]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/blob - mm/memory.c
59d42e50fa53683ce42b1ded64c96e4f0d98eb6c
[linux-2.6-omap-h63xx.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
54 #include <asm/tlb.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
57
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
65
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
69
70 unsigned long num_physpages;
71 /*
72  * A number of key systems in x86 including ioremap() rely on the assumption
73  * that high_memory defines the upper bound on direct map memory, then end
74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76  * and ZONE_HIGHMEM.
77  */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
80
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85 /*
86  * If a p?d_bad entry is found while walking page tables, report
87  * the error, before resetting entry to p?d_none.  Usually (but
88  * very seldom) called out from the p?d_none_or_clear_bad macros.
89  */
90
91 void pgd_clear_bad(pgd_t *pgd)
92 {
93         pgd_ERROR(*pgd);
94         pgd_clear(pgd);
95 }
96
97 void pud_clear_bad(pud_t *pud)
98 {
99         pud_ERROR(*pud);
100         pud_clear(pud);
101 }
102
103 void pmd_clear_bad(pmd_t *pmd)
104 {
105         pmd_ERROR(*pmd);
106         pmd_clear(pmd);
107 }
108
109 /*
110  * Note: this doesn't free the actual pages themselves. That
111  * has been handled earlier when unmapping all the memory regions.
112  */
113 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
114 {
115         struct page *page = pmd_page(*pmd);
116         pmd_clear(pmd);
117         pte_free_tlb(tlb, page);
118         dec_page_state(nr_page_table_pages);
119         tlb->mm->nr_ptes--;
120 }
121
122 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
123                                 unsigned long addr, unsigned long end,
124                                 unsigned long floor, unsigned long ceiling)
125 {
126         pmd_t *pmd;
127         unsigned long next;
128         unsigned long start;
129
130         start = addr;
131         pmd = pmd_offset(pud, addr);
132         do {
133                 next = pmd_addr_end(addr, end);
134                 if (pmd_none_or_clear_bad(pmd))
135                         continue;
136                 free_pte_range(tlb, pmd);
137         } while (pmd++, addr = next, addr != end);
138
139         start &= PUD_MASK;
140         if (start < floor)
141                 return;
142         if (ceiling) {
143                 ceiling &= PUD_MASK;
144                 if (!ceiling)
145                         return;
146         }
147         if (end - 1 > ceiling - 1)
148                 return;
149
150         pmd = pmd_offset(pud, start);
151         pud_clear(pud);
152         pmd_free_tlb(tlb, pmd);
153 }
154
155 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
156                                 unsigned long addr, unsigned long end,
157                                 unsigned long floor, unsigned long ceiling)
158 {
159         pud_t *pud;
160         unsigned long next;
161         unsigned long start;
162
163         start = addr;
164         pud = pud_offset(pgd, addr);
165         do {
166                 next = pud_addr_end(addr, end);
167                 if (pud_none_or_clear_bad(pud))
168                         continue;
169                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
170         } while (pud++, addr = next, addr != end);
171
172         start &= PGDIR_MASK;
173         if (start < floor)
174                 return;
175         if (ceiling) {
176                 ceiling &= PGDIR_MASK;
177                 if (!ceiling)
178                         return;
179         }
180         if (end - 1 > ceiling - 1)
181                 return;
182
183         pud = pud_offset(pgd, start);
184         pgd_clear(pgd);
185         pud_free_tlb(tlb, pud);
186 }
187
188 /*
189  * This function frees user-level page tables of a process.
190  *
191  * Must be called with pagetable lock held.
192  */
193 void free_pgd_range(struct mmu_gather **tlb,
194                         unsigned long addr, unsigned long end,
195                         unsigned long floor, unsigned long ceiling)
196 {
197         pgd_t *pgd;
198         unsigned long next;
199         unsigned long start;
200
201         /*
202          * The next few lines have given us lots of grief...
203          *
204          * Why are we testing PMD* at this top level?  Because often
205          * there will be no work to do at all, and we'd prefer not to
206          * go all the way down to the bottom just to discover that.
207          *
208          * Why all these "- 1"s?  Because 0 represents both the bottom
209          * of the address space and the top of it (using -1 for the
210          * top wouldn't help much: the masks would do the wrong thing).
211          * The rule is that addr 0 and floor 0 refer to the bottom of
212          * the address space, but end 0 and ceiling 0 refer to the top
213          * Comparisons need to use "end - 1" and "ceiling - 1" (though
214          * that end 0 case should be mythical).
215          *
216          * Wherever addr is brought up or ceiling brought down, we must
217          * be careful to reject "the opposite 0" before it confuses the
218          * subsequent tests.  But what about where end is brought down
219          * by PMD_SIZE below? no, end can't go down to 0 there.
220          *
221          * Whereas we round start (addr) and ceiling down, by different
222          * masks at different levels, in order to test whether a table
223          * now has no other vmas using it, so can be freed, we don't
224          * bother to round floor or end up - the tests don't need that.
225          */
226
227         addr &= PMD_MASK;
228         if (addr < floor) {
229                 addr += PMD_SIZE;
230                 if (!addr)
231                         return;
232         }
233         if (ceiling) {
234                 ceiling &= PMD_MASK;
235                 if (!ceiling)
236                         return;
237         }
238         if (end - 1 > ceiling - 1)
239                 end -= PMD_SIZE;
240         if (addr > end - 1)
241                 return;
242
243         start = addr;
244         pgd = pgd_offset((*tlb)->mm, addr);
245         do {
246                 next = pgd_addr_end(addr, end);
247                 if (pgd_none_or_clear_bad(pgd))
248                         continue;
249                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
250         } while (pgd++, addr = next, addr != end);
251
252         if (!(*tlb)->fullmm)
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         if (PageAnon(page))
401                 inc_mm_counter(dst_mm, anon_rss);
402         else
403                 inc_mm_counter(dst_mm, file_rss);
404         set_pte_at(dst_mm, addr, dst_pte, pte);
405         page_dup_rmap(page);
406 }
407
408 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
409                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
410                 unsigned long addr, unsigned long end)
411 {
412         pte_t *src_pte, *dst_pte;
413         unsigned long vm_flags = vma->vm_flags;
414         int progress = 0;
415
416 again:
417         dst_pte = pte_alloc_map(dst_mm, dst_pmd, addr);
418         if (!dst_pte)
419                 return -ENOMEM;
420         src_pte = pte_offset_map_nested(src_pmd, addr);
421
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) {
429                         progress = 0;
430                         if (need_resched() ||
431                             need_lockbreak(&src_mm->page_table_lock) ||
432                             need_lockbreak(&dst_mm->page_table_lock))
433                                 break;
434                 }
435                 if (pte_none(*src_pte)) {
436                         progress++;
437                         continue;
438                 }
439                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vm_flags, addr);
440                 progress += 8;
441         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
442         spin_unlock(&src_mm->page_table_lock);
443
444         pte_unmap_nested(src_pte - 1);
445         pte_unmap(dst_pte - 1);
446         cond_resched_lock(&dst_mm->page_table_lock);
447         if (addr != end)
448                 goto again;
449         return 0;
450 }
451
452 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
453                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
454                 unsigned long addr, unsigned long end)
455 {
456         pmd_t *src_pmd, *dst_pmd;
457         unsigned long next;
458
459         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
460         if (!dst_pmd)
461                 return -ENOMEM;
462         src_pmd = pmd_offset(src_pud, addr);
463         do {
464                 next = pmd_addr_end(addr, end);
465                 if (pmd_none_or_clear_bad(src_pmd))
466                         continue;
467                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
468                                                 vma, addr, next))
469                         return -ENOMEM;
470         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
471         return 0;
472 }
473
474 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
475                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
476                 unsigned long addr, unsigned long end)
477 {
478         pud_t *src_pud, *dst_pud;
479         unsigned long next;
480
481         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
482         if (!dst_pud)
483                 return -ENOMEM;
484         src_pud = pud_offset(src_pgd, addr);
485         do {
486                 next = pud_addr_end(addr, end);
487                 if (pud_none_or_clear_bad(src_pud))
488                         continue;
489                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
490                                                 vma, addr, next))
491                         return -ENOMEM;
492         } while (dst_pud++, src_pud++, addr = next, addr != end);
493         return 0;
494 }
495
496 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
497                 struct vm_area_struct *vma)
498 {
499         pgd_t *src_pgd, *dst_pgd;
500         unsigned long next;
501         unsigned long addr = vma->vm_start;
502         unsigned long end = vma->vm_end;
503
504         /*
505          * Don't copy ptes where a page fault will fill them correctly.
