[PATCH] mm: more CONFIG_DEBUG_VM
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
54 #include <asm/tlb.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
57
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
65
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
69
70 unsigned long num_physpages;
71 /*
72  * A number of key systems in x86 including ioremap() rely on the assumption
73  * that high_memory defines the upper bound on direct map memory, then end
74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76  * and ZONE_HIGHMEM.
77  */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
80
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85 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 (unlikely(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 #ifdef CONFIG_DEBUG_VM
400         if (unlikely(!pfn_valid(pfn))) {
401                 print_bad_pte(vma, pte, addr);
402                 return NULL;
403         }
404 #endif
405
406         /*
407          * NOTE! We still have PageReserved() pages in the page 
408          * tables. 
409          *
410          * The PAGE_ZERO() pages and various VDSO mappings can
411          * cause them to exist.
412          */
413         return pfn_to_page(pfn);
414 }
415
416 /*
417  * copy one vm_area from one task to the other. Assumes the page tables
418  * already present in the new task to be cleared in the whole range
419  * covered by this vma.
420  */
421
422 static inline void
423 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
424                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
425                 unsigned long addr, int *rss)
426 {
427         unsigned long vm_flags = vma->vm_flags;
428         pte_t pte = *src_pte;
429         struct page *page;
430
431         /* pte contains position in swap or file, so copy. */
432         if (unlikely(!pte_present(pte))) {
433                 if (!pte_file(pte)) {
434                         swap_duplicate(pte_to_swp_entry(pte));
435                         /* make sure dst_mm is on swapoff's mmlist. */
436                         if (unlikely(list_empty(&dst_mm->mmlist))) {
437                                 spin_lock(&mmlist_lock);
438                                 if (list_empty(&dst_mm->mmlist))
439                                         list_add(&dst_mm->mmlist,
440                                                  &src_mm->mmlist);
441                                 spin_unlock(&mmlist_lock);
442                         }
443                 }
444                 goto out_set_pte;
445         }
446
447         /*
448          * If it's a COW mapping, write protect it both
449          * in the parent and the child
450          */
451         if (is_cow_mapping(vm_flags)) {
452                 ptep_set_wrprotect(src_mm, addr, src_pte);
453                 pte = *src_pte;
454         }
455
456         /*
457          * If it's a shared mapping, mark it clean in
458          * the child
459          */
460         if (vm_flags & VM_SHARED)
461                 pte = pte_mkclean(pte);
462         pte = pte_mkold(pte);
463
464         page = vm_normal_page(vma, addr, pte);
465         if (page) {
466                 get_page(page);
467                 page_dup_rmap(page);
468                 rss[!!PageAnon(page)]++;
469         }
470
471 out_set_pte:
472         set_pte_at(dst_mm, addr, dst_pte, pte);
473 }
474
475 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
476                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
477                 unsigned long addr, unsigned long end)
478 {
479         pte_t *src_pte, *dst_pte;
480         spinlock_t *src_ptl, *dst_ptl;
481         int progress = 0;
482         int rss[2];
483
484 again:
485         rss[1] = rss[0] = 0;
486         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
487         if (!dst_pte)
488                 return -ENOMEM;
489         src_pte = pte_offset_map_nested(src_pmd, addr);
490         src_ptl = pte_lockptr(src_mm, src_pmd);
491         spin_lock(src_ptl);
492
493         do {
494                 /*
495                  * We are holding two locks at this point - either of them
496                  * could generate latencies in another task on another CPU.
497                  */
498                 if (progress >= 32) {
499                         progress = 0;
500                         if (need_resched() ||
501                             need_lockbreak(src_ptl) ||
502                             need_lockbreak(dst_ptl))
503                                 break;
504                 }
505                 if (pte_none(*src_pte)) {
506                         progress++;
507                         continue;
508                 }
509                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
510                 progress += 8;
511         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
512
513         spin_unlock(src_ptl);
514         pte_unmap_nested(src_pte - 1);
515         add_mm_rss(dst_mm, rss[0], rss[1]);
516         pte_unmap_unlock(dst_pte - 1, dst_ptl);
517         cond_resched();
518         if (addr != end)
519                 goto again;
520         return 0;
521 }
522
523 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
524                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
525                 unsigned long addr, unsigned long end)
526 {
527         pmd_t *src_pmd, *dst_pmd;
528         unsigned long next;
529
530         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
531         if (!dst_pmd)
532                 return -ENOMEM;
533         src_pmd = pmd_offset(src_pud, addr);
534         do {
535                 next = pmd_addr_end(addr, end);
536                 if (pmd_none_or_clear_bad(src_pmd))
537                         continue;
538                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
539                                                 vma, addr, next))
540                         return -ENOMEM;
541         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
542         return 0;
543 }
544
545 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
546                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
547                 unsigned long addr, unsigned long end)
548 {
549         pud_t *src_pud, *dst_pud;
550         unsigned long next;
551
552         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
553         if (!dst_pud)
554                 return -ENOMEM;
555         src_pud = pud_offset(src_pgd, addr);
556         do {
557                 next = pud_addr_end(addr, end);
558                 if (pud_none_or_clear_bad(src_pud))
559                         continue;
560                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
561                                                 vma, addr, next))
562                         return -ENOMEM;
563         } while (dst_pud++, src_pud++, addr = next, addr != end);
564         return 0;
565 }
566
567 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
568                 struct vm_area_struct *vma)
569 {
570         pgd_t *src_pgd, *dst_pgd;
571         unsigned long next;
572         unsigned long addr = vma->vm_start;
573         unsigned long end = vma->vm_end;
574
575         /*
576          * Don't copy ptes where a page fault will fill them correctly.
577          * Fork becomes much lighter when there are big shared or private
578          * readonly mappings. The tradeoff is that copy_page_range is more
579          * efficient than faulting.
580          */
581         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
582                 if (!vma->anon_vma)
583                         return 0;
584         }
585
586         if (is_vm_hugetlb_page(vma))
587                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
588
589         dst_pgd = pgd_offset(dst_mm, addr);
590         src_pgd = pgd_offset(src_mm, addr);
591         do {
592                 next = pgd_addr_end(addr, end);
593                 if (pgd_none_or_clear_bad(src_pgd))
594                         continue;
595                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
596                                                 vma, addr, next))
597                         return -ENOMEM;
598         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
599         return 0;
600 }
601
602 static unsigned long zap_pte_range(struct mmu_gather *tlb,
603                                 struct vm_area_struct *vma, pmd_t *pmd,
604                                 unsigned long addr, unsigned long end,
605                                 long *zap_work, struct zap_details *details)
606 {
607         struct mm_struct *mm = tlb->mm;
608         pte_t *pte;
609         spinlock_t *ptl;
610         int file_rss = 0;
611         int anon_rss = 0;
612
613         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
614         do {
615                 pte_t ptent = *pte;
616                 if (pte_none(ptent)) {
617                         (*zap_work)--;
618                         continue;
619                 }
620
621                 (*zap_work) -= PAGE_SIZE;
622
623                 if (pte_present(ptent)) {
624                         struct page *page;
625
626                         page = vm_normal_page(vma, addr, ptent);
627                         if (unlikely(details) && page) {
628                                 /*
629                                  * unmap_shared_mapping_pages() wants to
630                                  * invalidate cache without truncating:
631                                  * unmap shared but keep private pages.
632                                  */
633                                 if (details->check_mapping &&
634                                     details->check_mapping != page->mapping)
635                                         continue;
636                                 /*
637                                  * Each page->index must be checked when
638                                  * invalidating or truncating nonlinear.
639                                  */
640                                 if (details->nonlinear_vma &&
641                                     (page->index < details->first_index ||
642                                      page->index > details->last_index))
643                                         continue;
644                         }
645                         ptent = ptep_get_and_clear_full(mm, addr, pte,
646                                                         tlb->fullmm);
647                         tlb_remove_tlb_entry(tlb, pte, addr);
648                         if (unlikely(!page))
649                                 continue;
650                         if (unlikely(details) && details->nonlinear_vma
651                             && linear_page_index(details->nonlinear_vma,
652                                                 addr) != page->index)
653                                 set_pte_at(mm, addr, pte,
654                                            pgoff_to_pte(page->index));
655                         if (PageAnon(page))
656                                 anon_rss--;
657                         else {
658                                 if (pte_dirty(ptent))
659                                         set_page_dirty(page);
660                                 if (pte_young(ptent))
661                                         mark_page_accessed(page);
662                                 file_rss--;
663                         }
664                         page_remove_rmap(page);
665                         tlb_remove_page(tlb, page);
666                         continue;
667                 }
668                 /*
669                  * If details->check_mapping, we leave swap entries;
670                  * if details->nonlinear_vma, we leave file entries.
