mm: remove duplicated #include's
[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 #include <linux/mmu_notifier.h>
55
56 #include <asm/pgalloc.h>
57 #include <asm/uaccess.h>
58 #include <asm/tlb.h>
59 #include <asm/tlbflush.h>
60 #include <asm/pgtable.h>
61
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64
65 #include "internal.h"
66
67 #ifndef CONFIG_NEED_MULTIPLE_NODES
68 /* use the per-pgdat data instead for discontigmem - mbligh */
69 unsigned long max_mapnr;
70 struct page *mem_map;
71
72 EXPORT_SYMBOL(max_mapnr);
73 EXPORT_SYMBOL(mem_map);
74 #endif
75
76 unsigned long num_physpages;
77 /*
78  * A number of key systems in x86 including ioremap() rely on the assumption
79  * that high_memory defines the upper bound on direct map memory, then end
80  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
81  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
82  * and ZONE_HIGHMEM.
83  */
84 void * high_memory;
85
86 EXPORT_SYMBOL(num_physpages);
87 EXPORT_SYMBOL(high_memory);
88
89 /*
90  * Randomize the address space (stacks, mmaps, brk, etc.).
91  *
92  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
93  *   as ancient (libc5 based) binaries can segfault. )
94  */
95 int randomize_va_space __read_mostly =
96 #ifdef CONFIG_COMPAT_BRK
97                                         1;
98 #else
99                                         2;
100 #endif
101
102 static int __init disable_randmaps(char *s)
103 {
104         randomize_va_space = 0;
105         return 1;
106 }
107 __setup("norandmaps", disable_randmaps);
108
109
110 /*
111  * If a p?d_bad entry is found while walking page tables, report
112  * the error, before resetting entry to p?d_none.  Usually (but
113  * very seldom) called out from the p?d_none_or_clear_bad macros.
114  */
115
116 void pgd_clear_bad(pgd_t *pgd)
117 {
118         pgd_ERROR(*pgd);
119         pgd_clear(pgd);
120 }
121
122 void pud_clear_bad(pud_t *pud)
123 {
124         pud_ERROR(*pud);
125         pud_clear(pud);
126 }
127
128 void pmd_clear_bad(pmd_t *pmd)
129 {
130         pmd_ERROR(*pmd);
131         pmd_clear(pmd);
132 }
133
134 /*
135  * Note: this doesn't free the actual pages themselves. That
136  * has been handled earlier when unmapping all the memory regions.
137  */
138 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
139 {
140         pgtable_t token = pmd_pgtable(*pmd);
141         pmd_clear(pmd);
142         pte_free_tlb(tlb, token);
143         tlb->mm->nr_ptes--;
144 }
145
146 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
147                                 unsigned long addr, unsigned long end,
148                                 unsigned long floor, unsigned long ceiling)
149 {
150         pmd_t *pmd;
151         unsigned long next;
152         unsigned long start;
153
154         start = addr;
155         pmd = pmd_offset(pud, addr);
156         do {
157                 next = pmd_addr_end(addr, end);
158                 if (pmd_none_or_clear_bad(pmd))
159                         continue;
160                 free_pte_range(tlb, pmd);
161         } while (pmd++, addr = next, addr != end);
162
163         start &= PUD_MASK;
164         if (start < floor)
165                 return;
166         if (ceiling) {
167                 ceiling &= PUD_MASK;
168                 if (!ceiling)
169                         return;
170         }
171         if (end - 1 > ceiling - 1)
172                 return;
173
174         pmd = pmd_offset(pud, start);
175         pud_clear(pud);
176         pmd_free_tlb(tlb, pmd);
177 }
178
179 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
180                                 unsigned long addr, unsigned long end,
181                                 unsigned long floor, unsigned long ceiling)
182 {
183         pud_t *pud;
184         unsigned long next;
185         unsigned long start;
186
187         start = addr;
188         pud = pud_offset(pgd, addr);
189         do {
190                 next = pud_addr_end(addr, end);
191                 if (pud_none_or_clear_bad(pud))
192                         continue;
193                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
194         } while (pud++, addr = next, addr != end);
195
196         start &= PGDIR_MASK;
197         if (start < floor)
198                 return;
199         if (ceiling) {
200                 ceiling &= PGDIR_MASK;
201                 if (!ceiling)
202                         return;
203         }
204         if (end - 1 > ceiling - 1)
205                 return;
206
207         pud = pud_offset(pgd, start);
208         pgd_clear(pgd);
209         pud_free_tlb(tlb, pud);
210 }
211
212 /*
213  * This function frees user-level page tables of a process.
214  *
215  * Must be called with pagetable lock held.
216  */
217 void free_pgd_range(struct mmu_gather *tlb,
218                         unsigned long addr, unsigned long end,
219                         unsigned long floor, unsigned long ceiling)
220 {
221         pgd_t *pgd;
222         unsigned long next;
223         unsigned long start;
224
225         /*
226          * The next few lines have given us lots of grief...
227          *
228          * Why are we testing PMD* at this top level?  Because often
229          * there will be no work to do at all, and we'd prefer not to
230          * go all the way down to the bottom just to discover that.
231          *
232          * Why all these "- 1"s?  Because 0 represents both the bottom
233          * of the address space and the top of it (using -1 for the
234          * top wouldn't help much: the masks would do the wrong thing).
235          * The rule is that addr 0 and floor 0 refer to the bottom of
236          * the address space, but end 0 and ceiling 0 refer to the top
237          * Comparisons need to use "end - 1" and "ceiling - 1" (though
238          * that end 0 case should be mythical).
239          *
240          * Wherever addr is brought up or ceiling brought down, we must
241          * be careful to reject "the opposite 0" before it confuses the
242          * subsequent tests.  But what about where end is brought down
243          * by PMD_SIZE below? no, end can't go down to 0 there.
244          *
245          * Whereas we round start (addr) and ceiling down, by different
246          * masks at different levels, in order to test whether a table
247          * now has no other vmas using it, so can be freed, we don't
248          * bother to round floor or end up - the tests don't need that.
249          */
250
251         addr &= PMD_MASK;
252         if (addr < floor) {
253                 addr += PMD_SIZE;
254                 if (!addr)
255                         return;
256         }
257         if (ceiling) {
258                 ceiling &= PMD_MASK;
259                 if (!ceiling)
260                         return;
261         }
262         if (end - 1 > ceiling - 1)
263                 end -= PMD_SIZE;
264         if (addr > end - 1)
265                 return;
266
267         start = addr;
268         pgd = pgd_offset(tlb->mm, addr);
269         do {
270                 next = pgd_addr_end(addr, end);
271                 if (pgd_none_or_clear_bad(pgd))
272                         continue;
273                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
274         } while (pgd++, addr = next, addr != end);
275 }
276
277 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
278                 unsigned long floor, unsigned long ceiling)
279 {
280         while (vma) {
281                 struct vm_area_struct *next = vma->vm_next;
282                 unsigned long addr = vma->vm_start;
283
284                 /*
285                  * Hide vma from rmap and vmtruncate before freeing pgtables
286                  */
287                 anon_vma_unlink(vma);
288                 unlink_file_vma(vma);
289
290                 if (is_vm_hugetlb_page(vma)) {
291                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
292                                 floor, next? next->vm_start: ceiling);
293                 } else {
294                         /*
295                          * Optimization: gather nearby vmas into one call down
296                          */
297                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
298                                && !is_vm_hugetlb_page(next)) {
299                                 vma = next;
300                                 next = vma->vm_next;
301                                 anon_vma_unlink(vma);
302                                 unlink_file_vma(vma);
303                         }
304                         free_pgd_range(tlb, addr, vma->vm_end,
305                                 floor, next? next->vm_start: ceiling);
306                 }
307                 vma = next;
308         }
309 }
310
311 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
312 {
313         pgtable_t new = pte_alloc_one(mm, address);
314         if (!new)
315                 return -ENOMEM;
316
317         /*
318          * Ensure all pte setup (eg. pte page lock and page clearing) are
319          * visible before the pte is made visible to other CPUs by being
320          * put into page tables.
321          *
322          * The other side of the story is the pointer chasing in the page
323          * table walking code (when walking the page table without locking;
324          * ie. most of the time). Fortunately, these data accesses consist
325          * of a chain of data-dependent loads, meaning most CPUs (alpha
326          * being the notable exception) will already guarantee loads are
327          * seen in-order. See the alpha page table accessors for the
328          * smp_read_barrier_depends() barriers in page table walking code.
329          */
330         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
331
332         spin_lock(&mm->page_table_lock);
333         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
334                 mm->nr_ptes++;
335                 pmd_populate(mm, pmd, new);
336                 new = NULL;
337         }
338         spin_unlock(&mm->page_table_lock);
339         if (new)
340                 pte_free(mm, new);
341         return 0;
342 }
343
344 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
345 {
346         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
347         if (!new)
348                 return -ENOMEM;
349
350         smp_wmb(); /* See comment in __pte_alloc */
351
352         spin_lock(&init_mm.page_table_lock);
353         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
354                 pmd_populate_kernel(&init_mm, pmd, new);
355                 new = NULL;
356         }
357         spin_unlock(&init_mm.page_table_lock);
358         if (new)
359                 pte_free_kernel(&init_mm, new);
360         return 0;
361 }
362
363 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
364 {
365         if (file_rss)
366                 add_mm_counter(mm, file_rss, file_rss);
367         if (anon_rss)
368                 add_mm_counter(mm, anon_rss, anon_rss);
369 }
370
371 /*
372  * This function is called to print an error when a bad pte
373  * is found. For example, we might have a PFN-mapped pte in
374  * a region that doesn't allow it.
375  *
376  * The calling function must still handle the error.
377  */
378 static void print_bad_pte(struct vm_area_struct *vma, pte_t pte,
379                           unsigned long vaddr)
380 {
381         printk(KERN_ERR "Bad pte = %08llx, process = %s, "
382                         "vm_flags = %lx, vaddr = %lx\n",
383                 (long long)pte_val(pte),
384                 (vma->vm_mm == current->mm ? current->comm : "???"),
385                 vma->vm_flags, vaddr);
386         dump_stack();
387 }
388
389 static inline int is_cow_mapping(unsigned int flags)
390 {
391         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
392 }
393
394 /*
395  * vm_normal_page -- This function gets the "struct page" associated with a pte.
396  *
397  * "Special" mappings do not wish to be associated with a "struct page" (either
398  * it doesn't exist, or it exists but they don't want to touch it). In this
399  * case, NULL is returned here. "Normal" mappings do have a struct page.
400  *
401  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
402  * pte bit, in which case this function is trivial. Secondly, an architecture
403  * may not have a spare pte bit, which requires a more complicated scheme,
404  * described below.
405  *
406  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
407  * special mapping (even if there are underlying and valid "struct pages").
408  * COWed pages of a VM_PFNMAP are always normal.
409  *
410  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
411  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
412  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
413  * mapping will always honor the rule
414  *
415  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
416  *
417  * And for normal mappings this is false.
418  *
419  * This restricts such mappings to be a linear translation from virtual address
420  * to pfn. To get around this restriction, we allow arbitrary mappings so long
421  * as the vma is not a COW mapping; in that case, we know that all ptes are
422  * special (because none can have been COWed).
423  *
424  *
425  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
426  *
427  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
428  * page" backing, however the difference is that _all_ pages with a struct
429  * page (that is, those where pfn_valid is true) are refcounted and considered
430  * normal pages by the VM. The disadvantage is that pages are refcounted
431  * (which can be slower and simply not an option for some PFNMAP users). The
432  * advantage is that we don't have to follow the strict linearity rule of
433  * PFNMAP mappings in order to support COWable mappings.