506          * Fork becomes much lighter when there are big shared or private
507          * readonly mappings. The tradeoff is that copy_page_range is more
508          * efficient than faulting.
509          */
510         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_RESERVED))) {
511                 if (!vma->anon_vma)
512                         return 0;
513         }
514
515         if (is_vm_hugetlb_page(vma))
516                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
517
518         dst_pgd = pgd_offset(dst_mm, addr);
519         src_pgd = pgd_offset(src_mm, addr);
520         do {
521                 next = pgd_addr_end(addr, end);
522                 if (pgd_none_or_clear_bad(src_pgd))
523                         continue;
524                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
525                                                 vma, addr, next))
526                         return -ENOMEM;
527         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
528         return 0;
529 }
530
531 static void zap_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
532                                 unsigned long addr, unsigned long end,
533                                 struct zap_details *details)
534 {
535         pte_t *pte;
536
537         pte = pte_offset_map(pmd, addr);
538         do {
539                 pte_t ptent = *pte;
540                 if (pte_none(ptent))
541                         continue;
542                 if (pte_present(ptent)) {
543                         struct page *page = NULL;
544                         unsigned long pfn = pte_pfn(ptent);
545                         if (pfn_valid(pfn)) {
546                                 page = pfn_to_page(pfn);
547                                 if (PageReserved(page))
548                                         page = NULL;
549                         }
550                         if (unlikely(details) && page) {
551                                 /*
552                                  * unmap_shared_mapping_pages() wants to
553                                  * invalidate cache without truncating:
554                                  * unmap shared but keep private pages.
555                                  */
556                                 if (details->check_mapping &&
557                                     details->check_mapping != page->mapping)
558                                         continue;
559                                 /*
560                                  * Each page->index must be checked when
561                                  * invalidating or truncating nonlinear.
562                                  */
563                                 if (details->nonlinear_vma &&
564                                     (page->index < details->first_index ||
565                                      page->index > details->last_index))
566                                         continue;
567                         }
568                         ptent = ptep_get_and_clear_full(tlb->mm, addr, pte,
569                                                         tlb->fullmm);
570                         tlb_remove_tlb_entry(tlb, pte, addr);
571                         if (unlikely(!page))
572                                 continue;
573                         if (unlikely(details) && details->nonlinear_vma
574                             && linear_page_index(details->nonlinear_vma,
575                                                 addr) != page->index)
576                                 set_pte_at(tlb->mm, addr, pte,
577                                            pgoff_to_pte(page->index));
578                         if (PageAnon(page))
579                                 dec_mm_counter(tlb->mm, anon_rss);
580                         else {
581                                 if (pte_dirty(ptent))
582                                         set_page_dirty(page);
583                                 if (pte_young(ptent))
584                                         mark_page_accessed(page);
585                                 dec_mm_counter(tlb->mm, file_rss);
586                         }
587                         page_remove_rmap(page);
588                         tlb_remove_page(tlb, page);
589                         continue;
590                 }
591                 /*
592                  * If details->check_mapping, we leave swap entries;
593                  * if details->nonlinear_vma, we leave file entries.
594                  */
595                 if (unlikely(details))
596                         continue;
597                 if (!pte_file(ptent))
598                         free_swap_and_cache(pte_to_swp_entry(ptent));
599                 pte_clear_full(tlb->mm, addr, pte, tlb->fullmm);
600         } while (pte++, addr += PAGE_SIZE, addr != end);
601         pte_unmap(pte - 1);
602 }
603
604 static inline void zap_pmd_range(struct mmu_gather *tlb, pud_t *pud,
605                                 unsigned long addr, unsigned long end,
606                                 struct zap_details *details)
607 {
608         pmd_t *pmd;
609         unsigned long next;
610
611         pmd = pmd_offset(pud, addr);
612         do {
613                 next = pmd_addr_end(addr, end);
614                 if (pmd_none_or_clear_bad(pmd))
615                         continue;
616                 zap_pte_range(tlb, pmd, addr, next, details);
617         } while (pmd++, addr = next, addr != end);
618 }
619
620 static inline void zap_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
621                                 unsigned long addr, unsigned long end,
622                                 struct zap_details *details)
623 {
624         pud_t *pud;
625         unsigned long next;
626
627         pud = pud_offset(pgd, addr);
628         do {
629                 next = pud_addr_end(addr, end);
630                 if (pud_none_or_clear_bad(pud))
631                         continue;
632                 zap_pmd_range(tlb, pud, addr, next, details);
633         } while (pud++, addr = next, addr != end);
634 }
635
636 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
637                                 unsigned long addr, unsigned long end,
638                                 struct zap_details *details)
639 {
640         pgd_t *pgd;
641         unsigned long next;
642
643         if (details && !details->check_mapping && !details->nonlinear_vma)
644                 details = NULL;
645
646         BUG_ON(addr >= end);
647         tlb_start_vma(tlb, vma);
648         pgd = pgd_offset(vma->vm_mm, addr);
649         do {
650                 next = pgd_addr_end(addr, end);
651                 if (pgd_none_or_clear_bad(pgd))
652                         continue;
653                 zap_pud_range(tlb, pgd, addr, next, details);
654         } while (pgd++, addr = next, addr != end);
655         tlb_end_vma(tlb, vma);
656 }
657
658 #ifdef CONFIG_PREEMPT
659 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
660 #else
661 /* No preempt: go for improved straight-line efficiency */
662 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
663 #endif
664
665 /**
666  * unmap_vmas - unmap a range of memory covered by a list of vma's
667  * @tlbp: address of the caller's struct mmu_gather
668  * @mm: the controlling mm_struct
669  * @vma: the starting vma
670  * @start_addr: virtual address at which to start unmapping
671  * @end_addr: virtual address at which to end unmapping
672  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
673  * @details: details of nonlinear truncation or shared cache invalidation
674  *
675  * Returns the end address of the unmapping (restart addr if interrupted).
676  *
677  * Unmap all pages in the vma list.  Called under page_table_lock.
678  *
679  * We aim to not hold page_table_lock for too long (for scheduling latency
680  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
681  * return the ending mmu_gather to the caller.
682  *
683  * Only addresses between `start' and `end' will be unmapped.
684  *
685  * The VMA list must be sorted in ascending virtual address order.
686  *
687  * unmap_vmas() assumes that the caller will flush the whole unmapped address
688  * range after unmap_vmas() returns.  So the only responsibility here is to
689  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
690  * drops the lock and schedules.
