[PATCH] mm: rmap optimisation
[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 static inline int is_cow_mapping(unsigned int flags)
353 {
354         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
355 }
356
357 /*
358  * This function gets the "struct page" associated with a pte.
359  *
360  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
361  * will have each page table entry just pointing to a raw page frame
362  * number, and as far as the VM layer is concerned, those do not have
363  * pages associated with them - even if the PFN might point to memory
364  * that otherwise is perfectly fine and has a "struct page".
365  *
366  * The way we recognize those mappings is through the rules set up
367  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
368  * and the vm_pgoff will point to the first PFN mapped: thus every
369  * page that is a raw mapping will always honor the rule
370  *
371  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
372  *
373  * and if that isn't true, the page has been COW'ed (in which case it
374  * _does_ have a "struct page" associated with it even if it is in a
375  * VM_PFNMAP range).
376  */
377 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
378 {
379         unsigned long pfn = pte_pfn(pte);
380
381         if (vma->vm_flags & VM_PFNMAP) {
382                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
383                 if (pfn == vma->vm_pgoff + off)
384                         return NULL;
385                 if (!is_cow_mapping(vma->vm_flags))
386                         return NULL;
387         }
388
389         /*
390          * Add some anal sanity checks for now. Eventually,
391          * we should just do "return pfn_to_page(pfn)", but
392          * in the meantime we check that we get a valid pfn,
393          * and that the resulting page looks ok.
394          *
395          * Remove this test eventually!
396          */
397         if (unlikely(!pfn_valid(pfn))) {
398                 print_bad_pte(vma, pte, addr);
399                 return NULL;
400         }
401
402         /*
403          * NOTE! We still have PageReserved() pages in the page 
404          * tables. 
405          *
406          * The PAGE_ZERO() pages and various VDSO mappings can
407          * cause them to exist.
408          */
409         return pfn_to_page(pfn);
410 }
411
412 /*
413  * copy one vm_area from one task to the other. Assumes the page tables
414  * already present in the new task to be cleared in the whole range
415  * covered by this vma.
416  */
417
418 static inline void
419 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
420                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
421                 unsigned long addr, int *rss)
422 {
423         unsigned long vm_flags = vma->vm_flags;
424         pte_t pte = *src_pte;
425         struct page *page;
426
427         /* pte contains position in swap or file, so copy. */
428         if (unlikely(!pte_present(pte))) {
429                 if (!pte_file(pte)) {
430                         swap_duplicate(pte_to_swp_entry(pte));
431                         /* make sure dst_mm is on swapoff's mmlist. */
432                         if (unlikely(list_empty(&dst_mm->mmlist))) {
433                                 spin_lock(&mmlist_lock);
434                                 if (list_empty(&dst_mm->mmlist))
435                                         list_add(&dst_mm->mmlist,
436                                                  &src_mm->mmlist);
437                                 spin_unlock(&mmlist_lock);
438                         }
439                 }
440                 goto out_set_pte;
441         }
442
443         /*
444          * If it's a COW mapping, write protect it both
445          * in the parent and the child
446          */
447         if (is_cow_mapping(vm_flags)) {
448                 ptep_set_wrprotect(src_mm, addr, src_pte);
449                 pte = *src_pte;
450         }
451
452         /*
453          * If it's a shared mapping, mark it clean in
454          * the child
455          */
456         if (vm_flags & VM_SHARED)
457                 pte = pte_mkclean(pte);
458         pte = pte_mkold(pte);
459
460         page = vm_normal_page(vma, addr, pte);
461         if (page) {
462                 get_page(page);
463                 page_dup_rmap(page);
464                 rss[!!PageAnon(page)]++;
465         }
466
467 out_set_pte:
468         set_pte_at(dst_mm, addr, dst_pte, pte);
469 }
470
471 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
472                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
473                 unsigned long addr, unsigned long end)
474 {
475         pte_t *src_pte, *dst_pte;
476         spinlock_t *src_ptl, *dst_ptl;
477         int progress = 0;
478         int rss[2];
479
480 again:
481         rss[1] = rss[0] = 0;
482         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
483         if (!dst_pte)
484                 return -ENOMEM;
485         src_pte = pte_offset_map_nested(src_pmd, addr);
486         src_ptl = pte_lockptr(src_mm, src_pmd);
487         spin_lock(src_ptl);
488
489         do {
490                 /*
491                  * We are holding two locks at this point - either of them
492                  * could generate latencies in another task on another CPU.
493                  */
494                 if (progress >= 32) {
495                         progress = 0;
496                         if (need_resched() ||
497                             need_lockbreak(src_ptl) ||
498                             need_lockbreak(dst_ptl))
499                                 break;
500                 }
501                 if (pte_none(*src_pte)) {
502                         progress++;
503                         continue;
504                 }
505                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
506                 progress += 8;
507         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
508
509         spin_unlock(src_ptl);
510         pte_unmap_nested(src_pte - 1);
511         add_mm_rss(dst_mm, rss[0], rss[1]);
512         pte_unmap_unlock(dst_pte - 1, dst_ptl);
513         cond_resched();
514         if (addr != end)
515                 goto again;
516         return 0;
517 }
518
519 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
520                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
521                 unsigned long addr, unsigned long end)
522 {
523         pmd_t *src_pmd, *dst_pmd;
524         unsigned long next;
525
526         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
527         if (!dst_pmd)
528                 return -ENOMEM;
529         src_pmd = pmd_offset(src_pud, addr);
530         do {
531                 next = pmd_addr_end(addr, end);
532                 if (pmd_none_or_clear_bad(src_pmd))
533                         continue;
534                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
535                                                 vma, addr, next))
536                         return -ENOMEM;
537         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
538         return 0;
539 }
540
541 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
542                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
543                 unsigned long addr, unsigned long end)
544 {
545         pud_t *src_pud, *dst_pud;
546         unsigned long next;
547
548         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
549         if (!dst_pud)
550                 return -ENOMEM;
551         src_pud = pud_offset(src_pgd, addr);
552         do {
553                 next = pud_addr_end(addr, end);
554                 if (pud_none_or_clear_bad(src_pud))
555                         continue;
556                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
557                                                 vma, addr, next))
558                         return -ENOMEM;
559         } while (dst_pud++, src_pud++, addr = next, addr != end);
560         return 0;
561 }
562
563 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
564                 struct vm_area_struct *vma)
565 {
566         pgd_t *src_pgd, *dst_pgd;
567         unsigned long next;
568         unsigned long addr = vma->vm_start;
569         unsigned long end = vma->vm_end;
570
571         /*
572          * Don't copy ptes where a page fault will fill them correctly.
573          * Fork becomes much lighter when there are big shared or private
574          * readonly mappings. The tradeoff is that copy_page_range is more
575          * efficient than faulting.
576          */
577         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
578                 if (!vma->anon_vma)
579                         return 0;
580         }
581
582         if (is_vm_hugetlb_page(vma))
583                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
584
585         dst_pgd = pgd_offset(dst_mm, addr);
586         src_pgd = pgd_offset(src_mm, addr);
587         do {
588                 next = pgd_addr_end(addr, end);
589                 if (pgd_none_or_clear_bad(src_pgd))
590                         continue;
591                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
592                                                 vma, addr, next))
593                         return -ENOMEM;
594         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
595         return 0;
596 }
597
598 static unsigned long zap_pte_range(struct mmu_gather *tlb,
599                                 struct vm_area_struct *vma, pmd_t *pmd,
600                                 unsigned long addr, unsigned long end,
601                                 long *zap_work, struct zap_details *details)
602 {
603         struct mm_struct *mm = tlb->mm;
604         pte_t *pte;
605         spinlock_t *ptl;
606         int file_rss = 0;
607         int anon_rss = 0;
608
609         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
610         do {
611                 pte_t ptent = *pte;
612                 if (pte_none(ptent)) {
613                         (*zap_work)--;
614                         continue;
615                 }
616                 if (pte_present(ptent)) {
617                         struct page *page;
618
619                         (*zap_work) -= PAGE_SIZE;
620
621                         page = vm_normal_page(vma, addr, ptent);
622                         if (unlikely(details) && page) {
623                                 /*
624                                  * unmap_shared_mapping_pages() wants to
625                                  * invalidate cache without truncating:
626                                  * unmap shared but keep private pages.
627                                  */
628                                 if (details->check_mapping &&
629                                     details->check_mapping != page->mapping)
630                                         continue;
631                                 /*
632                                  * Each page->index must be checked when
633                                  * invalidating or truncating nonlinear.
