82f3f1c5cf171eb869269f85d590c69290040ccb
[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 #include "internal.h"
65
66 #ifndef CONFIG_NEED_MULTIPLE_NODES
67 /* use the per-pgdat data instead for discontigmem - mbligh */
68 unsigned long max_mapnr;
69 struct page *mem_map;
70
71 EXPORT_SYMBOL(max_mapnr);
72 EXPORT_SYMBOL(mem_map);
73 #endif
74
75 unsigned long num_physpages;
76 /*
77  * A number of key systems in x86 including ioremap() rely on the assumption
78  * that high_memory defines the upper bound on direct map memory, then end
79  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
80  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
81  * and ZONE_HIGHMEM.
82  */
83 void * high_memory;
84
85 EXPORT_SYMBOL(num_physpages);
86 EXPORT_SYMBOL(high_memory);
87
88 /*
89  * Randomize the address space (stacks, mmaps, brk, etc.).
90  *
91  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
92  *   as ancient (libc5 based) binaries can segfault. )
93  */
94 int randomize_va_space __read_mostly =
95 #ifdef CONFIG_COMPAT_BRK
96                                         1;
97 #else
98                                         2;
99 #endif
100
101 static int __init disable_randmaps(char *s)
102 {
103         randomize_va_space = 0;
104         return 1;
105 }
106 __setup("norandmaps", disable_randmaps);
107
108
109 /*
110  * If a p?d_bad entry is found while walking page tables, report
111  * the error, before resetting entry to p?d_none.  Usually (but
112  * very seldom) called out from the p?d_none_or_clear_bad macros.
113  */
114
115 void pgd_clear_bad(pgd_t *pgd)
116 {
117         pgd_ERROR(*pgd);
118         pgd_clear(pgd);
119 }
120
121 void pud_clear_bad(pud_t *pud)
122 {
123         pud_ERROR(*pud);
124         pud_clear(pud);
125 }
126
127 void pmd_clear_bad(pmd_t *pmd)
128 {
129         pmd_ERROR(*pmd);
130         pmd_clear(pmd);
131 }
132
133 /*
134  * Note: this doesn't free the actual pages themselves. That
135  * has been handled earlier when unmapping all the memory regions.
136  */
137 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
138 {
139         pgtable_t token = pmd_pgtable(*pmd);
140         pmd_clear(pmd);
141         pte_free_tlb(tlb, token);
142         tlb->mm->nr_ptes--;
143 }
144
145 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
146                                 unsigned long addr, unsigned long end,
147                                 unsigned long floor, unsigned long ceiling)
148 {
149         pmd_t *pmd;
150         unsigned long next;
151         unsigned long start;
152
153         start = addr;
154         pmd = pmd_offset(pud, addr);
155         do {
156                 next = pmd_addr_end(addr, end);
157                 if (pmd_none_or_clear_bad(pmd))
158                         continue;
159                 free_pte_range(tlb, pmd);
160         } while (pmd++, addr = next, addr != end);
161
162         start &= PUD_MASK;
163         if (start < floor)
164                 return;
165         if (ceiling) {
166                 ceiling &= PUD_MASK;
167                 if (!ceiling)
168                         return;
169         }
170         if (end - 1 > ceiling - 1)
171                 return;
172
173         pmd = pmd_offset(pud, start);
174         pud_clear(pud);
175         pmd_free_tlb(tlb, pmd);
176 }
177
178 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
179                                 unsigned long addr, unsigned long end,
180                                 unsigned long floor, unsigned long ceiling)
181 {
182         pud_t *pud;
183         unsigned long next;
184         unsigned long start;
185
186         start = addr;
187         pud = pud_offset(pgd, addr);
188         do {
189                 next = pud_addr_end(addr, end);
190                 if (pud_none_or_clear_bad(pud))
191                         continue;
192                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
193         } while (pud++, addr = next, addr != end);
194
195         start &= PGDIR_MASK;
196         if (start < floor)
197                 return;
198         if (ceiling) {
199                 ceiling &= PGDIR_MASK;
200                 if (!ceiling)
201                         return;
202         }
203         if (end - 1 > ceiling - 1)
204                 return;
205
206         pud = pud_offset(pgd, start);
207         pgd_clear(pgd);
208         pud_free_tlb(tlb, pud);
209 }
210
211 /*
212  * This function frees user-level page tables of a process.
213  *
214  * Must be called with pagetable lock held.
215  */
216 void free_pgd_range(struct mmu_gather *tlb,
217                         unsigned long addr, unsigned long end,
218                         unsigned long floor, unsigned long ceiling)
219 {
220         pgd_t *pgd;
221         unsigned long next;
222         unsigned long start;
223
224         /*
225          * The next few lines have given us lots of grief...
226          *
227          * Why are we testing PMD* at this top level?  Because often
228          * there will be no work to do at all, and we'd prefer not to
229          * go all the way down to the bottom just to discover that.
230          *
231          * Why all these "- 1"s?  Because 0 represents both the bottom
232          * of the address space and the top of it (using -1 for the
233          * top wouldn't help much: the masks would do the wrong thing).
234          * The rule is that addr 0 and floor 0 refer to the bottom of
235          * the address space, but end 0 and ceiling 0 refer to the top
236          * Comparisons need to use "end - 1" and "ceiling - 1" (though
237          * that end 0 case should be mythical).
238          *
239          * Wherever addr is brought up or ceiling brought down, we must
240          * be careful to reject "the opposite 0" before it confuses the
241          * subsequent tests.  But what about where end is brought down
242          * by PMD_SIZE below? no, end can't go down to 0 there.
243          *
244          * Whereas we round start (addr) and ceiling down, by different
245          * masks at different levels, in order to test whether a table
246          * now has no other vmas using it, so can be freed, we don't
247          * bother to round floor or end up - the tests don't need that.
248          */
249
250         addr &= PMD_MASK;
251         if (addr < floor) {
252                 addr += PMD_SIZE;
253                 if (!addr)
254                         return;
255         }
256         if (ceiling) {
257                 ceiling &= PMD_MASK;
258                 if (!ceiling)
259                         return;
260         }
261         if (end - 1 > ceiling - 1)
262                 end -= PMD_SIZE;
263         if (addr > end - 1)
264                 return;
265
266         start = addr;
267         pgd = pgd_offset(tlb->mm, addr);
268         do {
269                 next = pgd_addr_end(addr, end);
270                 if (pgd_none_or_clear_bad(pgd))
271                         continue;
272                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
273         } while (pgd++, addr = next, addr != end);
274 }
275
276 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
277                 unsigned long floor, unsigned long ceiling)
278 {
279         while (vma) {
280                 struct vm_area_struct *next = vma->vm_next;
281                 unsigned long addr = vma->vm_start;
282
283                 /*
284                  * Hide vma from rmap and vmtruncate before freeing pgtables
285                  */
286                 anon_vma_unlink(vma);
287                 unlink_file_vma(vma);
288
289                 if (is_vm_hugetlb_page(vma)) {
290                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
291                                 floor, next? next->vm_start: ceiling);
292                 } else {
293                         /*
294                          * Optimization: gather nearby vmas into one call down
295                          */
296                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
297                                && !is_vm_hugetlb_page(next)) {
298                                 vma = next;
299                                 next = vma->vm_next;
300                                 anon_vma_unlink(vma);
301                                 unlink_file_vma(vma);
302                         }
303                         free_pgd_range(tlb, addr, vma->vm_end,
304                                 floor, next? next->vm_start: ceiling);
305                 }
306                 vma = next;
307         }
308 }
309
310 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
311 {
312         pgtable_t new = pte_alloc_one(mm, address);
313         if (!new)
314                 return -ENOMEM;
315
316         /*
317          * Ensure all pte setup (eg. pte page lock and page clearing) are
318          * visible before the pte is made visible to other CPUs by being
319          * put into page tables.
320          *
321          * The other side of the story is the pointer chasing in the page
322          * table walking code (when walking the page table without locking;
323          * ie. most of the time). Fortunately, these data accesses consist
324          * of a chain of data-dependent loads, meaning most CPUs (alpha
325          * being the notable exception) will already guarantee loads are
326          * seen in-order. See the alpha page table accessors for the
327          * smp_read_barrier_depends() barriers in page table walking code.
328          */
329         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
330
331         spin_lock(&mm->page_table_lock);
332         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
333                 mm->nr_ptes++;
334                 pmd_populate(mm, pmd, new);
335                 new = NULL;
336         }
337         spin_unlock(&mm->page_table_lock);
338         if (new)
339                 pte_free(mm, new);
340         return 0;
341 }
342
343 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
344 {
345         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
346         if (!new)
347                 return -ENOMEM;
348
349         smp_wmb(); /* See comment in __pte_alloc */
350
351         spin_lock(&init_mm.page_table_lock);
352         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
353                 pmd_populate_kernel(&init_mm, pmd, new);
354                 new = NULL;
355         }
356         spin_unlock(&init_mm.page_table_lock);
357         if (new)
358                 pte_free_kernel(&init_mm, new);
359         return 0;
360 }
361
362 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
363 {
364         if (file_rss)
365                 add_mm_counter(mm, file_rss, file_rss);
366         if (anon_rss)
367                 add_mm_counter(mm, anon_rss, anon_rss);
368 }
369
370 /*
371  * This function is called to print an error when a bad pte
372  * is found. For example, we might have a PFN-mapped pte in
373  * a region that doesn't allow it.
374  *
375  * The calling function must still handle the error.
376  */
377 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
378 {
379         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
380                         "vm_flags = %lx, vaddr = %lx\n",
381                 (long long)pte_val(pte),
382                 (vma->vm_mm == current->mm ? current->comm : "???"),
383                 vma->vm_flags, vaddr);
384         dump_stack();
385 }
386
387 static inline int is_cow_mapping(unsigned int flags)
388 {
389         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
390 }
391
392 /*
393  * vm_normal_page -- This function gets the "struct page" associated with a pte.
394  *
395  * "Special" mappings do not wish to be associated with a "struct page" (either
396  * it doesn't exist, or it exists but they don't want to touch it). In this
397  * case, NULL is returned here. "Normal" mappings do have a struct page.
398  *
399  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
400  * pte bit, in which case this function is trivial. Secondly, an architecture
401  * may not have a spare pte bit, which requires a more complicated scheme,
402  * described below.
403  *
404  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
405  * special mapping (even if there are underlying and valid "struct pages").
406  * COWed pages of a VM_PFNMAP are always normal.
407  *
408  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
409  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
410  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
411  * mapping will always honor the rule
412  *
413  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
414  *
415  * And for normal mappings this is false.
