[PATCH] paravirt: pte clear not present
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/delayacct.h>
51 #include <linux/init.h>
52 #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);
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
510         do {
511                 /*
512                  * We are holding two locks at this point - either of them
513                  * could generate latencies in another task on another CPU.
514                  */
515                 if (progress >= 32) {
516                         progress = 0;
517                         if (need_resched() ||
518                             need_lockbreak(src_ptl) ||
519                             need_lockbreak(dst_ptl))
520                                 break;
521                 }
522                 if (pte_none(*src_pte)) {
523                         progress++;
524                         continue;
525                 }
526                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
527                 progress += 8;
528         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
529
530         spin_unlock(src_ptl);
531         pte_unmap_nested(src_pte - 1);
532         add_mm_rss(dst_mm, rss[0], rss[1]);
533         pte_unmap_unlock(dst_pte - 1, dst_ptl);
534         cond_resched();
535         if (addr != end)
536                 goto again;
537         return 0;
538 }
539
540 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
541                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
542                 unsigned long addr, unsigned long end)
543 {
544         pmd_t *src_pmd, *dst_pmd;
545         unsigned long next;
546
547         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
548         if (!dst_pmd)
549                 return -ENOMEM;
550         src_pmd = pmd_offset(src_pud, addr);
551         do {
552                 next = pmd_addr_end(addr, end);
553                 if (pmd_none_or_clear_bad(src_pmd))
554                         continue;
555                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
556                                                 vma, addr, next))
557                         return -ENOMEM;
558         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
559         return 0;
560 }
561
562 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
563                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
564                 unsigned long addr, unsigned long end)
565 {
566         pud_t *src_pud, *dst_pud;
567         unsigned long next;
568
569         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
570         if (!dst_pud)
571                 return -ENOMEM;
572         src_pud = pud_offset(src_pgd, addr);
573         do {
574                 next = pud_addr_end(addr, end);
575                 if (pud_none_or_clear_bad(src_pud))
576                         continue;
577                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
578                                                 vma, addr, next))
579                         return -ENOMEM;
580         } while (dst_pud++, src_pud++, addr = next, addr != end);
581         return 0;
582 }
583
584 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
585                 struct vm_area_struct *vma)
586 {
587         pgd_t *src_pgd, *dst_pgd;
588         unsigned long next;
589         unsigned long addr = vma->vm_start;
590         unsigned long end = vma->vm_end;
591
592         /*
593          * Don't copy ptes where a page fault will fill them correctly.
594          * Fork becomes much lighter when there are big shared or private
595          * readonly mappings. The tradeoff is that copy_page_range is more
596          * efficient than faulting.
597          */
598         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
599                 if (!vma->anon_vma)
600                         return 0;
601         }
602
603         if (is_vm_hugetlb_page(vma))
604                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
605
606         dst_pgd = pgd_offset(dst_mm, addr);
607         src_pgd = pgd_offset(src_mm, addr);
608         do {
609                 next = pgd_addr_end(addr, end);
610                 if (pgd_none_or_clear_bad(src_pgd))
611                         continue;
612                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
613                                                 vma, addr, next))
614                         return -ENOMEM;
615         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
616         return 0;
617 }
618
619 static unsigned long zap_pte_range(struct mmu_gather *tlb,
620                                 struct vm_area_struct *vma, pmd_t *pmd,
621                                 unsigned long addr, unsigned long end,
622                                 long *zap_work, struct zap_details *details)
623 {
624         struct mm_struct *mm = tlb->mm;
625         pte_t *pte;
626         spinlock_t *ptl;
627         int file_rss = 0;
628         int anon_rss = 0;
629
630         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
631         do {
632                 pte_t ptent = *pte;
633                 if (pte_none(ptent)) {
634                         (*zap_work)--;
635                         continue;
636                 }
637
638                 (*zap_work) -= PAGE_SIZE;
639
640                 if (pte_present(ptent)) {
641                         struct page *page;
642
643                         page = vm_normal_page(vma, addr, ptent);
644                         if (unlikely(details) && page) {
645                                 /*
646                                  * unmap_shared_mapping_pages() wants to
647                                  * invalidate cache without truncating:
648                                  * unmap shared but keep private pages.
649                                  */
650                                 if (details->check_mapping &&
651                                     details->check_mapping != page->mapping)
652                                         continue;
653                                 /*
654                                  * Each page->index must be checked when
655                                  * invalidating or truncating nonlinear.
656                                  */
657                                 if (details->nonlinear_vma &&
658                                     (page->index < details->first_index ||
659                                      page->index > details->last_index))
660                                         continue;
661                         }
662                         ptent = ptep_get_and_clear_full(mm, addr, pte,
663                                                         tlb->fullmm);
664                         tlb_remove_tlb_entry(tlb, pte, addr);
665                         if (unlikely(!page))
666                                 continue;
667                         if (unlikely(details) && details->nonlinear_vma
668                             && linear_page_index(details->nonlinear_vma,
669                                                 addr) != page->index)
670                                 set_pte_at(mm, addr, pte,
671                                            pgoff_to_pte(page->index));
672                         if (PageAnon(page))
673                                 anon_rss--;
674                         else {
675                                 if (pte_dirty(ptent))
676                                         set_page_dirty(page);
677                                 if (pte_young(ptent))
678                                         mark_page_accessed(page);
679                                 file_rss--;
680                         }
681                         page_remove_rmap(page);
682                         tlb_remove_page(tlb, page);
683                         continue;
684                 }
685                 /*
686                  * If details->check_mapping, we leave swap entries;
687                  * if details->nonlinear_vma, we leave file entries.
688                  */
689                 if (unlikely(details))
690                         continue;
691                 if (!pte_file(ptent))
692                         free_swap_and_cache(pte_to_swp_entry(ptent));
693                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
694         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
695
696         add_mm_rss(mm, file_rss, anon_rss);
697         pte_unmap_unlock(pte - 1, ptl);
698
699         return addr;
700 }
701
702 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
703                                 struct vm_area_struct *vma, pud_t *pud,
704                                 unsigned long addr, unsigned long end,
705                                 long *zap_work, struct zap_details *details)
706 {
707         pmd_t *pmd;
708         unsigned long next;
709
710         pmd = pmd_offset(pud, addr);
711         do {
712                 next = pmd_addr_end(addr, end);
713                 if (pmd_none_or_clear_bad(pmd)) {
714                         (*zap_work)--;
715                         continue;
716                 }
717                 next = zap_pte_range(tlb, vma, pmd, addr, next,
718                                                 zap_work, details);
719         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
720
721         return addr;
722 }
723
724 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
725                                 struct vm_area_struct *vma, pgd_t *pgd,
726                                 unsigned long addr, unsigned long end,
727                                 long *zap_work, struct zap_details *details)
728 {
729         pud_t *pud;
730         unsigned long next;
731
732         pud = pud_offset(pgd, addr);
733         do {
734                 next = pud_addr_end(addr, end);
735                 if (pud_none_or_clear_bad(pud)) {
736                         (*zap_work)--;
737                         continue;
738                 }
739                 next = zap_pmd_range(tlb, vma, pud, addr, next,
740                                                 zap_work, details);
741         } while (pud++, addr = next, (addr != end && *zap_work > 0));
742
743         return addr;
744 }
745
746 static unsigned long unmap_page_range(struct mmu_gather *tlb,
747                                 struct vm_area_struct *vma,
748                                 unsigned long addr, unsigned long end,
749                                 long *zap_work, struct zap_details *details)
750 {
751         pgd_t *pgd;
752         unsigned long next;
753
754         if (details && !details->check_mapping && !details->nonlinear_vma)
755                 details = NULL;
756
757         BUG_ON(addr >= end);
758         tlb_start_vma(tlb, vma);
759         pgd = pgd_offset(vma->vm_mm, addr);
760         do {
761                 next = pgd_addr_end(addr, end);
762                 if (pgd_none_or_clear_bad(pgd)) {
763                         (*zap_work)--;
764                         continue;
765                 }
766                 next = zap_pud_range(tlb, vma, pgd, addr, next,
767                                                 zap_work, details);
768         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
769         tlb_end_vma(tlb, vma);
770
771         return addr;
772 }
773
774 #ifdef CONFIG_PREEMPT
775 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
776 #else
777 /* No preempt: go for improved straight-line efficiency */
778 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
779 #endif
780
781 /**
782  * unmap_vmas - unmap a range of memory covered by a list of vma's
783  * @tlbp: address of the caller's struct mmu_gather
784  * @vma: the starting vma
785  * @start_addr: virtual address at which to start unmapping
786  * @end_addr: virtual address at which to end unmapping
787  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
788  * @details: details of nonlinear truncation or shared cache invalidation
789  *
790  * Returns the end address of the unmapping (restart addr if interrupted).
791  *
792  * Unmap all pages in the vma list.
793  *
794  * We aim to not hold locks for too long (for scheduling latency reasons).
795  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
796  * return the ending mmu_gather to the caller.
