]> git.karo-electronics.de Git - karo-tx-linux.git/blob - mm/memory.c
Be more robust about bad arguments in get_user_pages()
[karo-tx-linux.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/delayacct.h>
51 #include <linux/init.h>
52 #include <linux/writeback.h>
53
54 #include <asm/pgalloc.h>
55 #include <asm/uaccess.h>
56 #include <asm/tlb.h>
57 #include <asm/tlbflush.h>
58 #include <asm/pgtable.h>
59
60 #include <linux/swapops.h>
61 #include <linux/elf.h>
62
63 #ifndef CONFIG_NEED_MULTIPLE_NODES
64 /* use the per-pgdat data instead for discontigmem - mbligh */
65 unsigned long max_mapnr;
66 struct page *mem_map;
67
68 EXPORT_SYMBOL(max_mapnr);
69 EXPORT_SYMBOL(mem_map);
70 #endif
71
72 unsigned long num_physpages;
73 /*
74  * A number of key systems in x86 including ioremap() rely on the assumption
75  * that high_memory defines the upper bound on direct map memory, then end
76  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
77  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
78  * and ZONE_HIGHMEM.
79  */
80 void * high_memory;
81 unsigned long vmalloc_earlyreserve;
82
83 EXPORT_SYMBOL(num_physpages);
84 EXPORT_SYMBOL(high_memory);
85 EXPORT_SYMBOL(vmalloc_earlyreserve);
86
87 int randomize_va_space __read_mostly = 1;
88
89 static int __init disable_randmaps(char *s)
90 {
91         randomize_va_space = 0;
92         return 1;
93 }
94 __setup("norandmaps", disable_randmaps);
95
96
97 /*
98  * If a p?d_bad entry is found while walking page tables, report
99  * the error, before resetting entry to p?d_none.  Usually (but
100  * very seldom) called out from the p?d_none_or_clear_bad macros.
101  */
102
103 void pgd_clear_bad(pgd_t *pgd)
104 {
105         pgd_ERROR(*pgd);
106         pgd_clear(pgd);
107 }
108
109 void pud_clear_bad(pud_t *pud)
110 {
111         pud_ERROR(*pud);
112         pud_clear(pud);
113 }
114
115 void pmd_clear_bad(pmd_t *pmd)
116 {
117         pmd_ERROR(*pmd);
118         pmd_clear(pmd);
119 }
120
121 /*
122  * Note: this doesn't free the actual pages themselves. That
123  * has been handled earlier when unmapping all the memory regions.
124  */
125 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
126 {
127         struct page *page = pmd_page(*pmd);
128         pmd_clear(pmd);
129         pte_lock_deinit(page);
130         pte_free_tlb(tlb, page);
131         dec_zone_page_state(page, NR_PAGETABLE);
132         tlb->mm->nr_ptes--;
133 }
134
135 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
136                                 unsigned long addr, unsigned long end,
137                                 unsigned long floor, unsigned long ceiling)
138 {
139         pmd_t *pmd;
140         unsigned long next;
141         unsigned long start;
142
143         start = addr;
144         pmd = pmd_offset(pud, addr);
145         do {
146                 next = pmd_addr_end(addr, end);
147                 if (pmd_none_or_clear_bad(pmd))
148                         continue;
149                 free_pte_range(tlb, pmd);
150         } while (pmd++, addr = next, addr != end);
151
152         start &= PUD_MASK;
153         if (start < floor)
154                 return;
155         if (ceiling) {
156                 ceiling &= PUD_MASK;
157                 if (!ceiling)
158                         return;
159         }
160         if (end - 1 > ceiling - 1)
161                 return;
162
163         pmd = pmd_offset(pud, start);
164         pud_clear(pud);
165         pmd_free_tlb(tlb, pmd);
166 }
167
168 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
169                                 unsigned long addr, unsigned long end,
170                                 unsigned long floor, unsigned long ceiling)
171 {
172         pud_t *pud;
173         unsigned long next;
174         unsigned long start;
175
176         start = addr;
177         pud = pud_offset(pgd, addr);
178         do {
179                 next = pud_addr_end(addr, end);
180                 if (pud_none_or_clear_bad(pud))
181                         continue;
182                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
183         } while (pud++, addr = next, addr != end);
184
185         start &= PGDIR_MASK;
186         if (start < floor)
187                 return;
188         if (ceiling) {
189                 ceiling &= PGDIR_MASK;
190                 if (!ceiling)
191                         return;
192         }
193         if (end - 1 > ceiling - 1)
194                 return;
195
196         pud = pud_offset(pgd, start);
197         pgd_clear(pgd);
198         pud_free_tlb(tlb, pud);
199 }
200
201 /*
202  * This function frees user-level page tables of a process.
203  *
204  * Must be called with pagetable lock held.
205  */
206 void free_pgd_range(struct mmu_gather **tlb,
207                         unsigned long addr, unsigned long end,
208                         unsigned long floor, unsigned long ceiling)
209 {
210         pgd_t *pgd;
211         unsigned long next;
212         unsigned long start;
213
214         /*
215          * The next few lines have given us lots of grief...
216          *
217          * Why are we testing PMD* at this top level?  Because often
218          * there will be no work to do at all, and we'd prefer not to
219          * go all the way down to the bottom just to discover that.
220          *
221          * Why all these "- 1"s?  Because 0 represents both the bottom
222          * of the address space and the top of it (using -1 for the
223          * top wouldn't help much: the masks would do the wrong thing).
224          * The rule is that addr 0 and floor 0 refer to the bottom of
225          * the address space, but end 0 and ceiling 0 refer to the top
226          * Comparisons need to use "end - 1" and "ceiling - 1" (though
227          * that end 0 case should be mythical).
228          *
229          * Wherever addr is brought up or ceiling brought down, we must
230          * be careful to reject "the opposite 0" before it confuses the
231          * subsequent tests.  But what about where end is brought down
232          * by PMD_SIZE below? no, end can't go down to 0 there.
233          *
234          * Whereas we round start (addr) and ceiling down, by different
235          * masks at different levels, in order to test whether a table
236          * now has no other vmas using it, so can be freed, we don't
237          * bother to round floor or end up - the tests don't need that.
238          */
239
240         addr &= PMD_MASK;
241         if (addr < floor) {
242                 addr += PMD_SIZE;
243                 if (!addr)
244                         return;
245         }
246         if (ceiling) {
247                 ceiling &= PMD_MASK;
248                 if (!ceiling)
249                         return;
250         }
251         if (end - 1 > ceiling - 1)
252                 end -= PMD_SIZE;
253         if (addr > end - 1)
254                 return;
255
256         start = addr;
257         pgd = pgd_offset((*tlb)->mm, addr);
258         do {
259                 next = pgd_addr_end(addr, end);
260                 if (pgd_none_or_clear_bad(pgd))
261                         continue;
262                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
263         } while (pgd++, addr = next, addr != end);
264
265         if (!(*tlb)->fullmm)
266                 flush_tlb_pgtables((*tlb)->mm, start, end);
267 }
268
269 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
270                 unsigned long floor, unsigned long ceiling)
271 {
272         while (vma) {
273                 struct vm_area_struct *next = vma->vm_next;
274                 unsigned long addr = vma->vm_start;
275
276                 /*
277                  * Hide vma from rmap and vmtruncate before freeing pgtables
278                  */
279                 anon_vma_unlink(vma);
280                 unlink_file_vma(vma);
281
282                 if (is_vm_hugetlb_page(vma)) {
283                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
284                                 floor, next? next->vm_start: ceiling);
285                 } else {
286                         /*
287                          * Optimization: gather nearby vmas into one call down
288                          */
289                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
290                                && !is_vm_hugetlb_page(next)) {
291                                 vma = next;
292                                 next = vma->vm_next;
293                                 anon_vma_unlink(vma);
294                                 unlink_file_vma(vma);
295                         }
296                         free_pgd_range(tlb, addr, vma->vm_end,
297                                 floor, next? next->vm_start: ceiling);
298                 }
299                 vma = next;
300         }
301 }
302
303 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
304 {
305         struct page *new = pte_alloc_one(mm, address);
306         if (!new)
307                 return -ENOMEM;
308
309         pte_lock_init(new);
310         spin_lock(&mm->page_table_lock);
311         if (pmd_present(*pmd)) {        /* Another has populated it */
312                 pte_lock_deinit(new);
313                 pte_free(new);
314         } else {
315                 mm->nr_ptes++;
316                 inc_zone_page_state(new, NR_PAGETABLE);
317                 pmd_populate(mm, pmd, new);
318         }
319         spin_unlock(&mm->page_table_lock);
320         return 0;
321 }
322
323 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
324 {
325         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
326         if (!new)
327                 return -ENOMEM;
328
329         spin_lock(&init_mm.page_table_lock);
330         if (pmd_present(*pmd))          /* Another has populated it */
331                 pte_free_kernel(new);
332         else
333                 pmd_populate_kernel(&init_mm, pmd, new);
334         spin_unlock(&init_mm.page_table_lock);
335         return 0;
336 }
337
338 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
339 {
340         if (file_rss)
341                 add_mm_counter(mm, file_rss, file_rss);
342         if (anon_rss)
343                 add_mm_counter(mm, anon_rss, anon_rss);
344 }
345
346 /*
347  * This function is called to print an error when a bad pte
348  * is found. For example, we might have a PFN-mapped pte in
349  * a region that doesn't allow it.
350  *
351  * The calling function must still handle the error.
352  */
353 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
354 {
355         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
356                         "vm_flags = %lx, vaddr = %lx\n",
357                 (long long)pte_val(pte),
358                 (vma->vm_mm == current->mm ? current->comm : "???"),
359                 vma->vm_flags, vaddr);
360         dump_stack();
361 }
362
363 static inline int is_cow_mapping(unsigned int flags)
364 {
365         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
366 }
367
368 /*
369  * This function gets the "struct page" associated with a pte.
370  *
371  * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
372  * will have each page table entry just pointing to a raw page frame
373  * number, and as far as the VM layer is concerned, those do not have
374  * pages associated with them - even if the PFN might point to memory
375  * that otherwise is perfectly fine and has a "struct page".
376  *
377  * The way we recognize those mappings is through the rules set up
378  * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
379  * and the vm_pgoff will point to the first PFN mapped: thus every
380  * page that is a raw mapping will always honor the rule
381  *
382  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
383  *
384  * and if that isn't true, the page has been COW'ed (in which case it
385  * _does_ have a "struct page" associated with it even if it is in a
386  * VM_PFNMAP range).
387  */
388 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
389 {
390         unsigned long pfn = pte_pfn(pte);
391
392         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
393                 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
394                 if (pfn == vma->vm_pgoff + off)
395                         return NULL;
396                 if (!is_cow_mapping(vma->vm_flags))
397                         return NULL;
398         }
399
400         /*
401          * Add some anal sanity checks for now. Eventually,
402          * we should just do "return pfn_to_page(pfn)", but
403          * in the meantime we check that we get a valid pfn,
404          * and that the resulting page looks ok.
405          */
406         if (unlikely(!pfn_valid(pfn))) {
407                 print_bad_pte(vma, pte, addr);
408                 return NULL;
409         }
410
411         /*
412          * NOTE! We still have PageReserved() pages in the page 
413          * tables. 