416  *
417  * This restricts such mappings to be a linear translation from virtual address
418  * to pfn. To get around this restriction, we allow arbitrary mappings so long
419  * as the vma is not a COW mapping; in that case, we know that all ptes are
420  * special (because none can have been COWed).
421  *
422  *
423  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
424  *
425  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
426  * page" backing, however the difference is that _all_ pages with a struct
427  * page (that is, those where pfn_valid is true) are refcounted and considered
428  * normal pages by the VM. The disadvantage is that pages are refcounted
429  * (which can be slower and simply not an option for some PFNMAP users). The
430  * advantage is that we don't have to follow the strict linearity rule of
431  * PFNMAP mappings in order to support COWable mappings.
432  *
433  */
434 #ifdef __HAVE_ARCH_PTE_SPECIAL
435 # define HAVE_PTE_SPECIAL 1
436 #else
437 # define HAVE_PTE_SPECIAL 0
438 #endif
439 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
440                                 pte_t pte)
441 {
442         unsigned long pfn;
443
444         if (HAVE_PTE_SPECIAL) {
445                 if (likely(!pte_special(pte))) {
446                         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
447                         return pte_page(pte);
448                 }
449                 VM_BUG_ON(!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)));
450                 return NULL;
451         }
452
453         /* !HAVE_PTE_SPECIAL case follows: */
454
455         pfn = pte_pfn(pte);
456
457         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
458                 if (vma->vm_flags & VM_MIXEDMAP) {
459                         if (!pfn_valid(pfn))
460                                 return NULL;
461                         goto out;
462                 } else {
463                         unsigned long off;
464                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
465                         if (pfn == vma->vm_pgoff + off)
466                                 return NULL;
467                         if (!is_cow_mapping(vma->vm_flags))
468                                 return NULL;
469                 }
470         }
471
472         VM_BUG_ON(!pfn_valid(pfn));
473
474         /*
475          * NOTE! We still have PageReserved() pages in the page tables.
476          *
477          * eg. VDSO mappings can cause them to exist.
478          */
479 out:
480         return pfn_to_page(pfn);
481 }
482
483 /*
484  * copy one vm_area from one task to the other. Assumes the page tables
485  * already present in the new task to be cleared in the whole range
486  * covered by this vma.
487  */
488
489 static inline void
490 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
491                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
492                 unsigned long addr, int *rss)
493 {
494         unsigned long vm_flags = vma->vm_flags;
495         pte_t pte = *src_pte;
496         struct page *page;
497
498         /* pte contains position in swap or file, so copy. */
499         if (unlikely(!pte_present(pte))) {
500                 if (!pte_file(pte)) {
501                         swp_entry_t entry = pte_to_swp_entry(pte);
502
503                         swap_duplicate(entry);
504                         /* make sure dst_mm is on swapoff's mmlist. */
505                         if (unlikely(list_empty(&dst_mm->mmlist))) {
506                                 spin_lock(&mmlist_lock);
507                                 if (list_empty(&dst_mm->mmlist))
508                                         list_add(&dst_mm->mmlist,
509                                                  &src_mm->mmlist);
510                                 spin_unlock(&mmlist_lock);
511                         }
512                         if (is_write_migration_entry(entry) &&
513                                         is_cow_mapping(vm_flags)) {
514                                 /*
515                                  * COW mappings require pages in both parent
516                                  * and child to be set to read.
517                                  */
518                                 make_migration_entry_read(&entry);
519                                 pte = swp_entry_to_pte(entry);
520                                 set_pte_at(src_mm, addr, src_pte, pte);
521                         }
522                 }
523                 goto out_set_pte;
524         }
525
526         /*
527          * If it's a COW mapping, write protect it both
528          * in the parent and the child
529          */
530         if (is_cow_mapping(vm_flags)) {
531                 ptep_set_wrprotect(src_mm, addr, src_pte);
532                 pte = pte_wrprotect(pte);
533         }
534
535         /*
536          * If it's a shared mapping, mark it clean in
537          * the child
538          */
539         if (vm_flags & VM_SHARED)
540                 pte = pte_mkclean(pte);
541         pte = pte_mkold(pte);
542
543         page = vm_normal_page(vma, addr, pte);
544         if (page) {
545                 get_page(page);
546                 page_dup_rmap(page, vma, addr);
547                 rss[!!PageAnon(page)]++;
548         }
549
550 out_set_pte:
551         set_pte_at(dst_mm, addr, dst_pte, pte);
552 }
553
554 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
555                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
556                 unsigned long addr, unsigned long end)
557 {
558         pte_t *src_pte, *dst_pte;
559         spinlock_t *src_ptl, *dst_ptl;
560         int progress = 0;
561         int rss[2];
562
563 again:
564         rss[1] = rss[0] = 0;
565         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
566         if (!dst_pte)
567                 return -ENOMEM;
568         src_pte = pte_offset_map_nested(src_pmd, addr);
569         src_ptl = pte_lockptr(src_mm, src_pmd);
570         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
571         arch_enter_lazy_mmu_mode();
572
573         do {
574                 /*
575                  * We are holding two locks at this point - either of them
576                  * could generate latencies in another task on another CPU.
577                  */
578                 if (progress >= 32) {
579                         progress = 0;
580                         if (need_resched() ||
581                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
582                                 break;
583                 }
584                 if (pte_none(*src_pte)) {
585                         progress++;
586                         continue;
587                 }
588                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
589                 progress += 8;
590         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
591
592         arch_leave_lazy_mmu_mode();
593         spin_unlock(src_ptl);
594         pte_unmap_nested(src_pte - 1);
595         add_mm_rss(dst_mm, rss[0], rss[1]);
596         pte_unmap_unlock(dst_pte - 1, dst_ptl);
597         cond_resched();
598         if (addr != end)
599                 goto again;
600         return 0;
601 }
602
603 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
604                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
605                 unsigned long addr, unsigned long end)
606 {
607         pmd_t *src_pmd, *dst_pmd;
608         unsigned long next;
609
610         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
611         if (!dst_pmd)
612                 return -ENOMEM;
613         src_pmd = pmd_offset(src_pud, addr);
614         do {
615                 next = pmd_addr_end(addr, end);
616                 if (pmd_none_or_clear_bad(src_pmd))
617                         continue;
618                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
619                                                 vma, addr, next))
620                         return -ENOMEM;
621         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
622         return 0;
623 }
624
625 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
626                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
627                 unsigned long addr, unsigned long end)
628 {
629         pud_t *src_pud, *dst_pud;
630         unsigned long next;
631
632         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
633         if (!dst_pud)
634                 return -ENOMEM;
635         src_pud = pud_offset(src_pgd, addr);
636         do {
637                 next = pud_addr_end(addr, end);
638                 if (pud_none_or_clear_bad(src_pud))
639                         continue;
640                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
641                                                 vma, addr, next))
642                         return -ENOMEM;
643         } while (dst_pud++, src_pud++, addr = next, addr != end);
644         return 0;
645 }
646
647 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
648                 struct vm_area_struct *vma)
649 {
650         pgd_t *src_pgd, *dst_pgd;
651         unsigned long next;
652         unsigned long addr = vma->vm_start;
653         unsigned long end = vma->vm_end;
654
655         /*
656          * Don't copy ptes where a page fault will fill them correctly.
657          * Fork becomes much lighter when there are big shared or private
658          * readonly mappings. The tradeoff is that copy_page_range is more
659          * efficient than faulting.
660          */
661         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
662                 if (!vma->anon_vma)
663                         return 0;
664         }
665
666         if (is_vm_hugetlb_page(vma))
667                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
668
669         dst_pgd = pgd_offset(dst_mm, addr);
670         src_pgd = pgd_offset(src_mm, addr);
671         do {
672                 next = pgd_addr_end(addr, end);
673                 if (pgd_none_or_clear_bad(src_pgd))
674                         continue;
675                 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
676                                                 vma, addr, next))
677                         return -ENOMEM;
678         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
679         return 0;
680 }
681
682 static unsigned long zap_pte_range(struct mmu_gather *tlb,
683                                 struct vm_area_struct *vma, pmd_t *pmd,
684                                 unsigned long addr, unsigned long end,
685                                 long *zap_work, struct zap_details *details)
686 {
687         struct mm_struct *mm = tlb->mm;
688         pte_t *pte;
689         spinlock_t *ptl;
690         int file_rss = 0;
691         int anon_rss = 0;
692
693         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
694         arch_enter_lazy_mmu_mode();
695         do {
696                 pte_t ptent = *pte;
697                 if (pte_none(ptent)) {
698                         (*zap_work)--;
699                         continue;
700                 }
701
702                 (*zap_work) -= PAGE_SIZE;
703
704                 if (pte_present(ptent)) {
705                         struct page *page;
706
707                         page = vm_normal_page(vma, addr, ptent);
708                         if (unlikely(details) && page) {
709                                 /*
710                                  * unmap_shared_mapping_pages() wants to
711                                  * invalidate cache without truncating:
712                                  * unmap shared but keep private pages.
713                                  */
714                                 if (details->check_mapping &&
715                                     details->check_mapping != page->mapping)
716                                         continue;
717                                 /*
718                                  * Each page->index must be checked when
719                                  * invalidating or truncating nonlinear.
720                                  */
721                                 if (details->nonlinear_vma &&
722                                     (page->index < details->first_index ||
723                                      page->index > details->last_index))
724                                         continue;
725                         }
726                         ptent = ptep_get_and_clear_full(mm, addr, pte,
727                                                         tlb->fullmm);
728                         tlb_remove_tlb_entry(tlb, pte, addr);
729                         if (unlikely(!page))
730                                 continue;
731                         if (unlikely(details) && details->nonlinear_vma
732                             && linear_page_index(details->nonlinear_vma,
733                                                 addr) != page->index)
734                                 set_pte_at(mm, addr, pte,
735                                            pgoff_to_pte(page->index));
736                         if (PageAnon(page))
737                                 anon_rss--;
738                         else {
739                                 if (pte_dirty(ptent))
740                                         set_page_dirty(page);
741                                 if (pte_young(ptent))
742                                         SetPageReferenced(page);
743                                 file_rss--;
744                         }
745                         page_remove_rmap(page, vma);
746                         tlb_remove_page(tlb, page);
747                         continue;
748                 }
749                 /*
750                  * If details->check_mapping, we leave swap entries;
751                  * if details->nonlinear_vma, we leave file entries.