797  *
798  * Only addresses between `start' and `end' will be unmapped.
799  *
800  * The VMA list must be sorted in ascending virtual address order.
801  *
802  * unmap_vmas() assumes that the caller will flush the whole unmapped address
803  * range after unmap_vmas() returns.  So the only responsibility here is to
804  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
805  * drops the lock and schedules.
806  */
807 unsigned long unmap_vmas(struct mmu_gather **tlbp,
808                 struct vm_area_struct *vma, unsigned long start_addr,
809                 unsigned long end_addr, unsigned long *nr_accounted,
810                 struct zap_details *details)
811 {
812         long zap_work = ZAP_BLOCK_SIZE;
813         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
814         int tlb_start_valid = 0;
815         unsigned long start = start_addr;
816         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
817         int fullmm = (*tlbp)->fullmm;
818
819         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
820                 unsigned long end;
821
822                 start = max(vma->vm_start, start_addr);
823                 if (start >= vma->vm_end)
824                         continue;
825                 end = min(vma->vm_end, end_addr);
826                 if (end <= vma->vm_start)
827                         continue;
828
829                 if (vma->vm_flags & VM_ACCOUNT)
830                         *nr_accounted += (end - start) >> PAGE_SHIFT;
831
832                 while (start != end) {
833                         if (!tlb_start_valid) {
834                                 tlb_start = start;
835                                 tlb_start_valid = 1;
836                         }
837
838                         if (unlikely(is_vm_hugetlb_page(vma))) {
839                                 unmap_hugepage_range(vma, start, end);
840                                 zap_work -= (end - start) /
841                                                 (HPAGE_SIZE / PAGE_SIZE);
842                                 start = end;
843                         } else
844                                 start = unmap_page_range(*tlbp, vma,
845                                                 start, end, &zap_work, details);
846
847                         if (zap_work > 0) {
848                                 BUG_ON(start != end);
849                                 break;
850                         }
851
852                         tlb_finish_mmu(*tlbp, tlb_start, start);
853
854                         if (need_resched() ||
855                                 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
856                                 if (i_mmap_lock) {
857                                         *tlbp = NULL;
858                                         goto out;
859                                 }
860                                 cond_resched();
861                         }
862
863                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
864                         tlb_start_valid = 0;
865                         zap_work = ZAP_BLOCK_SIZE;
866                 }
867         }
868 out:
869         return start;   /* which is now the end (or restart) address */
870 }
871
872 /**
873  * zap_page_range - remove user pages in a given range
874  * @vma: vm_area_struct holding the applicable pages
875  * @address: starting address of pages to zap
876  * @size: number of bytes to zap
877  * @details: details of nonlinear truncation or shared cache invalidation
878  */
879 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
880                 unsigned long size, struct zap_details *details)
881 {
882         struct mm_struct *mm = vma->vm_mm;
883         struct mmu_gather *tlb;
884         unsigned long end = address + size;
885         unsigned long nr_accounted = 0;
886
887         lru_add_drain();
888         tlb = tlb_gather_mmu(mm, 0);
889         update_hiwater_rss(mm);
890         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
891         if (tlb)
892                 tlb_finish_mmu(tlb, address, end);
893         return end;
894 }
895
896 /*
897  * Do a quick page-table lookup for a single page.
898  */
899 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
900                         unsigned int flags)
901 {
902         pgd_t *pgd;
903         pud_t *pud;
904         pmd_t *pmd;
905         pte_t *ptep, pte;
906         spinlock_t *ptl;
907         struct page *page;
908         struct mm_struct *mm = vma->vm_mm;
909
910         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
911         if (!IS_ERR(page)) {
912                 BUG_ON(flags & FOLL_GET);
913                 goto out;
914         }
915
916         page = NULL;
917         pgd = pgd_offset(mm, address);
918         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
919                 goto no_page_table;
920
921         pud = pud_offset(pgd, address);
922         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
923                 goto no_page_table;
924         
925         pmd = pmd_offset(pud, address);
926         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
927                 goto no_page_table;
928
929         if (pmd_huge(*pmd)) {
930                 BUG_ON(flags & FOLL_GET);
931                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
932                 goto out;
933         }
934
935         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
936         if (!ptep)
937                 goto out;
938
939         pte = *ptep;
940         if (!pte_present(pte))
941                 goto unlock;
942         if ((flags & FOLL_WRITE) && !pte_write(pte))
943                 goto unlock;
944         page = vm_normal_page(vma, address, pte);
945         if (unlikely(!page))
946                 goto unlock;
947
948         if (flags & FOLL_GET)
949                 get_page(page);
950         if (flags & FOLL_TOUCH) {
951                 if ((flags & FOLL_WRITE) &&
952                     !pte_dirty(pte) && !PageDirty(page))
953                         set_page_dirty(page);
954                 mark_page_accessed(page);
955         }
956 unlock:
957         pte_unmap_unlock(ptep, ptl);
958 out:
959         return page;
960
961 no_page_table:
962         /*
963          * When core dumping an enormous anonymous area that nobody
964          * has touched so far, we don't want to allocate page tables.
965          */
966         if (flags & FOLL_ANON) {
967                 page = ZERO_PAGE(address);
968                 if (flags & FOLL_GET)
969                         get_page(page);
970                 BUG_ON(flags & FOLL_WRITE);
971         }
972         return page;
973 }
974
975 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
976                 unsigned long start, int len, int write, int force,
977                 struct page **pages, struct vm_area_struct **vmas)
978 {
979         int i;
980         unsigned int vm_flags;
981
982         /* 
983          * Require read or write permissions.
984          * If 'force' is set, we only require the "MAY" flags.
985          */
986         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
987         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
988         i = 0;
989
990         do {
991                 struct vm_area_struct *vma;
992                 unsigned int foll_flags;
993
994                 vma = find_extend_vma(mm, start);
995                 if (!vma && in_gate_area(tsk, start)) {
996                         unsigned long pg = start & PAGE_MASK;
997                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
998                         pgd_t *pgd;
999                         pud_t *pud;
1000                         pmd_t *pmd;
1001                         pte_t *pte;
1002                         if (write) /* user gate pages are read-only */
1003                                 return i ? : -EFAULT;
1004                         if (pg > TASK_SIZE)
1005                                 pgd = pgd_offset_k(pg);
1006                         else
1007                                 pgd = pgd_offset_gate(mm, pg);
1008                         BUG_ON(pgd_none(*pgd));
1009                         pud = pud_offset(pgd, pg);
1010                         BUG_ON(pud_none(*pud));
1011                         pmd = pmd_offset(pud, pg);
1012                         if (pmd_none(*pmd))
1013                                 return i ? : -EFAULT;
1014                         pte = pte_offset_map(pmd, pg);
1015                         if (pte_none(*pte)) {
1016                                 pte_unmap(pte);
1017                                 return i ? : -EFAULT;
1018                         }
1019                         if (pages) {
1020                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1021                                 pages[i] = page;
1022                                 if (page)
1023                                         get_page(page);
1024                         }
1025                         pte_unmap(pte);
1026                         if (vmas)
1027                                 vmas[i] = gate_vma;
1028                         i++;
1029                         start += PAGE_SIZE;
1030                         len--;
1031                         continue;
1032                 }
1033
1034                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1035                                 || !(vm_flags & vma->vm_flags))
1036                         return i ? : -EFAULT;
1037
1038                 if (is_vm_hugetlb_page(vma)) {
1039                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1040                                                 &start, &len, i);
1041                         continue;
1042                 }
1043
1044                 foll_flags = FOLL_TOUCH;
1045                 if (pages)
1046                         foll_flags |= FOLL_GET;
1047                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1048                     (!vma->vm_ops || !vma->vm_ops->nopage))
1049                         foll_flags |= FOLL_ANON;
1050
1051                 do {
1052                         struct page *page;
1053
1054                         if (write)
1055                                 foll_flags |= FOLL_WRITE;
1056
1057                         cond_resched();
1058                         while (!(page = follow_page(vma, start, foll_flags))) {
1059                                 int ret;
1060                                 ret = __handle_mm_fault(mm, vma, start,
1061                                                 foll_flags & FOLL_WRITE);
1062                                 /*
1063                                  * The VM_FAULT_WRITE bit tells us that do_wp_page has
1064                                  * broken COW when necessary, even if maybe_mkwrite
1065                                  * decided not to set pte_write. We can thus safely do
1066                                  * subsequent page lookups as if they were reads.
