109e9866237ec7992daa1e7c441e07dbfe976151
[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/delayacct.h>
51 #include <linux/init.h>
52
53 #include <asm/pgalloc.h>
54 #include <asm/uaccess.h>
55 #include <asm/tlb.h>
56 #include <asm/tlbflush.h>
57 #include <asm/pgtable.h>
58
59 #include <linux/swapops.h>
60 #include <linux/elf.h>
61
62 #ifndef CONFIG_NEED_MULTIPLE_NODES
63 /* use the per-pgdat data instead for discontigmem - mbligh */
64 unsigned long max_mapnr;
65 struct page *mem_map;
66
67 EXPORT_SYMBOL(max_mapnr);
68 EXPORT_SYMBOL(mem_map);
69 #endif
70
71 unsigned long num_physpages;
72 /*
73  * A number of key systems in x86 including ioremap() rely on the assumption
74  * that high_memory defines the upper bound on direct map memory, then end
75  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
76  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
77  * and ZONE_HIGHMEM.
78  */
79 void * high_memory;
80 unsigned long vmalloc_earlyreserve;
81
82 EXPORT_SYMBOL(num_physpages);
83 EXPORT_SYMBOL(high_memory);
84 EXPORT_SYMBOL(vmalloc_earlyreserve);
85
86 int randomize_va_space __read_mostly = 1;
87
88 static int __init disable_randmaps(char *s)
89 {
90         randomize_va_space = 0;
91         return 1;
92 }
93 __setup("norandmaps", disable_randmaps);
94
95
96 /*
97  * If a p?d_bad entry is found while walking page tables, report
98  * the error, before resetting entry to p?d_none.  Usually (but
99  * very seldom) called out from the p?d_none_or_clear_bad macros.
100  */
101
102 void pgd_clear_bad(pgd_t *pgd)
103 {
104         pgd_ERROR(*pgd);
105         pgd_clear(pgd);
106 }
107
108 void pud_clear_bad(pud_t *pud)
109 {
110         pud_ERROR(*pud);
111         pud_clear(pud);
112 }
113
114 void pmd_clear_bad(pmd_t *pmd)
115 {
116         pmd_ERROR(*pmd);
117         pmd_clear(pmd);
118 }
119
120 /*
121  * Note: this doesn't free the actual pages themselves. That
122  * has been handled earlier when unmapping all the memory regions.
123  */
124 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
125 {
126         struct page *page = pmd_page(*pmd);
127         pmd_clear(pmd);
128         pte_lock_deinit(page);
129         pte_free_tlb(tlb, page);
130         dec_zone_page_state(page, NR_PAGETABLE);
131         tlb->mm->nr_ptes--;
132 }
133
134 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
135                                 unsigned long addr, unsigned long end,
136                                 unsigned long floor, unsigned long ceiling)
137 {
138         pmd_t *pmd;
139         unsigned long next;
140         unsigned long start;
141
142         start = addr;
143         pmd = pmd_offset(pud, addr);
144         do {
145                 next = pmd_addr_end(addr, end);
146                 if (pmd_none_or_clear_bad(pmd))
147                         continue;
148                 free_pte_range(tlb, pmd);
149         } while (pmd++, addr = next, addr != end);
150
151         start &= PUD_MASK;
152         if (start < floor)
153                 return;
154         if (ceiling) {
155                 ceiling &= PUD_MASK;
156                 if (!ceiling)
157                         return;
158         }
159         if (end - 1 > ceiling - 1)
160                 return;
161
162         pmd = pmd_offset(pud, start);
163         pud_clear(pud);
164         pmd_free_tlb(tlb, pmd);
165 }
166
167 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
168                                 unsigned long addr, unsigned long end,
169                                 unsigned long floor, unsigned long ceiling)
170 {
171         pud_t *pud;
172         unsigned long next;
173         unsigned long start;
174
175         start = addr;
176         pud = pud_offset(pgd, addr);
177         do {
178                 next = pud_addr_end(addr, end);
179                 if (pud_none_or_clear_bad(pud))
180                         continue;
181                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
182         } while (pud++, addr = next, addr != end);
183
184         start &= PGDIR_MASK;
185         if (start < floor)
186                 return;
187         if (ceiling) {
188                 ceiling &= PGDIR_MASK;
189                 if (!ceiling)
190                         return;
191         }
192         if (end - 1 > ceiling - 1)
193                 return;
194
195         pud = pud_offset(pgd, start);
196         pgd_clear(pgd);
197         pud_free_tlb(tlb, pud);
198 }
199
200 /*
201  * This function frees user-level page tables of a process.
202  *
203  * Must be called with pagetable lock held.
204  */
205 void free_pgd_range(struct mmu_gather **tlb,
206                         unsigned long addr, unsigned long end,
207                         unsigned long floor, unsigned long ceiling)
208 {
209         pgd_t *pgd;
210         unsigned long next;
211         unsigned long start;
212
213         /*
214          * The next few lines have given us lots of grief...
215          *
216          * Why are we testing PMD* at this top level?  Because often
217          * there will be no work to do at all, and we'd prefer not to
218          * go all the way down to the bottom just to discover that.
219          *
220          * Why all these "- 1"s?  Because 0 represents both the bottom
221          * of the address space and the top of it (using -1 for the
222          * top wouldn't help much: the masks would do the wrong thing).
223          * The rule is that addr 0 and floor 0 refer to the bottom of
224          * the address space, but end 0 and ceiling 0 refer to the top
225          * Comparisons need to use "end - 1" and "ceiling - 1" (though
226          * that end 0 case should be mythical).
227          *
228          * Wherever addr is brought up or ceiling brought down, we must
229          * be careful to reject "the opposite 0" before it confuses the
230          * subsequent tests.  But what about where end is brought down
231          * by PMD_SIZE below? no, end can't go down to 0 there.
232          *
233          * Whereas we round start (addr) and ceiling down, by different
234          * masks at different levels, in order to test whether a table
235          * now has no other vmas using it, so can be freed, we don't
236          * bother to round floor or end up - the tests don't need that.
237          */
238
239         addr &= PMD_MASK;
240         if (addr < floor) {
241                 addr += PMD_SIZE;
242                 if (!addr)
243                         return;
244         }
245         if (ceiling) {
246                 ceiling &= PMD_MASK;
247                 if (!ceiling)
248                         return;
249         }
250         if (end - 1 > ceiling - 1)
251                 end -= PMD_SIZE;
252         if (addr > end - 1)
253                 return;
254
255         start = addr;
256         pgd = pgd_offset((*tlb)->mm, addr);
257         do {
258                 next = pgd_addr_end(addr, end);
259                 if (pgd_none_or_clear_bad(pgd))
260                         continue;
261                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
262         } while (pgd++, addr = next, addr != end);
263
264         if (!(*tlb)->fullmm)
265                 flush_tlb_pgtables((*tlb)->mm, start, end);
266 }
267
268 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
269                 unsigned long floor, unsigned long ceiling)
270 {
271         while (vma) {
272                 struct vm_area_struct *next = vma->vm_next;
273                 unsigned long addr = vma->vm_start;
274
275                 /*
276                  * Hide vma from rmap and vmtruncate before freeing pgtables
277                  */
278                 anon_vma_unlink(vma);
279                 unlink_file_vma(vma);
280
281                 if (is_vm_hugetlb_page(vma)) {
282                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
283                                 floor, next? next->vm_start: ceiling);
284                 } else {
285                         /*
286                          * Optimization: gather nearby vmas into one call down
287                          */
288                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
289                                && !is_vm_hugetlb_page(next)) {
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_zone_page_state(new, NR_PAGETABLE);
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         /*
400          * Add some anal sanity checks for now. Eventually,
401          * we should just do "return pfn_to_page(pfn)", but
402          * in the meantime we check that we get a valid pfn,
403          * and that the resulting page looks ok.
404          */
405         if (unlikely(!pfn_valid(pfn))) {
406                 print_bad_pte(vma, pte, addr);
407                 return NULL;
408         }
409
410         /*
411          * NOTE! We still have PageReserved() pages in the page 
412          * tables. 
413          *
414          * The PAGE_ZERO() pages and various VDSO mappings can
415          * cause them to exist.
416          */
417         return pfn_to_page(pfn);
418 }
419
420 /*
421  * copy one vm_area from one task to the other. Assumes the page tables
422  * already present in the new task to be cleared in the whole range
423  * covered by this vma.
424  */
425
426 static inline void
427 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
428                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
429                 unsigned long addr, int *rss)
430 {
431         unsigned long vm_flags = vma->vm_flags;
432         pte_t pte = *src_pte;
433         struct page *page;
434
435         /* pte contains position in swap or file, so copy. */
436         if (unlikely(!pte_present(pte))) {
437                 if (!pte_file(pte)) {
438                         swp_entry_t entry = pte_to_swp_entry(pte);
439
440                         swap_duplicate(entry);
441                         /* make sure dst_mm is on swapoff's mmlist. */
442                         if (unlikely(list_empty(&dst_mm->mmlist))) {
443                                 spin_lock(&mmlist_lock);
444                                 if (list_empty(&dst_mm->mmlist))
445                                         list_add(&dst_mm->mmlist,
446                                                  &src_mm->mmlist);
447                                 spin_unlock(&mmlist_lock);
448                         }
449                         if (is_write_migration_entry(entry) &&
450                                         is_cow_mapping(vm_flags)) {
451                                 /*
452                                  * COW mappings require pages in both parent
453                                  * and child to be set to read.
