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