HWPOISON: Fix page count leak in hwpoison late kill in do_swap_page
[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/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59
60 #include <asm/io.h>
61 #include <asm/pgalloc.h>
62 #include <asm/uaccess.h>
63 #include <asm/tlb.h>
64 #include <asm/tlbflush.h>
65 #include <asm/pgtable.h>
66
67 #include "internal.h"
68
69 #ifndef CONFIG_NEED_MULTIPLE_NODES
70 /* use the per-pgdat data instead for discontigmem - mbligh */
71 unsigned long max_mapnr;
72 struct page *mem_map;
73
74 EXPORT_SYMBOL(max_mapnr);
75 EXPORT_SYMBOL(mem_map);
76 #endif
77
78 unsigned long num_physpages;
79 /*
80  * A number of key systems in x86 including ioremap() rely on the assumption
81  * that high_memory defines the upper bound on direct map memory, then end
82  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
83  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
84  * and ZONE_HIGHMEM.
85  */
86 void * high_memory;
87
88 EXPORT_SYMBOL(num_physpages);
89 EXPORT_SYMBOL(high_memory);
90
91 /*
92  * Randomize the address space (stacks, mmaps, brk, etc.).
93  *
94  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
95  *   as ancient (libc5 based) binaries can segfault. )
96  */
97 int randomize_va_space __read_mostly =
98 #ifdef CONFIG_COMPAT_BRK
99                                         1;
100 #else
101                                         2;
102 #endif
103
104 static int __init disable_randmaps(char *s)
105 {
106         randomize_va_space = 0;
107         return 1;
108 }
109 __setup("norandmaps", disable_randmaps);
110
111 unsigned long zero_pfn __read_mostly;
112 unsigned long highest_memmap_pfn __read_mostly;
113
114 /*
115  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
116  */
117 static int __init init_zero_pfn(void)
118 {
119         zero_pfn = page_to_pfn(ZERO_PAGE(0));
120         return 0;
121 }
122 core_initcall(init_zero_pfn);
123
124 /*
125  * If a p?d_bad entry is found while walking page tables, report
126  * the error, before resetting entry to p?d_none.  Usually (but
127  * very seldom) called out from the p?d_none_or_clear_bad macros.
128  */
129
130 void pgd_clear_bad(pgd_t *pgd)
131 {
132         pgd_ERROR(*pgd);
133         pgd_clear(pgd);
134 }
135
136 void pud_clear_bad(pud_t *pud)
137 {
138         pud_ERROR(*pud);
139         pud_clear(pud);
140 }
141
142 void pmd_clear_bad(pmd_t *pmd)
143 {
144         pmd_ERROR(*pmd);
145         pmd_clear(pmd);
146 }
147
148 /*
149  * Note: this doesn't free the actual pages themselves. That
150  * has been handled earlier when unmapping all the memory regions.
151  */
152 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
153                            unsigned long addr)
154 {
155         pgtable_t token = pmd_pgtable(*pmd);
156         pmd_clear(pmd);
157         pte_free_tlb(tlb, token, addr);
158         tlb->mm->nr_ptes--;
159 }
160
161 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
162                                 unsigned long addr, unsigned long end,
163                                 unsigned long floor, unsigned long ceiling)
164 {
165         pmd_t *pmd;
166         unsigned long next;
167         unsigned long start;
168
169         start = addr;
170         pmd = pmd_offset(pud, addr);
171         do {
172                 next = pmd_addr_end(addr, end);
173                 if (pmd_none_or_clear_bad(pmd))
174                         continue;
175                 free_pte_range(tlb, pmd, addr);
176         } while (pmd++, addr = next, addr != end);
177
178         start &= PUD_MASK;
179         if (start < floor)
180                 return;
181         if (ceiling) {
182                 ceiling &= PUD_MASK;
183                 if (!ceiling)
184                         return;
185         }
186         if (end - 1 > ceiling - 1)
187                 return;
188
189         pmd = pmd_offset(pud, start);
190         pud_clear(pud);
191         pmd_free_tlb(tlb, pmd, start);
192 }
193
194 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
195                                 unsigned long addr, unsigned long end,
196                                 unsigned long floor, unsigned long ceiling)
197 {
198         pud_t *pud;
199         unsigned long next;
200         unsigned long start;
201
202         start = addr;
203         pud = pud_offset(pgd, addr);
204         do {
205                 next = pud_addr_end(addr, end);
206                 if (pud_none_or_clear_bad(pud))
207                         continue;
208                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
209         } while (pud++, addr = next, addr != end);
210
211         start &= PGDIR_MASK;
212         if (start < floor)
213                 return;
214         if (ceiling) {
215                 ceiling &= PGDIR_MASK;
216                 if (!ceiling)
217                         return;
218         }
219         if (end - 1 > ceiling - 1)
220                 return;
221
222         pud = pud_offset(pgd, start);
223         pgd_clear(pgd);
224         pud_free_tlb(tlb, pud, start);
225 }
226
227 /*
228  * This function frees user-level page tables of a process.
229  *
230  * Must be called with pagetable lock held.
231  */
232 void free_pgd_range(struct mmu_gather *tlb,
233                         unsigned long addr, unsigned long end,
234                         unsigned long floor, unsigned long ceiling)
235 {
236         pgd_t *pgd;
237         unsigned long next;
238         unsigned long start;
239
240         /*
241          * The next few lines have given us lots of grief...
242          *
243          * Why are we testing PMD* at this top level?  Because often
244          * there will be no work to do at all, and we'd prefer not to
245          * go all the way down to the bottom just to discover that.
246          *
247          * Why all these "- 1"s?  Because 0 represents both the bottom
248          * of the address space and the top of it (using -1 for the
249          * top wouldn't help much: the masks would do the wrong thing).
250          * The rule is that addr 0 and floor 0 refer to the bottom of
251          * the address space, but end 0 and ceiling 0 refer to the top
252          * Comparisons need to use "end - 1" and "ceiling - 1" (though
253          * that end 0 case should be mythical).
254          *
255          * Wherever addr is brought up or ceiling brought down, we must
256          * be careful to reject "the opposite 0" before it confuses the
257          * subsequent tests.  But what about where end is brought down
258          * by PMD_SIZE below? no, end can't go down to 0 there.
259          *
260          * Whereas we round start (addr) and ceiling down, by different
261          * masks at different levels, in order to test whether a table
262          * now has no other vmas using it, so can be freed, we don't
263          * bother to round floor or end up - the tests don't need that.
264          */
265
266         addr &= PMD_MASK;
267         if (addr < floor) {
268                 addr += PMD_SIZE;
269                 if (!addr)
270                         return;
271         }
272         if (ceiling) {
273                 ceiling &= PMD_MASK;
274                 if (!ceiling)
275                         return;
276         }
277         if (end - 1 > ceiling - 1)
278                 end -= PMD_SIZE;
279         if (addr > end - 1)
280                 return;
281
282         start = addr;
283         pgd = pgd_offset(tlb->mm, addr);
284         do {
285                 next = pgd_addr_end(addr, end);
286                 if (pgd_none_or_clear_bad(pgd))
287                         continue;
288                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
289         } while (pgd++, addr = next, addr != end);
290 }
291
292 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
293                 unsigned long floor, unsigned long ceiling)
294 {
295         while (vma) {
296                 struct vm_area_struct *next = vma->vm_next;
297                 unsigned long addr = vma->vm_start;
298
299                 /*
300                  * Hide vma from rmap and truncate_pagecache before freeing
301                  * pgtables
302                  */
303                 anon_vma_unlink(vma);
304                 unlink_file_vma(vma);
305
306                 if (is_vm_hugetlb_page(vma)) {
307                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
308                                 floor, next? next->vm_start: ceiling);
309                 } else {
310                         /*
311                          * Optimization: gather nearby vmas into one call down
312                          */
313                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
314                                && !is_vm_hugetlb_page(next)) {
315                                 vma = next;
316                                 next = vma->vm_next;
317                                 anon_vma_unlink(vma);
318                                 unlink_file_vma(vma);
319                         }
320                         free_pgd_range(tlb, addr, vma->vm_end,
321                                 floor, next? next->vm_start: ceiling);
322                 }
323                 vma = next;
324         }
325 }
326
327 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
328 {
329         pgtable_t new = pte_alloc_one(mm, address);
330         if (!new)
331                 return -ENOMEM;
332
333         /*
334          * Ensure all pte setup (eg. pte page lock and page clearing) are
335          * visible before the pte is made visible to other CPUs by being
336          * put into page tables.
337          *
338          * The other side of the story is the pointer chasing in the page
339          * table walking code (when walking the page table without locking;
340          * ie. most of the time). Fortunately, these data accesses consist
341          * of a chain of data-dependent loads, meaning most CPUs (alpha
342          * being the notable exception) will already guarantee loads are
343          * seen in-order. See the alpha page table accessors for the
344          * smp_read_barrier_depends() barriers in page table walking code.
345          */
346         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
347
348         spin_lock(&mm->page_table_lock);
349         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
350                 mm->nr_ptes++;
351                 pmd_populate(mm, pmd, new);
352                 new = NULL;
353         }
354         spin_unlock(&mm->page_table_lock);
355         if (new)
356                 pte_free(mm, new);
357         return 0;
358 }
359
360 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
361 {
362         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
363         if (!new)
364                 return -ENOMEM;
365
366         smp_wmb(); /* See comment in __pte_alloc */
367
368         spin_lock(&init_mm.page_table_lock);
369         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
370                 pmd_populate_kernel(&init_mm, pmd, new);
371                 new = NULL;
372         }
373         spin_unlock(&init_mm.page_table_lock);
374         if (new)
375                 pte_free_kernel(&init_mm, new);
376         return 0;
377 }
378
379 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
380 {
381         if (file_rss)
382                 add_mm_counter(mm, file_rss, file_rss);
383         if (anon_rss)
384                 add_mm_counter(mm, anon_rss, anon_rss);
385 }
386
387 /*
388  * This function is called to print an error when a bad pte
389  * is found. For example, we might have a PFN-mapped pte in
390  * a region that doesn't allow it.
391  *
392  * The calling function must still handle the error.
393  */
394 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
395                           pte_t pte, struct page *page)
396 {
397         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
398         pud_t *pud = pud_offset(pgd, addr);
399         pmd_t *pmd = pmd_offset(pud, addr);
400         struct address_space *mapping;
401         pgoff_t index;
402         static unsigned long resume;
403         static unsigned long nr_shown;
404         static unsigned long nr_unshown;
405
406         /*
407          * Allow a burst of 60 reports, then keep quiet for that minute;
408          * or allow a steady drip of one report per second.
409          */
410         if (nr_shown == 60) {
411                 if (time_before(jiffies, resume)) {
412                         nr_unshown++;
413                         return;
414                 }
415                 if (nr_unshown) {
416                         printk(KERN_ALERT
417                                 "BUG: Bad page map: %lu messages suppressed\n",
418                                 nr_unshown);
419                         nr_unshown = 0;
420                 }
421                 nr_shown = 0;
422         }
423         if (nr_shown++ == 0)
424                 resume = jiffies + 60 * HZ;
425
426         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
427         index = linear_page_index(vma, addr);
428
429         printk(KERN_ALERT
430                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
431                 current->comm,
432                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
433         if (page) {
434                 printk(KERN_ALERT
435                 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
436                 page, (void *)page->flags, page_count(page),
437                 page_mapcount(page), page->mapping, page->index);
438         }
439         printk(KERN_ALERT
440                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
441                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
442         /*
443          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
444          */
445         if (vma->vm_ops)
446                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
447                                 (unsigned long)vma->vm_ops->fault);
448         if (vma->vm_file && vma->vm_file->f_op)
449                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
450                                 (unsigned long)vma->vm_file->f_op->mmap);
451         dump_stack();
452         add_taint(TAINT_BAD_PAGE);
453 }
454
455 static inline int is_cow_mapping(unsigned int flags)
456 {
457         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
458 }
459
460 #ifndef is_zero_pfn
461 static inline int is_zero_pfn(unsigned long pfn)
462 {
463         return pfn == zero_pfn;
464 }
465 #endif
466
467 #ifndef my_zero_pfn
468 static inline unsigned long my_zero_pfn(unsigned long addr)
469 {
470         return zero_pfn;
471 }
472 #endif
473
474 /*
475  * vm_normal_page -- This function gets the "struct page" associated with a pte.
476  *
477  * "Special" mappings do not wish to be associated with a "struct page" (either
478  * it doesn't exist, or it exists but they don't want to touch it). In this
479  * case, NULL is returned here. "Normal" mappings do have a struct page.
480  *
481  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
482  * pte bit, in which case this function is trivial. Secondly, an architecture
483  * may not have a spare pte bit, which requires a more complicated scheme,
484  * described below.
485  *
486  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
487  * special mapping (even if there are underlying and valid "struct pages").
488  * COWed pages of a VM_PFNMAP are always normal.
489  *
490  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
491  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
492  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
493  * mapping will always honor the rule
494  *
495  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
496  *
497  * And for normal mappings this is false.
498  *
499  * This restricts such mappings to be a linear translation from virtual address
500  * to pfn. To get around this restriction, we allow arbitrary mappings so long
501  * as the vma is not a COW mapping; in that case, we know that all ptes are
502  * special (because none can have been COWed).
