e9ef599498b5eec6261c201992d54992b9bdcd6d
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
54 #include <asm/tlb.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
57
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
65
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
69
70 unsigned long num_physpages;
71 /*
72  * A number of key systems in x86 including ioremap() rely on the assumption
73  * that high_memory defines the upper bound on direct map memory, then end
74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76  * and ZONE_HIGHMEM.
77  */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
80
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85 /*
86  * If a p?d_bad entry is found while walking page tables, report
87  * the error, before resetting entry to p?d_none.  Usually (but
88  * very seldom) called out from the p?d_none_or_clear_bad macros.
89  */
90
91 void pgd_clear_bad(pgd_t *pgd)
92 {
93         pgd_ERROR(*pgd);
94         pgd_clear(pgd);
95 }
96
97 void pud_clear_bad(pud_t *pud)
98 {
99         pud_ERROR(*pud);
100         pud_clear(pud);
101 }
102
103 void pmd_clear_bad(pmd_t *pmd)
104 {
105         pmd_ERROR(*pmd);
106         pmd_clear(pmd);
107 }
108
109 /*
110  * Note: this doesn't free the actual pages themselves. That
111  * has been handled earlier when unmapping all the memory regions.
112  */
113 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
114 {
115         struct page *page = pmd_page(*pmd);
116         pmd_clear(pmd);
117         pte_lock_deinit(page);
118         pte_free_tlb(tlb, page);
119         dec_page_state(nr_page_table_pages);
120         tlb->mm->nr_ptes--;
121 }
122
123 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
124                                 unsigned long addr, unsigned long end,
125                                 unsigned long floor, unsigned long ceiling)
126 {
127         pmd_t *pmd;
128         unsigned long next;
129         unsigned long start;
130
131         start = addr;
132         pmd = pmd_offset(pud, addr);
133         do {
134                 next = pmd_addr_end(addr, end);
135                 if (pmd_none_or_clear_bad(pmd))
136                         continue;
137                 free_pte_range(tlb, pmd);
138         } while (pmd++, addr = next, addr != end);
139
140         start &= PUD_MASK;
141         if (start < floor)
142                 return;
143         if (ceiling) {
144                 ceiling &= PUD_MASK;
145                 if (!ceiling)
146                         return;
147         }
148         if (end - 1 > ceiling - 1)
149                 return;
150
151         pmd = pmd_offset(pud, start);
152         pud_clear(pud);
153         pmd_free_tlb(tlb, pmd);
154 }
155
156 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
157                                 unsigned long addr, unsigned long end,
158                                 unsigned long floor, unsigned long ceiling)
159 {
160         pud_t *pud;
161         unsigned long next;
162         unsigned long start;
163
164         start = addr;
165         pud = pud_offset(pgd, addr);
166         do {
167                 next = pud_addr_end(addr, end);
168                 if (pud_none_or_clear_bad(pud))
169                         continue;
170                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
171         } while (pud++, addr = next, addr != end);
172
173         start &= PGDIR_MASK;
174         if (start < floor)
175                 return;
176         if (ceiling) {
177                 ceiling &= PGDIR_MASK;
178                 if (!ceiling)
179                         return;
180         }
181         if (end - 1 > ceiling - 1)
182                 return;
183
184         pud = pud_offset(pgd, start);
185         pgd_clear(pgd);
186         pud_free_tlb(tlb, pud);
187 }
188
189 /*
190  * This function frees user-level page tables of a process.
191  *
192  * Must be called with pagetable lock held.
193  */
194 void free_pgd_range(struct mmu_gather **tlb,
195                         unsigned long addr, unsigned long end,
196                         unsigned long floor, unsigned long ceiling)
197 {
198         pgd_t *pgd;
199         unsigned long next;
200         unsigned long start;
201
202         /*
203          * The next few lines have given us lots of grief...
204          *
205          * Why are we testing PMD* at this top level?  Because often
206          * there will be no work to do at all, and we'd prefer not to
207          * go all the way down to the bottom just to discover that.
208          *
209          * Why all these "- 1"s?  Because 0 represents both the bottom
210          * of the address space and the top of it (using -1 for the
211          * top wouldn't help much: the masks would do the wrong thing).
212          * The rule is that addr 0 and floor 0 refer to the bottom of
213          * the address space, but end 0 and ceiling 0 refer to the top
214          * Comparisons need to use "end - 1" and "ceiling - 1" (though
215          * that end 0 case should be mythical).
216          *
217          * Wherever addr is brought up or ceiling brought down, we must
218          * be careful to reject "the opposite 0" before it confuses the
219          * subsequent tests.  But what about where end is brought down
220          * by PMD_SIZE below? no, end can't go down to 0 there.
221          *
222          * Whereas we round start (addr) and ceiling down, by different
223          * masks at different levels, in order to test whether a table
224          * now has no other vmas using it, so can be freed, we don't
225          * bother to round floor or end up - the tests don't need that.
226          */
227
228         addr &= PMD_MASK;
229         if (addr < floor) {
230                 addr += PMD_SIZE;
231                 if (!addr)
232                         return;
233         }
234         if (ceiling) {
235                 ceiling &= PMD_MASK;
236                 if (!ceiling)
237                         return;
238         }
239         if (end - 1 > ceiling - 1)
240                 end -= PMD_SIZE;
241         if (addr > end - 1)
242                 return;
243
244         start = addr;
245         pgd = pgd_offset((*tlb)->mm, addr);
246         do {
247                 next = pgd_addr_end(addr, end);
248                 if (pgd_none_or_clear_bad(pgd))
249                         continue;
250                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
251         } while (pgd++, addr = next, addr != end);
252
253         if (!(*tlb)->fullmm)
254                 flush_tlb_pgtables((*tlb)->mm, start, end);
255 }
256
257 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
258                 unsigned long floor, unsigned long ceiling)
259 {
260         while (vma) {
261                 struct vm_area_struct *next = vma->vm_next;
262                 unsigned long addr = vma->vm_start;
263
264                 /*
265                  * Hide vma from rmap and vmtruncate before freeing pgtables
266                  */
267                 anon_vma_unlink(vma);
268                 unlink_file_vma(vma);
269
270                 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
271                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
272                                 floor, next? next->vm_start: ceiling);
273                 } else {
274                         /*
275                          * Optimization: gather nearby vmas into one call down
276                          */
277                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
278                           && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
279                                                         HPAGE_SIZE)) {
280                                 vma = next;
281                                 next = vma->vm_next;
282                                 anon_vma_unlink(vma);
283                                 unlink_file_vma(vma);
284                         }
285                         free_pgd_range(tlb, addr, vma->vm_end,
286                                 floor, next? next->vm_start: ceiling);
287                 }
288                 vma = next;
289         }
290 }
291
292 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
293 {
294         struct page *new = pte_alloc_one(mm, address);
295         if (!new)
296                 return -ENOMEM;
297
298         pte_lock_init(new);
299         spin_lock(&mm->page_table_lock);
300         if (pmd_present(*pmd)) {        /* Another has populated it */
301                 pte_lock_deinit(new);
302                 pte_free(new);
303         } else {
304                 mm->nr_ptes++;
305                 inc_page_state(nr_page_table_pages);
306                 pmd_populate(mm, pmd, new);
307         }
308         spin_unlock(&mm->page_table_lock);
309         return 0;
310 }
311
312 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
313 {
314         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
315         if (!new)
316                 return -ENOMEM;
317
318         spin_lock(&init_mm.page_table_lock);
319         if (pmd_present(*pmd))          /* Another has populated it */
320                 pte_free_kernel(new);
321         else
322                 pmd_populate_kernel(&init_mm, pmd, new);
323         spin_unlock(&init_mm.page_table_lock);
324         return 0;
325 }
326
327 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
328 {
329         if (file_rss)
330                 add_mm_counter(mm, file_rss, file_rss);
331         if (anon_rss)
332                 add_mm_counter(mm, anon_rss, anon_rss);
333 }
334
335 /*
336  * This function is called to print an error when a pte in a
337  * !VM_RESERVED region is found pointing to an invalid pfn (which
338  * is an error.
339  *
340  * The calling function must still handle the error.
341  */
342 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
343 {
344         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
345                         "vm_flags = %lx, vaddr = %lx\n",
346                 (long long)pte_val(pte),
347                 (vma->vm_mm == current->mm ? current->comm : "???"),
348                 vma->vm_flags, vaddr);
349         dump_stack();
350 }
351
352 /*
353  * copy one vm_area from one task to the other. Assumes the page tables
354  * already present in the new task to be cleared in the whole range
355  * covered by this vma.