434  *
435  */
436 #ifdef __HAVE_ARCH_PTE_SPECIAL
437 # define HAVE_PTE_SPECIAL 1
438 #else
439 # define HAVE_PTE_SPECIAL 0
440 #endif
441 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
442                                 pte_t pte)
443 {
444         unsigned long pfn;
445
446         if (HAVE_PTE_SPECIAL) {
447                 if (likely(!pte_special(pte))) {
448                         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
449                         return pte_page(pte);
450                 }
451                 VM_BUG_ON(!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)));
452                 return NULL;
453         }
454
455         /* !HAVE_PTE_SPECIAL case follows: */
456
457         pfn = pte_pfn(pte);
458
459         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
460                 if (vma->vm_flags & VM_MIXEDMAP) {
461                         if (!pfn_valid(pfn))
462                                 return NULL;
463                         goto out;
464                 } else {
465                         unsigned long off;
466                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
467                         if (pfn == vma->vm_pgoff + off)
468                                 return NULL;
469                         if (!is_cow_mapping(vma->vm_flags))
470                                 return NULL;
471                 }
472         }
473
474         VM_BUG_ON(!pfn_valid(pfn));
475
476         /*
477          * NOTE! We still have PageReserved() pages in the page tables.
478          *
479          * eg. VDSO mappings can cause them to exist.
480          */
481 out:
482         return pfn_to_page(pfn);
483 }
484
485 /*
486  * copy one vm_area from one task to the other. Assumes the page tables
487  * already present in the new task to be cleared in the whole range
488  * covered by this vma.
489  */
490
491 static inline void
492 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
493                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
494                 unsigned long addr, int *rss)
495 {
496         unsigned long vm_flags = vma->vm_flags;
497         pte_t pte = *src_pte;
498         struct page *page;
499
500         /* pte contains position in swap or file, so copy. */
501         if (unlikely(!pte_present(pte))) {
502                 if (!pte_file(pte)) {
503                         swp_entry_t entry = pte_to_swp_entry(pte);
504
505                         swap_duplicate(entry);
506                         /* make sure dst_mm is on swapoff's mmlist. */
507                         if (unlikely(list_empty(&dst_mm->mmlist))) {
508                                 spin_lock(&mmlist_lock);
509                                 if (list_empty(&dst_mm->mmlist))
510                                         list_add(&dst_mm->mmlist,
511                                                  &src_mm->mmlist);
512                                 spin_unlock(&mmlist_lock);
513                         }
514                         if (is_write_migration_entry(entry) &&
515                                         is_cow_mapping(vm_flags)) {
516                                 /*
517                                  * COW mappings require pages in both parent
518                                  * and child to be set to read.
519                                  */
520                                 make_migration_entry_read(&entry);
521                                 pte = swp_entry_to_pte(entry);
522                                 set_pte_at(src_mm, addr, src_pte, pte);
523                         }
524                 }
525                 goto out_set_pte;
526         }
527
528         /*
529          * If it's a COW mapping, write protect it both
530          * in the parent and the child
531          */
532         if (is_cow_mapping(vm_flags)) {
533                 ptep_set_wrprotect(src_mm, addr, src_pte);
534                 pte = pte_wrprotect(pte);
535         }
536
537         /*
538          * If it's a shared mapping, mark it clean in
539          * the child
540          */
541         if (vm_flags & VM_SHARED)
542                 pte = pte_mkclean(pte);
543         pte = pte_mkold(pte);
544
545         page = vm_normal_page(vma, addr, pte);
546         if (page) {
547                 get_page(page);
548                 page_dup_rmap(page, vma, addr);
549                 rss[!!PageAnon(page)]++;
550         }
551
552 out_set_pte:
553         set_pte_at(dst_mm, addr, dst_pte, pte);
554 }
555
556 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
557                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
558                 unsigned long addr, unsigned long end)
559 {
560         pte_t *src_pte, *dst_pte;
561         spinlock_t *src_ptl, *dst_ptl;
562         int progress = 0;
563         int rss[2];
564
565 again:
566         rss[1] = rss[0] = 0;
567         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
568         if (!dst_pte)
569                 return -ENOMEM;
570         src_pte = pte_offset_map_nested(src_pmd, addr);
571         src_ptl = pte_lockptr(src_mm, src_pmd);
572         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
573         arch_enter_lazy_mmu_mode();
574
575         do {
576                 /*
577                  * We are holding two locks at this point - either of them
578                  * could generate latencies in another task on another CPU.
579                  */
580                 if (progress >= 32) {
581                         progress = 0;
582                         if (need_resched() ||
583                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
584                                 break;
585                 }
586                 if (pte_none(*src_pte)) {
587                         progress++;
588                         continue;
589                 }
590                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
591                 progress += 8;
592         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
593
594         arch_leave_lazy_mmu_mode();
595         spin_unlock(src_ptl);
596         pte_unmap_nested(src_pte - 1);
597         add_mm_rss(dst_mm, rss[0], rss[1]);
598         pte_unmap_unlock(dst_pte - 1, dst_ptl);
599         cond_resched();
600         if (addr != end)
601                 goto again;
602         return 0;
603 }
604
605 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
606                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
607                 unsigned long addr, unsigned long end)
608 {
609         pmd_t *src_pmd, *dst_pmd;
610         unsigned long next;
611
612         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
613         if (!dst_pmd)
614                 return -ENOMEM;
615         src_pmd = pmd_offset(src_pud, addr);
616         do {
617                 next = pmd_addr_end(addr, end);
618                 if (pmd_none_or_clear_bad(src_pmd))
619                         continue;
620                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
621                                                 vma, addr, next))
622                         return -ENOMEM;
623         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
624         return 0;
625 }
626
627 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
628                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
629                 unsigned long addr, unsigned long end)
630 {
631         pud_t *src_pud, *dst_pud;
632         unsigned long next;
633
634         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
635         if (!dst_pud)
636                 return -ENOMEM;
637         src_pud = pud_offset(src_pgd, addr);
638         do {
639                 next = pud_addr_end(addr, end);
640                 if (pud_none_or_clear_bad(src_pud))
641                         continue;
642                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
643                                                 vma, addr, next))
644                         return -ENOMEM;
645         } while (dst_pud++, src_pud++, addr = next, addr != end);
646         return 0;
647 }
648
649 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
650                 struct vm_area_struct *vma)
651 {
652         pgd_t *src_pgd, *dst_pgd;
653         unsigned long next;
654         unsigned long addr = vma->vm_start;
655         unsigned long end = vma->vm_end;
656         int ret;
657
658         /*
659          * Don't copy ptes where a page fault will fill them correctly.
660          * Fork becomes much lighter when there are big shared or private
661          * readonly mappings. The tradeoff is that copy_page_range is more
662          * efficient than faulting.
663          */
664         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
665                 if (!vma->anon_vma)
666                         return 0;
667         }
668
669         if (is_vm_hugetlb_page(vma))
670                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
671
672         /*
673          * We need to invalidate the secondary MMU mappings only when
674          * there could be a permission downgrade on the ptes of the
675          * parent mm. And a permission downgrade will only happen if
676          * is_cow_mapping() returns true.
677          */
678         if (is_cow_mapping(vma->vm_flags))
679                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
680
681         ret = 0;
682         dst_pgd = pgd_offset(dst_mm, addr);
683         src_pgd = pgd_offset(src_mm, addr);
684         do {
685                 next = pgd_addr_end(addr, end);
686                 if (pgd_none_or_clear_bad(src_pgd))
687                         continue;
688                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
689                                             vma, addr, next))) {
690                         ret = -ENOMEM;
691                         break;
692                 }
693         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
694
695         if (is_cow_mapping(vma->vm_flags))
696                 mmu_notifier_invalidate_range_end(src_mm,
697                                                   vma->vm_start, end);
698         return ret;
699 }
700
701 static unsigned long zap_pte_range(struct mmu_gather *tlb,
702                                 struct vm_area_struct *vma, pmd_t *pmd,
703                                 unsigned long addr, unsigned long end,
704                                 long *zap_work, struct zap_details *details)
705 {
706         struct mm_struct *mm = tlb->mm;
707         pte_t *pte;
708         spinlock_t *ptl;
709         int file_rss = 0;
710         int anon_rss = 0;
711
712         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
713         arch_enter_lazy_mmu_mode();
714         do {
715                 pte_t ptent = *pte;
716                 if (pte_none(ptent)) {
717                         (*zap_work)--;
718                         continue;
719                 }
720
721                 (*zap_work) -= PAGE_SIZE;
722
723                 if (pte_present(ptent)) {
724                         struct page *page;
725
726                         page = vm_normal_page(vma, addr, ptent);
727                         if (unlikely(details) && page) {
728                                 /*
729                                  * unmap_shared_mapping_pages() wants to
730                                  * invalidate cache without truncating:
731                                  * unmap shared but keep private pages.
732                                  */
733                                 if (details->check_mapping &&
734                                     details->check_mapping != page->mapping)
735                                         continue;
736                                 /*
737                                  * Each page->index must be checked when
738                                  * invalidating or truncating nonlinear.
739                                  */
740                                 if (details->nonlinear_vma &&
741                                     (page->index < details->first_index ||
742                                      page->index > details->last_index))
743                                         continue;
744                         }
745                         ptent = ptep_get_and_clear_full(mm, addr, pte,
746                                                         tlb->fullmm);
747                         tlb_remove_tlb_entry(tlb, pte, addr);
748                         if (unlikely(!page))
749                                 continue;
750                         if (unlikely(details) && details->nonlinear_vma
751                             && linear_page_index(details->nonlinear_vma,
752                                                 addr) != page->index)
753                                 set_pte_at(mm, addr, pte,
754                                            pgoff_to_pte(page->index));
755                         if (PageAnon(page))
756                                 anon_rss--;
757                         else {
758                                 if (pte_dirty(ptent))
759                                         set_page_dirty(page);
760                                 if (pte_young(ptent))
761                                         SetPageReferenced(page);
762                                 file_rss--;
763                         }
764                         page_remove_rmap(page, vma);
765                         tlb_remove_page(tlb, page);
766                         continue;
767                 }
768                 /*
769                  * If details->check_mapping, we leave swap entries;
770                  * if details->nonlinear_vma, we leave file entries.
771                  */
772                 if (unlikely(details))
773                         continue;
774                 if (!pte_file(ptent))
775                         free_swap_and_cache(pte_to_swp_entry(ptent));
776                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
777         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
778
779         add_mm_rss(mm, file_rss, anon_rss);
780         arch_leave_lazy_mmu_mode();
781         pte_unmap_unlock(pte - 1, ptl);
782
783         return addr;
784 }
785
786 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
787                                 struct vm_area_struct *vma, pud_t *pud,
788                                 unsigned long addr, unsigned long end,
789                                 long *zap_work, struct zap_details *details)
790 {
791         pmd_t *pmd;
792         unsigned long next;
793
794         pmd = pmd_offset(pud, addr);
795         do {
796                 next = pmd_addr_end(addr, end);
797                 if (pmd_none_or_clear_bad(pmd)) {
798                         (*zap_work)--;
799                         continue;
800                 }
801                 next = zap_pte_range(tlb, vma, pmd, addr, next,
802                                                 zap_work, details);
803         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
804
805         return addr;
806 }
807
808 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
809                                 struct vm_area_struct *vma, pgd_t *pgd,
810                                 unsigned long addr, unsigned long end,
811                                 long *zap_work, struct zap_details *details)
812 {
813         pud_t *pud;
814         unsigned long next;
815
816         pud = pud_offset(pgd, addr);
817         do {
818                 next = pud_addr_end(addr, end);
819                 if (pud_none_or_clear_bad(pud)) {
820                         (*zap_work)--;
821                         continue;
822                 }
823                 next = zap_pmd_range(tlb, vma, pud, addr, next,
824                                                 zap_work, details);
825         } while (pud++, addr = next, (addr != end && *zap_work > 0));
826
827         return addr;
828 }
829
830 static unsigned long unmap_page_range(struct mmu_gather *tlb,
831                                 struct vm_area_struct *vma,
832                                 unsigned long addr, unsigned long end,
833                                 long *zap_work, struct zap_details *details)
834 {
835         pgd_t *pgd;
836         unsigned long next;
837
838         if (details && !details->check_mapping && !details->nonlinear_vma)
839                 details = NULL;
840
841         BUG_ON(addr >= end);
842         tlb_start_vma(tlb, vma);
843         pgd = pgd_offset(vma->vm_mm, addr);
844         do {
845                 next = pgd_addr_end(addr, end);
846                 if (pgd_none_or_clear_bad(pgd)) {
847                         (*zap_work)--;
848                         continue;
849                 }
850                 next = zap_pud_range(tlb, vma, pgd, addr, next,
851                                                 zap_work, details);
852         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
853         tlb_end_vma(tlb, vma);
854
855         return addr;
856 }
857
858 #ifdef CONFIG_PREEMPT
859 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
860 #else
861 /* No preempt: go for improved straight-line efficiency */
862 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
863 #endif
864
865 /**
866  * unmap_vmas - unmap a range of memory covered by a list of vma's
867  * @tlbp: address of the caller's struct mmu_gather
868  * @vma: the starting vma
869  * @start_addr: virtual address at which to start unmapping
870  * @end_addr: virtual address at which to end unmapping
871  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
872  * @details: details of nonlinear truncation or shared cache invalidation
873  *
874  * Returns the end address of the unmapping (restart addr if interrupted).
