VM: add common helper function to create the page tables
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
54 #include <asm/tlb.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
57
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
65
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
69
70 unsigned long num_physpages;
71 /*
72  * A number of key systems in x86 including ioremap() rely on the assumption
73  * that high_memory defines the upper bound on direct map memory, then end
74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76  * and ZONE_HIGHMEM.
77  */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
80
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85 /*
86  * If a p?d_bad entry is found while walking page tables, report
87  * the error, before resetting entry to p?d_none.  Usually (but
88  * very seldom) called out from the p?d_none_or_clear_bad macros.
89  */
90
91 void pgd_clear_bad(pgd_t *pgd)
92 {
93         pgd_ERROR(*pgd);
94         pgd_clear(pgd);
95 }
96
97 void pud_clear_bad(pud_t *pud)
98 {
99         pud_ERROR(*pud);
100         pud_clear(pud);
101 }
102
103 void pmd_clear_bad(pmd_t *pmd)
104 {
105         pmd_ERROR(*pmd);
106         pmd_clear(pmd);
107 }
108
109 /*
110  * Note: this doesn't free the actual pages themselves. That
111  * has been handled earlier when unmapping all the memory regions.
112  */
113 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
114 {
115         struct page *page = pmd_page(*pmd);
116         pmd_clear(pmd);
117         pte_lock_deinit(page);
118         pte_free_tlb(tlb, page);
119         dec_page_state(nr_page_table_pages);
120         tlb->mm->nr_ptes--;
121 }
122
123 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
124                                 unsigned long addr, unsigned long end,
125                                 unsigned long floor, unsigned long ceiling)
126 {
127         pmd_t *pmd;
128         unsigned long next;
129         unsigned long start;
130
131         start = addr;
132         pmd = pmd_offset(pud, addr);
133         do {
134                 next = pmd_addr_end(addr, end);
135                 if (pmd_none_or_clear_bad(pmd))
136                         continue;
137                 free_pte_range(tlb, pmd);
138         } while (pmd++, addr = next, addr != end);
139
140         start &= PUD_MASK;
141         if (start < floor)
142                 return;
143         if (ceiling) {
144                 ceiling &= PUD_MASK;
145                 if (!ceiling)
146                         return;
147         }
148         if (end - 1 > ceiling - 1)
149                 return;
150
151         pmd = pmd_offset(pud, start);
152         pud_clear(pud);
153         pmd_free_tlb(tlb, pmd);
154 }
155
156 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
157                                 unsigned long addr, unsigned long end,
158                                 unsigned long floor, unsigned long ceiling)
159 {
160         pud_t *pud;
161         unsigned long next;
162         unsigned long start;
163
164         start = addr;
165         pud = pud_offset(pgd, addr);
166         do {
167                 next = pud_addr_end(addr, end);
168                 if (pud_none_or_clear_bad(pud))
169                         continue;
170                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
171         } while (pud++, addr = next, addr != end);
172
173         start &= PGDIR_MASK;
174         if (start < floor)
175                 return;
176         if (ceiling) {
177                 ceiling &= PGDIR_MASK;
178                 if (!ceiling)
179                         return;
180         }
181         if (end - 1 > ceiling - 1)
182                 return;
183
184         pud = pud_offset(pgd, start);
185         pgd_clear(pgd);
186         pud_free_tlb(tlb, pud);
187 }
188
189 /*
190  * This function frees user-level page tables of a process.
191  *
192  * Must be called with pagetable lock held.
193  */
194 void free_pgd_range(struct mmu_gather **tlb,
195                         unsigned long addr, unsigned long end,
196                         unsigned long floor, unsigned long ceiling)
197 {
198         pgd_t *pgd;
199         unsigned long next;
200         unsigned long start;
201
202         /*
203          * The next few lines have given us lots of grief...
204          *
205          * Why are we testing PMD* at this top level?  Because often
206          * there will be no work to do at all, and we'd prefer not to
207          * go all the way down to the bottom just to discover that.
208          *
209          * Why all these "- 1"s?  Because 0 represents both the bottom
210          * of the address space and the top of it (using -1 for the
211          * top wouldn't help much: the masks would do the wrong thing).
212          * The rule is that addr 0 and floor 0 refer to the bottom of
213          * the address space, but end 0 and ceiling 0 refer to the top
214          * Comparisons need to use "end - 1" and "ceiling - 1" (though
215          * that end 0 case should be mythical).
216          *
217          * Wherever addr is brought up or ceiling brought down, we must
218          * be careful to reject "the opposite 0" before it confuses the
219          * subsequent tests.  But what about where end is brought down
220          * by PMD_SIZE below? no, end can't go down to 0 there.
221          *
222          * Whereas we round start (addr) and ceiling down, by different
223          * masks at different levels, in order to test whether a table
224          * now has no other vmas using it, so can be freed, we don't
225          * bother to round floor or end up - the tests don't need that.
226          */
227
228         addr &= PMD_MASK;
229         if (addr < floor) {
230                 addr += PMD_SIZE;
231                 if (!addr)
232                         return;
233         }
234         if (ceiling) {
235                 ceiling &= PMD_MASK;
236                 if (!ceiling)
237                         return;
238         }
239         if (end - 1 > ceiling - 1)
240                 end -= PMD_SIZE;
241         if (addr > end - 1)
242                 return;
243
244         start = addr;
245         pgd = pgd_offset((*tlb)->mm, addr);
246         do {
247                 next = pgd_addr_end(addr, end);
248                 if (pgd_none_or_clear_bad(pgd))
249                         continue;
250                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
251         } while (pgd++, addr = next, addr != end);
252
253         if (!(*tlb)->fullmm)
254                 flush_tlb_pgtables((*tlb)->mm, start, end);
255 }
256
257 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
258                 unsigned long floor, unsigned long ceiling)
259 {
260         while (vma) {
261                 struct vm_area_struct *next = vma->vm_next;
262                 unsigned long addr = vma->vm_start;
263
264                 /*
265                  * Hide vma from rmap and vmtruncate before freeing pgtables
266                  */
267                 anon_vma_unlink(vma);
268                 unlink_file_vma(vma);
269
270                 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
271                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
272                                 floor, next? next->vm_start: ceiling);
273                 } else {
274                         /*
275                          * Optimization: gather nearby vmas into one call down
276                          */
277                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
278                           && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
279                                                         HPAGE_SIZE)) {
280                                 vma = next;
281                                 next = vma->vm_next;
282                                 anon_vma_unlink(vma);
283                                 unlink_file_vma(vma);
284                         }
285                         free_pgd_range(tlb, addr, vma->vm_end,
286                                 floor, next? next->vm_start: ceiling);
287                 }
288                 vma = next;
289         }
290 }
291
292 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
293 {
294         struct page *new = pte_alloc_one(mm, address);
295         if (!new)
296                 return -ENOMEM;
297
298         pte_lock_init(new);
299         spin_lock(&mm->page_table_lock);
300         if (pmd_present(*pmd)) {        /* Another has populated it */
301                 pte_lock_deinit(new);
302                 pte_free(new);
303         } else {
304                 mm->nr_ptes++;
305                 inc_page_state(nr_page_table_pages);
306                 pmd_populate(mm, pmd, new);
307         }
308         spin_unlock(&mm->page_table_lock);
309         return 0;
310 }
311
312 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
313 {
314         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
315         if (!new)
316                 return -ENOMEM;
317
318         spin_lock(&init_mm.page_table_lock);
319         if (pmd_present(*pmd))          /* Another has populated it */
320                 pte_free_kernel(new);
321         else
322                 pmd_populate_kernel(&init_mm, pmd, new);
323         spin_unlock(&init_mm.page_table_lock);
324         return 0;
325 }
326
327 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
328 {
329         if (file_rss)
330                 add_mm_counter(mm, file_rss, file_rss);
331         if (anon_rss)
332                 add_mm_counter(mm, anon_rss, anon_rss);
333 }
334
335 /*
336  * This function is called to print an error when a bad pte
337  * is found. For example, we might have a PFN-mapped pte in
338  * a region that doesn't allow it.
339  *
340  * The calling function must still handle the error.
341  */
342 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
343 {
344         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
345                         "vm_flags = %lx, vaddr = %lx\n",
346                 (long long)pte_val(pte),
347                 (vma->vm_mm == current->mm ? current->comm : "???"),
348                 vma->vm_flags, vaddr);
349         dump_stack();
350 }
351
352 /*
353  * This function gets the "struct page" associated with a pte.
354  *
355  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
356  * will have each page table entry just pointing to a raw page frame
357  * number, and as far as the VM layer is concerned, those do not have
358  * pages associated with them - even if the PFN might point to memory
359  * that otherwise is perfectly fine and has a "struct page".
360  *
361  * The way we recognize those mappings is through the rules set up
362  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
363  * and the vm_pgoff will point to the first PFN mapped: thus every
364  * page that is a raw mapping will always honor the rule
365  *
366  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
367  *
368  * and if that isn't true, the page has been COW'ed (in which case it
369  * _does_ have a "struct page" associated with it even if it is in a
370  * VM_PFNMAP range).
371  */
372 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
373 {
374         unsigned long pfn = pte_pfn(pte);
375
376         if (vma->vm_flags & VM_PFNMAP) {
377                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
378                 if (pfn == vma->vm_pgoff + off)
379                         return NULL;
380         }
381
382         /*
383          * Add some anal sanity checks for now. Eventually,
384          * we should just do "return pfn_to_page(pfn)", but
385          * in the meantime we check that we get a valid pfn,
386          * and that the resulting page looks ok.
