  * If a p?d_bad entry is found while walking page tables, report
87  * the error, before resetting entry to p?d_none.  Usually (but
88  * very seldom) called out from the p?d_none_or_clear_bad macros.
89  */
90
91 void pgd_clear_bad(pgd_t *pgd)
92 {
93         pgd_ERROR(*pgd);
94         pgd_clear(pgd);
95 }
96
97 void pud_clear_bad(pud_t *pud)
98 {
99         pud_ERROR(*pud);
100         pud_clear(pud);
101 }
102
103 void pmd_clear_bad(pmd_t *pmd)
104 {
105         pmd_ERROR(*pmd);
106         pmd_clear(pmd);
107 }
108
109 /*
110  * Note: this doesn't free the actual pages themselves. That
111  * has been handled earlier when unmapping all the memory regions.
112  */
113 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
114 {
115         struct page *page = pmd_page(*pmd);
116         pmd_clear(pmd);
117         pte_free_tlb(tlb, page);
118         dec_page_state(nr_page_table_pages);
119         tlb->mm->nr_ptes--;
120 }
121
122 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
123                                 unsigned long addr, unsigned long end,
124                                 unsigned long floor, unsigned long ceiling)
125 {
126         pmd_t *pmd;
127         unsigned long next;
128         unsigned long start;
129
130         start = addr;
131         pmd = pmd_offset(pud, addr);
132         do {
133                 next = pmd_addr_end(addr, end);
134                 if (pmd_none_or_clear_bad(pmd))
135                         continue;
136                 free_pte_range(tlb, pmd);
137         } while (pmd++, addr = next, addr != end);
138
139         start &= PUD_MASK;
140         if (start < floor)
141                 return;
142         if (ceiling) {
143                 ceiling &= PUD_MASK;
144                 if (!ceiling)
145                         return;
146         }
147         if (end - 1 > ceiling - 1)
148                 return;
149
150         pmd = pmd_offset(pud, start);
151         pud_clear(pud);
152         pmd_free_tlb(tlb, pmd);
153 }
154
155 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
156                                 unsigned long addr, unsigned long end,
157                                 unsigned long floor, unsigned long ceiling)
158 {
159         pud_t *pud;
160         unsigned long next;
161         unsigned long start;
162
163         start = addr;
164         pud = pud_offset(pgd, addr);
165         do {
166                 next = pud_addr_end(addr, end);
167                 if (pud_none_or_clear_bad(pud))
168                         continue;
169                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
170         } while (pud++, addr = next, addr != end);
171
172         start &= PGDIR_MASK;
173         if (start < floor)
174                 return;
175         if (ceiling) {
176                 ceiling &= PGDIR_MASK;
177                 if (!ceiling)
178                         return;
179         }
180         if (end - 1 > ceiling - 1)
181                 return;
182
183         pud = pud_offset(pgd, start);
184         pgd_clear(pgd);
185         pud_free_tlb(tlb, pud);
186 }
187
188 /*
189  * This function frees user-level page tables of a process.
190  *
191  * Must be called with pagetable lock held.
192  */
193 static inline 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                 /* Optimization: gather nearby vmas into a single call down */
264                 while (next && next->vm_start <= vma->vm_end + PMD_SIZE) {
265                         vma = next;
266                         next = vma->vm_next;
267                 }
268                 free_pgd_range(*tlb, addr, vma->vm_end,
269                                 floor, next? next->vm_start: ceiling);
270                 vma = next;
271         }
272 }
273
274 pte_t fastcall * pte_alloc_map(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
275 {
276         if (!pmd_present(*pmd)) {
277                 struct page *new;
278
279                 spin_unlock(&mm->page_table_lock);
280                 new = pte_alloc_one(mm, address);
281                 spin_lock(&mm->page_table_lock);
282                 if (!new)
283                         return NULL;
284                 /*
285                  * Because we dropped the lock, we should re-check the
286                  * entry, as somebody else could have populated it..
287                  */
288                 if (pmd_present(*pmd)) {
289                         pte_free(new);
290                         goto out;
291                 }
292                 mm->nr_ptes++;
293                 inc_page_state(nr_page_table_pages);
294                 pmd_populate(mm, pmd, new);
295         }
296 out:
297         return pte_offset_map(pmd, address);
298 }
299
300 pte_t fastcall * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
301 {
302         if (!pmd_present(*pmd)) {
303                 pte_t *new;
304
305                 spin_unlock(&mm->page_table_lock);
306                 new = pte_alloc_one_kernel(mm, address);
307                 spin_lock(&mm->page_table_lock);
308                 if (!new)
309                         return NULL;
310
311                 /*
312                  * Because we dropped the lock, we should re-check the
313                  * entry, as somebody else could have populated it..
314                  */
315                 if (pmd_present(*pmd)) {
316                         pte_free_kernel(new);
317                         goto out;
318                 }
319                 pmd_populate_kernel(mm, pmd, new);
320         }
321 out:
322         return pte_offset_kernel(pmd, address);
323 }
324
325 /*
326  * copy one vm_area from one task to the other. Assumes the page tables
327  * already present in the new task to be cleared in the whole range
328  * covered by this vma.
329  *
330  * dst->page_table_lock is held on entry and exit,
331  * but may be dropped within p[mg]d_alloc() and pte_alloc_map().
332  */
333
334 static inline void
335 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
336                 pte_t *dst_pte, pte_t *src_pte, unsigned long vm_flags,
337                 unsigned long addr)
338 {
339         pte_t pte = *src_pte;
340         struct page *page;
341         unsigned long pfn;
342
343         /* pte contains position in swap or file, so copy. */
344         if (unlikely(!pte_present(pte))) {
345                 if (!pte_file(pte)) {
346                         swap_duplicate(pte_to_swp_entry(pte));
347                         /* make sure dst_mm is on swapoff's mmlist. */
348                         if (unlikely(list_empty(&dst_mm->mmlist))) {
349                                 spin_lock(&mmlist_lock);
350                                 list_add(&dst_mm->mmlist, &src_mm->mmlist);
351                                 spin_unlock(&mmlist_lock);
352                         }
353                 }
354                 set_pte_at(dst_mm, addr, dst_pte, pte);
355                 return;
356         }
357
358         pfn = pte_pfn(pte);
359         /* the pte points outside of valid memory, the
360          * mapping is assumed to be good, meaningful
361          * and not mapped via rmap - duplicate the
362          * mapping as is.
363          */
364         page = NULL;
365         if (pfn_valid(pfn))
366                 page = pfn_to_page(pfn);
367
368         if (!page || PageReserved(page)) {
369                 set_pte_at(dst_mm, addr, dst_pte, pte);
370                 return;
371         }
372
373         /*
374          * If it's a COW mapping, write protect it both
375          * in the parent and the child
376          */
377         if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
378                 ptep_set_wrprotect(src_mm, addr, src_pte);
379                 pte = *src_pte;
380         }
381
382         /*
383          * If it's a shared mapping, mark it clean in
384          * the child
385          */
386         if (vm_flags & VM_SHARED)
387                 pte = pte_mkclean(pte);
388         pte = pte_mkold(pte);
389         get_page(page);
390         inc_mm_counter(dst_mm, rss);
391         if (PageAnon(page))
392                 inc_mm_counter(dst_mm, anon_rss);
393         set_pte_at(dst_mm, addr, dst_pte, pte);
394         page_dup_rmap(page);
395 }
396
397 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
398                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
399                 unsigned long addr, unsigned long end)
400 {
401         pte_t *src_pte, *dst_pte;
402         unsigned long vm_flags = vma->vm_flags;
403         int progress;
404
405 again:
406         dst_pte = pte_alloc_map(dst_mm, dst_pmd, addr);
407         if (!dst_pte)
408                 return -ENOMEM;
409         src_pte = pte_offset_map_nested(src_pmd, addr);
410
411         progress = 0;
412         spin_lock(&src_mm->page_table_lock);
413         do {
414                 /*
415                  * We are holding two locks at this point - either of them
416                  * could generate latencies in another task on another CPU.