454                                  */
455                                 make_migration_entry_read(&entry);
456                                 pte = swp_entry_to_pte(entry);
457                                 set_pte_at(src_mm, addr, src_pte, pte);
458                         }
459                 }
460                 goto out_set_pte;
461         }
462
463         /*
464          * If it's a COW mapping, write protect it both
465          * in the parent and the child
466          */
467         if (is_cow_mapping(vm_flags)) {
468                 ptep_set_wrprotect(src_mm, addr, src_pte);
469                 pte = *src_pte;
470         }
471
472         /*
473          * If it's a shared mapping, mark it clean in
474          * the child
475          */
476         if (vm_flags & VM_SHARED)
477                 pte = pte_mkclean(pte);
478         pte = pte_mkold(pte);
479
480         page = vm_normal_page(vma, addr, pte);
481         if (page) {
482                 get_page(page);
483                 page_dup_rmap(page);
484                 rss[!!PageAnon(page)]++;
485         }
486
487 out_set_pte:
488         set_pte_at(dst_mm, addr, dst_pte, pte);
489 }
490
491 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
492                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
493                 unsigned long addr, unsigned long end)
494 {
495         pte_t *src_pte, *dst_pte;
496         spinlock_t *src_ptl, *dst_ptl;
497         int progress = 0;
498         int rss[2];
499
500 again:
501         rss[1] = rss[0] = 0;
502         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
503         if (!dst_pte)
504                 return -ENOMEM;
505         src_pte = pte_offset_map_nested(src_pmd, addr);
506         src_ptl = pte_lockptr(src_mm, src_pmd);
507         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
508
509         do {
510                 /*
511                  * We are holding two locks at this point - either of them
512                  * could generate latencies in another task on another CPU.
513                  */
514                 if (progress >= 32) {
515                         progress = 0;
516                         if (need_resched() ||
517                             need_lockbreak(src_ptl) ||
518                             need_lockbreak(dst_ptl))
519                                 break;
520                 }
521                 if (pte_none(*src_pte)) {
522                         progress++;
523                         continue;
524                 }
525                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
526                 progress += 8;
527         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
528
529         spin_unlock(src_ptl);
530         pte_unmap_nested(src_pte - 1);
531         add_mm_rss(dst_mm, rss[0], rss[1]);
532         pte_unmap_unlock(dst_pte - 1, dst_ptl);
533         cond_resched();
534         if (addr != end)
535                 goto again;
536         return 0;
537 }
538
539 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
540                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
541                 unsigned long addr, unsigned long end)
542 {
543         pmd_t *src_pmd, *dst_pmd;
544         unsigned long next;
545
546         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
547         if (!dst_pmd)
548                 return -ENOMEM;
549         src_pmd = pmd_offset(src_pud, addr);
550         do {
551                 next = pmd_addr_end(addr, end);
552                 if (pmd_none_or_clear_bad(src_pmd))
553                         continue;
554                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
555                                                 vma, addr, next))
556                         return -ENOMEM;
557         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
558         return 0;
559 }
560
561 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
562                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
563                 unsigned long addr, unsigned long end)
564 {
565         pud_t *src_pud, *dst_pud;
566         unsigned long next;
567
568         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
569         if (!dst_pud)
570                 return -ENOMEM;
571         src_pud = pud_offset(src_pgd, addr);
572         do {
573                 next = pud_addr_end(addr, end);
574                 if (pud_none_or_clear_bad(src_pud))
575                         continue;
576                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
577                                                 vma, addr, next))
578                         return -ENOMEM;
579         } while (dst_pud++, src_pud++, addr = next, addr != end);
580         return 0;
581 }
582
583 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
584                 struct vm_area_struct *vma)
585 {
586         pgd_t *src_pgd, *dst_pgd;
587         unsigned long next;
588         unsigned long addr = vma->vm_start;
589         unsigned long end = vma->vm_end;
590
591         /*
592          * Don't copy ptes where a page fault will fill them correctly.
593          * Fork becomes much lighter when there are big shared or private
594          * readonly mappings. The tradeoff is that copy_page_range is more
595          * efficient than faulting.
596          */
597         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
598                 if (!vma->anon_vma)
599                         return 0;
600         }
601
602         if (is_vm_hugetlb_page(vma))
603                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
604
605         dst_pgd = pgd_offset(dst_mm, addr);
606         src_pgd = pgd_offset(src_mm, addr);
607         do {
608                 next = pgd_addr_end(addr, end);
609                 if (pgd_none_or_clear_bad(src_pgd))
610                         continue;
611                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
612                                                 vma, addr, next))
613                         return -ENOMEM;
614         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
615         return 0;
616 }
617
618 static unsigned long zap_pte_range(struct mmu_gather *tlb,
619                                 struct vm_area_struct *vma, pmd_t *pmd,
620                                 unsigned long addr, unsigned long end,
621                                 long *zap_work, struct zap_details *details)
622 {
623         struct mm_struct *mm = tlb->mm;
624         pte_t *pte;
625         spinlock_t *ptl;
626         int file_rss = 0;
627         int anon_rss = 0;
628
629         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
630         do {
631                 pte_t ptent = *pte;
632                 if (pte_none(ptent)) {
633                         (*zap_work)--;
634                         continue;
635                 }
636
637                 (*zap_work) -= PAGE_SIZE;
638
639                 if (pte_present(ptent)) {
640                         struct page *page;
641
642                         page = vm_normal_page(vma, addr, ptent);
643                         if (unlikely(details) && page) {
644                                 /*
645                                  * unmap_shared_mapping_pages() wants to
646                                  * invalidate cache without truncating:
647                                  * unmap shared but keep private pages.
648                                  */
649                                 if (details->check_mapping &&
650                                     details->check_mapping != page->mapping)
651                                         continue;
652                                 /*
653                                  * Each page->index must be checked when
654                                  * invalidating or truncating nonlinear.
655                                  */
656                                 if (details->nonlinear_vma &&
657                                     (page->index < details->first_index ||
658                                      page->index > details->last_index))
659                                         continue;
660                         }
661                         ptent = ptep_get_and_clear_full(mm, addr, pte,
662                                                         tlb->fullmm);
663                         tlb_remove_tlb_entry(tlb, pte, addr);
664                         if (unlikely(!page))
665                                 continue;
666                         if (unlikely(details) && details->nonlinear_vma
667                             && linear_page_index(details->nonlinear_vma,
668                                                 addr) != page->index)
669                                 set_pte_at(mm, addr, pte,
670                                            pgoff_to_pte(page->index));
671                         if (PageAnon(page))
672                                 anon_rss--;
673                         else {
674                                 if (pte_dirty(ptent))
675                                         set_page_dirty(page);
676                                 if (pte_young(ptent))
677                                         mark_page_accessed(page);
678                                 file_rss--;
679                         }
680                         page_remove_rmap(page);
681                         tlb_remove_page(tlb, page);
682                         continue;
683                 }
684                 /*
685                  * If details->check_mapping, we leave swap entries;
686                  * if details->nonlinear_vma, we leave file entries.
687                  */
688                 if (unlikely(details))
689                         continue;
690                 if (!pte_file(ptent))
691                         free_swap_and_cache(pte_to_swp_entry(ptent));
692                 pte_clear_full(mm, addr, pte, tlb->fullmm);
693         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
694
695         add_mm_rss(mm, file_rss, anon_rss);
696         pte_unmap_unlock(pte - 1, ptl);
697
698         return addr;
699 }
700
701 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
702                                 struct vm_area_struct *vma, pud_t *pud,
703                                 unsigned long addr, unsigned long end,
704                                 long *zap_work, struct zap_details *details)
705 {
706         pmd_t *pmd;
707         unsigned long next;
708
709         pmd = pmd_offset(pud, addr);
710         do {
711                 next = pmd_addr_end(addr, end);
712                 if (pmd_none_or_clear_bad(pmd)) {
713                         (*zap_work)--;
714                         continue;
715                 }
716                 next = zap_pte_range(tlb, vma, pmd, addr, next,
717                                                 zap_work, details);
718         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
719
720         return addr;
721 }
722
723 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
724                                 struct vm_area_struct *vma, pgd_t *pgd,
725                                 unsigned long addr, unsigned long end,
726                                 long *zap_work, struct zap_details *details)
727 {
728         pud_t *pud;
729         unsigned long next;
730
731         pud = pud_offset(pgd, addr);
732         do {
733                 next = pud_addr_end(addr, end);
734                 if (pud_none_or_clear_bad(pud)) {
735                         (*zap_work)--;
736                         continue;
737                 }
738                 next = zap_pmd_range(tlb, vma, pud, addr, next,
739                                                 zap_work, details);
740         } while (pud++, addr = next, (addr != end && *zap_work > 0));
741
742         return addr;
743 }
744
745 static unsigned long unmap_page_range(struct mmu_gather *tlb,
746                                 struct vm_area_struct *vma,
747                                 unsigned long addr, unsigned long end,
748                                 long *zap_work, struct zap_details *details)
749 {
750         pgd_t *pgd;
751         unsigned long next;
752
753         if (details && !details->check_mapping && !details->nonlinear_vma)
754                 details = NULL;
755
756         BUG_ON(addr >= end);
757         tlb_start_vma(tlb, vma);
758         pgd = pgd_offset(vma->vm_mm, addr);
759         do {
760                 next = pgd_addr_end(addr, end);
761                 if (pgd_none_or_clear_bad(pgd)) {
762                         (*zap_work)--;
763                         continue;
764                 }
765                 next = zap_pud_range(tlb, vma, pgd, addr, next,
766                                                 zap_work, details);
767         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
768         tlb_end_vma(tlb, vma);
769
770         return addr;
771 }
772
773 #ifdef CONFIG_PREEMPT
774 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
775 #else
776 /* No preempt: go for improved straight-line efficiency */
777 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
778 #endif
779
780 /**
781  * unmap_vmas - unmap a range of memory covered by a list of vma's
782  * @tlbp: address of the caller's struct mmu_gather
783  * @vma: the starting vma
784  * @start_addr: virtual address at which to start unmapping
785  * @end_addr: virtual address at which to end unmapping
786  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
787  * @details: details of nonlinear truncation or shared cache invalidation
788  *
789  * Returns the end address of the unmapping (restart addr if interrupted).