671                  */
672                 if (unlikely(details))
673                         continue;
674                 if (!pte_file(ptent))
675                         free_swap_and_cache(pte_to_swp_entry(ptent));
676                 pte_clear_full(mm, addr, pte, tlb->fullmm);
677         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
678
679         add_mm_rss(mm, file_rss, anon_rss);
680         pte_unmap_unlock(pte - 1, ptl);
681
682         return addr;
683 }
684
685 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
686                                 struct vm_area_struct *vma, pud_t *pud,
687                                 unsigned long addr, unsigned long end,
688                                 long *zap_work, struct zap_details *details)
689 {
690         pmd_t *pmd;
691         unsigned long next;
692
693         pmd = pmd_offset(pud, addr);
694         do {
695                 next = pmd_addr_end(addr, end);
696                 if (pmd_none_or_clear_bad(pmd)) {
697                         (*zap_work)--;
698                         continue;
699                 }
700                 next = zap_pte_range(tlb, vma, pmd, addr, next,
701                                                 zap_work, details);
702         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
703
704         return addr;
705 }
706
707 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
708                                 struct vm_area_struct *vma, pgd_t *pgd,
709                                 unsigned long addr, unsigned long end,
710                                 long *zap_work, struct zap_details *details)
711 {
712         pud_t *pud;
713         unsigned long next;
714
715         pud = pud_offset(pgd, addr);
716         do {
717                 next = pud_addr_end(addr, end);
718                 if (pud_none_or_clear_bad(pud)) {
719                         (*zap_work)--;
720                         continue;
721                 }
722                 next = zap_pmd_range(tlb, vma, pud, addr, next,
723                                                 zap_work, details);
724         } while (pud++, addr = next, (addr != end && *zap_work > 0));
725
726         return addr;
727 }
728
729 static unsigned long unmap_page_range(struct mmu_gather *tlb,
730                                 struct vm_area_struct *vma,
731                                 unsigned long addr, unsigned long end,
732                                 long *zap_work, struct zap_details *details)
733 {
734         pgd_t *pgd;
735         unsigned long next;
736
737         if (details && !details->check_mapping && !details->nonlinear_vma)
738                 details = NULL;
739
740         BUG_ON(addr >= end);
741         tlb_start_vma(tlb, vma);
742         pgd = pgd_offset(vma->vm_mm, addr);
743         do {
744                 next = pgd_addr_end(addr, end);
745                 if (pgd_none_or_clear_bad(pgd)) {
746                         (*zap_work)--;
747                         continue;
748                 }
749                 next = zap_pud_range(tlb, vma, pgd, addr, next,
750                                                 zap_work, details);
751         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
752         tlb_end_vma(tlb, vma);
753
754         return addr;
755 }
756
757 #ifdef CONFIG_PREEMPT
758 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
759 #else
760 /* No preempt: go for improved straight-line efficiency */
761 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
762 #endif
763
764 /**
765  * unmap_vmas - unmap a range of memory covered by a list of vma's
766  * @tlbp: address of the caller's struct mmu_gather
767  * @vma: the starting vma
768  * @start_addr: virtual address at which to start unmapping
769  * @end_addr: virtual address at which to end unmapping
770  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
771  * @details: details of nonlinear truncation or shared cache invalidation
772  *
773  * Returns the end address of the unmapping (restart addr if interrupted).
774  *
775  * Unmap all pages in the vma list.
776  *
777  * We aim to not hold locks for too long (for scheduling latency reasons).
778  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
779  * return the ending mmu_gather to the caller.
780  *
781  * Only addresses between `start' and `end' will be unmapped.
782  *
783  * The VMA list must be sorted in ascending virtual address order.
784  *
785  * unmap_vmas() assumes that the caller will flush the whole unmapped address
786  * range after unmap_vmas() returns.  So the only responsibility here is to
787  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
788  * drops the lock and schedules.
789  */
790 unsigned long unmap_vmas(struct mmu_gather **tlbp,
791                 struct vm_area_struct *vma, unsigned long start_addr,
792                 unsigned long end_addr, unsigned long *nr_accounted,
793                 struct zap_details *details)
794 {
795         long zap_work = ZAP_BLOCK_SIZE;
796         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
797         int tlb_start_valid = 0;
798         unsigned long start = start_addr;
799         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
800         int fullmm = (*tlbp)->fullmm;
801
802         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
803                 unsigned long end;
804
805                 start = max(vma->vm_start, start_addr);
806                 if (start >= vma->vm_end)
807                         continue;
808                 end = min(vma->vm_end, end_addr);
809                 if (end <= vma->vm_start)
810                         continue;
811
812                 if (vma->vm_flags & VM_ACCOUNT)
813                         *nr_accounted += (end - start) >> PAGE_SHIFT;
814
815                 while (start != end) {
816                         if (!tlb_start_valid) {
817                                 tlb_start = start;
818                                 tlb_start_valid = 1;
819                         }
820
821                         if (unlikely(is_vm_hugetlb_page(vma))) {
822                                 unmap_hugepage_range(vma, start, end);
823                                 zap_work -= (end - start) /
824                                                 (HPAGE_SIZE / PAGE_SIZE);
825                                 start = end;
826                         } else
827                                 start = unmap_page_range(*tlbp, vma,
828                                                 start, end, &zap_work, details);
829
830                         if (zap_work > 0) {
831                                 BUG_ON(start != end);
832                                 break;
833                         }
834
835                         tlb_finish_mmu(*tlbp, tlb_start, start);
836
837                         if (need_resched() ||
838                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
839                                 if (i_mmap_lock) {
840                                         *tlbp = NULL;
841                                         goto out;
842                                 }
843                                 cond_resched();
844                         }
845
846                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
847                         tlb_start_valid = 0;
848                         zap_work = ZAP_BLOCK_SIZE;
849                 }
850         }
851 out:
852         return start;   /* which is now the end (or restart) address */
853 }
854
855 /**
856  * zap_page_range - remove user pages in a given range
857  * @vma: vm_area_struct holding the applicable pages
858  * @address: starting address of pages to zap
859  * @size: number of bytes to zap
860  * @details: details of nonlinear truncation or shared cache invalidation
861  */
862 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
863                 unsigned long size, struct zap_details *details)
864 {
865         struct mm_struct *mm = vma->vm_mm;
866         struct mmu_gather *tlb;
867         unsigned long end = address + size;
868         unsigned long nr_accounted = 0;
869
870         lru_add_drain();
871         tlb = tlb_gather_mmu(mm, 0);
872         update_hiwater_rss(mm);
873         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
874         if (tlb)
875                 tlb_finish_mmu(tlb, address, end);
876         return end;
877 }
878
879 /*
880  * Do a quick page-table lookup for a single page.
881  */
882 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
883                         unsigned int flags)
884 {
885         pgd_t *pgd;
886         pud_t *pud;
887         pmd_t *pmd;
888         pte_t *ptep, pte;
889         spinlock_t *ptl;
890         struct page *page;
891         struct mm_struct *mm = vma->vm_mm;
892
893         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
894         if (!IS_ERR(page)) {
895                 BUG_ON(flags & FOLL_GET);
896                 goto out;
897         }
898
899         page = NULL;
900         pgd = pgd_offset(mm, address);
901         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
902                 goto no_page_table;
903
904         pud = pud_offset(pgd, address);
905         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
906                 goto no_page_table;
907         
908         pmd = pmd_offset(pud, address);
909         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
910                 goto no_page_table;
911
912         if (pmd_huge(*pmd)) {
913                 BUG_ON(flags & FOLL_GET);
914                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
915                 goto out;
916         }
917
918         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
919         if (!ptep)
920                 goto out;
921
922         pte = *ptep;
923         if (!pte_present(pte))
924                 goto unlock;
925         if ((flags & FOLL_WRITE) && !pte_write(pte))
926                 goto unlock;
927         page = vm_normal_page(vma, address, pte);
928         if (unlikely(!page))
929                 goto unlock;
930
931         if (flags & FOLL_GET)
932                 get_page(page);
933         if (flags & FOLL_TOUCH) {
934                 if ((flags & FOLL_WRITE) &&
935                     !pte_dirty(pte) && !PageDirty(page))
936                         set_page_dirty(page);
937                 mark_page_accessed(page);
938         }
939 unlock:
940         pte_unmap_unlock(ptep, ptl);
941 out:
942         return page;
943
944 no_page_table:
945         /*
946          * When core dumping an enormous anonymous area that nobody
947          * has touched so far, we don't want to allocate page tables.
