oom: stop allocating user memory if TIF_MEMDIE is set
[linux-2.6] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/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         /* 
987          * Require read or write permissions.
988          * If 'force' is set, we only require the "MAY" flags.
989          */
990         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
991         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
992         i = 0;
993
994         do {
995                 struct vm_area_struct *vma;
996                 unsigned int foll_flags;
997
998                 vma = find_extend_vma(mm, start);
999                 if (!vma && in_gate_area(tsk, start)) {
1000                         unsigned long pg = start & PAGE_MASK;
1001                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1002                         pgd_t *pgd;
1003                         pud_t *pud;
1004                         pmd_t *pmd;
1005                         pte_t *pte;
1006                         if (write) /* user gate pages are read-only */
1007                                 return i ? : -EFAULT;
1008                         if (pg > TASK_SIZE)
1009                                 pgd = pgd_offset_k(pg);
1010                         else
1011                                 pgd = pgd_offset_gate(mm, pg);
1012                         BUG_ON(pgd_none(*pgd));
1013                         pud = pud_offset(pgd, pg);
1014                         BUG_ON(pud_none(*pud));
1015                         pmd = pmd_offset(pud, pg);
1016                         if (pmd_none(*pmd))
1017                                 return i ? : -EFAULT;
1018                         pte = pte_offset_map(pmd, pg);
1019                         if (pte_none(*pte)) {
1020                                 pte_unmap(pte);
1021                                 return i ? : -EFAULT;
1022                         }
1023                         if (pages) {
1024                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1025                                 pages[i] = page;
1026                                 if (page)
1027                                         get_page(page);
1028                         }
1029                         pte_unmap(pte);
1030                         if (vmas)
1031                                 vmas[i] = gate_vma;
1032                         i++;
1033                         start += PAGE_SIZE;
1034                         len--;
1035                         continue;
1036                 }
1037
1038                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1039                                 || !(vm_flags & vma->vm_flags))
1040                         return i ? : -EFAULT;
1041
1042                 if (is_vm_hugetlb_page(vma)) {
1043                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1044                                                 &start, &len, i);
1045                         continue;
1046                 }
1047
1048                 foll_flags = FOLL_TOUCH;
1049                 if (pages)
1050                         foll_flags |= FOLL_GET;
1051                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1052                     (!vma->vm_ops || !vma->vm_ops->nopage))
1053                         foll_flags |= FOLL_ANON;
1054
1055                 do {
1056                         struct page *page;
1057
1058                         /*
1059                          * If tsk is ooming, cut off its access to large memory
1060                          * allocations. It has a pending SIGKILL, but it can't
1061                          * be processed until returning to user space.
1062                          */
1063                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1064                                 return -ENOMEM;
1065
1066                         if (write)
1067                                 foll_flags |= FOLL_WRITE;
1068
1069                         cond_resched();
1070                         while (!(page = follow_page(vma, start, foll_flags))) {
1071                                 int ret;
1072                                 ret = __handle_mm_fault(mm, vma, start,
1073                                                 foll_flags & FOLL_WRITE);
1074                                 /*
1075                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1076                                  * broken COW when necessary, even if maybe_mkwrite
1077                                  * decided not to set pte_write. We can thus safely do
1078                                  * subsequent page lookups as if they were reads.
1079                                  */
1080                                 if (ret & VM_FAULT_WRITE)
1081                                         foll_flags &= ~FOLL_WRITE;
1082                                 
1083                                 switch (ret & ~VM_FAULT_WRITE) {
1084                                 case VM_FAULT_MINOR:
1085                                         tsk->min_flt++;
1086                                         break;
1087                                 case VM_FAULT_MAJOR:
1088                                         tsk->maj_flt++;
1089                                         break;
1090                                 case VM_FAULT_SIGBUS:
1091                                         return i ? i : -EFAULT;
1092                                 case VM_FAULT_OOM:
1093                                         return i ? i : -ENOMEM;
1094                                 default:
1095                                         BUG();
1096                                 }
1097                                 cond_resched();
1098                         }
1099                         if (pages) {
1100                                 pages[i] = page;
1101
1102                                 flush_anon_page(vma, page, start);
1103                                 flush_dcache_page(page);
1104                         }
1105                         if (vmas)
1106                                 vmas[i] = vma;
1107                         i++;
1108                         start += PAGE_SIZE;
1109                         len--;
1110                 } while (len && start < vma->vm_end);
1111         } while (len);
1112         return i;
1113 }
1114 EXPORT_SYMBOL(get_user_pages);
1115
1116 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1117                         unsigned long addr, unsigned long end, pgprot_t prot)
1118 {
1119         pte_t *pte;
1120         spinlock_t *ptl;
1121         int err = 0;
1122
1123         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1124         if (!pte)
1125                 return -EAGAIN;
1126         arch_enter_lazy_mmu_mode();
1127         do {
1128                 struct page *page = ZERO_PAGE(addr);
1129                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1130
1131                 if (unlikely(!pte_none(*pte))) {
1132                         err = -EEXIST;
1133                         pte++;
1134                         break;
1135                 }
1136                 page_cache_get(page);
1137                 page_add_file_rmap(page);
1138                 inc_mm_counter(mm, file_rss);
1139                 set_pte_at(mm, addr, pte, zero_pte);
1140         } while (pte++, addr += PAGE_SIZE, addr != end);
1141         arch_leave_lazy_mmu_mode();
1142         pte_unmap_unlock(pte - 1, ptl);
1143         return err;
1144 }
1145
1146 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1147                         unsigned long addr, unsigned long end, pgprot_t prot)
1148 {
1149         pmd_t *pmd;
1150         unsigned long next;
1151         int err;
1152
1153         pmd = pmd_alloc(mm, pud, addr);
1154         if (!pmd)
1155                 return -EAGAIN;
1156         do {
1157                 next = pmd_addr_end(addr, end);
1158                 err = zeromap_pte_range(mm, pmd, addr, next, prot);
1159                 if (err)
1160                         break;
1161         } while (pmd++, addr = next, addr != end);
1162         return err;
1163 }
1164
1165 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1166                         unsigned long addr, unsigned long end, pgprot_t prot)
1167 {
1168         pud_t *pud;
1169         unsigned long next;
1170         int err;
1171
1172         pud = pud_alloc(mm, pgd, addr);
1173         if (!pud)
1174                 return -EAGAIN;
1175         do {
1176                 next = pud_addr_end(addr, end);
1177                 err = zeromap_pmd_range(mm, pud, addr, next, prot);
1178                 if (err)
1179                         break;
1180         } while (pud++, addr = next, addr != end);
1181         return err;
1182 }
1183
1184 int zeromap_page_range(struct vm_area_struct *vma,
1185                         unsigned long addr, unsigned long size, pgprot_t prot)
1186 {
1187         pgd_t *pgd;
1188         unsigned long next;
1189         unsigned long end = addr + size;
1190         struct mm_struct *mm = vma->vm_mm;
1191         int err;
1192
1193         BUG_ON(addr >= end);
1194         pgd = pgd_offset(mm, addr);
1195         flush_cache_range(vma, addr, end);
1196         do {
1197                 next = pgd_addr_end(addr, end);
1198                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1199                 if (err)
1200                         break;
1201         } while (pgd++, addr = next, addr != end);
1202         return err;
1203 }
1204
1205 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1206 {
1207         pgd_t * pgd = pgd_offset(mm, addr);
1208         pud_t * pud = pud_alloc(mm, pgd, addr);
1209         if (pud) {
1210                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1211                 if (pmd)
1212                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1213         }
1214         return NULL;
1215 }
1216
1217 /*
1218  * This is the old fallback for page remapping.
1219  *
1220  * For historical reasons, it only allows reserved pages. Only
1221  * old drivers should use this, and they needed to mark their
1222  * pages reserved for the old functions anyway.