417                  */
418                 if (progress >= 32 && (need_resched() ||
419                     need_lockbreak(&src_mm->page_table_lock) ||
420                     need_lockbreak(&dst_mm->page_table_lock)))
421                         break;
422                 if (pte_none(*src_pte)) {
423                         progress++;
424                         continue;
425                 }
426                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vm_flags, addr);
427                 progress += 8;
428         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
429         spin_unlock(&src_mm->page_table_lock);
430
431         pte_unmap_nested(src_pte - 1);
432         pte_unmap(dst_pte - 1);
433         cond_resched_lock(&dst_mm->page_table_lock);
434         if (addr != end)
435                 goto again;
436         return 0;
437 }
438
439 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
440                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
441                 unsigned long addr, unsigned long end)
442 {
443         pmd_t *src_pmd, *dst_pmd;
444         unsigned long next;
445
446         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
447         if (!dst_pmd)
448                 return -ENOMEM;
449         src_pmd = pmd_offset(src_pud, addr);
450         do {
451                 next = pmd_addr_end(addr, end);
452                 if (pmd_none_or_clear_bad(src_pmd))
453                         continue;
454                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
455                                                 vma, addr, next))
456                         return -ENOMEM;
457         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
458         return 0;
459 }
460
461 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
462                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
463                 unsigned long addr, unsigned long end)
464 {
465         pud_t *src_pud, *dst_pud;
466         unsigned long next;
467
468         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
469         if (!dst_pud)
470                 return -ENOMEM;
471         src_pud = pud_offset(src_pgd, addr);
472         do {
473                 next = pud_addr_end(addr, end);
474                 if (pud_none_or_clear_bad(src_pud))
475                         continue;
476                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
477                                                 vma, addr, next))
478                         return -ENOMEM;
479         } while (dst_pud++, src_pud++, addr = next, addr != end);
480         return 0;
481 }
482
483 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
484                 struct vm_area_struct *vma)
485 {
486         pgd_t *src_pgd, *dst_pgd;
487         unsigned long next;
488         unsigned long addr = vma->vm_start;
489         unsigned long end = vma->vm_end;
490
491         if (is_vm_hugetlb_page(vma))
492                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
493
494         dst_pgd = pgd_offset(dst_mm, addr);
495         src_pgd = pgd_offset(src_mm, addr);
496         do {
497                 next = pgd_addr_end(addr, end);
498                 if (pgd_none_or_clear_bad(src_pgd))
499                         continue;
500                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
501                                                 vma, addr, next))
502                         return -ENOMEM;
503         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
504         return 0;
505 }
506
507 static void zap_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
508                                 unsigned long addr, unsigned long end,
509                                 struct zap_details *details)
510 {
511         pte_t *pte;
512
513         pte = pte_offset_map(pmd, addr);
514         do {
515                 pte_t ptent = *pte;
516                 if (pte_none(ptent))
517                         continue;
518                 if (pte_present(ptent)) {
519                         struct page *page = NULL;
520                         unsigned long pfn = pte_pfn(ptent);
521                         if (pfn_valid(pfn)) {
522                                 page = pfn_to_page(pfn);
523                                 if (PageReserved(page))
524                                         page = NULL;
525                         }
526                         if (unlikely(details) && page) {
527                                 /*
528                                  * unmap_shared_mapping_pages() wants to
529                                  * invalidate cache without truncating:
530                                  * unmap shared but keep private pages.
531                                  */
532                                 if (details->check_mapping &&
533                                     details->check_mapping != page->mapping)
534                                         continue;
535                                 /*
536                                  * Each page->index must be checked when
537                                  * invalidating or truncating nonlinear.
538                                  */
539                                 if (details->nonlinear_vma &&
540                                     (page->index < details->first_index ||
541                                      page->index > details->last_index))
542                                         continue;
543                         }
544                         ptent = ptep_get_and_clear(tlb->mm, addr, pte);
545                         tlb_remove_tlb_entry(tlb, pte, addr);
546                         if (unlikely(!page))
547                                 continue;
548                         if (unlikely(details) && details->nonlinear_vma
549                             && linear_page_index(details->nonlinear_vma,
550                                                 addr) != page->index)
551                                 set_pte_at(tlb->mm, addr, pte,
552                                            pgoff_to_pte(page->index));
553                         if (pte_dirty(ptent))
554                                 set_page_dirty(page);
555                         if (PageAnon(page))
556                                 dec_mm_counter(tlb->mm, anon_rss);
557                         else if (pte_young(ptent))
558                                 mark_page_accessed(page);
559                         tlb->freed++;
560                         page_remove_rmap(page);
561                         tlb_remove_page(tlb, page);
562                         continue;
563                 }
564                 /*
565                  * If details->check_mapping, we leave swap entries;
566                  * if details->nonlinear_vma, we leave file entries.
567                  */
568                 if (unlikely(details))
569                         continue;
570                 if (!pte_file(ptent))
571                         free_swap_and_cache(pte_to_swp_entry(ptent));
572                 pte_clear(tlb->mm, addr, pte);
573         } while (pte++, addr += PAGE_SIZE, addr != end);
574         pte_unmap(pte - 1);
575 }
576
577 static inline void zap_pmd_range(struct mmu_gather *tlb, pud_t *pud,
578                                 unsigned long addr, unsigned long end,
579                                 struct zap_details *details)
580 {
581         pmd_t *pmd;
582         unsigned long next;
583
584         pmd = pmd_offset(pud, addr);
585         do {
586                 next = pmd_addr_end(addr, end);
587                 if (pmd_none_or_clear_bad(pmd))
588                         continue;
589                 zap_pte_range(tlb, pmd, addr, next, details);
590         } while (pmd++, addr = next, addr != end);
591 }
592
593 static inline void zap_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
594                                 unsigned long addr, unsigned long end,
595                                 struct zap_details *details)
596 {
597         pud_t *pud;
598         unsigned long next;
599
600         pud = pud_offset(pgd, addr);
601         do {
602                 next = pud_addr_end(addr, end);
603                 if (pud_none_or_clear_bad(pud))
604                         continue;
605                 zap_pmd_range(tlb, pud, addr, next, details);
606         } while (pud++, addr = next, addr != end);
607 }
608
609 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
610                                 unsigned long addr, unsigned long end,
611                                 struct zap_details *details)
612 {
613         pgd_t *pgd;
614         unsigned long next;
615
616         if (details && !details->check_mapping && !details->nonlinear_vma)
617                 details = NULL;
618
619         BUG_ON(addr >= end);
620         tlb_start_vma(tlb, vma);
621         pgd = pgd_offset(vma->vm_mm, addr);
622         do {
623                 next = pgd_addr_end(addr, end);
624                 if (pgd_none_or_clear_bad(pgd))
625                         continue;
626                 zap_pud_range(tlb, pgd, addr, next, details);
627         } while (pgd++, addr = next, addr != end);
628         tlb_end_vma(tlb, vma);
629 }
630
631 #ifdef CONFIG_PREEMPT
632 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
633 #else
634 /* No preempt: go for improved straight-line efficiency */
635 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
636 #endif
637
638 /**
639  * unmap_vmas - unmap a range of memory covered by a list of vma's
640  * @tlbp: address of the caller's struct mmu_gather
641  * @mm: the controlling mm_struct
642  * @vma: the starting vma
643  * @start_addr: virtual address at which to start unmapping
644  * @end_addr: virtual address at which to end unmapping
645  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
646  * @details: details of nonlinear truncation or shared cache invalidation
647  *
648  * Returns the end address of the unmapping (restart addr if interrupted).
649  *
650  * Unmap all pages in the vma list.  Called under page_table_lock.
651  *
652  * We aim to not hold page_table_lock for too long (for scheduling latency
653  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
654  * return the ending mmu_gather to the caller.
655  *
656  * Only addresses between `start' and `end' will be unmapped.
657  *
658  * The VMA list must be sorted in ascending virtual address order.
659  *
660  * unmap_vmas() assumes that the caller will flush the whole unmapped address
661  * range after unmap_vmas() returns.  So the only responsibility here is to
662  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
663  * drops the lock and schedules.
