mm: add vm_insert_mixed
[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 #include <linux/memcontrol.h>
54
55 #include <asm/pgalloc.h>
56 #include <asm/uaccess.h>
57 #include <asm/tlb.h>
58 #include <asm/tlbflush.h>
59 #include <asm/pgtable.h>
60
61 #include <linux/swapops.h>
62 #include <linux/elf.h>
63
64 #ifndef CONFIG_NEED_MULTIPLE_NODES
65 /* use the per-pgdat data instead for discontigmem - mbligh */
66 unsigned long max_mapnr;
67 struct page *mem_map;
68
69 EXPORT_SYMBOL(max_mapnr);
70 EXPORT_SYMBOL(mem_map);
71 #endif
72
73 unsigned long num_physpages;
74 /*
75  * A number of key systems in x86 including ioremap() rely on the assumption
76  * that high_memory defines the upper bound on direct map memory, then end
77  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
78  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
79  * and ZONE_HIGHMEM.
80  */
81 void * high_memory;
82
83 EXPORT_SYMBOL(num_physpages);
84 EXPORT_SYMBOL(high_memory);
85
86 /*
87  * Randomize the address space (stacks, mmaps, brk, etc.).
88  *
89  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
90  *   as ancient (libc5 based) binaries can segfault. )
91  */
92 int randomize_va_space __read_mostly =
93 #ifdef CONFIG_COMPAT_BRK
94                                         1;
95 #else
96                                         2;
97 #endif
98
99 static int __init disable_randmaps(char *s)
100 {
101         randomize_va_space = 0;
102         return 1;
103 }
104 __setup("norandmaps", disable_randmaps);
105
106
107 /*
108  * If a p?d_bad entry is found while walking page tables, report
109  * the error, before resetting entry to p?d_none.  Usually (but
110  * very seldom) called out from the p?d_none_or_clear_bad macros.
111  */
112
113 void pgd_clear_bad(pgd_t *pgd)
114 {
115         pgd_ERROR(*pgd);
116         pgd_clear(pgd);
117 }
118
119 void pud_clear_bad(pud_t *pud)
120 {
121         pud_ERROR(*pud);
122         pud_clear(pud);
123 }
124
125 void pmd_clear_bad(pmd_t *pmd)
126 {
127         pmd_ERROR(*pmd);
128         pmd_clear(pmd);
129 }
130
131 /*
132  * Note: this doesn't free the actual pages themselves. That
133  * has been handled earlier when unmapping all the memory regions.
134  */
135 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
136 {
137         pgtable_t token = pmd_pgtable(*pmd);
138         pmd_clear(pmd);
139         pte_free_tlb(tlb, token);
140         tlb->mm->nr_ptes--;
141 }
142
143 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
144                                 unsigned long addr, unsigned long end,
145                                 unsigned long floor, unsigned long ceiling)
146 {
147         pmd_t *pmd;
148         unsigned long next;
149         unsigned long start;
150
151         start = addr;
152         pmd = pmd_offset(pud, addr);
153         do {
154                 next = pmd_addr_end(addr, end);
155                 if (pmd_none_or_clear_bad(pmd))
156                         continue;
157                 free_pte_range(tlb, pmd);
158         } while (pmd++, addr = next, addr != end);
159
160         start &= PUD_MASK;
161         if (start < floor)
162                 return;
163         if (ceiling) {
164                 ceiling &= PUD_MASK;
165                 if (!ceiling)
166                         return;
167         }
168         if (end - 1 > ceiling - 1)
169                 return;
170
171         pmd = pmd_offset(pud, start);
172         pud_clear(pud);
173         pmd_free_tlb(tlb, pmd);
174 }
175
176 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
177                                 unsigned long addr, unsigned long end,
178                                 unsigned long floor, unsigned long ceiling)
179 {
180         pud_t *pud;
181         unsigned long next;
182         unsigned long start;
183
184         start = addr;
185         pud = pud_offset(pgd, addr);
186         do {
187                 next = pud_addr_end(addr, end);
188                 if (pud_none_or_clear_bad(pud))
189                         continue;
190                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
191         } while (pud++, addr = next, addr != end);
192
193         start &= PGDIR_MASK;
194         if (start < floor)
195                 return;
196         if (ceiling) {
197                 ceiling &= PGDIR_MASK;
198                 if (!ceiling)
199                         return;
200         }
201         if (end - 1 > ceiling - 1)
202                 return;
203
204         pud = pud_offset(pgd, start);
205         pgd_clear(pgd);
206         pud_free_tlb(tlb, pud);
207 }
208
209 /*
210  * This function frees user-level page tables of a process.
211  *
212  * Must be called with pagetable lock held.
213  */
214 void free_pgd_range(struct mmu_gather **tlb,
215                         unsigned long addr, unsigned long end,
216                         unsigned long floor, unsigned long ceiling)
217 {
218         pgd_t *pgd;
219         unsigned long next;
220         unsigned long start;
221
222         /*
223          * The next few lines have given us lots of grief...
224          *
225          * Why are we testing PMD* at this top level?  Because often
226          * there will be no work to do at all, and we'd prefer not to
227          * go all the way down to the bottom just to discover that.
228          *
229          * Why all these "- 1"s?  Because 0 represents both the bottom
230          * of the address space and the top of it (using -1 for the
231          * top wouldn't help much: the masks would do the wrong thing).
232          * The rule is that addr 0 and floor 0 refer to the bottom of
233          * the address space, but end 0 and ceiling 0 refer to the top
234          * Comparisons need to use "end - 1" and "ceiling - 1" (though
235          * that end 0 case should be mythical).
236          *
237          * Wherever addr is brought up or ceiling brought down, we must
238          * be careful to reject "the opposite 0" before it confuses the
239          * subsequent tests.  But what about where end is brought down
240          * by PMD_SIZE below? no, end can't go down to 0 there.
241          *
242          * Whereas we round start (addr) and ceiling down, by different
243          * masks at different levels, in order to test whether a table
244          * now has no other vmas using it, so can be freed, we don't
245          * bother to round floor or end up - the tests don't need that.
246          */
247
248         addr &= PMD_MASK;
249         if (addr < floor) {
250                 addr += PMD_SIZE;
251                 if (!addr)
252                         return;
253         }
254         if (ceiling) {
255                 ceiling &= PMD_MASK;
256                 if (!ceiling)
257                         return;
258         }
259         if (end - 1 > ceiling - 1)
260                 end -= PMD_SIZE;
261         if (addr > end - 1)
262                 return;
263
264         start = addr;
265         pgd = pgd_offset((*tlb)->mm, addr);
266         do {
267                 next = pgd_addr_end(addr, end);
268                 if (pgd_none_or_clear_bad(pgd))
269                         continue;
270                 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
271         } while (pgd++, addr = next, addr != end);
272 }
273
274 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
275                 unsigned long floor, unsigned long ceiling)
276 {
277         while (vma) {
278                 struct vm_area_struct *next = vma->vm_next;
279                 unsigned long addr = vma->vm_start;
280
281                 /*
282                  * Hide vma from rmap and vmtruncate before freeing pgtables
283                  */
284                 anon_vma_unlink(vma);
285                 unlink_file_vma(vma);
286
287                 if (is_vm_hugetlb_page(vma)) {
288                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
289                                 floor, next? next->vm_start: ceiling);
290                 } else {
291                         /*
292                          * Optimization: gather nearby vmas into one call down
293                          */
294                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
295                                && !is_vm_hugetlb_page(next)) {
296                                 vma = next;
297                                 next = vma->vm_next;
298                                 anon_vma_unlink(vma);
299                                 unlink_file_vma(vma);
300                         }
301                         free_pgd_range(tlb, addr, vma->vm_end,
302                                 floor, next? next->vm_start: ceiling);
303                 }
304                 vma = next;
305         }
306 }
307
308 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
309 {
310         pgtable_t new = pte_alloc_one(mm, address);
311         if (!new)
312                 return -ENOMEM;
313
314         spin_lock(&mm->page_table_lock);
315         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
316                 mm->nr_ptes++;
317                 pmd_populate(mm, pmd, new);
318                 new = NULL;
319         }
320         spin_unlock(&mm->page_table_lock);
321         if (new)
322                 pte_free(mm, new);
323         return 0;
324 }
325
326 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
327 {
328         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
329         if (!new)
330                 return -ENOMEM;
331
332         spin_lock(&init_mm.page_table_lock);
333         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
334                 pmd_populate_kernel(&init_mm, pmd, new);
335                 new = NULL;
336         }
337         spin_unlock(&init_mm.page_table_lock);
338         if (new)
339                 pte_free_kernel(&init_mm, new);
340         return 0;
341 }
342
343 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
344 {
345         if (file_rss)
346                 add_mm_counter(mm, file_rss, file_rss);
347         if (anon_rss)
348                 add_mm_counter(mm, anon_rss, anon_rss);
349 }
350
351 /*
352  * This function is called to print an error when a bad pte
353  * is found. For example, we might have a PFN-mapped pte in
354  * a region that doesn't allow it.
355  *
356  * The calling function must still handle the error.
357  */
358 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
359 {
360         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
361                         "vm_flags = %lx, vaddr = %lx\n",
362                 (long long)pte_val(pte),
363                 (vma->vm_mm == current->mm ? current->comm : "???"),
364                 vma->vm_flags, vaddr);
365         dump_stack();
366 }
367
368 static inline int is_cow_mapping(unsigned int flags)
369 {
370         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
371 }
372
373 /*
374  * vm_normal_page -- This function gets the "struct page" associated with a pte.
375  *
376  * "Special" mappings do not wish to be associated with a "struct page" (either
377  * it doesn't exist, or it exists but they don't want to touch it). In this
378  * case, NULL is returned here. "Normal" mappings do have a struct page.
379  *
380  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
381  * pte bit, in which case this function is trivial. Secondly, an architecture
382  * may not have a spare pte bit, which requires a more complicated scheme,
383  * described below.
384  *
385  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
386  * special mapping (even if there are underlying and valid "struct pages").
387  * COWed pages of a VM_PFNMAP are always normal.
388  *
389  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
390  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
391  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
392  * mapping will always honor the rule
393  *
394  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
395  *
396  * And for normal mappings this is false.
397  *
398  * This restricts such mappings to be a linear translation from virtual address
399  * to pfn. To get around this restriction, we allow arbitrary mappings so long
400  * as the vma is not a COW mapping; in that case, we know that all ptes are
401  * special (because none can have been COWed).
402  *
403  *
404  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
405  *
406  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
407  * page" backing, however the difference is that _all_ pages with a struct
408  * page (that is, those where pfn_valid is true) are refcounted and considered
409  * normal pages by the VM. The disadvantage is that pages are refcounted
410  * (which can be slower and simply not an option for some PFNMAP users). The
411  * advantage is that we don't have to follow the strict linearity rule of
412  * PFNMAP mappings in order to support COWable mappings.
