netfilter: fix double-free and use-after free
[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         /*
315          * Ensure all pte setup (eg. pte page lock and page clearing) are
316          * visible before the pte is made visible to other CPUs by being
317          * put into page tables.
318          *
319          * The other side of the story is the pointer chasing in the page
320          * table walking code (when walking the page table without locking;
321          * ie. most of the time). Fortunately, these data accesses consist
322          * of a chain of data-dependent loads, meaning most CPUs (alpha
323          * being the notable exception) will already guarantee loads are
324          * seen in-order. See the alpha page table accessors for the
325          * smp_read_barrier_depends() barriers in page table walking code.
326          */
327         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
328
329         spin_lock(&mm->page_table_lock);
330         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
331                 mm->nr_ptes++;
332                 pmd_populate(mm, pmd, new);
333                 new = NULL;
334         }
335         spin_unlock(&mm->page_table_lock);
336         if (new)
337                 pte_free(mm, new);
338         return 0;
339 }
340
341 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
342 {
343         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
344         if (!new)
345                 return -ENOMEM;
346
347         smp_wmb(); /* See comment in __pte_alloc */
348
349         spin_lock(&init_mm.page_table_lock);
350         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
351                 pmd_populate_kernel(&init_mm, pmd, new);
352                 new = NULL;
353         }
354         spin_unlock(&init_mm.page_table_lock);
355         if (new)
356                 pte_free_kernel(&init_mm, new);
357         return 0;
358 }
359
360 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
361 {
362         if (file_rss)
363                 add_mm_counter(mm, file_rss, file_rss);
364         if (anon_rss)
365                 add_mm_counter(mm, anon_rss, anon_rss);
366 }
367
368 /*
369  * This function is called to print an error when a bad pte
370  * is found. For example, we might have a PFN-mapped pte in
371  * a region that doesn't allow it.
372  *
373  * The calling function must still handle the error.
374  */
375 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
376 {
377         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
378                         "vm_flags = %lx, vaddr = %lx\n",
379                 (long long)pte_val(pte),
380                 (vma->vm_mm == current->mm ? current->comm : "???"),
381                 vma->vm_flags, vaddr);
382         dump_stack();
383 }
384
385 static inline int is_cow_mapping(unsigned int flags)
386 {
387         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
388 }
389
390 /*
391  * vm_normal_page -- This function gets the "struct page" associated with a pte.
392  *
393  * "Special" mappings do not wish to be associated with a "struct page" (either
394  * it doesn't exist, or it exists but they don't want to touch it). In this
395  * case, NULL is returned here. "Normal" mappings do have a struct page.
396  *
397  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
398  * pte bit, in which case this function is trivial. Secondly, an architecture
399  * may not have a spare pte bit, which requires a more complicated scheme,
400  * described below.
401  *
402  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
403  * special mapping (even if there are underlying and valid "struct pages").
404  * COWed pages of a VM_PFNMAP are always normal.
405  *
406  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
407  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
408  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
409  * mapping will always honor the rule
410  *
411  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
412  *
413  * And for normal mappings this is false.
414  *
415  * This restricts such mappings to be a linear translation from virtual address
416  * to pfn. To get around this restriction, we allow arbitrary mappings so long
417  * as the vma is not a COW mapping; in that case, we know that all ptes are
418  * special (because none can have been COWed).
419  *
420  *
421  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
422  *
423  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
424  * page" backing, however the difference is that _all_ pages with a struct
425  * page (that is, those where pfn_valid is true) are refcounted and considered
426  * normal pages by the VM. The disadvantage is that pages are refcounted
427  * (which can be slower and simply not an option for some PFNMAP users). The
428  * advantage is that we don't have to follow the strict linearity rule of
429  * PFNMAP mappings in order to support COWable mappings.
430  *
431  */
432 #ifdef __HAVE_ARCH_PTE_SPECIAL
433 # define HAVE_PTE_SPECIAL 1
434 #else
435 # define HAVE_PTE_SPECIAL 0
436 #endif
437 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
438                                 pte_t pte)
439 {
440         unsigned long pfn;
441
442         if (HAVE_PTE_SPECIAL) {
443                 if (likely(!pte_special(pte))) {
444                         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
445                         return pte_page(pte);
446                 }
447                 VM_BUG_ON(!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)));
448                 return NULL;
449         }
450
451         /* !HAVE_PTE_SPECIAL case follows: */
452
453         pfn = pte_pfn(pte);
454
455         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
456                 if (vma->vm_flags & VM_MIXEDMAP) {
457                         if (!pfn_valid(pfn))
458                                 return NULL;
459                         goto out;
460                 } else {
461                         unsigned long off;
462                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
463                         if (pfn == vma->vm_pgoff + off)
464                                 return NULL;
465                         if (!is_cow_mapping(vma->vm_flags))
466                                 return NULL;
467                 }
468         }
469
470         VM_BUG_ON(!pfn_valid(pfn));
471
472         /*
473          * NOTE! We still have PageReserved() pages in the page tables.
474          *
475          * eg. VDSO mappings can cause them to exist.
476          */
477 out:
478         return pfn_to_page(pfn);
479 }
480
481 /*
482  * copy one vm_area from one task to the other. Assumes the page tables
483  * already present in the new task to be cleared in the whole range
484  * covered by this vma.
485  */
486
487 static inline void
488 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
489                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
490                 unsigned long addr, int *rss)
491 {
492         unsigned long vm_flags = vma->vm_flags;
493         pte_t pte = *src_pte;
494         struct page *page;
495
496         /* pte contains position in swap or file, so copy. */
497         if (unlikely(!pte_present(pte))) {
498                 if (!pte_file(pte)) {
499                         swp_entry_t entry = pte_to_swp_entry(pte);
500
501                         swap_duplicate(entry);
502                         /* make sure dst_mm is on swapoff's mmlist. */
503                         if (unlikely(list_empty(&dst_mm->mmlist))) {
504                                 spin_lock(&mmlist_lock);
505                                 if (list_empty(&dst_mm->mmlist))
506                                         list_add(&dst_mm->mmlist,
507                                                  &src_mm->mmlist);
508                                 spin_unlock(&mmlist_lock);
509                         }
510                         if (is_write_migration_entry(entry) &&
511                                         is_cow_mapping(vm_flags)) {
512                                 /*
513                                  * COW mappings require pages in both parent
514                                  * and child to be set to read.
515                                  */
516                                 make_migration_entry_read(&entry);
517                                 pte = swp_entry_to_pte(entry);
518                                 set_pte_at(src_mm, addr, src_pte, pte);
519                         }
520                 }
521                 goto out_set_pte;
522         }
523
524         /*
525          * If it's a COW mapping, write protect it both
526          * in the parent and the child
527          */
528         if (is_cow_mapping(vm_flags)) {
529                 ptep_set_wrprotect(src_mm, addr, src_pte);
530                 pte = pte_wrprotect(pte);
531         }
532
533         /*
534          * If it's a shared mapping, mark it clean in
535          * the child
536          */
537         if (vm_flags & VM_SHARED)
538                 pte = pte_mkclean(pte);
539         pte = pte_mkold(pte);
540
541         page = vm_normal_page(vma, addr, pte);
542         if (page) {
543                 get_page(page);
544                 page_dup_rmap(page, vma, addr);
545                 rss[!!PageAnon(page)]++;
546         }
547
548 out_set_pte:
549         set_pte_at(dst_mm, addr, dst_pte, pte);
550 }
551
552 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
553                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
554                 unsigned long addr, unsigned long end)
555 {
556         pte_t *src_pte, *dst_pte;
557         spinlock_t *src_ptl, *dst_ptl;
558         int progress = 0;
559         int rss[2];
560
561 again:
562         rss[1] = rss[0] = 0;
563         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
564         if (!dst_pte)
565                 return -ENOMEM;
566         src_pte = pte_offset_map_nested(src_pmd, addr);
567         src_ptl = pte_lockptr(src_mm, src_pmd);
568         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
569         arch_enter_lazy_mmu_mode();
570
571         do {
572                 /*
573                  * We are holding two locks at this point - either of them
574                  * could generate latencies in another task on another CPU.
575                  */
576                 if (progress >= 32) {
577                         progress = 0;
578                         if (need_resched() ||
579                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
580                                 break;
581                 }
582                 if (pte_none(*src_pte)) {
583                         progress++;
584                         continue;
585                 }
586                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
587                 progress += 8;
588         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
589
590         arch_leave_lazy_mmu_mode();
591         spin_unlock(src_ptl);
592         pte_unmap_nested(src_pte - 1);
593         add_mm_rss(dst_mm, rss[0], rss[1]);
594         pte_unmap_unlock(dst_pte - 1, dst_ptl);
595         cond_resched();
596         if (addr != end)
597                 goto again;
598         return 0;
599 }
600
601 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
602                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
603                 unsigned long addr, unsigned long end)
604 {
605         pmd_t *src_pmd, *dst_pmd;
606         unsigned long next;
607
608         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
609         if (!dst_pmd)
610                 return -ENOMEM;
611         src_pmd = pmd_offset(src_pud, addr);
612         do {
613                 next = pmd_addr_end(addr, end);
614                 if (pmd_none_or_clear_bad(src_pmd))
615                         continue;
616                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
617                                                 vma, addr, next))
618                         return -ENOMEM;
619         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
620         return 0;
621 }
622
623 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
624                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
625                 unsigned long addr, unsigned long end)
626 {
627         pud_t *src_pud, *dst_pud;
628         unsigned long next;
629
630         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
631         if (!dst_pud)
632                 return -ENOMEM;
633         src_pud = pud_offset(src_pgd, addr);
634         do {
635                 next = pud_addr_end(addr, end);
636                 if (pud_none_or_clear_bad(src_pud))
637                         continue;
638                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
639                                                 vma, addr, next))
640                         return -ENOMEM;
641         } while (dst_pud++, src_pud++, addr = next, addr != end);
642         return 0;
643 }
644
645 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
646                 struct vm_area_struct *vma)
647 {
648         pgd_t *src_pgd, *dst_pgd;
649         unsigned long next;
650         unsigned long addr = vma->vm_start;
651         unsigned long end = vma->vm_end;
652
653         /*
654          * Don't copy ptes where a page fault will fill them correctly.
655          * Fork becomes much lighter when there are big shared or private
656          * readonly mappings. The tradeoff is that copy_page_range is more
657          * efficient than faulting.