691  */
692 unsigned long unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
693                 struct vm_area_struct *vma, unsigned long start_addr,
694                 unsigned long end_addr, unsigned long *nr_accounted,
695                 struct zap_details *details)
696 {
697         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
698         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
699         int tlb_start_valid = 0;
700         unsigned long start = start_addr;
701         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
702         int fullmm = (*tlbp)->fullmm;
703
704         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
705                 unsigned long end;
706
707                 start = max(vma->vm_start, start_addr);
708                 if (start >= vma->vm_end)
709                         continue;
710                 end = min(vma->vm_end, end_addr);
711                 if (end <= vma->vm_start)
712                         continue;
713
714                 if (vma->vm_flags & VM_ACCOUNT)
715                         *nr_accounted += (end - start) >> PAGE_SHIFT;
716
717                 while (start != end) {
718                         unsigned long block;
719
720                         if (!tlb_start_valid) {
721                                 tlb_start = start;
722                                 tlb_start_valid = 1;
723                         }
724
725                         if (is_vm_hugetlb_page(vma)) {
726                                 block = end - start;
727                                 unmap_hugepage_range(vma, start, end);
728                         } else {
729                                 block = min(zap_bytes, end - start);
730                                 unmap_page_range(*tlbp, vma, start,
731                                                 start + block, details);
732                         }
733
734                         start += block;
735                         zap_bytes -= block;
736                         if ((long)zap_bytes > 0)
737                                 continue;
738
739                         tlb_finish_mmu(*tlbp, tlb_start, start);
740
741                         if (need_resched() ||
742                                 need_lockbreak(&mm->page_table_lock) ||
743                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
744                                 if (i_mmap_lock) {
745                                         /* must reset count of rss freed */
746                                         *tlbp = tlb_gather_mmu(mm, fullmm);
747                                         goto out;
748                                 }
749                                 spin_unlock(&mm->page_table_lock);
750                                 cond_resched();
751                                 spin_lock(&mm->page_table_lock);
752                         }
753
754                         *tlbp = tlb_gather_mmu(mm, fullmm);
755                         tlb_start_valid = 0;
756                         zap_bytes = ZAP_BLOCK_SIZE;
757                 }
758         }
759 out:
760         return start;   /* which is now the end (or restart) address */
761 }
762
763 /**
764  * zap_page_range - remove user pages in a given range
765  * @vma: vm_area_struct holding the applicable pages
766  * @address: starting address of pages to zap
767  * @size: number of bytes to zap
768  * @details: details of nonlinear truncation or shared cache invalidation
769  */
770 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
771                 unsigned long size, struct zap_details *details)
772 {
773         struct mm_struct *mm = vma->vm_mm;
774         struct mmu_gather *tlb;
775         unsigned long end = address + size;
776         unsigned long nr_accounted = 0;
777
778         if (is_vm_hugetlb_page(vma)) {
779                 zap_hugepage_range(vma, address, size);
780                 return end;
781         }
782
783         lru_add_drain();
784         spin_lock(&mm->page_table_lock);
785         tlb = tlb_gather_mmu(mm, 0);
786         end = unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
787         tlb_finish_mmu(tlb, address, end);
788         spin_unlock(&mm->page_table_lock);
789         return end;
790 }
791
792 /*
793  * Do a quick page-table lookup for a single page.
794  * mm->page_table_lock must be held.
795  */
796 static struct page *__follow_page(struct mm_struct *mm, unsigned long address,
797                         int read, int write, int accessed)
798 {
799         pgd_t *pgd;
800         pud_t *pud;
801         pmd_t *pmd;
802         pte_t *ptep, pte;
803         unsigned long pfn;
804         struct page *page;
805
806         page = follow_huge_addr(mm, address, write);
807         if (! IS_ERR(page))
808                 return page;
809
810         pgd = pgd_offset(mm, address);
811         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
812                 goto out;
813
814         pud = pud_offset(pgd, address);
815         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
816                 goto out;
817         
818         pmd = pmd_offset(pud, address);
819         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
820                 goto out;
821         if (pmd_huge(*pmd))
822                 return follow_huge_pmd(mm, address, pmd, write);
823
824         ptep = pte_offset_map(pmd, address);
825         if (!ptep)
826                 goto out;
827
828         pte = *ptep;
829         pte_unmap(ptep);
830         if (pte_present(pte)) {
831                 if (write && !pte_write(pte))
832                         goto out;
833                 if (read && !pte_read(pte))
834                         goto out;
835                 pfn = pte_pfn(pte);
836                 if (pfn_valid(pfn)) {
837                         page = pfn_to_page(pfn);
838                         if (accessed) {
839                                 if (write && !pte_dirty(pte) &&!PageDirty(page))
840                                         set_page_dirty(page);
841                                 mark_page_accessed(page);
842                         }
843                         return page;
844                 }
845         }
846
847 out:
848         return NULL;
849 }
850
851 inline struct page *
852 follow_page(struct mm_struct *mm, unsigned long address, int write)
853 {
854         return __follow_page(mm, address, 0, write, 1);
855 }
856
857 /*
858  * check_user_page_readable() can be called frm niterrupt context by oprofile,
859  * so we need to avoid taking any non-irq-safe locks
860  */
861 int check_user_page_readable(struct mm_struct *mm, unsigned long address)
862 {
863         return __follow_page(mm, address, 1, 0, 0) != NULL;
864 }
865 EXPORT_SYMBOL(check_user_page_readable);
866
867 static inline int
868 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
869                          unsigned long address)
870 {
871         pgd_t *pgd;
872         pud_t *pud;
873         pmd_t *pmd;
874
875         /* Check if the vma is for an anonymous mapping. */
876         if (vma->vm_ops && vma->vm_ops->nopage)
877                 return 0;
878
879         /* Check if page directory entry exists. */
880         pgd = pgd_offset(mm, address);
881         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
882                 return 1;
883
884         pud = pud_offset(pgd, address);
885         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
886                 return 1;
887
888         /* Check if page middle directory entry exists. */
889         pmd = pmd_offset(pud, address);
890         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
891                 return 1;
892
893         /* There is a pte slot for 'address' in 'mm'. */
894         return 0;
895 }
896
897 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
898                 unsigned long start, int len, int write, int force,
899                 struct page **pages, struct vm_area_struct **vmas)
900 {
901         int i;
902         unsigned int flags;
903
904         /* 
905          * Require read or write permissions.
906          * If 'force' is set, we only require the "MAY" flags.
907          */
908         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
909         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
910         i = 0;
911
912         do {
913                 struct vm_area_struct * vma;
914
915                 vma = find_extend_vma(mm, start);
916                 if (!vma && in_gate_area(tsk, start)) {
917                         unsigned long pg = start & PAGE_MASK;
918                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
919                         pgd_t *pgd;
920                         pud_t *pud;
921                         pmd_t *pmd;
922                         pte_t *pte;
923                         if (write) /* user gate pages are read-only */
924                                 return i ? : -EFAULT;
925                         if (pg > TASK_SIZE)
926                                 pgd = pgd_offset_k(pg);
927                         else
928                                 pgd = pgd_offset_gate(mm, pg);
929                         BUG_ON(pgd_none(*pgd));
930                         pud = pud_offset(pgd, pg);
931                         BUG_ON(pud_none(*pud));
932                         pmd = pmd_offset(pud, pg);
933                         if (pmd_none(*pmd))
934                                 return i ? : -EFAULT;
935                         pte = pte_offset_map(pmd, pg);
936                         if (pte_none(*pte)) {
937                                 pte_unmap(pte);
938                                 return i ? : -EFAULT;
939                         }
940                         if (pages) {
941                                 pages[i] = pte_page(*pte);
942                                 get_page(pages[i]);
943                         }
944                         pte_unmap(pte);
945                         if (vmas)
946                                 vmas[i] = gate_vma;
947                         i++;
948                         start += PAGE_SIZE;
949                         len--;
950                         continue;
951                 }
952
953                 if (!vma || (vma->vm_flags & VM_IO)
954                                 || !(flags & vma->vm_flags))
955                         return i ? : -EFAULT;
956
957                 if (is_vm_hugetlb_page(vma)) {
958                         i = follow_hugetlb_page(mm, vma, pages, vmas,
959                                                 &start, &len, i);
960                         continue;
961                 }
962                 spin_lock(&mm->page_table_lock);
963                 do {
964                         int write_access = write;
965                         struct page *page;
966
967                         cond_resched_lock(&mm->page_table_lock);
968                         while (!(page = follow_page(mm, start, write_access))) {
969                                 int ret;
970
971                                 /*
972                                  * Shortcut for anonymous pages. We don't want
973                                  * to force the creation of pages tables for
974                                  * insanely big anonymously mapped areas that
975                                  * nobody touched so far. This is important
976                                  * for doing a core dump for these mappings.
977                                  */
978                                 if (!write && untouched_anonymous_page(mm,vma,start)) {
979                                         page = ZERO_PAGE(start);
980                                         break;
981                                 }
982                                 spin_unlock(&mm->page_table_lock);
983                                 ret = __handle_mm_fault(mm, vma, start, write_access);
984
985                                 /*
986                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
987                                  * broken COW when necessary, even if maybe_mkwrite
988                                  * decided not to set pte_write. We can thus safely do
989                                  * subsequent page lookups as if they were reads.
