manual update from upstream:
[linux-2.6] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
54 #include <asm/tlb.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
57
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
65
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
69
70 unsigned long num_physpages;
71 /*
72  * A number of key systems in x86 including ioremap() rely on the assumption
73  * that high_memory defines the upper bound on direct map memory, then end
74  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
75  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76  * and ZONE_HIGHMEM.
77  */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
80
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85 /*
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                                 if (list_empty(&dst_mm->mmlist))
376                                         list_add(&dst_mm->mmlist,
377                                                  &src_mm->mmlist);
378                                 spin_unlock(&mmlist_lock);
379                         }
380                 }
381                 goto out_set_pte;
382         }
383
384         /* If the region is VM_RESERVED, the mapping is not
385          * mapped via rmap - duplicate the pte as is.
386          */
387         if (vm_flags & VM_RESERVED)
388                 goto out_set_pte;
389
390         pfn = pte_pfn(pte);
391         /* If the pte points outside of valid memory but
392          * the region is not VM_RESERVED, we have a problem.
393          */
394         if (unlikely(!pfn_valid(pfn))) {
395                 print_bad_pte(vma, pte, addr);
396                 goto out_set_pte; /* try to do something sane */
397         }
398
399         page = pfn_to_page(pfn);
400
401         /*
402          * If it's a COW mapping, write protect it both
403          * in the parent and the child
404          */
405         if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
406                 ptep_set_wrprotect(src_mm, addr, src_pte);
407                 pte = *src_pte;
408         }
409
410         /*
411          * If it's a shared mapping, mark it clean in
412          * the child
413          */
414         if (vm_flags & VM_SHARED)
415                 pte = pte_mkclean(pte);
416         pte = pte_mkold(pte);
417         get_page(page);
418         page_dup_rmap(page);
419         rss[!!PageAnon(page)]++;
420
421 out_set_pte:
422         set_pte_at(dst_mm, addr, dst_pte, pte);
423 }
424
425 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
426                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
427                 unsigned long addr, unsigned long end)
428 {
429         pte_t *src_pte, *dst_pte;
430         spinlock_t *src_ptl, *dst_ptl;
431         int progress = 0;
432         int rss[2];
433
434 again:
435         rss[1] = rss[0] = 0;
436         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
437         if (!dst_pte)
438                 return -ENOMEM;
439         src_pte = pte_offset_map_nested(src_pmd, addr);
440         src_ptl = pte_lockptr(src_mm, src_pmd);
441         spin_lock(src_ptl);
442
443         do {
444                 /*
445                  * We are holding two locks at this point - either of them
446                  * could generate latencies in another task on another CPU.
447                  */
448                 if (progress >= 32) {
449                         progress = 0;
450                         if (need_resched() ||
451                             need_lockbreak(src_ptl) ||
452                             need_lockbreak(dst_ptl))
453                                 break;
454                 }
455                 if (pte_none(*src_pte)) {
456                         progress++;
457                         continue;
458                 }
459                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
460                 progress += 8;
461         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
462
463         spin_unlock(src_ptl);
464         pte_unmap_nested(src_pte - 1);
465         add_mm_rss(dst_mm, rss[0], rss[1]);
466         pte_unmap_unlock(dst_pte - 1, dst_ptl);
467         cond_resched();
468         if (addr != end)
469                 goto again;
470         return 0;
471 }
472
473 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
474                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
475                 unsigned long addr, unsigned long end)
476 {
477         pmd_t *src_pmd, *dst_pmd;
478         unsigned long next;
479
480         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
481         if (!dst_pmd)
482                 return -ENOMEM;
483         src_pmd = pmd_offset(src_pud, addr);
484         do {
485                 next = pmd_addr_end(addr, end);
486                 if (pmd_none_or_clear_bad(src_pmd))
487                         continue;
488                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
489                                                 vma, addr, next))
490                         return -ENOMEM;
491         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
492         return 0;
493 }
494
495 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
496                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
497                 unsigned long addr, unsigned long end)
498 {
499         pud_t *src_pud, *dst_pud;
500         unsigned long next;
501
502         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
503         if (!dst_pud)
504                 return -ENOMEM;
505         src_pud = pud_offset(src_pgd, addr);
506         do {
507                 next = pud_addr_end(addr, end);
508                 if (pud_none_or_clear_bad(src_pud))
509                         continue;
510                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
511                                                 vma, addr, next))
512                         return -ENOMEM;
513         } while (dst_pud++, src_pud++, addr = next, addr != end);
514         return 0;
515 }
516
517 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
518                 struct vm_area_struct *vma)
519 {
520         pgd_t *src_pgd, *dst_pgd;
521         unsigned long next;
522         unsigned long addr = vma->vm_start;
523         unsigned long end = vma->vm_end;
524
525         /*
526          * Don't copy ptes where a page fault will fill them correctly.
527          * Fork becomes much lighter when there are big shared or private
528          * readonly mappings. The tradeoff is that copy_page_range is more
529          * efficient than faulting.
530          */
531         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_RESERVED))) {
532                 if (!vma->anon_vma)
533                         return 0;
534         }
535
536         if (is_vm_hugetlb_page(vma))
537                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
538
539         dst_pgd = pgd_offset(dst_mm, addr);
540         src_pgd = pgd_offset(src_mm, addr);
541         do {
542                 next = pgd_addr_end(addr, end);
543                 if (pgd_none_or_clear_bad(src_pgd))
544                         continue;
545                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
546                                                 vma, addr, next))
547                         return -ENOMEM;
548         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
549         return 0;
550 }
551
552 static void zap_pte_range(struct mmu_gather *tlb,
553                                 struct vm_area_struct *vma, pmd_t *pmd,
554                                 unsigned long addr, unsigned long end,
555                                 struct zap_details *details)
556 {
557         struct mm_struct *mm = tlb->mm;
558         pte_t *pte;
559         spinlock_t *ptl;
560         int file_rss = 0;
561         int anon_rss = 0;
562
563         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
564         do {
565                 pte_t ptent = *pte;
566                 if (pte_none(ptent))
567                         continue;
568                 if (pte_present(ptent)) {
569                         struct page *page = NULL;
570                         if (!(vma->vm_flags & VM_RESERVED)) {
571                                 unsigned long pfn = pte_pfn(ptent);
572                                 if (unlikely(!pfn_valid(pfn)))
573                                         print_bad_pte(vma, ptent, addr);
574                                 else
575                                         page = pfn_to_page(pfn);
576                         }
577                         if (unlikely(details) && page) {
578                                 /*
579                                  * unmap_shared_mapping_pages() wants to
580                                  * invalidate cache without truncating:
581                                  * unmap shared but keep private pages.
582                                  */
583                                 if (details->check_mapping &&
584                                     details->check_mapping != page->mapping)
585                                         continue;
586                                 /*
587                                  * Each page->index must be checked when
588                                  * invalidating or truncating nonlinear.
589                                  */
590                                 if (details->nonlinear_vma &&
591                                     (page->index < details->first_index ||
592                                      page->index > details->last_index))
593                                         continue;
594                         }
595                         ptent = ptep_get_and_clear_full(mm, addr, pte,
596                                                         tlb->fullmm);
597                         tlb_remove_tlb_entry(tlb, pte, addr);
598                         if (unlikely(!page))
599                                 continue;
600                         if (unlikely(details) && details->nonlinear_vma
601                             && linear_page_index(details->nonlinear_vma,
602                                                 addr) != page->index)
603                                 set_pte_at(mm, addr, pte,
604                                            pgoff_to_pte(page->index));
605                         if (PageAnon(page))
606                                 anon_rss--;
607                         else {
608                                 if (pte_dirty(ptent))
609                                         set_page_dirty(page);
610                                 if (pte_young(ptent))
611                                         mark_page_accessed(page);
612                                 file_rss--;
613                         }
614                         page_remove_rmap(page);
615                         tlb_remove_page(tlb, page);
616                         continue;
617                 }
618                 /*
619                  * If details->check_mapping, we leave swap entries;
620                  * if details->nonlinear_vma, we leave file entries.