413  *
414  */
415 #ifdef __HAVE_ARCH_PTE_SPECIAL
416 # define HAVE_PTE_SPECIAL 1
417 #else
418 # define HAVE_PTE_SPECIAL 0
419 #endif
420 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
421                                 pte_t pte)
422 {
423         unsigned long pfn;
424
425         if (HAVE_PTE_SPECIAL) {
426                 if (likely(!pte_special(pte))) {
427                         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
428                         return pte_page(pte);
429                 }
430                 VM_BUG_ON(!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)));
431                 return NULL;
432         }
433
434         /* !HAVE_PTE_SPECIAL case follows: */
435
436         pfn = pte_pfn(pte);
437
438         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
439                 if (vma->vm_flags & VM_MIXEDMAP) {
440                         if (!pfn_valid(pfn))
441                                 return NULL;
442                         goto out;
443                 } else {
444                         unsigned long off;
445                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
446                         if (pfn == vma->vm_pgoff + off)
447                                 return NULL;
448                         if (!is_cow_mapping(vma->vm_flags))
449                                 return NULL;
450                 }
451         }
452
453         VM_BUG_ON(!pfn_valid(pfn));
454
455         /*
456          * NOTE! We still have PageReserved() pages in the page tables.
457          *
458          * eg. VDSO mappings can cause them to exist.
459          */
460 out:
461         return pfn_to_page(pfn);
462 }
463
464 /*
465  * copy one vm_area from one task to the other. Assumes the page tables
466  * already present in the new task to be cleared in the whole range
467  * covered by this vma.
468  */
469
470 static inline void
471 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
472                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
473                 unsigned long addr, int *rss)
474 {
475         unsigned long vm_flags = vma->vm_flags;
476         pte_t pte = *src_pte;
477         struct page *page;
478
479         /* pte contains position in swap or file, so copy. */
480         if (unlikely(!pte_present(pte))) {
481                 if (!pte_file(pte)) {
482                         swp_entry_t entry = pte_to_swp_entry(pte);
483
484                         swap_duplicate(entry);
485                         /* make sure dst_mm is on swapoff's mmlist. */
486                         if (unlikely(list_empty(&dst_mm->mmlist))) {
487                                 spin_lock(&mmlist_lock);
488                                 if (list_empty(&dst_mm->mmlist))
489                                         list_add(&dst_mm->mmlist,
490                                                  &src_mm->mmlist);
491                                 spin_unlock(&mmlist_lock);
492                         }
493                         if (is_write_migration_entry(entry) &&
494                                         is_cow_mapping(vm_flags)) {
495                                 /*
496                                  * COW mappings require pages in both parent
497                                  * and child to be set to read.
498                                  */
499                                 make_migration_entry_read(&entry);
500                                 pte = swp_entry_to_pte(entry);
501                                 set_pte_at(src_mm, addr, src_pte, pte);
502                         }
503                 }
504                 goto out_set_pte;
505         }
506
507         /*
508          * If it's a COW mapping, write protect it both
509          * in the parent and the child
510          */
511         if (is_cow_mapping(vm_flags)) {
512                 ptep_set_wrprotect(src_mm, addr, src_pte);
513                 pte = pte_wrprotect(pte);
514         }
515
516         /*
517          * If it's a shared mapping, mark it clean in
518          * the child
519          */
520         if (vm_flags & VM_SHARED)
521                 pte = pte_mkclean(pte);
522         pte = pte_mkold(pte);
523
524         page = vm_normal_page(vma, addr, pte);
525         if (page) {
526                 get_page(page);
527                 page_dup_rmap(page, vma, addr);
528                 rss[!!PageAnon(page)]++;
529         }
530
531 out_set_pte:
532         set_pte_at(dst_mm, addr, dst_pte, pte);
533 }
534
535 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
536                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
537                 unsigned long addr, unsigned long end)
538 {
539         pte_t *src_pte, *dst_pte;
540         spinlock_t *src_ptl, *dst_ptl;
541         int progress = 0;
542         int rss[2];
543
544 again:
545         rss[1] = rss[0] = 0;
546         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
547         if (!dst_pte)
548                 return -ENOMEM;
549         src_pte = pte_offset_map_nested(src_pmd, addr);
550         src_ptl = pte_lockptr(src_mm, src_pmd);
551         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
552         arch_enter_lazy_mmu_mode();
553
554         do {
555                 /*
556                  * We are holding two locks at this point - either of them
557                  * could generate latencies in another task on another CPU.
558                  */
559                 if (progress >= 32) {
560                         progress = 0;
561                         if (need_resched() ||
562                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
563                                 break;
564                 }
565                 if (pte_none(*src_pte)) {
566                         progress++;
567                         continue;
568                 }
569                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
570                 progress += 8;
571         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
572
573         arch_leave_lazy_mmu_mode();
574         spin_unlock(src_ptl);
575         pte_unmap_nested(src_pte - 1);
576         add_mm_rss(dst_mm, rss[0], rss[1]);
577         pte_unmap_unlock(dst_pte - 1, dst_ptl);
578         cond_resched();
579         if (addr != end)
580                 goto again;
581         return 0;
582 }
583
584 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
585                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
586                 unsigned long addr, unsigned long end)
587 {
588         pmd_t *src_pmd, *dst_pmd;
589         unsigned long next;
590
591         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
592         if (!dst_pmd)
593                 return -ENOMEM;
594         src_pmd = pmd_offset(src_pud, addr);
595         do {
596                 next = pmd_addr_end(addr, end);
597                 if (pmd_none_or_clear_bad(src_pmd))
598                         continue;
599                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
600                                                 vma, addr, next))
601                         return -ENOMEM;
602         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
603         return 0;
604 }
605
606 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
607                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
608                 unsigned long addr, unsigned long end)
609 {
610         pud_t *src_pud, *dst_pud;
611         unsigned long next;
612
613         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
614         if (!dst_pud)
615                 return -ENOMEM;
616         src_pud = pud_offset(src_pgd, addr);
617         do {
618                 next = pud_addr_end(addr, end);
619                 if (pud_none_or_clear_bad(src_pud))
620                         continue;
621                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
622                                                 vma, addr, next))
623                         return -ENOMEM;
624         } while (dst_pud++, src_pud++, addr = next, addr != end);
625         return 0;
626 }
627
628 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
629                 struct vm_area_struct *vma)
630 {
631         pgd_t *src_pgd, *dst_pgd;
632         unsigned long next;
633         unsigned long addr = vma->vm_start;
634         unsigned long end = vma->vm_end;
635
636         /*
637          * Don't copy ptes where a page fault will fill them correctly.
638          * Fork becomes much lighter when there are big shared or private
639          * readonly mappings. The tradeoff is that copy_page_range is more
640          * efficient than faulting.
641          */
642         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
643                 if (!vma->anon_vma)
644                         return 0;
645         }
646
647         if (is_vm_hugetlb_page(vma))
648                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
649
650         dst_pgd = pgd_offset(dst_mm, addr);
651         src_pgd = pgd_offset(src_mm, addr);
652         do {
653                 next = pgd_addr_end(addr, end);
654                 if (pgd_none_or_clear_bad(src_pgd))
655                         continue;
656                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
657                                                 vma, addr, next))
658                         return -ENOMEM;
659         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
660         return 0;
661 }
662
663 static unsigned long zap_pte_range(struct mmu_gather *tlb,
664                                 struct vm_area_struct *vma, pmd_t *pmd,
665                                 unsigned long addr, unsigned long end,
666                                 long *zap_work, struct zap_details *details)
667 {
668         struct mm_struct *mm = tlb->mm;
669         pte_t *pte;
670         spinlock_t *ptl;
671         int file_rss = 0;
672         int anon_rss = 0;
673
674         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
675         arch_enter_lazy_mmu_mode();
676         do {
677                 pte_t ptent = *pte;
678                 if (pte_none(ptent)) {
679                         (*zap_work)--;
680                         continue;
681                 }
682
683                 (*zap_work) -= PAGE_SIZE;
684
685                 if (pte_present(ptent)) {
686                         struct page *page;
687
688                         page = vm_normal_page(vma, addr, ptent);
689                         if (unlikely(details) && page) {
690                                 /*
691                                  * unmap_shared_mapping_pages() wants to
692                                  * invalidate cache without truncating:
693                                  * unmap shared but keep private pages.
694                                  */
695                                 if (details->check_mapping &&
696                                     details->check_mapping != page->mapping)
697                                         continue;
698                                 /*
699                                  * Each page->index must be checked when
700                                  * invalidating or truncating nonlinear.
701                                  */
702                                 if (details->nonlinear_vma &&
703                                     (page->index < details->first_index ||
704                                      page->index > details->last_index))
705                                         continue;
706                         }
707                         ptent = ptep_get_and_clear_full(mm, addr, pte,
708                                                         tlb->fullmm);
709                         tlb_remove_tlb_entry(tlb, pte, addr);
710                         if (unlikely(!page))
711                                 continue;
712                         if (unlikely(details) && details->nonlinear_vma
713                             && linear_page_index(details->nonlinear_vma,
714                                                 addr) != page->index)
715                                 set_pte_at(mm, addr, pte,
716                                            pgoff_to_pte(page->index));
717                         if (PageAnon(page))
718                                 anon_rss--;
719                         else {
720                                 if (pte_dirty(ptent))
721                                         set_page_dirty(page);
722                                 if (pte_young(ptent))
723                                         SetPageReferenced(page);
724                                 file_rss--;
725                         }
726                         page_remove_rmap(page, vma);
727                         tlb_remove_page(tlb, page);
728                         continue;
729                 }
730                 /*
731                  * If details->check_mapping, we leave swap entries;
732                  * if details->nonlinear_vma, we leave file entries.
733                  */
734                 if (unlikely(details))
735                         continue;
736                 if (!pte_file(ptent))
737                         free_swap_and_cache(pte_to_swp_entry(ptent));
738                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
739         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
740
741         add_mm_rss(mm, file_rss, anon_rss);
742         arch_leave_lazy_mmu_mode();
743         pte_unmap_unlock(pte - 1, ptl);
744
745         return addr;
746 }
747
748 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
749                                 struct vm_area_struct *vma, pud_t *pud,
750                                 unsigned long addr, unsigned long end,
751                                 long *zap_work, struct zap_details *details)
752 {
753         pmd_t *pmd;
754         unsigned long next;
755
756         pmd = pmd_offset(pud, addr);
757         do {
758                 next = pmd_addr_end(addr, end);
759                 if (pmd_none_or_clear_bad(pmd)) {
760                         (*zap_work)--;
761                         continue;
762                 }
763                 next = zap_pte_range(tlb, vma, pmd, addr, next,
764                                                 zap_work, details);
765         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
766
767         return addr;
768 }
769
770 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
771                                 struct vm_area_struct *vma, pgd_t *pgd,
772                                 unsigned long addr, unsigned long end,
773                                 long *zap_work, struct zap_details *details)
774 {
775         pud_t *pud;
776         unsigned long next;
777
778         pud = pud_offset(pgd, addr);
779         do {
780                 next = pud_addr_end(addr, end);
781                 if (pud_none_or_clear_bad(pud)) {
782                         (*zap_work)--;
783                         continue;
784                 }
785                 next = zap_pmd_range(tlb, vma, pud, addr, next,
786                                                 zap_work, details);
787         } while (pud++, addr = next, (addr != end && *zap_work > 0));
788
789         return addr;
790 }
791
792 static unsigned long unmap_page_range(struct mmu_gather *tlb,
793                                 struct vm_area_struct *vma,
794                                 unsigned long addr, unsigned long end,
795                                 long *zap_work, struct zap_details *details)
796 {
797         pgd_t *pgd;
798         unsigned long next;
799
800         if (details && !details->check_mapping && !details->nonlinear_vma)
801                 details = NULL;
802
803         BUG_ON(addr >= end);
804         tlb_start_vma(tlb, vma);
805         pgd = pgd_offset(vma->vm_mm, addr);
806         do {
807                 next = pgd_addr_end(addr, end);
808                 if (pgd_none_or_clear_bad(pgd)) {
809                         (*zap_work)--;
810                         continue;
811                 }
812                 next = zap_pud_range(tlb, vma, pgd, addr, next,
813                                                 zap_work, details);
814         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
815         tlb_end_vma(tlb, vma);
816
817         return addr;
818 }
819
820 #ifdef CONFIG_PREEMPT
821 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
822 #else
823 /* No preempt: go for improved straight-line efficiency */
824 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
825 #endif
826
827 /**
828  * unmap_vmas - unmap a range of memory covered by a list of vma's
829  * @tlbp: address of the caller's struct mmu_gather
830  * @vma: the starting vma
831  * @start_addr: virtual address at which to start unmapping
832  * @end_addr: virtual address at which to end unmapping
833  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
834  * @details: details of nonlinear truncation or shared cache invalidation
835  *
836  * Returns the end address of the unmapping (restart addr if interrupted).