990                                  */
991                                 if (ret & VM_FAULT_WRITE)
992                                         write_access = 0;
993                                 
994                                 switch (ret & ~VM_FAULT_WRITE) {
995                                 case VM_FAULT_MINOR:
996                                         tsk->min_flt++;
997                                         break;
998                                 case VM_FAULT_MAJOR:
999                                         tsk->maj_flt++;
1000                                         break;
1001                                 case VM_FAULT_SIGBUS:
1002                                         return i ? i : -EFAULT;
1003                                 case VM_FAULT_OOM:
1004                                         return i ? i : -ENOMEM;
1005                                 default:
1006                                         BUG();
1007                                 }
1008                                 spin_lock(&mm->page_table_lock);
1009                         }
1010                         if (pages) {
1011                                 pages[i] = page;
1012                                 flush_dcache_page(page);
1013                                 if (!PageReserved(page))
1014                                         page_cache_get(page);
1015                         }
1016                         if (vmas)
1017                                 vmas[i] = vma;
1018                         i++;
1019                         start += PAGE_SIZE;
1020                         len--;
1021                 } while (len && start < vma->vm_end);
1022                 spin_unlock(&mm->page_table_lock);
1023         } while (len);
1024         return i;
1025 }
1026 EXPORT_SYMBOL(get_user_pages);
1027
1028 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1029                         unsigned long addr, unsigned long end, pgprot_t prot)
1030 {
1031         pte_t *pte;
1032
1033         pte = pte_alloc_map(mm, pmd, addr);
1034         if (!pte)
1035                 return -ENOMEM;
1036         do {
1037                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(addr), prot));
1038                 BUG_ON(!pte_none(*pte));
1039                 set_pte_at(mm, addr, pte, zero_pte);
1040         } while (pte++, addr += PAGE_SIZE, addr != end);
1041         pte_unmap(pte - 1);
1042         return 0;
1043 }
1044
1045 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1046                         unsigned long addr, unsigned long end, pgprot_t prot)
1047 {
1048         pmd_t *pmd;
1049         unsigned long next;
1050
1051         pmd = pmd_alloc(mm, pud, addr);
1052         if (!pmd)
1053                 return -ENOMEM;
1054         do {
1055                 next = pmd_addr_end(addr, end);
1056                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1057                         return -ENOMEM;
1058         } while (pmd++, addr = next, addr != end);
1059         return 0;
1060 }
1061
1062 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1063                         unsigned long addr, unsigned long end, pgprot_t prot)
1064 {
1065         pud_t *pud;
1066         unsigned long next;
1067
1068         pud = pud_alloc(mm, pgd, addr);
1069         if (!pud)
1070                 return -ENOMEM;
1071         do {
1072                 next = pud_addr_end(addr, end);
1073                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1074                         return -ENOMEM;
1075         } while (pud++, addr = next, addr != end);
1076         return 0;
1077 }
1078
1079 int zeromap_page_range(struct vm_area_struct *vma,
1080                         unsigned long addr, unsigned long size, pgprot_t prot)
1081 {
1082         pgd_t *pgd;
1083         unsigned long next;
1084         unsigned long end = addr + size;
1085         struct mm_struct *mm = vma->vm_mm;
1086         int err;
1087
1088         BUG_ON(addr >= end);
1089         pgd = pgd_offset(mm, addr);
1090         flush_cache_range(vma, addr, end);
1091         spin_lock(&mm->page_table_lock);
1092         do {
1093                 next = pgd_addr_end(addr, end);
1094                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1095                 if (err)
1096                         break;
1097         } while (pgd++, addr = next, addr != end);
1098         spin_unlock(&mm->page_table_lock);
1099         return err;
1100 }
1101
1102 /*
1103  * maps a range of physical memory into the requested pages. the old
1104  * mappings are removed. any references to nonexistent pages results
1105  * in null mappings (currently treated as "copy-on-access")
1106  */
1107 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1108                         unsigned long addr, unsigned long end,
1109                         unsigned long pfn, pgprot_t prot)
1110 {
1111         pte_t *pte;
1112
1113         pte = pte_alloc_map(mm, pmd, addr);
1114         if (!pte)
1115                 return -ENOMEM;
1116         do {
1117                 BUG_ON(!pte_none(*pte));
1118                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
1119                         set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1120                 pfn++;
1121         } while (pte++, addr += PAGE_SIZE, addr != end);
1122         pte_unmap(pte - 1);
1123         return 0;
1124 }
1125
1126 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1127                         unsigned long addr, unsigned long end,
1128                         unsigned long pfn, pgprot_t prot)
1129 {
1130         pmd_t *pmd;
1131         unsigned long next;
1132
1133         pfn -= addr >> PAGE_SHIFT;
1134         pmd = pmd_alloc(mm, pud, addr);
1135         if (!pmd)
1136                 return -ENOMEM;
1137         do {
1138                 next = pmd_addr_end(addr, end);
1139                 if (remap_pte_range(mm, pmd, addr, next,
1140                                 pfn + (addr >> PAGE_SHIFT), prot))
1141                         return -ENOMEM;
1142         } while (pmd++, addr = next, addr != end);
1143         return 0;
1144 }
1145
1146 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1147                         unsigned long addr, unsigned long end,
1148                         unsigned long pfn, pgprot_t prot)
1149 {
1150         pud_t *pud;
1151         unsigned long next;
1152
1153         pfn -= addr >> PAGE_SHIFT;
1154         pud = pud_alloc(mm, pgd, addr);
1155         if (!pud)
1156                 return -ENOMEM;
1157         do {
1158                 next = pud_addr_end(addr, end);
1159                 if (remap_pmd_range(mm, pud, addr, next,
1160                                 pfn + (addr >> PAGE_SHIFT), prot))
1161                         return -ENOMEM;
1162         } while (pud++, addr = next, addr != end);
1163         return 0;
1164 }
1165
1166 /*  Note: this is only safe if the mm semaphore is held when called. */
1167 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1168                     unsigned long pfn, unsigned long size, pgprot_t prot)
1169 {
1170         pgd_t *pgd;
1171         unsigned long next;
1172         unsigned long end = addr + PAGE_ALIGN(size);
1173         struct mm_struct *mm = vma->vm_mm;
1174         int err;
1175
1176         /*
1177          * Physically remapped pages are special. Tell the
1178          * rest of the world about it:
1179          *   VM_IO tells people not to look at these pages
1180          *      (accesses can have side effects).
1181          *   VM_RESERVED tells swapout not to try to touch
1182          *      this region.
1183          */
1184         vma->vm_flags |= VM_IO | VM_RESERVED;
1185
1186         BUG_ON(addr >= end);
1187         pfn -= addr >> PAGE_SHIFT;
1188         pgd = pgd_offset(mm, addr);
1189         flush_cache_range(vma, addr, end);
1190         spin_lock(&mm->page_table_lock);
1191         do {
1192                 next = pgd_addr_end(addr, end);
1193                 err = remap_pud_range(mm, pgd, addr, next,
1194                                 pfn + (addr >> PAGE_SHIFT), prot);
1195                 if (err)
1196                         break;
1197         } while (pgd++, addr = next, addr != end);
1198         spin_unlock(&mm->page_table_lock);
1199         return err;
1200 }
1201 EXPORT_SYMBOL(remap_pfn_range);
1202
1203 /*
1204  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1205  * servicing faults for write access.  In the normal case, do always want
1206  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1207  * that do not have writing enabled, when used by access_process_vm.
1208  */
1209 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1210 {
1211         if (likely(vma->vm_flags & VM_WRITE))
1212                 pte = pte_mkwrite(pte);
1213         return pte;
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  * Note that this routine assumes that the protection checks have been
1222  * done by the caller (the low-level page fault routine in most cases).
1223  * Thus we can safely just mark it writable once we've done any necessary
1224  * COW.
1225  *
1226  * We also mark the page dirty at this point even though the page will
1227  * change only once the write actually happens. This avoids a few races,
1228  * and potentially makes it more efficient.
1229  *
1230  * We hold the mm semaphore and the page_table_lock on entry and exit
1231  * with the page_table_lock released.
1232  */
1233 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1234                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1235                 pte_t orig_pte)
1236 {
1237         struct page *old_page, *new_page;
1238         unsigned long pfn = pte_pfn(orig_pte);
1239         pte_t entry;
1240         int ret = VM_FAULT_MINOR;
1241
1242         if (unlikely(!pfn_valid(pfn))) {
1243                 /*
1244                  * Page table corrupted: show pte and kill process.