634                                  */
635                                 if (details->nonlinear_vma &&
636                                     (page->index < details->first_index ||
637                                      page->index > details->last_index))
638                                         continue;
639                         }
640                         ptent = ptep_get_and_clear_full(mm, addr, pte,
641                                                         tlb->fullmm);
642                         tlb_remove_tlb_entry(tlb, pte, addr);
643                         if (unlikely(!page))
644                                 continue;
645                         if (unlikely(details) && details->nonlinear_vma
646                             && linear_page_index(details->nonlinear_vma,
647                                                 addr) != page->index)
648                                 set_pte_at(mm, addr, pte,
649                                            pgoff_to_pte(page->index));
650                         if (PageAnon(page))
651                                 anon_rss--;
652                         else {
653                                 if (pte_dirty(ptent))
654                                         set_page_dirty(page);
655                                 if (pte_young(ptent))
656                                         mark_page_accessed(page);
657                                 file_rss--;
658                         }
659                         page_remove_rmap(page);
660                         tlb_remove_page(tlb, page);
661                         continue;
662                 }
663                 /*
664                  * If details->check_mapping, we leave swap entries;
665                  * if details->nonlinear_vma, we leave file entries.
666                  */
667                 if (unlikely(details))
668                         continue;
669                 if (!pte_file(ptent))
670                         free_swap_and_cache(pte_to_swp_entry(ptent));
671                 pte_clear_full(mm, addr, pte, tlb->fullmm);
672         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
673
674         add_mm_rss(mm, file_rss, anon_rss);
675         pte_unmap_unlock(pte - 1, ptl);
676
677         return addr;
678 }
679
680 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
681                                 struct vm_area_struct *vma, pud_t *pud,
682                                 unsigned long addr, unsigned long end,
683                                 long *zap_work, struct zap_details *details)
684 {
685         pmd_t *pmd;
686         unsigned long next;
687
688         pmd = pmd_offset(pud, addr);
689         do {
690                 next = pmd_addr_end(addr, end);
691                 if (pmd_none_or_clear_bad(pmd)) {
692                         (*zap_work)--;
693                         continue;
694                 }
695                 next = zap_pte_range(tlb, vma, pmd, addr, next,
696                                                 zap_work, details);
697         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
698
699         return addr;
700 }
701
702 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
703                                 struct vm_area_struct *vma, pgd_t *pgd,
704                                 unsigned long addr, unsigned long end,
705                                 long *zap_work, struct zap_details *details)
706 {
707         pud_t *pud;
708         unsigned long next;
709
710         pud = pud_offset(pgd, addr);
711         do {
712                 next = pud_addr_end(addr, end);
713                 if (pud_none_or_clear_bad(pud)) {
714                         (*zap_work)--;
715                         continue;
716                 }
717                 next = zap_pmd_range(tlb, vma, pud, addr, next,
718                                                 zap_work, details);
719         } while (pud++, addr = next, (addr != end && *zap_work > 0));
720
721         return addr;
722 }
723
724 static unsigned long unmap_page_range(struct mmu_gather *tlb,
725                                 struct vm_area_struct *vma,
726                                 unsigned long addr, unsigned long end,
727                                 long *zap_work, struct zap_details *details)
728 {
729         pgd_t *pgd;
730         unsigned long next;
731
732         if (details && !details->check_mapping && !details->nonlinear_vma)
733                 details = NULL;
734
735         BUG_ON(addr >= end);
736         tlb_start_vma(tlb, vma);
737         pgd = pgd_offset(vma->vm_mm, addr);
738         do {
739                 next = pgd_addr_end(addr, end);
740                 if (pgd_none_or_clear_bad(pgd)) {
741                         (*zap_work)--;
742                         continue;
743                 }
744                 next = zap_pud_range(tlb, vma, pgd, addr, next,
745                                                 zap_work, details);
746         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
747         tlb_end_vma(tlb, vma);
748
749         return addr;
750 }
751
752 #ifdef CONFIG_PREEMPT
753 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
754 #else
755 /* No preempt: go for improved straight-line efficiency */
756 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
757 #endif
758
759 /**
760  * unmap_vmas - unmap a range of memory covered by a list of vma's
761  * @tlbp: address of the caller's struct mmu_gather
762  * @vma: the starting vma
763  * @start_addr: virtual address at which to start unmapping
764  * @end_addr: virtual address at which to end unmapping
765  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
766  * @details: details of nonlinear truncation or shared cache invalidation
767  *
768  * Returns the end address of the unmapping (restart addr if interrupted).
769  *
770  * Unmap all pages in the vma list.
771  *
772  * We aim to not hold locks for too long (for scheduling latency reasons).
773  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
774  * return the ending mmu_gather to the caller.
775  *
776  * Only addresses between `start' and `end' will be unmapped.
777  *
778  * The VMA list must be sorted in ascending virtual address order.
779  *
780  * unmap_vmas() assumes that the caller will flush the whole unmapped address
781  * range after unmap_vmas() returns.  So the only responsibility here is to
782  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
783  * drops the lock and schedules.
784  */
785 unsigned long unmap_vmas(struct mmu_gather **tlbp,
786                 struct vm_area_struct *vma, unsigned long start_addr,
787                 unsigned long end_addr, unsigned long *nr_accounted,
788                 struct zap_details *details)
789 {
790         long zap_work = ZAP_BLOCK_SIZE;
791         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
792         int tlb_start_valid = 0;
793         unsigned long start = start_addr;
794         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
795         int fullmm = (*tlbp)->fullmm;
796
797         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
798                 unsigned long end;
799
800                 start = max(vma->vm_start, start_addr);
801                 if (start >= vma->vm_end)
802                         continue;
803                 end = min(vma->vm_end, end_addr);
804                 if (end <= vma->vm_start)
805                         continue;
806
807                 if (vma->vm_flags & VM_ACCOUNT)
808                         *nr_accounted += (end - start) >> PAGE_SHIFT;
809
810                 while (start != end) {
811                         if (!tlb_start_valid) {
812                                 tlb_start = start;
813                                 tlb_start_valid = 1;
814                         }
815
816                         if (unlikely(is_vm_hugetlb_page(vma))) {
817                                 unmap_hugepage_range(vma, start, end);
818                                 zap_work -= (end - start) /
819                                                 (HPAGE_SIZE / PAGE_SIZE);
820                                 start = end;
821                         } else
822                                 start = unmap_page_range(*tlbp, vma,
823                                                 start, end, &zap_work, details);
824
825                         if (zap_work > 0) {
826                                 BUG_ON(start != end);
827                                 break;
828                         }
829
830                         tlb_finish_mmu(*tlbp, tlb_start, start);
831
832                         if (need_resched() ||
833                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
834                                 if (i_mmap_lock) {
835                                         *tlbp = NULL;
836                                         goto out;
837                                 }
838                                 cond_resched();
839                         }
840
841                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
842                         tlb_start_valid = 0;
843                         zap_work = ZAP_BLOCK_SIZE;
844                 }
845         }
846 out:
847         return start;   /* which is now the end (or restart) address */
848 }
849
850 /**
851  * zap_page_range - remove user pages in a given range
852  * @vma: vm_area_struct holding the applicable pages
853  * @address: starting address of pages to zap
854  * @size: number of bytes to zap
855  * @details: details of nonlinear truncation or shared cache invalidation
856  */
857 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
858                 unsigned long size, struct zap_details *details)
859 {
860         struct mm_struct *mm = vma->vm_mm;
861         struct mmu_gather *tlb;
862         unsigned long end = address + size;
863         unsigned long nr_accounted = 0;
864
865         lru_add_drain();
866         tlb = tlb_gather_mmu(mm, 0);
867         update_hiwater_rss(mm);
868         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
869         if (tlb)
870                 tlb_finish_mmu(tlb, address, end);
871         return end;
872 }
873
874 /*
875  * Do a quick page-table lookup for a single page.
