ksm: let shared pages be swappable
[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 unsigned long
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                         if (swap_duplicate(entry) < 0)
590                                 return entry.val;
591
592                         /* make sure dst_mm is on swapoff's mmlist. */
593                         if (unlikely(list_empty(&dst_mm->mmlist))) {
594                                 spin_lock(&mmlist_lock);
595                                 if (list_empty(&dst_mm->mmlist))
596                                         list_add(&dst_mm->mmlist,
597                                                  &src_mm->mmlist);
598                                 spin_unlock(&mmlist_lock);
599                         }
600                         if (is_write_migration_entry(entry) &&
601                                         is_cow_mapping(vm_flags)) {
602                                 /*
603                                  * COW mappings require pages in both parent
604                                  * and child to be set to read.
605                                  */
606                                 make_migration_entry_read(&entry);
607                                 pte = swp_entry_to_pte(entry);
608                                 set_pte_at(src_mm, addr, src_pte, pte);
609                         }
610                 }
611                 goto out_set_pte;
612         }
613
614         /*
615          * If it's a COW mapping, write protect it both
616          * in the parent and the child
617          */
618         if (is_cow_mapping(vm_flags)) {
619                 ptep_set_wrprotect(src_mm, addr, src_pte);
620                 pte = pte_wrprotect(pte);
621         }
622
623         /*
624          * If it's a shared mapping, mark it clean in
625          * the child
626          */
627         if (vm_flags & VM_SHARED)
628                 pte = pte_mkclean(pte);
629         pte = pte_mkold(pte);
630
631         page = vm_normal_page(vma, addr, pte);
632         if (page) {
633                 get_page(page);
634                 page_dup_rmap(page);
635                 rss[PageAnon(page)]++;
636         }
637
638 out_set_pte:
639         set_pte_at(dst_mm, addr, dst_pte, pte);
640         return 0;
641 }
642
643 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
644                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
645                 unsigned long addr, unsigned long end)
646 {
647         pte_t *orig_src_pte, *orig_dst_pte;
648         pte_t *src_pte, *dst_pte;
649         spinlock_t *src_ptl, *dst_ptl;
650         int progress = 0;
651         int rss[2];
652         swp_entry_t entry = (swp_entry_t){0};
653
654 again:
655         rss[1] = rss[0] = 0;
656         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
657         if (!dst_pte)
658                 return -ENOMEM;
659         src_pte = pte_offset_map_nested(src_pmd, addr);
660         src_ptl = pte_lockptr(src_mm, src_pmd);
661         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
662         orig_src_pte = src_pte;
663         orig_dst_pte = dst_pte;
664         arch_enter_lazy_mmu_mode();
665
666         do {
667                 /*
668                  * We are holding two locks at this point - either of them
669                  * could generate latencies in another task on another CPU.
670                  */
671                 if (progress >= 32) {
672                         progress = 0;
673                         if (need_resched() ||
674                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
675                                 break;
676                 }
677                 if (pte_none(*src_pte)) {
678                         progress++;
679                         continue;
680                 }
681                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
682                                                         vma, addr, rss);
683                 if (entry.val)
684                         break;
685                 progress += 8;
686         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
687
688         arch_leave_lazy_mmu_mode();
689         spin_unlock(src_ptl);
690         pte_unmap_nested(orig_src_pte);
691         add_mm_rss(dst_mm, rss[0], rss[1]);
692         pte_unmap_unlock(orig_dst_pte, dst_ptl);
693         cond_resched();
694
695         if (entry.val) {
696                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
697                         return -ENOMEM;
698                 progress = 0;
699         }
700         if (addr != end)
701                 goto again;
702         return 0;
703 }
704
705 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
706                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
707                 unsigned long addr, unsigned long end)
708 {
709         pmd_t *src_pmd, *dst_pmd;
710         unsigned long next;
711
712         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
713         if (!dst_pmd)
714                 return -ENOMEM;
715         src_pmd = pmd_offset(src_pud, addr);
716         do {
717                 next = pmd_addr_end(addr, end);
718                 if (pmd_none_or_clear_bad(src_pmd))
719                         continue;
720                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
721                                                 vma, addr, next))
722                         return -ENOMEM;
723         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
724         return 0;
725 }
726
727 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
728                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
729                 unsigned long addr, unsigned long end)
730 {
731         pud_t *src_pud, *dst_pud;
732         unsigned long next;
733
734         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
735         if (!dst_pud)
736                 return -ENOMEM;
737         src_pud = pud_offset(src_pgd, addr);
738         do {
739                 next = pud_addr_end(addr, end);
740                 if (pud_none_or_clear_bad(src_pud))
741                         continue;
742                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
743                                                 vma, addr, next))
744                         return -ENOMEM;
745         } while (dst_pud++, src_pud++, addr = next, addr != end);
746         return 0;
747 }
748
749 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
750                 struct vm_area_struct *vma)
751 {
752         pgd_t *src_pgd, *dst_pgd;
753         unsigned long next;
754         unsigned long addr = vma->vm_start;
755         unsigned long end = vma->vm_end;
756         int ret;
757
758         /*
759          * Don't copy ptes where a page fault will fill them correctly.
760          * Fork becomes much lighter when there are big shared or private
761          * readonly mappings. The tradeoff is that copy_page_range is more
762          * efficient than faulting.
763          */
764         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
765                 if (!vma->anon_vma)
766                         return 0;
767         }
768
769         if (is_vm_hugetlb_page(vma))
770                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
771
772         if (unlikely(is_pfn_mapping(vma))) {
773                 /*
774                  * We do not free on error cases below as remove_vma
775                  * gets called on error from higher level routine
776                  */
777                 ret = track_pfn_vma_copy(vma);
778                 if (ret)
779                         return ret;
780         }
781
782         /*
783          * We need to invalidate the secondary MMU mappings only when
784          * there could be a permission downgrade on the ptes of the
785          * parent mm. And a permission downgrade will only happen if
786          * is_cow_mapping() returns true.
787          */
788         if (is_cow_mapping(vma->vm_flags))
789                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
790
791         ret = 0;
792         dst_pgd = pgd_offset(dst_mm, addr);
793         src_pgd = pgd_offset(src_mm, addr);
794         do {
795                 next = pgd_addr_end(addr, end);
796                 if (pgd_none_or_clear_bad(src_pgd))
797                         continue;
798                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
799                                             vma, addr, next))) {
800                         ret = -ENOMEM;
801                         break;
802                 }
803         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
804
805         if (is_cow_mapping(vma->vm_flags))
806                 mmu_notifier_invalidate_range_end(src_mm,
807                                                   vma->vm_start, end);
808         return ret;
809 }
810
811 static unsigned long zap_pte_range(struct mmu_gather *tlb,
812                                 struct vm_area_struct *vma, pmd_t *pmd,
813                                 unsigned long addr, unsigned long end,
814                                 long *zap_work, struct zap_details *details)
815 {
816         struct mm_struct *mm = tlb->mm;
817         pte_t *pte;
818         spinlock_t *ptl;
819         int file_rss = 0;
820         int anon_rss = 0;
821
822         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
823         arch_enter_lazy_mmu_mode();
824         do {
825                 pte_t ptent = *pte;
826                 if (pte_none(ptent)) {
827                         (*zap_work)--;
828                         continue;
829                 }
830
831                 (*zap_work) -= PAGE_SIZE;
832
833                 if (pte_present(ptent)) {
834                         struct page *page;
835
836                         page = vm_normal_page(vma, addr, ptent);
837                         if (unlikely(details) && page) {
838                                 /*
839                                  * unmap_shared_mapping_pages() wants to
840                                  * invalidate cache without truncating:
841                                  * unmap shared but keep private pages.
842                                  */
843                                 if (details->check_mapping &&
844                                     details->check_mapping != page->mapping)
845                                         continue;
846                                 /*
847                                  * Each page->index must be checked when
848                                  * invalidating or truncating nonlinear.
849                                  */
850                                 if (details->nonlinear_vma &&
851                                     (page->index < details->first_index ||
852                                      page->index > details->last_index))
853                                         continue;
854                         }
855                         ptent = ptep_get_and_clear_full(mm, addr, pte,
856                                                         tlb->fullmm);
857                         tlb_remove_tlb_entry(tlb, pte, addr);
858                         if (unlikely(!page))
859                                 continue;
860                         if (unlikely(details) && details->nonlinear_vma
861                             && linear_page_index(details->nonlinear_vma,
862                                                 addr) != page->index)
863                                 set_pte_at(mm, addr, pte,
864                                            pgoff_to_pte(page->index));
865                         if (PageAnon(page))
866                                 anon_rss--;
867                         else {
868                                 if (pte_dirty(ptent))
869                                         set_page_dirty(page);
870                                 if (pte_young(ptent) &&
871                                     likely(!VM_SequentialReadHint(vma)))
872                                         mark_page_accessed(page);
873                                 file_rss--;
874                         }
875                         page_remove_rmap(page);
876                         if (unlikely(page_mapcount(page) < 0))
877                                 print_bad_pte(vma, addr, ptent, page);
878                         tlb_remove_page(tlb, page);
879                         continue;
880                 }
881                 /*
882                  * If details->check_mapping, we leave swap entries;
883                  * if details->nonlinear_vma, we leave file entries.
884                  */
885                 if (unlikely(details))
886                         continue;
887                 if (pte_file(ptent)) {
888                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
889                                 print_bad_pte(vma, addr, ptent, NULL);
890                 } else if
891                   (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
892                         print_bad_pte(vma, addr, ptent, NULL);
893                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
894         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
895
896         add_mm_rss(mm, file_rss, anon_rss);
897         arch_leave_lazy_mmu_mode();
898         pte_unmap_unlock(pte - 1, ptl);
899
900         return addr;
901 }
902
903 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
904                                 struct vm_area_struct *vma, pud_t *pud,
905                                 unsigned long addr, unsigned long end,
906                                 long *zap_work, struct zap_details *details)
907 {
908         pmd_t *pmd;
909         unsigned long next;
910
911         pmd = pmd_offset(pud, addr);
912         do {
913                 next = pmd_addr_end(addr, end);
914                 if (pmd_none_or_clear_bad(pmd)) {
915                         (*zap_work)--;
916                         continue;
917                 }
918                 next = zap_pte_range(tlb, vma, pmd, addr, next,
919                                                 zap_work, details);
920         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
921
922         return addr;
923 }
924
925 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
926                                 struct vm_area_struct *vma, pgd_t *pgd,
927                                 unsigned long addr, unsigned long end,
928                                 long *zap_work, struct zap_details *details)
929 {
930         pud_t *pud;
931         unsigned long next;
932
933         pud = pud_offset(pgd, addr);
934         do {
935                 next = pud_addr_end(addr, end);
936                 if (pud_none_or_clear_bad(pud)) {
937                         (*zap_work)--;
938                         continue;
939                 }
940                 next = zap_pmd_range(tlb, vma, pud, addr, next,
941                                                 zap_work, details);
942         } while (pud++, addr = next, (addr != end && *zap_work > 0));
943
944         return addr;
945 }
946
947 static unsigned long unmap_page_range(struct mmu_gather *tlb,
948                                 struct vm_area_struct *vma,
949                                 unsigned long addr, unsigned long end,
950                                 long *zap_work, struct zap_details *details)
951 {
952         pgd_t *pgd;
953         unsigned long next;
954
955         if (details && !details->check_mapping && !details->nonlinear_vma)
956                 details = NULL;
957
958         BUG_ON(addr >= end);
959         tlb_start_vma(tlb, vma);
960         pgd = pgd_offset(vma->vm_mm, addr);
961         do {
962                 next = pgd_addr_end(addr, end);
963                 if (pgd_none_or_clear_bad(pgd)) {
964                         (*zap_work)--;
965                         continue;
966                 }
967                 next = zap_pud_range(tlb, vma, pgd, addr, next,
968                                                 zap_work, details);
969         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
970         tlb_end_vma(tlb, vma);
971
972         return addr;
973 }
974
975 #ifdef CONFIG_PREEMPT
976 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
977 #else
978 /* No preempt: go for improved straight-line efficiency */
979 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
980 #endif
981
982 /**
983  * unmap_vmas - unmap a range of memory covered by a list of vma's
984  * @tlbp: address of the caller's struct mmu_gather
985  * @vma: the starting vma
986  * @start_addr: virtual address at which to start unmapping
987  * @end_addr: virtual address at which to end unmapping
988  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
989  * @details: details of nonlinear truncation or shared cache invalidation
990  *
991  * Returns the end address of the unmapping (restart addr if interrupted).