658          */
659         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
660                 if (!vma->anon_vma)
661                         return 0;
662         }
663
664         if (is_vm_hugetlb_page(vma))
665                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
666
667         dst_pgd = pgd_offset(dst_mm, addr);
668         src_pgd = pgd_offset(src_mm, addr);
669         do {
670                 next = pgd_addr_end(addr, end);
671                 if (pgd_none_or_clear_bad(src_pgd))
672                         continue;
673                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
674                                                 vma, addr, next))
675                         return -ENOMEM;
676         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
677         return 0;
678 }
679
680 static unsigned long zap_pte_range(struct mmu_gather *tlb,
681                                 struct vm_area_struct *vma, pmd_t *pmd,
682                                 unsigned long addr, unsigned long end,
683                                 long *zap_work, struct zap_details *details)
684 {
685         struct mm_struct *mm = tlb->mm;
686         pte_t *pte;
687         spinlock_t *ptl;
688         int file_rss = 0;
689         int anon_rss = 0;
690
691         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
692         arch_enter_lazy_mmu_mode();
693         do {
694                 pte_t ptent = *pte;
695                 if (pte_none(ptent)) {
696                         (*zap_work)--;
697                         continue;
698                 }
699
700                 (*zap_work) -= PAGE_SIZE;
701
702                 if (pte_present(ptent)) {
703                         struct page *page;
704
705                         page = vm_normal_page(vma, addr, ptent);
706                         if (unlikely(details) && page) {
707                                 /*
708                                  * unmap_shared_mapping_pages() wants to
709                                  * invalidate cache without truncating:
710                                  * unmap shared but keep private pages.
711                                  */
712                                 if (details->check_mapping &&
713                                     details->check_mapping != page->mapping)
714                                         continue;
715                                 /*
716                                  * Each page->index must be checked when
717                                  * invalidating or truncating nonlinear.
718                                  */
719                                 if (details->nonlinear_vma &&
720                                     (page->index < details->first_index ||
721                                      page->index > details->last_index))
722                                         continue;
723                         }
724                         ptent = ptep_get_and_clear_full(mm, addr, pte,
725                                                         tlb->fullmm);
726                         tlb_remove_tlb_entry(tlb, pte, addr);
727                         if (unlikely(!page))
728                                 continue;
729                         if (unlikely(details) && details->nonlinear_vma
730                             && linear_page_index(details->nonlinear_vma,
731                                                 addr) != page->index)
732                                 set_pte_at(mm, addr, pte,
733                                            pgoff_to_pte(page->index));
734                         if (PageAnon(page))
735                                 anon_rss--;
736                         else {
737                                 if (pte_dirty(ptent))
738                                         set_page_dirty(page);
739                                 if (pte_young(ptent))
740                                         SetPageReferenced(page);
741                                 file_rss--;
742                         }
743                         page_remove_rmap(page, vma);
744                         tlb_remove_page(tlb, page);
745                         continue;
746                 }
747                 /*
748                  * If details->check_mapping, we leave swap entries;
749                  * if details->nonlinear_vma, we leave file entries.
750                  */
751                 if (unlikely(details))
752                         continue;
753                 if (!pte_file(ptent))
754                         free_swap_and_cache(pte_to_swp_entry(ptent));
755                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
756         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
757
758         add_mm_rss(mm, file_rss, anon_rss);
759         arch_leave_lazy_mmu_mode();
760         pte_unmap_unlock(pte - 1, ptl);
761
762         return addr;
763 }
764
765 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
766                                 struct vm_area_struct *vma, pud_t *pud,
767                                 unsigned long addr, unsigned long end,
768                                 long *zap_work, struct zap_details *details)
769 {
770         pmd_t *pmd;
771         unsigned long next;
772
773         pmd = pmd_offset(pud, addr);
774         do {
775                 next = pmd_addr_end(addr, end);
776                 if (pmd_none_or_clear_bad(pmd)) {
777                         (*zap_work)--;
778                         continue;
779                 }
780                 next = zap_pte_range(tlb, vma, pmd, addr, next,
781                                                 zap_work, details);
782         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
783
784         return addr;
785 }
786
787 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
788                                 struct vm_area_struct *vma, pgd_t *pgd,
789                                 unsigned long addr, unsigned long end,
790                                 long *zap_work, struct zap_details *details)
791 {
792         pud_t *pud;
793         unsigned long next;
794
795         pud = pud_offset(pgd, addr);
796         do {
797                 next = pud_addr_end(addr, end);
798                 if (pud_none_or_clear_bad(pud)) {
799                         (*zap_work)--;
800                         continue;
801                 }
802                 next = zap_pmd_range(tlb, vma, pud, addr, next,
803                                                 zap_work, details);
804         } while (pud++, addr = next, (addr != end && *zap_work > 0));
805
806         return addr;
807 }
808
809 static unsigned long unmap_page_range(struct mmu_gather *tlb,
810                                 struct vm_area_struct *vma,
811                                 unsigned long addr, unsigned long end,
812                                 long *zap_work, struct zap_details *details)
813 {
814         pgd_t *pgd;
815         unsigned long next;
816
817         if (details && !details->check_mapping && !details->nonlinear_vma)
818                 details = NULL;
819
820         BUG_ON(addr >= end);
821         tlb_start_vma(tlb, vma);
822         pgd = pgd_offset(vma->vm_mm, addr);
823         do {
824                 next = pgd_addr_end(addr, end);
825                 if (pgd_none_or_clear_bad(pgd)) {
826                         (*zap_work)--;
827                         continue;
828                 }
829                 next = zap_pud_range(tlb, vma, pgd, addr, next,
830                                                 zap_work, details);
831         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
832         tlb_end_vma(tlb, vma);
833
834         return addr;
835 }
836
837 #ifdef CONFIG_PREEMPT
838 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
839 #else
840 /* No preempt: go for improved straight-line efficiency */
841 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
842 #endif
843
844 /**
845  * unmap_vmas - unmap a range of memory covered by a list of vma's
846  * @tlbp: address of the caller's struct mmu_gather
847  * @vma: the starting vma
848  * @start_addr: virtual address at which to start unmapping
849  * @end_addr: virtual address at which to end unmapping
850  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
851  * @details: details of nonlinear truncation or shared cache invalidation
852  *
853  * Returns the end address of the unmapping (restart addr if interrupted).
854  *
855  * Unmap all pages in the vma list.
856  *
857  * We aim to not hold locks for too long (for scheduling latency reasons).
858  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
859  * return the ending mmu_gather to the caller.
860  *
861  * Only addresses between `start' and `end' will be unmapped.
862  *
863  * The VMA list must be sorted in ascending virtual address order.
864  *
865  * unmap_vmas() assumes that the caller will flush the whole unmapped address
866  * range after unmap_vmas() returns.  So the only responsibility here is to
867  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
868  * drops the lock and schedules.
869  */
870 unsigned long unmap_vmas(struct mmu_gather **tlbp,
871                 struct vm_area_struct *vma, unsigned long start_addr,
872                 unsigned long end_addr, unsigned long *nr_accounted,
873                 struct zap_details *details)
874 {
875         long zap_work = ZAP_BLOCK_SIZE;
876         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
877         int tlb_start_valid = 0;
878         unsigned long start = start_addr;
879         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
880         int fullmm = (*tlbp)->fullmm;
881
882         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
883                 unsigned long end;
884
885                 start = max(vma->vm_start, start_addr);
886                 if (start >= vma->vm_end)
887                         continue;
888                 end = min(vma->vm_end, end_addr);
889                 if (end <= vma->vm_start)
890                         continue;
891
892                 if (vma->vm_flags & VM_ACCOUNT)
893                         *nr_accounted += (end - start) >> PAGE_SHIFT;
894
895                 while (start != end) {
896                         if (!tlb_start_valid) {
897                                 tlb_start = start;
898                                 tlb_start_valid = 1;
899                         }
900
901                         if (unlikely(is_vm_hugetlb_page(vma))) {
902                                 unmap_hugepage_range(vma, start, end);
903                                 zap_work -= (end - start) /
904                                                 (HPAGE_SIZE / PAGE_SIZE);
905                                 start = end;
906                         } else
907                                 start = unmap_page_range(*tlbp, vma,
908                                                 start, end, &zap_work, details);
909
910                         if (zap_work > 0) {
911                                 BUG_ON(start != end);
912                                 break;
913                         }
914
915                         tlb_finish_mmu(*tlbp, tlb_start, start);
916
917                         if (need_resched() ||
918                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
919                                 if (i_mmap_lock) {
920                                         *tlbp = NULL;
921                                         goto out;
922                                 }
923                                 cond_resched();
924                         }
925
926                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
927                         tlb_start_valid = 0;
928                         zap_work = ZAP_BLOCK_SIZE;
929                 }
930         }
931 out:
932         return start;   /* which is now the end (or restart) address */
933 }
934
935 /**
936  * zap_page_range - remove user pages in a given range
937  * @vma: vm_area_struct holding the applicable pages
938  * @address: starting address of pages to zap
939  * @size: number of bytes to zap
940  * @details: details of nonlinear truncation or shared cache invalidation
941  */
942 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
943                 unsigned long size, struct zap_details *details)
944 {
945         struct mm_struct *mm = vma->vm_mm;
946         struct mmu_gather *tlb;
947         unsigned long end = address + size;
948         unsigned long nr_accounted = 0;
949
950         lru_add_drain();
951         tlb = tlb_gather_mmu(mm, 0);
952         update_hiwater_rss(mm);
953         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
954         if (tlb)
955                 tlb_finish_mmu(tlb, address, end);
956         return end;
957 }
958
959 /*
960  * Do a quick page-table lookup for a single page.
