Merge branch 'master'
[linux-2.6] / 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 int randomize_va_space __read_mostly = 1;
86
87 static int __init disable_randmaps(char *s)
88 {
89         randomize_va_space = 0;
90         return 0;
91 }
92 __setup("norandmaps", disable_randmaps);
93
94
95 /*
96  * If a p?d_bad entry is found while walking page tables, report
97  * the error, before resetting entry to p?d_none.  Usually (but
98  * very seldom) called out from the p?d_none_or_clear_bad macros.
99  */
100
101 void pgd_clear_bad(pgd_t *pgd)
102 {
103         pgd_ERROR(*pgd);
104         pgd_clear(pgd);
105 }
106
107 void pud_clear_bad(pud_t *pud)
108 {
109         pud_ERROR(*pud);
110         pud_clear(pud);
111 }
112
113 void pmd_clear_bad(pmd_t *pmd)
114 {
115         pmd_ERROR(*pmd);
116         pmd_clear(pmd);
117 }
118
119 /*
120  * Note: this doesn't free the actual pages themselves. That
121  * has been handled earlier when unmapping all the memory regions.
122  */
123 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
124 {
125         struct page *page = pmd_page(*pmd);
126         pmd_clear(pmd);
127         pte_lock_deinit(page);
128         pte_free_tlb(tlb, page);
129         dec_page_state(nr_page_table_pages);
130         tlb->mm->nr_ptes--;
131 }
132
133 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
134                                 unsigned long addr, unsigned long end,
135                                 unsigned long floor, unsigned long ceiling)
136 {
137         pmd_t *pmd;
138         unsigned long next;
139         unsigned long start;
140
141         start = addr;
142         pmd = pmd_offset(pud, addr);
143         do {
144                 next = pmd_addr_end(addr, end);
145                 if (pmd_none_or_clear_bad(pmd))
146                         continue;
147                 free_pte_range(tlb, pmd);
148         } while (pmd++, addr = next, addr != end);
149
150         start &= PUD_MASK;
151         if (start < floor)
152                 return;
153         if (ceiling) {
154                 ceiling &= PUD_MASK;
155                 if (!ceiling)
156                         return;
157         }
158         if (end - 1 > ceiling - 1)
159                 return;
160
161         pmd = pmd_offset(pud, start);
162         pud_clear(pud);
163         pmd_free_tlb(tlb, pmd);
164 }
165
166 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
167                                 unsigned long addr, unsigned long end,
168                                 unsigned long floor, unsigned long ceiling)
169 {
170         pud_t *pud;
171         unsigned long next;
172         unsigned long start;
173
174         start = addr;
175         pud = pud_offset(pgd, addr);
176         do {
177                 next = pud_addr_end(addr, end);
178                 if (pud_none_or_clear_bad(pud))
179                         continue;
180                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
181         } while (pud++, addr = next, addr != end);
182
183         start &= PGDIR_MASK;
184         if (start < floor)
185                 return;
186         if (ceiling) {
187                 ceiling &= PGDIR_MASK;
188                 if (!ceiling)
189                         return;
190         }
191         if (end - 1 > ceiling - 1)
192                 return;
193
194         pud = pud_offset(pgd, start);
195         pgd_clear(pgd);
196         pud_free_tlb(tlb, pud);
197 }
198
199 /*
200  * This function frees user-level page tables of a process.
201  *
202  * Must be called with pagetable lock held.
203  */
204 void free_pgd_range(struct mmu_gather **tlb,
205                         unsigned long addr, unsigned long end,
206                         unsigned long floor, unsigned long ceiling)
207 {
208         pgd_t *pgd;
209         unsigned long next;
210         unsigned long start;
211
212         /*
213          * The next few lines have given us lots of grief...
214          *
215          * Why are we testing PMD* at this top level?  Because often
216          * there will be no work to do at all, and we'd prefer not to
217          * go all the way down to the bottom just to discover that.
218          *
219          * Why all these "- 1"s?  Because 0 represents both the bottom
220          * of the address space and the top of it (using -1 for the
221          * top wouldn't help much: the masks would do the wrong thing).
222          * The rule is that addr 0 and floor 0 refer to the bottom of
223          * the address space, but end 0 and ceiling 0 refer to the top
224          * Comparisons need to use "end - 1" and "ceiling - 1" (though
225          * that end 0 case should be mythical).
226          *
227          * Wherever addr is brought up or ceiling brought down, we must
228          * be careful to reject "the opposite 0" before it confuses the
229          * subsequent tests.  But what about where end is brought down
230          * by PMD_SIZE below? no, end can't go down to 0 there.
231          *
232          * Whereas we round start (addr) and ceiling down, by different
233          * masks at different levels, in order to test whether a table
234          * now has no other vmas using it, so can be freed, we don't
235          * bother to round floor or end up - the tests don't need that.
236          */
237
238         addr &= PMD_MASK;
239         if (addr < floor) {
240                 addr += PMD_SIZE;
241                 if (!addr)
242                         return;
243         }
244         if (ceiling) {
245                 ceiling &= PMD_MASK;
246                 if (!ceiling)
247                         return;
248         }
249         if (end - 1 > ceiling - 1)
250                 end -= PMD_SIZE;
251         if (addr > end - 1)
252                 return;
253
254         start = addr;
255         pgd = pgd_offset((*tlb)->mm, addr);
256         do {
257                 next = pgd_addr_end(addr, end);
258                 if (pgd_none_or_clear_bad(pgd))
259                         continue;
260                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
261         } while (pgd++, addr = next, addr != end);
262
263         if (!(*tlb)->fullmm)
264                 flush_tlb_pgtables((*tlb)->mm, start, end);
265 }
266
267 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
268                 unsigned long floor, unsigned long ceiling)
269 {
270         while (vma) {
271                 struct vm_area_struct *next = vma->vm_next;
272                 unsigned long addr = vma->vm_start;
273
274                 /*
275                  * Hide vma from rmap and vmtruncate before freeing pgtables
276                  */
277                 anon_vma_unlink(vma);
278                 unlink_file_vma(vma);
279
280                 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
281                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
282                                 floor, next? next->vm_start: ceiling);
283                 } else {
284                         /*
285                          * Optimization: gather nearby vmas into one call down
286                          */
287                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
288                           && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
289                                                         HPAGE_SIZE)) {
290                                 vma = next;
291                                 next = vma->vm_next;
292                                 anon_vma_unlink(vma);
293                                 unlink_file_vma(vma);
294                         }
295                         free_pgd_range(tlb, addr, vma->vm_end,
296                                 floor, next? next->vm_start: ceiling);
297                 }
298                 vma = next;
299         }
300 }
301
302 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
303 {
304         struct page *new = pte_alloc_one(mm, address);
305         if (!new)
306                 return -ENOMEM;
307
308         pte_lock_init(new);
309         spin_lock(&mm->page_table_lock);
310         if (pmd_present(*pmd)) {        /* Another has populated it */
311                 pte_lock_deinit(new);
312                 pte_free(new);
313         } else {
314                 mm->nr_ptes++;
315                 inc_page_state(nr_page_table_pages);
316                 pmd_populate(mm, pmd, new);
317         }
318         spin_unlock(&mm->page_table_lock);
319         return 0;
320 }
321
322 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
323 {
324         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
325         if (!new)
326                 return -ENOMEM;
327
328         spin_lock(&init_mm.page_table_lock);
329         if (pmd_present(*pmd))          /* Another has populated it */
330                 pte_free_kernel(new);
331         else
332                 pmd_populate_kernel(&init_mm, pmd, new);
333         spin_unlock(&init_mm.page_table_lock);
334         return 0;
335 }
336
337 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
338 {
339         if (file_rss)
340                 add_mm_counter(mm, file_rss, file_rss);
341         if (anon_rss)
342                 add_mm_counter(mm, anon_rss, anon_rss);
343 }
344
345 /*
346  * This function is called to print an error when a bad pte
347  * is found. For example, we might have a PFN-mapped pte in
348  * a region that doesn't allow it.
349  *
350  * The calling function must still handle the error.
351  */
352 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
353 {
354         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
355                         "vm_flags = %lx, vaddr = %lx\n",
356                 (long long)pte_val(pte),
357                 (vma->vm_mm == current->mm ? current->comm : "???"),
358                 vma->vm_flags, vaddr);
359         dump_stack();
360 }
361
362 static inline int is_cow_mapping(unsigned int flags)
363 {
364         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
365 }
366
367 /*
368  * This function gets the "struct page" associated with a pte.
369  *
370  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
371  * will have each page table entry just pointing to a raw page frame
372  * number, and as far as the VM layer is concerned, those do not have
373  * pages associated with them - even if the PFN might point to memory
374  * that otherwise is perfectly fine and has a "struct page".
375  *
376  * The way we recognize those mappings is through the rules set up
377  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
378  * and the vm_pgoff will point to the first PFN mapped: thus every
379  * page that is a raw mapping will always honor the rule
380  *
381  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
382  *
383  * and if that isn't true, the page has been COW'ed (in which case it
384  * _does_ have a "struct page" associated with it even if it is in a
385  * VM_PFNMAP range).
386  */
387 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
388 {
389         unsigned long pfn = pte_pfn(pte);
390
391         if (vma->vm_flags & VM_PFNMAP) {
392                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
393                 if (pfn == vma->vm_pgoff + off)
394                         return NULL;
395                 if (!is_cow_mapping(vma->vm_flags))
396                         return NULL;
397         }
398
399         /*
400          * Add some anal sanity checks for now. Eventually,
401          * we should just do "return pfn_to_page(pfn)", but
402          * in the meantime we check that we get a valid pfn,
403          * and that the resulting page looks ok.
