badpage: remove vma from page_remove_rmap
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/delayacct.h>
51 #include <linux/init.h>
52 #include <linux/writeback.h>
53 #include <linux/memcontrol.h>
54 #include <linux/mmu_notifier.h>
55 #include <linux/kallsyms.h>
56 #include <linux/swapops.h>
57 #include <linux/elf.h>
58
59 #include <asm/pgalloc.h>
60 #include <asm/uaccess.h>
61 #include <asm/tlb.h>
62 #include <asm/tlbflush.h>
63 #include <asm/pgtable.h>
64
65 #include "internal.h"
66
67 #ifndef CONFIG_NEED_MULTIPLE_NODES
68 /* use the per-pgdat data instead for discontigmem - mbligh */
69 unsigned long max_mapnr;
70 struct page *mem_map;
71
72 EXPORT_SYMBOL(max_mapnr);
73 EXPORT_SYMBOL(mem_map);
74 #endif
75
76 unsigned long num_physpages;
77 /*
78  * A number of key systems in x86 including ioremap() rely on the assumption
79  * that high_memory defines the upper bound on direct map memory, then end
80  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
81  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
82  * and ZONE_HIGHMEM.
83  */
84 void * high_memory;
85
86 EXPORT_SYMBOL(num_physpages);
87 EXPORT_SYMBOL(high_memory);
88
89 /*
90  * Randomize the address space (stacks, mmaps, brk, etc.).
91  *
92  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
93  *   as ancient (libc5 based) binaries can segfault. )
94  */
95 int randomize_va_space __read_mostly =
96 #ifdef CONFIG_COMPAT_BRK
97                                         1;
98 #else
99                                         2;
100 #endif
101
102 static int __init disable_randmaps(char *s)
103 {
104         randomize_va_space = 0;
105         return 1;
106 }
107 __setup("norandmaps", disable_randmaps);
108
109
110 /*
111  * If a p?d_bad entry is found while walking page tables, report
112  * the error, before resetting entry to p?d_none.  Usually (but
113  * very seldom) called out from the p?d_none_or_clear_bad macros.
114  */
115
116 void pgd_clear_bad(pgd_t *pgd)
117 {
118         pgd_ERROR(*pgd);
119         pgd_clear(pgd);
120 }
121
122 void pud_clear_bad(pud_t *pud)
123 {
124         pud_ERROR(*pud);
125         pud_clear(pud);
126 }
127
128 void pmd_clear_bad(pmd_t *pmd)
129 {
130         pmd_ERROR(*pmd);
131         pmd_clear(pmd);
132 }
133
134 /*
135  * Note: this doesn't free the actual pages themselves. That
136  * has been handled earlier when unmapping all the memory regions.
137  */
138 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
139 {
140         pgtable_t token = pmd_pgtable(*pmd);
141         pmd_clear(pmd);
142         pte_free_tlb(tlb, token);
143         tlb->mm->nr_ptes--;
144 }
145
146 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
147                                 unsigned long addr, unsigned long end,
148                                 unsigned long floor, unsigned long ceiling)
149 {
150         pmd_t *pmd;
151         unsigned long next;
152         unsigned long start;
153
154         start = addr;
155         pmd = pmd_offset(pud, addr);
156         do {
157                 next = pmd_addr_end(addr, end);
158                 if (pmd_none_or_clear_bad(pmd))
159                         continue;
160                 free_pte_range(tlb, pmd);
161         } while (pmd++, addr = next, addr != end);
162
163         start &= PUD_MASK;
164         if (start < floor)
165                 return;
166         if (ceiling) {
167                 ceiling &= PUD_MASK;
168                 if (!ceiling)
169                         return;
170         }
171         if (end - 1 > ceiling - 1)
172                 return;
173
174         pmd = pmd_offset(pud, start);
175         pud_clear(pud);
176         pmd_free_tlb(tlb, pmd);
177 }
178
179 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
180                                 unsigned long addr, unsigned long end,
181                                 unsigned long floor, unsigned long ceiling)
182 {
183         pud_t *pud;
184         unsigned long next;
185         unsigned long start;
186
187         start = addr;
188         pud = pud_offset(pgd, addr);
189         do {
190                 next = pud_addr_end(addr, end);
191                 if (pud_none_or_clear_bad(pud))
192                         continue;
193                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
194         } while (pud++, addr = next, addr != end);
195
196         start &= PGDIR_MASK;
197         if (start < floor)
198                 return;
199         if (ceiling) {
200                 ceiling &= PGDIR_MASK;
201                 if (!ceiling)
202                         return;
203         }
204         if (end - 1 > ceiling - 1)
205                 return;
206
207         pud = pud_offset(pgd, start);
208         pgd_clear(pgd);
209         pud_free_tlb(tlb, pud);
210 }
211
212 /*
213  * This function frees user-level page tables of a process.