414          *
415          * The PAGE_ZERO() pages and various VDSO mappings can
416          * cause them to exist.
417          */
418         return pfn_to_page(pfn);
419 }
420
421 /*
422  * copy one vm_area from one task to the other. Assumes the page tables
423  * already present in the new task to be cleared in the whole range
424  * covered by this vma.
425  */
426
427 static inline void
428 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
429                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
430                 unsigned long addr, int *rss)
431 {
432         unsigned long vm_flags = vma->vm_flags;
433         pte_t pte = *src_pte;
434         struct page *page;
435
436         /* pte contains position in swap or file, so copy. */
437         if (unlikely(!pte_present(pte))) {
438                 if (!pte_file(pte)) {
439                         swp_entry_t entry = pte_to_swp_entry(pte);
440
441                         swap_duplicate(entry);
442                         /* make sure dst_mm is on swapoff's mmlist. */
443                         if (unlikely(list_empty(&dst_mm->mmlist))) {
444                                 spin_lock(&mmlist_lock);
445                                 if (list_empty(&dst_mm->mmlist))
446                                         list_add(&dst_mm->mmlist,
447                                                  &src_mm->mmlist);
448                                 spin_unlock(&mmlist_lock);
449                         }
450                         if (is_write_migration_entry(entry) &&
451                                         is_cow_mapping(vm_flags)) {
452                                 /*
453                                  * COW mappings require pages in both parent
454                                  * and child to be set to read.
455                                  */
456                                 make_migration_entry_read(&entry);
457                                 pte = swp_entry_to_pte(entry);
458                                 set_pte_at(src_mm, addr, src_pte, pte);
459                         }
460                 }
461                 goto out_set_pte;
462         }
463
464         /*
465          * If it's a COW mapping, write protect it both
466          * in the parent and the child
467          */
468         if (is_cow_mapping(vm_flags)) {
469                 ptep_set_wrprotect(src_mm, addr, src_pte);
470                 pte = pte_wrprotect(pte);
471         }
472
473         /*
474          * If it's a shared mapping, mark it clean in
475          * the child
476          */
477         if (vm_flags & VM_SHARED)
478                 pte = pte_mkclean(pte);
479         pte = pte_mkold(pte);
480
481         page = vm_normal_page(vma, addr, pte);
482         if (page) {
483                 get_page(page);
484                 page_dup_rmap(page, vma, addr);
485                 rss[!!PageAnon(page)]++;
486         }
487
488 out_set_pte:
489         set_pte_at(dst_mm, addr, dst_pte, pte);
490 }
491
492 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
493                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
494                 unsigned long addr, unsigned long end)
495 {
496         pte_t *src_pte, *dst_pte;
497         spinlock_t *src_ptl, *dst_ptl;
498         int progress = 0;
499         int rss[2];
500
501 again:
502         rss[1] = rss[0] = 0;
503         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
504         if (!dst_pte)
505                 return -ENOMEM;
506         src_pte = pte_offset_map_nested(src_pmd, addr);
507         src_ptl = pte_lockptr(src_mm, src_pmd);
508         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
509         arch_enter_lazy_mmu_mode();
510
511         do {
512                 /*
513                  * We are holding two locks at this point - either of them
514                  * could generate latencies in another task on another CPU.
515                  */
516                 if (progress >= 32) {
517                         progress = 0;
518                         if (need_resched() ||
519                             need_lockbreak(src_ptl) ||
520                             need_lockbreak(dst_ptl))
521                                 break;
522                 }
523                 if (pte_none(*src_pte)) {
524                         progress++;
525                         continue;
526                 }
527                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
528                 progress += 8;
529         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
530
531         arch_leave_lazy_mmu_mode();
532         spin_unlock(src_ptl);
533         pte_unmap_nested(src_pte - 1);
534         add_mm_rss(dst_mm, rss[0], rss[1]);
535         pte_unmap_unlock(dst_pte - 1, dst_ptl);
536         cond_resched();
537         if (addr != end)
538                 goto again;
539         return 0;
540 }
541
542 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
543                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
544                 unsigned long addr, unsigned long end)
545 {
546         pmd_t *src_pmd, *dst_pmd;
547         unsigned long next;
548
549         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
550         if (!dst_pmd)
551                 return -ENOMEM;
552         src_pmd = pmd_offset(src_pud, addr);
553         do {
554                 next = pmd_addr_end(addr, end);
555                 if (pmd_none_or_clear_bad(src_pmd))
556                         continue;
557                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
558                                                 vma, addr, next))
559                         return -ENOMEM;
560         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
561         return 0;
562 }
563
564 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
565                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
566                 unsigned long addr, unsigned long end)
567 {
568         pud_t *src_pud, *dst_pud;
569         unsigned long next;
570
571         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
572         if (!dst_pud)
573                 return -ENOMEM;
574         src_pud = pud_offset(src_pgd, addr);
575         do {
576                 next = pud_addr_end(addr, end);
577                 if (pud_none_or_clear_bad(src_pud))
578                         continue;
579                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
580                                                 vma, addr, next))
581                         return -ENOMEM;
582         } while (dst_pud++, src_pud++, addr = next, addr != end);
583         return 0;
584 }
585
586 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
587                 struct vm_area_struct *vma)
588 {
589         pgd_t *src_pgd, *dst_pgd;
590         unsigned long next;
591         unsigned long addr = vma->vm_start;
592         unsigned long end = vma->vm_end;
593
594         /*
595          * Don't copy ptes where a page fault will fill them correctly.
596          * Fork becomes much lighter when there are big shared or private
597          * readonly mappings. The tradeoff is that copy_page_range is more
598          * efficient than faulting.
599          */
600         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
601                 if (!vma->anon_vma)
602                         return 0;
603         }
604
605         if (is_vm_hugetlb_page(vma))
606                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
607
608         dst_pgd = pgd_offset(dst_mm, addr);
609         src_pgd = pgd_offset(src_mm, addr);
610         do {
611                 next = pgd_addr_end(addr, end);
612                 if (pgd_none_or_clear_bad(src_pgd))
613                         continue;
614                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
615                                                 vma, addr, next))
616                         return -ENOMEM;
617         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
618         return 0;
619 }
620
621 static unsigned long zap_pte_range(struct mmu_gather *tlb,
622                                 struct vm_area_struct *vma, pmd_t *pmd,
623                                 unsigned long addr, unsigned long end,
624                                 long *zap_work, struct zap_details *details)
625 {
626         struct mm_struct *mm = tlb->mm;
627         pte_t *pte;
628         spinlock_t *ptl;
629         int file_rss = 0;
630         int anon_rss = 0;
631
632         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
633         arch_enter_lazy_mmu_mode();
634         do {
635                 pte_t ptent = *pte;
636                 if (pte_none(ptent)) {
637                         (*zap_work)--;
638                         continue;
639                 }
640
641                 (*zap_work) -= PAGE_SIZE;
642
643                 if (pte_present(ptent)) {
644                         struct page *page;
645
646                         page = vm_normal_page(vma, addr, ptent);
647                         if (unlikely(details) && page) {
648                                 /*
649                                  * unmap_shared_mapping_pages() wants to
650                                  * invalidate cache without truncating:
651                                  * unmap shared but keep private pages.
652                                  */
653                                 if (details->check_mapping &&
654                                     details->check_mapping != page->mapping)
655                                         continue;
656                                 /*
657                                  * Each page->index must be checked when
658                                  * invalidating or truncating nonlinear.
659                                  */
660                                 if (details->nonlinear_vma &&
661                                     (page->index < details->first_index ||
662                                      page->index > details->last_index))
663                                         continue;
664                         }
665                         ptent = ptep_get_and_clear_full(mm, addr, pte,
666                                                         tlb->fullmm);
667                         tlb_remove_tlb_entry(tlb, pte, addr);
668                         if (unlikely(!page))
669                                 continue;
670                         if (unlikely(details) && details->nonlinear_vma
671                             && linear_page_index(details->nonlinear_vma,
672                                                 addr) != page->index)
673                                 set_pte_at(mm, addr, pte,
674                                            pgoff_to_pte(page->index));
675                         if (PageAnon(page))
676                                 anon_rss--;
677                         else {
678                                 if (pte_dirty(ptent))
679                                         set_page_dirty(page);
680                                 if (pte_young(ptent))
681                                         SetPageReferenced(page);
682                                 file_rss--;
683                         }
684                         page_remove_rmap(page, vma);
685                         tlb_remove_page(tlb, page);
686                         continue;
687                 }
688                 /*
689                  * If details->check_mapping, we leave swap entries;
690                  * if details->nonlinear_vma, we leave file entries.
691                  */
692                 if (unlikely(details))
693                         continue;
694                 if (!pte_file(ptent))
695                         free_swap_and_cache(pte_to_swp_entry(ptent));
696                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
697         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
698
699         add_mm_rss(mm, file_rss, anon_rss);
700         arch_leave_lazy_mmu_mode();
701         pte_unmap_unlock(pte - 1, ptl);
702
703         return addr;
704 }
705
706 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
707                                 struct vm_area_struct *vma, pud_t *pud,
708                                 unsigned long addr, unsigned long end,
709                                 long *zap_work, struct zap_details *details)
710 {
711         pmd_t *pmd;
712         unsigned long next;
713
714         pmd = pmd_offset(pud, addr);
715         do {
716                 next = pmd_addr_end(addr, end);
717                 if (pmd_none_or_clear_bad(pmd)) {
718                         (*zap_work)--;
719                         continue;
720                 }
721                 next = zap_pte_range(tlb, vma, pmd, addr, next,
722                                                 zap_work, details);
723         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
724
725         return addr;
726 }
727
728 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
729                                 struct vm_area_struct *vma, pgd_t *pgd,
730                                 unsigned long addr, unsigned long end,
731                                 long *zap_work, struct zap_details *details)
732 {
733         pud_t *pud;
734         unsigned long next;
735
736         pud = pud_offset(pgd, addr);
737         do {
738                 next = pud_addr_end(addr, end);
739                 if (pud_none_or_clear_bad(pud)) {
740                         (*zap_work)--;
741                         continue;
742                 }
743                 next = zap_pmd_range(tlb, vma, pud, addr, next,
744                                                 zap_work, details);
745         } while (pud++, addr = next, (addr != end && *zap_work > 0));
746
747         return addr;
748 }
749
750 static unsigned long unmap_page_range(struct mmu_gather *tlb,
751                                 struct vm_area_struct *vma,
752                                 unsigned long addr, unsigned long end,
753                                 long *zap_work, struct zap_details *details)
754 {
755         pgd_t *pgd;
756         unsigned long next;
757
758         if (details && !details->check_mapping && !details->nonlinear_vma)
759                 details = NULL;
760
761         BUG_ON(addr >= end);
762         tlb_start_vma(tlb, vma);
763         pgd = pgd_offset(vma->vm_mm, addr);
764         do {
765                 next = pgd_addr_end(addr, end);
766                 if (pgd_none_or_clear_bad(pgd)) {
767                         (*zap_work)--;
768                         continue;
769                 }
770                 next = zap_pud_range(tlb, vma, pgd, addr, next,
771                                                 zap_work, details);
772         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
773         tlb_end_vma(tlb, vma);
774
775         return addr;
776 }
777
778 #ifdef CONFIG_PREEMPT
779 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
780 #else
781 /* No preempt: go for improved straight-line efficiency */
782 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
783 #endif
784
785 /**
786  * unmap_vmas - unmap a range of memory covered by a list of vma's
787  * @tlbp: address of the caller's struct mmu_gather
788  * @vma: the starting vma
789  * @start_addr: virtual address at which to start unmapping
790  * @end_addr: virtual address at which to end unmapping
791  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
792  * @details: details of nonlinear truncation or shared cache invalidation
793  *
794  * Returns the end address of the unmapping (restart addr if interrupted).