621                  */
622                 if (unlikely(details))
623                         continue;
624                 if (!pte_file(ptent))
625                         free_swap_and_cache(pte_to_swp_entry(ptent));
626                 pte_clear_full(mm, addr, pte, tlb->fullmm);
627         } while (pte++, addr += PAGE_SIZE, addr != end);
628
629         add_mm_rss(mm, file_rss, anon_rss);
630         pte_unmap_unlock(pte - 1, ptl);
631 }
632
633 static inline void zap_pmd_range(struct mmu_gather *tlb,
634                                 struct vm_area_struct *vma, pud_t *pud,
635                                 unsigned long addr, unsigned long end,
636                                 struct zap_details *details)
637 {
638         pmd_t *pmd;
639         unsigned long next;
640
641         pmd = pmd_offset(pud, addr);
642         do {
643                 next = pmd_addr_end(addr, end);
644                 if (pmd_none_or_clear_bad(pmd))
645                         continue;
646                 zap_pte_range(tlb, vma, pmd, addr, next, details);
647         } while (pmd++, addr = next, addr != end);
648 }
649
650 static inline void zap_pud_range(struct mmu_gather *tlb,
651                                 struct vm_area_struct *vma, pgd_t *pgd,
652                                 unsigned long addr, unsigned long end,
653                                 struct zap_details *details)
654 {
655         pud_t *pud;
656         unsigned long next;
657
658         pud = pud_offset(pgd, addr);
659         do {
660                 next = pud_addr_end(addr, end);
661                 if (pud_none_or_clear_bad(pud))
662                         continue;
663                 zap_pmd_range(tlb, vma, pud, addr, next, details);
664         } while (pud++, addr = next, addr != end);
665 }
666
667 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
668                                 unsigned long addr, unsigned long end,
669                                 struct zap_details *details)
670 {
671         pgd_t *pgd;
672         unsigned long next;
673
674         if (details && !details->check_mapping && !details->nonlinear_vma)
675                 details = NULL;
676
677         BUG_ON(addr >= end);
678         tlb_start_vma(tlb, vma);
679         pgd = pgd_offset(vma->vm_mm, addr);
680         do {
681                 next = pgd_addr_end(addr, end);
682                 if (pgd_none_or_clear_bad(pgd))
683                         continue;
684                 zap_pud_range(tlb, vma, pgd, addr, next, details);
685         } while (pgd++, addr = next, addr != end);
686         tlb_end_vma(tlb, vma);
687 }
688
689 #ifdef CONFIG_PREEMPT
690 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
691 #else
692 /* No preempt: go for improved straight-line efficiency */
693 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
694 #endif
695
696 /**
697  * unmap_vmas - unmap a range of memory covered by a list of vma's
698  * @tlbp: address of the caller's struct mmu_gather
699  * @vma: the starting vma
700  * @start_addr: virtual address at which to start unmapping
701  * @end_addr: virtual address at which to end unmapping
702  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
703  * @details: details of nonlinear truncation or shared cache invalidation
704  *
705  * Returns the end address of the unmapping (restart addr if interrupted).
706  *
707  * Unmap all pages in the vma list.
708  *
709  * We aim to not hold locks for too long (for scheduling latency reasons).
710  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
711  * return the ending mmu_gather to the caller.
712  *
713  * Only addresses between `start' and `end' will be unmapped.
714  *
715  * The VMA list must be sorted in ascending virtual address order.
716  *
717  * unmap_vmas() assumes that the caller will flush the whole unmapped address
718  * range after unmap_vmas() returns.  So the only responsibility here is to
719  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
720  * drops the lock and schedules.
721  */
722 unsigned long unmap_vmas(struct mmu_gather **tlbp,
723                 struct vm_area_struct *vma, unsigned long start_addr,
724                 unsigned long end_addr, unsigned long *nr_accounted,
725                 struct zap_details *details)
726 {
727         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
728         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
729         int tlb_start_valid = 0;
730         unsigned long start = start_addr;
731         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
732         int fullmm = (*tlbp)->fullmm;
733
734         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
735                 unsigned long end;
736
737                 start = max(vma->vm_start, start_addr);
738                 if (start >= vma->vm_end)
739                         continue;
740                 end = min(vma->vm_end, end_addr);
741                 if (end <= vma->vm_start)
742                         continue;
743
744                 if (vma->vm_flags & VM_ACCOUNT)
745                         *nr_accounted += (end - start) >> PAGE_SHIFT;
746
747                 while (start != end) {
748                         unsigned long block;
749
750                         if (!tlb_start_valid) {
751                                 tlb_start = start;
752                                 tlb_start_valid = 1;
753                         }
754
755                         if (is_vm_hugetlb_page(vma)) {
756                                 block = end - start;
757                                 unmap_hugepage_range(vma, start, end);
758                         } else {
759                                 block = min(zap_bytes, end - start);
760                                 unmap_page_range(*tlbp, vma, start,
761                                                 start + block, details);
762                         }
763
764                         start += block;
765                         zap_bytes -= block;
766                         if ((long)zap_bytes > 0)
767                                 continue;
768
769                         tlb_finish_mmu(*tlbp, tlb_start, start);
770
771                         if (need_resched() ||
772                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
773                                 if (i_mmap_lock) {
774                                         *tlbp = NULL;
775                                         goto out;
776                                 }
777                                 cond_resched();
778                         }
779
780                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
781                         tlb_start_valid = 0;
782                         zap_bytes = ZAP_BLOCK_SIZE;
783                 }
784         }
785 out:
786         return start;   /* which is now the end (or restart) address */
787 }
788
789 /**
790  * zap_page_range - remove user pages in a given range
791  * @vma: vm_area_struct holding the applicable pages
792  * @address: starting address of pages to zap
793  * @size: number of bytes to zap
794  * @details: details of nonlinear truncation or shared cache invalidation
795  */
796 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
797                 unsigned long size, struct zap_details *details)
798 {
799         struct mm_struct *mm = vma->vm_mm;
800         struct mmu_gather *tlb;
801         unsigned long end = address + size;
802         unsigned long nr_accounted = 0;
803
804         lru_add_drain();
805         tlb = tlb_gather_mmu(mm, 0);
806         update_hiwater_rss(mm);
807         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
808         if (tlb)
809                 tlb_finish_mmu(tlb, address, end);
810         return end;
811 }
812
813 /*
814  * Do a quick page-table lookup for a single page.
815  */
816 struct page *follow_page(struct mm_struct *mm, unsigned long address,
817                         unsigned int flags)
818 {
819         pgd_t *pgd;
820         pud_t *pud;
821         pmd_t *pmd;
822         pte_t *ptep, pte;
823         spinlock_t *ptl;
824         unsigned long pfn;
825         struct page *page;
826
827         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
828         if (!IS_ERR(page)) {
829                 BUG_ON(flags & FOLL_GET);
830                 goto out;
831         }
832
833         page = NULL;
834         pgd = pgd_offset(mm, address);
835         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
836                 goto no_page_table;
837
838         pud = pud_offset(pgd, address);
839         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
840                 goto no_page_table;
841         
842         pmd = pmd_offset(pud, address);
843         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
844                 goto no_page_table;
845
846         if (pmd_huge(*pmd)) {
847                 BUG_ON(flags & FOLL_GET);
848                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
849                 goto out;
850         }
851
852         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
853         if (!ptep)
854                 goto out;
855
856         pte = *ptep;
857         if (!pte_present(pte))
858                 goto unlock;
859         if ((flags & FOLL_WRITE) && !pte_write(pte))
860                 goto unlock;
861         pfn = pte_pfn(pte);
862         if (!pfn_valid(pfn))
863                 goto unlock;
864
865         page = pfn_to_page(pfn);
866         if (flags & FOLL_GET)
867                 get_page(page);
868         if (flags & FOLL_TOUCH) {
869                 if ((flags & FOLL_WRITE) &&
870                     !pte_dirty(pte) && !PageDirty(page))
871                         set_page_dirty(page);
872                 mark_page_accessed(page);
873         }
874 unlock:
875         pte_unmap_unlock(ptep, ptl);
876 out:
877         return page;
878
879 no_page_table:
880         /*
881          * When core dumping an enormous anonymous area that nobody
882          * has touched so far, we don't want to allocate page tables.