948          */
949         if (flags & FOLL_ANON) {
950                 page = ZERO_PAGE(address);
951                 if (flags & FOLL_GET)
952                         get_page(page);
953                 BUG_ON(flags & FOLL_WRITE);
954         }
955         return page;
956 }
957
958 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
959                 unsigned long start, int len, int write, int force,
960                 struct page **pages, struct vm_area_struct **vmas)
961 {
962         int i;
963         unsigned int vm_flags;
964
965         /* 
966          * Require read or write permissions.
967          * If 'force' is set, we only require the "MAY" flags.
968          */
969         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
970         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
971         i = 0;
972
973         do {
974                 struct vm_area_struct *vma;
975                 unsigned int foll_flags;
976
977                 vma = find_extend_vma(mm, start);
978                 if (!vma && in_gate_area(tsk, start)) {
979                         unsigned long pg = start & PAGE_MASK;
980                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
981                         pgd_t *pgd;
982                         pud_t *pud;
983                         pmd_t *pmd;
984                         pte_t *pte;
985                         if (write) /* user gate pages are read-only */
986                                 return i ? : -EFAULT;
987                         if (pg > TASK_SIZE)
988                                 pgd = pgd_offset_k(pg);
989                         else
990                                 pgd = pgd_offset_gate(mm, pg);
991                         BUG_ON(pgd_none(*pgd));
992                         pud = pud_offset(pgd, pg);
993                         BUG_ON(pud_none(*pud));
994                         pmd = pmd_offset(pud, pg);
995                         if (pmd_none(*pmd))
996                                 return i ? : -EFAULT;
997                         pte = pte_offset_map(pmd, pg);
998                         if (pte_none(*pte)) {
999                                 pte_unmap(pte);
1000                                 return i ? : -EFAULT;
1001                         }
1002                         if (pages) {
1003                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1004                                 pages[i] = page;
1005                                 if (page)
1006                                         get_page(page);
1007                         }
1008                         pte_unmap(pte);
1009                         if (vmas)
1010                                 vmas[i] = gate_vma;
1011                         i++;
1012                         start += PAGE_SIZE;
1013                         len--;
1014                         continue;
1015                 }
1016
1017                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1018                                 || !(vm_flags & vma->vm_flags))
1019                         return i ? : -EFAULT;
1020
1021                 if (is_vm_hugetlb_page(vma)) {
1022                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1023                                                 &start, &len, i);
1024                         continue;
1025                 }
1026
1027                 foll_flags = FOLL_TOUCH;
1028                 if (pages)
1029                         foll_flags |= FOLL_GET;
1030                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1031                     (!vma->vm_ops || !vma->vm_ops->nopage))
1032                         foll_flags |= FOLL_ANON;
1033
1034                 do {
1035                         struct page *page;
1036
1037                         if (write)
1038                                 foll_flags |= FOLL_WRITE;
1039
1040                         cond_resched();
1041                         while (!(page = follow_page(vma, start, foll_flags))) {
1042                                 int ret;
1043                                 ret = __handle_mm_fault(mm, vma, start,
1044                                                 foll_flags & FOLL_WRITE);
1045                                 /*
1046                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1047                                  * broken COW when necessary, even if maybe_mkwrite
1048                                  * decided not to set pte_write. We can thus safely do
1049                                  * subsequent page lookups as if they were reads.
1050                                  */
1051                                 if (ret & VM_FAULT_WRITE)
1052                                         foll_flags &= ~FOLL_WRITE;
1053                                 
1054                                 switch (ret & ~VM_FAULT_WRITE) {
1055                                 case VM_FAULT_MINOR:
1056                                         tsk->min_flt++;
1057                                         break;
1058                                 case VM_FAULT_MAJOR:
1059                                         tsk->maj_flt++;
1060                                         break;
1061                                 case VM_FAULT_SIGBUS:
1062                                         return i ? i : -EFAULT;
1063                                 case VM_FAULT_OOM:
1064                                         return i ? i : -ENOMEM;
1065                                 default:
1066                                         BUG();
1067                                 }
1068                         }
1069                         if (pages) {
1070                                 pages[i] = page;
1071                                 flush_dcache_page(page);
1072                         }
1073                         if (vmas)
1074                                 vmas[i] = vma;
1075                         i++;
1076                         start += PAGE_SIZE;
1077                         len--;
1078                 } while (len && start < vma->vm_end);
1079         } while (len);
1080         return i;
1081 }
1082 EXPORT_SYMBOL(get_user_pages);
1083
1084 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1085                         unsigned long addr, unsigned long end, pgprot_t prot)
1086 {
1087         pte_t *pte;
1088         spinlock_t *ptl;
1089
1090         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1091         if (!pte)
1092                 return -ENOMEM;
1093         do {
1094                 struct page *page = ZERO_PAGE(addr);
1095                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1096                 page_cache_get(page);
1097                 page_add_file_rmap(page);
1098                 inc_mm_counter(mm, file_rss);
1099                 BUG_ON(!pte_none(*pte));
1100                 set_pte_at(mm, addr, pte, zero_pte);
1101         } while (pte++, addr += PAGE_SIZE, addr != end);
1102         pte_unmap_unlock(pte - 1, ptl);
1103         return 0;
1104 }
1105
1106 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1107                         unsigned long addr, unsigned long end, pgprot_t prot)
1108 {
1109         pmd_t *pmd;
1110         unsigned long next;
1111
1112         pmd = pmd_alloc(mm, pud, addr);
1113         if (!pmd)
1114                 return -ENOMEM;
1115         do {
1116                 next = pmd_addr_end(addr, end);
1117                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1118                         return -ENOMEM;
1119         } while (pmd++, addr = next, addr != end);
1120         return 0;
1121 }
1122
1123 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1124                         unsigned long addr, unsigned long end, pgprot_t prot)
1125 {
1126         pud_t *pud;
1127         unsigned long next;
1128
1129         pud = pud_alloc(mm, pgd, addr);
1130         if (!pud)
1131                 return -ENOMEM;
1132         do {
1133                 next = pud_addr_end(addr, end);
1134                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1135                         return -ENOMEM;
1136         } while (pud++, addr = next, addr != end);
1137         return 0;
1138 }
1139
1140 int zeromap_page_range(struct vm_area_struct *vma,
1141                         unsigned long addr, unsigned long size, pgprot_t prot)
1142 {
1143         pgd_t *pgd;
1144         unsigned long next;
1145         unsigned long end = addr + size;
1146         struct mm_struct *mm = vma->vm_mm;
1147         int err;
1148
1149         BUG_ON(addr >= end);
1150         pgd = pgd_offset(mm, addr);
1151         flush_cache_range(vma, addr, end);
1152         do {
1153                 next = pgd_addr_end(addr, end);
1154                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1155                 if (err)
1156                         break;
1157         } while (pgd++, addr = next, addr != end);
1158         return err;
1159 }
1160
1161 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1162 {
1163         pgd_t * pgd = pgd_offset(mm, addr);
1164         pud_t * pud = pud_alloc(mm, pgd, addr);
1165         if (pud) {
1166                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1167                 if (pmd)
1168                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1169         }
1170         return NULL;
1171 }
1172
1173 /*
1174  * This is the old fallback for page remapping.
1175  *
1176  * For historical reasons, it only allows reserved pages. Only
1177  * old drivers should use this, and they needed to mark their
1178  * pages reserved for the old functions anyway.
1179  */
1180 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1181 {
1182         int retval;
1183         pte_t *pte;
1184         spinlock_t *ptl;  
1185
1186         retval = -EINVAL;
1187         if (PageAnon(page))
1188                 goto out;
1189         retval = -ENOMEM;
1190         flush_dcache_page(page);
1191         pte = get_locked_pte(mm, addr, &ptl);
1192         if (!pte)
1193                 goto out;
1194         retval = -EBUSY;
1195         if (!pte_none(*pte))
1196                 goto out_unlock;
1197
1198         /* Ok, finally just insert the thing.. */
1199         get_page(page);
1200         inc_mm_counter(mm, file_rss);
1201         page_add_file_rmap(page);
1202         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1203
1204         retval = 0;
1205 out_unlock:
1206         pte_unmap_unlock(pte, ptl);
1207 out:
1208         return retval;
1209 }
1210
1211 /*
1212  * This allows drivers to insert individual pages they've allocated
1213  * into a user vma.