356  */
357
358 static inline void
359 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
360                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
361                 unsigned long addr, int *rss)
362 {
363         unsigned long vm_flags = vma->vm_flags;
364         pte_t pte = *src_pte;
365         struct page *page;
366         unsigned long pfn;
367
368         /* pte contains position in swap or file, so copy. */
369         if (unlikely(!pte_present(pte))) {
370                 if (!pte_file(pte)) {
371                         swap_duplicate(pte_to_swp_entry(pte));
372                         /* make sure dst_mm is on swapoff's mmlist. */
373                         if (unlikely(list_empty(&dst_mm->mmlist))) {
374                                 spin_lock(&mmlist_lock);
375                                 list_add(&dst_mm->mmlist, &src_mm->mmlist);
376                                 spin_unlock(&mmlist_lock);
377                         }
378                 }
379                 goto out_set_pte;
380         }
381
382         /* If the region is VM_RESERVED, the mapping is not
383          * mapped via rmap - duplicate the pte as is.
384          */
385         if (vm_flags & VM_RESERVED)
386                 goto out_set_pte;
387
388         pfn = pte_pfn(pte);
389         /* If the pte points outside of valid memory but
390          * the region is not VM_RESERVED, we have a problem.
391          */
392         if (unlikely(!pfn_valid(pfn))) {
393                 print_bad_pte(vma, pte, addr);
394                 goto out_set_pte; /* try to do something sane */
395         }
396
397         page = pfn_to_page(pfn);
398
399         /*
400          * If it's a COW mapping, write protect it both
401          * in the parent and the child
402          */
403         if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
404                 ptep_set_wrprotect(src_mm, addr, src_pte);
405                 pte = *src_pte;
406         }
407
408         /*
409          * If it's a shared mapping, mark it clean in
410          * the child
411          */
412         if (vm_flags & VM_SHARED)
413                 pte = pte_mkclean(pte);
414         pte = pte_mkold(pte);
415         get_page(page);
416         page_dup_rmap(page);
417         rss[!!PageAnon(page)]++;
418
419 out_set_pte:
420         set_pte_at(dst_mm, addr, dst_pte, pte);
421 }
422
423 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
424                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
425                 unsigned long addr, unsigned long end)
426 {
427         pte_t *src_pte, *dst_pte;
428         spinlock_t *src_ptl, *dst_ptl;
429         int progress = 0;
430         int rss[2];
431
432 again:
433         rss[1] = rss[0] = 0;
434         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
435         if (!dst_pte)
436                 return -ENOMEM;
437         src_pte = pte_offset_map_nested(src_pmd, addr);
438         src_ptl = pte_lockptr(src_mm, src_pmd);
439         spin_lock(src_ptl);
440
441         do {
442                 /*
443                  * We are holding two locks at this point - either of them
444                  * could generate latencies in another task on another CPU.
445                  */
446                 if (progress >= 32) {
447                         progress = 0;
448                         if (need_resched() ||
449                             need_lockbreak(src_ptl) ||
450                             need_lockbreak(dst_ptl))
451                                 break;
452                 }
453                 if (pte_none(*src_pte)) {
454                         progress++;
455                         continue;
456                 }
457                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
458                 progress += 8;
459         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
460
461         spin_unlock(src_ptl);
462         pte_unmap_nested(src_pte - 1);
463         add_mm_rss(dst_mm, rss[0], rss[1]);
464         pte_unmap_unlock(dst_pte - 1, dst_ptl);
465         cond_resched();
466         if (addr != end)
467                 goto again;
468         return 0;
469 }
470
471 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
472                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
473                 unsigned long addr, unsigned long end)
474 {
475         pmd_t *src_pmd, *dst_pmd;
476         unsigned long next;
477
478         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
479         if (!dst_pmd)
480                 return -ENOMEM;
481         src_pmd = pmd_offset(src_pud, addr);
482         do {
483                 next = pmd_addr_end(addr, end);
484                 if (pmd_none_or_clear_bad(src_pmd))
485                         continue;
486                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
487                                                 vma, addr, next))
488                         return -ENOMEM;
489         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
490         return 0;
491 }
492
493 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
494                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
495                 unsigned long addr, unsigned long end)
496 {
497         pud_t *src_pud, *dst_pud;
498         unsigned long next;
499
500         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
501         if (!dst_pud)
502                 return -ENOMEM;
503         src_pud = pud_offset(src_pgd, addr);
504         do {
505                 next = pud_addr_end(addr, end);
506                 if (pud_none_or_clear_bad(src_pud))
507                         continue;
508                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
509                                                 vma, addr, next))
510                         return -ENOMEM;
511         } while (dst_pud++, src_pud++, addr = next, addr != end);
512         return 0;
513 }
514
515 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
516                 struct vm_area_struct *vma)
517 {
518         pgd_t *src_pgd, *dst_pgd;
519         unsigned long next;
520         unsigned long addr = vma->vm_start;
521         unsigned long end = vma->vm_end;
522
523         /*
524          * Don't copy ptes where a page fault will fill them correctly.
525          * Fork becomes much lighter when there are big shared or private
526          * readonly mappings. The tradeoff is that copy_page_range is more
527          * efficient than faulting.
528          */
529         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_RESERVED))) {
530                 if (!vma->anon_vma)
531                         return 0;
532         }
533
534         if (is_vm_hugetlb_page(vma))
535                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
536
537         dst_pgd = pgd_offset(dst_mm, addr);
538         src_pgd = pgd_offset(src_mm, addr);
539         do {
540                 next = pgd_addr_end(addr, end);
541                 if (pgd_none_or_clear_bad(src_pgd))
542                         continue;
543                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
544                                                 vma, addr, next))
545                         return -ENOMEM;
546         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
547         return 0;
548 }
549
550 static void zap_pte_range(struct mmu_gather *tlb,
551                                 struct vm_area_struct *vma, pmd_t *pmd,
552                                 unsigned long addr, unsigned long end,
553                                 struct zap_details *details)
554 {
555         struct mm_struct *mm = tlb->mm;
556         pte_t *pte;
557         spinlock_t *ptl;
558         int file_rss = 0;
559         int anon_rss = 0;
560
561         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
562         do {
563                 pte_t ptent = *pte;
564                 if (pte_none(ptent))
565                         continue;
566                 if (pte_present(ptent)) {
567                         struct page *page = NULL;
568                         if (!(vma->vm_flags & VM_RESERVED)) {
569                                 unsigned long pfn = pte_pfn(ptent);
570                                 if (unlikely(!pfn_valid(pfn)))
571                                         print_bad_pte(vma, ptent, addr);
572                                 else
573                                         page = pfn_to_page(pfn);
574                         }
575                         if (unlikely(details) && page) {
576                                 /*
577                                  * unmap_shared_mapping_pages() wants to
578                                  * invalidate cache without truncating:
579                                  * unmap shared but keep private pages.
580                                  */
581                                 if (details->check_mapping &&
582                                     details->check_mapping != page->mapping)
583                                         continue;
584                                 /*
585                                  * Each page->index must be checked when
586                                  * invalidating or truncating nonlinear.
587                                  */
588                                 if (details->nonlinear_vma &&
589                                     (page->index < details->first_index ||
590                                      page->index > details->last_index))
591                                         continue;
592                         }
593                         ptent = ptep_get_and_clear_full(mm, addr, pte,
594                                                         tlb->fullmm);
595                         tlb_remove_tlb_entry(tlb, pte, addr);
596                         if (unlikely(!page))
597                                 continue;
598                         if (unlikely(details) && details->nonlinear_vma
599                             && linear_page_index(details->nonlinear_vma,
600                                                 addr) != page->index)
601                                 set_pte_at(mm, addr, pte,
602                                            pgoff_to_pte(page->index));
603                         if (PageAnon(page))
604                                 anon_rss--;
605                         else {
606                                 if (pte_dirty(ptent))
607                                         set_page_dirty(page);
608                                 if (pte_young(ptent))
609                                         mark_page_accessed(page);
610                                 file_rss--;
611                         }
612                         page_remove_rmap(page);
613                         tlb_remove_page(tlb, page);
614                         continue;
615                 }
616                 /*
617                  * If details->check_mapping, we leave swap entries;
618                  * if details->nonlinear_vma, we leave file entries.