1067                                  */
1068                                 if (ret & VM_FAULT_WRITE)
1069                                         foll_flags &= ~FOLL_WRITE;
1070                                 
1071                                 switch (ret & ~VM_FAULT_WRITE) {
1072                                 case VM_FAULT_MINOR:
1073                                         tsk->min_flt++;
1074                                         break;
1075                                 case VM_FAULT_MAJOR:
1076                                         tsk->maj_flt++;
1077                                         break;
1078                                 case VM_FAULT_SIGBUS:
1079                                         return i ? i : -EFAULT;
1080                                 case VM_FAULT_OOM:
1081                                         return i ? i : -ENOMEM;
1082                                 default:
1083                                         BUG();
1084                                 }
1085                         }
1086                         if (pages) {
1087                                 pages[i] = page;
1088
1089                                 flush_anon_page(page, start);
1090                                 flush_dcache_page(page);
1091                         }
1092                         if (vmas)
1093                                 vmas[i] = vma;
1094                         i++;
1095                         start += PAGE_SIZE;
1096                         len--;
1097                 } while (len && start < vma->vm_end);
1098         } while (len);
1099         return i;
1100 }
1101 EXPORT_SYMBOL(get_user_pages);
1102
1103 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1104                         unsigned long addr, unsigned long end, pgprot_t prot)
1105 {
1106         pte_t *pte;
1107         spinlock_t *ptl;
1108
1109         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1110         if (!pte)
1111                 return -ENOMEM;
1112         do {
1113                 struct page *page = ZERO_PAGE(addr);
1114                 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1115                 page_cache_get(page);
1116                 page_add_file_rmap(page);
1117                 inc_mm_counter(mm, file_rss);
1118                 BUG_ON(!pte_none(*pte));
1119                 set_pte_at(mm, addr, pte, zero_pte);
1120         } while (pte++, addr += PAGE_SIZE, addr != end);
1121         pte_unmap_unlock(pte - 1, ptl);
1122         return 0;
1123 }
1124
1125 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1126                         unsigned long addr, unsigned long end, pgprot_t prot)
1127 {
1128         pmd_t *pmd;
1129         unsigned long next;
1130
1131         pmd = pmd_alloc(mm, pud, addr);
1132         if (!pmd)
1133                 return -ENOMEM;
1134         do {
1135                 next = pmd_addr_end(addr, end);
1136                 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1137                         return -ENOMEM;
1138         } while (pmd++, addr = next, addr != end);
1139         return 0;
1140 }
1141
1142 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1143                         unsigned long addr, unsigned long end, pgprot_t prot)
1144 {
1145         pud_t *pud;
1146         unsigned long next;
1147
1148         pud = pud_alloc(mm, pgd, addr);
1149         if (!pud)
1150                 return -ENOMEM;
1151         do {
1152                 next = pud_addr_end(addr, end);
1153                 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1154                         return -ENOMEM;
1155         } while (pud++, addr = next, addr != end);
1156         return 0;
1157 }
1158
1159 int zeromap_page_range(struct vm_area_struct *vma,
1160                         unsigned long addr, unsigned long size, pgprot_t prot)
1161 {
1162         pgd_t *pgd;
1163         unsigned long next;
1164         unsigned long end = addr + size;
1165         struct mm_struct *mm = vma->vm_mm;
1166         int err;
1167
1168         BUG_ON(addr >= end);
1169         pgd = pgd_offset(mm, addr);
1170         flush_cache_range(vma, addr, end);
1171         do {
1172                 next = pgd_addr_end(addr, end);
1173                 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1174                 if (err)
1175                         break;
1176         } while (pgd++, addr = next, addr != end);
1177         return err;
1178 }
1179
1180 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1181 {
1182         pgd_t * pgd = pgd_offset(mm, addr);
1183         pud_t * pud = pud_alloc(mm, pgd, addr);
1184         if (pud) {
1185                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1186                 if (pmd)
1187                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1188         }
1189         return NULL;
1190 }
1191
1192 /*
1193  * This is the old fallback for page remapping.
1194  *
1195  * For historical reasons, it only allows reserved pages. Only
1196  * old drivers should use this, and they needed to mark their
1197  * pages reserved for the old functions anyway.
1198  */
1199 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1200 {
1201         int retval;
1202         pte_t *pte;
1203         spinlock_t *ptl;  
1204
1205         retval = -EINVAL;
1206         if (PageAnon(page))
1207                 goto out;
1208         retval = -ENOMEM;
1209         flush_dcache_page(page);
1210         pte = get_locked_pte(mm, addr, &ptl);
1211         if (!pte)
1212                 goto out;
1213         retval = -EBUSY;
1214         if (!pte_none(*pte))
1215                 goto out_unlock;
1216
1217         /* Ok, finally just insert the thing.. */
1218         get_page(page);
1219         inc_mm_counter(mm, file_rss);
1220         page_add_file_rmap(page);
1221         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1222
1223         retval = 0;
1224 out_unlock:
1225         pte_unmap_unlock(pte, ptl);
1226 out:
1227         return retval;
1228 }
1229
1230 /**
1231  * vm_insert_page - insert single page into user vma
1232  * @vma: user vma to map to
1233  * @addr: target user address of this page
1234  * @page: source kernel page
1235  *
1236  * This allows drivers to insert individual pages they've allocated
1237  * into a user vma.
1238  *
1239  * The page has to be a nice clean _individual_ kernel allocation.
1240  * If you allocate a compound page, you need to have marked it as
1241  * such (__GFP_COMP), or manually just split the page up yourself
1242  * (see split_page()).
1243  *
1244  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1245  * took an arbitrary page protection parameter. This doesn't allow
1246  * that. Your vma protection will have to be set up correctly, which
1247  * means that if you want a shared writable mapping, you'd better
1248  * ask for a shared writable mapping!
1249  *
1250  * The page does not need to be reserved.
1251  */
1252 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1253 {
1254         if (addr < vma->vm_start || addr >= vma->vm_end)
1255                 return -EFAULT;
1256         if (!page_count(page))
1257                 return -EINVAL;
1258         vma->vm_flags |= VM_INSERTPAGE;
1259         return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1260 }
1261 EXPORT_SYMBOL(vm_insert_page);
1262
1263 /*
1264  * maps a range of physical memory into the requested pages. the old
1265  * mappings are removed. any references to nonexistent pages results
1266  * in null mappings (currently treated as "copy-on-access")
1267  */
1268 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1269                         unsigned long addr, unsigned long end,
1270                         unsigned long pfn, pgprot_t prot)
1271 {
1272         pte_t *pte;
1273         spinlock_t *ptl;
1274
1275         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1276         if (!pte)
1277                 return -ENOMEM;
1278         do {
1279                 BUG_ON(!pte_none(*pte));
1280                 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1281                 pfn++;
1282         } while (pte++, addr += PAGE_SIZE, addr != end);
1283         pte_unmap_unlock(pte - 1, ptl);
1284         return 0;
1285 }
1286
1287 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1288                         unsigned long addr, unsigned long end,
1289                         unsigned long pfn, pgprot_t prot)
1290 {
1291         pmd_t *pmd;
1292         unsigned long next;
1293
1294         pfn -= addr >> PAGE_SHIFT;
1295         pmd = pmd_alloc(mm, pud, addr);
1296         if (!pmd)
1297                 return -ENOMEM;
1298         do {
1299                 next = pmd_addr_end(addr, end);
1300                 if (remap_pte_range(mm, pmd, addr, next,
1301                                 pfn + (addr >> PAGE_SHIFT), prot))
1302                         return -ENOMEM;
1303         } while (pmd++, addr = next, addr != end);
1304         return 0;
1305 }
1306
1307 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1308                         unsigned long addr, unsigned long end,
1309                         unsigned long pfn, pgprot_t prot)
1310 {
1311         pud_t *pud;
1312         unsigned long next;
1313
1314         pfn -= addr >> PAGE_SHIFT;
1315         pud = pud_alloc(mm, pgd, addr);
1316         if (!pud)
1317                 return -ENOMEM;
1318         do {
1319                 next = pud_addr_end(addr, end);
1320                 if (remap_pmd_range(mm, pud, addr, next,
1321                                 pfn + (addr >> PAGE_SHIFT), prot))
1322                         return -ENOMEM;
1323         } while (pud++, addr = next, addr != end);
1324         return 0;
1325 }
1326
1327 /**
1328  * remap_pfn_range - remap kernel memory to userspace
1329  * @vma: user vma to map to
1330  * @addr: target user address to start at
1331  * @pfn: physical address of kernel memory
1332  * @size: size of map area
1333  * @prot: page protection flags for this mapping
1334  *
1335  *  Note: this is only safe if the mm semaphore is held when called.
1336  */
1337 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1338                     unsigned long pfn, unsigned long size, pgprot_t prot)
1339 {
1340         pgd_t *pgd;
1341         unsigned long next;
1342         unsigned long end = addr + PAGE_ALIGN(size);
1343         struct mm_struct *mm = vma->vm_mm;
1344         int err;
1345
1346         /*
1347          * Physically remapped pages are special. Tell the
1348          * rest of the world about it:
1349          *   VM_IO tells people not to look at these pages
1350          *      (accesses can have side effects).
1351          *   VM_RESERVED is specified all over the place, because
1352          *      in 2.4 it kept swapout's vma scan off this vma; but
1353          *      in 2.6 the LRU scan won't even find its pages, so this
1354          *      flag means no more than count its pages in reserved_vm,
1355          *      and omit it from core dump, even when VM_IO turned off.