883          */
884         if (flags & FOLL_ANON) {
885                 page = ZERO_PAGE(address);
886                 if (flags & FOLL_GET)
887                         get_page(page);
888                 BUG_ON(flags & FOLL_WRITE);
889         }
890         return page;
891 }
892
893 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
894                 unsigned long start, int len, int write, int force,
895                 struct page **pages, struct vm_area_struct **vmas)
896 {
897         int i;
898         unsigned int vm_flags;
899
900         /* 
901          * Require read or write permissions.
902          * If 'force' is set, we only require the "MAY" flags.
903          */
904         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
905         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
906         i = 0;
907
908         do {
909                 struct vm_area_struct *vma;
910                 unsigned int foll_flags;
911
912                 vma = find_extend_vma(mm, start);
913                 if (!vma && in_gate_area(tsk, start)) {
914                         unsigned long pg = start & PAGE_MASK;
915                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
916                         pgd_t *pgd;
917                         pud_t *pud;
918                         pmd_t *pmd;
919                         pte_t *pte;
920                         if (write) /* user gate pages are read-only */
921                                 return i ? : -EFAULT;
922                         if (pg > TASK_SIZE)
923                                 pgd = pgd_offset_k(pg);
924                         else
925                                 pgd = pgd_offset_gate(mm, pg);
926                         BUG_ON(pgd_none(*pgd));
927                         pud = pud_offset(pgd, pg);
928                         BUG_ON(pud_none(*pud));
929                         pmd = pmd_offset(pud, pg);
930                         if (pmd_none(*pmd))
931                                 return i ? : -EFAULT;
932                         pte = pte_offset_map(pmd, pg);
933                         if (pte_none(*pte)) {
934                                 pte_unmap(pte);
935                                 return i ? : -EFAULT;
936                         }
937                         if (pages) {
938                                 pages[i] = pte_page(*pte);
939                                 get_page(pages[i]);
940                         }
941                         pte_unmap(pte);
942                         if (vmas)
943                                 vmas[i] = gate_vma;
944                         i++;
945                         start += PAGE_SIZE;
946                         len--;
947                         continue;
948                 }
949
950                 if (!vma || (vma->vm_flags & (VM_IO | VM_RESERVED))
951                                 || !(vm_flags & vma->vm_flags))
952                         return i ? : -EFAULT;
953
954                 if (is_vm_hugetlb_page(vma)) {
955                         i = follow_hugetlb_page(mm, vma, pages, vmas,
956                                                 &start, &len, i);
957                         continue;
958                 }
959
960                 foll_flags = FOLL_TOUCH;
961                 if (pages)
962                         foll_flags |= FOLL_GET;
963                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
964                     (!vma->vm_ops || !vma->vm_ops->nopage))
965                         foll_flags |= FOLL_ANON;
966
967                 do {
968                         struct page *page;
969
970                         if (write)
971                                 foll_flags |= FOLL_WRITE;
972
973                         cond_resched();
974                         while (!(page = follow_page(mm, start, foll_flags))) {
975                                 int ret;
976                                 ret = __handle_mm_fault(mm, vma, start,
977                                                 foll_flags & FOLL_WRITE);
978                                 /*
979                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
980                                  * broken COW when necessary, even if maybe_mkwrite
981                                  * decided not to set pte_write. We can thus safely do
982                                  * subsequent page lookups as if they were reads.
983                                  */
984                                 if (ret & VM_FAULT_WRITE)
985                                         foll_flags &= ~FOLL_WRITE;
986                                 
987                                 switch (ret & ~VM_FAULT_WRITE) {
988                                 case VM_FAULT_MINOR:
989                                         tsk->min_flt++;
990                                         break;
991                                 case VM_FAULT_MAJOR:
992                                         tsk->maj_flt++;
993                                         break;
994                                 case VM_FAULT_SIGBUS:
995                                         return i ? i : -EFAULT;
996                                 case VM_FAULT_OOM:
997                                         return i ? i : -ENOMEM;
998                                 default:
999                                         BUG();
1000                                 }
1001                         }
1002                         if (pages) {
1003                                 pages[i] = page;
1004                                 flush_dcache_page(page);
1005                         }
1006                         if (vmas)
1007                                 vmas[i] = vma;
1008                         i++;
1009                         start += PAGE_SIZE;
1010                         len--;
1011                 } while (len && start < vma->vm_end);
1012         } while (len);
1013         return i;
1014 }
1015 EXPORT_SYMBOL(get_user_pages);
1016
1017 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1018                         unsigned long addr, unsigned long end, pgprot_t prot)
1019 {
1020         pte_t *pte;
1021         spinlock_t *ptl;
1022
1023         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1024         if (!pte)
1025                 return -ENOMEM;
1026         do {
1027                 struct page *page = ZERO_PAGE(addr);
1028                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1029                 page_cache_get(page);
1030                 page_add_file_rmap(page);
1031                 inc_mm_counter(mm, file_rss);
1032                 BUG_ON(!pte_none(*pte));
1033                 set_pte_at(mm, addr, pte, zero_pte);
1034         } while (pte++, addr += PAGE_SIZE, addr != end);
1035         pte_unmap_unlock(pte - 1, ptl);
1036         return 0;
1037 }
1038
1039 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1040                         unsigned long addr, unsigned long end, pgprot_t prot)
1041 {
1042         pmd_t *pmd;
1043         unsigned long next;
1044
1045         pmd = pmd_alloc(mm, pud, addr);
1046         if (!pmd)
1047                 return -ENOMEM;
1048         do {
1049                 next = pmd_addr_end(addr, end);
1050                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1051                         return -ENOMEM;
1052         } while (pmd++, addr = next, addr != end);
1053         return 0;
1054 }
1055
1056 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1057                         unsigned long addr, unsigned long end, pgprot_t prot)
1058 {
1059         pud_t *pud;
1060         unsigned long next;
1061
1062         pud = pud_alloc(mm, pgd, addr);
1063         if (!pud)
1064                 return -ENOMEM;
1065         do {
1066                 next = pud_addr_end(addr, end);
1067                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1068                         return -ENOMEM;
1069         } while (pud++, addr = next, addr != end);
1070         return 0;
1071 }
1072
1073 int zeromap_page_range(struct vm_area_struct *vma,
1074                         unsigned long addr, unsigned long size, pgprot_t prot)
1075 {
1076         pgd_t *pgd;
1077         unsigned long next;
1078         unsigned long end = addr + size;
1079         struct mm_struct *mm = vma->vm_mm;