795  *
796  * Unmap all pages in the vma list.
797  *
798  * We aim to not hold locks for too long (for scheduling latency reasons).
799  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
800  * return the ending mmu_gather to the caller.
801  *
802  * Only addresses between `start' and `end' will be unmapped.
803  *
804  * The VMA list must be sorted in ascending virtual address order.
805  *
806  * unmap_vmas() assumes that the caller will flush the whole unmapped address
807  * range after unmap_vmas() returns.  So the only responsibility here is to
808  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
809  * drops the lock and schedules.
810  */
811 unsigned long unmap_vmas(struct mmu_gather **tlbp,
812                 struct vm_area_struct *vma, unsigned long start_addr,
813                 unsigned long end_addr, unsigned long *nr_accounted,
814                 struct zap_details *details)
815 {
816         long zap_work = ZAP_BLOCK_SIZE;
817         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
818         int tlb_start_valid = 0;
819         unsigned long start = start_addr;
820         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
821         int fullmm = (*tlbp)->fullmm;
822
823         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
824                 unsigned long end;
825
826                 start = max(vma->vm_start, start_addr);
827                 if (start >= vma->vm_end)
828                         continue;
829                 end = min(vma->vm_end, end_addr);
830                 if (end <= vma->vm_start)
831                         continue;
832
833                 if (vma->vm_flags & VM_ACCOUNT)
834                         *nr_accounted += (end - start) >> PAGE_SHIFT;
835
836                 while (start != end) {
837                         if (!tlb_start_valid) {
838                                 tlb_start = start;
839                                 tlb_start_valid = 1;
840                         }
841
842                         if (unlikely(is_vm_hugetlb_page(vma))) {
843                                 unmap_hugepage_range(vma, start, end);
844                                 zap_work -= (end - start) /
845                                                 (HPAGE_SIZE / PAGE_SIZE);
846                                 start = end;
847                         } else
848                                 start = unmap_page_range(*tlbp, vma,
849                                                 start, end, &zap_work, details);
850
851                         if (zap_work > 0) {
852                                 BUG_ON(start != end);
853                                 break;
854                         }
855
856                         tlb_finish_mmu(*tlbp, tlb_start, start);
857
858                         if (need_resched() ||
859                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
860                                 if (i_mmap_lock) {
861                                         *tlbp = NULL;
862                                         goto out;
863                                 }
864                                 cond_resched();
865                         }
866
867                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
868                         tlb_start_valid = 0;
869                         zap_work = ZAP_BLOCK_SIZE;
870                 }
871         }
872 out:
873         return start;   /* which is now the end (or restart) address */
874 }
875
876 /**
877  * zap_page_range - remove user pages in a given range
878  * @vma: vm_area_struct holding the applicable pages
879  * @address: starting address of pages to zap
880  * @size: number of bytes to zap
881  * @details: details of nonlinear truncation or shared cache invalidation
882  */
883 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
884                 unsigned long size, struct zap_details *details)
885 {
886         struct mm_struct *mm = vma->vm_mm;
887         struct mmu_gather *tlb;
888         unsigned long end = address + size;
889         unsigned long nr_accounted = 0;
890
891         lru_add_drain();
892         tlb = tlb_gather_mmu(mm, 0);
893         update_hiwater_rss(mm);
894         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
895         if (tlb)
896                 tlb_finish_mmu(tlb, address, end);
897         return end;
898 }
899
900 /*
901  * Do a quick page-table lookup for a single page.
902  */
903 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
904                         unsigned int flags)
905 {
906         pgd_t *pgd;
907         pud_t *pud;
908         pmd_t *pmd;
909         pte_t *ptep, pte;
910         spinlock_t *ptl;
911         struct page *page;
912         struct mm_struct *mm = vma->vm_mm;
913
914         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
915         if (!IS_ERR(page)) {
916                 BUG_ON(flags & FOLL_GET);
917                 goto out;
918         }
919
920         page = NULL;
921         pgd = pgd_offset(mm, address);
922         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
923                 goto no_page_table;
924
925         pud = pud_offset(pgd, address);
926         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
927                 goto no_page_table;
928         
929         pmd = pmd_offset(pud, address);
930         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
931                 goto no_page_table;
932
933         if (pmd_huge(*pmd)) {
934                 BUG_ON(flags & FOLL_GET);
935                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
936                 goto out;
937         }
938
939         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
940         if (!ptep)
941                 goto out;
942
943         pte = *ptep;
944         if (!pte_present(pte))
945                 goto unlock;
946         if ((flags & FOLL_WRITE) && !pte_write(pte))
947                 goto unlock;
948         page = vm_normal_page(vma, address, pte);
949         if (unlikely(!page))
950                 goto unlock;
951
952         if (flags & FOLL_GET)
953                 get_page(page);
954         if (flags & FOLL_TOUCH) {
955                 if ((flags & FOLL_WRITE) &&
956                     !pte_dirty(pte) && !PageDirty(page))
957                         set_page_dirty(page);
958                 mark_page_accessed(page);
959         }
960 unlock:
961         pte_unmap_unlock(ptep, ptl);
962 out:
963         return page;
964
965 no_page_table:
966         /*
967          * When core dumping an enormous anonymous area that nobody
968          * has touched so far, we don't want to allocate page tables.
969          */
970         if (flags & FOLL_ANON) {
971                 page = ZERO_PAGE(address);
972                 if (flags & FOLL_GET)
973                         get_page(page);
974                 BUG_ON(flags & FOLL_WRITE);
975         }
976         return page;
977 }
978
979 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
980                 unsigned long start, int len, int write, int force,
981                 struct page **pages, struct vm_area_struct **vmas)
982 {
983         int i;
984         unsigned int vm_flags;
985
986         if (len <= 0)
987                 return 0;
988         /* 
989          * Require read or write permissions.
990          * If 'force' is set, we only require the "MAY" flags.
991          */
992         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
993         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
994         i = 0;
995
996         do {
997                 struct vm_area_struct *vma;
998                 unsigned int foll_flags;
999
1000                 vma = find_extend_vma(mm, start);
1001                 if (!vma && in_gate_area(tsk, start)) {
1002                         unsigned long pg = start & PAGE_MASK;
1003                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1004                         pgd_t *pgd;
1005                         pud_t *pud;
1006                         pmd_t *pmd;
1007                         pte_t *pte;
1008                         if (write) /* user gate pages are read-only */
1009                                 return i ? : -EFAULT;
1010                         if (pg > TASK_SIZE)
1011                                 pgd = pgd_offset_k(pg);
1012                         else
1013                                 pgd = pgd_offset_gate(mm, pg);
1014                         BUG_ON(pgd_none(*pgd));
1015                         pud = pud_offset(pgd, pg);
1016                         BUG_ON(pud_none(*pud));
1017                         pmd = pmd_offset(pud, pg);
1018                         if (pmd_none(*pmd))
1019                                 return i ? : -EFAULT;
1020                         pte = pte_offset_map(pmd, pg);
1021                         if (pte_none(*pte)) {
1022                                 pte_unmap(pte);
1023                                 return i ? : -EFAULT;
1024                         }
1025                         if (pages) {
1026                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1027                                 pages[i] = page;
1028                                 if (page)
1029                                         get_page(page);
1030                         }
1031                         pte_unmap(pte);
1032                         if (vmas)
1033                                 vmas[i] = gate_vma;
1034                         i++;
1035                         start += PAGE_SIZE;
1036                         len--;
1037                         continue;
1038                 }
1039
1040                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1041                                 || !(vm_flags & vma->vm_flags))
1042                         return i ? : -EFAULT;
1043
1044                 if (is_vm_hugetlb_page(vma)) {
1045                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1046                                                 &start, &len, i);
1047                         continue;
1048                 }
1049
1050                 foll_flags = FOLL_TOUCH;
1051                 if (pages)
1052                         foll_flags |= FOLL_GET;
1053                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1054                     (!vma->vm_ops || !vma->vm_ops->nopage))
1055                         foll_flags |= FOLL_ANON;
1056
1057                 do {
1058                         struct page *page;
1059
1060                         if (write)
1061                                 foll_flags |= FOLL_WRITE;
1062
1063                         cond_resched();
1064                         while (!(page = follow_page(vma, start, foll_flags))) {
1065                                 int ret;
1066                                 ret = __handle_mm_fault(mm, vma, start,
1067                                                 foll_flags & FOLL_WRITE);
1068                                 /*
1069                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1070                                  * broken COW when necessary, even if maybe_mkwrite
1071                                  * decided not to set pte_write. We can thus safely do
1072                                  * subsequent page lookups as if they were reads.
1073                                  */
1074                                 if (ret & VM_FAULT_WRITE)
1075                                         foll_flags &= ~FOLL_WRITE;
1076                                 
1077                                 switch (ret & ~VM_FAULT_WRITE) {
1078                                 case VM_FAULT_MINOR:
1079                                         tsk->min_flt++;
1080                                         break;
1081                                 case VM_FAULT_MAJOR:
1082                                         tsk->maj_flt++;
1083                                         break;
1084                                 case VM_FAULT_SIGBUS:
1085                                         return i ? i : -EFAULT;
1086                                 case VM_FAULT_OOM:
1087                                         return i ? i : -ENOMEM;
1088                                 default:
1089                                         BUG();
1090                                 }
1091                                 cond_resched();
1092                         }
1093                         if (pages) {
1094                                 pages[i] = page;
1095
1096                                 flush_anon_page(vma, page, start);
1097                                 flush_dcache_page(page);
1098                         }
1099                         if (vmas)
1100                                 vmas[i] = vma;
1101                         i++;
1102                         start += PAGE_SIZE;
1103                         len--;
1104                 } while (len && start < vma->vm_end);
1105         } while (len);
1106         return i;
1107 }
1108 EXPORT_SYMBOL(get_user_pages);
1109
1110 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1111                         unsigned long addr, unsigned long end, pgprot_t prot)
1112 {
1113         pte_t *pte;
1114         spinlock_t *ptl;
1115         int err = 0;
1116
1117         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1118         if (!pte)
1119                 return -EAGAIN;
1120         arch_enter_lazy_mmu_mode();
1121         do {
1122                 struct page *page = ZERO_PAGE(addr);
1123                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1124
1125                 if (unlikely(!pte_none(*pte))) {
1126                         err = -EEXIST;
1127                         pte++;
1128                         break;
1129                 }
1130                 page_cache_get(page);
1131                 page_add_file_rmap(page);
1132                 inc_mm_counter(mm, file_rss);
1133                 set_pte_at(mm, addr, pte, zero_pte);
1134         } while (pte++, addr += PAGE_SIZE, addr != end);
1135         arch_leave_lazy_mmu_mode();
1136         pte_unmap_unlock(pte - 1, ptl);
1137         return err;
1138 }
1139
1140 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1141                         unsigned long addr, unsigned long end, pgprot_t prot)
1142 {
1143         pmd_t *pmd;
1144         unsigned long next;
1145         int err;
1146
1147         pmd = pmd_alloc(mm, pud, addr);
1148         if (!pmd)
1149                 return -EAGAIN;
1150         do {
1151                 next = pmd_addr_end(addr, end);
1152                 err = zeromap_pte_range(mm, pmd, addr, next, prot);
1153                 if (err)
1154                         break;
1155         } while (pmd++, addr = next, addr != end);
1156         return err;
1157 }
1158
1159 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1160                         unsigned long addr, unsigned long end, pgprot_t prot)
1161 {
1162         pud_t *pud;
1163         unsigned long next;
1164         int err;
1165
1166         pud = pud_alloc(mm, pgd, addr);
1167         if (!pud)
1168                 return -EAGAIN;
1169         do {
1170                 next = pud_addr_end(addr, end);
1171                 err = zeromap_pmd_range(mm, pud, addr, next, prot);
1172                 if (err)
1173                         break;
1174         } while (pud++, addr = next, addr != end);
1175         return err;
1176 }
1177
1178 int zeromap_page_range(struct vm_area_struct *vma,
1179                         unsigned long addr, unsigned long size, pgprot_t prot)
1180 {
1181         pgd_t *pgd;
1182         unsigned long next;
1183         unsigned long end = addr + size;
1184         struct mm_struct *mm = vma->vm_mm;
1185         int err;
1186
1187         BUG_ON(addr >= end);
1188         pgd = pgd_offset(mm, addr);
1189         flush_cache_range(vma, addr, end);
1190         do {
1191                 next = pgd_addr_end(addr, end);
1192                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1193                 if (err)
1194                         break;
1195         } while (pgd++, addr = next, addr != end);
1196         return err;
1197 }
1198
1199 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1200 {
1201         pgd_t * pgd = pgd_offset(mm, addr);
1202         pud_t * pud = pud_alloc(mm, pgd, addr);
1203         if (pud) {
1204                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1205                 if (pmd)
1206                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1207         }
1208         return NULL;
1209 }
1210
1211 /*
1212  * This is the old fallback for page remapping.