404          *
405          * Remove this test eventually!
406          */
407         if (unlikely(!pfn_valid(pfn))) {
408                 print_bad_pte(vma, pte, addr);
409                 return NULL;
410         }
411
412         /*
413          * NOTE! We still have PageReserved() pages in the page 
414          * tables. 
415          *
416          * The PAGE_ZERO() pages and various VDSO mappings can
417          * cause them to exist.
418          */
419         return pfn_to_page(pfn);
420 }
421
422 /*
423  * copy one vm_area from one task to the other. Assumes the page tables
424  * already present in the new task to be cleared in the whole range
425  * covered by this vma.
426  */
427
428 static inline void
429 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
430                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
431                 unsigned long addr, int *rss)
432 {
433         unsigned long vm_flags = vma->vm_flags;
434         pte_t pte = *src_pte;
435         struct page *page;
436
437         /* pte contains position in swap or file, so copy. */
438         if (unlikely(!pte_present(pte))) {
439                 if (!pte_file(pte)) {
440                         swap_duplicate(pte_to_swp_entry(pte));
441                         /* make sure dst_mm is on swapoff's mmlist. */
442                         if (unlikely(list_empty(&dst_mm->mmlist))) {
443                                 spin_lock(&mmlist_lock);
444                                 if (list_empty(&dst_mm->mmlist))
445                                         list_add(&dst_mm->mmlist,
446                                                  &src_mm->mmlist);
447                                 spin_unlock(&mmlist_lock);
448                         }
449                 }
450                 goto out_set_pte;
451         }
452
453         /*
454          * If it's a COW mapping, write protect it both
455          * in the parent and the child
456          */
457         if (is_cow_mapping(vm_flags)) {
458                 ptep_set_wrprotect(src_mm, addr, src_pte);
459                 pte = *src_pte;
460         }
461
462         /*
463          * If it's a shared mapping, mark it clean in
464          * the child
465          */
466         if (vm_flags & VM_SHARED)
467                 pte = pte_mkclean(pte);
468         pte = pte_mkold(pte);
469
470         page = vm_normal_page(vma, addr, pte);
471         if (page) {
472                 get_page(page);
473                 page_dup_rmap(page);
474                 rss[!!PageAnon(page)]++;
475         }
476
477 out_set_pte:
478         set_pte_at(dst_mm, addr, dst_pte, pte);
479 }
480
481 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
482                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
483                 unsigned long addr, unsigned long end)
484 {
485         pte_t *src_pte, *dst_pte;
486         spinlock_t *src_ptl, *dst_ptl;
487         int progress = 0;
488         int rss[2];
489
490 again:
491         rss[1] = rss[0] = 0;
492         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
493         if (!dst_pte)
494                 return -ENOMEM;
495         src_pte = pte_offset_map_nested(src_pmd, addr);
496         src_ptl = pte_lockptr(src_mm, src_pmd);
497         spin_lock(src_ptl);
498
499         do {
500                 /*
501                  * We are holding two locks at this point - either of them
502                  * could generate latencies in another task on another CPU.
503                  */
504                 if (progress >= 32) {
505                         progress = 0;
506                         if (need_resched() ||
507                             need_lockbreak(src_ptl) ||
508                             need_lockbreak(dst_ptl))
509                                 break;
510                 }
511                 if (pte_none(*src_pte)) {
512                         progress++;
513                         continue;
514                 }
515                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
516                 progress += 8;
517         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
518
519         spin_unlock(src_ptl);
520         pte_unmap_nested(src_pte - 1);
521         add_mm_rss(dst_mm, rss[0], rss[1]);
522         pte_unmap_unlock(dst_pte - 1, dst_ptl);
523         cond_resched();
524         if (addr != end)
525                 goto again;
526         return 0;
527 }
528
529 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
530                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
531                 unsigned long addr, unsigned long end)
532 {
533         pmd_t *src_pmd, *dst_pmd;
534         unsigned long next;
535
536         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
537         if (!dst_pmd)
538                 return -ENOMEM;
539         src_pmd = pmd_offset(src_pud, addr);
540         do {
541                 next = pmd_addr_end(addr, end);
542                 if (pmd_none_or_clear_bad(src_pmd))
543                         continue;
544                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
545                                                 vma, addr, next))
546                         return -ENOMEM;
547         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
548         return 0;
549 }
550
551 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
552                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
553                 unsigned long addr, unsigned long end)
554 {
555         pud_t *src_pud, *dst_pud;
556         unsigned long next;
557
558         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
559         if (!dst_pud)
560                 return -ENOMEM;
561         src_pud = pud_offset(src_pgd, addr);
562         do {
563                 next = pud_addr_end(addr, end);
564                 if (pud_none_or_clear_bad(src_pud))
565                         continue;
566                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
567                                                 vma, addr, next))
568                         return -ENOMEM;
569         } while (dst_pud++, src_pud++, addr = next, addr != end);
570         return 0;
571 }
572
573 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
574                 struct vm_area_struct *vma)
575 {
576         pgd_t *src_pgd, *dst_pgd;
577         unsigned long next;
578         unsigned long addr = vma->vm_start;
579         unsigned long end = vma->vm_end;
580
581         /*
582          * Don't copy ptes where a page fault will fill them correctly.
583          * Fork becomes much lighter when there are big shared or private
584          * readonly mappings. The tradeoff is that copy_page_range is more
585          * efficient than faulting.
586          */
587         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
588                 if (!vma->anon_vma)
589                         return 0;
590         }
591
592         if (is_vm_hugetlb_page(vma))
593                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
594
595         dst_pgd = pgd_offset(dst_mm, addr);
596         src_pgd = pgd_offset(src_mm, addr);
597         do {
598                 next = pgd_addr_end(addr, end);
599                 if (pgd_none_or_clear_bad(src_pgd))
600                         continue;
601                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
602                                                 vma, addr, next))
603                         return -ENOMEM;
604         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
605         return 0;
606 }
607
608 static unsigned long zap_pte_range(struct mmu_gather *tlb,
609                                 struct vm_area_struct *vma, pmd_t *pmd,
610                                 unsigned long addr, unsigned long end,
611                                 long *zap_work, struct zap_details *details)
612 {
613         struct mm_struct *mm = tlb->mm;
614         pte_t *pte;
615         spinlock_t *ptl;
616         int file_rss = 0;
617         int anon_rss = 0;
618
619         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
620         do {
621                 pte_t ptent = *pte;
622                 if (pte_none(ptent)) {
623                         (*zap_work)--;
624                         continue;
625                 }
626                 if (pte_present(ptent)) {
627                         struct page *page;
628
629                         (*zap_work) -= PAGE_SIZE;
630
631                         page = vm_normal_page(vma, addr, ptent);
632                         if (unlikely(details) && page) {
633                                 /*
634                                  * unmap_shared_mapping_pages() wants to
635                                  * invalidate cache without truncating:
636                                  * unmap shared but keep private pages.
637                                  */
638                                 if (details->check_mapping &&
639                                     details->check_mapping != page->mapping)
640                                         continue;
641                                 /*
642                                  * Each page->index must be checked when
643                                  * invalidating or truncating nonlinear.
644                                  */
645                                 if (details->nonlinear_vma &&
646                                     (page->index < details->first_index ||
647                                      page->index > details->last_index))
648                                         continue;
649                         }
650                         ptent = ptep_get_and_clear_full(mm, addr, pte,
651                                                         tlb->fullmm);
652                         tlb_remove_tlb_entry(tlb, pte, addr);
653                         if (unlikely(!page))
654                                 continue;
655                         if (unlikely(details) && details->nonlinear_vma
656                             && linear_page_index(details->nonlinear_vma,
657                                                 addr) != page->index)
658                                 set_pte_at(mm, addr, pte,
659                                            pgoff_to_pte(page->index));
660                         if (PageAnon(page))
661                                 anon_rss--;
662                         else {
663                                 if (pte_dirty(ptent))
664                                         set_page_dirty(page);
665                                 if (pte_young(ptent))
666                                         mark_page_accessed(page);
667                                 file_rss--;
668                         }
669                         page_remove_rmap(page);
670                         tlb_remove_page(tlb, page);
671                         continue;
672                 }
673                 /*
674                  * If details->check_mapping, we leave swap entries;
675                  * if details->nonlinear_vma, we leave file entries.