503  *
504  *
505  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
506  *
507  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
508  * page" backing, however the difference is that _all_ pages with a struct
509  * page (that is, those where pfn_valid is true) are refcounted and considered
510  * normal pages by the VM. The disadvantage is that pages are refcounted
511  * (which can be slower and simply not an option for some PFNMAP users). The
512  * advantage is that we don't have to follow the strict linearity rule of
513  * PFNMAP mappings in order to support COWable mappings.
514  *
515  */
516 #ifdef __HAVE_ARCH_PTE_SPECIAL
517 # define HAVE_PTE_SPECIAL 1
518 #else
519 # define HAVE_PTE_SPECIAL 0
520 #endif
521 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
522                                 pte_t pte)
523 {
524         unsigned long pfn = pte_pfn(pte);
525
526         if (HAVE_PTE_SPECIAL) {
527                 if (likely(!pte_special(pte)))
528                         goto check_pfn;
529                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
530                         return NULL;
531                 if (!is_zero_pfn(pfn))
532                         print_bad_pte(vma, addr, pte, NULL);
533                 return NULL;
534         }
535
536         /* !HAVE_PTE_SPECIAL case follows: */
537
538         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
539                 if (vma->vm_flags & VM_MIXEDMAP) {
540                         if (!pfn_valid(pfn))
541                                 return NULL;
542                         goto out;
543                 } else {
544                         unsigned long off;
545                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
546                         if (pfn == vma->vm_pgoff + off)
547                                 return NULL;
548                         if (!is_cow_mapping(vma->vm_flags))
549                                 return NULL;
550                 }
551         }
552
553         if (is_zero_pfn(pfn))
554                 return NULL;
555 check_pfn:
556         if (unlikely(pfn > highest_memmap_pfn)) {
557                 print_bad_pte(vma, addr, pte, NULL);
558                 return NULL;
559         }
560
561         /*
562          * NOTE! We still have PageReserved() pages in the page tables.
563          * eg. VDSO mappings can cause them to exist.
564          */
565 out:
566         return pfn_to_page(pfn);
567 }
568
569 /*
570  * copy one vm_area from one task to the other. Assumes the page tables
571  * already present in the new task to be cleared in the whole range
572  * covered by this vma.
573  */
574
575 static inline void
576 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
577                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
578                 unsigned long addr, int *rss)
579 {
580         unsigned long vm_flags = vma->vm_flags;
581         pte_t pte = *src_pte;
582         struct page *page;
583
584         /* pte contains position in swap or file, so copy. */
585         if (unlikely(!pte_present(pte))) {
586                 if (!pte_file(pte)) {
587                         swp_entry_t entry = pte_to_swp_entry(pte);
588
589                         swap_duplicate(entry);
590                         /* make sure dst_mm is on swapoff's mmlist. */
591                         if (unlikely(list_empty(&dst_mm->mmlist))) {
592                                 spin_lock(&mmlist_lock);
593                                 if (list_empty(&dst_mm->mmlist))
594                                         list_add(&dst_mm->mmlist,
595                                                  &src_mm->mmlist);
596                                 spin_unlock(&mmlist_lock);
597                         }
598                         if (is_write_migration_entry(entry) &&
599                                         is_cow_mapping(vm_flags)) {
600                                 /*
601                                  * COW mappings require pages in both parent
602                                  * and child to be set to read.
603                                  */
604                                 make_migration_entry_read(&entry);
605                                 pte = swp_entry_to_pte(entry);
606                                 set_pte_at(src_mm, addr, src_pte, pte);
607                         }
608                 }
609                 goto out_set_pte;
610         }
611
612         /*
613          * If it's a COW mapping, write protect it both
614          * in the parent and the child
615          */
616         if (is_cow_mapping(vm_flags)) {
617                 ptep_set_wrprotect(src_mm, addr, src_pte);
618                 pte = pte_wrprotect(pte);
619         }
620
621         /*
622          * If it's a shared mapping, mark it clean in
623          * the child
624          */
625         if (vm_flags & VM_SHARED)
626                 pte = pte_mkclean(pte);
627         pte = pte_mkold(pte);
628
629         page = vm_normal_page(vma, addr, pte);
630         if (page) {
631                 get_page(page);
632                 page_dup_rmap(page);
633                 rss[PageAnon(page)]++;
634         }
635
636 out_set_pte:
637         set_pte_at(dst_mm, addr, dst_pte, pte);
638 }
639
640 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
641                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
642                 unsigned long addr, unsigned long end)
643 {
644         pte_t *src_pte, *dst_pte;
645         spinlock_t *src_ptl, *dst_ptl;
646         int progress = 0;
647         int rss[2];
648
649 again:
650         rss[1] = rss[0] = 0;
651         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
652         if (!dst_pte)
653                 return -ENOMEM;
654         src_pte = pte_offset_map_nested(src_pmd, addr);
655         src_ptl = pte_lockptr(src_mm, src_pmd);
656         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
657         arch_enter_lazy_mmu_mode();
658
659         do {
660                 /*
661                  * We are holding two locks at this point - either of them
662                  * could generate latencies in another task on another CPU.
663                  */
664                 if (progress >= 32) {
665                         progress = 0;
666                         if (need_resched() ||
667                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
668                                 break;
669                 }
670                 if (pte_none(*src_pte)) {
671                         progress++;
672                         continue;
673                 }
674                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
675                 progress += 8;
676         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
677
678         arch_leave_lazy_mmu_mode();
679         spin_unlock(src_ptl);
680         pte_unmap_nested(src_pte - 1);
681         add_mm_rss(dst_mm, rss[0], rss[1]);
682         pte_unmap_unlock(dst_pte - 1, dst_ptl);
683         cond_resched();
684         if (addr != end)
685                 goto again;
686         return 0;
687 }
688
689 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
690                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
691                 unsigned long addr, unsigned long end)
692 {
693         pmd_t *src_pmd, *dst_pmd;
694         unsigned long next;
695
696         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
697         if (!dst_pmd)
698                 return -ENOMEM;
699         src_pmd = pmd_offset(src_pud, addr);
700         do {
701                 next = pmd_addr_end(addr, end);
702                 if (pmd_none_or_clear_bad(src_pmd))
703                         continue;
704                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
705                                                 vma, addr, next))
706                         return -ENOMEM;
707         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
708         return 0;
709 }
710
711 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
712                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
713                 unsigned long addr, unsigned long end)
714 {
715         pud_t *src_pud, *dst_pud;
716         unsigned long next;
717
718         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
719         if (!dst_pud)
720                 return -ENOMEM;
721         src_pud = pud_offset(src_pgd, addr);
722         do {
723                 next = pud_addr_end(addr, end);
724                 if (pud_none_or_clear_bad(src_pud))
725                         continue;
726                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
727                                                 vma, addr, next))
728                         return -ENOMEM;
729         } while (dst_pud++, src_pud++, addr = next, addr != end);
730         return 0;
731 }
732
733 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
734                 struct vm_area_struct *vma)
735 {
736         pgd_t *src_pgd, *dst_pgd;
737         unsigned long next;
738         unsigned long addr = vma->vm_start;
739         unsigned long end = vma->vm_end;
740         int ret;
741
742         /*
743          * Don't copy ptes where a page fault will fill them correctly.
744          * Fork becomes much lighter when there are big shared or private
745          * readonly mappings. The tradeoff is that copy_page_range is more
746          * efficient than faulting.
747          */
748         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
749                 if (!vma->anon_vma)
750                         return 0;
751         }
752
753         if (is_vm_hugetlb_page(vma))
754                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
755
756         if (unlikely(is_pfn_mapping(vma))) {
757                 /*
758                  * We do not free on error cases below as remove_vma
759                  * gets called on error from higher level routine
760                  */
761                 ret = track_pfn_vma_copy(vma);
762                 if (ret)
763                         return ret;
764         }
765
766         /*
767          * We need to invalidate the secondary MMU mappings only when
768          * there could be a permission downgrade on the ptes of the
769          * parent mm. And a permission downgrade will only happen if
770          * is_cow_mapping() returns true.
771          */
772         if (is_cow_mapping(vma->vm_flags))
773                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
774
775         ret = 0;
776         dst_pgd = pgd_offset(dst_mm, addr);
777         src_pgd = pgd_offset(src_mm, addr);
778         do {
779                 next = pgd_addr_end(addr, end);
780                 if (pgd_none_or_clear_bad(src_pgd))
781                         continue;
782                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
783                                             vma, addr, next))) {
784                         ret = -ENOMEM;
785                         break;
786                 }
787         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
788
789         if (is_cow_mapping(vma->vm_flags))
790                 mmu_notifier_invalidate_range_end(src_mm,
791                                                   vma->vm_start, end);
792         return ret;
793 }
794
795 static unsigned long zap_pte_range(struct mmu_gather *tlb,
796                                 struct vm_area_struct *vma, pmd_t *pmd,
797                                 unsigned long addr, unsigned long end,
798                                 long *zap_work, struct zap_details *details)
799 {
800         struct mm_struct *mm = tlb->mm;
801         pte_t *pte;
802         spinlock_t *ptl;
803         int file_rss = 0;
804         int anon_rss = 0;
805
806         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
807         arch_enter_lazy_mmu_mode();
808         do {
809                 pte_t ptent = *pte;
810                 if (pte_none(ptent)) {
811                         (*zap_work)--;
812                         continue;
813                 }
814
815                 (*zap_work) -= PAGE_SIZE;
816
817                 if (pte_present(ptent)) {
818                         struct page *page;
819
820                         page = vm_normal_page(vma, addr, ptent);
821                         if (unlikely(details) && page) {
822                                 /*
823                                  * unmap_shared_mapping_pages() wants to
824                                  * invalidate cache without truncating:
825                                  * unmap shared but keep private pages.
826                                  */
827                                 if (details->check_mapping &&
828                                     details->check_mapping != page->mapping)
829                                         continue;
830                                 /*
831                                  * Each page->index must be checked when
832                                  * invalidating or truncating nonlinear.
833                                  */
834                                 if (details->nonlinear_vma &&
835                                     (page->index < details->first_index ||
836                                      page->index > details->last_index))
837                                         continue;
838                         }
839                         ptent = ptep_get_and_clear_full(mm, addr, pte,
840                                                         tlb->fullmm);
841                         tlb_remove_tlb_entry(tlb, pte, addr);
842                         if (unlikely(!page))
843                                 continue;
844                         if (unlikely(details) && details->nonlinear_vma
845                             && linear_page_index(details->nonlinear_vma,
846                                                 addr) != page->index)
847                                 set_pte_at(mm, addr, pte,
848                                            pgoff_to_pte(page->index));
849                         if (PageAnon(page))
850                                 anon_rss--;
851                         else {
852                                 if (pte_dirty(ptent))
853                                         set_page_dirty(page);
854                                 if (pte_young(ptent) &&
855                                     likely(!VM_SequentialReadHint(vma)))
856                                         mark_page_accessed(page);
857                                 file_rss--;
858                         }
859                         page_remove_rmap(page);
860                         if (unlikely(page_mapcount(page) < 0))
861                                 print_bad_pte(vma, addr, ptent, page);
862                         tlb_remove_page(tlb, page);
863                         continue;
864                 }
865                 /*
866                  * If details->check_mapping, we leave swap entries;
867                  * if details->nonlinear_vma, we leave file entries.
868                  */
869                 if (unlikely(details))
870                         continue;
871                 if (pte_file(ptent)) {
872                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
873                                 print_bad_pte(vma, addr, ptent, NULL);
874                 } else if
875                   (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
876                         print_bad_pte(vma, addr, ptent, NULL);
877                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
878         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
879
880         add_mm_rss(mm, file_rss, anon_rss);
881         arch_leave_lazy_mmu_mode();
882         pte_unmap_unlock(pte - 1, ptl);
883
884         return addr;
885 }
886
887 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
888                                 struct vm_area_struct *vma, pud_t *pud,
889                                 unsigned long addr, unsigned long end,
890                                 long *zap_work, struct zap_details *details)
891 {
892         pmd_t *pmd;
893         unsigned long next;
894
895         pmd = pmd_offset(pud, addr);
896         do {
897                 next = pmd_addr_end(addr, end);
898                 if (pmd_none_or_clear_bad(pmd)) {
899                         (*zap_work)--;
900                         continue;
901                 }
902                 next = zap_pte_range(tlb, vma, pmd, addr, next,
903                                                 zap_work, details);
904         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
905
906         return addr;
907 }
908
909 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
910                                 struct vm_area_struct *vma, pgd_t *pgd,
911                                 unsigned long addr, unsigned long end,
912                                 long *zap_work, struct zap_details *details)
913 {
914         pud_t *pud;
915         unsigned long next;
916
917         pud = pud_offset(pgd, addr);
918         do {
919                 next = pud_addr_end(addr, end);
920                 if (pud_none_or_clear_bad(pud)) {
921                         (*zap_work)--;
922                         continue;
923                 }
924                 next = zap_pmd_range(tlb, vma, pud, addr, next,
925                                                 zap_work, details);
926         } while (pud++, addr = next, (addr != end && *zap_work > 0));
927
928         return addr;
929 }
930
931 static unsigned long unmap_page_range(struct mmu_gather *tlb,
932                                 struct vm_area_struct *vma,
933                                 unsigned long addr, unsigned long end,
934                                 long *zap_work, struct zap_details *details)
935 {
936         pgd_t *pgd;
937         unsigned long next;
938
939         if (details && !details->check_mapping && !details->nonlinear_vma)
940                 details = NULL;
941
942         BUG_ON(addr >= end);
943         tlb_start_vma(tlb, vma);
944         pgd = pgd_offset(vma->vm_mm, addr);
945         do {
946                 next = pgd_addr_end(addr, end);
947                 if (pgd_none_or_clear_bad(pgd)) {
948                         (*zap_work)--;
949                         continue;
950                 }
951                 next = zap_pud_range(tlb, vma, pgd, addr, next,
952                                                 zap_work, details);
953         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
954         tlb_end_vma(tlb, vma);
955
956         return addr;
957 }
958
959 #ifdef CONFIG_PREEMPT
960 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
961 #else
962 /* No preempt: go for improved straight-line efficiency */
963 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
964 #endif
965
966 /**
967  * unmap_vmas - unmap a range of memory covered by a list of vma's
968  * @tlbp: address of the caller's struct mmu_gather
969  * @vma: the starting vma
970  * @start_addr: virtual address at which to start unmapping
971  * @end_addr: virtual address at which to end unmapping
972  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
973  * @details: details of nonlinear truncation or shared cache invalidation
974  *
975  * Returns the end address of the unmapping (restart addr if interrupted).
