ksm: add mmu_notifier set_pte_at_notify()
[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 #include <linux/kallsyms.h>
56 #include <linux/swapops.h>
57 #include <linux/elf.h>
58
59 #include <asm/io.h>
60 #include <asm/pgalloc.h>
61 #include <asm/uaccess.h>
62 #include <asm/tlb.h>
63 #include <asm/tlbflush.h>
64 #include <asm/pgtable.h>
65
66 #include "internal.h"
67
68 #ifndef CONFIG_NEED_MULTIPLE_NODES
69 /* use the per-pgdat data instead for discontigmem - mbligh */
70 unsigned long max_mapnr;
71 struct page *mem_map;
72
73 EXPORT_SYMBOL(max_mapnr);
74 EXPORT_SYMBOL(mem_map);
75 #endif
76
77 unsigned long num_physpages;
78 /*
79  * A number of key systems in x86 including ioremap() rely on the assumption
80  * that high_memory defines the upper bound on direct map memory, then end
81  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
82  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
83  * and ZONE_HIGHMEM.
84  */
85 void * high_memory;
86
87 EXPORT_SYMBOL(num_physpages);
88 EXPORT_SYMBOL(high_memory);
89
90 /*
91  * Randomize the address space (stacks, mmaps, brk, etc.).
92  *
93  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
94  *   as ancient (libc5 based) binaries can segfault. )
95  */
96 int randomize_va_space __read_mostly =
97 #ifdef CONFIG_COMPAT_BRK
98                                         1;
99 #else
100                                         2;
101 #endif
102
103 static int __init disable_randmaps(char *s)
104 {
105         randomize_va_space = 0;
106         return 1;
107 }
108 __setup("norandmaps", disable_randmaps);
109
110
111 /*
112  * If a p?d_bad entry is found while walking page tables, report
113  * the error, before resetting entry to p?d_none.  Usually (but
114  * very seldom) called out from the p?d_none_or_clear_bad macros.
115  */
116
117 void pgd_clear_bad(pgd_t *pgd)
118 {
119         pgd_ERROR(*pgd);
120         pgd_clear(pgd);
121 }
122
123 void pud_clear_bad(pud_t *pud)
124 {
125         pud_ERROR(*pud);
126         pud_clear(pud);
127 }
128
129 void pmd_clear_bad(pmd_t *pmd)
130 {
131         pmd_ERROR(*pmd);
132         pmd_clear(pmd);
133 }
134
135 /*
136  * Note: this doesn't free the actual pages themselves. That
137  * has been handled earlier when unmapping all the memory regions.
138  */
139 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
140                            unsigned long addr)
141 {
142         pgtable_t token = pmd_pgtable(*pmd);
143         pmd_clear(pmd);
144         pte_free_tlb(tlb, token, addr);
145         tlb->mm->nr_ptes--;
146 }
147
148 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
149                                 unsigned long addr, unsigned long end,
150                                 unsigned long floor, unsigned long ceiling)
151 {
152         pmd_t *pmd;
153         unsigned long next;
154         unsigned long start;
155
156         start = addr;
157         pmd = pmd_offset(pud, addr);
158         do {
159                 next = pmd_addr_end(addr, end);
160                 if (pmd_none_or_clear_bad(pmd))
161                         continue;
162                 free_pte_range(tlb, pmd, addr);
163         } while (pmd++, addr = next, addr != end);
164
165         start &= PUD_MASK;
166         if (start < floor)
167                 return;
168         if (ceiling) {
169                 ceiling &= PUD_MASK;
170                 if (!ceiling)
171                         return;
172         }
173         if (end - 1 > ceiling - 1)
174                 return;
175
176         pmd = pmd_offset(pud, start);
177         pud_clear(pud);
178         pmd_free_tlb(tlb, pmd, start);
179 }
180
181 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
182                                 unsigned long addr, unsigned long end,
183                                 unsigned long floor, unsigned long ceiling)
184 {
185         pud_t *pud;
186         unsigned long next;
187         unsigned long start;
188
189         start = addr;
190         pud = pud_offset(pgd, addr);
191         do {
192                 next = pud_addr_end(addr, end);
193                 if (pud_none_or_clear_bad(pud))
194                         continue;
195                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
196         } while (pud++, addr = next, addr != end);
197
198         start &= PGDIR_MASK;
199         if (start < floor)
200                 return;
201         if (ceiling) {
202                 ceiling &= PGDIR_MASK;
203                 if (!ceiling)
204                         return;
205         }
206         if (end - 1 > ceiling - 1)
207                 return;
208
209         pud = pud_offset(pgd, start);
210         pgd_clear(pgd);
211         pud_free_tlb(tlb, pud, start);
212 }
213
214 /*
215  * This function frees user-level page tables of a process.
216  *
217  * Must be called with pagetable lock held.
218  */
219 void free_pgd_range(struct mmu_gather *tlb,
220                         unsigned long addr, unsigned long end,
221                         unsigned long floor, unsigned long ceiling)
222 {
223         pgd_t *pgd;
224         unsigned long next;
225         unsigned long start;
226
227         /*
228          * The next few lines have given us lots of grief...
229          *
230          * Why are we testing PMD* at this top level?  Because often
231          * there will be no work to do at all, and we'd prefer not to
232          * go all the way down to the bottom just to discover that.
233          *
234          * Why all these "- 1"s?  Because 0 represents both the bottom
235          * of the address space and the top of it (using -1 for the
236          * top wouldn't help much: the masks would do the wrong thing).
237          * The rule is that addr 0 and floor 0 refer to the bottom of
238          * the address space, but end 0 and ceiling 0 refer to the top
239          * Comparisons need to use "end - 1" and "ceiling - 1" (though
240          * that end 0 case should be mythical).
241          *
242          * Wherever addr is brought up or ceiling brought down, we must
243          * be careful to reject "the opposite 0" before it confuses the
244          * subsequent tests.  But what about where end is brought down
245          * by PMD_SIZE below? no, end can't go down to 0 there.
246          *
247          * Whereas we round start (addr) and ceiling down, by different
248          * masks at different levels, in order to test whether a table
249          * now has no other vmas using it, so can be freed, we don't
250          * bother to round floor or end up - the tests don't need that.
251          */
252
253         addr &= PMD_MASK;
254         if (addr < floor) {
255                 addr += PMD_SIZE;
256                 if (!addr)
257                         return;
258         }
259         if (ceiling) {
260                 ceiling &= PMD_MASK;
261                 if (!ceiling)
262                         return;
263         }
264         if (end - 1 > ceiling - 1)
265                 end -= PMD_SIZE;
266         if (addr > end - 1)
267                 return;
268
269         start = addr;
270         pgd = pgd_offset(tlb->mm, addr);
271         do {
272                 next = pgd_addr_end(addr, end);
273                 if (pgd_none_or_clear_bad(pgd))
274                         continue;
275                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
276         } while (pgd++, addr = next, addr != end);
277 }
278
279 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
280                 unsigned long floor, unsigned long ceiling)
281 {
282         while (vma) {
283                 struct vm_area_struct *next = vma->vm_next;
284                 unsigned long addr = vma->vm_start;
285
286                 /*
287                  * Hide vma from rmap and vmtruncate before freeing pgtables
288                  */
289                 anon_vma_unlink(vma);
290                 unlink_file_vma(vma);
291
292                 if (is_vm_hugetlb_page(vma)) {
293                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
294                                 floor, next? next->vm_start: ceiling);
295                 } else {
296                         /*
297                          * Optimization: gather nearby vmas into one call down
298                          */
299                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
300                                && !is_vm_hugetlb_page(next)) {
301                                 vma = next;
302                                 next = vma->vm_next;
303                                 anon_vma_unlink(vma);
304                                 unlink_file_vma(vma);
305                         }
306                         free_pgd_range(tlb, addr, vma->vm_end,
307                                 floor, next? next->vm_start: ceiling);
308                 }
309                 vma = next;
310         }
311 }
312
313 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
314 {
315         pgtable_t new = pte_alloc_one(mm, address);
316         if (!new)
317                 return -ENOMEM;
318
319         /*
320          * Ensure all pte setup (eg. pte page lock and page clearing) are
321          * visible before the pte is made visible to other CPUs by being
322          * put into page tables.
323          *
324          * The other side of the story is the pointer chasing in the page
325          * table walking code (when walking the page table without locking;
326          * ie. most of the time). Fortunately, these data accesses consist
327          * of a chain of data-dependent loads, meaning most CPUs (alpha
328          * being the notable exception) will already guarantee loads are
329          * seen in-order. See the alpha page table accessors for the
330          * smp_read_barrier_depends() barriers in page table walking code.
331          */
332         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
333
334         spin_lock(&mm->page_table_lock);
335         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
336                 mm->nr_ptes++;
337                 pmd_populate(mm, pmd, new);
338                 new = NULL;
339         }
340         spin_unlock(&mm->page_table_lock);
341         if (new)
342                 pte_free(mm, new);
343         return 0;
344 }
345
346 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
347 {
348         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
349         if (!new)
350                 return -ENOMEM;
351
352         smp_wmb(); /* See comment in __pte_alloc */
353
354         spin_lock(&init_mm.page_table_lock);
355         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
356                 pmd_populate_kernel(&init_mm, pmd, new);
357                 new = NULL;
358         }
359         spin_unlock(&init_mm.page_table_lock);
360         if (new)
361                 pte_free_kernel(&init_mm, new);
362         return 0;
363 }
364
365 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
366 {
367         if (file_rss)
368                 add_mm_counter(mm, file_rss, file_rss);
369         if (anon_rss)
370                 add_mm_counter(mm, anon_rss, anon_rss);
371 }
372
373 /*
374  * This function is called to print an error when a bad pte
375  * is found. For example, we might have a PFN-mapped pte in
376  * a region that doesn't allow it.
377  *
378  * The calling function must still handle the error.
379  */
380 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
381                           pte_t pte, struct page *page)
382 {
383         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
384         pud_t *pud = pud_offset(pgd, addr);
385         pmd_t *pmd = pmd_offset(pud, addr);
386         struct address_space *mapping;
387         pgoff_t index;
388         static unsigned long resume;
389         static unsigned long nr_shown;
390         static unsigned long nr_unshown;
391
392         /*
393          * Allow a burst of 60 reports, then keep quiet for that minute;
394          * or allow a steady drip of one report per second.
395          */
396         if (nr_shown == 60) {
397                 if (time_before(jiffies, resume)) {
398                         nr_unshown++;
399                         return;
400                 }
401                 if (nr_unshown) {
402                         printk(KERN_ALERT
403                                 "BUG: Bad page map: %lu messages suppressed\n",
404                                 nr_unshown);
405                         nr_unshown = 0;
406                 }
407                 nr_shown = 0;
408         }
409         if (nr_shown++ == 0)
410                 resume = jiffies + 60 * HZ;
411
412         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
413         index = linear_page_index(vma, addr);
414
415         printk(KERN_ALERT
416                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
417                 current->comm,
418                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
419         if (page) {
420                 printk(KERN_ALERT
421                 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
422                 page, (void *)page->flags, page_count(page),
423                 page_mapcount(page), page->mapping, page->index);
424         }
425         printk(KERN_ALERT
426                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
427                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
428         /*
429          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
430          */
431         if (vma->vm_ops)
432                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
433                                 (unsigned long)vma->vm_ops->fault);
434         if (vma->vm_file && vma->vm_file->f_op)
435                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
436                                 (unsigned long)vma->vm_file->f_op->mmap);
437         dump_stack();
438         add_taint(TAINT_BAD_PAGE);
439 }
440
441 static inline int is_cow_mapping(unsigned int flags)
442 {
443         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
444 }
445
446 /*
447  * vm_normal_page -- This function gets the "struct page" associated with a pte.
448  *
449  * "Special" mappings do not wish to be associated with a "struct page" (either
450  * it doesn't exist, or it exists but they don't want to touch it). In this
451  * case, NULL is returned here. "Normal" mappings do have a struct page.
452  *
453  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
454  * pte bit, in which case this function is trivial. Secondly, an architecture
455  * may not have a spare pte bit, which requires a more complicated scheme,
456  * described below.
457  *
458  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
459  * special mapping (even if there are underlying and valid "struct pages").
460  * COWed pages of a VM_PFNMAP are always normal.
461  *
462  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
463  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
464  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
465  * mapping will always honor the rule
466  *
467  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
468  *
469  * And for normal mappings this is false.
470  *
471  * This restricts such mappings to be a linear translation from virtual address
472  * to pfn. To get around this restriction, we allow arbitrary mappings so long
473  * as the vma is not a COW mapping; in that case, we know that all ptes are
474  * special (because none can have been COWed).
475  *
476  *
477  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
478  *
479  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
480  * page" backing, however the difference is that _all_ pages with a struct
481  * page (that is, those where pfn_valid is true) are refcounted and considered
482  * normal pages by the VM. The disadvantage is that pages are refcounted
483  * (which can be slower and simply not an option for some PFNMAP users). The
484  * advantage is that we don't have to follow the strict linearity rule of
485  * PFNMAP mappings in order to support COWable mappings.