619                  */
620                 if (unlikely(details))
621                         continue;
622                 if (!pte_file(ptent))
623                         free_swap_and_cache(pte_to_swp_entry(ptent));
624                 pte_clear_full(mm, addr, pte, tlb->fullmm);
625         } while (pte++, addr += PAGE_SIZE, addr != end);
626
627         add_mm_rss(mm, file_rss, anon_rss);
628         pte_unmap_unlock(pte - 1, ptl);
629 }
630
631 static inline void zap_pmd_range(struct mmu_gather *tlb,
632                                 struct vm_area_struct *vma, pud_t *pud,
633                                 unsigned long addr, unsigned long end,
634                                 struct zap_details *details)
635 {
636         pmd_t *pmd;
637         unsigned long next;
638
639         pmd = pmd_offset(pud, addr);
640         do {
641                 next = pmd_addr_end(addr, end);
642                 if (pmd_none_or_clear_bad(pmd))
643                         continue;
644                 zap_pte_range(tlb, vma, pmd, addr, next, details);
645         } while (pmd++, addr = next, addr != end);
646 }
647
648 static inline void zap_pud_range(struct mmu_gather *tlb,
649                                 struct vm_area_struct *vma, pgd_t *pgd,
650                                 unsigned long addr, unsigned long end,
651                                 struct zap_details *details)
652 {
653         pud_t *pud;
654         unsigned long next;
655
656         pud = pud_offset(pgd, addr);
657         do {
658                 next = pud_addr_end(addr, end);
659                 if (pud_none_or_clear_bad(pud))
660                         continue;
661                 zap_pmd_range(tlb, vma, pud, addr, next, details);
662         } while (pud++, addr = next, addr != end);
663 }
664
665 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
666                                 unsigned long addr, unsigned long end,
667                                 struct zap_details *details)
668 {
669         pgd_t *pgd;
670         unsigned long next;
671
672         if (details && !details->check_mapping && !details->nonlinear_vma)
673                 details = NULL;
674
675         BUG_ON(addr >= end);
676         tlb_start_vma(tlb, vma);
677         pgd = pgd_offset(vma->vm_mm, addr);
678         do {
679                 next = pgd_addr_end(addr, end);
680                 if (pgd_none_or_clear_bad(pgd))
681                         continue;
682                 zap_pud_range(tlb, vma, pgd, addr, next, details);
683         } while (pgd++, addr = next, addr != end);
684         tlb_end_vma(tlb, vma);
685 }
686
687 #ifdef CONFIG_PREEMPT
688 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
689 #else
690 /* No preempt: go for improved straight-line efficiency */
691 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
692 #endif
693
694 /**
695  * unmap_vmas - unmap a range of memory covered by a list of vma's
696  * @tlbp: address of the caller's struct mmu_gather
697  * @vma: the starting vma
698  * @start_addr: virtual address at which to start unmapping
699  * @end_addr: virtual address at which to end unmapping
700  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
701  * @details: details of nonlinear truncation or shared cache invalidation
702  *
703  * Returns the end address of the unmapping (restart addr if interrupted).
704  *
705  * Unmap all pages in the vma list.
706  *
707  * We aim to not hold locks for too long (for scheduling latency reasons).
708  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
709  * return the ending mmu_gather to the caller.
710  *
711  * Only addresses between `start' and `end' will be unmapped.
712  *
713  * The VMA list must be sorted in ascending virtual address order.
714  *
715  * unmap_vmas() assumes that the caller will flush the whole unmapped address
716  * range after unmap_vmas() returns.  So the only responsibility here is to
717  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
718  * drops the lock and schedules.
719  */
720 unsigned long unmap_vmas(struct mmu_gather **tlbp,
721                 struct vm_area_struct *vma, unsigned long start_addr,
722                 unsigned long end_addr, unsigned long *nr_accounted,
723                 struct zap_details *details)
724 {
725         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
726         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
727         int tlb_start_valid = 0;
728         unsigned long start = start_addr;
729         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
730         int fullmm = (*tlbp)->fullmm;
731
732         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
733                 unsigned long end;
734
735                 start = max(vma->vm_start, start_addr);
736                 if (start >= vma->vm_end)
737                         continue;
738                 end = min(vma->vm_end, end_addr);
739                 if (end <= vma->vm_start)
740                         continue;
741
742                 if (vma->vm_flags & VM_ACCOUNT)
743                         *nr_accounted += (end - start) >> PAGE_SHIFT;
744
745                 while (start != end) {
746                         unsigned long block;
747
748                         if (!tlb_start_valid) {
749                                 tlb_start = start;
750                                 tlb_start_valid = 1;
751                         }
752
753                         if (is_vm_hugetlb_page(vma)) {
754                                 block = end - start;
755                                 unmap_hugepage_range(vma, start, end);
756                         } else {
757                                 block = min(zap_bytes, end - start);
758                                 unmap_page_range(*tlbp, vma, start,
759                                                 start + block, details);
760                         }
761
762                         start += block;
763                         zap_bytes -= block;
764                         if ((long)zap_bytes > 0)
765                                 continue;
766
767                         tlb_finish_mmu(*tlbp, tlb_start, start);
768
769                         if (need_resched() ||
770                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
771                                 if (i_mmap_lock) {
772                                         *tlbp = NULL;
773                                         goto out;
774                                 }
775                                 cond_resched();
776                         }
777
778                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
779                         tlb_start_valid = 0;
780                         zap_bytes = ZAP_BLOCK_SIZE;
781                 }
782         }
783 out:
784         return start;   /* which is now the end (or restart) address */
785 }
786
787 /**
788  * zap_page_range - remove user pages in a given range
789  * @vma: vm_area_struct holding the applicable pages
790  * @address: starting address of pages to zap
791  * @size: number of bytes to zap
792  * @details: details of nonlinear truncation or shared cache invalidation
793  */
794 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
795                 unsigned long size, struct zap_details *details)
796 {
797         struct mm_struct *mm = vma->vm_mm;
798         struct mmu_gather *tlb;
799         unsigned long end = address + size;
800         unsigned long nr_accounted = 0;
801
802         lru_add_drain();
803         tlb = tlb_gather_mmu(mm, 0);
804         update_hiwater_rss(mm);
805         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
806         if (tlb)
807                 tlb_finish_mmu(tlb, address, end);
808         return end;
809 }
810
811 /*
812  * Do a quick page-table lookup for a single page.
813  */
814 struct page *follow_page(struct mm_struct *mm, unsigned long address,
815                         unsigned int flags)
816 {
817         pgd_t *pgd;
818         pud_t *pud;
819         pmd_t *pmd;
820         pte_t *ptep, pte;
821         spinlock_t *ptl;
822         unsigned long pfn;
823         struct page *page;
824
825         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
826         if (!IS_ERR(page)) {
827                 BUG_ON(flags & FOLL_GET);
828                 goto out;
829         }
830
831         page = NULL;
832         pgd = pgd_offset(mm, address);
833         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
834                 goto no_page_table;
835
836         pud = pud_offset(pgd, address);
837         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
838                 goto no_page_table;
839         
840         pmd = pmd_offset(pud, address);
841         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
842                 goto no_page_table;
843
844         if (pmd_huge(*pmd)) {
845                 BUG_ON(flags & FOLL_GET);
846                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
847                 goto out;
848         }
849
850         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
851         if (!ptep)
852                 goto out;
853
854         pte = *ptep;
855         if (!pte_present(pte))
856                 goto unlock;
857         if ((flags & FOLL_WRITE) && !pte_write(pte))
858                 goto unlock;
859         pfn = pte_pfn(pte);
860         if (!pfn_valid(pfn))
861                 goto unlock;
862
863         page = pfn_to_page(pfn);
864         if (flags & FOLL_GET)
865                 get_page(page);
866         if (flags & FOLL_TOUCH) {
867                 if ((flags & FOLL_WRITE) &&
868                     !pte_dirty(pte) && !PageDirty(page))
869                         set_page_dirty(page);
870                 mark_page_accessed(page);
871         }
872 unlock:
873         pte_unmap_unlock(ptep, ptl);
874 out:
875         return page;
876
877 no_page_table:
878         /*
879          * When core dumping an enormous anonymous area that nobody
880          * has touched so far, we don't want to allocate page tables.
881          */
882         if (flags & FOLL_ANON) {
883                 page = ZERO_PAGE(address);
884                 if (flags & FOLL_GET)
885                         get_page(page);
886                 BUG_ON(flags & FOLL_WRITE);
887         }
888         return page;
889 }
890
891 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
892                 unsigned long start, int len, int write, int force,
893                 struct page **pages, struct vm_area_struct **vmas)
894 {
895         int i;
896         unsigned int vm_flags;
897
898         /* 
899          * Require read or write permissions.
900          * If 'force' is set, we only require the "MAY" flags.