1356          *   VM_PFNMAP tells the core MM that the base pages are just
1357          *      raw PFN mappings, and do not have a "struct page" associated
1358          *      with them.
1359          *
1360          * There's a horrible special case to handle copy-on-write
1361          * behaviour that some programs depend on. We mark the "original"
1362          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1363          */
1364         if (is_cow_mapping(vma->vm_flags)) {
1365                 if (addr != vma->vm_start || end != vma->vm_end)
1366                         return -EINVAL;
1367                 vma->vm_pgoff = pfn;
1368         }
1369
1370         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1371
1372         BUG_ON(addr >= end);
1373         pfn -= addr >> PAGE_SHIFT;
1374         pgd = pgd_offset(mm, addr);
1375         flush_cache_range(vma, addr, end);
1376         do {
1377                 next = pgd_addr_end(addr, end);
1378                 err = remap_pud_range(mm, pgd, addr, next,
1379                                 pfn + (addr >> PAGE_SHIFT), prot);
1380                 if (err)
1381                         break;
1382         } while (pgd++, addr = next, addr != end);
1383         return err;
1384 }
1385 EXPORT_SYMBOL(remap_pfn_range);
1386
1387 /*
1388  * handle_pte_fault chooses page fault handler according to an entry
1389  * which was read non-atomically.  Before making any commitment, on
1390  * those architectures or configurations (e.g. i386 with PAE) which
1391  * might give a mix of unmatched parts, do_swap_page and do_file_page
1392  * must check under lock before unmapping the pte and proceeding
1393  * (but do_wp_page is only called after already making such a check;
1394  * and do_anonymous_page and do_no_page can safely check later on).
1395  */
1396 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1397                                 pte_t *page_table, pte_t orig_pte)
1398 {
1399         int same = 1;
1400 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1401         if (sizeof(pte_t) > sizeof(unsigned long)) {
1402                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1403                 spin_lock(ptl);
1404                 same = pte_same(*page_table, orig_pte);
1405                 spin_unlock(ptl);
1406         }
1407 #endif
1408         pte_unmap(page_table);
1409         return same;
1410 }
1411
1412 /*
1413  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1414  * servicing faults for write access.  In the normal case, do always want
1415  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1416  * that do not have writing enabled, when used by access_process_vm.
1417  */
1418 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1419 {
1420         if (likely(vma->vm_flags & VM_WRITE))
1421                 pte = pte_mkwrite(pte);
1422         return pte;
1423 }
1424
1425 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1426 {
1427         /*
1428          * If the source page was a PFN mapping, we don't have
1429          * a "struct page" for it. We do a best-effort copy by
1430          * just copying from the original user address. If that
1431          * fails, we just zero-fill it. Live with it.
1432          */
1433         if (unlikely(!src)) {
1434                 void *kaddr = kmap_atomic(dst, KM_USER0);
1435                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1436
1437                 /*
1438                  * This really shouldn't fail, because the page is there
1439                  * in the page tables. But it might just be unreadable,
1440                  * in which case we just give up and fill the result with
1441                  * zeroes.
1442                  */
1443                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1444                         memset(kaddr, 0, PAGE_SIZE);
1445                 kunmap_atomic(kaddr, KM_USER0);
1446                 return;
1447                 
1448         }
1449         copy_user_highpage(dst, src, va);
1450 }
1451
1452 /*
1453  * This routine handles present pages, when users try to write
1454  * to a shared page. It is done by copying the page to a new address
1455  * and decrementing the shared-page counter for the old page.
1456  *
1457  * Note that this routine assumes that the protection checks have been
1458  * done by the caller (the low-level page fault routine in most cases).
1459  * Thus we can safely just mark it writable once we've done any necessary
1460  * COW.
1461  *
1462  * We also mark the page dirty at this point even though the page will
1463  * change only once the write actually happens. This avoids a few races,
1464  * and potentially makes it more efficient.
1465  *
1466  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1467  * but allow concurrent faults), with pte both mapped and locked.
1468  * We return with mmap_sem still held, but pte unmapped and unlocked.
1469  */
1470 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1471                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1472                 spinlock_t *ptl, pte_t orig_pte)
1473 {
1474         struct page *old_page, *new_page;
1475         pte_t entry;
1476         int reuse = 0, ret = VM_FAULT_MINOR;
1477         struct page *dirty_page = NULL;
1478
1479         old_page = vm_normal_page(vma, address, orig_pte);
1480         if (!old_page)
1481                 goto gotten;
1482
1483         /*
1484          * Take out anonymous pages first, anonymous shared vmas are
1485          * not dirty accountable.
1486          */
1487         if (PageAnon(old_page)) {
1488                 if (!TestSetPageLocked(old_page)) {
1489                         reuse = can_share_swap_page(old_page);
1490                         unlock_page(old_page);
1491                 }
1492         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1493                                         (VM_WRITE|VM_SHARED))) {
1494                 /*
1495                  * Only catch write-faults on shared writable pages,
1496                  * read-only shared pages can get COWed by
1497                  * get_user_pages(.write=1, .force=1).
1498                  */
1499                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1500                         /*
1501                          * Notify the address space that the page is about to
1502                          * become writable so that it can prohibit this or wait
1503                          * for the page to get into an appropriate state.
1504                          *
1505                          * We do this without the lock held, so that it can
1506                          * sleep if it needs to.
1507                          */
1508                         page_cache_get(old_page);
1509                         pte_unmap_unlock(page_table, ptl);
1510
1511                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1512                                 goto unwritable_page;
1513
1514                         page_cache_release(old_page);
1515
1516                         /*
1517                          * Since we dropped the lock we need to revalidate
1518                          * the PTE as someone else may have changed it.  If
1519                          * they did, we just return, as we can count on the
1520                          * MMU to tell us if they didn't also make it writable.
1521                          */
1522                         page_table = pte_offset_map_lock(mm, pmd, address,
1523                                                          &ptl);
1524                         if (!pte_same(*page_table, orig_pte))
1525                                 goto unlock;
1526                 }
1527                 dirty_page = old_page;
1528                 get_page(dirty_page);
1529                 reuse = 1;
1530         }
1531
1532         if (reuse) {
1533                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1534                 entry = pte_mkyoung(orig_pte);
1535                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1536                 ptep_set_access_flags(vma, address, page_table, entry, 1);
1537                 update_mmu_cache(vma, address, entry);
1538                 lazy_mmu_prot_update(entry);
1539                 ret |= VM_FAULT_WRITE;
1540                 goto unlock;
1541         }
1542
1543         /*
1544          * Ok, we need to copy. Oh, well..
1545          */
1546         page_cache_get(old_page);
1547 gotten:
1548         pte_unmap_unlock(page_table, ptl);
1549
1550         if (unlikely(anon_vma_prepare(vma)))
1551                 goto oom;
1552         if (old_page == ZERO_PAGE(address)) {
1553                 new_page = alloc_zeroed_user_highpage(vma, address);
1554                 if (!new_page)
1555                         goto oom;
1556         } else {
1557                 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1558                 if (!new_page)
1559                         goto oom;
1560                 cow_user_page(new_page, old_page, address);
1561         }
1562
1563         /*
1564          * Re-check the pte - we dropped the lock
1565          */
1566         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1567         if (likely(pte_same(*page_table, orig_pte))) {
1568                 if (old_page) {
1569                         page_remove_rmap(old_page);
1570                         if (!PageAnon(old_page)) {
1571                                 dec_mm_counter(mm, file_rss);
1572                                 inc_mm_counter(mm, anon_rss);
1573                         }
1574                 } else
1575                         inc_mm_counter(mm, anon_rss);
1576                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1577                 entry = mk_pte(new_page, vma->vm_page_prot);
1578                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1579                 lazy_mmu_prot_update(entry);
1580                 /*
1581                  * Clear the pte entry and flush it first, before updating the
1582                  * pte with the new entry. This will avoid a race condition
1583                  * seen in the presence of one thread doing SMC and another
1584                  * thread doing COW.
1585                  */
1586                 ptep_clear_flush(vma, address, page_table);
1587                 set_pte_at(mm, address, page_table, entry);
1588                 update_mmu_cache(vma, address, entry);
1589                 lru_cache_add_active(new_page);
1590                 page_add_new_anon_rmap(new_page, vma, address);
1591
1592                 /* Free the old page.. */
1593                 new_page = old_page;
1594                 ret |= VM_FAULT_WRITE;
1595         }
1596         if (new_page)
1597                 page_cache_release(new_page);
1598         if (old_page)
1599                 page_cache_release(old_page);
1600 unlock:
1601         pte_unmap_unlock(page_table, ptl);
1602         if (dirty_page) {
1603                 set_page_dirty_balance(dirty_page);
1604                 put_page(dirty_page);
1605         }
1606         return ret;
1607 oom:
1608         if (old_page)
1609                 page_cache_release(old_page);
1610         return VM_FAULT_OOM;
1611
1612 unwritable_page:
1613         page_cache_release(old_page);
1614         return VM_FAULT_SIGBUS;
1615 }
1616
1617 /*
1618  * Helper functions for unmap_mapping_range().