486  *
487  */
488 #ifdef __HAVE_ARCH_PTE_SPECIAL
489 # define HAVE_PTE_SPECIAL 1
490 #else
491 # define HAVE_PTE_SPECIAL 0
492 #endif
493 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
494                                 pte_t pte)
495 {
496         unsigned long pfn = pte_pfn(pte);
497
498         if (HAVE_PTE_SPECIAL) {
499                 if (likely(!pte_special(pte)))
500                         goto check_pfn;
501                 if (!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)))
502                         print_bad_pte(vma, addr, pte, NULL);
503                 return NULL;
504         }
505
506         /* !HAVE_PTE_SPECIAL case follows: */
507
508         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
509                 if (vma->vm_flags & VM_MIXEDMAP) {
510                         if (!pfn_valid(pfn))
511                                 return NULL;
512                         goto out;
513                 } else {
514                         unsigned long off;
515                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
516                         if (pfn == vma->vm_pgoff + off)
517                                 return NULL;
518                         if (!is_cow_mapping(vma->vm_flags))
519                                 return NULL;
520                 }
521         }
522
523 check_pfn:
524         if (unlikely(pfn > highest_memmap_pfn)) {
525                 print_bad_pte(vma, addr, pte, NULL);
526                 return NULL;
527         }
528
529         /*
530          * NOTE! We still have PageReserved() pages in the page tables.
531          * eg. VDSO mappings can cause them to exist.
532          */
533 out:
534         return pfn_to_page(pfn);
535 }
536
537 /*
538  * copy one vm_area from one task to the other. Assumes the page tables
539  * already present in the new task to be cleared in the whole range
540  * covered by this vma.
541  */
542
543 static inline void
544 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
545                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
546                 unsigned long addr, int *rss)
547 {
548         unsigned long vm_flags = vma->vm_flags;
549         pte_t pte = *src_pte;
550         struct page *page;
551
552         /* pte contains position in swap or file, so copy. */
553         if (unlikely(!pte_present(pte))) {
554                 if (!pte_file(pte)) {
555                         swp_entry_t entry = pte_to_swp_entry(pte);
556
557                         swap_duplicate(entry);
558                         /* make sure dst_mm is on swapoff's mmlist. */
559                         if (unlikely(list_empty(&dst_mm->mmlist))) {
560                                 spin_lock(&mmlist_lock);
561                                 if (list_empty(&dst_mm->mmlist))
562                                         list_add(&dst_mm->mmlist,
563                                                  &src_mm->mmlist);
564                                 spin_unlock(&mmlist_lock);
565                         }
566                         if (is_write_migration_entry(entry) &&
567                                         is_cow_mapping(vm_flags)) {
568                                 /*
569                                  * COW mappings require pages in both parent
570                                  * and child to be set to read.
571                                  */
572                                 make_migration_entry_read(&entry);
573                                 pte = swp_entry_to_pte(entry);
574                                 set_pte_at(src_mm, addr, src_pte, pte);
575                         }
576                 }
577                 goto out_set_pte;
578         }
579
580         /*
581          * If it's a COW mapping, write protect it both
582          * in the parent and the child
583          */
584         if (is_cow_mapping(vm_flags)) {
585                 ptep_set_wrprotect(src_mm, addr, src_pte);
586                 pte = pte_wrprotect(pte);
587         }
588
589         /*
590          * If it's a shared mapping, mark it clean in
591          * the child
592          */
593         if (vm_flags & VM_SHARED)
594                 pte = pte_mkclean(pte);
595         pte = pte_mkold(pte);
596
597         page = vm_normal_page(vma, addr, pte);
598         if (page) {
599                 get_page(page);
600                 page_dup_rmap(page, vma, addr);
601                 rss[!!PageAnon(page)]++;
602         }
603
604 out_set_pte:
605         set_pte_at(dst_mm, addr, dst_pte, pte);
606 }
607
608 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
609                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
610                 unsigned long addr, unsigned long end)
611 {
612         pte_t *src_pte, *dst_pte;
613         spinlock_t *src_ptl, *dst_ptl;
614         int progress = 0;
615         int rss[2];
616
617 again:
618         rss[1] = rss[0] = 0;
619         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
620         if (!dst_pte)
621                 return -ENOMEM;
622         src_pte = pte_offset_map_nested(src_pmd, addr);
623         src_ptl = pte_lockptr(src_mm, src_pmd);
624         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
625         arch_enter_lazy_mmu_mode();
626
627         do {
628                 /*
629                  * We are holding two locks at this point - either of them
630                  * could generate latencies in another task on another CPU.
631                  */
632                 if (progress >= 32) {
633                         progress = 0;
634                         if (need_resched() ||
635                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
636                                 break;
637                 }
638                 if (pte_none(*src_pte)) {
639                         progress++;
640                         continue;
641                 }
642                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
643                 progress += 8;
644         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
645
646         arch_leave_lazy_mmu_mode();
647         spin_unlock(src_ptl);
648         pte_unmap_nested(src_pte - 1);
649         add_mm_rss(dst_mm, rss[0], rss[1]);
650         pte_unmap_unlock(dst_pte - 1, dst_ptl);
651         cond_resched();
652         if (addr != end)
653                 goto again;
654         return 0;
655 }
656
657 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
658                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
659                 unsigned long addr, unsigned long end)
660 {
661         pmd_t *src_pmd, *dst_pmd;
662         unsigned long next;
663
664         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
665         if (!dst_pmd)
666                 return -ENOMEM;
667         src_pmd = pmd_offset(src_pud, addr);
668         do {
669                 next = pmd_addr_end(addr, end);
670                 if (pmd_none_or_clear_bad(src_pmd))
671                         continue;
672                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
673                                                 vma, addr, next))
674                         return -ENOMEM;
675         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
676         return 0;
677 }
678
679 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
680                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
681                 unsigned long addr, unsigned long end)
682 {
683         pud_t *src_pud, *dst_pud;
684         unsigned long next;
685
686         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
687         if (!dst_pud)
688                 return -ENOMEM;
689         src_pud = pud_offset(src_pgd, addr);
690         do {
691                 next = pud_addr_end(addr, end);
692                 if (pud_none_or_clear_bad(src_pud))
693                         continue;
694                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
695                                                 vma, addr, next))
696                         return -ENOMEM;
697         } while (dst_pud++, src_pud++, addr = next, addr != end);
698         return 0;
699 }
700
701 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
702                 struct vm_area_struct *vma)
703 {
704         pgd_t *src_pgd, *dst_pgd;
705         unsigned long next;
706         unsigned long addr = vma->vm_start;
707         unsigned long end = vma->vm_end;
708         int ret;
709
710         /*
711          * Don't copy ptes where a page fault will fill them correctly.
712          * Fork becomes much lighter when there are big shared or private
713          * readonly mappings. The tradeoff is that copy_page_range is more
714          * efficient than faulting.
715          */
716         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
717                 if (!vma->anon_vma)
718                         return 0;
719         }
720
721         if (is_vm_hugetlb_page(vma))
722                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
723
724         if (unlikely(is_pfn_mapping(vma))) {
725                 /*
726                  * We do not free on error cases below as remove_vma
727                  * gets called on error from higher level routine
728                  */
729                 ret = track_pfn_vma_copy(vma);
730                 if (ret)
731                         return ret;
732         }
733
734         /*
735          * We need to invalidate the secondary MMU mappings only when
736          * there could be a permission downgrade on the ptes of the
737          * parent mm. And a permission downgrade will only happen if
738          * is_cow_mapping() returns true.
739          */
740         if (is_cow_mapping(vma->vm_flags))
741                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
742
743         ret = 0;
744         dst_pgd = pgd_offset(dst_mm, addr);
745         src_pgd = pgd_offset(src_mm, addr);
746         do {
747                 next = pgd_addr_end(addr, end);
748                 if (pgd_none_or_clear_bad(src_pgd))
749                         continue;
750                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
751                                             vma, addr, next))) {
752                         ret = -ENOMEM;
753                         break;
754                 }
755         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
756
757         if (is_cow_mapping(vma->vm_flags))
758                 mmu_notifier_invalidate_range_end(src_mm,
759                                                   vma->vm_start, end);
760         return ret;
761 }
762
763 static unsigned long zap_pte_range(struct mmu_gather *tlb,
764                                 struct vm_area_struct *vma, pmd_t *pmd,
765                                 unsigned long addr, unsigned long end,
766                                 long *zap_work, struct zap_details *details)
767 {
768         struct mm_struct *mm = tlb->mm;
769         pte_t *pte;
770         spinlock_t *ptl;
771         int file_rss = 0;
772         int anon_rss = 0;
773
774         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
775         arch_enter_lazy_mmu_mode();
776         do {
777                 pte_t ptent = *pte;
778                 if (pte_none(ptent)) {
779                         (*zap_work)--;
780                         continue;
781                 }
782
783                 (*zap_work) -= PAGE_SIZE;
784
785                 if (pte_present(ptent)) {
786                         struct page *page;
787
788                         page = vm_normal_page(vma, addr, ptent);
789                         if (unlikely(details) && page) {
790                                 /*
791                                  * unmap_shared_mapping_pages() wants to
792                                  * invalidate cache without truncating:
793                                  * unmap shared but keep private pages.
794                                  */
795                                 if (details->check_mapping &&
796                                     details->check_mapping != page->mapping)
797                                         continue;
798                                 /*
799                                  * Each page->index must be checked when
800                                  * invalidating or truncating nonlinear.
801                                  */
802                                 if (details->nonlinear_vma &&
803                                     (page->index < details->first_index ||
804                                      page->index > details->last_index))
805                                         continue;
806                         }
807                         ptent = ptep_get_and_clear_full(mm, addr, pte,
808                                                         tlb->fullmm);
809                         tlb_remove_tlb_entry(tlb, pte, addr);
810                         if (unlikely(!page))
811                                 continue;
812                         if (unlikely(details) && details->nonlinear_vma
813                             && linear_page_index(details->nonlinear_vma,
814                                                 addr) != page->index)
815                                 set_pte_at(mm, addr, pte,
816                                            pgoff_to_pte(page->index));
817                         if (PageAnon(page))
818                                 anon_rss--;
819                         else {
820                                 if (pte_dirty(ptent))
821                                         set_page_dirty(page);
822                                 if (pte_young(ptent) &&
823                                     likely(!VM_SequentialReadHint(vma)))
824                                         mark_page_accessed(page);
825                                 file_rss--;
826                         }
827                         page_remove_rmap(page);
828                         if (unlikely(page_mapcount(page) < 0))
829                                 print_bad_pte(vma, addr, ptent, page);
830                         tlb_remove_page(tlb, page);
831                         continue;
832                 }
833                 /*
834                  * If details->check_mapping, we leave swap entries;
835                  * if details->nonlinear_vma, we leave file entries.
836                  */
837                 if (unlikely(details))
838                         continue;
839                 if (pte_file(ptent)) {
840                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
841                                 print_bad_pte(vma, addr, ptent, NULL);
842                 } else if
843                   (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
844                         print_bad_pte(vma, addr, ptent, NULL);
845                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
846         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
847
848         add_mm_rss(mm, file_rss, anon_rss);
849         arch_leave_lazy_mmu_mode();
850         pte_unmap_unlock(pte - 1, ptl);
851
852         return addr;
853 }
854
855 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
856                                 struct vm_area_struct *vma, pud_t *pud,
857                                 unsigned long addr, unsigned long end,
858                                 long *zap_work, struct zap_details *details)
859 {
860         pmd_t *pmd;
861         unsigned long next;
862
863         pmd = pmd_offset(pud, addr);
864         do {
865                 next = pmd_addr_end(addr, end);
866                 if (pmd_none_or_clear_bad(pmd)) {
867                         (*zap_work)--;
868                         continue;
869                 }
870                 next = zap_pte_range(tlb, vma, pmd, addr, next,
871                                                 zap_work, details);
872         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
873
874         return addr;
875 }
876
877 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
878                                 struct vm_area_struct *vma, pgd_t *pgd,
879                                 unsigned long addr, unsigned long end,
880                                 long *zap_work, struct zap_details *details)
881 {
882         pud_t *pud;
883         unsigned long next;
884
885         pud = pud_offset(pgd, addr);
886         do {
887                 next = pud_addr_end(addr, end);
888                 if (pud_none_or_clear_bad(pud)) {
889                         (*zap_work)--;
890                         continue;
891                 }
892                 next = zap_pmd_range(tlb, vma, pud, addr, next,
893                                                 zap_work, details);
894         } while (pud++, addr = next, (addr != end && *zap_work > 0));
895
896         return addr;
897 }
898
899 static unsigned long unmap_page_range(struct mmu_gather *tlb,
900                                 struct vm_area_struct *vma,
901                                 unsigned long addr, unsigned long end,
902                                 long *zap_work, struct zap_details *details)
903 {
904         pgd_t *pgd;
905         unsigned long next;
906
907         if (details && !details->check_mapping && !details->nonlinear_vma)
908                 details = NULL;
909
910         BUG_ON(addr >= end);
911         tlb_start_vma(tlb, vma);
912         pgd = pgd_offset(vma->vm_mm, addr);
913         do {
914                 next = pgd_addr_end(addr, end);
915                 if (pgd_none_or_clear_bad(pgd)) {
916                         (*zap_work)--;
917                         continue;
918                 }
919                 next = zap_pud_range(tlb, vma, pgd, addr, next,
920                                                 zap_work, details);
921         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
922         tlb_end_vma(tlb, vma);
923
924         return addr;
925 }
926
927 #ifdef CONFIG_PREEMPT
928 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
929 #else
930 /* No preempt: go for improved straight-line efficiency */
931 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
932 #endif
933
934 /**
935  * unmap_vmas - unmap a range of memory covered by a list of vma's
936  * @tlbp: address of the caller's struct mmu_gather
937  * @vma: the starting vma
938  * @start_addr: virtual address at which to start unmapping
939  * @end_addr: virtual address at which to end unmapping
940  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
941  * @details: details of nonlinear truncation or shared cache invalidation
942  *
943  * Returns the end address of the unmapping (restart addr if interrupted).