664  */
665 unsigned long unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
666                 struct vm_area_struct *vma, unsigned long start_addr,
667                 unsigned long end_addr, unsigned long *nr_accounted,
668                 struct zap_details *details)
669 {
670         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
671         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
672         int tlb_start_valid = 0;
673         unsigned long start = start_addr;
674         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
675         int fullmm = tlb_is_full_mm(*tlbp);
676
677         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
678                 unsigned long end;
679
680                 start = max(vma->vm_start, start_addr);
681                 if (start >= vma->vm_end)
682                         continue;
683                 end = min(vma->vm_end, end_addr);
684                 if (end <= vma->vm_start)
685                         continue;
686
687                 if (vma->vm_flags & VM_ACCOUNT)
688                         *nr_accounted += (end - start) >> PAGE_SHIFT;
689
690                 while (start != end) {
691                         unsigned long block;
692
693                         if (!tlb_start_valid) {
694                                 tlb_start = start;
695                                 tlb_start_valid = 1;
696                         }
697
698                         if (is_vm_hugetlb_page(vma)) {
699                                 block = end - start;
700                                 unmap_hugepage_range(vma, start, end);
701                         } else {
702                                 block = min(zap_bytes, end - start);
703                                 unmap_page_range(*tlbp, vma, start,
704                                                 start + block, details);
705                         }
706
707                         start += block;
708                         zap_bytes -= block;
709                         if ((long)zap_bytes > 0)
710                                 continue;
711
712                         tlb_finish_mmu(*tlbp, tlb_start, start);
713
714                         if (need_resched() ||
715                                 need_lockbreak(&mm->page_table_lock) ||
716                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
717                                 if (i_mmap_lock) {
718                                         /* must reset count of rss freed */
719                                         *tlbp = tlb_gather_mmu(mm, fullmm);
720                                         goto out;
721                                 }
722                                 spin_unlock(&mm->page_table_lock);
723                                 cond_resched();
724                                 spin_lock(&mm->page_table_lock);
725                         }
726
727                         *tlbp = tlb_gather_mmu(mm, fullmm);
728                         tlb_start_valid = 0;
729                         zap_bytes = ZAP_BLOCK_SIZE;
730                 }
731         }
732 out:
733         return start;   /* which is now the end (or restart) address */
734 }
735
736 /**
737  * zap_page_range - remove user pages in a given range
738  * @vma: vm_area_struct holding the applicable pages
739  * @address: starting address of pages to zap
740  * @size: number of bytes to zap
741  * @details: details of nonlinear truncation or shared cache invalidation
742  */
743 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
744                 unsigned long size, struct zap_details *details)
745 {
746         struct mm_struct *mm = vma->vm_mm;
747         struct mmu_gather *tlb;
748         unsigned long end = address + size;
749         unsigned long nr_accounted = 0;
750
751         if (is_vm_hugetlb_page(vma)) {
752                 zap_hugepage_range(vma, address, size);
753                 return end;
754         }
755
756         lru_add_drain();
757         spin_lock(&mm->page_table_lock);
758         tlb = tlb_gather_mmu(mm, 0);
759         end = unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
760         tlb_finish_mmu(tlb, address, end);
761         spin_unlock(&mm->page_table_lock);
762         return end;
763 }
764
765 /*
766  * Do a quick page-table lookup for a single page.
767  * mm->page_table_lock must be held.
768  */
769 static struct page *
770 __follow_page(struct mm_struct *mm, unsigned long address, int read, int write)
771 {
772         pgd_t *pgd;
773         pud_t *pud;
774         pmd_t *pmd;
775         pte_t *ptep, pte;
776         unsigned long pfn;
777         struct page *page;
778
779         page = follow_huge_addr(mm, address, write);
780         if (! IS_ERR(page))
781                 return page;
782
783         pgd = pgd_offset(mm, address);
784         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
785                 goto out;
786
787         pud = pud_offset(pgd, address);
788         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
789                 goto out;
790         
791         pmd = pmd_offset(pud, address);
792         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
793                 goto out;
794         if (pmd_huge(*pmd))
795                 return follow_huge_pmd(mm, address, pmd, write);
796
797         ptep = pte_offset_map(pmd, address);
798         if (!ptep)
799                 goto out;
800
801         pte = *ptep;
802         pte_unmap(ptep);
803         if (pte_present(pte)) {
804                 if (write && !pte_write(pte))
805                         goto out;
806                 if (read && !pte_read(pte))
807                         goto out;
808                 pfn = pte_pfn(pte);
809                 if (pfn_valid(pfn)) {
810                         page = pfn_to_page(pfn);
811                         if (write && !pte_dirty(pte) && !PageDirty(page))
812                                 set_page_dirty(page);
813                         mark_page_accessed(page);
814                         return page;
815                 }
816         }
817
818 out:
819         return NULL;
820 }
821
822 struct page *
823 follow_page(struct mm_struct *mm, unsigned long address, int write)
824 {
825         return __follow_page(mm, address, /*read*/0, write);
826 }
827
828 int
829 check_user_page_readable(struct mm_struct *mm, unsigned long address)
830 {
831         return __follow_page(mm, address, /*read*/1, /*write*/0) != NULL;
832 }
833
834 EXPORT_SYMBOL(check_user_page_readable);
835
836 /* 
837  * Given a physical address, is there a useful struct page pointing to
838  * it?  This may become more complex in the future if we start dealing
839  * with IO-aperture pages for direct-IO.
840  */
841
842 static inline struct page *get_page_map(struct page *page)
843 {
844         if (!pfn_valid(page_to_pfn(page)))
845                 return NULL;
846         return page;
847 }
848
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
881 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
882                 unsigned long start, int len, int write, int force,
883                 struct page **pages, struct vm_area_struct **vmas)
884 {
885         int i;
886         unsigned int flags;
887
888         /* 
889          * Require read or write permissions.
890          * If 'force' is set, we only require the "MAY" flags.
891          */
892         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
893         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
894         i = 0;
895
896         do {
897                 struct vm_area_struct * vma;
898
899                 vma = find_extend_vma(mm, start);
900                 if (!vma && in_gate_area(tsk, start)) {
901                         unsigned long pg = start & PAGE_MASK;
902                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
903                         pgd_t *pgd;
904                         pud_t *pud;
905                         pmd_t *pmd;
906                         pte_t *pte;
907                         if (write) /* user gate pages are read-only */
908                                 return i ? : -EFAULT;
909                         if (pg > TASK_SIZE)
910                                 pgd = pgd_offset_k(pg);
911                         else
912                                 pgd = pgd_offset_gate(mm, pg);
913                         BUG_ON(pgd_none(*pgd));
914                         pud = pud_offset(pgd, pg);
915                         BUG_ON(pud_none(*pud));
916                         pmd = pmd_offset(pud, pg);
917                         BUG_ON(pmd_none(*pmd));
918                         pte = pte_offset_map(pmd, pg);
919                         BUG_ON(pte_none(*pte));
920                         if (pages) {
921                                 pages[i] = pte_page(*pte);
922                                 get_page(pages[i]);
923                         }
924                         pte_unmap(pte);
925                         if (vmas)
926                                 vmas[i] = gate_vma;
927                         i++;
928                         start += PAGE_SIZE;
929                         len--;
930                         continue;
931                 }
932
933                 if (!vma || (vma->vm_flags & VM_IO)
934                                 || !(flags & vma->vm_flags))
935                         return i ? : -EFAULT;
936
937                 if (is_vm_hugetlb_page(vma)) {
938                         i = follow_hugetlb_page(mm, vma, pages, vmas,
939                                                 &start, &len, i);
940                         continue;
941                 }
942                 spin_lock(&mm->page_table_lock);
943                 do {
944                         struct page *map;
945                         int lookup_write = write;
946
947                         cond_resched_lock(&mm->page_table_lock);
948                         while (!(map = follow_page(mm, start, lookup_write))) {
949                                 /*
950                                  * Shortcut for anonymous pages. We don't want
951                                  * to force the creation of pages tables for
952                                  * insanly big anonymously mapped areas that
953                                  * nobody touched so far. This is important
954                                  * for doing a core dump for these mappings.
955                                  */
956                                 if (!lookup_write &&
957                                     untouched_anonymous_page(mm,vma,start)) {
958                                         map = ZERO_PAGE(start);
959                                         break;
960                                 }
961                                 spin_unlock(&mm->page_table_lock);
962                                 switch (handle_mm_fault(mm,vma,start,write)) {
963                                 case VM_FAULT_MINOR:
964                                         tsk->min_flt++;
965                                         break;
966                                 case VM_FAULT_MAJOR:
967                                         tsk->maj_flt++;
968                                         break;
969                                 case VM_FAULT_SIGBUS:
970                                         return i ? i : -EFAULT;
971                                 case VM_FAULT_OOM:
972                                         return i ? i : -ENOMEM;
973                                 default:
974                                         BUG();
975                                 }
976                                 /*
977                                  * Now that we have performed a write fault
978                                  * and surely no longer have a shared page we
979                                  * shouldn't write, we shouldn't ignore an
980                                  * unwritable page in the page table if
981                                  * we are forcing write access.