790  *
791  * Unmap all pages in the vma list.
792  *
793  * We aim to not hold locks for too long (for scheduling latency reasons).
794  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
795  * return the ending mmu_gather to the caller.
796  *
797  * Only addresses between `start' and `end' will be unmapped.
798  *
799  * The VMA list must be sorted in ascending virtual address order.
800  *
801  * unmap_vmas() assumes that the caller will flush the whole unmapped address
802  * range after unmap_vmas() returns.  So the only responsibility here is to
803  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
804  * drops the lock and schedules.
805  */
806 unsigned long unmap_vmas(struct mmu_gather **tlbp,
807                 struct vm_area_struct *vma, unsigned long start_addr,
808                 unsigned long end_addr, unsigned long *nr_accounted,
809                 struct zap_details *details)
810 {
811         long zap_work = ZAP_BLOCK_SIZE;
812         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
813         int tlb_start_valid = 0;
814         unsigned long start = start_addr;
815         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
816         int fullmm = (*tlbp)->fullmm;
817
818         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
819                 unsigned long end;
820
821                 start = max(vma->vm_start, start_addr);
822                 if (start >= vma->vm_end)
823                         continue;
824                 end = min(vma->vm_end, end_addr);
825                 if (end <= vma->vm_start)
826                         continue;
827
828                 if (vma->vm_flags & VM_ACCOUNT)
829                         *nr_accounted += (end - start) >> PAGE_SHIFT;
830
831                 while (start != end) {
832                         if (!tlb_start_valid) {
833                                 tlb_start = start;
834                                 tlb_start_valid = 1;
835                         }
836
837                         if (unlikely(is_vm_hugetlb_page(vma))) {
838                                 unmap_hugepage_range(vma, start, end);
839                                 zap_work -= (end - start) /
840                                                 (HPAGE_SIZE / PAGE_SIZE);
841                                 start = end;
842                         } else
843                                 start = unmap_page_range(*tlbp, vma,
844                                                 start, end, &zap_work, details);
845
846                         if (zap_work > 0) {
847                                 BUG_ON(start != end);
848                                 break;
849                         }
850
851                         tlb_finish_mmu(*tlbp, tlb_start, start);
852
853                         if (need_resched() ||
854                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
855                                 if (i_mmap_lock) {
856                                         *tlbp = NULL;
857                                         goto out;
858                                 }
859                                 cond_resched();
860                         }
861
862                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
863                         tlb_start_valid = 0;
864                         zap_work = ZAP_BLOCK_SIZE;
865                 }
866         }
867 out:
868         return start;   /* which is now the end (or restart) address */
869 }
870
871 /**
872  * zap_page_range - remove user pages in a given range
873  * @vma: vm_area_struct holding the applicable pages
874  * @address: starting address of pages to zap
875  * @size: number of bytes to zap
876  * @details: details of nonlinear truncation or shared cache invalidation
877  */
878 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
879                 unsigned long size, struct zap_details *details)
880 {
881         struct mm_struct *mm = vma->vm_mm;
882         struct mmu_gather *tlb;
883         unsigned long end = address + size;
884         unsigned long nr_accounted = 0;
885
886         lru_add_drain();
887         tlb = tlb_gather_mmu(mm, 0);
888         update_hiwater_rss(mm);
889         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
890         if (tlb)
891                 tlb_finish_mmu(tlb, address, end);
892         return end;
893 }
894
895 /*
896  * Do a quick page-table lookup for a single page.
897  */
898 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
899                         unsigned int flags)
900 {
901         pgd_t *pgd;
902         pud_t *pud;
903         pmd_t *pmd;
904         pte_t *ptep, pte;
905         spinlock_t *ptl;
906         struct page *page;
907         struct mm_struct *mm = vma->vm_mm;
908
909         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
910         if (!IS_ERR(page)) {
911                 BUG_ON(flags & FOLL_GET);
912                 goto out;
913         }
914
915         page = NULL;
916         pgd = pgd_offset(mm, address);
917         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
918                 goto no_page_table;
919
920         pud = pud_offset(pgd, address);
921         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
922                 goto no_page_table;
923         
924         pmd = pmd_offset(pud, address);
925         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
926                 goto no_page_table;
927
928         if (pmd_huge(*pmd)) {
929                 BUG_ON(flags & FOLL_GET);
930                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
931                 goto out;
932         }
933
934         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
935         if (!ptep)
936                 goto out;
937
938         pte = *ptep;
939         if (!pte_present(pte))
940                 goto unlock;
941         if ((flags & FOLL_WRITE) && !pte_write(pte))
942                 goto unlock;
943         page = vm_normal_page(vma, address, pte);
944         if (unlikely(!page))
945                 goto unlock;
946
947         if (flags & FOLL_GET)
948                 get_page(page);
949         if (flags & FOLL_TOUCH) {
950                 if ((flags & FOLL_WRITE) &&
951                     !pte_dirty(pte) && !PageDirty(page))
952                         set_page_dirty(page);
953                 mark_page_accessed(page);
954         }
955 unlock:
956         pte_unmap_unlock(ptep, ptl);
957 out:
958         return page;
959
960 no_page_table:
961         /*
962          * When core dumping an enormous anonymous area that nobody
963          * has touched so far, we don't want to allocate page tables.
964          */
965         if (flags & FOLL_ANON) {
966                 page = ZERO_PAGE(address);
967                 if (flags & FOLL_GET)
968                         get_page(page);
969                 BUG_ON(flags & FOLL_WRITE);
970         }
971         return page;
972 }
973
974 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
975                 unsigned long start, int len, int write, int force,
976                 struct page **pages, struct vm_area_struct **vmas)
977 {
978         int i;
979         unsigned int vm_flags;
980
981         /* 
982          * Require read or write permissions.
983          * If 'force' is set, we only require the "MAY" flags.
984          */
985         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
986         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
987         i = 0;
988
989         do {
990                 struct vm_area_struct *vma;
991                 unsigned int foll_flags;
992
993                 vma = find_extend_vma(mm, start);
994                 if (!vma && in_gate_area(tsk, start)) {
995                         unsigned long pg = start & PAGE_MASK;
996                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
997                         pgd_t *pgd;
998                         pud_t *pud;
999                         pmd_t *pmd;
1000                         pte_t *pte;
1001                         if (write) /* user gate pages are read-only */
1002                                 return i ? : -EFAULT;
1003                         if (pg > TASK_SIZE)
1004                                 pgd = pgd_offset_k(pg);
1005                         else
1006                                 pgd = pgd_offset_gate(mm, pg);
1007                         BUG_ON(pgd_none(*pgd));
1008                         pud = pud_offset(pgd, pg);
1009                         BUG_ON(pud_none(*pud));
1010                         pmd = pmd_offset(pud, pg);
1011                         if (pmd_none(*pmd))
1012                                 return i ? : -EFAULT;
1013                         pte = pte_offset_map(pmd, pg);
1014                         if (pte_none(*pte)) {
1015                                 pte_unmap(pte);
1016                                 return i ? : -EFAULT;
1017                         }
1018                         if (pages) {
1019                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1020                                 pages[i] = page;
1021                                 if (page)
1022                                         get_page(page);
1023                         }
1024                         pte_unmap(pte);
1025                         if (vmas)
1026                                 vmas[i] = gate_vma;
1027                         i++;
1028                         start += PAGE_SIZE;
1029                         len--;
1030                         continue;
1031                 }
1032
1033                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1034                                 || !(vm_flags & vma->vm_flags))
1035                         return i ? : -EFAULT;
1036
1037                 if (is_vm_hugetlb_page(vma)) {
1038                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1039                                                 &start, &len, i);
1040                         continue;
1041                 }
1042
1043                 foll_flags = FOLL_TOUCH;
1044                 if (pages)
1045                         foll_flags |= FOLL_GET;
1046                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1047                     (!vma->vm_ops || !vma->vm_ops->nopage))
1048                         foll_flags |= FOLL_ANON;
1049
1050                 do {
1051                         struct page *page;
1052
1053                         if (write)
1054                                 foll_flags |= FOLL_WRITE;
1055
1056                         cond_resched();
1057                         while (!(page = follow_page(vma, start, foll_flags))) {
1058                                 int ret;
1059                                 ret = __handle_mm_fault(mm, vma, start,
1060                                                 foll_flags & FOLL_WRITE);
1061                                 /*
1062                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1063                                  * broken COW when necessary, even if maybe_mkwrite
1064                                  * decided not to set pte_write. We can thus safely do
1065                                  * subsequent page lookups as if they were reads.