876  */
877 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
878                         unsigned int flags)
879 {
880         pgd_t *pgd;
881         pud_t *pud;
882         pmd_t *pmd;
883         pte_t *ptep, pte;
884         spinlock_t *ptl;
885         struct page *page;
886         struct mm_struct *mm = vma->vm_mm;
887
888         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
889         if (!IS_ERR(page)) {
890                 BUG_ON(flags & FOLL_GET);
891                 goto out;
892         }
893
894         page = NULL;
895         pgd = pgd_offset(mm, address);
896         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
897                 goto no_page_table;
898
899         pud = pud_offset(pgd, address);
900         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
901                 goto no_page_table;
902         
903         pmd = pmd_offset(pud, address);
904         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
905                 goto no_page_table;
906
907         if (pmd_huge(*pmd)) {
908                 BUG_ON(flags & FOLL_GET);
909                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
910                 goto out;
911         }
912
913         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
914         if (!ptep)
915                 goto out;
916
917         pte = *ptep;
918         if (!pte_present(pte))
919                 goto unlock;
920         if ((flags & FOLL_WRITE) && !pte_write(pte))
921                 goto unlock;
922         page = vm_normal_page(vma, address, pte);
923         if (unlikely(!page))
924                 goto unlock;
925
926         if (flags & FOLL_GET)
927                 get_page(page);
928         if (flags & FOLL_TOUCH) {
929                 if ((flags & FOLL_WRITE) &&
930                     !pte_dirty(pte) && !PageDirty(page))
931                         set_page_dirty(page);
932                 mark_page_accessed(page);
933         }
934 unlock:
935         pte_unmap_unlock(ptep, ptl);
936 out:
937         return page;
938
939 no_page_table:
940         /*
941          * When core dumping an enormous anonymous area that nobody
942          * has touched so far, we don't want to allocate page tables.
943          */
944         if (flags & FOLL_ANON) {
945                 page = ZERO_PAGE(address);
946                 if (flags & FOLL_GET)
947                         get_page(page);
948                 BUG_ON(flags & FOLL_WRITE);
949         }
950         return page;
951 }
952
953 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
954                 unsigned long start, int len, int write, int force,
955                 struct page **pages, struct vm_area_struct **vmas)
956 {
957         int i;
958         unsigned int vm_flags;
959
960         /* 
961          * Require read or write permissions.
962          * If 'force' is set, we only require the "MAY" flags.
963          */
964         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
965         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
966         i = 0;
967
968         do {
969                 struct vm_area_struct *vma;
970                 unsigned int foll_flags;
971
972                 vma = find_extend_vma(mm, start);
973                 if (!vma && in_gate_area(tsk, start)) {
974                         unsigned long pg = start & PAGE_MASK;
975                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
976                         pgd_t *pgd;
977                         pud_t *pud;
978                         pmd_t *pmd;
979                         pte_t *pte;
980                         if (write) /* user gate pages are read-only */
981                                 return i ? : -EFAULT;
982                         if (pg > TASK_SIZE)
983                                 pgd = pgd_offset_k(pg);
984                         else
985                                 pgd = pgd_offset_gate(mm, pg);
986                         BUG_ON(pgd_none(*pgd));
987                         pud = pud_offset(pgd, pg);
988                         BUG_ON(pud_none(*pud));
989                         pmd = pmd_offset(pud, pg);
990                         if (pmd_none(*pmd))
991                                 return i ? : -EFAULT;
992                         pte = pte_offset_map(pmd, pg);
993                         if (pte_none(*pte)) {
994                                 pte_unmap(pte);
995                                 return i ? : -EFAULT;
996                         }
997                         if (pages) {
998                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
999                                 pages[i] = page;
1000                                 if (page)
1001                                         get_page(page);
1002                         }
1003                         pte_unmap(pte);
1004                         if (vmas)
1005                                 vmas[i] = gate_vma;
1006                         i++;
1007                         start += PAGE_SIZE;
1008                         len--;
1009                         continue;
1010                 }
1011
1012                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1013                                 || !(vm_flags & vma->vm_flags))
1014                         return i ? : -EFAULT;
1015
1016                 if (is_vm_hugetlb_page(vma)) {
1017                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1018                                                 &start, &len, i);
1019                         continue;
1020                 }
1021
1022                 foll_flags = FOLL_TOUCH;
1023                 if (pages)
1024                         foll_flags |= FOLL_GET;
1025                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1026                     (!vma->vm_ops || !vma->vm_ops->nopage))
1027                         foll_flags |= FOLL_ANON;
1028
1029                 do {
1030                         struct page *page;
1031
1032                         if (write)
1033                                 foll_flags |= FOLL_WRITE;
1034
1035                         cond_resched();
1036                         while (!(page = follow_page(vma, start, foll_flags))) {
1037                                 int ret;
1038                                 ret = __handle_mm_fault(mm, vma, start,
1039                                                 foll_flags & FOLL_WRITE);
1040                                 /*
1041                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1042                                  * broken COW when necessary, even if maybe_mkwrite
1043                                  * decided not to set pte_write. We can thus safely do
1044                                  * subsequent page lookups as if they were reads.
1045                                  */
1046                                 if (ret & VM_FAULT_WRITE)
1047                                         foll_flags &= ~FOLL_WRITE;
1048                                 
1049                                 switch (ret & ~VM_FAULT_WRITE) {
1050                                 case VM_FAULT_MINOR:
1051                                         tsk->min_flt++;
1052                                         break;
1053                                 case VM_FAULT_MAJOR:
1054                                         tsk->maj_flt++;
1055                                         break;
1056                                 case VM_FAULT_SIGBUS:
1057                                         return i ? i : -EFAULT;
1058                                 case VM_FAULT_OOM:
1059                                         return i ? i : -ENOMEM;
1060                                 default:
1061                                         BUG();
1062                                 }
1063                         }
1064                         if (pages) {
1065                                 pages[i] = page;
1066                                 flush_dcache_page(page);
1067                         }
1068                         if (vmas)
1069                                 vmas[i] = vma;
1070                         i++;
1071                         start += PAGE_SIZE;
1072                         len--;
1073                 } while (len && start < vma->vm_end);
1074         } while (len);
1075         return i;
1076 }
1077 EXPORT_SYMBOL(get_user_pages);
1078
1079 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1080                         unsigned long addr, unsigned long end, pgprot_t prot)
1081 {
1082         pte_t *pte;
1083         spinlock_t *ptl;
1084
1085         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1086         if (!pte)
1087                 return -ENOMEM;
1088         do {
1089                 struct page *page = ZERO_PAGE(addr);
1090                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1091                 page_cache_get(page);
1092                 page_add_file_rmap(page);
1093                 inc_mm_counter(mm, file_rss);
1094                 BUG_ON(!pte_none(*pte));
1095                 set_pte_at(mm, addr, pte, zero_pte);
1096         } while (pte++, addr += PAGE_SIZE, addr != end);
1097         pte_unmap_unlock(pte - 1, ptl);
1098         return 0;
1099 }
1100
1101 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1102                         unsigned long addr, unsigned long end, pgprot_t prot)
1103 {
1104         pmd_t *pmd;
1105         unsigned long next;
1106
1107         pmd = pmd_alloc(mm, pud, addr);
1108         if (!pmd)
1109                 return -ENOMEM;
1110         do {
1111                 next = pmd_addr_end(addr, end);
1112                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1113                         return -ENOMEM;
1114         } while (pmd++, addr = next, addr != end);
1115         return 0;
1116 }
1117
1118 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1119                         unsigned long addr, unsigned long end, pgprot_t prot)
1120 {
1121         pud_t *pud;
1122         unsigned long next;
1123
1124         pud = pud_alloc(mm, pgd, addr);
1125         if (!pud)
1126                 return -ENOMEM;
1127         do {
1128                 next = pud_addr_end(addr, end);
1129                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1130                         return -ENOMEM;
1131         } while (pud++, addr = next, addr != end);
1132         return 0;
1133 }
1134
1135 int zeromap_page_range(struct vm_area_struct *vma,
1136                         unsigned long addr, unsigned long size, pgprot_t prot)
1137 {
1138         pgd_t *pgd;
1139         unsigned long next;
1140         unsigned long end = addr + size;
1141         struct mm_struct *mm = vma->vm_mm;
1142         int err;
1143
1144         BUG_ON(addr >= end);
1145         pgd = pgd_offset(mm, addr);
1146         flush_cache_range(vma, addr, end);
1147         do {
1148                 next = pgd_addr_end(addr, end);
1149                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1150                 if (err)
1151                         break;
1152         } while (pgd++, addr = next, addr != end);
1153         return err;
1154 }
1155
1156 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1157 {
1158         pgd_t * pgd = pgd_offset(mm, addr);
1159         pud_t * pud = pud_alloc(mm, pgd, addr);
1160         if (pud) {
1161                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1162                 if (pmd)
1163                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1164         }
1165         return NULL;
1166 }
1167
1168 /*
1169  * This is the old fallback for page remapping.
1170  *
1171  * For historical reasons, it only allows reserved pages. Only
1172  * old drivers should use this, and they needed to mark their
1173  * pages reserved for the old functions anyway.