837  *
838  * Unmap all pages in the vma list.
839  *
840  * We aim to not hold locks for too long (for scheduling latency reasons).
841  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
842  * return the ending mmu_gather to the caller.
843  *
844  * Only addresses between `start' and `end' will be unmapped.
845  *
846  * The VMA list must be sorted in ascending virtual address order.
847  *
848  * unmap_vmas() assumes that the caller will flush the whole unmapped address
849  * range after unmap_vmas() returns.  So the only responsibility here is to
850  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
851  * drops the lock and schedules.
852  */
853 unsigned long unmap_vmas(struct mmu_gather **tlbp,
854                 struct vm_area_struct *vma, unsigned long start_addr,
855                 unsigned long end_addr, unsigned long *nr_accounted,
856                 struct zap_details *details)
857 {
858         long zap_work = ZAP_BLOCK_SIZE;
859         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
860         int tlb_start_valid = 0;
861         unsigned long start = start_addr;
862         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
863         int fullmm = (*tlbp)->fullmm;
864
865         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
866                 unsigned long end;
867
868                 start = max(vma->vm_start, start_addr);
869                 if (start >= vma->vm_end)
870                         continue;
871                 end = min(vma->vm_end, end_addr);
872                 if (end <= vma->vm_start)
873                         continue;
874
875                 if (vma->vm_flags & VM_ACCOUNT)
876                         *nr_accounted += (end - start) >> PAGE_SHIFT;
877
878                 while (start != end) {
879                         if (!tlb_start_valid) {
880                                 tlb_start = start;
881                                 tlb_start_valid = 1;
882                         }
883
884                         if (unlikely(is_vm_hugetlb_page(vma))) {
885                                 unmap_hugepage_range(vma, start, end);
886                                 zap_work -= (end - start) /
887                                                 (HPAGE_SIZE / PAGE_SIZE);
888                                 start = end;
889                         } else
890                                 start = unmap_page_range(*tlbp, vma,
891                                                 start, end, &zap_work, details);
892
893                         if (zap_work > 0) {
894                                 BUG_ON(start != end);
895                                 break;
896                         }
897
898                         tlb_finish_mmu(*tlbp, tlb_start, start);
899
900                         if (need_resched() ||
901                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
902                                 if (i_mmap_lock) {
903                                         *tlbp = NULL;
904                                         goto out;
905                                 }
906                                 cond_resched();
907                         }
908
909                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
910                         tlb_start_valid = 0;
911                         zap_work = ZAP_BLOCK_SIZE;
912                 }
913         }
914 out:
915         return start;   /* which is now the end (or restart) address */
916 }
917
918 /**
919  * zap_page_range - remove user pages in a given range
920  * @vma: vm_area_struct holding the applicable pages
921  * @address: starting address of pages to zap
922  * @size: number of bytes to zap
923  * @details: details of nonlinear truncation or shared cache invalidation
924  */
925 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
926                 unsigned long size, struct zap_details *details)
927 {
928         struct mm_struct *mm = vma->vm_mm;
929         struct mmu_gather *tlb;
930         unsigned long end = address + size;
931         unsigned long nr_accounted = 0;
932
933         lru_add_drain();
934         tlb = tlb_gather_mmu(mm, 0);
935         update_hiwater_rss(mm);
936         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
937         if (tlb)
938                 tlb_finish_mmu(tlb, address, end);
939         return end;
940 }
941
942 /*
943  * Do a quick page-table lookup for a single page.
944  */
945 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
946                         unsigned int flags)
947 {
948         pgd_t *pgd;
949         pud_t *pud;
950         pmd_t *pmd;
951         pte_t *ptep, pte;
952         spinlock_t *ptl;
953         struct page *page;
954         struct mm_struct *mm = vma->vm_mm;
955
956         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
957         if (!IS_ERR(page)) {
958                 BUG_ON(flags & FOLL_GET);
959                 goto out;
960         }
961
962         page = NULL;
963         pgd = pgd_offset(mm, address);
964         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
965                 goto no_page_table;
966
967         pud = pud_offset(pgd, address);
968         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
969                 goto no_page_table;
970         
971         pmd = pmd_offset(pud, address);
972         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
973                 goto no_page_table;
974
975         if (pmd_huge(*pmd)) {
976                 BUG_ON(flags & FOLL_GET);
977                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
978                 goto out;
979         }
980
981         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
982         if (!ptep)
983                 goto out;
984
985         pte = *ptep;
986         if (!pte_present(pte))
987                 goto unlock;
988         if ((flags & FOLL_WRITE) && !pte_write(pte))
989                 goto unlock;
990         page = vm_normal_page(vma, address, pte);
991         if (unlikely(!page))
992                 goto unlock;
993
994         if (flags & FOLL_GET)
995                 get_page(page);
996         if (flags & FOLL_TOUCH) {
997                 if ((flags & FOLL_WRITE) &&
998                     !pte_dirty(pte) && !PageDirty(page))
999                         set_page_dirty(page);
1000                 mark_page_accessed(page);
1001         }
1002 unlock:
1003         pte_unmap_unlock(ptep, ptl);
1004 out:
1005         return page;
1006
1007 no_page_table:
1008         /*
1009          * When core dumping an enormous anonymous area that nobody
1010          * has touched so far, we don't want to allocate page tables.
1011          */
1012         if (flags & FOLL_ANON) {
1013                 page = ZERO_PAGE(0);
1014                 if (flags & FOLL_GET)
1015                         get_page(page);
1016                 BUG_ON(flags & FOLL_WRITE);
1017         }
1018         return page;
1019 }
1020
1021 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1022                 unsigned long start, int len, int write, int force,
1023                 struct page **pages, struct vm_area_struct **vmas)
1024 {
1025         int i;
1026         unsigned int vm_flags;
1027
1028         if (len <= 0)
1029                 return 0;
1030         /* 
1031          * Require read or write permissions.
1032          * If 'force' is set, we only require the "MAY" flags.
1033          */
1034         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1035         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1036         i = 0;
1037
1038         do {
1039                 struct vm_area_struct *vma;
1040                 unsigned int foll_flags;
1041
1042                 vma = find_extend_vma(mm, start);
1043                 if (!vma && in_gate_area(tsk, start)) {
1044                         unsigned long pg = start & PAGE_MASK;
1045                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1046                         pgd_t *pgd;
1047                         pud_t *pud;
1048                         pmd_t *pmd;
1049                         pte_t *pte;
1050                         if (write) /* user gate pages are read-only */
1051                                 return i ? : -EFAULT;
1052                         if (pg > TASK_SIZE)
1053                                 pgd = pgd_offset_k(pg);
1054                         else
1055                                 pgd = pgd_offset_gate(mm, pg);
1056                         BUG_ON(pgd_none(*pgd));
1057                         pud = pud_offset(pgd, pg);
1058                         BUG_ON(pud_none(*pud));
1059                         pmd = pmd_offset(pud, pg);
1060                         if (pmd_none(*pmd))
1061                                 return i ? : -EFAULT;
1062                         pte = pte_offset_map(pmd, pg);
1063                         if (pte_none(*pte)) {
1064                                 pte_unmap(pte);
1065                                 return i ? : -EFAULT;
1066                         }
1067                         if (pages) {
1068                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1069                                 pages[i] = page;
1070                                 if (page)
1071                                         get_page(page);
1072                         }
1073                         pte_unmap(pte);
1074                         if (vmas)
1075                                 vmas[i] = gate_vma;
1076                         i++;
1077                         start += PAGE_SIZE;
1078                         len--;
1079                         continue;
1080                 }
1081
1082                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1083                                 || !(vm_flags & vma->vm_flags))
1084                         return i ? : -EFAULT;
1085
1086                 if (is_vm_hugetlb_page(vma)) {
1087                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1088                                                 &start, &len, i, write);
1089                         continue;
1090                 }
1091
1092                 foll_flags = FOLL_TOUCH;
1093                 if (pages)
1094                         foll_flags |= FOLL_GET;
1095                 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1096                     (!vma->vm_ops || !vma->vm_ops->fault))
1097                         foll_flags |= FOLL_ANON;
1098
1099                 do {
1100                         struct page *page;
1101
1102                         /*
1103                          * If tsk is ooming, cut off its access to large memory
1104                          * allocations. It has a pending SIGKILL, but it can't
1105                          * be processed until returning to user space.
1106                          */
1107                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1108                                 return -ENOMEM;
1109
1110                         if (write)
1111                                 foll_flags |= FOLL_WRITE;
1112
1113                         cond_resched();
1114                         while (!(page = follow_page(vma, start, foll_flags))) {
1115                                 int ret;
1116                                 ret = handle_mm_fault(mm, vma, start,
1117                                                 foll_flags & FOLL_WRITE);
1118                                 if (ret & VM_FAULT_ERROR) {
1119                                         if (ret & VM_FAULT_OOM)
1120                                                 return i ? i : -ENOMEM;
1121                                         else if (ret & VM_FAULT_SIGBUS)
1122                                                 return i ? i : -EFAULT;
1123                                         BUG();
1124                                 }
1125                                 if (ret & VM_FAULT_MAJOR)
1126                                         tsk->maj_flt++;
1127                                 else
1128                                         tsk->min_flt++;
1129
1130                                 /*
1131                                  * The VM_FAULT_WRITE bit tells us that
1132                                  * do_wp_page has broken COW when necessary,
1133                                  * even if maybe_mkwrite decided not to set
1134                                  * pte_write. We can thus safely do subsequent
1135                                  * page lookups as if they were reads.