1080         int err;
1081
1082         BUG_ON(addr >= end);
1083         pgd = pgd_offset(mm, addr);
1084         flush_cache_range(vma, addr, end);
1085         do {
1086                 next = pgd_addr_end(addr, end);
1087                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1088                 if (err)
1089                         break;
1090         } while (pgd++, addr = next, addr != end);
1091         return err;
1092 }
1093
1094 /*
1095  * maps a range of physical memory into the requested pages. the old
1096  * mappings are removed. any references to nonexistent pages results
1097  * in null mappings (currently treated as "copy-on-access")
1098  */
1099 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1100                         unsigned long addr, unsigned long end,
1101                         unsigned long pfn, pgprot_t prot)
1102 {
1103         pte_t *pte;
1104         spinlock_t *ptl;
1105
1106         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1107         if (!pte)
1108                 return -ENOMEM;
1109         do {
1110                 BUG_ON(!pte_none(*pte));
1111                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1112                 pfn++;
1113         } while (pte++, addr += PAGE_SIZE, addr != end);
1114         pte_unmap_unlock(pte - 1, ptl);
1115         return 0;
1116 }
1117
1118 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1119                         unsigned long addr, unsigned long end,
1120                         unsigned long pfn, pgprot_t prot)
1121 {
1122         pmd_t *pmd;
1123         unsigned long next;
1124
1125         pfn -= addr >> PAGE_SHIFT;
1126         pmd = pmd_alloc(mm, pud, addr);
1127         if (!pmd)
1128                 return -ENOMEM;
1129         do {
1130                 next = pmd_addr_end(addr, end);
1131                 if (remap_pte_range(mm, pmd, addr, next,
1132                                 pfn + (addr >> PAGE_SHIFT), prot))
1133                         return -ENOMEM;
1134         } while (pmd++, addr = next, addr != end);
1135         return 0;
1136 }
1137
1138 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1139                         unsigned long addr, unsigned long end,
1140                         unsigned long pfn, pgprot_t prot)
1141 {
1142         pud_t *pud;
1143         unsigned long next;
1144
1145         pfn -= addr >> PAGE_SHIFT;
1146         pud = pud_alloc(mm, pgd, addr);
1147         if (!pud)
1148                 return -ENOMEM;
1149         do {
1150                 next = pud_addr_end(addr, end);
1151                 if (remap_pmd_range(mm, pud, addr, next,
1152                                 pfn + (addr >> PAGE_SHIFT), prot))
1153                         return -ENOMEM;
1154         } while (pud++, addr = next, addr != end);
1155         return 0;
1156 }
1157
1158 /*  Note: this is only safe if the mm semaphore is held when called. */
1159 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1160                     unsigned long pfn, unsigned long size, pgprot_t prot)
1161 {
1162         pgd_t *pgd;
1163         unsigned long next;
1164         unsigned long end = addr + PAGE_ALIGN(size);
1165         struct mm_struct *mm = vma->vm_mm;
1166         int err;
1167
1168         /*
1169          * Physically remapped pages are special. Tell the
1170          * rest of the world about it:
1171          *   VM_IO tells people not to look at these pages
1172          *      (accesses can have side effects).
1173          *   VM_RESERVED tells the core MM not to "manage" these pages
1174          *      (e.g. refcount, mapcount, try to swap them out).
1175          */
1176         vma->vm_flags |= VM_IO | VM_RESERVED;
1177
1178         BUG_ON(addr >= end);
1179         pfn -= addr >> PAGE_SHIFT;
1180         pgd = pgd_offset(mm, addr);
1181         flush_cache_range(vma, addr, end);
1182         do {
1183                 next = pgd_addr_end(addr, end);
1184                 err = remap_pud_range(mm, pgd, addr, next,
1185                                 pfn + (addr >> PAGE_SHIFT), prot);
1186                 if (err)
1187                         break;
1188         } while (pgd++, addr = next, addr != end);
1189         return err;
1190 }
1191 EXPORT_SYMBOL(remap_pfn_range);
1192
1193 /*
1194  * handle_pte_fault chooses page fault handler according to an entry
1195  * which was read non-atomically.  Before making any commitment, on
1196  * those architectures or configurations (e.g. i386 with PAE) which
1197  * might give a mix of unmatched parts, do_swap_page and do_file_page
1198  * must check under lock before unmapping the pte and proceeding
1199  * (but do_wp_page is only called after already making such a check;
1200  * and do_anonymous_page and do_no_page can safely check later on).
1201  */
1202 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1203                                 pte_t *page_table, pte_t orig_pte)
1204 {
1205         int same = 1;
1206 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1207         if (sizeof(pte_t) > sizeof(unsigned long)) {
1208                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1209                 spin_lock(ptl);
1210                 same = pte_same(*page_table, orig_pte);
1211                 spin_unlock(ptl);
1212         }
1213 #endif
1214         pte_unmap(page_table);
1215         return same;
1216 }
1217
1218 /*
1219  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1220  * servicing faults for write access.  In the normal case, do always want
1221  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1222  * that do not have writing enabled, when used by access_process_vm.
1223  */
1224 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1225 {
1226         if (likely(vma->vm_flags & VM_WRITE))
1227                 pte = pte_mkwrite(pte);
1228         return pte;
1229 }
1230
1231 /*
1232  * This routine handles present pages, when users try to write
1233  * to a shared page. It is done by copying the page to a new address
1234  * and decrementing the shared-page counter for the old page.
1235  *
1236  * Note that this routine assumes that the protection checks have been
1237  * done by the caller (the low-level page fault routine in most cases).
1238  * Thus we can safely just mark it writable once we've done any necessary
1239  * COW.
1240  *
1241  * We also mark the page dirty at this point even though the page will
1242  * change only once the write actually happens. This avoids a few races,
1243  * and potentially makes it more efficient.
1244  *
1245  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1246  * but allow concurrent faults), with pte both mapped and locked.
1247  * We return with mmap_sem still held, but pte unmapped and unlocked.
1248  */
1249 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1250                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1251                 spinlock_t *ptl, pte_t orig_pte)
1252 {
1253         struct page *old_page, *new_page;
1254         unsigned long pfn = pte_pfn(orig_pte);
1255         pte_t entry;
1256         int ret = VM_FAULT_MINOR;
1257
1258         BUG_ON(vma->vm_flags & VM_RESERVED);
1259
1260         if (unlikely(!pfn_valid(pfn))) {
1261                 /*
1262                  * Page table corrupted: show pte and kill process.
1263                  */
1264                 print_bad_pte(vma, orig_pte, address);
1265                 ret = VM_FAULT_OOM;
1266                 goto unlock;
1267         }
1268         old_page = pfn_to_page(pfn);
1269
1270         if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1271                 int reuse = can_share_swap_page(old_page);
1272                 unlock_page(old_page);
1273                 if (reuse) {
1274                         flush_cache_page(vma, address, pfn);
1275                         entry = pte_mkyoung(orig_pte);
1276                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1277                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1278                         update_mmu_cache(vma, address, entry);
1279                         lazy_mmu_prot_update(entry);
1280                         ret |= VM_FAULT_WRITE;
1281                         goto unlock;
1282                 }
1283         }
1284
1285         /*
1286          * Ok, we need to copy. Oh, well..