1619  *
1620  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1621  *
1622  * We have to restart searching the prio_tree whenever we drop the lock,
1623  * since the iterator is only valid while the lock is held, and anyway
1624  * a later vma might be split and reinserted earlier while lock dropped.
1625  *
1626  * The list of nonlinear vmas could be handled more efficiently, using
1627  * a placeholder, but handle it in the same way until a need is shown.
1628  * It is important to search the prio_tree before nonlinear list: a vma
1629  * may become nonlinear and be shifted from prio_tree to nonlinear list
1630  * while the lock is dropped; but never shifted from list to prio_tree.
1631  *
1632  * In order to make forward progress despite restarting the search,
1633  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1634  * quickly skip it next time around.  Since the prio_tree search only
1635  * shows us those vmas affected by unmapping the range in question, we
1636  * can't efficiently keep all vmas in step with mapping->truncate_count:
1637  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1638  * mapping->truncate_count and vma->vm_truncate_count are protected by
1639  * i_mmap_lock.
1640  *
1641  * In order to make forward progress despite repeatedly restarting some
1642  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1643  * and restart from that address when we reach that vma again.  It might
1644  * have been split or merged, shrunk or extended, but never shifted: so
1645  * restart_addr remains valid so long as it remains in the vma's range.
1646  * unmap_mapping_range forces truncate_count to leap over page-aligned
1647  * values so we can save vma's restart_addr in its truncate_count field.
1648  */
1649 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1650
1651 static void reset_vma_truncate_counts(struct address_space *mapping)
1652 {
1653         struct vm_area_struct *vma;
1654         struct prio_tree_iter iter;
1655
1656         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1657                 vma->vm_truncate_count = 0;
1658         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1659                 vma->vm_truncate_count = 0;
1660 }
1661
1662 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1663                 unsigned long start_addr, unsigned long end_addr,
1664                 struct zap_details *details)
1665 {
1666         unsigned long restart_addr;
1667         int need_break;
1668
1669 again:
1670         restart_addr = vma->vm_truncate_count;
1671         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1672                 start_addr = restart_addr;
1673                 if (start_addr >= end_addr) {
1674                         /* Top of vma has been split off since last time */
1675                         vma->vm_truncate_count = details->truncate_count;
1676                         return 0;
1677                 }
1678         }
1679
1680         restart_addr = zap_page_range(vma, start_addr,
1681                                         end_addr - start_addr, details);
1682         need_break = need_resched() ||
1683                         need_lockbreak(details->i_mmap_lock);
1684
1685         if (restart_addr >= end_addr) {
1686                 /* We have now completed this vma: mark it so */
1687                 vma->vm_truncate_count = details->truncate_count;
1688                 if (!need_break)
1689                         return 0;
1690         } else {
1691                 /* Note restart_addr in vma's truncate_count field */
1692                 vma->vm_truncate_count = restart_addr;
1693                 if (!need_break)
1694                         goto again;
1695         }
1696
1697         spin_unlock(details->i_mmap_lock);
1698         cond_resched();
1699         spin_lock(details->i_mmap_lock);
1700         return -EINTR;
1701 }
1702
1703 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1704                                             struct zap_details *details)
1705 {
1706         struct vm_area_struct *vma;
1707         struct prio_tree_iter iter;
1708         pgoff_t vba, vea, zba, zea;
1709
1710 restart:
1711         vma_prio_tree_foreach(vma, &iter, root,
1712                         details->first_index, details->last_index) {
1713                 /* Skip quickly over those we have already dealt with */
1714                 if (vma->vm_truncate_count == details->truncate_count)
1715                         continue;
1716
1717                 vba = vma->vm_pgoff;
1718                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1719                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1720                 zba = details->first_index;
1721                 if (zba < vba)
1722                         zba = vba;
1723                 zea = details->last_index;
1724                 if (zea > vea)
1725                         zea = vea;
1726
1727                 if (unmap_mapping_range_vma(vma,
1728                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1729                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1730                                 details) < 0)
1731                         goto restart;
1732         }
1733 }
1734
1735 static inline void unmap_mapping_range_list(struct list_head *head,
1736                                             struct zap_details *details)
1737 {
1738         struct vm_area_struct *vma;
1739
1740         /*
1741          * In nonlinear VMAs there is no correspondence between virtual address
1742          * offset and file offset.  So we must perform an exhaustive search
1743          * across *all* the pages in each nonlinear VMA, not just the pages
1744          * whose virtual address lies outside the file truncation point.
1745          */
1746 restart:
1747         list_for_each_entry(vma, head, shared.vm_set.list) {
1748                 /* Skip quickly over those we have already dealt with */
1749                 if (vma->vm_truncate_count == details->truncate_count)
1750                         continue;
1751                 details->nonlinear_vma = vma;
1752                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1753                                         vma->vm_end, details) < 0)
1754                         goto restart;
1755         }
1756 }
1757
1758 /**
1759  * unmap_mapping_range - unmap the portion of all mmaps
1760  * in the specified address_space corresponding to the specified
1761  * page range in the underlying file.
1762  * @mapping: the address space containing mmaps to be unmapped.
1763  * @holebegin: byte in first page to unmap, relative to the start of
1764  * the underlying file.  This will be rounded down to a PAGE_SIZE
1765  * boundary.  Note that this is different from vmtruncate(), which
1766  * must keep the partial page.  In contrast, we must get rid of
1767  * partial pages.
1768  * @holelen: size of prospective hole in bytes.  This will be rounded
1769  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1770  * end of the file.
1771  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1772  * but 0 when invalidating pagecache, don't throw away private data.
1773  */
1774 void unmap_mapping_range(struct address_space *mapping,
1775                 loff_t const holebegin, loff_t const holelen, int even_cows)
1776 {
1777         struct zap_details details;
1778         pgoff_t hba = holebegin >> PAGE_SHIFT;
1779         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1780
1781         /* Check for overflow. */
1782         if (sizeof(holelen) > sizeof(hlen)) {
1783                 long long holeend =
1784                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1785                 if (holeend & ~(long long)ULONG_MAX)
1786                         hlen = ULONG_MAX - hba + 1;
1787         }
1788
1789         details.check_mapping = even_cows? NULL: mapping;
1790         details.nonlinear_vma = NULL;
1791         details.first_index = hba;
1792         details.last_index = hba + hlen - 1;
1793         if (details.last_index < details.first_index)
1794                 details.last_index = ULONG_MAX;
1795         details.i_mmap_lock = &mapping->i_mmap_lock;
1796
1797         spin_lock(&mapping->i_mmap_lock);
1798
1799         /* serialize i_size write against truncate_count write */
1800         smp_wmb();
1801         /* Protect against page faults, and endless unmapping loops */
1802         mapping->truncate_count++;
1803         /*
1804          * For archs where spin_lock has inclusive semantics like ia64
1805          * this smp_mb() will prevent to read pagetable contents
1806          * before the truncate_count increment is visible to
1807          * other cpus.
1808          */
1809         smp_mb();
1810         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1811                 if (mapping->truncate_count == 0)
1812                         reset_vma_truncate_counts(mapping);
1813                 mapping->truncate_count++;
1814         }
1815         details.truncate_count = mapping->truncate_count;
1816
1817         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1818                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1819         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1820                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1821         spin_unlock(&mapping->i_mmap_lock);
1822 }
1823 EXPORT_SYMBOL(unmap_mapping_range);
1824
1825 /**
1826  * vmtruncate - unmap mappings "freed" by truncate() syscall
1827  * @inode: inode of the file used
1828  * @offset: file offset to start truncating
1829  *
1830  * NOTE! We have to be ready to update the memory sharing
1831  * between the file and the memory map for a potential last
1832  * incomplete page.  Ugly, but necessary.
1833  */
1834 int vmtruncate(struct inode * inode, loff_t offset)
1835 {
1836         struct address_space *mapping = inode->i_mapping;
1837         unsigned long limit;
1838
1839         if (inode->i_size < offset)
1840                 goto do_expand;
1841         /*
1842          * truncation of in-use swapfiles is disallowed - it would cause
1843          * subsequent swapout to scribble on the now-freed blocks.
1844          */
1845         if (IS_SWAPFILE(inode))
1846                 goto out_busy;
1847         i_size_write(inode, offset);
1848         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1849         truncate_inode_pages(mapping, offset);
1850         goto out_truncate;
1851
1852 do_expand:
1853         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1854         if (limit != RLIM_INFINITY && offset > limit)
1855                 goto out_sig;
1856         if (offset > inode->i_sb->s_maxbytes)
1857                 goto out_big;
1858         i_size_write(inode, offset);
1859
1860 out_truncate:
1861         if (inode->i_op && inode->i_op->truncate)
1862                 inode->i_op->truncate(inode);
1863         return 0;
1864 out_sig:
1865         send_sig(SIGXFSZ, current, 0);
1866 out_big:
1867         return -EFBIG;
1868 out_busy:
1869         return -ETXTBSY;
1870 }
1871 EXPORT_SYMBOL(vmtruncate);
1872
1873 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1874 {
1875         struct address_space *mapping = inode->i_mapping;
1876
1877         /*
1878          * If the underlying filesystem is not going to provide
1879          * a way to truncate a range of blocks (punch a hole) -
1880          * we should return failure right now.