1223  */
1224 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1225 {
1226         int retval;
1227         pte_t *pte;
1228         spinlock_t *ptl;  
1229
1230         retval = -EINVAL;
1231         if (PageAnon(page))
1232                 goto out;
1233         retval = -ENOMEM;
1234         flush_dcache_page(page);
1235         pte = get_locked_pte(mm, addr, &ptl);
1236         if (!pte)
1237                 goto out;
1238         retval = -EBUSY;
1239         if (!pte_none(*pte))
1240                 goto out_unlock;
1241
1242         /* Ok, finally just insert the thing.. */
1243         get_page(page);
1244         inc_mm_counter(mm, file_rss);
1245         page_add_file_rmap(page);
1246         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1247
1248         retval = 0;
1249 out_unlock:
1250         pte_unmap_unlock(pte, ptl);
1251 out:
1252         return retval;
1253 }
1254
1255 /**
1256  * vm_insert_page - insert single page into user vma
1257  * @vma: user vma to map to
1258  * @addr: target user address of this page
1259  * @page: source kernel page
1260  *
1261  * This allows drivers to insert individual pages they've allocated
1262  * into a user vma.
1263  *
1264  * The page has to be a nice clean _individual_ kernel allocation.
1265  * If you allocate a compound page, you need to have marked it as
1266  * such (__GFP_COMP), or manually just split the page up yourself
1267  * (see split_page()).
1268  *
1269  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1270  * took an arbitrary page protection parameter. This doesn't allow
1271  * that. Your vma protection will have to be set up correctly, which
1272  * means that if you want a shared writable mapping, you'd better
1273  * ask for a shared writable mapping!
1274  *
1275  * The page does not need to be reserved.
1276  */
1277 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1278 {
1279         if (addr < vma->vm_start || addr >= vma->vm_end)
1280                 return -EFAULT;
1281         if (!page_count(page))
1282                 return -EINVAL;
1283         vma->vm_flags |= VM_INSERTPAGE;
1284         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1285 }
1286 EXPORT_SYMBOL(vm_insert_page);
1287
1288 /**
1289  * vm_insert_pfn - insert single pfn into user vma
1290  * @vma: user vma to map to
1291  * @addr: target user address of this page
1292  * @pfn: source kernel pfn
1293  *
1294  * Similar to vm_inert_page, this allows drivers to insert individual pages
1295  * they've allocated into a user vma. Same comments apply.
1296  *
1297  * This function should only be called from a vm_ops->fault handler, and
1298  * in that case the handler should return NULL.
1299  */
1300 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1301                 unsigned long pfn)
1302 {
1303         struct mm_struct *mm = vma->vm_mm;
1304         int retval;
1305         pte_t *pte, entry;
1306         spinlock_t *ptl;
1307
1308         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
1309         BUG_ON(is_cow_mapping(vma->vm_flags));
1310
1311         retval = -ENOMEM;
1312         pte = get_locked_pte(mm, addr, &ptl);
1313         if (!pte)
1314                 goto out;
1315         retval = -EBUSY;
1316         if (!pte_none(*pte))
1317                 goto out_unlock;
1318
1319         /* Ok, finally just insert the thing.. */
1320         entry = pfn_pte(pfn, vma->vm_page_prot);
1321         set_pte_at(mm, addr, pte, entry);
1322         update_mmu_cache(vma, addr, entry);
1323
1324         retval = 0;
1325 out_unlock:
1326         pte_unmap_unlock(pte, ptl);
1327
1328 out:
1329         return retval;
1330 }
1331 EXPORT_SYMBOL(vm_insert_pfn);
1332
1333 /*
1334  * maps a range of physical memory into the requested pages. the old
1335  * mappings are removed. any references to nonexistent pages results
1336  * in null mappings (currently treated as "copy-on-access")
1337  */
1338 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1339                         unsigned long addr, unsigned long end,
1340                         unsigned long pfn, pgprot_t prot)
1341 {
1342         pte_t *pte;
1343         spinlock_t *ptl;
1344
1345         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1346         if (!pte)
1347                 return -ENOMEM;
1348         arch_enter_lazy_mmu_mode();
1349         do {
1350                 BUG_ON(!pte_none(*pte));
1351                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1352                 pfn++;
1353         } while (pte++, addr += PAGE_SIZE, addr != end);
1354         arch_leave_lazy_mmu_mode();
1355         pte_unmap_unlock(pte - 1, ptl);
1356         return 0;
1357 }
1358
1359 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1360                         unsigned long addr, unsigned long end,
1361                         unsigned long pfn, pgprot_t prot)
1362 {
1363         pmd_t *pmd;
1364         unsigned long next;
1365
1366         pfn -= addr >> PAGE_SHIFT;
1367         pmd = pmd_alloc(mm, pud, addr);
1368         if (!pmd)
1369                 return -ENOMEM;
1370         do {
1371                 next = pmd_addr_end(addr, end);
1372                 if (remap_pte_range(mm, pmd, addr, next,
1373                                 pfn + (addr >> PAGE_SHIFT), prot))
1374                         return -ENOMEM;
1375         } while (pmd++, addr = next, addr != end);
1376         return 0;
1377 }
1378
1379 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1380                         unsigned long addr, unsigned long end,
1381                         unsigned long pfn, pgprot_t prot)
1382 {
1383         pud_t *pud;
1384         unsigned long next;
1385
1386         pfn -= addr >> PAGE_SHIFT;
1387         pud = pud_alloc(mm, pgd, addr);
1388         if (!pud)
1389                 return -ENOMEM;
1390         do {
1391                 next = pud_addr_end(addr, end);
1392                 if (remap_pmd_range(mm, pud, addr, next,
1393                                 pfn + (addr >> PAGE_SHIFT), prot))
1394                         return -ENOMEM;
1395         } while (pud++, addr = next, addr != end);
1396         return 0;
1397 }
1398
1399 /**
1400  * remap_pfn_range - remap kernel memory to userspace
1401  * @vma: user vma to map to
1402  * @addr: target user address to start at
1403  * @pfn: physical address of kernel memory
1404  * @size: size of map area
1405  * @prot: page protection flags for this mapping
1406  *
1407  *  Note: this is only safe if the mm semaphore is held when called.
1408  */
1409 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1410                     unsigned long pfn, unsigned long size, pgprot_t prot)
1411 {
1412         pgd_t *pgd;
1413         unsigned long next;
1414         unsigned long end = addr + PAGE_ALIGN(size);
1415         struct mm_struct *mm = vma->vm_mm;
1416         int err;
1417
1418         /*
1419          * Physically remapped pages are special. Tell the
1420          * rest of the world about it:
1421          *   VM_IO tells people not to look at these pages
1422          *      (accesses can have side effects).
1423          *   VM_RESERVED is specified all over the place, because
1424          *      in 2.4 it kept swapout's vma scan off this vma; but
1425          *      in 2.6 the LRU scan won't even find its pages, so this
1426          *      flag means no more than count its pages in reserved_vm,
1427          *      and omit it from core dump, even when VM_IO turned off.
1428          *   VM_PFNMAP tells the core MM that the base pages are just
1429          *      raw PFN mappings, and do not have a "struct page" associated
1430          *      with them.