1287          */
1288         page_cache_get(old_page);
1289         pte_unmap_unlock(page_table, ptl);
1290
1291         if (unlikely(anon_vma_prepare(vma)))
1292                 goto oom;
1293         if (old_page == ZERO_PAGE(address)) {
1294                 new_page = alloc_zeroed_user_highpage(vma, address);
1295                 if (!new_page)
1296                         goto oom;
1297         } else {
1298                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1299                 if (!new_page)
1300                         goto oom;
1301                 copy_user_highpage(new_page, old_page, address);
1302         }
1303
1304         /*
1305          * Re-check the pte - we dropped the lock
1306          */
1307         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1308         if (likely(pte_same(*page_table, orig_pte))) {
1309                 page_remove_rmap(old_page);
1310                 if (!PageAnon(old_page)) {
1311                         inc_mm_counter(mm, anon_rss);
1312                         dec_mm_counter(mm, file_rss);
1313                 }
1314                 flush_cache_page(vma, address, pfn);
1315                 entry = mk_pte(new_page, vma->vm_page_prot);
1316                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1317                 ptep_establish(vma, address, page_table, entry);
1318                 update_mmu_cache(vma, address, entry);
1319                 lazy_mmu_prot_update(entry);
1320                 lru_cache_add_active(new_page);
1321                 page_add_anon_rmap(new_page, vma, address);
1322
1323                 /* Free the old page.. */
1324                 new_page = old_page;
1325                 ret |= VM_FAULT_WRITE;
1326         }
1327         page_cache_release(new_page);
1328         page_cache_release(old_page);
1329 unlock:
1330         pte_unmap_unlock(page_table, ptl);
1331         return ret;
1332 oom:
1333         page_cache_release(old_page);
1334         return VM_FAULT_OOM;
1335 }
1336
1337 /*
1338  * Helper functions for unmap_mapping_range().
1339  *
1340  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1341  *
1342  * We have to restart searching the prio_tree whenever we drop the lock,
1343  * since the iterator is only valid while the lock is held, and anyway
1344  * a later vma might be split and reinserted earlier while lock dropped.
1345  *
1346  * The list of nonlinear vmas could be handled more efficiently, using
1347  * a placeholder, but handle it in the same way until a need is shown.
1348  * It is important to search the prio_tree before nonlinear list: a vma
1349  * may become nonlinear and be shifted from prio_tree to nonlinear list
1350  * while the lock is dropped; but never shifted from list to prio_tree.
1351  *
1352  * In order to make forward progress despite restarting the search,
1353  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1354  * quickly skip it next time around.  Since the prio_tree search only
1355  * shows us those vmas affected by unmapping the range in question, we
1356  * can't efficiently keep all vmas in step with mapping->truncate_count:
1357  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1358  * mapping->truncate_count and vma->vm_truncate_count are protected by
1359  * i_mmap_lock.
1360  *
1361  * In order to make forward progress despite repeatedly restarting some
1362  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1363  * and restart from that address when we reach that vma again.  It might
1364  * have been split or merged, shrunk or extended, but never shifted: so
1365  * restart_addr remains valid so long as it remains in the vma's range.
1366  * unmap_mapping_range forces truncate_count to leap over page-aligned
1367  * values so we can save vma's restart_addr in its truncate_count field.
1368  */
1369 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1370
1371 static void reset_vma_truncate_counts(struct address_space *mapping)
1372 {
1373         struct vm_area_struct *vma;
1374         struct prio_tree_iter iter;
1375
1376         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1377                 vma->vm_truncate_count = 0;
1378         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1379                 vma->vm_truncate_count = 0;
1380 }
1381
1382 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1383                 unsigned long start_addr, unsigned long end_addr,
1384                 struct zap_details *details)
1385 {
1386         unsigned long restart_addr;
1387         int need_break;
1388
1389 again:
1390         restart_addr = vma->vm_truncate_count;
1391         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1392                 start_addr = restart_addr;
1393                 if (start_addr >= end_addr) {
1394                         /* Top of vma has been split off since last time */
1395                         vma->vm_truncate_count = details->truncate_count;
1396                         return 0;
1397                 }
1398         }
1399
1400         restart_addr = zap_page_range(vma, start_addr,
1401                                         end_addr - start_addr, details);
1402         need_break = need_resched() ||
1403                         need_lockbreak(details->i_mmap_lock);
1404
1405         if (restart_addr >= end_addr) {
1406                 /* We have now completed this vma: mark it so */
1407                 vma->vm_truncate_count = details->truncate_count;
1408                 if (!need_break)
1409                         return 0;
1410         } else {
1411                 /* Note restart_addr in vma's truncate_count field */
1412                 vma->vm_truncate_count = restart_addr;
1413                 if (!need_break)
1414                         goto again;
1415         }
1416
1417         spin_unlock(details->i_mmap_lock);
1418         cond_resched();
1419         spin_lock(details->i_mmap_lock);
1420         return -EINTR;
1421 }
1422
1423 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1424                                             struct zap_details *details)
1425 {
1426         struct vm_area_struct *vma;
1427         struct prio_tree_iter iter;
1428         pgoff_t vba, vea, zba, zea;
1429
1430 restart:
1431         vma_prio_tree_foreach(vma, &iter, root,
1432                         details->first_index, details->last_index) {
1433                 /* Skip quickly over those we have already dealt with */
1434                 if (vma->vm_truncate_count == details->truncate_count)
1435                         continue;
1436
1437                 vba = vma->vm_pgoff;
1438                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1439                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1440                 zba = details->first_index;
1441                 if (zba < vba)
1442                         zba = vba;
1443                 zea = details->last_index;
1444                 if (zea > vea)
1445                         zea = vea;
1446
1447                 if (unmap_mapping_range_vma(vma,
1448                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1449                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1450                                 details) < 0)
1451                         goto restart;
1452         }
1453 }
1454
1455 static inline void unmap_mapping_range_list(struct list_head *head,
1456                                             struct zap_details *details)
1457 {
1458         struct vm_area_struct *vma;
1459
1460         /*
1461          * In nonlinear VMAs there is no correspondence between virtual address
1462          * offset and file offset.  So we must perform an exhaustive search
1463          * across *all* the pages in each nonlinear VMA, not just the pages
1464          * whose virtual address lies outside the file truncation point.
1465          */
1466 restart:
1467         list_for_each_entry(vma, head, shared.vm_set.list) {
1468                 /* Skip quickly over those we have already dealt with */
1469                 if (vma->vm_truncate_count == details->truncate_count)
1470                         continue;
1471                 details->nonlinear_vma = vma;
1472                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1473                                         vma->vm_end, details) < 0)
1474                         goto restart;
1475         }
1476 }
1477
1478 /**
1479  * unmap_mapping_range - unmap the portion of all mmaps
1480  * in the specified address_space corresponding to the specified
1481  * page range in the underlying file.
1482  * @mapping: the address space containing mmaps to be unmapped.
1483  * @holebegin: byte in first page to unmap, relative to the start of
1484  * the underlying file.  This will be rounded down to a PAGE_SIZE
1485  * boundary.  Note that this is different from vmtruncate(), which
1486  * must keep the partial page.  In contrast, we must get rid of
1487  * partial pages.
1488  * @holelen: size of prospective hole in bytes.  This will be rounded
1489  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1490  * end of the file.
1491  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1492  * but 0 when invalidating pagecache, don't throw away private data.
1493  */
1494 void unmap_mapping_range(struct address_space *mapping,
1495                 loff_t const holebegin, loff_t const holelen, int even_cows)
1496 {
1497         struct zap_details details;
1498         pgoff_t hba = holebegin >> PAGE_SHIFT;
1499         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1500
1501         /* Check for overflow. */
1502         if (sizeof(holelen) > sizeof(hlen)) {
1503                 long long holeend =
1504                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1505                 if (holeend & ~(long long)ULONG_MAX)
1506                         hlen = ULONG_MAX - hba + 1;
1507         }
1508
1509         details.check_mapping = even_cows? NULL: mapping;
1510         details.nonlinear_vma = NULL;
1511         details.first_index = hba;
1512         details.last_index = hba + hlen - 1;
1513         if (details.last_index < details.first_index)
1514                 details.last_index = ULONG_MAX;
1515         details.i_mmap_lock = &mapping->i_mmap_lock;
1516
1517         spin_lock(&mapping->i_mmap_lock);
1518
1519         /* serialize i_size write against truncate_count write */
1520         smp_wmb();
1521         /* Protect against page faults, and endless unmapping loops */
1522         mapping->truncate_count++;
1523         /*
1524          * For archs where spin_lock has inclusive semantics like ia64
1525          * this smp_mb() will prevent to read pagetable contents
1526          * before the truncate_count increment is visible to
1527          * other cpus.