901          */
902         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
903         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
904         i = 0;
905
906         do {
907                 struct vm_area_struct *vma;
908                 unsigned int foll_flags;
909
910                 vma = find_extend_vma(mm, start);
911                 if (!vma && in_gate_area(tsk, start)) {
912                         unsigned long pg = start & PAGE_MASK;
913                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
914                         pgd_t *pgd;
915                         pud_t *pud;
916                         pmd_t *pmd;
917                         pte_t *pte;
918                         if (write) /* user gate pages are read-only */
919                                 return i ? : -EFAULT;
920                         if (pg > TASK_SIZE)
921                                 pgd = pgd_offset_k(pg);
922                         else
923                                 pgd = pgd_offset_gate(mm, pg);
924                         BUG_ON(pgd_none(*pgd));
925                         pud = pud_offset(pgd, pg);
926                         BUG_ON(pud_none(*pud));
927                         pmd = pmd_offset(pud, pg);
928                         if (pmd_none(*pmd))
929                                 return i ? : -EFAULT;
930                         pte = pte_offset_map(pmd, pg);
931                         if (pte_none(*pte)) {
932                                 pte_unmap(pte);
933                                 return i ? : -EFAULT;
934                         }
935                         if (pages) {
936                                 pages[i] = pte_page(*pte);
937                                 get_page(pages[i]);
938                         }
939                         pte_unmap(pte);
940                         if (vmas)
941                                 vmas[i] = gate_vma;
942                         i++;
943                         start += PAGE_SIZE;
944                         len--;
945                         continue;
946                 }
947
948                 if (!vma || (vma->vm_flags & (VM_IO | VM_RESERVED))
949                                 || !(vm_flags & vma->vm_flags))
950                         return i ? : -EFAULT;
951
952                 if (is_vm_hugetlb_page(vma)) {
953                         i = follow_hugetlb_page(mm, vma, pages, vmas,
954                                                 &start, &len, i);
955                         continue;
956                 }
957
958                 foll_flags = FOLL_TOUCH;
959                 if (pages)
960                         foll_flags |= FOLL_GET;
961                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
962                     (!vma->vm_ops || !vma->vm_ops->nopage))
963                         foll_flags |= FOLL_ANON;
964
965                 do {
966                         struct page *page;
967
968                         if (write)
969                                 foll_flags |= FOLL_WRITE;
970
971                         cond_resched();
972                         while (!(page = follow_page(mm, start, foll_flags))) {
973                                 int ret;
974                                 ret = __handle_mm_fault(mm, vma, start,
975                                                 foll_flags & FOLL_WRITE);
976                                 /*
977                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
978                                  * broken COW when necessary, even if maybe_mkwrite
979                                  * decided not to set pte_write. We can thus safely do
980                                  * subsequent page lookups as if they were reads.
981                                  */
982                                 if (ret & VM_FAULT_WRITE)
983                                         foll_flags &= ~FOLL_WRITE;
984                                 
985                                 switch (ret & ~VM_FAULT_WRITE) {
986                                 case VM_FAULT_MINOR:
987                                         tsk->min_flt++;
988                                         break;
989                                 case VM_FAULT_MAJOR:
990                                         tsk->maj_flt++;
991                                         break;
992                                 case VM_FAULT_SIGBUS:
993                                         return i ? i : -EFAULT;
994                                 case VM_FAULT_OOM:
995                                         return i ? i : -ENOMEM;
996                                 default:
997                                         BUG();
998                                 }
999                         }
1000                         if (pages) {
1001                                 pages[i] = page;
1002                                 flush_dcache_page(page);
1003                         }
1004                         if (vmas)
1005                                 vmas[i] = vma;
1006                         i++;
1007                         start += PAGE_SIZE;
1008                         len--;
1009                 } while (len && start < vma->vm_end);
1010         } while (len);
1011         return i;
1012 }
1013 EXPORT_SYMBOL(get_user_pages);
1014
1015 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1016                         unsigned long addr, unsigned long end, pgprot_t prot)
1017 {
1018         pte_t *pte;
1019         spinlock_t *ptl;
1020
1021         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1022         if (!pte)
1023                 return -ENOMEM;
1024         do {
1025                 struct page *page = ZERO_PAGE(addr);
1026                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1027                 page_cache_get(page);
1028                 page_add_file_rmap(page);
1029                 inc_mm_counter(mm, file_rss);
1030                 BUG_ON(!pte_none(*pte));
1031                 set_pte_at(mm, addr, pte, zero_pte);
1032         } while (pte++, addr += PAGE_SIZE, addr != end);
1033         pte_unmap_unlock(pte - 1, ptl);
1034         return 0;
1035 }
1036
1037 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1038                         unsigned long addr, unsigned long end, pgprot_t prot)
1039 {
1040         pmd_t *pmd;
1041         unsigned long next;
1042
1043         pmd = pmd_alloc(mm, pud, addr);
1044         if (!pmd)
1045                 return -ENOMEM;
1046         do {
1047                 next = pmd_addr_end(addr, end);
1048                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1049                         return -ENOMEM;
1050         } while (pmd++, addr = next, addr != end);
1051         return 0;
1052 }
1053
1054 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1055                         unsigned long addr, unsigned long end, pgprot_t prot)
1056 {
1057         pud_t *pud;
1058         unsigned long next;
1059
1060         pud = pud_alloc(mm, pgd, addr);
1061         if (!pud)
1062                 return -ENOMEM;
1063         do {
1064                 next = pud_addr_end(addr, end);
1065                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1066                         return -ENOMEM;
1067         } while (pud++, addr = next, addr != end);
1068         return 0;
1069 }
1070
1071 int zeromap_page_range(struct vm_area_struct *vma,
1072                         unsigned long addr, unsigned long size, pgprot_t prot)
1073 {
1074         pgd_t *pgd;
1075         unsigned long next;
1076         unsigned long end = addr + size;
1077         struct mm_struct *mm = vma->vm_mm;
1078         int err;
1079
1080         BUG_ON(addr >= end);
1081         pgd = pgd_offset(mm, addr);
1082         flush_cache_range(vma, addr, end);
1083         do {
1084                 next = pgd_addr_end(addr, end);
1085                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1086                 if (err)
1087                         break;
1088         } while (pgd++, addr = next, addr != end);
1089         return err;
1090 }
1091
1092 /*
1093  * maps a range of physical memory into the requested pages. the old
1094  * mappings are removed. any references to nonexistent pages results
1095  * in null mappings (currently treated as "copy-on-access")
1096  */
1097 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1098                         unsigned long addr, unsigned long end,
1099                         unsigned long pfn, pgprot_t prot)
1100 {
1101         pte_t *pte;
1102         spinlock_t *ptl;
1103
1104         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1105         if (!pte)
1106                 return -ENOMEM;
1107         do {
1108                 BUG_ON(!pte_none(*pte));
1109                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1110                 pfn++;
1111         } while (pte++, addr += PAGE_SIZE, addr != end);
1112         pte_unmap_unlock(pte - 1, ptl);
1113         return 0;
1114 }
1115
1116 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1117                         unsigned long addr, unsigned long end,
1118                         unsigned long pfn, pgprot_t prot)
1119 {
1120         pmd_t *pmd;
1121         unsigned long next;
1122
1123         pfn -= addr >> PAGE_SHIFT;
1124         pmd = pmd_alloc(mm, pud, addr);
1125         if (!pmd)
1126                 return -ENOMEM;
1127         do {
1128                 next = pmd_addr_end(addr, end);
1129                 if (remap_pte_range(mm, pmd, addr, next,
1130                                 pfn + (addr >> PAGE_SHIFT), prot))
1131                         return -ENOMEM;
1132         } while (pmd++, addr = next, addr != end);
1133         return 0;
1134 }
1135
1136 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1137                         unsigned long addr, unsigned long end,
1138                         unsigned long pfn, pgprot_t prot)
1139 {
1140         pud_t *pud;
1141         unsigned long next;
1142
1143         pfn -= addr >> PAGE_SHIFT;
1144         pud = pud_alloc(mm, pgd, addr);
1145         if (!pud)
1146                 return -ENOMEM;
1147         do {
1148                 next = pud_addr_end(addr, end);
1149                 if (remap_pmd_range(mm, pud, addr, next,
1150                                 pfn + (addr >> PAGE_SHIFT), prot))
1151                         return -ENOMEM;
1152         } while (pud++, addr = next, addr != end);
1153         return 0;
1154 }
1155
1156 /*  Note: this is only safe if the mm semaphore is held when called. */
1157 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1158                     unsigned long pfn, unsigned long size, pgprot_t prot)
1159 {
1160         pgd_t *pgd;
1161         unsigned long next;
1162         unsigned long end = addr + PAGE_ALIGN(size);
1163         struct mm_struct *mm = vma->vm_mm;
1164         int err;
1165
1166         /*
1167          * Physically remapped pages are special. Tell the
1168          * rest of the world about it:
1169          *   VM_IO tells people not to look at these pages
1170          *      (accesses can have side effects).
1171          *   VM_RESERVED tells the core MM not to "manage" these pages
1172          *      (e.g. refcount, mapcount, try to swap them out).