1136                                  */
1137                                 if (ret & VM_FAULT_WRITE)
1138                                         foll_flags &= ~FOLL_WRITE;
1139
1140                                 cond_resched();
1141                         }
1142                         if (pages) {
1143                                 pages[i] = page;
1144
1145                                 flush_anon_page(vma, page, start);
1146                                 flush_dcache_page(page);
1147                         }
1148                         if (vmas)
1149                                 vmas[i] = vma;
1150                         i++;
1151                         start += PAGE_SIZE;
1152                         len--;
1153                 } while (len && start < vma->vm_end);
1154         } while (len);
1155         return i;
1156 }
1157 EXPORT_SYMBOL(get_user_pages);
1158
1159 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1160                         spinlock_t **ptl)
1161 {
1162         pgd_t * pgd = pgd_offset(mm, addr);
1163         pud_t * pud = pud_alloc(mm, pgd, addr);
1164         if (pud) {
1165                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1166                 if (pmd)
1167                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1168         }
1169         return NULL;
1170 }
1171
1172 /*
1173  * This is the old fallback for page remapping.
1174  *
1175  * For historical reasons, it only allows reserved pages. Only
1176  * old drivers should use this, and they needed to mark their
1177  * pages reserved for the old functions anyway.
1178  */
1179 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1180                         struct page *page, pgprot_t prot)
1181 {
1182         struct mm_struct *mm = vma->vm_mm;
1183         int retval;
1184         pte_t *pte;
1185         spinlock_t *ptl;
1186
1187         retval = mem_cgroup_charge(page, mm, GFP_KERNEL);
1188         if (retval)
1189                 goto out;
1190
1191         retval = -EINVAL;
1192         if (PageAnon(page))
1193                 goto out_uncharge;
1194         retval = -ENOMEM;
1195         flush_dcache_page(page);
1196         pte = get_locked_pte(mm, addr, &ptl);
1197         if (!pte)
1198                 goto out_uncharge;
1199         retval = -EBUSY;
1200         if (!pte_none(*pte))
1201                 goto out_unlock;
1202
1203         /* Ok, finally just insert the thing.. */
1204         get_page(page);
1205         inc_mm_counter(mm, file_rss);
1206         page_add_file_rmap(page);
1207         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1208
1209         retval = 0;
1210         pte_unmap_unlock(pte, ptl);
1211         return retval;
1212 out_unlock:
1213         pte_unmap_unlock(pte, ptl);
1214 out_uncharge:
1215         mem_cgroup_uncharge_page(page);
1216 out:
1217         return retval;
1218 }
1219
1220 /**
1221  * vm_insert_page - insert single page into user vma
1222  * @vma: user vma to map to
1223  * @addr: target user address of this page
1224  * @page: source kernel page
1225  *
1226  * This allows drivers to insert individual pages they've allocated
1227  * into a user vma.
1228  *
1229  * The page has to be a nice clean _individual_ kernel allocation.
1230  * If you allocate a compound page, you need to have marked it as
1231  * such (__GFP_COMP), or manually just split the page up yourself
1232  * (see split_page()).
1233  *
1234  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1235  * took an arbitrary page protection parameter. This doesn't allow
1236  * that. Your vma protection will have to be set up correctly, which
1237  * means that if you want a shared writable mapping, you'd better
1238  * ask for a shared writable mapping!
1239  *
1240  * The page does not need to be reserved.
1241  */
1242 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1243                         struct page *page)
1244 {
1245         if (addr < vma->vm_start || addr >= vma->vm_end)
1246                 return -EFAULT;
1247         if (!page_count(page))
1248                 return -EINVAL;
1249         vma->vm_flags |= VM_INSERTPAGE;
1250         return insert_page(vma, addr, page, vma->vm_page_prot);
1251 }
1252 EXPORT_SYMBOL(vm_insert_page);
1253
1254 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1255                         unsigned long pfn, pgprot_t prot)
1256 {
1257         struct mm_struct *mm = vma->vm_mm;
1258         int retval;
1259         pte_t *pte, entry;
1260         spinlock_t *ptl;
1261
1262         retval = -ENOMEM;
1263         pte = get_locked_pte(mm, addr, &ptl);
1264         if (!pte)
1265                 goto out;
1266         retval = -EBUSY;
1267         if (!pte_none(*pte))
1268                 goto out_unlock;
1269
1270         /* Ok, finally just insert the thing.. */
1271         entry = pte_mkspecial(pfn_pte(pfn, prot));
1272         set_pte_at(mm, addr, pte, entry);
1273         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1274
1275         retval = 0;
1276 out_unlock:
1277         pte_unmap_unlock(pte, ptl);
1278 out:
1279         return retval;
1280 }
1281
1282 /**
1283  * vm_insert_pfn - insert single pfn into user vma
1284  * @vma: user vma to map to
1285  * @addr: target user address of this page
1286  * @pfn: source kernel pfn
1287  *
1288  * Similar to vm_inert_page, this allows drivers to insert individual pages
1289  * they've allocated into a user vma. Same comments apply.
1290  *
1291  * This function should only be called from a vm_ops->fault handler, and
1292  * in that case the handler should return NULL.
1293  */
1294 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1295                         unsigned long pfn)
1296 {
1297         /*
1298          * Technically, architectures with pte_special can avoid all these
1299          * restrictions (same for remap_pfn_range).  However we would like
1300          * consistency in testing and feature parity among all, so we should
1301          * try to keep these invariants in place for everybody.
1302          */
1303         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1304         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1305                                                 (VM_PFNMAP|VM_MIXEDMAP));
1306         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1307         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1308
1309         if (addr < vma->vm_start || addr >= vma->vm_end)
1310                 return -EFAULT;
1311         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1312 }
1313 EXPORT_SYMBOL(vm_insert_pfn);
1314
1315 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1316                         unsigned long pfn)
1317 {
1318         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1319
1320         if (addr < vma->vm_start || addr >= vma->vm_end)
1321                 return -EFAULT;
1322
1323         /*
1324          * If we don't have pte special, then we have to use the pfn_valid()
1325          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1326          * refcount the page if pfn_valid is true (hence insert_page rather
1327          * than insert_pfn).
1328          */
1329         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1330                 struct page *page;
1331
1332                 page = pfn_to_page(pfn);
1333                 return insert_page(vma, addr, page, vma->vm_page_prot);
1334         }
1335         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1336 }
1337 EXPORT_SYMBOL(vm_insert_mixed);
1338
1339 /*
1340  * maps a range of physical memory into the requested pages. the old
1341  * mappings are removed. any references to nonexistent pages results
1342  * in null mappings (currently treated as "copy-on-access")
1343  */
1344 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1345                         unsigned long addr, unsigned long end,
1346                         unsigned long pfn, pgprot_t prot)
1347 {
1348         pte_t *pte;
1349         spinlock_t *ptl;
1350
1351         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1352         if (!pte)
1353                 return -ENOMEM;
1354         arch_enter_lazy_mmu_mode();
1355         do {
1356                 BUG_ON(!pte_none(*pte));
1357                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1358                 pfn++;
1359         } while (pte++, addr += PAGE_SIZE, addr != end);
1360         arch_leave_lazy_mmu_mode();
1361         pte_unmap_unlock(pte - 1, ptl);
1362         return 0;
1363 }
1364
1365 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1366                         unsigned long addr, unsigned long end,
1367                         unsigned long pfn, pgprot_t prot)
1368 {
1369         pmd_t *pmd;
1370         unsigned long next;
1371
1372         pfn -= addr >> PAGE_SHIFT;
1373         pmd = pmd_alloc(mm, pud, addr);
1374         if (!pmd)
1375                 return -ENOMEM;
1376         do {
1377                 next = pmd_addr_end(addr, end);
1378                 if (remap_pte_range(mm, pmd, addr, next,
1379                                 pfn + (addr >> PAGE_SHIFT), prot))
1380                         return -ENOMEM;
1381         } while (pmd++, addr = next, addr != end);
1382         return 0;
1383 }
1384
1385 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1386                         unsigned long addr, unsigned long end,
1387                         unsigned long pfn, pgprot_t prot)
1388 {
1389         pud_t *pud;
1390         unsigned long next;
1391
1392         pfn -= addr >> PAGE_SHIFT;
1393         pud = pud_alloc(mm, pgd, addr);
1394         if (!pud)
1395                 return -ENOMEM;
1396         do {
1397                 next = pud_addr_end(addr, end);
1398                 if (remap_pmd_range(mm, pud, addr, next,
1399                                 pfn + (addr >> PAGE_SHIFT), prot))
1400                         return -ENOMEM;
1401         } while (pud++, addr = next, addr != end);
1402         return 0;
1403 }
1404
1405 /**
1406  * remap_pfn_range - remap kernel memory to userspace
1407  * @vma: user vma to map to
1408  * @addr: target user address to start at
1409  * @pfn: physical address of kernel memory
1410  * @size: size of map area
1411  * @prot: page protection flags for this mapping
1412  *
1413  *  Note: this is only safe if the mm semaphore is held when called.
1414  */
1415 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1416                     unsigned long pfn, unsigned long size, pgprot_t prot)
1417 {
1418         pgd_t *pgd;
1419         unsigned long next;
1420         unsigned long end = addr + PAGE_ALIGN(size);
1421         struct mm_struct *mm = vma->vm_mm;
1422         int err;
1423
1424         /*
1425          * Physically remapped pages are special. Tell the
1426          * rest of the world about it:
1427          *   VM_IO tells people not to look at these pages
1428          *      (accesses can have side effects).
1429          *   VM_RESERVED is specified all over the place, because
1430          *      in 2.4 it kept swapout's vma scan off this vma; but
1431          *      in 2.6 the LRU scan won't even find its pages, so this
1432          *      flag means no more than count its pages in reserved_vm,
1433          *      and omit it from core dump, even when VM_IO turned off.
1434          *   VM_PFNMAP tells the core MM that the base pages are just
1435          *      raw PFN mappings, and do not have a "struct page" associated
1436          *      with them.
1437          *
1438          * There's a horrible special case to handle copy-on-write
1439          * behaviour that some programs depend on. We mark the "original"
1440          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1441          */
1442         if (is_cow_mapping(vma->vm_flags)) {
1443                 if (addr != vma->vm_start || end != vma->vm_end)
1444                         return -EINVAL;
1445                 vma->vm_pgoff = pfn;
1446         }
1447
1448         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1449
1450         BUG_ON(addr >= end);
1451         pfn -= addr >> PAGE_SHIFT;
1452         pgd = pgd_offset(mm, addr);
1453         flush_cache_range(vma, addr, end);
1454         do {
1455                 next = pgd_addr_end(addr, end);
1456                 err = remap_pud_range(mm, pgd, addr, next,
1457                                 pfn + (addr >> PAGE_SHIFT), prot);
1458                 if (err)
1459                         break;
1460         } while (pgd++, addr = next, addr != end);
1461         return err;
1462 }
1463 EXPORT_SYMBOL(remap_pfn_range);
1464
1465 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1466                                      unsigned long addr, unsigned long end,
1467                                      pte_fn_t fn, void *data)
1468 {
1469         pte_t *pte;
1470         int err;
1471         pgtable_t token;
1472         spinlock_t *uninitialized_var(ptl);
1473
1474         pte = (mm == &init_mm) ?