982                                  */
983                                 lookup_write = write && !force;
984                                 spin_lock(&mm->page_table_lock);
985                         }
986                         if (pages) {
987                                 pages[i] = get_page_map(map);
988                                 if (!pages[i]) {
989                                         spin_unlock(&mm->page_table_lock);
990                                         while (i--)
991                                                 page_cache_release(pages[i]);
992                                         i = -EFAULT;
993                                         goto out;
994                                 }
995                                 flush_dcache_page(pages[i]);
996                                 if (!PageReserved(pages[i]))
997                                         page_cache_get(pages[i]);
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 out:
1008         return i;
1009 }
1010
1011 EXPORT_SYMBOL(get_user_pages);
1012
1013 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1014                         unsigned long addr, unsigned long end, pgprot_t prot)
1015 {
1016         pte_t *pte;
1017
1018         pte = pte_alloc_map(mm, pmd, addr);
1019         if (!pte)
1020                 return -ENOMEM;
1021         do {
1022                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(addr), prot));
1023                 BUG_ON(!pte_none(*pte));
1024                 set_pte_at(mm, addr, pte, zero_pte);
1025         } while (pte++, addr += PAGE_SIZE, addr != end);
1026         pte_unmap(pte - 1);
1027         return 0;
1028 }
1029
1030 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1031                         unsigned long addr, unsigned long end, pgprot_t prot)
1032 {
1033         pmd_t *pmd;
1034         unsigned long next;
1035
1036         pmd = pmd_alloc(mm, pud, addr);
1037         if (!pmd)
1038                 return -ENOMEM;
1039         do {
1040                 next = pmd_addr_end(addr, end);
1041                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1042                         return -ENOMEM;
1043         } while (pmd++, addr = next, addr != end);
1044         return 0;
1045 }
1046
1047 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1048                         unsigned long addr, unsigned long end, pgprot_t prot)
1049 {
1050         pud_t *pud;
1051         unsigned long next;
1052
1053         pud = pud_alloc(mm, pgd, addr);
1054         if (!pud)
1055                 return -ENOMEM;
1056         do {
1057                 next = pud_addr_end(addr, end);
1058                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1059                         return -ENOMEM;
1060         } while (pud++, addr = next, addr != end);
1061         return 0;
1062 }
1063
1064 int zeromap_page_range(struct vm_area_struct *vma,
1065                         unsigned long addr, unsigned long size, pgprot_t prot)
1066 {
1067         pgd_t *pgd;
1068         unsigned long next;
1069         unsigned long end = addr + size;
1070         struct mm_struct *mm = vma->vm_mm;
1071         int err;
1072
1073         BUG_ON(addr >= end);
1074         pgd = pgd_offset(mm, addr);
1075         flush_cache_range(vma, addr, end);
1076         spin_lock(&mm->page_table_lock);
1077         do {
1078                 next = pgd_addr_end(addr, end);
1079                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1080                 if (err)
1081                         break;
1082         } while (pgd++, addr = next, addr != end);
1083         spin_unlock(&mm->page_table_lock);
1084         return err;
1085 }
1086
1087 /*
1088  * maps a range of physical memory into the requested pages. the old
1089  * mappings are removed. any references to nonexistent pages results
1090  * in null mappings (currently treated as "copy-on-access")
1091  */
1092 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1093                         unsigned long addr, unsigned long end,
1094                         unsigned long pfn, pgprot_t prot)
1095 {
1096         pte_t *pte;
1097
1098         pte = pte_alloc_map(mm, pmd, addr);
1099         if (!pte)
1100                 return -ENOMEM;
1101         do {
1102                 BUG_ON(!pte_none(*pte));
1103                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
1104                         set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1105                 pfn++;
1106         } while (pte++, addr += PAGE_SIZE, addr != end);
1107         pte_unmap(pte - 1);
1108         return 0;
1109 }
1110
1111 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1112                         unsigned long addr, unsigned long end,
1113                         unsigned long pfn, pgprot_t prot)
1114 {
1115         pmd_t *pmd;
1116         unsigned long next;
1117
1118         pfn -= addr >> PAGE_SHIFT;
1119         pmd = pmd_alloc(mm, pud, addr);
1120         if (!pmd)
1121                 return -ENOMEM;
1122         do {
1123                 next = pmd_addr_end(addr, end);
1124                 if (remap_pte_range(mm, pmd, addr, next,
1125                                 pfn + (addr >> PAGE_SHIFT), prot))
1126                         return -ENOMEM;
1127         } while (pmd++, addr = next, addr != end);
1128         return 0;
1129 }
1130
1131 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1132                         unsigned long addr, unsigned long end,
1133                         unsigned long pfn, pgprot_t prot)
1134 {
1135         pud_t *pud;
1136         unsigned long next;
1137
1138         pfn -= addr >> PAGE_SHIFT;
1139         pud = pud_alloc(mm, pgd, addr);
1140         if (!pud)
1141                 return -ENOMEM;
1142         do {
1143                 next = pud_addr_end(addr, end);
1144                 if (remap_pmd_range(mm, pud, addr, next,
1145                                 pfn + (addr >> PAGE_SHIFT), prot))
1146                         return -ENOMEM;
1147         } while (pud++, addr = next, addr != end);
1148         return 0;
1149 }
1150
1151 /*  Note: this is only safe if the mm semaphore is held when called. */
1152 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1153                     unsigned long pfn, unsigned long size, pgprot_t prot)
1154 {
1155         pgd_t *pgd;
1156         unsigned long next;
1157         unsigned long end = addr + size;
1158         struct mm_struct *mm = vma->vm_mm;
1159         int err;
1160
1161         /*
1162          * Physically remapped pages are special. Tell the
1163          * rest of the world about it:
1164          *   VM_IO tells people not to look at these pages
1165          *      (accesses can have side effects).
1166          *   VM_RESERVED tells swapout not to try to touch
1167          *      this region.
1168          */
1169         vma->vm_flags |= VM_IO | VM_RESERVED;
1170
1171         BUG_ON(addr >= end);
1172         pfn -= addr >> PAGE_SHIFT;
1173         pgd = pgd_offset(mm, addr);
1174         flush_cache_range(vma, addr, end);
1175         spin_lock(&mm->page_table_lock);
1176         do {
1177                 next = pgd_addr_end(addr, end);
1178                 err = remap_pud_range(mm, pgd, addr, next,
1179                                 pfn + (addr >> PAGE_SHIFT), prot);
1180                 if (err)
1181                         break;
1182         } while (pgd++, addr = next, addr != end);
1183         spin_unlock(&mm->page_table_lock);
1184         return err;
1185 }
1186 EXPORT_SYMBOL(remap_pfn_range);
1187
1188 /*
1189  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1190  * servicing faults for write access.  In the normal case, do always want
1191  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1192  * that do not have writing enabled, when used by access_process_vm.
1193  */
1194 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1195 {
1196         if (likely(vma->vm_flags & VM_WRITE))
1197                 pte = pte_mkwrite(pte);
1198         return pte;
1199 }
1200
1201 /*
1202  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1203  */
1204 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address, 
1205                 pte_t *page_table)
1206 {
1207         pte_t entry;
1208
1209         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1210                               vma);
1211         ptep_establish(vma, address, page_table, entry);
1212         update_mmu_cache(vma, address, entry);
1213         lazy_mmu_prot_update(entry);
1214 }
1215
1216 /*
1217  * This routine handles present pages, when users try to write
1218  * to a shared page. It is done by copying the page to a new address
1219  * and decrementing the shared-page counter for the old page.
1220  *
1221  * Goto-purists beware: the only reason for goto's here is that it results
1222  * in better assembly code.. The "default" path will see no jumps at all.
1223  *
1224  * Note that this routine assumes that the protection checks have been
1225  * done by the caller (the low-level page fault routine in most cases).
1226  * Thus we can safely just mark it writable once we've done any necessary
1227  * COW.
1228  *
1229  * We also mark the page dirty at this point even though the page will
1230  * change only once the write actually happens. This avoids a few races,
1231  * and potentially makes it more efficient.
1232  *
1233  * We hold the mm semaphore and the page_table_lock on entry and exit
1234  * with the page_table_lock released.