992  *
993  * Unmap all pages in the vma list.
994  *
995  * We aim to not hold locks for too long (for scheduling latency reasons).
996  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
997  * return the ending mmu_gather to the caller.
998  *
999  * Only addresses between `start' and `end' will be unmapped.
1000  *
1001  * The VMA list must be sorted in ascending virtual address order.
1002  *
1003  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1004  * range after unmap_vmas() returns.  So the only responsibility here is to
1005  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1006  * drops the lock and schedules.
1007  */
1008 unsigned long unmap_vmas(struct mmu_gather **tlbp,
1009                 struct vm_area_struct *vma, unsigned long start_addr,
1010                 unsigned long end_addr, unsigned long *nr_accounted,
1011                 struct zap_details *details)
1012 {
1013         long zap_work = ZAP_BLOCK_SIZE;
1014         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
1015         int tlb_start_valid = 0;
1016         unsigned long start = start_addr;
1017         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1018         int fullmm = (*tlbp)->fullmm;
1019         struct mm_struct *mm = vma->vm_mm;
1020
1021         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1022         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1023                 unsigned long end;
1024
1025                 start = max(vma->vm_start, start_addr);
1026                 if (start >= vma->vm_end)
1027                         continue;
1028                 end = min(vma->vm_end, end_addr);
1029                 if (end <= vma->vm_start)
1030                         continue;
1031
1032                 if (vma->vm_flags & VM_ACCOUNT)
1033                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1034
1035                 if (unlikely(is_pfn_mapping(vma)))
1036                         untrack_pfn_vma(vma, 0, 0);
1037
1038                 while (start != end) {
1039                         if (!tlb_start_valid) {
1040                                 tlb_start = start;
1041                                 tlb_start_valid = 1;
1042                         }
1043
1044                         if (unlikely(is_vm_hugetlb_page(vma))) {
1045                                 /*
1046                                  * It is undesirable to test vma->vm_file as it
1047                                  * should be non-null for valid hugetlb area.
1048                                  * However, vm_file will be NULL in the error
1049                                  * cleanup path of do_mmap_pgoff. When
1050                                  * hugetlbfs ->mmap method fails,
1051                                  * do_mmap_pgoff() nullifies vma->vm_file
1052                                  * before calling this function to clean up.
1053                                  * Since no pte has actually been setup, it is
1054                                  * safe to do nothing in this case.
1055                                  */
1056                                 if (vma->vm_file) {
1057                                         unmap_hugepage_range(vma, start, end, NULL);
1058                                         zap_work -= (end - start) /
1059                                         pages_per_huge_page(hstate_vma(vma));
1060                                 }
1061
1062                                 start = end;
1063                         } else
1064                                 start = unmap_page_range(*tlbp, vma,
1065                                                 start, end, &zap_work, details);
1066
1067                         if (zap_work > 0) {
1068                                 BUG_ON(start != end);
1069                                 break;
1070                         }
1071
1072                         tlb_finish_mmu(*tlbp, tlb_start, start);
1073
1074                         if (need_resched() ||
1075                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1076                                 if (i_mmap_lock) {
1077                                         *tlbp = NULL;
1078                                         goto out;
1079                                 }
1080                                 cond_resched();
1081                         }
1082
1083                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1084                         tlb_start_valid = 0;
1085                         zap_work = ZAP_BLOCK_SIZE;
1086                 }
1087         }
1088 out:
1089         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1090         return start;   /* which is now the end (or restart) address */
1091 }
1092
1093 /**
1094  * zap_page_range - remove user pages in a given range
1095  * @vma: vm_area_struct holding the applicable pages
1096  * @address: starting address of pages to zap
1097  * @size: number of bytes to zap
1098  * @details: details of nonlinear truncation or shared cache invalidation
1099  */
1100 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1101                 unsigned long size, struct zap_details *details)
1102 {
1103         struct mm_struct *mm = vma->vm_mm;
1104         struct mmu_gather *tlb;
1105         unsigned long end = address + size;
1106         unsigned long nr_accounted = 0;
1107
1108         lru_add_drain();
1109         tlb = tlb_gather_mmu(mm, 0);
1110         update_hiwater_rss(mm);
1111         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1112         if (tlb)
1113                 tlb_finish_mmu(tlb, address, end);
1114         return end;
1115 }
1116
1117 /**
1118  * zap_vma_ptes - remove ptes mapping the vma
1119  * @vma: vm_area_struct holding ptes to be zapped
1120  * @address: starting address of pages to zap
1121  * @size: number of bytes to zap
1122  *
1123  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1124  *
1125  * The entire address range must be fully contained within the vma.
1126  *
1127  * Returns 0 if successful.
1128  */
1129 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1130                 unsigned long size)
1131 {
1132         if (address < vma->vm_start || address + size > vma->vm_end ||
1133                         !(vma->vm_flags & VM_PFNMAP))
1134                 return -1;
1135         zap_page_range(vma, address, size, NULL);
1136         return 0;
1137 }
1138 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1139
1140 /*
1141  * Do a quick page-table lookup for a single page.
1142  */
1143 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1144                         unsigned int flags)
1145 {
1146         pgd_t *pgd;
1147         pud_t *pud;
1148         pmd_t *pmd;
1149         pte_t *ptep, pte;
1150         spinlock_t *ptl;
1151         struct page *page;
1152         struct mm_struct *mm = vma->vm_mm;
1153
1154         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1155         if (!IS_ERR(page)) {
1156                 BUG_ON(flags & FOLL_GET);
1157                 goto out;
1158         }
1159
1160         page = NULL;
1161         pgd = pgd_offset(mm, address);
1162         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1163                 goto no_page_table;
1164
1165         pud = pud_offset(pgd, address);
1166         if (pud_none(*pud))
1167                 goto no_page_table;
1168         if (pud_huge(*pud)) {
1169                 BUG_ON(flags & FOLL_GET);
1170                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1171                 goto out;
1172         }
1173         if (unlikely(pud_bad(*pud)))
1174                 goto no_page_table;
1175
1176         pmd = pmd_offset(pud, address);
1177         if (pmd_none(*pmd))
1178                 goto no_page_table;
1179         if (pmd_huge(*pmd)) {
1180                 BUG_ON(flags & FOLL_GET);
1181                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1182                 goto out;
1183         }
1184         if (unlikely(pmd_bad(*pmd)))
1185                 goto no_page_table;
1186
1187         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1188
1189         pte = *ptep;
1190         if (!pte_present(pte))
1191                 goto no_page;
1192         if ((flags & FOLL_WRITE) && !pte_write(pte))
1193                 goto unlock;
1194
1195         page = vm_normal_page(vma, address, pte);
1196         if (unlikely(!page)) {
1197                 if ((flags & FOLL_DUMP) ||
1198                     !is_zero_pfn(pte_pfn(pte)))
1199                         goto bad_page;
1200                 page = pte_page(pte);
1201         }
1202
1203         if (flags & FOLL_GET)
1204                 get_page(page);
1205         if (flags & FOLL_TOUCH) {
1206                 if ((flags & FOLL_WRITE) &&
1207                     !pte_dirty(pte) && !PageDirty(page))
1208                         set_page_dirty(page);
1209                 /*
1210                  * pte_mkyoung() would be more correct here, but atomic care
1211                  * is needed to avoid losing the dirty bit: it is easier to use
1212                  * mark_page_accessed().
1213                  */
1214                 mark_page_accessed(page);
1215         }
1216 unlock:
1217         pte_unmap_unlock(ptep, ptl);
1218 out:
1219         return page;
1220
1221 bad_page:
1222         pte_unmap_unlock(ptep, ptl);
1223         return ERR_PTR(-EFAULT);
1224
1225 no_page:
1226         pte_unmap_unlock(ptep, ptl);
1227         if (!pte_none(pte))
1228                 return page;
1229
1230 no_page_table:
1231         /*
1232          * When core dumping an enormous anonymous area that nobody
1233          * has touched so far, we don't want to allocate unnecessary pages or
1234          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1235          * then get_dump_page() will return NULL to leave a hole in the dump.
1236          * But we can only make this optimization where a hole would surely
1237          * be zero-filled if handle_mm_fault() actually did handle it.
1238          */
1239         if ((flags & FOLL_DUMP) &&
1240             (!vma->vm_ops || !vma->vm_ops->fault))
1241                 return ERR_PTR(-EFAULT);
1242         return page;
1243 }
1244
1245 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1246                      unsigned long start, int nr_pages, unsigned int gup_flags,
1247                      struct page **pages, struct vm_area_struct **vmas)
1248 {
1249         int i;
1250         unsigned long vm_flags;
1251
1252         if (nr_pages <= 0)
1253                 return 0;
1254
1255         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1256
1257         /* 
1258          * Require read or write permissions.
1259          * If FOLL_FORCE is set, we only require the "MAY" flags.
1260          */
1261         vm_flags  = (gup_flags & FOLL_WRITE) ?
1262                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1263         vm_flags &= (gup_flags & FOLL_FORCE) ?
1264                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1265         i = 0;
1266
1267         do {
1268                 struct vm_area_struct *vma;
1269
1270                 vma = find_extend_vma(mm, start);
1271                 if (!vma && in_gate_area(tsk, start)) {
1272                         unsigned long pg = start & PAGE_MASK;
1273                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1274                         pgd_t *pgd;
1275                         pud_t *pud;
1276                         pmd_t *pmd;
1277                         pte_t *pte;
1278
1279                         /* user gate pages are read-only */
1280                         if (gup_flags & FOLL_WRITE)
1281                                 return i ? : -EFAULT;
1282                         if (pg > TASK_SIZE)
1283                                 pgd = pgd_offset_k(pg);
1284                         else
1285                                 pgd = pgd_offset_gate(mm, pg);
1286                         BUG_ON(pgd_none(*pgd));
1287                         pud = pud_offset(pgd, pg);
1288                         BUG_ON(pud_none(*pud));
1289                         pmd = pmd_offset(pud, pg);
1290                         if (pmd_none(*pmd))
1291                                 return i ? : -EFAULT;
1292                         pte = pte_offset_map(pmd, pg);
1293                         if (pte_none(*pte)) {
1294                                 pte_unmap(pte);
1295                                 return i ? : -EFAULT;
1296                         }
1297                         if (pages) {
1298                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1299                                 pages[i] = page;
1300                                 if (page)
1301                                         get_page(page);
1302                         }
1303                         pte_unmap(pte);
1304                         if (vmas)
1305                                 vmas[i] = gate_vma;
1306                         i++;
1307                         start += PAGE_SIZE;
1308                         nr_pages--;
1309                         continue;
1310                 }
1311
1312                 if (!vma ||
1313                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1314                     !(vm_flags & vma->vm_flags))
1315                         return i ? : -EFAULT;
1316
1317                 if (is_vm_hugetlb_page(vma)) {
1318                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1319                                         &start, &nr_pages, i, gup_flags);
1320                         continue;
1321                 }
1322
1323                 do {
1324                         struct page *page;
1325                         unsigned int foll_flags = gup_flags;
1326
1327                         /*
1328                          * If we have a pending SIGKILL, don't keep faulting
1329                          * pages and potentially allocating memory.