387          *
388          * Remove this test eventually!
389          */
390         if (unlikely(!pfn_valid(pfn))) {
391                 print_bad_pte(vma, pte, addr);
392                 return NULL;
393         }
394
395         /*
396          * NOTE! We still have PageReserved() pages in the page 
397          * tables. 
398          *
399          * The PAGE_ZERO() pages and various VDSO mappings can
400          * cause them to exist.
401          */
402         return pfn_to_page(pfn);
403 }
404
405 /*
406  * copy one vm_area from one task to the other. Assumes the page tables
407  * already present in the new task to be cleared in the whole range
408  * covered by this vma.
409  */
410
411 static inline void
412 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
413                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
414                 unsigned long addr, int *rss)
415 {
416         unsigned long vm_flags = vma->vm_flags;
417         pte_t pte = *src_pte;
418         struct page *page;
419
420         /* pte contains position in swap or file, so copy. */
421         if (unlikely(!pte_present(pte))) {
422                 if (!pte_file(pte)) {
423                         swap_duplicate(pte_to_swp_entry(pte));
424                         /* make sure dst_mm is on swapoff's mmlist. */
425                         if (unlikely(list_empty(&dst_mm->mmlist))) {
426                                 spin_lock(&mmlist_lock);
427                                 if (list_empty(&dst_mm->mmlist))
428                                         list_add(&dst_mm->mmlist,
429                                                  &src_mm->mmlist);
430                                 spin_unlock(&mmlist_lock);
431                         }
432                 }
433                 goto out_set_pte;
434         }
435
436         /*
437          * If it's a COW mapping, write protect it both
438          * in the parent and the child
439          */
440         if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
441                 ptep_set_wrprotect(src_mm, addr, src_pte);
442                 pte = *src_pte;
443         }
444
445         /*
446          * If it's a shared mapping, mark it clean in
447          * the child
448          */
449         if (vm_flags & VM_SHARED)
450                 pte = pte_mkclean(pte);
451         pte = pte_mkold(pte);
452
453         page = vm_normal_page(vma, addr, pte);
454         if (page) {
455                 get_page(page);
456                 page_dup_rmap(page);
457                 rss[!!PageAnon(page)]++;
458         }
459
460 out_set_pte:
461         set_pte_at(dst_mm, addr, dst_pte, pte);
462 }
463
464 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
465                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
466                 unsigned long addr, unsigned long end)
467 {
468         pte_t *src_pte, *dst_pte;
469         spinlock_t *src_ptl, *dst_ptl;
470         int progress = 0;
471         int rss[2];
472
473 again:
474         rss[1] = rss[0] = 0;
475         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
476         if (!dst_pte)
477                 return -ENOMEM;
478         src_pte = pte_offset_map_nested(src_pmd, addr);
479         src_ptl = pte_lockptr(src_mm, src_pmd);
480         spin_lock(src_ptl);
481
482         do {
483                 /*
484                  * We are holding two locks at this point - either of them
485                  * could generate latencies in another task on another CPU.
486                  */
487                 if (progress >= 32) {
488                         progress = 0;
489                         if (need_resched() ||
490                             need_lockbreak(src_ptl) ||
491                             need_lockbreak(dst_ptl))
492                                 break;
493                 }
494                 if (pte_none(*src_pte)) {
495                         progress++;
496                         continue;
497                 }
498                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
499                 progress += 8;
500         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
501
502         spin_unlock(src_ptl);
503         pte_unmap_nested(src_pte - 1);
504         add_mm_rss(dst_mm, rss[0], rss[1]);
505         pte_unmap_unlock(dst_pte - 1, dst_ptl);
506         cond_resched();
507         if (addr != end)
508                 goto again;
509         return 0;
510 }
511
512 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
513                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
514                 unsigned long addr, unsigned long end)
515 {
516         pmd_t *src_pmd, *dst_pmd;
517         unsigned long next;
518
519         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
520         if (!dst_pmd)
521                 return -ENOMEM;
522         src_pmd = pmd_offset(src_pud, addr);
523         do {
524                 next = pmd_addr_end(addr, end);
525                 if (pmd_none_or_clear_bad(src_pmd))
526                         continue;
527                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
528                                                 vma, addr, next))
529                         return -ENOMEM;
530         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
531         return 0;
532 }
533
534 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
535                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
536                 unsigned long addr, unsigned long end)
537 {
538         pud_t *src_pud, *dst_pud;
539         unsigned long next;
540
541         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
542         if (!dst_pud)
543                 return -ENOMEM;
544         src_pud = pud_offset(src_pgd, addr);
545         do {
546                 next = pud_addr_end(addr, end);
547                 if (pud_none_or_clear_bad(src_pud))
548                         continue;
549                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
550                                                 vma, addr, next))
551                         return -ENOMEM;
552         } while (dst_pud++, src_pud++, addr = next, addr != end);
553         return 0;
554 }
555
556 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
557                 struct vm_area_struct *vma)
558 {
559         pgd_t *src_pgd, *dst_pgd;
560         unsigned long next;
561         unsigned long addr = vma->vm_start;
562         unsigned long end = vma->vm_end;
563
564         /*
565          * Don't copy ptes where a page fault will fill them correctly.
566          * Fork becomes much lighter when there are big shared or private
567          * readonly mappings. The tradeoff is that copy_page_range is more
568          * efficient than faulting.
569          */
570         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP))) {
571                 if (!vma->anon_vma)
572                         return 0;
573         }
574
575         if (is_vm_hugetlb_page(vma))
576                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
577
578         dst_pgd = pgd_offset(dst_mm, addr);
579         src_pgd = pgd_offset(src_mm, addr);
580         do {
581                 next = pgd_addr_end(addr, end);
582                 if (pgd_none_or_clear_bad(src_pgd))
583                         continue;
584                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
585                                                 vma, addr, next))
586                         return -ENOMEM;
587         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
588         return 0;
589 }
590
591 static unsigned long zap_pte_range(struct mmu_gather *tlb,
592                                 struct vm_area_struct *vma, pmd_t *pmd,
593                                 unsigned long addr, unsigned long end,
594                                 long *zap_work, struct zap_details *details)
595 {
596         struct mm_struct *mm = tlb->mm;
597         pte_t *pte;
598         spinlock_t *ptl;
599         int file_rss = 0;
600         int anon_rss = 0;
601
602         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
603         do {
604                 pte_t ptent = *pte;
605                 if (pte_none(ptent)) {
606                         (*zap_work)--;
607                         continue;
608                 }
609                 if (pte_present(ptent)) {
610                         struct page *page;
611
612                         (*zap_work) -= PAGE_SIZE;
613
614                         page = vm_normal_page(vma, addr, ptent);
615                         if (unlikely(details) && page) {
616                                 /*
617                                  * unmap_shared_mapping_pages() wants to
618                                  * invalidate cache without truncating:
619                                  * unmap shared but keep private pages.
620                                  */
621                                 if (details->check_mapping &&
622                                     details->check_mapping != page->mapping)
623                                         continue;
624                                 /*
625                                  * Each page->index must be checked when
626                                  * invalidating or truncating nonlinear.
627                                  */
628                                 if (details->nonlinear_vma &&
629                                     (page->index < details->first_index ||
630                                      page->index > details->last_index))
631                                         continue;
632                         }
633                         ptent = ptep_get_and_clear_full(mm, addr, pte,
634                                                         tlb->fullmm);
635                         tlb_remove_tlb_entry(tlb, pte, addr);
636                         if (unlikely(!page))
637                                 continue;
638                         if (unlikely(details) && details->nonlinear_vma
639                             && linear_page_index(details->nonlinear_vma,
640                                                 addr) != page->index)
641                                 set_pte_at(mm, addr, pte,
642                                            pgoff_to_pte(page->index));
643                         if (PageAnon(page))
644                                 anon_rss--;
645                         else {
646                                 if (pte_dirty(ptent))
647                                         set_page_dirty(page);
648                                 if (pte_young(ptent))
649                                         mark_page_accessed(page);
650                                 file_rss--;
651                         }
652                         page_remove_rmap(page);
653                         tlb_remove_page(tlb, page);
654                         continue;
655                 }
656                 /*
657                  * If details->check_mapping, we leave swap entries;
658                  * if details->nonlinear_vma, we leave file entries.