1173          */
1174         vma->vm_flags |= VM_IO | VM_RESERVED;
1175
1176         BUG_ON(addr >= end);
1177         pfn -= addr >> PAGE_SHIFT;
1178         pgd = pgd_offset(mm, addr);
1179         flush_cache_range(vma, addr, end);
1180         do {
1181                 next = pgd_addr_end(addr, end);
1182                 err = remap_pud_range(mm, pgd, addr, next,
1183                                 pfn + (addr >> PAGE_SHIFT), prot);
1184                 if (err)
1185                         break;
1186         } while (pgd++, addr = next, addr != end);
1187         return err;
1188 }
1189 EXPORT_SYMBOL(remap_pfn_range);
1190
1191 /*
1192  * handle_pte_fault chooses page fault handler according to an entry
1193  * which was read non-atomically.  Before making any commitment, on
1194  * those architectures or configurations (e.g. i386 with PAE) which
1195  * might give a mix of unmatched parts, do_swap_page and do_file_page
1196  * must check under lock before unmapping the pte and proceeding
1197  * (but do_wp_page is only called after already making such a check;
1198  * and do_anonymous_page and do_no_page can safely check later on).
1199  */
1200 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1201                                 pte_t *page_table, pte_t orig_pte)
1202 {
1203         int same = 1;
1204 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1205         if (sizeof(pte_t) > sizeof(unsigned long)) {
1206                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1207                 spin_lock(ptl);
1208                 same = pte_same(*page_table, orig_pte);
1209                 spin_unlock(ptl);
1210         }
1211 #endif
1212         pte_unmap(page_table);
1213         return same;
1214 }
1215
1216 /*
1217  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1218  * servicing faults for write access.  In the normal case, do always want
1219  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1220  * that do not have writing enabled, when used by access_process_vm.
1221  */
1222 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1223 {
1224         if (likely(vma->vm_flags & VM_WRITE))
1225                 pte = pte_mkwrite(pte);
1226         return pte;
1227 }
1228
1229 /*
1230  * This routine handles present pages, when users try to write
1231  * to a shared page. It is done by copying the page to a new address
1232  * and decrementing the shared-page counter for the old page.
1233  *
1234  * Note that this routine assumes that the protection checks have been
1235  * done by the caller (the low-level page fault routine in most cases).
1236  * Thus we can safely just mark it writable once we've done any necessary
1237  * COW.
1238  *
1239  * We also mark the page dirty at this point even though the page will
1240  * change only once the write actually happens. This avoids a few races,
1241  * and potentially makes it more efficient.
1242  *
1243  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1244  * but allow concurrent faults), with pte both mapped and locked.
1245  * We return with mmap_sem still held, but pte unmapped and unlocked.
1246  */
1247 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1248                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1249                 spinlock_t *ptl, pte_t orig_pte)
1250 {
1251         struct page *old_page, *new_page;
1252         unsigned long pfn = pte_pfn(orig_pte);
1253         pte_t entry;
1254         int ret = VM_FAULT_MINOR;
1255
1256         BUG_ON(vma->vm_flags & VM_RESERVED);
1257
1258         if (unlikely(!pfn_valid(pfn))) {
1259                 /*
1260                  * Page table corrupted: show pte and kill process.
1261                  */
1262                 print_bad_pte(vma, orig_pte, address);
1263                 ret = VM_FAULT_OOM;
1264                 goto unlock;
1265         }
1266         old_page = pfn_to_page(pfn);
1267
1268         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1269                 int reuse = can_share_swap_page(old_page);
1270                 unlock_page(old_page);
1271                 if (reuse) {
1272                         flush_cache_page(vma, address, pfn);
1273                         entry = pte_mkyoung(orig_pte);
1274                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1275                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1276                         update_mmu_cache(vma, address, entry);
1277                         lazy_mmu_prot_update(entry);
1278                         ret |= VM_FAULT_WRITE;
1279                         goto unlock;
1280                 }
1281         }
1282
1283         /*
1284          * Ok, we need to copy. Oh, well..
1285          */
1286         page_cache_get(old_page);
1287         pte_unmap_unlock(page_table, ptl);
1288
1289         if (unlikely(anon_vma_prepare(vma)))
1290                 goto oom;
1291         if (old_page == ZERO_PAGE(address)) {
1292                 new_page = alloc_zeroed_user_highpage(vma, address);
1293                 if (!new_page)
1294                         goto oom;
1295         } else {
1296                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1297                 if (!new_page)
1298                         goto oom;
1299                 copy_user_highpage(new_page, old_page, address);
1300         }
1301
1302         /*
1303          * Re-check the pte - we dropped the lock
1304          */
1305         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1306         if (likely(pte_same(*page_table, orig_pte))) {
1307                 page_remove_rmap(old_page);
1308                 if (!PageAnon(old_page)) {
1309                         inc_mm_counter(mm, anon_rss);
1310                         dec_mm_counter(mm, file_rss);
1311                 }
1312                 flush_cache_page(vma, address, pfn);
1313                 entry = mk_pte(new_page, vma->vm_page_prot);
1314                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1315                 ptep_establish(vma, address, page_table, entry);
1316                 update_mmu_cache(vma, address, entry);
1317                 lazy_mmu_prot_update(entry);
1318                 lru_cache_add_active(new_page);
1319                 page_add_anon_rmap(new_page, vma, address);
1320
1321                 /* Free the old page.. */
1322                 new_page = old_page;
1323                 ret |= VM_FAULT_WRITE;
1324         }
1325         page_cache_release(new_page);
1326         page_cache_release(old_page);
1327 unlock:
1328         pte_unmap_unlock(page_table, ptl);
1329         return ret;
1330 oom:
1331         page_cache_release(old_page);
1332         return VM_FAULT_OOM;
1333 }
1334
1335 /*
1336  * Helper functions for unmap_mapping_range().
1337  *
1338  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1339  *
1340  * We have to restart searching the prio_tree whenever we drop the lock,
1341  * since the iterator is only valid while the lock is held, and anyway
1342  * a later vma might be split and reinserted earlier while lock dropped.
1343  *
1344  * The list of nonlinear vmas could be handled more efficiently, using
1345  * a placeholder, but handle it in the same way until a need is shown.
1346  * It is important to search the prio_tree before nonlinear list: a vma
1347  * may become nonlinear and be shifted from prio_tree to nonlinear list
1348  * while the lock is dropped; but never shifted from list to prio_tree.
1349  *
1350  * In order to make forward progress despite restarting the search,
1351  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1352  * quickly skip it next time around.  Since the prio_tree search only
1353  * shows us those vmas affected by unmapping the range in question, we
1354  * can't efficiently keep all vmas in step with mapping->truncate_count:
1355  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1356  * mapping->truncate_count and vma->vm_truncate_count are protected by
1357  * i_mmap_lock.
1358  *
1359  * In order to make forward progress despite repeatedly restarting some
1360  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1361  * and restart from that address when we reach that vma again.  It might
1362  * have been split or merged, shrunk or extended, but never shifted: so
1363  * restart_addr remains valid so long as it remains in the vma's range.
1364  * unmap_mapping_range forces truncate_count to leap over page-aligned
1365  * values so we can save vma's restart_addr in its truncate_count field.
1366  */
1367 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1368
1369 static void reset_vma_truncate_counts(struct address_space *mapping)
1370 {
1371         struct vm_area_struct *vma;
1372         struct prio_tree_iter iter;
1373
1374         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1375                 vma->vm_truncate_count = 0;
1376         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1377                 vma->vm_truncate_count = 0;
1378 }
1379
1380 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1381                 unsigned long start_addr, unsigned long end_addr,
1382                 struct zap_details *details)
1383 {
1384         unsigned long restart_addr;
1385         int need_break;
1386
1387 again:
1388         restart_addr = vma->vm_truncate_count;
1389         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1390                 start_addr = restart_addr;
1391                 if (start_addr >= end_addr) {
1392                         /* Top of vma has been split off since last time */
1393                         vma->vm_truncate_count = details->truncate_count;
1394                         return 0;
1395                 }
1396         }
1397
1398         restart_addr = zap_page_range(vma, start_addr,
1399                                         end_addr - start_addr, details);
1400         need_break = need_resched() ||
1401                         need_lockbreak(details->i_mmap_lock);
1402
1403         if (restart_addr >= end_addr) {
1404                 /* We have now completed this vma: mark it so */
1405                 vma->vm_truncate_count = details->truncate_count;
1406                 if (!need_break)
1407                         return 0;
1408         } else {
1409                 /* Note restart_addr in vma's truncate_count field */
1410                 vma->vm_truncate_count = restart_addr;
1411                 if (!need_break)
1412                         goto again;
1413         }
1414
1415         spin_unlock(details->i_mmap_lock);
1416         cond_resched();
1417         spin_lock(details->i_mmap_lock);
1418         return -EINTR;
1419 }
1420
1421 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1422                                             struct zap_details *details)
1423 {
1424         struct vm_area_struct *vma;
1425         struct prio_tree_iter iter;
1426         pgoff_t vba, vea, zba, zea;
1427
1428 restart:
1429         vma_prio_tree_foreach(vma, &iter, root,
1430                         details->first_index, details->last_index) {
1431                 /* Skip quickly over those we have already dealt with */
1432                 if (vma->vm_truncate_count == details->truncate_count)
1433                         continue;
1434
1435                 vba = vma->vm_pgoff;
1436                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1437                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1438                 zba = details->first_index;
1439                 if (zba < vba)
1440                         zba = vba;
1441                 zea = details->last_index;
1442                 if (zea > vea)
1443                         zea = vea;
1444
1445                 if (unmap_mapping_range_vma(vma,
1446                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1447                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1448                                 details) < 0)
1449                         goto restart;
1450         }
1451 }
1452
1453 static inline void unmap_mapping_range_list(struct list_head *head,
1454                                             struct zap_details *details)
1455 {
1456         struct vm_area_struct *vma;
1457
1458         /*
1459          * In nonlinear VMAs there is no correspondence between virtual address
1460          * offset and file offset.  So we must perform an exhaustive search
1461          * across *all* the pages in each nonlinear VMA, not just the pages
1462          * whose virtual address lies outside the file truncation point.