1330                          */
1331                         if (unlikely(fatal_signal_pending(current)))
1332                                 return i ? i : -ERESTARTSYS;
1333
1334                         cond_resched();
1335                         while (!(page = follow_page(vma, start, foll_flags))) {
1336                                 int ret;
1337
1338                                 ret = handle_mm_fault(mm, vma, start,
1339                                         (foll_flags & FOLL_WRITE) ?
1340                                         FAULT_FLAG_WRITE : 0);
1341
1342                                 if (ret & VM_FAULT_ERROR) {
1343                                         if (ret & VM_FAULT_OOM)
1344                                                 return i ? i : -ENOMEM;
1345                                         if (ret &
1346                                             (VM_FAULT_HWPOISON|VM_FAULT_SIGBUS))
1347                                                 return i ? i : -EFAULT;
1348                                         BUG();
1349                                 }
1350                                 if (ret & VM_FAULT_MAJOR)
1351                                         tsk->maj_flt++;
1352                                 else
1353                                         tsk->min_flt++;
1354
1355                                 /*
1356                                  * The VM_FAULT_WRITE bit tells us that
1357                                  * do_wp_page has broken COW when necessary,
1358                                  * even if maybe_mkwrite decided not to set
1359                                  * pte_write. We can thus safely do subsequent
1360                                  * page lookups as if they were reads. But only
1361                                  * do so when looping for pte_write is futile:
1362                                  * in some cases userspace may also be wanting
1363                                  * to write to the gotten user page, which a
1364                                  * read fault here might prevent (a readonly
1365                                  * page might get reCOWed by userspace write).
1366                                  */
1367                                 if ((ret & VM_FAULT_WRITE) &&
1368                                     !(vma->vm_flags & VM_WRITE))
1369                                         foll_flags &= ~FOLL_WRITE;
1370
1371                                 cond_resched();
1372                         }
1373                         if (IS_ERR(page))
1374                                 return i ? i : PTR_ERR(page);
1375                         if (pages) {
1376                                 pages[i] = page;
1377
1378                                 flush_anon_page(vma, page, start);
1379                                 flush_dcache_page(page);
1380                         }
1381                         if (vmas)
1382                                 vmas[i] = vma;
1383                         i++;
1384                         start += PAGE_SIZE;
1385                         nr_pages--;
1386                 } while (nr_pages && start < vma->vm_end);
1387         } while (nr_pages);
1388         return i;
1389 }
1390
1391 /**
1392  * get_user_pages() - pin user pages in memory
1393  * @tsk:        task_struct of target task
1394  * @mm:         mm_struct of target mm
1395  * @start:      starting user address
1396  * @nr_pages:   number of pages from start to pin
1397  * @write:      whether pages will be written to by the caller
1398  * @force:      whether to force write access even if user mapping is
1399  *              readonly. This will result in the page being COWed even
1400  *              in MAP_SHARED mappings. You do not want this.
1401  * @pages:      array that receives pointers to the pages pinned.
1402  *              Should be at least nr_pages long. Or NULL, if caller
1403  *              only intends to ensure the pages are faulted in.
1404  * @vmas:       array of pointers to vmas corresponding to each page.
1405  *              Or NULL if the caller does not require them.
1406  *
1407  * Returns number of pages pinned. This may be fewer than the number
1408  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1409  * were pinned, returns -errno. Each page returned must be released
1410  * with a put_page() call when it is finished with. vmas will only
1411  * remain valid while mmap_sem is held.
1412  *
1413  * Must be called with mmap_sem held for read or write.
1414  *
1415  * get_user_pages walks a process's page tables and takes a reference to
1416  * each struct page that each user address corresponds to at a given
1417  * instant. That is, it takes the page that would be accessed if a user
1418  * thread accesses the given user virtual address at that instant.
1419  *
1420  * This does not guarantee that the page exists in the user mappings when
1421  * get_user_pages returns, and there may even be a completely different
1422  * page there in some cases (eg. if mmapped pagecache has been invalidated
1423  * and subsequently re faulted). However it does guarantee that the page
1424  * won't be freed completely. And mostly callers simply care that the page
1425  * contains data that was valid *at some point in time*. Typically, an IO
1426  * or similar operation cannot guarantee anything stronger anyway because
1427  * locks can't be held over the syscall boundary.
1428  *
1429  * If write=0, the page must not be written to. If the page is written to,
1430  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1431  * after the page is finished with, and before put_page is called.
1432  *
1433  * get_user_pages is typically used for fewer-copy IO operations, to get a
1434  * handle on the memory by some means other than accesses via the user virtual
1435  * addresses. The pages may be submitted for DMA to devices or accessed via
1436  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1437  * use the correct cache flushing APIs.
1438  *
1439  * See also get_user_pages_fast, for performance critical applications.
1440  */
1441 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1442                 unsigned long start, int nr_pages, int write, int force,
1443                 struct page **pages, struct vm_area_struct **vmas)
1444 {
1445         int flags = FOLL_TOUCH;
1446
1447         if (pages)
1448                 flags |= FOLL_GET;
1449         if (write)
1450                 flags |= FOLL_WRITE;
1451         if (force)
1452                 flags |= FOLL_FORCE;
1453
1454         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
1455 }
1456 EXPORT_SYMBOL(get_user_pages);
1457
1458 /**
1459  * get_dump_page() - pin user page in memory while writing it to core dump
1460  * @addr: user address
1461  *
1462  * Returns struct page pointer of user page pinned for dump,
1463  * to be freed afterwards by page_cache_release() or put_page().
1464  *
1465  * Returns NULL on any kind of failure - a hole must then be inserted into
1466  * the corefile, to preserve alignment with its headers; and also returns
1467  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1468  * allowing a hole to be left in the corefile to save diskspace.
1469  *
1470  * Called without mmap_sem, but after all other threads have been killed.
1471  */
1472 #ifdef CONFIG_ELF_CORE
1473 struct page *get_dump_page(unsigned long addr)
1474 {
1475         struct vm_area_struct *vma;
1476         struct page *page;
1477
1478         if (__get_user_pages(current, current->mm, addr, 1,
1479                         FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma) < 1)
1480                 return NULL;
1481         flush_cache_page(vma, addr, page_to_pfn(page));
1482         return page;
1483 }
1484 #endif /* CONFIG_ELF_CORE */
1485
1486 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1487                         spinlock_t **ptl)
1488 {
1489         pgd_t * pgd = pgd_offset(mm, addr);
1490         pud_t * pud = pud_alloc(mm, pgd, addr);
1491         if (pud) {
1492                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1493                 if (pmd)
1494                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1495         }
1496         return NULL;
1497 }
1498
1499 /*
1500  * This is the old fallback for page remapping.
1501  *
1502  * For historical reasons, it only allows reserved pages. Only
1503  * old drivers should use this, and they needed to mark their
1504  * pages reserved for the old functions anyway.
1505  */
1506 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1507                         struct page *page, pgprot_t prot)
1508 {
1509         struct mm_struct *mm = vma->vm_mm;
1510         int retval;
1511         pte_t *pte;
1512         spinlock_t *ptl;
1513
1514         retval = -EINVAL;
1515         if (PageAnon(page))
1516                 goto out;
1517         retval = -ENOMEM;
1518         flush_dcache_page(page);
1519         pte = get_locked_pte(mm, addr, &ptl);
1520         if (!pte)
1521                 goto out;
1522         retval = -EBUSY;
1523         if (!pte_none(*pte))
1524                 goto out_unlock;
1525
1526         /* Ok, finally just insert the thing.. */
1527         get_page(page);
1528         inc_mm_counter(mm, file_rss);
1529         page_add_file_rmap(page);
1530         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1531
1532         retval = 0;
1533         pte_unmap_unlock(pte, ptl);
1534         return retval;
1535 out_unlock:
1536         pte_unmap_unlock(pte, ptl);
1537 out:
1538         return retval;
1539 }
1540
1541 /**
1542  * vm_insert_page - insert single page into user vma
1543  * @vma: user vma to map to
1544  * @addr: target user address of this page
1545  * @page: source kernel page
1546  *
1547  * This allows drivers to insert individual pages they've allocated
1548  * into a user vma.
1549  *
1550  * The page has to be a nice clean _individual_ kernel allocation.
1551  * If you allocate a compound page, you need to have marked it as
1552  * such (__GFP_COMP), or manually just split the page up yourself
1553  * (see split_page()).
1554  *
1555  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1556  * took an arbitrary page protection parameter. This doesn't allow
1557  * that. Your vma protection will have to be set up correctly, which
1558  * means that if you want a shared writable mapping, you'd better
1559  * ask for a shared writable mapping!
1560  *
1561  * The page does not need to be reserved.
1562  */
1563 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1564                         struct page *page)
1565 {
1566         if (addr < vma->vm_start || addr >= vma->vm_end)
1567                 return -EFAULT;
1568         if (!page_count(page))
1569                 return -EINVAL;
1570         vma->vm_flags |= VM_INSERTPAGE;
1571         return insert_page(vma, addr, page, vma->vm_page_prot);
1572 }
1573 EXPORT_SYMBOL(vm_insert_page);
1574
1575 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1576                         unsigned long pfn, pgprot_t prot)
1577 {
1578         struct mm_struct *mm = vma->vm_mm;
1579         int retval;
1580         pte_t *pte, entry;
1581         spinlock_t *ptl;
1582
1583         retval = -ENOMEM;
1584         pte = get_locked_pte(mm, addr, &ptl);
1585         if (!pte)
1586                 goto out;
1587         retval = -EBUSY;
1588         if (!pte_none(*pte))
1589                 goto out_unlock;
1590
1591         /* Ok, finally just insert the thing.. */
1592         entry = pte_mkspecial(pfn_pte(pfn, prot));
1593         set_pte_at(mm, addr, pte, entry);
1594         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1595
1596         retval = 0;
1597 out_unlock:
1598         pte_unmap_unlock(pte, ptl);
1599 out:
1600         return retval;
1601 }
1602
1603 /**
1604  * vm_insert_pfn - insert single pfn into user vma
1605  * @vma: user vma to map to
1606  * @addr: target user address of this page
1607  * @pfn: source kernel pfn
1608  *
1609  * Similar to vm_inert_page, this allows drivers to insert individual pages
1610  * they've allocated into a user vma. Same comments apply.
1611  *
1612  * This function should only be called from a vm_ops->fault handler, and
1613  * in that case the handler should return NULL.
1614  *
1615  * vma cannot be a COW mapping.
1616  *
1617  * As this is called only for pages that do not currently exist, we
1618  * do not need to flush old virtual caches or the TLB.
1619  */
1620 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1621                         unsigned long pfn)
1622 {
1623         int ret;
1624         pgprot_t pgprot = vma->vm_page_prot;
1625         /*
1626          * Technically, architectures with pte_special can avoid all these
1627          * restrictions (same for remap_pfn_range).  However we would like
1628          * consistency in testing and feature parity among all, so we should
1629          * try to keep these invariants in place for everybody.