875  *
876  * Unmap all pages in the vma list.
877  *
878  * We aim to not hold locks for too long (for scheduling latency reasons).
879  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
880  * return the ending mmu_gather to the caller.
881  *
882  * Only addresses between `start' and `end' will be unmapped.
883  *
884  * The VMA list must be sorted in ascending virtual address order.
885  *
886  * unmap_vmas() assumes that the caller will flush the whole unmapped address
887  * range after unmap_vmas() returns.  So the only responsibility here is to
888  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
889  * drops the lock and schedules.
890  */
891 unsigned long unmap_vmas(struct mmu_gather **tlbp,
892                 struct vm_area_struct *vma, unsigned long start_addr,
893                 unsigned long end_addr, unsigned long *nr_accounted,
894                 struct zap_details *details)
895 {
896         long zap_work = ZAP_BLOCK_SIZE;
897         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
898         int tlb_start_valid = 0;
899         unsigned long start = start_addr;
900         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
901         int fullmm = (*tlbp)->fullmm;
902         struct mm_struct *mm = vma->vm_mm;
903
904         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
905         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
906                 unsigned long end;
907
908                 start = max(vma->vm_start, start_addr);
909                 if (start >= vma->vm_end)
910                         continue;
911                 end = min(vma->vm_end, end_addr);
912                 if (end <= vma->vm_start)
913                         continue;
914
915                 if (vma->vm_flags & VM_ACCOUNT)
916                         *nr_accounted += (end - start) >> PAGE_SHIFT;
917
918                 while (start != end) {
919                         if (!tlb_start_valid) {
920                                 tlb_start = start;
921                                 tlb_start_valid = 1;
922                         }
923
924                         if (unlikely(is_vm_hugetlb_page(vma))) {
925                                 /*
926                                  * It is undesirable to test vma->vm_file as it
927                                  * should be non-null for valid hugetlb area.
928                                  * However, vm_file will be NULL in the error
929                                  * cleanup path of do_mmap_pgoff. When
930                                  * hugetlbfs ->mmap method fails,
931                                  * do_mmap_pgoff() nullifies vma->vm_file
932                                  * before calling this function to clean up.
933                                  * Since no pte has actually been setup, it is
934                                  * safe to do nothing in this case.
935                                  */
936                                 if (vma->vm_file) {
937                                         unmap_hugepage_range(vma, start, end, NULL);
938                                         zap_work -= (end - start) /
939                                         pages_per_huge_page(hstate_vma(vma));
940                                 }
941
942                                 start = end;
943                         } else
944                                 start = unmap_page_range(*tlbp, vma,
945                                                 start, end, &zap_work, details);
946
947                         if (zap_work > 0) {
948                                 BUG_ON(start != end);
949                                 break;
950                         }
951
952                         tlb_finish_mmu(*tlbp, tlb_start, start);
953
954                         if (need_resched() ||
955                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
956                                 if (i_mmap_lock) {
957                                         *tlbp = NULL;
958                                         goto out;
959                                 }
960                                 cond_resched();
961                         }
962
963                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
964                         tlb_start_valid = 0;
965                         zap_work = ZAP_BLOCK_SIZE;
966                 }
967         }
968 out:
969         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
970         return start;   /* which is now the end (or restart) address */
971 }
972
973 /**
974  * zap_page_range - remove user pages in a given range
975  * @vma: vm_area_struct holding the applicable pages
976  * @address: starting address of pages to zap
977  * @size: number of bytes to zap
978  * @details: details of nonlinear truncation or shared cache invalidation
979  */
980 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
981                 unsigned long size, struct zap_details *details)
982 {
983         struct mm_struct *mm = vma->vm_mm;
984         struct mmu_gather *tlb;
985         unsigned long end = address + size;
986         unsigned long nr_accounted = 0;
987
988         lru_add_drain();
989         tlb = tlb_gather_mmu(mm, 0);
990         update_hiwater_rss(mm);
991         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
992         if (tlb)
993                 tlb_finish_mmu(tlb, address, end);
994         return end;
995 }
996
997 /**
998  * zap_vma_ptes - remove ptes mapping the vma
999  * @vma: vm_area_struct holding ptes to be zapped
1000  * @address: starting address of pages to zap
1001  * @size: number of bytes to zap
1002  *
1003  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1004  *
1005  * The entire address range must be fully contained within the vma.
1006  *
1007  * Returns 0 if successful.
1008  */
1009 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1010                 unsigned long size)
1011 {
1012         if (address < vma->vm_start || address + size > vma->vm_end ||
1013                         !(vma->vm_flags & VM_PFNMAP))
1014                 return -1;
1015         zap_page_range(vma, address, size, NULL);
1016         return 0;
1017 }
1018 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1019
1020 /*
1021  * Do a quick page-table lookup for a single page.
1022  */
1023 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1024                         unsigned int flags)
1025 {
1026         pgd_t *pgd;
1027         pud_t *pud;
1028         pmd_t *pmd;
1029         pte_t *ptep, pte;
1030         spinlock_t *ptl;
1031         struct page *page;
1032         struct mm_struct *mm = vma->vm_mm;
1033
1034         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1035         if (!IS_ERR(page)) {
1036                 BUG_ON(flags & FOLL_GET);
1037                 goto out;
1038         }
1039
1040         page = NULL;
1041         pgd = pgd_offset(mm, address);
1042         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1043                 goto no_page_table;
1044
1045         pud = pud_offset(pgd, address);
1046         if (pud_none(*pud))
1047                 goto no_page_table;
1048         if (pud_huge(*pud)) {
1049                 BUG_ON(flags & FOLL_GET);
1050                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1051                 goto out;
1052         }
1053         if (unlikely(pud_bad(*pud)))
1054                 goto no_page_table;
1055
1056         pmd = pmd_offset(pud, address);
1057         if (pmd_none(*pmd))
1058                 goto no_page_table;
1059         if (pmd_huge(*pmd)) {
1060                 BUG_ON(flags & FOLL_GET);
1061                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1062                 goto out;
1063         }
1064         if (unlikely(pmd_bad(*pmd)))
1065                 goto no_page_table;
1066
1067         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1068
1069         pte = *ptep;
1070         if (!pte_present(pte))
1071                 goto no_page;
1072         if ((flags & FOLL_WRITE) && !pte_write(pte))
1073                 goto unlock;
1074         page = vm_normal_page(vma, address, pte);
1075         if (unlikely(!page))
1076                 goto bad_page;
1077
1078         if (flags & FOLL_GET)
1079                 get_page(page);
1080         if (flags & FOLL_TOUCH) {
1081                 if ((flags & FOLL_WRITE) &&
1082                     !pte_dirty(pte) && !PageDirty(page))
1083                         set_page_dirty(page);
1084                 mark_page_accessed(page);
1085         }
1086 unlock:
1087         pte_unmap_unlock(ptep, ptl);
1088 out:
1089         return page;
1090
1091 bad_page:
1092         pte_unmap_unlock(ptep, ptl);
1093         return ERR_PTR(-EFAULT);
1094
1095 no_page:
1096         pte_unmap_unlock(ptep, ptl);
1097         if (!pte_none(pte))
1098                 return page;
1099         /* Fall through to ZERO_PAGE handling */
1100 no_page_table:
1101         /*
1102          * When core dumping an enormous anonymous area that nobody
1103          * has touched so far, we don't want to allocate page tables.
1104          */
1105         if (flags & FOLL_ANON) {
1106                 page = ZERO_PAGE(0);
1107                 if (flags & FOLL_GET)
1108                         get_page(page);
1109                 BUG_ON(flags & FOLL_WRITE);
1110         }
1111         return page;
1112 }
1113
1114 /* Can we do the FOLL_ANON optimization? */
1115 static inline int use_zero_page(struct vm_area_struct *vma)
1116 {
1117         /*
1118          * We don't want to optimize FOLL_ANON for make_pages_present()
1119          * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1120          * we want to get the page from the page tables to make sure
1121          * that we serialize and update with any other user of that
1122          * mapping.
1123          */
1124         if (vma->vm_flags & (VM_LOCKED | VM_SHARED))
1125                 return 0;
1126         /*
1127          * And if we have a fault routine, it's not an anonymous region.
1128          */
1129         return !vma->vm_ops || !vma->vm_ops->fault;
1130 }
1131
1132
1133
1134 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1135                      unsigned long start, int len, int flags,
1136                 struct page **pages, struct vm_area_struct **vmas)
1137 {
1138         int i;
1139         unsigned int vm_flags = 0;
1140         int write = !!(flags & GUP_FLAGS_WRITE);
1141         int force = !!(flags & GUP_FLAGS_FORCE);
1142         int ignore = !!(flags & GUP_FLAGS_IGNORE_VMA_PERMISSIONS);
1143
1144         if (len <= 0)
1145                 return 0;
1146         /* 
1147          * Require read or write permissions.
1148          * If 'force' is set, we only require the "MAY" flags.
1149          */
1150         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1151         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1152         i = 0;
1153
1154         do {
1155                 struct vm_area_struct *vma;
1156                 unsigned int foll_flags;
1157
1158                 vma = find_extend_vma(mm, start);
1159                 if (!vma && in_gate_area(tsk, start)) {
1160                         unsigned long pg = start & PAGE_MASK;
1161                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1162                         pgd_t *pgd;
1163                         pud_t *pud;
1164                         pmd_t *pmd;
1165                         pte_t *pte;
1166
1167                         /* user gate pages are read-only */
1168                         if (!ignore && write)
1169                                 return i ? : -EFAULT;
1170                         if (pg > TASK_SIZE)
1171                                 pgd = pgd_offset_k(pg);
1172                         else
1173                                 pgd = pgd_offset_gate(mm, pg);
1174                         BUG_ON(pgd_none(*pgd));
1175                         pud = pud_offset(pgd, pg);
1176                         BUG_ON(pud_none(*pud));
1177                         pmd = pmd_offset(pud, pg);
1178                         if (pmd_none(*pmd))
1179                                 return i ? : -EFAULT;
1180                         pte = pte_offset_map(pmd, pg);
1181                         if (pte_none(*pte)) {
1182                                 pte_unmap(pte);
1183                                 return i ? : -EFAULT;
1184                         }
1185                         if (pages) {
1186                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1187                                 pages[i] = page;
1188                                 if (page)
1189                                         get_page(page);
1190                         }
1191                         pte_unmap(pte);
1192                         if (vmas)
1193                                 vmas[i] = gate_vma;
1194                         i++;
1195                         start += PAGE_SIZE;
1196                         len--;
1197                         continue;
1198                 }
1199
1200                 if (!vma ||
1201                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1202                     (!ignore && !(vm_flags & vma->vm_flags)))
1203                         return i ? : -EFAULT;
1204
1205                 if (is_vm_hugetlb_page(vma)) {
1206                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1207                                                 &start, &len, i, write);
1208                         continue;
1209                 }
1210
1211                 foll_flags = FOLL_TOUCH;
1212                 if (pages)
1213                         foll_flags |= FOLL_GET;
1214                 if (!write && use_zero_page(vma))
1215                         foll_flags |= FOLL_ANON;
1216
1217                 do {
1218                         struct page *page;
1219
1220                         /*
1221                          * If tsk is ooming, cut off its access to large memory
1222                          * allocations. It has a pending SIGKILL, but it can't
1223                          * be processed until returning to user space.