1881          */
1882         if (!inode->i_op || !inode->i_op->truncate_range)
1883                 return -ENOSYS;
1884
1885         mutex_lock(&inode->i_mutex);
1886         down_write(&inode->i_alloc_sem);
1887         unmap_mapping_range(mapping, offset, (end - offset), 1);
1888         truncate_inode_pages_range(mapping, offset, end);
1889         inode->i_op->truncate_range(inode, offset, end);
1890         up_write(&inode->i_alloc_sem);
1891         mutex_unlock(&inode->i_mutex);
1892
1893         return 0;
1894 }
1895 EXPORT_UNUSED_SYMBOL(vmtruncate_range);  /*  June 2006  */
1896
1897 /**
1898  * swapin_readahead - swap in pages in hope we need them soon
1899  * @entry: swap entry of this memory
1900  * @addr: address to start
1901  * @vma: user vma this addresses belong to
1902  *
1903  * Primitive swap readahead code. We simply read an aligned block of
1904  * (1 << page_cluster) entries in the swap area. This method is chosen
1905  * because it doesn't cost us any seek time.  We also make sure to queue
1906  * the 'original' request together with the readahead ones...
1907  *
1908  * This has been extended to use the NUMA policies from the mm triggering
1909  * the readahead.
1910  *
1911  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1912  */
1913 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1914 {
1915 #ifdef CONFIG_NUMA
1916         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1917 #endif
1918         int i, num;
1919         struct page *new_page;
1920         unsigned long offset;
1921
1922         /*
1923          * Get the number of handles we should do readahead io to.
1924          */
1925         num = valid_swaphandles(entry, &offset);
1926         for (i = 0; i < num; offset++, i++) {
1927                 /* Ok, do the async read-ahead now */
1928                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1929                                                            offset), vma, addr);
1930                 if (!new_page)
1931                         break;
1932                 page_cache_release(new_page);
1933 #ifdef CONFIG_NUMA
1934                 /*
1935                  * Find the next applicable VMA for the NUMA policy.
1936                  */
1937                 addr += PAGE_SIZE;
1938                 if (addr == 0)
1939                         vma = NULL;
1940                 if (vma) {
1941                         if (addr >= vma->vm_end) {
1942                                 vma = next_vma;
1943                                 next_vma = vma ? vma->vm_next : NULL;
1944                         }
1945                         if (vma && addr < vma->vm_start)
1946                                 vma = NULL;
1947                 } else {
1948                         if (next_vma && addr >= next_vma->vm_start) {
1949                                 vma = next_vma;
1950                                 next_vma = vma->vm_next;
1951                         }
1952                 }
1953 #endif
1954         }
1955         lru_add_drain();        /* Push any new pages onto the LRU now */
1956 }
1957
1958 /*
1959  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1960  * but allow concurrent faults), and pte mapped but not yet locked.
1961  * We return with mmap_sem still held, but pte unmapped and unlocked.
1962  */
1963 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1964                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1965                 int write_access, pte_t orig_pte)
1966 {
1967         spinlock_t *ptl;
1968         struct page *page;
1969         swp_entry_t entry;
1970         pte_t pte;
1971         int ret = VM_FAULT_MINOR;
1972
1973         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1974                 goto out;
1975
1976         entry = pte_to_swp_entry(orig_pte);
1977         if (is_migration_entry(entry)) {
1978                 migration_entry_wait(mm, pmd, address);
1979                 goto out;
1980         }
1981         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
1982         page = lookup_swap_cache(entry);
1983         if (!page) {
1984                 swapin_readahead(entry, address, vma);
1985                 page = read_swap_cache_async(entry, vma, address);
1986                 if (!page) {
1987                         /*
1988                          * Back out if somebody else faulted in this pte
1989                          * while we released the pte lock.
1990                          */
1991                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1992                         if (likely(pte_same(*page_table, orig_pte)))
1993                                 ret = VM_FAULT_OOM;
1994                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
1995                         goto unlock;
1996                 }
1997
1998                 /* Had to read the page from swap area: Major fault */
1999                 ret = VM_FAULT_MAJOR;
2000                 count_vm_event(PGMAJFAULT);
2001                 grab_swap_token();
2002         }
2003
2004         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2005         mark_page_accessed(page);
2006         lock_page(page);
2007
2008         /*
2009          * Back out if somebody else already faulted in this pte.
2010          */
2011         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2012         if (unlikely(!pte_same(*page_table, orig_pte)))
2013                 goto out_nomap;
2014
2015         if (unlikely(!PageUptodate(page))) {
2016                 ret = VM_FAULT_SIGBUS;
2017                 goto out_nomap;
2018         }
2019
2020         /* The page isn't present yet, go ahead with the fault. */
2021
2022         inc_mm_counter(mm, anon_rss);
2023         pte = mk_pte(page, vma->vm_page_prot);
2024         if (write_access && can_share_swap_page(page)) {
2025                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2026                 write_access = 0;
2027         }
2028
2029         flush_icache_page(vma, page);
2030         set_pte_at(mm, address, page_table, pte);
2031         page_add_anon_rmap(page, vma, address);
2032
2033         swap_free(entry);
2034         if (vm_swap_full())
2035                 remove_exclusive_swap_page(page);
2036         unlock_page(page);
2037
2038         if (write_access) {
2039                 if (do_wp_page(mm, vma, address,
2040                                 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
2041                         ret = VM_FAULT_OOM;
2042                 goto out;
2043         }
2044
2045         /* No need to invalidate - it was non-present before */
2046         update_mmu_cache(vma, address, pte);
2047         lazy_mmu_prot_update(pte);
2048 unlock:
2049         pte_unmap_unlock(page_table, ptl);
2050 out:
2051         return ret;
2052 out_nomap:
2053         pte_unmap_unlock(page_table, ptl);
2054         unlock_page(page);
2055         page_cache_release(page);
2056         return ret;
2057 }
2058
2059 /*
2060  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2061  * but allow concurrent faults), and pte mapped but not yet locked.
2062  * We return with mmap_sem still held, but pte unmapped and unlocked.
2063  */
2064 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2065                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2066                 int write_access)
2067 {
2068         struct page *page;
2069         spinlock_t *ptl;
2070         pte_t entry;
2071
2072         if (write_access) {
2073                 /* Allocate our own private page. */
2074                 pte_unmap(page_table);
2075
2076                 if (unlikely(anon_vma_prepare(vma)))
2077                         goto oom;
2078                 page = alloc_zeroed_user_highpage(vma, address);
2079                 if (!page)
2080                         goto oom;
2081
2082                 entry = mk_pte(page, vma->vm_page_prot);
2083                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2084
2085                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2086                 if (!pte_none(*page_table))
2087                         goto release;
2088                 inc_mm_counter(mm, anon_rss);
2089                 lru_cache_add_active(page);
2090                 page_add_new_anon_rmap(page, vma, address);
2091         } else {
2092                 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2093                 page = ZERO_PAGE(address);
2094                 page_cache_get(page);
2095                 entry = mk_pte(page, vma->vm_page_prot);
2096
2097                 ptl = pte_lockptr(mm, pmd);
2098                 spin_lock(ptl);
2099                 if (!pte_none(*page_table))
2100                         goto release;
2101                 inc_mm_counter(mm, file_rss);
2102                 page_add_file_rmap(page);
2103         }
2104
2105         set_pte_at(mm, address, page_table, entry);
2106
2107         /* No need to invalidate - it was non-present before */
2108         update_mmu_cache(vma, address, entry);
2109         lazy_mmu_prot_update(entry);
2110 unlock:
2111         pte_unmap_unlock(page_table, ptl);
2112         return VM_FAULT_MINOR;
2113 release:
2114         page_cache_release(page);
2115         goto unlock;
2116 oom:
2117         return VM_FAULT_OOM;
2118 }
2119
2120 /*
2121  * do_no_page() tries to create a new page mapping. It aggressively
2122  * tries to share with existing pages, but makes a separate copy if
2123  * the "write_access" parameter is true in order to avoid the next
2124  * page fault.
2125  *
2126  * As this is called only for pages that do not currently exist, we
2127  * do not need to flush old virtual caches or the TLB.
2128  *
2129  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2130  * but allow concurrent faults), and pte mapped but not yet locked.
2131  * We return with mmap_sem still held, but pte unmapped and unlocked.