1066                                  */
1067                                 if (ret & VM_FAULT_WRITE)
1068                                         foll_flags &= ~FOLL_WRITE;
1069                                 
1070                                 switch (ret & ~VM_FAULT_WRITE) {
1071                                 case VM_FAULT_MINOR:
1072                                         tsk->min_flt++;
1073                                         break;
1074                                 case VM_FAULT_MAJOR:
1075                                         tsk->maj_flt++;
1076                                         break;
1077                                 case VM_FAULT_SIGBUS:
1078                                         return i ? i : -EFAULT;
1079                                 case VM_FAULT_OOM:
1080                                         return i ? i : -ENOMEM;
1081                                 default:
1082                                         BUG();
1083                                 }
1084                         }
1085                         if (pages) {
1086                                 pages[i] = page;
1087
1088                                 flush_anon_page(page, start);
1089                                 flush_dcache_page(page);
1090                         }
1091                         if (vmas)
1092                                 vmas[i] = vma;
1093                         i++;
1094                         start += PAGE_SIZE;
1095                         len--;
1096                 } while (len && start < vma->vm_end);
1097         } while (len);
1098         return i;
1099 }
1100 EXPORT_SYMBOL(get_user_pages);
1101
1102 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1103                         unsigned long addr, unsigned long end, pgprot_t prot)
1104 {
1105         pte_t *pte;
1106         spinlock_t *ptl;
1107
1108         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1109         if (!pte)
1110                 return -ENOMEM;
1111         do {
1112                 struct page *page = ZERO_PAGE(addr);
1113                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1114                 page_cache_get(page);
1115                 page_add_file_rmap(page);
1116                 inc_mm_counter(mm, file_rss);
1117                 BUG_ON(!pte_none(*pte));
1118                 set_pte_at(mm, addr, pte, zero_pte);
1119         } while (pte++, addr += PAGE_SIZE, addr != end);
1120         pte_unmap_unlock(pte - 1, ptl);
1121         return 0;
1122 }
1123
1124 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1125                         unsigned long addr, unsigned long end, pgprot_t prot)
1126 {
1127         pmd_t *pmd;
1128         unsigned long next;
1129
1130         pmd = pmd_alloc(mm, pud, addr);
1131         if (!pmd)
1132                 return -ENOMEM;
1133         do {
1134                 next = pmd_addr_end(addr, end);
1135                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1136                         return -ENOMEM;
1137         } while (pmd++, addr = next, addr != end);
1138         return 0;
1139 }
1140
1141 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1142                         unsigned long addr, unsigned long end, pgprot_t prot)
1143 {
1144         pud_t *pud;
1145         unsigned long next;
1146
1147         pud = pud_alloc(mm, pgd, addr);
1148         if (!pud)
1149                 return -ENOMEM;
1150         do {
1151                 next = pud_addr_end(addr, end);
1152                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1153                         return -ENOMEM;
1154         } while (pud++, addr = next, addr != end);
1155         return 0;
1156 }
1157
1158 int zeromap_page_range(struct vm_area_struct *vma,
1159                         unsigned long addr, unsigned long size, pgprot_t prot)
1160 {
1161         pgd_t *pgd;
1162         unsigned long next;
1163         unsigned long end = addr + size;
1164         struct mm_struct *mm = vma->vm_mm;
1165         int err;
1166
1167         BUG_ON(addr >= end);
1168         pgd = pgd_offset(mm, addr);
1169         flush_cache_range(vma, addr, end);
1170         do {
1171                 next = pgd_addr_end(addr, end);
1172                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1173                 if (err)
1174                         break;
1175         } while (pgd++, addr = next, addr != end);
1176         return err;
1177 }
1178
1179 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1180 {
1181         pgd_t * pgd = pgd_offset(mm, addr);
1182         pud_t * pud = pud_alloc(mm, pgd, addr);
1183         if (pud) {
1184                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1185                 if (pmd)
1186                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1187         }
1188         return NULL;
1189 }
1190
1191 /*
1192  * This is the old fallback for page remapping.
1193  *
1194  * For historical reasons, it only allows reserved pages. Only
1195  * old drivers should use this, and they needed to mark their
1196  * pages reserved for the old functions anyway.
1197  */
1198 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1199 {
1200         int retval;
1201         pte_t *pte;
1202         spinlock_t *ptl;  
1203
1204         retval = -EINVAL;
1205         if (PageAnon(page))
1206                 goto out;
1207         retval = -ENOMEM;
1208         flush_dcache_page(page);
1209         pte = get_locked_pte(mm, addr, &ptl);
1210         if (!pte)
1211                 goto out;
1212         retval = -EBUSY;
1213         if (!pte_none(*pte))
1214                 goto out_unlock;
1215
1216         /* Ok, finally just insert the thing.. */
1217         get_page(page);
1218         inc_mm_counter(mm, file_rss);
1219         page_add_file_rmap(page);
1220         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1221
1222         retval = 0;
1223 out_unlock:
1224         pte_unmap_unlock(pte, ptl);
1225 out:
1226         return retval;
1227 }
1228
1229 /*
1230  * This allows drivers to insert individual pages they've allocated
1231  * into a user vma.
1232  *
1233  * The page has to be a nice clean _individual_ kernel allocation.
1234  * If you allocate a compound page, you need to have marked it as
1235  * such (__GFP_COMP), or manually just split the page up yourself
1236  * (see split_page()).
1237  *
1238  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1239  * took an arbitrary page protection parameter. This doesn't allow
1240  * that. Your vma protection will have to be set up correctly, which
1241  * means that if you want a shared writable mapping, you'd better
1242  * ask for a shared writable mapping!
1243  *
1244  * The page does not need to be reserved.
1245  */
1246 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1247 {
1248         if (addr < vma->vm_start || addr >= vma->vm_end)
1249                 return -EFAULT;
1250         if (!page_count(page))
1251                 return -EINVAL;
1252         vma->vm_flags |= VM_INSERTPAGE;
1253         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1254 }
1255 EXPORT_SYMBOL(vm_insert_page);
1256
1257 /*
1258  * maps a range of physical memory into the requested pages. the old
1259  * mappings are removed. any references to nonexistent pages results
1260  * in null mappings (currently treated as "copy-on-access")
1261  */
1262 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1263                         unsigned long addr, unsigned long end,
1264                         unsigned long pfn, pgprot_t prot)
1265 {
1266         pte_t *pte;
1267         spinlock_t *ptl;
1268
1269         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1270         if (!pte)
1271                 return -ENOMEM;
1272         do {
1273                 BUG_ON(!pte_none(*pte));
1274                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1275                 pfn++;
1276         } while (pte++, addr += PAGE_SIZE, addr != end);
1277         pte_unmap_unlock(pte - 1, ptl);
1278         return 0;
1279 }
1280
1281 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1282                         unsigned long addr, unsigned long end,
1283                         unsigned long pfn, pgprot_t prot)
1284 {
1285         pmd_t *pmd;
1286         unsigned long next;
1287
1288         pfn -= addr >> PAGE_SHIFT;
1289         pmd = pmd_alloc(mm, pud, addr);
1290         if (!pmd)
1291                 return -ENOMEM;
1292         do {
1293                 next = pmd_addr_end(addr, end);
1294                 if (remap_pte_range(mm, pmd, addr, next,
1295                                 pfn + (addr >> PAGE_SHIFT), prot))
1296                         return -ENOMEM;
1297         } while (pmd++, addr = next, addr != end);
1298         return 0;
1299 }
1300
1301 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1302                         unsigned long addr, unsigned long end,
1303                         unsigned long pfn, pgprot_t prot)
1304 {
1305         pud_t *pud;
1306         unsigned long next;
1307
1308         pfn -= addr >> PAGE_SHIFT;
1309         pud = pud_alloc(mm, pgd, addr);
1310         if (!pud)
1311                 return -ENOMEM;
1312         do {
1313                 next = pud_addr_end(addr, end);
1314                 if (remap_pmd_range(mm, pud, addr, next,
1315                                 pfn + (addr >> PAGE_SHIFT), prot))
1316                         return -ENOMEM;
1317         } while (pud++, addr = next, addr != end);
1318         return 0;
1319 }
1320
1321 /*  Note: this is only safe if the mm semaphore is held when called. */
1322 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1323                     unsigned long pfn, unsigned long size, pgprot_t prot)
1324 {
1325         pgd_t *pgd;
1326         unsigned long next;
1327         unsigned long end = addr + PAGE_ALIGN(size);
1328         struct mm_struct *mm = vma->vm_mm;
1329         int err;
1330
1331         /*
1332          * Physically remapped pages are special. Tell the
1333          * rest of the world about it:
1334          *   VM_IO tells people not to look at these pages
1335          *      (accesses can have side effects).
1336          *   VM_RESERVED is specified all over the place, because
1337          *      in 2.4 it kept swapout's vma scan off this vma; but
1338          *      in 2.6 the LRU scan won't even find its pages, so this
1339          *      flag means no more than count its pages in reserved_vm,
1340          *      and omit it from core dump, even when VM_IO turned off.
1341          *   VM_PFNMAP tells the core MM that the base pages are just
1342          *      raw PFN mappings, and do not have a "struct page" associated
1343          *      with them.
1344          *
1345          * There's a horrible special case to handle copy-on-write
1346          * behaviour that some programs depend on. We mark the "original"
1347          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1348          */
1349         if (is_cow_mapping(vma->vm_flags)) {
1350                 if (addr != vma->vm_start || end != vma->vm_end)
1351                         return -EINVAL;
1352                 vma->vm_pgoff = pfn;
1353         }
1354
1355         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1356
1357         BUG_ON(addr >= end);
1358         pfn -= addr >> PAGE_SHIFT;
1359         pgd = pgd_offset(mm, addr);
1360         flush_cache_range(vma, addr, end);
1361         do {
1362                 next = pgd_addr_end(addr, end);
1363                 err = remap_pud_range(mm, pgd, addr, next,
1364                                 pfn + (addr >> PAGE_SHIFT), prot);
1365                 if (err)
1366                         break;
1367         } while (pgd++, addr = next, addr != end);
1368         return err;
1369 }
1370 EXPORT_SYMBOL(remap_pfn_range);
1371
1372 /*
1373  * handle_pte_fault chooses page fault handler according to an entry
1374  * which was read non-atomically.  Before making any commitment, on
1375  * those architectures or configurations (e.g. i386 with PAE) which
1376  * might give a mix of unmatched parts, do_swap_page and do_file_page
1377  * must check under lock before unmapping the pte and proceeding
1378  * (but do_wp_page is only called after already making such a check;
1379  * and do_anonymous_page and do_no_page can safely check later on).