1431          *
1432          * There's a horrible special case to handle copy-on-write
1433          * behaviour that some programs depend on. We mark the "original"
1434          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1435          */
1436         if (is_cow_mapping(vma->vm_flags)) {
1437                 if (addr != vma->vm_start || end != vma->vm_end)
1438                         return -EINVAL;
1439                 vma->vm_pgoff = pfn;
1440         }
1441
1442         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1443
1444         BUG_ON(addr >= end);
1445         pfn -= addr >> PAGE_SHIFT;
1446         pgd = pgd_offset(mm, addr);
1447         flush_cache_range(vma, addr, end);
1448         do {
1449                 next = pgd_addr_end(addr, end);
1450                 err = remap_pud_range(mm, pgd, addr, next,
1451                                 pfn + (addr >> PAGE_SHIFT), prot);
1452                 if (err)
1453                         break;
1454         } while (pgd++, addr = next, addr != end);
1455         return err;
1456 }
1457 EXPORT_SYMBOL(remap_pfn_range);
1458
1459 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1460                                      unsigned long addr, unsigned long end,
1461                                      pte_fn_t fn, void *data)
1462 {
1463         pte_t *pte;
1464         int err;
1465         struct page *pmd_page;
1466         spinlock_t *uninitialized_var(ptl);
1467
1468         pte = (mm == &init_mm) ?
1469                 pte_alloc_kernel(pmd, addr) :
1470                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1471         if (!pte)
1472                 return -ENOMEM;
1473
1474         BUG_ON(pmd_huge(*pmd));
1475
1476         pmd_page = pmd_page(*pmd);
1477
1478         do {
1479                 err = fn(pte, pmd_page, addr, data);
1480                 if (err)
1481                         break;
1482         } while (pte++, addr += PAGE_SIZE, addr != end);
1483
1484         if (mm != &init_mm)
1485                 pte_unmap_unlock(pte-1, ptl);
1486         return err;
1487 }
1488
1489 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1490                                      unsigned long addr, unsigned long end,
1491                                      pte_fn_t fn, void *data)
1492 {
1493         pmd_t *pmd;
1494         unsigned long next;
1495         int err;
1496
1497         pmd = pmd_alloc(mm, pud, addr);
1498         if (!pmd)
1499                 return -ENOMEM;
1500         do {
1501                 next = pmd_addr_end(addr, end);
1502                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1503                 if (err)
1504                         break;
1505         } while (pmd++, addr = next, addr != end);
1506         return err;
1507 }
1508
1509 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1510                                      unsigned long addr, unsigned long end,
1511                                      pte_fn_t fn, void *data)
1512 {
1513         pud_t *pud;
1514         unsigned long next;
1515         int err;
1516
1517         pud = pud_alloc(mm, pgd, addr);
1518         if (!pud)
1519                 return -ENOMEM;
1520         do {
1521                 next = pud_addr_end(addr, end);
1522                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1523                 if (err)
1524                         break;
1525         } while (pud++, addr = next, addr != end);
1526         return err;
1527 }
1528
1529 /*
1530  * Scan a region of virtual memory, filling in page tables as necessary
1531  * and calling a provided function on each leaf page table.
1532  */
1533 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1534                         unsigned long size, pte_fn_t fn, void *data)
1535 {
1536         pgd_t *pgd;
1537         unsigned long next;
1538         unsigned long end = addr + size;
1539         int err;
1540
1541         BUG_ON(addr >= end);
1542         pgd = pgd_offset(mm, addr);
1543         do {
1544                 next = pgd_addr_end(addr, end);
1545                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1546                 if (err)
1547                         break;
1548         } while (pgd++, addr = next, addr != end);
1549         return err;
1550 }
1551 EXPORT_SYMBOL_GPL(apply_to_page_range);
1552
1553 /*
1554  * handle_pte_fault chooses page fault handler according to an entry
1555  * which was read non-atomically.  Before making any commitment, on
1556  * those architectures or configurations (e.g. i386 with PAE) which
1557  * might give a mix of unmatched parts, do_swap_page and do_file_page
1558  * must check under lock before unmapping the pte and proceeding
1559  * (but do_wp_page is only called after already making such a check;
1560  * and do_anonymous_page and do_no_page can safely check later on).
1561  */
1562 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1563                                 pte_t *page_table, pte_t orig_pte)
1564 {
1565         int same = 1;
1566 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1567         if (sizeof(pte_t) > sizeof(unsigned long)) {
1568                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1569                 spin_lock(ptl);
1570                 same = pte_same(*page_table, orig_pte);
1571                 spin_unlock(ptl);
1572         }
1573 #endif
1574         pte_unmap(page_table);
1575         return same;
1576 }
1577
1578 /*
1579  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1580  * servicing faults for write access.  In the normal case, do always want
1581  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1582  * that do not have writing enabled, when used by access_process_vm.
1583  */
1584 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1585 {
1586         if (likely(vma->vm_flags & VM_WRITE))
1587                 pte = pte_mkwrite(pte);
1588         return pte;
1589 }
1590
1591 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1592 {
1593         /*
1594          * If the source page was a PFN mapping, we don't have
1595          * a "struct page" for it. We do a best-effort copy by
1596          * just copying from the original user address. If that
1597          * fails, we just zero-fill it. Live with it.
1598          */
1599         if (unlikely(!src)) {
1600                 void *kaddr = kmap_atomic(dst, KM_USER0);
1601                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1602
1603                 /*
1604                  * This really shouldn't fail, because the page is there
1605                  * in the page tables. But it might just be unreadable,
1606                  * in which case we just give up and fill the result with
1607                  * zeroes.
1608                  */
1609                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1610                         memset(kaddr, 0, PAGE_SIZE);
1611                 kunmap_atomic(kaddr, KM_USER0);
1612                 flush_dcache_page(dst);
1613                 return;
1614
1615         }
1616         copy_user_highpage(dst, src, va, vma);
1617 }
1618
1619 /*
1620  * This routine handles present pages, when users try to write
1621  * to a shared page. It is done by copying the page to a new address
1622  * and decrementing the shared-page counter for the old page.
1623  *
1624  * Note that this routine assumes that the protection checks have been
1625  * done by the caller (the low-level page fault routine in most cases).
1626  * Thus we can safely just mark it writable once we've done any necessary
1627  * COW.
1628  *
1629  * We also mark the page dirty at this point even though the page will
1630  * change only once the write actually happens. This avoids a few races,
1631  * and potentially makes it more efficient.
1632  *
1633  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1634  * but allow concurrent faults), with pte both mapped and locked.
1635  * We return with mmap_sem still held, but pte unmapped and unlocked.
1636  */
1637 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1638                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1639                 spinlock_t *ptl, pte_t orig_pte)
1640 {
1641         struct page *old_page, *new_page;
1642         pte_t entry;
1643         int reuse = 0, ret = VM_FAULT_MINOR;
1644         struct page *dirty_page = NULL;
1645
1646         old_page = vm_normal_page(vma, address, orig_pte);
1647         if (!old_page)
1648                 goto gotten;
1649
1650         /*
1651          * Take out anonymous pages first, anonymous shared vmas are
1652          * not dirty accountable.
1653          */
1654         if (PageAnon(old_page)) {
1655                 if (!TestSetPageLocked(old_page)) {
1656                         reuse = can_share_swap_page(old_page);
1657                         unlock_page(old_page);
1658                 }
1659         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1660                                         (VM_WRITE|VM_SHARED))) {
1661                 /*
1662                  * Only catch write-faults on shared writable pages,
1663                  * read-only shared pages can get COWed by
1664                  * get_user_pages(.write=1, .force=1).
1665                  */
1666                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1667                         /*
1668                          * Notify the address space that the page is about to
1669                          * become writable so that it can prohibit this or wait
1670                          * for the page to get into an appropriate state.
1671                          *
1672                          * We do this without the lock held, so that it can
1673                          * sleep if it needs to.
1674                          */
1675                         page_cache_get(old_page);
1676                         pte_unmap_unlock(page_table, ptl);
1677
1678                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1679                                 goto unwritable_page;
1680
1681                         /*
1682                          * Since we dropped the lock we need to revalidate
1683                          * the PTE as someone else may have changed it.  If
1684                          * they did, we just return, as we can count on the
1685                          * MMU to tell us if they didn't also make it writable.