1213  *
1214  * For historical reasons, it only allows reserved pages. Only
1215  * old drivers should use this, and they needed to mark their
1216  * pages reserved for the old functions anyway.
1217  */
1218 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1219 {
1220         int retval;
1221         pte_t *pte;
1222         spinlock_t *ptl;  
1223
1224         retval = -EINVAL;
1225         if (PageAnon(page))
1226                 goto out;
1227         retval = -ENOMEM;
1228         flush_dcache_page(page);
1229         pte = get_locked_pte(mm, addr, &ptl);
1230         if (!pte)
1231                 goto out;
1232         retval = -EBUSY;
1233         if (!pte_none(*pte))
1234                 goto out_unlock;
1235
1236         /* Ok, finally just insert the thing.. */
1237         get_page(page);
1238         inc_mm_counter(mm, file_rss);
1239         page_add_file_rmap(page);
1240         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1241
1242         retval = 0;
1243 out_unlock:
1244         pte_unmap_unlock(pte, ptl);
1245 out:
1246         return retval;
1247 }
1248
1249 /**
1250  * vm_insert_page - insert single page into user vma
1251  * @vma: user vma to map to
1252  * @addr: target user address of this page
1253  * @page: source kernel page
1254  *
1255  * This allows drivers to insert individual pages they've allocated
1256  * into a user vma.
1257  *
1258  * The page has to be a nice clean _individual_ kernel allocation.
1259  * If you allocate a compound page, you need to have marked it as
1260  * such (__GFP_COMP), or manually just split the page up yourself
1261  * (see split_page()).
1262  *
1263  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1264  * took an arbitrary page protection parameter. This doesn't allow
1265  * that. Your vma protection will have to be set up correctly, which
1266  * means that if you want a shared writable mapping, you'd better
1267  * ask for a shared writable mapping!
1268  *
1269  * The page does not need to be reserved.
1270  */
1271 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1272 {
1273         if (addr < vma->vm_start || addr >= vma->vm_end)
1274                 return -EFAULT;
1275         if (!page_count(page))
1276                 return -EINVAL;
1277         vma->vm_flags |= VM_INSERTPAGE;
1278         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1279 }
1280 EXPORT_SYMBOL(vm_insert_page);
1281
1282 /**
1283  * vm_insert_pfn - insert single pfn into user vma
1284  * @vma: user vma to map to
1285  * @addr: target user address of this page
1286  * @pfn: source kernel pfn
1287  *
1288  * Similar to vm_inert_page, this allows drivers to insert individual pages
1289  * they've allocated into a user vma. Same comments apply.
1290  *
1291  * This function should only be called from a vm_ops->fault handler, and
1292  * in that case the handler should return NULL.
1293  */
1294 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1295                 unsigned long pfn)
1296 {
1297         struct mm_struct *mm = vma->vm_mm;
1298         int retval;
1299         pte_t *pte, entry;
1300         spinlock_t *ptl;
1301
1302         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
1303         BUG_ON(is_cow_mapping(vma->vm_flags));
1304
1305         retval = -ENOMEM;
1306         pte = get_locked_pte(mm, addr, &ptl);
1307         if (!pte)
1308                 goto out;
1309         retval = -EBUSY;
1310         if (!pte_none(*pte))
1311                 goto out_unlock;
1312
1313         /* Ok, finally just insert the thing.. */
1314         entry = pfn_pte(pfn, vma->vm_page_prot);
1315         set_pte_at(mm, addr, pte, entry);
1316         update_mmu_cache(vma, addr, entry);
1317
1318         retval = 0;
1319 out_unlock:
1320         pte_unmap_unlock(pte, ptl);
1321
1322 out:
1323         return retval;
1324 }
1325 EXPORT_SYMBOL(vm_insert_pfn);
1326
1327 /*
1328  * maps a range of physical memory into the requested pages. the old
1329  * mappings are removed. any references to nonexistent pages results
1330  * in null mappings (currently treated as "copy-on-access")
1331  */
1332 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1333                         unsigned long addr, unsigned long end,
1334                         unsigned long pfn, pgprot_t prot)
1335 {
1336         pte_t *pte;
1337         spinlock_t *ptl;
1338
1339         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1340         if (!pte)
1341                 return -ENOMEM;
1342         arch_enter_lazy_mmu_mode();
1343         do {
1344                 BUG_ON(!pte_none(*pte));
1345                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1346                 pfn++;
1347         } while (pte++, addr += PAGE_SIZE, addr != end);
1348         arch_leave_lazy_mmu_mode();
1349         pte_unmap_unlock(pte - 1, ptl);
1350         return 0;
1351 }
1352
1353 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1354                         unsigned long addr, unsigned long end,
1355                         unsigned long pfn, pgprot_t prot)
1356 {
1357         pmd_t *pmd;
1358         unsigned long next;
1359
1360         pfn -= addr >> PAGE_SHIFT;
1361         pmd = pmd_alloc(mm, pud, addr);
1362         if (!pmd)
1363                 return -ENOMEM;
1364         do {
1365                 next = pmd_addr_end(addr, end);
1366                 if (remap_pte_range(mm, pmd, addr, next,
1367                                 pfn + (addr >> PAGE_SHIFT), prot))
1368                         return -ENOMEM;
1369         } while (pmd++, addr = next, addr != end);
1370         return 0;
1371 }
1372
1373 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1374                         unsigned long addr, unsigned long end,
1375                         unsigned long pfn, pgprot_t prot)
1376 {
1377         pud_t *pud;
1378         unsigned long next;
1379
1380         pfn -= addr >> PAGE_SHIFT;
1381         pud = pud_alloc(mm, pgd, addr);
1382         if (!pud)
1383                 return -ENOMEM;
1384         do {
1385                 next = pud_addr_end(addr, end);
1386                 if (remap_pmd_range(mm, pud, addr, next,
1387                                 pfn + (addr >> PAGE_SHIFT), prot))
1388                         return -ENOMEM;
1389         } while (pud++, addr = next, addr != end);
1390         return 0;
1391 }
1392
1393 /**
1394  * remap_pfn_range - remap kernel memory to userspace
1395  * @vma: user vma to map to
1396  * @addr: target user address to start at
1397  * @pfn: physical address of kernel memory
1398  * @size: size of map area
1399  * @prot: page protection flags for this mapping
1400  *
1401  *  Note: this is only safe if the mm semaphore is held when called.
1402  */
1403 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1404                     unsigned long pfn, unsigned long size, pgprot_t prot)
1405 {
1406         pgd_t *pgd;
1407         unsigned long next;
1408         unsigned long end = addr + PAGE_ALIGN(size);
1409         struct mm_struct *mm = vma->vm_mm;
1410         int err;
1411
1412         /*
1413          * Physically remapped pages are special. Tell the
1414          * rest of the world about it:
1415          *   VM_IO tells people not to look at these pages
1416          *      (accesses can have side effects).
1417          *   VM_RESERVED is specified all over the place, because
1418          *      in 2.4 it kept swapout's vma scan off this vma; but
1419          *      in 2.6 the LRU scan won't even find its pages, so this
1420          *      flag means no more than count its pages in reserved_vm,
1421          *      and omit it from core dump, even when VM_IO turned off.
1422          *   VM_PFNMAP tells the core MM that the base pages are just
1423          *      raw PFN mappings, and do not have a "struct page" associated
1424          *      with them.
1425          *
1426          * There's a horrible special case to handle copy-on-write
1427          * behaviour that some programs depend on. We mark the "original"
1428          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1429          */
1430         if (is_cow_mapping(vma->vm_flags)) {
1431                 if (addr != vma->vm_start || end != vma->vm_end)
1432                         return -EINVAL;
1433                 vma->vm_pgoff = pfn;
1434         }
1435
1436         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1437
1438         BUG_ON(addr >= end);
1439         pfn -= addr >> PAGE_SHIFT;
1440         pgd = pgd_offset(mm, addr);
1441         flush_cache_range(vma, addr, end);
1442         do {
1443                 next = pgd_addr_end(addr, end);
1444                 err = remap_pud_range(mm, pgd, addr, next,
1445                                 pfn + (addr >> PAGE_SHIFT), prot);
1446                 if (err)
1447                         break;
1448         } while (pgd++, addr = next, addr != end);
1449         return err;
1450 }
1451 EXPORT_SYMBOL(remap_pfn_range);
1452
1453 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1454                                      unsigned long addr, unsigned long end,
1455                                      pte_fn_t fn, void *data)
1456 {
1457         pte_t *pte;
1458         int err;
1459         struct page *pmd_page;
1460         spinlock_t *uninitialized_var(ptl);
1461
1462         pte = (mm == &init_mm) ?