1463          */
1464 restart:
1465         list_for_each_entry(vma, head, shared.vm_set.list) {
1466                 /* Skip quickly over those we have already dealt with */
1467                 if (vma->vm_truncate_count == details->truncate_count)
1468                         continue;
1469                 details->nonlinear_vma = vma;
1470                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1471                                         vma->vm_end, details) < 0)
1472                         goto restart;
1473         }
1474 }
1475
1476 /**
1477  * unmap_mapping_range - unmap the portion of all mmaps
1478  * in the specified address_space corresponding to the specified
1479  * page range in the underlying file.
1480  * @mapping: the address space containing mmaps to be unmapped.
1481  * @holebegin: byte in first page to unmap, relative to the start of
1482  * the underlying file.  This will be rounded down to a PAGE_SIZE
1483  * boundary.  Note that this is different from vmtruncate(), which
1484  * must keep the partial page.  In contrast, we must get rid of
1485  * partial pages.
1486  * @holelen: size of prospective hole in bytes.  This will be rounded
1487  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1488  * end of the file.
1489  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1490  * but 0 when invalidating pagecache, don't throw away private data.
1491  */
1492 void unmap_mapping_range(struct address_space *mapping,
1493                 loff_t const holebegin, loff_t const holelen, int even_cows)
1494 {
1495         struct zap_details details;
1496         pgoff_t hba = holebegin >> PAGE_SHIFT;
1497         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1498
1499         /* Check for overflow. */
1500         if (sizeof(holelen) > sizeof(hlen)) {
1501                 long long holeend =
1502                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1503                 if (holeend & ~(long long)ULONG_MAX)
1504                         hlen = ULONG_MAX - hba + 1;
1505         }
1506
1507         details.check_mapping = even_cows? NULL: mapping;
1508         details.nonlinear_vma = NULL;
1509         details.first_index = hba;
1510         details.last_index = hba + hlen - 1;
1511         if (details.last_index < details.first_index)
1512                 details.last_index = ULONG_MAX;
1513         details.i_mmap_lock = &mapping->i_mmap_lock;
1514
1515         spin_lock(&mapping->i_mmap_lock);
1516
1517         /* serialize i_size write against truncate_count write */
1518         smp_wmb();
1519         /* Protect against page faults, and endless unmapping loops */
1520         mapping->truncate_count++;
1521         /*
1522          * For archs where spin_lock has inclusive semantics like ia64
1523          * this smp_mb() will prevent to read pagetable contents
1524          * before the truncate_count increment is visible to
1525          * other cpus.
1526          */
1527         smp_mb();
1528         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1529                 if (mapping->truncate_count == 0)
1530                         reset_vma_truncate_counts(mapping);
1531                 mapping->truncate_count++;
1532         }
1533         details.truncate_count = mapping->truncate_count;
1534
1535         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1536                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1537         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1538                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1539         spin_unlock(&mapping->i_mmap_lock);
1540 }
1541 EXPORT_SYMBOL(unmap_mapping_range);
1542
1543 /*
1544  * Handle all mappings that got truncated by a "truncate()"
1545  * system call.
1546  *
1547  * NOTE! We have to be ready to update the memory sharing
1548  * between the file and the memory map for a potential last
1549  * incomplete page.  Ugly, but necessary.
1550  */
1551 int vmtruncate(struct inode * inode, loff_t offset)
1552 {
1553         struct address_space *mapping = inode->i_mapping;
1554         unsigned long limit;
1555
1556         if (inode->i_size < offset)
1557                 goto do_expand;
1558         /*
1559          * truncation of in-use swapfiles is disallowed - it would cause
1560          * subsequent swapout to scribble on the now-freed blocks.
1561          */
1562         if (IS_SWAPFILE(inode))
1563                 goto out_busy;
1564         i_size_write(inode, offset);
1565         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1566         truncate_inode_pages(mapping, offset);
1567         goto out_truncate;
1568
1569 do_expand:
1570         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1571         if (limit != RLIM_INFINITY && offset > limit)
1572                 goto out_sig;
1573         if (offset > inode->i_sb->s_maxbytes)
1574                 goto out_big;
1575         i_size_write(inode, offset);
1576
1577 out_truncate:
1578         if (inode->i_op && inode->i_op->truncate)
1579                 inode->i_op->truncate(inode);
1580         return 0;
1581 out_sig:
1582         send_sig(SIGXFSZ, current, 0);
1583 out_big:
1584         return -EFBIG;
1585 out_busy:
1586         return -ETXTBSY;
1587 }
1588
1589 EXPORT_SYMBOL(vmtruncate);
1590
1591 /* 
1592  * Primitive swap readahead code. We simply read an aligned block of
1593  * (1 << page_cluster) entries in the swap area. This method is chosen
1594  * because it doesn't cost us any seek time.  We also make sure to queue
1595  * the 'original' request together with the readahead ones...  
1596  *
1597  * This has been extended to use the NUMA policies from the mm triggering
1598  * the readahead.
1599  *
1600  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1601  */
1602 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1603 {
1604 #ifdef CONFIG_NUMA
1605         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1606 #endif
1607         int i, num;
1608         struct page *new_page;
1609         unsigned long offset;
1610
1611         /*
1612          * Get the number of handles we should do readahead io to.
1613          */
1614         num = valid_swaphandles(entry, &offset);
1615         for (i = 0; i < num; offset++, i++) {
1616                 /* Ok, do the async read-ahead now */
1617                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1618                                                            offset), vma, addr);
1619                 if (!new_page)
1620                         break;
1621                 page_cache_release(new_page);
1622 #ifdef CONFIG_NUMA
1623                 /*
1624                  * Find the next applicable VMA for the NUMA policy.
1625                  */
1626                 addr += PAGE_SIZE;
1627                 if (addr == 0)
1628                         vma = NULL;
1629                 if (vma) {
1630                         if (addr >= vma->vm_end) {
1631                                 vma = next_vma;
1632                                 next_vma = vma ? vma->vm_next : NULL;
1633                         }
1634                         if (vma && addr < vma->vm_start)
1635                                 vma = NULL;
1636                 } else {
1637                         if (next_vma && addr >= next_vma->vm_start) {
1638                                 vma = next_vma;
1639                                 next_vma = vma->vm_next;
1640                         }
1641                 }
1642 #endif
1643         }
1644         lru_add_drain();        /* Push any new pages onto the LRU now */
1645 }
1646
1647 /*
1648  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1649  * but allow concurrent faults), and pte mapped but not yet locked.
1650  * We return with mmap_sem still held, but pte unmapped and unlocked.
1651  */
1652 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1653                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1654                 int write_access, pte_t orig_pte)
1655 {
1656         spinlock_t *ptl;
1657         struct page *page;
1658         swp_entry_t entry;
1659         pte_t pte;
1660         int ret = VM_FAULT_MINOR;
1661
1662         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1663                 goto out;
1664
1665         entry = pte_to_swp_entry(orig_pte);
1666         page = lookup_swap_cache(entry);
1667         if (!page) {
1668                 swapin_readahead(entry, address, vma);
1669                 page = read_swap_cache_async(entry, vma, address);
1670                 if (!page) {
1671                         /*
1672                          * Back out if somebody else faulted in this pte
1673                          * while we released the pte lock.