752                  */
753                 if (unlikely(details))
754                         continue;
755                 if (!pte_file(ptent))
756                         free_swap_and_cache(pte_to_swp_entry(ptent));
757                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
758         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
759
760         add_mm_rss(mm, file_rss, anon_rss);
761         arch_leave_lazy_mmu_mode();
762         pte_unmap_unlock(pte - 1, ptl);
763
764         return addr;
765 }
766
767 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
768                                 struct vm_area_struct *vma, pud_t *pud,
769                                 unsigned long addr, unsigned long end,
770                                 long *zap_work, struct zap_details *details)
771 {
772         pmd_t *pmd;
773         unsigned long next;
774
775         pmd = pmd_offset(pud, addr);
776         do {
777                 next = pmd_addr_end(addr, end);
778                 if (pmd_none_or_clear_bad(pmd)) {
779                         (*zap_work)--;
780                         continue;
781                 }
782                 next = zap_pte_range(tlb, vma, pmd, addr, next,
783                                                 zap_work, details);
784         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
785
786         return addr;
787 }
788
789 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
790                                 struct vm_area_struct *vma, pgd_t *pgd,
791                                 unsigned long addr, unsigned long end,
792                                 long *zap_work, struct zap_details *details)
793 {
794         pud_t *pud;
795         unsigned long next;
796
797         pud = pud_offset(pgd, addr);
798         do {
799                 next = pud_addr_end(addr, end);
800                 if (pud_none_or_clear_bad(pud)) {
801                         (*zap_work)--;
802                         continue;
803                 }
804                 next = zap_pmd_range(tlb, vma, pud, addr, next,
805                                                 zap_work, details);
806         } while (pud++, addr = next, (addr != end && *zap_work > 0));
807
808         return addr;
809 }
810
811 static unsigned long unmap_page_range(struct mmu_gather *tlb,
812                                 struct vm_area_struct *vma,
813                                 unsigned long addr, unsigned long end,
814                                 long *zap_work, struct zap_details *details)
815 {
816         pgd_t *pgd;
817         unsigned long next;
818
819         if (details && !details->check_mapping && !details->nonlinear_vma)
820                 details = NULL;
821
822         BUG_ON(addr >= end);
823         tlb_start_vma(tlb, vma);
824         pgd = pgd_offset(vma->vm_mm, addr);
825         do {
826                 next = pgd_addr_end(addr, end);
827                 if (pgd_none_or_clear_bad(pgd)) {
828                         (*zap_work)--;
829                         continue;
830                 }
831                 next = zap_pud_range(tlb, vma, pgd, addr, next,
832                                                 zap_work, details);
833         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
834         tlb_end_vma(tlb, vma);
835
836         return addr;
837 }
838
839 #ifdef CONFIG_PREEMPT
840 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
841 #else
842 /* No preempt: go for improved straight-line efficiency */
843 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
844 #endif
845
846 /**
847  * unmap_vmas - unmap a range of memory covered by a list of vma's
848  * @tlbp: address of the caller's struct mmu_gather
849  * @vma: the starting vma
850  * @start_addr: virtual address at which to start unmapping
851  * @end_addr: virtual address at which to end unmapping
852  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
853  * @details: details of nonlinear truncation or shared cache invalidation
854  *
855  * Returns the end address of the unmapping (restart addr if interrupted).
856  *
857  * Unmap all pages in the vma list.
858  *
859  * We aim to not hold locks for too long (for scheduling latency reasons).
860  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
861  * return the ending mmu_gather to the caller.
862  *
863  * Only addresses between `start' and `end' will be unmapped.
864  *
865  * The VMA list must be sorted in ascending virtual address order.
866  *
867  * unmap_vmas() assumes that the caller will flush the whole unmapped address
868  * range after unmap_vmas() returns.  So the only responsibility here is to
869  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
870  * drops the lock and schedules.
871  */
872 unsigned long unmap_vmas(struct mmu_gather **tlbp,
873                 struct vm_area_struct *vma, unsigned long start_addr,
874                 unsigned long end_addr, unsigned long *nr_accounted,
875                 struct zap_details *details)
876 {
877         long zap_work = ZAP_BLOCK_SIZE;
878         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
879         int tlb_start_valid = 0;
880         unsigned long start = start_addr;
881         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
882         int fullmm = (*tlbp)->fullmm;
883
884         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
885                 unsigned long end;
886
887                 start = max(vma->vm_start, start_addr);
888                 if (start >= vma->vm_end)
889                         continue;
890                 end = min(vma->vm_end, end_addr);
891                 if (end <= vma->vm_start)
892                         continue;
893
894                 if (vma->vm_flags & VM_ACCOUNT)
895                         *nr_accounted += (end - start) >> PAGE_SHIFT;
896
897                 while (start != end) {
898                         if (!tlb_start_valid) {
899                                 tlb_start = start;
900                                 tlb_start_valid = 1;
901                         }
902
903                         if (unlikely(is_vm_hugetlb_page(vma))) {
904                                 unmap_hugepage_range(vma, start, end);
905                                 zap_work -= (end - start) /
906                                                 (HPAGE_SIZE / PAGE_SIZE);
907                                 start = end;
908                         } else
909                                 start = unmap_page_range(*tlbp, vma,
910                                                 start, end, &zap_work, details);
911
912                         if (zap_work > 0) {
913                                 BUG_ON(start != end);
914                                 break;
915                         }
916
917                         tlb_finish_mmu(*tlbp, tlb_start, start);
918
919                         if (need_resched() ||
920                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
921                                 if (i_mmap_lock) {
922                                         *tlbp = NULL;
923                                         goto out;
924                                 }
925                                 cond_resched();
926                         }
927
928                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
929                         tlb_start_valid = 0;
930                         zap_work = ZAP_BLOCK_SIZE;
931                 }
932         }
933 out:
934         return start;   /* which is now the end (or restart) address */
935 }
936
937 /**
938  * zap_page_range - remove user pages in a given range
939  * @vma: vm_area_struct holding the applicable pages
940  * @address: starting address of pages to zap
941  * @size: number of bytes to zap
942  * @details: details of nonlinear truncation or shared cache invalidation
943  */
944 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
945                 unsigned long size, struct zap_details *details)
946 {
947         struct mm_struct *mm = vma->vm_mm;
948         struct mmu_gather *tlb;
949         unsigned long end = address + size;
950         unsigned long nr_accounted = 0;
951
952         lru_add_drain();
953         tlb = tlb_gather_mmu(mm, 0);
954         update_hiwater_rss(mm);
955         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
956         if (tlb)
957                 tlb_finish_mmu(tlb, address, end);
958         return end;
959 }
960
961 /*
962  * Do a quick page-table lookup for a single page.
963  */
964 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
965                         unsigned int flags)
966 {
967         pgd_t *pgd;
968         pud_t *pud;
969         pmd_t *pmd;
970         pte_t *ptep, pte;
971         spinlock_t *ptl;
972         struct page *page;
973         struct mm_struct *mm = vma->vm_mm;
974
975         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
976         if (!IS_ERR(page)) {
977                 BUG_ON(flags & FOLL_GET);
978                 goto out;
979         }
980
981         page = NULL;
982         pgd = pgd_offset(mm, address);
983         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
984                 goto no_page_table;
985
986         pud = pud_offset(pgd, address);
987         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
988                 goto no_page_table;
989         
990         pmd = pmd_offset(pud, address);
991         if (pmd_none(*pmd))
992                 goto no_page_table;
993
994         if (pmd_huge(*pmd)) {
995                 BUG_ON(flags & FOLL_GET);
996                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
997                 goto out;
998         }
999
1000         if (unlikely(pmd_bad(*pmd)))
1001                 goto no_page_table;
1002
1003         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1004
1005         pte = *ptep;
1006         if (!pte_present(pte))
1007                 goto no_page;
1008         if ((flags & FOLL_WRITE) && !pte_write(pte))
1009                 goto unlock;
1010         page = vm_normal_page(vma, address, pte);
1011         if (unlikely(!page))
1012                 goto bad_page;
1013
1014         if (flags & FOLL_GET)
1015                 get_page(page);
1016         if (flags & FOLL_TOUCH) {
1017                 if ((flags & FOLL_WRITE) &&
1018                     !pte_dirty(pte) && !PageDirty(page))
1019                         set_page_dirty(page);
1020                 mark_page_accessed(page);
1021         }
1022 unlock:
1023         pte_unmap_unlock(ptep, ptl);
1024 out:
1025         return page;
1026
1027 bad_page:
1028         pte_unmap_unlock(ptep, ptl);
1029         return ERR_PTR(-EFAULT);
1030
1031 no_page:
1032         pte_unmap_unlock(ptep, ptl);
1033         if (!pte_none(pte))
1034                 return page;
1035         /* Fall through to ZERO_PAGE handling */
1036 no_page_table:
1037         /*
1038          * When core dumping an enormous anonymous area that nobody
1039          * has touched so far, we don't want to allocate page tables.
1040          */
1041         if (flags & FOLL_ANON) {
1042                 page = ZERO_PAGE(0);
1043                 if (flags & FOLL_GET)
1044                         get_page(page);
1045                 BUG_ON(flags & FOLL_WRITE);
1046         }
1047         return page;
1048 }
1049
1050 /* Can we do the FOLL_ANON optimization? */
1051 static inline int use_zero_page(struct vm_area_struct *vma)
1052 {
1053         /*
1054          * We don't want to optimize FOLL_ANON for make_pages_present()
1055          * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1056          * we want to get the page from the page tables to make sure
1057          * that we serialize and update with any other user of that
1058          * mapping.
1059          */
1060         if (vma->vm_flags & (VM_LOCKED | VM_SHARED))
1061                 return 0;
1062         /*
1063          * And if we have a fault routine, it's not an anonymous region.
1064          */
1065         return !vma->vm_ops || !vma->vm_ops->fault;
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  * vma cannot be a COW mapping.
1343  *
1344  * As this is called only for pages that do not currently exist, we
1345  * do not need to flush old virtual caches or the TLB.
1346  */
1347 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1348                         unsigned long pfn)
1349 {
1350         /*
1351          * Technically, architectures with pte_special can avoid all these
1352          * restrictions (same for remap_pfn_range).  However we would like
1353          * consistency in testing and feature parity among all, so we should
1354          * try to keep these invariants in place for everybody.