976  *
977  * Unmap all pages in the vma list.
978  *
979  * We aim to not hold locks for too long (for scheduling latency reasons).
980  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
981  * return the ending mmu_gather to the caller.
982  *
983  * Only addresses between `start' and `end' will be unmapped.
984  *
985  * The VMA list must be sorted in ascending virtual address order.
986  *
987  * unmap_vmas() assumes that the caller will flush the whole unmapped address
988  * range after unmap_vmas() returns.  So the only responsibility here is to
989  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
990  * drops the lock and schedules.
991  */
992 unsigned long unmap_vmas(struct mmu_gather **tlbp,
993                 struct vm_area_struct *vma, unsigned long start_addr,
994                 unsigned long end_addr, unsigned long *nr_accounted,
995                 struct zap_details *details)
996 {
997         long zap_work = ZAP_BLOCK_SIZE;
998         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
999         int tlb_start_valid = 0;
1000         unsigned long start = start_addr;
1001         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1002         int fullmm = (*tlbp)->fullmm;
1003         struct mm_struct *mm = vma->vm_mm;
1004
1005         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1006         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1007                 unsigned long end;
1008
1009                 start = max(vma->vm_start, start_addr);
1010                 if (start >= vma->vm_end)
1011                         continue;
1012                 end = min(vma->vm_end, end_addr);
1013                 if (end <= vma->vm_start)
1014                         continue;
1015
1016                 if (vma->vm_flags & VM_ACCOUNT)
1017                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1018
1019                 if (unlikely(is_pfn_mapping(vma)))
1020                         untrack_pfn_vma(vma, 0, 0);
1021
1022                 while (start != end) {
1023                         if (!tlb_start_valid) {
1024                                 tlb_start = start;
1025                                 tlb_start_valid = 1;
1026                         }
1027
1028                         if (unlikely(is_vm_hugetlb_page(vma))) {
1029                                 /*
1030                                  * It is undesirable to test vma->vm_file as it
1031                                  * should be non-null for valid hugetlb area.
1032                                  * However, vm_file will be NULL in the error
1033                                  * cleanup path of do_mmap_pgoff. When
1034                                  * hugetlbfs ->mmap method fails,
1035                                  * do_mmap_pgoff() nullifies vma->vm_file
1036                                  * before calling this function to clean up.
1037                                  * Since no pte has actually been setup, it is
1038                                  * safe to do nothing in this case.
1039                                  */
1040                                 if (vma->vm_file) {
1041                                         unmap_hugepage_range(vma, start, end, NULL);
1042                                         zap_work -= (end - start) /
1043                                         pages_per_huge_page(hstate_vma(vma));
1044                                 }
1045
1046                                 start = end;
1047                         } else
1048                                 start = unmap_page_range(*tlbp, vma,
1049                                                 start, end, &zap_work, details);
1050
1051                         if (zap_work > 0) {
1052                                 BUG_ON(start != end);
1053                                 break;
1054                         }
1055
1056                         tlb_finish_mmu(*tlbp, tlb_start, start);
1057
1058                         if (need_resched() ||
1059                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1060                                 if (i_mmap_lock) {
1061                                         *tlbp = NULL;
1062                                         goto out;
1063                                 }
1064                                 cond_resched();
1065                         }
1066
1067                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1068                         tlb_start_valid = 0;
1069                         zap_work = ZAP_BLOCK_SIZE;
1070                 }
1071         }
1072 out:
1073         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1074         return start;   /* which is now the end (or restart) address */
1075 }
1076
1077 /**
1078  * zap_page_range - remove user pages in a given range
1079  * @vma: vm_area_struct holding the applicable pages
1080  * @address: starting address of pages to zap
1081  * @size: number of bytes to zap
1082  * @details: details of nonlinear truncation or shared cache invalidation
1083  */
1084 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1085                 unsigned long size, struct zap_details *details)
1086 {
1087         struct mm_struct *mm = vma->vm_mm;
1088         struct mmu_gather *tlb;
1089         unsigned long end = address + size;
1090         unsigned long nr_accounted = 0;
1091
1092         lru_add_drain();
1093         tlb = tlb_gather_mmu(mm, 0);
1094         update_hiwater_rss(mm);
1095         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1096         if (tlb)
1097                 tlb_finish_mmu(tlb, address, end);
1098         return end;
1099 }
1100
1101 /**
1102  * zap_vma_ptes - remove ptes mapping the vma
1103  * @vma: vm_area_struct holding ptes to be zapped
1104  * @address: starting address of pages to zap
1105  * @size: number of bytes to zap
1106  *
1107  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1108  *
1109  * The entire address range must be fully contained within the vma.
1110  *
1111  * Returns 0 if successful.
1112  */
1113 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1114                 unsigned long size)
1115 {
1116         if (address < vma->vm_start || address + size > vma->vm_end ||
1117                         !(vma->vm_flags & VM_PFNMAP))
1118                 return -1;
1119         zap_page_range(vma, address, size, NULL);
1120         return 0;
1121 }
1122 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1123
1124 /*
1125  * Do a quick page-table lookup for a single page.
1126  */
1127 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1128                         unsigned int flags)
1129 {
1130         pgd_t *pgd;
1131         pud_t *pud;
1132         pmd_t *pmd;
1133         pte_t *ptep, pte;
1134         spinlock_t *ptl;
1135         struct page *page;
1136         struct mm_struct *mm = vma->vm_mm;
1137
1138         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1139         if (!IS_ERR(page)) {
1140                 BUG_ON(flags & FOLL_GET);
1141                 goto out;
1142         }
1143
1144         page = NULL;
1145         pgd = pgd_offset(mm, address);
1146         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1147                 goto no_page_table;
1148
1149         pud = pud_offset(pgd, address);
1150         if (pud_none(*pud))
1151                 goto no_page_table;
1152         if (pud_huge(*pud)) {
1153                 BUG_ON(flags & FOLL_GET);
1154                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1155                 goto out;
1156         }
1157         if (unlikely(pud_bad(*pud)))
1158                 goto no_page_table;
1159
1160         pmd = pmd_offset(pud, address);
1161         if (pmd_none(*pmd))
1162                 goto no_page_table;
1163         if (pmd_huge(*pmd)) {
1164                 BUG_ON(flags & FOLL_GET);
1165                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1166                 goto out;
1167         }
1168         if (unlikely(pmd_bad(*pmd)))
1169                 goto no_page_table;
1170
1171         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1172
1173         pte = *ptep;
1174         if (!pte_present(pte))
1175                 goto no_page;
1176         if ((flags & FOLL_WRITE) && !pte_write(pte))
1177                 goto unlock;
1178
1179         page = vm_normal_page(vma, address, pte);
1180         if (unlikely(!page)) {
1181                 if ((flags & FOLL_DUMP) ||
1182                     !is_zero_pfn(pte_pfn(pte)))
1183                         goto bad_page;
1184                 page = pte_page(pte);
1185         }
1186
1187         if (flags & FOLL_GET)
1188                 get_page(page);
1189         if (flags & FOLL_TOUCH) {
1190                 if ((flags & FOLL_WRITE) &&
1191                     !pte_dirty(pte) && !PageDirty(page))
1192                         set_page_dirty(page);
1193                 /*
1194                  * pte_mkyoung() would be more correct here, but atomic care
1195                  * is needed to avoid losing the dirty bit: it is easier to use
1196                  * mark_page_accessed().
1197                  */
1198                 mark_page_accessed(page);
1199         }
1200 unlock:
1201         pte_unmap_unlock(ptep, ptl);
1202 out:
1203         return page;
1204
1205 bad_page:
1206         pte_unmap_unlock(ptep, ptl);
1207         return ERR_PTR(-EFAULT);
1208
1209 no_page:
1210         pte_unmap_unlock(ptep, ptl);
1211         if (!pte_none(pte))
1212                 return page;
1213
1214 no_page_table:
1215         /*
1216          * When core dumping an enormous anonymous area that nobody
1217          * has touched so far, we don't want to allocate unnecessary pages or
1218          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1219          * then get_dump_page() will return NULL to leave a hole in the dump.
1220          * But we can only make this optimization where a hole would surely
1221          * be zero-filled if handle_mm_fault() actually did handle it.
1222          */
1223         if ((flags & FOLL_DUMP) &&
1224             (!vma->vm_ops || !vma->vm_ops->fault))
1225                 return ERR_PTR(-EFAULT);
1226         return page;
1227 }
1228
1229 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1230                      unsigned long start, int nr_pages, unsigned int gup_flags,
1231                      struct page **pages, struct vm_area_struct **vmas)
1232 {
1233         int i;
1234         unsigned long vm_flags;
1235
1236         if (nr_pages <= 0)
1237                 return 0;
1238
1239         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1240
1241         /* 
1242          * Require read or write permissions.
1243          * If FOLL_FORCE is set, we only require the "MAY" flags.
1244          */
1245         vm_flags  = (gup_flags & FOLL_WRITE) ?
1246                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1247         vm_flags &= (gup_flags & FOLL_FORCE) ?
1248                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1249         i = 0;
1250
1251         do {
1252                 struct vm_area_struct *vma;
1253
1254                 vma = find_extend_vma(mm, start);
1255                 if (!vma && in_gate_area(tsk, start)) {
1256                         unsigned long pg = start & PAGE_MASK;
1257                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1258                         pgd_t *pgd;
1259                         pud_t *pud;
1260                         pmd_t *pmd;
1261                         pte_t *pte;
1262
1263                         /* user gate pages are read-only */
1264                         if (gup_flags & FOLL_WRITE)
1265                                 return i ? : -EFAULT;
1266                         if (pg > TASK_SIZE)
1267                                 pgd = pgd_offset_k(pg);
1268                         else
1269                                 pgd = pgd_offset_gate(mm, pg);
1270                         BUG_ON(pgd_none(*pgd));
1271                         pud = pud_offset(pgd, pg);
1272                         BUG_ON(pud_none(*pud));
1273                         pmd = pmd_offset(pud, pg);
1274                         if (pmd_none(*pmd))
1275                                 return i ? : -EFAULT;
1276                         pte = pte_offset_map(pmd, pg);
1277                         if (pte_none(*pte)) {
1278                                 pte_unmap(pte);
1279                                 return i ? : -EFAULT;
1280                         }
1281                         if (pages) {
1282                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1283                                 pages[i] = page;
1284                                 if (page)
1285                                         get_page(page);
1286                         }
1287                         pte_unmap(pte);
1288                         if (vmas)
1289                                 vmas[i] = gate_vma;
1290                         i++;
1291                         start += PAGE_SIZE;
1292                         nr_pages--;
1293                         continue;
1294                 }
1295
1296                 if (!vma ||
1297                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1298                     !(vm_flags & vma->vm_flags))
1299                         return i ? : -EFAULT;
1300
1301                 if (is_vm_hugetlb_page(vma)) {
1302                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1303                                         &start, &nr_pages, i, gup_flags);
1304                         continue;
1305                 }
1306
1307                 do {
1308                         struct page *page;
1309                         unsigned int foll_flags = gup_flags;
1310
1311                         /*
1312                          * If we have a pending SIGKILL, don't keep faulting
1313                          * pages and potentially allocating memory.
1314                          */
1315                         if (unlikely(fatal_signal_pending(current)))
1316                                 return i ? i : -ERESTARTSYS;
1317
1318                         cond_resched();
1319                         while (!(page = follow_page(vma, start, foll_flags))) {
1320                                 int ret;
1321
1322                                 ret = handle_mm_fault(mm, vma, start,
1323                                         (foll_flags & FOLL_WRITE) ?
1324                                         FAULT_FLAG_WRITE : 0);
1325
1326                                 if (ret & VM_FAULT_ERROR) {
1327                                         if (ret & VM_FAULT_OOM)
1328                                                 return i ? i : -ENOMEM;
1329                                         if (ret &
1330                                             (VM_FAULT_HWPOISON|VM_FAULT_SIGBUS))
1331                                                 return i ? i : -EFAULT;
1332                                         BUG();
1333                                 }
1334                                 if (ret & VM_FAULT_MAJOR)
1335                                         tsk->maj_flt++;
1336                                 else
1337                                         tsk->min_flt++;
1338
1339                                 /*
1340                                  * The VM_FAULT_WRITE bit tells us that
1341                                  * do_wp_page has broken COW when necessary,
1342                                  * even if maybe_mkwrite decided not to set
1343                                  * pte_write. We can thus safely do subsequent
1344                                  * page lookups as if they were reads. But only
1345                                  * do so when looping for pte_write is futile:
1346                                  * in some cases userspace may also be wanting
1347                                  * to write to the gotten user page, which a
1348                                  * read fault here might prevent (a readonly
1349                                  * page might get reCOWed by userspace write).