1463                 pte_alloc_kernel(pmd, addr) :
1464                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1465         if (!pte)
1466                 return -ENOMEM;
1467
1468         BUG_ON(pmd_huge(*pmd));
1469
1470         pmd_page = pmd_page(*pmd);
1471
1472         do {
1473                 err = fn(pte, pmd_page, addr, data);
1474                 if (err)
1475                         break;
1476         } while (pte++, addr += PAGE_SIZE, addr != end);
1477
1478         if (mm != &init_mm)
1479                 pte_unmap_unlock(pte-1, ptl);
1480         return err;
1481 }
1482
1483 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1484                                      unsigned long addr, unsigned long end,
1485                                      pte_fn_t fn, void *data)
1486 {
1487         pmd_t *pmd;
1488         unsigned long next;
1489         int err;
1490
1491         pmd = pmd_alloc(mm, pud, addr);
1492         if (!pmd)
1493                 return -ENOMEM;
1494         do {
1495                 next = pmd_addr_end(addr, end);
1496                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1497                 if (err)
1498                         break;
1499         } while (pmd++, addr = next, addr != end);
1500         return err;
1501 }
1502
1503 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1504                                      unsigned long addr, unsigned long end,
1505                                      pte_fn_t fn, void *data)
1506 {
1507         pud_t *pud;
1508         unsigned long next;
1509         int err;
1510
1511         pud = pud_alloc(mm, pgd, addr);
1512         if (!pud)
1513                 return -ENOMEM;
1514         do {
1515                 next = pud_addr_end(addr, end);
1516                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1517                 if (err)
1518                         break;
1519         } while (pud++, addr = next, addr != end);
1520         return err;
1521 }
1522
1523 /*
1524  * Scan a region of virtual memory, filling in page tables as necessary
1525  * and calling a provided function on each leaf page table.
1526  */
1527 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1528                         unsigned long size, pte_fn_t fn, void *data)
1529 {
1530         pgd_t *pgd;
1531         unsigned long next;
1532         unsigned long end = addr + size;
1533         int err;
1534
1535         BUG_ON(addr >= end);
1536         pgd = pgd_offset(mm, addr);
1537         do {
1538                 next = pgd_addr_end(addr, end);
1539                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1540                 if (err)
1541                         break;
1542         } while (pgd++, addr = next, addr != end);
1543         return err;
1544 }
1545 EXPORT_SYMBOL_GPL(apply_to_page_range);
1546
1547 /*
1548  * handle_pte_fault chooses page fault handler according to an entry
1549  * which was read non-atomically.  Before making any commitment, on
1550  * those architectures or configurations (e.g. i386 with PAE) which
1551  * might give a mix of unmatched parts, do_swap_page and do_file_page
1552  * must check under lock before unmapping the pte and proceeding
1553  * (but do_wp_page is only called after already making such a check;
1554  * and do_anonymous_page and do_no_page can safely check later on).
1555  */
1556 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1557                                 pte_t *page_table, pte_t orig_pte)
1558 {
1559         int same = 1;
1560 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1561         if (sizeof(pte_t) > sizeof(unsigned long)) {
1562                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1563                 spin_lock(ptl);
1564                 same = pte_same(*page_table, orig_pte);
1565                 spin_unlock(ptl);
1566         }
1567 #endif
1568         pte_unmap(page_table);
1569         return same;
1570 }
1571
1572 /*
1573  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1574  * servicing faults for write access.  In the normal case, do always want
1575  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1576  * that do not have writing enabled, when used by access_process_vm.
1577  */
1578 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1579 {
1580         if (likely(vma->vm_flags & VM_WRITE))
1581                 pte = pte_mkwrite(pte);
1582         return pte;
1583 }
1584
1585 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1586 {
1587         /*
1588          * If the source page was a PFN mapping, we don't have
1589          * a "struct page" for it. We do a best-effort copy by
1590          * just copying from the original user address. If that
1591          * fails, we just zero-fill it. Live with it.
1592          */
1593         if (unlikely(!src)) {
1594                 void *kaddr = kmap_atomic(dst, KM_USER0);
1595                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1596
1597                 /*
1598                  * This really shouldn't fail, because the page is there
1599                  * in the page tables. But it might just be unreadable,
1600                  * in which case we just give up and fill the result with
1601                  * zeroes.
1602                  */
1603                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1604                         memset(kaddr, 0, PAGE_SIZE);
1605                 kunmap_atomic(kaddr, KM_USER0);
1606                 flush_dcache_page(dst);
1607                 return;
1608
1609         }
1610         copy_user_highpage(dst, src, va, vma);
1611 }
1612
1613 /*
1614  * This routine handles present pages, when users try to write
1615  * to a shared page. It is done by copying the page to a new address
1616  * and decrementing the shared-page counter for the old page.
1617  *
1618  * Note that this routine assumes that the protection checks have been
1619  * done by the caller (the low-level page fault routine in most cases).
1620  * Thus we can safely just mark it writable once we've done any necessary
1621  * COW.
1622  *
1623  * We also mark the page dirty at this point even though the page will
1624  * change only once the write actually happens. This avoids a few races,
1625  * and potentially makes it more efficient.
1626  *
1627  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1628  * but allow concurrent faults), with pte both mapped and locked.
1629  * We return with mmap_sem still held, but pte unmapped and unlocked.
1630  */
1631 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1632                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1633                 spinlock_t *ptl, pte_t orig_pte)
1634 {
1635         struct page *old_page, *new_page;
1636         pte_t entry;
1637         int reuse = 0, ret = VM_FAULT_MINOR;
1638         struct page *dirty_page = NULL;
1639
1640         old_page = vm_normal_page(vma, address, orig_pte);
1641         if (!old_page)
1642                 goto gotten;
1643
1644         /*
1645          * Take out anonymous pages first, anonymous shared vmas are
1646          * not dirty accountable.
1647          */
1648         if (PageAnon(old_page)) {
1649                 if (!TestSetPageLocked(old_page)) {
1650                         reuse = can_share_swap_page(old_page);
1651                         unlock_page(old_page);
1652                 }
1653         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1654                                         (VM_WRITE|VM_SHARED))) {
1655                 /*
1656                  * Only catch write-faults on shared writable pages,
1657                  * read-only shared pages can get COWed by
1658                  * get_user_pages(.write=1, .force=1).
1659                  */
1660                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1661                         /*
1662                          * Notify the address space that the page is about to
1663                          * become writable so that it can prohibit this or wait
1664                          * for the page to get into an appropriate state.
1665                          *
1666                          * We do this without the lock held, so that it can
1667                          * sleep if it needs to.
1668                          */
1669                         page_cache_get(old_page);
1670                         pte_unmap_unlock(page_table, ptl);
1671
1672                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1673                                 goto unwritable_page;
1674
1675                         /*
1676                          * Since we dropped the lock we need to revalidate
1677                          * the PTE as someone else may have changed it.  If
1678                          * they did, we just return, as we can count on the
1679                          * MMU to tell us if they didn't also make it writable.
1680                          */
1681                         page_table = pte_offset_map_lock(mm, pmd, address,
1682                                                          &ptl);
1683                         page_cache_release(old_page);
1684                         if (!pte_same(*page_table, orig_pte))
1685                                 goto unlock;
1686                 }
1687                 dirty_page = old_page;
1688                 get_page(dirty_page);
1689                 reuse = 1;
1690         }
1691
1692         if (reuse) {
1693                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1694                 entry = pte_mkyoung(orig_pte);
1695                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1696                 if (ptep_set_access_flags(vma, address, page_table, entry,1)) {
1697                         update_mmu_cache(vma, address, entry);
1698                         lazy_mmu_prot_update(entry);
1699                 }
1700                 ret |= VM_FAULT_WRITE;
1701                 goto unlock;
1702         }
1703
1704         /*
1705          * Ok, we need to copy. Oh, well..
1706          */
1707         page_cache_get(old_page);
1708 gotten:
1709         pte_unmap_unlock(page_table, ptl);
1710
1711         if (unlikely(anon_vma_prepare(vma)))
1712                 goto oom;
1713         if (old_page == ZERO_PAGE(address)) {
1714                 new_page = alloc_zeroed_user_highpage(vma, address);
1715                 if (!new_page)
1716                         goto oom;
1717         } else {
1718                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1719                 if (!new_page)
1720                         goto oom;
1721                 cow_user_page(new_page, old_page, address, vma);
1722         }
1723
1724         /*
1725          * Re-check the pte - we dropped the lock
1726          */
1727         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1728         if (likely(pte_same(*page_table, orig_pte))) {
1729                 if (old_page) {
1730                         page_remove_rmap(old_page, vma);
1731                         if (!PageAnon(old_page)) {
1732                                 dec_mm_counter(mm, file_rss);
1733                                 inc_mm_counter(mm, anon_rss);
1734                         }
1735                 } else
1736                         inc_mm_counter(mm, anon_rss);
1737                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1738                 entry = mk_pte(new_page, vma->vm_page_prot);
1739                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1740                 lazy_mmu_prot_update(entry);
1741                 /*
1742                  * Clear the pte entry and flush it first, before updating the
1743                  * pte with the new entry. This will avoid a race condition
1744                  * seen in the presence of one thread doing SMC and another
1745                  * thread doing COW.
1746                  */
1747                 ptep_clear_flush(vma, address, page_table);
1748                 set_pte_at(mm, address, page_table, entry);
1749                 update_mmu_cache(vma, address, entry);
1750                 lru_cache_add_active(new_page);
1751                 page_add_new_anon_rmap(new_page, vma, address);
1752
1753                 /* Free the old page.. */
1754                 new_page = old_page;
1755                 ret |= VM_FAULT_WRITE;
1756         }
1757         if (new_page)
1758                 page_cache_release(new_page);
1759         if (old_page)
1760                 page_cache_release(old_page);
1761 unlock:
1762         pte_unmap_unlock(page_table, ptl);
1763         if (dirty_page) {
1764                 set_page_dirty_balance(dirty_page);
1765                 put_page(dirty_page);
1766         }
1767         return ret;
1768 oom:
1769         if (old_page)
1770                 page_cache_release(old_page);
1771         return VM_FAULT_OOM;
1772
1773 unwritable_page:
1774         page_cache_release(old_page);
1775         return VM_FAULT_SIGBUS;
1776 }
1777
1778 /*
1779  * Helper functions for unmap_mapping_range().
1780  *
1781  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1782  *
1783  * We have to restart searching the prio_tree whenever we drop the lock,
1784  * since the iterator is only valid while the lock is held, and anyway
1785  * a later vma might be split and reinserted earlier while lock dropped.
1786  *
1787  * The list of nonlinear vmas could be handled more efficiently, using
1788  * a placeholder, but handle it in the same way until a need is shown.
1789  * It is important to search the prio_tree before nonlinear list: a vma
1790  * may become nonlinear and be shifted from prio_tree to nonlinear list
1791  * while the lock is dropped; but never shifted from list to prio_tree.
1792  *
1793  * In order to make forward progress despite restarting the search,
1794  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1795  * quickly skip it next time around.  Since the prio_tree search only
1796  * shows us those vmas affected by unmapping the range in question, we
1797  * can't efficiently keep all vmas in step with mapping->truncate_count:
1798  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1799  * mapping->truncate_count and vma->vm_truncate_count are protected by
1800  * i_mmap_lock.
1801  *
1802  * In order to make forward progress despite repeatedly restarting some
1803  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1804  * and restart from that address when we reach that vma again.  It might
1805  * have been split or merged, shrunk or extended, but never shifted: so
1806  * restart_addr remains valid so long as it remains in the vma's range.
1807  * unmap_mapping_range forces truncate_count to leap over page-aligned
1808  * values so we can save vma's restart_addr in its truncate_count field.