1245                  */
1246                 pte_ERROR(orig_pte);
1247                 ret = VM_FAULT_OOM;
1248                 goto unlock;
1249         }
1250         old_page = pfn_to_page(pfn);
1251
1252         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1253                 int reuse = can_share_swap_page(old_page);
1254                 unlock_page(old_page);
1255                 if (reuse) {
1256                         flush_cache_page(vma, address, pfn);
1257                         entry = pte_mkyoung(orig_pte);
1258                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1259                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1260                         update_mmu_cache(vma, address, entry);
1261                         lazy_mmu_prot_update(entry);
1262                         ret |= VM_FAULT_WRITE;
1263                         goto unlock;
1264                 }
1265         }
1266
1267         /*
1268          * Ok, we need to copy. Oh, well..
1269          */
1270         if (!PageReserved(old_page))
1271                 page_cache_get(old_page);
1272         pte_unmap(page_table);
1273         spin_unlock(&mm->page_table_lock);
1274
1275         if (unlikely(anon_vma_prepare(vma)))
1276                 goto oom;
1277         if (old_page == ZERO_PAGE(address)) {
1278                 new_page = alloc_zeroed_user_highpage(vma, address);
1279                 if (!new_page)
1280                         goto oom;
1281         } else {
1282                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1283                 if (!new_page)
1284                         goto oom;
1285                 copy_user_highpage(new_page, old_page, address);
1286         }
1287
1288         /*
1289          * Re-check the pte - we dropped the lock
1290          */
1291         spin_lock(&mm->page_table_lock);
1292         page_table = pte_offset_map(pmd, address);
1293         if (likely(pte_same(*page_table, orig_pte))) {
1294                 if (PageReserved(old_page))
1295                         inc_mm_counter(mm, anon_rss);
1296                 else {
1297                         page_remove_rmap(old_page);
1298                         if (!PageAnon(old_page)) {
1299                                 inc_mm_counter(mm, anon_rss);
1300                                 dec_mm_counter(mm, file_rss);
1301                         }
1302                 }
1303                 flush_cache_page(vma, address, pfn);
1304                 entry = mk_pte(new_page, vma->vm_page_prot);
1305                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1306                 ptep_establish(vma, address, page_table, entry);
1307                 update_mmu_cache(vma, address, entry);
1308                 lazy_mmu_prot_update(entry);
1309
1310                 lru_cache_add_active(new_page);
1311                 page_add_anon_rmap(new_page, vma, address);
1312
1313                 /* Free the old page.. */
1314                 new_page = old_page;
1315                 ret |= VM_FAULT_WRITE;
1316         }
1317         page_cache_release(new_page);
1318         page_cache_release(old_page);
1319 unlock:
1320         pte_unmap(page_table);
1321         spin_unlock(&mm->page_table_lock);
1322         return ret;
1323 oom:
1324         page_cache_release(old_page);
1325         return VM_FAULT_OOM;
1326 }
1327
1328 /*
1329  * Helper functions for unmap_mapping_range().
1330  *
1331  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1332  *
1333  * We have to restart searching the prio_tree whenever we drop the lock,
1334  * since the iterator is only valid while the lock is held, and anyway
1335  * a later vma might be split and reinserted earlier while lock dropped.
1336  *
1337  * The list of nonlinear vmas could be handled more efficiently, using
1338  * a placeholder, but handle it in the same way until a need is shown.
1339  * It is important to search the prio_tree before nonlinear list: a vma
1340  * may become nonlinear and be shifted from prio_tree to nonlinear list
1341  * while the lock is dropped; but never shifted from list to prio_tree.
1342  *
1343  * In order to make forward progress despite restarting the search,
1344  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1345  * quickly skip it next time around.  Since the prio_tree search only
1346  * shows us those vmas affected by unmapping the range in question, we
1347  * can't efficiently keep all vmas in step with mapping->truncate_count:
1348  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1349  * mapping->truncate_count and vma->vm_truncate_count are protected by
1350  * i_mmap_lock.
1351  *
1352  * In order to make forward progress despite repeatedly restarting some
1353  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1354  * and restart from that address when we reach that vma again.  It might
1355  * have been split or merged, shrunk or extended, but never shifted: so
1356  * restart_addr remains valid so long as it remains in the vma's range.
1357  * unmap_mapping_range forces truncate_count to leap over page-aligned
1358  * values so we can save vma's restart_addr in its truncate_count field.
1359  */
1360 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1361
1362 static void reset_vma_truncate_counts(struct address_space *mapping)
1363 {
1364         struct vm_area_struct *vma;
1365         struct prio_tree_iter iter;
1366
1367         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1368                 vma->vm_truncate_count = 0;
1369         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1370                 vma->vm_truncate_count = 0;
1371 }
1372
1373 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1374                 unsigned long start_addr, unsigned long end_addr,
1375                 struct zap_details *details)
1376 {
1377         unsigned long restart_addr;
1378         int need_break;
1379
1380 again:
1381         restart_addr = vma->vm_truncate_count;
1382         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1383                 start_addr = restart_addr;
1384                 if (start_addr >= end_addr) {
1385                         /* Top of vma has been split off since last time */
1386                         vma->vm_truncate_count = details->truncate_count;
1387                         return 0;
1388                 }
1389         }
1390
1391         restart_addr = zap_page_range(vma, start_addr,
1392                                         end_addr - start_addr, details);
1393
1394         /*
1395          * We cannot rely on the break test in unmap_vmas:
1396          * on the one hand, we don't want to restart our loop
1397          * just because that broke out for the page_table_lock;
1398          * on the other hand, it does no test when vma is small.
1399          */
1400         need_break = need_resched() ||
1401                         need_lockbreak(details->i_mmap_lock);
1402
1403         if (restart_addr >= end_addr) {
1404                 /* We have now completed this vma: mark it so */
1405                 vma->vm_truncate_count = details->truncate_count;
1406                 if (!need_break)
1407                         return 0;
1408         } else {
1409                 /* Note restart_addr in vma's truncate_count field */
1410                 vma->vm_truncate_count = restart_addr;
1411                 if (!need_break)
1412                         goto again;
1413         }
1414
1415         spin_unlock(details->i_mmap_lock);
1416         cond_resched();
1417         spin_lock(details->i_mmap_lock);
1418         return -EINTR;
1419 }
1420
1421 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1422                                             struct zap_details *details)
1423 {
1424         struct vm_area_struct *vma;
1425         struct prio_tree_iter iter;
1426         pgoff_t vba, vea, zba, zea;
1427
1428 restart:
1429         vma_prio_tree_foreach(vma, &iter, root,
1430                         details->first_index, details->last_index) {
1431                 /* Skip quickly over those we have already dealt with */
1432                 if (vma->vm_truncate_count == details->truncate_count)
1433                         continue;
1434
1435                 vba = vma->vm_pgoff;
1436                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1437                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1438                 zba = details->first_index;
1439                 if (zba < vba)
1440                         zba = vba;
1441                 zea = details->last_index;
1442                 if (zea > vea)
1443                         zea = vea;
1444
1445                 if (unmap_mapping_range_vma(vma,
1446                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1447                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1448                                 details) < 0)
1449                         goto restart;
1450         }
1451 }
1452
1453 static inline void unmap_mapping_range_list(struct list_head *head,
1454                                             struct zap_details *details)
1455 {
1456         struct vm_area_struct *vma;
1457
1458         /*
1459          * In nonlinear VMAs there is no correspondence between virtual address
1460          * offset and file offset.  So we must perform an exhaustive search
1461          * across *all* the pages in each nonlinear VMA, not just the pages
1462          * whose virtual address lies outside the file truncation point.
1463          */
1464 restart:
1465         list_for_each_entry(vma, head, shared.vm_set.list) {
1466                 /* Skip quickly over those we have already dealt with */
1467                 if (vma->vm_truncate_count == details->truncate_count)
1468                         continue;
1469                 details->nonlinear_vma = vma;
1470                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1471                                         vma->vm_end, details) < 0)
1472                         goto restart;
1473         }
1474 }
1475
1476 /**
1477  * unmap_mapping_range - unmap the portion of all mmaps
1478  * in the specified address_space corresponding to the specified
1479  * page range in the underlying file.