1214  *
1215  * The page has to be a nice clean _individual_ kernel allocation.
1216  * If you allocate a compound page, you need to have marked it as
1217  * such (__GFP_COMP), or manually just split the page up yourself
1218  * (see split_page()).
1219  *
1220  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1221  * took an arbitrary page protection parameter. This doesn't allow
1222  * that. Your vma protection will have to be set up correctly, which
1223  * means that if you want a shared writable mapping, you'd better
1224  * ask for a shared writable mapping!
1225  *
1226  * The page does not need to be reserved.
1227  */
1228 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1229 {
1230         if (addr < vma->vm_start || addr >= vma->vm_end)
1231                 return -EFAULT;
1232         if (!page_count(page))
1233                 return -EINVAL;
1234         vma->vm_flags |= VM_INSERTPAGE;
1235         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1236 }
1237 EXPORT_SYMBOL(vm_insert_page);
1238
1239 /*
1240  * maps a range of physical memory into the requested pages. the old
1241  * mappings are removed. any references to nonexistent pages results
1242  * in null mappings (currently treated as "copy-on-access")
1243  */
1244 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1245                         unsigned long addr, unsigned long end,
1246                         unsigned long pfn, pgprot_t prot)
1247 {
1248         pte_t *pte;
1249         spinlock_t *ptl;
1250
1251         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1252         if (!pte)
1253                 return -ENOMEM;
1254         do {
1255                 BUG_ON(!pte_none(*pte));
1256                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1257                 pfn++;
1258         } while (pte++, addr += PAGE_SIZE, addr != end);
1259         pte_unmap_unlock(pte - 1, ptl);
1260         return 0;
1261 }
1262
1263 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1264                         unsigned long addr, unsigned long end,
1265                         unsigned long pfn, pgprot_t prot)
1266 {
1267         pmd_t *pmd;
1268         unsigned long next;
1269
1270         pfn -= addr >> PAGE_SHIFT;
1271         pmd = pmd_alloc(mm, pud, addr);
1272         if (!pmd)
1273                 return -ENOMEM;
1274         do {
1275                 next = pmd_addr_end(addr, end);
1276                 if (remap_pte_range(mm, pmd, addr, next,
1277                                 pfn + (addr >> PAGE_SHIFT), prot))
1278                         return -ENOMEM;
1279         } while (pmd++, addr = next, addr != end);
1280         return 0;
1281 }
1282
1283 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1284                         unsigned long addr, unsigned long end,
1285                         unsigned long pfn, pgprot_t prot)
1286 {
1287         pud_t *pud;
1288         unsigned long next;
1289
1290         pfn -= addr >> PAGE_SHIFT;
1291         pud = pud_alloc(mm, pgd, addr);
1292         if (!pud)
1293                 return -ENOMEM;
1294         do {
1295                 next = pud_addr_end(addr, end);
1296                 if (remap_pmd_range(mm, pud, addr, next,
1297                                 pfn + (addr >> PAGE_SHIFT), prot))
1298                         return -ENOMEM;
1299         } while (pud++, addr = next, addr != end);
1300         return 0;
1301 }
1302
1303 /*  Note: this is only safe if the mm semaphore is held when called. */
1304 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1305                     unsigned long pfn, unsigned long size, pgprot_t prot)
1306 {
1307         pgd_t *pgd;
1308         unsigned long next;
1309         unsigned long end = addr + PAGE_ALIGN(size);
1310         struct mm_struct *mm = vma->vm_mm;
1311         int err;
1312
1313         /*
1314          * Physically remapped pages are special. Tell the
1315          * rest of the world about it:
1316          *   VM_IO tells people not to look at these pages
1317          *      (accesses can have side effects).
1318          *   VM_RESERVED is specified all over the place, because
1319          *      in 2.4 it kept swapout's vma scan off this vma; but
1320          *      in 2.6 the LRU scan won't even find its pages, so this
1321          *      flag means no more than count its pages in reserved_vm,
1322          *      and omit it from core dump, even when VM_IO turned off.
1323          *   VM_PFNMAP tells the core MM that the base pages are just
1324          *      raw PFN mappings, and do not have a "struct page" associated
1325          *      with them.
1326          *
1327          * There's a horrible special case to handle copy-on-write
1328          * behaviour that some programs depend on. We mark the "original"
1329          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1330          */
1331         if (is_cow_mapping(vma->vm_flags)) {
1332                 if (addr != vma->vm_start || end != vma->vm_end)
1333                         return -EINVAL;
1334                 vma->vm_pgoff = pfn;
1335         }
1336
1337         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1338
1339         BUG_ON(addr >= end);
1340         pfn -= addr >> PAGE_SHIFT;
1341         pgd = pgd_offset(mm, addr);
1342         flush_cache_range(vma, addr, end);
1343         do {
1344                 next = pgd_addr_end(addr, end);
1345                 err = remap_pud_range(mm, pgd, addr, next,
1346                                 pfn + (addr >> PAGE_SHIFT), prot);
1347                 if (err)
1348                         break;
1349         } while (pgd++, addr = next, addr != end);
1350         return err;
1351 }
1352 EXPORT_SYMBOL(remap_pfn_range);
1353
1354 /*
1355  * handle_pte_fault chooses page fault handler according to an entry
1356  * which was read non-atomically.  Before making any commitment, on
1357  * those architectures or configurations (e.g. i386 with PAE) which
1358  * might give a mix of unmatched parts, do_swap_page and do_file_page
1359  * must check under lock before unmapping the pte and proceeding
1360  * (but do_wp_page is only called after already making such a check;
1361  * and do_anonymous_page and do_no_page can safely check later on).
1362  */
1363 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1364                                 pte_t *page_table, pte_t orig_pte)
1365 {
1366         int same = 1;
1367 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1368         if (sizeof(pte_t) > sizeof(unsigned long)) {
1369                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1370                 spin_lock(ptl);
1371                 same = pte_same(*page_table, orig_pte);
1372                 spin_unlock(ptl);
1373         }
1374 #endif
1375         pte_unmap(page_table);
1376         return same;
1377 }
1378
1379 /*
1380  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1381  * servicing faults for write access.  In the normal case, do always want
1382  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1383  * that do not have writing enabled, when used by access_process_vm.
1384  */
1385 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1386 {
1387         if (likely(vma->vm_flags & VM_WRITE))
1388                 pte = pte_mkwrite(pte);
1389         return pte;
1390 }
1391
1392 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1393 {
1394         /*
1395          * If the source page was a PFN mapping, we don't have
1396          * a "struct page" for it. We do a best-effort copy by
1397          * just copying from the original user address. If that
1398          * fails, we just zero-fill it. Live with it.
1399          */
1400         if (unlikely(!src)) {
1401                 void *kaddr = kmap_atomic(dst, KM_USER0);
1402                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1403
1404                 /*
1405                  * This really shouldn't fail, because the page is there
1406                  * in the page tables. But it might just be unreadable,
1407                  * in which case we just give up and fill the result with
1408                  * zeroes.
1409                  */
1410                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1411                         memset(kaddr, 0, PAGE_SIZE);
1412                 kunmap_atomic(kaddr, KM_USER0);
1413                 return;
1414                 
1415         }
1416         copy_user_highpage(dst, src, va);
1417 }
1418
1419 /*
1420  * This routine handles present pages, when users try to write
1421  * to a shared page. It is done by copying the page to a new address
1422  * and decrementing the shared-page counter for the old page.
1423  *
1424  * Note that this routine assumes that the protection checks have been
1425  * done by the caller (the low-level page fault routine in most cases).
1426  * Thus we can safely just mark it writable once we've done any necessary
1427  * COW.
1428  *
1429  * We also mark the page dirty at this point even though the page will
1430  * change only once the write actually happens. This avoids a few races,
1431  * and potentially makes it more efficient.
1432  *
1433  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1434  * but allow concurrent faults), with pte both mapped and locked.
1435  * We return with mmap_sem still held, but pte unmapped and unlocked.