961  */
962 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
963                         unsigned int flags)
964 {
965         pgd_t *pgd;
966         pud_t *pud;
967         pmd_t *pmd;
968         pte_t *ptep, pte;
969         spinlock_t *ptl;
970         struct page *page;
971         struct mm_struct *mm = vma->vm_mm;
972
973         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
974         if (!IS_ERR(page)) {
975                 BUG_ON(flags & FOLL_GET);
976                 goto out;
977         }
978
979         page = NULL;
980         pgd = pgd_offset(mm, address);
981         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
982                 goto no_page_table;
983
984         pud = pud_offset(pgd, address);
985         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
986                 goto no_page_table;
987         
988         pmd = pmd_offset(pud, address);
989         if (pmd_none(*pmd))
990                 goto no_page_table;
991
992         if (pmd_huge(*pmd)) {
993                 BUG_ON(flags & FOLL_GET);
994                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
995                 goto out;
996         }
997
998         if (unlikely(pmd_bad(*pmd)))
999                 goto no_page_table;
1000
1001         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1002
1003         pte = *ptep;
1004         if (!pte_present(pte))
1005                 goto no_page;
1006         if ((flags & FOLL_WRITE) && !pte_write(pte))
1007                 goto unlock;
1008         page = vm_normal_page(vma, address, pte);
1009         if (unlikely(!page))
1010                 goto bad_page;
1011
1012         if (flags & FOLL_GET)
1013                 get_page(page);
1014         if (flags & FOLL_TOUCH) {
1015                 if ((flags & FOLL_WRITE) &&
1016                     !pte_dirty(pte) && !PageDirty(page))
1017                         set_page_dirty(page);
1018                 mark_page_accessed(page);
1019         }
1020 unlock:
1021         pte_unmap_unlock(ptep, ptl);
1022 out:
1023         return page;
1024
1025 bad_page:
1026         pte_unmap_unlock(ptep, ptl);
1027         return ERR_PTR(-EFAULT);
1028
1029 no_page:
1030         pte_unmap_unlock(ptep, ptl);
1031         if (!pte_none(pte))
1032                 return page;
1033         /* Fall through to ZERO_PAGE handling */
1034 no_page_table:
1035         /*
1036          * When core dumping an enormous anonymous area that nobody
1037          * has touched so far, we don't want to allocate page tables.
1038          */
1039         if (flags & FOLL_ANON) {
1040                 page = ZERO_PAGE(0);
1041                 if (flags & FOLL_GET)
1042                         get_page(page);
1043                 BUG_ON(flags & FOLL_WRITE);
1044         }
1045         return page;
1046 }
1047
1048 /* Can we do the FOLL_ANON optimization? */
1049 static inline int use_zero_page(struct vm_area_struct *vma)
1050 {
1051         /*
1052          * We don't want to optimize FOLL_ANON for make_pages_present()
1053          * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1054          * we want to get the page from the page tables to make sure
1055          * that we serialize and update with any other user of that
1056          * mapping.
1057          */
1058         if (vma->vm_flags & (VM_LOCKED | VM_SHARED))
1059                 return 0;
1060         /*
1061          * And if we have a fault or a nopfn routine, it's not an
1062          * anonymous region.
1063          */
1064         return !vma->vm_ops ||
1065                 (!vma->vm_ops->fault && !vma->vm_ops->nopfn);
1066 }
1067
1068 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1069                 unsigned long start, int len, int write, int force,
1070                 struct page **pages, struct vm_area_struct **vmas)
1071 {
1072         int i;
1073         unsigned int vm_flags;
1074
1075         if (len <= 0)
1076                 return 0;
1077         /* 
1078          * Require read or write permissions.
1079          * If 'force' is set, we only require the "MAY" flags.
1080          */
1081         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1082         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1083         i = 0;
1084
1085         do {
1086                 struct vm_area_struct *vma;
1087                 unsigned int foll_flags;
1088
1089                 vma = find_extend_vma(mm, start);
1090                 if (!vma && in_gate_area(tsk, start)) {
1091                         unsigned long pg = start & PAGE_MASK;
1092                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1093                         pgd_t *pgd;
1094                         pud_t *pud;
1095                         pmd_t *pmd;
1096                         pte_t *pte;
1097                         if (write) /* user gate pages are read-only */
1098                                 return i ? : -EFAULT;
1099                         if (pg > TASK_SIZE)
1100                                 pgd = pgd_offset_k(pg);
1101                         else
1102                                 pgd = pgd_offset_gate(mm, pg);
1103                         BUG_ON(pgd_none(*pgd));
1104                         pud = pud_offset(pgd, pg);
1105                         BUG_ON(pud_none(*pud));
1106                         pmd = pmd_offset(pud, pg);
1107                         if (pmd_none(*pmd))
1108                                 return i ? : -EFAULT;
1109                         pte = pte_offset_map(pmd, pg);
1110                         if (pte_none(*pte)) {
1111                                 pte_unmap(pte);
1112                                 return i ? : -EFAULT;
1113                         }
1114                         if (pages) {
1115                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1116                                 pages[i] = page;
1117                                 if (page)
1118                                         get_page(page);
1119                         }
1120                         pte_unmap(pte);
1121                         if (vmas)
1122                                 vmas[i] = gate_vma;
1123                         i++;
1124                         start += PAGE_SIZE;
1125                         len--;
1126                         continue;
1127                 }
1128
1129                 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1130                                 || !(vm_flags & vma->vm_flags))
1131                         return i ? : -EFAULT;
1132
1133                 if (is_vm_hugetlb_page(vma)) {
1134                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1135                                                 &start, &len, i, write);
1136                         continue;
1137                 }
1138
1139                 foll_flags = FOLL_TOUCH;
1140                 if (pages)
1141                         foll_flags |= FOLL_GET;
1142                 if (!write && use_zero_page(vma))
1143                         foll_flags |= FOLL_ANON;
1144
1145                 do {
1146                         struct page *page;
1147
1148                         /*
1149                          * If tsk is ooming, cut off its access to large memory
1150                          * allocations. It has a pending SIGKILL, but it can't
1151                          * be processed until returning to user space.
1152                          */
1153                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1154                                 return i ? i : -ENOMEM;
1155
1156                         if (write)
1157                                 foll_flags |= FOLL_WRITE;
1158
1159                         cond_resched();
1160                         while (!(page = follow_page(vma, start, foll_flags))) {
1161                                 int ret;
1162                                 ret = handle_mm_fault(mm, vma, start,
1163                                                 foll_flags & FOLL_WRITE);
1164                                 if (ret & VM_FAULT_ERROR) {
1165                                         if (ret & VM_FAULT_OOM)
1166                                                 return i ? i : -ENOMEM;
1167                                         else if (ret & VM_FAULT_SIGBUS)
1168                                                 return i ? i : -EFAULT;
1169                                         BUG();
1170                                 }
1171                                 if (ret & VM_FAULT_MAJOR)
1172                                         tsk->maj_flt++;
1173                                 else
1174                                         tsk->min_flt++;
1175
1176                                 /*
1177                                  * The VM_FAULT_WRITE bit tells us that
1178                                  * do_wp_page has broken COW when necessary,
1179                                  * even if maybe_mkwrite decided not to set
1180                                  * pte_write. We can thus safely do subsequent
1181                                  * page lookups as if they were reads.
1182                                  */
1183                                 if (ret & VM_FAULT_WRITE)
1184                                         foll_flags &= ~FOLL_WRITE;
1185
1186                                 cond_resched();
1187                         }
1188                         if (IS_ERR(page))
1189                                 return i ? i : PTR_ERR(page);
1190                         if (pages) {
1191                                 pages[i] = page;
1192
1193                                 flush_anon_page(vma, page, start);
1194                                 flush_dcache_page(page);
1195                         }
1196                         if (vmas)
1197                                 vmas[i] = vma;
1198                         i++;
1199                         start += PAGE_SIZE;
1200                         len--;
1201                 } while (len && start < vma->vm_end);
1202         } while (len);
1203         return i;
1204 }
1205 EXPORT_SYMBOL(get_user_pages);
1206
1207 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1208                         spinlock_t **ptl)
1209 {
1210         pgd_t * pgd = pgd_offset(mm, addr);
1211         pud_t * pud = pud_alloc(mm, pgd, addr);
1212         if (pud) {
1213                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1214                 if (pmd)
1215                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1216         }
1217         return NULL;
1218 }
1219
1220 /*
1221  * This is the old fallback for page remapping.
1222  *
1223  * For historical reasons, it only allows reserved pages. Only
1224  * old drivers should use this, and they needed to mark their
1225  * pages reserved for the old functions anyway.
1226  */
1227 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1228                         struct page *page, pgprot_t prot)
1229 {
1230         struct mm_struct *mm = vma->vm_mm;
1231         int retval;
1232         pte_t *pte;
1233         spinlock_t *ptl;
1234
1235         retval = mem_cgroup_charge(page, mm, GFP_KERNEL);
1236         if (retval)
1237                 goto out;
1238
1239         retval = -EINVAL;
1240         if (PageAnon(page))
1241                 goto out_uncharge;
1242         retval = -ENOMEM;
1243         flush_dcache_page(page);
1244         pte = get_locked_pte(mm, addr, &ptl);
1245         if (!pte)
1246                 goto out_uncharge;
1247         retval = -EBUSY;
1248         if (!pte_none(*pte))
1249                 goto out_unlock;
1250
1251         /* Ok, finally just insert the thing.. */
1252         get_page(page);
1253         inc_mm_counter(mm, file_rss);
1254         page_add_file_rmap(page);
1255         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1256
1257         retval = 0;
1258         pte_unmap_unlock(pte, ptl);
1259         return retval;
1260 out_unlock:
1261         pte_unmap_unlock(pte, ptl);
1262 out_uncharge:
1263         mem_cgroup_uncharge_page(page);
1264 out:
1265         return retval;
1266 }
1267
1268 /**
1269  * vm_insert_page - insert single page into user vma
1270  * @vma: user vma to map to
1271  * @addr: target user address of this page
1272  * @page: source kernel page
1273  *
1274  * This allows drivers to insert individual pages they've allocated
1275  * into a user vma.
1276  *
1277  * The page has to be a nice clean _individual_ kernel allocation.
1278  * If you allocate a compound page, you need to have marked it as
1279  * such (__GFP_COMP), or manually just split the page up yourself
1280  * (see split_page()).
1281  *
1282  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1283  * took an arbitrary page protection parameter. This doesn't allow
1284  * that. Your vma protection will have to be set up correctly, which
1285  * means that if you want a shared writable mapping, you'd better
1286  * ask for a shared writable mapping!
1287  *
1288  * The page does not need to be reserved.
1289  */
1290 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1291                         struct page *page)
1292 {
1293         if (addr < vma->vm_start || addr >= vma->vm_end)
1294                 return -EFAULT;
1295         if (!page_count(page))
1296                 return -EINVAL;
1297         vma->vm_flags |= VM_INSERTPAGE;
1298         return insert_page(vma, addr, page, vma->vm_page_prot);
1299 }
1300 EXPORT_SYMBOL(vm_insert_page);
1301
1302 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1303                         unsigned long pfn, pgprot_t prot)
1304 {
1305         struct mm_struct *mm = vma->vm_mm;
1306         int retval;
1307         pte_t *pte, entry;
1308         spinlock_t *ptl;
1309
1310         retval = -ENOMEM;
1311         pte = get_locked_pte(mm, addr, &ptl);
1312         if (!pte)
1313                 goto out;
1314         retval = -EBUSY;
1315         if (!pte_none(*pte))
1316                 goto out_unlock;
1317
1318         /* Ok, finally just insert the thing.. */
1319         entry = pte_mkspecial(pfn_pte(pfn, prot));
1320         set_pte_at(mm, addr, pte, entry);
1321         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1322
1323         retval = 0;
1324 out_unlock:
1325         pte_unmap_unlock(pte, ptl);
1326 out:
1327         return retval;
1328 }
1329
1330 /**
1331  * vm_insert_pfn - insert single pfn into user vma
1332  * @vma: user vma to map to
1333  * @addr: target user address of this page
1334  * @pfn: source kernel pfn
1335  *
1336  * Similar to vm_inert_page, this allows drivers to insert individual pages
1337  * they've allocated into a user vma. Same comments apply.