1475                 pte_alloc_kernel(pmd, addr) :
1476                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1477         if (!pte)
1478                 return -ENOMEM;
1479
1480         BUG_ON(pmd_huge(*pmd));
1481
1482         token = pmd_pgtable(*pmd);
1483
1484         do {
1485                 err = fn(pte, token, addr, data);
1486                 if (err)
1487                         break;
1488         } while (pte++, addr += PAGE_SIZE, addr != end);
1489
1490         if (mm != &init_mm)
1491                 pte_unmap_unlock(pte-1, ptl);
1492         return err;
1493 }
1494
1495 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1496                                      unsigned long addr, unsigned long end,
1497                                      pte_fn_t fn, void *data)
1498 {
1499         pmd_t *pmd;
1500         unsigned long next;
1501         int err;
1502
1503         pmd = pmd_alloc(mm, pud, addr);
1504         if (!pmd)
1505                 return -ENOMEM;
1506         do {
1507                 next = pmd_addr_end(addr, end);
1508                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1509                 if (err)
1510                         break;
1511         } while (pmd++, addr = next, addr != end);
1512         return err;
1513 }
1514
1515 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1516                                      unsigned long addr, unsigned long end,
1517                                      pte_fn_t fn, void *data)
1518 {
1519         pud_t *pud;
1520         unsigned long next;
1521         int err;
1522
1523         pud = pud_alloc(mm, pgd, addr);
1524         if (!pud)
1525                 return -ENOMEM;
1526         do {
1527                 next = pud_addr_end(addr, end);
1528                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1529                 if (err)
1530                         break;
1531         } while (pud++, addr = next, addr != end);
1532         return err;
1533 }
1534
1535 /*
1536  * Scan a region of virtual memory, filling in page tables as necessary
1537  * and calling a provided function on each leaf page table.
1538  */
1539 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1540                         unsigned long size, pte_fn_t fn, void *data)
1541 {
1542         pgd_t *pgd;
1543         unsigned long next;
1544         unsigned long end = addr + size;
1545         int err;
1546
1547         BUG_ON(addr >= end);
1548         pgd = pgd_offset(mm, addr);
1549         do {
1550                 next = pgd_addr_end(addr, end);
1551                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1552                 if (err)
1553                         break;
1554         } while (pgd++, addr = next, addr != end);
1555         return err;
1556 }
1557 EXPORT_SYMBOL_GPL(apply_to_page_range);
1558
1559 /*
1560  * handle_pte_fault chooses page fault handler according to an entry
1561  * which was read non-atomically.  Before making any commitment, on
1562  * those architectures or configurations (e.g. i386 with PAE) which
1563  * might give a mix of unmatched parts, do_swap_page and do_file_page
1564  * must check under lock before unmapping the pte and proceeding
1565  * (but do_wp_page is only called after already making such a check;
1566  * and do_anonymous_page and do_no_page can safely check later on).
1567  */
1568 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1569                                 pte_t *page_table, pte_t orig_pte)
1570 {
1571         int same = 1;
1572 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1573         if (sizeof(pte_t) > sizeof(unsigned long)) {
1574                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1575                 spin_lock(ptl);
1576                 same = pte_same(*page_table, orig_pte);
1577                 spin_unlock(ptl);
1578         }
1579 #endif
1580         pte_unmap(page_table);
1581         return same;
1582 }
1583
1584 /*
1585  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1586  * servicing faults for write access.  In the normal case, do always want
1587  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1588  * that do not have writing enabled, when used by access_process_vm.
1589  */
1590 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1591 {
1592         if (likely(vma->vm_flags & VM_WRITE))
1593                 pte = pte_mkwrite(pte);
1594         return pte;
1595 }
1596
1597 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1598 {
1599         /*
1600          * If the source page was a PFN mapping, we don't have
1601          * a "struct page" for it. We do a best-effort copy by
1602          * just copying from the original user address. If that
1603          * fails, we just zero-fill it. Live with it.
1604          */
1605         if (unlikely(!src)) {
1606                 void *kaddr = kmap_atomic(dst, KM_USER0);
1607                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1608
1609                 /*
1610                  * This really shouldn't fail, because the page is there
1611                  * in the page tables. But it might just be unreadable,
1612                  * in which case we just give up and fill the result with
1613                  * zeroes.
1614                  */
1615                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1616                         memset(kaddr, 0, PAGE_SIZE);
1617                 kunmap_atomic(kaddr, KM_USER0);
1618                 flush_dcache_page(dst);
1619         } else
1620                 copy_user_highpage(dst, src, va, vma);
1621 }
1622
1623 /*
1624  * This routine handles present pages, when users try to write
1625  * to a shared page. It is done by copying the page to a new address
1626  * and decrementing the shared-page counter for the old page.
1627  *
1628  * Note that this routine assumes that the protection checks have been
1629  * done by the caller (the low-level page fault routine in most cases).
1630  * Thus we can safely just mark it writable once we've done any necessary
1631  * COW.
1632  *
1633  * We also mark the page dirty at this point even though the page will
1634  * change only once the write actually happens. This avoids a few races,
1635  * and potentially makes it more efficient.
1636  *
1637  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1638  * but allow concurrent faults), with pte both mapped and locked.
1639  * We return with mmap_sem still held, but pte unmapped and unlocked.
1640  */
1641 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1642                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1643                 spinlock_t *ptl, pte_t orig_pte)
1644 {
1645         struct page *old_page, *new_page;
1646         pte_t entry;
1647         int reuse = 0, ret = 0;
1648         int page_mkwrite = 0;
1649         struct page *dirty_page = NULL;
1650
1651         old_page = vm_normal_page(vma, address, orig_pte);
1652         if (!old_page)
1653                 goto gotten;
1654
1655         /*
1656          * Take out anonymous pages first, anonymous shared vmas are
1657          * not dirty accountable.
1658          */
1659         if (PageAnon(old_page)) {
1660                 if (!TestSetPageLocked(old_page)) {
1661                         reuse = can_share_swap_page(old_page);
1662                         unlock_page(old_page);
1663                 }
1664         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1665                                         (VM_WRITE|VM_SHARED))) {
1666                 /*
1667                  * Only catch write-faults on shared writable pages,
1668                  * read-only shared pages can get COWed by
1669                  * get_user_pages(.write=1, .force=1).
1670                  */
1671                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1672                         /*
1673                          * Notify the address space that the page is about to
1674                          * become writable so that it can prohibit this or wait
1675                          * for the page to get into an appropriate state.
1676                          *
1677                          * We do this without the lock held, so that it can
1678                          * sleep if it needs to.
1679                          */
1680                         page_cache_get(old_page);
1681                         pte_unmap_unlock(page_table, ptl);
1682
1683                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1684                                 goto unwritable_page;
1685
1686                         /*
1687                          * Since we dropped the lock we need to revalidate
1688                          * the PTE as someone else may have changed it.  If
1689                          * they did, we just return, as we can count on the
1690                          * MMU to tell us if they didn't also make it writable.
1691                          */
1692                         page_table = pte_offset_map_lock(mm, pmd, address,
1693                                                          &ptl);
1694                         page_cache_release(old_page);
1695                         if (!pte_same(*page_table, orig_pte))
1696                                 goto unlock;
1697
1698                         page_mkwrite = 1;
1699                 }
1700                 dirty_page = old_page;
1701                 get_page(dirty_page);
1702                 reuse = 1;
1703         }
1704
1705         if (reuse) {
1706                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1707                 entry = pte_mkyoung(orig_pte);
1708                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1709                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1710                         update_mmu_cache(vma, address, entry);
1711                 ret |= VM_FAULT_WRITE;
1712                 goto unlock;
1713         }
1714
1715         /*
1716          * Ok, we need to copy. Oh, well..
1717          */
1718         page_cache_get(old_page);
1719 gotten:
1720         pte_unmap_unlock(page_table, ptl);
1721
1722         if (unlikely(anon_vma_prepare(vma)))
1723                 goto oom;
1724         VM_BUG_ON(old_page == ZERO_PAGE(0));
1725         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1726         if (!new_page)
1727                 goto oom;
1728         cow_user_page(new_page, old_page, address, vma);
1729         __SetPageUptodate(new_page);
1730
1731         if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
1732                 goto oom_free_new;
1733
1734         /*
1735          * Re-check the pte - we dropped the lock
1736          */
1737         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1738         if (likely(pte_same(*page_table, orig_pte))) {
1739                 if (old_page) {
1740                         page_remove_rmap(old_page, vma);
1741                         if (!PageAnon(old_page)) {
1742                                 dec_mm_counter(mm, file_rss);
1743                                 inc_mm_counter(mm, anon_rss);
1744                         }
1745                 } else
1746                         inc_mm_counter(mm, anon_rss);
1747                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1748                 entry = mk_pte(new_page, vma->vm_page_prot);
1749                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1750                 /*
1751                  * Clear the pte entry and flush it first, before updating the
1752                  * pte with the new entry. This will avoid a race condition
1753                  * seen in the presence of one thread doing SMC and another
1754                  * thread doing COW.
1755                  */
1756                 ptep_clear_flush(vma, address, page_table);
1757                 set_pte_at(mm, address, page_table, entry);
1758                 update_mmu_cache(vma, address, entry);
1759                 lru_cache_add_active(new_page);
1760                 page_add_new_anon_rmap(new_page, vma, address);
1761
1762                 /* Free the old page.. */
1763                 new_page = old_page;
1764                 ret |= VM_FAULT_WRITE;
1765         } else
1766                 mem_cgroup_uncharge_page(new_page);
1767
1768         if (new_page)
1769                 page_cache_release(new_page);
1770         if (old_page)
1771                 page_cache_release(old_page);
1772 unlock:
1773         pte_unmap_unlock(page_table, ptl);
1774         if (dirty_page) {
1775                 if (vma->vm_file)
1776                         file_update_time(vma->vm_file);
1777
1778                 /*
1779                  * Yes, Virginia, this is actually required to prevent a race
1780                  * with clear_page_dirty_for_io() from clearing the page dirty
1781                  * bit after it clear all dirty ptes, but before a racing
1782                  * do_wp_page installs a dirty pte.
1783                  *
1784                  * do_no_page is protected similarly.
1785                  */
1786                 wait_on_page_locked(dirty_page);
1787                 set_page_dirty_balance(dirty_page, page_mkwrite);
1788                 put_page(dirty_page);
1789         }
1790         return ret;
1791 oom_free_new:
1792         page_cache_release(new_page);
1793 oom:
1794         if (old_page)
1795                 page_cache_release(old_page);
1796         return VM_FAULT_OOM;
1797
1798 unwritable_page:
1799         page_cache_release(old_page);
1800         return VM_FAULT_SIGBUS;
1801 }
1802
1803 /*
1804  * Helper functions for unmap_mapping_range().
1805  *
1806  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1807  *
1808  * We have to restart searching the prio_tree whenever we drop the lock,
1809  * since the iterator is only valid while the lock is held, and anyway
1810  * a later vma might be split and reinserted earlier while lock dropped.