1235  */
1236 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1237         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1238 {
1239         struct page *old_page, *new_page;
1240         unsigned long pfn = pte_pfn(pte);
1241         pte_t entry;
1242
1243         if (unlikely(!pfn_valid(pfn))) {
1244                 /*
1245                  * This should really halt the system so it can be debugged or
1246                  * at least the kernel stops what it's doing before it corrupts
1247                  * data, but for the moment just pretend this is OOM.
1248                  */
1249                 pte_unmap(page_table);
1250                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1251                                 address);
1252                 spin_unlock(&mm->page_table_lock);
1253                 return VM_FAULT_OOM;
1254         }
1255         old_page = pfn_to_page(pfn);
1256
1257         if (!TestSetPageLocked(old_page)) {
1258                 int reuse = can_share_swap_page(old_page);
1259                 unlock_page(old_page);
1260                 if (reuse) {
1261                         flush_cache_page(vma, address, pfn);
1262                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1263                                               vma);
1264                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1265                         update_mmu_cache(vma, address, entry);
1266                         lazy_mmu_prot_update(entry);
1267                         pte_unmap(page_table);
1268                         spin_unlock(&mm->page_table_lock);
1269                         return VM_FAULT_MINOR;
1270                 }
1271         }
1272         pte_unmap(page_table);
1273
1274         /*
1275          * Ok, we need to copy. Oh, well..
1276          */
1277         if (!PageReserved(old_page))
1278                 page_cache_get(old_page);
1279         spin_unlock(&mm->page_table_lock);
1280
1281         if (unlikely(anon_vma_prepare(vma)))
1282                 goto no_new_page;
1283         if (old_page == ZERO_PAGE(address)) {
1284                 new_page = alloc_zeroed_user_highpage(vma, address);
1285                 if (!new_page)
1286                         goto no_new_page;
1287         } else {
1288                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1289                 if (!new_page)
1290                         goto no_new_page;
1291                 copy_user_highpage(new_page, old_page, address);
1292         }
1293         /*
1294          * Re-check the pte - we dropped the lock
1295          */
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         }
1313         pte_unmap(page_table);
1314         page_cache_release(new_page);
1315         page_cache_release(old_page);
1316         spin_unlock(&mm->page_table_lock);
1317         return VM_FAULT_MINOR;
1318
1319 no_new_page:
1320         page_cache_release(old_page);
1321         return VM_FAULT_OOM;
1322 }
1323
1324 /*
1325  * Helper functions for unmap_mapping_range().
1326  *
1327  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1328  *
1329  * We have to restart searching the prio_tree whenever we drop the lock,
1330  * since the iterator is only valid while the lock is held, and anyway
1331  * a later vma might be split and reinserted earlier while lock dropped.
1332  *
1333  * The list of nonlinear vmas could be handled more efficiently, using
1334  * a placeholder, but handle it in the same way until a need is shown.
1335  * It is important to search the prio_tree before nonlinear list: a vma
1336  * may become nonlinear and be shifted from prio_tree to nonlinear list
1337  * while the lock is dropped; but never shifted from list to prio_tree.
1338  *
1339  * In order to make forward progress despite restarting the search,
1340  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1341  * quickly skip it next time around.  Since the prio_tree search only
1342  * shows us those vmas affected by unmapping the range in question, we
1343  * can't efficiently keep all vmas in step with mapping->truncate_count:
1344  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1345  * mapping->truncate_count and vma->vm_truncate_count are protected by
1346  * i_mmap_lock.
1347  *
1348  * In order to make forward progress despite repeatedly restarting some
1349  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1350  * and restart from that address when we reach that vma again.  It might
1351  * have been split or merged, shrunk or extended, but never shifted: so
1352  * restart_addr remains valid so long as it remains in the vma's range.
1353  * unmap_mapping_range forces truncate_count to leap over page-aligned
1354  * values so we can save vma's restart_addr in its truncate_count field.
1355  */
1356 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1357
1358 static void reset_vma_truncate_counts(struct address_space *mapping)
1359 {
1360         struct vm_area_struct *vma;
1361         struct prio_tree_iter iter;
1362
1363         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1364                 vma->vm_truncate_count = 0;
1365         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1366                 vma->vm_truncate_count = 0;
1367 }
1368
1369 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1370                 unsigned long start_addr, unsigned long end_addr,
1371                 struct zap_details *details)
1372 {
1373         unsigned long restart_addr;
1374         int need_break;
1375
1376 again:
1377         restart_addr = vma->vm_truncate_count;
1378         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1379                 start_addr = restart_addr;
1380                 if (start_addr >= end_addr) {
1381                         /* Top of vma has been split off since last time */
1382                         vma->vm_truncate_count = details->truncate_count;
1383                         return 0;
1384                 }
1385         }
1386
1387         restart_addr = zap_page_range(vma, start_addr,
1388                                         end_addr - start_addr, details);
1389
1390         /*
1391          * We cannot rely on the break test in unmap_vmas:
1392          * on the one hand, we don't want to restart our loop
1393          * just because that broke out for the page_table_lock;
1394          * on the other hand, it does no test when vma is small.
1395          */
1396         need_break = need_resched() ||
1397                         need_lockbreak(details->i_mmap_lock);
1398
1399         if (restart_addr >= end_addr) {
1400                 /* We have now completed this vma: mark it so */
1401                 vma->vm_truncate_count = details->truncate_count;
1402                 if (!need_break)
1403                         return 0;
1404         } else {
1405                 /* Note restart_addr in vma's truncate_count field */
1406                 vma->vm_truncate_count = restart_addr;
1407                 if (!need_break)
1408                         goto again;
1409         }
1410
1411         spin_unlock(details->i_mmap_lock);
1412         cond_resched();
1413         spin_lock(details->i_mmap_lock);
1414         return -EINTR;
1415 }
1416
1417 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1418                                             struct zap_details *details)
1419 {
1420         struct vm_area_struct *vma;
1421         struct prio_tree_iter iter;
1422         pgoff_t vba, vea, zba, zea;
1423
1424 restart:
1425         vma_prio_tree_foreach(vma, &iter, root,
1426                         details->first_index, details->last_index) {
1427                 /* Skip quickly over those we have already dealt with */
1428                 if (vma->vm_truncate_count == details->truncate_count)
1429                         continue;
1430
1431                 vba = vma->vm_pgoff;
1432                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1433                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1434                 zba = details->first_index;
1435                 if (zba < vba)
1436                         zba = vba;
1437                 zea = details->last_index;
1438                 if (zea > vea)
1439                         zea = vea;
1440
1441                 if (unmap_mapping_range_vma(vma,
1442                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1443                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1444                                 details) < 0)
1445                         goto restart;
1446         }
1447 }
1448
1449 static inline void unmap_mapping_range_list(struct list_head *head,
1450                                             struct zap_details *details)
1451 {
1452         struct vm_area_struct *vma;
1453
1454         /*
1455          * In nonlinear VMAs there is no correspondence between virtual address
1456          * offset and file offset.  So we must perform an exhaustive search
1457          * across *all* the pages in each nonlinear VMA, not just the pages
1458          * whose virtual address lies outside the file truncation point.
1459          */
1460 restart:
1461         list_for_each_entry(vma, head, shared.vm_set.list) {
1462                 /* Skip quickly over those we have already dealt with */
1463                 if (vma->vm_truncate_count == details->truncate_count)
1464                         continue;
1465                 details->nonlinear_vma = vma;
1466                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1467                                         vma->vm_end, details) < 0)
1468                         goto restart;
1469         }
1470 }
1471
1472 /**
1473  * unmap_mapping_range - unmap the portion of all mmaps
1474  * in the specified address_space corresponding to the specified
1475  * page range in the underlying file.
1476  * @address_space: the address space containing mmaps to be unmapped.
1477  * @holebegin: byte in first page to unmap, relative to the start of
1478  * the underlying file.  This will be rounded down to a PAGE_SIZE
1479  * boundary.  Note that this is different from vmtruncate(), which
1480  * must keep the partial page.  In contrast, we must get rid of
1481  * partial pages.
1482  * @holelen: size of prospective hole in bytes.  This will be rounded
1483  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1484  * end of the file.
1485  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1486  * but 0 when invalidating pagecache, don't throw away private data.