1174  */
1175 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1176 {
1177         int retval;
1178         pte_t *pte;
1179         spinlock_t *ptl;  
1180
1181         retval = -EINVAL;
1182         if (PageAnon(page))
1183                 goto out;
1184         retval = -ENOMEM;
1185         flush_dcache_page(page);
1186         pte = get_locked_pte(mm, addr, &ptl);
1187         if (!pte)
1188                 goto out;
1189         retval = -EBUSY;
1190         if (!pte_none(*pte))
1191                 goto out_unlock;
1192
1193         /* Ok, finally just insert the thing.. */
1194         get_page(page);
1195         inc_mm_counter(mm, file_rss);
1196         page_add_file_rmap(page);
1197         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1198
1199         retval = 0;
1200 out_unlock:
1201         pte_unmap_unlock(pte, ptl);
1202 out:
1203         return retval;
1204 }
1205
1206 /*
1207  * This allows drivers to insert individual pages they've allocated
1208  * into a user vma.
1209  *
1210  * The page has to be a nice clean _individual_ kernel allocation.
1211  * If you allocate a compound page, you need to have marked it as
1212  * such (__GFP_COMP), or manually just split the page up yourself
1213  * (which is mainly an issue of doing "set_page_count(page, 1)" for
1214  * each sub-page, and then freeing them one by one when you free
1215  * them rather than freeing it as a compound page).
1216  *
1217  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1218  * took an arbitrary page protection parameter. This doesn't allow
1219  * that. Your vma protection will have to be set up correctly, which
1220  * means that if you want a shared writable mapping, you'd better
1221  * ask for a shared writable mapping!
1222  *
1223  * The page does not need to be reserved.
1224  */
1225 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1226 {
1227         if (addr < vma->vm_start || addr >= vma->vm_end)
1228                 return -EFAULT;
1229         if (!page_count(page))
1230                 return -EINVAL;
1231         vma->vm_flags |= VM_INSERTPAGE;
1232         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1233 }
1234 EXPORT_SYMBOL(vm_insert_page);
1235
1236 /*
1237  * maps a range of physical memory into the requested pages. the old
1238  * mappings are removed. any references to nonexistent pages results
1239  * in null mappings (currently treated as "copy-on-access")
1240  */
1241 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1242                         unsigned long addr, unsigned long end,
1243                         unsigned long pfn, pgprot_t prot)
1244 {
1245         pte_t *pte;
1246         spinlock_t *ptl;
1247
1248         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1249         if (!pte)
1250                 return -ENOMEM;
1251         do {
1252                 BUG_ON(!pte_none(*pte));
1253                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1254                 pfn++;
1255         } while (pte++, addr += PAGE_SIZE, addr != end);
1256         pte_unmap_unlock(pte - 1, ptl);
1257         return 0;
1258 }
1259
1260 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1261                         unsigned long addr, unsigned long end,
1262                         unsigned long pfn, pgprot_t prot)
1263 {
1264         pmd_t *pmd;
1265         unsigned long next;
1266
1267         pfn -= addr >> PAGE_SHIFT;
1268         pmd = pmd_alloc(mm, pud, addr);
1269         if (!pmd)
1270                 return -ENOMEM;
1271         do {
1272                 next = pmd_addr_end(addr, end);
1273                 if (remap_pte_range(mm, pmd, addr, next,
1274                                 pfn + (addr >> PAGE_SHIFT), prot))
1275                         return -ENOMEM;
1276         } while (pmd++, addr = next, addr != end);
1277         return 0;
1278 }
1279
1280 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1281                         unsigned long addr, unsigned long end,
1282                         unsigned long pfn, pgprot_t prot)
1283 {
1284         pud_t *pud;
1285         unsigned long next;
1286
1287         pfn -= addr >> PAGE_SHIFT;
1288         pud = pud_alloc(mm, pgd, addr);
1289         if (!pud)
1290                 return -ENOMEM;
1291         do {
1292                 next = pud_addr_end(addr, end);
1293                 if (remap_pmd_range(mm, pud, addr, next,
1294                                 pfn + (addr >> PAGE_SHIFT), prot))
1295                         return -ENOMEM;
1296         } while (pud++, addr = next, addr != end);
1297         return 0;
1298 }
1299
1300 /*  Note: this is only safe if the mm semaphore is held when called. */
1301 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1302                     unsigned long pfn, unsigned long size, pgprot_t prot)
1303 {
1304         pgd_t *pgd;
1305         unsigned long next;
1306         unsigned long end = addr + PAGE_ALIGN(size);
1307         struct mm_struct *mm = vma->vm_mm;
1308         int err;
1309
1310         /*
1311          * Physically remapped pages are special. Tell the
1312          * rest of the world about it:
1313          *   VM_IO tells people not to look at these pages
1314          *      (accesses can have side effects).
1315          *   VM_RESERVED is specified all over the place, because
1316          *      in 2.4 it kept swapout's vma scan off this vma; but
1317          *      in 2.6 the LRU scan won't even find its pages, so this
1318          *      flag means no more than count its pages in reserved_vm,
1319          *      and omit it from core dump, even when VM_IO turned off.
1320          *   VM_PFNMAP tells the core MM that the base pages are just
1321          *      raw PFN mappings, and do not have a "struct page" associated
1322          *      with them.
1323          *
1324          * There's a horrible special case to handle copy-on-write
1325          * behaviour that some programs depend on. We mark the "original"
1326          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1327          */
1328         if (is_cow_mapping(vma->vm_flags)) {
1329                 if (addr != vma->vm_start || end != vma->vm_end)
1330                         return -EINVAL;
1331                 vma->vm_pgoff = pfn;
1332         }
1333
1334         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1335
1336         BUG_ON(addr >= end);
1337         pfn -= addr >> PAGE_SHIFT;
1338         pgd = pgd_offset(mm, addr);
1339         flush_cache_range(vma, addr, end);
1340         do {
1341                 next = pgd_addr_end(addr, end);
1342                 err = remap_pud_range(mm, pgd, addr, next,
1343                                 pfn + (addr >> PAGE_SHIFT), prot);
1344                 if (err)
1345                         break;
1346         } while (pgd++, addr = next, addr != end);
1347         return err;
1348 }
1349 EXPORT_SYMBOL(remap_pfn_range);
1350
1351 /*
1352  * handle_pte_fault chooses page fault handler according to an entry
1353  * which was read non-atomically.  Before making any commitment, on
1354  * those architectures or configurations (e.g. i386 with PAE) which
1355  * might give a mix of unmatched parts, do_swap_page and do_file_page
1356  * must check under lock before unmapping the pte and proceeding
1357  * (but do_wp_page is only called after already making such a check;
1358  * and do_anonymous_page and do_no_page can safely check later on).
1359  */
1360 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1361                                 pte_t *page_table, pte_t orig_pte)
1362 {
1363         int same = 1;
1364 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1365         if (sizeof(pte_t) > sizeof(unsigned long)) {
1366                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1367                 spin_lock(ptl);
1368                 same = pte_same(*page_table, orig_pte);
1369                 spin_unlock(ptl);
1370         }
1371 #endif
1372         pte_unmap(page_table);
1373         return same;
1374 }
1375
1376 /*
1377  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1378  * servicing faults for write access.  In the normal case, do always want
1379  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1380  * that do not have writing enabled, when used by access_process_vm.
1381  */
1382 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1383 {
1384         if (likely(vma->vm_flags & VM_WRITE))
1385                 pte = pte_mkwrite(pte);
1386         return pte;
1387 }
1388
1389 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1390 {
1391         /*
1392          * If the source page was a PFN mapping, we don't have
1393          * a "struct page" for it. We do a best-effort copy by
1394          * just copying from the original user address. If that
1395          * fails, we just zero-fill it. Live with it.
1396          */
1397         if (unlikely(!src)) {
1398                 void *kaddr = kmap_atomic(dst, KM_USER0);
1399                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1400
1401                 /*
1402                  * This really shouldn't fail, because the page is there
1403                  * in the page tables. But it might just be unreadable,
1404                  * in which case we just give up and fill the result with
1405                  * zeroes.
1406                  */
1407                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1408                         memset(kaddr, 0, PAGE_SIZE);
1409                 kunmap_atomic(kaddr, KM_USER0);
1410                 return;
1411                 
1412         }
1413         copy_user_highpage(dst, src, va);
1414 }
1415
1416 /*
1417  * This routine handles present pages, when users try to write
1418  * to a shared page. It is done by copying the page to a new address
1419  * and decrementing the shared-page counter for the old page.
1420  *
1421  * Note that this routine assumes that the protection checks have been
1422  * done by the caller (the low-level page fault routine in most cases).