1224                          */
1225                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1226                                 return i ? i : -ENOMEM;
1227
1228                         if (write)
1229                                 foll_flags |= FOLL_WRITE;
1230
1231                         cond_resched();
1232                         while (!(page = follow_page(vma, start, foll_flags))) {
1233                                 int ret;
1234                                 ret = handle_mm_fault(mm, vma, start,
1235                                                 foll_flags & FOLL_WRITE);
1236                                 if (ret & VM_FAULT_ERROR) {
1237                                         if (ret & VM_FAULT_OOM)
1238                                                 return i ? i : -ENOMEM;
1239                                         else if (ret & VM_FAULT_SIGBUS)
1240                                                 return i ? i : -EFAULT;
1241                                         BUG();
1242                                 }
1243                                 if (ret & VM_FAULT_MAJOR)
1244                                         tsk->maj_flt++;
1245                                 else
1246                                         tsk->min_flt++;
1247
1248                                 /*
1249                                  * The VM_FAULT_WRITE bit tells us that
1250                                  * do_wp_page has broken COW when necessary,
1251                                  * even if maybe_mkwrite decided not to set
1252                                  * pte_write. We can thus safely do subsequent
1253                                  * page lookups as if they were reads.
1254                                  */
1255                                 if (ret & VM_FAULT_WRITE)
1256                                         foll_flags &= ~FOLL_WRITE;
1257
1258                                 cond_resched();
1259                         }
1260                         if (IS_ERR(page))
1261                                 return i ? i : PTR_ERR(page);
1262                         if (pages) {
1263                                 pages[i] = page;
1264
1265                                 flush_anon_page(vma, page, start);
1266                                 flush_dcache_page(page);
1267                         }
1268                         if (vmas)
1269                                 vmas[i] = vma;
1270                         i++;
1271                         start += PAGE_SIZE;
1272                         len--;
1273                 } while (len && start < vma->vm_end);
1274         } while (len);
1275         return i;
1276 }
1277
1278 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1279                 unsigned long start, int len, int write, int force,
1280                 struct page **pages, struct vm_area_struct **vmas)
1281 {
1282         int flags = 0;
1283
1284         if (write)
1285                 flags |= GUP_FLAGS_WRITE;
1286         if (force)
1287                 flags |= GUP_FLAGS_FORCE;
1288
1289         return __get_user_pages(tsk, mm,
1290                                 start, len, flags,
1291                                 pages, vmas);
1292 }
1293
1294 EXPORT_SYMBOL(get_user_pages);
1295
1296 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1297                         spinlock_t **ptl)
1298 {
1299         pgd_t * pgd = pgd_offset(mm, addr);
1300         pud_t * pud = pud_alloc(mm, pgd, addr);
1301         if (pud) {
1302                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1303                 if (pmd)
1304                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1305         }
1306         return NULL;
1307 }
1308
1309 /*
1310  * This is the old fallback for page remapping.
1311  *
1312  * For historical reasons, it only allows reserved pages. Only
1313  * old drivers should use this, and they needed to mark their
1314  * pages reserved for the old functions anyway.
1315  */
1316 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1317                         struct page *page, pgprot_t prot)
1318 {
1319         struct mm_struct *mm = vma->vm_mm;
1320         int retval;
1321         pte_t *pte;
1322         spinlock_t *ptl;
1323
1324         retval = -EINVAL;
1325         if (PageAnon(page))
1326                 goto out;
1327         retval = -ENOMEM;
1328         flush_dcache_page(page);
1329         pte = get_locked_pte(mm, addr, &ptl);
1330         if (!pte)
1331                 goto out;
1332         retval = -EBUSY;
1333         if (!pte_none(*pte))
1334                 goto out_unlock;
1335
1336         /* Ok, finally just insert the thing.. */
1337         get_page(page);
1338         inc_mm_counter(mm, file_rss);
1339         page_add_file_rmap(page);
1340         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1341
1342         retval = 0;
1343         pte_unmap_unlock(pte, ptl);
1344         return retval;
1345 out_unlock:
1346         pte_unmap_unlock(pte, ptl);
1347 out:
1348         return retval;
1349 }
1350
1351 /**
1352  * vm_insert_page - insert single page into user vma
1353  * @vma: user vma to map to
1354  * @addr: target user address of this page
1355  * @page: source kernel page
1356  *
1357  * This allows drivers to insert individual pages they've allocated
1358  * into a user vma.
1359  *
1360  * The page has to be a nice clean _individual_ kernel allocation.
1361  * If you allocate a compound page, you need to have marked it as
1362  * such (__GFP_COMP), or manually just split the page up yourself
1363  * (see split_page()).
1364  *
1365  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1366  * took an arbitrary page protection parameter. This doesn't allow
1367  * that. Your vma protection will have to be set up correctly, which
1368  * means that if you want a shared writable mapping, you'd better
1369  * ask for a shared writable mapping!
1370  *
1371  * The page does not need to be reserved.
1372  */
1373 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1374                         struct page *page)
1375 {
1376         if (addr < vma->vm_start || addr >= vma->vm_end)
1377                 return -EFAULT;
1378         if (!page_count(page))
1379                 return -EINVAL;
1380         vma->vm_flags |= VM_INSERTPAGE;
1381         return insert_page(vma, addr, page, vma->vm_page_prot);
1382 }
1383 EXPORT_SYMBOL(vm_insert_page);
1384
1385 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1386                         unsigned long pfn, pgprot_t prot)
1387 {
1388         struct mm_struct *mm = vma->vm_mm;
1389         int retval;
1390         pte_t *pte, entry;
1391         spinlock_t *ptl;
1392
1393         retval = -ENOMEM;
1394         pte = get_locked_pte(mm, addr, &ptl);
1395         if (!pte)
1396                 goto out;
1397         retval = -EBUSY;
1398         if (!pte_none(*pte))
1399                 goto out_unlock;
1400
1401         /* Ok, finally just insert the thing.. */
1402         entry = pte_mkspecial(pfn_pte(pfn, prot));
1403         set_pte_at(mm, addr, pte, entry);
1404         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1405
1406         retval = 0;
1407 out_unlock:
1408         pte_unmap_unlock(pte, ptl);
1409 out:
1410         return retval;
1411 }
1412
1413 /**
1414  * vm_insert_pfn - insert single pfn into user vma
1415  * @vma: user vma to map to
1416  * @addr: target user address of this page
1417  * @pfn: source kernel pfn
1418  *
1419  * Similar to vm_inert_page, this allows drivers to insert individual pages
1420  * they've allocated into a user vma. Same comments apply.
1421  *
1422  * This function should only be called from a vm_ops->fault handler, and
1423  * in that case the handler should return NULL.
1424  *
1425  * vma cannot be a COW mapping.
1426  *
1427  * As this is called only for pages that do not currently exist, we
1428  * do not need to flush old virtual caches or the TLB.
1429  */
1430 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1431                         unsigned long pfn)
1432 {
1433         /*
1434          * Technically, architectures with pte_special can avoid all these
1435          * restrictions (same for remap_pfn_range).  However we would like
1436          * consistency in testing and feature parity among all, so we should
1437          * try to keep these invariants in place for everybody.
1438          */
1439         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1440         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1441                                                 (VM_PFNMAP|VM_MIXEDMAP));
1442         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1443         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1444
1445         if (addr < vma->vm_start || addr >= vma->vm_end)
1446                 return -EFAULT;
1447         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1448 }
1449 EXPORT_SYMBOL(vm_insert_pfn);
1450
1451 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1452                         unsigned long pfn)
1453 {
1454         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1455
1456         if (addr < vma->vm_start || addr >= vma->vm_end)
1457                 return -EFAULT;
1458
1459         /*
1460          * If we don't have pte special, then we have to use the pfn_valid()
1461          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1462          * refcount the page if pfn_valid is true (hence insert_page rather
1463          * than insert_pfn).
1464          */
1465         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1466                 struct page *page;
1467
1468                 page = pfn_to_page(pfn);
1469                 return insert_page(vma, addr, page, vma->vm_page_prot);
1470         }
1471         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1472 }
1473 EXPORT_SYMBOL(vm_insert_mixed);
1474
1475 /*
1476  * maps a range of physical memory into the requested pages. the old
1477  * mappings are removed. any references to nonexistent pages results
1478  * in null mappings (currently treated as "copy-on-access")
1479  */
1480 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1481                         unsigned long addr, unsigned long end,
1482                         unsigned long pfn, pgprot_t prot)
1483 {
1484         pte_t *pte;
1485         spinlock_t *ptl;
1486
1487         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1488         if (!pte)
1489                 return -ENOMEM;
1490         arch_enter_lazy_mmu_mode();
1491         do {
1492                 BUG_ON(!pte_none(*pte));
1493                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1494                 pfn++;
1495         } while (pte++, addr += PAGE_SIZE, addr != end);
1496         arch_leave_lazy_mmu_mode();
1497         pte_unmap_unlock(pte - 1, ptl);
1498         return 0;
1499 }
1500
1501 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1502                         unsigned long addr, unsigned long end,
1503                         unsigned long pfn, pgprot_t prot)
1504 {
1505         pmd_t *pmd;
1506         unsigned long next;
1507
1508         pfn -= addr >> PAGE_SHIFT;
1509         pmd = pmd_alloc(mm, pud, addr);
1510         if (!pmd)
1511                 return -ENOMEM;
1512         do {
1513                 next = pmd_addr_end(addr, end);
1514                 if (remap_pte_range(mm, pmd, addr, next,
1515                                 pfn + (addr >> PAGE_SHIFT), prot))
1516                         return -ENOMEM;
1517         } while (pmd++, addr = next, addr != end);
1518         return 0;
1519 }
1520
1521 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1522                         unsigned long addr, unsigned long end,
1523                         unsigned long pfn, pgprot_t prot)
1524 {
1525         pud_t *pud;
1526         unsigned long next;
1527
1528         pfn -= addr >> PAGE_SHIFT;
1529         pud = pud_alloc(mm, pgd, addr);
1530         if (!pud)
1531                 return -ENOMEM;
1532         do {
1533                 next = pud_addr_end(addr, end);
1534                 if (remap_pmd_range(mm, pud, addr, next,
1535                                 pfn + (addr >> PAGE_SHIFT), prot))
1536                         return -ENOMEM;
1537         } while (pud++, addr = next, addr != end);
1538         return 0;
1539 }
1540
1541 /**
1542  * remap_pfn_range - remap kernel memory to userspace
1543  * @vma: user vma to map to
1544  * @addr: target user address to start at
1545  * @pfn: physical address of kernel memory
1546  * @size: size of map area
1547  * @prot: page protection flags for this mapping
1548  *
1549  *  Note: this is only safe if the mm semaphore is held when called.
1550  */
1551 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1552                     unsigned long pfn, unsigned long size, pgprot_t prot)
1553 {
1554         pgd_t *pgd;
1555         unsigned long next;
1556         unsigned long end = addr + PAGE_ALIGN(size);
1557         struct mm_struct *mm = vma->vm_mm;
1558         int err;
1559
1560         /*
1561          * Physically remapped pages are special. Tell the
1562          * rest of the world about it:
1563          *   VM_IO tells people not to look at these pages
1564          *      (accesses can have side effects).
1565          *   VM_RESERVED is specified all over the place, because
1566          *      in 2.4 it kept swapout's vma scan off this vma; but
1567          *      in 2.6 the LRU scan won't even find its pages, so this
1568          *      flag means no more than count its pages in reserved_vm,
1569          *      and omit it from core dump, even when VM_IO turned off.
1570          *   VM_PFNMAP tells the core MM that the base pages are just
1571          *      raw PFN mappings, and do not have a "struct page" associated
1572          *      with them.