1686                          */
1687                         page_table = pte_offset_map_lock(mm, pmd, address,
1688                                                          &ptl);
1689                         page_cache_release(old_page);
1690                         if (!pte_same(*page_table, orig_pte))
1691                                 goto unlock;
1692                 }
1693                 dirty_page = old_page;
1694                 get_page(dirty_page);
1695                 reuse = 1;
1696         }
1697
1698         if (reuse) {
1699                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1700                 entry = pte_mkyoung(orig_pte);
1701                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1702                 if (ptep_set_access_flags(vma, address, page_table, entry,1)) {
1703                         update_mmu_cache(vma, address, entry);
1704                         lazy_mmu_prot_update(entry);
1705                 }
1706                 ret |= VM_FAULT_WRITE;
1707                 goto unlock;
1708         }
1709
1710         /*
1711          * Ok, we need to copy. Oh, well..
1712          */
1713         page_cache_get(old_page);
1714 gotten:
1715         pte_unmap_unlock(page_table, ptl);
1716
1717         if (unlikely(anon_vma_prepare(vma)))
1718                 goto oom;
1719         if (old_page == ZERO_PAGE(address)) {
1720                 new_page = alloc_zeroed_user_highpage(vma, address);
1721                 if (!new_page)
1722                         goto oom;
1723         } else {
1724                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1725                 if (!new_page)
1726                         goto oom;
1727                 cow_user_page(new_page, old_page, address, vma);
1728         }
1729
1730         /*
1731          * Re-check the pte - we dropped the lock
1732          */
1733         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1734         if (likely(pte_same(*page_table, orig_pte))) {
1735                 if (old_page) {
1736                         page_remove_rmap(old_page, vma);
1737                         if (!PageAnon(old_page)) {
1738                                 dec_mm_counter(mm, file_rss);
1739                                 inc_mm_counter(mm, anon_rss);
1740                         }
1741                 } else
1742                         inc_mm_counter(mm, anon_rss);
1743                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1744                 entry = mk_pte(new_page, vma->vm_page_prot);
1745                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1746                 lazy_mmu_prot_update(entry);
1747                 /*
1748                  * Clear the pte entry and flush it first, before updating the
1749                  * pte with the new entry. This will avoid a race condition
1750                  * seen in the presence of one thread doing SMC and another
1751                  * thread doing COW.
1752                  */
1753                 ptep_clear_flush(vma, address, page_table);
1754                 set_pte_at(mm, address, page_table, entry);
1755                 update_mmu_cache(vma, address, entry);
1756                 lru_cache_add_active(new_page);
1757                 page_add_new_anon_rmap(new_page, vma, address);
1758
1759                 /* Free the old page.. */
1760                 new_page = old_page;
1761                 ret |= VM_FAULT_WRITE;
1762         }
1763         if (new_page)
1764                 page_cache_release(new_page);
1765         if (old_page)
1766                 page_cache_release(old_page);
1767 unlock:
1768         pte_unmap_unlock(page_table, ptl);
1769         if (dirty_page) {
1770                 set_page_dirty_balance(dirty_page);
1771                 put_page(dirty_page);
1772         }
1773         return ret;
1774 oom:
1775         if (old_page)
1776                 page_cache_release(old_page);
1777         return VM_FAULT_OOM;
1778
1779 unwritable_page:
1780         page_cache_release(old_page);
1781         return VM_FAULT_SIGBUS;
1782 }
1783
1784 /*
1785  * Helper functions for unmap_mapping_range().
1786  *
1787  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1788  *
1789  * We have to restart searching the prio_tree whenever we drop the lock,
1790  * since the iterator is only valid while the lock is held, and anyway
1791  * a later vma might be split and reinserted earlier while lock dropped.
1792  *
1793  * The list of nonlinear vmas could be handled more efficiently, using
1794  * a placeholder, but handle it in the same way until a need is shown.
1795  * It is important to search the prio_tree before nonlinear list: a vma
1796  * may become nonlinear and be shifted from prio_tree to nonlinear list
1797  * while the lock is dropped; but never shifted from list to prio_tree.
1798  *
1799  * In order to make forward progress despite restarting the search,
1800  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1801  * quickly skip it next time around.  Since the prio_tree search only
1802  * shows us those vmas affected by unmapping the range in question, we
1803  * can't efficiently keep all vmas in step with mapping->truncate_count:
1804  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1805  * mapping->truncate_count and vma->vm_truncate_count are protected by
1806  * i_mmap_lock.
1807  *
1808  * In order to make forward progress despite repeatedly restarting some
1809  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1810  * and restart from that address when we reach that vma again.  It might
1811  * have been split or merged, shrunk or extended, but never shifted: so
1812  * restart_addr remains valid so long as it remains in the vma's range.
1813  * unmap_mapping_range forces truncate_count to leap over page-aligned
1814  * values so we can save vma's restart_addr in its truncate_count field.
1815  */
1816 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1817
1818 static void reset_vma_truncate_counts(struct address_space *mapping)
1819 {
1820         struct vm_area_struct *vma;
1821         struct prio_tree_iter iter;
1822
1823         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1824                 vma->vm_truncate_count = 0;
1825         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1826                 vma->vm_truncate_count = 0;
1827 }
1828
1829 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1830                 unsigned long start_addr, unsigned long end_addr,
1831                 struct zap_details *details)
1832 {
1833         unsigned long restart_addr;
1834         int need_break;
1835
1836 again:
1837         restart_addr = vma->vm_truncate_count;
1838         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1839                 start_addr = restart_addr;
1840                 if (start_addr >= end_addr) {
1841                         /* Top of vma has been split off since last time */
1842                         vma->vm_truncate_count = details->truncate_count;
1843                         return 0;
1844                 }
1845         }
1846
1847         restart_addr = zap_page_range(vma, start_addr,
1848                                         end_addr - start_addr, details);
1849         need_break = need_resched() ||
1850                         need_lockbreak(details->i_mmap_lock);
1851
1852         if (restart_addr >= end_addr) {
1853                 /* We have now completed this vma: mark it so */
1854                 vma->vm_truncate_count = details->truncate_count;
1855                 if (!need_break)
1856                         return 0;
1857         } else {
1858                 /* Note restart_addr in vma's truncate_count field */
1859                 vma->vm_truncate_count = restart_addr;
1860                 if (!need_break)
1861                         goto again;
1862         }
1863
1864         spin_unlock(details->i_mmap_lock);
1865         cond_resched();
1866         spin_lock(details->i_mmap_lock);
1867         return -EINTR;
1868 }
1869
1870 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1871                                             struct zap_details *details)
1872 {
1873         struct vm_area_struct *vma;
1874         struct prio_tree_iter iter;
1875         pgoff_t vba, vea, zba, zea;
1876
1877 restart:
1878         vma_prio_tree_foreach(vma, &iter, root,
1879                         details->first_index, details->last_index) {
1880                 /* Skip quickly over those we have already dealt with */
1881                 if (vma->vm_truncate_count == details->truncate_count)
1882                         continue;
1883
1884                 vba = vma->vm_pgoff;
1885                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1886                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1887                 zba = details->first_index;
1888                 if (zba < vba)
1889                         zba = vba;
1890                 zea = details->last_index;
1891                 if (zea > vea)
1892                         zea = vea;
1893
1894                 if (unmap_mapping_range_vma(vma,
1895                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1896                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1897                                 details) < 0)
1898                         goto restart;
1899         }
1900 }
1901
1902 static inline void unmap_mapping_range_list(struct list_head *head,
1903                                             struct zap_details *details)
1904 {
1905         struct vm_area_struct *vma;
1906
1907         /*
1908          * In nonlinear VMAs there is no correspondence between virtual address
1909          * offset and file offset.  So we must perform an exhaustive search
1910          * across *all* the pages in each nonlinear VMA, not just the pages
1911          * whose virtual address lies outside the file truncation point.