1423  * Thus we can safely just mark it writable once we've done any necessary
1424  * COW.
1425  *
1426  * We also mark the page dirty at this point even though the page will
1427  * change only once the write actually happens. This avoids a few races,
1428  * and potentially makes it more efficient.
1429  *
1430  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1431  * but allow concurrent faults), with pte both mapped and locked.
1432  * We return with mmap_sem still held, but pte unmapped and unlocked.
1433  */
1434 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1435                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1436                 spinlock_t *ptl, pte_t orig_pte)
1437 {
1438         struct page *old_page, *new_page;
1439         pte_t entry;
1440         int ret = VM_FAULT_MINOR;
1441
1442         old_page = vm_normal_page(vma, address, orig_pte);
1443         if (!old_page)
1444                 goto gotten;
1445
1446         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1447                 int reuse = can_share_swap_page(old_page);
1448                 unlock_page(old_page);
1449                 if (reuse) {
1450                         flush_cache_page(vma, address, pte_pfn(orig_pte));
1451                         entry = pte_mkyoung(orig_pte);
1452                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1453                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1454                         update_mmu_cache(vma, address, entry);
1455                         lazy_mmu_prot_update(entry);
1456                         ret |= VM_FAULT_WRITE;
1457                         goto unlock;
1458                 }
1459         }
1460
1461         /*
1462          * Ok, we need to copy. Oh, well..
1463          */
1464         page_cache_get(old_page);
1465 gotten:
1466         pte_unmap_unlock(page_table, ptl);
1467
1468         if (unlikely(anon_vma_prepare(vma)))
1469                 goto oom;
1470         if (old_page == ZERO_PAGE(address)) {
1471                 new_page = alloc_zeroed_user_highpage(vma, address);
1472                 if (!new_page)
1473                         goto oom;
1474         } else {
1475                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1476                 if (!new_page)
1477                         goto oom;
1478                 cow_user_page(new_page, old_page, address);
1479         }
1480
1481         /*
1482          * Re-check the pte - we dropped the lock
1483          */
1484         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1485         if (likely(pte_same(*page_table, orig_pte))) {
1486                 if (old_page) {
1487                         page_remove_rmap(old_page);
1488                         if (!PageAnon(old_page)) {
1489                                 dec_mm_counter(mm, file_rss);
1490                                 inc_mm_counter(mm, anon_rss);
1491                         }
1492                 } else
1493                         inc_mm_counter(mm, anon_rss);
1494                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1495                 entry = mk_pte(new_page, vma->vm_page_prot);
1496                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1497                 ptep_establish(vma, address, page_table, entry);
1498                 update_mmu_cache(vma, address, entry);
1499                 lazy_mmu_prot_update(entry);
1500                 lru_cache_add_active(new_page);
1501                 page_add_new_anon_rmap(new_page, vma, address);
1502
1503                 /* Free the old page.. */
1504                 new_page = old_page;
1505                 ret |= VM_FAULT_WRITE;
1506         }
1507         if (new_page)
1508                 page_cache_release(new_page);
1509         if (old_page)
1510                 page_cache_release(old_page);
1511 unlock:
1512         pte_unmap_unlock(page_table, ptl);
1513         return ret;
1514 oom:
1515         if (old_page)
1516                 page_cache_release(old_page);
1517         return VM_FAULT_OOM;
1518 }
1519
1520 /*
1521  * Helper functions for unmap_mapping_range().
1522  *
1523  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1524  *
1525  * We have to restart searching the prio_tree whenever we drop the lock,
1526  * since the iterator is only valid while the lock is held, and anyway
1527  * a later vma might be split and reinserted earlier while lock dropped.
1528  *
1529  * The list of nonlinear vmas could be handled more efficiently, using
1530  * a placeholder, but handle it in the same way until a need is shown.
1531  * It is important to search the prio_tree before nonlinear list: a vma
1532  * may become nonlinear and be shifted from prio_tree to nonlinear list
1533  * while the lock is dropped; but never shifted from list to prio_tree.
1534  *
1535  * In order to make forward progress despite restarting the search,
1536  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1537  * quickly skip it next time around.  Since the prio_tree search only
1538  * shows us those vmas affected by unmapping the range in question, we
1539  * can't efficiently keep all vmas in step with mapping->truncate_count:
1540  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1541  * mapping->truncate_count and vma->vm_truncate_count are protected by
1542  * i_mmap_lock.
1543  *
1544  * In order to make forward progress despite repeatedly restarting some
1545  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1546  * and restart from that address when we reach that vma again.  It might
1547  * have been split or merged, shrunk or extended, but never shifted: so
1548  * restart_addr remains valid so long as it remains in the vma's range.
1549  * unmap_mapping_range forces truncate_count to leap over page-aligned
1550  * values so we can save vma's restart_addr in its truncate_count field.
1551  */
1552 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1553
1554 static void reset_vma_truncate_counts(struct address_space *mapping)
1555 {
1556         struct vm_area_struct *vma;
1557         struct prio_tree_iter iter;
1558
1559         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1560                 vma->vm_truncate_count = 0;
1561         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1562                 vma->vm_truncate_count = 0;
1563 }
1564
1565 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1566                 unsigned long start_addr, unsigned long end_addr,
1567                 struct zap_details *details)
1568 {
1569         unsigned long restart_addr;
1570         int need_break;
1571
1572 again:
1573         restart_addr = vma->vm_truncate_count;
1574         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1575                 start_addr = restart_addr;
1576                 if (start_addr >= end_addr) {
1577                         /* Top of vma has been split off since last time */
1578                         vma->vm_truncate_count = details->truncate_count;
1579                         return 0;
1580                 }
1581         }
1582
1583         restart_addr = zap_page_range(vma, start_addr,
1584                                         end_addr - start_addr, details);
1585         need_break = need_resched() ||
1586                         need_lockbreak(details->i_mmap_lock);
1587
1588         if (restart_addr >= end_addr) {
1589                 /* We have now completed this vma: mark it so */
1590                 vma->vm_truncate_count = details->truncate_count;
1591                 if (!need_break)
1592                         return 0;
1593         } else {
1594                 /* Note restart_addr in vma's truncate_count field */
1595                 vma->vm_truncate_count = restart_addr;
1596                 if (!need_break)
1597                         goto again;
1598         }
1599
1600         spin_unlock(details->i_mmap_lock);
1601         cond_resched();
1602         spin_lock(details->i_mmap_lock);
1603         return -EINTR;
1604 }
1605
1606 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1607                                             struct zap_details *details)
1608 {
1609         struct vm_area_struct *vma;
1610         struct prio_tree_iter iter;
1611         pgoff_t vba, vea, zba, zea;
1612
1613 restart:
1614         vma_prio_tree_foreach(vma, &iter, root,
1615                         details->first_index, details->last_index) {
1616                 /* Skip quickly over those we have already dealt with */
1617                 if (vma->vm_truncate_count == details->truncate_count)
1618                         continue;
1619
1620                 vba = vma->vm_pgoff;
1621                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1622                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1623                 zba = details->first_index;
1624                 if (zba < vba)
1625                         zba = vba;
1626                 zea = details->last_index;
1627                 if (zea > vea)
1628                         zea = vea;
1629
1630                 if (unmap_mapping_range_vma(vma,
1631                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1632                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1633                                 details) < 0)
1634                         goto restart;
1635         }
1636 }
1637
1638 static inline void unmap_mapping_range_list(struct list_head *head,
1639                                             struct zap_details *details)
1640 {
1641         struct vm_area_struct *vma;
1642
1643         /*
1644          * In nonlinear VMAs there is no correspondence between virtual address
1645          * offset and file offset.  So we must perform an exhaustive search
1646          * across *all* the pages in each nonlinear VMA, not just the pages
1647          * whose virtual address lies outside the file truncation point.
1648          */
1649 restart:
1650         list_for_each_entry(vma, head, shared.vm_set.list) {
1651                 /* Skip quickly over those we have already dealt with */
1652                 if (vma->vm_truncate_count == details->truncate_count)
1653                         continue;
1654                 details->nonlinear_vma = vma;
1655                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1656                                         vma->vm_end, details) < 0)
1657                         goto restart;
1658         }
1659 }
1660
1661 /**
1662  * unmap_mapping_range - unmap the portion of all mmaps
1663  * in the specified address_space corresponding to the specified
1664  * page range in the underlying file.
1665  * @mapping: the address space containing mmaps to be unmapped.