1350                                  */
1351                                 if ((ret & VM_FAULT_WRITE) &&
1352                                     !(vma->vm_flags & VM_WRITE))
1353                                         foll_flags &= ~FOLL_WRITE;
1354
1355                                 cond_resched();
1356                         }
1357                         if (IS_ERR(page))
1358                                 return i ? i : PTR_ERR(page);
1359                         if (pages) {
1360                                 pages[i] = page;
1361
1362                                 flush_anon_page(vma, page, start);
1363                                 flush_dcache_page(page);
1364                         }
1365                         if (vmas)
1366                                 vmas[i] = vma;
1367                         i++;
1368                         start += PAGE_SIZE;
1369                         nr_pages--;
1370                 } while (nr_pages && start < vma->vm_end);
1371         } while (nr_pages);
1372         return i;
1373 }
1374
1375 /**
1376  * get_user_pages() - pin user pages in memory
1377  * @tsk:        task_struct of target task
1378  * @mm:         mm_struct of target mm
1379  * @start:      starting user address
1380  * @nr_pages:   number of pages from start to pin
1381  * @write:      whether pages will be written to by the caller
1382  * @force:      whether to force write access even if user mapping is
1383  *              readonly. This will result in the page being COWed even
1384  *              in MAP_SHARED mappings. You do not want this.
1385  * @pages:      array that receives pointers to the pages pinned.
1386  *              Should be at least nr_pages long. Or NULL, if caller
1387  *              only intends to ensure the pages are faulted in.
1388  * @vmas:       array of pointers to vmas corresponding to each page.
1389  *              Or NULL if the caller does not require them.
1390  *
1391  * Returns number of pages pinned. This may be fewer than the number
1392  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1393  * were pinned, returns -errno. Each page returned must be released
1394  * with a put_page() call when it is finished with. vmas will only
1395  * remain valid while mmap_sem is held.
1396  *
1397  * Must be called with mmap_sem held for read or write.
1398  *
1399  * get_user_pages walks a process's page tables and takes a reference to
1400  * each struct page that each user address corresponds to at a given
1401  * instant. That is, it takes the page that would be accessed if a user
1402  * thread accesses the given user virtual address at that instant.
1403  *
1404  * This does not guarantee that the page exists in the user mappings when
1405  * get_user_pages returns, and there may even be a completely different
1406  * page there in some cases (eg. if mmapped pagecache has been invalidated
1407  * and subsequently re faulted). However it does guarantee that the page
1408  * won't be freed completely. And mostly callers simply care that the page
1409  * contains data that was valid *at some point in time*. Typically, an IO
1410  * or similar operation cannot guarantee anything stronger anyway because
1411  * locks can't be held over the syscall boundary.
1412  *
1413  * If write=0, the page must not be written to. If the page is written to,
1414  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1415  * after the page is finished with, and before put_page is called.
1416  *
1417  * get_user_pages is typically used for fewer-copy IO operations, to get a
1418  * handle on the memory by some means other than accesses via the user virtual
1419  * addresses. The pages may be submitted for DMA to devices or accessed via
1420  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1421  * use the correct cache flushing APIs.
1422  *
1423  * See also get_user_pages_fast, for performance critical applications.
1424  */
1425 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1426                 unsigned long start, int nr_pages, int write, int force,
1427                 struct page **pages, struct vm_area_struct **vmas)
1428 {
1429         int flags = FOLL_TOUCH;
1430
1431         if (pages)
1432                 flags |= FOLL_GET;
1433         if (write)
1434                 flags |= FOLL_WRITE;
1435         if (force)
1436                 flags |= FOLL_FORCE;
1437
1438         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
1439 }
1440 EXPORT_SYMBOL(get_user_pages);
1441
1442 /**
1443  * get_dump_page() - pin user page in memory while writing it to core dump
1444  * @addr: user address
1445  *
1446  * Returns struct page pointer of user page pinned for dump,
1447  * to be freed afterwards by page_cache_release() or put_page().
1448  *
1449  * Returns NULL on any kind of failure - a hole must then be inserted into
1450  * the corefile, to preserve alignment with its headers; and also returns
1451  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1452  * allowing a hole to be left in the corefile to save diskspace.
1453  *
1454  * Called without mmap_sem, but after all other threads have been killed.
1455  */
1456 #ifdef CONFIG_ELF_CORE
1457 struct page *get_dump_page(unsigned long addr)
1458 {
1459         struct vm_area_struct *vma;
1460         struct page *page;
1461
1462         if (__get_user_pages(current, current->mm, addr, 1,
1463                         FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma) < 1)
1464                 return NULL;
1465         flush_cache_page(vma, addr, page_to_pfn(page));
1466         return page;
1467 }
1468 #endif /* CONFIG_ELF_CORE */
1469
1470 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1471                         spinlock_t **ptl)
1472 {
1473         pgd_t * pgd = pgd_offset(mm, addr);
1474         pud_t * pud = pud_alloc(mm, pgd, addr);
1475         if (pud) {
1476                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1477                 if (pmd)
1478                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1479         }
1480         return NULL;
1481 }
1482
1483 /*
1484  * This is the old fallback for page remapping.
1485  *
1486  * For historical reasons, it only allows reserved pages. Only
1487  * old drivers should use this, and they needed to mark their
1488  * pages reserved for the old functions anyway.
1489  */
1490 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1491                         struct page *page, pgprot_t prot)
1492 {
1493         struct mm_struct *mm = vma->vm_mm;
1494         int retval;
1495         pte_t *pte;
1496         spinlock_t *ptl;
1497
1498         retval = -EINVAL;
1499         if (PageAnon(page))
1500                 goto out;
1501         retval = -ENOMEM;
1502         flush_dcache_page(page);
1503         pte = get_locked_pte(mm, addr, &ptl);
1504         if (!pte)
1505                 goto out;
1506         retval = -EBUSY;
1507         if (!pte_none(*pte))
1508                 goto out_unlock;
1509
1510         /* Ok, finally just insert the thing.. */
1511         get_page(page);
1512         inc_mm_counter(mm, file_rss);
1513         page_add_file_rmap(page);
1514         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1515
1516         retval = 0;
1517         pte_unmap_unlock(pte, ptl);
1518         return retval;
1519 out_unlock:
1520         pte_unmap_unlock(pte, ptl);
1521 out:
1522         return retval;
1523 }
1524
1525 /**
1526  * vm_insert_page - insert single page into user vma
1527  * @vma: user vma to map to
1528  * @addr: target user address of this page
1529  * @page: source kernel page
1530  *
1531  * This allows drivers to insert individual pages they've allocated
1532  * into a user vma.
1533  *
1534  * The page has to be a nice clean _individual_ kernel allocation.
1535  * If you allocate a compound page, you need to have marked it as
1536  * such (__GFP_COMP), or manually just split the page up yourself
1537  * (see split_page()).
1538  *
1539  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1540  * took an arbitrary page protection parameter. This doesn't allow
1541  * that. Your vma protection will have to be set up correctly, which
1542  * means that if you want a shared writable mapping, you'd better
1543  * ask for a shared writable mapping!
1544  *
1545  * The page does not need to be reserved.
1546  */
1547 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1548                         struct page *page)
1549 {
1550         if (addr < vma->vm_start || addr >= vma->vm_end)
1551                 return -EFAULT;
1552         if (!page_count(page))
1553                 return -EINVAL;
1554         vma->vm_flags |= VM_INSERTPAGE;
1555         return insert_page(vma, addr, page, vma->vm_page_prot);
1556 }
1557 EXPORT_SYMBOL(vm_insert_page);
1558
1559 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1560                         unsigned long pfn, pgprot_t prot)
1561 {
1562         struct mm_struct *mm = vma->vm_mm;
1563         int retval;
1564         pte_t *pte, entry;
1565         spinlock_t *ptl;
1566
1567         retval = -ENOMEM;
1568         pte = get_locked_pte(mm, addr, &ptl);
1569         if (!pte)
1570                 goto out;
1571         retval = -EBUSY;
1572         if (!pte_none(*pte))
1573                 goto out_unlock;
1574
1575         /* Ok, finally just insert the thing.. */
1576         entry = pte_mkspecial(pfn_pte(pfn, prot));
1577         set_pte_at(mm, addr, pte, entry);
1578         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1579
1580         retval = 0;
1581 out_unlock:
1582         pte_unmap_unlock(pte, ptl);
1583 out:
1584         return retval;
1585 }
1586
1587 /**
1588  * vm_insert_pfn - insert single pfn into user vma
1589  * @vma: user vma to map to
1590  * @addr: target user address of this page
1591  * @pfn: source kernel pfn
1592  *
1593  * Similar to vm_inert_page, this allows drivers to insert individual pages
1594  * they've allocated into a user vma. Same comments apply.
1595  *
1596  * This function should only be called from a vm_ops->fault handler, and
1597  * in that case the handler should return NULL.
1598  *
1599  * vma cannot be a COW mapping.
1600  *
1601  * As this is called only for pages that do not currently exist, we
1602  * do not need to flush old virtual caches or the TLB.
1603  */
1604 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1605                         unsigned long pfn)
1606 {
1607         int ret;
1608         pgprot_t pgprot = vma->vm_page_prot;
1609         /*
1610          * Technically, architectures with pte_special can avoid all these
1611          * restrictions (same for remap_pfn_range).  However we would like
1612          * consistency in testing and feature parity among all, so we should
1613          * try to keep these invariants in place for everybody.
1614          */
1615         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1616         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1617                                                 (VM_PFNMAP|VM_MIXEDMAP));
1618         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1619         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1620
1621         if (addr < vma->vm_start || addr >= vma->vm_end)
1622                 return -EFAULT;
1623         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1624                 return -EINVAL;
1625
1626         ret = insert_pfn(vma, addr, pfn, pgprot);
1627
1628         if (ret)
1629                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1630
1631         return ret;
1632 }
1633 EXPORT_SYMBOL(vm_insert_pfn);
1634
1635 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1636                         unsigned long pfn)
1637 {
1638         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1639
1640         if (addr < vma->vm_start || addr >= vma->vm_end)
1641                 return -EFAULT;
1642
1643         /*
1644          * If we don't have pte special, then we have to use the pfn_valid()
1645          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1646          * refcount the page if pfn_valid is true (hence insert_page rather
1647          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1648          * without pte special, it would there be refcounted as a normal page.
1649          */
1650         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1651                 struct page *page;
1652
1653                 page = pfn_to_page(pfn);
1654                 return insert_page(vma, addr, page, vma->vm_page_prot);
1655         }
1656         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1657 }
1658 EXPORT_SYMBOL(vm_insert_mixed);
1659
1660 /*
1661  * maps a range of physical memory into the requested pages. the old
1662  * mappings are removed. any references to nonexistent pages results
1663  * in null mappings (currently treated as "copy-on-access")
1664  */
1665 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1666                         unsigned long addr, unsigned long end,
1667                         unsigned long pfn, pgprot_t prot)
1668 {
1669         pte_t *pte;
1670         spinlock_t *ptl;
1671
1672         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1673         if (!pte)
1674                 return -ENOMEM;
1675         arch_enter_lazy_mmu_mode();
1676         do {
1677                 BUG_ON(!pte_none(*pte));
1678                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1679                 pfn++;
1680         } while (pte++, addr += PAGE_SIZE, addr != end);
1681         arch_leave_lazy_mmu_mode();
1682         pte_unmap_unlock(pte - 1, ptl);
1683         return 0;
1684 }
1685
1686 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1687                         unsigned long addr, unsigned long end,
1688                         unsigned long pfn, pgprot_t prot)
1689 {
1690         pmd_t *pmd;
1691         unsigned long next;
1692
1693         pfn -= addr >> PAGE_SHIFT;
1694         pmd = pmd_alloc(mm, pud, addr);
1695         if (!pmd)
1696                 return -ENOMEM;
1697         do {
1698                 next = pmd_addr_end(addr, end);
1699                 if (remap_pte_range(mm, pmd, addr, next,
1700                                 pfn + (addr >> PAGE_SHIFT), prot))
1701                         return -ENOMEM;
1702         } while (pmd++, addr = next, addr != end);
1703         return 0;
1704 }
1705
1706 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1707                         unsigned long addr, unsigned long end,
1708                         unsigned long pfn, pgprot_t prot)
1709 {
1710         pud_t *pud;
1711         unsigned long next;
1712
1713         pfn -= addr >> PAGE_SHIFT;
1714         pud = pud_alloc(mm, pgd, addr);
1715         if (!pud)
1716                 return -ENOMEM;
1717         do {
1718                 next = pud_addr_end(addr, end);
1719                 if (remap_pmd_range(mm, pud, addr, next,
1720                                 pfn + (addr >> PAGE_SHIFT), prot))
1721                         return -ENOMEM;
1722         } while (pud++, addr = next, addr != end);
1723         return 0;
1724 }
1725
1726 /**
1727  * remap_pfn_range - remap kernel memory to userspace
1728  * @vma: user vma to map to
1729  * @addr: target user address to start at
1730  * @pfn: physical address of kernel memory
1731  * @size: size of map area
1732  * @prot: page protection flags for this mapping
1733  *
1734  *  Note: this is only safe if the mm semaphore is held when called.