1912          */
1913 restart:
1914         list_for_each_entry(vma, head, shared.vm_set.list) {
1915                 /* Skip quickly over those we have already dealt with */
1916                 if (vma->vm_truncate_count == details->truncate_count)
1917                         continue;
1918                 details->nonlinear_vma = vma;
1919                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1920                                         vma->vm_end, details) < 0)
1921                         goto restart;
1922         }
1923 }
1924
1925 /**
1926  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
1927  * @mapping: the address space containing mmaps to be unmapped.
1928  * @holebegin: byte in first page to unmap, relative to the start of
1929  * the underlying file.  This will be rounded down to a PAGE_SIZE
1930  * boundary.  Note that this is different from vmtruncate(), which
1931  * must keep the partial page.  In contrast, we must get rid of
1932  * partial pages.
1933  * @holelen: size of prospective hole in bytes.  This will be rounded
1934  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1935  * end of the file.
1936  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1937  * but 0 when invalidating pagecache, don't throw away private data.
1938  */
1939 void unmap_mapping_range(struct address_space *mapping,
1940                 loff_t const holebegin, loff_t const holelen, int even_cows)
1941 {
1942         struct zap_details details;
1943         pgoff_t hba = holebegin >> PAGE_SHIFT;
1944         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1945
1946         /* Check for overflow. */
1947         if (sizeof(holelen) > sizeof(hlen)) {
1948                 long long holeend =
1949                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1950                 if (holeend & ~(long long)ULONG_MAX)
1951                         hlen = ULONG_MAX - hba + 1;
1952         }
1953
1954         details.check_mapping = even_cows? NULL: mapping;
1955         details.nonlinear_vma = NULL;
1956         details.first_index = hba;
1957         details.last_index = hba + hlen - 1;
1958         if (details.last_index < details.first_index)
1959                 details.last_index = ULONG_MAX;
1960         details.i_mmap_lock = &mapping->i_mmap_lock;
1961
1962         spin_lock(&mapping->i_mmap_lock);
1963
1964         /* serialize i_size write against truncate_count write */
1965         smp_wmb();
1966         /* Protect against page faults, and endless unmapping loops */
1967         mapping->truncate_count++;
1968         /*
1969          * For archs where spin_lock has inclusive semantics like ia64
1970          * this smp_mb() will prevent to read pagetable contents
1971          * before the truncate_count increment is visible to
1972          * other cpus.
1973          */
1974         smp_mb();
1975         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1976                 if (mapping->truncate_count == 0)
1977                         reset_vma_truncate_counts(mapping);
1978                 mapping->truncate_count++;
1979         }
1980         details.truncate_count = mapping->truncate_count;
1981
1982         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1983                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1984         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1985                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1986         spin_unlock(&mapping->i_mmap_lock);
1987 }
1988 EXPORT_SYMBOL(unmap_mapping_range);
1989
1990 /**
1991  * vmtruncate - unmap mappings "freed" by truncate() syscall
1992  * @inode: inode of the file used
1993  * @offset: file offset to start truncating
1994  *
1995  * NOTE! We have to be ready to update the memory sharing
1996  * between the file and the memory map for a potential last
1997  * incomplete page.  Ugly, but necessary.
1998  */
1999 int vmtruncate(struct inode * inode, loff_t offset)
2000 {
2001         struct address_space *mapping = inode->i_mapping;
2002         unsigned long limit;
2003
2004         if (inode->i_size < offset)
2005                 goto do_expand;
2006         /*
2007          * truncation of in-use swapfiles is disallowed - it would cause
2008          * subsequent swapout to scribble on the now-freed blocks.
2009          */
2010         if (IS_SWAPFILE(inode))
2011                 goto out_busy;
2012         i_size_write(inode, offset);
2013         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2014         truncate_inode_pages(mapping, offset);
2015         goto out_truncate;
2016
2017 do_expand:
2018         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2019         if (limit != RLIM_INFINITY && offset > limit)
2020                 goto out_sig;
2021         if (offset > inode->i_sb->s_maxbytes)
2022                 goto out_big;
2023         i_size_write(inode, offset);
2024
2025 out_truncate:
2026         if (inode->i_op && inode->i_op->truncate)
2027                 inode->i_op->truncate(inode);
2028         return 0;
2029 out_sig:
2030         send_sig(SIGXFSZ, current, 0);
2031 out_big:
2032         return -EFBIG;
2033 out_busy:
2034         return -ETXTBSY;
2035 }
2036 EXPORT_SYMBOL(vmtruncate);
2037
2038 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2039 {
2040         struct address_space *mapping = inode->i_mapping;
2041
2042         /*
2043          * If the underlying filesystem is not going to provide
2044          * a way to truncate a range of blocks (punch a hole) -
2045          * we should return failure right now.
2046          */
2047         if (!inode->i_op || !inode->i_op->truncate_range)
2048                 return -ENOSYS;
2049
2050         mutex_lock(&inode->i_mutex);
2051         down_write(&inode->i_alloc_sem);
2052         unmap_mapping_range(mapping, offset, (end - offset), 1);
2053         truncate_inode_pages_range(mapping, offset, end);
2054         inode->i_op->truncate_range(inode, offset, end);
2055         up_write(&inode->i_alloc_sem);
2056         mutex_unlock(&inode->i_mutex);
2057
2058         return 0;
2059 }
2060
2061 /**
2062  * swapin_readahead - swap in pages in hope we need them soon
2063  * @entry: swap entry of this memory
2064  * @addr: address to start
2065  * @vma: user vma this addresses belong to
2066  *
2067  * Primitive swap readahead code. We simply read an aligned block of
2068  * (1 << page_cluster) entries in the swap area. This method is chosen
2069  * because it doesn't cost us any seek time.  We also make sure to queue
2070  * the 'original' request together with the readahead ones...
2071  *
2072  * This has been extended to use the NUMA policies from the mm triggering
2073  * the readahead.
2074  *
2075  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
2076  */
2077 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
2078 {
2079 #ifdef CONFIG_NUMA
2080         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
2081 #endif
2082         int i, num;
2083         struct page *new_page;
2084         unsigned long offset;
2085
2086         /*
2087          * Get the number of handles we should do readahead io to.
2088          */
2089         num = valid_swaphandles(entry, &offset);
2090         for (i = 0; i < num; offset++, i++) {
2091                 /* Ok, do the async read-ahead now */
2092                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
2093                                                            offset), vma, addr);
2094                 if (!new_page)
2095                         break;
2096                 page_cache_release(new_page);
2097 #ifdef CONFIG_NUMA
2098                 /*
2099                  * Find the next applicable VMA for the NUMA policy.
2100                  */
2101                 addr += PAGE_SIZE;
2102                 if (addr == 0)
2103                         vma = NULL;
2104                 if (vma) {
2105                         if (addr >= vma->vm_end) {
2106                                 vma = next_vma;
2107                                 next_vma = vma ? vma->vm_next : NULL;
2108                         }
2109                         if (vma && addr < vma->vm_start)
2110                                 vma = NULL;
2111                 } else {
2112                         if (next_vma && addr >= next_vma->vm_start) {
2113                                 vma = next_vma;
2114                                 next_vma = vma->vm_next;
2115                         }
2116                 }
2117 #endif
2118         }
2119         lru_add_drain();        /* Push any new pages onto the LRU now */
2120 }
2121
2122 /*
2123  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2124  * but allow concurrent faults), and pte mapped but not yet locked.
2125  * We return with mmap_sem still held, but pte unmapped and unlocked.