214  *
215  * Must be called with pagetable lock held.
216  */
217 void free_pgd_range(struct mmu_gather *tlb,
218                         unsigned long addr, unsigned long end,
219                         unsigned long floor, unsigned long ceiling)
220 {
221         pgd_t *pgd;
222         unsigned long next;
223         unsigned long start;
224
225         /*
226          * The next few lines have given us lots of grief...
227          *
228          * Why are we testing PMD* at this top level?  Because often
229          * there will be no work to do at all, and we'd prefer not to
230          * go all the way down to the bottom just to discover that.
231          *
232          * Why all these "- 1"s?  Because 0 represents both the bottom
233          * of the address space and the top of it (using -1 for the
234          * top wouldn't help much: the masks would do the wrong thing).
235          * The rule is that addr 0 and floor 0 refer to the bottom of
236          * the address space, but end 0 and ceiling 0 refer to the top
237          * Comparisons need to use "end - 1" and "ceiling - 1" (though
238          * that end 0 case should be mythical).
239          *
240          * Wherever addr is brought up or ceiling brought down, we must
241          * be careful to reject "the opposite 0" before it confuses the
242          * subsequent tests.  But what about where end is brought down
243          * by PMD_SIZE below? no, end can't go down to 0 there.
244          *
245          * Whereas we round start (addr) and ceiling down, by different
246          * masks at different levels, in order to test whether a table
247          * now has no other vmas using it, so can be freed, we don't
248          * bother to round floor or end up - the tests don't need that.
249          */
250
251         addr &= PMD_MASK;
252         if (addr < floor) {
253                 addr += PMD_SIZE;
254                 if (!addr)
255                         return;
256         }
257         if (ceiling) {
258                 ceiling &= PMD_MASK;
259                 if (!ceiling)
260                         return;
261         }
262         if (end - 1 > ceiling - 1)
263                 end -= PMD_SIZE;
264         if (addr > end - 1)
265                 return;
266
267         start = addr;
268         pgd = pgd_offset(tlb->mm, addr);
269         do {
270                 next = pgd_addr_end(addr, end);
271                 if (pgd_none_or_clear_bad(pgd))
272                         continue;
273                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
274         } while (pgd++, addr = next, addr != end);
275 }
276
277 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
278                 unsigned long floor, unsigned long ceiling)
279 {
280         while (vma) {
281                 struct vm_area_struct *next = vma->vm_next;
282                 unsigned long addr = vma->vm_start;
283
284                 /*
285                  * Hide vma from rmap and vmtruncate before freeing pgtables
286                  */
287                 anon_vma_unlink(vma);
288                 unlink_file_vma(vma);
289
290                 if (is_vm_hugetlb_page(vma)) {
291                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
292                                 floor, next? next->vm_start: ceiling);
293                 } else {
294                         /*
295                          * Optimization: gather nearby vmas into one call down
296                          */
297                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
298                                && !is_vm_hugetlb_page(next)) {
299                                 vma = next;
300                                 next = vma->vm_next;
301                                 anon_vma_unlink(vma);
302                                 unlink_file_vma(vma);
303                         }
304                         free_pgd_range(tlb, addr, vma->vm_end,
305                                 floor, next? next->vm_start: ceiling);
306                 }
307                 vma = next;
308         }
309 }
310
311 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
312 {
313         pgtable_t new = pte_alloc_one(mm, address);
314         if (!new)
315                 return -ENOMEM;
316
317         /*
318          * Ensure all pte setup (eg. pte page lock and page clearing) are
319          * visible before the pte is made visible to other CPUs by being
320          * put into page tables.
321          *
322          * The other side of the story is the pointer chasing in the page
323          * table walking code (when walking the page table without locking;
324          * ie. most of the time). Fortunately, these data accesses consist
325          * of a chain of data-dependent loads, meaning most CPUs (alpha
326          * being the notable exception) will already guarantee loads are
327          * seen in-order. See the alpha page table accessors for the
328          * smp_read_barrier_depends() barriers in page table walking code.