944  *
945  * Unmap all pages in the vma list.
946  *
947  * We aim to not hold locks for too long (for scheduling latency reasons).
948  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
949  * return the ending mmu_gather to the caller.
950  *
951  * Only addresses between `start' and `end' will be unmapped.
952  *
953  * The VMA list must be sorted in ascending virtual address order.
954  *
955  * unmap_vmas() assumes that the caller will flush the whole unmapped address
956  * range after unmap_vmas() returns.  So the only responsibility here is to
957  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
958  * drops the lock and schedules.
959  */
960 unsigned long unmap_vmas(struct mmu_gather **tlbp,
961                 struct vm_area_struct *vma, unsigned long start_addr,
962                 unsigned long end_addr, unsigned long *nr_accounted,
963                 struct zap_details *details)
964 {
965         long zap_work = ZAP_BLOCK_SIZE;
966         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
967         int tlb_start_valid = 0;
968         unsigned long start = start_addr;
969         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
970         int fullmm = (*tlbp)->fullmm;
971         struct mm_struct *mm = vma->vm_mm;
972
973         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
974         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
975                 unsigned long end;
976
977                 start = max(vma->vm_start, start_addr);
978                 if (start >= vma->vm_end)
979                         continue;
980                 end = min(vma->vm_end, end_addr);
981                 if (end <= vma->vm_start)
982                         continue;
983
984                 if (vma->vm_flags & VM_ACCOUNT)
985                         *nr_accounted += (end - start) >> PAGE_SHIFT;
986
987                 if (unlikely(is_pfn_mapping(vma)))
988                         untrack_pfn_vma(vma, 0, 0);
989
990                 while (start != end) {
991                         if (!tlb_start_valid) {
992                                 tlb_start = start;
993                                 tlb_start_valid = 1;
994                         }
995
996                         if (unlikely(is_vm_hugetlb_page(vma))) {
997                                 /*
998                                  * It is undesirable to test vma->vm_file as it
999                                  * should be non-null for valid hugetlb area.
1000                                  * However, vm_file will be NULL in the error
1001                                  * cleanup path of do_mmap_pgoff. When
1002                                  * hugetlbfs ->mmap method fails,
1003                                  * do_mmap_pgoff() nullifies vma->vm_file
1004                                  * before calling this function to clean up.
1005                                  * Since no pte has actually been setup, it is
1006                                  * safe to do nothing in this case.
1007                                  */
1008                                 if (vma->vm_file) {
1009                                         unmap_hugepage_range(vma, start, end, NULL);
1010                                         zap_work -= (end - start) /
1011                                         pages_per_huge_page(hstate_vma(vma));
1012                                 }
1013
1014                                 start = end;
1015                         } else
1016                                 start = unmap_page_range(*tlbp, vma,
1017                                                 start, end, &zap_work, details);
1018
1019                         if (zap_work > 0) {
1020                                 BUG_ON(start != end);
1021                                 break;
1022                         }
1023
1024                         tlb_finish_mmu(*tlbp, tlb_start, start);
1025
1026                         if (need_resched() ||
1027                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1028                                 if (i_mmap_lock) {
1029                                         *tlbp = NULL;
1030                                         goto out;
1031                                 }
1032                                 cond_resched();
1033                         }
1034
1035                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1036                         tlb_start_valid = 0;
1037                         zap_work = ZAP_BLOCK_SIZE;
1038                 }
1039         }
1040 out:
1041         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1042         return start;   /* which is now the end (or restart) address */
1043 }
1044
1045 /**
1046  * zap_page_range - remove user pages in a given range
1047  * @vma: vm_area_struct holding the applicable pages
1048  * @address: starting address of pages to zap
1049  * @size: number of bytes to zap
1050  * @details: details of nonlinear truncation or shared cache invalidation
1051  */
1052 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1053                 unsigned long size, struct zap_details *details)
1054 {
1055         struct mm_struct *mm = vma->vm_mm;
1056         struct mmu_gather *tlb;
1057         unsigned long end = address + size;
1058         unsigned long nr_accounted = 0;
1059
1060         lru_add_drain();
1061         tlb = tlb_gather_mmu(mm, 0);
1062         update_hiwater_rss(mm);
1063         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1064         if (tlb)
1065                 tlb_finish_mmu(tlb, address, end);
1066         return end;
1067 }
1068
1069 /**
1070  * zap_vma_ptes - remove ptes mapping the vma
1071  * @vma: vm_area_struct holding ptes to be zapped
1072  * @address: starting address of pages to zap
1073  * @size: number of bytes to zap
1074  *
1075  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1076  *
1077  * The entire address range must be fully contained within the vma.
1078  *
1079  * Returns 0 if successful.
1080  */
1081 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1082                 unsigned long size)
1083 {
1084         if (address < vma->vm_start || address + size > vma->vm_end ||
1085                         !(vma->vm_flags & VM_PFNMAP))
1086                 return -1;
1087         zap_page_range(vma, address, size, NULL);
1088         return 0;
1089 }
1090 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1091
1092 /*
1093  * Do a quick page-table lookup for a single page.
1094  */
1095 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1096                         unsigned int flags)
1097 {
1098         pgd_t *pgd;
1099         pud_t *pud;
1100         pmd_t *pmd;
1101         pte_t *ptep, pte;
1102         spinlock_t *ptl;
1103         struct page *page;
1104         struct mm_struct *mm = vma->vm_mm;
1105
1106         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1107         if (!IS_ERR(page)) {
1108                 BUG_ON(flags & FOLL_GET);
1109                 goto out;
1110         }
1111
1112         page = NULL;
1113         pgd = pgd_offset(mm, address);
1114         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1115                 goto no_page_table;
1116
1117         pud = pud_offset(pgd, address);
1118         if (pud_none(*pud))
1119                 goto no_page_table;
1120         if (pud_huge(*pud)) {
1121                 BUG_ON(flags & FOLL_GET);
1122                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1123                 goto out;
1124         }
1125         if (unlikely(pud_bad(*pud)))
1126                 goto no_page_table;
1127
1128         pmd = pmd_offset(pud, address);
1129         if (pmd_none(*pmd))
1130                 goto no_page_table;
1131         if (pmd_huge(*pmd)) {
1132                 BUG_ON(flags & FOLL_GET);
1133                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1134                 goto out;
1135         }
1136         if (unlikely(pmd_bad(*pmd)))
1137                 goto no_page_table;
1138
1139         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1140
1141         pte = *ptep;
1142         if (!pte_present(pte))
1143                 goto no_page;
1144         if ((flags & FOLL_WRITE) && !pte_write(pte))
1145                 goto unlock;
1146         page = vm_normal_page(vma, address, pte);
1147         if (unlikely(!page))
1148                 goto bad_page;
1149
1150         if (flags & FOLL_GET)
1151                 get_page(page);
1152         if (flags & FOLL_TOUCH) {
1153                 if ((flags & FOLL_WRITE) &&
1154                     !pte_dirty(pte) && !PageDirty(page))
1155                         set_page_dirty(page);
1156                 /*
1157                  * pte_mkyoung() would be more correct here, but atomic care
1158                  * is needed to avoid losing the dirty bit: it is easier to use
1159                  * mark_page_accessed().
1160                  */
1161                 mark_page_accessed(page);
1162         }
1163 unlock:
1164         pte_unmap_unlock(ptep, ptl);
1165 out:
1166         return page;
1167
1168 bad_page:
1169         pte_unmap_unlock(ptep, ptl);
1170         return ERR_PTR(-EFAULT);
1171
1172 no_page:
1173         pte_unmap_unlock(ptep, ptl);
1174         if (!pte_none(pte))
1175                 return page;
1176         /* Fall through to ZERO_PAGE handling */
1177 no_page_table:
1178         /*
1179          * When core dumping an enormous anonymous area that nobody
1180          * has touched so far, we don't want to allocate page tables.
1181          */
1182         if (flags & FOLL_ANON) {
1183                 page = ZERO_PAGE(0);
1184                 if (flags & FOLL_GET)
1185                         get_page(page);
1186                 BUG_ON(flags & FOLL_WRITE);
1187         }
1188         return page;
1189 }
1190
1191 /* Can we do the FOLL_ANON optimization? */
1192 static inline int use_zero_page(struct vm_area_struct *vma)
1193 {
1194         /*
1195          * We don't want to optimize FOLL_ANON for make_pages_present()
1196          * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1197          * we want to get the page from the page tables to make sure
1198          * that we serialize and update with any other user of that
1199          * mapping.
1200          */
1201         if (vma->vm_flags & (VM_LOCKED | VM_SHARED))
1202                 return 0;
1203         /*
1204          * And if we have a fault routine, it's not an anonymous region.
1205          */
1206         return !vma->vm_ops || !vma->vm_ops->fault;
1207 }
1208
1209
1210
1211 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1212                      unsigned long start, int nr_pages, int flags,
1213                      struct page **pages, struct vm_area_struct **vmas)
1214 {
1215         int i;
1216         unsigned int vm_flags = 0;
1217         int write = !!(flags & GUP_FLAGS_WRITE);
1218         int force = !!(flags & GUP_FLAGS_FORCE);
1219         int ignore = !!(flags & GUP_FLAGS_IGNORE_VMA_PERMISSIONS);
1220         int ignore_sigkill = !!(flags & GUP_FLAGS_IGNORE_SIGKILL);
1221
1222         if (nr_pages <= 0)
1223                 return 0;
1224         /* 
1225          * Require read or write permissions.
1226          * If 'force' is set, we only require the "MAY" flags.
1227          */
1228         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1229         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1230         i = 0;
1231
1232         do {
1233                 struct vm_area_struct *vma;
1234                 unsigned int foll_flags;
1235
1236                 vma = find_extend_vma(mm, start);
1237                 if (!vma && in_gate_area(tsk, start)) {
1238                         unsigned long pg = start & PAGE_MASK;
1239                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1240                         pgd_t *pgd;
1241                         pud_t *pud;
1242                         pmd_t *pmd;
1243                         pte_t *pte;
1244
1245                         /* user gate pages are read-only */
1246                         if (!ignore && write)
1247                                 return i ? : -EFAULT;
1248                         if (pg > TASK_SIZE)
1249                                 pgd = pgd_offset_k(pg);
1250                         else
1251                                 pgd = pgd_offset_gate(mm, pg);
1252                         BUG_ON(pgd_none(*pgd));
1253                         pud = pud_offset(pgd, pg);
1254                         BUG_ON(pud_none(*pud));
1255                         pmd = pmd_offset(pud, pg);
1256                         if (pmd_none(*pmd))
1257                                 return i ? : -EFAULT;
1258                         pte = pte_offset_map(pmd, pg);
1259                         if (pte_none(*pte)) {
1260                                 pte_unmap(pte);
1261                                 return i ? : -EFAULT;
1262                         }
1263                         if (pages) {
1264                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1265                                 pages[i] = page;
1266                                 if (page)
1267                                         get_page(page);
1268                         }
1269                         pte_unmap(pte);
1270                         if (vmas)
1271                                 vmas[i] = gate_vma;
1272                         i++;
1273                         start += PAGE_SIZE;
1274                         nr_pages--;
1275                         continue;
1276                 }
1277
1278                 if (!vma ||
1279                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1280                     (!ignore && !(vm_flags & vma->vm_flags)))
1281                         return i ? : -EFAULT;
1282
1283                 if (is_vm_hugetlb_page(vma)) {
1284                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1285                                                 &start, &nr_pages, i, write);
1286                         continue;
1287                 }
1288
1289                 foll_flags = FOLL_TOUCH;
1290                 if (pages)
1291                         foll_flags |= FOLL_GET;
1292                 if (!write && use_zero_page(vma))
1293                         foll_flags |= FOLL_ANON;
1294
1295                 do {
1296                         struct page *page;
1297
1298                         /*
1299                          * If we have a pending SIGKILL, don't keep faulting
1300                          * pages and potentially allocating memory, unless
1301                          * current is handling munlock--e.g., on exit. In
1302                          * that case, we are not allocating memory.  Rather,
1303                          * we're only unlocking already resident/mapped pages.
1304                          */
1305                         if (unlikely(!ignore_sigkill &&
1306                                         fatal_signal_pending(current)))
1307                                 return i ? i : -ERESTARTSYS;
1308
1309                         if (write)
1310                                 foll_flags |= FOLL_WRITE;
1311
1312                         cond_resched();
1313                         while (!(page = follow_page(vma, start, foll_flags))) {
1314                                 int ret;
1315
1316                                 ret = handle_mm_fault(mm, vma, start,
1317                                         (foll_flags & FOLL_WRITE) ?