2126  */
2127 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2128                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2129                 int write_access, pte_t orig_pte)
2130 {
2131         spinlock_t *ptl;
2132         struct page *page;
2133         swp_entry_t entry;
2134         pte_t pte;
2135         int ret = VM_FAULT_MINOR;
2136
2137         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2138                 goto out;
2139
2140         entry = pte_to_swp_entry(orig_pte);
2141         if (is_migration_entry(entry)) {
2142                 migration_entry_wait(mm, pmd, address);
2143                 goto out;
2144         }
2145         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2146         page = lookup_swap_cache(entry);
2147         if (!page) {
2148                 grab_swap_token(); /* Contend for token _before_ read-in */
2149                 swapin_readahead(entry, address, vma);
2150                 page = read_swap_cache_async(entry, vma, address);
2151                 if (!page) {
2152                         /*
2153                          * Back out if somebody else faulted in this pte
2154                          * while we released the pte lock.
2155                          */
2156                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2157                         if (likely(pte_same(*page_table, orig_pte)))
2158                                 ret = VM_FAULT_OOM;
2159                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2160                         goto unlock;
2161                 }
2162
2163                 /* Had to read the page from swap area: Major fault */
2164                 ret = VM_FAULT_MAJOR;
2165                 count_vm_event(PGMAJFAULT);
2166         }
2167
2168         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2169         mark_page_accessed(page);
2170         lock_page(page);
2171
2172         /*
2173          * Back out if somebody else already faulted in this pte.
2174          */
2175         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2176         if (unlikely(!pte_same(*page_table, orig_pte)))
2177                 goto out_nomap;
2178
2179         if (unlikely(!PageUptodate(page))) {
2180                 ret = VM_FAULT_SIGBUS;
2181                 goto out_nomap;
2182         }
2183
2184         /* The page isn't present yet, go ahead with the fault. */
2185
2186         inc_mm_counter(mm, anon_rss);
2187         pte = mk_pte(page, vma->vm_page_prot);
2188         if (write_access && can_share_swap_page(page)) {
2189                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2190                 write_access = 0;
2191         }
2192
2193         flush_icache_page(vma, page);
2194         set_pte_at(mm, address, page_table, pte);
2195         page_add_anon_rmap(page, vma, address);
2196
2197         swap_free(entry);
2198         if (vm_swap_full())
2199                 remove_exclusive_swap_page(page);
2200         unlock_page(page);
2201
2202         if (write_access) {
2203                 if (do_wp_page(mm, vma, address,
2204                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
2205                         ret = VM_FAULT_OOM;
2206                 goto out;
2207         }
2208
2209         /* No need to invalidate - it was non-present before */
2210         update_mmu_cache(vma, address, pte);
2211         lazy_mmu_prot_update(pte);
2212 unlock:
2213         pte_unmap_unlock(page_table, ptl);
2214 out:
2215         return ret;
2216 out_nomap:
2217         pte_unmap_unlock(page_table, ptl);
2218         unlock_page(page);
2219         page_cache_release(page);
2220         return ret;
2221 }
2222
2223 /*
2224  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2225  * but allow concurrent faults), and pte mapped but not yet locked.
2226  * We return with mmap_sem still held, but pte unmapped and unlocked.
2227  */
2228 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2229                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2230                 int write_access)
2231 {
2232         struct page *page;
2233         spinlock_t *ptl;
2234         pte_t entry;
2235
2236         if (write_access) {
2237                 /* Allocate our own private page. */
2238                 pte_unmap(page_table);
2239
2240                 if (unlikely(anon_vma_prepare(vma)))
2241                         goto oom;
2242                 page = alloc_zeroed_user_highpage(vma, address);
2243                 if (!page)
2244                         goto oom;
2245
2246                 entry = mk_pte(page, vma->vm_page_prot);
2247                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2248
2249                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2250                 if (!pte_none(*page_table))
2251                         goto release;
2252                 inc_mm_counter(mm, anon_rss);
2253                 lru_cache_add_active(page);
2254                 page_add_new_anon_rmap(page, vma, address);
2255         } else {
2256                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2257                 page = ZERO_PAGE(address);
2258                 page_cache_get(page);
2259                 entry = mk_pte(page, vma->vm_page_prot);
2260
2261                 ptl = pte_lockptr(mm, pmd);
2262                 spin_lock(ptl);
2263                 if (!pte_none(*page_table))
2264                         goto release;
2265                 inc_mm_counter(mm, file_rss);
2266                 page_add_file_rmap(page);
2267         }
2268
2269         set_pte_at(mm, address, page_table, entry);
2270
2271         /* No need to invalidate - it was non-present before */
2272         update_mmu_cache(vma, address, entry);
2273         lazy_mmu_prot_update(entry);
2274 unlock:
2275         pte_unmap_unlock(page_table, ptl);
2276         return VM_FAULT_MINOR;
2277 release:
2278         page_cache_release(page);
2279         goto unlock;
2280 oom:
2281         return VM_FAULT_OOM;
2282 }
2283
2284 /*
2285  * do_no_page() tries to create a new page mapping. It aggressively
2286  * tries to share with existing pages, but makes a separate copy if
2287  * the "write_access" parameter is true in order to avoid the next
2288  * page fault.
2289  *
2290  * As this is called only for pages that do not currently exist, we
2291  * do not need to flush old virtual caches or the TLB.
2292  *
2293  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2294  * but allow concurrent faults), and pte mapped but not yet locked.
2295  * We return with mmap_sem still held, but pte unmapped and unlocked.
2296  */
2297 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2298                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2299                 int write_access)
2300 {
2301         spinlock_t *ptl;
2302         struct page *new_page;
2303         struct address_space *mapping = NULL;
2304         pte_t entry;
2305         unsigned int sequence = 0;
2306         int ret = VM_FAULT_MINOR;
2307         int anon = 0;
2308         struct page *dirty_page = NULL;
2309
2310         pte_unmap(page_table);
2311         BUG_ON(vma->vm_flags & VM_PFNMAP);
2312
2313         if (vma->vm_file) {
2314                 mapping = vma->vm_file->f_mapping;
2315                 sequence = mapping->truncate_count;
2316                 smp_rmb(); /* serializes i_size against truncate_count */
2317         }
2318 retry:
2319         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2320         /*
2321          * No smp_rmb is needed here as long as there's a full
2322          * spin_lock/unlock sequence inside the ->nopage callback
2323          * (for the pagecache lookup) that acts as an implicit
2324          * smp_mb() and prevents the i_size read to happen
2325          * after the next truncate_count read.
2326          */
2327
2328         /* no page was available -- either SIGBUS, OOM or REFAULT */
2329         if (unlikely(new_page == NOPAGE_SIGBUS))
2330                 return VM_FAULT_SIGBUS;
2331         else if (unlikely(new_page == NOPAGE_OOM))
2332                 return VM_FAULT_OOM;
2333         else if (unlikely(new_page == NOPAGE_REFAULT))
2334                 return VM_FAULT_MINOR;
2335
2336         /*
2337          * Should we do an early C-O-W break?
2338          */
2339         if (write_access) {
2340                 if (!(vma->vm_flags & VM_SHARED)) {
2341                         struct page *page;
2342
2343                         if (unlikely(anon_vma_prepare(vma)))
2344                                 goto oom;
2345                         page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2346                         if (!page)
2347                                 goto oom;
2348                         copy_user_highpage(page, new_page, address, vma);
2349                         page_cache_release(new_page);
2350                         new_page = page;
2351                         anon = 1;
2352
2353                 } else {
2354                         /* if the page will be shareable, see if the backing
2355                          * address space wants to know that the page is about
2356                          * to become writable */
2357                         if (vma->vm_ops->page_mkwrite &&
2358                             vma->vm_ops->page_mkwrite(vma, new_page) < 0
2359                             ) {
2360                                 page_cache_release(new_page);
2361                                 return VM_FAULT_SIGBUS;
2362                         }
2363                 }
2364         }
2365
2366         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2367         /*
2368          * For a file-backed vma, someone could have truncated or otherwise
2369          * invalidated this page.  If unmap_mapping_range got called,
2370          * retry getting the page.