1735  */
1736 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1737                     unsigned long pfn, unsigned long size, pgprot_t prot)
1738 {
1739         pgd_t *pgd;
1740         unsigned long next;
1741         unsigned long end = addr + PAGE_ALIGN(size);
1742         struct mm_struct *mm = vma->vm_mm;
1743         int err;
1744
1745         /*
1746          * Physically remapped pages are special. Tell the
1747          * rest of the world about it:
1748          *   VM_IO tells people not to look at these pages
1749          *      (accesses can have side effects).
1750          *   VM_RESERVED is specified all over the place, because
1751          *      in 2.4 it kept swapout's vma scan off this vma; but
1752          *      in 2.6 the LRU scan won't even find its pages, so this
1753          *      flag means no more than count its pages in reserved_vm,
1754          *      and omit it from core dump, even when VM_IO turned off.
1755          *   VM_PFNMAP tells the core MM that the base pages are just
1756          *      raw PFN mappings, and do not have a "struct page" associated
1757          *      with them.
1758          *
1759          * There's a horrible special case to handle copy-on-write
1760          * behaviour that some programs depend on. We mark the "original"
1761          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1762          */
1763         if (addr == vma->vm_start && end == vma->vm_end) {
1764                 vma->vm_pgoff = pfn;
1765                 vma->vm_flags |= VM_PFN_AT_MMAP;
1766         } else if (is_cow_mapping(vma->vm_flags))
1767                 return -EINVAL;
1768
1769         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1770
1771         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1772         if (err) {
1773                 /*
1774                  * To indicate that track_pfn related cleanup is not
1775                  * needed from higher level routine calling unmap_vmas
1776                  */
1777                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1778                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1779                 return -EINVAL;
1780         }
1781
1782         BUG_ON(addr >= end);
1783         pfn -= addr >> PAGE_SHIFT;
1784         pgd = pgd_offset(mm, addr);
1785         flush_cache_range(vma, addr, end);
1786         do {
1787                 next = pgd_addr_end(addr, end);
1788                 err = remap_pud_range(mm, pgd, addr, next,
1789                                 pfn + (addr >> PAGE_SHIFT), prot);
1790                 if (err)
1791                         break;
1792         } while (pgd++, addr = next, addr != end);
1793
1794         if (err)
1795                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1796
1797         return err;
1798 }
1799 EXPORT_SYMBOL(remap_pfn_range);
1800
1801 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1802                                      unsigned long addr, unsigned long end,
1803                                      pte_fn_t fn, void *data)
1804 {
1805         pte_t *pte;
1806         int err;
1807         pgtable_t token;
1808         spinlock_t *uninitialized_var(ptl);
1809
1810         pte = (mm == &init_mm) ?
1811                 pte_alloc_kernel(pmd, addr) :
1812                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1813         if (!pte)
1814                 return -ENOMEM;
1815
1816         BUG_ON(pmd_huge(*pmd));
1817
1818         arch_enter_lazy_mmu_mode();
1819
1820         token = pmd_pgtable(*pmd);
1821
1822         do {
1823                 err = fn(pte, token, addr, data);
1824                 if (err)
1825                         break;
1826         } while (pte++, addr += PAGE_SIZE, addr != end);
1827
1828         arch_leave_lazy_mmu_mode();
1829
1830         if (mm != &init_mm)
1831                 pte_unmap_unlock(pte-1, ptl);
1832         return err;
1833 }
1834
1835 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1836                                      unsigned long addr, unsigned long end,
1837                                      pte_fn_t fn, void *data)
1838 {
1839         pmd_t *pmd;
1840         unsigned long next;
1841         int err;
1842
1843         BUG_ON(pud_huge(*pud));
1844
1845         pmd = pmd_alloc(mm, pud, addr);
1846         if (!pmd)
1847                 return -ENOMEM;
1848         do {
1849                 next = pmd_addr_end(addr, end);
1850                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1851                 if (err)
1852                         break;
1853         } while (pmd++, addr = next, addr != end);
1854         return err;
1855 }
1856
1857 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1858                                      unsigned long addr, unsigned long end,
1859                                      pte_fn_t fn, void *data)
1860 {
1861         pud_t *pud;
1862         unsigned long next;
1863         int err;
1864
1865         pud = pud_alloc(mm, pgd, addr);
1866         if (!pud)
1867                 return -ENOMEM;
1868         do {
1869                 next = pud_addr_end(addr, end);
1870                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1871                 if (err)
1872                         break;
1873         } while (pud++, addr = next, addr != end);
1874         return err;
1875 }
1876
1877 /*
1878  * Scan a region of virtual memory, filling in page tables as necessary
1879  * and calling a provided function on each leaf page table.
1880  */
1881 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1882                         unsigned long size, pte_fn_t fn, void *data)
1883 {
1884         pgd_t *pgd;
1885         unsigned long next;
1886         unsigned long start = addr, end = addr + size;
1887         int err;
1888
1889         BUG_ON(addr >= end);
1890         mmu_notifier_invalidate_range_start(mm, start, end);
1891         pgd = pgd_offset(mm, addr);
1892         do {
1893                 next = pgd_addr_end(addr, end);
1894                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1895                 if (err)
1896                         break;
1897         } while (pgd++, addr = next, addr != end);
1898         mmu_notifier_invalidate_range_end(mm, start, end);
1899         return err;
1900 }
1901 EXPORT_SYMBOL_GPL(apply_to_page_range);
1902
1903 /*
1904  * handle_pte_fault chooses page fault handler according to an entry
1905  * which was read non-atomically.  Before making any commitment, on
1906  * those architectures or configurations (e.g. i386 with PAE) which
1907  * might give a mix of unmatched parts, do_swap_page and do_file_page
1908  * must check under lock before unmapping the pte and proceeding
1909  * (but do_wp_page is only called after already making such a check;
1910  * and do_anonymous_page and do_no_page can safely check later on).
1911  */
1912 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1913                                 pte_t *page_table, pte_t orig_pte)
1914 {
1915         int same = 1;
1916 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1917         if (sizeof(pte_t) > sizeof(unsigned long)) {
1918                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1919                 spin_lock(ptl);
1920                 same = pte_same(*page_table, orig_pte);
1921                 spin_unlock(ptl);
1922         }
1923 #endif
1924         pte_unmap(page_table);
1925         return same;
1926 }
1927
1928 /*
1929  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1930  * servicing faults for write access.  In the normal case, do always want
1931  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1932  * that do not have writing enabled, when used by access_process_vm.
1933  */
1934 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1935 {
1936         if (likely(vma->vm_flags & VM_WRITE))
1937                 pte = pte_mkwrite(pte);
1938         return pte;
1939 }
1940
1941 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1942 {
1943         /*
1944          * If the source page was a PFN mapping, we don't have
1945          * a "struct page" for it. We do a best-effort copy by
1946          * just copying from the original user address. If that
1947          * fails, we just zero-fill it. Live with it.
1948          */
1949         if (unlikely(!src)) {
1950                 void *kaddr = kmap_atomic(dst, KM_USER0);
1951                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1952
1953                 /*
1954                  * This really shouldn't fail, because the page is there
1955                  * in the page tables. But it might just be unreadable,
1956                  * in which case we just give up and fill the result with
1957                  * zeroes.
1958                  */
1959                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1960                         memset(kaddr, 0, PAGE_SIZE);
1961                 kunmap_atomic(kaddr, KM_USER0);
1962                 flush_dcache_page(dst);
1963         } else
1964                 copy_user_highpage(dst, src, va, vma);
1965 }
1966
1967 /*
1968  * This routine handles present pages, when users try to write
1969  * to a shared page. It is done by copying the page to a new address
1970  * and decrementing the shared-page counter for the old page.
1971  *
1972  * Note that this routine assumes that the protection checks have been
1973  * done by the caller (the low-level page fault routine in most cases).
1974  * Thus we can safely just mark it writable once we've done any necessary
1975  * COW.
1976  *
1977  * We also mark the page dirty at this point even though the page will
1978  * change only once the write actually happens. This avoids a few races,
1979  * and potentially makes it more efficient.
1980  *
1981  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1982  * but allow concurrent faults), with pte both mapped and locked.
1983  * We return with mmap_sem still held, but pte unmapped and unlocked.
1984  */
1985 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1986                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1987                 spinlock_t *ptl, pte_t orig_pte)
1988 {
1989         struct page *old_page, *new_page;
1990         pte_t entry;
1991         int reuse = 0, ret = 0;
1992         int page_mkwrite = 0;
1993         struct page *dirty_page = NULL;
1994
1995         old_page = vm_normal_page(vma, address, orig_pte);
1996         if (!old_page) {
1997                 /*
1998                  * VM_MIXEDMAP !pfn_valid() case
1999                  *
2000                  * We should not cow pages in a shared writeable mapping.
2001                  * Just mark the pages writable as we can't do any dirty
2002                  * accounting on raw pfn maps.
2003                  */
2004                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2005                                      (VM_WRITE|VM_SHARED))
2006                         goto reuse;
2007                 goto gotten;
2008         }
2009
2010         /*
2011          * Take out anonymous pages first, anonymous shared vmas are
2012          * not dirty accountable.
2013          */
2014         if (PageAnon(old_page) && !PageKsm(old_page)) {
2015                 if (!trylock_page(old_page)) {
2016                         page_cache_get(old_page);
2017                         pte_unmap_unlock(page_table, ptl);
2018                         lock_page(old_page);
2019                         page_table = pte_offset_map_lock(mm, pmd, address,
2020                                                          &ptl);
2021                         if (!pte_same(*page_table, orig_pte)) {
2022                                 unlock_page(old_page);
2023                                 page_cache_release(old_page);
2024                                 goto unlock;
2025                         }
2026                         page_cache_release(old_page);
2027                 }
2028                 reuse = reuse_swap_page(old_page);
2029                 unlock_page(old_page);
2030         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2031                                         (VM_WRITE|VM_SHARED))) {
2032                 /*
2033                  * Only catch write-faults on shared writable pages,
2034                  * read-only shared pages can get COWed by
2035                  * get_user_pages(.write=1, .force=1).
2036                  */
2037                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2038                         struct vm_fault vmf;
2039                         int tmp;
2040
2041                         vmf.virtual_address = (void __user *)(address &
2042                                                                 PAGE_MASK);
2043                         vmf.pgoff = old_page->index;
2044                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2045                         vmf.page = old_page;
2046
2047                         /*
2048                          * Notify the address space that the page is about to
2049                          * become writable so that it can prohibit this or wait
2050                          * for the page to get into an appropriate state.
2051                          *
2052                          * We do this without the lock held, so that it can
2053                          * sleep if it needs to.
2054                          */
2055                         page_cache_get(old_page);
2056                         pte_unmap_unlock(page_table, ptl);
2057
2058                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2059                         if (unlikely(tmp &
2060                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2061                                 ret = tmp;
2062                                 goto unwritable_page;
2063                         }
2064                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2065                                 lock_page(old_page);
2066                                 if (!old_page->mapping) {
2067                                         ret = 0; /* retry the fault */
2068                                         unlock_page(old_page);
2069                                         goto unwritable_page;
2070                                 }
2071                         } else
2072                                 VM_BUG_ON(!PageLocked(old_page));
2073
2074                         /*
2075                          * Since we dropped the lock we need to revalidate
2076                          * the PTE as someone else may have changed it.  If
2077                          * they did, we just return, as we can count on the
2078                          * MMU to tell us if they didn't also make it writable.
2079                          */
2080                         page_table = pte_offset_map_lock(mm, pmd, address,
2081                                                          &ptl);
2082                         if (!pte_same(*page_table, orig_pte)) {
2083                                 unlock_page(old_page);
2084                                 page_cache_release(old_page);
2085                                 goto unlock;
2086                         }
2087
2088                         page_mkwrite = 1;
2089                 }
2090                 dirty_page = old_page;
2091                 get_page(dirty_page);
2092                 reuse = 1;
2093         }
2094
2095         if (reuse) {
2096 reuse:
2097                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2098                 entry = pte_mkyoung(orig_pte);
2099                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2100                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2101                         update_mmu_cache(vma, address, entry);
2102                 ret |= VM_FAULT_WRITE;
2103                 goto unlock;
2104         }
2105
2106         /*
2107          * Ok, we need to copy. Oh, well..
2108          */
2109         page_cache_get(old_page);
2110 gotten:
2111         pte_unmap_unlock(page_table, ptl);
2112
2113         if (unlikely(anon_vma_prepare(vma)))
2114                 goto oom;
2115
2116         if (is_zero_pfn(pte_pfn(orig_pte))) {
2117                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2118                 if (!new_page)
2119                         goto oom;
2120         } else {
2121                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2122                 if (!new_page)
2123                         goto oom;
2124                 cow_user_page(new_page, old_page, address, vma);
2125         }
2126         __SetPageUptodate(new_page);
2127
2128         /*
2129          * Don't let another task, with possibly unlocked vma,
2130          * keep the mlocked page.