1573          *
1574          * There's a horrible special case to handle copy-on-write
1575          * behaviour that some programs depend on. We mark the "original"
1576          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1577          */
1578         if (is_cow_mapping(vma->vm_flags)) {
1579                 if (addr != vma->vm_start || end != vma->vm_end)
1580                         return -EINVAL;
1581                 vma->vm_pgoff = pfn;
1582         }
1583
1584         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1585
1586         BUG_ON(addr >= end);
1587         pfn -= addr >> PAGE_SHIFT;
1588         pgd = pgd_offset(mm, addr);
1589         flush_cache_range(vma, addr, end);
1590         do {
1591                 next = pgd_addr_end(addr, end);
1592                 err = remap_pud_range(mm, pgd, addr, next,
1593                                 pfn + (addr >> PAGE_SHIFT), prot);
1594                 if (err)
1595                         break;
1596         } while (pgd++, addr = next, addr != end);
1597         return err;
1598 }
1599 EXPORT_SYMBOL(remap_pfn_range);
1600
1601 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1602                                      unsigned long addr, unsigned long end,
1603                                      pte_fn_t fn, void *data)
1604 {
1605         pte_t *pte;
1606         int err;
1607         pgtable_t token;
1608         spinlock_t *uninitialized_var(ptl);
1609
1610         pte = (mm == &init_mm) ?
1611                 pte_alloc_kernel(pmd, addr) :
1612                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1613         if (!pte)
1614                 return -ENOMEM;
1615
1616         BUG_ON(pmd_huge(*pmd));
1617
1618         token = pmd_pgtable(*pmd);
1619
1620         do {
1621                 err = fn(pte, token, addr, data);
1622                 if (err)
1623                         break;
1624         } while (pte++, addr += PAGE_SIZE, addr != end);
1625
1626         if (mm != &init_mm)
1627                 pte_unmap_unlock(pte-1, ptl);
1628         return err;
1629 }
1630
1631 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1632                                      unsigned long addr, unsigned long end,
1633                                      pte_fn_t fn, void *data)
1634 {
1635         pmd_t *pmd;
1636         unsigned long next;
1637         int err;
1638
1639         BUG_ON(pud_huge(*pud));
1640
1641         pmd = pmd_alloc(mm, pud, addr);
1642         if (!pmd)
1643                 return -ENOMEM;
1644         do {
1645                 next = pmd_addr_end(addr, end);
1646                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1647                 if (err)
1648                         break;
1649         } while (pmd++, addr = next, addr != end);
1650         return err;
1651 }
1652
1653 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1654                                      unsigned long addr, unsigned long end,
1655                                      pte_fn_t fn, void *data)
1656 {
1657         pud_t *pud;
1658         unsigned long next;
1659         int err;
1660
1661         pud = pud_alloc(mm, pgd, addr);
1662         if (!pud)
1663                 return -ENOMEM;
1664         do {
1665                 next = pud_addr_end(addr, end);
1666                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1667                 if (err)
1668                         break;
1669         } while (pud++, addr = next, addr != end);
1670         return err;
1671 }
1672
1673 /*
1674  * Scan a region of virtual memory, filling in page tables as necessary
1675  * and calling a provided function on each leaf page table.
1676  */
1677 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1678                         unsigned long size, pte_fn_t fn, void *data)
1679 {
1680         pgd_t *pgd;
1681         unsigned long next;
1682         unsigned long start = addr, end = addr + size;
1683         int err;
1684
1685         BUG_ON(addr >= end);
1686         mmu_notifier_invalidate_range_start(mm, start, end);
1687         pgd = pgd_offset(mm, addr);
1688         do {
1689                 next = pgd_addr_end(addr, end);
1690                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1691                 if (err)
1692                         break;
1693         } while (pgd++, addr = next, addr != end);
1694         mmu_notifier_invalidate_range_end(mm, start, end);
1695         return err;
1696 }
1697 EXPORT_SYMBOL_GPL(apply_to_page_range);
1698
1699 /*
1700  * handle_pte_fault chooses page fault handler according to an entry
1701  * which was read non-atomically.  Before making any commitment, on
1702  * those architectures or configurations (e.g. i386 with PAE) which
1703  * might give a mix of unmatched parts, do_swap_page and do_file_page
1704  * must check under lock before unmapping the pte and proceeding
1705  * (but do_wp_page is only called after already making such a check;
1706  * and do_anonymous_page and do_no_page can safely check later on).
1707  */
1708 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1709                                 pte_t *page_table, pte_t orig_pte)
1710 {
1711         int same = 1;
1712 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1713         if (sizeof(pte_t) > sizeof(unsigned long)) {
1714                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1715                 spin_lock(ptl);
1716                 same = pte_same(*page_table, orig_pte);
1717                 spin_unlock(ptl);
1718         }
1719 #endif
1720         pte_unmap(page_table);
1721         return same;
1722 }
1723
1724 /*
1725  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1726  * servicing faults for write access.  In the normal case, do always want
1727  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1728  * that do not have writing enabled, when used by access_process_vm.
1729  */
1730 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1731 {
1732         if (likely(vma->vm_flags & VM_WRITE))
1733                 pte = pte_mkwrite(pte);
1734         return pte;
1735 }
1736
1737 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1738 {
1739         /*
1740          * If the source page was a PFN mapping, we don't have
1741          * a "struct page" for it. We do a best-effort copy by
1742          * just copying from the original user address. If that
1743          * fails, we just zero-fill it. Live with it.
1744          */
1745         if (unlikely(!src)) {
1746                 void *kaddr = kmap_atomic(dst, KM_USER0);
1747                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1748
1749                 /*
1750                  * This really shouldn't fail, because the page is there
1751                  * in the page tables. But it might just be unreadable,
1752                  * in which case we just give up and fill the result with
1753                  * zeroes.
1754                  */
1755                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1756                         memset(kaddr, 0, PAGE_SIZE);
1757                 kunmap_atomic(kaddr, KM_USER0);
1758                 flush_dcache_page(dst);
1759         } else
1760                 copy_user_highpage(dst, src, va, vma);
1761 }
1762
1763 /*
1764  * This routine handles present pages, when users try to write
1765  * to a shared page. It is done by copying the page to a new address
1766  * and decrementing the shared-page counter for the old page.
1767  *
1768  * Note that this routine assumes that the protection checks have been
1769  * done by the caller (the low-level page fault routine in most cases).
1770  * Thus we can safely just mark it writable once we've done any necessary
1771  * COW.
1772  *
1773  * We also mark the page dirty at this point even though the page will
1774  * change only once the write actually happens. This avoids a few races,
1775  * and potentially makes it more efficient.
1776  *
1777  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1778  * but allow concurrent faults), with pte both mapped and locked.
1779  * We return with mmap_sem still held, but pte unmapped and unlocked.
1780  */
1781 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1782                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1783                 spinlock_t *ptl, pte_t orig_pte)
1784 {
1785         struct page *old_page, *new_page;
1786         pte_t entry;
1787         int reuse = 0, ret = 0;
1788         int page_mkwrite = 0;
1789         struct page *dirty_page = NULL;
1790
1791         old_page = vm_normal_page(vma, address, orig_pte);
1792         if (!old_page) {
1793                 /*
1794                  * VM_MIXEDMAP !pfn_valid() case
1795                  *
1796                  * We should not cow pages in a shared writeable mapping.
1797                  * Just mark the pages writable as we can't do any dirty
1798                  * accounting on raw pfn maps.
1799                  */
1800                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1801                                      (VM_WRITE|VM_SHARED))
1802                         goto reuse;
1803                 goto gotten;
1804         }
1805
1806         /*
1807          * Take out anonymous pages first, anonymous shared vmas are
1808          * not dirty accountable.
1809          */
1810         if (PageAnon(old_page)) {
1811                 if (trylock_page(old_page)) {
1812                         reuse = can_share_swap_page(old_page);
1813                         unlock_page(old_page);
1814                 }
1815         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1816                                         (VM_WRITE|VM_SHARED))) {
1817                 /*
1818                  * Only catch write-faults on shared writable pages,
1819                  * read-only shared pages can get COWed by
1820                  * get_user_pages(.write=1, .force=1).
1821                  */
1822                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1823                         /*
1824                          * Notify the address space that the page is about to
1825                          * become writable so that it can prohibit this or wait
1826                          * for the page to get into an appropriate state.
1827                          *
1828                          * We do this without the lock held, so that it can
1829                          * sleep if it needs to.
1830                          */
1831                         page_cache_get(old_page);
1832                         pte_unmap_unlock(page_table, ptl);
1833
1834                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1835                                 goto unwritable_page;
1836
1837                         /*
1838                          * Since we dropped the lock we need to revalidate
1839                          * the PTE as someone else may have changed it.  If
1840                          * they did, we just return, as we can count on the
1841                          * MMU to tell us if they didn't also make it writable.
1842                          */
1843                         page_table = pte_offset_map_lock(mm, pmd, address,
1844                                                          &ptl);
1845                         page_cache_release(old_page);
1846                         if (!pte_same(*page_table, orig_pte))
1847                                 goto unlock;
1848
1849                         page_mkwrite = 1;
1850                 }
1851                 dirty_page = old_page;
1852                 get_page(dirty_page);
1853                 reuse = 1;
1854         }
1855
1856         if (reuse) {
1857 reuse:
1858                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1859                 entry = pte_mkyoung(orig_pte);
1860                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1861                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1862                         update_mmu_cache(vma, address, entry);
1863                 ret |= VM_FAULT_WRITE;
1864                 goto unlock;
1865         }
1866
1867         /*
1868          * Ok, we need to copy. Oh, well..
1869          */
1870         page_cache_get(old_page);
1871 gotten:
1872         pte_unmap_unlock(page_table, ptl);
1873
1874         if (unlikely(anon_vma_prepare(vma)))
1875                 goto oom;
1876         VM_BUG_ON(old_page == ZERO_PAGE(0));
1877         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1878         if (!new_page)
1879                 goto oom;
1880         /*
1881          * Don't let another task, with possibly unlocked vma,
1882          * keep the mlocked page.
1883          */
1884         if (vma->vm_flags & VM_LOCKED) {
1885                 lock_page(old_page);    /* for LRU manipulation */
1886                 clear_page_mlock(old_page);
1887                 unlock_page(old_page);
1888         }
1889         cow_user_page(new_page, old_page, address, vma);
1890         __SetPageUptodate(new_page);
1891
1892         if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
1893                 goto oom_free_new;
1894
1895         /*
1896          * Re-check the pte - we dropped the lock
1897          */
1898         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1899         if (likely(pte_same(*page_table, orig_pte))) {
1900                 if (old_page) {
1901                         if (!PageAnon(old_page)) {
1902                                 dec_mm_counter(mm, file_rss);
1903                                 inc_mm_counter(mm, anon_rss);
1904                         }
1905                 } else
1906                         inc_mm_counter(mm, anon_rss);
1907                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1908                 entry = mk_pte(new_page, vma->vm_page_prot);
1909                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1910                 /*
1911                  * Clear the pte entry and flush it first, before updating the
1912                  * pte with the new entry. This will avoid a race condition
1913                  * seen in the presence of one thread doing SMC and another
1914                  * thread doing COW.
1915                  */
1916                 ptep_clear_flush_notify(vma, address, page_table);
1917                 SetPageSwapBacked(new_page);
1918                 lru_cache_add_active_or_unevictable(new_page, vma);
1919                 page_add_new_anon_rmap(new_page, vma, address);
1920
1921 //TODO:  is this safe?  do_anonymous_page() does it this way.
1922                 set_pte_at(mm, address, page_table, entry);
1923                 update_mmu_cache(vma, address, entry);
1924                 if (old_page) {
1925                         /*
1926                          * Only after switching the pte to the new page may
1927                          * we remove the mapcount here. Otherwise another
1928                          * process may come and find the rmap count decremented
1929                          * before the pte is switched to the new page, and
1930                          * "reuse" the old page writing into it while our pte
1931                          * here still points into it and can be read by other
1932                          * threads.
1933                          *
1934                          * The critical issue is to order this
1935                          * page_remove_rmap with the ptp_clear_flush above.
1936                          * Those stores are ordered by (if nothing else,)
1937                          * the barrier present in the atomic_add_negative
1938                          * in page_remove_rmap.
1939                          *
1940                          * Then the TLB flush in ptep_clear_flush ensures that
1941                          * no process can access the old page before the
1942                          * decremented mapcount is visible. And the old page
1943                          * cannot be reused until after the decremented
1944                          * mapcount is visible. So transitively, TLBs to
1945                          * old page will be flushed before it can be reused.