1355          */
1356         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1357         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1358                                                 (VM_PFNMAP|VM_MIXEDMAP));
1359         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1360         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1361
1362         if (addr < vma->vm_start || addr >= vma->vm_end)
1363                 return -EFAULT;
1364         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1365 }
1366 EXPORT_SYMBOL(vm_insert_pfn);
1367
1368 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1369                         unsigned long pfn)
1370 {
1371         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1372
1373         if (addr < vma->vm_start || addr >= vma->vm_end)
1374                 return -EFAULT;
1375
1376         /*
1377          * If we don't have pte special, then we have to use the pfn_valid()
1378          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1379          * refcount the page if pfn_valid is true (hence insert_page rather
1380          * than insert_pfn).
1381          */
1382         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1383                 struct page *page;
1384
1385                 page = pfn_to_page(pfn);
1386                 return insert_page(vma, addr, page, vma->vm_page_prot);
1387         }
1388         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1389 }
1390 EXPORT_SYMBOL(vm_insert_mixed);
1391
1392 /*
1393  * maps a range of physical memory into the requested pages. the old
1394  * mappings are removed. any references to nonexistent pages results
1395  * in null mappings (currently treated as "copy-on-access")
1396  */
1397 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1398                         unsigned long addr, unsigned long end,
1399                         unsigned long pfn, pgprot_t prot)
1400 {
1401         pte_t *pte;
1402         spinlock_t *ptl;
1403
1404         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1405         if (!pte)
1406                 return -ENOMEM;
1407         arch_enter_lazy_mmu_mode();
1408         do {
1409                 BUG_ON(!pte_none(*pte));
1410                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1411                 pfn++;
1412         } while (pte++, addr += PAGE_SIZE, addr != end);
1413         arch_leave_lazy_mmu_mode();
1414         pte_unmap_unlock(pte - 1, ptl);
1415         return 0;
1416 }
1417
1418 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1419                         unsigned long addr, unsigned long end,
1420                         unsigned long pfn, pgprot_t prot)
1421 {
1422         pmd_t *pmd;
1423         unsigned long next;
1424
1425         pfn -= addr >> PAGE_SHIFT;
1426         pmd = pmd_alloc(mm, pud, addr);
1427         if (!pmd)
1428                 return -ENOMEM;
1429         do {
1430                 next = pmd_addr_end(addr, end);
1431                 if (remap_pte_range(mm, pmd, addr, next,
1432                                 pfn + (addr >> PAGE_SHIFT), prot))
1433                         return -ENOMEM;
1434         } while (pmd++, addr = next, addr != end);
1435         return 0;
1436 }
1437
1438 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1439                         unsigned long addr, unsigned long end,
1440                         unsigned long pfn, pgprot_t prot)
1441 {
1442         pud_t *pud;
1443         unsigned long next;
1444
1445         pfn -= addr >> PAGE_SHIFT;
1446         pud = pud_alloc(mm, pgd, addr);
1447         if (!pud)
1448                 return -ENOMEM;
1449         do {
1450                 next = pud_addr_end(addr, end);
1451                 if (remap_pmd_range(mm, pud, addr, next,
1452                                 pfn + (addr >> PAGE_SHIFT), prot))
1453                         return -ENOMEM;
1454         } while (pud++, addr = next, addr != end);
1455         return 0;
1456 }
1457
1458 /**
1459  * remap_pfn_range - remap kernel memory to userspace
1460  * @vma: user vma to map to
1461  * @addr: target user address to start at
1462  * @pfn: physical address of kernel memory
1463  * @size: size of map area
1464  * @prot: page protection flags for this mapping
1465  *
1466  *  Note: this is only safe if the mm semaphore is held when called.
1467  */
1468 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1469                     unsigned long pfn, unsigned long size, pgprot_t prot)
1470 {
1471         pgd_t *pgd;
1472         unsigned long next;
1473         unsigned long end = addr + PAGE_ALIGN(size);
1474         struct mm_struct *mm = vma->vm_mm;
1475         int err;
1476
1477         /*
1478          * Physically remapped pages are special. Tell the
1479          * rest of the world about it:
1480          *   VM_IO tells people not to look at these pages
1481          *      (accesses can have side effects).
1482          *   VM_RESERVED is specified all over the place, because
1483          *      in 2.4 it kept swapout's vma scan off this vma; but
1484          *      in 2.6 the LRU scan won't even find its pages, so this
1485          *      flag means no more than count its pages in reserved_vm,
1486          *      and omit it from core dump, even when VM_IO turned off.
1487          *   VM_PFNMAP tells the core MM that the base pages are just
1488          *      raw PFN mappings, and do not have a "struct page" associated
1489          *      with them.
1490          *
1491          * There's a horrible special case to handle copy-on-write
1492          * behaviour that some programs depend on. We mark the "original"
1493          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1494          */
1495         if (is_cow_mapping(vma->vm_flags)) {
1496                 if (addr != vma->vm_start || end != vma->vm_end)
1497                         return -EINVAL;
1498                 vma->vm_pgoff = pfn;
1499         }
1500
1501         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1502
1503         BUG_ON(addr >= end);
1504         pfn -= addr >> PAGE_SHIFT;
1505         pgd = pgd_offset(mm, addr);
1506         flush_cache_range(vma, addr, end);
1507         do {
1508                 next = pgd_addr_end(addr, end);
1509                 err = remap_pud_range(mm, pgd, addr, next,
1510                                 pfn + (addr >> PAGE_SHIFT), prot);
1511                 if (err)
1512                         break;
1513         } while (pgd++, addr = next, addr != end);
1514         return err;
1515 }
1516 EXPORT_SYMBOL(remap_pfn_range);
1517
1518 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1519                                      unsigned long addr, unsigned long end,
1520                                      pte_fn_t fn, void *data)
1521 {
1522         pte_t *pte;
1523         int err;
1524         pgtable_t token;
1525         spinlock_t *uninitialized_var(ptl);
1526
1527         pte = (mm == &init_mm) ?
1528                 pte_alloc_kernel(pmd, addr) :
1529                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1530         if (!pte)
1531                 return -ENOMEM;
1532
1533         BUG_ON(pmd_huge(*pmd));
1534
1535         token = pmd_pgtable(*pmd);
1536
1537         do {
1538                 err = fn(pte, token, addr, data);
1539                 if (err)
1540                         break;
1541         } while (pte++, addr += PAGE_SIZE, addr != end);
1542
1543         if (mm != &init_mm)
1544                 pte_unmap_unlock(pte-1, ptl);
1545         return err;
1546 }
1547
1548 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1549                                      unsigned long addr, unsigned long end,
1550                                      pte_fn_t fn, void *data)
1551 {
1552         pmd_t *pmd;
1553         unsigned long next;
1554         int err;
1555
1556         pmd = pmd_alloc(mm, pud, addr);
1557         if (!pmd)
1558                 return -ENOMEM;
1559         do {
1560                 next = pmd_addr_end(addr, end);
1561                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1562                 if (err)
1563                         break;
1564         } while (pmd++, addr = next, addr != end);
1565         return err;
1566 }
1567
1568 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1569                                      unsigned long addr, unsigned long end,
1570                                      pte_fn_t fn, void *data)
1571 {
1572         pud_t *pud;
1573         unsigned long next;
1574         int err;
1575
1576         pud = pud_alloc(mm, pgd, addr);
1577         if (!pud)
1578                 return -ENOMEM;
1579         do {
1580                 next = pud_addr_end(addr, end);
1581                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1582                 if (err)
1583                         break;
1584         } while (pud++, addr = next, addr != end);
1585         return err;
1586 }
1587
1588 /*
1589  * Scan a region of virtual memory, filling in page tables as necessary
1590  * and calling a provided function on each leaf page table.
1591  */
1592 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1593                         unsigned long size, pte_fn_t fn, void *data)
1594 {
1595         pgd_t *pgd;
1596         unsigned long next;
1597         unsigned long end = addr + size;
1598         int err;
1599
1600         BUG_ON(addr >= end);
1601         pgd = pgd_offset(mm, addr);
1602         do {
1603                 next = pgd_addr_end(addr, end);
1604                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1605                 if (err)
1606                         break;
1607         } while (pgd++, addr = next, addr != end);
1608         return err;
1609 }
1610 EXPORT_SYMBOL_GPL(apply_to_page_range);
1611
1612 /*
1613  * handle_pte_fault chooses page fault handler according to an entry
1614  * which was read non-atomically.  Before making any commitment, on
1615  * those architectures or configurations (e.g. i386 with PAE) which
1616  * might give a mix of unmatched parts, do_swap_page and do_file_page
1617  * must check under lock before unmapping the pte and proceeding
1618  * (but do_wp_page is only called after already making such a check;
1619  * and do_anonymous_page and do_no_page can safely check later on).
1620  */
1621 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1622                                 pte_t *page_table, pte_t orig_pte)
1623 {
1624         int same = 1;
1625 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1626         if (sizeof(pte_t) > sizeof(unsigned long)) {
1627                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1628                 spin_lock(ptl);
1629                 same = pte_same(*page_table, orig_pte);
1630                 spin_unlock(ptl);
1631         }
1632 #endif
1633         pte_unmap(page_table);
1634         return same;
1635 }
1636
1637 /*
1638  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1639  * servicing faults for write access.  In the normal case, do always want
1640  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1641  * that do not have writing enabled, when used by access_process_vm.
1642  */
1643 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1644 {
1645         if (likely(vma->vm_flags & VM_WRITE))
1646                 pte = pte_mkwrite(pte);
1647         return pte;
1648 }
1649
1650 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1651 {
1652         /*
1653          * If the source page was a PFN mapping, we don't have
1654          * a "struct page" for it. We do a best-effort copy by
1655          * just copying from the original user address. If that
1656          * fails, we just zero-fill it. Live with it.
1657          */
1658         if (unlikely(!src)) {
1659                 void *kaddr = kmap_atomic(dst, KM_USER0);
1660                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1661
1662                 /*
1663                  * This really shouldn't fail, because the page is there
1664                  * in the page tables. But it might just be unreadable,
1665                  * in which case we just give up and fill the result with
1666                  * zeroes.
1667                  */
1668                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1669                         memset(kaddr, 0, PAGE_SIZE);
1670                 kunmap_atomic(kaddr, KM_USER0);
1671                 flush_dcache_page(dst);
1672         } else
1673                 copy_user_highpage(dst, src, va, vma);
1674 }
1675
1676 /*
1677  * This routine handles present pages, when users try to write
1678  * to a shared page. It is done by copying the page to a new address
1679  * and decrementing the shared-page counter for the old page.
1680  *
1681  * Note that this routine assumes that the protection checks have been
1682  * done by the caller (the low-level page fault routine in most cases).
1683  * Thus we can safely just mark it writable once we've done any necessary
1684  * COW.