1487  */
1488 void unmap_mapping_range(struct address_space *mapping,
1489                 loff_t const holebegin, loff_t const holelen, int even_cows)
1490 {
1491         struct zap_details details;
1492         pgoff_t hba = holebegin >> PAGE_SHIFT;
1493         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1494
1495         /* Check for overflow. */
1496         if (sizeof(holelen) > sizeof(hlen)) {
1497                 long long holeend =
1498                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1499                 if (holeend & ~(long long)ULONG_MAX)
1500                         hlen = ULONG_MAX - hba + 1;
1501         }
1502
1503         details.check_mapping = even_cows? NULL: mapping;
1504         details.nonlinear_vma = NULL;
1505         details.first_index = hba;
1506         details.last_index = hba + hlen - 1;
1507         if (details.last_index < details.first_index)
1508                 details.last_index = ULONG_MAX;
1509         details.i_mmap_lock = &mapping->i_mmap_lock;
1510
1511         spin_lock(&mapping->i_mmap_lock);
1512
1513         /* serialize i_size write against truncate_count write */
1514         smp_wmb();
1515         /* Protect against page faults, and endless unmapping loops */
1516         mapping->truncate_count++;
1517         /*
1518          * For archs where spin_lock has inclusive semantics like ia64
1519          * this smp_mb() will prevent to read pagetable contents
1520          * before the truncate_count increment is visible to
1521          * other cpus.
1522          */
1523         smp_mb();
1524         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1525                 if (mapping->truncate_count == 0)
1526                         reset_vma_truncate_counts(mapping);
1527                 mapping->truncate_count++;
1528         }
1529         details.truncate_count = mapping->truncate_count;
1530
1531         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1532                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1533         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1534                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1535         spin_unlock(&mapping->i_mmap_lock);
1536 }
1537 EXPORT_SYMBOL(unmap_mapping_range);
1538
1539 /*
1540  * Handle all mappings that got truncated by a "truncate()"
1541  * system call.
1542  *
1543  * NOTE! We have to be ready to update the memory sharing
1544  * between the file and the memory map for a potential last
1545  * incomplete page.  Ugly, but necessary.
1546  */
1547 int vmtruncate(struct inode * inode, loff_t offset)
1548 {
1549         struct address_space *mapping = inode->i_mapping;
1550         unsigned long limit;
1551
1552         if (inode->i_size < offset)
1553                 goto do_expand;
1554         /*
1555          * truncation of in-use swapfiles is disallowed - it would cause
1556          * subsequent swapout to scribble on the now-freed blocks.
1557          */
1558         if (IS_SWAPFILE(inode))
1559                 goto out_busy;
1560         i_size_write(inode, offset);
1561         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1562         truncate_inode_pages(mapping, offset);
1563         goto out_truncate;
1564
1565 do_expand:
1566         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1567         if (limit != RLIM_INFINITY && offset > limit)
1568                 goto out_sig;
1569         if (offset > inode->i_sb->s_maxbytes)
1570                 goto out_big;
1571         i_size_write(inode, offset);
1572
1573 out_truncate:
1574         if (inode->i_op && inode->i_op->truncate)
1575                 inode->i_op->truncate(inode);
1576         return 0;
1577 out_sig:
1578         send_sig(SIGXFSZ, current, 0);
1579 out_big:
1580         return -EFBIG;
1581 out_busy:
1582         return -ETXTBSY;
1583 }
1584
1585 EXPORT_SYMBOL(vmtruncate);
1586
1587 /* 
1588  * Primitive swap readahead code. We simply read an aligned block of
1589  * (1 << page_cluster) entries in the swap area. This method is chosen
1590  * because it doesn't cost us any seek time.  We also make sure to queue
1591  * the 'original' request together with the readahead ones...  
1592  *
1593  * This has been extended to use the NUMA policies from the mm triggering
1594  * the readahead.
1595  *
1596  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1597  */
1598 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1599 {
1600 #ifdef CONFIG_NUMA
1601         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1602 #endif
1603         int i, num;
1604         struct page *new_page;
1605         unsigned long offset;
1606
1607         /*
1608          * Get the number of handles we should do readahead io to.
1609          */
1610         num = valid_swaphandles(entry, &offset);
1611         for (i = 0; i < num; offset++, i++) {
1612                 /* Ok, do the async read-ahead now */
1613                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1614                                                            offset), vma, addr);
1615                 if (!new_page)
1616                         break;
1617                 page_cache_release(new_page);
1618 #ifdef CONFIG_NUMA
1619                 /*
1620                  * Find the next applicable VMA for the NUMA policy.
1621                  */
1622                 addr += PAGE_SIZE;
1623                 if (addr == 0)
1624                         vma = NULL;
1625                 if (vma) {
1626                         if (addr >= vma->vm_end) {
1627                                 vma = next_vma;
1628                                 next_vma = vma ? vma->vm_next : NULL;
1629                         }
1630                         if (vma && addr < vma->vm_start)
1631                                 vma = NULL;
1632                 } else {
1633                         if (next_vma && addr >= next_vma->vm_start) {
1634                                 vma = next_vma;
1635                                 next_vma = vma->vm_next;
1636                         }
1637                 }
1638 #endif
1639         }
1640         lru_add_drain();        /* Push any new pages onto the LRU now */
1641 }
1642
1643 /*
1644  * We hold the mm semaphore and the page_table_lock on entry and
1645  * should release the pagetable lock on exit..
1646  */
1647 static int do_swap_page(struct mm_struct * mm,
1648         struct vm_area_struct * vma, unsigned long address,
1649         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1650 {
1651         struct page *page;
1652         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1653         pte_t pte;
1654         int ret = VM_FAULT_MINOR;
1655
1656         pte_unmap(page_table);
1657         spin_unlock(&mm->page_table_lock);
1658         page = lookup_swap_cache(entry);
1659         if (!page) {
1660                 swapin_readahead(entry, address, vma);
1661                 page = read_swap_cache_async(entry, vma, address);
1662                 if (!page) {
1663                         /*
1664                          * Back out if somebody else faulted in this pte while
1665                          * we released the page table lock.
1666                          */
1667                         spin_lock(&mm->page_table_lock);
1668                         page_table = pte_offset_map(pmd, address);
1669                         if (likely(pte_same(*page_table, orig_pte)))
1670                                 ret = VM_FAULT_OOM;
1671                         else
1672                                 ret = VM_FAULT_MINOR;
1673                         pte_unmap(page_table);
1674                         spin_unlock(&mm->page_table_lock);
1675                         goto out;
1676                 }
1677
1678                 /* Had to read the page from swap area: Major fault */
1679                 ret = VM_FAULT_MAJOR;
1680                 inc_page_state(pgmajfault);
1681                 grab_swap_token();
1682         }
1683
1684         mark_page_accessed(page);
1685         lock_page(page);
1686
1687         /*
1688          * Back out if somebody else faulted in this pte while we
1689          * released the page table lock.
1690          */
1691         spin_lock(&mm->page_table_lock);
1692         page_table = pte_offset_map(pmd, address);
1693         if (unlikely(!pte_same(*page_table, orig_pte))) {
1694                 pte_unmap(page_table);
1695                 spin_unlock(&mm->page_table_lock);
1696                 unlock_page(page);
1697                 page_cache_release(page);
1698                 ret = VM_FAULT_MINOR;
1699                 goto out;
1700         }
1701
1702         /* The page isn't present yet, go ahead with the fault. */
1703                 
1704         swap_free(entry);
1705         if (vm_swap_full())
1706                 remove_exclusive_swap_page(page);
1707
1708         inc_mm_counter(mm, rss);
1709         pte = mk_pte(page, vma->vm_page_prot);
1710         if (write_access && can_share_swap_page(page)) {
1711                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1712                 write_access = 0;
1713         }
1714         unlock_page(page);
1715
1716         flush_icache_page(vma, page);
1717         set_pte_at(mm, address, page_table, pte);
1718         page_add_anon_rmap(page, vma, address);
1719
1720         if (write_access) {
1721                 if (do_wp_page(mm, vma, address,
1722                                 page_table, pmd, pte) == VM_FAULT_OOM)
1723                         ret = VM_FAULT_OOM;
1724                 goto out;
1725         }
1726
1727         /* No need to invalidate - it was non-present before */
1728         update_mmu_cache(vma, address, pte);
1729         lazy_mmu_prot_update(pte);
1730         pte_unmap(page_table);
1731         spin_unlock(&mm->page_table_lock);
1732 out:
1733         return ret;
1734 }
1735
1736 /*
1737  * We are called with the MM semaphore and page_table_lock
1738  * spinlock held to protect against concurrent faults in
1739  * multithreaded programs. 