1811  *
1812  * The list of nonlinear vmas could be handled more efficiently, using
1813  * a placeholder, but handle it in the same way until a need is shown.
1814  * It is important to search the prio_tree before nonlinear list: a vma
1815  * may become nonlinear and be shifted from prio_tree to nonlinear list
1816  * while the lock is dropped; but never shifted from list to prio_tree.
1817  *
1818  * In order to make forward progress despite restarting the search,
1819  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1820  * quickly skip it next time around.  Since the prio_tree search only
1821  * shows us those vmas affected by unmapping the range in question, we
1822  * can't efficiently keep all vmas in step with mapping->truncate_count:
1823  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1824  * mapping->truncate_count and vma->vm_truncate_count are protected by
1825  * i_mmap_lock.
1826  *
1827  * In order to make forward progress despite repeatedly restarting some
1828  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1829  * and restart from that address when we reach that vma again.  It might
1830  * have been split or merged, shrunk or extended, but never shifted: so
1831  * restart_addr remains valid so long as it remains in the vma's range.
1832  * unmap_mapping_range forces truncate_count to leap over page-aligned
1833  * values so we can save vma's restart_addr in its truncate_count field.
1834  */
1835 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1836
1837 static void reset_vma_truncate_counts(struct address_space *mapping)
1838 {
1839         struct vm_area_struct *vma;
1840         struct prio_tree_iter iter;
1841
1842         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1843                 vma->vm_truncate_count = 0;
1844         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1845                 vma->vm_truncate_count = 0;
1846 }
1847
1848 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1849                 unsigned long start_addr, unsigned long end_addr,
1850                 struct zap_details *details)
1851 {
1852         unsigned long restart_addr;
1853         int need_break;
1854
1855         /*
1856          * files that support invalidating or truncating portions of the
1857          * file from under mmaped areas must have their ->fault function
1858          * return a locked page (and set VM_FAULT_LOCKED in the return).
1859          * This provides synchronisation against concurrent unmapping here.
1860          */
1861
1862 again:
1863         restart_addr = vma->vm_truncate_count;
1864         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1865                 start_addr = restart_addr;
1866                 if (start_addr >= end_addr) {
1867                         /* Top of vma has been split off since last time */
1868                         vma->vm_truncate_count = details->truncate_count;
1869                         return 0;
1870                 }
1871         }
1872
1873         restart_addr = zap_page_range(vma, start_addr,
1874                                         end_addr - start_addr, details);
1875         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
1876
1877         if (restart_addr >= end_addr) {
1878                 /* We have now completed this vma: mark it so */
1879                 vma->vm_truncate_count = details->truncate_count;
1880                 if (!need_break)
1881                         return 0;
1882         } else {
1883                 /* Note restart_addr in vma's truncate_count field */
1884                 vma->vm_truncate_count = restart_addr;
1885                 if (!need_break)
1886                         goto again;
1887         }
1888
1889         spin_unlock(details->i_mmap_lock);
1890         cond_resched();
1891         spin_lock(details->i_mmap_lock);
1892         return -EINTR;
1893 }
1894
1895 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1896                                             struct zap_details *details)
1897 {
1898         struct vm_area_struct *vma;
1899         struct prio_tree_iter iter;
1900         pgoff_t vba, vea, zba, zea;
1901
1902 restart:
1903         vma_prio_tree_foreach(vma, &iter, root,
1904                         details->first_index, details->last_index) {
1905                 /* Skip quickly over those we have already dealt with */
1906                 if (vma->vm_truncate_count == details->truncate_count)
1907                         continue;
1908
1909                 vba = vma->vm_pgoff;
1910                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1911                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1912                 zba = details->first_index;
1913                 if (zba < vba)
1914                         zba = vba;
1915                 zea = details->last_index;
1916                 if (zea > vea)
1917                         zea = vea;
1918
1919                 if (unmap_mapping_range_vma(vma,
1920                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1921                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1922                                 details) < 0)
1923                         goto restart;
1924         }
1925 }
1926
1927 static inline void unmap_mapping_range_list(struct list_head *head,
1928                                             struct zap_details *details)
1929 {
1930         struct vm_area_struct *vma;
1931
1932         /*
1933          * In nonlinear VMAs there is no correspondence between virtual address
1934          * offset and file offset.  So we must perform an exhaustive search
1935          * across *all* the pages in each nonlinear VMA, not just the pages
1936          * whose virtual address lies outside the file truncation point.
1937          */
1938 restart:
1939         list_for_each_entry(vma, head, shared.vm_set.list) {
1940                 /* Skip quickly over those we have already dealt with */
1941                 if (vma->vm_truncate_count == details->truncate_count)
1942                         continue;
1943                 details->nonlinear_vma = vma;
1944                 if (unmap_mapping_range_vma(vma, vma->vm_start,
1945                                         vma->vm_end, details) < 0)
1946                         goto restart;
1947         }
1948 }
1949
1950 /**
1951  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
1952  * @mapping: the address space containing mmaps to be unmapped.
1953  * @holebegin: byte in first page to unmap, relative to the start of
1954  * the underlying file.  This will be rounded down to a PAGE_SIZE
1955  * boundary.  Note that this is different from vmtruncate(), which
1956  * must keep the partial page.  In contrast, we must get rid of
1957  * partial pages.
1958  * @holelen: size of prospective hole in bytes.  This will be rounded
1959  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1960  * end of the file.
1961  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1962  * but 0 when invalidating pagecache, don't throw away private data.
1963  */
1964 void unmap_mapping_range(struct address_space *mapping,
1965                 loff_t const holebegin, loff_t const holelen, int even_cows)
1966 {
1967         struct zap_details details;
1968         pgoff_t hba = holebegin >> PAGE_SHIFT;
1969         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1970
1971         /* Check for overflow. */
1972         if (sizeof(holelen) > sizeof(hlen)) {
1973                 long long holeend =
1974                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1975                 if (holeend & ~(long long)ULONG_MAX)
1976                         hlen = ULONG_MAX - hba + 1;
1977         }
1978
1979         details.check_mapping = even_cows? NULL: mapping;
1980         details.nonlinear_vma = NULL;
1981         details.first_index = hba;
1982         details.last_index = hba + hlen - 1;
1983         if (details.last_index < details.first_index)
1984                 details.last_index = ULONG_MAX;
1985         details.i_mmap_lock = &mapping->i_mmap_lock;
1986
1987         spin_lock(&mapping->i_mmap_lock);
1988
1989         /* Protect against endless unmapping loops */
1990         mapping->truncate_count++;
1991         if (unlikely(is_restart_addr(mapping->truncate_count))) {
1992                 if (mapping->truncate_count == 0)
1993                         reset_vma_truncate_counts(mapping);
1994                 mapping->truncate_count++;
1995         }
1996         details.truncate_count = mapping->truncate_count;
1997
1998         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1999                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2000         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2001                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2002         spin_unlock(&mapping->i_mmap_lock);
2003 }
2004 EXPORT_SYMBOL(unmap_mapping_range);
2005
2006 /**
2007  * vmtruncate - unmap mappings "freed" by truncate() syscall
2008  * @inode: inode of the file used
2009  * @offset: file offset to start truncating
2010  *
2011  * NOTE! We have to be ready to update the memory sharing
2012  * between the file and the memory map for a potential last
2013  * incomplete page.  Ugly, but necessary.
2014  */
2015 int vmtruncate(struct inode * inode, loff_t offset)
2016 {
2017         if (inode->i_size < offset) {
2018                 unsigned long limit;
2019
2020                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2021                 if (limit != RLIM_INFINITY && offset > limit)
2022                         goto out_sig;
2023                 if (offset > inode->i_sb->s_maxbytes)
2024                         goto out_big;
2025                 i_size_write(inode, offset);
2026         } else {
2027                 struct address_space *mapping = inode->i_mapping;
2028
2029                 /*
2030                  * truncation of in-use swapfiles is disallowed - it would
2031                  * cause subsequent swapout to scribble on the now-freed
2032                  * blocks.
2033                  */
2034                 if (IS_SWAPFILE(inode))
2035                         return -ETXTBSY;
2036                 i_size_write(inode, offset);
2037
2038                 /*
2039                  * unmap_mapping_range is called twice, first simply for
2040                  * efficiency so that truncate_inode_pages does fewer
2041                  * single-page unmaps.  However after this first call, and
2042                  * before truncate_inode_pages finishes, it is possible for
2043                  * private pages to be COWed, which remain after
2044                  * truncate_inode_pages finishes, hence the second
2045                  * unmap_mapping_range call must be made for correctness.
2046                  */
2047                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2048                 truncate_inode_pages(mapping, offset);
2049                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2050         }
2051
2052         if (inode->i_op && inode->i_op->truncate)
2053                 inode->i_op->truncate(inode);
2054         return 0;
2055
2056 out_sig:
2057         send_sig(SIGXFSZ, current, 0);
2058 out_big:
2059         return -EFBIG;
2060 }
2061 EXPORT_SYMBOL(vmtruncate);
2062
2063 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2064 {
2065         struct address_space *mapping = inode->i_mapping;
2066
2067         /*
2068          * If the underlying filesystem is not going to provide
2069          * a way to truncate a range of blocks (punch a hole) -
2070          * we should return failure right now.
2071          */
2072         if (!inode->i_op || !inode->i_op->truncate_range)
2073                 return -ENOSYS;
2074
2075         mutex_lock(&inode->i_mutex);
2076         down_write(&inode->i_alloc_sem);
2077         unmap_mapping_range(mapping, offset, (end - offset), 1);
2078         truncate_inode_pages_range(mapping, offset, end);
2079         unmap_mapping_range(mapping, offset, (end - offset), 1);
2080         inode->i_op->truncate_range(inode, offset, end);
2081         up_write(&inode->i_alloc_sem);
2082         mutex_unlock(&inode->i_mutex);
2083
2084         return 0;
2085 }
2086
2087 /*
2088  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2089  * but allow concurrent faults), and pte mapped but not yet locked.
2090  * We return with mmap_sem still held, but pte unmapped and unlocked.
2091  */
2092 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2093                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2094                 int write_access, pte_t orig_pte)
2095 {
2096         spinlock_t *ptl;
2097         struct page *page;
2098         swp_entry_t entry;
2099         pte_t pte;
2100         int ret = 0;
2101
2102         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2103                 goto out;
2104
2105         entry = pte_to_swp_entry(orig_pte);
2106         if (is_migration_entry(entry)) {
2107                 migration_entry_wait(mm, pmd, address);
2108                 goto out;
2109         }
2110         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2111         page = lookup_swap_cache(entry);
2112         if (!page) {
2113                 grab_swap_token(); /* Contend for token _before_ read-in */
2114                 page = swapin_readahead(entry,
2115                                         GFP_HIGHUSER_MOVABLE, vma, address);
2116                 if (!page) {
2117                         /*
2118                          * Back out if somebody else faulted in this pte
2119                          * while we released the pte lock.