1630          */
1631         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1632         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1633                                                 (VM_PFNMAP|VM_MIXEDMAP));
1634         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1635         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1636
1637         if (addr < vma->vm_start || addr >= vma->vm_end)
1638                 return -EFAULT;
1639         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1640                 return -EINVAL;
1641
1642         ret = insert_pfn(vma, addr, pfn, pgprot);
1643
1644         if (ret)
1645                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1646
1647         return ret;
1648 }
1649 EXPORT_SYMBOL(vm_insert_pfn);
1650
1651 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1652                         unsigned long pfn)
1653 {
1654         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1655
1656         if (addr < vma->vm_start || addr >= vma->vm_end)
1657                 return -EFAULT;
1658
1659         /*
1660          * If we don't have pte special, then we have to use the pfn_valid()
1661          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1662          * refcount the page if pfn_valid is true (hence insert_page rather
1663          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1664          * without pte special, it would there be refcounted as a normal page.
1665          */
1666         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1667                 struct page *page;
1668
1669                 page = pfn_to_page(pfn);
1670                 return insert_page(vma, addr, page, vma->vm_page_prot);
1671         }
1672         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1673 }
1674 EXPORT_SYMBOL(vm_insert_mixed);
1675
1676 /*
1677  * maps a range of physical memory into the requested pages. the old
1678  * mappings are removed. any references to nonexistent pages results
1679  * in null mappings (currently treated as "copy-on-access")
1680  */
1681 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1682                         unsigned long addr, unsigned long end,
1683                         unsigned long pfn, pgprot_t prot)
1684 {
1685         pte_t *pte;
1686         spinlock_t *ptl;
1687
1688         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1689         if (!pte)
1690                 return -ENOMEM;
1691         arch_enter_lazy_mmu_mode();
1692         do {
1693                 BUG_ON(!pte_none(*pte));
1694                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1695                 pfn++;
1696         } while (pte++, addr += PAGE_SIZE, addr != end);
1697         arch_leave_lazy_mmu_mode();
1698         pte_unmap_unlock(pte - 1, ptl);
1699         return 0;
1700 }
1701
1702 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1703                         unsigned long addr, unsigned long end,
1704                         unsigned long pfn, pgprot_t prot)
1705 {
1706         pmd_t *pmd;
1707         unsigned long next;
1708
1709         pfn -= addr >> PAGE_SHIFT;
1710         pmd = pmd_alloc(mm, pud, addr);
1711         if (!pmd)
1712                 return -ENOMEM;
1713         do {
1714                 next = pmd_addr_end(addr, end);
1715                 if (remap_pte_range(mm, pmd, addr, next,
1716                                 pfn + (addr >> PAGE_SHIFT), prot))
1717                         return -ENOMEM;
1718         } while (pmd++, addr = next, addr != end);
1719         return 0;
1720 }
1721
1722 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1723                         unsigned long addr, unsigned long end,
1724                         unsigned long pfn, pgprot_t prot)
1725 {
1726         pud_t *pud;
1727         unsigned long next;
1728
1729         pfn -= addr >> PAGE_SHIFT;
1730         pud = pud_alloc(mm, pgd, addr);
1731         if (!pud)
1732                 return -ENOMEM;
1733         do {
1734                 next = pud_addr_end(addr, end);
1735                 if (remap_pmd_range(mm, pud, addr, next,
1736                                 pfn + (addr >> PAGE_SHIFT), prot))
1737                         return -ENOMEM;
1738         } while (pud++, addr = next, addr != end);
1739         return 0;
1740 }
1741
1742 /**
1743  * remap_pfn_range - remap kernel memory to userspace
1744  * @vma: user vma to map to
1745  * @addr: target user address to start at
1746  * @pfn: physical address of kernel memory
1747  * @size: size of map area
1748  * @prot: page protection flags for this mapping
1749  *
1750  *  Note: this is only safe if the mm semaphore is held when called.
1751  */
1752 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1753                     unsigned long pfn, unsigned long size, pgprot_t prot)
1754 {
1755         pgd_t *pgd;
1756         unsigned long next;
1757         unsigned long end = addr + PAGE_ALIGN(size);
1758         struct mm_struct *mm = vma->vm_mm;
1759         int err;
1760
1761         /*
1762          * Physically remapped pages are special. Tell the
1763          * rest of the world about it:
1764          *   VM_IO tells people not to look at these pages
1765          *      (accesses can have side effects).
1766          *   VM_RESERVED is specified all over the place, because
1767          *      in 2.4 it kept swapout's vma scan off this vma; but
1768          *      in 2.6 the LRU scan won't even find its pages, so this
1769          *      flag means no more than count its pages in reserved_vm,
1770          *      and omit it from core dump, even when VM_IO turned off.
1771          *   VM_PFNMAP tells the core MM that the base pages are just
1772          *      raw PFN mappings, and do not have a "struct page" associated
1773          *      with them.
1774          *
1775          * There's a horrible special case to handle copy-on-write
1776          * behaviour that some programs depend on. We mark the "original"
1777          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1778          */
1779         if (addr == vma->vm_start && end == vma->vm_end) {
1780                 vma->vm_pgoff = pfn;
1781                 vma->vm_flags |= VM_PFN_AT_MMAP;
1782         } else if (is_cow_mapping(vma->vm_flags))
1783                 return -EINVAL;
1784
1785         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1786
1787         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1788         if (err) {
1789                 /*
1790                  * To indicate that track_pfn related cleanup is not
1791                  * needed from higher level routine calling unmap_vmas
1792                  */
1793                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1794                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1795                 return -EINVAL;
1796         }
1797
1798         BUG_ON(addr >= end);
1799         pfn -= addr >> PAGE_SHIFT;
1800         pgd = pgd_offset(mm, addr);
1801         flush_cache_range(vma, addr, end);
1802         do {
1803                 next = pgd_addr_end(addr, end);
1804                 err = remap_pud_range(mm, pgd, addr, next,
1805                                 pfn + (addr >> PAGE_SHIFT), prot);
1806                 if (err)
1807                         break;
1808         } while (pgd++, addr = next, addr != end);
1809
1810         if (err)
1811                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1812
1813         return err;
1814 }
1815 EXPORT_SYMBOL(remap_pfn_range);
1816
1817 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1818                                      unsigned long addr, unsigned long end,
1819                                      pte_fn_t fn, void *data)
1820 {
1821         pte_t *pte;
1822         int err;
1823         pgtable_t token;
1824         spinlock_t *uninitialized_var(ptl);
1825
1826         pte = (mm == &init_mm) ?
1827                 pte_alloc_kernel(pmd, addr) :
1828                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1829         if (!pte)
1830                 return -ENOMEM;
1831
1832         BUG_ON(pmd_huge(*pmd));
1833
1834         arch_enter_lazy_mmu_mode();
1835
1836         token = pmd_pgtable(*pmd);
1837
1838         do {
1839                 err = fn(pte++, token, addr, data);
1840                 if (err)
1841                         break;
1842         } while (addr += PAGE_SIZE, addr != end);
1843
1844         arch_leave_lazy_mmu_mode();
1845
1846         if (mm != &init_mm)
1847                 pte_unmap_unlock(pte-1, ptl);
1848         return err;
1849 }
1850
1851 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1852                                      unsigned long addr, unsigned long end,
1853                                      pte_fn_t fn, void *data)
1854 {
1855         pmd_t *pmd;
1856         unsigned long next;
1857         int err;
1858
1859         BUG_ON(pud_huge(*pud));
1860
1861         pmd = pmd_alloc(mm, pud, addr);
1862         if (!pmd)
1863                 return -ENOMEM;
1864         do {
1865                 next = pmd_addr_end(addr, end);
1866                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1867                 if (err)
1868                         break;
1869         } while (pmd++, addr = next, addr != end);
1870         return err;
1871 }
1872
1873 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1874                                      unsigned long addr, unsigned long end,
1875                                      pte_fn_t fn, void *data)
1876 {
1877         pud_t *pud;
1878         unsigned long next;
1879         int err;
1880
1881         pud = pud_alloc(mm, pgd, addr);
1882         if (!pud)
1883                 return -ENOMEM;
1884         do {
1885                 next = pud_addr_end(addr, end);
1886                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1887                 if (err)
1888                         break;
1889         } while (pud++, addr = next, addr != end);
1890         return err;
1891 }
1892
1893 /*
1894  * Scan a region of virtual memory, filling in page tables as necessary
1895  * and calling a provided function on each leaf page table.
1896  */
1897 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1898                         unsigned long size, pte_fn_t fn, void *data)
1899 {
1900         pgd_t *pgd;
1901         unsigned long next;
1902         unsigned long start = addr, end = addr + size;
1903         int err;
1904
1905         BUG_ON(addr >= end);
1906         mmu_notifier_invalidate_range_start(mm, start, end);
1907         pgd = pgd_offset(mm, addr);
1908         do {
1909                 next = pgd_addr_end(addr, end);
1910                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1911                 if (err)
1912                         break;
1913         } while (pgd++, addr = next, addr != end);
1914         mmu_notifier_invalidate_range_end(mm, start, end);
1915         return err;
1916 }
1917 EXPORT_SYMBOL_GPL(apply_to_page_range);
1918
1919 /*
1920  * handle_pte_fault chooses page fault handler according to an entry
1921  * which was read non-atomically.  Before making any commitment, on
1922  * those architectures or configurations (e.g. i386 with PAE) which
1923  * might give a mix of unmatched parts, do_swap_page and do_file_page
1924  * must check under lock before unmapping the pte and proceeding
1925  * (but do_wp_page is only called after already making such a check;
1926  * and do_anonymous_page and do_no_page can safely check later on).
1927  */
1928 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1929                                 pte_t *page_table, pte_t orig_pte)
1930 {
1931         int same = 1;
1932 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1933         if (sizeof(pte_t) > sizeof(unsigned long)) {
1934                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1935                 spin_lock(ptl);
1936                 same = pte_same(*page_table, orig_pte);
1937                 spin_unlock(ptl);
1938         }
1939 #endif
1940         pte_unmap(page_table);
1941         return same;
1942 }
1943
1944 /*
1945  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1946  * servicing faults for write access.  In the normal case, do always want
1947  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1948  * that do not have writing enabled, when used by access_process_vm.
1949  */
1950 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1951 {
1952         if (likely(vma->vm_flags & VM_WRITE))
1953                 pte = pte_mkwrite(pte);
1954         return pte;
1955 }
1956
1957 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1958 {
1959         /*
1960          * If the source page was a PFN mapping, we don't have
1961          * a "struct page" for it. We do a best-effort copy by
1962          * just copying from the original user address. If that
1963          * fails, we just zero-fill it. Live with it.
1964          */
1965         if (unlikely(!src)) {
1966                 void *kaddr = kmap_atomic(dst, KM_USER0);
1967                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1968
1969                 /*
1970                  * This really shouldn't fail, because the page is there
1971                  * in the page tables. But it might just be unreadable,
1972                  * in which case we just give up and fill the result with
1973                  * zeroes.
1974                  */
1975                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1976                         memset(kaddr, 0, PAGE_SIZE);
1977                 kunmap_atomic(kaddr, KM_USER0);
1978                 flush_dcache_page(dst);
1979         } else
1980                 copy_user_highpage(dst, src, va, vma);
1981 }
1982
1983 /*
1984  * This routine handles present pages, when users try to write
1985  * to a shared page. It is done by copying the page to a new address
1986  * and decrementing the shared-page counter for the old page.
1987  *
1988  * Note that this routine assumes that the protection checks have been
1989  * done by the caller (the low-level page fault routine in most cases).