329          */
330         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
331
332         spin_lock(&mm->page_table_lock);
333         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
334                 mm->nr_ptes++;
335                 pmd_populate(mm, pmd, new);
336                 new = NULL;
337         }
338         spin_unlock(&mm->page_table_lock);
339         if (new)
340                 pte_free(mm, new);
341         return 0;
342 }
343
344 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
345 {
346         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
347         if (!new)
348                 return -ENOMEM;
349
350         smp_wmb(); /* See comment in __pte_alloc */
351
352         spin_lock(&init_mm.page_table_lock);
353         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
354                 pmd_populate_kernel(&init_mm, pmd, new);
355                 new = NULL;
356         }
357         spin_unlock(&init_mm.page_table_lock);
358         if (new)
359                 pte_free_kernel(&init_mm, new);
360         return 0;
361 }
362
363 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
364 {
365         if (file_rss)
366                 add_mm_counter(mm, file_rss, file_rss);
367         if (anon_rss)
368                 add_mm_counter(mm, anon_rss, anon_rss);
369 }
370
371 /*
372  * This function is called to print an error when a bad pte
373  * is found. For example, we might have a PFN-mapped pte in
374  * a region that doesn't allow it.
375  *
376  * The calling function must still handle the error.
377  */
378 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
379                           pte_t pte, struct page *page)
380 {
381         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
382         pud_t *pud = pud_offset(pgd, addr);
383         pmd_t *pmd = pmd_offset(pud, addr);
384         struct address_space *mapping;
385         pgoff_t index;
386
387         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
388         index = linear_page_index(vma, addr);
389
390         printk(KERN_EMERG "Bad page map in process %s  pte:%08llx pmd:%08llx\n",
391                 current->comm,
392                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
393         if (page) {
394                 printk(KERN_EMERG
395                 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
396                 page, (void *)page->flags, page_count(page),
397                 page_mapcount(page), page->mapping, page->index);
398         }
399         printk(KERN_EMERG
400                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
401                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
402         /*
403          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
404          */
405         if (vma->vm_ops)
406                 print_symbol(KERN_EMERG "vma->vm_ops->fault: %s\n",
407                                 (unsigned long)vma->vm_ops->fault);
408         if (vma->vm_file && vma->vm_file->f_op)
409                 print_symbol(KERN_EMERG "vma->vm_file->f_op->mmap: %s\n",
410                                 (unsigned long)vma->vm_file->f_op->mmap);
411         dump_stack();
412         add_taint(TAINT_BAD_PAGE);
413 }
414
415 static inline int is_cow_mapping(unsigned int flags)
416 {
417         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
418 }
419
420 /*
421  * vm_normal_page -- This function gets the "struct page" associated with a pte.
422  *
423  * "Special" mappings do not wish to be associated with a "struct page" (either
424  * it doesn't exist, or it exists but they don't want to touch it). In this
425  * case, NULL is returned here. "Normal" mappings do have a struct page.
426  *
427  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
428  * pte bit, in which case this function is trivial. Secondly, an architecture
429  * may not have a spare pte bit, which requires a more complicated scheme,
430  * described below.
431  *
432  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
433  * special mapping (even if there are underlying and valid "struct pages").
434  * COWed pages of a VM_PFNMAP are always normal.
435  *
436  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
437  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
438  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
439  * mapping will always honor the rule
440  *
441  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
442  *
443  * And for normal mappings this is false.
444  *
445  * This restricts such mappings to be a linear translation from virtual address
446  * to pfn. To get around this restriction, we allow arbitrary mappings so long
447  * as the vma is not a COW mapping; in that case, we know that all ptes are
448  * special (because none can have been COWed).
449  *
450  *
451  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
452  *
453  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
454  * page" backing, however the difference is that _all_ pages with a struct
455  * page (that is, those where pfn_valid is true) are refcounted and considered
456  * normal pages by the VM. The disadvantage is that pages are refcounted
457  * (which can be slower and simply not an option for some PFNMAP users). The
458  * advantage is that we don't have to follow the strict linearity rule of
459  * PFNMAP mappings in order to support COWable mappings.
460  *
461  */
462 #ifdef __HAVE_ARCH_PTE_SPECIAL
463 # define HAVE_PTE_SPECIAL 1
464 #else
465 # define HAVE_PTE_SPECIAL 0
466 #endif
467 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
468                                 pte_t pte)
469 {
470         unsigned long pfn = pte_pfn(pte);
471
472         if (HAVE_PTE_SPECIAL) {
473                 if (likely(!pte_special(pte)))
474                         goto check_pfn;
475                 if (!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)))
476                         print_bad_pte(vma, addr, pte, NULL);
477                 return NULL;
478         }
479
480         /* !HAVE_PTE_SPECIAL case follows: */
481
482         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
483                 if (vma->vm_flags & VM_MIXEDMAP) {
484                         if (!pfn_valid(pfn))
485                                 return NULL;
486                         goto out;
487                 } else {
488                         unsigned long off;
489                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
490                         if (pfn == vma->vm_pgoff + off)
491                                 return NULL;
492                         if (!is_cow_mapping(vma->vm_flags))
493                                 return NULL;
494                 }
495         }
496
497 check_pfn:
498         if (unlikely(pfn > highest_memmap_pfn)) {
499                 print_bad_pte(vma, addr, pte, NULL);
500                 return NULL;
501         }
502
503         /*
504          * NOTE! We still have PageReserved() pages in the page tables.