2120                          */
2121                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2122                         if (likely(pte_same(*page_table, orig_pte)))
2123                                 ret = VM_FAULT_OOM;
2124                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2125                         goto unlock;
2126                 }
2127
2128                 /* Had to read the page from swap area: Major fault */
2129                 ret = VM_FAULT_MAJOR;
2130                 count_vm_event(PGMAJFAULT);
2131         }
2132
2133         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2134                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2135                 ret = VM_FAULT_OOM;
2136                 goto out;
2137         }
2138
2139         mark_page_accessed(page);
2140         lock_page(page);
2141         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2142
2143         /*
2144          * Back out if somebody else already faulted in this pte.
2145          */
2146         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2147         if (unlikely(!pte_same(*page_table, orig_pte)))
2148                 goto out_nomap;
2149
2150         if (unlikely(!PageUptodate(page))) {
2151                 ret = VM_FAULT_SIGBUS;
2152                 goto out_nomap;
2153         }
2154
2155         /* The page isn't present yet, go ahead with the fault. */
2156
2157         inc_mm_counter(mm, anon_rss);
2158         pte = mk_pte(page, vma->vm_page_prot);
2159         if (write_access && can_share_swap_page(page)) {
2160                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2161                 write_access = 0;
2162         }
2163
2164         flush_icache_page(vma, page);
2165         set_pte_at(mm, address, page_table, pte);
2166         page_add_anon_rmap(page, vma, address);
2167
2168         swap_free(entry);
2169         if (vm_swap_full())
2170                 remove_exclusive_swap_page(page);
2171         unlock_page(page);
2172
2173         if (write_access) {
2174                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2175                 if (ret & VM_FAULT_ERROR)
2176                         ret &= VM_FAULT_ERROR;
2177                 goto out;
2178         }
2179
2180         /* No need to invalidate - it was non-present before */
2181         update_mmu_cache(vma, address, pte);
2182 unlock:
2183         pte_unmap_unlock(page_table, ptl);
2184 out:
2185         return ret;
2186 out_nomap:
2187         mem_cgroup_uncharge_page(page);
2188         pte_unmap_unlock(page_table, ptl);
2189         unlock_page(page);
2190         page_cache_release(page);
2191         return ret;
2192 }
2193
2194 /*
2195  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2196  * but allow concurrent faults), and pte mapped but not yet locked.
2197  * We return with mmap_sem still held, but pte unmapped and unlocked.
2198  */
2199 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2200                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2201                 int write_access)
2202 {
2203         struct page *page;
2204         spinlock_t *ptl;
2205         pte_t entry;
2206
2207         /* Allocate our own private page. */
2208         pte_unmap(page_table);
2209
2210         if (unlikely(anon_vma_prepare(vma)))
2211                 goto oom;
2212         page = alloc_zeroed_user_highpage_movable(vma, address);
2213         if (!page)
2214                 goto oom;
2215         __SetPageUptodate(page);
2216
2217         if (mem_cgroup_charge(page, mm, GFP_KERNEL))
2218                 goto oom_free_page;
2219
2220         entry = mk_pte(page, vma->vm_page_prot);
2221         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2222
2223         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2224         if (!pte_none(*page_table))
2225                 goto release;
2226         inc_mm_counter(mm, anon_rss);
2227         lru_cache_add_active(page);
2228         page_add_new_anon_rmap(page, vma, address);
2229         set_pte_at(mm, address, page_table, entry);
2230
2231         /* No need to invalidate - it was non-present before */
2232         update_mmu_cache(vma, address, entry);
2233 unlock:
2234         pte_unmap_unlock(page_table, ptl);
2235         return 0;
2236 release:
2237         mem_cgroup_uncharge_page(page);
2238         page_cache_release(page);
2239         goto unlock;
2240 oom_free_page:
2241         page_cache_release(page);
2242 oom:
2243         return VM_FAULT_OOM;
2244 }
2245
2246 /*
2247  * __do_fault() tries to create a new page mapping. It aggressively
2248  * tries to share with existing pages, but makes a separate copy if
2249  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2250  * the next page fault.
2251  *
2252  * As this is called only for pages that do not currently exist, we
2253  * do not need to flush old virtual caches or the TLB.
2254  *
2255  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2256  * but allow concurrent faults), and pte neither mapped nor locked.
2257  * We return with mmap_sem still held, but pte unmapped and unlocked.
2258  */
2259 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2260                 unsigned long address, pmd_t *pmd,
2261                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2262 {
2263         pte_t *page_table;
2264         spinlock_t *ptl;
2265         struct page *page;
2266         pte_t entry;
2267         int anon = 0;
2268         struct page *dirty_page = NULL;
2269         struct vm_fault vmf;
2270         int ret;
2271         int page_mkwrite = 0;
2272
2273         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2274         vmf.pgoff = pgoff;
2275         vmf.flags = flags;
2276         vmf.page = NULL;
2277
2278         BUG_ON(vma->vm_flags & VM_PFNMAP);
2279
2280         ret = vma->vm_ops->fault(vma, &vmf);
2281         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2282                 return ret;
2283
2284         /*
2285          * For consistency in subsequent calls, make the faulted page always
2286          * locked.
2287          */
2288         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2289                 lock_page(vmf.page);
2290         else
2291                 VM_BUG_ON(!PageLocked(vmf.page));
2292
2293         /*
2294          * Should we do an early C-O-W break?
2295          */
2296         page = vmf.page;
2297         if (flags & FAULT_FLAG_WRITE) {
2298                 if (!(vma->vm_flags & VM_SHARED)) {
2299                         anon = 1;
2300                         if (unlikely(anon_vma_prepare(vma))) {
2301                                 ret = VM_FAULT_OOM;
2302                                 goto out;
2303                         }
2304                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2305                                                 vma, address);
2306                         if (!page) {
2307                                 ret = VM_FAULT_OOM;
2308                                 goto out;
2309                         }
2310                         copy_user_highpage(page, vmf.page, address, vma);
2311                         __SetPageUptodate(page);
2312                 } else {
2313                         /*
2314                          * If the page will be shareable, see if the backing
2315                          * address space wants to know that the page is about
2316                          * to become writable
2317                          */
2318                         if (vma->vm_ops->page_mkwrite) {
2319                                 unlock_page(page);
2320                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2321                                         ret = VM_FAULT_SIGBUS;
2322                                         anon = 1; /* no anon but release vmf.page */
2323                                         goto out_unlocked;
2324                                 }
2325                                 lock_page(page);
2326                                 /*
2327                                  * XXX: this is not quite right (racy vs
2328                                  * invalidate) to unlock and relock the page
2329                                  * like this, however a better fix requires
2330                                  * reworking page_mkwrite locking API, which
2331                                  * is better done later.
2332                                  */
2333                                 if (!page->mapping) {
2334                                         ret = 0;
2335                                         anon = 1; /* no anon but release vmf.page */
2336                                         goto out;
2337                                 }
2338                                 page_mkwrite = 1;
2339                         }
2340                 }
2341
2342         }
2343
2344         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2345                 ret = VM_FAULT_OOM;
2346                 goto out;
2347         }
2348
2349         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2350
2351         /*
2352          * This silly early PAGE_DIRTY setting removes a race
2353          * due to the bad i386 page protection. But it's valid
2354          * for other architectures too.
2355          *
2356          * Note that if write_access is true, we either now have
2357          * an exclusive copy of the page, or this is a shared mapping,
2358          * so we can make it writable and dirty to avoid having to
2359          * handle that later.
2360          */
2361         /* Only go through if we didn't race with anybody else... */
2362         if (likely(pte_same(*page_table, orig_pte))) {
2363                 flush_icache_page(vma, page);
2364                 entry = mk_pte(page, vma->vm_page_prot);
2365                 if (flags & FAULT_FLAG_WRITE)
2366                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2367                 set_pte_at(mm, address, page_table, entry);
2368                 if (anon) {
2369                         inc_mm_counter(mm, anon_rss);
2370                         lru_cache_add_active(page);
2371                         page_add_new_anon_rmap(page, vma, address);
2372                 } else {
2373                         inc_mm_counter(mm, file_rss);
2374                         page_add_file_rmap(page);
2375                         if (flags & FAULT_FLAG_WRITE) {
2376                                 dirty_page = page;
2377                                 get_page(dirty_page);
2378                         }
2379                 }
2380
2381                 /* no need to invalidate: a not-present page won't be cached */
2382                 update_mmu_cache(vma, address, entry);
2383         } else {
2384                 mem_cgroup_uncharge_page(page);
2385                 if (anon)
2386                         page_cache_release(page);
2387                 else
2388                         anon = 1; /* no anon but release faulted_page */
2389         }
2390
2391         pte_unmap_unlock(page_table, ptl);
2392
2393 out:
2394         unlock_page(vmf.page);
2395 out_unlocked:
2396         if (anon)
2397                 page_cache_release(vmf.page);
2398         else if (dirty_page) {
2399                 if (vma->vm_file)
2400                         file_update_time(vma->vm_file);
2401
2402                 set_page_dirty_balance(dirty_page, page_mkwrite);
2403                 put_page(dirty_page);
2404         }
2405
2406         return ret;
2407 }
2408
2409 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2410                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2411                 int write_access, pte_t orig_pte)
2412 {
2413         pgoff_t pgoff = (((address & PAGE_MASK)
2414                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2415         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2416
2417         pte_unmap(page_table);
2418         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2419 }
2420
2421
2422 /*
2423  * do_no_pfn() tries to create a new page mapping for a page without
2424  * a struct_page backing it
2425  *
2426  * As this is called only for pages that do not currently exist, we
2427  * do not need to flush old virtual caches or the TLB.
2428  *
2429  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2430  * but allow concurrent faults), and pte mapped but not yet locked.
2431  * We return with mmap_sem still held, but pte unmapped and unlocked.
2432  *
2433  * It is expected that the ->nopfn handler always returns the same pfn
2434  * for a given virtual mapping.
2435  *
2436  * Mark this `noinline' to prevent it from bloating the main pagefault code.
2437  */
2438 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2439                      unsigned long address, pte_t *page_table, pmd_t *pmd,
2440                      int write_access)
2441 {
2442         spinlock_t *ptl;
2443         pte_t entry;
2444         unsigned long pfn;
2445
2446         pte_unmap(page_table);
2447         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2448         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2449
2450         pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2451
2452         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2453
2454         if (unlikely(pfn == NOPFN_OOM))
2455                 return VM_FAULT_OOM;
2456         else if (unlikely(pfn == NOPFN_SIGBUS))
2457                 return VM_FAULT_SIGBUS;
2458         else if (unlikely(pfn == NOPFN_REFAULT))
2459                 return 0;
2460
2461         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2462
2463         /* Only go through if we didn't race with anybody else... */
2464         if (pte_none(*page_table)) {
2465                 entry = pfn_pte(pfn, vma->vm_page_prot);
2466                 if (write_access)
2467                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2468                 set_pte_at(mm, address, page_table, entry);
2469         }
2470         pte_unmap_unlock(page_table, ptl);
2471         return 0;
2472 }
2473
2474 /*
2475  * Fault of a previously existing named mapping. Repopulate the pte
2476  * from the encoded file_pte if possible. This enables swappable
2477  * nonlinear vmas.
2478  *
2479  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2480  * but allow concurrent faults), and pte mapped but not yet locked.