1338  *
1339  * This function should only be called from a vm_ops->fault handler, and
1340  * in that case the handler should return NULL.
1341  */
1342 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1343                         unsigned long pfn)
1344 {
1345         /*
1346          * Technically, architectures with pte_special can avoid all these
1347          * restrictions (same for remap_pfn_range).  However we would like
1348          * consistency in testing and feature parity among all, so we should
1349          * try to keep these invariants in place for everybody.
1350          */
1351         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1352         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1353                                                 (VM_PFNMAP|VM_MIXEDMAP));
1354         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1355         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1356
1357         if (addr < vma->vm_start || addr >= vma->vm_end)
1358                 return -EFAULT;
1359         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1360 }
1361 EXPORT_SYMBOL(vm_insert_pfn);
1362
1363 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1364                         unsigned long pfn)
1365 {
1366         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1367
1368         if (addr < vma->vm_start || addr >= vma->vm_end)
1369                 return -EFAULT;
1370
1371         /*
1372          * If we don't have pte special, then we have to use the pfn_valid()
1373          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1374          * refcount the page if pfn_valid is true (hence insert_page rather
1375          * than insert_pfn).
1376          */
1377         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1378                 struct page *page;
1379
1380                 page = pfn_to_page(pfn);
1381                 return insert_page(vma, addr, page, vma->vm_page_prot);
1382         }
1383         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1384 }
1385 EXPORT_SYMBOL(vm_insert_mixed);
1386
1387 /*
1388  * maps a range of physical memory into the requested pages. the old
1389  * mappings are removed. any references to nonexistent pages results
1390  * in null mappings (currently treated as "copy-on-access")
1391  */
1392 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1393                         unsigned long addr, unsigned long end,
1394                         unsigned long pfn, pgprot_t prot)
1395 {
1396         pte_t *pte;
1397         spinlock_t *ptl;
1398
1399         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1400         if (!pte)
1401                 return -ENOMEM;
1402         arch_enter_lazy_mmu_mode();
1403         do {
1404                 BUG_ON(!pte_none(*pte));
1405                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1406                 pfn++;
1407         } while (pte++, addr += PAGE_SIZE, addr != end);
1408         arch_leave_lazy_mmu_mode();
1409         pte_unmap_unlock(pte - 1, ptl);
1410         return 0;
1411 }
1412
1413 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1414                         unsigned long addr, unsigned long end,
1415                         unsigned long pfn, pgprot_t prot)
1416 {
1417         pmd_t *pmd;
1418         unsigned long next;
1419
1420         pfn -= addr >> PAGE_SHIFT;
1421         pmd = pmd_alloc(mm, pud, addr);
1422         if (!pmd)
1423                 return -ENOMEM;
1424         do {
1425                 next = pmd_addr_end(addr, end);
1426                 if (remap_pte_range(mm, pmd, addr, next,
1427                                 pfn + (addr >> PAGE_SHIFT), prot))
1428                         return -ENOMEM;
1429         } while (pmd++, addr = next, addr != end);
1430         return 0;
1431 }
1432
1433 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1434                         unsigned long addr, unsigned long end,
1435                         unsigned long pfn, pgprot_t prot)
1436 {
1437         pud_t *pud;
1438         unsigned long next;
1439
1440         pfn -= addr >> PAGE_SHIFT;
1441         pud = pud_alloc(mm, pgd, addr);
1442         if (!pud)
1443                 return -ENOMEM;
1444         do {
1445                 next = pud_addr_end(addr, end);
1446                 if (remap_pmd_range(mm, pud, addr, next,
1447                                 pfn + (addr >> PAGE_SHIFT), prot))
1448                         return -ENOMEM;
1449         } while (pud++, addr = next, addr != end);
1450         return 0;
1451 }
1452
1453 /**
1454  * remap_pfn_range - remap kernel memory to userspace
1455  * @vma: user vma to map to
1456  * @addr: target user address to start at
1457  * @pfn: physical address of kernel memory
1458  * @size: size of map area
1459  * @prot: page protection flags for this mapping
1460  *
1461  *  Note: this is only safe if the mm semaphore is held when called.
1462  */
1463 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1464                     unsigned long pfn, unsigned long size, pgprot_t prot)
1465 {
1466         pgd_t *pgd;
1467         unsigned long next;
1468         unsigned long end = addr + PAGE_ALIGN(size);
1469         struct mm_struct *mm = vma->vm_mm;
1470         int err;
1471
1472         /*
1473          * Physically remapped pages are special. Tell the
1474          * rest of the world about it:
1475          *   VM_IO tells people not to look at these pages
1476          *      (accesses can have side effects).
1477          *   VM_RESERVED is specified all over the place, because
1478          *      in 2.4 it kept swapout's vma scan off this vma; but
1479          *      in 2.6 the LRU scan won't even find its pages, so this
1480          *      flag means no more than count its pages in reserved_vm,
1481          *      and omit it from core dump, even when VM_IO turned off.
1482          *   VM_PFNMAP tells the core MM that the base pages are just
1483          *      raw PFN mappings, and do not have a "struct page" associated
1484          *      with them.
1485          *
1486          * There's a horrible special case to handle copy-on-write
1487          * behaviour that some programs depend on. We mark the "original"
1488          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1489          */
1490         if (is_cow_mapping(vma->vm_flags)) {
1491                 if (addr != vma->vm_start || end != vma->vm_end)
1492                         return -EINVAL;
1493                 vma->vm_pgoff = pfn;
1494         }
1495
1496         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1497
1498         BUG_ON(addr >= end);
1499         pfn -= addr >> PAGE_SHIFT;
1500         pgd = pgd_offset(mm, addr);
1501         flush_cache_range(vma, addr, end);
1502         do {
1503                 next = pgd_addr_end(addr, end);
1504                 err = remap_pud_range(mm, pgd, addr, next,
1505                                 pfn + (addr >> PAGE_SHIFT), prot);
1506                 if (err)
1507                         break;
1508         } while (pgd++, addr = next, addr != end);
1509         return err;
1510 }
1511 EXPORT_SYMBOL(remap_pfn_range);
1512
1513 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1514                                      unsigned long addr, unsigned long end,
1515                                      pte_fn_t fn, void *data)
1516 {
1517         pte_t *pte;
1518         int err;
1519         pgtable_t token;
1520         spinlock_t *uninitialized_var(ptl);
1521
1522         pte = (mm == &init_mm) ?
1523                 pte_alloc_kernel(pmd, addr) :
1524                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1525         if (!pte)
1526                 return -ENOMEM;
1527
1528         BUG_ON(pmd_huge(*pmd));
1529
1530         token = pmd_pgtable(*pmd);
1531
1532         do {
1533                 err = fn(pte, token, addr, data);
1534                 if (err)
1535                         break;
1536         } while (pte++, addr += PAGE_SIZE, addr != end);
1537
1538         if (mm != &init_mm)
1539                 pte_unmap_unlock(pte-1, ptl);
1540         return err;
1541 }
1542
1543 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1544                                      unsigned long addr, unsigned long end,
1545                                      pte_fn_t fn, void *data)
1546 {
1547         pmd_t *pmd;
1548         unsigned long next;
1549         int err;
1550
1551         pmd = pmd_alloc(mm, pud, addr);
1552         if (!pmd)
1553                 return -ENOMEM;
1554         do {
1555                 next = pmd_addr_end(addr, end);
1556                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1557                 if (err)
1558                         break;
1559         } while (pmd++, addr = next, addr != end);
1560         return err;
1561 }
1562
1563 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1564                                      unsigned long addr, unsigned long end,
1565                                      pte_fn_t fn, void *data)
1566 {
1567         pud_t *pud;
1568         unsigned long next;
1569         int err;
1570
1571         pud = pud_alloc(mm, pgd, addr);
1572         if (!pud)
1573                 return -ENOMEM;
1574         do {
1575                 next = pud_addr_end(addr, end);
1576                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1577                 if (err)
1578                         break;
1579         } while (pud++, addr = next, addr != end);
1580         return err;
1581 }
1582
1583 /*
1584  * Scan a region of virtual memory, filling in page tables as necessary
1585  * and calling a provided function on each leaf page table.
1586  */
1587 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1588                         unsigned long size, pte_fn_t fn, void *data)
1589 {
1590         pgd_t *pgd;
1591         unsigned long next;
1592         unsigned long end = addr + size;
1593         int err;
1594
1595         BUG_ON(addr >= end);
1596         pgd = pgd_offset(mm, addr);
1597         do {
1598                 next = pgd_addr_end(addr, end);
1599                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1600                 if (err)
1601                         break;
1602         } while (pgd++, addr = next, addr != end);
1603         return err;
1604 }
1605 EXPORT_SYMBOL_GPL(apply_to_page_range);
1606
1607 /*
1608  * handle_pte_fault chooses page fault handler according to an entry
1609  * which was read non-atomically.  Before making any commitment, on
1610  * those architectures or configurations (e.g. i386 with PAE) which
1611  * might give a mix of unmatched parts, do_swap_page and do_file_page
1612  * must check under lock before unmapping the pte and proceeding
1613  * (but do_wp_page is only called after already making such a check;
1614  * and do_anonymous_page and do_no_page can safely check later on).
1615  */
1616 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1617                                 pte_t *page_table, pte_t orig_pte)
1618 {
1619         int same = 1;
1620 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1621         if (sizeof(pte_t) > sizeof(unsigned long)) {
1622                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1623                 spin_lock(ptl);
1624                 same = pte_same(*page_table, orig_pte);
1625                 spin_unlock(ptl);
1626         }
1627 #endif
1628         pte_unmap(page_table);
1629         return same;
1630 }
1631
1632 /*
1633  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1634  * servicing faults for write access.  In the normal case, do always want
1635  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1636  * that do not have writing enabled, when used by access_process_vm.