1685  *
1686  * We also mark the page dirty at this point even though the page will
1687  * change only once the write actually happens. This avoids a few races,
1688  * and potentially makes it more efficient.
1689  *
1690  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1691  * but allow concurrent faults), with pte both mapped and locked.
1692  * We return with mmap_sem still held, but pte unmapped and unlocked.
1693  */
1694 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1695                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1696                 spinlock_t *ptl, pte_t orig_pte)
1697 {
1698         struct page *old_page, *new_page;
1699         pte_t entry;
1700         int reuse = 0, ret = 0;
1701         int page_mkwrite = 0;
1702         struct page *dirty_page = NULL;
1703
1704         old_page = vm_normal_page(vma, address, orig_pte);
1705         if (!old_page) {
1706                 /*
1707                  * VM_MIXEDMAP !pfn_valid() case
1708                  *
1709                  * We should not cow pages in a shared writeable mapping.
1710                  * Just mark the pages writable as we can't do any dirty
1711                  * accounting on raw pfn maps.
1712                  */
1713                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1714                                      (VM_WRITE|VM_SHARED))
1715                         goto reuse;
1716                 goto gotten;
1717         }
1718
1719         /*
1720          * Take out anonymous pages first, anonymous shared vmas are
1721          * not dirty accountable.
1722          */
1723         if (PageAnon(old_page)) {
1724                 if (!TestSetPageLocked(old_page)) {
1725                         reuse = can_share_swap_page(old_page);
1726                         unlock_page(old_page);
1727                 }
1728         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1729                                         (VM_WRITE|VM_SHARED))) {
1730                 /*
1731                  * Only catch write-faults on shared writable pages,
1732                  * read-only shared pages can get COWed by
1733                  * get_user_pages(.write=1, .force=1).
1734                  */
1735                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1736                         /*
1737                          * Notify the address space that the page is about to
1738                          * become writable so that it can prohibit this or wait
1739                          * for the page to get into an appropriate state.
1740                          *
1741                          * We do this without the lock held, so that it can
1742                          * sleep if it needs to.
1743                          */
1744                         page_cache_get(old_page);
1745                         pte_unmap_unlock(page_table, ptl);
1746
1747                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1748                                 goto unwritable_page;
1749
1750                         /*
1751                          * Since we dropped the lock we need to revalidate
1752                          * the PTE as someone else may have changed it.  If
1753                          * they did, we just return, as we can count on the
1754                          * MMU to tell us if they didn't also make it writable.
1755                          */
1756                         page_table = pte_offset_map_lock(mm, pmd, address,
1757                                                          &ptl);
1758                         page_cache_release(old_page);
1759                         if (!pte_same(*page_table, orig_pte))
1760                                 goto unlock;
1761
1762                         page_mkwrite = 1;
1763                 }
1764                 dirty_page = old_page;
1765                 get_page(dirty_page);
1766                 reuse = 1;
1767         }
1768
1769         if (reuse) {
1770 reuse:
1771                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1772                 entry = pte_mkyoung(orig_pte);
1773                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1774                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1775                         update_mmu_cache(vma, address, entry);
1776                 ret |= VM_FAULT_WRITE;
1777                 goto unlock;
1778         }
1779
1780         /*
1781          * Ok, we need to copy. Oh, well..
1782          */
1783         page_cache_get(old_page);
1784 gotten:
1785         pte_unmap_unlock(page_table, ptl);
1786
1787         if (unlikely(anon_vma_prepare(vma)))
1788                 goto oom;
1789         VM_BUG_ON(old_page == ZERO_PAGE(0));
1790         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1791         if (!new_page)
1792                 goto oom;
1793         cow_user_page(new_page, old_page, address, vma);
1794         __SetPageUptodate(new_page);
1795
1796         if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
1797                 goto oom_free_new;
1798
1799         /*
1800          * Re-check the pte - we dropped the lock
1801          */
1802         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1803         if (likely(pte_same(*page_table, orig_pte))) {
1804                 if (old_page) {
1805                         if (!PageAnon(old_page)) {
1806                                 dec_mm_counter(mm, file_rss);
1807                                 inc_mm_counter(mm, anon_rss);
1808                         }
1809                 } else
1810                         inc_mm_counter(mm, anon_rss);
1811                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1812                 entry = mk_pte(new_page, vma->vm_page_prot);
1813                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1814                 /*
1815                  * Clear the pte entry and flush it first, before updating the
1816                  * pte with the new entry. This will avoid a race condition
1817                  * seen in the presence of one thread doing SMC and another
1818                  * thread doing COW.
1819                  */
1820                 ptep_clear_flush(vma, address, page_table);
1821                 set_pte_at(mm, address, page_table, entry);
1822                 update_mmu_cache(vma, address, entry);
1823                 lru_cache_add_active(new_page);
1824                 page_add_new_anon_rmap(new_page, vma, address);
1825
1826                 if (old_page) {
1827                         /*
1828                          * Only after switching the pte to the new page may
1829                          * we remove the mapcount here. Otherwise another
1830                          * process may come and find the rmap count decremented
1831                          * before the pte is switched to the new page, and
1832                          * "reuse" the old page writing into it while our pte
1833                          * here still points into it and can be read by other
1834                          * threads.
1835                          *
1836                          * The critical issue is to order this
1837                          * page_remove_rmap with the ptp_clear_flush above.
1838                          * Those stores are ordered by (if nothing else,)
1839                          * the barrier present in the atomic_add_negative
1840                          * in page_remove_rmap.
1841                          *
1842                          * Then the TLB flush in ptep_clear_flush ensures that
1843                          * no process can access the old page before the
1844                          * decremented mapcount is visible. And the old page
1845                          * cannot be reused until after the decremented
1846                          * mapcount is visible. So transitively, TLBs to
1847                          * old page will be flushed before it can be reused.
1848                          */
1849                         page_remove_rmap(old_page, vma);
1850                 }
1851
1852                 /* Free the old page.. */
1853                 new_page = old_page;
1854                 ret |= VM_FAULT_WRITE;
1855         } else
1856                 mem_cgroup_uncharge_page(new_page);
1857
1858         if (new_page)
1859                 page_cache_release(new_page);
1860         if (old_page)
1861                 page_cache_release(old_page);
1862 unlock:
1863         pte_unmap_unlock(page_table, ptl);
1864         if (dirty_page) {
1865                 if (vma->vm_file)
1866                         file_update_time(vma->vm_file);
1867
1868                 /*
1869                  * Yes, Virginia, this is actually required to prevent a race
1870                  * with clear_page_dirty_for_io() from clearing the page dirty
1871                  * bit after it clear all dirty ptes, but before a racing
1872                  * do_wp_page installs a dirty pte.
1873                  *
1874                  * do_no_page is protected similarly.
1875                  */
1876                 wait_on_page_locked(dirty_page);
1877                 set_page_dirty_balance(dirty_page, page_mkwrite);
1878                 put_page(dirty_page);
1879         }
1880         return ret;
1881 oom_free_new:
1882         page_cache_release(new_page);
1883 oom:
1884         if (old_page)
1885                 page_cache_release(old_page);
1886         return VM_FAULT_OOM;
1887
1888 unwritable_page:
1889         page_cache_release(old_page);
1890         return VM_FAULT_SIGBUS;
1891 }
1892
1893 /*
1894  * Helper functions for unmap_mapping_range().
1895  *
1896  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1897  *
1898  * We have to restart searching the prio_tree whenever we drop the lock,
1899  * since the iterator is only valid while the lock is held, and anyway
1900  * a later vma might be split and reinserted earlier while lock dropped.
1901  *
1902  * The list of nonlinear vmas could be handled more efficiently, using
1903  * a placeholder, but handle it in the same way until a need is shown.
1904  * It is important to search the prio_tree before nonlinear list: a vma
1905  * may become nonlinear and be shifted from prio_tree to nonlinear list
1906  * while the lock is dropped; but never shifted from list to prio_tree.
1907  *
1908  * In order to make forward progress despite restarting the search,
1909  * vm_truncate_count is used to mark a vma as now dealt with, so we can
1910  * quickly skip it next time around.  Since the prio_tree search only
1911  * shows us those vmas affected by unmapping the range in question, we
1912  * can't efficiently keep all vmas in step with mapping->truncate_count:
1913  * so instead reset them all whenever it wraps back to 0 (then go to 1).
1914  * mapping->truncate_count and vma->vm_truncate_count are protected by
1915  * i_mmap_lock.
1916  *
1917  * In order to make forward progress despite repeatedly restarting some
1918  * large vma, note the restart_addr from unmap_vmas when it breaks out:
1919  * and restart from that address when we reach that vma again.  It might
1920  * have been split or merged, shrunk or extended, but never shifted: so
1921  * restart_addr remains valid so long as it remains in the vma's range.
1922  * unmap_mapping_range forces truncate_count to leap over page-aligned
1923  * values so we can save vma's restart_addr in its truncate_count field.
1924  */
1925 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1926
1927 static void reset_vma_truncate_counts(struct address_space *mapping)
1928 {
1929         struct vm_area_struct *vma;
1930         struct prio_tree_iter iter;
1931
1932         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1933                 vma->vm_truncate_count = 0;
1934         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1935                 vma->vm_truncate_count = 0;
1936 }
1937
1938 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1939                 unsigned long start_addr, unsigned long end_addr,
1940                 struct zap_details *details)
1941 {
1942         unsigned long restart_addr;
1943         int need_break;
1944
1945         /*
1946          * files that support invalidating or truncating portions of the
1947          * file from under mmaped areas must have their ->fault function
1948          * return a locked page (and set VM_FAULT_LOCKED in the return).
1949          * This provides synchronisation against concurrent unmapping here.