659                  */
660                 if (unlikely(details))
661                         continue;
662                 if (!pte_file(ptent))
663                         free_swap_and_cache(pte_to_swp_entry(ptent));
664                 pte_clear_full(mm, addr, pte, tlb->fullmm);
665         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
666
667         add_mm_rss(mm, file_rss, anon_rss);
668         pte_unmap_unlock(pte - 1, ptl);
669
670         return addr;
671 }
672
673 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
674                                 struct vm_area_struct *vma, pud_t *pud,
675                                 unsigned long addr, unsigned long end,
676                                 long *zap_work, struct zap_details *details)
677 {
678         pmd_t *pmd;
679         unsigned long next;
680
681         pmd = pmd_offset(pud, addr);
682         do {
683                 next = pmd_addr_end(addr, end);
684                 if (pmd_none_or_clear_bad(pmd)) {
685                         (*zap_work)--;
686                         continue;
687                 }
688                 next = zap_pte_range(tlb, vma, pmd, addr, next,
689                                                 zap_work, details);
690         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
691
692         return addr;
693 }
694
695 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
696                                 struct vm_area_struct *vma, pgd_t *pgd,
697                                 unsigned long addr, unsigned long end,
698                                 long *zap_work, struct zap_details *details)
699 {
700         pud_t *pud;
701         unsigned long next;
702
703         pud = pud_offset(pgd, addr);
704         do {
705                 next = pud_addr_end(addr, end);
706                 if (pud_none_or_clear_bad(pud)) {
707                         (*zap_work)--;
708                         continue;
709                 }
710                 next = zap_pmd_range(tlb, vma, pud, addr, next,
711                                                 zap_work, details);
712         } while (pud++, addr = next, (addr != end && *zap_work > 0));
713
714         return addr;
715 }
716
717 static unsigned long unmap_page_range(struct mmu_gather *tlb,
718                                 struct vm_area_struct *vma,
719                                 unsigned long addr, unsigned long end,
720                                 long *zap_work, struct zap_details *details)
721 {
722         pgd_t *pgd;
723         unsigned long next;
724
725         if (details && !details->check_mapping && !details->nonlinear_vma)
726                 details = NULL;
727
728         BUG_ON(addr >= end);
729         tlb_start_vma(tlb, vma);
730         pgd = pgd_offset(vma->vm_mm, addr);
731         do {
732                 next = pgd_addr_end(addr, end);
733                 if (pgd_none_or_clear_bad(pgd)) {
734                         (*zap_work)--;
735                         continue;
736                 }
737                 next = zap_pud_range(tlb, vma, pgd, addr, next,
738                                                 zap_work, details);
739         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
740         tlb_end_vma(tlb, vma);
741
742         return addr;
743 }
744
745 #ifdef CONFIG_PREEMPT
746 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
747 #else
748 /* No preempt: go for improved straight-line efficiency */
749 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
750 #endif
751
752 /**
753  * unmap_vmas - unmap a range of memory covered by a list of vma's
754  * @tlbp: address of the caller's struct mmu_gather
755  * @vma: the starting vma
756  * @start_addr: virtual address at which to start unmapping
757  * @end_addr: virtual address at which to end unmapping
758  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
759  * @details: details of nonlinear truncation or shared cache invalidation
760  *
761  * Returns the end address of the unmapping (restart addr if interrupted).
762  *
763  * Unmap all pages in the vma list.
764  *
765  * We aim to not hold locks for too long (for scheduling latency reasons).
766  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
767  * return the ending mmu_gather to the caller.
768  *
769  * Only addresses between `start' and `end' will be unmapped.
770  *
771  * The VMA list must be sorted in ascending virtual address order.
772  *
773  * unmap_vmas() assumes that the caller will flush the whole unmapped address
774  * range after unmap_vmas() returns.  So the only responsibility here is to
775  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
776  * drops the lock and schedules.
777  */
778 unsigned long unmap_vmas(struct mmu_gather **tlbp,
779                 struct vm_area_struct *vma, unsigned long start_addr,
780                 unsigned long end_addr, unsigned long *nr_accounted,
781                 struct zap_details *details)
782 {
783         long zap_work = ZAP_BLOCK_SIZE;
784         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
785         int tlb_start_valid = 0;
786         unsigned long start = start_addr;
787         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
788         int fullmm = (*tlbp)->fullmm;
789
790         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
791                 unsigned long end;
792
793                 start = max(vma->vm_start, start_addr);
794                 if (start >= vma->vm_end)
795                         continue;
796                 end = min(vma->vm_end, end_addr);
797                 if (end <= vma->vm_start)
798                         continue;
799
800                 if (vma->vm_flags & VM_ACCOUNT)
801                         *nr_accounted += (end - start) >> PAGE_SHIFT;
802
803                 while (start != end) {
804                         if (!tlb_start_valid) {
805                                 tlb_start = start;
806                                 tlb_start_valid = 1;
807                         }
808
809                         if (unlikely(is_vm_hugetlb_page(vma))) {
810                                 unmap_hugepage_range(vma, start, end);
811                                 zap_work -= (end - start) /
812                                                 (HPAGE_SIZE / PAGE_SIZE);
813                                 start = end;
814                         } else
815                                 start = unmap_page_range(*tlbp, vma,
816                                                 start, end, &zap_work, details);
817
818                         if (zap_work > 0) {
819                                 BUG_ON(start != end);
820                                 break;
821                         }
822
823                         tlb_finish_mmu(*tlbp, tlb_start, start);
824
825                         if (need_resched() ||
826                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
827                                 if (i_mmap_lock) {
828                                         *tlbp = NULL;
829                                         goto out;
830                                 }
831                                 cond_resched();
832                         }
833
834                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
835                         tlb_start_valid = 0;
836                         zap_work = ZAP_BLOCK_SIZE;
837                 }
838         }
839 out:
840         return start;   /* which is now the end (or restart) address */
841 }
842
843 /**
844  * zap_page_range - remove user pages in a given range
845  * @vma: vm_area_struct holding the applicable pages
846  * @address: starting address of pages to zap
847  * @size: number of bytes to zap
848  * @details: details of nonlinear truncation or shared cache invalidation
849  */
850 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
851                 unsigned long size, struct zap_details *details)
852 {
853         struct mm_struct *mm = vma->vm_mm;
854         struct mmu_gather *tlb;
855         unsigned long end = address + size;
856         unsigned long nr_accounted = 0;
857
858         lru_add_drain();
859         tlb = tlb_gather_mmu(mm, 0);
860         update_hiwater_rss(mm);
861         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
862         if (tlb)
863                 tlb_finish_mmu(tlb, address, end);
864         return end;
865 }
866
867 /*
868  * Do a quick page-table lookup for a single page.
869  */
870 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
871                         unsigned int flags)
872 {
873         pgd_t *pgd;
874         pud_t *pud;
875         pmd_t *pmd;
876         pte_t *ptep, pte;
877         spinlock_t *ptl;
878         struct page *page;
879         struct mm_struct *mm = vma->vm_mm;
880
881         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
882         if (!IS_ERR(page)) {
883                 BUG_ON(flags & FOLL_GET);
884                 goto out;
885         }
886
887         page = NULL;
888         pgd = pgd_offset(mm, address);
889         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
890                 goto no_page_table;
891
892         pud = pud_offset(pgd, address);
893         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
894                 goto no_page_table;
895         
896         pmd = pmd_offset(pud, address);
897         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
898                 goto no_page_table;
899
900         if (pmd_huge(*pmd)) {
901                 BUG_ON(flags & FOLL_GET);
902                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
903                 goto out;
904         }
905
906         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
907         if (!ptep)
908                 goto out;
909
910         pte = *ptep;
911         if (!pte_present(pte))
912                 goto unlock;
913         if ((flags & FOLL_WRITE) && !pte_write(pte))
914                 goto unlock;
915         page = vm_normal_page(vma, address, pte);
916         if (unlikely(!page))
917                 goto unlock;
918
919         if (flags & FOLL_GET)
920                 get_page(page);
921         if (flags & FOLL_TOUCH) {
922                 if ((flags & FOLL_WRITE) &&
923                     !pte_dirty(pte) && !PageDirty(page))
924                         set_page_dirty(page);
925                 mark_page_accessed(page);
926         }
927 unlock:
928         pte_unmap_unlock(ptep, ptl);
929 out:
930         return page;
931
932 no_page_table:
933         /*
934          * When core dumping an enormous anonymous area that nobody
935          * has touched so far, we don't want to allocate page tables.
936          */
937         if (flags & FOLL_ANON) {
938                 page = ZERO_PAGE(address);
939                 if (flags & FOLL_GET)
940                         get_page(page);
941                 BUG_ON(flags & FOLL_WRITE);
942         }
943         return page;
944 }
945
946 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
947                 unsigned long start, int len, int write, int force,
948                 struct page **pages, struct vm_area_struct **vmas)
949 {
950         int i;
951         unsigned int vm_flags;
952
953         /* 
954          * Require read or write permissions.
955          * If 'force' is set, we only require the "MAY" flags.
956          */
957         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
958         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
959         i = 0;
960
961         do {
962                 struct vm_area_struct *vma;
963                 unsigned int foll_flags;
964
965                 vma = find_extend_vma(mm, start);
966                 if (!vma && in_gate_area(tsk, start)) {
967                         unsigned long pg = start & PAGE_MASK;
968                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
969                         pgd_t *pgd;
970                         pud_t *pud;
971                         pmd_t *pmd;
972                         pte_t *pte;
973                         if (write) /* user gate pages are read-only */
974                                 return i ? : -EFAULT;
975                         if (pg > TASK_SIZE)
976                                 pgd = pgd_offset_k(pg);
977                         else
978                                 pgd = pgd_offset_gate(mm, pg);
979                         BUG_ON(pgd_none(*pgd));
980                         pud = pud_offset(pgd, pg);
981                         BUG_ON(pud_none(*pud));
982                         pmd = pmd_offset(pud, pg);
983                         if (pmd_none(*pmd))
984                                 return i ? : -EFAULT;
985                         pte = pte_offset_map(pmd, pg);
986                         if (pte_none(*pte)) {
987                                 pte_unmap(pte);
988                                 return i ? : -EFAULT;
989                         }
990                         if (pages) {
991                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
992                                 pages[i] = page;
993                                 if (page)
994                                         get_page(page);
995                         }
996                         pte_unmap(pte);
997                         if (vmas)
998                                 vmas[i] = gate_vma;
999                         i++;
1000                         start += PAGE_SIZE;
1001                         len--;
1002                         continue;
1003                 }
1004
1005                 if (!vma || (vma->vm_flags & VM_IO)
1006                                 || !(vm_flags & vma->vm_flags))
1007                         return i ? : -EFAULT;
1008
1009                 if (is_vm_hugetlb_page(vma)) {
1010                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1011                                                 &start, &len, i);
1012                         continue;
1013                 }
1014
1015                 foll_flags = FOLL_TOUCH;
1016                 if (pages)
1017                         foll_flags |= FOLL_GET;
1018                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1019                     (!vma->vm_ops || !vma->vm_ops->nopage))
1020                         foll_flags |= FOLL_ANON;
1021
1022                 do {
1023                         struct page *page;
1024
1025                         if (write)
1026                                 foll_flags |= FOLL_WRITE;
1027
1028                         cond_resched();
1029                         while (!(page = follow_page(vma, start, foll_flags))) {
1030                                 int ret;
1031                                 ret = __handle_mm_fault(mm, vma, start,
1032                                                 foll_flags & FOLL_WRITE);
1033                                 /*
1034                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1035                                  * broken COW when necessary, even if maybe_mkwrite
1036                                  * decided not to set pte_write. We can thus safely do
1037                                  * subsequent page lookups as if they were reads.