2131          */
2132         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2133                 lock_page(old_page);    /* for LRU manipulation */
2134                 clear_page_mlock(old_page);
2135                 unlock_page(old_page);
2136         }
2137
2138         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2139                 goto oom_free_new;
2140
2141         /*
2142          * Re-check the pte - we dropped the lock
2143          */
2144         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2145         if (likely(pte_same(*page_table, orig_pte))) {
2146                 if (old_page) {
2147                         if (!PageAnon(old_page)) {
2148                                 dec_mm_counter(mm, file_rss);
2149                                 inc_mm_counter(mm, anon_rss);
2150                         }
2151                 } else
2152                         inc_mm_counter(mm, anon_rss);
2153                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2154                 entry = mk_pte(new_page, vma->vm_page_prot);
2155                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2156                 /*
2157                  * Clear the pte entry and flush it first, before updating the
2158                  * pte with the new entry. This will avoid a race condition
2159                  * seen in the presence of one thread doing SMC and another
2160                  * thread doing COW.
2161                  */
2162                 ptep_clear_flush(vma, address, page_table);
2163                 page_add_new_anon_rmap(new_page, vma, address);
2164                 /*
2165                  * We call the notify macro here because, when using secondary
2166                  * mmu page tables (such as kvm shadow page tables), we want the
2167                  * new page to be mapped directly into the secondary page table.
2168                  */
2169                 set_pte_at_notify(mm, address, page_table, entry);
2170                 update_mmu_cache(vma, address, entry);
2171                 if (old_page) {
2172                         /*
2173                          * Only after switching the pte to the new page may
2174                          * we remove the mapcount here. Otherwise another
2175                          * process may come and find the rmap count decremented
2176                          * before the pte is switched to the new page, and
2177                          * "reuse" the old page writing into it while our pte
2178                          * here still points into it and can be read by other
2179                          * threads.
2180                          *
2181                          * The critical issue is to order this
2182                          * page_remove_rmap with the ptp_clear_flush above.
2183                          * Those stores are ordered by (if nothing else,)
2184                          * the barrier present in the atomic_add_negative
2185                          * in page_remove_rmap.
2186                          *
2187                          * Then the TLB flush in ptep_clear_flush ensures that
2188                          * no process can access the old page before the
2189                          * decremented mapcount is visible. And the old page
2190                          * cannot be reused until after the decremented
2191                          * mapcount is visible. So transitively, TLBs to
2192                          * old page will be flushed before it can be reused.
2193                          */
2194                         page_remove_rmap(old_page);
2195                 }
2196
2197                 /* Free the old page.. */
2198                 new_page = old_page;
2199                 ret |= VM_FAULT_WRITE;
2200         } else
2201                 mem_cgroup_uncharge_page(new_page);
2202
2203         if (new_page)
2204                 page_cache_release(new_page);
2205         if (old_page)
2206                 page_cache_release(old_page);
2207 unlock:
2208         pte_unmap_unlock(page_table, ptl);
2209         if (dirty_page) {
2210                 /*
2211                  * Yes, Virginia, this is actually required to prevent a race
2212                  * with clear_page_dirty_for_io() from clearing the page dirty
2213                  * bit after it clear all dirty ptes, but before a racing
2214                  * do_wp_page installs a dirty pte.
2215                  *
2216                  * do_no_page is protected similarly.
2217                  */
2218                 if (!page_mkwrite) {
2219                         wait_on_page_locked(dirty_page);
2220                         set_page_dirty_balance(dirty_page, page_mkwrite);
2221                 }
2222                 put_page(dirty_page);
2223                 if (page_mkwrite) {
2224                         struct address_space *mapping = dirty_page->mapping;
2225
2226                         set_page_dirty(dirty_page);
2227                         unlock_page(dirty_page);
2228                         page_cache_release(dirty_page);
2229                         if (mapping)    {
2230                                 /*
2231                                  * Some device drivers do not set page.mapping
2232                                  * but still dirty their pages
2233                                  */
2234                                 balance_dirty_pages_ratelimited(mapping);
2235                         }
2236                 }
2237
2238                 /* file_update_time outside page_lock */
2239                 if (vma->vm_file)
2240                         file_update_time(vma->vm_file);
2241         }
2242         return ret;
2243 oom_free_new:
2244         page_cache_release(new_page);
2245 oom:
2246         if (old_page) {
2247                 if (page_mkwrite) {
2248                         unlock_page(old_page);
2249                         page_cache_release(old_page);
2250                 }
2251                 page_cache_release(old_page);
2252         }
2253         return VM_FAULT_OOM;
2254
2255 unwritable_page:
2256         page_cache_release(old_page);
2257         return ret;
2258 }
2259
2260 /*
2261  * Helper functions for unmap_mapping_range().
2262  *
2263  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2264  *
2265  * We have to restart searching the prio_tree whenever we drop the lock,
2266  * since the iterator is only valid while the lock is held, and anyway
2267  * a later vma might be split and reinserted earlier while lock dropped.
2268  *
2269  * The list of nonlinear vmas could be handled more efficiently, using
2270  * a placeholder, but handle it in the same way until a need is shown.
2271  * It is important to search the prio_tree before nonlinear list: a vma
2272  * may become nonlinear and be shifted from prio_tree to nonlinear list
2273  * while the lock is dropped; but never shifted from list to prio_tree.
2274  *
2275  * In order to make forward progress despite restarting the search,
2276  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2277  * quickly skip it next time around.  Since the prio_tree search only
2278  * shows us those vmas affected by unmapping the range in question, we
2279  * can't efficiently keep all vmas in step with mapping->truncate_count:
2280  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2281  * mapping->truncate_count and vma->vm_truncate_count are protected by
2282  * i_mmap_lock.
2283  *
2284  * In order to make forward progress despite repeatedly restarting some
2285  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2286  * and restart from that address when we reach that vma again.  It might
2287  * have been split or merged, shrunk or extended, but never shifted: so
2288  * restart_addr remains valid so long as it remains in the vma's range.
2289  * unmap_mapping_range forces truncate_count to leap over page-aligned
2290  * values so we can save vma's restart_addr in its truncate_count field.
2291  */
2292 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2293
2294 static void reset_vma_truncate_counts(struct address_space *mapping)
2295 {
2296         struct vm_area_struct *vma;
2297         struct prio_tree_iter iter;
2298
2299         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2300                 vma->vm_truncate_count = 0;
2301         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2302                 vma->vm_truncate_count = 0;
2303 }
2304
2305 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2306                 unsigned long start_addr, unsigned long end_addr,
2307                 struct zap_details *details)
2308 {
2309         unsigned long restart_addr;
2310         int need_break;
2311
2312         /*
2313          * files that support invalidating or truncating portions of the
2314          * file from under mmaped areas must have their ->fault function
2315          * return a locked page (and set VM_FAULT_LOCKED in the return).
2316          * This provides synchronisation against concurrent unmapping here.
2317          */
2318
2319 again:
2320         restart_addr = vma->vm_truncate_count;
2321         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2322                 start_addr = restart_addr;
2323                 if (start_addr >= end_addr) {
2324                         /* Top of vma has been split off since last time */
2325                         vma->vm_truncate_count = details->truncate_count;
2326                         return 0;
2327                 }
2328         }
2329
2330         restart_addr = zap_page_range(vma, start_addr,
2331                                         end_addr - start_addr, details);
2332         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2333
2334         if (restart_addr >= end_addr) {
2335                 /* We have now completed this vma: mark it so */
2336                 vma->vm_truncate_count = details->truncate_count;
2337                 if (!need_break)
2338                         return 0;
2339         } else {
2340                 /* Note restart_addr in vma's truncate_count field */
2341                 vma->vm_truncate_count = restart_addr;
2342                 if (!need_break)
2343                         goto again;
2344         }
2345
2346         spin_unlock(details->i_mmap_lock);
2347         cond_resched();
2348         spin_lock(details->i_mmap_lock);
2349         return -EINTR;
2350 }
2351
2352 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2353                                             struct zap_details *details)
2354 {
2355         struct vm_area_struct *vma;
2356         struct prio_tree_iter iter;
2357         pgoff_t vba, vea, zba, zea;
2358
2359 restart:
2360         vma_prio_tree_foreach(vma, &iter, root,
2361                         details->first_index, details->last_index) {
2362                 /* Skip quickly over those we have already dealt with */
2363                 if (vma->vm_truncate_count == details->truncate_count)
2364                         continue;
2365
2366                 vba = vma->vm_pgoff;
2367                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2368                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2369                 zba = details->first_index;
2370                 if (zba < vba)
2371                         zba = vba;
2372                 zea = details->last_index;
2373                 if (zea > vea)
2374                         zea = vea;
2375
2376                 if (unmap_mapping_range_vma(vma,
2377                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2378                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2379                                 details) < 0)
2380                         goto restart;
2381         }
2382 }
2383
2384 static inline void unmap_mapping_range_list(struct list_head *head,
2385                                             struct zap_details *details)
2386 {
2387         struct vm_area_struct *vma;
2388
2389         /*
2390          * In nonlinear VMAs there is no correspondence between virtual address
2391          * offset and file offset.  So we must perform an exhaustive search
2392          * across *all* the pages in each nonlinear VMA, not just the pages
2393          * whose virtual address lies outside the file truncation point.
2394          */
2395 restart:
2396         list_for_each_entry(vma, head, shared.vm_set.list) {
2397                 /* Skip quickly over those we have already dealt with */
2398                 if (vma->vm_truncate_count == details->truncate_count)
2399                         continue;
2400                 details->nonlinear_vma = vma;
2401                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2402                                         vma->vm_end, details) < 0)
2403                         goto restart;
2404         }
2405 }
2406
2407 /**
2408  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2409  * @mapping: the address space containing mmaps to be unmapped.
2410  * @holebegin: byte in first page to unmap, relative to the start of
2411  * the underlying file.  This will be rounded down to a PAGE_SIZE
2412  * boundary.  Note that this is different from truncate_pagecache(), which
2413  * must keep the partial page.  In contrast, we must get rid of
2414  * partial pages.
2415  * @holelen: size of prospective hole in bytes.  This will be rounded
2416  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2417  * end of the file.
2418  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2419  * but 0 when invalidating pagecache, don't throw away private data.
2420  */
2421 void unmap_mapping_range(struct address_space *mapping,
2422                 loff_t const holebegin, loff_t const holelen, int even_cows)
2423 {
2424         struct zap_details details;
2425         pgoff_t hba = holebegin >> PAGE_SHIFT;
2426         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2427
2428         /* Check for overflow. */
2429         if (sizeof(holelen) > sizeof(hlen)) {
2430                 long long holeend =
2431                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2432                 if (holeend & ~(long long)ULONG_MAX)
2433                         hlen = ULONG_MAX - hba + 1;
2434         }
2435
2436         details.check_mapping = even_cows? NULL: mapping;
2437         details.nonlinear_vma = NULL;
2438         details.first_index = hba;
2439         details.last_index = hba + hlen - 1;
2440         if (details.last_index < details.first_index)
2441                 details.last_index = ULONG_MAX;
2442         details.i_mmap_lock = &mapping->i_mmap_lock;
2443
2444         spin_lock(&mapping->i_mmap_lock);
2445
2446         /* Protect against endless unmapping loops */
2447         mapping->truncate_count++;
2448         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2449                 if (mapping->truncate_count == 0)
2450                         reset_vma_truncate_counts(mapping);
2451                 mapping->truncate_count++;
2452         }
2453         details.truncate_count = mapping->truncate_count;
2454
2455         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2456                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2457         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2458                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2459         spin_unlock(&mapping->i_mmap_lock);
2460 }
2461 EXPORT_SYMBOL(unmap_mapping_range);
2462
2463 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2464 {
2465         struct address_space *mapping = inode->i_mapping;
2466
2467         /*
2468          * If the underlying filesystem is not going to provide
2469          * a way to truncate a range of blocks (punch a hole) -
2470          * we should return failure right now.
2471          */
2472         if (!inode->i_op->truncate_range)
2473                 return -ENOSYS;
2474
2475         mutex_lock(&inode->i_mutex);
2476         down_write(&inode->i_alloc_sem);
2477         unmap_mapping_range(mapping, offset, (end - offset), 1);
2478         truncate_inode_pages_range(mapping, offset, end);
2479         unmap_mapping_range(mapping, offset, (end - offset), 1);
2480         inode->i_op->truncate_range(inode, offset, end);
2481         up_write(&inode->i_alloc_sem);
2482         mutex_unlock(&inode->i_mutex);
2483
2484         return 0;
2485 }
2486
2487 /*
2488  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2489  * but allow concurrent faults), and pte mapped but not yet locked.
2490  * We return with mmap_sem still held, but pte unmapped and unlocked.
2491  */
2492 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2493                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2494                 unsigned int flags, pte_t orig_pte)
2495 {
2496         spinlock_t *ptl;
2497         struct page *page;
2498         swp_entry_t entry;
2499         pte_t pte;
2500         struct mem_cgroup *ptr = NULL;
2501         int ret = 0;
2502
2503         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2504                 goto out;
2505
2506         entry = pte_to_swp_entry(orig_pte);
2507         if (unlikely(non_swap_entry(entry))) {
2508                 if (is_migration_entry(entry)) {
2509                         migration_entry_wait(mm, pmd, address);
2510                 } else if (is_hwpoison_entry(entry)) {
2511                         ret = VM_FAULT_HWPOISON;
2512                 } else {
2513                         print_bad_pte(vma, address, orig_pte, NULL);
2514                         ret = VM_FAULT_OOM;
2515                 }
2516                 goto out;
2517         }
2518         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2519         page = lookup_swap_cache(entry);
2520         if (!page) {
2521                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2522                 page = swapin_readahead(entry,
2523                                         GFP_HIGHUSER_MOVABLE, vma, address);
2524                 if (!page) {
2525                         /*
2526                          * Back out if somebody else faulted in this pte
2527                          * while we released the pte lock.