1990  * Thus we can safely just mark it writable once we've done any necessary
1991  * COW.
1992  *
1993  * We also mark the page dirty at this point even though the page will
1994  * change only once the write actually happens. This avoids a few races,
1995  * and potentially makes it more efficient.
1996  *
1997  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1998  * but allow concurrent faults), with pte both mapped and locked.
1999  * We return with mmap_sem still held, but pte unmapped and unlocked.
2000  */
2001 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2002                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2003                 spinlock_t *ptl, pte_t orig_pte)
2004 {
2005         struct page *old_page, *new_page;
2006         pte_t entry;
2007         int reuse = 0, ret = 0;
2008         int page_mkwrite = 0;
2009         struct page *dirty_page = NULL;
2010
2011         old_page = vm_normal_page(vma, address, orig_pte);
2012         if (!old_page) {
2013                 /*
2014                  * VM_MIXEDMAP !pfn_valid() case
2015                  *
2016                  * We should not cow pages in a shared writeable mapping.
2017                  * Just mark the pages writable as we can't do any dirty
2018                  * accounting on raw pfn maps.
2019                  */
2020                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2021                                      (VM_WRITE|VM_SHARED))
2022                         goto reuse;
2023                 goto gotten;
2024         }
2025
2026         /*
2027          * Take out anonymous pages first, anonymous shared vmas are
2028          * not dirty accountable.
2029          */
2030         if (PageAnon(old_page) && !PageKsm(old_page)) {
2031                 if (!trylock_page(old_page)) {
2032                         page_cache_get(old_page);
2033                         pte_unmap_unlock(page_table, ptl);
2034                         lock_page(old_page);
2035                         page_table = pte_offset_map_lock(mm, pmd, address,
2036                                                          &ptl);
2037                         if (!pte_same(*page_table, orig_pte)) {
2038                                 unlock_page(old_page);
2039                                 page_cache_release(old_page);
2040                                 goto unlock;
2041                         }
2042                         page_cache_release(old_page);
2043                 }
2044                 reuse = reuse_swap_page(old_page);
2045                 unlock_page(old_page);
2046         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2047                                         (VM_WRITE|VM_SHARED))) {
2048                 /*
2049                  * Only catch write-faults on shared writable pages,
2050                  * read-only shared pages can get COWed by
2051                  * get_user_pages(.write=1, .force=1).
2052                  */
2053                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2054                         struct vm_fault vmf;
2055                         int tmp;
2056
2057                         vmf.virtual_address = (void __user *)(address &
2058                                                                 PAGE_MASK);
2059                         vmf.pgoff = old_page->index;
2060                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2061                         vmf.page = old_page;
2062
2063                         /*
2064                          * Notify the address space that the page is about to
2065                          * become writable so that it can prohibit this or wait
2066                          * for the page to get into an appropriate state.
2067                          *
2068                          * We do this without the lock held, so that it can
2069                          * sleep if it needs to.
2070                          */
2071                         page_cache_get(old_page);
2072                         pte_unmap_unlock(page_table, ptl);
2073
2074                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2075                         if (unlikely(tmp &
2076                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2077                                 ret = tmp;
2078                                 goto unwritable_page;
2079                         }
2080                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2081                                 lock_page(old_page);
2082                                 if (!old_page->mapping) {
2083                                         ret = 0; /* retry the fault */
2084                                         unlock_page(old_page);
2085                                         goto unwritable_page;
2086                                 }
2087                         } else
2088                                 VM_BUG_ON(!PageLocked(old_page));
2089
2090                         /*
2091                          * Since we dropped the lock we need to revalidate
2092                          * the PTE as someone else may have changed it.  If
2093                          * they did, we just return, as we can count on the
2094                          * MMU to tell us if they didn't also make it writable.
2095                          */
2096                         page_table = pte_offset_map_lock(mm, pmd, address,
2097                                                          &ptl);
2098                         if (!pte_same(*page_table, orig_pte)) {
2099                                 unlock_page(old_page);
2100                                 page_cache_release(old_page);
2101                                 goto unlock;
2102                         }
2103
2104                         page_mkwrite = 1;
2105                 }
2106                 dirty_page = old_page;
2107                 get_page(dirty_page);
2108                 reuse = 1;
2109         }
2110
2111         if (reuse) {
2112 reuse:
2113                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2114                 entry = pte_mkyoung(orig_pte);
2115                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2116                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2117                         update_mmu_cache(vma, address, entry);
2118                 ret |= VM_FAULT_WRITE;
2119                 goto unlock;
2120         }
2121
2122         /*
2123          * Ok, we need to copy. Oh, well..
2124          */
2125         page_cache_get(old_page);
2126 gotten:
2127         pte_unmap_unlock(page_table, ptl);
2128
2129         if (unlikely(anon_vma_prepare(vma)))
2130                 goto oom;
2131
2132         if (is_zero_pfn(pte_pfn(orig_pte))) {
2133                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2134                 if (!new_page)
2135                         goto oom;
2136         } else {
2137                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2138                 if (!new_page)
2139                         goto oom;
2140                 cow_user_page(new_page, old_page, address, vma);
2141         }
2142         __SetPageUptodate(new_page);
2143
2144         /*
2145          * Don't let another task, with possibly unlocked vma,
2146          * keep the mlocked page.
2147          */
2148         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2149                 lock_page(old_page);    /* for LRU manipulation */
2150                 clear_page_mlock(old_page);
2151                 unlock_page(old_page);
2152         }
2153
2154         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2155                 goto oom_free_new;
2156
2157         /*
2158          * Re-check the pte - we dropped the lock
2159          */
2160         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2161         if (likely(pte_same(*page_table, orig_pte))) {
2162                 if (old_page) {
2163                         if (!PageAnon(old_page)) {
2164                                 dec_mm_counter(mm, file_rss);
2165                                 inc_mm_counter(mm, anon_rss);
2166                         }
2167                 } else
2168                         inc_mm_counter(mm, anon_rss);
2169                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2170                 entry = mk_pte(new_page, vma->vm_page_prot);
2171                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2172                 /*
2173                  * Clear the pte entry and flush it first, before updating the
2174                  * pte with the new entry. This will avoid a race condition
2175                  * seen in the presence of one thread doing SMC and another
2176                  * thread doing COW.
2177                  */
2178                 ptep_clear_flush(vma, address, page_table);
2179                 page_add_new_anon_rmap(new_page, vma, address);
2180                 /*
2181                  * We call the notify macro here because, when using secondary
2182                  * mmu page tables (such as kvm shadow page tables), we want the
2183                  * new page to be mapped directly into the secondary page table.
2184                  */
2185                 set_pte_at_notify(mm, address, page_table, entry);
2186                 update_mmu_cache(vma, address, entry);
2187                 if (old_page) {
2188                         /*
2189                          * Only after switching the pte to the new page may
2190                          * we remove the mapcount here. Otherwise another
2191                          * process may come and find the rmap count decremented
2192                          * before the pte is switched to the new page, and
2193                          * "reuse" the old page writing into it while our pte
2194                          * here still points into it and can be read by other
2195                          * threads.
2196                          *
2197                          * The critical issue is to order this
2198                          * page_remove_rmap with the ptp_clear_flush above.
2199                          * Those stores are ordered by (if nothing else,)
2200                          * the barrier present in the atomic_add_negative
2201                          * in page_remove_rmap.
2202                          *
2203                          * Then the TLB flush in ptep_clear_flush ensures that
2204                          * no process can access the old page before the
2205                          * decremented mapcount is visible. And the old page
2206                          * cannot be reused until after the decremented
2207                          * mapcount is visible. So transitively, TLBs to
2208                          * old page will be flushed before it can be reused.
2209                          */
2210                         page_remove_rmap(old_page);
2211                 }
2212
2213                 /* Free the old page.. */
2214                 new_page = old_page;
2215                 ret |= VM_FAULT_WRITE;
2216         } else
2217                 mem_cgroup_uncharge_page(new_page);
2218
2219         if (new_page)
2220                 page_cache_release(new_page);
2221         if (old_page)
2222                 page_cache_release(old_page);
2223 unlock:
2224         pte_unmap_unlock(page_table, ptl);
2225         if (dirty_page) {
2226                 /*
2227                  * Yes, Virginia, this is actually required to prevent a race
2228                  * with clear_page_dirty_for_io() from clearing the page dirty
2229                  * bit after it clear all dirty ptes, but before a racing
2230                  * do_wp_page installs a dirty pte.
2231                  *
2232                  * do_no_page is protected similarly.
2233                  */
2234                 if (!page_mkwrite) {
2235                         wait_on_page_locked(dirty_page);
2236                         set_page_dirty_balance(dirty_page, page_mkwrite);
2237                 }
2238                 put_page(dirty_page);
2239                 if (page_mkwrite) {
2240                         struct address_space *mapping = dirty_page->mapping;
2241
2242                         set_page_dirty(dirty_page);
2243                         unlock_page(dirty_page);
2244                         page_cache_release(dirty_page);
2245                         if (mapping)    {
2246                                 /*
2247                                  * Some device drivers do not set page.mapping
2248                                  * but still dirty their pages
2249                                  */
2250                                 balance_dirty_pages_ratelimited(mapping);
2251                         }
2252                 }
2253
2254                 /* file_update_time outside page_lock */
2255                 if (vma->vm_file)
2256                         file_update_time(vma->vm_file);
2257         }
2258         return ret;
2259 oom_free_new:
2260         page_cache_release(new_page);
2261 oom:
2262         if (old_page) {
2263                 if (page_mkwrite) {
2264                         unlock_page(old_page);
2265                         page_cache_release(old_page);
2266                 }
2267                 page_cache_release(old_page);
2268         }
2269         return VM_FAULT_OOM;
2270
2271 unwritable_page:
2272         page_cache_release(old_page);
2273         return ret;
2274 }
2275
2276 /*
2277  * Helper functions for unmap_mapping_range().
2278  *
2279  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2280  *
2281  * We have to restart searching the prio_tree whenever we drop the lock,
2282  * since the iterator is only valid while the lock is held, and anyway
2283  * a later vma might be split and reinserted earlier while lock dropped.
2284  *
2285  * The list of nonlinear vmas could be handled more efficiently, using
2286  * a placeholder, but handle it in the same way until a need is shown.
2287  * It is important to search the prio_tree before nonlinear list: a vma
2288  * may become nonlinear and be shifted from prio_tree to nonlinear list
2289  * while the lock is dropped; but never shifted from list to prio_tree.
2290  *
2291  * In order to make forward progress despite restarting the search,
2292  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2293  * quickly skip it next time around.  Since the prio_tree search only
2294  * shows us those vmas affected by unmapping the range in question, we
2295  * can't efficiently keep all vmas in step with mapping->truncate_count:
2296  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2297  * mapping->truncate_count and vma->vm_truncate_count are protected by
2298  * i_mmap_lock.
2299  *
2300  * In order to make forward progress despite repeatedly restarting some
2301  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2302  * and restart from that address when we reach that vma again.  It might
2303  * have been split or merged, shrunk or extended, but never shifted: so
2304  * restart_addr remains valid so long as it remains in the vma's range.
2305  * unmap_mapping_range forces truncate_count to leap over page-aligned
2306  * values so we can save vma's restart_addr in its truncate_count field.