1318                                         FAULT_FLAG_WRITE : 0);
1319
1320                                 if (ret & VM_FAULT_ERROR) {
1321                                         if (ret & VM_FAULT_OOM)
1322                                                 return i ? i : -ENOMEM;
1323                                         else if (ret & VM_FAULT_SIGBUS)
1324                                                 return i ? i : -EFAULT;
1325                                         BUG();
1326                                 }
1327                                 if (ret & VM_FAULT_MAJOR)
1328                                         tsk->maj_flt++;
1329                                 else
1330                                         tsk->min_flt++;
1331
1332                                 /*
1333                                  * The VM_FAULT_WRITE bit tells us that
1334                                  * do_wp_page has broken COW when necessary,
1335                                  * even if maybe_mkwrite decided not to set
1336                                  * pte_write. We can thus safely do subsequent
1337                                  * page lookups as if they were reads. But only
1338                                  * do so when looping for pte_write is futile:
1339                                  * in some cases userspace may also be wanting
1340                                  * to write to the gotten user page, which a
1341                                  * read fault here might prevent (a readonly
1342                                  * page might get reCOWed by userspace write).
1343                                  */
1344                                 if ((ret & VM_FAULT_WRITE) &&
1345                                     !(vma->vm_flags & VM_WRITE))
1346                                         foll_flags &= ~FOLL_WRITE;
1347
1348                                 cond_resched();
1349                         }
1350                         if (IS_ERR(page))
1351                                 return i ? i : PTR_ERR(page);
1352                         if (pages) {
1353                                 pages[i] = page;
1354
1355                                 flush_anon_page(vma, page, start);
1356                                 flush_dcache_page(page);
1357                         }
1358                         if (vmas)
1359                                 vmas[i] = vma;
1360                         i++;
1361                         start += PAGE_SIZE;
1362                         nr_pages--;
1363                 } while (nr_pages && start < vma->vm_end);
1364         } while (nr_pages);
1365         return i;
1366 }
1367
1368 /**
1369  * get_user_pages() - pin user pages in memory
1370  * @tsk:        task_struct of target task
1371  * @mm:         mm_struct of target mm
1372  * @start:      starting user address
1373  * @nr_pages:   number of pages from start to pin
1374  * @write:      whether pages will be written to by the caller
1375  * @force:      whether to force write access even if user mapping is
1376  *              readonly. This will result in the page being COWed even
1377  *              in MAP_SHARED mappings. You do not want this.
1378  * @pages:      array that receives pointers to the pages pinned.
1379  *              Should be at least nr_pages long. Or NULL, if caller
1380  *              only intends to ensure the pages are faulted in.
1381  * @vmas:       array of pointers to vmas corresponding to each page.
1382  *              Or NULL if the caller does not require them.
1383  *
1384  * Returns number of pages pinned. This may be fewer than the number
1385  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1386  * were pinned, returns -errno. Each page returned must be released
1387  * with a put_page() call when it is finished with. vmas will only
1388  * remain valid while mmap_sem is held.
1389  *
1390  * Must be called with mmap_sem held for read or write.
1391  *
1392  * get_user_pages walks a process's page tables and takes a reference to
1393  * each struct page that each user address corresponds to at a given
1394  * instant. That is, it takes the page that would be accessed if a user
1395  * thread accesses the given user virtual address at that instant.
1396  *
1397  * This does not guarantee that the page exists in the user mappings when
1398  * get_user_pages returns, and there may even be a completely different
1399  * page there in some cases (eg. if mmapped pagecache has been invalidated
1400  * and subsequently re faulted). However it does guarantee that the page
1401  * won't be freed completely. And mostly callers simply care that the page
1402  * contains data that was valid *at some point in time*. Typically, an IO
1403  * or similar operation cannot guarantee anything stronger anyway because
1404  * locks can't be held over the syscall boundary.
1405  *
1406  * If write=0, the page must not be written to. If the page is written to,
1407  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1408  * after the page is finished with, and before put_page is called.
1409  *
1410  * get_user_pages is typically used for fewer-copy IO operations, to get a
1411  * handle on the memory by some means other than accesses via the user virtual
1412  * addresses. The pages may be submitted for DMA to devices or accessed via
1413  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1414  * use the correct cache flushing APIs.
1415  *
1416  * See also get_user_pages_fast, for performance critical applications.
1417  */
1418 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1419                 unsigned long start, int nr_pages, int write, int force,
1420                 struct page **pages, struct vm_area_struct **vmas)
1421 {
1422         int flags = 0;
1423
1424         if (write)
1425                 flags |= GUP_FLAGS_WRITE;
1426         if (force)
1427                 flags |= GUP_FLAGS_FORCE;
1428
1429         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
1430 }
1431
1432 EXPORT_SYMBOL(get_user_pages);
1433
1434 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1435                         spinlock_t **ptl)
1436 {
1437         pgd_t * pgd = pgd_offset(mm, addr);
1438         pud_t * pud = pud_alloc(mm, pgd, addr);
1439         if (pud) {
1440                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1441                 if (pmd)
1442                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1443         }
1444         return NULL;
1445 }
1446
1447 /*
1448  * This is the old fallback for page remapping.
1449  *
1450  * For historical reasons, it only allows reserved pages. Only
1451  * old drivers should use this, and they needed to mark their
1452  * pages reserved for the old functions anyway.
1453  */
1454 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1455                         struct page *page, pgprot_t prot)
1456 {
1457         struct mm_struct *mm = vma->vm_mm;
1458         int retval;
1459         pte_t *pte;
1460         spinlock_t *ptl;
1461
1462         retval = -EINVAL;
1463         if (PageAnon(page))
1464                 goto out;
1465         retval = -ENOMEM;
1466         flush_dcache_page(page);
1467         pte = get_locked_pte(mm, addr, &ptl);
1468         if (!pte)
1469                 goto out;
1470         retval = -EBUSY;
1471         if (!pte_none(*pte))
1472                 goto out_unlock;
1473
1474         /* Ok, finally just insert the thing.. */
1475         get_page(page);
1476         inc_mm_counter(mm, file_rss);
1477         page_add_file_rmap(page);
1478         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1479
1480         retval = 0;
1481         pte_unmap_unlock(pte, ptl);
1482         return retval;
1483 out_unlock:
1484         pte_unmap_unlock(pte, ptl);
1485 out:
1486         return retval;
1487 }
1488
1489 /**
1490  * vm_insert_page - insert single page into user vma
1491  * @vma: user vma to map to
1492  * @addr: target user address of this page
1493  * @page: source kernel page
1494  *
1495  * This allows drivers to insert individual pages they've allocated
1496  * into a user vma.
1497  *
1498  * The page has to be a nice clean _individual_ kernel allocation.
1499  * If you allocate a compound page, you need to have marked it as
1500  * such (__GFP_COMP), or manually just split the page up yourself
1501  * (see split_page()).
1502  *
1503  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1504  * took an arbitrary page protection parameter. This doesn't allow
1505  * that. Your vma protection will have to be set up correctly, which
1506  * means that if you want a shared writable mapping, you'd better
1507  * ask for a shared writable mapping!
1508  *
1509  * The page does not need to be reserved.
1510  */
1511 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1512                         struct page *page)
1513 {
1514         if (addr < vma->vm_start || addr >= vma->vm_end)
1515                 return -EFAULT;
1516         if (!page_count(page))
1517                 return -EINVAL;
1518         vma->vm_flags |= VM_INSERTPAGE;
1519         return insert_page(vma, addr, page, vma->vm_page_prot);
1520 }
1521 EXPORT_SYMBOL(vm_insert_page);
1522
1523 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1524                         unsigned long pfn, pgprot_t prot)
1525 {
1526         struct mm_struct *mm = vma->vm_mm;
1527         int retval;
1528         pte_t *pte, entry;
1529         spinlock_t *ptl;
1530
1531         retval = -ENOMEM;
1532         pte = get_locked_pte(mm, addr, &ptl);
1533         if (!pte)
1534                 goto out;
1535         retval = -EBUSY;
1536         if (!pte_none(*pte))
1537                 goto out_unlock;
1538
1539         /* Ok, finally just insert the thing.. */
1540         entry = pte_mkspecial(pfn_pte(pfn, prot));
1541         set_pte_at(mm, addr, pte, entry);
1542         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1543
1544         retval = 0;
1545 out_unlock:
1546         pte_unmap_unlock(pte, ptl);
1547 out:
1548         return retval;
1549 }
1550
1551 /**
1552  * vm_insert_pfn - insert single pfn into user vma
1553  * @vma: user vma to map to
1554  * @addr: target user address of this page
1555  * @pfn: source kernel pfn
1556  *
1557  * Similar to vm_inert_page, this allows drivers to insert individual pages
1558  * they've allocated into a user vma. Same comments apply.
1559  *
1560  * This function should only be called from a vm_ops->fault handler, and
1561  * in that case the handler should return NULL.
1562  *
1563  * vma cannot be a COW mapping.
1564  *
1565  * As this is called only for pages that do not currently exist, we
1566  * do not need to flush old virtual caches or the TLB.
1567  */
1568 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1569                         unsigned long pfn)
1570 {
1571         int ret;
1572         pgprot_t pgprot = vma->vm_page_prot;
1573         /*
1574          * Technically, architectures with pte_special can avoid all these
1575          * restrictions (same for remap_pfn_range).  However we would like
1576          * consistency in testing and feature parity among all, so we should
1577          * try to keep these invariants in place for everybody.
1578          */
1579         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1580         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1581                                                 (VM_PFNMAP|VM_MIXEDMAP));
1582         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1583         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1584
1585         if (addr < vma->vm_start || addr >= vma->vm_end)
1586                 return -EFAULT;
1587         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1588                 return -EINVAL;
1589
1590         ret = insert_pfn(vma, addr, pfn, pgprot);
1591
1592         if (ret)
1593                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1594
1595         return ret;
1596 }
1597 EXPORT_SYMBOL(vm_insert_pfn);
1598
1599 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1600                         unsigned long pfn)
1601 {
1602         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1603
1604         if (addr < vma->vm_start || addr >= vma->vm_end)
1605                 return -EFAULT;
1606
1607         /*
1608          * If we don't have pte special, then we have to use the pfn_valid()
1609          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1610          * refcount the page if pfn_valid is true (hence insert_page rather
1611          * than insert_pfn).
1612          */
1613         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1614                 struct page *page;
1615
1616                 page = pfn_to_page(pfn);
1617                 return insert_page(vma, addr, page, vma->vm_page_prot);
1618         }
1619         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1620 }
1621 EXPORT_SYMBOL(vm_insert_mixed);
1622
1623 /*
1624  * maps a range of physical memory into the requested pages. the old
1625  * mappings are removed. any references to nonexistent pages results
1626  * in null mappings (currently treated as "copy-on-access")
1627  */
1628 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1629                         unsigned long addr, unsigned long end,
1630                         unsigned long pfn, pgprot_t prot)
1631 {
1632         pte_t *pte;
1633         spinlock_t *ptl;
1634
1635         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1636         if (!pte)
1637                 return -ENOMEM;
1638         arch_enter_lazy_mmu_mode();
1639         do {
1640                 BUG_ON(!pte_none(*pte));
1641                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1642                 pfn++;
1643         } while (pte++, addr += PAGE_SIZE, addr != end);
1644         arch_leave_lazy_mmu_mode();
1645         pte_unmap_unlock(pte - 1, ptl);
1646         return 0;
1647 }
1648
1649 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1650                         unsigned long addr, unsigned long end,
1651                         unsigned long pfn, pgprot_t prot)
1652 {
1653         pmd_t *pmd;
1654         unsigned long next;
1655
1656         pfn -= addr >> PAGE_SHIFT;
1657         pmd = pmd_alloc(mm, pud, addr);
1658         if (!pmd)
1659                 return -ENOMEM;
1660         do {
1661                 next = pmd_addr_end(addr, end);
1662                 if (remap_pte_range(mm, pmd, addr, next,
1663                                 pfn + (addr >> PAGE_SHIFT), prot))
1664                         return -ENOMEM;
1665         } while (pmd++, addr = next, addr != end);
1666         return 0;
1667 }
1668
1669 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1670                         unsigned long addr, unsigned long end,
1671                         unsigned long pfn, pgprot_t prot)
1672 {
1673         pud_t *pud;
1674         unsigned long next;
1675
1676         pfn -= addr >> PAGE_SHIFT;
1677         pud = pud_alloc(mm, pgd, addr);
1678         if (!pud)
1679                 return -ENOMEM;
1680         do {
1681                 next = pud_addr_end(addr, end);
1682                 if (remap_pmd_range(mm, pud, addr, next,
1683                                 pfn + (addr >> PAGE_SHIFT), prot))
1684                         return -ENOMEM;
1685         } while (pud++, addr = next, addr != end);
1686         return 0;
1687 }
1688
1689 /**
1690  * remap_pfn_range - remap kernel memory to userspace
1691  * @vma: user vma to map to
1692  * @addr: target user address to start at
1693  * @pfn: physical address of kernel memory
1694  * @size: size of map area
1695  * @prot: page protection flags for this mapping
1696  *
1697  *  Note: this is only safe if the mm semaphore is held when called.
1698  */
1699 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1700                     unsigned long pfn, unsigned long size, pgprot_t prot)
1701 {
1702         pgd_t *pgd;
1703         unsigned long next;
1704         unsigned long end = addr + PAGE_ALIGN(size);
1705         struct mm_struct *mm = vma->vm_mm;
1706         int err;
1707
1708         /*
1709          * Physically remapped pages are special. Tell the
1710          * rest of the world about it:
1711          *   VM_IO tells people not to look at these pages
1712          *      (accesses can have side effects).
1713          *   VM_RESERVED is specified all over the place, because
1714          *      in 2.4 it kept swapout's vma scan off this vma; but
1715          *      in 2.6 the LRU scan won't even find its pages, so this
1716          *      flag means no more than count its pages in reserved_vm,
1717          *      and omit it from core dump, even when VM_IO turned off.