1946                          */
1947                         page_remove_rmap(old_page, vma);
1948                 }
1949
1950                 /* Free the old page.. */
1951                 new_page = old_page;
1952                 ret |= VM_FAULT_WRITE;
1953         } else
1954                 mem_cgroup_uncharge_page(new_page);
1955
1956         if (new_page)
1957                 page_cache_release(new_page);
1958         if (old_page)
1959                 page_cache_release(old_page);
1960 unlock:
1961         pte_unmap_unlock(page_table, ptl);
1962         if (dirty_page) {
1963                 if (vma->vm_file)
1964                         file_update_time(vma->vm_file);
1965
1966                 /*
1967                  * Yes, Virginia, this is actually required to prevent a race
1968                  * with clear_page_dirty_for_io() from clearing the page dirty
1969                  * bit after it clear all dirty ptes, but before a racing
1970                  * do_wp_page installs a dirty pte.
1971                  *
1972                  * do_no_page is protected similarly.
1973                  */
1974                 wait_on_page_locked(dirty_page);
1975                 set_page_dirty_balance(dirty_page, page_mkwrite);
1976                 put_page(dirty_page);
1977         }
1978         return ret;
1979 oom_free_new:
1980         page_cache_release(new_page);
1981 oom:
1982         if (old_page)
1983                 page_cache_release(old_page);
1984         return VM_FAULT_OOM;
1985
1986 unwritable_page:
1987         page_cache_release(old_page);
1988         return VM_FAULT_SIGBUS;
1989 }
1990
1991 /*
1992  * Helper functions for unmap_mapping_range().
1993  *
1994  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1995  *
1996  * We have to restart searching the prio_tree whenever we drop the lock,
1997  * since the iterator is only valid while the lock is held, and anyway
1998  * a later vma might be split and reinserted earlier while lock dropped.
1999  *
2000  * The list of nonlinear vmas could be handled more efficiently, using
2001  * a placeholder, but handle it in the same way until a need is shown.
2002  * It is important to search the prio_tree before nonlinear list: a vma
2003  * may become nonlinear and be shifted from prio_tree to nonlinear list
2004  * while the lock is dropped; but never shifted from list to prio_tree.
2005  *
2006  * In order to make forward progress despite restarting the search,
2007  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2008  * quickly skip it next time around.  Since the prio_tree search only
2009  * shows us those vmas affected by unmapping the range in question, we
2010  * can't efficiently keep all vmas in step with mapping->truncate_count:
2011  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2012  * mapping->truncate_count and vma->vm_truncate_count are protected by
2013  * i_mmap_lock.
2014  *
2015  * In order to make forward progress despite repeatedly restarting some
2016  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2017  * and restart from that address when we reach that vma again.  It might
2018  * have been split or merged, shrunk or extended, but never shifted: so
2019  * restart_addr remains valid so long as it remains in the vma's range.
2020  * unmap_mapping_range forces truncate_count to leap over page-aligned
2021  * values so we can save vma's restart_addr in its truncate_count field.
2022  */
2023 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2024
2025 static void reset_vma_truncate_counts(struct address_space *mapping)
2026 {
2027         struct vm_area_struct *vma;
2028         struct prio_tree_iter iter;
2029
2030         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2031                 vma->vm_truncate_count = 0;
2032         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2033                 vma->vm_truncate_count = 0;
2034 }
2035
2036 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2037                 unsigned long start_addr, unsigned long end_addr,
2038                 struct zap_details *details)
2039 {
2040         unsigned long restart_addr;
2041         int need_break;
2042
2043         /*
2044          * files that support invalidating or truncating portions of the
2045          * file from under mmaped areas must have their ->fault function
2046          * return a locked page (and set VM_FAULT_LOCKED in the return).
2047          * This provides synchronisation against concurrent unmapping here.
2048          */
2049
2050 again:
2051         restart_addr = vma->vm_truncate_count;
2052         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2053                 start_addr = restart_addr;
2054                 if (start_addr >= end_addr) {
2055                         /* Top of vma has been split off since last time */
2056                         vma->vm_truncate_count = details->truncate_count;
2057                         return 0;
2058                 }
2059         }
2060
2061         restart_addr = zap_page_range(vma, start_addr,
2062                                         end_addr - start_addr, details);
2063         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2064
2065         if (restart_addr >= end_addr) {
2066                 /* We have now completed this vma: mark it so */
2067                 vma->vm_truncate_count = details->truncate_count;
2068                 if (!need_break)
2069                         return 0;
2070         } else {
2071                 /* Note restart_addr in vma's truncate_count field */
2072                 vma->vm_truncate_count = restart_addr;
2073                 if (!need_break)
2074                         goto again;
2075         }
2076
2077         spin_unlock(details->i_mmap_lock);
2078         cond_resched();
2079         spin_lock(details->i_mmap_lock);
2080         return -EINTR;
2081 }
2082
2083 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2084                                             struct zap_details *details)
2085 {
2086         struct vm_area_struct *vma;
2087         struct prio_tree_iter iter;
2088         pgoff_t vba, vea, zba, zea;
2089
2090 restart:
2091         vma_prio_tree_foreach(vma, &iter, root,
2092                         details->first_index, details->last_index) {
2093                 /* Skip quickly over those we have already dealt with */
2094                 if (vma->vm_truncate_count == details->truncate_count)
2095                         continue;
2096
2097                 vba = vma->vm_pgoff;
2098                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2099                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2100                 zba = details->first_index;
2101                 if (zba < vba)
2102                         zba = vba;
2103                 zea = details->last_index;
2104                 if (zea > vea)
2105                         zea = vea;
2106
2107                 if (unmap_mapping_range_vma(vma,
2108                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2109                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2110                                 details) < 0)
2111                         goto restart;
2112         }
2113 }
2114
2115 static inline void unmap_mapping_range_list(struct list_head *head,
2116                                             struct zap_details *details)
2117 {
2118         struct vm_area_struct *vma;
2119
2120         /*
2121          * In nonlinear VMAs there is no correspondence between virtual address
2122          * offset and file offset.  So we must perform an exhaustive search
2123          * across *all* the pages in each nonlinear VMA, not just the pages
2124          * whose virtual address lies outside the file truncation point.
2125          */
2126 restart:
2127         list_for_each_entry(vma, head, shared.vm_set.list) {
2128                 /* Skip quickly over those we have already dealt with */
2129                 if (vma->vm_truncate_count == details->truncate_count)
2130                         continue;
2131                 details->nonlinear_vma = vma;
2132                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2133                                         vma->vm_end, details) < 0)
2134                         goto restart;
2135         }
2136 }
2137
2138 /**
2139  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2140  * @mapping: the address space containing mmaps to be unmapped.
2141  * @holebegin: byte in first page to unmap, relative to the start of
2142  * the underlying file.  This will be rounded down to a PAGE_SIZE
2143  * boundary.  Note that this is different from vmtruncate(), which
2144  * must keep the partial page.  In contrast, we must get rid of
2145  * partial pages.
2146  * @holelen: size of prospective hole in bytes.  This will be rounded
2147  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2148  * end of the file.
2149  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2150  * but 0 when invalidating pagecache, don't throw away private data.
2151  */
2152 void unmap_mapping_range(struct address_space *mapping,
2153                 loff_t const holebegin, loff_t const holelen, int even_cows)
2154 {
2155         struct zap_details details;
2156         pgoff_t hba = holebegin >> PAGE_SHIFT;
2157         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2158
2159         /* Check for overflow. */
2160         if (sizeof(holelen) > sizeof(hlen)) {
2161                 long long holeend =
2162                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2163                 if (holeend & ~(long long)ULONG_MAX)
2164                         hlen = ULONG_MAX - hba + 1;
2165         }
2166
2167         details.check_mapping = even_cows? NULL: mapping;
2168         details.nonlinear_vma = NULL;
2169         details.first_index = hba;
2170         details.last_index = hba + hlen - 1;
2171         if (details.last_index < details.first_index)
2172                 details.last_index = ULONG_MAX;
2173         details.i_mmap_lock = &mapping->i_mmap_lock;
2174
2175         spin_lock(&mapping->i_mmap_lock);
2176
2177         /* Protect against endless unmapping loops */
2178         mapping->truncate_count++;
2179         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2180                 if (mapping->truncate_count == 0)
2181                         reset_vma_truncate_counts(mapping);
2182                 mapping->truncate_count++;
2183         }
2184         details.truncate_count = mapping->truncate_count;
2185
2186         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2187                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2188         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2189                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2190         spin_unlock(&mapping->i_mmap_lock);
2191 }
2192 EXPORT_SYMBOL(unmap_mapping_range);
2193
2194 /**
2195  * vmtruncate - unmap mappings "freed" by truncate() syscall
2196  * @inode: inode of the file used
2197  * @offset: file offset to start truncating
2198  *
2199  * NOTE! We have to be ready to update the memory sharing
2200  * between the file and the memory map for a potential last
2201  * incomplete page.  Ugly, but necessary.
2202  */
2203 int vmtruncate(struct inode * inode, loff_t offset)
2204 {
2205         if (inode->i_size < offset) {
2206                 unsigned long limit;
2207
2208                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2209                 if (limit != RLIM_INFINITY && offset > limit)
2210                         goto out_sig;
2211                 if (offset > inode->i_sb->s_maxbytes)
2212                         goto out_big;
2213                 i_size_write(inode, offset);
2214         } else {
2215                 struct address_space *mapping = inode->i_mapping;
2216
2217                 /*
2218                  * truncation of in-use swapfiles is disallowed - it would
2219                  * cause subsequent swapout to scribble on the now-freed
2220                  * blocks.
2221                  */
2222                 if (IS_SWAPFILE(inode))
2223                         return -ETXTBSY;
2224                 i_size_write(inode, offset);
2225
2226                 /*
2227                  * unmap_mapping_range is called twice, first simply for
2228                  * efficiency so that truncate_inode_pages does fewer
2229                  * single-page unmaps.  However after this first call, and
2230                  * before truncate_inode_pages finishes, it is possible for
2231                  * private pages to be COWed, which remain after
2232                  * truncate_inode_pages finishes, hence the second
2233                  * unmap_mapping_range call must be made for correctness.
2234                  */
2235                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2236                 truncate_inode_pages(mapping, offset);
2237                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2238         }
2239
2240         if (inode->i_op && inode->i_op->truncate)
2241                 inode->i_op->truncate(inode);
2242         return 0;
2243
2244 out_sig:
2245         send_sig(SIGXFSZ, current, 0);
2246 out_big:
2247         return -EFBIG;
2248 }
2249 EXPORT_SYMBOL(vmtruncate);
2250
2251 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2252 {
2253         struct address_space *mapping = inode->i_mapping;
2254
2255         /*
2256          * If the underlying filesystem is not going to provide
2257          * a way to truncate a range of blocks (punch a hole) -
2258          * we should return failure right now.
2259          */
2260         if (!inode->i_op || !inode->i_op->truncate_range)
2261                 return -ENOSYS;
2262
2263         mutex_lock(&inode->i_mutex);
2264         down_write(&inode->i_alloc_sem);
2265         unmap_mapping_range(mapping, offset, (end - offset), 1);
2266         truncate_inode_pages_range(mapping, offset, end);
2267         unmap_mapping_range(mapping, offset, (end - offset), 1);
2268         inode->i_op->truncate_range(inode, offset, end);
2269         up_write(&inode->i_alloc_sem);
2270         mutex_unlock(&inode->i_mutex);
2271
2272         return 0;
2273 }
2274
2275 /*
2276  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2277  * but allow concurrent faults), and pte mapped but not yet locked.
2278  * We return with mmap_sem still held, but pte unmapped and unlocked.
2279  */
2280 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2281                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2282                 int write_access, pte_t orig_pte)
2283 {
2284         spinlock_t *ptl;
2285         struct page *page;
2286         swp_entry_t entry;
2287         pte_t pte;
2288         int ret = 0;
2289
2290         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2291                 goto out;
2292
2293         entry = pte_to_swp_entry(orig_pte);
2294         if (is_migration_entry(entry)) {
2295                 migration_entry_wait(mm, pmd, address);
2296                 goto out;
2297         }
2298         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2299         page = lookup_swap_cache(entry);
2300         if (!page) {
2301                 grab_swap_token(); /* Contend for token _before_ read-in */
2302                 page = swapin_readahead(entry,
2303                                         GFP_HIGHUSER_MOVABLE, vma, address);
2304                 if (!page) {
2305                         /*
2306                          * Back out if somebody else faulted in this pte
2307                          * while we released the pte lock.