505          * eg. VDSO mappings can cause them to exist.
506          */
507 out:
508         return pfn_to_page(pfn);
509 }
510
511 /*
512  * copy one vm_area from one task to the other. Assumes the page tables
513  * already present in the new task to be cleared in the whole range
514  * covered by this vma.
515  */
516
517 static inline void
518 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
519                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
520                 unsigned long addr, int *rss)
521 {
522         unsigned long vm_flags = vma->vm_flags;
523         pte_t pte = *src_pte;
524         struct page *page;
525
526         /* pte contains position in swap or file, so copy. */
527         if (unlikely(!pte_present(pte))) {
528                 if (!pte_file(pte)) {
529                         swp_entry_t entry = pte_to_swp_entry(pte);
530
531                         swap_duplicate(entry);
532                         /* make sure dst_mm is on swapoff's mmlist. */
533                         if (unlikely(list_empty(&dst_mm->mmlist))) {
534                                 spin_lock(&mmlist_lock);
535                                 if (list_empty(&dst_mm->mmlist))
536                                         list_add(&dst_mm->mmlist,
537                                                  &src_mm->mmlist);
538                                 spin_unlock(&mmlist_lock);
539                         }
540                         if (is_write_migration_entry(entry) &&
541                                         is_cow_mapping(vm_flags)) {
542                                 /*
543                                  * COW mappings require pages in both parent
544                                  * and child to be set to read.
545                                  */
546                                 make_migration_entry_read(&entry);
547                                 pte = swp_entry_to_pte(entry);
548                                 set_pte_at(src_mm, addr, src_pte, pte);
549                         }
550                 }
551                 goto out_set_pte;
552         }
553
554         /*
555          * If it's a COW mapping, write protect it both
556          * in the parent and the child
557          */
558         if (is_cow_mapping(vm_flags)) {
559                 ptep_set_wrprotect(src_mm, addr, src_pte);
560                 pte = pte_wrprotect(pte);
561         }
562
563         /*
564          * If it's a shared mapping, mark it clean in
565          * the child
566          */
567         if (vm_flags & VM_SHARED)
568                 pte = pte_mkclean(pte);
569         pte = pte_mkold(pte);
570
571         page = vm_normal_page(vma, addr, pte);
572         if (page) {
573                 get_page(page);
574                 page_dup_rmap(page, vma, addr);
575                 rss[!!PageAnon(page)]++;
576         }
577
578 out_set_pte:
579         set_pte_at(dst_mm, addr, dst_pte, pte);
580 }
581
582 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
583                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
584                 unsigned long addr, unsigned long end)
585 {
586         pte_t *src_pte, *dst_pte;
587         spinlock_t *src_ptl, *dst_ptl;
588         int progress = 0;
589         int rss[2];
590
591 again:
592         rss[1] = rss[0] = 0;
593         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
594         if (!dst_pte)
595                 return -ENOMEM;
596         src_pte = pte_offset_map_nested(src_pmd, addr);
597         src_ptl = pte_lockptr(src_mm, src_pmd);
598         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
599         arch_enter_lazy_mmu_mode();
600
601         do {
602                 /*
603                  * We are holding two locks at this point - either of them
604                  * could generate latencies in another task on another CPU.
605                  */
606                 if (progress >= 32) {
607                         progress = 0;
608                         if (need_resched() ||
609                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
610                                 break;
611                 }
612                 if (pte_none(*src_pte)) {
613                         progress++;
614                         continue;
615                 }
616                 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
617                 progress += 8;
618         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
619
620         arch_leave_lazy_mmu_mode();
621         spin_unlock(src_ptl);
622         pte_unmap_nested(src_pte - 1);
623         add_mm_rss(dst_mm, rss[0], rss[1]);
624         pte_unmap_unlock(dst_pte - 1, dst_ptl);
625         cond_resched();
626         if (addr != end)
627                 goto again;
628         return 0;
629 }
630
631 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
632                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
633                 unsigned long addr, unsigned long end)
634 {
635         pmd_t *src_pmd, *dst_pmd;
636         unsigned long next;
637
638         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
639         if (!dst_pmd)
640                 return -ENOMEM;
641         src_pmd = pmd_offset(src_pud, addr);
642         do {
643                 next = pmd_addr_end(addr, end);
644                 if (pmd_none_or_clear_bad(src_pmd))
645                         continue;
646                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
647                                                 vma, addr, next))
648                         return -ENOMEM;
649         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
650         return 0;
651 }
652
653 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
654                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
655                 unsigned long addr, unsigned long end)
656 {
657         pud_t *src_pud, *dst_pud;
658         unsigned long next;
659
660         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
661         if (!dst_pud)
662                 return -ENOMEM;
663         src_pud = pud_offset(src_pgd, addr);
664         do {
665                 next = pud_addr_end(addr, end);
666                 if (pud_none_or_clear_bad(src_pud))
667                         continue;
668                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
669                                                 vma, addr, next))
670                         return -ENOMEM;
671         } while (dst_pud++, src_pud++, addr = next, addr != end);
672         return 0;
673 }
674
675 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
676                 struct vm_area_struct *vma)
677 {
678         pgd_t *src_pgd, *dst_pgd;
679         unsigned long next;
680         unsigned long addr = vma->vm_start;
681         unsigned long end = vma->vm_end;
682         int ret;
683
684         /*
685          * Don't copy ptes where a page fault will fill them correctly.