1718          *   VM_PFNMAP tells the core MM that the base pages are just
1719          *      raw PFN mappings, and do not have a "struct page" associated
1720          *      with them.
1721          *
1722          * There's a horrible special case to handle copy-on-write
1723          * behaviour that some programs depend on. We mark the "original"
1724          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1725          */
1726         if (addr == vma->vm_start && end == vma->vm_end) {
1727                 vma->vm_pgoff = pfn;
1728                 vma->vm_flags |= VM_PFN_AT_MMAP;
1729         } else if (is_cow_mapping(vma->vm_flags))
1730                 return -EINVAL;
1731
1732         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1733
1734         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1735         if (err) {
1736                 /*
1737                  * To indicate that track_pfn related cleanup is not
1738                  * needed from higher level routine calling unmap_vmas
1739                  */
1740                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1741                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1742                 return -EINVAL;
1743         }
1744
1745         BUG_ON(addr >= end);
1746         pfn -= addr >> PAGE_SHIFT;
1747         pgd = pgd_offset(mm, addr);
1748         flush_cache_range(vma, addr, end);
1749         do {
1750                 next = pgd_addr_end(addr, end);
1751                 err = remap_pud_range(mm, pgd, addr, next,
1752                                 pfn + (addr >> PAGE_SHIFT), prot);
1753                 if (err)
1754                         break;
1755         } while (pgd++, addr = next, addr != end);
1756
1757         if (err)
1758                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1759
1760         return err;
1761 }
1762 EXPORT_SYMBOL(remap_pfn_range);
1763
1764 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1765                                      unsigned long addr, unsigned long end,
1766                                      pte_fn_t fn, void *data)
1767 {
1768         pte_t *pte;
1769         int err;
1770         pgtable_t token;
1771         spinlock_t *uninitialized_var(ptl);
1772
1773         pte = (mm == &init_mm) ?
1774                 pte_alloc_kernel(pmd, addr) :
1775                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1776         if (!pte)
1777                 return -ENOMEM;
1778
1779         BUG_ON(pmd_huge(*pmd));
1780
1781         arch_enter_lazy_mmu_mode();
1782
1783         token = pmd_pgtable(*pmd);
1784
1785         do {
1786                 err = fn(pte, token, addr, data);
1787                 if (err)
1788                         break;
1789         } while (pte++, addr += PAGE_SIZE, addr != end);
1790
1791         arch_leave_lazy_mmu_mode();
1792
1793         if (mm != &init_mm)
1794                 pte_unmap_unlock(pte-1, ptl);
1795         return err;
1796 }
1797
1798 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1799                                      unsigned long addr, unsigned long end,
1800                                      pte_fn_t fn, void *data)
1801 {
1802         pmd_t *pmd;
1803         unsigned long next;
1804         int err;
1805
1806         BUG_ON(pud_huge(*pud));
1807
1808         pmd = pmd_alloc(mm, pud, addr);
1809         if (!pmd)
1810                 return -ENOMEM;
1811         do {
1812                 next = pmd_addr_end(addr, end);
1813                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1814                 if (err)
1815                         break;
1816         } while (pmd++, addr = next, addr != end);
1817         return err;
1818 }
1819
1820 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1821                                      unsigned long addr, unsigned long end,
1822                                      pte_fn_t fn, void *data)
1823 {
1824         pud_t *pud;
1825         unsigned long next;
1826         int err;
1827
1828         pud = pud_alloc(mm, pgd, addr);
1829         if (!pud)
1830                 return -ENOMEM;
1831         do {
1832                 next = pud_addr_end(addr, end);
1833                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1834                 if (err)
1835                         break;
1836         } while (pud++, addr = next, addr != end);
1837         return err;
1838 }
1839
1840 /*
1841  * Scan a region of virtual memory, filling in page tables as necessary
1842  * and calling a provided function on each leaf page table.
1843  */
1844 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1845                         unsigned long size, pte_fn_t fn, void *data)
1846 {
1847         pgd_t *pgd;
1848         unsigned long next;
1849         unsigned long start = addr, end = addr + size;
1850         int err;
1851
1852         BUG_ON(addr >= end);
1853         mmu_notifier_invalidate_range_start(mm, start, end);
1854         pgd = pgd_offset(mm, addr);
1855         do {
1856                 next = pgd_addr_end(addr, end);
1857                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1858                 if (err)
1859                         break;
1860         } while (pgd++, addr = next, addr != end);
1861         mmu_notifier_invalidate_range_end(mm, start, end);
1862         return err;
1863 }
1864 EXPORT_SYMBOL_GPL(apply_to_page_range);
1865
1866 /*
1867  * handle_pte_fault chooses page fault handler according to an entry
1868  * which was read non-atomically.  Before making any commitment, on
1869  * those architectures or configurations (e.g. i386 with PAE) which
1870  * might give a mix of unmatched parts, do_swap_page and do_file_page
1871  * must check under lock before unmapping the pte and proceeding
1872  * (but do_wp_page is only called after already making such a check;
1873  * and do_anonymous_page and do_no_page can safely check later on).
1874  */
1875 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1876                                 pte_t *page_table, pte_t orig_pte)
1877 {
1878         int same = 1;
1879 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1880         if (sizeof(pte_t) > sizeof(unsigned long)) {
1881                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1882                 spin_lock(ptl);
1883                 same = pte_same(*page_table, orig_pte);
1884                 spin_unlock(ptl);
1885         }
1886 #endif
1887         pte_unmap(page_table);
1888         return same;
1889 }
1890
1891 /*
1892  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1893  * servicing faults for write access.  In the normal case, do always want
1894  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1895  * that do not have writing enabled, when used by access_process_vm.
1896  */
1897 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1898 {
1899         if (likely(vma->vm_flags & VM_WRITE))
1900                 pte = pte_mkwrite(pte);
1901         return pte;
1902 }
1903
1904 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1905 {
1906         /*
1907          * If the source page was a PFN mapping, we don't have
1908          * a "struct page" for it. We do a best-effort copy by
1909          * just copying from the original user address. If that
1910          * fails, we just zero-fill it. Live with it.
1911          */
1912         if (unlikely(!src)) {
1913                 void *kaddr = kmap_atomic(dst, KM_USER0);
1914                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1915
1916                 /*
1917                  * This really shouldn't fail, because the page is there
1918                  * in the page tables. But it might just be unreadable,
1919                  * in which case we just give up and fill the result with
1920                  * zeroes.
1921                  */
1922                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1923                         memset(kaddr, 0, PAGE_SIZE);
1924                 kunmap_atomic(kaddr, KM_USER0);
1925                 flush_dcache_page(dst);
1926         } else
1927                 copy_user_highpage(dst, src, va, vma);
1928 }
1929
1930 /*
1931  * This routine handles present pages, when users try to write
1932  * to a shared page. It is done by copying the page to a new address
1933  * and decrementing the shared-page counter for the old page.
1934  *
1935  * Note that this routine assumes that the protection checks have been
1936  * done by the caller (the low-level page fault routine in most cases).
1937  * Thus we can safely just mark it writable once we've done any necessary
1938  * COW.
1939  *
1940  * We also mark the page dirty at this point even though the page will
1941  * change only once the write actually happens. This avoids a few races,
1942  * and potentially makes it more efficient.
1943  *
1944  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1945  * but allow concurrent faults), with pte both mapped and locked.
1946  * We return with mmap_sem still held, but pte unmapped and unlocked.
1947  */
1948 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1949                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1950                 spinlock_t *ptl, pte_t orig_pte)
1951 {
1952         struct page *old_page, *new_page;
1953         pte_t entry;
1954         int reuse = 0, ret = 0;
1955         int page_mkwrite = 0;
1956         struct page *dirty_page = NULL;
1957
1958         old_page = vm_normal_page(vma, address, orig_pte);
1959         if (!old_page) {
1960                 /*
1961                  * VM_MIXEDMAP !pfn_valid() case
1962                  *
1963                  * We should not cow pages in a shared writeable mapping.
1964                  * Just mark the pages writable as we can't do any dirty
1965                  * accounting on raw pfn maps.
1966                  */
1967                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1968                                      (VM_WRITE|VM_SHARED))
1969                         goto reuse;
1970                 goto gotten;
1971         }
1972
1973         /*
1974          * Take out anonymous pages first, anonymous shared vmas are
1975          * not dirty accountable.
1976          */
1977         if (PageAnon(old_page)) {
1978                 if (!trylock_page(old_page)) {
1979                         page_cache_get(old_page);
1980                         pte_unmap_unlock(page_table, ptl);
1981                         lock_page(old_page);
1982                         page_table = pte_offset_map_lock(mm, pmd, address,
1983                                                          &ptl);
1984                         if (!pte_same(*page_table, orig_pte)) {
1985                                 unlock_page(old_page);
1986                                 page_cache_release(old_page);
1987                                 goto unlock;
1988                         }
1989                         page_cache_release(old_page);
1990                 }
1991                 reuse = reuse_swap_page(old_page);
1992                 unlock_page(old_page);
1993         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1994                                         (VM_WRITE|VM_SHARED))) {
1995                 /*
1996                  * Only catch write-faults on shared writable pages,
1997                  * read-only shared pages can get COWed by
1998                  * get_user_pages(.write=1, .force=1).
1999                  */
2000                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2001                         struct vm_fault vmf;
2002                         int tmp;
2003
2004                         vmf.virtual_address = (void __user *)(address &
2005                                                                 PAGE_MASK);
2006                         vmf.pgoff = old_page->index;
2007                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2008                         vmf.page = old_page;
2009
2010                         /*
2011                          * Notify the address space that the page is about to
2012                          * become writable so that it can prohibit this or wait
2013                          * for the page to get into an appropriate state.
2014                          *
2015                          * We do this without the lock held, so that it can
2016                          * sleep if it needs to.
2017                          */
2018                         page_cache_get(old_page);
2019                         pte_unmap_unlock(page_table, ptl);
2020
2021                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2022                         if (unlikely(tmp &
2023                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2024                                 ret = tmp;
2025                                 goto unwritable_page;
2026                         }
2027                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2028                                 lock_page(old_page);
2029                                 if (!old_page->mapping) {
2030                                         ret = 0; /* retry the fault */
2031                                         unlock_page(old_page);
2032                                         goto unwritable_page;
2033                                 }
2034                         } else
2035                                 VM_BUG_ON(!PageLocked(old_page));
2036
2037                         /*
2038                          * Since we dropped the lock we need to revalidate
2039                          * the PTE as someone else may have changed it.  If
2040                          * they did, we just return, as we can count on the
2041                          * MMU to tell us if they didn't also make it writable.
2042                          */
2043                         page_table = pte_offset_map_lock(mm, pmd, address,
2044                                                          &ptl);
2045                         if (!pte_same(*page_table, orig_pte)) {
2046                                 unlock_page(old_page);
2047                                 page_cache_release(old_page);
2048                                 goto unlock;
2049                         }
2050
2051                         page_mkwrite = 1;
2052                 }
2053                 dirty_page = old_page;
2054                 get_page(dirty_page);
2055                 reuse = 1;
2056         }
2057
2058         if (reuse) {
2059 reuse:
2060                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2061                 entry = pte_mkyoung(orig_pte);
2062                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2063                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2064                         update_mmu_cache(vma, address, entry);
2065                 ret |= VM_FAULT_WRITE;
2066                 goto unlock;
2067         }
2068
2069         /*
2070          * Ok, we need to copy. Oh, well..
2071          */
2072         page_cache_get(old_page);
2073 gotten:
2074         pte_unmap_unlock(page_table, ptl);
2075
2076         if (unlikely(anon_vma_prepare(vma)))
2077                 goto oom;
2078         VM_BUG_ON(old_page == ZERO_PAGE(0));
2079         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2080         if (!new_page)
2081                 goto oom;
2082         /*
2083          * Don't let another task, with possibly unlocked vma,
2084          * keep the mlocked page.
2085          */
2086         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2087                 lock_page(old_page);    /* for LRU manipulation */
2088                 clear_page_mlock(old_page);
2089                 unlock_page(old_page);
2090         }
2091         cow_user_page(new_page, old_page, address, vma);
2092         __SetPageUptodate(new_page);
2093
2094         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2095                 goto oom_free_new;
2096
2097         /*
2098          * Re-check the pte - we dropped the lock
2099          */
2100         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2101         if (likely(pte_same(*page_table, orig_pte))) {
2102                 if (old_page) {
2103                         if (!PageAnon(old_page)) {
2104                                 dec_mm_counter(mm, file_rss);
2105                                 inc_mm_counter(mm, anon_rss);
2106                         }
2107                 } else
2108                         inc_mm_counter(mm, anon_rss);
2109                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2110                 entry = mk_pte(new_page, vma->vm_page_prot);
2111                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2112                 /*
2113                  * Clear the pte entry and flush it first, before updating the
2114                  * pte with the new entry. This will avoid a race condition
2115                  * seen in the presence of one thread doing SMC and another
2116                  * thread doing COW.
2117                  */
2118                 ptep_clear_flush(vma, address, page_table);
2119                 page_add_new_anon_rmap(new_page, vma, address);
2120                 /*
2121                  * We call the notify macro here because, when using secondary
2122                  * mmu page tables (such as kvm shadow page tables), we want the
2123                  * new page to be mapped directly into the secondary page table.