2371          */
2372         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2373                 pte_unmap_unlock(page_table, ptl);
2374                 page_cache_release(new_page);
2375                 cond_resched();
2376                 sequence = mapping->truncate_count;
2377                 smp_rmb();
2378                 goto retry;
2379         }
2380
2381         /*
2382          * This silly early PAGE_DIRTY setting removes a race
2383          * due to the bad i386 page protection. But it's valid
2384          * for other architectures too.
2385          *
2386          * Note that if write_access is true, we either now have
2387          * an exclusive copy of the page, or this is a shared mapping,
2388          * so we can make it writable and dirty to avoid having to
2389          * handle that later.
2390          */
2391         /* Only go through if we didn't race with anybody else... */
2392         if (pte_none(*page_table)) {
2393                 flush_icache_page(vma, new_page);
2394                 entry = mk_pte(new_page, vma->vm_page_prot);
2395                 if (write_access)
2396                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2397                 set_pte_at(mm, address, page_table, entry);
2398                 if (anon) {
2399                         inc_mm_counter(mm, anon_rss);
2400                         lru_cache_add_active(new_page);
2401                         page_add_new_anon_rmap(new_page, vma, address);
2402                 } else {
2403                         inc_mm_counter(mm, file_rss);
2404                         page_add_file_rmap(new_page);
2405                         if (write_access) {
2406                                 dirty_page = new_page;
2407                                 get_page(dirty_page);
2408                         }
2409                 }
2410         } else {
2411                 /* One of our sibling threads was faster, back out. */
2412                 page_cache_release(new_page);
2413                 goto unlock;
2414         }
2415
2416         /* no need to invalidate: a not-present page shouldn't be cached */
2417         update_mmu_cache(vma, address, entry);
2418         lazy_mmu_prot_update(entry);
2419 unlock:
2420         pte_unmap_unlock(page_table, ptl);
2421         if (dirty_page) {
2422                 set_page_dirty_balance(dirty_page);
2423                 put_page(dirty_page);
2424         }
2425         return ret;
2426 oom:
2427         page_cache_release(new_page);
2428         return VM_FAULT_OOM;
2429 }
2430
2431 /*
2432  * do_no_pfn() tries to create a new page mapping for a page without
2433  * a struct_page backing it
2434  *
2435  * As this is called only for pages that do not currently exist, we
2436  * do not need to flush old virtual caches or the TLB.
2437  *
2438  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2439  * but allow concurrent faults), and pte mapped but not yet locked.
2440  * We return with mmap_sem still held, but pte unmapped and unlocked.
2441  *
2442  * It is expected that the ->nopfn handler always returns the same pfn
2443  * for a given virtual mapping.
2444  *
2445  * Mark this `noinline' to prevent it from bloating the main pagefault code.
2446  */
2447 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2448                      unsigned long address, pte_t *page_table, pmd_t *pmd,
2449                      int write_access)
2450 {
2451         spinlock_t *ptl;
2452         pte_t entry;
2453         unsigned long pfn;
2454         int ret = VM_FAULT_MINOR;
2455
2456         pte_unmap(page_table);
2457         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
2458         BUG_ON(is_cow_mapping(vma->vm_flags));
2459
2460         pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2461         if (unlikely(pfn == NOPFN_OOM))
2462                 return VM_FAULT_OOM;
2463         else if (unlikely(pfn == NOPFN_SIGBUS))
2464                 return VM_FAULT_SIGBUS;
2465         else if (unlikely(pfn == NOPFN_REFAULT))
2466                 return VM_FAULT_MINOR;
2467
2468         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2469
2470         /* Only go through if we didn't race with anybody else... */
2471         if (pte_none(*page_table)) {
2472                 entry = pfn_pte(pfn, vma->vm_page_prot);
2473                 if (write_access)
2474                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2475                 set_pte_at(mm, address, page_table, entry);
2476         }
2477         pte_unmap_unlock(page_table, ptl);
2478         return ret;
2479 }
2480
2481 /*
2482  * Fault of a previously existing named mapping. Repopulate the pte
2483  * from the encoded file_pte if possible. This enables swappable
2484  * nonlinear vmas.
2485  *
2486  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2487  * but allow concurrent faults), and pte mapped but not yet locked.
2488  * We return with mmap_sem still held, but pte unmapped and unlocked.
2489  */
2490 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2491                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2492                 int write_access, pte_t orig_pte)
2493 {
2494         pgoff_t pgoff;
2495         int err;
2496
2497         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2498                 return VM_FAULT_MINOR;
2499
2500         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2501                 /*
2502                  * Page table corrupted: show pte and kill process.
2503                  */
2504                 print_bad_pte(vma, orig_pte, address);
2505                 return VM_FAULT_OOM;
2506         }
2507         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2508
2509         pgoff = pte_to_pgoff(orig_pte);
2510         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2511                                         vma->vm_page_prot, pgoff, 0);
2512         if (err == -ENOMEM)
2513                 return VM_FAULT_OOM;
2514         if (err)
2515                 return VM_FAULT_SIGBUS;
2516         return VM_FAULT_MAJOR;
2517 }
2518
2519 /*
2520  * These routines also need to handle stuff like marking pages dirty
2521  * and/or accessed for architectures that don't do it in hardware (most
2522  * RISC architectures).  The early dirtying is also good on the i386.
2523  *
2524  * There is also a hook called "update_mmu_cache()" that architectures
2525  * with external mmu caches can use to update those (ie the Sparc or
2526  * PowerPC hashed page tables that act as extended TLBs).
2527  *
2528  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2529  * but allow concurrent faults), and pte mapped but not yet locked.
2530  * We return with mmap_sem still held, but pte unmapped and unlocked.
2531  */
2532 static inline int handle_pte_fault(struct mm_struct *mm,
2533                 struct vm_area_struct *vma, unsigned long address,
2534                 pte_t *pte, pmd_t *pmd, int write_access)
2535 {
2536         pte_t entry;
2537         spinlock_t *ptl;
2538
2539         entry = *pte;
2540         if (!pte_present(entry)) {
2541                 if (pte_none(entry)) {
2542                         if (vma->vm_ops) {
2543                                 if (vma->vm_ops->nopage)
2544                                         return do_no_page(mm, vma, address,
2545                                                           pte, pmd,
2546                                                           write_access);
2547                                 if (unlikely(vma->vm_ops->nopfn))
2548                                         return do_no_pfn(mm, vma, address, pte,
2549                                                          pmd, write_access);
2550                         }
2551                         return do_anonymous_page(mm, vma, address,
2552                                                  pte, pmd, write_access);
2553                 }
2554                 if (pte_file(entry))
2555                         return do_file_page(mm, vma, address,
2556                                         pte, pmd, write_access, entry);
2557                 return do_swap_page(mm, vma, address,
2558                                         pte, pmd, write_access, entry);
2559         }
2560
2561         ptl = pte_lockptr(mm, pmd);
2562         spin_lock(ptl);
2563         if (unlikely(!pte_same(*pte, entry)))
2564                 goto unlock;
2565         if (write_access) {
2566                 if (!pte_write(entry))
2567                         return do_wp_page(mm, vma, address,
2568                                         pte, pmd, ptl, entry);
2569                 entry = pte_mkdirty(entry);
2570         }
2571         entry = pte_mkyoung(entry);
2572         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2573                 update_mmu_cache(vma, address, entry);
2574                 lazy_mmu_prot_update(entry);
2575         } else {
2576                 /*
2577                  * This is needed only for protection faults but the arch code
2578                  * is not yet telling us if this is a protection fault or not.
2579                  * This still avoids useless tlb flushes for .text page faults
2580                  * with threads.