2307  */
2308 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2309
2310 static void reset_vma_truncate_counts(struct address_space *mapping)
2311 {
2312         struct vm_area_struct *vma;
2313         struct prio_tree_iter iter;
2314
2315         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2316                 vma->vm_truncate_count = 0;
2317         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2318                 vma->vm_truncate_count = 0;
2319 }
2320
2321 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2322                 unsigned long start_addr, unsigned long end_addr,
2323                 struct zap_details *details)
2324 {
2325         unsigned long restart_addr;
2326         int need_break;
2327
2328         /*
2329          * files that support invalidating or truncating portions of the
2330          * file from under mmaped areas must have their ->fault function
2331          * return a locked page (and set VM_FAULT_LOCKED in the return).
2332          * This provides synchronisation against concurrent unmapping here.
2333          */
2334
2335 again:
2336         restart_addr = vma->vm_truncate_count;
2337         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2338                 start_addr = restart_addr;
2339                 if (start_addr >= end_addr) {
2340                         /* Top of vma has been split off since last time */
2341                         vma->vm_truncate_count = details->truncate_count;
2342                         return 0;
2343                 }
2344         }
2345
2346         restart_addr = zap_page_range(vma, start_addr,
2347                                         end_addr - start_addr, details);
2348         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2349
2350         if (restart_addr >= end_addr) {
2351                 /* We have now completed this vma: mark it so */
2352                 vma->vm_truncate_count = details->truncate_count;
2353                 if (!need_break)
2354                         return 0;
2355         } else {
2356                 /* Note restart_addr in vma's truncate_count field */
2357                 vma->vm_truncate_count = restart_addr;
2358                 if (!need_break)
2359                         goto again;
2360         }
2361
2362         spin_unlock(details->i_mmap_lock);
2363         cond_resched();
2364         spin_lock(details->i_mmap_lock);
2365         return -EINTR;
2366 }
2367
2368 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2369                                             struct zap_details *details)
2370 {
2371         struct vm_area_struct *vma;
2372         struct prio_tree_iter iter;
2373         pgoff_t vba, vea, zba, zea;
2374
2375 restart:
2376         vma_prio_tree_foreach(vma, &iter, root,
2377                         details->first_index, details->last_index) {
2378                 /* Skip quickly over those we have already dealt with */
2379                 if (vma->vm_truncate_count == details->truncate_count)
2380                         continue;
2381
2382                 vba = vma->vm_pgoff;
2383                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2384                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2385                 zba = details->first_index;
2386                 if (zba < vba)
2387                         zba = vba;
2388                 zea = details->last_index;
2389                 if (zea > vea)
2390                         zea = vea;
2391
2392                 if (unmap_mapping_range_vma(vma,
2393                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2394                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2395                                 details) < 0)
2396                         goto restart;
2397         }
2398 }
2399
2400 static inline void unmap_mapping_range_list(struct list_head *head,
2401                                             struct zap_details *details)
2402 {
2403         struct vm_area_struct *vma;
2404
2405         /*
2406          * In nonlinear VMAs there is no correspondence between virtual address
2407          * offset and file offset.  So we must perform an exhaustive search
2408          * across *all* the pages in each nonlinear VMA, not just the pages
2409          * whose virtual address lies outside the file truncation point.
2410          */
2411 restart:
2412         list_for_each_entry(vma, head, shared.vm_set.list) {
2413                 /* Skip quickly over those we have already dealt with */
2414                 if (vma->vm_truncate_count == details->truncate_count)
2415                         continue;
2416                 details->nonlinear_vma = vma;
2417                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2418                                         vma->vm_end, details) < 0)
2419                         goto restart;
2420         }
2421 }
2422
2423 /**
2424  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2425  * @mapping: the address space containing mmaps to be unmapped.
2426  * @holebegin: byte in first page to unmap, relative to the start of
2427  * the underlying file.  This will be rounded down to a PAGE_SIZE
2428  * boundary.  Note that this is different from truncate_pagecache(), which
2429  * must keep the partial page.  In contrast, we must get rid of
2430  * partial pages.
2431  * @holelen: size of prospective hole in bytes.  This will be rounded
2432  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2433  * end of the file.
2434  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2435  * but 0 when invalidating pagecache, don't throw away private data.
2436  */
2437 void unmap_mapping_range(struct address_space *mapping,
2438                 loff_t const holebegin, loff_t const holelen, int even_cows)
2439 {
2440         struct zap_details details;
2441         pgoff_t hba = holebegin >> PAGE_SHIFT;
2442         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2443
2444         /* Check for overflow. */
2445         if (sizeof(holelen) > sizeof(hlen)) {
2446                 long long holeend =
2447                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2448                 if (holeend & ~(long long)ULONG_MAX)
2449                         hlen = ULONG_MAX - hba + 1;
2450         }
2451
2452         details.check_mapping = even_cows? NULL: mapping;
2453         details.nonlinear_vma = NULL;
2454         details.first_index = hba;
2455         details.last_index = hba + hlen - 1;
2456         if (details.last_index < details.first_index)
2457                 details.last_index = ULONG_MAX;
2458         details.i_mmap_lock = &mapping->i_mmap_lock;
2459
2460         spin_lock(&mapping->i_mmap_lock);
2461
2462         /* Protect against endless unmapping loops */
2463         mapping->truncate_count++;
2464         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2465                 if (mapping->truncate_count == 0)
2466                         reset_vma_truncate_counts(mapping);
2467                 mapping->truncate_count++;
2468         }
2469         details.truncate_count = mapping->truncate_count;
2470
2471         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2472                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2473         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2474                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2475         spin_unlock(&mapping->i_mmap_lock);
2476 }
2477 EXPORT_SYMBOL(unmap_mapping_range);
2478
2479 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2480 {
2481         struct address_space *mapping = inode->i_mapping;
2482
2483         /*
2484          * If the underlying filesystem is not going to provide
2485          * a way to truncate a range of blocks (punch a hole) -
2486          * we should return failure right now.
2487          */
2488         if (!inode->i_op->truncate_range)
2489                 return -ENOSYS;
2490
2491         mutex_lock(&inode->i_mutex);
2492         down_write(&inode->i_alloc_sem);
2493         unmap_mapping_range(mapping, offset, (end - offset), 1);
2494         truncate_inode_pages_range(mapping, offset, end);
2495         unmap_mapping_range(mapping, offset, (end - offset), 1);
2496         inode->i_op->truncate_range(inode, offset, end);
2497         up_write(&inode->i_alloc_sem);
2498         mutex_unlock(&inode->i_mutex);
2499
2500         return 0;
2501 }
2502
2503 /*
2504  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2505  * but allow concurrent faults), and pte mapped but not yet locked.
2506  * We return with mmap_sem still held, but pte unmapped and unlocked.
2507  */
2508 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2509                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2510                 unsigned int flags, pte_t orig_pte)
2511 {
2512         spinlock_t *ptl;
2513         struct page *page;
2514         swp_entry_t entry;
2515         pte_t pte;
2516         struct mem_cgroup *ptr = NULL;
2517         int ret = 0;
2518
2519         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2520                 goto out;
2521
2522         entry = pte_to_swp_entry(orig_pte);
2523         if (unlikely(non_swap_entry(entry))) {
2524                 if (is_migration_entry(entry)) {
2525                         migration_entry_wait(mm, pmd, address);
2526                 } else if (is_hwpoison_entry(entry)) {
2527                         ret = VM_FAULT_HWPOISON;
2528                 } else {
2529                         print_bad_pte(vma, address, orig_pte, NULL);
2530                         ret = VM_FAULT_SIGBUS;
2531                 }
2532                 goto out;
2533         }
2534         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2535         page = lookup_swap_cache(entry);
2536         if (!page) {
2537                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2538                 page = swapin_readahead(entry,
2539                                         GFP_HIGHUSER_MOVABLE, vma, address);
2540                 if (!page) {
2541                         /*
2542                          * Back out if somebody else faulted in this pte
2543                          * while we released the pte lock.
2544                          */
2545                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2546                         if (likely(pte_same(*page_table, orig_pte)))
2547                                 ret = VM_FAULT_OOM;
2548                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2549                         goto unlock;
2550                 }
2551
2552                 /* Had to read the page from swap area: Major fault */
2553                 ret = VM_FAULT_MAJOR;
2554                 count_vm_event(PGMAJFAULT);
2555         } else if (PageHWPoison(page)) {
2556                 ret = VM_FAULT_HWPOISON;
2557                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2558                 goto out_release;
2559         }
2560
2561         lock_page(page);
2562         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2563
2564         page = ksm_might_need_to_copy(page, vma, address);
2565         if (!page) {
2566                 ret = VM_FAULT_OOM;
2567                 goto out;
2568         }
2569
2570         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2571                 ret = VM_FAULT_OOM;
2572                 goto out_page;
2573         }
2574
2575         /*
2576          * Back out if somebody else already faulted in this pte.
2577          */
2578         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2579         if (unlikely(!pte_same(*page_table, orig_pte)))
2580                 goto out_nomap;
2581
2582         if (unlikely(!PageUptodate(page))) {
2583                 ret = VM_FAULT_SIGBUS;
2584                 goto out_nomap;
2585         }
2586
2587         /*
2588          * The page isn't present yet, go ahead with the fault.
2589          *
2590          * Be careful about the sequence of operations here.
2591          * To get its accounting right, reuse_swap_page() must be called
2592          * while the page is counted on swap but not yet in mapcount i.e.
2593          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2594          * must be called after the swap_free(), or it will never succeed.
2595          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2596          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2597          * in page->private. In this case, a record in swap_cgroup  is silently
2598          * discarded at swap_free().
2599          */
2600
2601         inc_mm_counter(mm, anon_rss);
2602         pte = mk_pte(page, vma->vm_page_prot);
2603         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2604                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2605                 flags &= ~FAULT_FLAG_WRITE;
2606         }
2607         flush_icache_page(vma, page);
2608         set_pte_at(mm, address, page_table, pte);
2609         page_add_anon_rmap(page, vma, address);
2610         /* It's better to call commit-charge after rmap is established */
2611         mem_cgroup_commit_charge_swapin(page, ptr);
2612
2613         swap_free(entry);
2614         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2615                 try_to_free_swap(page);
2616         unlock_page(page);
2617
2618         if (flags & FAULT_FLAG_WRITE) {
2619                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2620                 if (ret & VM_FAULT_ERROR)
2621                         ret &= VM_FAULT_ERROR;
2622                 goto out;
2623         }
2624
2625         /* No need to invalidate - it was non-present before */
2626         update_mmu_cache(vma, address, pte);
2627 unlock:
2628         pte_unmap_unlock(page_table, ptl);
2629 out:
2630         return ret;
2631 out_nomap:
2632         mem_cgroup_cancel_charge_swapin(ptr);
2633         pte_unmap_unlock(page_table, ptl);
2634 out_page:
2635         unlock_page(page);
2636 out_release:
2637         page_cache_release(page);
2638         return ret;
2639 }
2640
2641 /*
2642  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2643  * but allow concurrent faults), and pte mapped but not yet locked.
2644  * We return with mmap_sem still held, but pte unmapped and unlocked.