1528          */
1529         smp_mb();
1530         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1531                 if (mapping->truncate_count == 0)
1532                         reset_vma_truncate_counts(mapping);
1533                 mapping->truncate_count++;
1534         }
1535         details.truncate_count = mapping->truncate_count;
1536
1537         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1538                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1539         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1540                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1541         spin_unlock(&mapping->i_mmap_lock);
1542 }
1543 EXPORT_SYMBOL(unmap_mapping_range);
1544
1545 /*
1546  * Handle all mappings that got truncated by a "truncate()"
1547  * system call.
1548  *
1549  * NOTE! We have to be ready to update the memory sharing
1550  * between the file and the memory map for a potential last
1551  * incomplete page.  Ugly, but necessary.
1552  */
1553 int vmtruncate(struct inode * inode, loff_t offset)
1554 {
1555         struct address_space *mapping = inode->i_mapping;
1556         unsigned long limit;
1557
1558         if (inode->i_size < offset)
1559                 goto do_expand;
1560         /*
1561          * truncation of in-use swapfiles is disallowed - it would cause
1562          * subsequent swapout to scribble on the now-freed blocks.
1563          */
1564         if (IS_SWAPFILE(inode))
1565                 goto out_busy;
1566         i_size_write(inode, offset);
1567         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1568         truncate_inode_pages(mapping, offset);
1569         goto out_truncate;
1570
1571 do_expand:
1572         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1573         if (limit != RLIM_INFINITY && offset > limit)
1574                 goto out_sig;
1575         if (offset > inode->i_sb->s_maxbytes)
1576                 goto out_big;
1577         i_size_write(inode, offset);
1578
1579 out_truncate:
1580         if (inode->i_op && inode->i_op->truncate)
1581                 inode->i_op->truncate(inode);
1582         return 0;
1583 out_sig:
1584         send_sig(SIGXFSZ, current, 0);
1585 out_big:
1586         return -EFBIG;
1587 out_busy:
1588         return -ETXTBSY;
1589 }
1590
1591 EXPORT_SYMBOL(vmtruncate);
1592
1593 /* 
1594  * Primitive swap readahead code. We simply read an aligned block of
1595  * (1 << page_cluster) entries in the swap area. This method is chosen
1596  * because it doesn't cost us any seek time.  We also make sure to queue
1597  * the 'original' request together with the readahead ones...  
1598  *
1599  * This has been extended to use the NUMA policies from the mm triggering
1600  * the readahead.
1601  *
1602  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1603  */
1604 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1605 {
1606 #ifdef CONFIG_NUMA
1607         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1608 #endif
1609         int i, num;
1610         struct page *new_page;
1611         unsigned long offset;
1612
1613         /*
1614          * Get the number of handles we should do readahead io to.
1615          */
1616         num = valid_swaphandles(entry, &offset);
1617         for (i = 0; i < num; offset++, i++) {
1618                 /* Ok, do the async read-ahead now */
1619                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1620                                                            offset), vma, addr);
1621                 if (!new_page)
1622                         break;
1623                 page_cache_release(new_page);
1624 #ifdef CONFIG_NUMA
1625                 /*
1626                  * Find the next applicable VMA for the NUMA policy.
1627                  */
1628                 addr += PAGE_SIZE;
1629                 if (addr == 0)
1630                         vma = NULL;
1631                 if (vma) {
1632                         if (addr >= vma->vm_end) {
1633                                 vma = next_vma;
1634                                 next_vma = vma ? vma->vm_next : NULL;
1635                         }
1636                         if (vma && addr < vma->vm_start)
1637                                 vma = NULL;
1638                 } else {
1639                         if (next_vma && addr >= next_vma->vm_start) {
1640                                 vma = next_vma;
1641                                 next_vma = vma->vm_next;
1642                         }
1643                 }
1644 #endif
1645         }
1646         lru_add_drain();        /* Push any new pages onto the LRU now */
1647 }
1648
1649 /*
1650  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1651  * but allow concurrent faults), and pte mapped but not yet locked.
1652  * We return with mmap_sem still held, but pte unmapped and unlocked.
1653  */
1654 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1655                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1656                 int write_access, pte_t orig_pte)
1657 {
1658         spinlock_t *ptl;
1659         struct page *page;
1660         swp_entry_t entry;
1661         pte_t pte;
1662         int ret = VM_FAULT_MINOR;
1663
1664         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1665                 goto out;
1666
1667         entry = pte_to_swp_entry(orig_pte);
1668         page = lookup_swap_cache(entry);
1669         if (!page) {
1670                 swapin_readahead(entry, address, vma);
1671                 page = read_swap_cache_async(entry, vma, address);
1672                 if (!page) {
1673                         /*
1674                          * Back out if somebody else faulted in this pte
1675                          * while we released the pte lock.
1676                          */
1677                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1678                         if (likely(pte_same(*page_table, orig_pte)))
1679                                 ret = VM_FAULT_OOM;
1680                         goto unlock;
1681                 }
1682
1683                 /* Had to read the page from swap area: Major fault */
1684                 ret = VM_FAULT_MAJOR;
1685                 inc_page_state(pgmajfault);
1686                 grab_swap_token();
1687         }
1688
1689         mark_page_accessed(page);
1690         lock_page(page);
1691
1692         /*
1693          * Back out if somebody else already faulted in this pte.
1694          */
1695         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1696         if (unlikely(!pte_same(*page_table, orig_pte)))
1697                 goto out_nomap;
1698
1699         if (unlikely(!PageUptodate(page))) {
1700                 ret = VM_FAULT_SIGBUS;
1701                 goto out_nomap;
1702         }
1703
1704         /* The page isn't present yet, go ahead with the fault. */
1705
1706         inc_mm_counter(mm, anon_rss);
1707         pte = mk_pte(page, vma->vm_page_prot);
1708         if (write_access && can_share_swap_page(page)) {
1709                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1710                 write_access = 0;
1711         }
1712
1713         flush_icache_page(vma, page);
1714         set_pte_at(mm, address, page_table, pte);
1715         page_add_anon_rmap(page, vma, address);
1716
1717         swap_free(entry);
1718         if (vm_swap_full())
1719                 remove_exclusive_swap_page(page);
1720         unlock_page(page);
1721
1722         if (write_access) {
1723                 if (do_wp_page(mm, vma, address,
1724                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1725                         ret = VM_FAULT_OOM;
1726                 goto out;
1727         }
1728
1729         /* No need to invalidate - it was non-present before */
1730         update_mmu_cache(vma, address, pte);
1731         lazy_mmu_prot_update(pte);
1732 unlock:
1733         pte_unmap_unlock(page_table, ptl);
1734 out:
1735         return ret;
1736 out_nomap:
1737         pte_unmap_unlock(page_table, ptl);
1738         unlock_page(page);
1739         page_cache_release(page);
1740         return ret;
1741 }
1742
1743 /*
1744  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1745  * but allow concurrent faults), and pte mapped but not yet locked.
1746  * We return with mmap_sem still held, but pte unmapped and unlocked.