1637  */
1638 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1639 {
1640         if (likely(vma->vm_flags & VM_WRITE))
1641                 pte = pte_mkwrite(pte);
1642         return pte;
1643 }
1644
1645 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1646 {
1647         /*
1648          * If the source page was a PFN mapping, we don't have
1649          * a "struct page" for it. We do a best-effort copy by
1650          * just copying from the original user address. If that
1651          * fails, we just zero-fill it. Live with it.
1652          */
1653         if (unlikely(!src)) {
1654                 void *kaddr = kmap_atomic(dst, KM_USER0);
1655                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1656
1657                 /*
1658                  * This really shouldn't fail, because the page is there
1659                  * in the page tables. But it might just be unreadable,
1660                  * in which case we just give up and fill the result with
1661                  * zeroes.
1662                  */
1663                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1664                         memset(kaddr, 0, PAGE_SIZE);
1665                 kunmap_atomic(kaddr, KM_USER0);
1666                 flush_dcache_page(dst);
1667         } else
1668                 copy_user_highpage(dst, src, va, vma);
1669 }
1670
1671 /*
1672  * This routine handles present pages, when users try to write
1673  * to a shared page. It is done by copying the page to a new address
1674  * and decrementing the shared-page counter for the old page.
1675  *
1676  * Note that this routine assumes that the protection checks have been
1677  * done by the caller (the low-level page fault routine in most cases).
1678  * Thus we can safely just mark it writable once we've done any necessary
1679  * COW.
1680  *
1681  * We also mark the page dirty at this point even though the page will
1682  * change only once the write actually happens. This avoids a few races,
1683  * and potentially makes it more efficient.
1684  *
1685  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1686  * but allow concurrent faults), with pte both mapped and locked.
1687  * We return with mmap_sem still held, but pte unmapped and unlocked.
1688  */
1689 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1690                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1691                 spinlock_t *ptl, pte_t orig_pte)
1692 {
1693         struct page *old_page, *new_page;
1694         pte_t entry;
1695         int reuse = 0, ret = 0;
1696         int page_mkwrite = 0;
1697         struct page *dirty_page = NULL;
1698
1699         old_page = vm_normal_page(vma, address, orig_pte);
1700         if (!old_page) {
1701                 /*
1702                  * VM_MIXEDMAP !pfn_valid() case
1703                  *
1704                  * We should not cow pages in a shared writeable mapping.
1705                  * Just mark the pages writable as we can't do any dirty
1706                  * accounting on raw pfn maps.
1707                  */
1708                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1709                                      (VM_WRITE|VM_SHARED))
1710                         goto reuse;
1711                 goto gotten;
1712         }
1713
1714         /*
1715          * Take out anonymous pages first, anonymous shared vmas are
1716          * not dirty accountable.
1717          */
1718         if (PageAnon(old_page)) {
1719                 if (!TestSetPageLocked(old_page)) {
1720                         reuse = can_share_swap_page(old_page);
1721                         unlock_page(old_page);
1722                 }
1723         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1724                                         (VM_WRITE|VM_SHARED))) {
1725                 /*
1726                  * Only catch write-faults on shared writable pages,
1727                  * read-only shared pages can get COWed by
1728                  * get_user_pages(.write=1, .force=1).
1729                  */
1730                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1731                         /*
1732                          * Notify the address space that the page is about to
1733                          * become writable so that it can prohibit this or wait
1734                          * for the page to get into an appropriate state.
1735                          *
1736                          * We do this without the lock held, so that it can
1737                          * sleep if it needs to.
1738                          */
1739                         page_cache_get(old_page);
1740                         pte_unmap_unlock(page_table, ptl);
1741
1742                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1743                                 goto unwritable_page;
1744
1745                         /*
1746                          * Since we dropped the lock we need to revalidate
1747                          * the PTE as someone else may have changed it.  If
1748                          * they did, we just return, as we can count on the
1749                          * MMU to tell us if they didn't also make it writable.
1750                          */
1751                         page_table = pte_offset_map_lock(mm, pmd, address,
1752                                                          &ptl);
1753                         page_cache_release(old_page);
1754                         if (!pte_same(*page_table, orig_pte))
1755                                 goto unlock;
1756
1757                         page_mkwrite = 1;
1758                 }
1759                 dirty_page = old_page;
1760                 get_page(dirty_page);
1761                 reuse = 1;
1762         }
1763
1764         if (reuse) {
1765 reuse:
1766                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1767                 entry = pte_mkyoung(orig_pte);
1768                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1769                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1770                         update_mmu_cache(vma, address, entry);
1771                 ret |= VM_FAULT_WRITE;
1772                 goto unlock;
1773         }
1774
1775         /*
1776          * Ok, we need to copy. Oh, well..
1777          */
1778         page_cache_get(old_page);
1779 gotten:
1780         pte_unmap_unlock(page_table, ptl);
1781
1782         if (unlikely(anon_vma_prepare(vma)))
1783                 goto oom;
1784         VM_BUG_ON(old_page == ZERO_PAGE(0));
1785         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1786         if (!new_page)
1787                 goto oom;
1788         cow_user_page(new_page, old_page, address, vma);
1789         __SetPageUptodate(new_page);
1790
1791         if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
1792                 goto oom_free_new;
1793
1794         /*
1795          * Re-check the pte - we dropped the lock
1796          */
1797         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1798         if (likely(pte_same(*page_table, orig_pte))) {
1799                 if (old_page) {
1800                         if (!PageAnon(old_page)) {
1801                                 dec_mm_counter(mm, file_rss);
1802                                 inc_mm_counter(mm, anon_rss);
1803                         }
1804                 } else
1805                         inc_mm_counter(mm, anon_rss);
1806                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1807                 entry = mk_pte(new_page, vma->vm_page_prot);
1808                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1809                 /*
1810                  * Clear the pte entry and flush it first, before updating the
1811                  * pte with the new entry. This will avoid a race condition
1812                  * seen in the presence of one thread doing SMC and another
1813                  * thread doing COW.
1814                  */
1815                 ptep_clear_flush(vma, address, page_table);
1816                 set_pte_at(mm, address, page_table, entry);
1817                 update_mmu_cache(vma, address, entry);
1818                 lru_cache_add_active(new_page);
1819                 page_add_new_anon_rmap(new_page, vma, address);
1820
1821                 if (old_page) {
1822                         /*
1823                          * Only after switching the pte to the new page may
1824                          * we remove the mapcount here. Otherwise another
1825                          * process may come and find the rmap count decremented
1826                          * before the pte is switched to the new page, and
1827                          * "reuse" the old page writing into it while our pte
1828                          * here still points into it and can be read by other
1829                          * threads.
1830                          *
1831                          * The critical issue is to order this
1832                          * page_remove_rmap with the ptp_clear_flush above.
1833                          * Those stores are ordered by (if nothing else,)
1834                          * the barrier present in the atomic_add_negative
1835                          * in page_remove_rmap.
1836                          *
1837                          * Then the TLB flush in ptep_clear_flush ensures that
1838                          * no process can access the old page before the
1839                          * decremented mapcount is visible. And the old page
1840                          * cannot be reused until after the decremented
1841                          * mapcount is visible. So transitively, TLBs to
1842                          * old page will be flushed before it can be reused.
1843                          */
1844                         page_remove_rmap(old_page, vma);
1845                 }
1846
1847                 /* Free the old page.. */
1848                 new_page = old_page;
1849                 ret |= VM_FAULT_WRITE;
1850         } else
1851                 mem_cgroup_uncharge_page(new_page);
1852
1853         if (new_page)
1854                 page_cache_release(new_page);
1855         if (old_page)
1856                 page_cache_release(old_page);
1857 unlock:
1858         pte_unmap_unlock(page_table, ptl);
1859         if (dirty_page) {
1860                 if (vma->vm_file)
1861                         file_update_time(vma->vm_file);
1862
1863                 /*
1864                  * Yes, Virginia, this is actually required to prevent a race
1865                  * with clear_page_dirty_for_io() from clearing the page dirty
1866                  * bit after it clear all dirty ptes, but before a racing
1867                  * do_wp_page installs a dirty pte.
1868                  *
1869                  * do_no_page is protected similarly.
1870                  */
1871                 wait_on_page_locked(dirty_page);
1872                 set_page_dirty_balance(dirty_page, page_mkwrite);
1873                 put_page(dirty_page);
1874         }
1875         return ret;
1876 oom_free_new:
1877         page_cache_release(new_page);
1878 oom:
1879         if (old_page)
1880                 page_cache_release(old_page);
1881         return VM_FAULT_OOM;
1882
1883 unwritable_page:
1884         page_cache_release(old_page);
1885         return VM_FAULT_SIGBUS;
1886 }
1887
1888 /*
1889  * Helper functions for unmap_mapping_range().
1890  *
1891  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1892  *
1893  * We have to restart searching the prio_tree whenever we drop the lock,
1894  * since the iterator is only valid while the lock is held, and anyway
1895  * a later vma might be split and reinserted earlier while lock dropped.
1896  *
1897  * The list of nonlinear vmas could be handled more efficiently, using
1898  * a placeholder, but handle it in the same way until a need is shown.
1899  * It is important to search the prio_tree before nonlinear list: a vma
1900  * may become nonlinear and be shifted from prio_tree to nonlinear list
1901  * while the lock is dropped; but never shifted from list to prio_tree.
1902  *
1903  * In order to make forward progress despite restarting the search,
1904  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1905  * quickly skip it next time around.  Since the prio_tree search only
1906  * shows us those vmas affected by unmapping the range in question, we
1907  * can't efficiently keep all vmas in step with mapping->truncate_count:
1908  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1909  * mapping->truncate_count and vma->vm_truncate_count are protected by
1910  * i_mmap_lock.
1911  *
1912  * In order to make forward progress despite repeatedly restarting some
1913  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1914  * and restart from that address when we reach that vma again.  It might
1915  * have been split or merged, shrunk or extended, but never shifted: so
1916  * restart_addr remains valid so long as it remains in the vma's range.
1917  * unmap_mapping_range forces truncate_count to leap over page-aligned
1918  * values so we can save vma's restart_addr in its truncate_count field.
1919  */
1920 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1921
1922 static void reset_vma_truncate_counts(struct address_space *mapping)
1923 {
1924         struct vm_area_struct *vma;
1925         struct prio_tree_iter iter;
1926
1927         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1928                 vma->vm_truncate_count = 0;
1929         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1930                 vma->vm_truncate_count = 0;
1931 }
1932
1933 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1934                 unsigned long start_addr, unsigned long end_addr,
1935                 struct zap_details *details)
1936 {
1937         unsigned long restart_addr;
1938         int need_break;
1939
1940         /*
1941          * files that support invalidating or truncating portions of the
1942          * file from under mmaped areas must have their ->fault function
1943          * return a locked page (and set VM_FAULT_LOCKED in the return).