1436  */
1437 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1438                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1439                 spinlock_t *ptl, pte_t orig_pte)
1440 {
1441         struct page *old_page, *new_page;
1442         pte_t entry;
1443         int ret = VM_FAULT_MINOR;
1444
1445         old_page = vm_normal_page(vma, address, orig_pte);
1446         if (!old_page)
1447                 goto gotten;
1448
1449         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1450                 int reuse = can_share_swap_page(old_page);
1451                 unlock_page(old_page);
1452                 if (reuse) {
1453                         flush_cache_page(vma, address, pte_pfn(orig_pte));
1454                         entry = pte_mkyoung(orig_pte);
1455                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1456                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1457                         update_mmu_cache(vma, address, entry);
1458                         lazy_mmu_prot_update(entry);
1459                         ret |= VM_FAULT_WRITE;
1460                         goto unlock;
1461                 }
1462         }
1463
1464         /*
1465          * Ok, we need to copy. Oh, well..
1466          */
1467         page_cache_get(old_page);
1468 gotten:
1469         pte_unmap_unlock(page_table, ptl);
1470
1471         if (unlikely(anon_vma_prepare(vma)))
1472                 goto oom;
1473         if (old_page == ZERO_PAGE(address)) {
1474                 new_page = alloc_zeroed_user_highpage(vma, address);
1475                 if (!new_page)
1476                         goto oom;
1477         } else {
1478                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1479                 if (!new_page)
1480                         goto oom;
1481                 cow_user_page(new_page, old_page, address);
1482         }
1483
1484         /*
1485          * Re-check the pte - we dropped the lock
1486          */
1487         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1488         if (likely(pte_same(*page_table, orig_pte))) {
1489                 if (old_page) {
1490                         page_remove_rmap(old_page);
1491                         if (!PageAnon(old_page)) {
1492                                 dec_mm_counter(mm, file_rss);
1493                                 inc_mm_counter(mm, anon_rss);
1494                         }
1495                 } else
1496                         inc_mm_counter(mm, anon_rss);
1497                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1498                 entry = mk_pte(new_page, vma->vm_page_prot);
1499                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1500                 ptep_establish(vma, address, page_table, entry);
1501                 update_mmu_cache(vma, address, entry);
1502                 lazy_mmu_prot_update(entry);
1503                 lru_cache_add_active(new_page);
1504                 page_add_new_anon_rmap(new_page, vma, address);
1505
1506                 /* Free the old page.. */
1507                 new_page = old_page;
1508                 ret |= VM_FAULT_WRITE;
1509         }
1510         if (new_page)
1511                 page_cache_release(new_page);
1512         if (old_page)
1513                 page_cache_release(old_page);
1514 unlock:
1515         pte_unmap_unlock(page_table, ptl);
1516         return ret;
1517 oom:
1518         if (old_page)
1519                 page_cache_release(old_page);
1520         return VM_FAULT_OOM;
1521 }
1522
1523 /*
1524  * Helper functions for unmap_mapping_range().
1525  *
1526  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1527  *
1528  * We have to restart searching the prio_tree whenever we drop the lock,
1529  * since the iterator is only valid while the lock is held, and anyway
1530  * a later vma might be split and reinserted earlier while lock dropped.
1531  *
1532  * The list of nonlinear vmas could be handled more efficiently, using
1533  * a placeholder, but handle it in the same way until a need is shown.
1534  * It is important to search the prio_tree before nonlinear list: a vma
1535  * may become nonlinear and be shifted from prio_tree to nonlinear list
1536  * while the lock is dropped; but never shifted from list to prio_tree.
1537  *
1538  * In order to make forward progress despite restarting the search,
1539  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1540  * quickly skip it next time around.  Since the prio_tree search only
1541  * shows us those vmas affected by unmapping the range in question, we
1542  * can't efficiently keep all vmas in step with mapping->truncate_count:
1543  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1544  * mapping->truncate_count and vma->vm_truncate_count are protected by
1545  * i_mmap_lock.
1546  *
1547  * In order to make forward progress despite repeatedly restarting some
1548  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1549  * and restart from that address when we reach that vma again.  It might
1550  * have been split or merged, shrunk or extended, but never shifted: so
1551  * restart_addr remains valid so long as it remains in the vma's range.
1552  * unmap_mapping_range forces truncate_count to leap over page-aligned
1553  * values so we can save vma's restart_addr in its truncate_count field.
1554  */
1555 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1556
1557 static void reset_vma_truncate_counts(struct address_space *mapping)
1558 {
1559         struct vm_area_struct *vma;
1560         struct prio_tree_iter iter;
1561
1562         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1563                 vma->vm_truncate_count = 0;
1564         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1565                 vma->vm_truncate_count = 0;
1566 }
1567
1568 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1569                 unsigned long start_addr, unsigned long end_addr,
1570                 struct zap_details *details)
1571 {
1572         unsigned long restart_addr;
1573         int need_break;
1574
1575 again:
1576         restart_addr = vma->vm_truncate_count;
1577         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1578                 start_addr = restart_addr;
1579                 if (start_addr >= end_addr) {
1580                         /* Top of vma has been split off since last time */
1581                         vma->vm_truncate_count = details->truncate_count;
1582                         return 0;
1583                 }
1584         }
1585
1586         restart_addr = zap_page_range(vma, start_addr,
1587                                         end_addr - start_addr, details);
1588         need_break = need_resched() ||
1589                         need_lockbreak(details->i_mmap_lock);
1590
1591         if (restart_addr >= end_addr) {
1592                 /* We have now completed this vma: mark it so */
1593                 vma->vm_truncate_count = details->truncate_count;
1594                 if (!need_break)
1595                         return 0;
1596         } else {
1597                 /* Note restart_addr in vma's truncate_count field */
1598                 vma->vm_truncate_count = restart_addr;
1599                 if (!need_break)
1600                         goto again;
1601         }
1602
1603         spin_unlock(details->i_mmap_lock);
1604         cond_resched();
1605         spin_lock(details->i_mmap_lock);
1606         return -EINTR;
1607 }
1608
1609 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1610                                             struct zap_details *details)
1611 {
1612         struct vm_area_struct *vma;
1613         struct prio_tree_iter iter;
1614         pgoff_t vba, vea, zba, zea;
1615
1616 restart:
1617         vma_prio_tree_foreach(vma, &iter, root,
1618                         details->first_index, details->last_index) {
1619                 /* Skip quickly over those we have already dealt with */
1620                 if (vma->vm_truncate_count == details->truncate_count)
1621                         continue;
1622
1623                 vba = vma->vm_pgoff;
1624                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1625                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1626                 zba = details->first_index;
1627                 if (zba < vba)
1628                         zba = vba;
1629                 zea = details->last_index;
1630                 if (zea > vea)
1631                         zea = vea;
1632
1633                 if (unmap_mapping_range_vma(vma,
1634                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1635                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1636                                 details) < 0)
1637                         goto restart;
1638         }
1639 }
1640
1641 static inline void unmap_mapping_range_list(struct list_head *head,
1642                                             struct zap_details *details)
1643 {
1644         struct vm_area_struct *vma;
1645
1646         /*
1647          * In nonlinear VMAs there is no correspondence between virtual address
1648          * offset and file offset.  So we must perform an exhaustive search
1649          * across *all* the pages in each nonlinear VMA, not just the pages
1650          * whose virtual address lies outside the file truncation point.
1651          */
1652 restart:
1653         list_for_each_entry(vma, head, shared.vm_set.list) {
1654                 /* Skip quickly over those we have already dealt with */
1655                 if (vma->vm_truncate_count == details->truncate_count)
1656                         continue;
1657                 details->nonlinear_vma = vma;
1658                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1659                                         vma->vm_end, details) < 0)
1660                         goto restart;
1661         }
1662 }
1663
1664 /**
1665  * unmap_mapping_range - unmap the portion of all mmaps
1666  * in the specified address_space corresponding to the specified
1667  * page range in the underlying file.
1668  * @mapping: the address space containing mmaps to be unmapped.
1669  * @holebegin: byte in first page to unmap, relative to the start of
1670  * the underlying file.  This will be rounded down to a PAGE_SIZE
1671  * boundary.  Note that this is different from vmtruncate(), which
1672  * must keep the partial page.  In contrast, we must get rid of
1673  * partial pages.
1674  * @holelen: size of prospective hole in bytes.  This will be rounded
1675  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1676  * end of the file.
1677  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1678  * but 0 when invalidating pagecache, don't throw away private data.