1740  */
1741 static int
1742 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1743                 pte_t *page_table, pmd_t *pmd, int write_access,
1744                 unsigned long addr)
1745 {
1746         pte_t entry;
1747         struct page * page = ZERO_PAGE(addr);
1748
1749         /* Read-only mapping of ZERO_PAGE. */
1750         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1751
1752         /* ..except if it's a write access */
1753         if (write_access) {
1754                 /* Allocate our own private page. */
1755                 pte_unmap(page_table);
1756                 spin_unlock(&mm->page_table_lock);
1757
1758                 if (unlikely(anon_vma_prepare(vma)))
1759                         goto no_mem;
1760                 page = alloc_zeroed_user_highpage(vma, addr);
1761                 if (!page)
1762                         goto no_mem;
1763
1764                 spin_lock(&mm->page_table_lock);
1765                 page_table = pte_offset_map(pmd, addr);
1766
1767                 if (!pte_none(*page_table)) {
1768                         pte_unmap(page_table);
1769                         page_cache_release(page);
1770                         spin_unlock(&mm->page_table_lock);
1771                         goto out;
1772                 }
1773                 inc_mm_counter(mm, rss);
1774                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1775                                                          vma->vm_page_prot)),
1776                                       vma);
1777                 lru_cache_add_active(page);
1778                 SetPageReferenced(page);
1779                 page_add_anon_rmap(page, vma, addr);
1780         }
1781
1782         set_pte_at(mm, addr, page_table, entry);
1783         pte_unmap(page_table);
1784
1785         /* No need to invalidate - it was non-present before */
1786         update_mmu_cache(vma, addr, entry);
1787         lazy_mmu_prot_update(entry);
1788         spin_unlock(&mm->page_table_lock);
1789 out:
1790         return VM_FAULT_MINOR;
1791 no_mem:
1792         return VM_FAULT_OOM;
1793 }
1794
1795 /*
1796  * do_no_page() tries to create a new page mapping. It aggressively
1797  * tries to share with existing pages, but makes a separate copy if
1798  * the "write_access" parameter is true in order to avoid the next
1799  * page fault.
1800  *
1801  * As this is called only for pages that do not currently exist, we
1802  * do not need to flush old virtual caches or the TLB.
1803  *
1804  * This is called with the MM semaphore held and the page table
1805  * spinlock held. Exit with the spinlock released.
1806  */
1807 static int
1808 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1809         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1810 {
1811         struct page * new_page;
1812         struct address_space *mapping = NULL;
1813         pte_t entry;
1814         unsigned int sequence = 0;
1815         int ret = VM_FAULT_MINOR;
1816         int anon = 0;
1817
1818         if (!vma->vm_ops || !vma->vm_ops->nopage)
1819                 return do_anonymous_page(mm, vma, page_table,
1820                                         pmd, write_access, address);
1821         pte_unmap(page_table);
1822         spin_unlock(&mm->page_table_lock);
1823
1824         if (vma->vm_file) {
1825                 mapping = vma->vm_file->f_mapping;
1826                 sequence = mapping->truncate_count;
1827                 smp_rmb(); /* serializes i_size against truncate_count */
1828         }
1829 retry:
1830         cond_resched();
1831         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1832         /*
1833          * No smp_rmb is needed here as long as there's a full
1834          * spin_lock/unlock sequence inside the ->nopage callback
1835          * (for the pagecache lookup) that acts as an implicit
1836          * smp_mb() and prevents the i_size read to happen
1837          * after the next truncate_count read.
1838          */
1839
1840         /* no page was available -- either SIGBUS or OOM */
1841         if (new_page == NOPAGE_SIGBUS)
1842                 return VM_FAULT_SIGBUS;
1843         if (new_page == NOPAGE_OOM)
1844                 return VM_FAULT_OOM;
1845
1846         /*
1847          * Should we do an early C-O-W break?
1848          */
1849         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1850                 struct page *page;
1851
1852                 if (unlikely(anon_vma_prepare(vma)))
1853                         goto oom;
1854                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1855                 if (!page)
1856                         goto oom;
1857                 copy_user_highpage(page, new_page, address);
1858                 page_cache_release(new_page);
1859                 new_page = page;
1860                 anon = 1;
1861         }
1862
1863         spin_lock(&mm->page_table_lock);
1864         /*
1865          * For a file-backed vma, someone could have truncated or otherwise
1866          * invalidated this page.  If unmap_mapping_range got called,
1867          * retry getting the page.
1868          */
1869         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1870                 sequence = mapping->truncate_count;
1871                 spin_unlock(&mm->page_table_lock);
1872                 page_cache_release(new_page);
1873                 goto retry;
1874         }
1875         page_table = pte_offset_map(pmd, address);
1876
1877         /*
1878          * This silly early PAGE_DIRTY setting removes a race
1879          * due to the bad i386 page protection. But it's valid
1880          * for other architectures too.
1881          *
1882          * Note that if write_access is true, we either now have
1883          * an exclusive copy of the page, or this is a shared mapping,
1884          * so we can make it writable and dirty to avoid having to
1885          * handle that later.
1886          */
1887         /* Only go through if we didn't race with anybody else... */
1888         if (pte_none(*page_table)) {
1889                 if (!PageReserved(new_page))
1890                         inc_mm_counter(mm, rss);
1891
1892                 flush_icache_page(vma, new_page);
1893                 entry = mk_pte(new_page, vma->vm_page_prot);
1894                 if (write_access)
1895                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1896                 set_pte_at(mm, address, page_table, entry);
1897                 if (anon) {
1898                         lru_cache_add_active(new_page);
1899                         page_add_anon_rmap(new_page, vma, address);
1900                 } else
1901                         page_add_file_rmap(new_page);
1902                 pte_unmap(page_table);
1903         } else {
1904                 /* One of our sibling threads was faster, back out. */
1905                 pte_unmap(page_table);
1906                 page_cache_release(new_page);
1907                 spin_unlock(&mm->page_table_lock);
1908                 goto out;
1909         }
1910
1911         /* no need to invalidate: a not-present page shouldn't be cached */
1912         update_mmu_cache(vma, address, entry);
1913         lazy_mmu_prot_update(entry);
1914         spin_unlock(&mm->page_table_lock);
1915 out:
1916         return ret;
1917 oom:
1918         page_cache_release(new_page);
1919         ret = VM_FAULT_OOM;
1920         goto out;
1921 }
1922
1923 /*
1924  * Fault of a previously existing named mapping. Repopulate the pte
1925  * from the encoded file_pte if possible. This enables swappable
1926  * nonlinear vmas.
1927  */
1928 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1929         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1930 {
1931         unsigned long pgoff;
1932         int err;
1933
1934         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1935         /*
1936          * Fall back to the linear mapping if the fs does not support
1937          * ->populate:
1938          */
1939         if (!vma->vm_ops || !vma->vm_ops->populate || 
1940                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1941                 pte_clear(mm, address, pte);
1942                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1943         }
1944
1945         pgoff = pte_to_pgoff(*pte);
1946
1947         pte_unmap(pte);
1948         spin_unlock(&mm->page_table_lock);
1949
1950         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1951         if (err == -ENOMEM)
1952                 return VM_FAULT_OOM;
1953         if (err)
1954                 return VM_FAULT_SIGBUS;
1955         return VM_FAULT_MAJOR;
1956 }
1957
1958 /*
1959  * These routines also need to handle stuff like marking pages dirty
1960  * and/or accessed for architectures that don't do it in hardware (most
1961  * RISC architectures).  The early dirtying is also good on the i386.
1962  *
1963  * There is also a hook called "update_mmu_cache()" that architectures
1964  * with external mmu caches can use to update those (ie the Sparc or
1965  * PowerPC hashed page tables that act as extended TLBs).
1966  *
1967  * Note the "page_table_lock". It is to protect against kswapd removing
1968  * pages from under us. Note that kswapd only ever _removes_ pages, never
1969  * adds them. As such, once we have noticed that the page is not present,
1970  * we can drop the lock early.
1971  *
1972  * The adding of pages is protected by the MM semaphore (which we hold),
1973  * so we don't need to worry about a page being suddenly been added into
1974  * our VM.
1975  *
1976  * We enter with the pagetable spinlock held, we are supposed to
1977  * release it when done.