1950          */
1951
1952 again:
1953         restart_addr = vma->vm_truncate_count;
1954         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1955                 start_addr = restart_addr;
1956                 if (start_addr >= end_addr) {
1957                         /* Top of vma has been split off since last time */
1958                         vma->vm_truncate_count = details->truncate_count;
1959                         return 0;
1960                 }
1961         }
1962
1963         restart_addr = zap_page_range(vma, start_addr,
1964                                         end_addr - start_addr, details);
1965         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
1966
1967         if (restart_addr >= end_addr) {
1968                 /* We have now completed this vma: mark it so */
1969                 vma->vm_truncate_count = details->truncate_count;
1970                 if (!need_break)
1971                         return 0;
1972         } else {
1973                 /* Note restart_addr in vma's truncate_count field */
1974                 vma->vm_truncate_count = restart_addr;
1975                 if (!need_break)
1976                         goto again;
1977         }
1978
1979         spin_unlock(details->i_mmap_lock);
1980         cond_resched();
1981         spin_lock(details->i_mmap_lock);
1982         return -EINTR;
1983 }
1984
1985 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1986                                             struct zap_details *details)
1987 {
1988         struct vm_area_struct *vma;
1989         struct prio_tree_iter iter;
1990         pgoff_t vba, vea, zba, zea;
1991
1992 restart:
1993         vma_prio_tree_foreach(vma, &iter, root,
1994                         details->first_index, details->last_index) {
1995                 /* Skip quickly over those we have already dealt with */
1996                 if (vma->vm_truncate_count == details->truncate_count)
1997                         continue;
1998
1999                 vba = vma->vm_pgoff;
2000                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2001                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2002                 zba = details->first_index;
2003                 if (zba < vba)
2004                         zba = vba;
2005                 zea = details->last_index;
2006                 if (zea > vea)
2007                         zea = vea;
2008
2009                 if (unmap_mapping_range_vma(vma,
2010                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2011                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2012                                 details) < 0)
2013                         goto restart;
2014         }
2015 }
2016
2017 static inline void unmap_mapping_range_list(struct list_head *head,
2018                                             struct zap_details *details)
2019 {
2020         struct vm_area_struct *vma;
2021
2022         /*
2023          * In nonlinear VMAs there is no correspondence between virtual address
2024          * offset and file offset.  So we must perform an exhaustive search
2025          * across *all* the pages in each nonlinear VMA, not just the pages
2026          * whose virtual address lies outside the file truncation point.
2027          */
2028 restart:
2029         list_for_each_entry(vma, head, shared.vm_set.list) {
2030                 /* Skip quickly over those we have already dealt with */
2031                 if (vma->vm_truncate_count == details->truncate_count)
2032                         continue;
2033                 details->nonlinear_vma = vma;
2034                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2035                                         vma->vm_end, details) < 0)
2036                         goto restart;
2037         }
2038 }
2039
2040 /**
2041  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2042  * @mapping: the address space containing mmaps to be unmapped.
2043  * @holebegin: byte in first page to unmap, relative to the start of
2044  * the underlying file.  This will be rounded down to a PAGE_SIZE
2045  * boundary.  Note that this is different from vmtruncate(), which
2046  * must keep the partial page.  In contrast, we must get rid of
2047  * partial pages.
2048  * @holelen: size of prospective hole in bytes.  This will be rounded
2049  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2050  * end of the file.
2051  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2052  * but 0 when invalidating pagecache, don't throw away private data.
2053  */
2054 void unmap_mapping_range(struct address_space *mapping,
2055                 loff_t const holebegin, loff_t const holelen, int even_cows)
2056 {
2057         struct zap_details details;
2058         pgoff_t hba = holebegin >> PAGE_SHIFT;
2059         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2060
2061         /* Check for overflow. */
2062         if (sizeof(holelen) > sizeof(hlen)) {
2063                 long long holeend =
2064                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2065                 if (holeend & ~(long long)ULONG_MAX)
2066                         hlen = ULONG_MAX - hba + 1;
2067         }
2068
2069         details.check_mapping = even_cows? NULL: mapping;
2070         details.nonlinear_vma = NULL;
2071         details.first_index = hba;
2072         details.last_index = hba + hlen - 1;
2073         if (details.last_index < details.first_index)
2074                 details.last_index = ULONG_MAX;
2075         details.i_mmap_lock = &mapping->i_mmap_lock;
2076
2077         spin_lock(&mapping->i_mmap_lock);
2078
2079         /* Protect against endless unmapping loops */
2080         mapping->truncate_count++;
2081         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2082                 if (mapping->truncate_count == 0)
2083                         reset_vma_truncate_counts(mapping);
2084                 mapping->truncate_count++;
2085         }
2086         details.truncate_count = mapping->truncate_count;
2087
2088         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2089                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2090         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2091                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2092         spin_unlock(&mapping->i_mmap_lock);
2093 }
2094 EXPORT_SYMBOL(unmap_mapping_range);
2095
2096 /**
2097  * vmtruncate - unmap mappings "freed" by truncate() syscall
2098  * @inode: inode of the file used
2099  * @offset: file offset to start truncating
2100  *
2101  * NOTE! We have to be ready to update the memory sharing
2102  * between the file and the memory map for a potential last
2103  * incomplete page.  Ugly, but necessary.
2104  */
2105 int vmtruncate(struct inode * inode, loff_t offset)
2106 {
2107         if (inode->i_size < offset) {
2108                 unsigned long limit;
2109
2110                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2111                 if (limit != RLIM_INFINITY && offset > limit)
2112                         goto out_sig;
2113                 if (offset > inode->i_sb->s_maxbytes)
2114                         goto out_big;
2115                 i_size_write(inode, offset);
2116         } else {
2117                 struct address_space *mapping = inode->i_mapping;
2118
2119                 /*
2120                  * truncation of in-use swapfiles is disallowed - it would
2121                  * cause subsequent swapout to scribble on the now-freed
2122                  * blocks.
2123                  */
2124                 if (IS_SWAPFILE(inode))
2125                         return -ETXTBSY;
2126                 i_size_write(inode, offset);
2127
2128                 /*
2129                  * unmap_mapping_range is called twice, first simply for
2130                  * efficiency so that truncate_inode_pages does fewer
2131                  * single-page unmaps.  However after this first call, and
2132                  * before truncate_inode_pages finishes, it is possible for
2133                  * private pages to be COWed, which remain after
2134                  * truncate_inode_pages finishes, hence the second
2135                  * unmap_mapping_range call must be made for correctness.
2136                  */
2137                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2138                 truncate_inode_pages(mapping, offset);
2139                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2140         }
2141
2142         if (inode->i_op && inode->i_op->truncate)
2143                 inode->i_op->truncate(inode);
2144         return 0;
2145
2146 out_sig:
2147         send_sig(SIGXFSZ, current, 0);
2148 out_big:
2149         return -EFBIG;
2150 }
2151 EXPORT_SYMBOL(vmtruncate);
2152
2153 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2154 {
2155         struct address_space *mapping = inode->i_mapping;
2156
2157         /*
2158          * If the underlying filesystem is not going to provide
2159          * a way to truncate a range of blocks (punch a hole) -
2160          * we should return failure right now.
2161          */
2162         if (!inode->i_op || !inode->i_op->truncate_range)
2163                 return -ENOSYS;
2164
2165         mutex_lock(&inode->i_mutex);
2166         down_write(&inode->i_alloc_sem);
2167         unmap_mapping_range(mapping, offset, (end - offset), 1);
2168         truncate_inode_pages_range(mapping, offset, end);
2169         unmap_mapping_range(mapping, offset, (end - offset), 1);
2170         inode->i_op->truncate_range(inode, offset, end);
2171         up_write(&inode->i_alloc_sem);
2172         mutex_unlock(&inode->i_mutex);
2173
2174         return 0;
2175 }
2176
2177 /*
2178  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2179  * but allow concurrent faults), and pte mapped but not yet locked.
2180  * We return with mmap_sem still held, but pte unmapped and unlocked.
2181  */
2182 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2183                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2184                 int write_access, pte_t orig_pte)
2185 {
2186         spinlock_t *ptl;
2187         struct page *page;
2188         swp_entry_t entry;
2189         pte_t pte;
2190         int ret = 0;
2191
2192         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2193                 goto out;
2194
2195         entry = pte_to_swp_entry(orig_pte);
2196         if (is_migration_entry(entry)) {
2197                 migration_entry_wait(mm, pmd, address);
2198                 goto out;
2199         }
2200         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2201         page = lookup_swap_cache(entry);
2202         if (!page) {
2203                 grab_swap_token(); /* Contend for token _before_ read-in */
2204                 page = swapin_readahead(entry,
2205                                         GFP_HIGHUSER_MOVABLE, vma, address);
2206                 if (!page) {
2207                         /*
2208                          * Back out if somebody else faulted in this pte
2209                          * while we released the pte lock.
2210                          */
2211                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2212                         if (likely(pte_same(*page_table, orig_pte)))
2213                                 ret = VM_FAULT_OOM;
2214                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2215                         goto unlock;
2216                 }
2217
2218                 /* Had to read the page from swap area: Major fault */
2219                 ret = VM_FAULT_MAJOR;
2220                 count_vm_event(PGMAJFAULT);
2221         }
2222
2223         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2224                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2225                 ret = VM_FAULT_OOM;
2226                 goto out;
2227         }
2228
2229         mark_page_accessed(page);
2230         lock_page(page);
2231         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2232
2233         /*
2234          * Back out if somebody else already faulted in this pte.
2235          */
2236         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2237         if (unlikely(!pte_same(*page_table, orig_pte)))
2238                 goto out_nomap;
2239
2240         if (unlikely(!PageUptodate(page))) {
2241                 ret = VM_FAULT_SIGBUS;
2242                 goto out_nomap;
2243         }
2244
2245         /* The page isn't present yet, go ahead with the fault. */
2246
2247         inc_mm_counter(mm, anon_rss);
2248         pte = mk_pte(page, vma->vm_page_prot);
2249         if (write_access && can_share_swap_page(page)) {
2250                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2251                 write_access = 0;
2252         }
2253
2254         flush_icache_page(vma, page);
2255         set_pte_at(mm, address, page_table, pte);
2256         page_add_anon_rmap(page, vma, address);
2257
2258         swap_free(entry);
2259         if (vm_swap_full())
2260                 remove_exclusive_swap_page(page);
2261         unlock_page(page);
2262
2263         if (write_access) {
2264                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2265                 if (ret & VM_FAULT_ERROR)
2266                         ret &= VM_FAULT_ERROR;
2267                 goto out;
2268         }
2269
2270         /* No need to invalidate - it was non-present before */
2271         update_mmu_cache(vma, address, pte);
2272 unlock:
2273         pte_unmap_unlock(page_table, ptl);
2274 out:
2275         return ret;
2276 out_nomap:
2277         mem_cgroup_uncharge_page(page);
2278         pte_unmap_unlock(page_table, ptl);
2279         unlock_page(page);
2280         page_cache_release(page);
2281         return ret;
2282 }
2283
2284 /*
2285  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2286  * but allow concurrent faults), and pte mapped but not yet locked.
2287  * We return with mmap_sem still held, but pte unmapped and unlocked.