1038                                  */
1039                                 if (ret & VM_FAULT_WRITE)
1040                                         foll_flags &= ~FOLL_WRITE;
1041                                 
1042                                 switch (ret & ~VM_FAULT_WRITE) {
1043                                 case VM_FAULT_MINOR:
1044                                         tsk->min_flt++;
1045                                         break;
1046                                 case VM_FAULT_MAJOR:
1047                                         tsk->maj_flt++;
1048                                         break;
1049                                 case VM_FAULT_SIGBUS:
1050                                         return i ? i : -EFAULT;
1051                                 case VM_FAULT_OOM:
1052                                         return i ? i : -ENOMEM;
1053                                 default:
1054                                         BUG();
1055                                 }
1056                         }
1057                         if (pages) {
1058                                 pages[i] = page;
1059                                 flush_dcache_page(page);
1060                         }
1061                         if (vmas)
1062                                 vmas[i] = vma;
1063                         i++;
1064                         start += PAGE_SIZE;
1065                         len--;
1066                 } while (len && start < vma->vm_end);
1067         } while (len);
1068         return i;
1069 }
1070 EXPORT_SYMBOL(get_user_pages);
1071
1072 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1073                         unsigned long addr, unsigned long end, pgprot_t prot)
1074 {
1075         pte_t *pte;
1076         spinlock_t *ptl;
1077
1078         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1079         if (!pte)
1080                 return -ENOMEM;
1081         do {
1082                 struct page *page = ZERO_PAGE(addr);
1083                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1084                 page_cache_get(page);
1085                 page_add_file_rmap(page);
1086                 inc_mm_counter(mm, file_rss);
1087                 BUG_ON(!pte_none(*pte));
1088                 set_pte_at(mm, addr, pte, zero_pte);
1089         } while (pte++, addr += PAGE_SIZE, addr != end);
1090         pte_unmap_unlock(pte - 1, ptl);
1091         return 0;
1092 }
1093
1094 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1095                         unsigned long addr, unsigned long end, pgprot_t prot)
1096 {
1097         pmd_t *pmd;
1098         unsigned long next;
1099
1100         pmd = pmd_alloc(mm, pud, addr);
1101         if (!pmd)
1102                 return -ENOMEM;
1103         do {
1104                 next = pmd_addr_end(addr, end);
1105                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1106                         return -ENOMEM;
1107         } while (pmd++, addr = next, addr != end);
1108         return 0;
1109 }
1110
1111 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1112                         unsigned long addr, unsigned long end, pgprot_t prot)
1113 {
1114         pud_t *pud;
1115         unsigned long next;
1116
1117         pud = pud_alloc(mm, pgd, addr);
1118         if (!pud)
1119                 return -ENOMEM;
1120         do {
1121                 next = pud_addr_end(addr, end);
1122                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1123                         return -ENOMEM;
1124         } while (pud++, addr = next, addr != end);
1125         return 0;
1126 }
1127
1128 int zeromap_page_range(struct vm_area_struct *vma,
1129                         unsigned long addr, unsigned long size, pgprot_t prot)
1130 {
1131         pgd_t *pgd;
1132         unsigned long next;
1133         unsigned long end = addr + size;
1134         struct mm_struct *mm = vma->vm_mm;
1135         int err;
1136
1137         BUG_ON(addr >= end);
1138         pgd = pgd_offset(mm, addr);
1139         flush_cache_range(vma, addr, end);
1140         do {
1141                 next = pgd_addr_end(addr, end);
1142                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1143                 if (err)
1144                         break;
1145         } while (pgd++, addr = next, addr != end);
1146         return err;
1147 }
1148
1149 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1150 {
1151         pgd_t * pgd = pgd_offset(mm, addr);
1152         pud_t * pud = pud_alloc(mm, pgd, addr);
1153         if (pud) {
1154                 pmd_t * pmd = pmd_alloc(mm, pgd, addr);
1155                 if (pmd)
1156                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1157         }
1158         return NULL;
1159 }
1160
1161 /*
1162  * This is the old fallback for page remapping.
1163  *
1164  * For historical reasons, it only allows reserved pages. Only
1165  * old drivers should use this, and they needed to mark their
1166  * pages reserved for the old functions anyway.
1167  */
1168 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1169 {
1170         int retval;
1171         pte_t *pte;
1172         spinlock_t *ptl;  
1173
1174         retval = -EINVAL;
1175         if (PageAnon(page) || !PageReserved(page))
1176                 goto out;
1177         retval = -ENOMEM;
1178         flush_dcache_page(page);
1179         pte = get_locked_pte(mm, addr, &ptl);
1180         if (!pte)
1181                 goto out;
1182         retval = -EBUSY;
1183         if (!pte_none(*pte))
1184                 goto out_unlock;
1185
1186         /* Ok, finally just insert the thing.. */
1187         get_page(page);
1188         inc_mm_counter(mm, file_rss);
1189         page_add_file_rmap(page);
1190         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1191
1192         retval = 0;
1193 out_unlock:
1194         pte_unmap_unlock(pte, ptl);
1195 out:
1196         return retval;
1197 }
1198
1199 /*
1200  * Somebody does a pfn remapping that doesn't actually work as a vma.
1201  *
1202  * Do it as individual pages instead, and warn about it. It's bad form,
1203  * and very inefficient.
1204  */
1205 static int incomplete_pfn_remap(struct vm_area_struct *vma,
1206                 unsigned long start, unsigned long end,
1207                 unsigned long pfn, pgprot_t prot)
1208 {
1209         static int warn = 10;
1210         struct page *page;
1211         int retval;
1212
1213         if (!(vma->vm_flags & VM_INCOMPLETE)) {
1214                 if (warn) {
1215                         warn--;
1216                         printk("%s does an incomplete pfn remapping", current->comm);
1217                         dump_stack();
1218                 }
1219         }
1220         vma->vm_flags |= VM_INCOMPLETE | VM_IO | VM_RESERVED;
1221
1222         if (start < vma->vm_start || end > vma->vm_end)
1223                 return -EINVAL;
1224
1225         if (!pfn_valid(pfn))
1226                 return -EINVAL;
1227
1228         retval = 0;
1229         page = pfn_to_page(pfn);
1230         while (start < end) {
1231                 retval = insert_page(vma->vm_mm, start, page, prot);
1232                 if (retval < 0)
1233                         break;
1234                 start += PAGE_SIZE;
1235                 page++;
1236         }
1237         return retval;
1238 }
1239
1240 /*
1241  * maps a range of physical memory into the requested pages. the old
1242  * mappings are removed. any references to nonexistent pages results
1243  * in null mappings (currently treated as "copy-on-access")
1244  */
1245 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1246                         unsigned long addr, unsigned long end,
1247                         unsigned long pfn, pgprot_t prot)
1248 {
1249         pte_t *pte;
1250         spinlock_t *ptl;
1251
1252         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1253         if (!pte)
1254                 return -ENOMEM;
1255         do {
1256                 BUG_ON(!pte_none(*pte));
1257                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1258                 pfn++;
1259         } while (pte++, addr += PAGE_SIZE, addr != end);
1260         pte_unmap_unlock(pte - 1, ptl);
1261         return 0;
1262 }
1263
1264 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1265                         unsigned long addr, unsigned long end,
1266                         unsigned long pfn, pgprot_t prot)
1267 {
1268         pmd_t *pmd;
1269         unsigned long next;
1270
1271         pfn -= addr >> PAGE_SHIFT;
1272         pmd = pmd_alloc(mm, pud, addr);
1273         if (!pmd)
1274                 return -ENOMEM;
1275         do {
1276                 next = pmd_addr_end(addr, end);
1277                 if (remap_pte_range(mm, pmd, addr, next,
1278                                 pfn + (addr >> PAGE_SHIFT), prot))
1279                         return -ENOMEM;
1280         } while (pmd++, addr = next, addr != end);
1281         return 0;
1282 }
1283
1284 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1285                         unsigned long addr, unsigned long end,
1286                         unsigned long pfn, pgprot_t prot)
1287 {
1288         pud_t *pud;
1289         unsigned long next;
1290
1291         pfn -= addr >> PAGE_SHIFT;
1292         pud = pud_alloc(mm, pgd, addr);
1293         if (!pud)
1294                 return -ENOMEM;
1295         do {
1296                 next = pud_addr_end(addr, end);
1297                 if (remap_pmd_range(mm, pud, addr, next,
1298                                 pfn + (addr >> PAGE_SHIFT), prot))
1299                         return -ENOMEM;
1300         } while (pud++, addr = next, addr != end);
1301         return 0;
1302 }
1303
1304 /*  Note: this is only safe if the mm semaphore is held when called. */
1305 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1306                     unsigned long pfn, unsigned long size, pgprot_t prot)
1307 {
1308         pgd_t *pgd;
1309         unsigned long next;
1310         unsigned long end = addr + PAGE_ALIGN(size);
1311         struct mm_struct *mm = vma->vm_mm;
1312         int err;
1313
1314         if (addr != vma->vm_start || end != vma->vm_end)
1315                 return incomplete_pfn_remap(vma, addr, end, pfn, prot);
1316
1317         /*
1318          * Physically remapped pages are special. Tell the
1319          * rest of the world about it:
1320          *   VM_IO tells people not to look at these pages
1321          *      (accesses can have side effects).