1666  * @holebegin: byte in first page to unmap, relative to the start of
1667  * the underlying file.  This will be rounded down to a PAGE_SIZE
1668  * boundary.  Note that this is different from vmtruncate(), which
1669  * must keep the partial page.  In contrast, we must get rid of
1670  * partial pages.
1671  * @holelen: size of prospective hole in bytes.  This will be rounded
1672  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1673  * end of the file.
1674  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1675  * but 0 when invalidating pagecache, don't throw away private data.
1676  */
1677 void unmap_mapping_range(struct address_space *mapping,
1678                 loff_t const holebegin, loff_t const holelen, int even_cows)
1679 {
1680         struct zap_details details;
1681         pgoff_t hba = holebegin >> PAGE_SHIFT;
1682         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1683
1684         /* Check for overflow. */
1685         if (sizeof(holelen) > sizeof(hlen)) {
1686                 long long holeend =
1687                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1688                 if (holeend & ~(long long)ULONG_MAX)
1689                         hlen = ULONG_MAX - hba + 1;
1690         }
1691
1692         details.check_mapping = even_cows? NULL: mapping;
1693         details.nonlinear_vma = NULL;
1694         details.first_index = hba;
1695         details.last_index = hba + hlen - 1;
1696         if (details.last_index < details.first_index)
1697                 details.last_index = ULONG_MAX;
1698         details.i_mmap_lock = &mapping->i_mmap_lock;
1699
1700         spin_lock(&mapping->i_mmap_lock);
1701
1702         /* serialize i_size write against truncate_count write */
1703         smp_wmb();
1704         /* Protect against page faults, and endless unmapping loops */
1705         mapping->truncate_count++;
1706         /*
1707          * For archs where spin_lock has inclusive semantics like ia64
1708          * this smp_mb() will prevent to read pagetable contents
1709          * before the truncate_count increment is visible to
1710          * other cpus.
1711          */
1712         smp_mb();
1713         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1714                 if (mapping->truncate_count == 0)
1715                         reset_vma_truncate_counts(mapping);
1716                 mapping->truncate_count++;
1717         }
1718         details.truncate_count = mapping->truncate_count;
1719
1720         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1721                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1722         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1723                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1724         spin_unlock(&mapping->i_mmap_lock);
1725 }
1726 EXPORT_SYMBOL(unmap_mapping_range);
1727
1728 /*
1729  * Handle all mappings that got truncated by a "truncate()"
1730  * system call.
1731  *
1732  * NOTE! We have to be ready to update the memory sharing
1733  * between the file and the memory map for a potential last
1734  * incomplete page.  Ugly, but necessary.
1735  */
1736 int vmtruncate(struct inode * inode, loff_t offset)
1737 {
1738         struct address_space *mapping = inode->i_mapping;
1739         unsigned long limit;
1740
1741         if (inode->i_size < offset)
1742                 goto do_expand;
1743         /*
1744          * truncation of in-use swapfiles is disallowed - it would cause
1745          * subsequent swapout to scribble on the now-freed blocks.
1746          */
1747         if (IS_SWAPFILE(inode))
1748                 goto out_busy;
1749         i_size_write(inode, offset);
1750         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1751         truncate_inode_pages(mapping, offset);
1752         goto out_truncate;
1753
1754 do_expand:
1755         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1756         if (limit != RLIM_INFINITY && offset > limit)
1757                 goto out_sig;
1758         if (offset > inode->i_sb->s_maxbytes)
1759                 goto out_big;
1760         i_size_write(inode, offset);
1761
1762 out_truncate:
1763         if (inode->i_op && inode->i_op->truncate)
1764                 inode->i_op->truncate(inode);
1765         return 0;
1766 out_sig:
1767         send_sig(SIGXFSZ, current, 0);
1768 out_big:
1769         return -EFBIG;
1770 out_busy:
1771         return -ETXTBSY;
1772 }
1773 EXPORT_SYMBOL(vmtruncate);
1774
1775 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1776 {
1777         struct address_space *mapping = inode->i_mapping;
1778
1779         /*
1780          * If the underlying filesystem is not going to provide
1781          * a way to truncate a range of blocks (punch a hole) -
1782          * we should return failure right now.
1783          */
1784         if (!inode->i_op || !inode->i_op->truncate_range)
1785                 return -ENOSYS;
1786
1787         down(&inode->i_sem);
1788         down_write(&inode->i_alloc_sem);
1789         unmap_mapping_range(mapping, offset, (end - offset), 1);
1790         truncate_inode_pages_range(mapping, offset, end);
1791         inode->i_op->truncate_range(inode, offset, end);
1792         up_write(&inode->i_alloc_sem);
1793         up(&inode->i_sem);
1794
1795         return 0;
1796 }
1797 EXPORT_SYMBOL(vmtruncate_range);
1798
1799 /* 
1800  * Primitive swap readahead code. We simply read an aligned block of
1801  * (1 << page_cluster) entries in the swap area. This method is chosen
1802  * because it doesn't cost us any seek time.  We also make sure to queue
1803  * the 'original' request together with the readahead ones...  
1804  *
1805  * This has been extended to use the NUMA policies from the mm triggering
1806  * the readahead.
1807  *
1808  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1809  */
1810 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1811 {
1812 #ifdef CONFIG_NUMA
1813         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1814 #endif
1815         int i, num;
1816         struct page *new_page;
1817         unsigned long offset;
1818
1819         /*
1820          * Get the number of handles we should do readahead io to.
1821          */
1822         num = valid_swaphandles(entry, &offset);
1823         for (i = 0; i < num; offset++, i++) {
1824                 /* Ok, do the async read-ahead now */
1825                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1826                                                            offset), vma, addr);
1827                 if (!new_page)
1828                         break;
1829                 page_cache_release(new_page);
1830 #ifdef CONFIG_NUMA
1831                 /*
1832                  * Find the next applicable VMA for the NUMA policy.
1833                  */
1834                 addr += PAGE_SIZE;
1835                 if (addr == 0)
1836                         vma = NULL;
1837                 if (vma) {
1838                         if (addr >= vma->vm_end) {
1839                                 vma = next_vma;
1840                                 next_vma = vma ? vma->vm_next : NULL;
1841                         }
1842                         if (vma && addr < vma->vm_start)
1843                                 vma = NULL;
1844                 } else {
1845                         if (next_vma && addr >= next_vma->vm_start) {
1846                                 vma = next_vma;
1847                                 next_vma = vma->vm_next;
1848                         }
1849                 }
1850 #endif
1851         }
1852         lru_add_drain();        /* Push any new pages onto the LRU now */
1853 }
1854
1855 /*
1856  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1857  * but allow concurrent faults), and pte mapped but not yet locked.
1858  * We return with mmap_sem still held, but pte unmapped and unlocked.
1859  */
1860 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1861                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1862                 int write_access, pte_t orig_pte)
1863 {
1864         spinlock_t *ptl;
1865         struct page *page;
1866         swp_entry_t entry;
1867         pte_t pte;
1868         int ret = VM_FAULT_MINOR;
1869
1870         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1871                 goto out;
1872
1873         entry = pte_to_swp_entry(orig_pte);
1874         page = lookup_swap_cache(entry);
1875         if (!page) {
1876                 swapin_readahead(entry, address, vma);
1877                 page = read_swap_cache_async(entry, vma, address);
1878                 if (!page) {
1879                         /*
1880                          * Back out if somebody else faulted in this pte
1881                          * while we released the pte lock.
1882                          */
1883                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1884                         if (likely(pte_same(*page_table, orig_pte)))
1885                                 ret = VM_FAULT_OOM;
1886                         goto unlock;
1887                 }
1888
1889                 /* Had to read the page from swap area: Major fault */
1890                 ret = VM_FAULT_MAJOR;
1891                 inc_page_state(pgmajfault);
1892                 grab_swap_token();
1893         }
1894
1895         mark_page_accessed(page);
1896         lock_page(page);
1897
1898         /*
1899          * Back out if somebody else already faulted in this pte.