1809  */
1810 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1811
1812 static void reset_vma_truncate_counts(struct address_space *mapping)
1813 {
1814         struct vm_area_struct *vma;
1815         struct prio_tree_iter iter;
1816
1817         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1818                 vma->vm_truncate_count = 0;
1819         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1820                 vma->vm_truncate_count = 0;
1821 }
1822
1823 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1824                 unsigned long start_addr, unsigned long end_addr,
1825                 struct zap_details *details)
1826 {
1827         unsigned long restart_addr;
1828         int need_break;
1829
1830 again:
1831         restart_addr = vma->vm_truncate_count;
1832         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1833                 start_addr = restart_addr;
1834                 if (start_addr >= end_addr) {
1835                         /* Top of vma has been split off since last time */
1836                         vma->vm_truncate_count = details->truncate_count;
1837                         return 0;
1838                 }
1839         }
1840
1841         restart_addr = zap_page_range(vma, start_addr,
1842                                         end_addr - start_addr, details);
1843         need_break = need_resched() ||
1844                         need_lockbreak(details->i_mmap_lock);
1845
1846         if (restart_addr >= end_addr) {
1847                 /* We have now completed this vma: mark it so */
1848                 vma->vm_truncate_count = details->truncate_count;
1849                 if (!need_break)
1850                         return 0;
1851         } else {
1852                 /* Note restart_addr in vma's truncate_count field */
1853                 vma->vm_truncate_count = restart_addr;
1854                 if (!need_break)
1855                         goto again;
1856         }
1857
1858         spin_unlock(details->i_mmap_lock);
1859         cond_resched();
1860         spin_lock(details->i_mmap_lock);
1861         return -EINTR;
1862 }
1863
1864 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1865                                             struct zap_details *details)
1866 {
1867         struct vm_area_struct *vma;
1868         struct prio_tree_iter iter;
1869         pgoff_t vba, vea, zba, zea;
1870
1871 restart:
1872         vma_prio_tree_foreach(vma, &iter, root,
1873                         details->first_index, details->last_index) {
1874                 /* Skip quickly over those we have already dealt with */
1875                 if (vma->vm_truncate_count == details->truncate_count)
1876                         continue;
1877
1878                 vba = vma->vm_pgoff;
1879                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1880                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1881                 zba = details->first_index;
1882                 if (zba < vba)
1883                         zba = vba;
1884                 zea = details->last_index;
1885                 if (zea > vea)
1886                         zea = vea;
1887
1888                 if (unmap_mapping_range_vma(vma,
1889                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1890                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1891                                 details) < 0)
1892                         goto restart;
1893         }
1894 }
1895
1896 static inline void unmap_mapping_range_list(struct list_head *head,
1897                                             struct zap_details *details)
1898 {
1899         struct vm_area_struct *vma;
1900
1901         /*
1902          * In nonlinear VMAs there is no correspondence between virtual address
1903          * offset and file offset.  So we must perform an exhaustive search
1904          * across *all* the pages in each nonlinear VMA, not just the pages
1905          * whose virtual address lies outside the file truncation point.
1906          */
1907 restart:
1908         list_for_each_entry(vma, head, shared.vm_set.list) {
1909                 /* Skip quickly over those we have already dealt with */
1910                 if (vma->vm_truncate_count == details->truncate_count)
1911                         continue;
1912                 details->nonlinear_vma = vma;
1913                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1914                                         vma->vm_end, details) < 0)
1915                         goto restart;
1916         }
1917 }
1918
1919 /**
1920  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
1921  * @mapping: the address space containing mmaps to be unmapped.
1922  * @holebegin: byte in first page to unmap, relative to the start of
1923  * the underlying file.  This will be rounded down to a PAGE_SIZE
1924  * boundary.  Note that this is different from vmtruncate(), which
1925  * must keep the partial page.  In contrast, we must get rid of
1926  * partial pages.
1927  * @holelen: size of prospective hole in bytes.  This will be rounded
1928  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1929  * end of the file.
1930  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1931  * but 0 when invalidating pagecache, don't throw away private data.
1932  */
1933 void unmap_mapping_range(struct address_space *mapping,
1934                 loff_t const holebegin, loff_t const holelen, int even_cows)
1935 {
1936         struct zap_details details;
1937         pgoff_t hba = holebegin >> PAGE_SHIFT;
1938         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1939
1940         /* Check for overflow. */
1941         if (sizeof(holelen) > sizeof(hlen)) {
1942                 long long holeend =
1943                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1944                 if (holeend & ~(long long)ULONG_MAX)
1945                         hlen = ULONG_MAX - hba + 1;
1946         }
1947
1948         details.check_mapping = even_cows? NULL: mapping;
1949         details.nonlinear_vma = NULL;
1950         details.first_index = hba;
1951         details.last_index = hba + hlen - 1;
1952         if (details.last_index < details.first_index)
1953                 details.last_index = ULONG_MAX;
1954         details.i_mmap_lock = &mapping->i_mmap_lock;
1955
1956         spin_lock(&mapping->i_mmap_lock);
1957
1958         /* serialize i_size write against truncate_count write */
1959         smp_wmb();
1960         /* Protect against page faults, and endless unmapping loops */
1961         mapping->truncate_count++;
1962         /*
1963          * For archs where spin_lock has inclusive semantics like ia64
1964          * this smp_mb() will prevent to read pagetable contents
1965          * before the truncate_count increment is visible to
1966          * other cpus.
1967          */
1968         smp_mb();
1969         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1970                 if (mapping->truncate_count == 0)
1971                         reset_vma_truncate_counts(mapping);
1972                 mapping->truncate_count++;
1973         }
1974         details.truncate_count = mapping->truncate_count;
1975
1976         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1977                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1978         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1979                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1980         spin_unlock(&mapping->i_mmap_lock);
1981 }
1982 EXPORT_SYMBOL(unmap_mapping_range);
1983
1984 /**
1985  * vmtruncate - unmap mappings "freed" by truncate() syscall
1986  * @inode: inode of the file used
1987  * @offset: file offset to start truncating
1988  *
1989  * NOTE! We have to be ready to update the memory sharing
1990  * between the file and the memory map for a potential last
1991  * incomplete page.  Ugly, but necessary.
1992  */
1993 int vmtruncate(struct inode * inode, loff_t offset)
1994 {
1995         struct address_space *mapping = inode->i_mapping;
1996         unsigned long limit;
1997
1998         if (inode->i_size < offset)
1999                 goto do_expand;
2000         /*
2001          * truncation of in-use swapfiles is disallowed - it would cause
2002          * subsequent swapout to scribble on the now-freed blocks.
2003          */
2004         if (IS_SWAPFILE(inode))
2005                 goto out_busy;
2006         i_size_write(inode, offset);
2007         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2008         truncate_inode_pages(mapping, offset);
2009         goto out_truncate;
2010
2011 do_expand:
2012         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2013         if (limit != RLIM_INFINITY && offset > limit)
2014                 goto out_sig;
2015         if (offset > inode->i_sb->s_maxbytes)
2016                 goto out_big;
2017         i_size_write(inode, offset);
2018
2019 out_truncate:
2020         if (inode->i_op && inode->i_op->truncate)
2021                 inode->i_op->truncate(inode);
2022         return 0;
2023 out_sig:
2024         send_sig(SIGXFSZ, current, 0);
2025 out_big:
2026         return -EFBIG;
2027 out_busy:
2028         return -ETXTBSY;
2029 }
2030 EXPORT_SYMBOL(vmtruncate);
2031
2032 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2033 {
2034         struct address_space *mapping = inode->i_mapping;
2035
2036         /*
2037          * If the underlying filesystem is not going to provide
2038          * a way to truncate a range of blocks (punch a hole) -
2039          * we should return failure right now.
2040          */
2041         if (!inode->i_op || !inode->i_op->truncate_range)
2042                 return -ENOSYS;
2043
2044         mutex_lock(&inode->i_mutex);
2045         down_write(&inode->i_alloc_sem);
2046         unmap_mapping_range(mapping, offset, (end - offset), 1);
2047         truncate_inode_pages_range(mapping, offset, end);
2048         inode->i_op->truncate_range(inode, offset, end);
2049         up_write(&inode->i_alloc_sem);
2050         mutex_unlock(&inode->i_mutex);
2051
2052         return 0;
2053 }
2054
2055 /**
2056  * swapin_readahead - swap in pages in hope we need them soon
2057  * @entry: swap entry of this memory
2058  * @addr: address to start
2059  * @vma: user vma this addresses belong to
2060  *
2061  * Primitive swap readahead code. We simply read an aligned block of
2062  * (1 << page_cluster) entries in the swap area. This method is chosen
2063  * because it doesn't cost us any seek time.  We also make sure to queue
2064  * the 'original' request together with the readahead ones...
2065  *
2066  * This has been extended to use the NUMA policies from the mm triggering
2067  * the readahead.
2068  *
2069  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
2070  */
2071 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
2072 {
2073 #ifdef CONFIG_NUMA
2074         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
2075 #endif
2076         int i, num;
2077         struct page *new_page;
2078         unsigned long offset;
2079
2080         /*
2081          * Get the number of handles we should do readahead io to.
2082          */
2083         num = valid_swaphandles(entry, &offset);
2084         for (i = 0; i < num; offset++, i++) {
2085                 /* Ok, do the async read-ahead now */
2086                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
2087                                                            offset), vma, addr);
2088                 if (!new_page)
2089                         break;
2090                 page_cache_release(new_page);
2091 #ifdef CONFIG_NUMA
2092                 /*
2093                  * Find the next applicable VMA for the NUMA policy.
2094                  */
2095                 addr += PAGE_SIZE;
2096                 if (addr == 0)
2097                         vma = NULL;
2098                 if (vma) {
2099                         if (addr >= vma->vm_end) {
2100                                 vma = next_vma;
2101                                 next_vma = vma ? vma->vm_next : NULL;
2102                         }
2103                         if (vma && addr < vma->vm_start)
2104                                 vma = NULL;
2105                 } else {
2106                         if (next_vma && addr >= next_vma->vm_start) {
2107                                 vma = next_vma;
2108                                 next_vma = vma->vm_next;
2109                         }
2110                 }
2111 #endif
2112         }
2113         lru_add_drain();        /* Push any new pages onto the LRU now */
2114 }
2115
2116 /*
2117  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2118  * but allow concurrent faults), and pte mapped but not yet locked.
2119  * We return with mmap_sem still held, but pte unmapped and unlocked.
2120  */
2121 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2122                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2123                 int write_access, pte_t orig_pte)
2124 {
2125         spinlock_t *ptl;
2126         struct page *page;
2127         swp_entry_t entry;
2128         pte_t pte;
2129         int ret = VM_FAULT_MINOR;
2130
2131         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2132                 goto out;
2133
2134         entry = pte_to_swp_entry(orig_pte);
2135         if (is_migration_entry(entry)) {
2136                 migration_entry_wait(mm, pmd, address);
2137                 goto out;
2138         }
2139         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2140         page = lookup_swap_cache(entry);
2141         if (!page) {
2142                 grab_swap_token(); /* Contend for token _before_ read-in */
2143                 swapin_readahead(entry, address, vma);
2144                 page = read_swap_cache_async(entry, vma, address);
2145                 if (!page) {
2146                         /*
2147                          * Back out if somebody else faulted in this pte
2148                          * while we released the pte lock.