1944          * This provides synchronisation against concurrent unmapping here.
1945          */
1946
1947 again:
1948         restart_addr = vma->vm_truncate_count;
1949         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1950                 start_addr = restart_addr;
1951                 if (start_addr >= end_addr) {
1952                         /* Top of vma has been split off since last time */
1953                         vma->vm_truncate_count = details->truncate_count;
1954                         return 0;
1955                 }
1956         }
1957
1958         restart_addr = zap_page_range(vma, start_addr,
1959                                         end_addr - start_addr, details);
1960         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
1961
1962         if (restart_addr >= end_addr) {
1963                 /* We have now completed this vma: mark it so */
1964                 vma->vm_truncate_count = details->truncate_count;
1965                 if (!need_break)
1966                         return 0;
1967         } else {
1968                 /* Note restart_addr in vma's truncate_count field */
1969                 vma->vm_truncate_count = restart_addr;
1970                 if (!need_break)
1971                         goto again;
1972         }
1973
1974         spin_unlock(details->i_mmap_lock);
1975         cond_resched();
1976         spin_lock(details->i_mmap_lock);
1977         return -EINTR;
1978 }
1979
1980 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1981                                             struct zap_details *details)
1982 {
1983         struct vm_area_struct *vma;
1984         struct prio_tree_iter iter;
1985         pgoff_t vba, vea, zba, zea;
1986
1987 restart:
1988         vma_prio_tree_foreach(vma, &iter, root,
1989                         details->first_index, details->last_index) {
1990                 /* Skip quickly over those we have already dealt with */
1991                 if (vma->vm_truncate_count == details->truncate_count)
1992                         continue;
1993
1994                 vba = vma->vm_pgoff;
1995                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1996                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1997                 zba = details->first_index;
1998                 if (zba < vba)
1999                         zba = vba;
2000                 zea = details->last_index;
2001                 if (zea > vea)
2002                         zea = vea;
2003
2004                 if (unmap_mapping_range_vma(vma,
2005                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2006                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2007                                 details) < 0)
2008                         goto restart;
2009         }
2010 }
2011
2012 static inline void unmap_mapping_range_list(struct list_head *head,
2013                                             struct zap_details *details)
2014 {
2015         struct vm_area_struct *vma;
2016
2017         /*
2018          * In nonlinear VMAs there is no correspondence between virtual address
2019          * offset and file offset.  So we must perform an exhaustive search
2020          * across *all* the pages in each nonlinear VMA, not just the pages
2021          * whose virtual address lies outside the file truncation point.
2022          */
2023 restart:
2024         list_for_each_entry(vma, head, shared.vm_set.list) {
2025                 /* Skip quickly over those we have already dealt with */
2026                 if (vma->vm_truncate_count == details->truncate_count)
2027                         continue;
2028                 details->nonlinear_vma = vma;
2029                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2030                                         vma->vm_end, details) < 0)
2031                         goto restart;
2032         }
2033 }
2034
2035 /**
2036  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2037  * @mapping: the address space containing mmaps to be unmapped.
2038  * @holebegin: byte in first page to unmap, relative to the start of
2039  * the underlying file.  This will be rounded down to a PAGE_SIZE
2040  * boundary.  Note that this is different from vmtruncate(), which
2041  * must keep the partial page.  In contrast, we must get rid of
2042  * partial pages.
2043  * @holelen: size of prospective hole in bytes.  This will be rounded
2044  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2045  * end of the file.
2046  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2047  * but 0 when invalidating pagecache, don't throw away private data.
2048  */
2049 void unmap_mapping_range(struct address_space *mapping,
2050                 loff_t const holebegin, loff_t const holelen, int even_cows)
2051 {
2052         struct zap_details details;
2053         pgoff_t hba = holebegin >> PAGE_SHIFT;
2054         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2055
2056         /* Check for overflow. */
2057         if (sizeof(holelen) > sizeof(hlen)) {
2058                 long long holeend =
2059                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2060                 if (holeend & ~(long long)ULONG_MAX)
2061                         hlen = ULONG_MAX - hba + 1;
2062         }
2063
2064         details.check_mapping = even_cows? NULL: mapping;
2065         details.nonlinear_vma = NULL;
2066         details.first_index = hba;
2067         details.last_index = hba + hlen - 1;
2068         if (details.last_index < details.first_index)
2069                 details.last_index = ULONG_MAX;
2070         details.i_mmap_lock = &mapping->i_mmap_lock;
2071
2072         spin_lock(&mapping->i_mmap_lock);
2073
2074         /* Protect against endless unmapping loops */
2075         mapping->truncate_count++;
2076         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2077                 if (mapping->truncate_count == 0)
2078                         reset_vma_truncate_counts(mapping);
2079                 mapping->truncate_count++;
2080         }
2081         details.truncate_count = mapping->truncate_count;
2082
2083         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2084                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2085         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2086                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2087         spin_unlock(&mapping->i_mmap_lock);
2088 }
2089 EXPORT_SYMBOL(unmap_mapping_range);
2090
2091 /**
2092  * vmtruncate - unmap mappings "freed" by truncate() syscall
2093  * @inode: inode of the file used
2094  * @offset: file offset to start truncating
2095  *
2096  * NOTE! We have to be ready to update the memory sharing
2097  * between the file and the memory map for a potential last
2098  * incomplete page.  Ugly, but necessary.
2099  */
2100 int vmtruncate(struct inode * inode, loff_t offset)
2101 {
2102         if (inode->i_size < offset) {
2103                 unsigned long limit;
2104
2105                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2106                 if (limit != RLIM_INFINITY && offset > limit)
2107                         goto out_sig;
2108                 if (offset > inode->i_sb->s_maxbytes)
2109                         goto out_big;
2110                 i_size_write(inode, offset);
2111         } else {
2112                 struct address_space *mapping = inode->i_mapping;
2113
2114                 /*
2115                  * truncation of in-use swapfiles is disallowed - it would
2116                  * cause subsequent swapout to scribble on the now-freed
2117                  * blocks.
2118                  */
2119                 if (IS_SWAPFILE(inode))
2120                         return -ETXTBSY;
2121                 i_size_write(inode, offset);
2122
2123                 /*
2124                  * unmap_mapping_range is called twice, first simply for
2125                  * efficiency so that truncate_inode_pages does fewer
2126                  * single-page unmaps.  However after this first call, and
2127                  * before truncate_inode_pages finishes, it is possible for
2128                  * private pages to be COWed, which remain after
2129                  * truncate_inode_pages finishes, hence the second
2130                  * unmap_mapping_range call must be made for correctness.
2131                  */
2132                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2133                 truncate_inode_pages(mapping, offset);
2134                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2135         }
2136
2137         if (inode->i_op && inode->i_op->truncate)
2138                 inode->i_op->truncate(inode);
2139         return 0;
2140
2141 out_sig:
2142         send_sig(SIGXFSZ, current, 0);
2143 out_big:
2144         return -EFBIG;
2145 }
2146 EXPORT_SYMBOL(vmtruncate);
2147
2148 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2149 {
2150         struct address_space *mapping = inode->i_mapping;
2151
2152         /*
2153          * If the underlying filesystem is not going to provide
2154          * a way to truncate a range of blocks (punch a hole) -
2155          * we should return failure right now.
2156          */
2157         if (!inode->i_op || !inode->i_op->truncate_range)
2158                 return -ENOSYS;
2159
2160         mutex_lock(&inode->i_mutex);
2161         down_write(&inode->i_alloc_sem);
2162         unmap_mapping_range(mapping, offset, (end - offset), 1);
2163         truncate_inode_pages_range(mapping, offset, end);
2164         unmap_mapping_range(mapping, offset, (end - offset), 1);
2165         inode->i_op->truncate_range(inode, offset, end);
2166         up_write(&inode->i_alloc_sem);
2167         mutex_unlock(&inode->i_mutex);
2168
2169         return 0;
2170 }
2171
2172 /*
2173  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2174  * but allow concurrent faults), and pte mapped but not yet locked.
2175  * We return with mmap_sem still held, but pte unmapped and unlocked.
2176  */
2177 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2178                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2179                 int write_access, pte_t orig_pte)
2180 {
2181         spinlock_t *ptl;
2182         struct page *page;
2183         swp_entry_t entry;
2184         pte_t pte;
2185         int ret = 0;
2186
2187         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2188                 goto out;
2189
2190         entry = pte_to_swp_entry(orig_pte);
2191         if (is_migration_entry(entry)) {
2192                 migration_entry_wait(mm, pmd, address);
2193                 goto out;
2194         }
2195         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2196         page = lookup_swap_cache(entry);
2197         if (!page) {
2198                 grab_swap_token(); /* Contend for token _before_ read-in */
2199                 page = swapin_readahead(entry,
2200                                         GFP_HIGHUSER_MOVABLE, vma, address);
2201                 if (!page) {
2202                         /*
2203                          * Back out if somebody else faulted in this pte
2204                          * while we released the pte lock.
2205                          */
2206                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2207                         if (likely(pte_same(*page_table, orig_pte)))
2208                                 ret = VM_FAULT_OOM;
2209                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2210                         goto unlock;
2211                 }
2212
2213                 /* Had to read the page from swap area: Major fault */
2214                 ret = VM_FAULT_MAJOR;
2215                 count_vm_event(PGMAJFAULT);
2216         }
2217
2218         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2219                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2220                 ret = VM_FAULT_OOM;
2221                 goto out;
2222         }
2223
2224         mark_page_accessed(page);
2225         lock_page(page);
2226         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2227
2228         /*
2229          * Back out if somebody else already faulted in this pte.