2124                  */
2125                 set_pte_at_notify(mm, address, page_table, entry);
2126                 update_mmu_cache(vma, address, entry);
2127                 if (old_page) {
2128                         /*
2129                          * Only after switching the pte to the new page may
2130                          * we remove the mapcount here. Otherwise another
2131                          * process may come and find the rmap count decremented
2132                          * before the pte is switched to the new page, and
2133                          * "reuse" the old page writing into it while our pte
2134                          * here still points into it and can be read by other
2135                          * threads.
2136                          *
2137                          * The critical issue is to order this
2138                          * page_remove_rmap with the ptp_clear_flush above.
2139                          * Those stores are ordered by (if nothing else,)
2140                          * the barrier present in the atomic_add_negative
2141                          * in page_remove_rmap.
2142                          *
2143                          * Then the TLB flush in ptep_clear_flush ensures that
2144                          * no process can access the old page before the
2145                          * decremented mapcount is visible. And the old page
2146                          * cannot be reused until after the decremented
2147                          * mapcount is visible. So transitively, TLBs to
2148                          * old page will be flushed before it can be reused.
2149                          */
2150                         page_remove_rmap(old_page);
2151                 }
2152
2153                 /* Free the old page.. */
2154                 new_page = old_page;
2155                 ret |= VM_FAULT_WRITE;
2156         } else
2157                 mem_cgroup_uncharge_page(new_page);
2158
2159         if (new_page)
2160                 page_cache_release(new_page);
2161         if (old_page)
2162                 page_cache_release(old_page);
2163 unlock:
2164         pte_unmap_unlock(page_table, ptl);
2165         if (dirty_page) {
2166                 /*
2167                  * Yes, Virginia, this is actually required to prevent a race
2168                  * with clear_page_dirty_for_io() from clearing the page dirty
2169                  * bit after it clear all dirty ptes, but before a racing
2170                  * do_wp_page installs a dirty pte.
2171                  *
2172                  * do_no_page is protected similarly.
2173                  */
2174                 if (!page_mkwrite) {
2175                         wait_on_page_locked(dirty_page);
2176                         set_page_dirty_balance(dirty_page, page_mkwrite);
2177                 }
2178                 put_page(dirty_page);
2179                 if (page_mkwrite) {
2180                         struct address_space *mapping = dirty_page->mapping;
2181
2182                         set_page_dirty(dirty_page);
2183                         unlock_page(dirty_page);
2184                         page_cache_release(dirty_page);
2185                         if (mapping)    {
2186                                 /*
2187                                  * Some device drivers do not set page.mapping
2188                                  * but still dirty their pages
2189                                  */
2190                                 balance_dirty_pages_ratelimited(mapping);
2191                         }
2192                 }
2193
2194                 /* file_update_time outside page_lock */
2195                 if (vma->vm_file)
2196                         file_update_time(vma->vm_file);
2197         }
2198         return ret;
2199 oom_free_new:
2200         page_cache_release(new_page);
2201 oom:
2202         if (old_page) {
2203                 if (page_mkwrite) {
2204                         unlock_page(old_page);
2205                         page_cache_release(old_page);
2206                 }
2207                 page_cache_release(old_page);
2208         }
2209         return VM_FAULT_OOM;
2210
2211 unwritable_page:
2212         page_cache_release(old_page);
2213         return ret;
2214 }
2215
2216 /*
2217  * Helper functions for unmap_mapping_range().
2218  *
2219  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2220  *
2221  * We have to restart searching the prio_tree whenever we drop the lock,
2222  * since the iterator is only valid while the lock is held, and anyway
2223  * a later vma might be split and reinserted earlier while lock dropped.
2224  *
2225  * The list of nonlinear vmas could be handled more efficiently, using
2226  * a placeholder, but handle it in the same way until a need is shown.
2227  * It is important to search the prio_tree before nonlinear list: a vma
2228  * may become nonlinear and be shifted from prio_tree to nonlinear list
2229  * while the lock is dropped; but never shifted from list to prio_tree.
2230  *
2231  * In order to make forward progress despite restarting the search,
2232  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2233  * quickly skip it next time around.  Since the prio_tree search only
2234  * shows us those vmas affected by unmapping the range in question, we
2235  * can't efficiently keep all vmas in step with mapping->truncate_count:
2236  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2237  * mapping->truncate_count and vma->vm_truncate_count are protected by
2238  * i_mmap_lock.
2239  *
2240  * In order to make forward progress despite repeatedly restarting some
2241  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2242  * and restart from that address when we reach that vma again.  It might
2243  * have been split or merged, shrunk or extended, but never shifted: so
2244  * restart_addr remains valid so long as it remains in the vma's range.
2245  * unmap_mapping_range forces truncate_count to leap over page-aligned
2246  * values so we can save vma's restart_addr in its truncate_count field.
2247  */
2248 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2249
2250 static void reset_vma_truncate_counts(struct address_space *mapping)
2251 {
2252         struct vm_area_struct *vma;
2253         struct prio_tree_iter iter;
2254
2255         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2256                 vma->vm_truncate_count = 0;
2257         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2258                 vma->vm_truncate_count = 0;
2259 }
2260
2261 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2262                 unsigned long start_addr, unsigned long end_addr,
2263                 struct zap_details *details)
2264 {
2265         unsigned long restart_addr;
2266         int need_break;
2267
2268         /*
2269          * files that support invalidating or truncating portions of the
2270          * file from under mmaped areas must have their ->fault function
2271          * return a locked page (and set VM_FAULT_LOCKED in the return).
2272          * This provides synchronisation against concurrent unmapping here.
2273          */
2274
2275 again:
2276         restart_addr = vma->vm_truncate_count;
2277         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2278                 start_addr = restart_addr;
2279                 if (start_addr >= end_addr) {
2280                         /* Top of vma has been split off since last time */
2281                         vma->vm_truncate_count = details->truncate_count;
2282                         return 0;
2283                 }
2284         }
2285
2286         restart_addr = zap_page_range(vma, start_addr,
2287                                         end_addr - start_addr, details);
2288         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2289
2290         if (restart_addr >= end_addr) {
2291                 /* We have now completed this vma: mark it so */
2292                 vma->vm_truncate_count = details->truncate_count;
2293                 if (!need_break)
2294                         return 0;
2295         } else {
2296                 /* Note restart_addr in vma's truncate_count field */
2297                 vma->vm_truncate_count = restart_addr;
2298                 if (!need_break)
2299                         goto again;
2300         }
2301
2302         spin_unlock(details->i_mmap_lock);
2303         cond_resched();
2304         spin_lock(details->i_mmap_lock);
2305         return -EINTR;
2306 }
2307
2308 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2309                                             struct zap_details *details)
2310 {
2311         struct vm_area_struct *vma;
2312         struct prio_tree_iter iter;
2313         pgoff_t vba, vea, zba, zea;
2314
2315 restart:
2316         vma_prio_tree_foreach(vma, &iter, root,
2317                         details->first_index, details->last_index) {
2318                 /* Skip quickly over those we have already dealt with */
2319                 if (vma->vm_truncate_count == details->truncate_count)
2320                         continue;
2321
2322                 vba = vma->vm_pgoff;
2323                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2324                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2325                 zba = details->first_index;
2326                 if (zba < vba)
2327                         zba = vba;
2328                 zea = details->last_index;
2329                 if (zea > vea)
2330                         zea = vea;
2331
2332                 if (unmap_mapping_range_vma(vma,
2333                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2334                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2335                                 details) < 0)
2336                         goto restart;
2337         }
2338 }
2339
2340 static inline void unmap_mapping_range_list(struct list_head *head,
2341                                             struct zap_details *details)
2342 {
2343         struct vm_area_struct *vma;
2344
2345         /*
2346          * In nonlinear VMAs there is no correspondence between virtual address
2347          * offset and file offset.  So we must perform an exhaustive search
2348          * across *all* the pages in each nonlinear VMA, not just the pages
2349          * whose virtual address lies outside the file truncation point.
2350          */
2351 restart:
2352         list_for_each_entry(vma, head, shared.vm_set.list) {
2353                 /* Skip quickly over those we have already dealt with */
2354                 if (vma->vm_truncate_count == details->truncate_count)
2355                         continue;
2356                 details->nonlinear_vma = vma;
2357                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2358                                         vma->vm_end, details) < 0)
2359                         goto restart;
2360         }
2361 }
2362
2363 /**
2364  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2365  * @mapping: the address space containing mmaps to be unmapped.
2366  * @holebegin: byte in first page to unmap, relative to the start of
2367  * the underlying file.  This will be rounded down to a PAGE_SIZE
2368  * boundary.  Note that this is different from vmtruncate(), which
2369  * must keep the partial page.  In contrast, we must get rid of
2370  * partial pages.
2371  * @holelen: size of prospective hole in bytes.  This will be rounded
2372  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2373  * end of the file.
2374  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2375  * but 0 when invalidating pagecache, don't throw away private data.
2376  */
2377 void unmap_mapping_range(struct address_space *mapping,
2378                 loff_t const holebegin, loff_t const holelen, int even_cows)
2379 {
2380         struct zap_details details;
2381         pgoff_t hba = holebegin >> PAGE_SHIFT;
2382         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2383
2384         /* Check for overflow. */
2385         if (sizeof(holelen) > sizeof(hlen)) {
2386                 long long holeend =
2387                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2388                 if (holeend & ~(long long)ULONG_MAX)
2389                         hlen = ULONG_MAX - hba + 1;
2390         }
2391
2392         details.check_mapping = even_cows? NULL: mapping;
2393         details.nonlinear_vma = NULL;
2394         details.first_index = hba;
2395         details.last_index = hba + hlen - 1;
2396         if (details.last_index < details.first_index)
2397                 details.last_index = ULONG_MAX;
2398         details.i_mmap_lock = &mapping->i_mmap_lock;
2399
2400         spin_lock(&mapping->i_mmap_lock);
2401
2402         /* Protect against endless unmapping loops */
2403         mapping->truncate_count++;
2404         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2405                 if (mapping->truncate_count == 0)
2406                         reset_vma_truncate_counts(mapping);
2407                 mapping->truncate_count++;
2408         }
2409         details.truncate_count = mapping->truncate_count;
2410
2411         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2412                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2413         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2414                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2415         spin_unlock(&mapping->i_mmap_lock);
2416 }
2417 EXPORT_SYMBOL(unmap_mapping_range);
2418
2419 /**
2420  * vmtruncate - unmap mappings "freed" by truncate() syscall
2421  * @inode: inode of the file used
2422  * @offset: file offset to start truncating
2423  *
2424  * NOTE! We have to be ready to update the memory sharing
2425  * between the file and the memory map for a potential last
2426  * incomplete page.  Ugly, but necessary.
2427  */
2428 int vmtruncate(struct inode * inode, loff_t offset)
2429 {
2430         if (inode->i_size < offset) {
2431                 unsigned long limit;
2432
2433                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2434                 if (limit != RLIM_INFINITY && offset > limit)
2435                         goto out_sig;
2436                 if (offset > inode->i_sb->s_maxbytes)
2437                         goto out_big;
2438                 i_size_write(inode, offset);
2439         } else {
2440                 struct address_space *mapping = inode->i_mapping;
2441
2442                 /*
2443                  * truncation of in-use swapfiles is disallowed - it would
2444                  * cause subsequent swapout to scribble on the now-freed
2445                  * blocks.
2446                  */
2447                 if (IS_SWAPFILE(inode))
2448                         return -ETXTBSY;
2449                 i_size_write(inode, offset);
2450
2451                 /*
2452                  * unmap_mapping_range is called twice, first simply for
2453                  * efficiency so that truncate_inode_pages does fewer
2454                  * single-page unmaps.  However after this first call, and
2455                  * before truncate_inode_pages finishes, it is possible for
2456                  * private pages to be COWed, which remain after
2457                  * truncate_inode_pages finishes, hence the second
2458                  * unmap_mapping_range call must be made for correctness.
2459                  */
2460                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2461                 truncate_inode_pages(mapping, offset);
2462                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2463         }
2464
2465         if (inode->i_op->truncate)
2466                 inode->i_op->truncate(inode);
2467         return 0;
2468
2469 out_sig:
2470         send_sig(SIGXFSZ, current, 0);
2471 out_big:
2472         return -EFBIG;
2473 }
2474 EXPORT_SYMBOL(vmtruncate);
2475
2476 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2477 {
2478         struct address_space *mapping = inode->i_mapping;
2479
2480         /*
2481          * If the underlying filesystem is not going to provide
2482          * a way to truncate a range of blocks (punch a hole) -
2483          * we should return failure right now.
2484          */
2485         if (!inode->i_op->truncate_range)
2486                 return -ENOSYS;
2487
2488         mutex_lock(&inode->i_mutex);
2489         down_write(&inode->i_alloc_sem);
2490         unmap_mapping_range(mapping, offset, (end - offset), 1);
2491         truncate_inode_pages_range(mapping, offset, end);
2492         unmap_mapping_range(mapping, offset, (end - offset), 1);
2493         inode->i_op->truncate_range(inode, offset, end);
2494         up_write(&inode->i_alloc_sem);
2495         mutex_unlock(&inode->i_mutex);
2496
2497         return 0;
2498 }
2499
2500 /*
2501  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2502  * but allow concurrent faults), and pte mapped but not yet locked.
2503  * We return with mmap_sem still held, but pte unmapped and unlocked.