2149                          */
2150                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2151                         if (likely(pte_same(*page_table, orig_pte)))
2152                                 ret = VM_FAULT_OOM;
2153                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2154                         goto unlock;
2155                 }
2156
2157                 /* Had to read the page from swap area: Major fault */
2158                 ret = VM_FAULT_MAJOR;
2159                 count_vm_event(PGMAJFAULT);
2160         }
2161
2162         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2163         mark_page_accessed(page);
2164         lock_page(page);
2165
2166         /*
2167          * Back out if somebody else already faulted in this pte.
2168          */
2169         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2170         if (unlikely(!pte_same(*page_table, orig_pte)))
2171                 goto out_nomap;
2172
2173         if (unlikely(!PageUptodate(page))) {
2174                 ret = VM_FAULT_SIGBUS;
2175                 goto out_nomap;
2176         }
2177
2178         /* The page isn't present yet, go ahead with the fault. */
2179
2180         inc_mm_counter(mm, anon_rss);
2181         pte = mk_pte(page, vma->vm_page_prot);
2182         if (write_access && can_share_swap_page(page)) {
2183                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2184                 write_access = 0;
2185         }
2186
2187         flush_icache_page(vma, page);
2188         set_pte_at(mm, address, page_table, pte);
2189         page_add_anon_rmap(page, vma, address);
2190
2191         swap_free(entry);
2192         if (vm_swap_full())
2193                 remove_exclusive_swap_page(page);
2194         unlock_page(page);
2195
2196         if (write_access) {
2197                 if (do_wp_page(mm, vma, address,
2198                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
2199                         ret = VM_FAULT_OOM;
2200                 goto out;
2201         }
2202
2203         /* No need to invalidate - it was non-present before */
2204         update_mmu_cache(vma, address, pte);
2205         lazy_mmu_prot_update(pte);
2206 unlock:
2207         pte_unmap_unlock(page_table, ptl);
2208 out:
2209         return ret;
2210 out_nomap:
2211         pte_unmap_unlock(page_table, ptl);
2212         unlock_page(page);
2213         page_cache_release(page);
2214         return ret;
2215 }
2216
2217 /*
2218  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2219  * but allow concurrent faults), and pte mapped but not yet locked.
2220  * We return with mmap_sem still held, but pte unmapped and unlocked.
2221  */
2222 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2223                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2224                 int write_access)
2225 {
2226         struct page *page;
2227         spinlock_t *ptl;
2228         pte_t entry;
2229
2230         if (write_access) {
2231                 /* Allocate our own private page. */
2232                 pte_unmap(page_table);
2233
2234                 if (unlikely(anon_vma_prepare(vma)))
2235                         goto oom;
2236                 page = alloc_zeroed_user_highpage(vma, address);
2237                 if (!page)
2238                         goto oom;
2239
2240                 entry = mk_pte(page, vma->vm_page_prot);
2241                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2242
2243                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2244                 if (!pte_none(*page_table))
2245                         goto release;
2246                 inc_mm_counter(mm, anon_rss);
2247                 lru_cache_add_active(page);
2248                 page_add_new_anon_rmap(page, vma, address);
2249         } else {
2250                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2251                 page = ZERO_PAGE(address);
2252                 page_cache_get(page);
2253                 entry = mk_pte(page, vma->vm_page_prot);
2254
2255                 ptl = pte_lockptr(mm, pmd);
2256                 spin_lock(ptl);
2257                 if (!pte_none(*page_table))
2258                         goto release;
2259                 inc_mm_counter(mm, file_rss);
2260                 page_add_file_rmap(page);
2261         }
2262
2263         set_pte_at(mm, address, page_table, entry);
2264
2265         /* No need to invalidate - it was non-present before */
2266         update_mmu_cache(vma, address, entry);
2267         lazy_mmu_prot_update(entry);
2268 unlock:
2269         pte_unmap_unlock(page_table, ptl);
2270         return VM_FAULT_MINOR;
2271 release:
2272         page_cache_release(page);
2273         goto unlock;
2274 oom:
2275         return VM_FAULT_OOM;
2276 }
2277
2278 /*
2279  * do_no_page() tries to create a new page mapping. It aggressively
2280  * tries to share with existing pages, but makes a separate copy if
2281  * the "write_access" parameter is true in order to avoid the next
2282  * page fault.
2283  *
2284  * As this is called only for pages that do not currently exist, we
2285  * do not need to flush old virtual caches or the TLB.
2286  *
2287  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2288  * but allow concurrent faults), and pte mapped but not yet locked.
2289  * We return with mmap_sem still held, but pte unmapped and unlocked.
2290  */
2291 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2292                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2293                 int write_access)
2294 {
2295         spinlock_t *ptl;
2296         struct page *new_page;
2297         struct address_space *mapping = NULL;
2298         pte_t entry;
2299         unsigned int sequence = 0;
2300         int ret = VM_FAULT_MINOR;
2301         int anon = 0;
2302         struct page *dirty_page = NULL;
2303
2304         pte_unmap(page_table);
2305         BUG_ON(vma->vm_flags & VM_PFNMAP);
2306
2307         if (vma->vm_file) {
2308                 mapping = vma->vm_file->f_mapping;
2309                 sequence = mapping->truncate_count;
2310                 smp_rmb(); /* serializes i_size against truncate_count */
2311         }
2312 retry:
2313         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2314         /*
2315          * No smp_rmb is needed here as long as there's a full
2316          * spin_lock/unlock sequence inside the ->nopage callback
2317          * (for the pagecache lookup) that acts as an implicit
2318          * smp_mb() and prevents the i_size read to happen
2319          * after the next truncate_count read.
2320          */
2321
2322         /* no page was available -- either SIGBUS, OOM or REFAULT */
2323         if (unlikely(new_page == NOPAGE_SIGBUS))
2324                 return VM_FAULT_SIGBUS;
2325         else if (unlikely(new_page == NOPAGE_OOM))
2326                 return VM_FAULT_OOM;
2327         else if (unlikely(new_page == NOPAGE_REFAULT))
2328                 return VM_FAULT_MINOR;
2329
2330         /*
2331          * Should we do an early C-O-W break?
2332          */
2333         if (write_access) {
2334                 if (!(vma->vm_flags & VM_SHARED)) {
2335                         struct page *page;
2336
2337                         if (unlikely(anon_vma_prepare(vma)))
2338                                 goto oom;
2339                         page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2340                         if (!page)
2341                                 goto oom;
2342                         copy_user_highpage(page, new_page, address, vma);
2343                         page_cache_release(new_page);
2344                         new_page = page;
2345                         anon = 1;
2346
2347                 } else {
2348                         /* if the page will be shareable, see if the backing
2349                          * address space wants to know that the page is about
2350                          * to become writable */
2351                         if (vma->vm_ops->page_mkwrite &&
2352                             vma->vm_ops->page_mkwrite(vma, new_page) < 0
2353                             ) {
2354                                 page_cache_release(new_page);
2355                                 return VM_FAULT_SIGBUS;
2356                         }
2357                 }
2358         }
2359
2360         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2361         /*
2362          * For a file-backed vma, someone could have truncated or otherwise
2363          * invalidated this page.  If unmap_mapping_range got called,
2364          * retry getting the page.
2365          */
2366         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2367                 pte_unmap_unlock(page_table, ptl);
2368                 page_cache_release(new_page);
2369                 cond_resched();
2370                 sequence = mapping->truncate_count;
2371                 smp_rmb();
2372                 goto retry;
2373         }
2374
2375         /*
2376          * This silly early PAGE_DIRTY setting removes a race
2377          * due to the bad i386 page protection. But it's valid
2378          * for other architectures too.
2379          *
2380          * Note that if write_access is true, we either now have
2381          * an exclusive copy of the page, or this is a shared mapping,
2382          * so we can make it writable and dirty to avoid having to
2383          * handle that later.
2384          */
2385         /* Only go through if we didn't race with anybody else... */
2386         if (pte_none(*page_table)) {
2387                 flush_icache_page(vma, new_page);
2388                 entry = mk_pte(new_page, vma->vm_page_prot);
2389                 if (write_access)
2390                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2391                 set_pte_at(mm, address, page_table, entry);
2392                 if (anon) {
2393                         inc_mm_counter(mm, anon_rss);
2394                         lru_cache_add_active(new_page);
2395                         page_add_new_anon_rmap(new_page, vma, address);
2396                 } else {
2397                         inc_mm_counter(mm, file_rss);
2398                         page_add_file_rmap(new_page);
2399                         if (write_access) {
2400                                 dirty_page = new_page;
2401                                 get_page(dirty_page);
2402                         }
2403                 }
2404         } else {
2405                 /* One of our sibling threads was faster, back out. */
2406                 page_cache_release(new_page);
2407                 goto unlock;
2408         }
2409
2410         /* no need to invalidate: a not-present page shouldn't be cached */
2411         update_mmu_cache(vma, address, entry);
2412         lazy_mmu_prot_update(entry);
2413 unlock:
2414         pte_unmap_unlock(page_table, ptl);
2415         if (dirty_page) {
2416                 set_page_dirty_balance(dirty_page);
2417                 put_page(dirty_page);
2418         }
2419         return ret;
2420 oom:
2421         page_cache_release(new_page);
2422         return VM_FAULT_OOM;
2423 }
2424
2425 /*
2426  * do_no_pfn() tries to create a new page mapping for a page without
2427  * a struct_page backing it
2428  *
2429  * As this is called only for pages that do not currently exist, we
2430  * do not need to flush old virtual caches or the TLB.
2431  *
2432  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2433  * but allow concurrent faults), and pte mapped but not yet locked.
2434  * We return with mmap_sem still held, but pte unmapped and unlocked.
2435  *
2436  * It is expected that the ->nopfn handler always returns the same pfn
2437  * for a given virtual mapping.
2438  *
2439  * Mark this `noinline' to prevent it from bloating the main pagefault code.
2440  */
2441 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2442                      unsigned long address, pte_t *page_table, pmd_t *pmd,
2443                      int write_access)
2444 {
2445         spinlock_t *ptl;
2446         pte_t entry;
2447         unsigned long pfn;
2448         int ret = VM_FAULT_MINOR;
2449
2450         pte_unmap(page_table);
2451         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
2452         BUG_ON(is_cow_mapping(vma->vm_flags));
2453
2454         pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2455         if (unlikely(pfn == NOPFN_OOM))
2456                 return VM_FAULT_OOM;
2457         else if (unlikely(pfn == NOPFN_SIGBUS))
2458                 return VM_FAULT_SIGBUS;
2459         else if (unlikely(pfn == NOPFN_REFAULT))
2460                 return VM_FAULT_MINOR;
2461
2462         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2463
2464         /* Only go through if we didn't race with anybody else... */
2465         if (pte_none(*page_table)) {
2466                 entry = pfn_pte(pfn, vma->vm_page_prot);
2467                 if (write_access)
2468                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2469                 set_pte_at(mm, address, page_table, entry);
2470         }
2471         pte_unmap_unlock(page_table, ptl);
2472         return ret;
2473 }
2474
2475 /*
2476  * Fault of a previously existing named mapping. Repopulate the pte
2477  * from the encoded file_pte if possible. This enables swappable
2478  * nonlinear vmas.
2479  *
2480  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2481  * but allow concurrent faults), and pte mapped but not yet locked.
2482  * We return with mmap_sem still held, but pte unmapped and unlocked.