676                  */
677                 if (unlikely(details))
678                         continue;
679                 if (!pte_file(ptent))
680                         free_swap_and_cache(pte_to_swp_entry(ptent));
681                 pte_clear_full(mm, addr, pte, tlb->fullmm);
682         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
683
684         add_mm_rss(mm, file_rss, anon_rss);
685         pte_unmap_unlock(pte - 1, ptl);
686
687         return addr;
688 }
689
690 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
691                                 struct vm_area_struct *vma, pud_t *pud,
692                                 unsigned long addr, unsigned long end,
693                                 long *zap_work, struct zap_details *details)
694 {
695         pmd_t *pmd;
696         unsigned long next;
697
698         pmd = pmd_offset(pud, addr);
699         do {
700                 next = pmd_addr_end(addr, end);
701                 if (pmd_none_or_clear_bad(pmd)) {
702                         (*zap_work)--;
703                         continue;
704                 }
705                 next = zap_pte_range(tlb, vma, pmd, addr, next,
706                                                 zap_work, details);
707         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
708
709         return addr;
710 }
711
712 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
713                                 struct vm_area_struct *vma, pgd_t *pgd,
714                                 unsigned long addr, unsigned long end,
715                                 long *zap_work, struct zap_details *details)
716 {
717         pud_t *pud;
718         unsigned long next;
719
720         pud = pud_offset(pgd, addr);
721         do {
722                 next = pud_addr_end(addr, end);
723                 if (pud_none_or_clear_bad(pud)) {
724                         (*zap_work)--;
725                         continue;
726                 }
727                 next = zap_pmd_range(tlb, vma, pud, addr, next,
728                                                 zap_work, details);
729         } while (pud++, addr = next, (addr != end && *zap_work > 0));
730
731         return addr;
732 }
733
734 static unsigned long unmap_page_range(struct mmu_gather *tlb,
735                                 struct vm_area_struct *vma,
736                                 unsigned long addr, unsigned long end,
737                                 long *zap_work, struct zap_details *details)
738 {
739         pgd_t *pgd;
740         unsigned long next;
741
742         if (details && !details->check_mapping && !details->nonlinear_vma)
743                 details = NULL;
744
745         BUG_ON(addr >= end);
746         tlb_start_vma(tlb, vma);
747         pgd = pgd_offset(vma->vm_mm, addr);
748         do {
749                 next = pgd_addr_end(addr, end);
750                 if (pgd_none_or_clear_bad(pgd)) {
751                         (*zap_work)--;
752                         continue;
753                 }
754                 next = zap_pud_range(tlb, vma, pgd, addr, next,
755                                                 zap_work, details);
756         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
757         tlb_end_vma(tlb, vma);
758
759         return addr;
760 }
761
762 #ifdef CONFIG_PREEMPT
763 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
764 #else
765 /* No preempt: go for improved straight-line efficiency */
766 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
767 #endif
768
769 /**
770  * unmap_vmas - unmap a range of memory covered by a list of vma's
771  * @tlbp: address of the caller's struct mmu_gather
772  * @vma: the starting vma
773  * @start_addr: virtual address at which to start unmapping
774  * @end_addr: virtual address at which to end unmapping
775  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
776  * @details: details of nonlinear truncation or shared cache invalidation
777  *
778  * Returns the end address of the unmapping (restart addr if interrupted).
779  *
780  * Unmap all pages in the vma list.
781  *
782  * We aim to not hold locks for too long (for scheduling latency reasons).
783  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
784  * return the ending mmu_gather to the caller.
785  *
786  * Only addresses between `start' and `end' will be unmapped.
787  *
788  * The VMA list must be sorted in ascending virtual address order.
789  *
790  * unmap_vmas() assumes that the caller will flush the whole unmapped address
791  * range after unmap_vmas() returns.  So the only responsibility here is to
792  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
793  * drops the lock and schedules.
794  */
795 unsigned long unmap_vmas(struct mmu_gather **tlbp,
796                 struct vm_area_struct *vma, unsigned long start_addr,
797                 unsigned long end_addr, unsigned long *nr_accounted,
798                 struct zap_details *details)
799 {
800         long zap_work = ZAP_BLOCK_SIZE;
801         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
802         int tlb_start_valid = 0;
803         unsigned long start = start_addr;
804         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
805         int fullmm = (*tlbp)->fullmm;
806
807         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
808                 unsigned long end;
809
810                 start = max(vma->vm_start, start_addr);
811                 if (start >= vma->vm_end)
812                         continue;
813                 end = min(vma->vm_end, end_addr);
814                 if (end <= vma->vm_start)
815                         continue;
816
817                 if (vma->vm_flags & VM_ACCOUNT)
818                         *nr_accounted += (end - start) >> PAGE_SHIFT;
819
820                 while (start != end) {
821                         if (!tlb_start_valid) {
822                                 tlb_start = start;
823                                 tlb_start_valid = 1;
824                         }
825
826                         if (unlikely(is_vm_hugetlb_page(vma))) {
827                                 unmap_hugepage_range(vma, start, end);
828                                 zap_work -= (end - start) /
829                                                 (HPAGE_SIZE / PAGE_SIZE);
830                                 start = end;
831                         } else
832                                 start = unmap_page_range(*tlbp, vma,
833                                                 start, end, &zap_work, details);
834
835                         if (zap_work > 0) {
836                                 BUG_ON(start != end);
837                                 break;
838                         }
839
840                         tlb_finish_mmu(*tlbp, tlb_start, start);
841
842                         if (need_resched() ||
843                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
844                                 if (i_mmap_lock) {
845                                         *tlbp = NULL;
846                                         goto out;
847                                 }
848                                 cond_resched();
849                         }
850
851                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
852                         tlb_start_valid = 0;
853                         zap_work = ZAP_BLOCK_SIZE;
854                 }
855         }
856 out:
857         return start;   /* which is now the end (or restart) address */
858 }
859
860 /**
861  * zap_page_range - remove user pages in a given range
862  * @vma: vm_area_struct holding the applicable pages
863  * @address: starting address of pages to zap
864  * @size: number of bytes to zap
865  * @details: details of nonlinear truncation or shared cache invalidation
866  */
867 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
868                 unsigned long size, struct zap_details *details)
869 {
870         struct mm_struct *mm = vma->vm_mm;
871         struct mmu_gather *tlb;
872         unsigned long end = address + size;
873         unsigned long nr_accounted = 0;
874
875         lru_add_drain();
876         tlb = tlb_gather_mmu(mm, 0);
877         update_hiwater_rss(mm);
878         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
879         if (tlb)
880                 tlb_finish_mmu(tlb, address, end);
881         return end;
882 }
883
884 /*
885  * Do a quick page-table lookup for a single page.
886  */
887 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
888                         unsigned int flags)
889 {
890         pgd_t *pgd;
891         pud_t *pud;
892         pmd_t *pmd;
893         pte_t *ptep, pte;
894         spinlock_t *ptl;
895         struct page *page;
896         struct mm_struct *mm = vma->vm_mm;
897
898         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
899         if (!IS_ERR(page)) {
900                 BUG_ON(flags & FOLL_GET);
901                 goto out;
902         }
903
904         page = NULL;
905         pgd = pgd_offset(mm, address);
906         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
907                 goto no_page_table;
908
909         pud = pud_offset(pgd, address);
910         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
911                 goto no_page_table;
912         
913         pmd = pmd_offset(pud, address);
914         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
915                 goto no_page_table;
916
917         if (pmd_huge(*pmd)) {
918                 BUG_ON(flags & FOLL_GET);
919                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
920                 goto out;
921         }
922
923         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
924         if (!ptep)
925                 goto out;
926
927         pte = *ptep;
928         if (!pte_present(pte))
929                 goto unlock;
930         if ((flags & FOLL_WRITE) && !pte_write(pte))
931                 goto unlock;
932         page = vm_normal_page(vma, address, pte);
933         if (unlikely(!page))
934                 goto unlock;
935
936         if (flags & FOLL_GET)
937                 get_page(page);
938         if (flags & FOLL_TOUCH) {
939                 if ((flags & FOLL_WRITE) &&
940                     !pte_dirty(pte) && !PageDirty(page))
941                         set_page_dirty(page);
942                 mark_page_accessed(page);
943         }
944 unlock:
945         pte_unmap_unlock(ptep, ptl);
946 out:
947         return page;
948
949 no_page_table:
950         /*
951          * When core dumping an enormous anonymous area that nobody
952          * has touched so far, we don't want to allocate page tables.
953          */
954         if (flags & FOLL_ANON) {
955                 page = ZERO_PAGE(address);
956                 if (flags & FOLL_GET)
957                         get_page(page);
958                 BUG_ON(flags & FOLL_WRITE);
959         }
960         return page;
961 }
962
963 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
964                 unsigned long start, int len, int write, int force,
965                 struct page **pages, struct vm_area_struct **vmas)
966 {
967         int i;
968         unsigned int vm_flags;
969
970         /* 
971          * Require read or write permissions.
972          * If 'force' is set, we only require the "MAY" flags.
973          */
974         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
975         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
976         i = 0;
977
978         do {
979                 struct vm_area_struct *vma;
980                 unsigned int foll_flags;
981
982                 vma = find_extend_vma(mm, start);
983                 if (!vma && in_gate_area(tsk, start)) {
984                         unsigned long pg = start & PAGE_MASK;
985                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
986                         pgd_t *pgd;
987                         pud_t *pud;
988                         pmd_t *pmd;
989                         pte_t *pte;
990                         if (write) /* user gate pages are read-only */
991                                 return i ? : -EFAULT;
992                         if (pg > TASK_SIZE)
993                                 pgd = pgd_offset_k(pg);
994                         else
995                                 pgd = pgd_offset_gate(mm, pg);
996                         BUG_ON(pgd_none(*pgd));
997                         pud = pud_offset(pgd, pg);
998                         BUG_ON(pud_none(*pud));
999                         pmd = pmd_offset(pud, pg);
1000                         if (pmd_none(*pmd))
1001                                 return i ? : -EFAULT;
1002                         pte = pte_offset_map(pmd, pg);
1003                         if (pte_none(*pte)) {
1004                                 pte_unmap(pte);
1005                                 return i ? : -EFAULT;
1006                         }
1007                         if (pages) {
1008                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1009                                 pages[i] = page;
1010                                 if (page)
1011                                         get_page(page);
1012                         }
1013                         pte_unmap(pte);
1014                         if (vmas)
1015                                 vmas[i] = gate_vma;
1016                         i++;
1017                         start += PAGE_SIZE;
1018                         len--;
1019                         continue;
1020                 }
1021
1022                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1023                                 || !(vm_flags & vma->vm_flags))
1024                         return i ? : -EFAULT;
1025
1026                 if (is_vm_hugetlb_page(vma)) {
1027                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1028                                                 &start, &len, i);
1029                         continue;
1030                 }
1031
1032                 foll_flags = FOLL_TOUCH;
1033                 if (pages)
1034                         foll_flags |= FOLL_GET;
1035                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1036                     (!vma->vm_ops || !vma->vm_ops->nopage))
1037                         foll_flags |= FOLL_ANON;
1038
1039                 do {
1040                         struct page *page;
1041
1042                         if (write)
1043                                 foll_flags |= FOLL_WRITE;
1044
1045                         cond_resched();
1046                         while (!(page = follow_page(vma, start, foll_flags))) {
1047                                 int ret;
1048                                 ret = __handle_mm_fault(mm, vma, start,
1049                                                 foll_flags & FOLL_WRITE);
1050                                 /*
1051                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1052                                  * broken COW when necessary, even if maybe_mkwrite
1053                                  * decided not to set pte_write. We can thus safely do
1054                                  * subsequent page lookups as if they were reads.