2528                          */
2529                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2530                         if (likely(pte_same(*page_table, orig_pte)))
2531                                 ret = VM_FAULT_OOM;
2532                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2533                         goto unlock;
2534                 }
2535
2536                 /* Had to read the page from swap area: Major fault */
2537                 ret = VM_FAULT_MAJOR;
2538                 count_vm_event(PGMAJFAULT);
2539         } else if (PageHWPoison(page)) {
2540                 ret = VM_FAULT_HWPOISON;
2541                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2542                 goto out_release;
2543         }
2544
2545         lock_page(page);
2546         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2547
2548         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2549                 ret = VM_FAULT_OOM;
2550                 goto out_page;
2551         }
2552
2553         /*
2554          * Back out if somebody else already faulted in this pte.
2555          */
2556         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2557         if (unlikely(!pte_same(*page_table, orig_pte)))
2558                 goto out_nomap;
2559
2560         if (unlikely(!PageUptodate(page))) {
2561                 ret = VM_FAULT_SIGBUS;
2562                 goto out_nomap;
2563         }
2564
2565         /*
2566          * The page isn't present yet, go ahead with the fault.
2567          *
2568          * Be careful about the sequence of operations here.
2569          * To get its accounting right, reuse_swap_page() must be called
2570          * while the page is counted on swap but not yet in mapcount i.e.
2571          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2572          * must be called after the swap_free(), or it will never succeed.
2573          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2574          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2575          * in page->private. In this case, a record in swap_cgroup  is silently
2576          * discarded at swap_free().
2577          */
2578
2579         inc_mm_counter(mm, anon_rss);
2580         pte = mk_pte(page, vma->vm_page_prot);
2581         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2582                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2583                 flags &= ~FAULT_FLAG_WRITE;
2584         }
2585         flush_icache_page(vma, page);
2586         set_pte_at(mm, address, page_table, pte);
2587         page_add_anon_rmap(page, vma, address);
2588         /* It's better to call commit-charge after rmap is established */
2589         mem_cgroup_commit_charge_swapin(page, ptr);
2590
2591         swap_free(entry);
2592         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2593                 try_to_free_swap(page);
2594         unlock_page(page);
2595
2596         if (flags & FAULT_FLAG_WRITE) {
2597                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2598                 if (ret & VM_FAULT_ERROR)
2599                         ret &= VM_FAULT_ERROR;
2600                 goto out;
2601         }
2602
2603         /* No need to invalidate - it was non-present before */
2604         update_mmu_cache(vma, address, pte);
2605 unlock:
2606         pte_unmap_unlock(page_table, ptl);
2607 out:
2608         return ret;
2609 out_nomap:
2610         mem_cgroup_cancel_charge_swapin(ptr);
2611         pte_unmap_unlock(page_table, ptl);
2612 out_page:
2613         unlock_page(page);
2614 out_release:
2615         page_cache_release(page);
2616         return ret;
2617 }
2618
2619 /*
2620  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2621  * but allow concurrent faults), and pte mapped but not yet locked.
2622  * We return with mmap_sem still held, but pte unmapped and unlocked.
2623  */
2624 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2625                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2626                 unsigned int flags)
2627 {
2628         struct page *page;
2629         spinlock_t *ptl;
2630         pte_t entry;
2631
2632         if (!(flags & FAULT_FLAG_WRITE)) {
2633                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2634                                                 vma->vm_page_prot));
2635                 ptl = pte_lockptr(mm, pmd);
2636                 spin_lock(ptl);
2637                 if (!pte_none(*page_table))
2638                         goto unlock;
2639                 goto setpte;
2640         }
2641
2642         /* Allocate our own private page. */
2643         pte_unmap(page_table);
2644
2645         if (unlikely(anon_vma_prepare(vma)))
2646                 goto oom;
2647         page = alloc_zeroed_user_highpage_movable(vma, address);
2648         if (!page)
2649                 goto oom;
2650         __SetPageUptodate(page);
2651
2652         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2653                 goto oom_free_page;
2654
2655         entry = mk_pte(page, vma->vm_page_prot);
2656         if (vma->vm_flags & VM_WRITE)
2657                 entry = pte_mkwrite(pte_mkdirty(entry));
2658
2659         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2660         if (!pte_none(*page_table))
2661                 goto release;
2662
2663         inc_mm_counter(mm, anon_rss);
2664         page_add_new_anon_rmap(page, vma, address);
2665 setpte:
2666         set_pte_at(mm, address, page_table, entry);
2667
2668         /* No need to invalidate - it was non-present before */
2669         update_mmu_cache(vma, address, entry);
2670 unlock:
2671         pte_unmap_unlock(page_table, ptl);
2672         return 0;
2673 release:
2674         mem_cgroup_uncharge_page(page);
2675         page_cache_release(page);
2676         goto unlock;
2677 oom_free_page:
2678         page_cache_release(page);
2679 oom:
2680         return VM_FAULT_OOM;
2681 }
2682
2683 /*
2684  * __do_fault() tries to create a new page mapping. It aggressively
2685  * tries to share with existing pages, but makes a separate copy if
2686  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2687  * the next page fault.
2688  *
2689  * As this is called only for pages that do not currently exist, we
2690  * do not need to flush old virtual caches or the TLB.
2691  *
2692  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2693  * but allow concurrent faults), and pte neither mapped nor locked.
2694  * We return with mmap_sem still held, but pte unmapped and unlocked.
2695  */
2696 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2697                 unsigned long address, pmd_t *pmd,
2698                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2699 {
2700         pte_t *page_table;
2701         spinlock_t *ptl;
2702         struct page *page;
2703         pte_t entry;
2704         int anon = 0;
2705         int charged = 0;
2706         struct page *dirty_page = NULL;
2707         struct vm_fault vmf;
2708         int ret;
2709         int page_mkwrite = 0;
2710
2711         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2712         vmf.pgoff = pgoff;
2713         vmf.flags = flags;
2714         vmf.page = NULL;
2715
2716         ret = vma->vm_ops->fault(vma, &vmf);
2717         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2718                 return ret;
2719
2720         if (unlikely(PageHWPoison(vmf.page))) {
2721                 if (ret & VM_FAULT_LOCKED)
2722                         unlock_page(vmf.page);
2723                 return VM_FAULT_HWPOISON;
2724         }
2725
2726         /*
2727          * For consistency in subsequent calls, make the faulted page always
2728          * locked.
2729          */
2730         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2731                 lock_page(vmf.page);
2732         else
2733                 VM_BUG_ON(!PageLocked(vmf.page));
2734
2735         /*
2736          * Should we do an early C-O-W break?
2737          */
2738         page = vmf.page;
2739         if (flags & FAULT_FLAG_WRITE) {
2740                 if (!(vma->vm_flags & VM_SHARED)) {
2741                         anon = 1;
2742                         if (unlikely(anon_vma_prepare(vma))) {
2743                                 ret = VM_FAULT_OOM;
2744                                 goto out;
2745                         }
2746                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2747                                                 vma, address);
2748                         if (!page) {
2749                                 ret = VM_FAULT_OOM;
2750                                 goto out;
2751                         }
2752                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2753                                 ret = VM_FAULT_OOM;
2754                                 page_cache_release(page);
2755                                 goto out;
2756                         }
2757                         charged = 1;
2758                         /*
2759                          * Don't let another task, with possibly unlocked vma,
2760                          * keep the mlocked page.
2761                          */
2762                         if (vma->vm_flags & VM_LOCKED)
2763                                 clear_page_mlock(vmf.page);
2764                         copy_user_highpage(page, vmf.page, address, vma);
2765                         __SetPageUptodate(page);
2766                 } else {
2767                         /*
2768                          * If the page will be shareable, see if the backing
2769                          * address space wants to know that the page is about
2770                          * to become writable
2771                          */
2772                         if (vma->vm_ops->page_mkwrite) {
2773                                 int tmp;
2774
2775                                 unlock_page(page);
2776                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2777                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2778                                 if (unlikely(tmp &
2779                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2780                                         ret = tmp;
2781                                         goto unwritable_page;
2782                                 }
2783                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2784                                         lock_page(page);
2785                                         if (!page->mapping) {
2786                                                 ret = 0; /* retry the fault */
2787                                                 unlock_page(page);
2788                                                 goto unwritable_page;
2789                                         }
2790                                 } else
2791                                         VM_BUG_ON(!PageLocked(page));
2792                                 page_mkwrite = 1;
2793                         }
2794                 }
2795
2796         }
2797
2798         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2799
2800         /*
2801          * This silly early PAGE_DIRTY setting removes a race
2802          * due to the bad i386 page protection. But it's valid
2803          * for other architectures too.
2804          *
2805          * Note that if FAULT_FLAG_WRITE is set, we either now have
2806          * an exclusive copy of the page, or this is a shared mapping,
2807          * so we can make it writable and dirty to avoid having to
2808          * handle that later.
2809          */
2810         /* Only go through if we didn't race with anybody else... */
2811         if (likely(pte_same(*page_table, orig_pte))) {
2812                 flush_icache_page(vma, page);
2813                 entry = mk_pte(page, vma->vm_page_prot);
2814                 if (flags & FAULT_FLAG_WRITE)
2815                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2816                 if (anon) {
2817                         inc_mm_counter(mm, anon_rss);
2818                         page_add_new_anon_rmap(page, vma, address);
2819                 } else {
2820                         inc_mm_counter(mm, file_rss);
2821                         page_add_file_rmap(page);
2822                         if (flags & FAULT_FLAG_WRITE) {
2823                                 dirty_page = page;
2824                                 get_page(dirty_page);
2825                         }
2826                 }
2827                 set_pte_at(mm, address, page_table, entry);
2828
2829                 /* no need to invalidate: a not-present page won't be cached */
2830                 update_mmu_cache(vma, address, entry);
2831         } else {
2832                 if (charged)
2833                         mem_cgroup_uncharge_page(page);
2834                 if (anon)
2835                         page_cache_release(page);
2836                 else
2837                         anon = 1; /* no anon but release faulted_page */
2838         }
2839
2840         pte_unmap_unlock(page_table, ptl);
2841
2842 out:
2843         if (dirty_page) {
2844                 struct address_space *mapping = page->mapping;
2845
2846                 if (set_page_dirty(dirty_page))
2847                         page_mkwrite = 1;
2848                 unlock_page(dirty_page);
2849                 put_page(dirty_page);
2850                 if (page_mkwrite && mapping) {
2851                         /*
2852                          * Some device drivers do not set page.mapping but still
2853                          * dirty their pages
2854                          */
2855                         balance_dirty_pages_ratelimited(mapping);
2856                 }
2857
2858                 /* file_update_time outside page_lock */
2859                 if (vma->vm_file)
2860                         file_update_time(vma->vm_file);
2861         } else {
2862                 unlock_page(vmf.page);
2863                 if (anon)
2864                         page_cache_release(vmf.page);
2865         }
2866
2867         return ret;
2868
2869 unwritable_page:
2870         page_cache_release(page);
2871         return ret;
2872 }
2873
2874 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2875                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2876                 unsigned int flags, pte_t orig_pte)
2877 {
2878         pgoff_t pgoff = (((address & PAGE_MASK)
2879                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2880
2881         pte_unmap(page_table);
2882         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2883 }
2884
2885 /*
2886  * Fault of a previously existing named mapping. Repopulate the pte
2887  * from the encoded file_pte if possible. This enables swappable
2888  * nonlinear vmas.
2889  *
2890  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2891  * but allow concurrent faults), and pte mapped but not yet locked.
2892  * We return with mmap_sem still held, but pte unmapped and unlocked.
2893  */
2894 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2895                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2896                 unsigned int flags, pte_t orig_pte)
2897 {
2898         pgoff_t pgoff;
2899
2900         flags |= FAULT_FLAG_NONLINEAR;
2901
2902         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2903                 return 0;
2904
2905         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2906                 /*
2907                  * Page table corrupted: show pte and kill process.
2908                  */
2909                 print_bad_pte(vma, address, orig_pte, NULL);
2910                 return VM_FAULT_OOM;
2911         }
2912
2913         pgoff = pte_to_pgoff(orig_pte);
2914         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2915 }
2916
2917 /*
2918  * These routines also need to handle stuff like marking pages dirty
2919  * and/or accessed for architectures that don't do it in hardware (most
2920  * RISC architectures).  The early dirtying is also good on the i386.
2921  *
2922  * There is also a hook called "update_mmu_cache()" that architectures
2923  * with external mmu caches can use to update those (ie the Sparc or
2924  * PowerPC hashed page tables that act as extended TLBs).
2925  *
2926  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2927  * but allow concurrent faults), and pte mapped but not yet locked.
2928  * We return with mmap_sem still held, but pte unmapped and unlocked.