1380  */
1381 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1382                                 pte_t *page_table, pte_t orig_pte)
1383 {
1384         int same = 1;
1385 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1386         if (sizeof(pte_t) > sizeof(unsigned long)) {
1387                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1388                 spin_lock(ptl);
1389                 same = pte_same(*page_table, orig_pte);
1390                 spin_unlock(ptl);
1391         }
1392 #endif
1393         pte_unmap(page_table);
1394         return same;
1395 }
1396
1397 /*
1398  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1399  * servicing faults for write access.  In the normal case, do always want
1400  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1401  * that do not have writing enabled, when used by access_process_vm.
1402  */
1403 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1404 {
1405         if (likely(vma->vm_flags & VM_WRITE))
1406                 pte = pte_mkwrite(pte);
1407         return pte;
1408 }
1409
1410 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1411 {
1412         /*
1413          * If the source page was a PFN mapping, we don't have
1414          * a "struct page" for it. We do a best-effort copy by
1415          * just copying from the original user address. If that
1416          * fails, we just zero-fill it. Live with it.
1417          */
1418         if (unlikely(!src)) {
1419                 void *kaddr = kmap_atomic(dst, KM_USER0);
1420                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1421
1422                 /*
1423                  * This really shouldn't fail, because the page is there
1424                  * in the page tables. But it might just be unreadable,
1425                  * in which case we just give up and fill the result with
1426                  * zeroes.
1427                  */
1428                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1429                         memset(kaddr, 0, PAGE_SIZE);
1430                 kunmap_atomic(kaddr, KM_USER0);
1431                 return;
1432                 
1433         }
1434         copy_user_highpage(dst, src, va);
1435 }
1436
1437 /*
1438  * This routine handles present pages, when users try to write
1439  * to a shared page. It is done by copying the page to a new address
1440  * and decrementing the shared-page counter for the old page.
1441  *
1442  * Note that this routine assumes that the protection checks have been
1443  * done by the caller (the low-level page fault routine in most cases).
1444  * Thus we can safely just mark it writable once we've done any necessary
1445  * COW.
1446  *
1447  * We also mark the page dirty at this point even though the page will
1448  * change only once the write actually happens. This avoids a few races,
1449  * and potentially makes it more efficient.
1450  *
1451  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1452  * but allow concurrent faults), with pte both mapped and locked.
1453  * We return with mmap_sem still held, but pte unmapped and unlocked.
1454  */
1455 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1456                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1457                 spinlock_t *ptl, pte_t orig_pte)
1458 {
1459         struct page *old_page, *new_page;
1460         pte_t entry;
1461         int reuse, ret = VM_FAULT_MINOR;
1462
1463         old_page = vm_normal_page(vma, address, orig_pte);
1464         if (!old_page)
1465                 goto gotten;
1466
1467         if (unlikely((vma->vm_flags & (VM_SHARED|VM_WRITE)) ==
1468                                 (VM_SHARED|VM_WRITE))) {
1469                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1470                         /*
1471                          * Notify the address space that the page is about to
1472                          * become writable so that it can prohibit this or wait
1473                          * for the page to get into an appropriate state.
1474                          *
1475                          * We do this without the lock held, so that it can
1476                          * sleep if it needs to.
1477                          */
1478                         page_cache_get(old_page);
1479                         pte_unmap_unlock(page_table, ptl);
1480
1481                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1482                                 goto unwritable_page;
1483
1484                         page_cache_release(old_page);
1485
1486                         /*
1487                          * Since we dropped the lock we need to revalidate
1488                          * the PTE as someone else may have changed it.  If
1489                          * they did, we just return, as we can count on the
1490                          * MMU to tell us if they didn't also make it writable.
1491                          */
1492                         page_table = pte_offset_map_lock(mm, pmd, address,
1493                                                          &ptl);
1494                         if (!pte_same(*page_table, orig_pte))
1495                                 goto unlock;
1496                 }
1497
1498                 reuse = 1;
1499         } else if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1500                 reuse = can_share_swap_page(old_page);
1501                 unlock_page(old_page);
1502         } else {
1503                 reuse = 0;
1504         }
1505
1506         if (reuse) {
1507                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1508                 entry = pte_mkyoung(orig_pte);
1509                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1510                 ptep_set_access_flags(vma, address, page_table, entry, 1);
1511                 update_mmu_cache(vma, address, entry);
1512                 lazy_mmu_prot_update(entry);
1513                 ret |= VM_FAULT_WRITE;
1514                 goto unlock;
1515         }
1516
1517         /*
1518          * Ok, we need to copy. Oh, well..
1519          */
1520         page_cache_get(old_page);
1521 gotten:
1522         pte_unmap_unlock(page_table, ptl);
1523
1524         if (unlikely(anon_vma_prepare(vma)))
1525                 goto oom;
1526         if (old_page == ZERO_PAGE(address)) {
1527                 new_page = alloc_zeroed_user_highpage(vma, address);
1528                 if (!new_page)
1529                         goto oom;
1530         } else {
1531                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1532                 if (!new_page)
1533                         goto oom;
1534                 cow_user_page(new_page, old_page, address);
1535         }
1536
1537         /*
1538          * Re-check the pte - we dropped the lock
1539          */
1540         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1541         if (likely(pte_same(*page_table, orig_pte))) {
1542                 if (old_page) {
1543                         page_remove_rmap(old_page);
1544                         if (!PageAnon(old_page)) {
1545                                 dec_mm_counter(mm, file_rss);
1546                                 inc_mm_counter(mm, anon_rss);
1547                         }
1548                 } else
1549                         inc_mm_counter(mm, anon_rss);
1550                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1551                 entry = mk_pte(new_page, vma->vm_page_prot);
1552                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1553                 lazy_mmu_prot_update(entry);
1554                 ptep_establish(vma, address, page_table, entry);
1555                 update_mmu_cache(vma, address, entry);
1556                 lru_cache_add_active(new_page);
1557                 page_add_new_anon_rmap(new_page, vma, address);
1558
1559                 /* Free the old page.. */
1560                 new_page = old_page;
1561                 ret |= VM_FAULT_WRITE;
1562         }
1563         if (new_page)
1564                 page_cache_release(new_page);
1565         if (old_page)
1566                 page_cache_release(old_page);
1567 unlock:
1568         pte_unmap_unlock(page_table, ptl);
1569         return ret;
1570 oom:
1571         if (old_page)
1572                 page_cache_release(old_page);
1573         return VM_FAULT_OOM;
1574
1575 unwritable_page:
1576         page_cache_release(old_page);
1577         return VM_FAULT_SIGBUS;
1578 }
1579
1580 /*
1581  * Helper functions for unmap_mapping_range().
1582  *
1583  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1584  *
1585  * We have to restart searching the prio_tree whenever we drop the lock,
1586  * since the iterator is only valid while the lock is held, and anyway
1587  * a later vma might be split and reinserted earlier while lock dropped.
1588  *
1589  * The list of nonlinear vmas could be handled more efficiently, using
1590  * a placeholder, but handle it in the same way until a need is shown.
1591  * It is important to search the prio_tree before nonlinear list: a vma
1592  * may become nonlinear and be shifted from prio_tree to nonlinear list
1593  * while the lock is dropped; but never shifted from list to prio_tree.
1594  *
1595  * In order to make forward progress despite restarting the search,
1596  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1597  * quickly skip it next time around.  Since the prio_tree search only
1598  * shows us those vmas affected by unmapping the range in question, we
1599  * can't efficiently keep all vmas in step with mapping->truncate_count:
1600  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1601  * mapping->truncate_count and vma->vm_truncate_count are protected by
1602  * i_mmap_lock.
1603  *
1604  * In order to make forward progress despite repeatedly restarting some
1605  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1606  * and restart from that address when we reach that vma again.  It might
1607  * have been split or merged, shrunk or extended, but never shifted: so
1608  * restart_addr remains valid so long as it remains in the vma's range.
1609  * unmap_mapping_range forces truncate_count to leap over page-aligned
1610  * values so we can save vma's restart_addr in its truncate_count field.