1055                                  */
1056                                 if (ret & VM_FAULT_WRITE)
1057                                         foll_flags &= ~FOLL_WRITE;
1058                                 
1059                                 switch (ret & ~VM_FAULT_WRITE) {
1060                                 case VM_FAULT_MINOR:
1061                                         tsk->min_flt++;
1062                                         break;
1063                                 case VM_FAULT_MAJOR:
1064                                         tsk->maj_flt++;
1065                                         break;
1066                                 case VM_FAULT_SIGBUS:
1067                                         return i ? i : -EFAULT;
1068                                 case VM_FAULT_OOM:
1069                                         return i ? i : -ENOMEM;
1070                                 default:
1071                                         BUG();
1072                                 }
1073                         }
1074                         if (pages) {
1075                                 pages[i] = page;
1076                                 flush_dcache_page(page);
1077                         }
1078                         if (vmas)
1079                                 vmas[i] = vma;
1080                         i++;
1081                         start += PAGE_SIZE;
1082                         len--;
1083                 } while (len && start < vma->vm_end);
1084         } while (len);
1085         return i;
1086 }
1087 EXPORT_SYMBOL(get_user_pages);
1088
1089 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1090                         unsigned long addr, unsigned long end, pgprot_t prot)
1091 {
1092         pte_t *pte;
1093         spinlock_t *ptl;
1094
1095         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1096         if (!pte)
1097                 return -ENOMEM;
1098         do {
1099                 struct page *page = ZERO_PAGE(addr);
1100                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1101                 page_cache_get(page);
1102                 page_add_file_rmap(page);
1103                 inc_mm_counter(mm, file_rss);
1104                 BUG_ON(!pte_none(*pte));
1105                 set_pte_at(mm, addr, pte, zero_pte);
1106         } while (pte++, addr += PAGE_SIZE, addr != end);
1107         pte_unmap_unlock(pte - 1, ptl);
1108         return 0;
1109 }
1110
1111 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1112                         unsigned long addr, unsigned long end, pgprot_t prot)
1113 {
1114         pmd_t *pmd;
1115         unsigned long next;
1116
1117         pmd = pmd_alloc(mm, pud, addr);
1118         if (!pmd)
1119                 return -ENOMEM;
1120         do {
1121                 next = pmd_addr_end(addr, end);
1122                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1123                         return -ENOMEM;
1124         } while (pmd++, addr = next, addr != end);
1125         return 0;
1126 }
1127
1128 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1129                         unsigned long addr, unsigned long end, pgprot_t prot)
1130 {
1131         pud_t *pud;
1132         unsigned long next;
1133
1134         pud = pud_alloc(mm, pgd, addr);
1135         if (!pud)
1136                 return -ENOMEM;
1137         do {
1138                 next = pud_addr_end(addr, end);
1139                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1140                         return -ENOMEM;
1141         } while (pud++, addr = next, addr != end);
1142         return 0;
1143 }
1144
1145 int zeromap_page_range(struct vm_area_struct *vma,
1146                         unsigned long addr, unsigned long size, pgprot_t prot)
1147 {
1148         pgd_t *pgd;
1149         unsigned long next;
1150         unsigned long end = addr + size;
1151         struct mm_struct *mm = vma->vm_mm;
1152         int err;
1153
1154         BUG_ON(addr >= end);
1155         pgd = pgd_offset(mm, addr);
1156         flush_cache_range(vma, addr, end);
1157         do {
1158                 next = pgd_addr_end(addr, end);
1159                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1160                 if (err)
1161                         break;
1162         } while (pgd++, addr = next, addr != end);
1163         return err;
1164 }
1165
1166 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1167 {
1168         pgd_t * pgd = pgd_offset(mm, addr);
1169         pud_t * pud = pud_alloc(mm, pgd, addr);
1170         if (pud) {
1171                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1172                 if (pmd)
1173                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1174         }
1175         return NULL;
1176 }
1177
1178 /*
1179  * This is the old fallback for page remapping.
1180  *
1181  * For historical reasons, it only allows reserved pages. Only
1182  * old drivers should use this, and they needed to mark their
1183  * pages reserved for the old functions anyway.
1184  */
1185 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1186 {
1187         int retval;
1188         pte_t *pte;
1189         spinlock_t *ptl;  
1190
1191         retval = -EINVAL;
1192         if (PageAnon(page))
1193                 goto out;
1194         retval = -ENOMEM;
1195         flush_dcache_page(page);
1196         pte = get_locked_pte(mm, addr, &ptl);
1197         if (!pte)
1198                 goto out;
1199         retval = -EBUSY;
1200         if (!pte_none(*pte))
1201                 goto out_unlock;
1202
1203         /* Ok, finally just insert the thing.. */
1204         get_page(page);
1205         inc_mm_counter(mm, file_rss);
1206         page_add_file_rmap(page);
1207         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1208
1209         retval = 0;
1210 out_unlock:
1211         pte_unmap_unlock(pte, ptl);
1212 out:
1213         return retval;
1214 }
1215
1216 /*
1217  * This allows drivers to insert individual pages they've allocated
1218  * into a user vma.
1219  *
1220  * The page has to be a nice clean _individual_ kernel allocation.
1221  * If you allocate a compound page, you need to have marked it as
1222  * such (__GFP_COMP), or manually just split the page up yourself
1223  * (which is mainly an issue of doing "set_page_count(page, 1)" for
1224  * each sub-page, and then freeing them one by one when you free
1225  * them rather than freeing it as a compound page).
1226  *
1227  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1228  * took an arbitrary page protection parameter. This doesn't allow
1229  * that. Your vma protection will have to be set up correctly, which
1230  * means that if you want a shared writable mapping, you'd better
1231  * ask for a shared writable mapping!
1232  *
1233  * The page does not need to be reserved.
1234  */
1235 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1236 {
1237         if (addr < vma->vm_start || addr >= vma->vm_end)
1238                 return -EFAULT;
1239         if (!page_count(page))
1240                 return -EINVAL;
1241         vma->vm_flags |= VM_INSERTPAGE;
1242         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1243 }
1244 EXPORT_SYMBOL(vm_insert_page);
1245
1246 /*
1247  * maps a range of physical memory into the requested pages. the old
1248  * mappings are removed. any references to nonexistent pages results
1249  * in null mappings (currently treated as "copy-on-access")
1250  */
1251 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1252                         unsigned long addr, unsigned long end,
1253                         unsigned long pfn, pgprot_t prot)
1254 {
1255         pte_t *pte;
1256         spinlock_t *ptl;
1257
1258         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1259         if (!pte)
1260                 return -ENOMEM;
1261         do {
1262                 BUG_ON(!pte_none(*pte));
1263                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1264                 pfn++;
1265         } while (pte++, addr += PAGE_SIZE, addr != end);
1266         pte_unmap_unlock(pte - 1, ptl);
1267         return 0;
1268 }
1269
1270 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1271                         unsigned long addr, unsigned long end,
1272                         unsigned long pfn, pgprot_t prot)
1273 {
1274         pmd_t *pmd;
1275         unsigned long next;
1276
1277         pfn -= addr >> PAGE_SHIFT;
1278         pmd = pmd_alloc(mm, pud, addr);
1279         if (!pmd)
1280                 return -ENOMEM;
1281         do {
1282                 next = pmd_addr_end(addr, end);
1283                 if (remap_pte_range(mm, pmd, addr, next,
1284                                 pfn + (addr >> PAGE_SHIFT), prot))
1285                         return -ENOMEM;
1286         } while (pmd++, addr = next, addr != end);
1287         return 0;
1288 }
1289
1290 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1291                         unsigned long addr, unsigned long end,
1292                         unsigned long pfn, pgprot_t prot)
1293 {
1294         pud_t *pud;
1295         unsigned long next;
1296
1297         pfn -= addr >> PAGE_SHIFT;
1298         pud = pud_alloc(mm, pgd, addr);
1299         if (!pud)
1300                 return -ENOMEM;
1301         do {
1302                 next = pud_addr_end(addr, end);
1303                 if (remap_pmd_range(mm, pud, addr, next,
1304                                 pfn + (addr >> PAGE_SHIFT), prot))
1305                         return -ENOMEM;
1306         } while (pud++, addr = next, addr != end);
1307         return 0;
1308 }
1309
1310 /*  Note: this is only safe if the mm semaphore is held when called. */
1311 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1312                     unsigned long pfn, unsigned long size, pgprot_t prot)
1313 {
1314         pgd_t *pgd;
1315         unsigned long next;
1316         unsigned long end = addr + PAGE_ALIGN(size);
1317         struct mm_struct *mm = vma->vm_mm;
1318         int err;
1319
1320         /*
1321          * Physically remapped pages are special. Tell the
1322          * rest of the world about it:
1323          *   VM_IO tells people not to look at these pages
1324          *      (accesses can have side effects).