2308                          */
2309                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2310                         if (likely(pte_same(*page_table, orig_pte)))
2311                                 ret = VM_FAULT_OOM;
2312                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2313                         goto unlock;
2314                 }
2315
2316                 /* Had to read the page from swap area: Major fault */
2317                 ret = VM_FAULT_MAJOR;
2318                 count_vm_event(PGMAJFAULT);
2319         }
2320
2321         mark_page_accessed(page);
2322
2323         lock_page(page);
2324         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2325
2326         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2327                 ret = VM_FAULT_OOM;
2328                 unlock_page(page);
2329                 goto out;
2330         }
2331
2332         /*
2333          * Back out if somebody else already faulted in this pte.
2334          */
2335         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2336         if (unlikely(!pte_same(*page_table, orig_pte)))
2337                 goto out_nomap;
2338
2339         if (unlikely(!PageUptodate(page))) {
2340                 ret = VM_FAULT_SIGBUS;
2341                 goto out_nomap;
2342         }
2343
2344         /* The page isn't present yet, go ahead with the fault. */
2345
2346         inc_mm_counter(mm, anon_rss);
2347         pte = mk_pte(page, vma->vm_page_prot);
2348         if (write_access && can_share_swap_page(page)) {
2349                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2350                 write_access = 0;
2351         }
2352
2353         flush_icache_page(vma, page);
2354         set_pte_at(mm, address, page_table, pte);
2355         page_add_anon_rmap(page, vma, address);
2356
2357         swap_free(entry);
2358         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2359                 remove_exclusive_swap_page(page);
2360         unlock_page(page);
2361
2362         if (write_access) {
2363                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2364                 if (ret & VM_FAULT_ERROR)
2365                         ret &= VM_FAULT_ERROR;
2366                 goto out;
2367         }
2368
2369         /* No need to invalidate - it was non-present before */
2370         update_mmu_cache(vma, address, pte);
2371 unlock:
2372         pte_unmap_unlock(page_table, ptl);
2373 out:
2374         return ret;
2375 out_nomap:
2376         mem_cgroup_uncharge_page(page);
2377         pte_unmap_unlock(page_table, ptl);
2378         unlock_page(page);
2379         page_cache_release(page);
2380         return ret;
2381 }
2382
2383 /*
2384  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2385  * but allow concurrent faults), and pte mapped but not yet locked.
2386  * We return with mmap_sem still held, but pte unmapped and unlocked.
2387  */
2388 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2389                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2390                 int write_access)
2391 {
2392         struct page *page;
2393         spinlock_t *ptl;
2394         pte_t entry;
2395
2396         /* Allocate our own private page. */
2397         pte_unmap(page_table);
2398
2399         if (unlikely(anon_vma_prepare(vma)))
2400                 goto oom;
2401         page = alloc_zeroed_user_highpage_movable(vma, address);
2402         if (!page)
2403                 goto oom;
2404         __SetPageUptodate(page);
2405
2406         if (mem_cgroup_charge(page, mm, GFP_KERNEL))
2407                 goto oom_free_page;
2408
2409         entry = mk_pte(page, vma->vm_page_prot);
2410         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2411
2412         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2413         if (!pte_none(*page_table))
2414                 goto release;
2415         inc_mm_counter(mm, anon_rss);
2416         SetPageSwapBacked(page);
2417         lru_cache_add_active_or_unevictable(page, vma);
2418         page_add_new_anon_rmap(page, vma, address);
2419         set_pte_at(mm, address, page_table, entry);
2420
2421         /* No need to invalidate - it was non-present before */
2422         update_mmu_cache(vma, address, entry);
2423 unlock:
2424         pte_unmap_unlock(page_table, ptl);
2425         return 0;
2426 release:
2427         mem_cgroup_uncharge_page(page);
2428         page_cache_release(page);
2429         goto unlock;
2430 oom_free_page:
2431         page_cache_release(page);
2432 oom:
2433         return VM_FAULT_OOM;
2434 }
2435
2436 /*
2437  * __do_fault() tries to create a new page mapping. It aggressively
2438  * tries to share with existing pages, but makes a separate copy if
2439  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2440  * the next page fault.
2441  *
2442  * As this is called only for pages that do not currently exist, we
2443  * do not need to flush old virtual caches or the TLB.
2444  *
2445  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2446  * but allow concurrent faults), and pte neither mapped nor locked.
2447  * We return with mmap_sem still held, but pte unmapped and unlocked.
2448  */
2449 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2450                 unsigned long address, pmd_t *pmd,
2451                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2452 {
2453         pte_t *page_table;
2454         spinlock_t *ptl;
2455         struct page *page;
2456         pte_t entry;
2457         int anon = 0;
2458         int charged = 0;
2459         struct page *dirty_page = NULL;
2460         struct vm_fault vmf;
2461         int ret;
2462         int page_mkwrite = 0;
2463
2464         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2465         vmf.pgoff = pgoff;
2466         vmf.flags = flags;
2467         vmf.page = NULL;
2468
2469         ret = vma->vm_ops->fault(vma, &vmf);
2470         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2471                 return ret;
2472
2473         /*
2474          * For consistency in subsequent calls, make the faulted page always
2475          * locked.
2476          */
2477         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2478                 lock_page(vmf.page);
2479         else
2480                 VM_BUG_ON(!PageLocked(vmf.page));
2481
2482         /*
2483          * Should we do an early C-O-W break?
2484          */
2485         page = vmf.page;
2486         if (flags & FAULT_FLAG_WRITE) {
2487                 if (!(vma->vm_flags & VM_SHARED)) {
2488                         anon = 1;
2489                         if (unlikely(anon_vma_prepare(vma))) {
2490                                 ret = VM_FAULT_OOM;
2491                                 goto out;
2492                         }
2493                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2494                                                 vma, address);
2495                         if (!page) {
2496                                 ret = VM_FAULT_OOM;
2497                                 goto out;
2498                         }
2499                         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2500                                 ret = VM_FAULT_OOM;
2501                                 page_cache_release(page);
2502                                 goto out;
2503                         }
2504                         charged = 1;
2505                         /*
2506                          * Don't let another task, with possibly unlocked vma,
2507                          * keep the mlocked page.
2508                          */
2509                         if (vma->vm_flags & VM_LOCKED)
2510                                 clear_page_mlock(vmf.page);
2511                         copy_user_highpage(page, vmf.page, address, vma);
2512                         __SetPageUptodate(page);
2513                 } else {
2514                         /*
2515                          * If the page will be shareable, see if the backing
2516                          * address space wants to know that the page is about
2517                          * to become writable
2518                          */
2519                         if (vma->vm_ops->page_mkwrite) {
2520                                 unlock_page(page);
2521                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2522                                         ret = VM_FAULT_SIGBUS;
2523                                         anon = 1; /* no anon but release vmf.page */
2524                                         goto out_unlocked;
2525                                 }
2526                                 lock_page(page);
2527                                 /*
2528                                  * XXX: this is not quite right (racy vs
2529                                  * invalidate) to unlock and relock the page
2530                                  * like this, however a better fix requires
2531                                  * reworking page_mkwrite locking API, which
2532                                  * is better done later.
2533                                  */
2534                                 if (!page->mapping) {
2535                                         ret = 0;
2536                                         anon = 1; /* no anon but release vmf.page */
2537                                         goto out;
2538                                 }
2539                                 page_mkwrite = 1;
2540                         }
2541                 }
2542
2543         }
2544
2545         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2546
2547         /*
2548          * This silly early PAGE_DIRTY setting removes a race
2549          * due to the bad i386 page protection. But it's valid
2550          * for other architectures too.
2551          *
2552          * Note that if write_access is true, we either now have
2553          * an exclusive copy of the page, or this is a shared mapping,
2554          * so we can make it writable and dirty to avoid having to
2555          * handle that later.
2556          */
2557         /* Only go through if we didn't race with anybody else... */
2558         if (likely(pte_same(*page_table, orig_pte))) {
2559                 flush_icache_page(vma, page);
2560                 entry = mk_pte(page, vma->vm_page_prot);
2561                 if (flags & FAULT_FLAG_WRITE)
2562                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2563                 if (anon) {
2564                         inc_mm_counter(mm, anon_rss);
2565                         SetPageSwapBacked(page);
2566                         lru_cache_add_active_or_unevictable(page, vma);
2567                         page_add_new_anon_rmap(page, vma, address);
2568                 } else {
2569                         inc_mm_counter(mm, file_rss);
2570                         page_add_file_rmap(page);
2571                         if (flags & FAULT_FLAG_WRITE) {
2572                                 dirty_page = page;
2573                                 get_page(dirty_page);
2574                         }
2575                 }
2576 //TODO:  is this safe?  do_anonymous_page() does it this way.
2577                 set_pte_at(mm, address, page_table, entry);
2578
2579                 /* no need to invalidate: a not-present page won't be cached */
2580                 update_mmu_cache(vma, address, entry);
2581         } else {
2582                 if (charged)
2583                         mem_cgroup_uncharge_page(page);
2584                 if (anon)
2585                         page_cache_release(page);
2586                 else
2587                         anon = 1; /* no anon but release faulted_page */
2588         }
2589
2590         pte_unmap_unlock(page_table, ptl);
2591
2592 out:
2593         unlock_page(vmf.page);
2594 out_unlocked:
2595         if (anon)
2596                 page_cache_release(vmf.page);
2597         else if (dirty_page) {
2598                 if (vma->vm_file)
2599                         file_update_time(vma->vm_file);
2600
2601                 set_page_dirty_balance(dirty_page, page_mkwrite);
2602                 put_page(dirty_page);
2603         }
2604
2605         return ret;
2606 }
2607
2608 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2609                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2610                 int write_access, pte_t orig_pte)
2611 {
2612         pgoff_t pgoff = (((address & PAGE_MASK)
2613                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2614         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2615
2616         pte_unmap(page_table);
2617         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2618 }
2619
2620 /*
2621  * Fault of a previously existing named mapping. Repopulate the pte
2622  * from the encoded file_pte if possible. This enables swappable
2623  * nonlinear vmas.
2624  *
2625  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2626  * but allow concurrent faults), and pte mapped but not yet locked.
2627  * We return with mmap_sem still held, but pte unmapped and unlocked.
2628  */
2629 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2630                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2631                 int write_access, pte_t orig_pte)
2632 {
2633         unsigned int flags = FAULT_FLAG_NONLINEAR |
2634                                 (write_access ? FAULT_FLAG_WRITE : 0);
2635         pgoff_t pgoff;
2636
2637         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2638                 return 0;
2639
2640         if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2641                         !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2642                 /*
2643                  * Page table corrupted: show pte and kill process.
2644                  */
2645                 print_bad_pte(vma, orig_pte, address);
2646                 return VM_FAULT_OOM;
2647         }
2648
2649         pgoff = pte_to_pgoff(orig_pte);
2650         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2651 }
2652
2653 /*
2654  * These routines also need to handle stuff like marking pages dirty
2655  * and/or accessed for architectures that don't do it in hardware (most
2656  * RISC architectures).  The early dirtying is also good on the i386.
2657  *
2658  * There is also a hook called "update_mmu_cache()" that architectures
2659  * with external mmu caches can use to update those (ie the Sparc or
2660  * PowerPC hashed page tables that act as extended TLBs).
2661  *
2662  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2663  * but allow concurrent faults), and pte mapped but not yet locked.
2664  * We return with mmap_sem still held, but pte unmapped and unlocked.