1674                          */
1675                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1676                         if (likely(pte_same(*page_table, orig_pte)))
1677                                 ret = VM_FAULT_OOM;
1678                         goto unlock;
1679                 }
1680
1681                 /* Had to read the page from swap area: Major fault */
1682                 ret = VM_FAULT_MAJOR;
1683                 inc_page_state(pgmajfault);
1684                 grab_swap_token();
1685         }
1686
1687         mark_page_accessed(page);
1688         lock_page(page);
1689
1690         /*
1691          * Back out if somebody else already faulted in this pte.
1692          */
1693         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1694         if (unlikely(!pte_same(*page_table, orig_pte)))
1695                 goto out_nomap;
1696
1697         if (unlikely(!PageUptodate(page))) {
1698                 ret = VM_FAULT_SIGBUS;
1699                 goto out_nomap;
1700         }
1701
1702         /* The page isn't present yet, go ahead with the fault. */
1703
1704         inc_mm_counter(mm, anon_rss);
1705         pte = mk_pte(page, vma->vm_page_prot);
1706         if (write_access && can_share_swap_page(page)) {
1707                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1708                 write_access = 0;
1709         }
1710
1711         flush_icache_page(vma, page);
1712         set_pte_at(mm, address, page_table, pte);
1713         page_add_anon_rmap(page, vma, address);
1714
1715         swap_free(entry);
1716         if (vm_swap_full())
1717                 remove_exclusive_swap_page(page);
1718         unlock_page(page);
1719
1720         if (write_access) {
1721                 if (do_wp_page(mm, vma, address,
1722                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1723                         ret = VM_FAULT_OOM;
1724                 goto out;
1725         }
1726
1727         /* No need to invalidate - it was non-present before */
1728         update_mmu_cache(vma, address, pte);
1729         lazy_mmu_prot_update(pte);
1730 unlock:
1731         pte_unmap_unlock(page_table, ptl);
1732 out:
1733         return ret;
1734 out_nomap:
1735         pte_unmap_unlock(page_table, ptl);
1736         unlock_page(page);
1737         page_cache_release(page);
1738         return ret;
1739 }
1740
1741 /*
1742  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1743  * but allow concurrent faults), and pte mapped but not yet locked.
1744  * We return with mmap_sem still held, but pte unmapped and unlocked.
1745  */
1746 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1747                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1748                 int write_access)
1749 {
1750         struct page *page;
1751         spinlock_t *ptl;
1752         pte_t entry;
1753
1754         if (write_access) {
1755                 /* Allocate our own private page. */
1756                 pte_unmap(page_table);
1757
1758                 if (unlikely(anon_vma_prepare(vma)))
1759                         goto oom;
1760                 page = alloc_zeroed_user_highpage(vma, address);
1761                 if (!page)
1762                         goto oom;
1763
1764                 entry = mk_pte(page, vma->vm_page_prot);
1765                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1766
1767                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1768                 if (!pte_none(*page_table))
1769                         goto release;
1770                 inc_mm_counter(mm, anon_rss);
1771                 lru_cache_add_active(page);
1772                 SetPageReferenced(page);
1773                 page_add_anon_rmap(page, vma, address);
1774         } else {
1775                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1776                 page = ZERO_PAGE(address);
1777                 page_cache_get(page);
1778                 entry = mk_pte(page, vma->vm_page_prot);
1779
1780                 ptl = pte_lockptr(mm, pmd);
1781                 spin_lock(ptl);
1782                 if (!pte_none(*page_table))
1783                         goto release;
1784                 inc_mm_counter(mm, file_rss);
1785                 page_add_file_rmap(page);
1786         }
1787
1788         set_pte_at(mm, address, page_table, entry);
1789
1790         /* No need to invalidate - it was non-present before */
1791         update_mmu_cache(vma, address, entry);
1792         lazy_mmu_prot_update(entry);
1793 unlock:
1794         pte_unmap_unlock(page_table, ptl);
1795         return VM_FAULT_MINOR;
1796 release:
1797         page_cache_release(page);
1798         goto unlock;
1799 oom:
1800         return VM_FAULT_OOM;
1801 }
1802
1803 /*
1804  * do_no_page() tries to create a new page mapping. It aggressively
1805  * tries to share with existing pages, but makes a separate copy if
1806  * the "write_access" parameter is true in order to avoid the next
1807  * page fault.
1808  *
1809  * As this is called only for pages that do not currently exist, we
1810  * do not need to flush old virtual caches or the TLB.
1811  *
1812  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1813  * but allow concurrent faults), and pte mapped but not yet locked.
1814  * We return with mmap_sem still held, but pte unmapped and unlocked.
1815  */
1816 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1817                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1818                 int write_access)
1819 {
1820         spinlock_t *ptl;
1821         struct page *new_page;
1822         struct address_space *mapping = NULL;
1823         pte_t entry;
1824         unsigned int sequence = 0;
1825         int ret = VM_FAULT_MINOR;
1826         int anon = 0;
1827
1828         pte_unmap(page_table);
1829
1830         if (vma->vm_file) {
1831                 mapping = vma->vm_file->f_mapping;
1832                 sequence = mapping->truncate_count;
1833                 smp_rmb(); /* serializes i_size against truncate_count */
1834         }
1835 retry:
1836         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1837         /*
1838          * No smp_rmb is needed here as long as there's a full
1839          * spin_lock/unlock sequence inside the ->nopage callback
1840          * (for the pagecache lookup) that acts as an implicit
1841          * smp_mb() and prevents the i_size read to happen
1842          * after the next truncate_count read.
1843          */
1844
1845         /* no page was available -- either SIGBUS or OOM */
1846         if (new_page == NOPAGE_SIGBUS)
1847                 return VM_FAULT_SIGBUS;
1848         if (new_page == NOPAGE_OOM)
1849                 return VM_FAULT_OOM;
1850
1851         /*
1852          * Should we do an early C-O-W break?
1853          */
1854         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1855                 struct page *page;
1856
1857                 if (unlikely(anon_vma_prepare(vma)))
1858                         goto oom;
1859                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1860                 if (!page)
1861                         goto oom;
1862                 copy_user_highpage(page, new_page, address);
1863                 page_cache_release(new_page);
1864                 new_page = page;
1865                 anon = 1;
1866         }
1867
1868         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1869         /*
1870          * For a file-backed vma, someone could have truncated or otherwise
1871          * invalidated this page.  If unmap_mapping_range got called,
1872          * retry getting the page.
1873          */
1874         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1875                 pte_unmap_unlock(page_table, ptl);
1876                 page_cache_release(new_page);
1877                 cond_resched();
1878                 sequence = mapping->truncate_count;
1879                 smp_rmb();
1880                 goto retry;
1881         }
1882
1883         /*
1884          * This silly early PAGE_DIRTY setting removes a race
1885          * due to the bad i386 page protection. But it's valid
1886          * for other architectures too.
1887          *
1888          * Note that if write_access is true, we either now have
1889          * an exclusive copy of the page, or this is a shared mapping,
1890          * so we can make it writable and dirty to avoid having to
1891          * handle that later.
1892          */
1893         /* Only go through if we didn't race with anybody else... */
1894         if (pte_none(*page_table)) {
1895                 flush_icache_page(vma, new_page);
1896                 entry = mk_pte(new_page, vma->vm_page_prot);
1897                 if (write_access)
1898                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1899                 set_pte_at(mm, address, page_table, entry);
1900                 if (anon) {
1901                         inc_mm_counter(mm, anon_rss);
1902                         lru_cache_add_active(new_page);
1903                         page_add_anon_rmap(new_page, vma, address);
1904                 } else if (!(vma->vm_flags & VM_RESERVED)) {
1905                         inc_mm_counter(mm, file_rss);
1906                         page_add_file_rmap(new_page);
1907                 }
1908         } else {
1909                 /* One of our sibling threads was faster, back out. */
1910                 page_cache_release(new_page);
1911                 goto unlock;
1912         }
1913
1914         /* no need to invalidate: a not-present page shouldn't be cached */
1915         update_mmu_cache(vma, address, entry);
1916         lazy_mmu_prot_update(entry);
1917 unlock:
1918         pte_unmap_unlock(page_table, ptl);
1919         return ret;
1920 oom:
1921         page_cache_release(new_page);
1922         return VM_FAULT_OOM;
1923 }
1924
1925 /*
1926  * Fault of a previously existing named mapping. Repopulate the pte
1927  * from the encoded file_pte if possible. This enables swappable
1928  * nonlinear vmas.
1929  *
1930  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1931  * but allow concurrent faults), and pte mapped but not yet locked.
1932  * We return with mmap_sem still held, but pte unmapped and unlocked.
1933  */
1934 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
1935                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1936                 int write_access, pte_t orig_pte)
1937 {
1938         pgoff_t pgoff;
1939         int err;
1940
1941         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1942                 return VM_FAULT_MINOR;
1943
1944         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
1945                 /*
1946                  * Page table corrupted: show pte and kill process.