686          * Fork becomes much lighter when there are big shared or private
687          * readonly mappings. The tradeoff is that copy_page_range is more
688          * efficient than faulting.
689          */
690         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
691                 if (!vma->anon_vma)
692                         return 0;
693         }
694
695         if (is_vm_hugetlb_page(vma))
696                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
697
698         if (unlikely(is_pfn_mapping(vma))) {
699                 /*
700                  * We do not free on error cases below as remove_vma
701                  * gets called on error from higher level routine
702                  */
703                 ret = track_pfn_vma_copy(vma);
704                 if (ret)
705                         return ret;
706         }
707
708         /*
709          * We need to invalidate the secondary MMU mappings only when
710          * there could be a permission downgrade on the ptes of the
711          * parent mm. And a permission downgrade will only happen if
712          * is_cow_mapping() returns true.
713          */
714         if (is_cow_mapping(vma->vm_flags))
715                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
716
717         ret = 0;
718         dst_pgd = pgd_offset(dst_mm, addr);
719         src_pgd = pgd_offset(src_mm, addr);
720         do {
721                 next = pgd_addr_end(addr, end);
722                 if (pgd_none_or_clear_bad(src_pgd))
723                         continue;
724                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
725                                             vma, addr, next))) {
726                         ret = -ENOMEM;
727                         break;
728                 }
729         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
730
731         if (is_cow_mapping(vma->vm_flags))
732                 mmu_notifier_invalidate_range_end(src_mm,
733                                                   vma->vm_start, end);
734         return ret;
735 }
736
737 static unsigned long zap_pte_range(struct mmu_gather *tlb,
738                                 struct vm_area_struct *vma, pmd_t *pmd,
739                                 unsigned long addr, unsigned long end,
740                                 long *zap_work, struct zap_details *details)
741 {
742         struct mm_struct *mm = tlb->mm;
743         pte_t *pte;
744         spinlock_t *ptl;
745         int file_rss = 0;
746         int anon_rss = 0;
747
748         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
749         arch_enter_lazy_mmu_mode();
750         do {
751                 pte_t ptent = *pte;
752                 if (pte_none(ptent)) {
753                         (*zap_work)--;
754                         continue;
755                 }
756
757                 (*zap_work) -= PAGE_SIZE;
758
759                 if (pte_present(ptent)) {
760                         struct page *page;
761
762                         page = vm_normal_page(vma, addr, ptent);
763                         if (unlikely(details) && page) {
764                                 /*
765                                  * unmap_shared_mapping_pages() wants to
766                                  * invalidate cache without truncating:
767                                  * unmap shared but keep private pages.
768                                  */
769                                 if (details->check_mapping &&
770                                     details->check_mapping != page->mapping)
771                                         continue;
772                                 /*
773                                  * Each page->index must be checked when
774                                  * invalidating or truncating nonlinear.
775                                  */
776                                 if (details->nonlinear_vma &&
777                                     (page->index < details->first_index ||
778                                      page->index > details->last_index))
779                                         continue;
780                         }
781                         ptent = ptep_get_and_clear_full(mm, addr, pte,
782                                                         tlb->fullmm);
783                         tlb_remove_tlb_entry(tlb, pte, addr);
784                         if (unlikely(!page))
785                                 continue;
786                         if (unlikely(details) && details->nonlinear_vma
787                             && linear_page_index(details->nonlinear_vma,
788                                                 addr) != page->index)
789                                 set_pte_at(mm, addr, pte,
790                                            pgoff_to_pte(page->index));
791                         if (PageAnon(page))
792                                 anon_rss--;
793                         else {
794                                 if (pte_dirty(ptent))
795                                         set_page_dirty(page);
796                                 if (pte_young(ptent) &&
797                                     likely(!VM_SequentialReadHint(vma)))
798                                         mark_page_accessed(page);
799                                 file_rss--;
800                         }
801                         page_remove_rmap(page);
802                         if (unlikely(page_mapcount(page) < 0))
803                                 print_bad_pte(vma, addr, ptent, page);
804                         tlb_remove_page(tlb, page);
805                         continue;
806                 }
807                 /*
808                  * If details->check_mapping, we leave swap entries;
809                  * if details->nonlinear_vma, we leave file entries.