2665  */
2666 static inline int handle_pte_fault(struct mm_struct *mm,
2667                 struct vm_area_struct *vma, unsigned long address,
2668                 pte_t *pte, pmd_t *pmd, int write_access)
2669 {
2670         pte_t entry;
2671         spinlock_t *ptl;
2672
2673         entry = *pte;
2674         if (!pte_present(entry)) {
2675                 if (pte_none(entry)) {
2676                         if (vma->vm_ops) {
2677                                 if (likely(vma->vm_ops->fault))
2678                                         return do_linear_fault(mm, vma, address,
2679                                                 pte, pmd, write_access, entry);
2680                         }
2681                         return do_anonymous_page(mm, vma, address,
2682                                                  pte, pmd, write_access);
2683                 }
2684                 if (pte_file(entry))
2685                         return do_nonlinear_fault(mm, vma, address,
2686                                         pte, pmd, write_access, entry);
2687                 return do_swap_page(mm, vma, address,
2688                                         pte, pmd, write_access, entry);
2689         }
2690
2691         ptl = pte_lockptr(mm, pmd);
2692         spin_lock(ptl);
2693         if (unlikely(!pte_same(*pte, entry)))
2694                 goto unlock;
2695         if (write_access) {
2696                 if (!pte_write(entry))
2697                         return do_wp_page(mm, vma, address,
2698                                         pte, pmd, ptl, entry);
2699                 entry = pte_mkdirty(entry);
2700         }
2701         entry = pte_mkyoung(entry);
2702         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2703                 update_mmu_cache(vma, address, entry);
2704         } else {
2705                 /*
2706                  * This is needed only for protection faults but the arch code
2707                  * is not yet telling us if this is a protection fault or not.
2708                  * This still avoids useless tlb flushes for .text page faults
2709                  * with threads.
2710                  */
2711                 if (write_access)
2712                         flush_tlb_page(vma, address);
2713         }
2714 unlock:
2715         pte_unmap_unlock(pte, ptl);
2716         return 0;
2717 }
2718
2719 /*
2720  * By the time we get here, we already hold the mm semaphore
2721  */
2722 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2723                 unsigned long address, int write_access)
2724 {
2725         pgd_t *pgd;
2726         pud_t *pud;
2727         pmd_t *pmd;
2728         pte_t *pte;
2729
2730         __set_current_state(TASK_RUNNING);
2731
2732         count_vm_event(PGFAULT);
2733
2734         if (unlikely(is_vm_hugetlb_page(vma)))
2735                 return hugetlb_fault(mm, vma, address, write_access);
2736
2737         pgd = pgd_offset(mm, address);
2738         pud = pud_alloc(mm, pgd, address);
2739         if (!pud)
2740                 return VM_FAULT_OOM;
2741         pmd = pmd_alloc(mm, pud, address);
2742         if (!pmd)
2743                 return VM_FAULT_OOM;
2744         pte = pte_alloc_map(mm, pmd, address);
2745         if (!pte)
2746                 return VM_FAULT_OOM;
2747
2748         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2749 }
2750
2751 #ifndef __PAGETABLE_PUD_FOLDED
2752 /*
2753  * Allocate page upper directory.
2754  * We've already handled the fast-path in-line.
2755  */
2756 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2757 {
2758         pud_t *new = pud_alloc_one(mm, address);
2759         if (!new)
2760                 return -ENOMEM;
2761
2762         smp_wmb(); /* See comment in __pte_alloc */
2763
2764         spin_lock(&mm->page_table_lock);
2765         if (pgd_present(*pgd))          /* Another has populated it */
2766                 pud_free(mm, new);
2767         else
2768                 pgd_populate(mm, pgd, new);
2769         spin_unlock(&mm->page_table_lock);
2770         return 0;
2771 }
2772 #endif /* __PAGETABLE_PUD_FOLDED */
2773
2774 #ifndef __PAGETABLE_PMD_FOLDED
2775 /*
2776  * Allocate page middle directory.
2777  * We've already handled the fast-path in-line.
2778  */
2779 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2780 {
2781         pmd_t *new = pmd_alloc_one(mm, address);
2782         if (!new)
2783                 return -ENOMEM;
2784
2785         smp_wmb(); /* See comment in __pte_alloc */
2786
2787         spin_lock(&mm->page_table_lock);
2788 #ifndef __ARCH_HAS_4LEVEL_HACK
2789         if (pud_present(*pud))          /* Another has populated it */
2790                 pmd_free(mm, new);
2791         else
2792                 pud_populate(mm, pud, new);
2793 #else
2794         if (pgd_present(*pud))          /* Another has populated it */
2795                 pmd_free(mm, new);
2796         else
2797                 pgd_populate(mm, pud, new);
2798 #endif /* __ARCH_HAS_4LEVEL_HACK */
2799         spin_unlock(&mm->page_table_lock);
2800         return 0;
2801 }
2802 #endif /* __PAGETABLE_PMD_FOLDED */
2803
2804 int make_pages_present(unsigned long addr, unsigned long end)
2805 {
2806         int ret, len, write;
2807         struct vm_area_struct * vma;
2808
2809         vma = find_vma(current->mm, addr);
2810         if (!vma)
2811                 return -ENOMEM;
2812         write = (vma->vm_flags & VM_WRITE) != 0;
2813         BUG_ON(addr >= end);
2814         BUG_ON(end > vma->vm_end);
2815         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2816         ret = get_user_pages(current, current->mm, addr,
2817                         len, write, 0, NULL, NULL);
2818         if (ret < 0)
2819                 return ret;
2820         return ret == len ? 0 : -EFAULT;
2821 }
2822
2823 #if !defined(__HAVE_ARCH_GATE_AREA)
2824
2825 #if defined(AT_SYSINFO_EHDR)
2826 static struct vm_area_struct gate_vma;
2827
2828 static int __init gate_vma_init(void)
2829 {
2830         gate_vma.vm_mm = NULL;
2831         gate_vma.vm_start = FIXADDR_USER_START;
2832         gate_vma.vm_end = FIXADDR_USER_END;
2833         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2834         gate_vma.vm_page_prot = __P101;
2835         /*
2836          * Make sure the vDSO gets into every core dump.
2837          * Dumping its contents makes post-mortem fully interpretable later
2838          * without matching up the same kernel and hardware config to see
2839          * what PC values meant.
2840          */
2841         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2842         return 0;
2843 }
2844 __initcall(gate_vma_init);
2845 #endif
2846
2847 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2848 {
2849 #ifdef AT_SYSINFO_EHDR
2850         return &gate_vma;
2851 #else
2852         return NULL;
2853 #endif
2854 }
2855
2856 int in_gate_area_no_task(unsigned long addr)
2857 {
2858 #ifdef AT_SYSINFO_EHDR
2859         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2860                 return 1;
2861 #endif
2862         return 0;
2863 }
2864
2865 #endif  /* __HAVE_ARCH_GATE_AREA */
2866
2867 #ifdef CONFIG_HAVE_IOREMAP_PROT
2868 static resource_size_t follow_phys(struct vm_area_struct *vma,
2869                         unsigned long address, unsigned int flags,
2870                         unsigned long *prot)
2871 {
2872         pgd_t *pgd;
2873         pud_t *pud;
2874         pmd_t *pmd;
2875         pte_t *ptep, pte;
2876         spinlock_t *ptl;
2877         resource_size_t phys_addr = 0;
2878         struct mm_struct *mm = vma->vm_mm;
2879
2880         VM_BUG_ON(!(vma->vm_flags & (VM_IO | VM_PFNMAP)));
2881
2882         pgd = pgd_offset(mm, address);
2883         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
2884                 goto no_page_table;
2885
2886         pud = pud_offset(pgd, address);
2887         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
2888                 goto no_page_table;
2889
2890         pmd = pmd_offset(pud, address);
2891         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
2892                 goto no_page_table;
2893
2894         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
2895         if (pmd_huge(*pmd))
2896                 goto no_page_table;
2897
2898         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
2899         if (!ptep)
2900                 goto out;
2901
2902         pte = *ptep;
2903         if (!pte_present(pte))
2904                 goto unlock;
2905         if ((flags & FOLL_WRITE) && !pte_write(pte))
2906                 goto unlock;
2907         phys_addr = pte_pfn(pte);
2908         phys_addr <<= PAGE_SHIFT; /* Shift here to avoid overflow on PAE */
2909
2910         *prot = pgprot_val(pte_pgprot(pte));
2911
2912 unlock:
2913         pte_unmap_unlock(ptep, ptl);
2914 out:
2915         return phys_addr;
2916 no_page_table:
2917         return 0;
2918 }
2919
2920 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
2921                         void *buf, int len, int write)
2922 {
2923         resource_size_t phys_addr;
2924         unsigned long prot = 0;
2925         void *maddr;
2926         int offset = addr & (PAGE_SIZE-1);
2927
2928         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
2929                 return -EINVAL;
2930
2931         phys_addr = follow_phys(vma, addr, write, &prot);
2932
2933         if (!phys_addr)
2934                 return -EINVAL;
2935
2936         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
2937         if (write)
2938                 memcpy_toio(maddr + offset, buf, len);
2939         else
2940                 memcpy_fromio(buf, maddr + offset, len);
2941         iounmap(maddr);
2942
2943         return len;
2944 }
2945 #endif
2946
2947 /*
2948  * Access another process' address space.
2949  * Source/target buffer must be kernel space,
2950  * Do not walk the page table directly, use get_user_pages
2951  */
2952 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2953 {
2954         struct mm_struct *mm;
2955         struct vm_area_struct *vma;
2956         void *old_buf = buf;
2957
2958         mm = get_task_mm(tsk);
2959         if (!mm)
2960                 return 0;
2961
2962         down_read(&mm->mmap_sem);
2963         /* ignore errors, just check how much was successfully transferred */
2964         while (len) {
2965                 int bytes, ret, offset;
2966                 void *maddr;
2967                 struct page *page = NULL;
2968
2969                 ret = get_user_pages(tsk, mm, addr, 1,
2970                                 write, 1, &page, &vma);
2971                 if (ret <= 0) {
2972                         /*
2973                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
2974                          * we can access using slightly different code.
2975                          */
2976 #ifdef CONFIG_HAVE_IOREMAP_PROT
2977                         vma = find_vma(mm, addr);
2978                         if (!vma)
2979                                 break;
2980                         if (vma->vm_ops && vma->vm_ops->access)
2981                                 ret = vma->vm_ops->access(vma, addr, buf,
2982                                                           len, write);
2983                         if (ret <= 0)
2984 #endif
2985                                 break;
2986                         bytes = ret;
2987                 } else {
2988                         bytes = len;
2989                         offset = addr & (PAGE_SIZE-1);
2990                         if (bytes > PAGE_SIZE-offset)
2991                                 bytes = PAGE_SIZE-offset;
2992
2993                         maddr = kmap(page);
2994                         if (write) {
2995                                 copy_to_user_page(vma, page, addr,
2996                                                   maddr + offset, buf, bytes);
2997                                 set_page_dirty_lock(page);
2998                         } else {
2999                                 copy_from_user_page(vma, page, addr,
3000                                                     buf, maddr + offset, bytes);
3001                         }
3002                         kunmap(page);
3003                         page_cache_release(page);
3004                 }
3005                 len -= bytes;
3006                 buf += bytes;
3007                 addr += bytes;
3008         }
3009         up_read(&mm->mmap_sem);
3010         mmput(mm);
3011
3012         return buf - old_buf;
3013 }
3014
3015 /*
3016  * Print the name of a VMA.
3017  */
3018 void print_vma_addr(char *prefix, unsigned long ip)
3019 {
3020         struct mm_struct *mm = current->mm;
3021         struct vm_area_struct *vma;
3022
3023         /*
3024          * Do not print if we are in atomic
3025          * contexts (in exception stacks, etc.):
3026          */
3027         if (preempt_count())
3028                 return;
3029
3030         down_read(&mm->mmap_sem);
3031         vma = find_vma(mm, ip);
3032         if (vma && vma->vm_file) {
3033                 struct file *f = vma->vm_file;
3034                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3035                 if (buf) {
3036                         char *p, *s;
3037
3038                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3039                         if (IS_ERR(p))
3040                                 p = "?";
3041                         s = strrchr(p, '/');
3042                         if (s)
3043                                 p = s+1;
3044                         printk("%s%s[%lx+%lx]", prefix, p,
3045                                         vma->vm_start,
3046                                         vma->vm_end - vma->vm_start);
3047                         free_page((unsigned long)buf);
3048                 }
3049         }
3050         up_read(&current->mm->mmap_sem);
3051 }