1322          *   VM_RESERVED is specified all over the place, because
1323          *      in 2.4 it kept swapout's vma scan off this vma; but
1324          *      in 2.6 the LRU scan won't even find its pages, so this
1325          *      flag means no more than count its pages in reserved_vm,
1326          *      and omit it from core dump, even when VM_IO turned off.
1327          *   VM_PFNMAP tells the core MM that the base pages are just
1328          *      raw PFN mappings, and do not have a "struct page" associated
1329          *      with them.
1330          */
1331         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1332         vma->vm_pgoff = pfn;
1333
1334         BUG_ON(addr >= end);
1335         pfn -= addr >> PAGE_SHIFT;
1336         pgd = pgd_offset(mm, addr);
1337         flush_cache_range(vma, addr, end);
1338         do {
1339                 next = pgd_addr_end(addr, end);
1340                 err = remap_pud_range(mm, pgd, addr, next,
1341                                 pfn + (addr >> PAGE_SHIFT), prot);
1342                 if (err)
1343                         break;
1344         } while (pgd++, addr = next, addr != end);
1345         return err;
1346 }
1347 EXPORT_SYMBOL(remap_pfn_range);
1348
1349 /*
1350  * handle_pte_fault chooses page fault handler according to an entry
1351  * which was read non-atomically.  Before making any commitment, on
1352  * those architectures or configurations (e.g. i386 with PAE) which
1353  * might give a mix of unmatched parts, do_swap_page and do_file_page
1354  * must check under lock before unmapping the pte and proceeding
1355  * (but do_wp_page is only called after already making such a check;
1356  * and do_anonymous_page and do_no_page can safely check later on).
1357  */
1358 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1359                                 pte_t *page_table, pte_t orig_pte)
1360 {
1361         int same = 1;
1362 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1363         if (sizeof(pte_t) > sizeof(unsigned long)) {
1364                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1365                 spin_lock(ptl);
1366                 same = pte_same(*page_table, orig_pte);
1367                 spin_unlock(ptl);
1368         }
1369 #endif
1370         pte_unmap(page_table);
1371         return same;
1372 }
1373
1374 /*
1375  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1376  * servicing faults for write access.  In the normal case, do always want
1377  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1378  * that do not have writing enabled, when used by access_process_vm.
1379  */
1380 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1381 {
1382         if (likely(vma->vm_flags & VM_WRITE))
1383                 pte = pte_mkwrite(pte);
1384         return pte;
1385 }
1386
1387 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1388 {
1389         /*
1390          * If the source page was a PFN mapping, we don't have
1391          * a "struct page" for it. We do a best-effort copy by
1392          * just copying from the original user address. If that
1393          * fails, we just zero-fill it. Live with it.
1394          */
1395         if (unlikely(!src)) {
1396                 void *kaddr = kmap_atomic(dst, KM_USER0);
1397                 unsigned long left = __copy_from_user_inatomic(kaddr, (void __user *)va, PAGE_SIZE);
1398                 if (left)
1399                         memset(kaddr, 0, PAGE_SIZE);
1400                 kunmap_atomic(kaddr, KM_USER0);
1401                 return;
1402                 
1403         }
1404         copy_user_highpage(dst, src, va);
1405 }
1406
1407 /*
1408  * This routine handles present pages, when users try to write
1409  * to a shared page. It is done by copying the page to a new address
1410  * and decrementing the shared-page counter for the old page.
1411  *
1412  * Note that this routine assumes that the protection checks have been
1413  * done by the caller (the low-level page fault routine in most cases).
1414  * Thus we can safely just mark it writable once we've done any necessary
1415  * COW.
1416  *
1417  * We also mark the page dirty at this point even though the page will
1418  * change only once the write actually happens. This avoids a few races,
1419  * and potentially makes it more efficient.
1420  *
1421  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1422  * but allow concurrent faults), with pte both mapped and locked.
1423  * We return with mmap_sem still held, but pte unmapped and unlocked.
1424  */
1425 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1426                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1427                 spinlock_t *ptl, pte_t orig_pte)
1428 {
1429         struct page *old_page, *src_page, *new_page;
1430         pte_t entry;
1431         int ret = VM_FAULT_MINOR;
1432
1433         old_page = vm_normal_page(vma, address, orig_pte);
1434         src_page = old_page;
1435         if (!old_page)
1436                 goto gotten;
1437
1438         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1439                 int reuse = can_share_swap_page(old_page);
1440                 unlock_page(old_page);
1441                 if (reuse) {
1442                         flush_cache_page(vma, address, pte_pfn(orig_pte));
1443                         entry = pte_mkyoung(orig_pte);
1444                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1445                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1446                         update_mmu_cache(vma, address, entry);
1447                         lazy_mmu_prot_update(entry);
1448                         ret |= VM_FAULT_WRITE;
1449                         goto unlock;
1450                 }
1451         }
1452
1453         /*
1454          * Ok, we need to copy. Oh, well..
1455          */
1456         page_cache_get(old_page);
1457 gotten:
1458         pte_unmap_unlock(page_table, ptl);
1459
1460         if (unlikely(anon_vma_prepare(vma)))
1461                 goto oom;
1462         if (src_page == ZERO_PAGE(address)) {
1463                 new_page = alloc_zeroed_user_highpage(vma, address);
1464                 if (!new_page)
1465                         goto oom;
1466         } else {
1467                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1468                 if (!new_page)
1469                         goto oom;
1470                 cow_user_page(new_page, src_page, address);
1471         }
1472
1473         /*
1474          * Re-check the pte - we dropped the lock
1475          */
1476         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1477         if (likely(pte_same(*page_table, orig_pte))) {
1478                 if (old_page) {
1479                         page_remove_rmap(old_page);
1480                         if (!PageAnon(old_page)) {
1481                                 dec_mm_counter(mm, file_rss);
1482                                 inc_mm_counter(mm, anon_rss);
1483                         }
1484                 } else
1485                         inc_mm_counter(mm, anon_rss);
1486                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1487                 entry = mk_pte(new_page, vma->vm_page_prot);
1488                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1489                 ptep_establish(vma, address, page_table, entry);
1490                 update_mmu_cache(vma, address, entry);
1491                 lazy_mmu_prot_update(entry);
1492                 lru_cache_add_active(new_page);
1493                 page_add_anon_rmap(new_page, vma, address);
1494
1495                 /* Free the old page.. */
1496                 new_page = old_page;
1497                 ret |= VM_FAULT_WRITE;
1498         }
1499         if (new_page)
1500                 page_cache_release(new_page);
1501         if (old_page)
1502                 page_cache_release(old_page);
1503 unlock:
1504         pte_unmap_unlock(page_table, ptl);
1505         return ret;
1506 oom:
1507         if (old_page)
1508                 page_cache_release(old_page);
1509         return VM_FAULT_OOM;
1510 }
1511
1512 /*
1513  * Helper functions for unmap_mapping_range().
1514  *
1515  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1516  *
1517  * We have to restart searching the prio_tree whenever we drop the lock,
1518  * since the iterator is only valid while the lock is held, and anyway
1519  * a later vma might be split and reinserted earlier while lock dropped.
1520  *
1521  * The list of nonlinear vmas could be handled more efficiently, using
1522  * a placeholder, but handle it in the same way until a need is shown.
1523  * It is important to search the prio_tree before nonlinear list: a vma
1524  * may become nonlinear and be shifted from prio_tree to nonlinear list
1525  * while the lock is dropped; but never shifted from list to prio_tree.
1526  *
1527  * In order to make forward progress despite restarting the search,
1528  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1529  * quickly skip it next time around.  Since the prio_tree search only
1530  * shows us those vmas affected by unmapping the range in question, we
1531  * can't efficiently keep all vmas in step with mapping->truncate_count:
1532  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1533  * mapping->truncate_count and vma->vm_truncate_count are protected by
1534  * i_mmap_lock.
1535  *
1536  * In order to make forward progress despite repeatedly restarting some
1537  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1538  * and restart from that address when we reach that vma again.  It might
1539  * have been split or merged, shrunk or extended, but never shifted: so
1540  * restart_addr remains valid so long as it remains in the vma's range.
1541  * unmap_mapping_range forces truncate_count to leap over page-aligned
1542  * values so we can save vma's restart_addr in its truncate_count field.