2288  */
2289 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2290                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2291                 int write_access)
2292 {
2293         struct page *page;
2294         spinlock_t *ptl;
2295         pte_t entry;
2296
2297         /* Allocate our own private page. */
2298         pte_unmap(page_table);
2299
2300         if (unlikely(anon_vma_prepare(vma)))
2301                 goto oom;
2302         page = alloc_zeroed_user_highpage_movable(vma, address);
2303         if (!page)
2304                 goto oom;
2305         __SetPageUptodate(page);
2306
2307         if (mem_cgroup_charge(page, mm, GFP_KERNEL))
2308                 goto oom_free_page;
2309
2310         entry = mk_pte(page, vma->vm_page_prot);
2311         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2312
2313         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2314         if (!pte_none(*page_table))
2315                 goto release;
2316         inc_mm_counter(mm, anon_rss);
2317         lru_cache_add_active(page);
2318         page_add_new_anon_rmap(page, vma, address);
2319         set_pte_at(mm, address, page_table, entry);
2320
2321         /* No need to invalidate - it was non-present before */
2322         update_mmu_cache(vma, address, entry);
2323 unlock:
2324         pte_unmap_unlock(page_table, ptl);
2325         return 0;
2326 release:
2327         mem_cgroup_uncharge_page(page);
2328         page_cache_release(page);
2329         goto unlock;
2330 oom_free_page:
2331         page_cache_release(page);
2332 oom:
2333         return VM_FAULT_OOM;
2334 }
2335
2336 /*
2337  * __do_fault() tries to create a new page mapping. It aggressively
2338  * tries to share with existing pages, but makes a separate copy if
2339  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2340  * the next page fault.
2341  *
2342  * As this is called only for pages that do not currently exist, we
2343  * do not need to flush old virtual caches or the TLB.
2344  *
2345  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2346  * but allow concurrent faults), and pte neither mapped nor locked.
2347  * We return with mmap_sem still held, but pte unmapped and unlocked.
2348  */
2349 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2350                 unsigned long address, pmd_t *pmd,
2351                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2352 {
2353         pte_t *page_table;
2354         spinlock_t *ptl;
2355         struct page *page;
2356         pte_t entry;
2357         int anon = 0;
2358         struct page *dirty_page = NULL;
2359         struct vm_fault vmf;
2360         int ret;
2361         int page_mkwrite = 0;
2362
2363         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2364         vmf.pgoff = pgoff;
2365         vmf.flags = flags;
2366         vmf.page = NULL;
2367
2368         ret = vma->vm_ops->fault(vma, &vmf);
2369         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2370                 return ret;
2371
2372         /*
2373          * For consistency in subsequent calls, make the faulted page always
2374          * locked.
2375          */
2376         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2377                 lock_page(vmf.page);
2378         else
2379                 VM_BUG_ON(!PageLocked(vmf.page));
2380
2381         /*
2382          * Should we do an early C-O-W break?
2383          */
2384         page = vmf.page;
2385         if (flags & FAULT_FLAG_WRITE) {
2386                 if (!(vma->vm_flags & VM_SHARED)) {
2387                         anon = 1;
2388                         if (unlikely(anon_vma_prepare(vma))) {
2389                                 ret = VM_FAULT_OOM;
2390                                 goto out;
2391                         }
2392                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2393                                                 vma, address);
2394                         if (!page) {
2395                                 ret = VM_FAULT_OOM;
2396                                 goto out;
2397                         }
2398                         copy_user_highpage(page, vmf.page, address, vma);
2399                         __SetPageUptodate(page);
2400                 } else {
2401                         /*
2402                          * If the page will be shareable, see if the backing
2403                          * address space wants to know that the page is about
2404                          * to become writable
2405                          */
2406                         if (vma->vm_ops->page_mkwrite) {
2407                                 unlock_page(page);
2408                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2409                                         ret = VM_FAULT_SIGBUS;
2410                                         anon = 1; /* no anon but release vmf.page */
2411                                         goto out_unlocked;
2412                                 }
2413                                 lock_page(page);
2414                                 /*
2415                                  * XXX: this is not quite right (racy vs
2416                                  * invalidate) to unlock and relock the page
2417                                  * like this, however a better fix requires
2418                                  * reworking page_mkwrite locking API, which
2419                                  * is better done later.
2420                                  */
2421                                 if (!page->mapping) {
2422                                         ret = 0;
2423                                         anon = 1; /* no anon but release vmf.page */
2424                                         goto out;
2425                                 }
2426                                 page_mkwrite = 1;
2427                         }
2428                 }
2429
2430         }
2431
2432         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2433                 ret = VM_FAULT_OOM;
2434                 goto out;
2435         }
2436
2437         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2438
2439         /*
2440          * This silly early PAGE_DIRTY setting removes a race
2441          * due to the bad i386 page protection. But it's valid
2442          * for other architectures too.
2443          *
2444          * Note that if write_access is true, we either now have
2445          * an exclusive copy of the page, or this is a shared mapping,
2446          * so we can make it writable and dirty to avoid having to
2447          * handle that later.
2448          */
2449         /* Only go through if we didn't race with anybody else... */
2450         if (likely(pte_same(*page_table, orig_pte))) {
2451                 flush_icache_page(vma, page);
2452                 entry = mk_pte(page, vma->vm_page_prot);
2453                 if (flags & FAULT_FLAG_WRITE)
2454                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2455                 set_pte_at(mm, address, page_table, entry);
2456                 if (anon) {
2457                         inc_mm_counter(mm, anon_rss);
2458                         lru_cache_add_active(page);
2459                         page_add_new_anon_rmap(page, vma, address);
2460                 } else {
2461                         inc_mm_counter(mm, file_rss);
2462                         page_add_file_rmap(page);
2463                         if (flags & FAULT_FLAG_WRITE) {
2464                                 dirty_page = page;
2465                                 get_page(dirty_page);
2466                         }
2467                 }
2468
2469                 /* no need to invalidate: a not-present page won't be cached */
2470                 update_mmu_cache(vma, address, entry);
2471         } else {
2472                 mem_cgroup_uncharge_page(page);
2473                 if (anon)
2474                         page_cache_release(page);
2475                 else
2476                         anon = 1; /* no anon but release faulted_page */
2477         }
2478
2479         pte_unmap_unlock(page_table, ptl);
2480
2481 out:
2482         unlock_page(vmf.page);
2483 out_unlocked:
2484         if (anon)
2485                 page_cache_release(vmf.page);
2486         else if (dirty_page) {
2487                 if (vma->vm_file)
2488                         file_update_time(vma->vm_file);
2489
2490                 set_page_dirty_balance(dirty_page, page_mkwrite);
2491                 put_page(dirty_page);
2492         }
2493
2494         return ret;
2495 }
2496
2497 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2498                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2499                 int write_access, pte_t orig_pte)
2500 {
2501         pgoff_t pgoff = (((address & PAGE_MASK)
2502                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2503         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2504
2505         pte_unmap(page_table);
2506         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2507 }
2508
2509 /*
2510  * Fault of a previously existing named mapping. Repopulate the pte
2511  * from the encoded file_pte if possible. This enables swappable
2512  * nonlinear vmas.
2513  *
2514  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2515  * but allow concurrent faults), and pte mapped but not yet locked.
2516  * We return with mmap_sem still held, but pte unmapped and unlocked.
2517  */
2518 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2519                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2520                 int write_access, pte_t orig_pte)
2521 {
2522         unsigned int flags = FAULT_FLAG_NONLINEAR |
2523                                 (write_access ? FAULT_FLAG_WRITE : 0);
2524         pgoff_t pgoff;
2525
2526         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2527                 return 0;
2528
2529         if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2530                         !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2531                 /*
2532                  * Page table corrupted: show pte and kill process.
2533                  */
2534                 print_bad_pte(vma, orig_pte, address);
2535                 return VM_FAULT_OOM;
2536         }
2537
2538         pgoff = pte_to_pgoff(orig_pte);
2539         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2540 }
2541
2542 /*
2543  * These routines also need to handle stuff like marking pages dirty
2544  * and/or accessed for architectures that don't do it in hardware (most
2545  * RISC architectures).  The early dirtying is also good on the i386.
2546  *
2547  * There is also a hook called "update_mmu_cache()" that architectures
2548  * with external mmu caches can use to update those (ie the Sparc or
2549  * PowerPC hashed page tables that act as extended TLBs).
2550  *
2551  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2552  * but allow concurrent faults), and pte mapped but not yet locked.
2553  * We return with mmap_sem still held, but pte unmapped and unlocked.
2554  */
2555 static inline int handle_pte_fault(struct mm_struct *mm,
2556                 struct vm_area_struct *vma, unsigned long address,
2557                 pte_t *pte, pmd_t *pmd, int write_access)
2558 {
2559         pte_t entry;
2560         spinlock_t *ptl;
2561
2562         entry = *pte;
2563         if (!pte_present(entry)) {
2564                 if (pte_none(entry)) {
2565                         if (vma->vm_ops) {
2566                                 if (likely(vma->vm_ops->fault))
2567                                         return do_linear_fault(mm, vma, address,
2568                                                 pte, pmd, write_access, entry);
2569                         }
2570                         return do_anonymous_page(mm, vma, address,
2571                                                  pte, pmd, write_access);
2572                 }
2573                 if (pte_file(entry))
2574                         return do_nonlinear_fault(mm, vma, address,
2575                                         pte, pmd, write_access, entry);
2576                 return do_swap_page(mm, vma, address,
2577                                         pte, pmd, write_access, entry);
2578         }
2579
2580         ptl = pte_lockptr(mm, pmd);
2581         spin_lock(ptl);
2582         if (unlikely(!pte_same(*pte, entry)))
2583                 goto unlock;
2584         if (write_access) {
2585                 if (!pte_write(entry))
2586                         return do_wp_page(mm, vma, address,
2587                                         pte, pmd, ptl, entry);
2588                 entry = pte_mkdirty(entry);
2589         }
2590         entry = pte_mkyoung(entry);
2591         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2592                 update_mmu_cache(vma, address, entry);
2593         } else {
2594                 /*
2595                  * This is needed only for protection faults but the arch code
2596                  * is not yet telling us if this is a protection fault or not.
2597                  * This still avoids useless tlb flushes for .text page faults
2598                  * with threads.