1900          */
1901         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1902         if (unlikely(!pte_same(*page_table, orig_pte)))
1903                 goto out_nomap;
1904
1905         if (unlikely(!PageUptodate(page))) {
1906                 ret = VM_FAULT_SIGBUS;
1907                 goto out_nomap;
1908         }
1909
1910         /* The page isn't present yet, go ahead with the fault. */
1911
1912         inc_mm_counter(mm, anon_rss);
1913         pte = mk_pte(page, vma->vm_page_prot);
1914         if (write_access && can_share_swap_page(page)) {
1915                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1916                 write_access = 0;
1917         }
1918
1919         flush_icache_page(vma, page);
1920         set_pte_at(mm, address, page_table, pte);
1921         page_add_anon_rmap(page, vma, address);
1922
1923         swap_free(entry);
1924         if (vm_swap_full())
1925                 remove_exclusive_swap_page(page);
1926         unlock_page(page);
1927
1928         if (write_access) {
1929                 if (do_wp_page(mm, vma, address,
1930                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1931                         ret = VM_FAULT_OOM;
1932                 goto out;
1933         }
1934
1935         /* No need to invalidate - it was non-present before */
1936         update_mmu_cache(vma, address, pte);
1937         lazy_mmu_prot_update(pte);
1938 unlock:
1939         pte_unmap_unlock(page_table, ptl);
1940 out:
1941         return ret;
1942 out_nomap:
1943         pte_unmap_unlock(page_table, ptl);
1944         unlock_page(page);
1945         page_cache_release(page);
1946         return ret;
1947 }
1948
1949 /*
1950  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1951  * but allow concurrent faults), and pte mapped but not yet locked.
1952  * We return with mmap_sem still held, but pte unmapped and unlocked.
1953  */
1954 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1955                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1956                 int write_access)
1957 {
1958         struct page *page;
1959         spinlock_t *ptl;
1960         pte_t entry;
1961
1962         if (write_access) {
1963                 /* Allocate our own private page. */
1964                 pte_unmap(page_table);
1965
1966                 if (unlikely(anon_vma_prepare(vma)))
1967                         goto oom;
1968                 page = alloc_zeroed_user_highpage(vma, address);
1969                 if (!page)
1970                         goto oom;
1971
1972                 entry = mk_pte(page, vma->vm_page_prot);
1973                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1974
1975                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1976                 if (!pte_none(*page_table))
1977                         goto release;
1978                 inc_mm_counter(mm, anon_rss);
1979                 lru_cache_add_active(page);
1980                 SetPageReferenced(page);
1981                 page_add_new_anon_rmap(page, vma, address);
1982         } else {
1983                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1984                 page = ZERO_PAGE(address);
1985                 page_cache_get(page);
1986                 entry = mk_pte(page, vma->vm_page_prot);
1987
1988                 ptl = pte_lockptr(mm, pmd);
1989                 spin_lock(ptl);
1990                 if (!pte_none(*page_table))
1991                         goto release;
1992                 inc_mm_counter(mm, file_rss);
1993                 page_add_file_rmap(page);
1994         }
1995
1996         set_pte_at(mm, address, page_table, entry);
1997
1998         /* No need to invalidate - it was non-present before */
1999         update_mmu_cache(vma, address, entry);
2000         lazy_mmu_prot_update(entry);
2001 unlock:
2002         pte_unmap_unlock(page_table, ptl);
2003         return VM_FAULT_MINOR;
2004 release:
2005         page_cache_release(page);
2006         goto unlock;
2007 oom:
2008         return VM_FAULT_OOM;
2009 }
2010
2011 /*
2012  * do_no_page() tries to create a new page mapping. It aggressively
2013  * tries to share with existing pages, but makes a separate copy if
2014  * the "write_access" parameter is true in order to avoid the next
2015  * page fault.
2016  *
2017  * As this is called only for pages that do not currently exist, we
2018  * do not need to flush old virtual caches or the TLB.
2019  *
2020  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2021  * but allow concurrent faults), and pte mapped but not yet locked.
2022  * We return with mmap_sem still held, but pte unmapped and unlocked.
2023  */
2024 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2025                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2026                 int write_access)
2027 {
2028         spinlock_t *ptl;
2029         struct page *new_page;
2030         struct address_space *mapping = NULL;
2031         pte_t entry;
2032         unsigned int sequence = 0;
2033         int ret = VM_FAULT_MINOR;
2034         int anon = 0;
2035
2036         pte_unmap(page_table);
2037         BUG_ON(vma->vm_flags & VM_PFNMAP);
2038
2039         if (vma->vm_file) {
2040                 mapping = vma->vm_file->f_mapping;
2041                 sequence = mapping->truncate_count;
2042                 smp_rmb(); /* serializes i_size against truncate_count */
2043         }
2044 retry:
2045         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2046         /*
2047          * No smp_rmb is needed here as long as there's a full
2048          * spin_lock/unlock sequence inside the ->nopage callback
2049          * (for the pagecache lookup) that acts as an implicit
2050          * smp_mb() and prevents the i_size read to happen
2051          * after the next truncate_count read.
2052          */
2053
2054         /* no page was available -- either SIGBUS or OOM */
2055         if (new_page == NOPAGE_SIGBUS)
2056                 return VM_FAULT_SIGBUS;
2057         if (new_page == NOPAGE_OOM)
2058                 return VM_FAULT_OOM;
2059
2060         /*
2061          * Should we do an early C-O-W break?
2062          */
2063         if (write_access && !(vma->vm_flags & VM_SHARED)) {
2064                 struct page *page;
2065
2066                 if (unlikely(anon_vma_prepare(vma)))
2067                         goto oom;
2068                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2069                 if (!page)
2070                         goto oom;
2071                 copy_user_highpage(page, new_page, address);
2072                 page_cache_release(new_page);
2073                 new_page = page;
2074                 anon = 1;
2075         }
2076
2077         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2078         /*
2079          * For a file-backed vma, someone could have truncated or otherwise
2080          * invalidated this page.  If unmap_mapping_range got called,
2081          * retry getting the page.
2082          */
2083         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2084                 pte_unmap_unlock(page_table, ptl);
2085                 page_cache_release(new_page);
2086                 cond_resched();
2087                 sequence = mapping->truncate_count;
2088                 smp_rmb();
2089                 goto retry;
2090         }
2091
2092         /*
2093          * This silly early PAGE_DIRTY setting removes a race
2094          * due to the bad i386 page protection. But it's valid
2095          * for other architectures too.
2096          *
2097          * Note that if write_access is true, we either now have
2098          * an exclusive copy of the page, or this is a shared mapping,
2099          * so we can make it writable and dirty to avoid having to
2100          * handle that later.
2101          */
2102         /* Only go through if we didn't race with anybody else... */
2103         if (pte_none(*page_table)) {
2104                 flush_icache_page(vma, new_page);
2105                 entry = mk_pte(new_page, vma->vm_page_prot);
2106                 if (write_access)
2107                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2108                 set_pte_at(mm, address, page_table, entry);
2109                 if (anon) {
2110                         inc_mm_counter(mm, anon_rss);
2111                         lru_cache_add_active(new_page);
2112                         page_add_new_anon_rmap(new_page, vma, address);
2113                 } else {
2114                         inc_mm_counter(mm, file_rss);
2115                         page_add_file_rmap(new_page);
2116                 }
2117         } else {
2118                 /* One of our sibling threads was faster, back out. */
2119                 page_cache_release(new_page);
2120                 goto unlock;
2121         }
2122
2123         /* no need to invalidate: a not-present page shouldn't be cached */
2124         update_mmu_cache(vma, address, entry);
2125         lazy_mmu_prot_update(entry);
2126 unlock:
2127         pte_unmap_unlock(page_table, ptl);
2128         return ret;
2129 oom:
2130         page_cache_release(new_page);
2131         return VM_FAULT_OOM;
2132 }
2133
2134 /*
2135  * Fault of a previously existing named mapping. Repopulate the pte
2136  * from the encoded file_pte if possible. This enables swappable
2137  * nonlinear vmas.
2138  *
2139  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2140  * but allow concurrent faults), and pte mapped but not yet locked.
2141  * We return with mmap_sem still held, but pte unmapped and unlocked.
2142  */
2143 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2144                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2145                 int write_access, pte_t orig_pte)
2146 {
2147         pgoff_t pgoff;
2148         int err;
2149
2150         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2151                 return VM_FAULT_MINOR;
2152
2153         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2154                 /*
2155                  * Page table corrupted: show pte and kill process.
2156                  */
2157                 print_bad_pte(vma, orig_pte, address);
2158                 return VM_FAULT_OOM;
2159         }
2160         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2161
2162         pgoff = pte_to_pgoff(orig_pte);
2163         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2164                                         vma->vm_page_prot, pgoff, 0);
2165         if (err == -ENOMEM)
2166                 return VM_FAULT_OOM;
2167         if (err)
2168                 return VM_FAULT_SIGBUS;
2169         return VM_FAULT_MAJOR;
2170 }
2171
2172 /*
2173  * These routines also need to handle stuff like marking pages dirty
2174  * and/or accessed for architectures that don't do it in hardware (most
2175  * RISC architectures).  The early dirtying is also good on the i386.