2581                  */
2582                 if (write_access)
2583                         flush_tlb_page(vma, address);
2584         }
2585 unlock:
2586         pte_unmap_unlock(pte, ptl);
2587         return VM_FAULT_MINOR;
2588 }
2589
2590 /*
2591  * By the time we get here, we already hold the mm semaphore
2592  */
2593 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2594                 unsigned long address, int write_access)
2595 {
2596         pgd_t *pgd;
2597         pud_t *pud;
2598         pmd_t *pmd;
2599         pte_t *pte;
2600
2601         __set_current_state(TASK_RUNNING);
2602
2603         count_vm_event(PGFAULT);
2604
2605         if (unlikely(is_vm_hugetlb_page(vma)))
2606                 return hugetlb_fault(mm, vma, address, write_access);
2607
2608         pgd = pgd_offset(mm, address);
2609         pud = pud_alloc(mm, pgd, address);
2610         if (!pud)
2611                 return VM_FAULT_OOM;
2612         pmd = pmd_alloc(mm, pud, address);
2613         if (!pmd)
2614                 return VM_FAULT_OOM;
2615         pte = pte_alloc_map(mm, pmd, address);
2616         if (!pte)
2617                 return VM_FAULT_OOM;
2618
2619         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2620 }
2621
2622 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2623
2624 #ifndef __PAGETABLE_PUD_FOLDED
2625 /*
2626  * Allocate page upper directory.
2627  * We've already handled the fast-path in-line.
2628  */
2629 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2630 {
2631         pud_t *new = pud_alloc_one(mm, address);
2632         if (!new)
2633                 return -ENOMEM;
2634
2635         spin_lock(&mm->page_table_lock);
2636         if (pgd_present(*pgd))          /* Another has populated it */
2637                 pud_free(new);
2638         else
2639                 pgd_populate(mm, pgd, new);
2640         spin_unlock(&mm->page_table_lock);
2641         return 0;
2642 }
2643 #endif /* __PAGETABLE_PUD_FOLDED */
2644
2645 #ifndef __PAGETABLE_PMD_FOLDED
2646 /*
2647  * Allocate page middle directory.
2648  * We've already handled the fast-path in-line.
2649  */
2650 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2651 {
2652         pmd_t *new = pmd_alloc_one(mm, address);
2653         if (!new)
2654                 return -ENOMEM;
2655
2656         spin_lock(&mm->page_table_lock);
2657 #ifndef __ARCH_HAS_4LEVEL_HACK
2658         if (pud_present(*pud))          /* Another has populated it */
2659                 pmd_free(new);
2660         else
2661                 pud_populate(mm, pud, new);
2662 #else
2663         if (pgd_present(*pud))          /* Another has populated it */
2664                 pmd_free(new);
2665         else
2666                 pgd_populate(mm, pud, new);
2667 #endif /* __ARCH_HAS_4LEVEL_HACK */
2668         spin_unlock(&mm->page_table_lock);
2669         return 0;
2670 }
2671 #endif /* __PAGETABLE_PMD_FOLDED */
2672
2673 int make_pages_present(unsigned long addr, unsigned long end)
2674 {
2675         int ret, len, write;
2676         struct vm_area_struct * vma;
2677
2678         vma = find_vma(current->mm, addr);
2679         if (!vma)
2680                 return -1;
2681         write = (vma->vm_flags & VM_WRITE) != 0;
2682         BUG_ON(addr >= end);
2683         BUG_ON(end > vma->vm_end);
2684         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2685         ret = get_user_pages(current, current->mm, addr,
2686                         len, write, 0, NULL, NULL);
2687         if (ret < 0)
2688                 return ret;
2689         return ret == len ? 0 : -1;
2690 }
2691
2692 /* 
2693  * Map a vmalloc()-space virtual address to the physical page.
2694  */
2695 struct page * vmalloc_to_page(void * vmalloc_addr)
2696 {
2697         unsigned long addr = (unsigned long) vmalloc_addr;
2698         struct page *page = NULL;
2699         pgd_t *pgd = pgd_offset_k(addr);
2700         pud_t *pud;
2701         pmd_t *pmd;
2702         pte_t *ptep, pte;
2703   
2704         if (!pgd_none(*pgd)) {
2705                 pud = pud_offset(pgd, addr);
2706                 if (!pud_none(*pud)) {
2707                         pmd = pmd_offset(pud, addr);
2708                         if (!pmd_none(*pmd)) {
2709                                 ptep = pte_offset_map(pmd, addr);
2710                                 pte = *ptep;
2711                                 if (pte_present(pte))
2712                                         page = pte_page(pte);
2713                                 pte_unmap(ptep);
2714                         }
2715                 }
2716         }
2717         return page;
2718 }
2719
2720 EXPORT_SYMBOL(vmalloc_to_page);
2721
2722 /*
2723  * Map a vmalloc()-space virtual address to the physical page frame number.
2724  */
2725 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2726 {
2727         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2728 }
2729
2730 EXPORT_SYMBOL(vmalloc_to_pfn);
2731
2732 #if !defined(__HAVE_ARCH_GATE_AREA)
2733
2734 #if defined(AT_SYSINFO_EHDR)
2735 static struct vm_area_struct gate_vma;
2736
2737 static int __init gate_vma_init(void)
2738 {
2739         gate_vma.vm_mm = NULL;
2740         gate_vma.vm_start = FIXADDR_USER_START;
2741         gate_vma.vm_end = FIXADDR_USER_END;
2742         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2743         gate_vma.vm_page_prot = __P101;
2744         /*
2745          * Make sure the vDSO gets into every core dump.
2746          * Dumping its contents makes post-mortem fully interpretable later
2747          * without matching up the same kernel and hardware config to see
2748          * what PC values meant.
2749          */
2750         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2751         return 0;
2752 }
2753 __initcall(gate_vma_init);
2754 #endif
2755
2756 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2757 {
2758 #ifdef AT_SYSINFO_EHDR
2759         return &gate_vma;
2760 #else
2761         return NULL;
2762 #endif
2763 }
2764
2765 int in_gate_area_no_task(unsigned long addr)
2766 {
2767 #ifdef AT_SYSINFO_EHDR
2768         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2769                 return 1;
2770 #endif
2771         return 0;
2772 }
2773
2774 #endif  /* __HAVE_ARCH_GATE_AREA */
2775
2776 /*
2777  * Access another process' address space.
2778  * Source/target buffer must be kernel space,
2779  * Do not walk the page table directly, use get_user_pages
2780  */
2781 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2782 {
2783         struct mm_struct *mm;
2784         struct vm_area_struct *vma;
2785         struct page *page;
2786         void *old_buf = buf;
2787
2788         mm = get_task_mm(tsk);
2789         if (!mm)
2790                 return 0;
2791
2792         down_read(&mm->mmap_sem);
2793         /* ignore errors, just check how much was sucessfully transfered */
2794         while (len) {
2795                 int bytes, ret, offset;
2796                 void *maddr;
2797
2798                 ret = get_user_pages(tsk, mm, addr, 1,
2799                                 write, 1, &page, &vma);
2800                 if (ret <= 0)
2801                         break;
2802
2803                 bytes = len;
2804                 offset = addr & (PAGE_SIZE-1);
2805                 if (bytes > PAGE_SIZE-offset)
2806                         bytes = PAGE_SIZE-offset;
2807
2808                 maddr = kmap(page);
2809                 if (write) {
2810                         copy_to_user_page(vma, page, addr,
2811                                           maddr + offset, buf, bytes);
2812                         set_page_dirty_lock(page);
2813                 } else {
2814                         copy_from_user_page(vma, page, addr,
2815                                             buf, maddr + offset, bytes);
2816                 }
2817                 kunmap(page);
2818                 page_cache_release(page);
2819                 len -= bytes;
2820                 buf += bytes;
2821                 addr += bytes;
2822         }
2823         up_read(&mm->mmap_sem);
2824         mmput(mm);
2825
2826         return buf - old_buf;
2827 }