2132  */
2133 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2134                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2135                 int write_access)
2136 {
2137         spinlock_t *ptl;
2138         struct page *new_page;
2139         struct address_space *mapping = NULL;
2140         pte_t entry;
2141         unsigned int sequence = 0;
2142         int ret = VM_FAULT_MINOR;
2143         int anon = 0;
2144         struct page *dirty_page = NULL;
2145
2146         pte_unmap(page_table);
2147         BUG_ON(vma->vm_flags & VM_PFNMAP);
2148
2149         if (vma->vm_file) {
2150                 mapping = vma->vm_file->f_mapping;
2151                 sequence = mapping->truncate_count;
2152                 smp_rmb(); /* serializes i_size against truncate_count */
2153         }
2154 retry:
2155         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2156         /*
2157          * No smp_rmb is needed here as long as there's a full
2158          * spin_lock/unlock sequence inside the ->nopage callback
2159          * (for the pagecache lookup) that acts as an implicit
2160          * smp_mb() and prevents the i_size read to happen
2161          * after the next truncate_count read.
2162          */
2163
2164         /* no page was available -- either SIGBUS or OOM */
2165         if (new_page == NOPAGE_SIGBUS)
2166                 return VM_FAULT_SIGBUS;
2167         if (new_page == NOPAGE_OOM)
2168                 return VM_FAULT_OOM;
2169
2170         /*
2171          * Should we do an early C-O-W break?
2172          */
2173         if (write_access) {
2174                 if (!(vma->vm_flags & VM_SHARED)) {
2175                         struct page *page;
2176
2177                         if (unlikely(anon_vma_prepare(vma)))
2178                                 goto oom;
2179                         page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2180                         if (!page)
2181                                 goto oom;
2182                         copy_user_highpage(page, new_page, address);
2183                         page_cache_release(new_page);
2184                         new_page = page;
2185                         anon = 1;
2186
2187                 } else {
2188                         /* if the page will be shareable, see if the backing
2189                          * address space wants to know that the page is about
2190                          * to become writable */
2191                         if (vma->vm_ops->page_mkwrite &&
2192                             vma->vm_ops->page_mkwrite(vma, new_page) < 0
2193                             ) {
2194                                 page_cache_release(new_page);
2195                                 return VM_FAULT_SIGBUS;
2196                         }
2197                 }
2198         }
2199
2200         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2201         /*
2202          * For a file-backed vma, someone could have truncated or otherwise
2203          * invalidated this page.  If unmap_mapping_range got called,
2204          * retry getting the page.
2205          */
2206         if (mapping && unlikely(sequence != mapping->truncate_count)) {
2207                 pte_unmap_unlock(page_table, ptl);
2208                 page_cache_release(new_page);
2209                 cond_resched();
2210                 sequence = mapping->truncate_count;
2211                 smp_rmb();
2212                 goto retry;
2213         }
2214
2215         /*
2216          * This silly early PAGE_DIRTY setting removes a race
2217          * due to the bad i386 page protection. But it's valid
2218          * for other architectures too.
2219          *
2220          * Note that if write_access is true, we either now have
2221          * an exclusive copy of the page, or this is a shared mapping,
2222          * so we can make it writable and dirty to avoid having to
2223          * handle that later.
2224          */
2225         /* Only go through if we didn't race with anybody else... */
2226         if (pte_none(*page_table)) {
2227                 flush_icache_page(vma, new_page);
2228                 entry = mk_pte(new_page, vma->vm_page_prot);
2229                 if (write_access)
2230                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2231                 set_pte_at(mm, address, page_table, entry);
2232                 if (anon) {
2233                         inc_mm_counter(mm, anon_rss);
2234                         lru_cache_add_active(new_page);
2235                         page_add_new_anon_rmap(new_page, vma, address);
2236                 } else {
2237                         inc_mm_counter(mm, file_rss);
2238                         page_add_file_rmap(new_page);
2239                         if (write_access) {
2240                                 dirty_page = new_page;
2241                                 get_page(dirty_page);
2242                         }
2243                 }
2244         } else {
2245                 /* One of our sibling threads was faster, back out. */
2246                 page_cache_release(new_page);
2247                 goto unlock;
2248         }
2249
2250         /* no need to invalidate: a not-present page shouldn't be cached */
2251         update_mmu_cache(vma, address, entry);
2252         lazy_mmu_prot_update(entry);
2253 unlock:
2254         pte_unmap_unlock(page_table, ptl);
2255         if (dirty_page) {
2256                 set_page_dirty_balance(dirty_page);
2257                 put_page(dirty_page);
2258         }
2259         return ret;
2260 oom:
2261         page_cache_release(new_page);
2262         return VM_FAULT_OOM;
2263 }
2264
2265 /*
2266  * do_no_pfn() tries to create a new page mapping for a page without
2267  * a struct_page backing it
2268  *
2269  * As this is called only for pages that do not currently exist, we
2270  * do not need to flush old virtual caches or the TLB.
2271  *
2272  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2273  * but allow concurrent faults), and pte mapped but not yet locked.
2274  * We return with mmap_sem still held, but pte unmapped and unlocked.
2275  *
2276  * It is expected that the ->nopfn handler always returns the same pfn
2277  * for a given virtual mapping.
2278  *
2279  * Mark this `noinline' to prevent it from bloating the main pagefault code.
2280  */
2281 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2282                      unsigned long address, pte_t *page_table, pmd_t *pmd,
2283                      int write_access)
2284 {
2285         spinlock_t *ptl;
2286         pte_t entry;
2287         unsigned long pfn;
2288         int ret = VM_FAULT_MINOR;
2289
2290         pte_unmap(page_table);
2291         BUG_ON(!(vma->vm_flags & VM_PFNMAP));
2292         BUG_ON(is_cow_mapping(vma->vm_flags));
2293
2294         pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2295         if (pfn == NOPFN_OOM)
2296                 return VM_FAULT_OOM;
2297         if (pfn == NOPFN_SIGBUS)
2298                 return VM_FAULT_SIGBUS;
2299
2300         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2301
2302         /* Only go through if we didn't race with anybody else... */
2303         if (pte_none(*page_table)) {
2304                 entry = pfn_pte(pfn, vma->vm_page_prot);
2305                 if (write_access)
2306                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2307                 set_pte_at(mm, address, page_table, entry);
2308         }
2309         pte_unmap_unlock(page_table, ptl);
2310         return ret;
2311 }
2312
2313 /*
2314  * Fault of a previously existing named mapping. Repopulate the pte
2315  * from the encoded file_pte if possible. This enables swappable
2316  * nonlinear vmas.
2317  *
2318  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2319  * but allow concurrent faults), and pte mapped but not yet locked.
2320  * We return with mmap_sem still held, but pte unmapped and unlocked.
2321  */
2322 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2323                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2324                 int write_access, pte_t orig_pte)
2325 {
2326         pgoff_t pgoff;
2327         int err;
2328
2329         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2330                 return VM_FAULT_MINOR;
2331
2332         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2333                 /*
2334                  * Page table corrupted: show pte and kill process.
2335                  */
2336                 print_bad_pte(vma, orig_pte, address);
2337                 return VM_FAULT_OOM;
2338         }
2339         /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2340
2341         pgoff = pte_to_pgoff(orig_pte);
2342         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2343                                         vma->vm_page_prot, pgoff, 0);
2344         if (err == -ENOMEM)
2345                 return VM_FAULT_OOM;
2346         if (err)
2347                 return VM_FAULT_SIGBUS;
2348         return VM_FAULT_MAJOR;
2349 }
2350
2351 /*
2352  * These routines also need to handle stuff like marking pages dirty
2353  * and/or accessed for architectures that don't do it in hardware (most
2354  * RISC architectures).  The early dirtying is also good on the i386.
2355  *
2356  * There is also a hook called "update_mmu_cache()" that architectures
2357  * with external mmu caches can use to update those (ie the Sparc or
2358  * PowerPC hashed page tables that act as extended TLBs).
2359  *
2360  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2361  * but allow concurrent faults), and pte mapped but not yet locked.
2362  * We return with mmap_sem still held, but pte unmapped and unlocked.
2363  */
2364 static inline int handle_pte_fault(struct mm_struct *mm,
2365                 struct vm_area_struct *vma, unsigned long address,
2366                 pte_t *pte, pmd_t *pmd, int write_access)
2367 {
2368         pte_t entry;
2369         pte_t old_entry;
2370         spinlock_t *ptl;
2371
2372         old_entry = entry = *pte;
2373         if (!pte_present(entry)) {
2374                 if (pte_none(entry)) {
2375                         if (vma->vm_ops) {
2376                                 if (vma->vm_ops->nopage)
2377                                         return do_no_page(mm, vma, address,
2378                                                           pte, pmd,
2379                                                           write_access);
2380                                 if (unlikely(vma->vm_ops->nopfn))
2381                                         return do_no_pfn(mm, vma, address, pte,
2382                                                          pmd, write_access);
2383                         }
2384                         return do_anonymous_page(mm, vma, address,
2385                                                  pte, pmd, write_access);
2386                 }
2387                 if (pte_file(entry))
2388                         return do_file_page(mm, vma, address,
2389                                         pte, pmd, write_access, entry);
2390                 return do_swap_page(mm, vma, address,
2391                                         pte, pmd, write_access, entry);
2392         }
2393
2394         ptl = pte_lockptr(mm, pmd);
2395         spin_lock(ptl);
2396         if (unlikely(!pte_same(*pte, entry)))
2397                 goto unlock;
2398         if (write_access) {
2399                 if (!pte_write(entry))
2400                         return do_wp_page(mm, vma, address,
2401                                         pte, pmd, ptl, entry);
2402                 entry = pte_mkdirty(entry);
2403         }
2404         entry = pte_mkyoung(entry);
2405         if (!pte_same(old_entry, entry)) {
2406                 ptep_set_access_flags(vma, address, pte, entry, write_access);
2407                 update_mmu_cache(vma, address, entry);
2408                 lazy_mmu_prot_update(entry);
2409         } else {
2410                 /*
2411                  * This is needed only for protection faults but the arch code
2412                  * is not yet telling us if this is a protection fault or not.