2176  *
2177  * There is also a hook called "update_mmu_cache()" that architectures
2178  * with external mmu caches can use to update those (ie the Sparc or
2179  * PowerPC hashed page tables that act as extended TLBs).
2180  *
2181  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2182  * but allow concurrent faults), and pte mapped but not yet locked.
2183  * We return with mmap_sem still held, but pte unmapped and unlocked.
2184  */
2185 static inline int handle_pte_fault(struct mm_struct *mm,
2186                 struct vm_area_struct *vma, unsigned long address,
2187                 pte_t *pte, pmd_t *pmd, int write_access)
2188 {
2189         pte_t entry;
2190         pte_t old_entry;
2191         spinlock_t *ptl;
2192
2193         old_entry = entry = *pte;
2194         if (!pte_present(entry)) {
2195                 if (pte_none(entry)) {
2196                         if (!vma->vm_ops || !vma->vm_ops->nopage)
2197                                 return do_anonymous_page(mm, vma, address,
2198                                         pte, pmd, write_access);
2199                         return do_no_page(mm, vma, address,
2200                                         pte, pmd, write_access);
2201                 }
2202                 if (pte_file(entry))
2203                         return do_file_page(mm, vma, address,
2204                                         pte, pmd, write_access, entry);
2205                 return do_swap_page(mm, vma, address,
2206                                         pte, pmd, write_access, entry);
2207         }
2208
2209         ptl = pte_lockptr(mm, pmd);
2210         spin_lock(ptl);
2211         if (unlikely(!pte_same(*pte, entry)))
2212                 goto unlock;
2213         if (write_access) {
2214                 if (!pte_write(entry))
2215                         return do_wp_page(mm, vma, address,
2216                                         pte, pmd, ptl, entry);
2217                 entry = pte_mkdirty(entry);
2218         }
2219         entry = pte_mkyoung(entry);
2220         if (!pte_same(old_entry, entry)) {
2221                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2222                 update_mmu_cache(vma, address, entry);
2223                 lazy_mmu_prot_update(entry);
2224         } else {
2225                 /*
2226                  * This is needed only for protection faults but the arch code
2227                  * is not yet telling us if this is a protection fault or not.
2228                  * This still avoids useless tlb flushes for .text page faults
2229                  * with threads.
2230                  */
2231                 if (write_access)
2232                         flush_tlb_page(vma, address);
2233         }
2234 unlock:
2235         pte_unmap_unlock(pte, ptl);
2236         return VM_FAULT_MINOR;
2237 }
2238
2239 /*
2240  * By the time we get here, we already hold the mm semaphore
2241  */
2242 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2243                 unsigned long address, int write_access)
2244 {
2245         pgd_t *pgd;
2246         pud_t *pud;
2247         pmd_t *pmd;
2248         pte_t *pte;
2249
2250         __set_current_state(TASK_RUNNING);
2251
2252         inc_page_state(pgfault);
2253
2254         if (unlikely(is_vm_hugetlb_page(vma)))
2255                 return hugetlb_fault(mm, vma, address, write_access);
2256
2257         pgd = pgd_offset(mm, address);
2258         pud = pud_alloc(mm, pgd, address);
2259         if (!pud)
2260                 return VM_FAULT_OOM;
2261         pmd = pmd_alloc(mm, pud, address);
2262         if (!pmd)
2263                 return VM_FAULT_OOM;
2264         pte = pte_alloc_map(mm, pmd, address);
2265         if (!pte)
2266                 return VM_FAULT_OOM;
2267
2268         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2269 }
2270
2271 #ifndef __PAGETABLE_PUD_FOLDED
2272 /*
2273  * Allocate page upper directory.
2274  * We've already handled the fast-path in-line.
2275  */
2276 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2277 {
2278         pud_t *new = pud_alloc_one(mm, address);
2279         if (!new)
2280                 return -ENOMEM;
2281
2282         spin_lock(&mm->page_table_lock);
2283         if (pgd_present(*pgd))          /* Another has populated it */
2284                 pud_free(new);
2285         else
2286                 pgd_populate(mm, pgd, new);
2287         spin_unlock(&mm->page_table_lock);
2288         return 0;
2289 }
2290 #else
2291 /* Workaround for gcc 2.96 */
2292 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2293 {
2294         return 0;
2295 }
2296 #endif /* __PAGETABLE_PUD_FOLDED */
2297
2298 #ifndef __PAGETABLE_PMD_FOLDED
2299 /*
2300  * Allocate page middle directory.
2301  * We've already handled the fast-path in-line.
2302  */
2303 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2304 {
2305         pmd_t *new = pmd_alloc_one(mm, address);
2306         if (!new)
2307                 return -ENOMEM;
2308
2309         spin_lock(&mm->page_table_lock);
2310 #ifndef __ARCH_HAS_4LEVEL_HACK
2311         if (pud_present(*pud))          /* Another has populated it */
2312                 pmd_free(new);
2313         else
2314                 pud_populate(mm, pud, new);
2315 #else
2316         if (pgd_present(*pud))          /* Another has populated it */
2317                 pmd_free(new);
2318         else
2319                 pgd_populate(mm, pud, new);
2320 #endif /* __ARCH_HAS_4LEVEL_HACK */
2321         spin_unlock(&mm->page_table_lock);
2322         return 0;
2323 }
2324 #else
2325 /* Workaround for gcc 2.96 */
2326 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2327 {
2328         return 0;
2329 }
2330 #endif /* __PAGETABLE_PMD_FOLDED */
2331
2332 int make_pages_present(unsigned long addr, unsigned long end)
2333 {
2334         int ret, len, write;
2335         struct vm_area_struct * vma;
2336
2337         vma = find_vma(current->mm, addr);
2338         if (!vma)
2339                 return -1;
2340         write = (vma->vm_flags & VM_WRITE) != 0;
2341         if (addr >= end)
2342                 BUG();
2343         if (end > vma->vm_end)
2344                 BUG();
2345         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2346         ret = get_user_pages(current, current->mm, addr,
2347                         len, write, 0, NULL, NULL);
2348         if (ret < 0)
2349                 return ret;
2350         return ret == len ? 0 : -1;
2351 }
2352
2353 /* 
2354  * Map a vmalloc()-space virtual address to the physical page.
2355  */
2356 struct page * vmalloc_to_page(void * vmalloc_addr)
2357 {
2358         unsigned long addr = (unsigned long) vmalloc_addr;
2359         struct page *page = NULL;
2360         pgd_t *pgd = pgd_offset_k(addr);
2361         pud_t *pud;
2362         pmd_t *pmd;
2363         pte_t *ptep, pte;
2364   
2365         if (!pgd_none(*pgd)) {
2366                 pud = pud_offset(pgd, addr);
2367                 if (!pud_none(*pud)) {
2368                         pmd = pmd_offset(pud, addr);
2369                         if (!pmd_none(*pmd)) {
2370                                 ptep = pte_offset_map(pmd, addr);
2371                                 pte = *ptep;
2372                                 if (pte_present(pte))
2373                                         page = pte_page(pte);
2374                                 pte_unmap(ptep);
2375                         }
2376                 }
2377         }
2378         return page;
2379 }
2380
2381 EXPORT_SYMBOL(vmalloc_to_page);
2382
2383 /*
2384  * Map a vmalloc()-space virtual address to the physical page frame number.
2385  */
2386 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2387 {
2388         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2389 }
2390
2391 EXPORT_SYMBOL(vmalloc_to_pfn);
2392
2393 #if !defined(__HAVE_ARCH_GATE_AREA)
2394
2395 #if defined(AT_SYSINFO_EHDR)
2396 static struct vm_area_struct gate_vma;
2397
2398 static int __init gate_vma_init(void)
2399 {
2400         gate_vma.vm_mm = NULL;
2401         gate_vma.vm_start = FIXADDR_USER_START;
2402         gate_vma.vm_end = FIXADDR_USER_END;
2403         gate_vma.vm_page_prot = PAGE_READONLY;
2404         gate_vma.vm_flags = 0;
2405         return 0;
2406 }
2407 __initcall(gate_vma_init);
2408 #endif
2409
2410 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2411 {
2412 #ifdef AT_SYSINFO_EHDR
2413         return &gate_vma;
2414 #else
2415         return NULL;
2416 #endif
2417 }
2418
2419 int in_gate_area_no_task(unsigned long addr)
2420 {
2421 #ifdef AT_SYSINFO_EHDR
2422         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2423                 return 1;
2424 #endif
2425         return 0;
2426 }
2427
2428 #endif  /* __HAVE_ARCH_GATE_AREA */