1679  */
1680 void unmap_mapping_range(struct address_space *mapping,
1681                 loff_t const holebegin, loff_t const holelen, int even_cows)
1682 {
1683         struct zap_details details;
1684         pgoff_t hba = holebegin >> PAGE_SHIFT;
1685         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1686
1687         /* Check for overflow. */
1688         if (sizeof(holelen) > sizeof(hlen)) {
1689                 long long holeend =
1690                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1691                 if (holeend & ~(long long)ULONG_MAX)
1692                         hlen = ULONG_MAX - hba + 1;
1693         }
1694
1695         details.check_mapping = even_cows? NULL: mapping;
1696         details.nonlinear_vma = NULL;
1697         details.first_index = hba;
1698         details.last_index = hba + hlen - 1;
1699         if (details.last_index < details.first_index)
1700                 details.last_index = ULONG_MAX;
1701         details.i_mmap_lock = &mapping->i_mmap_lock;
1702
1703         spin_lock(&mapping->i_mmap_lock);
1704
1705         /* serialize i_size write against truncate_count write */
1706         smp_wmb();
1707         /* Protect against page faults, and endless unmapping loops */
1708         mapping->truncate_count++;
1709         /*
1710          * For archs where spin_lock has inclusive semantics like ia64
1711          * this smp_mb() will prevent to read pagetable contents
1712          * before the truncate_count increment is visible to
1713          * other cpus.
1714          */
1715         smp_mb();
1716         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1717                 if (mapping->truncate_count == 0)
1718                         reset_vma_truncate_counts(mapping);
1719                 mapping->truncate_count++;
1720         }
1721         details.truncate_count = mapping->truncate_count;
1722
1723         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1724                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1725         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1726                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1727         spin_unlock(&mapping->i_mmap_lock);
1728 }
1729 EXPORT_SYMBOL(unmap_mapping_range);
1730
1731 /*
1732  * Handle all mappings that got truncated by a "truncate()"
1733  * system call.
1734  *
1735  * NOTE! We have to be ready to update the memory sharing
1736  * between the file and the memory map for a potential last
1737  * incomplete page.  Ugly, but necessary.
1738  */
1739 int vmtruncate(struct inode * inode, loff_t offset)
1740 {
1741         struct address_space *mapping = inode->i_mapping;
1742         unsigned long limit;
1743
1744         if (inode->i_size < offset)
1745                 goto do_expand;
1746         /*
1747          * truncation of in-use swapfiles is disallowed - it would cause
1748          * subsequent swapout to scribble on the now-freed blocks.
1749          */
1750         if (IS_SWAPFILE(inode))
1751                 goto out_busy;
1752         i_size_write(inode, offset);
1753         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1754         truncate_inode_pages(mapping, offset);
1755         goto out_truncate;
1756
1757 do_expand:
1758         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1759         if (limit != RLIM_INFINITY && offset > limit)
1760                 goto out_sig;
1761         if (offset > inode->i_sb->s_maxbytes)
1762                 goto out_big;
1763         i_size_write(inode, offset);
1764
1765 out_truncate:
1766         if (inode->i_op && inode->i_op->truncate)
1767                 inode->i_op->truncate(inode);
1768         return 0;
1769 out_sig:
1770         send_sig(SIGXFSZ, current, 0);
1771 out_big:
1772         return -EFBIG;
1773 out_busy:
1774         return -ETXTBSY;
1775 }
1776 EXPORT_SYMBOL(vmtruncate);
1777
1778 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1779 {
1780         struct address_space *mapping = inode->i_mapping;
1781
1782         /*
1783          * If the underlying filesystem is not going to provide
1784          * a way to truncate a range of blocks (punch a hole) -
1785          * we should return failure right now.
1786          */
1787         if (!inode->i_op || !inode->i_op->truncate_range)
1788                 return -ENOSYS;
1789
1790         mutex_lock(&inode->i_mutex);
1791         down_write(&inode->i_alloc_sem);
1792         unmap_mapping_range(mapping, offset, (end - offset), 1);
1793         truncate_inode_pages_range(mapping, offset, end);
1794         inode->i_op->truncate_range(inode, offset, end);
1795         up_write(&inode->i_alloc_sem);
1796         mutex_unlock(&inode->i_mutex);
1797
1798         return 0;
1799 }
1800 EXPORT_SYMBOL(vmtruncate_range);
1801
1802 /* 
1803  * Primitive swap readahead code. We simply read an aligned block of
1804  * (1 << page_cluster) entries in the swap area. This method is chosen
1805  * because it doesn't cost us any seek time.  We also make sure to queue
1806  * the 'original' request together with the readahead ones...  
1807  *
1808  * This has been extended to use the NUMA policies from the mm triggering
1809  * the readahead.
1810  *
1811  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1812  */
1813 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1814 {
1815 #ifdef CONFIG_NUMA
1816         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1817 #endif
1818         int i, num;
1819         struct page *new_page;
1820         unsigned long offset;
1821
1822         /*
1823          * Get the number of handles we should do readahead io to.
1824          */
1825         num = valid_swaphandles(entry, &offset);
1826         for (i = 0; i < num; offset++, i++) {
1827                 /* Ok, do the async read-ahead now */
1828                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1829                                                            offset), vma, addr);
1830                 if (!new_page)
1831                         break;
1832                 page_cache_release(new_page);
1833 #ifdef CONFIG_NUMA
1834                 /*
1835                  * Find the next applicable VMA for the NUMA policy.
1836                  */
1837                 addr += PAGE_SIZE;
1838                 if (addr == 0)
1839                         vma = NULL;
1840                 if (vma) {
1841                         if (addr >= vma->vm_end) {
1842                                 vma = next_vma;
1843                                 next_vma = vma ? vma->vm_next : NULL;
1844                         }
1845                         if (vma && addr < vma->vm_start)
1846                                 vma = NULL;
1847                 } else {
1848                         if (next_vma && addr >= next_vma->vm_start) {
1849                                 vma = next_vma;
1850                                 next_vma = vma->vm_next;
1851                         }
1852                 }
1853 #endif
1854         }
1855         lru_add_drain();        /* Push any new pages onto the LRU now */
1856 }
1857
1858 /*
1859  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1860  * but allow concurrent faults), and pte mapped but not yet locked.
1861  * We return with mmap_sem still held, but pte unmapped and unlocked.
1862  */
1863 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1864                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1865                 int write_access, pte_t orig_pte)
1866 {
1867         spinlock_t *ptl;
1868         struct page *page;
1869         swp_entry_t entry;
1870         pte_t pte;
1871         int ret = VM_FAULT_MINOR;
1872
1873         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1874                 goto out;
1875
1876         entry = pte_to_swp_entry(orig_pte);
1877 again:
1878         page = lookup_swap_cache(entry);
1879         if (!page) {
1880                 swapin_readahead(entry, address, vma);
1881                 page = read_swap_cache_async(entry, vma, address);
1882                 if (!page) {
1883                         /*
1884                          * Back out if somebody else faulted in this pte
1885                          * while we released the pte lock.
1886                          */
1887                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1888                         if (likely(pte_same(*page_table, orig_pte)))
1889                                 ret = VM_FAULT_OOM;
1890                         goto unlock;
1891                 }
1892
1893                 /* Had to read the page from swap area: Major fault */
1894                 ret = VM_FAULT_MAJOR;
1895                 inc_page_state(pgmajfault);
1896                 grab_swap_token();
1897         }
1898
1899         mark_page_accessed(page);
1900         lock_page(page);
1901         if (!PageSwapCache(page)) {
1902                 /* Page migration has occured */
1903                 unlock_page(page);
1904                 page_cache_release(page);
1905                 goto again;
1906         }
1907
1908         /*
1909          * Back out if somebody else already faulted in this pte.
1910          */
1911         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1912         if (unlikely(!pte_same(*page_table, orig_pte)))
1913                 goto out_nomap;
1914
1915         if (unlikely(!PageUptodate(page))) {
1916                 ret = VM_FAULT_SIGBUS;
1917                 goto out_nomap;
1918         }
1919
1920         /* The page isn't present yet, go ahead with the fault. */
1921
1922         inc_mm_counter(mm, anon_rss);
1923         pte = mk_pte(page, vma->vm_page_prot);
1924         if (write_access && can_share_swap_page(page)) {
1925                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1926                 write_access = 0;
1927         }
1928
1929         flush_icache_page(vma, page);
1930         set_pte_at(mm, address, page_table, pte);
1931         page_add_anon_rmap(page, vma, address);
1932
1933         swap_free(entry);
1934         if (vm_swap_full())
1935                 remove_exclusive_swap_page(page);
1936         unlock_page(page);
1937
1938         if (write_access) {
1939                 if (do_wp_page(mm, vma, address,
1940                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1941                         ret = VM_FAULT_OOM;
1942                 goto out;
1943         }
1944
1945         /* No need to invalidate - it was non-present before */
1946         update_mmu_cache(vma, address, pte);
1947         lazy_mmu_prot_update(pte);
1948 unlock:
1949         pte_unmap_unlock(page_table, ptl);
1950 out:
1951         return ret;
1952 out_nomap:
1953         pte_unmap_unlock(page_table, ptl);
1954         unlock_page(page);
1955         page_cache_release(page);
1956         return ret;
1957 }
1958
1959 /*
1960  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1961  * but allow concurrent faults), and pte mapped but not yet locked.