2230          */
2231         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2232         if (unlikely(!pte_same(*page_table, orig_pte)))
2233                 goto out_nomap;
2234
2235         if (unlikely(!PageUptodate(page))) {
2236                 ret = VM_FAULT_SIGBUS;
2237                 goto out_nomap;
2238         }
2239
2240         /* The page isn't present yet, go ahead with the fault. */
2241
2242         inc_mm_counter(mm, anon_rss);
2243         pte = mk_pte(page, vma->vm_page_prot);
2244         if (write_access && can_share_swap_page(page)) {
2245                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2246                 write_access = 0;
2247         }
2248
2249         flush_icache_page(vma, page);
2250         set_pte_at(mm, address, page_table, pte);
2251         page_add_anon_rmap(page, vma, address);
2252
2253         swap_free(entry);
2254         if (vm_swap_full())
2255                 remove_exclusive_swap_page(page);
2256         unlock_page(page);
2257
2258         if (write_access) {
2259                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2260                 if (ret & VM_FAULT_ERROR)
2261                         ret &= VM_FAULT_ERROR;
2262                 goto out;
2263         }
2264
2265         /* No need to invalidate - it was non-present before */
2266         update_mmu_cache(vma, address, pte);
2267 unlock:
2268         pte_unmap_unlock(page_table, ptl);
2269 out:
2270         return ret;
2271 out_nomap:
2272         mem_cgroup_uncharge_page(page);
2273         pte_unmap_unlock(page_table, ptl);
2274         unlock_page(page);
2275         page_cache_release(page);
2276         return ret;
2277 }
2278
2279 /*
2280  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2281  * but allow concurrent faults), and pte mapped but not yet locked.
2282  * We return with mmap_sem still held, but pte unmapped and unlocked.
2283  */
2284 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2285                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2286                 int write_access)
2287 {
2288         struct page *page;
2289         spinlock_t *ptl;
2290         pte_t entry;
2291
2292         /* Allocate our own private page. */
2293         pte_unmap(page_table);
2294
2295         if (unlikely(anon_vma_prepare(vma)))
2296                 goto oom;
2297         page = alloc_zeroed_user_highpage_movable(vma, address);
2298         if (!page)
2299                 goto oom;
2300         __SetPageUptodate(page);
2301
2302         if (mem_cgroup_charge(page, mm, GFP_KERNEL))
2303                 goto oom_free_page;
2304
2305         entry = mk_pte(page, vma->vm_page_prot);
2306         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2307
2308         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2309         if (!pte_none(*page_table))
2310                 goto release;
2311         inc_mm_counter(mm, anon_rss);
2312         lru_cache_add_active(page);
2313         page_add_new_anon_rmap(page, vma, address);
2314         set_pte_at(mm, address, page_table, entry);
2315
2316         /* No need to invalidate - it was non-present before */
2317         update_mmu_cache(vma, address, entry);
2318 unlock:
2319         pte_unmap_unlock(page_table, ptl);
2320         return 0;
2321 release:
2322         mem_cgroup_uncharge_page(page);
2323         page_cache_release(page);
2324         goto unlock;
2325 oom_free_page:
2326         page_cache_release(page);
2327 oom:
2328         return VM_FAULT_OOM;
2329 }
2330
2331 /*
2332  * __do_fault() tries to create a new page mapping. It aggressively
2333  * tries to share with existing pages, but makes a separate copy if
2334  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2335  * the next page fault.
2336  *
2337  * As this is called only for pages that do not currently exist, we
2338  * do not need to flush old virtual caches or the TLB.
2339  *
2340  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2341  * but allow concurrent faults), and pte neither mapped nor locked.
2342  * We return with mmap_sem still held, but pte unmapped and unlocked.
2343  */
2344 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2345                 unsigned long address, pmd_t *pmd,
2346                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2347 {
2348         pte_t *page_table;
2349         spinlock_t *ptl;
2350         struct page *page;
2351         pte_t entry;
2352         int anon = 0;
2353         struct page *dirty_page = NULL;
2354         struct vm_fault vmf;
2355         int ret;
2356         int page_mkwrite = 0;
2357
2358         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2359         vmf.pgoff = pgoff;
2360         vmf.flags = flags;
2361         vmf.page = NULL;
2362
2363         ret = vma->vm_ops->fault(vma, &vmf);
2364         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2365                 return ret;
2366
2367         /*
2368          * For consistency in subsequent calls, make the faulted page always
2369          * locked.
2370          */
2371         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2372                 lock_page(vmf.page);
2373         else
2374                 VM_BUG_ON(!PageLocked(vmf.page));
2375
2376         /*
2377          * Should we do an early C-O-W break?
2378          */
2379         page = vmf.page;
2380         if (flags & FAULT_FLAG_WRITE) {
2381                 if (!(vma->vm_flags & VM_SHARED)) {
2382                         anon = 1;
2383                         if (unlikely(anon_vma_prepare(vma))) {
2384                                 ret = VM_FAULT_OOM;
2385                                 goto out;
2386                         }
2387                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2388                                                 vma, address);
2389                         if (!page) {
2390                                 ret = VM_FAULT_OOM;
2391                                 goto out;
2392                         }
2393                         copy_user_highpage(page, vmf.page, address, vma);
2394                         __SetPageUptodate(page);
2395                 } else {
2396                         /*
2397                          * If the page will be shareable, see if the backing
2398                          * address space wants to know that the page is about
2399                          * to become writable
2400                          */
2401                         if (vma->vm_ops->page_mkwrite) {
2402                                 unlock_page(page);
2403                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2404                                         ret = VM_FAULT_SIGBUS;
2405                                         anon = 1; /* no anon but release vmf.page */
2406                                         goto out_unlocked;
2407                                 }
2408                                 lock_page(page);
2409                                 /*
2410                                  * XXX: this is not quite right (racy vs
2411                                  * invalidate) to unlock and relock the page
2412                                  * like this, however a better fix requires
2413                                  * reworking page_mkwrite locking API, which
2414                                  * is better done later.
2415                                  */
2416                                 if (!page->mapping) {
2417                                         ret = 0;
2418                                         anon = 1; /* no anon but release vmf.page */
2419                                         goto out;
2420                                 }
2421                                 page_mkwrite = 1;
2422                         }
2423                 }
2424
2425         }
2426
2427         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2428                 ret = VM_FAULT_OOM;
2429                 goto out;
2430         }
2431
2432         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2433
2434         /*
2435          * This silly early PAGE_DIRTY setting removes a race
2436          * due to the bad i386 page protection. But it's valid
2437          * for other architectures too.
2438          *
2439          * Note that if write_access is true, we either now have
2440          * an exclusive copy of the page, or this is a shared mapping,
2441          * so we can make it writable and dirty to avoid having to
2442          * handle that later.
2443          */
2444         /* Only go through if we didn't race with anybody else... */
2445         if (likely(pte_same(*page_table, orig_pte))) {
2446                 flush_icache_page(vma, page);
2447                 entry = mk_pte(page, vma->vm_page_prot);
2448                 if (flags & FAULT_FLAG_WRITE)
2449                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2450                 set_pte_at(mm, address, page_table, entry);
2451                 if (anon) {
2452                         inc_mm_counter(mm, anon_rss);
2453                         lru_cache_add_active(page);
2454                         page_add_new_anon_rmap(page, vma, address);
2455                 } else {
2456                         inc_mm_counter(mm, file_rss);
2457                         page_add_file_rmap(page);
2458                         if (flags & FAULT_FLAG_WRITE) {
2459                                 dirty_page = page;
2460                                 get_page(dirty_page);
2461                         }
2462                 }
2463
2464                 /* no need to invalidate: a not-present page won't be cached */
2465                 update_mmu_cache(vma, address, entry);
2466         } else {
2467                 mem_cgroup_uncharge_page(page);
2468                 if (anon)
2469                         page_cache_release(page);
2470                 else
2471                         anon = 1; /* no anon but release faulted_page */
2472         }
2473
2474         pte_unmap_unlock(page_table, ptl);
2475
2476 out:
2477         unlock_page(vmf.page);
2478 out_unlocked:
2479         if (anon)
2480                 page_cache_release(vmf.page);
2481         else if (dirty_page) {
2482                 if (vma->vm_file)
2483                         file_update_time(vma->vm_file);
2484
2485                 set_page_dirty_balance(dirty_page, page_mkwrite);
2486                 put_page(dirty_page);
2487         }
2488
2489         return ret;
2490 }
2491
2492 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2493                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2494                 int write_access, pte_t orig_pte)
2495 {
2496         pgoff_t pgoff = (((address & PAGE_MASK)
2497                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2498         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2499
2500         pte_unmap(page_table);
2501         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2502 }
2503
2504
2505 /*
2506  * do_no_pfn() tries to create a new page mapping for a page without
2507  * a struct_page backing it
2508  *
2509  * As this is called only for pages that do not currently exist, we
2510  * do not need to flush old virtual caches or the TLB.
2511  *
2512  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2513  * but allow concurrent faults), and pte mapped but not yet locked.
2514  * We return with mmap_sem still held, but pte unmapped and unlocked.
2515  *
2516  * It is expected that the ->nopfn handler always returns the same pfn
2517  * for a given virtual mapping.
2518  *
2519  * Mark this `noinline' to prevent it from bloating the main pagefault code.
2520  */
2521 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2522                      unsigned long address, pte_t *page_table, pmd_t *pmd,
2523                      int write_access)
2524 {
2525         spinlock_t *ptl;
2526         pte_t entry;
2527         unsigned long pfn;
2528
2529         pte_unmap(page_table);
2530         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2531         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2532
2533         pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2534
2535         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2536
2537         if (unlikely(pfn == NOPFN_OOM))
2538                 return VM_FAULT_OOM;
2539         else if (unlikely(pfn == NOPFN_SIGBUS))
2540                 return VM_FAULT_SIGBUS;
2541         else if (unlikely(pfn == NOPFN_REFAULT))
2542                 return 0;
2543
2544         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2545
2546         /* Only go through if we didn't race with anybody else... */
2547         if (pte_none(*page_table)) {
2548                 entry = pfn_pte(pfn, vma->vm_page_prot);
2549                 if (write_access)
2550                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2551                 set_pte_at(mm, address, page_table, entry);
2552         }
2553         pte_unmap_unlock(page_table, ptl);
2554         return 0;
2555 }
2556
2557 /*
2558  * Fault of a previously existing named mapping. Repopulate the pte
2559  * from the encoded file_pte if possible. This enables swappable
2560  * nonlinear vmas.
2561  *
2562  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2563  * but allow concurrent faults), and pte mapped but not yet locked.
2564  * We return with mmap_sem still held, but pte unmapped and unlocked.
2565  */
2566 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2567                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2568                 int write_access, pte_t orig_pte)
2569 {
2570         unsigned int flags = FAULT_FLAG_NONLINEAR |
2571                                 (write_access ? FAULT_FLAG_WRITE : 0);
2572         pgoff_t pgoff;
2573
2574         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2575                 return 0;
2576
2577         if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2578                         !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2579                 /*
2580                  * Page table corrupted: show pte and kill process.