1611  */
1612 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1613
1614 static void reset_vma_truncate_counts(struct address_space *mapping)
1615 {
1616         struct vm_area_struct *vma;
1617         struct prio_tree_iter iter;
1618
1619         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1620                 vma->vm_truncate_count = 0;
1621         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1622                 vma->vm_truncate_count = 0;
1623 }
1624
1625 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1626                 unsigned long start_addr, unsigned long end_addr,
1627                 struct zap_details *details)
1628 {
1629         unsigned long restart_addr;
1630         int need_break;
1631
1632 again:
1633         restart_addr = vma->vm_truncate_count;
1634         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1635                 start_addr = restart_addr;
1636                 if (start_addr >= end_addr) {
1637                         /* Top of vma has been split off since last time */
1638                         vma->vm_truncate_count = details->truncate_count;
1639                         return 0;
1640                 }
1641         }
1642
1643         restart_addr = zap_page_range(vma, start_addr,
1644                                         end_addr - start_addr, details);
1645         need_break = need_resched() ||
1646                         need_lockbreak(details->i_mmap_lock);
1647
1648         if (restart_addr >= end_addr) {
1649                 /* We have now completed this vma: mark it so */
1650                 vma->vm_truncate_count = details->truncate_count;
1651                 if (!need_break)
1652                         return 0;
1653         } else {
1654                 /* Note restart_addr in vma's truncate_count field */
1655                 vma->vm_truncate_count = restart_addr;
1656                 if (!need_break)
1657                         goto again;
1658         }
1659
1660         spin_unlock(details->i_mmap_lock);
1661         cond_resched();
1662         spin_lock(details->i_mmap_lock);
1663         return -EINTR;
1664 }
1665
1666 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1667                                             struct zap_details *details)
1668 {
1669         struct vm_area_struct *vma;
1670         struct prio_tree_iter iter;
1671         pgoff_t vba, vea, zba, zea;
1672
1673 restart:
1674         vma_prio_tree_foreach(vma, &iter, root,
1675                         details->first_index, details->last_index) {
1676                 /* Skip quickly over those we have already dealt with */
1677                 if (vma->vm_truncate_count == details->truncate_count)
1678                         continue;
1679
1680                 vba = vma->vm_pgoff;
1681                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1682                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1683                 zba = details->first_index;
1684                 if (zba < vba)
1685                         zba = vba;
1686                 zea = details->last_index;
1687                 if (zea > vea)
1688                         zea = vea;
1689
1690                 if (unmap_mapping_range_vma(vma,
1691                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1692                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1693                                 details) < 0)
1694                         goto restart;
1695         }
1696 }
1697
1698 static inline void unmap_mapping_range_list(struct list_head *head,
1699                                             struct zap_details *details)
1700 {
1701         struct vm_area_struct *vma;
1702
1703         /*
1704          * In nonlinear VMAs there is no correspondence between virtual address
1705          * offset and file offset.  So we must perform an exhaustive search
1706          * across *all* the pages in each nonlinear VMA, not just the pages
1707          * whose virtual address lies outside the file truncation point.
1708          */
1709 restart:
1710         list_for_each_entry(vma, head, shared.vm_set.list) {
1711                 /* Skip quickly over those we have already dealt with */
1712                 if (vma->vm_truncate_count == details->truncate_count)
1713                         continue;
1714                 details->nonlinear_vma = vma;
1715                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1716                                         vma->vm_end, details) < 0)
1717                         goto restart;
1718         }
1719 }
1720
1721 /**
1722  * unmap_mapping_range - unmap the portion of all mmaps
1723  * in the specified address_space corresponding to the specified
1724  * page range in the underlying file.
1725  * @mapping: the address space containing mmaps to be unmapped.
1726  * @holebegin: byte in first page to unmap, relative to the start of
1727  * the underlying file.  This will be rounded down to a PAGE_SIZE
1728  * boundary.  Note that this is different from vmtruncate(), which
1729  * must keep the partial page.  In contrast, we must get rid of
1730  * partial pages.
1731  * @holelen: size of prospective hole in bytes.  This will be rounded
1732  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1733  * end of the file.
1734  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1735  * but 0 when invalidating pagecache, don't throw away private data.
1736  */
1737 void unmap_mapping_range(struct address_space *mapping,
1738                 loff_t const holebegin, loff_t const holelen, int even_cows)
1739 {
1740         struct zap_details details;
1741         pgoff_t hba = holebegin >> PAGE_SHIFT;
1742         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1743
1744         /* Check for overflow. */
1745         if (sizeof(holelen) > sizeof(hlen)) {
1746                 long long holeend =
1747                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1748                 if (holeend & ~(long long)ULONG_MAX)
1749                         hlen = ULONG_MAX - hba + 1;
1750         }
1751
1752         details.check_mapping = even_cows? NULL: mapping;
1753         details.nonlinear_vma = NULL;
1754         details.first_index = hba;
1755         details.last_index = hba + hlen - 1;
1756         if (details.last_index < details.first_index)
1757                 details.last_index = ULONG_MAX;
1758         details.i_mmap_lock = &mapping->i_mmap_lock;
1759
1760         spin_lock(&mapping->i_mmap_lock);
1761
1762         /* serialize i_size write against truncate_count write */
1763         smp_wmb();
1764         /* Protect against page faults, and endless unmapping loops */
1765         mapping->truncate_count++;
1766         /*
1767          * For archs where spin_lock has inclusive semantics like ia64
1768          * this smp_mb() will prevent to read pagetable contents
1769          * before the truncate_count increment is visible to
1770          * other cpus.
1771          */
1772         smp_mb();
1773         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1774                 if (mapping->truncate_count == 0)
1775                         reset_vma_truncate_counts(mapping);
1776                 mapping->truncate_count++;
1777         }
1778         details.truncate_count = mapping->truncate_count;
1779
1780         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1781                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1782         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1783                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1784         spin_unlock(&mapping->i_mmap_lock);
1785 }
1786 EXPORT_SYMBOL(unmap_mapping_range);
1787
1788 /*
1789  * Handle all mappings that got truncated by a "truncate()"
1790  * system call.
1791  *
1792  * NOTE! We have to be ready to update the memory sharing
1793  * between the file and the memory map for a potential last
1794  * incomplete page.  Ugly, but necessary.
1795  */
1796 int vmtruncate(struct inode * inode, loff_t offset)
1797 {
1798         struct address_space *mapping = inode->i_mapping;
1799         unsigned long limit;
1800
1801         if (inode->i_size < offset)
1802                 goto do_expand;
1803         /*
1804          * truncation of in-use swapfiles is disallowed - it would cause
1805          * subsequent swapout to scribble on the now-freed blocks.
1806          */
1807         if (IS_SWAPFILE(inode))
1808                 goto out_busy;
1809         i_size_write(inode, offset);
1810         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1811         truncate_inode_pages(mapping, offset);
1812         goto out_truncate;
1813
1814 do_expand:
1815         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1816         if (limit != RLIM_INFINITY && offset > limit)
1817                 goto out_sig;
1818         if (offset > inode->i_sb->s_maxbytes)
1819                 goto out_big;
1820         i_size_write(inode, offset);
1821
1822 out_truncate:
1823         if (inode->i_op && inode->i_op->truncate)
1824                 inode->i_op->truncate(inode);
1825         return 0;
1826 out_sig:
1827         send_sig(SIGXFSZ, current, 0);
1828 out_big:
1829         return -EFBIG;
1830 out_busy:
1831         return -ETXTBSY;
1832 }
1833 EXPORT_SYMBOL(vmtruncate);
1834
1835 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1836 {
1837         struct address_space *mapping = inode->i_mapping;
1838
1839         /*
1840          * If the underlying filesystem is not going to provide
1841          * a way to truncate a range of blocks (punch a hole) -
1842          * we should return failure right now.
1843          */
1844         if (!inode->i_op || !inode->i_op->truncate_range)
1845                 return -ENOSYS;
1846
1847         mutex_lock(&inode->i_mutex);
1848         down_write(&inode->i_alloc_sem);
1849         unmap_mapping_range(mapping, offset, (end - offset), 1);
1850         truncate_inode_pages_range(mapping, offset, end);
1851         inode->i_op->truncate_range(inode, offset, end);
1852         up_write(&inode->i_alloc_sem);
1853         mutex_unlock(&inode->i_mutex);
1854
1855         return 0;
1856 }
1857 EXPORT_UNUSED_SYMBOL(vmtruncate_range);  /*  June 2006  */
1858
1859 /* 
1860  * Primitive swap readahead code. We simply read an aligned block of
1861  * (1 << page_cluster) entries in the swap area. This method is chosen
1862  * because it doesn't cost us any seek time.  We also make sure to queue
1863  * the 'original' request together with the readahead ones...  
1864  *
1865  * This has been extended to use the NUMA policies from the mm triggering
1866  * the readahead.
1867  *
1868  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1869  */
1870 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1871 {
1872 #ifdef CONFIG_NUMA
1873         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1874 #endif
1875         int i, num;
1876         struct page *new_page;
1877         unsigned long offset;
1878
1879         /*
1880          * Get the number of handles we should do readahead io to.
1881          */
1882         num = valid_swaphandles(entry, &offset);
1883         for (i = 0; i < num; offset++, i++) {
1884                 /* Ok, do the async read-ahead now */
1885                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1886                                                            offset), vma, addr);
1887                 if (!new_page)
1888                         break;
1889                 page_cache_release(new_page);
1890 #ifdef CONFIG_NUMA
1891                 /*
1892                  * Find the next applicable VMA for the NUMA policy.
1893                  */
1894                 addr += PAGE_SIZE;
1895                 if (addr == 0)
1896                         vma = NULL;
1897                 if (vma) {
1898                         if (addr >= vma->vm_end) {
1899                                 vma = next_vma;
1900                                 next_vma = vma ? vma->vm_next : NULL;
1901                         }
1902                         if (vma && addr < vma->vm_start)
1903                                 vma = NULL;
1904                 } else {
1905                         if (next_vma && addr >= next_vma->vm_start) {
1906                                 vma = next_vma;
1907                                 next_vma = vma->vm_next;
1908                         }
1909                 }
1910 #endif
1911         }
1912         lru_add_drain();        /* Push any new pages onto the LRU now */
1913 }
1914
1915 /*
1916  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1917  * but allow concurrent faults), and pte mapped but not yet locked.
1918  * We return with mmap_sem still held, but pte unmapped and unlocked.