2483  */
2484 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2485                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2486                 int write_access, pte_t orig_pte)
2487 {
2488         pgoff_t pgoff;
2489         int err;
2490
2491         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2492                 return VM_FAULT_MINOR;
2493
2494         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2495                 /*
2496                  * Page table corrupted: show pte and kill process.
2497                  */
2498                 print_bad_pte(vma, orig_pte, address);
2499                 return VM_FAULT_OOM;
2500         }
2501         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2502
2503         pgoff = pte_to_pgoff(orig_pte);
2504         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2505                                         vma->vm_page_prot, pgoff, 0);
2506         if (err == -ENOMEM)
2507                 return VM_FAULT_OOM;
2508         if (err)
2509                 return VM_FAULT_SIGBUS;
2510         return VM_FAULT_MAJOR;
2511 }
2512
2513 /*
2514  * These routines also need to handle stuff like marking pages dirty
2515  * and/or accessed for architectures that don't do it in hardware (most
2516  * RISC architectures).  The early dirtying is also good on the i386.
2517  *
2518  * There is also a hook called "update_mmu_cache()" that architectures
2519  * with external mmu caches can use to update those (ie the Sparc or
2520  * PowerPC hashed page tables that act as extended TLBs).
2521  *
2522  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2523  * but allow concurrent faults), and pte mapped but not yet locked.
2524  * We return with mmap_sem still held, but pte unmapped and unlocked.
2525  */
2526 static inline int handle_pte_fault(struct mm_struct *mm,
2527                 struct vm_area_struct *vma, unsigned long address,
2528                 pte_t *pte, pmd_t *pmd, int write_access)
2529 {
2530         pte_t entry;
2531         spinlock_t *ptl;
2532
2533         entry = *pte;
2534         if (!pte_present(entry)) {
2535                 if (pte_none(entry)) {
2536                         if (vma->vm_ops) {
2537                                 if (vma->vm_ops->nopage)
2538                                         return do_no_page(mm, vma, address,
2539                                                           pte, pmd,
2540                                                           write_access);
2541                                 if (unlikely(vma->vm_ops->nopfn))
2542                                         return do_no_pfn(mm, vma, address, pte,
2543                                                          pmd, write_access);
2544                         }
2545                         return do_anonymous_page(mm, vma, address,
2546                                                  pte, pmd, write_access);
2547                 }
2548                 if (pte_file(entry))
2549                         return do_file_page(mm, vma, address,
2550                                         pte, pmd, write_access, entry);
2551                 return do_swap_page(mm, vma, address,
2552                                         pte, pmd, write_access, entry);
2553         }
2554
2555         ptl = pte_lockptr(mm, pmd);
2556         spin_lock(ptl);
2557         if (unlikely(!pte_same(*pte, entry)))
2558                 goto unlock;
2559         if (write_access) {
2560                 if (!pte_write(entry))
2561                         return do_wp_page(mm, vma, address,
2562                                         pte, pmd, ptl, entry);
2563                 entry = pte_mkdirty(entry);
2564         }
2565         entry = pte_mkyoung(entry);
2566         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2567                 update_mmu_cache(vma, address, entry);
2568                 lazy_mmu_prot_update(entry);
2569         } else {
2570                 /*
2571                  * This is needed only for protection faults but the arch code
2572                  * is not yet telling us if this is a protection fault or not.
2573                  * This still avoids useless tlb flushes for .text page faults
2574                  * with threads.
2575                  */
2576                 if (write_access)
2577                         flush_tlb_page(vma, address);
2578         }
2579 unlock:
2580         pte_unmap_unlock(pte, ptl);
2581         return VM_FAULT_MINOR;
2582 }
2583
2584 /*
2585  * By the time we get here, we already hold the mm semaphore
2586  */
2587 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2588                 unsigned long address, int write_access)
2589 {
2590         pgd_t *pgd;
2591         pud_t *pud;
2592         pmd_t *pmd;
2593         pte_t *pte;
2594
2595         __set_current_state(TASK_RUNNING);
2596
2597         count_vm_event(PGFAULT);
2598
2599         if (unlikely(is_vm_hugetlb_page(vma)))
2600                 return hugetlb_fault(mm, vma, address, write_access);
2601
2602         pgd = pgd_offset(mm, address);
2603         pud = pud_alloc(mm, pgd, address);
2604         if (!pud)
2605                 return VM_FAULT_OOM;
2606         pmd = pmd_alloc(mm, pud, address);
2607         if (!pmd)
2608                 return VM_FAULT_OOM;
2609         pte = pte_alloc_map(mm, pmd, address);
2610         if (!pte)
2611                 return VM_FAULT_OOM;
2612
2613         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2614 }
2615
2616 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2617
2618 #ifndef __PAGETABLE_PUD_FOLDED
2619 /*
2620  * Allocate page upper directory.
2621  * We've already handled the fast-path in-line.
2622  */
2623 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2624 {
2625         pud_t *new = pud_alloc_one(mm, address);
2626         if (!new)
2627                 return -ENOMEM;
2628
2629         spin_lock(&mm->page_table_lock);
2630         if (pgd_present(*pgd))          /* Another has populated it */
2631                 pud_free(new);
2632         else
2633                 pgd_populate(mm, pgd, new);
2634         spin_unlock(&mm->page_table_lock);
2635         return 0;
2636 }
2637 #endif /* __PAGETABLE_PUD_FOLDED */
2638
2639 #ifndef __PAGETABLE_PMD_FOLDED
2640 /*
2641  * Allocate page middle directory.
2642  * We've already handled the fast-path in-line.
2643  */
2644 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2645 {
2646         pmd_t *new = pmd_alloc_one(mm, address);
2647         if (!new)
2648                 return -ENOMEM;
2649
2650         spin_lock(&mm->page_table_lock);
2651 #ifndef __ARCH_HAS_4LEVEL_HACK
2652         if (pud_present(*pud))          /* Another has populated it */
2653                 pmd_free(new);
2654         else
2655                 pud_populate(mm, pud, new);
2656 #else
2657         if (pgd_present(*pud))          /* Another has populated it */
2658                 pmd_free(new);
2659         else
2660                 pgd_populate(mm, pud, new);
2661 #endif /* __ARCH_HAS_4LEVEL_HACK */
2662         spin_unlock(&mm->page_table_lock);
2663         return 0;
2664 }
2665 #endif /* __PAGETABLE_PMD_FOLDED */
2666
2667 int make_pages_present(unsigned long addr, unsigned long end)
2668 {
2669         int ret, len, write;
2670         struct vm_area_struct * vma;
2671
2672         vma = find_vma(current->mm, addr);
2673         if (!vma)
2674                 return -1;
2675         write = (vma->vm_flags & VM_WRITE) != 0;
2676         BUG_ON(addr >= end);
2677         BUG_ON(end > vma->vm_end);
2678         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2679         ret = get_user_pages(current, current->mm, addr,
2680                         len, write, 0, NULL, NULL);
2681         if (ret < 0)
2682                 return ret;
2683         return ret == len ? 0 : -1;
2684 }
2685
2686 /* 
2687  * Map a vmalloc()-space virtual address to the physical page.
2688  */
2689 struct page * vmalloc_to_page(void * vmalloc_addr)
2690 {
2691         unsigned long addr = (unsigned long) vmalloc_addr;
2692         struct page *page = NULL;
2693         pgd_t *pgd = pgd_offset_k(addr);
2694         pud_t *pud;
2695         pmd_t *pmd;
2696         pte_t *ptep, pte;
2697   
2698         if (!pgd_none(*pgd)) {
2699                 pud = pud_offset(pgd, addr);
2700                 if (!pud_none(*pud)) {
2701                         pmd = pmd_offset(pud, addr);
2702                         if (!pmd_none(*pmd)) {
2703                                 ptep = pte_offset_map(pmd, addr);
2704                                 pte = *ptep;
2705                                 if (pte_present(pte))
2706                                         page = pte_page(pte);
2707                                 pte_unmap(ptep);
2708                         }
2709                 }
2710         }
2711         return page;
2712 }
2713
2714 EXPORT_SYMBOL(vmalloc_to_page);
2715
2716 /*
2717  * Map a vmalloc()-space virtual address to the physical page frame number.
2718  */
2719 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2720 {
2721         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2722 }
2723
2724 EXPORT_SYMBOL(vmalloc_to_pfn);
2725
2726 #if !defined(__HAVE_ARCH_GATE_AREA)
2727
2728 #if defined(AT_SYSINFO_EHDR)
2729 static struct vm_area_struct gate_vma;
2730
2731 static int __init gate_vma_init(void)
2732 {
2733         gate_vma.vm_mm = NULL;
2734         gate_vma.vm_start = FIXADDR_USER_START;
2735         gate_vma.vm_end = FIXADDR_USER_END;
2736         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2737         gate_vma.vm_page_prot = __P101;
2738         /*
2739          * Make sure the vDSO gets into every core dump.
2740          * Dumping its contents makes post-mortem fully interpretable later
2741          * without matching up the same kernel and hardware config to see
2742          * what PC values meant.
2743          */
2744         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2745         return 0;
2746 }
2747 __initcall(gate_vma_init);
2748 #endif
2749
2750 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2751 {
2752 #ifdef AT_SYSINFO_EHDR
2753         return &gate_vma;
2754 #else
2755         return NULL;
2756 #endif
2757 }
2758
2759 int in_gate_area_no_task(unsigned long addr)
2760 {
2761 #ifdef AT_SYSINFO_EHDR
2762         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2763                 return 1;
2764 #endif
2765         return 0;
2766 }
2767
2768 #endif  /* __HAVE_ARCH_GATE_AREA */
2769
2770 /*
2771  * Access another process' address space.
2772  * Source/target buffer must be kernel space,
2773  * Do not walk the page table directly, use get_user_pages
2774  */
2775 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2776 {
2777         struct mm_struct *mm;
2778         struct vm_area_struct *vma;
2779         struct page *page;
2780         void *old_buf = buf;
2781
2782         mm = get_task_mm(tsk);
2783         if (!mm)
2784                 return 0;
2785
2786         down_read(&mm->mmap_sem);
2787         /* ignore errors, just check how much was sucessfully transfered */
2788         while (len) {
2789                 int bytes, ret, offset;
2790                 void *maddr;
2791
2792                 ret = get_user_pages(tsk, mm, addr, 1,
2793                                 write, 1, &page, &vma);
2794                 if (ret <= 0)
2795                         break;
2796
2797                 bytes = len;
2798                 offset = addr & (PAGE_SIZE-1);
2799                 if (bytes > PAGE_SIZE-offset)
2800                         bytes = PAGE_SIZE-offset;
2801
2802                 maddr = kmap(page);
2803                 if (write) {
2804                         copy_to_user_page(vma, page, addr,
2805                                           maddr + offset, buf, bytes);
2806                         set_page_dirty_lock(page);
2807                 } else {
2808                         copy_from_user_page(vma, page, addr,
2809                                             buf, maddr + offset, bytes);
2810                 }
2811                 kunmap(page);
2812                 page_cache_release(page);
2813                 len -= bytes;
2814                 buf += bytes;
2815                 addr += bytes;
2816         }
2817         up_read(&mm->mmap_sem);
2818         mmput(mm);
2819
2820         return buf - old_buf;
2821 }