2645  */
2646 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2647                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2648                 unsigned int flags)
2649 {
2650         struct page *page;
2651         spinlock_t *ptl;
2652         pte_t entry;
2653
2654         if (!(flags & FAULT_FLAG_WRITE)) {
2655                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2656                                                 vma->vm_page_prot));
2657                 ptl = pte_lockptr(mm, pmd);
2658                 spin_lock(ptl);
2659                 if (!pte_none(*page_table))
2660                         goto unlock;
2661                 goto setpte;
2662         }
2663
2664         /* Allocate our own private page. */
2665         pte_unmap(page_table);
2666
2667         if (unlikely(anon_vma_prepare(vma)))
2668                 goto oom;
2669         page = alloc_zeroed_user_highpage_movable(vma, address);
2670         if (!page)
2671                 goto oom;
2672         __SetPageUptodate(page);
2673
2674         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2675                 goto oom_free_page;
2676
2677         entry = mk_pte(page, vma->vm_page_prot);
2678         if (vma->vm_flags & VM_WRITE)
2679                 entry = pte_mkwrite(pte_mkdirty(entry));
2680
2681         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2682         if (!pte_none(*page_table))
2683                 goto release;
2684
2685         inc_mm_counter(mm, anon_rss);
2686         page_add_new_anon_rmap(page, vma, address);
2687 setpte:
2688         set_pte_at(mm, address, page_table, entry);
2689
2690         /* No need to invalidate - it was non-present before */
2691         update_mmu_cache(vma, address, entry);
2692 unlock:
2693         pte_unmap_unlock(page_table, ptl);
2694         return 0;
2695 release:
2696         mem_cgroup_uncharge_page(page);
2697         page_cache_release(page);
2698         goto unlock;
2699 oom_free_page:
2700         page_cache_release(page);
2701 oom:
2702         return VM_FAULT_OOM;
2703 }
2704
2705 /*
2706  * __do_fault() tries to create a new page mapping. It aggressively
2707  * tries to share with existing pages, but makes a separate copy if
2708  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2709  * the next page fault.
2710  *
2711  * As this is called only for pages that do not currently exist, we
2712  * do not need to flush old virtual caches or the TLB.
2713  *
2714  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2715  * but allow concurrent faults), and pte neither mapped nor locked.
2716  * We return with mmap_sem still held, but pte unmapped and unlocked.
2717  */
2718 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2719                 unsigned long address, pmd_t *pmd,
2720                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2721 {
2722         pte_t *page_table;
2723         spinlock_t *ptl;
2724         struct page *page;
2725         pte_t entry;
2726         int anon = 0;
2727         int charged = 0;
2728         struct page *dirty_page = NULL;
2729         struct vm_fault vmf;
2730         int ret;
2731         int page_mkwrite = 0;
2732
2733         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2734         vmf.pgoff = pgoff;
2735         vmf.flags = flags;
2736         vmf.page = NULL;
2737
2738         ret = vma->vm_ops->fault(vma, &vmf);
2739         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2740                 return ret;
2741
2742         if (unlikely(PageHWPoison(vmf.page))) {
2743                 if (ret & VM_FAULT_LOCKED)
2744                         unlock_page(vmf.page);
2745                 return VM_FAULT_HWPOISON;
2746         }
2747
2748         /*
2749          * For consistency in subsequent calls, make the faulted page always
2750          * locked.
2751          */
2752         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2753                 lock_page(vmf.page);
2754         else
2755                 VM_BUG_ON(!PageLocked(vmf.page));
2756
2757         /*
2758          * Should we do an early C-O-W break?
2759          */
2760         page = vmf.page;
2761         if (flags & FAULT_FLAG_WRITE) {
2762                 if (!(vma->vm_flags & VM_SHARED)) {
2763                         anon = 1;
2764                         if (unlikely(anon_vma_prepare(vma))) {
2765                                 ret = VM_FAULT_OOM;
2766                                 goto out;
2767                         }
2768                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2769                                                 vma, address);
2770                         if (!page) {
2771                                 ret = VM_FAULT_OOM;
2772                                 goto out;
2773                         }
2774                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2775                                 ret = VM_FAULT_OOM;
2776                                 page_cache_release(page);
2777                                 goto out;
2778                         }
2779                         charged = 1;
2780                         /*
2781                          * Don't let another task, with possibly unlocked vma,
2782                          * keep the mlocked page.
2783                          */
2784                         if (vma->vm_flags & VM_LOCKED)
2785                                 clear_page_mlock(vmf.page);
2786                         copy_user_highpage(page, vmf.page, address, vma);
2787                         __SetPageUptodate(page);
2788                 } else {
2789                         /*
2790                          * If the page will be shareable, see if the backing
2791                          * address space wants to know that the page is about
2792                          * to become writable
2793                          */
2794                         if (vma->vm_ops->page_mkwrite) {
2795                                 int tmp;
2796
2797                                 unlock_page(page);
2798                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2799                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2800                                 if (unlikely(tmp &
2801                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2802                                         ret = tmp;
2803                                         goto unwritable_page;
2804                                 }
2805                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2806                                         lock_page(page);
2807                                         if (!page->mapping) {
2808                                                 ret = 0; /* retry the fault */
2809                                                 unlock_page(page);
2810                                                 goto unwritable_page;
2811                                         }
2812                                 } else
2813                                         VM_BUG_ON(!PageLocked(page));
2814                                 page_mkwrite = 1;
2815                         }
2816                 }
2817
2818         }
2819
2820         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2821
2822         /*
2823          * This silly early PAGE_DIRTY setting removes a race
2824          * due to the bad i386 page protection. But it's valid
2825          * for other architectures too.
2826          *
2827          * Note that if FAULT_FLAG_WRITE is set, we either now have
2828          * an exclusive copy of the page, or this is a shared mapping,
2829          * so we can make it writable and dirty to avoid having to
2830          * handle that later.
2831          */
2832         /* Only go through if we didn't race with anybody else... */
2833         if (likely(pte_same(*page_table, orig_pte))) {
2834                 flush_icache_page(vma, page);
2835                 entry = mk_pte(page, vma->vm_page_prot);
2836                 if (flags & FAULT_FLAG_WRITE)
2837                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2838                 if (anon) {
2839                         inc_mm_counter(mm, anon_rss);
2840                         page_add_new_anon_rmap(page, vma, address);
2841                 } else {
2842                         inc_mm_counter(mm, file_rss);
2843                         page_add_file_rmap(page);
2844                         if (flags & FAULT_FLAG_WRITE) {
2845                                 dirty_page = page;
2846                                 get_page(dirty_page);
2847                         }
2848                 }
2849                 set_pte_at(mm, address, page_table, entry);
2850
2851                 /* no need to invalidate: a not-present page won't be cached */
2852                 update_mmu_cache(vma, address, entry);
2853         } else {
2854                 if (charged)
2855                         mem_cgroup_uncharge_page(page);
2856                 if (anon)
2857                         page_cache_release(page);
2858                 else
2859                         anon = 1; /* no anon but release faulted_page */
2860         }
2861
2862         pte_unmap_unlock(page_table, ptl);
2863
2864 out:
2865         if (dirty_page) {
2866                 struct address_space *mapping = page->mapping;
2867
2868                 if (set_page_dirty(dirty_page))
2869                         page_mkwrite = 1;
2870                 unlock_page(dirty_page);
2871                 put_page(dirty_page);
2872                 if (page_mkwrite && mapping) {
2873                         /*
2874                          * Some device drivers do not set page.mapping but still
2875                          * dirty their pages
2876                          */
2877                         balance_dirty_pages_ratelimited(mapping);
2878                 }
2879
2880                 /* file_update_time outside page_lock */
2881                 if (vma->vm_file)
2882                         file_update_time(vma->vm_file);
2883         } else {
2884                 unlock_page(vmf.page);
2885                 if (anon)
2886                         page_cache_release(vmf.page);
2887         }
2888
2889         return ret;
2890
2891 unwritable_page:
2892         page_cache_release(page);
2893         return ret;
2894 }
2895
2896 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2897                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2898                 unsigned int flags, pte_t orig_pte)
2899 {
2900         pgoff_t pgoff = (((address & PAGE_MASK)
2901                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2902
2903         pte_unmap(page_table);
2904         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2905 }
2906
2907 /*
2908  * Fault of a previously existing named mapping. Repopulate the pte
2909  * from the encoded file_pte if possible. This enables swappable
2910  * nonlinear vmas.
2911  *
2912  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2913  * but allow concurrent faults), and pte mapped but not yet locked.
2914  * We return with mmap_sem still held, but pte unmapped and unlocked.
2915  */
2916 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2917                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2918                 unsigned int flags, pte_t orig_pte)
2919 {
2920         pgoff_t pgoff;
2921
2922         flags |= FAULT_FLAG_NONLINEAR;
2923
2924         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2925                 return 0;
2926
2927         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2928                 /*
2929                  * Page table corrupted: show pte and kill process.
2930                  */
2931                 print_bad_pte(vma, address, orig_pte, NULL);
2932                 return VM_FAULT_SIGBUS;
2933         }
2934
2935         pgoff = pte_to_pgoff(orig_pte);
2936         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2937 }
2938
2939 /*
2940  * These routines also need to handle stuff like marking pages dirty
2941  * and/or accessed for architectures that don't do it in hardware (most
2942  * RISC architectures).  The early dirtying is also good on the i386.
2943  *
2944  * There is also a hook called "update_mmu_cache()" that architectures
2945  * with external mmu caches can use to update those (ie the Sparc or
2946  * PowerPC hashed page tables that act as extended TLBs).
2947  *
2948  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2949  * but allow concurrent faults), and pte mapped but not yet locked.
2950  * We return with mmap_sem still held, but pte unmapped and unlocked.
2951  */
2952 static inline int handle_pte_fault(struct mm_struct *mm,
2953                 struct vm_area_struct *vma, unsigned long address,
2954                 pte_t *pte, pmd_t *pmd, unsigned int flags)
2955 {
2956         pte_t entry;
2957         spinlock_t *ptl;
2958
2959         entry = *pte;
2960         if (!pte_present(entry)) {
2961                 if (pte_none(entry)) {
2962                         if (vma->vm_ops) {
2963                                 if (likely(vma->vm_ops->fault))
2964                                         return do_linear_fault(mm, vma, address,
2965                                                 pte, pmd, flags, entry);
2966                         }
2967                         return do_anonymous_page(mm, vma, address,
2968                                                  pte, pmd, flags);
2969                 }
2970                 if (pte_file(entry))
2971                         return do_nonlinear_fault(mm, vma, address,
2972                                         pte, pmd, flags, entry);
2973                 return do_swap_page(mm, vma, address,
2974                                         pte, pmd, flags, entry);
2975         }
2976
2977         ptl = pte_lockptr(mm, pmd);
2978         spin_lock(ptl);
2979         if (unlikely(!pte_same(*pte, entry)))
2980                 goto unlock;
2981         if (flags & FAULT_FLAG_WRITE) {
2982                 if (!pte_write(entry))
2983                         return do_wp_page(mm, vma, address,
2984                                         pte, pmd, ptl, entry);
2985                 entry = pte_mkdirty(entry);
2986         }
2987         entry = pte_mkyoung(entry);
2988         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
2989                 update_mmu_cache(vma, address, entry);
2990         } else {
2991                 /*
2992                  * This is needed only for protection faults but the arch code
2993                  * is not yet telling us if this is a protection fault or not.
2994                  * This still avoids useless tlb flushes for .text page faults
2995                  * with threads.
2996                  */
2997                 if (flags & FAULT_FLAG_WRITE)
2998                         flush_tlb_page(vma, address);
2999         }
3000 unlock:
3001         pte_unmap_unlock(pte, ptl);
3002         return 0;
3003 }
3004
3005 /*
3006  * By the time we get here, we already hold the mm semaphore
3007  */
3008 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3009                 unsigned long address, unsigned int flags)
3010 {
3011         pgd_t *pgd;
3012         pud_t *pud;
3013         pmd_t *pmd;
3014         pte_t *pte;
3015
3016         __set_current_state(TASK_RUNNING);
3017
3018         count_vm_event(PGFAULT);
3019
3020         if (unlikely(is_vm_hugetlb_page(vma)))
3021                 return hugetlb_fault(mm, vma, address, flags);
3022
3023         pgd = pgd_offset(mm, address);
3024         pud = pud_alloc(mm, pgd, address);
3025         if (!pud)
3026                 return VM_FAULT_OOM;
3027         pmd = pmd_alloc(mm, pud, address);
3028         if (!pmd)
3029                 return VM_FAULT_OOM;
3030         pte = pte_alloc_map(mm, pmd, address);
3031         if (!pte)
3032                 return VM_FAULT_OOM;
3033
3034         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3035 }
3036
3037 #ifndef __PAGETABLE_PUD_FOLDED
3038 /*
3039  * Allocate page upper directory.