2413                  * This still avoids useless tlb flushes for .text page faults
2414                  * with threads.
2415                  */
2416                 if (write_access)
2417                         flush_tlb_page(vma, address);
2418         }
2419 unlock:
2420         pte_unmap_unlock(pte, ptl);
2421         return VM_FAULT_MINOR;
2422 }
2423
2424 /*
2425  * By the time we get here, we already hold the mm semaphore
2426  */
2427 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2428                 unsigned long address, int write_access)
2429 {
2430         pgd_t *pgd;
2431         pud_t *pud;
2432         pmd_t *pmd;
2433         pte_t *pte;
2434
2435         __set_current_state(TASK_RUNNING);
2436
2437         count_vm_event(PGFAULT);
2438
2439         if (unlikely(is_vm_hugetlb_page(vma)))
2440                 return hugetlb_fault(mm, vma, address, write_access);
2441
2442         pgd = pgd_offset(mm, address);
2443         pud = pud_alloc(mm, pgd, address);
2444         if (!pud)
2445                 return VM_FAULT_OOM;
2446         pmd = pmd_alloc(mm, pud, address);
2447         if (!pmd)
2448                 return VM_FAULT_OOM;
2449         pte = pte_alloc_map(mm, pmd, address);
2450         if (!pte)
2451                 return VM_FAULT_OOM;
2452
2453         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2454 }
2455
2456 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2457
2458 #ifndef __PAGETABLE_PUD_FOLDED
2459 /*
2460  * Allocate page upper directory.
2461  * We've already handled the fast-path in-line.
2462  */
2463 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2464 {
2465         pud_t *new = pud_alloc_one(mm, address);
2466         if (!new)
2467                 return -ENOMEM;
2468
2469         spin_lock(&mm->page_table_lock);
2470         if (pgd_present(*pgd))          /* Another has populated it */
2471                 pud_free(new);
2472         else
2473                 pgd_populate(mm, pgd, new);
2474         spin_unlock(&mm->page_table_lock);
2475         return 0;
2476 }
2477 #else
2478 /* Workaround for gcc 2.96 */
2479 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2480 {
2481         return 0;
2482 }
2483 #endif /* __PAGETABLE_PUD_FOLDED */
2484
2485 #ifndef __PAGETABLE_PMD_FOLDED
2486 /*
2487  * Allocate page middle directory.
2488  * We've already handled the fast-path in-line.
2489  */
2490 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2491 {
2492         pmd_t *new = pmd_alloc_one(mm, address);
2493         if (!new)
2494                 return -ENOMEM;
2495
2496         spin_lock(&mm->page_table_lock);
2497 #ifndef __ARCH_HAS_4LEVEL_HACK
2498         if (pud_present(*pud))          /* Another has populated it */
2499                 pmd_free(new);
2500         else
2501                 pud_populate(mm, pud, new);
2502 #else
2503         if (pgd_present(*pud))          /* Another has populated it */
2504                 pmd_free(new);
2505         else
2506                 pgd_populate(mm, pud, new);
2507 #endif /* __ARCH_HAS_4LEVEL_HACK */
2508         spin_unlock(&mm->page_table_lock);
2509         return 0;
2510 }
2511 #else
2512 /* Workaround for gcc 2.96 */
2513 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2514 {
2515         return 0;
2516 }
2517 #endif /* __PAGETABLE_PMD_FOLDED */
2518
2519 int make_pages_present(unsigned long addr, unsigned long end)
2520 {
2521         int ret, len, write;
2522         struct vm_area_struct * vma;
2523
2524         vma = find_vma(current->mm, addr);
2525         if (!vma)
2526                 return -1;
2527         write = (vma->vm_flags & VM_WRITE) != 0;
2528         BUG_ON(addr >= end);
2529         BUG_ON(end > vma->vm_end);
2530         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2531         ret = get_user_pages(current, current->mm, addr,
2532                         len, write, 0, NULL, NULL);
2533         if (ret < 0)
2534                 return ret;
2535         return ret == len ? 0 : -1;
2536 }
2537
2538 /* 
2539  * Map a vmalloc()-space virtual address to the physical page.
2540  */
2541 struct page * vmalloc_to_page(void * vmalloc_addr)
2542 {
2543         unsigned long addr = (unsigned long) vmalloc_addr;
2544         struct page *page = NULL;
2545         pgd_t *pgd = pgd_offset_k(addr);
2546         pud_t *pud;
2547         pmd_t *pmd;
2548         pte_t *ptep, pte;
2549   
2550         if (!pgd_none(*pgd)) {
2551                 pud = pud_offset(pgd, addr);
2552                 if (!pud_none(*pud)) {
2553                         pmd = pmd_offset(pud, addr);
2554                         if (!pmd_none(*pmd)) {
2555                                 ptep = pte_offset_map(pmd, addr);
2556                                 pte = *ptep;
2557                                 if (pte_present(pte))
2558                                         page = pte_page(pte);
2559                                 pte_unmap(ptep);
2560                         }
2561                 }
2562         }
2563         return page;
2564 }
2565
2566 EXPORT_SYMBOL(vmalloc_to_page);
2567
2568 /*
2569  * Map a vmalloc()-space virtual address to the physical page frame number.
2570  */
2571 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2572 {
2573         return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2574 }
2575
2576 EXPORT_SYMBOL(vmalloc_to_pfn);
2577
2578 #if !defined(__HAVE_ARCH_GATE_AREA)
2579
2580 #if defined(AT_SYSINFO_EHDR)
2581 static struct vm_area_struct gate_vma;
2582
2583 static int __init gate_vma_init(void)
2584 {
2585         gate_vma.vm_mm = NULL;
2586         gate_vma.vm_start = FIXADDR_USER_START;
2587         gate_vma.vm_end = FIXADDR_USER_END;
2588         gate_vma.vm_page_prot = PAGE_READONLY;
2589         gate_vma.vm_flags = 0;
2590         return 0;
2591 }
2592 __initcall(gate_vma_init);
2593 #endif
2594
2595 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2596 {
2597 #ifdef AT_SYSINFO_EHDR
2598         return &gate_vma;
2599 #else
2600         return NULL;
2601 #endif
2602 }
2603
2604 int in_gate_area_no_task(unsigned long addr)
2605 {
2606 #ifdef AT_SYSINFO_EHDR
2607         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2608                 return 1;
2609 #endif
2610         return 0;
2611 }
2612
2613 #endif  /* __HAVE_ARCH_GATE_AREA */
2614
2615 /*
2616  * Access another process' address space.
2617  * Source/target buffer must be kernel space,
2618  * Do not walk the page table directly, use get_user_pages
2619  */
2620 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2621 {
2622         struct mm_struct *mm;
2623         struct vm_area_struct *vma;
2624         struct page *page;
2625         void *old_buf = buf;
2626
2627         mm = get_task_mm(tsk);
2628         if (!mm)
2629                 return 0;
2630
2631         down_read(&mm->mmap_sem);
2632         /* ignore errors, just check how much was sucessfully transfered */
2633         while (len) {
2634                 int bytes, ret, offset;
2635                 void *maddr;
2636
2637                 ret = get_user_pages(tsk, mm, addr, 1,
2638                                 write, 1, &page, &vma);
2639                 if (ret <= 0)
2640                         break;
2641
2642                 bytes = len;
2643                 offset = addr & (PAGE_SIZE-1);
2644                 if (bytes > PAGE_SIZE-offset)
2645                         bytes = PAGE_SIZE-offset;
2646
2647                 maddr = kmap(page);
2648                 if (write) {
2649                         copy_to_user_page(vma, page, addr,
2650                                           maddr + offset, buf, bytes);
2651                         set_page_dirty_lock(page);
2652                 } else {
2653                         copy_from_user_page(vma, page, addr,
2654                                             buf, maddr + offset, bytes);
2655                 }
2656                 kunmap(page);
2657                 page_cache_release(page);
2658                 len -= bytes;
2659                 buf += bytes;
2660                 addr += bytes;
2661         }
2662         up_read(&mm->mmap_sem);
2663         mmput(mm);
2664
2665         return buf - old_buf;
2666 }