1325          *   VM_RESERVED is specified all over the place, because
1326          *      in 2.4 it kept swapout's vma scan off this vma; but
1327          *      in 2.6 the LRU scan won't even find its pages, so this
1328          *      flag means no more than count its pages in reserved_vm,
1329          *      and omit it from core dump, even when VM_IO turned off.
1330          *   VM_PFNMAP tells the core MM that the base pages are just
1331          *      raw PFN mappings, and do not have a "struct page" associated
1332          *      with them.
1333          *
1334          * There's a horrible special case to handle copy-on-write
1335          * behaviour that some programs depend on. We mark the "original"
1336          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1337          */
1338         if (is_cow_mapping(vma->vm_flags)) {
1339                 if (addr != vma->vm_start || end != vma->vm_end)
1340                         return -EINVAL;
1341                 vma->vm_pgoff = pfn;
1342         }
1343
1344         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1345
1346         BUG_ON(addr >= end);
1347         pfn -= addr >> PAGE_SHIFT;
1348         pgd = pgd_offset(mm, addr);
1349         flush_cache_range(vma, addr, end);
1350         do {
1351                 next = pgd_addr_end(addr, end);
1352                 err = remap_pud_range(mm, pgd, addr, next,
1353                                 pfn + (addr >> PAGE_SHIFT), prot);
1354                 if (err)
1355                         break;
1356         } while (pgd++, addr = next, addr != end);
1357         return err;
1358 }
1359 EXPORT_SYMBOL(remap_pfn_range);
1360
1361 /*
1362  * handle_pte_fault chooses page fault handler according to an entry
1363  * which was read non-atomically.  Before making any commitment, on
1364  * those architectures or configurations (e.g. i386 with PAE) which
1365  * might give a mix of unmatched parts, do_swap_page and do_file_page
1366  * must check under lock before unmapping the pte and proceeding
1367  * (but do_wp_page is only called after already making such a check;
1368  * and do_anonymous_page and do_no_page can safely check later on).
1369  */
1370 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1371                                 pte_t *page_table, pte_t orig_pte)
1372 {
1373         int same = 1;
1374 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1375         if (sizeof(pte_t) > sizeof(unsigned long)) {
1376                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1377                 spin_lock(ptl);
1378                 same = pte_same(*page_table, orig_pte);
1379                 spin_unlock(ptl);
1380         }
1381 #endif
1382         pte_unmap(page_table);
1383         return same;
1384 }
1385
1386 /*
1387  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1388  * servicing faults for write access.  In the normal case, do always want
1389  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1390  * that do not have writing enabled, when used by access_process_vm.
1391  */
1392 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1393 {
1394         if (likely(vma->vm_flags & VM_WRITE))
1395                 pte = pte_mkwrite(pte);
1396         return pte;
1397 }
1398
1399 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1400 {
1401         /*
1402          * If the source page was a PFN mapping, we don't have
1403          * a "struct page" for it. We do a best-effort copy by
1404          * just copying from the original user address. If that
1405          * fails, we just zero-fill it. Live with it.
1406          */
1407         if (unlikely(!src)) {
1408                 void *kaddr = kmap_atomic(dst, KM_USER0);
1409                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1410
1411                 /*
1412                  * This really shouldn't fail, because the page is there
1413                  * in the page tables. But it might just be unreadable,
1414                  * in which case we just give up and fill the result with
1415                  * zeroes.
1416                  */
1417                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1418                         memset(kaddr, 0, PAGE_SIZE);
1419                 kunmap_atomic(kaddr, KM_USER0);
1420                 return;
1421                 
1422         }
1423         copy_user_highpage(dst, src, va);
1424 }
1425
1426 /*
1427  * This routine handles present pages, when users try to write
1428  * to a shared page. It is done by copying the page to a new address
1429  * and decrementing the shared-page counter for the old page.
1430  *
1431  * Note that this routine assumes that the protection checks have been
1432  * done by the caller (the low-level page fault routine in most cases).
1433  * Thus we can safely just mark it writable once we've done any necessary
1434  * COW.
1435  *
1436  * We also mark the page dirty at this point even though the page will
1437  * change only once the write actually happens. This avoids a few races,
1438  * and potentially makes it more efficient.
1439  *
1440  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1441  * but allow concurrent faults), with pte both mapped and locked.
1442  * We return with mmap_sem still held, but pte unmapped and unlocked.
1443  */
1444 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1445                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1446                 spinlock_t *ptl, pte_t orig_pte)
1447 {
1448         struct page *old_page, *new_page;
1449         pte_t entry;
1450         int ret = VM_FAULT_MINOR;
1451
1452         old_page = vm_normal_page(vma, address, orig_pte);
1453         if (!old_page)
1454                 goto gotten;
1455
1456         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1457                 int reuse = can_share_swap_page(old_page);
1458                 unlock_page(old_page);
1459                 if (reuse) {
1460                         flush_cache_page(vma, address, pte_pfn(orig_pte));
1461                         entry = pte_mkyoung(orig_pte);
1462                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1463                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1464                         update_mmu_cache(vma, address, entry);
1465                         lazy_mmu_prot_update(entry);
1466                         ret |= VM_FAULT_WRITE;
1467                         goto unlock;
1468                 }
1469         }
1470
1471         /*
1472          * Ok, we need to copy. Oh, well..
1473          */
1474         page_cache_get(old_page);
1475 gotten:
1476         pte_unmap_unlock(page_table, ptl);
1477
1478         if (unlikely(anon_vma_prepare(vma)))
1479                 goto oom;
1480         if (old_page == ZERO_PAGE(address)) {
1481                 new_page = alloc_zeroed_user_highpage(vma, address);
1482                 if (!new_page)
1483                         goto oom;
1484         } else {
1485                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1486                 if (!new_page)
1487                         goto oom;
1488                 cow_user_page(new_page, old_page, address);
1489         }
1490
1491         /*
1492          * Re-check the pte - we dropped the lock
1493          */
1494         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1495         if (likely(pte_same(*page_table, orig_pte))) {
1496                 if (old_page) {
1497                         page_remove_rmap(old_page);
1498                         if (!PageAnon(old_page)) {
1499                                 dec_mm_counter(mm, file_rss);
1500                                 inc_mm_counter(mm, anon_rss);
1501                         }
1502                 } else
1503                         inc_mm_counter(mm, anon_rss);
1504                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1505                 entry = mk_pte(new_page, vma->vm_page_prot);
1506                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1507                 ptep_establish(vma, address, page_table, entry);
1508                 update_mmu_cache(vma, address, entry);
1509                 lazy_mmu_prot_update(entry);
1510                 lru_cache_add_active(new_page);
1511                 page_add_new_anon_rmap(new_page, vma, address);
1512
1513                 /* Free the old page.. */
1514                 new_page = old_page;
1515                 ret |= VM_FAULT_WRITE;
1516         }
1517         if (new_page)
1518                 page_cache_release(new_page);
1519         if (old_page)
1520                 page_cache_release(old_page);
1521 unlock:
1522         pte_unmap_unlock(page_table, ptl);
1523         return ret;
1524 oom:
1525         if (old_page)
1526                 page_cache_release(old_page);
1527         return VM_FAULT_OOM;
1528 }
1529
1530 /*
1531  * Helper functions for unmap_mapping_range().
1532  *
1533  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1534  *
1535  * We have to restart searching the prio_tree whenever we drop the lock,
1536  * since the iterator is only valid while the lock is held, and anyway
1537  * a later vma might be split and reinserted earlier while lock dropped.
1538  *
1539  * The list of nonlinear vmas could be handled more efficiently, using
1540  * a placeholder, but handle it in the same way until a need is shown.
1541  * It is important to search the prio_tree before nonlinear list: a vma
1542  * may become nonlinear and be shifted from prio_tree to nonlinear list
1543  * while the lock is dropped; but never shifted from list to prio_tree.
1544  *
1545  * In order to make forward progress despite restarting the search,
1546  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1547  * quickly skip it next time around.  Since the prio_tree search only
1548  * shows us those vmas affected by unmapping the range in question, we
1549  * can't efficiently keep all vmas in step with mapping->truncate_count:
1550  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1551  * mapping->truncate_count and vma->vm_truncate_count are protected by
1552  * i_mmap_lock.
1553  *
1554  * In order to make forward progress despite repeatedly restarting some
1555  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1556  * and restart from that address when we reach that vma again.  It might
1557  * have been split or merged, shrunk or extended, but never shifted: so
1558  * restart_addr remains valid so long as it remains in the vma's range.
1559  * unmap_mapping_range forces truncate_count to leap over page-aligned
1560  * values so we can save vma's restart_addr in its truncate_count field.