1962  * We return with mmap_sem still held, but pte unmapped and unlocked.
1963  */
1964 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1965                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1966                 int write_access)
1967 {
1968         struct page *page;
1969         spinlock_t *ptl;
1970         pte_t entry;
1971
1972         if (write_access) {
1973                 /* Allocate our own private page. */
1974                 pte_unmap(page_table);
1975
1976                 if (unlikely(anon_vma_prepare(vma)))
1977                         goto oom;
1978                 page = alloc_zeroed_user_highpage(vma, address);
1979                 if (!page)
1980                         goto oom;
1981
1982                 entry = mk_pte(page, vma->vm_page_prot);
1983                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1984
1985                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1986                 if (!pte_none(*page_table))
1987                         goto release;
1988                 inc_mm_counter(mm, anon_rss);
1989                 lru_cache_add_active(page);
1990                 page_add_new_anon_rmap(page, vma, address);
1991         } else {
1992                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1993                 page = ZERO_PAGE(address);
1994                 page_cache_get(page);
1995                 entry = mk_pte(page, vma->vm_page_prot);
1996
1997                 ptl = pte_lockptr(mm, pmd);
1998                 spin_lock(ptl);
1999                 if (!pte_none(*page_table))
2000                         goto release;
2001                 inc_mm_counter(mm, file_rss);
2002                 page_add_file_rmap(page);
2003         }
2004
2005         set_pte_at(mm, address, page_table, entry);
2006
2007         /* No need to invalidate - it was non-present before */
2008         update_mmu_cache(vma, address, entry);
2009         lazy_mmu_prot_update(entry);
2010 unlock:
2011         pte_unmap_unlock(page_table, ptl);
2012         return VM_FAULT_MINOR;
2013 release:
2014         page_cache_release(page);
2015         goto unlock;
2016 oom:
2017         return VM_FAULT_OOM;
2018 }
2019
2020 /*
2021  * do_no_page() tries to create a new page mapping. It aggressively
2022  * tries to share with existing pages, but makes a separate copy if
2023  * the "write_access" parameter is true in order to avoid the next
2024  * page fault.
2025  *
2026  * As this is called only for pages that do not currently exist, we
2027  * do not need to flush old virtual caches or the TLB.
2028  *
2029  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2030  * but allow concurrent faults), and pte mapped but not yet locked.
2031  * We return with mmap_sem still held, but pte unmapped and unlocked.
2032  */
2033 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2034                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2035                 int write_access)
2036 {
2037         spinlock_t *ptl;
2038         struct page *new_page;
2039         struct address_space *mapping = NULL;
2040         pte_t entry;
2041         unsigned int sequence = 0;
2042         int ret = VM_FAULT_MINOR;
2043         int anon = 0;
2044
2045         pte_unmap(page_table);
2046         BUG_ON(vma->vm_flags & VM_PFNMAP);
2047
2048         if (vma->vm_file) {
2049                 mapping = vma->vm_file->f_mapping;
2050                 sequence = mapping->truncate_count;
2051                 smp_rmb(); /* serializes i_size against truncate_count */
2052         }
2053 retry:
2054         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2055         /*
2056          * No smp_rmb is needed here as long as there's a full
2057          * spin_lock/unlock sequence inside the ->nopage callback
2058          * (for the pagecache lookup) that acts as an implicit
2059          * smp_mb() and prevents the i_size read to happen
2060          * after the next truncate_count read.
2061          */
2062
2063         /* no page was available -- either SIGBUS or OOM */
2064         if (new_page == NOPAGE_SIGBUS)
2065                 return VM_FAULT_SIGBUS;
2066         if (new_page == NOPAGE_OOM)
2067                 return VM_FAULT_OOM;
2068
2069         /*
2070          * Should we do an early C-O-W break?
2071          */
2072         if (write_access && !(vma->vm_flags & VM_SHARED)) {
2073                 struct page *page;
2074
2075                 if (unlikely(anon_vma_prepare(vma)))
2076                         goto oom;
2077                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2078                 if (!page)
2079                         goto oom;
2080                 copy_user_highpage(page, new_page, address);
2081                 page_cache_release(new_page);
2082                 new_page = page;
2083                 anon = 1;
2084         }
2085
2086         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2087         /*
2088          * For a file-backed vma, someone could have truncated or otherwise
2089          * invalidated this page.  If unmap_mapping_range got called,
2090          * retry getting the page.
2091          */
2092         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2093                 pte_unmap_unlock(page_table, ptl);
2094                 page_cache_release(new_page);
2095                 cond_resched();
2096                 sequence = mapping->truncate_count;
2097                 smp_rmb();
2098                 goto retry;
2099         }
2100
2101         /*
2102          * This silly early PAGE_DIRTY setting removes a race
2103          * due to the bad i386 page protection. But it's valid
2104          * for other architectures too.
2105          *
2106          * Note that if write_access is true, we either now have
2107          * an exclusive copy of the page, or this is a shared mapping,
2108          * so we can make it writable and dirty to avoid having to
2109          * handle that later.
2110          */
2111         /* Only go through if we didn't race with anybody else... */
2112         if (pte_none(*page_table)) {
2113                 flush_icache_page(vma, new_page);
2114                 entry = mk_pte(new_page, vma->vm_page_prot);
2115                 if (write_access)
2116                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2117                 set_pte_at(mm, address, page_table, entry);
2118                 if (anon) {
2119                         inc_mm_counter(mm, anon_rss);
2120                         lru_cache_add_active(new_page);
2121                         page_add_new_anon_rmap(new_page, vma, address);
2122                 } else {
2123                         inc_mm_counter(mm, file_rss);
2124                         page_add_file_rmap(new_page);
2125                 }
2126         } else {
2127                 /* One of our sibling threads was faster, back out. */
2128                 page_cache_release(new_page);
2129                 goto unlock;
2130         }
2131
2132         /* no need to invalidate: a not-present page shouldn't be cached */
2133         update_mmu_cache(vma, address, entry);
2134         lazy_mmu_prot_update(entry);
2135 unlock:
2136         pte_unmap_unlock(page_table, ptl);
2137         return ret;
2138 oom:
2139         page_cache_release(new_page);
2140         return VM_FAULT_OOM;
2141 }
2142
2143 /*
2144  * Fault of a previously existing named mapping. Repopulate the pte
2145  * from the encoded file_pte if possible. This enables swappable
2146  * nonlinear vmas.
2147  *
2148  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2149  * but allow concurrent faults), and pte mapped but not yet locked.
2150  * We return with mmap_sem still held, but pte unmapped and unlocked.
2151  */
2152 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2153                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2154                 int write_access, pte_t orig_pte)
2155 {
2156         pgoff_t pgoff;
2157         int err;
2158
2159         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2160                 return VM_FAULT_MINOR;
2161
2162         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2163                 /*
2164                  * Page table corrupted: show pte and kill process.
2165                  */
2166                 print_bad_pte(vma, orig_pte, address);
2167                 return VM_FAULT_OOM;
2168         }
2169         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2170
2171         pgoff = pte_to_pgoff(orig_pte);
2172         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2173                                         vma->vm_page_prot, pgoff, 0);
2174         if (err == -ENOMEM)
2175                 return VM_FAULT_OOM;
2176         if (err)
2177                 return VM_FAULT_SIGBUS;
2178         return VM_FAULT_MAJOR;
2179 }
2180
2181 /*
2182  * These routines also need to handle stuff like marking pages dirty
2183  * and/or accessed for architectures that don't do it in hardware (most
2184  * RISC architectures).  The early dirtying is also good on the i386.
2185  *
2186  * There is also a hook called "update_mmu_cache()" that architectures
2187  * with external mmu caches can use to update those (ie the Sparc or
2188  * PowerPC hashed page tables that act as extended TLBs).