1543  */
1544 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1545
1546 static void reset_vma_truncate_counts(struct address_space *mapping)
1547 {
1548         struct vm_area_struct *vma;
1549         struct prio_tree_iter iter;
1550
1551         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1552                 vma->vm_truncate_count = 0;
1553         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1554                 vma->vm_truncate_count = 0;
1555 }
1556
1557 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1558                 unsigned long start_addr, unsigned long end_addr,
1559                 struct zap_details *details)
1560 {
1561         unsigned long restart_addr;
1562         int need_break;
1563
1564 again:
1565         restart_addr = vma->vm_truncate_count;
1566         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1567                 start_addr = restart_addr;
1568                 if (start_addr >= end_addr) {
1569                         /* Top of vma has been split off since last time */
1570                         vma->vm_truncate_count = details->truncate_count;
1571                         return 0;
1572                 }
1573         }
1574
1575         restart_addr = zap_page_range(vma, start_addr,
1576                                         end_addr - start_addr, details);
1577         need_break = need_resched() ||
1578                         need_lockbreak(details->i_mmap_lock);
1579
1580         if (restart_addr >= end_addr) {
1581                 /* We have now completed this vma: mark it so */
1582                 vma->vm_truncate_count = details->truncate_count;
1583                 if (!need_break)
1584                         return 0;
1585         } else {
1586                 /* Note restart_addr in vma's truncate_count field */
1587                 vma->vm_truncate_count = restart_addr;
1588                 if (!need_break)
1589                         goto again;
1590         }
1591
1592         spin_unlock(details->i_mmap_lock);
1593         cond_resched();
1594         spin_lock(details->i_mmap_lock);
1595         return -EINTR;
1596 }
1597
1598 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1599                                             struct zap_details *details)
1600 {
1601         struct vm_area_struct *vma;
1602         struct prio_tree_iter iter;
1603         pgoff_t vba, vea, zba, zea;
1604
1605 restart:
1606         vma_prio_tree_foreach(vma, &iter, root,
1607                         details->first_index, details->last_index) {
1608                 /* Skip quickly over those we have already dealt with */
1609                 if (vma->vm_truncate_count == details->truncate_count)
1610                         continue;
1611
1612                 vba = vma->vm_pgoff;
1613                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1614                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1615                 zba = details->first_index;
1616                 if (zba < vba)
1617                         zba = vba;
1618                 zea = details->last_index;
1619                 if (zea > vea)
1620                         zea = vea;
1621
1622                 if (unmap_mapping_range_vma(vma,
1623                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1624                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1625                                 details) < 0)
1626                         goto restart;
1627         }
1628 }
1629
1630 static inline void unmap_mapping_range_list(struct list_head *head,
1631                                             struct zap_details *details)
1632 {
1633         struct vm_area_struct *vma;
1634
1635         /*
1636          * In nonlinear VMAs there is no correspondence between virtual address
1637          * offset and file offset.  So we must perform an exhaustive search
1638          * across *all* the pages in each nonlinear VMA, not just the pages
1639          * whose virtual address lies outside the file truncation point.
1640          */
1641 restart:
1642         list_for_each_entry(vma, head, shared.vm_set.list) {
1643                 /* Skip quickly over those we have already dealt with */
1644                 if (vma->vm_truncate_count == details->truncate_count)
1645                         continue;
1646                 details->nonlinear_vma = vma;
1647                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1648                                         vma->vm_end, details) < 0)
1649                         goto restart;
1650         }
1651 }
1652
1653 /**
1654  * unmap_mapping_range - unmap the portion of all mmaps
1655  * in the specified address_space corresponding to the specified
1656  * page range in the underlying file.
1657  * @mapping: the address space containing mmaps to be unmapped.
1658  * @holebegin: byte in first page to unmap, relative to the start of
1659  * the underlying file.  This will be rounded down to a PAGE_SIZE
1660  * boundary.  Note that this is different from vmtruncate(), which
1661  * must keep the partial page.  In contrast, we must get rid of
1662  * partial pages.
1663  * @holelen: size of prospective hole in bytes.  This will be rounded
1664  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1665  * end of the file.
1666  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1667  * but 0 when invalidating pagecache, don't throw away private data.
1668  */
1669 void unmap_mapping_range(struct address_space *mapping,
1670                 loff_t const holebegin, loff_t const holelen, int even_cows)
1671 {
1672         struct zap_details details;
1673         pgoff_t hba = holebegin >> PAGE_SHIFT;
1674         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1675
1676         /* Check for overflow. */
1677         if (sizeof(holelen) > sizeof(hlen)) {
1678                 long long holeend =
1679                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1680                 if (holeend & ~(long long)ULONG_MAX)
1681                         hlen = ULONG_MAX - hba + 1;
1682         }
1683
1684         details.check_mapping = even_cows? NULL: mapping;
1685         details.nonlinear_vma = NULL;
1686         details.first_index = hba;
1687         details.last_index = hba + hlen - 1;
1688         if (details.last_index < details.first_index)
1689                 details.last_index = ULONG_MAX;
1690         details.i_mmap_lock = &mapping->i_mmap_lock;
1691
1692         spin_lock(&mapping->i_mmap_lock);
1693
1694         /* serialize i_size write against truncate_count write */
1695         smp_wmb();
1696         /* Protect against page faults, and endless unmapping loops */
1697         mapping->truncate_count++;
1698         /*
1699          * For archs where spin_lock has inclusive semantics like ia64
1700          * this smp_mb() will prevent to read pagetable contents
1701          * before the truncate_count increment is visible to
1702          * other cpus.
1703          */
1704         smp_mb();
1705         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1706                 if (mapping->truncate_count == 0)
1707                         reset_vma_truncate_counts(mapping);
1708                 mapping->truncate_count++;
1709         }
1710         details.truncate_count = mapping->truncate_count;
1711
1712         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1713                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1714         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1715                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1716         spin_unlock(&mapping->i_mmap_lock);
1717 }
1718 EXPORT_SYMBOL(unmap_mapping_range);
1719
1720 /*
1721  * Handle all mappings that got truncated by a "truncate()"
1722  * system call.
1723  *
1724  * NOTE! We have to be ready to update the memory sharing
1725  * between the file and the memory map for a potential last
1726  * incomplete page.  Ugly, but necessary.
1727  */
1728 int vmtruncate(struct inode * inode, loff_t offset)
1729 {
1730         struct address_space *mapping = inode->i_mapping;
1731         unsigned long limit;
1732
1733         if (inode->i_size < offset)
1734                 goto do_expand;
1735         /*
1736          * truncation of in-use swapfiles is disallowed - it would cause
1737          * subsequent swapout to scribble on the now-freed blocks.
1738          */
1739         if (IS_SWAPFILE(inode))
1740                 goto out_busy;
1741         i_size_write(inode, offset);
1742         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1743         truncate_inode_pages(mapping, offset);
1744         goto out_truncate;
1745
1746 do_expand:
1747         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1748         if (limit != RLIM_INFINITY && offset > limit)
1749                 goto out_sig;
1750         if (offset > inode->i_sb->s_maxbytes)
1751                 goto out_big;
1752         i_size_write(inode, offset);
1753
1754 out_truncate:
1755         if (inode->i_op && inode->i_op->truncate)
1756                 inode->i_op->truncate(inode);
1757         return 0;
1758 out_sig:
1759         send_sig(SIGXFSZ, current, 0);
1760 out_big:
1761         return -EFBIG;
1762 out_busy:
1763         return -ETXTBSY;
1764 }
1765
1766 EXPORT_SYMBOL(vmtruncate);
1767
1768 /* 
1769  * Primitive swap readahead code. We simply read an aligned block of
1770  * (1 << page_cluster) entries in the swap area. This method is chosen
1771  * because it doesn't cost us any seek time.  We also make sure to queue
1772  * the 'original' request together with the readahead ones...  
1773  *
1774  * This has been extended to use the NUMA policies from the mm triggering
1775  * the readahead.
1776  *
1777  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1778  */
1779 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1780 {
1781 #ifdef CONFIG_NUMA
1782         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1783 #endif
1784         int i, num;
1785         struct page *new_page;
1786         unsigned long offset;
1787
1788         /*
1789          * Get the number of handles we should do readahead io to.
1790          */
1791         num = valid_swaphandles(entry, &offset);
1792         for (i = 0; i < num; offset++, i++) {
1793                 /* Ok, do the async read-ahead now */
1794                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1795                                                            offset), vma, addr);
1796                 if (!new_page)
1797                         break;
1798                 page_cache_release(new_page);
1799 #ifdef CONFIG_NUMA
1800                 /*
1801                  * Find the next applicable VMA for the NUMA policy.
1802                  */
1803                 addr += PAGE_SIZE;
1804                 if (addr == 0)
1805                         vma = NULL;
1806                 if (vma) {
1807                         if (addr >= vma->vm_end) {
1808                                 vma = next_vma;
1809                                 next_vma = vma ? vma->vm_next : NULL;
1810                         }
1811                         if (vma && addr < vma->vm_start)
1812                                 vma = NULL;
1813                 } else {
1814                         if (next_vma && addr >= next_vma->vm_start) {
1815                                 vma = next_vma;
1816                                 next_vma = vma->vm_next;
1817                         }
1818                 }
1819 #endif
1820         }
1821         lru_add_drain();        /* Push any new pages onto the LRU now */
1822 }
1823
1824 /*
1825  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1826  * but allow concurrent faults), and pte mapped but not yet locked.
1827  * We return with mmap_sem still held, but pte unmapped and unlocked.