3040  * We've already handled the fast-path in-line.
3041  */
3042 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3043 {
3044         pud_t *new = pud_alloc_one(mm, address);
3045         if (!new)
3046                 return -ENOMEM;
3047
3048         smp_wmb(); /* See comment in __pte_alloc */
3049
3050         spin_lock(&mm->page_table_lock);
3051         if (pgd_present(*pgd))          /* Another has populated it */
3052                 pud_free(mm, new);
3053         else
3054                 pgd_populate(mm, pgd, new);
3055         spin_unlock(&mm->page_table_lock);
3056         return 0;
3057 }
3058 #endif /* __PAGETABLE_PUD_FOLDED */
3059
3060 #ifndef __PAGETABLE_PMD_FOLDED
3061 /*
3062  * Allocate page middle directory.
3063  * We've already handled the fast-path in-line.
3064  */
3065 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3066 {
3067         pmd_t *new = pmd_alloc_one(mm, address);
3068         if (!new)
3069                 return -ENOMEM;
3070
3071         smp_wmb(); /* See comment in __pte_alloc */
3072
3073         spin_lock(&mm->page_table_lock);
3074 #ifndef __ARCH_HAS_4LEVEL_HACK
3075         if (pud_present(*pud))          /* Another has populated it */
3076                 pmd_free(mm, new);
3077         else
3078                 pud_populate(mm, pud, new);
3079 #else
3080         if (pgd_present(*pud))          /* Another has populated it */
3081                 pmd_free(mm, new);
3082         else
3083                 pgd_populate(mm, pud, new);
3084 #endif /* __ARCH_HAS_4LEVEL_HACK */
3085         spin_unlock(&mm->page_table_lock);
3086         return 0;
3087 }
3088 #endif /* __PAGETABLE_PMD_FOLDED */
3089
3090 int make_pages_present(unsigned long addr, unsigned long end)
3091 {
3092         int ret, len, write;
3093         struct vm_area_struct * vma;
3094
3095         vma = find_vma(current->mm, addr);
3096         if (!vma)
3097                 return -ENOMEM;
3098         write = (vma->vm_flags & VM_WRITE) != 0;
3099         BUG_ON(addr >= end);
3100         BUG_ON(end > vma->vm_end);
3101         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3102         ret = get_user_pages(current, current->mm, addr,
3103                         len, write, 0, NULL, NULL);
3104         if (ret < 0)
3105                 return ret;
3106         return ret == len ? 0 : -EFAULT;
3107 }
3108
3109 #if !defined(__HAVE_ARCH_GATE_AREA)
3110
3111 #if defined(AT_SYSINFO_EHDR)
3112 static struct vm_area_struct gate_vma;
3113
3114 static int __init gate_vma_init(void)
3115 {
3116         gate_vma.vm_mm = NULL;
3117         gate_vma.vm_start = FIXADDR_USER_START;
3118         gate_vma.vm_end = FIXADDR_USER_END;
3119         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3120         gate_vma.vm_page_prot = __P101;
3121         /*
3122          * Make sure the vDSO gets into every core dump.
3123          * Dumping its contents makes post-mortem fully interpretable later
3124          * without matching up the same kernel and hardware config to see
3125          * what PC values meant.
3126          */
3127         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3128         return 0;
3129 }
3130 __initcall(gate_vma_init);
3131 #endif
3132
3133 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3134 {
3135 #ifdef AT_SYSINFO_EHDR
3136         return &gate_vma;
3137 #else
3138         return NULL;
3139 #endif
3140 }
3141
3142 int in_gate_area_no_task(unsigned long addr)
3143 {
3144 #ifdef AT_SYSINFO_EHDR
3145         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3146                 return 1;
3147 #endif
3148         return 0;
3149 }
3150
3151 #endif  /* __HAVE_ARCH_GATE_AREA */
3152
3153 static int follow_pte(struct mm_struct *mm, unsigned long address,
3154                 pte_t **ptepp, spinlock_t **ptlp)
3155 {
3156         pgd_t *pgd;
3157         pud_t *pud;
3158         pmd_t *pmd;
3159         pte_t *ptep;
3160
3161         pgd = pgd_offset(mm, address);
3162         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3163                 goto out;
3164
3165         pud = pud_offset(pgd, address);
3166         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3167                 goto out;
3168
3169         pmd = pmd_offset(pud, address);
3170         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3171                 goto out;
3172
3173         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3174         if (pmd_huge(*pmd))
3175                 goto out;
3176
3177         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3178         if (!ptep)
3179                 goto out;
3180         if (!pte_present(*ptep))
3181                 goto unlock;
3182         *ptepp = ptep;
3183         return 0;
3184 unlock:
3185         pte_unmap_unlock(ptep, *ptlp);
3186 out:
3187         return -EINVAL;
3188 }
3189
3190 /**
3191  * follow_pfn - look up PFN at a user virtual address
3192  * @vma: memory mapping
3193  * @address: user virtual address
3194  * @pfn: location to store found PFN
3195  *
3196  * Only IO mappings and raw PFN mappings are allowed.
3197  *
3198  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3199  */
3200 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3201         unsigned long *pfn)
3202 {
3203         int ret = -EINVAL;
3204         spinlock_t *ptl;
3205         pte_t *ptep;
3206
3207         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3208                 return ret;
3209
3210         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3211         if (ret)
3212                 return ret;
3213         *pfn = pte_pfn(*ptep);
3214         pte_unmap_unlock(ptep, ptl);
3215         return 0;
3216 }
3217 EXPORT_SYMBOL(follow_pfn);
3218
3219 #ifdef CONFIG_HAVE_IOREMAP_PROT
3220 int follow_phys(struct vm_area_struct *vma,
3221                 unsigned long address, unsigned int flags,
3222                 unsigned long *prot, resource_size_t *phys)
3223 {
3224         int ret = -EINVAL;
3225         pte_t *ptep, pte;
3226         spinlock_t *ptl;
3227
3228         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3229                 goto out;
3230
3231         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3232                 goto out;
3233         pte = *ptep;
3234
3235         if ((flags & FOLL_WRITE) && !pte_write(pte))
3236                 goto unlock;
3237
3238         *prot = pgprot_val(pte_pgprot(pte));
3239         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3240
3241         ret = 0;
3242 unlock:
3243         pte_unmap_unlock(ptep, ptl);
3244 out:
3245         return ret;
3246 }
3247
3248 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3249                         void *buf, int len, int write)
3250 {
3251         resource_size_t phys_addr;
3252         unsigned long prot = 0;
3253         void __iomem *maddr;
3254         int offset = addr & (PAGE_SIZE-1);
3255
3256         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3257                 return -EINVAL;
3258
3259         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3260         if (write)
3261                 memcpy_toio(maddr + offset, buf, len);
3262         else
3263                 memcpy_fromio(buf, maddr + offset, len);
3264         iounmap(maddr);
3265
3266         return len;
3267 }
3268 #endif
3269
3270 /*
3271  * Access another process' address space.
3272  * Source/target buffer must be kernel space,
3273  * Do not walk the page table directly, use get_user_pages
3274  */
3275 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3276 {
3277         struct mm_struct *mm;
3278         struct vm_area_struct *vma;
3279         void *old_buf = buf;
3280
3281         mm = get_task_mm(tsk);
3282         if (!mm)
3283                 return 0;
3284
3285         down_read(&mm->mmap_sem);
3286         /* ignore errors, just check how much was successfully transferred */
3287         while (len) {
3288                 int bytes, ret, offset;
3289                 void *maddr;
3290                 struct page *page = NULL;
3291
3292                 ret = get_user_pages(tsk, mm, addr, 1,
3293                                 write, 1, &page, &vma);
3294                 if (ret <= 0) {
3295                         /*
3296                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3297                          * we can access using slightly different code.
3298                          */
3299 #ifdef CONFIG_HAVE_IOREMAP_PROT
3300                         vma = find_vma(mm, addr);
3301                         if (!vma)
3302                                 break;
3303                         if (vma->vm_ops && vma->vm_ops->access)
3304                                 ret = vma->vm_ops->access(vma, addr, buf,
3305                                                           len, write);
3306                         if (ret <= 0)
3307 #endif
3308                                 break;
3309                         bytes = ret;
3310                 } else {
3311                         bytes = len;
3312                         offset = addr & (PAGE_SIZE-1);
3313                         if (bytes > PAGE_SIZE-offset)
3314                                 bytes = PAGE_SIZE-offset;
3315
3316                         maddr = kmap(page);
3317                         if (write) {
3318                                 copy_to_user_page(vma, page, addr,
3319                                                   maddr + offset, buf, bytes);
3320                                 set_page_dirty_lock(page);
3321                         } else {
3322                                 copy_from_user_page(vma, page, addr,
3323                                                     buf, maddr + offset, bytes);
3324                         }
3325                         kunmap(page);
3326                         page_cache_release(page);
3327                 }
3328                 len -= bytes;
3329                 buf += bytes;
3330                 addr += bytes;
3331         }
3332         up_read(&mm->mmap_sem);
3333         mmput(mm);
3334
3335         return buf - old_buf;
3336 }
3337
3338 /*
3339  * Print the name of a VMA.
3340  */
3341 void print_vma_addr(char *prefix, unsigned long ip)
3342 {
3343         struct mm_struct *mm = current->mm;
3344         struct vm_area_struct *vma;
3345
3346         /*
3347          * Do not print if we are in atomic
3348          * contexts (in exception stacks, etc.):
3349          */
3350         if (preempt_count())
3351                 return;
3352
3353         down_read(&mm->mmap_sem);
3354         vma = find_vma(mm, ip);
3355         if (vma && vma->vm_file) {
3356                 struct file *f = vma->vm_file;
3357                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3358                 if (buf) {
3359                         char *p, *s;
3360
3361                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3362                         if (IS_ERR(p))
3363                                 p = "?";
3364                         s = strrchr(p, '/');
3365                         if (s)
3366                                 p = s+1;
3367                         printk("%s%s[%lx+%lx]", prefix, p,
3368                                         vma->vm_start,
3369                                         vma->vm_end - vma->vm_start);
3370                         free_page((unsigned long)buf);
3371                 }
3372         }
3373         up_read(&current->mm->mmap_sem);
3374 }
3375
3376 #ifdef CONFIG_PROVE_LOCKING
3377 void might_fault(void)
3378 {
3379         /*
3380          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3381          * holding the mmap_sem, this is safe because kernel memory doesn't
3382          * get paged out, therefore we'll never actually fault, and the
3383          * below annotations will generate false positives.
3384          */
3385         if (segment_eq(get_fs(), KERNEL_DS))
3386                 return;
3387
3388         might_sleep();
3389         /*
3390          * it would be nicer only to annotate paths which are not under
3391          * pagefault_disable, however that requires a larger audit and
3392          * providing helpers like get_user_atomic.
3393          */
3394         if (!in_atomic() && current->mm)
3395                 might_lock_read(&current->mm->mmap_sem);
3396 }
3397 EXPORT_SYMBOL(might_fault);
3398 #endif