810                  */
811                 if (unlikely(details))
812                         continue;
813                 if (pte_file(ptent)) {
814                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
815                                 print_bad_pte(vma, addr, ptent, NULL);
816                 } else if
817                   (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
818                         print_bad_pte(vma, addr, ptent, NULL);
819                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
820         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
821
822         add_mm_rss(mm, file_rss, anon_rss);
823         arch_leave_lazy_mmu_mode();
824         pte_unmap_unlock(pte - 1, ptl);
825
826         return addr;
827 }
828
829 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
830                                 struct vm_area_struct *vma, pud_t *pud,
831                                 unsigned long addr, unsigned long end,
832                                 long *zap_work, struct zap_details *details)
833 {
834         pmd_t *pmd;
835         unsigned long next;
836
837         pmd = pmd_offset(pud, addr);
838         do {
839                 next = pmd_addr_end(addr, end);
840                 if (pmd_none_or_clear_bad(pmd)) {
841                         (*zap_work)--;
842                         continue;
843                 }
844                 next = zap_pte_range(tlb, vma, pmd, addr, next,
845                                                 zap_work, details);
846         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
847
848         return addr;
849 }
850
851 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
852                                 struct vm_area_struct *vma, pgd_t *pgd,
853                                 unsigned long addr, unsigned long end,
854                                 long *zap_work, struct zap_details *details)
855 {
856         pud_t *pud;
857         unsigned long next;
858
859         pud = pud_offset(pgd, addr);
860         do {
861                 next = pud_addr_end(addr, end);
862                 if (pud_none_or_clear_bad(pud)) {
863                         (*zap_work)--;
864                         continue;
865                 }
866                 next = zap_pmd_range(tlb, vma, pud, addr, next,
867                                                 zap_work, details);
868         } while (pud++, addr = next, (addr != end && *zap_work > 0));
869
870         return addr;
871 }
872
873 static unsigned long unmap_page_range(struct mmu_gather *tlb,
874                                 struct vm_area_struct *vma,
875                                 unsigned long addr, unsigned long end,
876                                 long *zap_work, struct zap_details *details)
877 {
878         pgd_t *pgd;
879         unsigned long next;
880
881         if (details && !details->check_mapping && !details->nonlinear_vma)
882                 details = NULL;
883
884         BUG_ON(addr >= end);
885         tlb_start_vma(tlb, vma);
886         pgd = pgd_offset(vma->vm_mm, addr);
887         do {
888                 next = pgd_addr_end(addr, end);
889                 if (pgd_none_or_clear_bad(pgd)) {
890                         (*zap_work)--;
891                         continue;
892                 }
893                 next = zap_pud_range(tlb, vma, pgd, addr, next,
894                                                 zap_work, details);
895         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
896         tlb_end_vma(tlb, vma);
897
898         return addr;
899 }
900
901 #ifdef CONFIG_PREEMPT
902 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
903 #else
904 /* No preempt: go for improved straight-line efficiency */
905 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
906 #endif
907
908 /**
909  * unmap_vmas - unmap a range of memory covered by a list of vma's
910  * @tlbp: address of the caller's struct mmu_gather
911  * @vma: the starting vma
912  * @start_addr: virtual address at which to start unmapping
913  * @end_addr: virtual address at which to end unmapping
914  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
915  * @details: details of nonlinear truncation or shared cache invalidation
916  *
917  * Returns the end address of the unmapping (restart addr if interrupted).
918  *
919  * Unmap all pages in the vma list.
920  *
921  * We aim to not hold locks for too long (for scheduling latency reasons).
922  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
923  * return the ending mmu_gather to the caller.
924  *
925  * Only addresses between `start' and `end' will be unmapped.
926  *
927  * The VMA list must be sorted in ascending virtual address order.
928  *
929  * unmap_vmas() assumes that the caller will flush the whole unmapped address
930  * range after unmap_vmas() returns.  So the only responsibility here is to
931  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
932  * drops the lock and schedules.