1947                  */
1948                 print_bad_pte(vma, orig_pte, address);
1949                 return VM_FAULT_OOM;
1950         }
1951         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
1952
1953         pgoff = pte_to_pgoff(orig_pte);
1954         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
1955                                         vma->vm_page_prot, pgoff, 0);
1956         if (err == -ENOMEM)
1957                 return VM_FAULT_OOM;
1958         if (err)
1959                 return VM_FAULT_SIGBUS;
1960         return VM_FAULT_MAJOR;
1961 }
1962
1963 /*
1964  * These routines also need to handle stuff like marking pages dirty
1965  * and/or accessed for architectures that don't do it in hardware (most
1966  * RISC architectures).  The early dirtying is also good on the i386.
1967  *
1968  * There is also a hook called "update_mmu_cache()" that architectures
1969  * with external mmu caches can use to update those (ie the Sparc or
1970  * PowerPC hashed page tables that act as extended TLBs).
1971  *
1972  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1973  * but allow concurrent faults), and pte mapped but not yet locked.
1974  * We return with mmap_sem still held, but pte unmapped and unlocked.
1975  */
1976 static inline int handle_pte_fault(struct mm_struct *mm,
1977                 struct vm_area_struct *vma, unsigned long address,
1978                 pte_t *pte, pmd_t *pmd, int write_access)
1979 {
1980         pte_t entry;
1981         spinlock_t *ptl;
1982
1983         entry = *pte;
1984         if (!pte_present(entry)) {
1985                 if (pte_none(entry)) {
1986                         if (!vma->vm_ops || !vma->vm_ops->nopage)
1987                                 return do_anonymous_page(mm, vma, address,
1988                                         pte, pmd, write_access);
1989                         return do_no_page(mm, vma, address,
1990                                         pte, pmd, write_access);
1991                 }
1992                 if (pte_file(entry))
1993                         return do_file_page(mm, vma, address,
1994                                         pte, pmd, write_access, entry);
1995                 return do_swap_page(mm, vma, address,
1996                                         pte, pmd, write_access, entry);
1997         }
1998
1999         ptl = pte_lockptr(mm, pmd);
2000         spin_lock(ptl);
2001         if (unlikely(!pte_same(*pte, entry)))
2002                 goto unlock;
2003         if (write_access) {
2004                 if (!pte_write(entry))
2005                         return do_wp_page(mm, vma, address,
2006                                         pte, pmd, ptl, entry);
2007                 entry = pte_mkdirty(entry);
2008         }
2009         entry = pte_mkyoung(entry);
2010         ptep_set_access_flags(vma, address, pte, entry, write_access);
2011         update_mmu_cache(vma, address, entry);
2012         lazy_mmu_prot_update(entry);
2013 unlock:
2014         pte_unmap_unlock(pte, ptl);
2015         return VM_FAULT_MINOR;
2016 }
2017
2018 /*
2019  * By the time we get here, we already hold the mm semaphore
2020  */
2021 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2022                 unsigned long address, int write_access)
2023 {
2024         pgd_t *pgd;
2025         pud_t *pud;
2026         pmd_t *pmd;
2027         pte_t *pte;
2028
2029         __set_current_state(TASK_RUNNING);
2030
2031         inc_page_state(pgfault);
2032
2033         if (unlikely(is_vm_hugetlb_page(vma)))
2034                 return hugetlb_fault(mm, vma, address, write_access);
2035
2036         pgd = pgd_offset(mm, address);
2037         pud = pud_alloc(mm, pgd, address);
2038         if (!pud)
2039                 return VM_FAULT_OOM;
2040         pmd = pmd_alloc(mm, pud, address);
2041         if (!pmd)
2042                 return VM_FAULT_OOM;
2043         pte = pte_alloc_map(mm, pmd, address);
2044         if (!pte)
2045                 return VM_FAULT_OOM;
2046
2047         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2048 }
2049
2050 #ifndef __PAGETABLE_PUD_FOLDED
2051 /*
2052  * Allocate page upper directory.
2053  * We've already handled the fast-path in-line.
2054  */
2055 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2056 {
2057         pud_t *new = pud_alloc_one(mm, address);
2058         if (!new)
2059                 return -ENOMEM;
2060
2061         spin_lock(&mm->page_table_lock);
2062         if (pgd_present(*pgd))          /* Another has populated it */
2063                 pud_free(new);
2064         else
2065                 pgd_populate(mm, pgd, new);
2066         spin_unlock(&mm->page_table_lock);
2067         return 0;
2068 }
2069 #endif /* __PAGETABLE_PUD_FOLDED */
2070
2071 #ifndef __PAGETABLE_PMD_FOLDED
2072 /*
2073  * Allocate page middle directory.
2074  * We've already handled the fast-path in-line.
2075  */
2076 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2077 {
2078         pmd_t *new = pmd_alloc_one(mm, address);
2079         if (!new)
2080                 return -ENOMEM;
2081
2082         spin_lock(&mm->page_table_lock);
2083 #ifndef __ARCH_HAS_4LEVEL_HACK
2084         if (pud_present(*pud))          /* Another has populated it */
2085                 pmd_free(new);
2086         else
2087                 pud_populate(mm, pud, new);
2088 #else
2089         if (pgd_present(*pud))          /* Another has populated it */
2090                 pmd_free(new);
2091         else
2092                 pgd_populate(mm, pud, new);
2093 #endif /* __ARCH_HAS_4LEVEL_HACK */
2094         spin_unlock(&mm->page_table_lock);
2095         return 0;
2096 }
2097 #endif /* __PAGETABLE_PMD_FOLDED */
2098
2099 int make_pages_present(unsigned long addr, unsigned long end)
2100 {
2101         int ret, len, write;
2102         struct vm_area_struct * vma;
2103
2104         vma = find_vma(current->mm, addr);
2105         if (!vma)
2106                 return -1;
2107         write = (vma->vm_flags & VM_WRITE) != 0;
2108         if (addr >= end)
2109                 BUG();
2110         if (end > vma->vm_end)
2111                 BUG();
2112         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2113         ret = get_user_pages(current, current->mm, addr,
2114                         len, write, 0, NULL, NULL);
2115         if (ret < 0)
2116                 return ret;
2117         return ret == len ? 0 : -1;
2118 }
2119
2120 /* 
2121  * Map a vmalloc()-space virtual address to the physical page.
2122  */
2123 struct page * vmalloc_to_page(void * vmalloc_addr)
2124 {
2125         unsigned long addr = (unsigned long) vmalloc_addr;
2126         struct page *page = NULL;
2127         pgd_t *pgd = pgd_offset_k(addr);
2128         pud_t *pud;
2129         pmd_t *pmd;
2130         pte_t *ptep, pte;
2131   
2132         if (!pgd_none(*pgd)) {
2133                 pud = pud_offset(pgd, addr);
2134                 if (!pud_none(*pud)) {
2135                         pmd = pmd_offset(pud, addr);
2136                         if (!pmd_none(*pmd)) {
2137                                 ptep = pte_offset_map(pmd, addr);
2138                                 pte = *ptep;
2139                                 if (pte_present(pte))
2140                                         page = pte_page(pte);
2141                                 pte_unmap(ptep);
2142                         }
2143                 }
2144         }
2145         return page;
2146 }
2147
2148 EXPORT_SYMBOL(vmalloc_to_page);
2149
2150 /*
2151  * Map a vmalloc()-space virtual address to the physical page frame number.
2152  */
2153 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2154 {
2155         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2156 }
2157
2158 EXPORT_SYMBOL(vmalloc_to_pfn);
2159
2160 #if !defined(__HAVE_ARCH_GATE_AREA)
2161
2162 #if defined(AT_SYSINFO_EHDR)
2163 static struct vm_area_struct gate_vma;
2164
2165 static int __init gate_vma_init(void)
2166 {
2167         gate_vma.vm_mm = NULL;
2168         gate_vma.vm_start = FIXADDR_USER_START;
2169         gate_vma.vm_end = FIXADDR_USER_END;
2170         gate_vma.vm_page_prot = PAGE_READONLY;
2171         gate_vma.vm_flags = VM_RESERVED;
2172         return 0;
2173 }
2174 __initcall(gate_vma_init);
2175 #endif
2176
2177 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2178 {
2179 #ifdef AT_SYSINFO_EHDR
2180         return &gate_vma;
2181 #else
2182         return NULL;
2183 #endif
2184 }
2185
2186 int in_gate_area_no_task(unsigned long addr)
2187 {
2188 #ifdef AT_SYSINFO_EHDR
2189         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2190                 return 1;
2191 #endif
2192         return 0;
2193 }
2194
2195 #endif  /* __HAVE_ARCH_GATE_AREA */