1561  */
1562 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1563
1564 static void reset_vma_truncate_counts(struct address_space *mapping)
1565 {
1566         struct vm_area_struct *vma;
1567         struct prio_tree_iter iter;
1568
1569         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1570                 vma->vm_truncate_count = 0;
1571         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1572                 vma->vm_truncate_count = 0;
1573 }
1574
1575 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1576                 unsigned long start_addr, unsigned long end_addr,
1577                 struct zap_details *details)
1578 {
1579         unsigned long restart_addr;
1580         int need_break;
1581
1582 again:
1583         restart_addr = vma->vm_truncate_count;
1584         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1585                 start_addr = restart_addr;
1586                 if (start_addr >= end_addr) {
1587                         /* Top of vma has been split off since last time */
1588                         vma->vm_truncate_count = details->truncate_count;
1589                         return 0;
1590                 }
1591         }
1592
1593         restart_addr = zap_page_range(vma, start_addr,
1594                                         end_addr - start_addr, details);
1595         need_break = need_resched() ||
1596                         need_lockbreak(details->i_mmap_lock);
1597
1598         if (restart_addr >= end_addr) {
1599                 /* We have now completed this vma: mark it so */
1600                 vma->vm_truncate_count = details->truncate_count;
1601                 if (!need_break)
1602                         return 0;
1603         } else {
1604                 /* Note restart_addr in vma's truncate_count field */
1605                 vma->vm_truncate_count = restart_addr;
1606                 if (!need_break)
1607                         goto again;
1608         }
1609
1610         spin_unlock(details->i_mmap_lock);
1611         cond_resched();
1612         spin_lock(details->i_mmap_lock);
1613         return -EINTR;
1614 }
1615
1616 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1617                                             struct zap_details *details)
1618 {
1619         struct vm_area_struct *vma;
1620         struct prio_tree_iter iter;
1621         pgoff_t vba, vea, zba, zea;
1622
1623 restart:
1624         vma_prio_tree_foreach(vma, &iter, root,
1625                         details->first_index, details->last_index) {
1626                 /* Skip quickly over those we have already dealt with */
1627                 if (vma->vm_truncate_count == details->truncate_count)
1628                         continue;
1629
1630                 vba = vma->vm_pgoff;
1631                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1632                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1633                 zba = details->first_index;
1634                 if (zba < vba)
1635                         zba = vba;
1636                 zea = details->last_index;
1637                 if (zea > vea)
1638                         zea = vea;
1639
1640                 if (unmap_mapping_range_vma(vma,
1641                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1642                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1643                                 details) < 0)
1644                         goto restart;
1645         }
1646 }
1647
1648 static inline void unmap_mapping_range_list(struct list_head *head,
1649                                             struct zap_details *details)
1650 {
1651         struct vm_area_struct *vma;
1652
1653         /*
1654          * In nonlinear VMAs there is no correspondence between virtual address
1655          * offset and file offset.  So we must perform an exhaustive search
1656          * across *all* the pages in each nonlinear VMA, not just the pages
1657          * whose virtual address lies outside the file truncation point.
1658          */
1659 restart:
1660         list_for_each_entry(vma, head, shared.vm_set.list) {
1661                 /* Skip quickly over those we have already dealt with */
1662                 if (vma->vm_truncate_count == details->truncate_count)
1663                         continue;
1664                 details->nonlinear_vma = vma;
1665                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1666                                         vma->vm_end, details) < 0)
1667                         goto restart;
1668         }
1669 }
1670
1671 /**
1672  * unmap_mapping_range - unmap the portion of all mmaps
1673  * in the specified address_space corresponding to the specified
1674  * page range in the underlying file.
1675  * @mapping: the address space containing mmaps to be unmapped.
1676  * @holebegin: byte in first page to unmap, relative to the start of
1677  * the underlying file.  This will be rounded down to a PAGE_SIZE
1678  * boundary.  Note that this is different from vmtruncate(), which
1679  * must keep the partial page.  In contrast, we must get rid of
1680  * partial pages.
1681  * @holelen: size of prospective hole in bytes.  This will be rounded
1682  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1683  * end of the file.
1684  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1685  * but 0 when invalidating pagecache, don't throw away private data.
1686  */
1687 void unmap_mapping_range(struct address_space *mapping,
1688                 loff_t const holebegin, loff_t const holelen, int even_cows)
1689 {
1690         struct zap_details details;
1691         pgoff_t hba = holebegin >> PAGE_SHIFT;
1692         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1693
1694         /* Check for overflow. */
1695         if (sizeof(holelen) > sizeof(hlen)) {
1696                 long long holeend =
1697                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1698                 if (holeend & ~(long long)ULONG_MAX)
1699                         hlen = ULONG_MAX - hba + 1;
1700         }
1701
1702         details.check_mapping = even_cows? NULL: mapping;
1703         details.nonlinear_vma = NULL;
1704         details.first_index = hba;
1705         details.last_index = hba + hlen - 1;
1706         if (details.last_index < details.first_index)
1707                 details.last_index = ULONG_MAX;
1708         details.i_mmap_lock = &mapping->i_mmap_lock;
1709
1710         spin_lock(&mapping->i_mmap_lock);
1711
1712         /* serialize i_size write against truncate_count write */
1713         smp_wmb();
1714         /* Protect against page faults, and endless unmapping loops */
1715         mapping->truncate_count++;
1716         /*
1717          * For archs where spin_lock has inclusive semantics like ia64
1718          * this smp_mb() will prevent to read pagetable contents
1719          * before the truncate_count increment is visible to
1720          * other cpus.
1721          */
1722         smp_mb();
1723         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1724                 if (mapping->truncate_count == 0)
1725                         reset_vma_truncate_counts(mapping);
1726                 mapping->truncate_count++;
1727         }
1728         details.truncate_count = mapping->truncate_count;
1729
1730         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1731                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1732         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1733                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1734         spin_unlock(&mapping->i_mmap_lock);
1735 }
1736 EXPORT_SYMBOL(unmap_mapping_range);
1737
1738 /*
1739  * Handle all mappings that got truncated by a "truncate()"
1740  * system call.
1741  *
1742  * NOTE! We have to be ready to update the memory sharing
1743  * between the file and the memory map for a potential last
1744  * incomplete page.  Ugly, but necessary.
1745  */
1746 int vmtruncate(struct inode * inode, loff_t offset)
1747 {
1748         struct address_space *mapping = inode->i_mapping;
1749         unsigned long limit;
1750
1751         if (inode->i_size < offset)
1752                 goto do_expand;
1753         /*
1754          * truncation of in-use swapfiles is disallowed - it would cause
1755          * subsequent swapout to scribble on the now-freed blocks.
1756          */
1757         if (IS_SWAPFILE(inode))
1758                 goto out_busy;
1759         i_size_write(inode, offset);
1760         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1761         truncate_inode_pages(mapping, offset);
1762         goto out_truncate;
1763
1764 do_expand:
1765         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1766         if (limit != RLIM_INFINITY && offset > limit)
1767                 goto out_sig;
1768         if (offset > inode->i_sb->s_maxbytes)
1769                 goto out_big;
1770         i_size_write(inode, offset);
1771
1772 out_truncate:
1773         if (inode->i_op && inode->i_op->truncate)
1774                 inode->i_op->truncate(inode);
1775         return 0;
1776 out_sig:
1777         send_sig(SIGXFSZ, current, 0);
1778 out_big:
1779         return -EFBIG;
1780 out_busy:
1781         return -ETXTBSY;
1782 }
1783 EXPORT_SYMBOL(vmtruncate);
1784
1785 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1786 {
1787         struct address_space *mapping = inode->i_mapping;
1788
1789         /*
1790          * If the underlying filesystem is not going to provide
1791          * a way to truncate a range of blocks (punch a hole) -
1792          * we should return failure right now.
1793          */
1794         if (!inode->i_op || !inode->i_op->truncate_range)
1795                 return -ENOSYS;
1796
1797         mutex_lock(&inode->i_mutex);
1798         down_write(&inode->i_alloc_sem);
1799         unmap_mapping_range(mapping, offset, (end - offset), 1);
1800         truncate_inode_pages_range(mapping, offset, end);
1801         inode->i_op->truncate_range(inode, offset, end);
1802         up_write(&inode->i_alloc_sem);
1803         mutex_unlock(&inode->i_mutex);
1804
1805         return 0;
1806 }
1807 EXPORT_SYMBOL(vmtruncate_range);
1808
1809 /* 
1810  * Primitive swap readahead code. We simply read an aligned block of
1811  * (1 << page_cluster) entries in the swap area. This method is chosen
1812  * because it doesn't cost us any seek time.  We also make sure to queue
1813  * the 'original' request together with the readahead ones...  
1814  *
1815  * This has been extended to use the NUMA policies from the mm triggering
1816  * the readahead.
1817  *
1818  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1819  */
1820 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1821 {
1822 #ifdef CONFIG_NUMA
1823         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1824 #endif
1825         int i, num;
1826         struct page *new_page;
1827         unsigned long offset;
1828
1829         /*
1830          * Get the number of handles we should do readahead io to.
1831          */
1832         num = valid_swaphandles(entry, &offset);
1833         for (i = 0; i < num; offset++, i++) {
1834                 /* Ok, do the async read-ahead now */
1835                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1836                                                            offset), vma, addr);
1837                 if (!new_page)
1838                         break;
1839                 page_cache_release(new_page);
1840 #ifdef CONFIG_NUMA
1841                 /*
1842                  * Find the next applicable VMA for the NUMA policy.
1843                  */
1844                 addr += PAGE_SIZE;
1845                 if (addr == 0)
1846                         vma = NULL;
1847                 if (vma) {
1848                         if (addr >= vma->vm_end) {
1849                                 vma = next_vma;
1850                                 next_vma = vma ? vma->vm_next : NULL;
1851                         }
1852                         if (vma && addr < vma->vm_start)
1853                                 vma = NULL;
1854                 } else {
1855                         if (next_vma && addr >= next_vma->vm_start) {
1856                                 vma = next_vma;
1857                                 next_vma = vma->vm_next;
1858                         }
1859                 }
1860 #endif
1861         }
1862         lru_add_drain();        /* Push any new pages onto the LRU now */
1863 }
1864
1865 /*
1866  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1867  * but allow concurrent faults), and pte mapped but not yet locked.
1868  * We return with mmap_sem still held, but pte unmapped and unlocked.