2504  */
2505 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2506                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2507                 unsigned int flags, pte_t orig_pte)
2508 {
2509         spinlock_t *ptl;
2510         struct page *page;
2511         swp_entry_t entry;
2512         pte_t pte;
2513         struct mem_cgroup *ptr = NULL;
2514         int ret = 0;
2515
2516         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2517                 goto out;
2518
2519         entry = pte_to_swp_entry(orig_pte);
2520         if (is_migration_entry(entry)) {
2521                 migration_entry_wait(mm, pmd, address);
2522                 goto out;
2523         }
2524         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2525         page = lookup_swap_cache(entry);
2526         if (!page) {
2527                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2528                 page = swapin_readahead(entry,
2529                                         GFP_HIGHUSER_MOVABLE, vma, address);
2530                 if (!page) {
2531                         /*
2532                          * Back out if somebody else faulted in this pte
2533                          * while we released the pte lock.
2534                          */
2535                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2536                         if (likely(pte_same(*page_table, orig_pte)))
2537                                 ret = VM_FAULT_OOM;
2538                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2539                         goto unlock;
2540                 }
2541
2542                 /* Had to read the page from swap area: Major fault */
2543                 ret = VM_FAULT_MAJOR;
2544                 count_vm_event(PGMAJFAULT);
2545         }
2546
2547         lock_page(page);
2548         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2549
2550         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2551                 ret = VM_FAULT_OOM;
2552                 goto out_page;
2553         }
2554
2555         /*
2556          * Back out if somebody else already faulted in this pte.
2557          */
2558         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2559         if (unlikely(!pte_same(*page_table, orig_pte)))
2560                 goto out_nomap;
2561
2562         if (unlikely(!PageUptodate(page))) {
2563                 ret = VM_FAULT_SIGBUS;
2564                 goto out_nomap;
2565         }
2566
2567         /*
2568          * The page isn't present yet, go ahead with the fault.
2569          *
2570          * Be careful about the sequence of operations here.
2571          * To get its accounting right, reuse_swap_page() must be called
2572          * while the page is counted on swap but not yet in mapcount i.e.
2573          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2574          * must be called after the swap_free(), or it will never succeed.
2575          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2576          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2577          * in page->private. In this case, a record in swap_cgroup  is silently
2578          * discarded at swap_free().
2579          */
2580
2581         inc_mm_counter(mm, anon_rss);
2582         pte = mk_pte(page, vma->vm_page_prot);
2583         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2584                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2585                 flags &= ~FAULT_FLAG_WRITE;
2586         }
2587         flush_icache_page(vma, page);
2588         set_pte_at(mm, address, page_table, pte);
2589         page_add_anon_rmap(page, vma, address);
2590         /* It's better to call commit-charge after rmap is established */
2591         mem_cgroup_commit_charge_swapin(page, ptr);
2592
2593         swap_free(entry);
2594         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2595                 try_to_free_swap(page);
2596         unlock_page(page);
2597
2598         if (flags & FAULT_FLAG_WRITE) {
2599                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2600                 if (ret & VM_FAULT_ERROR)
2601                         ret &= VM_FAULT_ERROR;
2602                 goto out;
2603         }
2604
2605         /* No need to invalidate - it was non-present before */
2606         update_mmu_cache(vma, address, pte);
2607 unlock:
2608         pte_unmap_unlock(page_table, ptl);
2609 out:
2610         return ret;
2611 out_nomap:
2612         mem_cgroup_cancel_charge_swapin(ptr);
2613         pte_unmap_unlock(page_table, ptl);
2614 out_page:
2615         unlock_page(page);
2616         page_cache_release(page);
2617         return ret;
2618 }
2619
2620 /*
2621  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2622  * but allow concurrent faults), and pte mapped but not yet locked.
2623  * We return with mmap_sem still held, but pte unmapped and unlocked.
2624  */
2625 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2626                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2627                 unsigned int flags)
2628 {
2629         struct page *page;
2630         spinlock_t *ptl;
2631         pte_t entry;
2632
2633         /* Allocate our own private page. */
2634         pte_unmap(page_table);
2635
2636         if (unlikely(anon_vma_prepare(vma)))
2637                 goto oom;
2638         page = alloc_zeroed_user_highpage_movable(vma, address);
2639         if (!page)
2640                 goto oom;
2641         __SetPageUptodate(page);
2642
2643         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2644                 goto oom_free_page;
2645
2646         entry = mk_pte(page, vma->vm_page_prot);
2647         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2648
2649         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2650         if (!pte_none(*page_table))
2651                 goto release;
2652         inc_mm_counter(mm, anon_rss);
2653         page_add_new_anon_rmap(page, vma, address);
2654         set_pte_at(mm, address, page_table, entry);
2655
2656         /* No need to invalidate - it was non-present before */
2657         update_mmu_cache(vma, address, entry);
2658 unlock:
2659         pte_unmap_unlock(page_table, ptl);
2660         return 0;
2661 release:
2662         mem_cgroup_uncharge_page(page);
2663         page_cache_release(page);
2664         goto unlock;
2665 oom_free_page:
2666         page_cache_release(page);
2667 oom:
2668         return VM_FAULT_OOM;
2669 }
2670
2671 /*
2672  * __do_fault() tries to create a new page mapping. It aggressively
2673  * tries to share with existing pages, but makes a separate copy if
2674  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2675  * the next page fault.
2676  *
2677  * As this is called only for pages that do not currently exist, we
2678  * do not need to flush old virtual caches or the TLB.
2679  *
2680  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2681  * but allow concurrent faults), and pte neither mapped nor locked.
2682  * We return with mmap_sem still held, but pte unmapped and unlocked.
2683  */
2684 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2685                 unsigned long address, pmd_t *pmd,
2686                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2687 {
2688         pte_t *page_table;
2689         spinlock_t *ptl;
2690         struct page *page;
2691         pte_t entry;
2692         int anon = 0;
2693         int charged = 0;
2694         struct page *dirty_page = NULL;
2695         struct vm_fault vmf;
2696         int ret;
2697         int page_mkwrite = 0;
2698
2699         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2700         vmf.pgoff = pgoff;
2701         vmf.flags = flags;
2702         vmf.page = NULL;
2703
2704         ret = vma->vm_ops->fault(vma, &vmf);
2705         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2706                 return ret;
2707
2708         /*
2709          * For consistency in subsequent calls, make the faulted page always
2710          * locked.
2711          */
2712         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2713                 lock_page(vmf.page);
2714         else
2715                 VM_BUG_ON(!PageLocked(vmf.page));
2716
2717         /*
2718          * Should we do an early C-O-W break?
2719          */
2720         page = vmf.page;
2721         if (flags & FAULT_FLAG_WRITE) {
2722                 if (!(vma->vm_flags & VM_SHARED)) {
2723                         anon = 1;
2724                         if (unlikely(anon_vma_prepare(vma))) {
2725                                 ret = VM_FAULT_OOM;
2726                                 goto out;
2727                         }
2728                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2729                                                 vma, address);
2730                         if (!page) {
2731                                 ret = VM_FAULT_OOM;
2732                                 goto out;
2733                         }
2734                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2735                                 ret = VM_FAULT_OOM;
2736                                 page_cache_release(page);
2737                                 goto out;
2738                         }
2739                         charged = 1;
2740                         /*
2741                          * Don't let another task, with possibly unlocked vma,
2742                          * keep the mlocked page.
2743                          */
2744                         if (vma->vm_flags & VM_LOCKED)
2745                                 clear_page_mlock(vmf.page);
2746                         copy_user_highpage(page, vmf.page, address, vma);
2747                         __SetPageUptodate(page);
2748                 } else {
2749                         /*
2750                          * If the page will be shareable, see if the backing
2751                          * address space wants to know that the page is about
2752                          * to become writable
2753                          */
2754                         if (vma->vm_ops->page_mkwrite) {
2755                                 int tmp;
2756
2757                                 unlock_page(page);
2758                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2759                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2760                                 if (unlikely(tmp &
2761                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2762                                         ret = tmp;
2763                                         goto unwritable_page;
2764                                 }
2765                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2766                                         lock_page(page);
2767                                         if (!page->mapping) {
2768                                                 ret = 0; /* retry the fault */
2769                                                 unlock_page(page);
2770                                                 goto unwritable_page;
2771                                         }
2772                                 } else
2773                                         VM_BUG_ON(!PageLocked(page));
2774                                 page_mkwrite = 1;
2775                         }
2776                 }
2777
2778         }
2779
2780         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2781
2782         /*
2783          * This silly early PAGE_DIRTY setting removes a race
2784          * due to the bad i386 page protection. But it's valid
2785          * for other architectures too.
2786          *
2787          * Note that if FAULT_FLAG_WRITE is set, we either now have
2788          * an exclusive copy of the page, or this is a shared mapping,
2789          * so we can make it writable and dirty to avoid having to
2790          * handle that later.
2791          */
2792         /* Only go through if we didn't race with anybody else... */
2793         if (likely(pte_same(*page_table, orig_pte))) {
2794                 flush_icache_page(vma, page);
2795                 entry = mk_pte(page, vma->vm_page_prot);
2796                 if (flags & FAULT_FLAG_WRITE)
2797                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2798                 if (anon) {
2799                         inc_mm_counter(mm, anon_rss);
2800                         page_add_new_anon_rmap(page, vma, address);
2801                 } else {
2802                         inc_mm_counter(mm, file_rss);
2803                         page_add_file_rmap(page);
2804                         if (flags & FAULT_FLAG_WRITE) {
2805                                 dirty_page = page;
2806                                 get_page(dirty_page);
2807                         }
2808                 }
2809                 set_pte_at(mm, address, page_table, entry);
2810
2811                 /* no need to invalidate: a not-present page won't be cached */
2812                 update_mmu_cache(vma, address, entry);
2813         } else {
2814                 if (charged)
2815                         mem_cgroup_uncharge_page(page);
2816                 if (anon)
2817                         page_cache_release(page);
2818                 else
2819                         anon = 1; /* no anon but release faulted_page */
2820         }
2821
2822         pte_unmap_unlock(page_table, ptl);
2823
2824 out:
2825         if (dirty_page) {
2826                 struct address_space *mapping = page->mapping;
2827
2828                 if (set_page_dirty(dirty_page))
2829                         page_mkwrite = 1;
2830                 unlock_page(dirty_page);
2831                 put_page(dirty_page);
2832                 if (page_mkwrite && mapping) {
2833                         /*
2834                          * Some device drivers do not set page.mapping but still
2835                          * dirty their pages
2836                          */
2837                         balance_dirty_pages_ratelimited(mapping);
2838                 }
2839
2840                 /* file_update_time outside page_lock */
2841                 if (vma->vm_file)
2842                         file_update_time(vma->vm_file);
2843         } else {
2844                 unlock_page(vmf.page);
2845                 if (anon)
2846                         page_cache_release(vmf.page);
2847         }
2848
2849         return ret;
2850
2851 unwritable_page:
2852         page_cache_release(page);
2853         return ret;
2854 }
2855
2856 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2857                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2858                 unsigned int flags, pte_t orig_pte)
2859 {
2860         pgoff_t pgoff = (((address & PAGE_MASK)
2861                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2862
2863         pte_unmap(page_table);
2864         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2865 }
2866
2867 /*
2868  * Fault of a previously existing named mapping. Repopulate the pte
2869  * from the encoded file_pte if possible. This enables swappable
2870  * nonlinear vmas.
2871  *
2872  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2873  * but allow concurrent faults), and pte mapped but not yet locked.
2874  * We return with mmap_sem still held, but pte unmapped and unlocked.
2875  */
2876 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2877                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2878                 unsigned int flags, pte_t orig_pte)
2879 {
2880         pgoff_t pgoff;
2881
2882         flags |= FAULT_FLAG_NONLINEAR;
2883
2884         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2885                 return 0;
2886
2887         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2888                 /*
2889                  * Page table corrupted: show pte and kill process.
2890                  */
2891                 print_bad_pte(vma, address, orig_pte, NULL);
2892                 return VM_FAULT_OOM;
2893         }
2894
2895         pgoff = pte_to_pgoff(orig_pte);
2896         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2897 }
2898
2899 /*
2900  * These routines also need to handle stuff like marking pages dirty
2901  * and/or accessed for architectures that don't do it in hardware (most
2902  * RISC architectures).  The early dirtying is also good on the i386.
2903  *
2904  * There is also a hook called "update_mmu_cache()" that architectures
2905  * with external mmu caches can use to update those (ie the Sparc or
2906  * PowerPC hashed page tables that act as extended TLBs).
2907  *
2908  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2909  * but allow concurrent faults), and pte mapped but not yet locked.
2910  * We return with mmap_sem still held, but pte unmapped and unlocked.