1480  * @mapping: the address space containing mmaps to be unmapped.
1481  * @holebegin: byte in first page to unmap, relative to the start of
1482  * the underlying file.  This will be rounded down to a PAGE_SIZE
1483  * boundary.  Note that this is different from vmtruncate(), which
1484  * must keep the partial page.  In contrast, we must get rid of
1485  * partial pages.
1486  * @holelen: size of prospective hole in bytes.  This will be rounded
1487  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1488  * end of the file.
1489  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1490  * but 0 when invalidating pagecache, don't throw away private data.
1491  */
1492 void unmap_mapping_range(struct address_space *mapping,
1493                 loff_t const holebegin, loff_t const holelen, int even_cows)
1494 {
1495         struct zap_details details;
1496         pgoff_t hba = holebegin >> PAGE_SHIFT;
1497         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1498
1499         /* Check for overflow. */
1500         if (sizeof(holelen) > sizeof(hlen)) {
1501                 long long holeend =
1502                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1503                 if (holeend & ~(long long)ULONG_MAX)
1504                         hlen = ULONG_MAX - hba + 1;
1505         }
1506
1507         details.check_mapping = even_cows? NULL: mapping;
1508         details.nonlinear_vma = NULL;
1509         details.first_index = hba;
1510         details.last_index = hba + hlen - 1;
1511         if (details.last_index < details.first_index)
1512                 details.last_index = ULONG_MAX;
1513         details.i_mmap_lock = &mapping->i_mmap_lock;
1514
1515         spin_lock(&mapping->i_mmap_lock);
1516
1517         /* serialize i_size write against truncate_count write */
1518         smp_wmb();
1519         /* Protect against page faults, and endless unmapping loops */
1520         mapping->truncate_count++;
1521         /*
1522          * For archs where spin_lock has inclusive semantics like ia64
1523          * this smp_mb() will prevent to read pagetable contents
1524          * before the truncate_count increment is visible to
1525          * other cpus.
1526          */
1527         smp_mb();
1528         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1529                 if (mapping->truncate_count == 0)
1530                         reset_vma_truncate_counts(mapping);
1531                 mapping->truncate_count++;
1532         }
1533         details.truncate_count = mapping->truncate_count;
1534
1535         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1536                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1537         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1538                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1539         spin_unlock(&mapping->i_mmap_lock);
1540 }
1541 EXPORT_SYMBOL(unmap_mapping_range);
1542
1543 /*
1544  * Handle all mappings that got truncated by a "truncate()"
1545  * system call.
1546  *
1547  * NOTE! We have to be ready to update the memory sharing
1548  * between the file and the memory map for a potential last
1549  * incomplete page.  Ugly, but necessary.
1550  */
1551 int vmtruncate(struct inode * inode, loff_t offset)
1552 {
1553         struct address_space *mapping = inode->i_mapping;
1554         unsigned long limit;
1555
1556         if (inode->i_size < offset)
1557                 goto do_expand;
1558         /*
1559          * truncation of in-use swapfiles is disallowed - it would cause
1560          * subsequent swapout to scribble on the now-freed blocks.
1561          */
1562         if (IS_SWAPFILE(inode))
1563                 goto out_busy;
1564         i_size_write(inode, offset);
1565         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1566         truncate_inode_pages(mapping, offset);
1567         goto out_truncate;
1568
1569 do_expand:
1570         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1571         if (limit != RLIM_INFINITY && offset > limit)
1572                 goto out_sig;
1573         if (offset > inode->i_sb->s_maxbytes)
1574                 goto out_big;
1575         i_size_write(inode, offset);
1576
1577 out_truncate:
1578         if (inode->i_op && inode->i_op->truncate)
1579                 inode->i_op->truncate(inode);
1580         return 0;
1581 out_sig:
1582         send_sig(SIGXFSZ, current, 0);
1583 out_big:
1584         return -EFBIG;
1585 out_busy:
1586         return -ETXTBSY;
1587 }
1588
1589 EXPORT_SYMBOL(vmtruncate);
1590
1591 /* 
1592  * Primitive swap readahead code. We simply read an aligned block of
1593  * (1 << page_cluster) entries in the swap area. This method is chosen
1594  * because it doesn't cost us any seek time.  We also make sure to queue
1595  * the 'original' request together with the readahead ones...  
1596  *
1597  * This has been extended to use the NUMA policies from the mm triggering
1598  * the readahead.
1599  *
1600  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1601  */
1602 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1603 {
1604 #ifdef CONFIG_NUMA
1605         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1606 #endif
1607         int i, num;
1608         struct page *new_page;
1609         unsigned long offset;
1610
1611         /*
1612          * Get the number of handles we should do readahead io to.
1613          */
1614         num = valid_swaphandles(entry, &offset);
1615         for (i = 0; i < num; offset++, i++) {
1616                 /* Ok, do the async read-ahead now */
1617                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1618                                                            offset), vma, addr);
1619                 if (!new_page)
1620                         break;
1621                 page_cache_release(new_page);
1622 #ifdef CONFIG_NUMA
1623                 /*
1624                  * Find the next applicable VMA for the NUMA policy.
1625                  */
1626                 addr += PAGE_SIZE;
1627                 if (addr == 0)
1628                         vma = NULL;
1629                 if (vma) {
1630                         if (addr >= vma->vm_end) {
1631                                 vma = next_vma;
1632                                 next_vma = vma ? vma->vm_next : NULL;
1633                         }
1634                         if (vma && addr < vma->vm_start)
1635                                 vma = NULL;
1636                 } else {
1637                         if (next_vma && addr >= next_vma->vm_start) {
1638                                 vma = next_vma;
1639                                 next_vma = vma->vm_next;
1640                         }
1641                 }
1642 #endif
1643         }
1644         lru_add_drain();        /* Push any new pages onto the LRU now */
1645 }
1646
1647 /*
1648  * We hold the mm semaphore and the page_table_lock on entry and
1649  * should release the pagetable lock on exit..
1650  */
1651 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1652                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1653                 int write_access, pte_t orig_pte)
1654 {
1655         struct page *page;
1656         swp_entry_t entry;
1657         pte_t pte;
1658         int ret = VM_FAULT_MINOR;
1659
1660         pte_unmap(page_table);
1661         spin_unlock(&mm->page_table_lock);
1662
1663         entry = pte_to_swp_entry(orig_pte);
1664         page = lookup_swap_cache(entry);
1665         if (!page) {
1666                 swapin_readahead(entry, address, vma);
1667                 page = read_swap_cache_async(entry, vma, address);
1668                 if (!page) {
1669                         /*
1670                          * Back out if somebody else faulted in this pte while
1671                          * we released the page table lock.
1672                          */
1673                         spin_lock(&mm->page_table_lock);
1674                         page_table = pte_offset_map(pmd, address);
1675                         if (likely(pte_same(*page_table, orig_pte)))
1676                                 ret = VM_FAULT_OOM;
1677                         goto unlock;
1678                 }
1679
1680                 /* Had to read the page from swap area: Major fault */
1681                 ret = VM_FAULT_MAJOR;
1682                 inc_page_state(pgmajfault);
1683                 grab_swap_token();
1684         }
1685
1686         mark_page_accessed(page);
1687         lock_page(page);
1688
1689         /*
1690          * Back out if somebody else faulted in this pte while we
1691          * released the page table lock.