2599                  */
2600                 if (write_access)
2601                         flush_tlb_page(vma, address);
2602         }
2603 unlock:
2604         pte_unmap_unlock(pte, ptl);
2605         return 0;
2606 }
2607
2608 /*
2609  * By the time we get here, we already hold the mm semaphore
2610  */
2611 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2612                 unsigned long address, int write_access)
2613 {
2614         pgd_t *pgd;
2615         pud_t *pud;
2616         pmd_t *pmd;
2617         pte_t *pte;
2618
2619         __set_current_state(TASK_RUNNING);
2620
2621         count_vm_event(PGFAULT);
2622
2623         if (unlikely(is_vm_hugetlb_page(vma)))
2624                 return hugetlb_fault(mm, vma, address, write_access);
2625
2626         pgd = pgd_offset(mm, address);
2627         pud = pud_alloc(mm, pgd, address);
2628         if (!pud)
2629                 return VM_FAULT_OOM;
2630         pmd = pmd_alloc(mm, pud, address);
2631         if (!pmd)
2632                 return VM_FAULT_OOM;
2633         pte = pte_alloc_map(mm, pmd, address);
2634         if (!pte)
2635                 return VM_FAULT_OOM;
2636
2637         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2638 }
2639
2640 #ifndef __PAGETABLE_PUD_FOLDED
2641 /*
2642  * Allocate page upper directory.
2643  * We've already handled the fast-path in-line.
2644  */
2645 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2646 {
2647         pud_t *new = pud_alloc_one(mm, address);
2648         if (!new)
2649                 return -ENOMEM;
2650
2651         smp_wmb(); /* See comment in __pte_alloc */
2652
2653         spin_lock(&mm->page_table_lock);
2654         if (pgd_present(*pgd))          /* Another has populated it */
2655                 pud_free(mm, new);
2656         else
2657                 pgd_populate(mm, pgd, new);
2658         spin_unlock(&mm->page_table_lock);
2659         return 0;
2660 }
2661 #endif /* __PAGETABLE_PUD_FOLDED */
2662
2663 #ifndef __PAGETABLE_PMD_FOLDED
2664 /*
2665  * Allocate page middle directory.
2666  * We've already handled the fast-path in-line.
2667  */
2668 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2669 {
2670         pmd_t *new = pmd_alloc_one(mm, address);
2671         if (!new)
2672                 return -ENOMEM;
2673
2674         smp_wmb(); /* See comment in __pte_alloc */
2675
2676         spin_lock(&mm->page_table_lock);
2677 #ifndef __ARCH_HAS_4LEVEL_HACK
2678         if (pud_present(*pud))          /* Another has populated it */
2679                 pmd_free(mm, new);
2680         else
2681                 pud_populate(mm, pud, new);
2682 #else
2683         if (pgd_present(*pud))          /* Another has populated it */
2684                 pmd_free(mm, new);
2685         else
2686                 pgd_populate(mm, pud, new);
2687 #endif /* __ARCH_HAS_4LEVEL_HACK */
2688         spin_unlock(&mm->page_table_lock);
2689         return 0;
2690 }
2691 #endif /* __PAGETABLE_PMD_FOLDED */
2692
2693 int make_pages_present(unsigned long addr, unsigned long end)
2694 {
2695         int ret, len, write;
2696         struct vm_area_struct * vma;
2697
2698         vma = find_vma(current->mm, addr);
2699         if (!vma)
2700                 return -1;
2701         write = (vma->vm_flags & VM_WRITE) != 0;
2702         BUG_ON(addr >= end);
2703         BUG_ON(end > vma->vm_end);
2704         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2705         ret = get_user_pages(current, current->mm, addr,
2706                         len, write, 0, NULL, NULL);
2707         if (ret < 0)
2708                 return ret;
2709         return ret == len ? 0 : -1;
2710 }
2711
2712 #if !defined(__HAVE_ARCH_GATE_AREA)
2713
2714 #if defined(AT_SYSINFO_EHDR)
2715 static struct vm_area_struct gate_vma;
2716
2717 static int __init gate_vma_init(void)
2718 {
2719         gate_vma.vm_mm = NULL;
2720         gate_vma.vm_start = FIXADDR_USER_START;
2721         gate_vma.vm_end = FIXADDR_USER_END;
2722         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2723         gate_vma.vm_page_prot = __P101;
2724         /*
2725          * Make sure the vDSO gets into every core dump.
2726          * Dumping its contents makes post-mortem fully interpretable later
2727          * without matching up the same kernel and hardware config to see
2728          * what PC values meant.
2729          */
2730         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2731         return 0;
2732 }
2733 __initcall(gate_vma_init);
2734 #endif
2735
2736 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2737 {
2738 #ifdef AT_SYSINFO_EHDR
2739         return &gate_vma;
2740 #else
2741         return NULL;
2742 #endif
2743 }
2744
2745 int in_gate_area_no_task(unsigned long addr)
2746 {
2747 #ifdef AT_SYSINFO_EHDR
2748         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2749                 return 1;
2750 #endif
2751         return 0;
2752 }
2753
2754 #endif  /* __HAVE_ARCH_GATE_AREA */
2755
2756 #ifdef CONFIG_HAVE_IOREMAP_PROT
2757 static resource_size_t follow_phys(struct vm_area_struct *vma,
2758                         unsigned long address, unsigned int flags,
2759                         unsigned long *prot)
2760 {
2761         pgd_t *pgd;
2762         pud_t *pud;
2763         pmd_t *pmd;
2764         pte_t *ptep, pte;
2765         spinlock_t *ptl;
2766         resource_size_t phys_addr = 0;
2767         struct mm_struct *mm = vma->vm_mm;
2768
2769         VM_BUG_ON(!(vma->vm_flags & (VM_IO | VM_PFNMAP)));
2770
2771         pgd = pgd_offset(mm, address);
2772         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
2773                 goto no_page_table;
2774
2775         pud = pud_offset(pgd, address);
2776         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
2777                 goto no_page_table;
2778
2779         pmd = pmd_offset(pud, address);
2780         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
2781                 goto no_page_table;
2782
2783         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
2784         if (pmd_huge(*pmd))
2785                 goto no_page_table;
2786
2787         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
2788         if (!ptep)
2789                 goto out;
2790
2791         pte = *ptep;
2792         if (!pte_present(pte))
2793                 goto unlock;
2794         if ((flags & FOLL_WRITE) && !pte_write(pte))
2795                 goto unlock;
2796         phys_addr = pte_pfn(pte);
2797         phys_addr <<= PAGE_SHIFT; /* Shift here to avoid overflow on PAE */
2798
2799         *prot = pgprot_val(pte_pgprot(pte));
2800
2801 unlock:
2802         pte_unmap_unlock(ptep, ptl);
2803 out:
2804         return phys_addr;
2805 no_page_table:
2806         return 0;
2807 }
2808
2809 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
2810                         void *buf, int len, int write)
2811 {
2812         resource_size_t phys_addr;
2813         unsigned long prot = 0;
2814         void *maddr;
2815         int offset = addr & (PAGE_SIZE-1);
2816
2817         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
2818                 return -EINVAL;
2819
2820         phys_addr = follow_phys(vma, addr, write, &prot);
2821
2822         if (!phys_addr)
2823                 return -EINVAL;
2824
2825         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
2826         if (write)
2827                 memcpy_toio(maddr + offset, buf, len);
2828         else
2829                 memcpy_fromio(buf, maddr + offset, len);
2830         iounmap(maddr);
2831
2832         return len;
2833 }
2834 #endif
2835
2836 /*
2837  * Access another process' address space.
2838  * Source/target buffer must be kernel space,
2839  * Do not walk the page table directly, use get_user_pages
2840  */
2841 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2842 {
2843         struct mm_struct *mm;
2844         struct vm_area_struct *vma;
2845         void *old_buf = buf;
2846
2847         mm = get_task_mm(tsk);
2848         if (!mm)
2849                 return 0;
2850
2851         down_read(&mm->mmap_sem);
2852         /* ignore errors, just check how much was successfully transferred */
2853         while (len) {
2854                 int bytes, ret, offset;
2855                 void *maddr;
2856                 struct page *page = NULL;
2857
2858                 ret = get_user_pages(tsk, mm, addr, 1,
2859                                 write, 1, &page, &vma);
2860                 if (ret <= 0) {
2861                         /*
2862                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
2863                          * we can access using slightly different code.
2864                          */
2865 #ifdef CONFIG_HAVE_IOREMAP_PROT
2866                         vma = find_vma(mm, addr);
2867                         if (!vma)
2868                                 break;
2869                         if (vma->vm_ops && vma->vm_ops->access)
2870                                 ret = vma->vm_ops->access(vma, addr, buf,
2871                                                           len, write);
2872                         if (ret <= 0)
2873 #endif
2874                                 break;
2875                         bytes = ret;
2876                 } else {
2877                         bytes = len;
2878                         offset = addr & (PAGE_SIZE-1);
2879                         if (bytes > PAGE_SIZE-offset)
2880                                 bytes = PAGE_SIZE-offset;
2881
2882                         maddr = kmap(page);
2883                         if (write) {
2884                                 copy_to_user_page(vma, page, addr,
2885                                                   maddr + offset, buf, bytes);
2886                                 set_page_dirty_lock(page);
2887                         } else {
2888                                 copy_from_user_page(vma, page, addr,
2889                                                     buf, maddr + offset, bytes);
2890                         }
2891                         kunmap(page);
2892                         page_cache_release(page);
2893                 }
2894                 len -= bytes;
2895                 buf += bytes;
2896                 addr += bytes;
2897         }
2898         up_read(&mm->mmap_sem);
2899         mmput(mm);
2900
2901         return buf - old_buf;
2902 }
2903
2904 /*
2905  * Print the name of a VMA.
2906  */
2907 void print_vma_addr(char *prefix, unsigned long ip)
2908 {
2909         struct mm_struct *mm = current->mm;
2910         struct vm_area_struct *vma;
2911
2912         /*
2913          * Do not print if we are in atomic
2914          * contexts (in exception stacks, etc.):
2915          */
2916         if (preempt_count())
2917                 return;
2918
2919         down_read(&mm->mmap_sem);
2920         vma = find_vma(mm, ip);
2921         if (vma && vma->vm_file) {
2922                 struct file *f = vma->vm_file;
2923                 char *buf = (char *)__get_free_page(GFP_KERNEL);
2924                 if (buf) {
2925                         char *p, *s;
2926
2927                         p = d_path(&f->f_path, buf, PAGE_SIZE);
2928                         if (IS_ERR(p))
2929                                 p = "?";
2930                         s = strrchr(p, '/');
2931                         if (s)
2932                                 p = s+1;
2933                         printk("%s%s[%lx+%lx]", prefix, p,
2934                                         vma->vm_start,
2935                                         vma->vm_end - vma->vm_start);
2936                         free_page((unsigned long)buf);
2937                 }
2938         }
2939         up_read(&current->mm->mmap_sem);
2940 }