933  */
934 unsigned long unmap_vmas(struct mmu_gather **tlbp,
935                 struct vm_area_struct *vma, unsigned long start_addr,
936                 unsigned long end_addr, unsigned long *nr_accounted,
937                 struct zap_details *details)
938 {
939         long zap_work = ZAP_BLOCK_SIZE;
940         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
941         int tlb_start_valid = 0;
942         unsigned long start = start_addr;
943         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
944         int fullmm = (*tlbp)->fullmm;
945         struct mm_struct *mm = vma->vm_mm;
946
947         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
948         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
949                 unsigned long end;
950
951                 start = max(vma->vm_start, start_addr);
952                 if (start >= vma->vm_end)
953                         continue;
954                 end = min(vma->vm_end, end_addr);
955                 if (end <= vma->vm_start)
956                         continue;
957
958                 if (vma->vm_flags & VM_ACCOUNT)
959                         *nr_accounted += (end - start) >> PAGE_SHIFT;
960
961                 if (unlikely(is_pfn_mapping(vma)))
962                         untrack_pfn_vma(vma, 0, 0);
963
964                 while (start != end) {
965                         if (!tlb_start_valid) {
966                                 tlb_start = start;
967                                 tlb_start_valid = 1;
968                         }
969
970                         if (unlikely(is_vm_hugetlb_page(vma))) {
971                                 /*
972                                  * It is undesirable to test vma->vm_file as it
973                                  * should be non-null for valid hugetlb area.
974                                  * However, vm_file will be NULL in the error
975                                  * cleanup path of do_mmap_pgoff. When
976                                  * hugetlbfs ->mmap method fails,
977                                  * do_mmap_pgoff() nullifies vma->vm_file
978                                  * before calling this function to clean up.
979                                  * Since no pte has actually been setup, it is
980                                  * safe to do nothing in this case.
981                                  */
982                                 if (vma->vm_file) {
983                                         unmap_hugepage_range(vma, start, end, NULL);
984                                         zap_work -= (end - start) /
985                                         pages_per_huge_page(hstate_vma(vma));
986                                 }
987
988                                 start = end;
989                         } else
990                                 start = unmap_page_range(*tlbp, vma,
991                                                 start, end, &zap_work, details);
992
993                         if (zap_work > 0) {
994                                 BUG_ON(start != end);
995                                 break;
996                         }
997
998                         tlb_finish_mmu(*tlbp, tlb_start, start);
999
1000                         if (need_resched() ||
1001                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1002                                 if (i_mmap_lock) {
1003                                         *tlbp = NULL;
1004                                         goto out;
1005                                 }
1006                                 cond_resched();
1007                         }
1008
1009                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1010                         tlb_start_valid = 0;
1011                         zap_work = ZAP_BLOCK_SIZE;
1012                 }
1013         }
1014 out:
1015         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1016         return start;   /* which is now the end (or restart) address */
1017 }
1018
1019 /**
1020  * zap_page_range - remove user pages in a given range
1021  * @vma: vm_area_struct holding the applicable pages
1022  * @address: starting address of pages to zap
1023  * @size: number of bytes to zap
1024  * @details: details of nonlinear truncation or shared cache invalidation
1025  */
1026 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1027                 unsigned long size, struct zap_details *details)
1028 {
1029         struct mm_struct *mm = vma->vm_mm;
1030         struct mmu_gather *tlb;
1031         unsigned long end = address + size;
1032         unsigned long nr_accounted = 0;
1033
1034         lru_add_drain();
1035         tlb = tlb_gather_mmu(mm, 0);
1036         update_hiwater_rss(mm);
1037         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1038         if (tlb)
1039                 tlb_finish_mmu(tlb, address, end);
1040         return end;
1041 }
1042
1043 /**
1044  * zap_vma_ptes - remove ptes mapping the vma
1045  * @vma: vm_area_struct holding ptes to be zapped
1046  * @address: starting address of pages to zap
1047  * @size: number of bytes to zap
1048  *
1049  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1050  *
1051  * The entire address range must be fully contained within the vma.
1052  *
1053  * Returns 0 if successful.
1054  */
1055 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1056                 unsigned long size)
1057 {
1058         if (address < vma->vm_start || address + size > vma->vm_end ||
1059                         !(vma->vm_flags & VM_PFNMAP))
1060                 return -1;
1061         zap_page_range(vma, address, size, NULL);
1062         return 0;
1063 }
1064 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1065
1066 /*
1067  * Do a quick page-table lookup for a single page.