2929  */
2930 static inline int handle_pte_fault(struct mm_struct *mm,
2931                 struct vm_area_struct *vma, unsigned long address,
2932                 pte_t *pte, pmd_t *pmd, unsigned int flags)
2933 {
2934         pte_t entry;
2935         spinlock_t *ptl;
2936
2937         entry = *pte;
2938         if (!pte_present(entry)) {
2939                 if (pte_none(entry)) {
2940                         if (vma->vm_ops) {
2941                                 if (likely(vma->vm_ops->fault))
2942                                         return do_linear_fault(mm, vma, address,
2943                                                 pte, pmd, flags, entry);
2944                         }
2945                         return do_anonymous_page(mm, vma, address,
2946                                                  pte, pmd, flags);
2947                 }
2948                 if (pte_file(entry))
2949                         return do_nonlinear_fault(mm, vma, address,
2950                                         pte, pmd, flags, entry);
2951                 return do_swap_page(mm, vma, address,
2952                                         pte, pmd, flags, entry);
2953         }
2954
2955         ptl = pte_lockptr(mm, pmd);
2956         spin_lock(ptl);
2957         if (unlikely(!pte_same(*pte, entry)))
2958                 goto unlock;
2959         if (flags & FAULT_FLAG_WRITE) {
2960                 if (!pte_write(entry))
2961                         return do_wp_page(mm, vma, address,
2962                                         pte, pmd, ptl, entry);
2963                 entry = pte_mkdirty(entry);
2964         }
2965         entry = pte_mkyoung(entry);
2966         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
2967                 update_mmu_cache(vma, address, entry);
2968         } else {
2969                 /*
2970                  * This is needed only for protection faults but the arch code
2971                  * is not yet telling us if this is a protection fault or not.
2972                  * This still avoids useless tlb flushes for .text page faults
2973                  * with threads.
2974                  */
2975                 if (flags & FAULT_FLAG_WRITE)
2976                         flush_tlb_page(vma, address);
2977         }
2978 unlock:
2979         pte_unmap_unlock(pte, ptl);
2980         return 0;
2981 }
2982
2983 /*
2984  * By the time we get here, we already hold the mm semaphore
2985  */
2986 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2987                 unsigned long address, unsigned int flags)
2988 {
2989         pgd_t *pgd;
2990         pud_t *pud;
2991         pmd_t *pmd;
2992         pte_t *pte;
2993
2994         __set_current_state(TASK_RUNNING);
2995
2996         count_vm_event(PGFAULT);
2997
2998         if (unlikely(is_vm_hugetlb_page(vma)))
2999                 return hugetlb_fault(mm, vma, address, flags);
3000
3001         pgd = pgd_offset(mm, address);
3002         pud = pud_alloc(mm, pgd, address);
3003         if (!pud)
3004                 return VM_FAULT_OOM;
3005         pmd = pmd_alloc(mm, pud, address);
3006         if (!pmd)
3007                 return VM_FAULT_OOM;
3008         pte = pte_alloc_map(mm, pmd, address);
3009         if (!pte)
3010                 return VM_FAULT_OOM;
3011
3012         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3013 }
3014
3015 #ifndef __PAGETABLE_PUD_FOLDED
3016 /*
3017  * Allocate page upper directory.
3018  * We've already handled the fast-path in-line.
3019  */
3020 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3021 {
3022         pud_t *new = pud_alloc_one(mm, address);
3023         if (!new)
3024                 return -ENOMEM;
3025
3026         smp_wmb(); /* See comment in __pte_alloc */
3027
3028         spin_lock(&mm->page_table_lock);
3029         if (pgd_present(*pgd))          /* Another has populated it */
3030                 pud_free(mm, new);
3031         else
3032                 pgd_populate(mm, pgd, new);
3033         spin_unlock(&mm->page_table_lock);
3034         return 0;
3035 }
3036 #endif /* __PAGETABLE_PUD_FOLDED */
3037
3038 #ifndef __PAGETABLE_PMD_FOLDED
3039 /*
3040  * Allocate page middle directory.
3041  * We've already handled the fast-path in-line.
3042  */
3043 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3044 {
3045         pmd_t *new = pmd_alloc_one(mm, address);
3046         if (!new)
3047                 return -ENOMEM;
3048
3049         smp_wmb(); /* See comment in __pte_alloc */
3050
3051         spin_lock(&mm->page_table_lock);
3052 #ifndef __ARCH_HAS_4LEVEL_HACK
3053         if (pud_present(*pud))          /* Another has populated it */
3054                 pmd_free(mm, new);
3055         else
3056                 pud_populate(mm, pud, new);
3057 #else
3058         if (pgd_present(*pud))          /* Another has populated it */
3059                 pmd_free(mm, new);
3060         else
3061                 pgd_populate(mm, pud, new);
3062 #endif /* __ARCH_HAS_4LEVEL_HACK */
3063         spin_unlock(&mm->page_table_lock);
3064         return 0;
3065 }
3066 #endif /* __PAGETABLE_PMD_FOLDED */
3067
3068 int make_pages_present(unsigned long addr, unsigned long end)
3069 {
3070         int ret, len, write;
3071         struct vm_area_struct * vma;
3072
3073         vma = find_vma(current->mm, addr);
3074         if (!vma)
3075                 return -ENOMEM;
3076         write = (vma->vm_flags & VM_WRITE) != 0;
3077         BUG_ON(addr >= end);
3078         BUG_ON(end > vma->vm_end);
3079         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3080         ret = get_user_pages(current, current->mm, addr,
3081                         len, write, 0, NULL, NULL);
3082         if (ret < 0)
3083                 return ret;
3084         return ret == len ? 0 : -EFAULT;
3085 }
3086
3087 #if !defined(__HAVE_ARCH_GATE_AREA)
3088
3089 #if defined(AT_SYSINFO_EHDR)
3090 static struct vm_area_struct gate_vma;
3091
3092 static int __init gate_vma_init(void)
3093 {
3094         gate_vma.vm_mm = NULL;
3095         gate_vma.vm_start = FIXADDR_USER_START;
3096         gate_vma.vm_end = FIXADDR_USER_END;
3097         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3098         gate_vma.vm_page_prot = __P101;
3099         /*
3100          * Make sure the vDSO gets into every core dump.
3101          * Dumping its contents makes post-mortem fully interpretable later
3102          * without matching up the same kernel and hardware config to see
3103          * what PC values meant.
3104          */
3105         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3106         return 0;
3107 }
3108 __initcall(gate_vma_init);
3109 #endif
3110
3111 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3112 {
3113 #ifdef AT_SYSINFO_EHDR
3114         return &gate_vma;
3115 #else
3116         return NULL;
3117 #endif
3118 }
3119
3120 int in_gate_area_no_task(unsigned long addr)
3121 {
3122 #ifdef AT_SYSINFO_EHDR
3123         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3124                 return 1;
3125 #endif
3126         return 0;
3127 }
3128
3129 #endif  /* __HAVE_ARCH_GATE_AREA */
3130
3131 static int follow_pte(struct mm_struct *mm, unsigned long address,
3132                 pte_t **ptepp, spinlock_t **ptlp)
3133 {
3134         pgd_t *pgd;
3135         pud_t *pud;
3136         pmd_t *pmd;
3137         pte_t *ptep;
3138
3139         pgd = pgd_offset(mm, address);
3140         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3141                 goto out;
3142
3143         pud = pud_offset(pgd, address);
3144         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3145                 goto out;
3146
3147         pmd = pmd_offset(pud, address);
3148         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3149                 goto out;
3150
3151         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3152         if (pmd_huge(*pmd))
3153                 goto out;
3154
3155         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3156         if (!ptep)
3157                 goto out;
3158         if (!pte_present(*ptep))
3159                 goto unlock;
3160         *ptepp = ptep;
3161         return 0;
3162 unlock:
3163         pte_unmap_unlock(ptep, *ptlp);
3164 out:
3165         return -EINVAL;
3166 }
3167
3168 /**
3169  * follow_pfn - look up PFN at a user virtual address
3170  * @vma: memory mapping
3171  * @address: user virtual address
3172  * @pfn: location to store found PFN
3173  *
3174  * Only IO mappings and raw PFN mappings are allowed.
3175  *
3176  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3177  */
3178 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3179         unsigned long *pfn)
3180 {
3181         int ret = -EINVAL;
3182         spinlock_t *ptl;
3183         pte_t *ptep;
3184
3185         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3186                 return ret;
3187
3188         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3189         if (ret)
3190                 return ret;
3191         *pfn = pte_pfn(*ptep);
3192         pte_unmap_unlock(ptep, ptl);
3193         return 0;
3194 }
3195 EXPORT_SYMBOL(follow_pfn);
3196
3197 #ifdef CONFIG_HAVE_IOREMAP_PROT
3198 int follow_phys(struct vm_area_struct *vma,
3199                 unsigned long address, unsigned int flags,
3200                 unsigned long *prot, resource_size_t *phys)
3201 {
3202         int ret = -EINVAL;
3203         pte_t *ptep, pte;
3204         spinlock_t *ptl;
3205
3206         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3207                 goto out;
3208
3209         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3210                 goto out;
3211         pte = *ptep;
3212
3213         if ((flags & FOLL_WRITE) && !pte_write(pte))
3214                 goto unlock;
3215
3216         *prot = pgprot_val(pte_pgprot(pte));
3217         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3218
3219         ret = 0;
3220 unlock:
3221         pte_unmap_unlock(ptep, ptl);
3222 out:
3223         return ret;
3224 }
3225
3226 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3227                         void *buf, int len, int write)
3228 {
3229         resource_size_t phys_addr;
3230         unsigned long prot = 0;
3231         void __iomem *maddr;
3232         int offset = addr & (PAGE_SIZE-1);
3233
3234         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3235                 return -EINVAL;
3236
3237         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3238         if (write)
3239                 memcpy_toio(maddr + offset, buf, len);
3240         else
3241                 memcpy_fromio(buf, maddr + offset, len);
3242         iounmap(maddr);
3243
3244         return len;
3245 }
3246 #endif
3247
3248 /*
3249  * Access another process' address space.
3250  * Source/target buffer must be kernel space,
3251  * Do not walk the page table directly, use get_user_pages
3252  */
3253 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3254 {
3255         struct mm_struct *mm;
3256         struct vm_area_struct *vma;
3257         void *old_buf = buf;
3258
3259         mm = get_task_mm(tsk);
3260         if (!mm)
3261                 return 0;
3262
3263         down_read(&mm->mmap_sem);
3264         /* ignore errors, just check how much was successfully transferred */
3265         while (len) {
3266                 int bytes, ret, offset;
3267                 void *maddr;
3268                 struct page *page = NULL;
3269
3270                 ret = get_user_pages(tsk, mm, addr, 1,
3271                                 write, 1, &page, &vma);
3272                 if (ret <= 0) {
3273                         /*
3274                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3275                          * we can access using slightly different code.
3276                          */
3277 #ifdef CONFIG_HAVE_IOREMAP_PROT
3278                         vma = find_vma(mm, addr);
3279                         if (!vma)
3280                                 break;
3281                         if (vma->vm_ops && vma->vm_ops->access)
3282                                 ret = vma->vm_ops->access(vma, addr, buf,
3283                                                           len, write);
3284                         if (ret <= 0)
3285 #endif
3286                                 break;
3287                         bytes = ret;
3288                 } else {
3289                         bytes = len;
3290                         offset = addr & (PAGE_SIZE-1);
3291                         if (bytes > PAGE_SIZE-offset)
3292                                 bytes = PAGE_SIZE-offset;
3293
3294                         maddr = kmap(page);
3295                         if (write) {
3296                                 copy_to_user_page(vma, page, addr,
3297                                                   maddr + offset, buf, bytes);
3298                                 set_page_dirty_lock(page);
3299                         } else {
3300                                 copy_from_user_page(vma, page, addr,
3301                                                     buf, maddr + offset, bytes);
3302                         }
3303                         kunmap(page);
3304                         page_cache_release(page);
3305                 }
3306                 len -= bytes;
3307                 buf += bytes;
3308                 addr += bytes;
3309         }
3310         up_read(&mm->mmap_sem);
3311         mmput(mm);
3312
3313         return buf - old_buf;
3314 }
3315
3316 /*
3317  * Print the name of a VMA.
3318  */
3319 void print_vma_addr(char *prefix, unsigned long ip)
3320 {
3321         struct mm_struct *mm = current->mm;
3322         struct vm_area_struct *vma;
3323
3324         /*
3325          * Do not print if we are in atomic
3326          * contexts (in exception stacks, etc.):
3327          */
3328         if (preempt_count())
3329                 return;
3330
3331         down_read(&mm->mmap_sem);
3332         vma = find_vma(mm, ip);
3333         if (vma && vma->vm_file) {
3334                 struct file *f = vma->vm_file;
3335                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3336                 if (buf) {
3337                         char *p, *s;
3338
3339                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3340                         if (IS_ERR(p))
3341                                 p = "?";
3342                         s = strrchr(p, '/');
3343                         if (s)
3344                                 p = s+1;
3345                         printk("%s%s[%lx+%lx]", prefix, p,
3346                                         vma->vm_start,
3347                                         vma->vm_end - vma->vm_start);
3348                         free_page((unsigned long)buf);
3349                 }
3350         }
3351         up_read(&current->mm->mmap_sem);
3352 }
3353
3354 #ifdef CONFIG_PROVE_LOCKING
3355 void might_fault(void)
3356 {
3357         /*
3358          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3359          * holding the mmap_sem, this is safe because kernel memory doesn't
3360          * get paged out, therefore we'll never actually fault, and the
3361          * below annotations will generate false positives.
3362          */
3363         if (segment_eq(get_fs(), KERNEL_DS))
3364                 return;
3365
3366         might_sleep();
3367         /*
3368          * it would be nicer only to annotate paths which are not under
3369          * pagefault_disable, however that requires a larger audit and
3370          * providing helpers like get_user_atomic.
3371          */
3372         if (!in_atomic() && current->mm)
3373                 might_lock_read(&current->mm->mmap_sem);
3374 }
3375 EXPORT_SYMBOL(might_fault);
3376 #endif