1692          */
1693         spin_lock(&mm->page_table_lock);
1694         page_table = pte_offset_map(pmd, address);
1695         if (unlikely(!pte_same(*page_table, orig_pte))) {
1696                 ret = VM_FAULT_MINOR;
1697                 goto out_nomap;
1698         }
1699
1700         if (unlikely(!PageUptodate(page))) {
1701                 ret = VM_FAULT_SIGBUS;
1702                 goto out_nomap;
1703         }
1704
1705         /* The page isn't present yet, go ahead with the fault. */
1706
1707         inc_mm_counter(mm, anon_rss);
1708         pte = mk_pte(page, vma->vm_page_prot);
1709         if (write_access && can_share_swap_page(page)) {
1710                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1711                 write_access = 0;
1712         }
1713
1714         flush_icache_page(vma, page);
1715         set_pte_at(mm, address, page_table, pte);
1716         page_add_anon_rmap(page, vma, address);
1717
1718         swap_free(entry);
1719         if (vm_swap_full())
1720                 remove_exclusive_swap_page(page);
1721         unlock_page(page);
1722
1723         if (write_access) {
1724                 if (do_wp_page(mm, vma, address,
1725                                 page_table, pmd, pte) == VM_FAULT_OOM)
1726                         ret = VM_FAULT_OOM;
1727                 goto out;
1728         }
1729
1730         /* No need to invalidate - it was non-present before */
1731         update_mmu_cache(vma, address, pte);
1732         lazy_mmu_prot_update(pte);
1733 unlock:
1734         pte_unmap(page_table);
1735         spin_unlock(&mm->page_table_lock);
1736 out:
1737         return ret;
1738 out_nomap:
1739         pte_unmap(page_table);
1740         spin_unlock(&mm->page_table_lock);
1741         unlock_page(page);
1742         page_cache_release(page);
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 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1752                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1753                 int write_access)
1754 {
1755         pte_t entry;
1756
1757         /* Mapping of ZERO_PAGE - vm_page_prot is readonly */
1758         entry = mk_pte(ZERO_PAGE(addr), vma->vm_page_prot);
1759
1760         if (write_access) {
1761                 struct page *page;
1762
1763                 /* Allocate our own private page. */
1764                 pte_unmap(page_table);
1765                 spin_unlock(&mm->page_table_lock);
1766
1767                 if (unlikely(anon_vma_prepare(vma)))
1768                         goto oom;
1769                 page = alloc_zeroed_user_highpage(vma, address);
1770                 if (!page)
1771                         goto oom;
1772
1773                 spin_lock(&mm->page_table_lock);
1774                 page_table = pte_offset_map(pmd, address);
1775
1776                 if (!pte_none(*page_table)) {
1777                         page_cache_release(page);
1778                         goto unlock;
1779                 }
1780                 inc_mm_counter(mm, anon_rss);
1781                 entry = mk_pte(page, vma->vm_page_prot);
1782                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1783                 lru_cache_add_active(page);
1784                 SetPageReferenced(page);
1785                 page_add_anon_rmap(page, vma, address);
1786         }
1787
1788         set_pte_at(mm, address, page_table, entry);
1789
1790         /* No need to invalidate - it was non-present before */
1791         update_mmu_cache(vma, address, entry);
1792         lazy_mmu_prot_update(entry);
1793 unlock:
1794         pte_unmap(page_table);
1795         spin_unlock(&mm->page_table_lock);
1796         return VM_FAULT_MINOR;
1797 oom:
1798         return VM_FAULT_OOM;
1799 }
1800
1801 /*
1802  * do_no_page() tries to create a new page mapping. It aggressively
1803  * tries to share with existing pages, but makes a separate copy if
1804  * the "write_access" parameter is true in order to avoid the next
1805  * page fault.
1806  *
1807  * As this is called only for pages that do not currently exist, we
1808  * do not need to flush old virtual caches or the TLB.
1809  *
1810  * This is called with the MM semaphore held and the page table
1811  * spinlock held. Exit with the spinlock released.
1812  */
1813 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1814                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1815                 int write_access)
1816 {
1817         struct page *new_page;
1818         struct address_space *mapping = NULL;
1819         pte_t entry;
1820         unsigned int sequence = 0;
1821         int ret = VM_FAULT_MINOR;
1822         int anon = 0;
1823
1824         pte_unmap(page_table);
1825         spin_unlock(&mm->page_table_lock);
1826
1827         if (vma->vm_file) {
1828                 mapping = vma->vm_file->f_mapping;
1829                 sequence = mapping->truncate_count;
1830                 smp_rmb(); /* serializes i_size against truncate_count */
1831         }
1832 retry:
1833         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1834         /*
1835          * No smp_rmb is needed here as long as there's a full
1836          * spin_lock/unlock sequence inside the ->nopage callback
1837          * (for the pagecache lookup) that acts as an implicit
1838          * smp_mb() and prevents the i_size read to happen
1839          * after the next truncate_count read.
1840          */
1841
1842         /* no page was available -- either SIGBUS or OOM */
1843         if (new_page == NOPAGE_SIGBUS)
1844                 return VM_FAULT_SIGBUS;
1845         if (new_page == NOPAGE_OOM)
1846                 return VM_FAULT_OOM;
1847
1848         /*
1849          * Should we do an early C-O-W break?
1850          */
1851         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1852                 struct page *page;
1853
1854                 if (unlikely(anon_vma_prepare(vma)))
1855                         goto oom;
1856                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1857                 if (!page)
1858                         goto oom;
1859                 copy_user_highpage(page, new_page, address);
1860                 page_cache_release(new_page);
1861                 new_page = page;
1862                 anon = 1;
1863         }
1864
1865         spin_lock(&mm->page_table_lock);
1866         /*
1867          * For a file-backed vma, someone could have truncated or otherwise
1868          * invalidated this page.  If unmap_mapping_range got called,
1869          * retry getting the page.
1870          */
1871         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1872                 spin_unlock(&mm->page_table_lock);
1873                 page_cache_release(new_page);
1874                 cond_resched();
1875                 sequence = mapping->truncate_count;
1876                 smp_rmb();
1877                 goto retry;
1878         }
1879         page_table = pte_offset_map(pmd, address);
1880
1881         /*
1882          * This silly early PAGE_DIRTY setting removes a race
1883          * due to the bad i386 page protection. But it's valid
1884          * for other architectures too.
1885          *
1886          * Note that if write_access is true, we either now have
1887          * an exclusive copy of the page, or this is a shared mapping,
1888          * so we can make it writable and dirty to avoid having to
1889          * handle that later.
1890          */
1891         /* Only go through if we didn't race with anybody else... */
1892         if (pte_none(*page_table)) {
1893                 flush_icache_page(vma, new_page);
1894                 entry = mk_pte(new_page, vma->vm_page_prot);
1895                 if (write_access)
1896                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1897                 set_pte_at(mm, address, page_table, entry);
1898                 if (anon) {
1899                         inc_mm_counter(mm, anon_rss);
1900                         lru_cache_add_active(new_page);
1901                         page_add_anon_rmap(new_page, vma, address);
1902                 } else if (!PageReserved(new_page)) {
1903                         inc_mm_counter(mm, file_rss);
1904                         page_add_file_rmap(new_page);
1905                 }
1906         } else {
1907                 /* One of our sibling threads was faster, back out. */
1908                 page_cache_release(new_page);
1909                 goto unlock;
1910         }
1911
1912         /* no need to invalidate: a not-present page shouldn't be cached */
1913         update_mmu_cache(vma, address, entry);
1914         lazy_mmu_prot_update(entry);
1915 unlock:
1916         pte_unmap(page_table);
1917         spin_unlock(&mm->page_table_lock);
1918         return ret;
1919 oom:
1920         page_cache_release(new_page);
1921         return VM_FAULT_OOM;
1922 }
1923
1924 /*
1925  * Fault of a previously existing named mapping. Repopulate the pte
1926  * from the encoded file_pte if possible. This enables swappable
1927  * nonlinear vmas.
1928  */
1929 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
1930                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1931                 int write_access, pte_t orig_pte)
1932 {
1933         pgoff_t pgoff;
1934         int err;
1935
1936         pte_unmap(page_table);
1937         spin_unlock(&mm->page_table_lock);
1938
1939         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
1940                 /*
1941                  * Page table corrupted: show pte and kill process.