1869  */
1870 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1871                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1872                 int write_access, pte_t orig_pte)
1873 {
1874         spinlock_t *ptl;
1875         struct page *page;
1876         swp_entry_t entry;
1877         pte_t pte;
1878         int ret = VM_FAULT_MINOR;
1879
1880         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1881                 goto out;
1882
1883         entry = pte_to_swp_entry(orig_pte);
1884 again:
1885         page = lookup_swap_cache(entry);
1886         if (!page) {
1887                 swapin_readahead(entry, address, vma);
1888                 page = read_swap_cache_async(entry, vma, address);
1889                 if (!page) {
1890                         /*
1891                          * Back out if somebody else faulted in this pte
1892                          * while we released the pte lock.
1893                          */
1894                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1895                         if (likely(pte_same(*page_table, orig_pte)))
1896                                 ret = VM_FAULT_OOM;
1897                         goto unlock;
1898                 }
1899
1900                 /* Had to read the page from swap area: Major fault */
1901                 ret = VM_FAULT_MAJOR;
1902                 inc_page_state(pgmajfault);
1903                 grab_swap_token();
1904         }
1905
1906         mark_page_accessed(page);
1907         lock_page(page);
1908         if (!PageSwapCache(page)) {
1909                 /* Page migration has occured */
1910                 unlock_page(page);
1911                 page_cache_release(page);
1912                 goto again;
1913         }
1914
1915         /*
1916          * Back out if somebody else already faulted in this pte.
1917          */
1918         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1919         if (unlikely(!pte_same(*page_table, orig_pte)))
1920                 goto out_nomap;
1921
1922         if (unlikely(!PageUptodate(page))) {
1923                 ret = VM_FAULT_SIGBUS;
1924                 goto out_nomap;
1925         }
1926
1927         /* The page isn't present yet, go ahead with the fault. */
1928
1929         inc_mm_counter(mm, anon_rss);
1930         pte = mk_pte(page, vma->vm_page_prot);
1931         if (write_access && can_share_swap_page(page)) {
1932                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1933                 write_access = 0;
1934         }
1935
1936         flush_icache_page(vma, page);
1937         set_pte_at(mm, address, page_table, pte);
1938         page_add_anon_rmap(page, vma, address);
1939
1940         swap_free(entry);
1941         if (vm_swap_full())
1942                 remove_exclusive_swap_page(page);
1943         unlock_page(page);
1944
1945         if (write_access) {
1946                 if (do_wp_page(mm, vma, address,
1947                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1948                         ret = VM_FAULT_OOM;
1949                 goto out;
1950         }
1951
1952         /* No need to invalidate - it was non-present before */
1953         update_mmu_cache(vma, address, pte);
1954         lazy_mmu_prot_update(pte);
1955 unlock:
1956         pte_unmap_unlock(page_table, ptl);
1957 out:
1958         return ret;
1959 out_nomap:
1960         pte_unmap_unlock(page_table, ptl);
1961         unlock_page(page);
1962         page_cache_release(page);
1963         return ret;
1964 }
1965
1966 /*
1967  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1968  * but allow concurrent faults), and pte mapped but not yet locked.
1969  * We return with mmap_sem still held, but pte unmapped and unlocked.
1970  */
1971 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1972                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1973                 int write_access)
1974 {
1975         struct page *page;
1976         spinlock_t *ptl;
1977         pte_t entry;
1978
1979         if (write_access) {
1980                 /* Allocate our own private page. */
1981                 pte_unmap(page_table);
1982
1983                 if (unlikely(anon_vma_prepare(vma)))
1984                         goto oom;
1985                 page = alloc_zeroed_user_highpage(vma, address);
1986                 if (!page)
1987                         goto oom;
1988
1989                 entry = mk_pte(page, vma->vm_page_prot);
1990                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1991
1992                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1993                 if (!pte_none(*page_table))
1994                         goto release;
1995                 inc_mm_counter(mm, anon_rss);
1996                 lru_cache_add_active(page);
1997                 page_add_new_anon_rmap(page, vma, address);
1998         } else {
1999                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2000                 page = ZERO_PAGE(address);
2001                 page_cache_get(page);
2002                 entry = mk_pte(page, vma->vm_page_prot);
2003
2004                 ptl = pte_lockptr(mm, pmd);
2005                 spin_lock(ptl);
2006                 if (!pte_none(*page_table))
2007                         goto release;
2008                 inc_mm_counter(mm, file_rss);
2009                 page_add_file_rmap(page);
2010         }
2011
2012         set_pte_at(mm, address, page_table, entry);
2013
2014         /* No need to invalidate - it was non-present before */
2015         update_mmu_cache(vma, address, entry);
2016         lazy_mmu_prot_update(entry);
2017 unlock:
2018         pte_unmap_unlock(page_table, ptl);
2019         return VM_FAULT_MINOR;
2020 release:
2021         page_cache_release(page);
2022         goto unlock;
2023 oom:
2024         return VM_FAULT_OOM;
2025 }
2026
2027 /*
2028  * do_no_page() tries to create a new page mapping. It aggressively
2029  * tries to share with existing pages, but makes a separate copy if
2030  * the "write_access" parameter is true in order to avoid the next
2031  * page fault.
2032  *
2033  * As this is called only for pages that do not currently exist, we
2034  * do not need to flush old virtual caches or the TLB.
2035  *
2036  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2037  * but allow concurrent faults), and pte mapped but not yet locked.
2038  * We return with mmap_sem still held, but pte unmapped and unlocked.
2039  */
2040 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2041                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2042                 int write_access)
2043 {
2044         spinlock_t *ptl;
2045         struct page *new_page;
2046         struct address_space *mapping = NULL;
2047         pte_t entry;
2048         unsigned int sequence = 0;
2049         int ret = VM_FAULT_MINOR;
2050         int anon = 0;
2051
2052         pte_unmap(page_table);
2053         BUG_ON(vma->vm_flags & VM_PFNMAP);
2054
2055         if (vma->vm_file) {
2056                 mapping = vma->vm_file->f_mapping;
2057                 sequence = mapping->truncate_count;
2058                 smp_rmb(); /* serializes i_size against truncate_count */
2059         }
2060 retry:
2061         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2062         /*
2063          * No smp_rmb is needed here as long as there's a full
2064          * spin_lock/unlock sequence inside the ->nopage callback
2065          * (for the pagecache lookup) that acts as an implicit
2066          * smp_mb() and prevents the i_size read to happen
2067          * after the next truncate_count read.
2068          */
2069
2070         /* no page was available -- either SIGBUS or OOM */
2071         if (new_page == NOPAGE_SIGBUS)
2072                 return VM_FAULT_SIGBUS;
2073         if (new_page == NOPAGE_OOM)
2074                 return VM_FAULT_OOM;
2075
2076         /*
2077          * Should we do an early C-O-W break?
2078          */
2079         if (write_access && !(vma->vm_flags & VM_SHARED)) {
2080                 struct page *page;
2081
2082                 if (unlikely(anon_vma_prepare(vma)))
2083                         goto oom;
2084                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2085                 if (!page)
2086                         goto oom;
2087                 copy_user_highpage(page, new_page, address);
2088                 page_cache_release(new_page);
2089                 new_page = page;
2090                 anon = 1;
2091         }
2092
2093         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2094         /*
2095          * For a file-backed vma, someone could have truncated or otherwise
2096          * invalidated this page.  If unmap_mapping_range got called,
2097          * retry getting the page.
2098          */
2099         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2100                 pte_unmap_unlock(page_table, ptl);
2101                 page_cache_release(new_page);
2102                 cond_resched();
2103                 sequence = mapping->truncate_count;
2104                 smp_rmb();
2105                 goto retry;
2106         }
2107
2108         /*
2109          * This silly early PAGE_DIRTY setting removes a race
2110          * due to the bad i386 page protection. But it's valid
2111          * for other architectures too.
2112          *
2113          * Note that if write_access is true, we either now have
2114          * an exclusive copy of the page, or this is a shared mapping,
2115          * so we can make it writable and dirty to avoid having to
2116          * handle that later.
2117          */
2118         /* Only go through if we didn't race with anybody else... */
2119         if (pte_none(*page_table)) {
2120                 flush_icache_page(vma, new_page);
2121                 entry = mk_pte(new_page, vma->vm_page_prot);
2122                 if (write_access)
2123                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2124                 set_pte_at(mm, address, page_table, entry);
2125                 if (anon) {
2126                         inc_mm_counter(mm, anon_rss);
2127                         lru_cache_add_active(new_page);
2128                         page_add_new_anon_rmap(new_page, vma, address);
2129                 } else {
2130                         inc_mm_counter(mm, file_rss);
2131                         page_add_file_rmap(new_page);
2132                 }
2133         } else {
2134                 /* One of our sibling threads was faster, back out. */
2135                 page_cache_release(new_page);
2136                 goto unlock;
2137         }
2138
2139         /* no need to invalidate: a not-present page shouldn't be cached */
2140         update_mmu_cache(vma, address, entry);
2141         lazy_mmu_prot_update(entry);
2142 unlock:
2143         pte_unmap_unlock(page_table, ptl);
2144         return ret;
2145 oom:
2146         page_cache_release(new_page);
2147         return VM_FAULT_OOM;
2148 }
2149
2150 /*
2151  * Fault of a previously existing named mapping. Repopulate the pte
2152  * from the encoded file_pte if possible. This enables swappable
2153  * nonlinear vmas.
2154  *
2155  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2156  * but allow concurrent faults), and pte mapped but not yet locked.
2157  * We return with mmap_sem still held, but pte unmapped and unlocked.