1978  */
1979 static inline int handle_pte_fault(struct mm_struct *mm,
1980         struct vm_area_struct * vma, unsigned long address,
1981         int write_access, pte_t *pte, pmd_t *pmd)
1982 {
1983         pte_t entry;
1984
1985         entry = *pte;
1986         if (!pte_present(entry)) {
1987                 /*
1988                  * If it truly wasn't present, we know that kswapd
1989                  * and the PTE updates will not touch it later. So
1990                  * drop the lock.
1991                  */
1992                 if (pte_none(entry))
1993                         return do_no_page(mm, vma, address, write_access, pte, pmd);
1994                 if (pte_file(entry))
1995                         return do_file_page(mm, vma, address, write_access, pte, pmd);
1996                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1997         }
1998
1999         if (write_access) {
2000                 if (!pte_write(entry))
2001                         return do_wp_page(mm, vma, address, pte, pmd, entry);
2002
2003                 entry = pte_mkdirty(entry);
2004         }
2005         entry = pte_mkyoung(entry);
2006         ptep_set_access_flags(vma, address, pte, entry, write_access);
2007         update_mmu_cache(vma, address, entry);
2008         lazy_mmu_prot_update(entry);
2009         pte_unmap(pte);
2010         spin_unlock(&mm->page_table_lock);
2011         return VM_FAULT_MINOR;
2012 }
2013
2014 /*
2015  * By the time we get here, we already hold the mm semaphore
2016  */
2017 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
2018                 unsigned long address, int write_access)
2019 {
2020         pgd_t *pgd;
2021         pud_t *pud;
2022         pmd_t *pmd;
2023         pte_t *pte;
2024
2025         __set_current_state(TASK_RUNNING);
2026
2027         inc_page_state(pgfault);
2028
2029         if (is_vm_hugetlb_page(vma))
2030                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
2031
2032         /*
2033          * We need the page table lock to synchronize with kswapd
2034          * and the SMP-safe atomic PTE updates.
2035          */
2036         pgd = pgd_offset(mm, address);
2037         spin_lock(&mm->page_table_lock);
2038
2039         pud = pud_alloc(mm, pgd, address);
2040         if (!pud)
2041                 goto oom;
2042
2043         pmd = pmd_alloc(mm, pud, address);
2044         if (!pmd)
2045                 goto oom;
2046
2047         pte = pte_alloc_map(mm, pmd, address);
2048         if (!pte)
2049                 goto oom;
2050         
2051         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
2052
2053  oom:
2054         spin_unlock(&mm->page_table_lock);
2055         return VM_FAULT_OOM;
2056 }
2057
2058 #ifndef __PAGETABLE_PUD_FOLDED
2059 /*
2060  * Allocate page upper directory.
2061  *
2062  * We've already handled the fast-path in-line, and we own the
2063  * page table lock.
2064  */
2065 pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2066 {
2067         pud_t *new;
2068
2069         spin_unlock(&mm->page_table_lock);
2070         new = pud_alloc_one(mm, address);
2071         spin_lock(&mm->page_table_lock);
2072         if (!new)
2073                 return NULL;
2074
2075         /*
2076          * Because we dropped the lock, we should re-check the
2077          * entry, as somebody else could have populated it..
2078          */
2079         if (pgd_present(*pgd)) {
2080                 pud_free(new);
2081                 goto out;
2082         }
2083         pgd_populate(mm, pgd, new);
2084  out:
2085         return pud_offset(pgd, address);
2086 }
2087 #endif /* __PAGETABLE_PUD_FOLDED */
2088
2089 #ifndef __PAGETABLE_PMD_FOLDED
2090 /*
2091  * Allocate page middle directory.
2092  *
2093  * We've already handled the fast-path in-line, and we own the
2094  * page table lock.
2095  */
2096 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2097 {
2098         pmd_t *new;
2099
2100         spin_unlock(&mm->page_table_lock);
2101         new = pmd_alloc_one(mm, address);
2102         spin_lock(&mm->page_table_lock);
2103         if (!new)
2104                 return NULL;
2105
2106         /*
2107          * Because we dropped the lock, we should re-check the
2108          * entry, as somebody else could have populated it..
2109          */
2110 #ifndef __ARCH_HAS_4LEVEL_HACK
2111         if (pud_present(*pud)) {
2112                 pmd_free(new);
2113                 goto out;
2114         }
2115         pud_populate(mm, pud, new);
2116 #else
2117         if (pgd_present(*pud)) {
2118                 pmd_free(new);
2119                 goto out;
2120         }
2121         pgd_populate(mm, pud, new);
2122 #endif /* __ARCH_HAS_4LEVEL_HACK */
2123
2124  out:
2125         return pmd_offset(pud, address);
2126 }
2127 #endif /* __PAGETABLE_PMD_FOLDED */
2128
2129 int make_pages_present(unsigned long addr, unsigned long end)
2130 {
2131         int ret, len, write;
2132         struct vm_area_struct * vma;
2133
2134         vma = find_vma(current->mm, addr);
2135         if (!vma)
2136                 return -1;
2137         write = (vma->vm_flags & VM_WRITE) != 0;
2138         if (addr >= end)
2139                 BUG();
2140         if (end > vma->vm_end)
2141                 BUG();
2142         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2143         ret = get_user_pages(current, current->mm, addr,
2144                         len, write, 0, NULL, NULL);
2145         if (ret < 0)
2146                 return ret;
2147         return ret == len ? 0 : -1;
2148 }
2149
2150 /* 
2151  * Map a vmalloc()-space virtual address to the physical page.
2152  */
2153 struct page * vmalloc_to_page(void * vmalloc_addr)
2154 {
2155         unsigned long addr = (unsigned long) vmalloc_addr;
2156         struct page *page = NULL;
2157         pgd_t *pgd = pgd_offset_k(addr);
2158         pud_t *pud;
2159         pmd_t *pmd;
2160         pte_t *ptep, pte;
2161   
2162         if (!pgd_none(*pgd)) {
2163                 pud = pud_offset(pgd, addr);
2164                 if (!pud_none(*pud)) {
2165                         pmd = pmd_offset(pud, addr);
2166                         if (!pmd_none(*pmd)) {
2167                                 ptep = pte_offset_map(pmd, addr);
2168                                 pte = *ptep;
2169                                 if (pte_present(pte))
2170                                         page = pte_page(pte);
2171                                 pte_unmap(ptep);
2172                         }
2173                 }
2174         }
2175         return page;
2176 }
2177
2178 EXPORT_SYMBOL(vmalloc_to_page);
2179
2180 /*
2181  * Map a vmalloc()-space virtual address to the physical page frame number.
2182  */
2183 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2184 {
2185         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2186 }
2187
2188 EXPORT_SYMBOL(vmalloc_to_pfn);
2189
2190 /*
2191  * update_mem_hiwater
2192  *      - update per process rss and vm high water data
2193  */
2194 void update_mem_hiwater(struct task_struct *tsk)
2195 {
2196         if (tsk->mm) {
2197                 unsigned long rss = get_mm_counter(tsk->mm, rss);
2198
2199                 if (tsk->mm->hiwater_rss < rss)
2200                         tsk->mm->hiwater_rss = rss;
2201                 if (tsk->mm->hiwater_vm < tsk->mm->total_vm)
2202                         tsk->mm->hiwater_vm = tsk->mm->total_vm;
2203         }
2204 }
2205
2206 #if !defined(__HAVE_ARCH_GATE_AREA)
2207
2208 #if defined(AT_SYSINFO_EHDR)
2209 struct vm_area_struct gate_vma;
2210
2211 static int __init gate_vma_init(void)
2212 {
2213         gate_vma.vm_mm = NULL;
2214         gate_vma.vm_start = FIXADDR_USER_START;
2215         gate_vma.vm_end = FIXADDR_USER_END;
2216         gate_vma.vm_page_prot = PAGE_READONLY;
2217         gate_vma.vm_flags = 0;
2218         return 0;
2219 }
2220 __initcall(gate_vma_init);
2221 #endif
2222
2223 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2224 {
2225 #ifdef AT_SYSINFO_EHDR
2226         return &gate_vma;
2227 #else
2228         return NULL;
2229 #endif
2230 }
2231
2232 int in_gate_area_no_task(unsigned long addr)
2233 {
2234 #ifdef AT_SYSINFO_EHDR
2235         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2236                 return 1;
2237 #endif
2238         return 0;
2239 }
2240
2241 #endif  /* __HAVE_ARCH_GATE_AREA */