1747  */
1748 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1749                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1750                 int write_access)
1751 {
1752         struct page *page;
1753         spinlock_t *ptl;
1754         pte_t entry;
1755
1756         if (write_access) {
1757                 /* Allocate our own private page. */
1758                 pte_unmap(page_table);
1759
1760                 if (unlikely(anon_vma_prepare(vma)))
1761                         goto oom;
1762                 page = alloc_zeroed_user_highpage(vma, address);
1763                 if (!page)
1764                         goto oom;
1765
1766                 entry = mk_pte(page, vma->vm_page_prot);
1767                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1768
1769                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1770                 if (!pte_none(*page_table))
1771                         goto release;
1772                 inc_mm_counter(mm, anon_rss);
1773                 lru_cache_add_active(page);
1774                 SetPageReferenced(page);
1775                 page_add_anon_rmap(page, vma, address);
1776         } else {
1777                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1778                 page = ZERO_PAGE(address);
1779                 page_cache_get(page);
1780                 entry = mk_pte(page, vma->vm_page_prot);
1781
1782                 ptl = pte_lockptr(mm, pmd);
1783                 spin_lock(ptl);
1784                 if (!pte_none(*page_table))
1785                         goto release;
1786                 inc_mm_counter(mm, file_rss);
1787                 page_add_file_rmap(page);
1788         }
1789
1790         set_pte_at(mm, address, page_table, entry);
1791
1792         /* No need to invalidate - it was non-present before */
1793         update_mmu_cache(vma, address, entry);
1794         lazy_mmu_prot_update(entry);
1795 unlock:
1796         pte_unmap_unlock(page_table, ptl);
1797         return VM_FAULT_MINOR;
1798 release:
1799         page_cache_release(page);
1800         goto unlock;
1801 oom:
1802         return VM_FAULT_OOM;
1803 }
1804
1805 /*
1806  * do_no_page() tries to create a new page mapping. It aggressively
1807  * tries to share with existing pages, but makes a separate copy if
1808  * the "write_access" parameter is true in order to avoid the next
1809  * page fault.
1810  *
1811  * As this is called only for pages that do not currently exist, we
1812  * do not need to flush old virtual caches or the TLB.
1813  *
1814  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1815  * but allow concurrent faults), and pte mapped but not yet locked.
1816  * We return with mmap_sem still held, but pte unmapped and unlocked.
1817  */
1818 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1819                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1820                 int write_access)
1821 {
1822         spinlock_t *ptl;
1823         struct page *new_page;
1824         struct address_space *mapping = NULL;
1825         pte_t entry;
1826         unsigned int sequence = 0;
1827         int ret = VM_FAULT_MINOR;
1828         int anon = 0;
1829
1830         pte_unmap(page_table);
1831
1832         if (vma->vm_file) {
1833                 mapping = vma->vm_file->f_mapping;
1834                 sequence = mapping->truncate_count;
1835                 smp_rmb(); /* serializes i_size against truncate_count */
1836         }
1837 retry:
1838         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1839         /*
1840          * No smp_rmb is needed here as long as there's a full
1841          * spin_lock/unlock sequence inside the ->nopage callback
1842          * (for the pagecache lookup) that acts as an implicit
1843          * smp_mb() and prevents the i_size read to happen
1844          * after the next truncate_count read.
1845          */
1846
1847         /* no page was available -- either SIGBUS or OOM */
1848         if (new_page == NOPAGE_SIGBUS)
1849                 return VM_FAULT_SIGBUS;
1850         if (new_page == NOPAGE_OOM)
1851                 return VM_FAULT_OOM;
1852
1853         /*
1854          * Should we do an early C-O-W break?
1855          */
1856         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1857                 struct page *page;
1858
1859                 if (unlikely(anon_vma_prepare(vma)))
1860                         goto oom;
1861                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1862                 if (!page)
1863                         goto oom;
1864                 copy_user_highpage(page, new_page, address);
1865                 page_cache_release(new_page);
1866                 new_page = page;
1867                 anon = 1;
1868         }
1869
1870         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1871         /*
1872          * For a file-backed vma, someone could have truncated or otherwise
1873          * invalidated this page.  If unmap_mapping_range got called,
1874          * retry getting the page.
1875          */
1876         if (mapping && unlikely(sequence != mapping->truncate_count)) {
1877                 pte_unmap_unlock(page_table, ptl);
1878                 page_cache_release(new_page);
1879                 cond_resched();
1880                 sequence = mapping->truncate_count;
1881                 smp_rmb();
1882                 goto retry;
1883         }
1884
1885         /*
1886          * This silly early PAGE_DIRTY setting removes a race
1887          * due to the bad i386 page protection. But it's valid
1888          * for other architectures too.
1889          *
1890          * Note that if write_access is true, we either now have
1891          * an exclusive copy of the page, or this is a shared mapping,
1892          * so we can make it writable and dirty to avoid having to
1893          * handle that later.
1894          */
1895         /* Only go through if we didn't race with anybody else... */
1896         if (pte_none(*page_table)) {
1897                 flush_icache_page(vma, new_page);
1898                 entry = mk_pte(new_page, vma->vm_page_prot);
1899                 if (write_access)
1900                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1901                 set_pte_at(mm, address, page_table, entry);
1902                 if (anon) {
1903                         inc_mm_counter(mm, anon_rss);
1904                         lru_cache_add_active(new_page);
1905                         page_add_anon_rmap(new_page, vma, address);
1906                 } else if (!(vma->vm_flags & VM_RESERVED)) {
1907                         inc_mm_counter(mm, file_rss);
1908                         page_add_file_rmap(new_page);
1909                 }
1910         } else {
1911                 /* One of our sibling threads was faster, back out. */
1912                 page_cache_release(new_page);
1913                 goto unlock;
1914         }
1915
1916         /* no need to invalidate: a not-present page shouldn't be cached */
1917         update_mmu_cache(vma, address, entry);
1918         lazy_mmu_prot_update(entry);
1919 unlock:
1920         pte_unmap_unlock(page_table, ptl);
1921         return ret;
1922 oom:
1923         page_cache_release(new_page);
1924         return VM_FAULT_OOM;
1925 }
1926
1927 /*
1928  * Fault of a previously existing named mapping. Repopulate the pte
1929  * from the encoded file_pte if possible. This enables swappable
1930  * nonlinear vmas.
1931  *
1932  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1933  * but allow concurrent faults), and pte mapped but not yet locked.
1934  * We return with mmap_sem still held, but pte unmapped and unlocked.
1935  */
1936 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
1937                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1938                 int write_access, pte_t orig_pte)
1939 {
1940         pgoff_t pgoff;
1941         int err;
1942
1943         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1944                 return VM_FAULT_MINOR;
1945
1946         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
1947                 /*
1948                  * Page table corrupted: show pte and kill process.
1949                  */
1950                 print_bad_pte(vma, orig_pte, address);
1951                 return VM_FAULT_OOM;
1952         }
1953         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
1954
1955         pgoff = pte_to_pgoff(orig_pte);
1956         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
1957                                         vma->vm_page_prot, pgoff, 0);
1958         if (err == -ENOMEM)
1959                 return VM_FAULT_OOM;
1960         if (err)
1961                 return VM_FAULT_SIGBUS;
1962         return VM_FAULT_MAJOR;
1963 }
1964
1965 /*
1966  * These routines also need to handle stuff like marking pages dirty
1967  * and/or accessed for architectures that don't do it in hardware (most
1968  * RISC architectures).  The early dirtying is also good on the i386.
1969  *
1970  * There is also a hook called "update_mmu_cache()" that architectures
1971  * with external mmu caches can use to update those (ie the Sparc or
1972  * PowerPC hashed page tables that act as extended TLBs).
1973  *
1974  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1975  * but allow concurrent faults), and pte mapped but not yet locked.
1976  * We return with mmap_sem still held, but pte unmapped and unlocked.