1828  */
1829 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1830                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1831                 int write_access, pte_t orig_pte)
1832 {
1833         spinlock_t *ptl;
1834         struct page *page;
1835         swp_entry_t entry;
1836         pte_t pte;
1837         int ret = VM_FAULT_MINOR;
1838
1839         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1840                 goto out;
1841
1842         entry = pte_to_swp_entry(orig_pte);
1843         page = lookup_swap_cache(entry);
1844         if (!page) {
1845                 swapin_readahead(entry, address, vma);
1846                 page = read_swap_cache_async(entry, vma, address);
1847                 if (!page) {
1848                         /*
1849                          * Back out if somebody else faulted in this pte
1850                          * while we released the pte lock.
1851                          */
1852                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1853                         if (likely(pte_same(*page_table, orig_pte)))
1854                                 ret = VM_FAULT_OOM;
1855                         goto unlock;
1856                 }
1857
1858                 /* Had to read the page from swap area: Major fault */
1859                 ret = VM_FAULT_MAJOR;
1860                 inc_page_state(pgmajfault);
1861                 grab_swap_token();
1862         }
1863
1864         mark_page_accessed(page);
1865         lock_page(page);
1866
1867         /*
1868          * Back out if somebody else already faulted in this pte.
1869          */
1870         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1871         if (unlikely(!pte_same(*page_table, orig_pte)))
1872                 goto out_nomap;
1873
1874         if (unlikely(!PageUptodate(page))) {
1875                 ret = VM_FAULT_SIGBUS;
1876                 goto out_nomap;
1877         }
1878
1879         /* The page isn't present yet, go ahead with the fault. */
1880
1881         inc_mm_counter(mm, anon_rss);
1882         pte = mk_pte(page, vma->vm_page_prot);
1883         if (write_access && can_share_swap_page(page)) {
1884                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1885                 write_access = 0;
1886         }
1887
1888         flush_icache_page(vma, page);
1889         set_pte_at(mm, address, page_table, pte);
1890         page_add_anon_rmap(page, vma, address);
1891
1892         swap_free(entry);
1893         if (vm_swap_full())
1894                 remove_exclusive_swap_page(page);
1895         unlock_page(page);
1896
1897         if (write_access) {
1898                 if (do_wp_page(mm, vma, address,
1899                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1900                         ret = VM_FAULT_OOM;
1901                 goto out;
1902         }
1903
1904         /* No need to invalidate - it was non-present before */
1905         update_mmu_cache(vma, address, pte);
1906         lazy_mmu_prot_update(pte);
1907 unlock:
1908         pte_unmap_unlock(page_table, ptl);
1909 out:
1910         return ret;
1911 out_nomap:
1912         pte_unmap_unlock(page_table, ptl);
1913         unlock_page(page);
1914         page_cache_release(page);
1915         return ret;
1916 }
1917
1918 /*
1919  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1920  * but allow concurrent faults), and pte mapped but not yet locked.
1921  * We return with mmap_sem still held, but pte unmapped and unlocked.
1922  */
1923 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1924                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1925                 int write_access)
1926 {
1927         struct page *page;
1928         spinlock_t *ptl;
1929         pte_t entry;
1930
1931         if (write_access) {
1932                 /* Allocate our own private page. */
1933                 pte_unmap(page_table);
1934
1935                 if (unlikely(anon_vma_prepare(vma)))
1936                         goto oom;
1937                 page = alloc_zeroed_user_highpage(vma, address);
1938                 if (!page)
1939                         goto oom;
1940
1941                 entry = mk_pte(page, vma->vm_page_prot);
1942                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1943
1944                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1945                 if (!pte_none(*page_table))
1946                         goto release;
1947                 inc_mm_counter(mm, anon_rss);
1948                 lru_cache_add_active(page);
1949                 SetPageReferenced(page);
1950                 page_add_anon_rmap(page, vma, address);
1951         } else {
1952                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1953                 page = ZERO_PAGE(address);
1954                 page_cache_get(page);
1955                 entry = mk_pte(page, vma->vm_page_prot);
1956
1957                 ptl = pte_lockptr(mm, pmd);
1958                 spin_lock(ptl);
1959                 if (!pte_none(*page_table))
1960                         goto release;
1961                 inc_mm_counter(mm, file_rss);
1962                 page_add_file_rmap(page);
1963         }
1964
1965         set_pte_at(mm, address, page_table, entry);
1966
1967         /* No need to invalidate - it was non-present before */
1968         update_mmu_cache(vma, address, entry);
1969         lazy_mmu_prot_update(entry);
1970 unlock:
1971         pte_unmap_unlock(page_table, ptl);
1972         return VM_FAULT_MINOR;
1973 release:
1974         page_cache_release(page);
1975         goto unlock;
1976 oom:
1977         return VM_FAULT_OOM;
1978 }
1979
1980 /*
1981  * do_no_page() tries to create a new page mapping. It aggressively
1982  * tries to share with existing pages, but makes a separate copy if
1983  * the "write_access" parameter is true in order to avoid the next
1984  * page fault.
1985  *
1986  * As this is called only for pages that do not currently exist, we
1987  * do not need to flush old virtual caches or the TLB.
1988  *
1989  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1990  * but allow concurrent faults), and pte mapped but not yet locked.
1991  * We return with mmap_sem still held, but pte unmapped and unlocked.
1992  */
1993 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1994                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1995                 int write_access)
1996 {
1997         spinlock_t *ptl;
1998         struct page *new_page;
1999         struct address_space *mapping = NULL;
2000         pte_t entry;
2001         unsigned int sequence = 0;
2002         int ret = VM_FAULT_MINOR;
2003         int anon = 0;
2004
2005         pte_unmap(page_table);
2006         if (vma->vm_file) {
2007                 mapping = vma->vm_file->f_mapping;
2008                 sequence = mapping->truncate_count;
2009                 smp_rmb(); /* serializes i_size against truncate_count */
2010         }
2011 retry:
2012         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2013         /*
2014          * No smp_rmb is needed here as long as there's a full
2015          * spin_lock/unlock sequence inside the ->nopage callback
2016          * (for the pagecache lookup) that acts as an implicit
2017          * smp_mb() and prevents the i_size read to happen
2018          * after the next truncate_count read.
2019          */
2020
2021         /* no page was available -- either SIGBUS or OOM */
2022         if (new_page == NOPAGE_SIGBUS)
2023                 return VM_FAULT_SIGBUS;
2024         if (new_page == NOPAGE_OOM)
2025                 return VM_FAULT_OOM;
2026
2027         /*
2028          * Should we do an early C-O-W break?
2029          */
2030         if (write_access && !(vma->vm_flags & VM_SHARED)) {
2031                 struct page *page;
2032
2033                 if (unlikely(anon_vma_prepare(vma)))
2034                         goto oom;
2035                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2036                 if (!page)
2037                         goto oom;
2038                 cow_user_page(page, new_page, address);
2039                 page_cache_release(new_page);
2040                 new_page = page;
2041                 anon = 1;
2042         }
2043
2044         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2045         /*
2046          * For a file-backed vma, someone could have truncated or otherwise
2047          * invalidated this page.  If unmap_mapping_range got called,
2048          * retry getting the page.
2049          */
2050         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2051                 pte_unmap_unlock(page_table, ptl);
2052                 page_cache_release(new_page);
2053                 cond_resched();
2054                 sequence = mapping->truncate_count;
2055                 smp_rmb();
2056                 goto retry;
2057         }
2058
2059         /*
2060          * This silly early PAGE_DIRTY setting removes a race
2061          * due to the bad i386 page protection. But it's valid
2062          * for other architectures too.
2063          *
2064          * Note that if write_access is true, we either now have
2065          * an exclusive copy of the page, or this is a shared mapping,
2066          * so we can make it writable and dirty to avoid having to
2067          * handle that later.
2068          */
2069         /* Only go through if we didn't race with anybody else... */
2070         if (pte_none(*page_table)) {
2071                 flush_icache_page(vma, new_page);
2072                 entry = mk_pte(new_page, vma->vm_page_prot);
2073                 if (write_access)
2074                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2075                 set_pte_at(mm, address, page_table, entry);
2076                 if (anon) {
2077                         inc_mm_counter(mm, anon_rss);
2078                         lru_cache_add_active(new_page);
2079                         page_add_anon_rmap(new_page, vma, address);
2080                 } else {
2081                         inc_mm_counter(mm, file_rss);
2082                         page_add_file_rmap(new_page);
2083                 }
2084         } else {
2085                 /* One of our sibling threads was faster, back out. */
2086                 page_cache_release(new_page);
2087                 goto unlock;
2088         }
2089
2090         /* no need to invalidate: a not-present page shouldn't be cached */
2091         update_mmu_cache(vma, address, entry);
2092         lazy_mmu_prot_update(entry);
2093 unlock:
2094         pte_unmap_unlock(page_table, ptl);
2095         return ret;
2096 oom:
2097         page_cache_release(new_page);
2098         return VM_FAULT_OOM;
2099 }
2100
2101 /*
2102  * Fault of a previously existing named mapping. Repopulate the pte
2103  * from the encoded file_pte if possible. This enables swappable
2104  * nonlinear vmas.
2105  *
2106  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2107  * but allow concurrent faults), and pte mapped but not yet locked.
2108  * We return with mmap_sem still held, but pte unmapped and unlocked.