2158  */
2159 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2160                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2161                 int write_access, pte_t orig_pte)
2162 {
2163         pgoff_t pgoff;
2164         int err;
2165
2166         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2167                 return VM_FAULT_MINOR;
2168
2169         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2170                 /*
2171                  * Page table corrupted: show pte and kill process.
2172                  */
2173                 print_bad_pte(vma, orig_pte, address);
2174                 return VM_FAULT_OOM;
2175         }
2176         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2177
2178         pgoff = pte_to_pgoff(orig_pte);
2179         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2180                                         vma->vm_page_prot, pgoff, 0);
2181         if (err == -ENOMEM)
2182                 return VM_FAULT_OOM;
2183         if (err)
2184                 return VM_FAULT_SIGBUS;
2185         return VM_FAULT_MAJOR;
2186 }
2187
2188 /*
2189  * These routines also need to handle stuff like marking pages dirty
2190  * and/or accessed for architectures that don't do it in hardware (most
2191  * RISC architectures).  The early dirtying is also good on the i386.
2192  *
2193  * There is also a hook called "update_mmu_cache()" that architectures
2194  * with external mmu caches can use to update those (ie the Sparc or
2195  * PowerPC hashed page tables that act as extended TLBs).
2196  *
2197  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2198  * but allow concurrent faults), and pte mapped but not yet locked.
2199  * We return with mmap_sem still held, but pte unmapped and unlocked.
2200  */
2201 static inline int handle_pte_fault(struct mm_struct *mm,
2202                 struct vm_area_struct *vma, unsigned long address,
2203                 pte_t *pte, pmd_t *pmd, int write_access)
2204 {
2205         pte_t entry;
2206         pte_t old_entry;
2207         spinlock_t *ptl;
2208
2209         old_entry = entry = *pte;
2210         if (!pte_present(entry)) {
2211                 if (pte_none(entry)) {
2212                         if (!vma->vm_ops || !vma->vm_ops->nopage)
2213                                 return do_anonymous_page(mm, vma, address,
2214                                         pte, pmd, write_access);
2215                         return do_no_page(mm, vma, address,
2216                                         pte, pmd, write_access);
2217                 }
2218                 if (pte_file(entry))
2219                         return do_file_page(mm, vma, address,
2220                                         pte, pmd, write_access, entry);
2221                 return do_swap_page(mm, vma, address,
2222                                         pte, pmd, write_access, entry);
2223         }
2224
2225         ptl = pte_lockptr(mm, pmd);
2226         spin_lock(ptl);
2227         if (unlikely(!pte_same(*pte, entry)))
2228                 goto unlock;
2229         if (write_access) {
2230                 if (!pte_write(entry))
2231                         return do_wp_page(mm, vma, address,
2232                                         pte, pmd, ptl, entry);
2233                 entry = pte_mkdirty(entry);
2234         }
2235         entry = pte_mkyoung(entry);
2236         if (!pte_same(old_entry, entry)) {
2237                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2238                 update_mmu_cache(vma, address, entry);
2239                 lazy_mmu_prot_update(entry);
2240         } else {
2241                 /*
2242                  * This is needed only for protection faults but the arch code
2243                  * is not yet telling us if this is a protection fault or not.
2244                  * This still avoids useless tlb flushes for .text page faults
2245                  * with threads.
2246                  */
2247                 if (write_access)
2248                         flush_tlb_page(vma, address);
2249         }
2250 unlock:
2251         pte_unmap_unlock(pte, ptl);
2252         return VM_FAULT_MINOR;
2253 }
2254
2255 /*
2256  * By the time we get here, we already hold the mm semaphore
2257  */
2258 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2259                 unsigned long address, int write_access)
2260 {
2261         pgd_t *pgd;
2262         pud_t *pud;
2263         pmd_t *pmd;
2264         pte_t *pte;
2265
2266         __set_current_state(TASK_RUNNING);
2267
2268         inc_page_state(pgfault);
2269
2270         if (unlikely(is_vm_hugetlb_page(vma)))
2271                 return hugetlb_fault(mm, vma, address, write_access);
2272
2273         pgd = pgd_offset(mm, address);
2274         pud = pud_alloc(mm, pgd, address);
2275         if (!pud)
2276                 return VM_FAULT_OOM;
2277         pmd = pmd_alloc(mm, pud, address);
2278         if (!pmd)
2279                 return VM_FAULT_OOM;
2280         pte = pte_alloc_map(mm, pmd, address);
2281         if (!pte)
2282                 return VM_FAULT_OOM;
2283
2284         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2285 }
2286
2287 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2288
2289 #ifndef __PAGETABLE_PUD_FOLDED
2290 /*
2291  * Allocate page upper directory.
2292  * We've already handled the fast-path in-line.
2293  */
2294 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2295 {
2296         pud_t *new = pud_alloc_one(mm, address);
2297         if (!new)
2298                 return -ENOMEM;
2299
2300         spin_lock(&mm->page_table_lock);
2301         if (pgd_present(*pgd))          /* Another has populated it */
2302                 pud_free(new);
2303         else
2304                 pgd_populate(mm, pgd, new);
2305         spin_unlock(&mm->page_table_lock);
2306         return 0;
2307 }
2308 #else
2309 /* Workaround for gcc 2.96 */
2310 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2311 {
2312         return 0;
2313 }
2314 #endif /* __PAGETABLE_PUD_FOLDED */
2315
2316 #ifndef __PAGETABLE_PMD_FOLDED
2317 /*
2318  * Allocate page middle directory.
2319  * We've already handled the fast-path in-line.
2320  */
2321 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2322 {
2323         pmd_t *new = pmd_alloc_one(mm, address);
2324         if (!new)
2325                 return -ENOMEM;
2326
2327         spin_lock(&mm->page_table_lock);
2328 #ifndef __ARCH_HAS_4LEVEL_HACK
2329         if (pud_present(*pud))          /* Another has populated it */
2330                 pmd_free(new);
2331         else
2332                 pud_populate(mm, pud, new);
2333 #else
2334         if (pgd_present(*pud))          /* Another has populated it */
2335                 pmd_free(new);
2336         else
2337                 pgd_populate(mm, pud, new);
2338 #endif /* __ARCH_HAS_4LEVEL_HACK */
2339         spin_unlock(&mm->page_table_lock);
2340         return 0;
2341 }
2342 #else
2343 /* Workaround for gcc 2.96 */
2344 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2345 {
2346         return 0;
2347 }
2348 #endif /* __PAGETABLE_PMD_FOLDED */
2349
2350 int make_pages_present(unsigned long addr, unsigned long end)
2351 {
2352         int ret, len, write;
2353         struct vm_area_struct * vma;
2354
2355         vma = find_vma(current->mm, addr);
2356         if (!vma)
2357                 return -1;
2358         write = (vma->vm_flags & VM_WRITE) != 0;
2359         if (addr >= end)
2360                 BUG();
2361         if (end > vma->vm_end)
2362                 BUG();
2363         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2364         ret = get_user_pages(current, current->mm, addr,
2365                         len, write, 0, NULL, NULL);
2366         if (ret < 0)
2367                 return ret;
2368         return ret == len ? 0 : -1;
2369 }
2370
2371 /* 
2372  * Map a vmalloc()-space virtual address to the physical page.
2373  */
2374 struct page * vmalloc_to_page(void * vmalloc_addr)
2375 {
2376         unsigned long addr = (unsigned long) vmalloc_addr;
2377         struct page *page = NULL;
2378         pgd_t *pgd = pgd_offset_k(addr);
2379         pud_t *pud;
2380         pmd_t *pmd;
2381         pte_t *ptep, pte;
2382   
2383         if (!pgd_none(*pgd)) {
2384                 pud = pud_offset(pgd, addr);
2385                 if (!pud_none(*pud)) {
2386                         pmd = pmd_offset(pud, addr);
2387                         if (!pmd_none(*pmd)) {
2388                                 ptep = pte_offset_map(pmd, addr);
2389                                 pte = *ptep;
2390                                 if (pte_present(pte))
2391                                         page = pte_page(pte);
2392                                 pte_unmap(ptep);
2393                         }
2394                 }
2395         }
2396         return page;
2397 }
2398
2399 EXPORT_SYMBOL(vmalloc_to_page);
2400
2401 /*
2402  * Map a vmalloc()-space virtual address to the physical page frame number.
2403  */
2404 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2405 {
2406         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2407 }
2408
2409 EXPORT_SYMBOL(vmalloc_to_pfn);
2410
2411 #if !defined(__HAVE_ARCH_GATE_AREA)
2412
2413 #if defined(AT_SYSINFO_EHDR)
2414 static struct vm_area_struct gate_vma;
2415
2416 static int __init gate_vma_init(void)
2417 {
2418         gate_vma.vm_mm = NULL;
2419         gate_vma.vm_start = FIXADDR_USER_START;
2420         gate_vma.vm_end = FIXADDR_USER_END;
2421         gate_vma.vm_page_prot = PAGE_READONLY;
2422         gate_vma.vm_flags = 0;
2423         return 0;
2424 }
2425 __initcall(gate_vma_init);
2426 #endif
2427
2428 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2429 {
2430 #ifdef AT_SYSINFO_EHDR
2431         return &gate_vma;
2432 #else
2433         return NULL;
2434 #endif
2435 }
2436
2437 int in_gate_area_no_task(unsigned long addr)
2438 {
2439 #ifdef AT_SYSINFO_EHDR
2440         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2441                 return 1;
2442 #endif
2443         return 0;
2444 }
2445
2446 #endif  /* __HAVE_ARCH_GATE_AREA */