1977  */
1978 static inline int handle_pte_fault(struct mm_struct *mm,
1979                 struct vm_area_struct *vma, unsigned long address,
1980                 pte_t *pte, pmd_t *pmd, int write_access)
1981 {
1982         pte_t entry;
1983         pte_t old_entry;
1984         spinlock_t *ptl;
1985
1986         old_entry = entry = *pte;
1987         if (!pte_present(entry)) {
1988                 if (pte_none(entry)) {
1989                         if (!vma->vm_ops || !vma->vm_ops->nopage)
1990                                 return do_anonymous_page(mm, vma, address,
1991                                         pte, pmd, write_access);
1992                         return do_no_page(mm, vma, address,
1993                                         pte, pmd, write_access);
1994                 }
1995                 if (pte_file(entry))
1996                         return do_file_page(mm, vma, address,
1997                                         pte, pmd, write_access, entry);
1998                 return do_swap_page(mm, vma, address,
1999                                         pte, pmd, write_access, entry);
2000         }
2001
2002         ptl = pte_lockptr(mm, pmd);
2003         spin_lock(ptl);
2004         if (unlikely(!pte_same(*pte, entry)))
2005                 goto unlock;
2006         if (write_access) {
2007                 if (!pte_write(entry))
2008                         return do_wp_page(mm, vma, address,
2009                                         pte, pmd, ptl, entry);
2010                 entry = pte_mkdirty(entry);
2011         }
2012         entry = pte_mkyoung(entry);
2013         if (!pte_same(old_entry, entry)) {
2014                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2015                 update_mmu_cache(vma, address, entry);
2016                 lazy_mmu_prot_update(entry);
2017         } else {
2018                 /*
2019                  * This is needed only for protection faults but the arch code
2020                  * is not yet telling us if this is a protection fault or not.
2021                  * This still avoids useless tlb flushes for .text page faults
2022                  * with threads.
2023                  */
2024                 if (write_access)
2025                         flush_tlb_page(vma, address);
2026         }
2027 unlock:
2028         pte_unmap_unlock(pte, ptl);
2029         return VM_FAULT_MINOR;
2030 }
2031
2032 /*
2033  * By the time we get here, we already hold the mm semaphore
2034  */
2035 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2036                 unsigned long address, int write_access)
2037 {
2038         pgd_t *pgd;
2039         pud_t *pud;
2040         pmd_t *pmd;
2041         pte_t *pte;
2042
2043         __set_current_state(TASK_RUNNING);
2044
2045         inc_page_state(pgfault);
2046
2047         if (unlikely(is_vm_hugetlb_page(vma)))
2048                 return hugetlb_fault(mm, vma, address, write_access);
2049
2050         pgd = pgd_offset(mm, address);
2051         pud = pud_alloc(mm, pgd, address);
2052         if (!pud)
2053                 return VM_FAULT_OOM;
2054         pmd = pmd_alloc(mm, pud, address);
2055         if (!pmd)
2056                 return VM_FAULT_OOM;
2057         pte = pte_alloc_map(mm, pmd, address);
2058         if (!pte)
2059                 return VM_FAULT_OOM;
2060
2061         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2062 }
2063
2064 #ifndef __PAGETABLE_PUD_FOLDED
2065 /*
2066  * Allocate page upper directory.
2067  * We've already handled the fast-path in-line.
2068  */
2069 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2070 {
2071         pud_t *new = pud_alloc_one(mm, address);
2072         if (!new)
2073                 return -ENOMEM;
2074
2075         spin_lock(&mm->page_table_lock);
2076         if (pgd_present(*pgd))          /* Another has populated it */
2077                 pud_free(new);
2078         else
2079                 pgd_populate(mm, pgd, new);
2080         spin_unlock(&mm->page_table_lock);
2081         return 0;
2082 }
2083 #endif /* __PAGETABLE_PUD_FOLDED */
2084
2085 #ifndef __PAGETABLE_PMD_FOLDED
2086 /*
2087  * Allocate page middle directory.
2088  * We've already handled the fast-path in-line.
2089  */
2090 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2091 {
2092         pmd_t *new = pmd_alloc_one(mm, address);
2093         if (!new)
2094                 return -ENOMEM;
2095
2096         spin_lock(&mm->page_table_lock);
2097 #ifndef __ARCH_HAS_4LEVEL_HACK
2098         if (pud_present(*pud))          /* Another has populated it */
2099                 pmd_free(new);
2100         else
2101                 pud_populate(mm, pud, new);
2102 #else
2103         if (pgd_present(*pud))          /* Another has populated it */
2104                 pmd_free(new);
2105         else
2106                 pgd_populate(mm, pud, new);
2107 #endif /* __ARCH_HAS_4LEVEL_HACK */
2108         spin_unlock(&mm->page_table_lock);
2109         return 0;
2110 }
2111 #endif /* __PAGETABLE_PMD_FOLDED */
2112
2113 int make_pages_present(unsigned long addr, unsigned long end)
2114 {
2115         int ret, len, write;
2116         struct vm_area_struct * vma;
2117
2118         vma = find_vma(current->mm, addr);
2119         if (!vma)
2120                 return -1;
2121         write = (vma->vm_flags & VM_WRITE) != 0;
2122         if (addr >= end)
2123                 BUG();
2124         if (end > vma->vm_end)
2125                 BUG();
2126         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2127         ret = get_user_pages(current, current->mm, addr,
2128                         len, write, 0, NULL, NULL);
2129         if (ret < 0)
2130                 return ret;
2131         return ret == len ? 0 : -1;
2132 }
2133
2134 /* 
2135  * Map a vmalloc()-space virtual address to the physical page.
2136  */
2137 struct page * vmalloc_to_page(void * vmalloc_addr)
2138 {
2139         unsigned long addr = (unsigned long) vmalloc_addr;
2140         struct page *page = NULL;
2141         pgd_t *pgd = pgd_offset_k(addr);
2142         pud_t *pud;
2143         pmd_t *pmd;
2144         pte_t *ptep, pte;
2145   
2146         if (!pgd_none(*pgd)) {
2147                 pud = pud_offset(pgd, addr);
2148                 if (!pud_none(*pud)) {
2149                         pmd = pmd_offset(pud, addr);
2150                         if (!pmd_none(*pmd)) {
2151                                 ptep = pte_offset_map(pmd, addr);
2152                                 pte = *ptep;
2153                                 if (pte_present(pte))
2154                                         page = pte_page(pte);
2155                                 pte_unmap(ptep);
2156                         }
2157                 }
2158         }
2159         return page;
2160 }
2161
2162 EXPORT_SYMBOL(vmalloc_to_page);
2163
2164 /*
2165  * Map a vmalloc()-space virtual address to the physical page frame number.
2166  */
2167 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2168 {
2169         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2170 }
2171
2172 EXPORT_SYMBOL(vmalloc_to_pfn);
2173
2174 #if !defined(__HAVE_ARCH_GATE_AREA)
2175
2176 #if defined(AT_SYSINFO_EHDR)
2177 static struct vm_area_struct gate_vma;
2178
2179 static int __init gate_vma_init(void)
2180 {
2181         gate_vma.vm_mm = NULL;
2182         gate_vma.vm_start = FIXADDR_USER_START;
2183         gate_vma.vm_end = FIXADDR_USER_END;
2184         gate_vma.vm_page_prot = PAGE_READONLY;
2185         gate_vma.vm_flags = VM_RESERVED;
2186         return 0;
2187 }
2188 __initcall(gate_vma_init);
2189 #endif
2190
2191 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2192 {
2193 #ifdef AT_SYSINFO_EHDR
2194         return &gate_vma;
2195 #else
2196         return NULL;
2197 #endif
2198 }
2199
2200 int in_gate_area_no_task(unsigned long addr)
2201 {
2202 #ifdef AT_SYSINFO_EHDR
2203         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2204                 return 1;
2205 #endif
2206         return 0;
2207 }
2208
2209 #endif  /* __HAVE_ARCH_GATE_AREA */