2189  *
2190  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2191  * but allow concurrent faults), and pte mapped but not yet locked.
2192  * We return with mmap_sem still held, but pte unmapped and unlocked.
2193  */
2194 static inline int handle_pte_fault(struct mm_struct *mm,
2195                 struct vm_area_struct *vma, unsigned long address,
2196                 pte_t *pte, pmd_t *pmd, int write_access)
2197 {
2198         pte_t entry;
2199         pte_t old_entry;
2200         spinlock_t *ptl;
2201
2202         old_entry = entry = *pte;
2203         if (!pte_present(entry)) {
2204                 if (pte_none(entry)) {
2205                         if (!vma->vm_ops || !vma->vm_ops->nopage)
2206                                 return do_anonymous_page(mm, vma, address,
2207                                         pte, pmd, write_access);
2208                         return do_no_page(mm, vma, address,
2209                                         pte, pmd, write_access);
2210                 }
2211                 if (pte_file(entry))
2212                         return do_file_page(mm, vma, address,
2213                                         pte, pmd, write_access, entry);
2214                 return do_swap_page(mm, vma, address,
2215                                         pte, pmd, write_access, entry);
2216         }
2217
2218         ptl = pte_lockptr(mm, pmd);
2219         spin_lock(ptl);
2220         if (unlikely(!pte_same(*pte, entry)))
2221                 goto unlock;
2222         if (write_access) {
2223                 if (!pte_write(entry))
2224                         return do_wp_page(mm, vma, address,
2225                                         pte, pmd, ptl, entry);
2226                 entry = pte_mkdirty(entry);
2227         }
2228         entry = pte_mkyoung(entry);
2229         if (!pte_same(old_entry, entry)) {
2230                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2231                 update_mmu_cache(vma, address, entry);
2232                 lazy_mmu_prot_update(entry);
2233         } else {
2234                 /*
2235                  * This is needed only for protection faults but the arch code
2236                  * is not yet telling us if this is a protection fault or not.
2237                  * This still avoids useless tlb flushes for .text page faults
2238                  * with threads.
2239                  */
2240                 if (write_access)
2241                         flush_tlb_page(vma, address);
2242         }
2243 unlock:
2244         pte_unmap_unlock(pte, ptl);
2245         return VM_FAULT_MINOR;
2246 }
2247
2248 /*
2249  * By the time we get here, we already hold the mm semaphore
2250  */
2251 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2252                 unsigned long address, int write_access)
2253 {
2254         pgd_t *pgd;
2255         pud_t *pud;
2256         pmd_t *pmd;
2257         pte_t *pte;
2258
2259         __set_current_state(TASK_RUNNING);
2260
2261         inc_page_state(pgfault);
2262
2263         if (unlikely(is_vm_hugetlb_page(vma)))
2264                 return hugetlb_fault(mm, vma, address, write_access);
2265
2266         pgd = pgd_offset(mm, address);
2267         pud = pud_alloc(mm, pgd, address);
2268         if (!pud)
2269                 return VM_FAULT_OOM;
2270         pmd = pmd_alloc(mm, pud, address);
2271         if (!pmd)
2272                 return VM_FAULT_OOM;
2273         pte = pte_alloc_map(mm, pmd, address);
2274         if (!pte)
2275                 return VM_FAULT_OOM;
2276
2277         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2278 }
2279
2280 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2281
2282 #ifndef __PAGETABLE_PUD_FOLDED
2283 /*
2284  * Allocate page upper directory.
2285  * We've already handled the fast-path in-line.
2286  */
2287 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2288 {
2289         pud_t *new = pud_alloc_one(mm, address);
2290         if (!new)
2291                 return -ENOMEM;
2292
2293         spin_lock(&mm->page_table_lock);
2294         if (pgd_present(*pgd))          /* Another has populated it */
2295                 pud_free(new);
2296         else
2297                 pgd_populate(mm, pgd, new);
2298         spin_unlock(&mm->page_table_lock);
2299         return 0;
2300 }
2301 #else
2302 /* Workaround for gcc 2.96 */
2303 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2304 {
2305         return 0;
2306 }
2307 #endif /* __PAGETABLE_PUD_FOLDED */
2308
2309 #ifndef __PAGETABLE_PMD_FOLDED
2310 /*
2311  * Allocate page middle directory.
2312  * We've already handled the fast-path in-line.
2313  */
2314 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2315 {
2316         pmd_t *new = pmd_alloc_one(mm, address);
2317         if (!new)
2318                 return -ENOMEM;
2319
2320         spin_lock(&mm->page_table_lock);
2321 #ifndef __ARCH_HAS_4LEVEL_HACK
2322         if (pud_present(*pud))          /* Another has populated it */
2323                 pmd_free(new);
2324         else
2325                 pud_populate(mm, pud, new);
2326 #else
2327         if (pgd_present(*pud))          /* Another has populated it */
2328                 pmd_free(new);
2329         else
2330                 pgd_populate(mm, pud, new);
2331 #endif /* __ARCH_HAS_4LEVEL_HACK */
2332         spin_unlock(&mm->page_table_lock);
2333         return 0;
2334 }
2335 #else
2336 /* Workaround for gcc 2.96 */
2337 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2338 {
2339         return 0;
2340 }
2341 #endif /* __PAGETABLE_PMD_FOLDED */
2342
2343 int make_pages_present(unsigned long addr, unsigned long end)
2344 {
2345         int ret, len, write;
2346         struct vm_area_struct * vma;
2347
2348         vma = find_vma(current->mm, addr);
2349         if (!vma)
2350                 return -1;
2351         write = (vma->vm_flags & VM_WRITE) != 0;
2352         if (addr >= end)
2353                 BUG();
2354         if (end > vma->vm_end)
2355                 BUG();
2356         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2357         ret = get_user_pages(current, current->mm, addr,
2358                         len, write, 0, NULL, NULL);
2359         if (ret < 0)
2360                 return ret;
2361         return ret == len ? 0 : -1;
2362 }
2363
2364 /* 
2365  * Map a vmalloc()-space virtual address to the physical page.
2366  */
2367 struct page * vmalloc_to_page(void * vmalloc_addr)
2368 {
2369         unsigned long addr = (unsigned long) vmalloc_addr;
2370         struct page *page = NULL;
2371         pgd_t *pgd = pgd_offset_k(addr);
2372         pud_t *pud;
2373         pmd_t *pmd;
2374         pte_t *ptep, pte;
2375   
2376         if (!pgd_none(*pgd)) {
2377                 pud = pud_offset(pgd, addr);
2378                 if (!pud_none(*pud)) {
2379                         pmd = pmd_offset(pud, addr);
2380                         if (!pmd_none(*pmd)) {
2381                                 ptep = pte_offset_map(pmd, addr);
2382                                 pte = *ptep;
2383                                 if (pte_present(pte))
2384                                         page = pte_page(pte);
2385                                 pte_unmap(ptep);
2386                         }
2387                 }
2388         }
2389         return page;
2390 }
2391
2392 EXPORT_SYMBOL(vmalloc_to_page);
2393
2394 /*
2395  * Map a vmalloc()-space virtual address to the physical page frame number.
2396  */
2397 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2398 {
2399         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2400 }
2401
2402 EXPORT_SYMBOL(vmalloc_to_pfn);
2403
2404 #if !defined(__HAVE_ARCH_GATE_AREA)
2405
2406 #if defined(AT_SYSINFO_EHDR)
2407 static struct vm_area_struct gate_vma;
2408
2409 static int __init gate_vma_init(void)
2410 {
2411         gate_vma.vm_mm = NULL;
2412         gate_vma.vm_start = FIXADDR_USER_START;
2413         gate_vma.vm_end = FIXADDR_USER_END;
2414         gate_vma.vm_page_prot = PAGE_READONLY;
2415         gate_vma.vm_flags = 0;
2416         return 0;
2417 }
2418 __initcall(gate_vma_init);
2419 #endif
2420
2421 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2422 {
2423 #ifdef AT_SYSINFO_EHDR
2424         return &gate_vma;
2425 #else
2426         return NULL;
2427 #endif
2428 }
2429
2430 int in_gate_area_no_task(unsigned long addr)
2431 {
2432 #ifdef AT_SYSINFO_EHDR
2433         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2434                 return 1;
2435 #endif
2436         return 0;
2437 }
2438
2439 #endif  /* __HAVE_ARCH_GATE_AREA */