2911  */
2912 static inline int handle_pte_fault(struct mm_struct *mm,
2913                 struct vm_area_struct *vma, unsigned long address,
2914                 pte_t *pte, pmd_t *pmd, unsigned int flags)
2915 {
2916         pte_t entry;
2917         spinlock_t *ptl;
2918
2919         entry = *pte;
2920         if (!pte_present(entry)) {
2921                 if (pte_none(entry)) {
2922                         if (vma->vm_ops) {
2923                                 if (likely(vma->vm_ops->fault))
2924                                         return do_linear_fault(mm, vma, address,
2925                                                 pte, pmd, flags, entry);
2926                         }
2927                         return do_anonymous_page(mm, vma, address,
2928                                                  pte, pmd, flags);
2929                 }
2930                 if (pte_file(entry))
2931                         return do_nonlinear_fault(mm, vma, address,
2932                                         pte, pmd, flags, entry);
2933                 return do_swap_page(mm, vma, address,
2934                                         pte, pmd, flags, entry);
2935         }
2936
2937         ptl = pte_lockptr(mm, pmd);
2938         spin_lock(ptl);
2939         if (unlikely(!pte_same(*pte, entry)))
2940                 goto unlock;
2941         if (flags & FAULT_FLAG_WRITE) {
2942                 if (!pte_write(entry))
2943                         return do_wp_page(mm, vma, address,
2944                                         pte, pmd, ptl, entry);
2945                 entry = pte_mkdirty(entry);
2946         }
2947         entry = pte_mkyoung(entry);
2948         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
2949                 update_mmu_cache(vma, address, entry);
2950         } else {
2951                 /*
2952                  * This is needed only for protection faults but the arch code
2953                  * is not yet telling us if this is a protection fault or not.
2954                  * This still avoids useless tlb flushes for .text page faults
2955                  * with threads.
2956                  */
2957                 if (flags & FAULT_FLAG_WRITE)
2958                         flush_tlb_page(vma, address);
2959         }
2960 unlock:
2961         pte_unmap_unlock(pte, ptl);
2962         return 0;
2963 }
2964
2965 /*
2966  * By the time we get here, we already hold the mm semaphore
2967  */
2968 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2969                 unsigned long address, unsigned int flags)
2970 {
2971         pgd_t *pgd;
2972         pud_t *pud;
2973         pmd_t *pmd;
2974         pte_t *pte;
2975
2976         __set_current_state(TASK_RUNNING);
2977
2978         count_vm_event(PGFAULT);
2979
2980         if (unlikely(is_vm_hugetlb_page(vma)))
2981                 return hugetlb_fault(mm, vma, address, flags);
2982
2983         pgd = pgd_offset(mm, address);
2984         pud = pud_alloc(mm, pgd, address);
2985         if (!pud)
2986                 return VM_FAULT_OOM;
2987         pmd = pmd_alloc(mm, pud, address);
2988         if (!pmd)
2989                 return VM_FAULT_OOM;
2990         pte = pte_alloc_map(mm, pmd, address);
2991         if (!pte)
2992                 return VM_FAULT_OOM;
2993
2994         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
2995 }
2996
2997 #ifndef __PAGETABLE_PUD_FOLDED
2998 /*
2999  * Allocate page upper directory.
3000  * We've already handled the fast-path in-line.
3001  */
3002 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3003 {
3004         pud_t *new = pud_alloc_one(mm, address);
3005         if (!new)
3006                 return -ENOMEM;
3007
3008         smp_wmb(); /* See comment in __pte_alloc */
3009
3010         spin_lock(&mm->page_table_lock);
3011         if (pgd_present(*pgd))          /* Another has populated it */
3012                 pud_free(mm, new);
3013         else
3014                 pgd_populate(mm, pgd, new);
3015         spin_unlock(&mm->page_table_lock);
3016         return 0;
3017 }
3018 #endif /* __PAGETABLE_PUD_FOLDED */
3019
3020 #ifndef __PAGETABLE_PMD_FOLDED
3021 /*
3022  * Allocate page middle directory.
3023  * We've already handled the fast-path in-line.
3024  */
3025 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3026 {
3027         pmd_t *new = pmd_alloc_one(mm, address);
3028         if (!new)
3029                 return -ENOMEM;
3030
3031         smp_wmb(); /* See comment in __pte_alloc */
3032
3033         spin_lock(&mm->page_table_lock);
3034 #ifndef __ARCH_HAS_4LEVEL_HACK
3035         if (pud_present(*pud))          /* Another has populated it */
3036                 pmd_free(mm, new);
3037         else
3038                 pud_populate(mm, pud, new);
3039 #else
3040         if (pgd_present(*pud))          /* Another has populated it */
3041                 pmd_free(mm, new);
3042         else
3043                 pgd_populate(mm, pud, new);
3044 #endif /* __ARCH_HAS_4LEVEL_HACK */
3045         spin_unlock(&mm->page_table_lock);
3046         return 0;
3047 }
3048 #endif /* __PAGETABLE_PMD_FOLDED */
3049
3050 int make_pages_present(unsigned long addr, unsigned long end)
3051 {
3052         int ret, len, write;
3053         struct vm_area_struct * vma;
3054
3055         vma = find_vma(current->mm, addr);
3056         if (!vma)
3057                 return -ENOMEM;
3058         write = (vma->vm_flags & VM_WRITE) != 0;
3059         BUG_ON(addr >= end);
3060         BUG_ON(end > vma->vm_end);
3061         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3062         ret = get_user_pages(current, current->mm, addr,
3063                         len, write, 0, NULL, NULL);
3064         if (ret < 0)
3065                 return ret;
3066         return ret == len ? 0 : -EFAULT;
3067 }
3068
3069 #if !defined(__HAVE_ARCH_GATE_AREA)
3070
3071 #if defined(AT_SYSINFO_EHDR)
3072 static struct vm_area_struct gate_vma;
3073
3074 static int __init gate_vma_init(void)
3075 {
3076         gate_vma.vm_mm = NULL;
3077         gate_vma.vm_start = FIXADDR_USER_START;
3078         gate_vma.vm_end = FIXADDR_USER_END;
3079         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3080         gate_vma.vm_page_prot = __P101;
3081         /*
3082          * Make sure the vDSO gets into every core dump.
3083          * Dumping its contents makes post-mortem fully interpretable later
3084          * without matching up the same kernel and hardware config to see
3085          * what PC values meant.
3086          */
3087         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3088         return 0;
3089 }
3090 __initcall(gate_vma_init);
3091 #endif
3092
3093 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3094 {
3095 #ifdef AT_SYSINFO_EHDR
3096         return &gate_vma;
3097 #else
3098         return NULL;
3099 #endif
3100 }
3101
3102 int in_gate_area_no_task(unsigned long addr)
3103 {
3104 #ifdef AT_SYSINFO_EHDR
3105         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3106                 return 1;
3107 #endif
3108         return 0;
3109 }
3110
3111 #endif  /* __HAVE_ARCH_GATE_AREA */
3112
3113 static int follow_pte(struct mm_struct *mm, unsigned long address,
3114                 pte_t **ptepp, spinlock_t **ptlp)
3115 {
3116         pgd_t *pgd;
3117         pud_t *pud;
3118         pmd_t *pmd;
3119         pte_t *ptep;
3120
3121         pgd = pgd_offset(mm, address);
3122         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3123                 goto out;
3124
3125         pud = pud_offset(pgd, address);
3126         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3127                 goto out;
3128
3129         pmd = pmd_offset(pud, address);
3130         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3131                 goto out;
3132
3133         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3134         if (pmd_huge(*pmd))
3135                 goto out;
3136
3137         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3138         if (!ptep)
3139                 goto out;
3140         if (!pte_present(*ptep))
3141                 goto unlock;
3142         *ptepp = ptep;
3143         return 0;
3144 unlock:
3145         pte_unmap_unlock(ptep, *ptlp);
3146 out:
3147         return -EINVAL;
3148 }
3149
3150 /**
3151  * follow_pfn - look up PFN at a user virtual address
3152  * @vma: memory mapping
3153  * @address: user virtual address
3154  * @pfn: location to store found PFN
3155  *
3156  * Only IO mappings and raw PFN mappings are allowed.
3157  *
3158  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3159  */
3160 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3161         unsigned long *pfn)
3162 {
3163         int ret = -EINVAL;
3164         spinlock_t *ptl;
3165         pte_t *ptep;
3166
3167         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3168                 return ret;
3169
3170         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3171         if (ret)
3172                 return ret;
3173         *pfn = pte_pfn(*ptep);
3174         pte_unmap_unlock(ptep, ptl);
3175         return 0;
3176 }
3177 EXPORT_SYMBOL(follow_pfn);
3178
3179 #ifdef CONFIG_HAVE_IOREMAP_PROT
3180 int follow_phys(struct vm_area_struct *vma,
3181                 unsigned long address, unsigned int flags,
3182                 unsigned long *prot, resource_size_t *phys)
3183 {
3184         int ret = -EINVAL;
3185         pte_t *ptep, pte;
3186         spinlock_t *ptl;
3187
3188         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3189                 goto out;
3190
3191         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3192                 goto out;
3193         pte = *ptep;
3194
3195         if ((flags & FOLL_WRITE) && !pte_write(pte))
3196                 goto unlock;
3197
3198         *prot = pgprot_val(pte_pgprot(pte));
3199         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3200
3201         ret = 0;
3202 unlock:
3203         pte_unmap_unlock(ptep, ptl);
3204 out:
3205         return ret;
3206 }
3207
3208 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3209                         void *buf, int len, int write)
3210 {
3211         resource_size_t phys_addr;
3212         unsigned long prot = 0;
3213         void __iomem *maddr;
3214         int offset = addr & (PAGE_SIZE-1);
3215
3216         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3217                 return -EINVAL;
3218
3219         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3220         if (write)
3221                 memcpy_toio(maddr + offset, buf, len);
3222         else
3223                 memcpy_fromio(buf, maddr + offset, len);
3224         iounmap(maddr);
3225
3226         return len;
3227 }
3228 #endif
3229
3230 /*
3231  * Access another process' address space.
3232  * Source/target buffer must be kernel space,
3233  * Do not walk the page table directly, use get_user_pages
3234  */
3235 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3236 {
3237         struct mm_struct *mm;
3238         struct vm_area_struct *vma;
3239         void *old_buf = buf;
3240
3241         mm = get_task_mm(tsk);
3242         if (!mm)
3243                 return 0;
3244
3245         down_read(&mm->mmap_sem);
3246         /* ignore errors, just check how much was successfully transferred */
3247         while (len) {
3248                 int bytes, ret, offset;
3249                 void *maddr;
3250                 struct page *page = NULL;
3251
3252                 ret = get_user_pages(tsk, mm, addr, 1,
3253                                 write, 1, &page, &vma);
3254                 if (ret <= 0) {
3255                         /*
3256                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3257                          * we can access using slightly different code.
3258                          */
3259 #ifdef CONFIG_HAVE_IOREMAP_PROT
3260                         vma = find_vma(mm, addr);
3261                         if (!vma)
3262                                 break;
3263                         if (vma->vm_ops && vma->vm_ops->access)
3264                                 ret = vma->vm_ops->access(vma, addr, buf,
3265                                                           len, write);
3266                         if (ret <= 0)
3267 #endif
3268                                 break;
3269                         bytes = ret;
3270                 } else {
3271                         bytes = len;
3272                         offset = addr & (PAGE_SIZE-1);
3273                         if (bytes > PAGE_SIZE-offset)
3274                                 bytes = PAGE_SIZE-offset;
3275
3276                         maddr = kmap(page);
3277                         if (write) {
3278                                 copy_to_user_page(vma, page, addr,
3279                                                   maddr + offset, buf, bytes);
3280                                 set_page_dirty_lock(page);
3281                         } else {
3282                                 copy_from_user_page(vma, page, addr,
3283                                                     buf, maddr + offset, bytes);
3284                         }
3285                         kunmap(page);
3286                         page_cache_release(page);
3287                 }
3288                 len -= bytes;
3289                 buf += bytes;
3290                 addr += bytes;
3291         }
3292         up_read(&mm->mmap_sem);
3293         mmput(mm);
3294
3295         return buf - old_buf;
3296 }
3297
3298 /*
3299  * Print the name of a VMA.
3300  */
3301 void print_vma_addr(char *prefix, unsigned long ip)
3302 {
3303         struct mm_struct *mm = current->mm;
3304         struct vm_area_struct *vma;
3305
3306         /*
3307          * Do not print if we are in atomic
3308          * contexts (in exception stacks, etc.):
3309          */
3310         if (preempt_count())
3311                 return;
3312
3313         down_read(&mm->mmap_sem);
3314         vma = find_vma(mm, ip);
3315         if (vma && vma->vm_file) {
3316                 struct file *f = vma->vm_file;
3317                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3318                 if (buf) {
3319                         char *p, *s;
3320
3321                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3322                         if (IS_ERR(p))
3323                                 p = "?";
3324                         s = strrchr(p, '/');
3325                         if (s)
3326                                 p = s+1;
3327                         printk("%s%s[%lx+%lx]", prefix, p,
3328                                         vma->vm_start,
3329                                         vma->vm_end - vma->vm_start);
3330                         free_page((unsigned long)buf);
3331                 }
3332         }
3333         up_read(&current->mm->mmap_sem);
3334 }
3335
3336 #ifdef CONFIG_PROVE_LOCKING
3337 void might_fault(void)
3338 {
3339         /*
3340          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3341          * holding the mmap_sem, this is safe because kernel memory doesn't
3342          * get paged out, therefore we'll never actually fault, and the
3343          * below annotations will generate false positives.
3344          */
3345         if (segment_eq(get_fs(), KERNEL_DS))
3346                 return;
3347
3348         might_sleep();
3349         /*
3350          * it would be nicer only to annotate paths which are not under
3351          * pagefault_disable, however that requires a larger audit and
3352          * providing helpers like get_user_atomic.
3353          */
3354         if (!in_atomic() && current->mm)
3355                 might_lock_read(&current->mm->mmap_sem);
3356 }
3357 EXPORT_SYMBOL(might_fault);
3358 #endif