2481  * We return with mmap_sem still held, but pte unmapped and unlocked.
2482  */
2483 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2484                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2485                 int write_access, pte_t orig_pte)
2486 {
2487         unsigned int flags = FAULT_FLAG_NONLINEAR |
2488                                 (write_access ? FAULT_FLAG_WRITE : 0);
2489         pgoff_t pgoff;
2490
2491         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2492                 return 0;
2493
2494         if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2495                         !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2496                 /*
2497                  * Page table corrupted: show pte and kill process.
2498                  */
2499                 print_bad_pte(vma, orig_pte, address);
2500                 return VM_FAULT_OOM;
2501         }
2502
2503         pgoff = pte_to_pgoff(orig_pte);
2504         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2505 }
2506
2507 /*
2508  * These routines also need to handle stuff like marking pages dirty
2509  * and/or accessed for architectures that don't do it in hardware (most
2510  * RISC architectures).  The early dirtying is also good on the i386.
2511  *
2512  * There is also a hook called "update_mmu_cache()" that architectures
2513  * with external mmu caches can use to update those (ie the Sparc or
2514  * PowerPC hashed page tables that act as extended TLBs).
2515  *
2516  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2517  * but allow concurrent faults), and pte mapped but not yet locked.
2518  * We return with mmap_sem still held, but pte unmapped and unlocked.
2519  */
2520 static inline int handle_pte_fault(struct mm_struct *mm,
2521                 struct vm_area_struct *vma, unsigned long address,
2522                 pte_t *pte, pmd_t *pmd, int write_access)
2523 {
2524         pte_t entry;
2525         spinlock_t *ptl;
2526
2527         entry = *pte;
2528         if (!pte_present(entry)) {
2529                 if (pte_none(entry)) {
2530                         if (vma->vm_ops) {
2531                                 if (likely(vma->vm_ops->fault))
2532                                         return do_linear_fault(mm, vma, address,
2533                                                 pte, pmd, write_access, entry);
2534                                 if (unlikely(vma->vm_ops->nopfn))
2535                                         return do_no_pfn(mm, vma, address, pte,
2536                                                          pmd, write_access);
2537                         }
2538                         return do_anonymous_page(mm, vma, address,
2539                                                  pte, pmd, write_access);
2540                 }
2541                 if (pte_file(entry))
2542                         return do_nonlinear_fault(mm, vma, address,
2543                                         pte, pmd, write_access, entry);
2544                 return do_swap_page(mm, vma, address,
2545                                         pte, pmd, write_access, entry);
2546         }
2547
2548         ptl = pte_lockptr(mm, pmd);
2549         spin_lock(ptl);
2550         if (unlikely(!pte_same(*pte, entry)))
2551                 goto unlock;
2552         if (write_access) {
2553                 if (!pte_write(entry))
2554                         return do_wp_page(mm, vma, address,
2555                                         pte, pmd, ptl, entry);
2556                 entry = pte_mkdirty(entry);
2557         }
2558         entry = pte_mkyoung(entry);
2559         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2560                 update_mmu_cache(vma, address, entry);
2561         } else {
2562                 /*
2563                  * This is needed only for protection faults but the arch code
2564                  * is not yet telling us if this is a protection fault or not.
2565                  * This still avoids useless tlb flushes for .text page faults
2566                  * with threads.
2567                  */
2568                 if (write_access)
2569                         flush_tlb_page(vma, address);
2570         }
2571 unlock:
2572         pte_unmap_unlock(pte, ptl);
2573         return 0;
2574 }
2575
2576 /*
2577  * By the time we get here, we already hold the mm semaphore
2578  */
2579 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2580                 unsigned long address, int write_access)
2581 {
2582         pgd_t *pgd;
2583         pud_t *pud;
2584         pmd_t *pmd;
2585         pte_t *pte;
2586
2587         __set_current_state(TASK_RUNNING);
2588
2589         count_vm_event(PGFAULT);
2590
2591         if (unlikely(is_vm_hugetlb_page(vma)))
2592                 return hugetlb_fault(mm, vma, address, write_access);
2593
2594         pgd = pgd_offset(mm, address);
2595         pud = pud_alloc(mm, pgd, address);
2596         if (!pud)
2597                 return VM_FAULT_OOM;
2598         pmd = pmd_alloc(mm, pud, address);
2599         if (!pmd)
2600                 return VM_FAULT_OOM;
2601         pte = pte_alloc_map(mm, pmd, address);
2602         if (!pte)
2603                 return VM_FAULT_OOM;
2604
2605         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2606 }
2607
2608 #ifndef __PAGETABLE_PUD_FOLDED
2609 /*
2610  * Allocate page upper directory.
2611  * We've already handled the fast-path in-line.
2612  */
2613 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2614 {
2615         pud_t *new = pud_alloc_one(mm, address);
2616         if (!new)
2617                 return -ENOMEM;
2618
2619         spin_lock(&mm->page_table_lock);
2620         if (pgd_present(*pgd))          /* Another has populated it */
2621                 pud_free(mm, new);
2622         else
2623                 pgd_populate(mm, pgd, new);
2624         spin_unlock(&mm->page_table_lock);
2625         return 0;
2626 }
2627 #endif /* __PAGETABLE_PUD_FOLDED */
2628
2629 #ifndef __PAGETABLE_PMD_FOLDED
2630 /*
2631  * Allocate page middle directory.
2632  * We've already handled the fast-path in-line.
2633  */
2634 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2635 {
2636         pmd_t *new = pmd_alloc_one(mm, address);
2637         if (!new)
2638                 return -ENOMEM;
2639
2640         spin_lock(&mm->page_table_lock);
2641 #ifndef __ARCH_HAS_4LEVEL_HACK
2642         if (pud_present(*pud))          /* Another has populated it */
2643                 pmd_free(mm, new);
2644         else
2645                 pud_populate(mm, pud, new);
2646 #else
2647         if (pgd_present(*pud))          /* Another has populated it */
2648                 pmd_free(mm, new);
2649         else
2650                 pgd_populate(mm, pud, new);
2651 #endif /* __ARCH_HAS_4LEVEL_HACK */
2652         spin_unlock(&mm->page_table_lock);
2653         return 0;
2654 }
2655 #endif /* __PAGETABLE_PMD_FOLDED */
2656
2657 int make_pages_present(unsigned long addr, unsigned long end)
2658 {
2659         int ret, len, write;
2660         struct vm_area_struct * vma;
2661
2662         vma = find_vma(current->mm, addr);
2663         if (!vma)
2664                 return -1;
2665         write = (vma->vm_flags & VM_WRITE) != 0;
2666         BUG_ON(addr >= end);
2667         BUG_ON(end > vma->vm_end);
2668         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2669         ret = get_user_pages(current, current->mm, addr,
2670                         len, write, 0, NULL, NULL);
2671         if (ret < 0)
2672                 return ret;
2673         return ret == len ? 0 : -1;
2674 }
2675
2676 #if !defined(__HAVE_ARCH_GATE_AREA)
2677
2678 #if defined(AT_SYSINFO_EHDR)
2679 static struct vm_area_struct gate_vma;
2680
2681 static int __init gate_vma_init(void)
2682 {
2683         gate_vma.vm_mm = NULL;
2684         gate_vma.vm_start = FIXADDR_USER_START;
2685         gate_vma.vm_end = FIXADDR_USER_END;
2686         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2687         gate_vma.vm_page_prot = __P101;
2688         /*
2689          * Make sure the vDSO gets into every core dump.
2690          * Dumping its contents makes post-mortem fully interpretable later
2691          * without matching up the same kernel and hardware config to see
2692          * what PC values meant.
2693          */
2694         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2695         return 0;
2696 }
2697 __initcall(gate_vma_init);
2698 #endif
2699
2700 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2701 {
2702 #ifdef AT_SYSINFO_EHDR
2703         return &gate_vma;
2704 #else
2705         return NULL;
2706 #endif
2707 }
2708
2709 int in_gate_area_no_task(unsigned long addr)
2710 {
2711 #ifdef AT_SYSINFO_EHDR
2712         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2713                 return 1;
2714 #endif
2715         return 0;
2716 }
2717
2718 #endif  /* __HAVE_ARCH_GATE_AREA */
2719
2720 /*
2721  * Access another process' address space.
2722  * Source/target buffer must be kernel space,
2723  * Do not walk the page table directly, use get_user_pages
2724  */
2725 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2726 {
2727         struct mm_struct *mm;
2728         struct vm_area_struct *vma;
2729         struct page *page;
2730         void *old_buf = buf;
2731
2732         mm = get_task_mm(tsk);
2733         if (!mm)
2734                 return 0;
2735
2736         down_read(&mm->mmap_sem);
2737         /* ignore errors, just check how much was successfully transferred */
2738         while (len) {
2739                 int bytes, ret, offset;
2740                 void *maddr;
2741
2742                 ret = get_user_pages(tsk, mm, addr, 1,
2743                                 write, 1, &page, &vma);
2744                 if (ret <= 0)
2745                         break;
2746
2747                 bytes = len;
2748                 offset = addr & (PAGE_SIZE-1);
2749                 if (bytes > PAGE_SIZE-offset)
2750                         bytes = PAGE_SIZE-offset;
2751
2752                 maddr = kmap(page);
2753                 if (write) {
2754                         copy_to_user_page(vma, page, addr,
2755                                           maddr + offset, buf, bytes);
2756                         set_page_dirty_lock(page);
2757                 } else {
2758                         copy_from_user_page(vma, page, addr,
2759                                             buf, maddr + offset, bytes);
2760                 }
2761                 kunmap(page);
2762                 page_cache_release(page);
2763                 len -= bytes;
2764                 buf += bytes;
2765                 addr += bytes;
2766         }
2767         up_read(&mm->mmap_sem);
2768         mmput(mm);
2769
2770         return buf - old_buf;
2771 }
2772
2773 /*
2774  * Print the name of a VMA.
2775  */
2776 void print_vma_addr(char *prefix, unsigned long ip)
2777 {
2778         struct mm_struct *mm = current->mm;
2779         struct vm_area_struct *vma;
2780
2781         /*
2782          * Do not print if we are in atomic
2783          * contexts (in exception stacks, etc.):
2784          */
2785         if (preempt_count())
2786                 return;
2787
2788         down_read(&mm->mmap_sem);
2789         vma = find_vma(mm, ip);
2790         if (vma && vma->vm_file) {
2791                 struct file *f = vma->vm_file;
2792                 char *buf = (char *)__get_free_page(GFP_KERNEL);
2793                 if (buf) {
2794                         char *p, *s;
2795
2796                         p = d_path(&f->f_path, buf, PAGE_SIZE);
2797                         if (IS_ERR(p))
2798                                 p = "?";
2799                         s = strrchr(p, '/');
2800                         if (s)
2801                                 p = s+1;
2802                         printk("%s%s[%lx+%lx]", prefix, p,
2803                                         vma->vm_start,
2804                                         vma->vm_end - vma->vm_start);
2805                         free_page((unsigned long)buf);
2806                 }
2807         }
2808         up_read(&current->mm->mmap_sem);
2809 }