2581                  */
2582                 print_bad_pte(vma, orig_pte, address);
2583                 return VM_FAULT_OOM;
2584         }
2585
2586         pgoff = pte_to_pgoff(orig_pte);
2587         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2588 }
2589
2590 /*
2591  * These routines also need to handle stuff like marking pages dirty
2592  * and/or accessed for architectures that don't do it in hardware (most
2593  * RISC architectures).  The early dirtying is also good on the i386.
2594  *
2595  * There is also a hook called "update_mmu_cache()" that architectures
2596  * with external mmu caches can use to update those (ie the Sparc or
2597  * PowerPC hashed page tables that act as extended TLBs).
2598  *
2599  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2600  * but allow concurrent faults), and pte mapped but not yet locked.
2601  * We return with mmap_sem still held, but pte unmapped and unlocked.
2602  */
2603 static inline int handle_pte_fault(struct mm_struct *mm,
2604                 struct vm_area_struct *vma, unsigned long address,
2605                 pte_t *pte, pmd_t *pmd, int write_access)
2606 {
2607         pte_t entry;
2608         spinlock_t *ptl;
2609
2610         entry = *pte;
2611         if (!pte_present(entry)) {
2612                 if (pte_none(entry)) {
2613                         if (vma->vm_ops) {
2614                                 if (likely(vma->vm_ops->fault))
2615                                         return do_linear_fault(mm, vma, address,
2616                                                 pte, pmd, write_access, entry);
2617                                 if (unlikely(vma->vm_ops->nopfn))
2618                                         return do_no_pfn(mm, vma, address, pte,
2619                                                          pmd, write_access);
2620                         }
2621                         return do_anonymous_page(mm, vma, address,
2622                                                  pte, pmd, write_access);
2623                 }
2624                 if (pte_file(entry))
2625                         return do_nonlinear_fault(mm, vma, address,
2626                                         pte, pmd, write_access, entry);
2627                 return do_swap_page(mm, vma, address,
2628                                         pte, pmd, write_access, entry);
2629         }
2630
2631         ptl = pte_lockptr(mm, pmd);
2632         spin_lock(ptl);
2633         if (unlikely(!pte_same(*pte, entry)))
2634                 goto unlock;
2635         if (write_access) {
2636                 if (!pte_write(entry))
2637                         return do_wp_page(mm, vma, address,
2638                                         pte, pmd, ptl, entry);
2639                 entry = pte_mkdirty(entry);
2640         }
2641         entry = pte_mkyoung(entry);
2642         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2643                 update_mmu_cache(vma, address, entry);
2644         } else {
2645                 /*
2646                  * This is needed only for protection faults but the arch code
2647                  * is not yet telling us if this is a protection fault or not.
2648                  * This still avoids useless tlb flushes for .text page faults
2649                  * with threads.
2650                  */
2651                 if (write_access)
2652                         flush_tlb_page(vma, address);
2653         }
2654 unlock:
2655         pte_unmap_unlock(pte, ptl);
2656         return 0;
2657 }
2658
2659 /*
2660  * By the time we get here, we already hold the mm semaphore
2661  */
2662 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2663                 unsigned long address, int write_access)
2664 {
2665         pgd_t *pgd;
2666         pud_t *pud;
2667         pmd_t *pmd;
2668         pte_t *pte;
2669
2670         __set_current_state(TASK_RUNNING);
2671
2672         count_vm_event(PGFAULT);
2673
2674         if (unlikely(is_vm_hugetlb_page(vma)))
2675                 return hugetlb_fault(mm, vma, address, write_access);
2676
2677         pgd = pgd_offset(mm, address);
2678         pud = pud_alloc(mm, pgd, address);
2679         if (!pud)
2680                 return VM_FAULT_OOM;
2681         pmd = pmd_alloc(mm, pud, address);
2682         if (!pmd)
2683                 return VM_FAULT_OOM;
2684         pte = pte_alloc_map(mm, pmd, address);
2685         if (!pte)
2686                 return VM_FAULT_OOM;
2687
2688         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2689 }
2690
2691 #ifndef __PAGETABLE_PUD_FOLDED
2692 /*
2693  * Allocate page upper directory.
2694  * We've already handled the fast-path in-line.
2695  */
2696 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2697 {
2698         pud_t *new = pud_alloc_one(mm, address);
2699         if (!new)
2700                 return -ENOMEM;
2701
2702         smp_wmb(); /* See comment in __pte_alloc */
2703
2704         spin_lock(&mm->page_table_lock);
2705         if (pgd_present(*pgd))          /* Another has populated it */
2706                 pud_free(mm, new);
2707         else
2708                 pgd_populate(mm, pgd, new);
2709         spin_unlock(&mm->page_table_lock);
2710         return 0;
2711 }
2712 #endif /* __PAGETABLE_PUD_FOLDED */
2713
2714 #ifndef __PAGETABLE_PMD_FOLDED
2715 /*
2716  * Allocate page middle directory.
2717  * We've already handled the fast-path in-line.
2718  */
2719 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2720 {
2721         pmd_t *new = pmd_alloc_one(mm, address);
2722         if (!new)
2723                 return -ENOMEM;
2724
2725         smp_wmb(); /* See comment in __pte_alloc */
2726
2727         spin_lock(&mm->page_table_lock);
2728 #ifndef __ARCH_HAS_4LEVEL_HACK
2729         if (pud_present(*pud))          /* Another has populated it */
2730                 pmd_free(mm, new);
2731         else
2732                 pud_populate(mm, pud, new);
2733 #else
2734         if (pgd_present(*pud))          /* Another has populated it */
2735                 pmd_free(mm, new);
2736         else
2737                 pgd_populate(mm, pud, new);
2738 #endif /* __ARCH_HAS_4LEVEL_HACK */
2739         spin_unlock(&mm->page_table_lock);
2740         return 0;
2741 }
2742 #endif /* __PAGETABLE_PMD_FOLDED */
2743
2744 int make_pages_present(unsigned long addr, unsigned long end)
2745 {
2746         int ret, len, write;
2747         struct vm_area_struct * vma;
2748
2749         vma = find_vma(current->mm, addr);
2750         if (!vma)
2751                 return -1;
2752         write = (vma->vm_flags & VM_WRITE) != 0;
2753         BUG_ON(addr >= end);
2754         BUG_ON(end > vma->vm_end);
2755         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2756         ret = get_user_pages(current, current->mm, addr,
2757                         len, write, 0, NULL, NULL);
2758         if (ret < 0)
2759                 return ret;
2760         return ret == len ? 0 : -1;
2761 }
2762
2763 #if !defined(__HAVE_ARCH_GATE_AREA)
2764
2765 #if defined(AT_SYSINFO_EHDR)
2766 static struct vm_area_struct gate_vma;
2767
2768 static int __init gate_vma_init(void)
2769 {
2770         gate_vma.vm_mm = NULL;
2771         gate_vma.vm_start = FIXADDR_USER_START;
2772         gate_vma.vm_end = FIXADDR_USER_END;
2773         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2774         gate_vma.vm_page_prot = __P101;
2775         /*
2776          * Make sure the vDSO gets into every core dump.
2777          * Dumping its contents makes post-mortem fully interpretable later
2778          * without matching up the same kernel and hardware config to see
2779          * what PC values meant.
2780          */
2781         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2782         return 0;
2783 }
2784 __initcall(gate_vma_init);
2785 #endif
2786
2787 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2788 {
2789 #ifdef AT_SYSINFO_EHDR
2790         return &gate_vma;
2791 #else
2792         return NULL;
2793 #endif
2794 }
2795
2796 int in_gate_area_no_task(unsigned long addr)
2797 {
2798 #ifdef AT_SYSINFO_EHDR
2799         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2800                 return 1;
2801 #endif
2802         return 0;
2803 }
2804
2805 #endif  /* __HAVE_ARCH_GATE_AREA */
2806
2807 /*
2808  * Access another process' address space.
2809  * Source/target buffer must be kernel space,
2810  * Do not walk the page table directly, use get_user_pages
2811  */
2812 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2813 {
2814         struct mm_struct *mm;
2815         struct vm_area_struct *vma;
2816         struct page *page;
2817         void *old_buf = buf;
2818
2819         mm = get_task_mm(tsk);
2820         if (!mm)
2821                 return 0;
2822
2823         down_read(&mm->mmap_sem);
2824         /* ignore errors, just check how much was successfully transferred */
2825         while (len) {
2826                 int bytes, ret, offset;
2827                 void *maddr;
2828
2829                 ret = get_user_pages(tsk, mm, addr, 1,
2830                                 write, 1, &page, &vma);
2831                 if (ret <= 0)
2832                         break;
2833
2834                 bytes = len;
2835                 offset = addr & (PAGE_SIZE-1);
2836                 if (bytes > PAGE_SIZE-offset)
2837                         bytes = PAGE_SIZE-offset;
2838
2839                 maddr = kmap(page);
2840                 if (write) {
2841                         copy_to_user_page(vma, page, addr,
2842                                           maddr + offset, buf, bytes);
2843                         set_page_dirty_lock(page);
2844                 } else {
2845                         copy_from_user_page(vma, page, addr,
2846                                             buf, maddr + offset, bytes);
2847                 }
2848                 kunmap(page);
2849                 page_cache_release(page);
2850                 len -= bytes;
2851                 buf += bytes;
2852                 addr += bytes;
2853         }
2854         up_read(&mm->mmap_sem);
2855         mmput(mm);
2856
2857         return buf - old_buf;
2858 }
2859
2860 /*
2861  * Print the name of a VMA.
2862  */
2863 void print_vma_addr(char *prefix, unsigned long ip)
2864 {
2865         struct mm_struct *mm = current->mm;
2866         struct vm_area_struct *vma;
2867
2868         /*
2869          * Do not print if we are in atomic
2870          * contexts (in exception stacks, etc.):
2871          */
2872         if (preempt_count())
2873                 return;
2874
2875         down_read(&mm->mmap_sem);
2876         vma = find_vma(mm, ip);
2877         if (vma && vma->vm_file) {
2878                 struct file *f = vma->vm_file;
2879                 char *buf = (char *)__get_free_page(GFP_KERNEL);
2880                 if (buf) {
2881                         char *p, *s;
2882
2883                         p = d_path(&f->f_path, buf, PAGE_SIZE);
2884                         if (IS_ERR(p))
2885                                 p = "?";
2886                         s = strrchr(p, '/');
2887                         if (s)
2888                                 p = s+1;
2889                         printk("%s%s[%lx+%lx]", prefix, p,
2890                                         vma->vm_start,
2891                                         vma->vm_end - vma->vm_start);
2892                         free_page((unsigned long)buf);
2893                 }
2894         }
2895         up_read(&current->mm->mmap_sem);
2896 }