1919  */
1920 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1921                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1922                 int write_access, pte_t orig_pte)
1923 {
1924         spinlock_t *ptl;
1925         struct page *page;
1926         swp_entry_t entry;
1927         pte_t pte;
1928         int ret = VM_FAULT_MINOR;
1929
1930         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1931                 goto out;
1932
1933         entry = pte_to_swp_entry(orig_pte);
1934         if (is_migration_entry(entry)) {
1935                 migration_entry_wait(mm, pmd, address);
1936                 goto out;
1937         }
1938         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
1939         page = lookup_swap_cache(entry);
1940         if (!page) {
1941                 swapin_readahead(entry, address, vma);
1942                 page = read_swap_cache_async(entry, vma, address);
1943                 if (!page) {
1944                         /*
1945                          * Back out if somebody else faulted in this pte
1946                          * while we released the pte lock.
1947                          */
1948                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1949                         if (likely(pte_same(*page_table, orig_pte)))
1950                                 ret = VM_FAULT_OOM;
1951                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
1952                         goto unlock;
1953                 }
1954
1955                 /* Had to read the page from swap area: Major fault */
1956                 ret = VM_FAULT_MAJOR;
1957                 count_vm_event(PGMAJFAULT);
1958                 grab_swap_token();
1959         }
1960
1961         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
1962         mark_page_accessed(page);
1963         lock_page(page);
1964
1965         /*
1966          * Back out if somebody else already faulted in this pte.
1967          */
1968         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1969         if (unlikely(!pte_same(*page_table, orig_pte)))
1970                 goto out_nomap;
1971
1972         if (unlikely(!PageUptodate(page))) {
1973                 ret = VM_FAULT_SIGBUS;
1974                 goto out_nomap;
1975         }
1976
1977         /* The page isn't present yet, go ahead with the fault. */
1978
1979         inc_mm_counter(mm, anon_rss);
1980         pte = mk_pte(page, vma->vm_page_prot);
1981         if (write_access && can_share_swap_page(page)) {
1982                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1983                 write_access = 0;
1984         }
1985
1986         flush_icache_page(vma, page);
1987         set_pte_at(mm, address, page_table, pte);
1988         page_add_anon_rmap(page, vma, address);
1989
1990         swap_free(entry);
1991         if (vm_swap_full())
1992                 remove_exclusive_swap_page(page);
1993         unlock_page(page);
1994
1995         if (write_access) {
1996                 if (do_wp_page(mm, vma, address,
1997                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1998                         ret = VM_FAULT_OOM;
1999                 goto out;
2000         }
2001
2002         /* No need to invalidate - it was non-present before */
2003         update_mmu_cache(vma, address, pte);
2004         lazy_mmu_prot_update(pte);
2005 unlock:
2006         pte_unmap_unlock(page_table, ptl);
2007 out:
2008         return ret;
2009 out_nomap:
2010         pte_unmap_unlock(page_table, ptl);
2011         unlock_page(page);
2012         page_cache_release(page);
2013         return ret;
2014 }
2015
2016 /*
2017  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2018  * but allow concurrent faults), and pte mapped but not yet locked.
2019  * We return with mmap_sem still held, but pte unmapped and unlocked.
2020  */
2021 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2022                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2023                 int write_access)
2024 {
2025         struct page *page;
2026         spinlock_t *ptl;
2027         pte_t entry;
2028
2029         if (write_access) {
2030                 /* Allocate our own private page. */
2031                 pte_unmap(page_table);
2032
2033                 if (unlikely(anon_vma_prepare(vma)))
2034                         goto oom;
2035                 page = alloc_zeroed_user_highpage(vma, address);
2036                 if (!page)
2037                         goto oom;
2038
2039                 entry = mk_pte(page, vma->vm_page_prot);
2040                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2041
2042                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2043                 if (!pte_none(*page_table))
2044                         goto release;
2045                 inc_mm_counter(mm, anon_rss);
2046                 lru_cache_add_active(page);
2047                 page_add_new_anon_rmap(page, vma, address);
2048         } else {
2049                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2050                 page = ZERO_PAGE(address);
2051                 page_cache_get(page);
2052                 entry = mk_pte(page, vma->vm_page_prot);
2053
2054                 ptl = pte_lockptr(mm, pmd);
2055                 spin_lock(ptl);
2056                 if (!pte_none(*page_table))
2057                         goto release;
2058                 inc_mm_counter(mm, file_rss);
2059                 page_add_file_rmap(page);
2060         }
2061
2062         set_pte_at(mm, address, page_table, entry);
2063
2064         /* No need to invalidate - it was non-present before */
2065         update_mmu_cache(vma, address, entry);
2066         lazy_mmu_prot_update(entry);
2067 unlock:
2068         pte_unmap_unlock(page_table, ptl);
2069         return VM_FAULT_MINOR;
2070 release:
2071         page_cache_release(page);
2072         goto unlock;
2073 oom:
2074         return VM_FAULT_OOM;
2075 }
2076
2077 /*
2078  * do_no_page() tries to create a new page mapping. It aggressively
2079  * tries to share with existing pages, but makes a separate copy if
2080  * the "write_access" parameter is true in order to avoid the next
2081  * page fault.
2082  *
2083  * As this is called only for pages that do not currently exist, we
2084  * do not need to flush old virtual caches or the TLB.
2085  *
2086  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2087  * but allow concurrent faults), and pte mapped but not yet locked.
2088  * We return with mmap_sem still held, but pte unmapped and unlocked.
2089  */
2090 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2091                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2092                 int write_access)
2093 {
2094         spinlock_t *ptl;
2095         struct page *new_page;
2096         struct address_space *mapping = NULL;
2097         pte_t entry;
2098         unsigned int sequence = 0;
2099         int ret = VM_FAULT_MINOR;
2100         int anon = 0;
2101
2102         pte_unmap(page_table);
2103         BUG_ON(vma->vm_flags & VM_PFNMAP);
2104
2105         if (vma->vm_file) {
2106                 mapping = vma->vm_file->f_mapping;
2107                 sequence = mapping->truncate_count;
2108                 smp_rmb(); /* serializes i_size against truncate_count */
2109         }
2110 retry:
2111         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2112         /*
2113          * No smp_rmb is needed here as long as there's a full
2114          * spin_lock/unlock sequence inside the ->nopage callback
2115          * (for the pagecache lookup) that acts as an implicit
2116          * smp_mb() and prevents the i_size read to happen
2117          * after the next truncate_count read.
2118          */
2119
2120         /* no page was available -- either SIGBUS or OOM */
2121         if (new_page == NOPAGE_SIGBUS)
2122                 return VM_FAULT_SIGBUS;
2123         if (new_page == NOPAGE_OOM)
2124                 return VM_FAULT_OOM;
2125
2126         /*
2127          * Should we do an early C-O-W break?
2128          */
2129         if (write_access) {
2130                 if (!(vma->vm_flags & VM_SHARED)) {
2131                         struct page *page;
2132
2133                         if (unlikely(anon_vma_prepare(vma)))
2134                                 goto oom;
2135                         page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2136                         if (!page)
2137                                 goto oom;
2138                         copy_user_highpage(page, new_page, address);
2139                         page_cache_release(new_page);
2140                         new_page = page;
2141                         anon = 1;
2142
2143                 } else {
2144                         /* if the page will be shareable, see if the backing
2145                          * address space wants to know that the page is about
2146                          * to become writable */
2147                         if (vma->vm_ops->page_mkwrite &&
2148                             vma->vm_ops->page_mkwrite(vma, new_page) < 0
2149                             ) {
2150                                 page_cache_release(new_page);
2151                                 return VM_FAULT_SIGBUS;
2152                         }
2153                 }
2154         }
2155
2156         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2157         /*
2158          * For a file-backed vma, someone could have truncated or otherwise
2159          * invalidated this page.  If unmap_mapping_range got called,
2160          * retry getting the page.
2161          */
2162         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2163                 pte_unmap_unlock(page_table, ptl);
2164                 page_cache_release(new_page);
2165                 cond_resched();
2166                 sequence = mapping->truncate_count;
2167                 smp_rmb();
2168                 goto retry;
2169         }
2170
2171         /*
2172          * This silly early PAGE_DIRTY setting removes a race
2173          * due to the bad i386 page protection. But it's valid
2174          * for other architectures too.
2175          *
2176          * Note that if write_access is true, we either now have
2177          * an exclusive copy of the page, or this is a shared mapping,
2178          * so we can make it writable and dirty to avoid having to
2179          * handle that later.
2180          */
2181         /* Only go through if we didn't race with anybody else... */
2182         if (pte_none(*page_table)) {
2183                 flush_icache_page(vma, new_page);
2184                 entry = mk_pte(new_page, vma->vm_page_prot);
2185                 if (write_access)
2186                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2187                 set_pte_at(mm, address, page_table, entry);
2188                 if (anon) {
2189                         inc_mm_counter(mm, anon_rss);
2190                         lru_cache_add_active(new_page);
2191                         page_add_new_anon_rmap(new_page, vma, address);
2192                 } else {
2193                         inc_mm_counter(mm, file_rss);
2194                         page_add_file_rmap(new_page);
2195                 }
2196         } else {
2197                 /* One of our sibling threads was faster, back out. */
2198                 page_cache_release(new_page);
2199                 goto unlock;
2200         }
2201
2202         /* no need to invalidate: a not-present page shouldn't be cached */
2203         update_mmu_cache(vma, address, entry);
2204         lazy_mmu_prot_update(entry);
2205 unlock:
2206         pte_unmap_unlock(page_table, ptl);
2207         return ret;
2208 oom:
2209         page_cache_release(new_page);
2210         return VM_FAULT_OOM;
2211 }
2212
2213 /*
2214  * Fault of a previously existing named mapping. Repopulate the pte
2215  * from the encoded file_pte if possible. This enables swappable
2216  * nonlinear vmas.
2217  *
2218  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2219  * but allow concurrent faults), and pte mapped but not yet locked.