2109  */
2110 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2111                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2112                 int write_access, pte_t orig_pte)
2113 {
2114         pgoff_t pgoff;
2115         int err;
2116
2117         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2118                 return VM_FAULT_MINOR;
2119
2120         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2121                 /*
2122                  * Page table corrupted: show pte and kill process.
2123                  */
2124                 print_bad_pte(vma, orig_pte, address);
2125                 return VM_FAULT_OOM;
2126         }
2127         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2128
2129         pgoff = pte_to_pgoff(orig_pte);
2130         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2131                                         vma->vm_page_prot, pgoff, 0);
2132         if (err == -ENOMEM)
2133                 return VM_FAULT_OOM;
2134         if (err)
2135                 return VM_FAULT_SIGBUS;
2136         return VM_FAULT_MAJOR;
2137 }
2138
2139 /*
2140  * These routines also need to handle stuff like marking pages dirty
2141  * and/or accessed for architectures that don't do it in hardware (most
2142  * RISC architectures).  The early dirtying is also good on the i386.
2143  *
2144  * There is also a hook called "update_mmu_cache()" that architectures
2145  * with external mmu caches can use to update those (ie the Sparc or
2146  * PowerPC hashed page tables that act as extended TLBs).
2147  *
2148  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2149  * but allow concurrent faults), and pte mapped but not yet locked.
2150  * We return with mmap_sem still held, but pte unmapped and unlocked.
2151  */
2152 static inline int handle_pte_fault(struct mm_struct *mm,
2153                 struct vm_area_struct *vma, unsigned long address,
2154                 pte_t *pte, pmd_t *pmd, int write_access)
2155 {
2156         pte_t entry;
2157         pte_t old_entry;
2158         spinlock_t *ptl;
2159
2160         old_entry = entry = *pte;
2161         if (!pte_present(entry)) {
2162                 if (pte_none(entry)) {
2163                         if (!vma->vm_ops || !vma->vm_ops->nopage)
2164                                 return do_anonymous_page(mm, vma, address,
2165                                         pte, pmd, write_access);
2166                         return do_no_page(mm, vma, address,
2167                                         pte, pmd, write_access);
2168                 }
2169                 if (pte_file(entry))
2170                         return do_file_page(mm, vma, address,
2171                                         pte, pmd, write_access, entry);
2172                 return do_swap_page(mm, vma, address,
2173                                         pte, pmd, write_access, entry);
2174         }
2175
2176         ptl = pte_lockptr(mm, pmd);
2177         spin_lock(ptl);
2178         if (unlikely(!pte_same(*pte, entry)))
2179                 goto unlock;
2180         if (write_access) {
2181                 if (!pte_write(entry))
2182                         return do_wp_page(mm, vma, address,
2183                                         pte, pmd, ptl, entry);
2184                 entry = pte_mkdirty(entry);
2185         }
2186         entry = pte_mkyoung(entry);
2187         if (!pte_same(old_entry, entry)) {
2188                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2189                 update_mmu_cache(vma, address, entry);
2190                 lazy_mmu_prot_update(entry);
2191         } else {
2192                 /*
2193                  * This is needed only for protection faults but the arch code
2194                  * is not yet telling us if this is a protection fault or not.
2195                  * This still avoids useless tlb flushes for .text page faults
2196                  * with threads.
2197                  */
2198                 if (write_access)
2199                         flush_tlb_page(vma, address);
2200         }
2201 unlock:
2202         pte_unmap_unlock(pte, ptl);
2203         return VM_FAULT_MINOR;
2204 }
2205
2206 /*
2207  * By the time we get here, we already hold the mm semaphore
2208  */
2209 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2210                 unsigned long address, int write_access)
2211 {
2212         pgd_t *pgd;
2213         pud_t *pud;
2214         pmd_t *pmd;
2215         pte_t *pte;
2216
2217         __set_current_state(TASK_RUNNING);
2218
2219         inc_page_state(pgfault);
2220
2221         if (unlikely(is_vm_hugetlb_page(vma)))
2222                 return hugetlb_fault(mm, vma, address, write_access);
2223
2224         pgd = pgd_offset(mm, address);
2225         pud = pud_alloc(mm, pgd, address);
2226         if (!pud)
2227                 return VM_FAULT_OOM;
2228         pmd = pmd_alloc(mm, pud, address);
2229         if (!pmd)
2230                 return VM_FAULT_OOM;
2231         pte = pte_alloc_map(mm, pmd, address);
2232         if (!pte)
2233                 return VM_FAULT_OOM;
2234
2235         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2236 }
2237
2238 #ifndef __PAGETABLE_PUD_FOLDED
2239 /*
2240  * Allocate page upper directory.
2241  * We've already handled the fast-path in-line.
2242  */
2243 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2244 {
2245         pud_t *new = pud_alloc_one(mm, address);
2246         if (!new)
2247                 return -ENOMEM;
2248
2249         spin_lock(&mm->page_table_lock);
2250         if (pgd_present(*pgd))          /* Another has populated it */
2251                 pud_free(new);
2252         else
2253                 pgd_populate(mm, pgd, new);
2254         spin_unlock(&mm->page_table_lock);
2255         return 0;
2256 }
2257 #else
2258 /* Workaround for gcc 2.96 */
2259 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2260 {
2261         return 0;
2262 }
2263 #endif /* __PAGETABLE_PUD_FOLDED */
2264
2265 #ifndef __PAGETABLE_PMD_FOLDED
2266 /*
2267  * Allocate page middle directory.
2268  * We've already handled the fast-path in-line.
2269  */
2270 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2271 {
2272         pmd_t *new = pmd_alloc_one(mm, address);
2273         if (!new)
2274                 return -ENOMEM;
2275
2276         spin_lock(&mm->page_table_lock);
2277 #ifndef __ARCH_HAS_4LEVEL_HACK
2278         if (pud_present(*pud))          /* Another has populated it */
2279                 pmd_free(new);
2280         else
2281                 pud_populate(mm, pud, new);
2282 #else
2283         if (pgd_present(*pud))          /* Another has populated it */
2284                 pmd_free(new);
2285         else
2286                 pgd_populate(mm, pud, new);
2287 #endif /* __ARCH_HAS_4LEVEL_HACK */
2288         spin_unlock(&mm->page_table_lock);
2289         return 0;
2290 }
2291 #else
2292 /* Workaround for gcc 2.96 */
2293 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2294 {
2295         return 0;
2296 }
2297 #endif /* __PAGETABLE_PMD_FOLDED */
2298
2299 int make_pages_present(unsigned long addr, unsigned long end)
2300 {
2301         int ret, len, write;
2302         struct vm_area_struct * vma;
2303
2304         vma = find_vma(current->mm, addr);
2305         if (!vma)
2306                 return -1;
2307         write = (vma->vm_flags & VM_WRITE) != 0;
2308         if (addr >= end)
2309                 BUG();
2310         if (end > vma->vm_end)
2311                 BUG();
2312         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2313         ret = get_user_pages(current, current->mm, addr,
2314                         len, write, 0, NULL, NULL);
2315         if (ret < 0)
2316                 return ret;
2317         return ret == len ? 0 : -1;
2318 }
2319
2320 /* 
2321  * Map a vmalloc()-space virtual address to the physical page.
2322  */
2323 struct page * vmalloc_to_page(void * vmalloc_addr)
2324 {
2325         unsigned long addr = (unsigned long) vmalloc_addr;
2326         struct page *page = NULL;
2327         pgd_t *pgd = pgd_offset_k(addr);
2328         pud_t *pud;
2329         pmd_t *pmd;
2330         pte_t *ptep, pte;
2331   
2332         if (!pgd_none(*pgd)) {
2333                 pud = pud_offset(pgd, addr);
2334                 if (!pud_none(*pud)) {
2335                         pmd = pmd_offset(pud, addr);
2336                         if (!pmd_none(*pmd)) {
2337                                 ptep = pte_offset_map(pmd, addr);
2338                                 pte = *ptep;
2339                                 if (pte_present(pte))
2340                                         page = pte_page(pte);
2341                                 pte_unmap(ptep);
2342                         }
2343                 }
2344         }
2345         return page;
2346 }
2347
2348 EXPORT_SYMBOL(vmalloc_to_page);
2349
2350 /*
2351  * Map a vmalloc()-space virtual address to the physical page frame number.
2352  */
2353 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2354 {
2355         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2356 }
2357
2358 EXPORT_SYMBOL(vmalloc_to_pfn);
2359
2360 #if !defined(__HAVE_ARCH_GATE_AREA)
2361
2362 #if defined(AT_SYSINFO_EHDR)
2363 static struct vm_area_struct gate_vma;
2364
2365 static int __init gate_vma_init(void)
2366 {
2367         gate_vma.vm_mm = NULL;
2368         gate_vma.vm_start = FIXADDR_USER_START;
2369         gate_vma.vm_end = FIXADDR_USER_END;
2370         gate_vma.vm_page_prot = PAGE_READONLY;
2371         gate_vma.vm_flags = 0;
2372         return 0;
2373 }
2374 __initcall(gate_vma_init);
2375 #endif
2376
2377 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2378 {
2379 #ifdef AT_SYSINFO_EHDR
2380         return &gate_vma;
2381 #else
2382         return NULL;
2383 #endif
2384 }
2385
2386 int in_gate_area_no_task(unsigned long addr)
2387 {
2388 #ifdef AT_SYSINFO_EHDR
2389         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2390                 return 1;
2391 #endif
2392         return 0;
2393 }
2394
2395 #endif  /* __HAVE_ARCH_GATE_AREA */