1942                  */
1943                 pte_ERROR(orig_pte);
1944                 return VM_FAULT_OOM;
1945         }
1946         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
1947
1948         pgoff = pte_to_pgoff(orig_pte);
1949         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
1950                                         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                 pte_t *pte, pmd_t *pmd, int write_access)
1982 {
1983         pte_t entry;
1984
1985         entry = *pte;
1986         if (!pte_present(entry)) {
1987                 if (pte_none(entry)) {
1988                         if (!vma->vm_ops || !vma->vm_ops->nopage)
1989                                 return do_anonymous_page(mm, vma, address,
1990                                         pte, pmd, write_access);
1991                         return do_no_page(mm, vma, address,
1992                                         pte, pmd, write_access);
1993                 }
1994                 if (pte_file(entry))
1995                         return do_file_page(mm, vma, address,
1996                                         pte, pmd, write_access, entry);
1997                 return do_swap_page(mm, vma, address,
1998                                         pte, pmd, write_access, entry);
1999         }
2000
2001         if (write_access) {
2002                 if (!pte_write(entry))
2003                         return do_wp_page(mm, vma, address, pte, pmd, entry);
2004                 entry = pte_mkdirty(entry);
2005         }
2006         entry = pte_mkyoung(entry);
2007         ptep_set_access_flags(vma, address, pte, entry, write_access);
2008         update_mmu_cache(vma, address, entry);
2009         lazy_mmu_prot_update(entry);
2010         pte_unmap(pte);
2011         spin_unlock(&mm->page_table_lock);
2012         return VM_FAULT_MINOR;
2013 }
2014
2015 /*
2016  * By the time we get here, we already hold the mm semaphore
2017  */
2018 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2019                 unsigned long address, int write_access)
2020 {
2021         pgd_t *pgd;
2022         pud_t *pud;
2023         pmd_t *pmd;
2024         pte_t *pte;
2025
2026         __set_current_state(TASK_RUNNING);
2027
2028         inc_page_state(pgfault);
2029
2030         if (unlikely(is_vm_hugetlb_page(vma)))
2031                 return hugetlb_fault(mm, vma, address, write_access);
2032
2033         /*
2034          * We need the page table lock to synchronize with kswapd
2035          * and the SMP-safe atomic PTE updates.
2036          */
2037         pgd = pgd_offset(mm, address);
2038         spin_lock(&mm->page_table_lock);
2039
2040         pud = pud_alloc(mm, pgd, address);
2041         if (!pud)
2042                 goto oom;
2043
2044         pmd = pmd_alloc(mm, pud, address);
2045         if (!pmd)
2046                 goto oom;
2047
2048         pte = pte_alloc_map(mm, pmd, address);
2049         if (!pte)
2050                 goto oom;
2051         
2052         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2053
2054  oom:
2055         spin_unlock(&mm->page_table_lock);
2056         return VM_FAULT_OOM;
2057 }
2058
2059 #ifndef __PAGETABLE_PUD_FOLDED
2060 /*
2061  * Allocate page upper directory.
2062  *
2063  * We've already handled the fast-path in-line, and we own the
2064  * page table lock.
2065  */
2066 pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2067 {
2068         pud_t *new;
2069
2070         spin_unlock(&mm->page_table_lock);
2071         new = pud_alloc_one(mm, address);
2072         spin_lock(&mm->page_table_lock);
2073         if (!new)
2074                 return NULL;
2075
2076         /*
2077          * Because we dropped the lock, we should re-check the
2078          * entry, as somebody else could have populated it..
2079          */
2080         if (pgd_present(*pgd)) {
2081                 pud_free(new);
2082                 goto out;
2083         }
2084         pgd_populate(mm, pgd, new);
2085  out:
2086         return pud_offset(pgd, address);
2087 }
2088 #endif /* __PAGETABLE_PUD_FOLDED */
2089
2090 #ifndef __PAGETABLE_PMD_FOLDED
2091 /*
2092  * Allocate page middle directory.
2093  *
2094  * We've already handled the fast-path in-line, and we own the
2095  * page table lock.
2096  */
2097 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2098 {
2099         pmd_t *new;
2100
2101         spin_unlock(&mm->page_table_lock);
2102         new = pmd_alloc_one(mm, address);
2103         spin_lock(&mm->page_table_lock);
2104         if (!new)
2105                 return NULL;
2106
2107         /*
2108          * Because we dropped the lock, we should re-check the
2109          * entry, as somebody else could have populated it..
2110          */
2111 #ifndef __ARCH_HAS_4LEVEL_HACK
2112         if (pud_present(*pud)) {
2113                 pmd_free(new);
2114                 goto out;
2115         }
2116         pud_populate(mm, pud, new);
2117 #else
2118         if (pgd_present(*pud)) {
2119                 pmd_free(new);
2120                 goto out;
2121         }
2122         pgd_populate(mm, pud, new);
2123 #endif /* __ARCH_HAS_4LEVEL_HACK */
2124
2125  out:
2126         return pmd_offset(pud, address);
2127 }
2128 #endif /* __PAGETABLE_PMD_FOLDED */
2129
2130 int make_pages_present(unsigned long addr, unsigned long end)
2131 {
2132         int ret, len, write;
2133         struct vm_area_struct * vma;
2134
2135         vma = find_vma(current->mm, addr);
2136         if (!vma)
2137                 return -1;
2138         write = (vma->vm_flags & VM_WRITE) != 0;
2139         if (addr >= end)
2140                 BUG();
2141         if (end > vma->vm_end)
2142                 BUG();
2143         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2144         ret = get_user_pages(current, current->mm, addr,
2145                         len, write, 0, NULL, NULL);
2146         if (ret < 0)
2147                 return ret;
2148         return ret == len ? 0 : -1;
2149 }
2150
2151 /* 
2152  * Map a vmalloc()-space virtual address to the physical page.
2153  */
2154 struct page * vmalloc_to_page(void * vmalloc_addr)
2155 {
2156         unsigned long addr = (unsigned long) vmalloc_addr;
2157         struct page *page = NULL;
2158         pgd_t *pgd = pgd_offset_k(addr);
2159         pud_t *pud;
2160         pmd_t *pmd;
2161         pte_t *ptep, pte;
2162   
2163         if (!pgd_none(*pgd)) {
2164                 pud = pud_offset(pgd, addr);
2165                 if (!pud_none(*pud)) {
2166                         pmd = pmd_offset(pud, addr);
2167                         if (!pmd_none(*pmd)) {
2168                                 ptep = pte_offset_map(pmd, addr);
2169                                 pte = *ptep;
2170                                 if (pte_present(pte))
2171                                         page = pte_page(pte);
2172                                 pte_unmap(ptep);
2173                         }
2174                 }
2175         }
2176         return page;
2177 }
2178
2179 EXPORT_SYMBOL(vmalloc_to_page);
2180
2181 /*
2182  * Map a vmalloc()-space virtual address to the physical page frame number.
2183  */
2184 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2185 {
2186         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2187 }
2188
2189 EXPORT_SYMBOL(vmalloc_to_pfn);
2190
2191 /*
2192  * update_mem_hiwater
2193  *      - update per process rss and vm high water data
2194  */
2195 void update_mem_hiwater(struct task_struct *tsk)
2196 {
2197         if (tsk->mm) {
2198                 unsigned long rss = get_mm_rss(tsk->mm);
2199
2200                 if (tsk->mm->hiwater_rss < rss)
2201                         tsk->mm->hiwater_rss = rss;
2202                 if (tsk->mm->hiwater_vm < tsk->mm->total_vm)
2203                         tsk->mm->hiwater_vm = tsk->mm->total_vm;
2204         }
2205 }
2206
2207 #if !defined(__HAVE_ARCH_GATE_AREA)
2208
2209 #if defined(AT_SYSINFO_EHDR)
2210 static struct vm_area_struct gate_vma;
2211
2212 static int __init gate_vma_init(void)
2213 {
2214         gate_vma.vm_mm = NULL;
2215         gate_vma.vm_start = FIXADDR_USER_START;
2216         gate_vma.vm_end = FIXADDR_USER_END;
2217         gate_vma.vm_page_prot = PAGE_READONLY;
2218         gate_vma.vm_flags = 0;
2219         return 0;
2220 }
2221 __initcall(gate_vma_init);
2222 #endif
2223
2224 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2225 {
2226 #ifdef AT_SYSINFO_EHDR
2227         return &gate_vma;
2228 #else
2229         return NULL;
2230 #endif
2231 }
2232
2233 int in_gate_area_no_task(unsigned long addr)
2234 {
2235 #ifdef AT_SYSINFO_EHDR
2236         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2237                 return 1;
2238 #endif
2239         return 0;
2240 }
2241
2242 #endif  /* __HAVE_ARCH_GATE_AREA */