2220  * We return with mmap_sem still held, but pte unmapped and unlocked.
2221  */
2222 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2223                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2224                 int write_access, pte_t orig_pte)
2225 {
2226         pgoff_t pgoff;
2227         int err;
2228
2229         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2230                 return VM_FAULT_MINOR;
2231
2232         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2233                 /*
2234                  * Page table corrupted: show pte and kill process.
2235                  */
2236                 print_bad_pte(vma, orig_pte, address);
2237                 return VM_FAULT_OOM;
2238         }
2239         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2240
2241         pgoff = pte_to_pgoff(orig_pte);
2242         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2243                                         vma->vm_page_prot, pgoff, 0);
2244         if (err == -ENOMEM)
2245                 return VM_FAULT_OOM;
2246         if (err)
2247                 return VM_FAULT_SIGBUS;
2248         return VM_FAULT_MAJOR;
2249 }
2250
2251 /*
2252  * These routines also need to handle stuff like marking pages dirty
2253  * and/or accessed for architectures that don't do it in hardware (most
2254  * RISC architectures).  The early dirtying is also good on the i386.
2255  *
2256  * There is also a hook called "update_mmu_cache()" that architectures
2257  * with external mmu caches can use to update those (ie the Sparc or
2258  * PowerPC hashed page tables that act as extended TLBs).
2259  *
2260  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2261  * but allow concurrent faults), and pte mapped but not yet locked.
2262  * We return with mmap_sem still held, but pte unmapped and unlocked.
2263  */
2264 static inline int handle_pte_fault(struct mm_struct *mm,
2265                 struct vm_area_struct *vma, unsigned long address,
2266                 pte_t *pte, pmd_t *pmd, int write_access)
2267 {
2268         pte_t entry;
2269         pte_t old_entry;
2270         spinlock_t *ptl;
2271
2272         old_entry = entry = *pte;
2273         if (!pte_present(entry)) {
2274                 if (pte_none(entry)) {
2275                         if (!vma->vm_ops || !vma->vm_ops->nopage)
2276                                 return do_anonymous_page(mm, vma, address,
2277                                         pte, pmd, write_access);
2278                         return do_no_page(mm, vma, address,
2279                                         pte, pmd, write_access);
2280                 }
2281                 if (pte_file(entry))
2282                         return do_file_page(mm, vma, address,
2283                                         pte, pmd, write_access, entry);
2284                 return do_swap_page(mm, vma, address,
2285                                         pte, pmd, write_access, entry);
2286         }
2287
2288         ptl = pte_lockptr(mm, pmd);
2289         spin_lock(ptl);
2290         if (unlikely(!pte_same(*pte, entry)))
2291                 goto unlock;
2292         if (write_access) {
2293                 if (!pte_write(entry))
2294                         return do_wp_page(mm, vma, address,
2295                                         pte, pmd, ptl, entry);
2296                 entry = pte_mkdirty(entry);
2297         }
2298         entry = pte_mkyoung(entry);
2299         if (!pte_same(old_entry, entry)) {
2300                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2301                 update_mmu_cache(vma, address, entry);
2302                 lazy_mmu_prot_update(entry);
2303         } else {
2304                 /*
2305                  * This is needed only for protection faults but the arch code
2306                  * is not yet telling us if this is a protection fault or not.
2307                  * This still avoids useless tlb flushes for .text page faults
2308                  * with threads.
2309                  */
2310                 if (write_access)
2311                         flush_tlb_page(vma, address);
2312         }
2313 unlock:
2314         pte_unmap_unlock(pte, ptl);
2315         return VM_FAULT_MINOR;
2316 }
2317
2318 /*
2319  * By the time we get here, we already hold the mm semaphore
2320  */
2321 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2322                 unsigned long address, int write_access)
2323 {
2324         pgd_t *pgd;
2325         pud_t *pud;
2326         pmd_t *pmd;
2327         pte_t *pte;
2328
2329         __set_current_state(TASK_RUNNING);
2330
2331         count_vm_event(PGFAULT);
2332
2333         if (unlikely(is_vm_hugetlb_page(vma)))
2334                 return hugetlb_fault(mm, vma, address, write_access);
2335
2336         pgd = pgd_offset(mm, address);
2337         pud = pud_alloc(mm, pgd, address);
2338         if (!pud)
2339                 return VM_FAULT_OOM;
2340         pmd = pmd_alloc(mm, pud, address);
2341         if (!pmd)
2342                 return VM_FAULT_OOM;
2343         pte = pte_alloc_map(mm, pmd, address);
2344         if (!pte)
2345                 return VM_FAULT_OOM;
2346
2347         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2348 }
2349
2350 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2351
2352 #ifndef __PAGETABLE_PUD_FOLDED
2353 /*
2354  * Allocate page upper directory.
2355  * We've already handled the fast-path in-line.
2356  */
2357 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2358 {
2359         pud_t *new = pud_alloc_one(mm, address);
2360         if (!new)
2361                 return -ENOMEM;
2362
2363         spin_lock(&mm->page_table_lock);
2364         if (pgd_present(*pgd))          /* Another has populated it */
2365                 pud_free(new);
2366         else
2367                 pgd_populate(mm, pgd, new);
2368         spin_unlock(&mm->page_table_lock);
2369         return 0;
2370 }
2371 #else
2372 /* Workaround for gcc 2.96 */
2373 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2374 {
2375         return 0;
2376 }
2377 #endif /* __PAGETABLE_PUD_FOLDED */
2378
2379 #ifndef __PAGETABLE_PMD_FOLDED
2380 /*
2381  * Allocate page middle directory.
2382  * We've already handled the fast-path in-line.
2383  */
2384 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2385 {
2386         pmd_t *new = pmd_alloc_one(mm, address);
2387         if (!new)
2388                 return -ENOMEM;
2389
2390         spin_lock(&mm->page_table_lock);
2391 #ifndef __ARCH_HAS_4LEVEL_HACK
2392         if (pud_present(*pud))          /* Another has populated it */
2393                 pmd_free(new);
2394         else
2395                 pud_populate(mm, pud, new);
2396 #else
2397         if (pgd_present(*pud))          /* Another has populated it */
2398                 pmd_free(new);
2399         else
2400                 pgd_populate(mm, pud, new);
2401 #endif /* __ARCH_HAS_4LEVEL_HACK */
2402         spin_unlock(&mm->page_table_lock);
2403         return 0;
2404 }
2405 #else
2406 /* Workaround for gcc 2.96 */
2407 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2408 {
2409         return 0;
2410 }
2411 #endif /* __PAGETABLE_PMD_FOLDED */
2412
2413 int make_pages_present(unsigned long addr, unsigned long end)
2414 {
2415         int ret, len, write;
2416         struct vm_area_struct * vma;
2417
2418         vma = find_vma(current->mm, addr);
2419         if (!vma)
2420                 return -1;
2421         write = (vma->vm_flags & VM_WRITE) != 0;
2422         BUG_ON(addr >= end);
2423         BUG_ON(end > vma->vm_end);
2424         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2425         ret = get_user_pages(current, current->mm, addr,
2426                         len, write, 0, NULL, NULL);
2427         if (ret < 0)
2428                 return ret;
2429         return ret == len ? 0 : -1;
2430 }
2431
2432 /* 
2433  * Map a vmalloc()-space virtual address to the physical page.
2434  */
2435 struct page * vmalloc_to_page(void * vmalloc_addr)
2436 {
2437         unsigned long addr = (unsigned long) vmalloc_addr;
2438         struct page *page = NULL;
2439         pgd_t *pgd = pgd_offset_k(addr);
2440         pud_t *pud;
2441         pmd_t *pmd;
2442         pte_t *ptep, pte;
2443   
2444         if (!pgd_none(*pgd)) {
2445                 pud = pud_offset(pgd, addr);
2446                 if (!pud_none(*pud)) {
2447                         pmd = pmd_offset(pud, addr);
2448                         if (!pmd_none(*pmd)) {
2449                                 ptep = pte_offset_map(pmd, addr);
2450                                 pte = *ptep;
2451                                 if (pte_present(pte))
2452                                         page = pte_page(pte);
2453                                 pte_unmap(ptep);
2454                         }
2455                 }
2456         }
2457         return page;
2458 }
2459
2460 EXPORT_SYMBOL(vmalloc_to_page);
2461
2462 /*
2463  * Map a vmalloc()-space virtual address to the physical page frame number.
2464  */
2465 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2466 {
2467         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2468 }
2469
2470 EXPORT_SYMBOL(vmalloc_to_pfn);
2471
2472 #if !defined(__HAVE_ARCH_GATE_AREA)
2473
2474 #if defined(AT_SYSINFO_EHDR)
2475 static struct vm_area_struct gate_vma;
2476
2477 static int __init gate_vma_init(void)
2478 {
2479         gate_vma.vm_mm = NULL;
2480         gate_vma.vm_start = FIXADDR_USER_START;
2481         gate_vma.vm_end = FIXADDR_USER_END;
2482         gate_vma.vm_page_prot = PAGE_READONLY;
2483         gate_vma.vm_flags = 0;
2484         return 0;
2485 }
2486 __initcall(gate_vma_init);
2487 #endif
2488
2489 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2490 {
2491 #ifdef AT_SYSINFO_EHDR
2492         return &gate_vma;
2493 #else
2494         return NULL;
2495 #endif
2496 }
2497
2498 int in_gate_area_no_task(unsigned long addr)
2499 {
2500 #ifdef AT_SYSINFO_EHDR
2501         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2502                 return 1;
2503 #endif
2504         return 0;
2505 }
2506
2507 #endif  /* __HAVE_ARCH_GATE_AREA */