1068  */
1069 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1070                         unsigned int flags)
1071 {
1072         pgd_t *pgd;
1073         pud_t *pud;
1074         pmd_t *pmd;
1075         pte_t *ptep, pte;
1076         spinlock_t *ptl;
1077         struct page *page;
1078         struct mm_struct *mm = vma->vm_mm;
1079
1080         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1081         if (!IS_ERR(page)) {
1082                 BUG_ON(flags & FOLL_GET);
1083                 goto out;
1084         }
1085
1086         page = NULL;
1087         pgd = pgd_offset(mm, address);
1088         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1089                 goto no_page_table;
1090
1091         pud = pud_offset(pgd, address);
1092         if (pud_none(*pud))
1093                 goto no_page_table;
1094         if (pud_huge(*pud)) {
1095                 BUG_ON(flags & FOLL_GET);
1096                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1097                 goto out;
1098         }
1099         if (unlikely(pud_bad(*pud)))
1100                 goto no_page_table;
1101
1102         pmd = pmd_offset(pud, address);
1103         if (pmd_none(*pmd))
1104                 goto no_page_table;
1105         if (pmd_huge(*pmd)) {
1106                 BUG_ON(flags & FOLL_GET);
1107                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1108                 goto out;
1109         }
1110         if (unlikely(pmd_bad(*pmd)))
1111                 goto no_page_table;
1112
1113         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1114
1115         pte = *ptep;
1116         if (!pte_present(pte))
1117                 goto no_page;
1118         if ((flags & FOLL_WRITE) && !pte_write(pte))
1119                 goto unlock;
1120         page = vm_normal_page(vma, address, pte);
1121         if (unlikely(!page))
1122                 goto bad_page;
1123
1124         if (flags & FOLL_GET)
1125                 get_page(page);
1126         if (flags & FOLL_TOUCH) {
1127                 if ((flags & FOLL_WRITE) &&
1128                     !pte_dirty(pte) && !PageDirty(page))
1129                         set_page_dirty(page);
1130                 mark_page_accessed(page);
1131         }
1132 unlock:
1133         pte_unmap_unlock(ptep, ptl);
1134 out:
1135         return page;
1136
1137 bad_page:
1138         pte_unmap_unlock(ptep, ptl);
1139         return ERR_PTR(-EFAULT);
1140
1141 no_page:
1142         pte_unmap_unlock(ptep, ptl);
1143         if (!pte_none(pte))
1144                 return page;
1145         /* Fall through to ZERO_PAGE handling */
1146 no_page_table:
1147         /*
1148          * When core dumping an enormous anonymous area that nobody
1149          * has touched so far, we don't want to allocate page tables.
1150          */
1151         if (flags & FOLL_ANON) {
1152                 page = ZERO_PAGE(0);
1153                 if (flags & FOLL_GET)
1154                         get_page(page);
1155                 BUG_ON(flags & FOLL_WRITE);
1156         }
1157         return page;
1158 }
1159
1160 /* Can we do the FOLL_ANON optimization? */
1161 static inline int use_zero_page(struct vm_area_struct *vma)
1162 {
1163         /*
1164          * We don't want to optimize FOLL_ANON for make_pages_present()
1165          * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1166          * we want to get the page from the page tables to make sure
1167          * that we serialize and update with any other user of that
1168          * mapping.
1169          */
1170         if (vma->vm_flags & (VM_LOCKED | VM_SHARED))
1171                 return 0;
1172         /*
1173          * And if we have a fault routine, it's not an anonymous region.
1174          */
1175         return !vma->vm_ops || !vma->vm_ops->fault;
1176 }
1177
1178
1179
1180 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1181                      unsigned long start, int len, int flags,
1182                 struct page **pages, struct vm_area_struct **vmas)
1183 {
1184         int i;
1185         unsigned int vm_flags = 0;
1186         int write = !!(flags & GUP_FLAGS_WRITE);
1187         int force = !!(flags & GUP_FLAGS_FORCE);
1188         int ignore = !!(flags & GUP_FLAGS_IGNORE_VMA_PERMISSIONS);
1189
1190         if (len <= 0)
1191                 return 0;
1192         /* 
1193          * Require read or write permissions.
1194          * If 'force' is set, we only require the "MAY" flags.
1195          */
1196         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1197         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1198         i = 0;
1199
1200         do {
1201                 struct vm_area_struct *vma;
1202                 unsigned int foll_flags;
1203
1204                 vma = find_extend_vma(mm, start);
1205                 if (!vma && in_gate_area(tsk, start)) {
1206                         unsigned long pg = start & PAGE_MASK;
1207                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1208                         pgd_t *pgd;
1209                         pud_t *pud;
1210                         pmd_t *pmd;
1211                         pte_t *pte;
1212
1213                         /* user gate pages are read-only */
1214                         if (!ignore && write)
1215                                 return i ? : -EFAULT;
1216                         if (pg > TASK_SIZE)
1217                                 pgd = pgd_offset_k(pg);
1218                         else
1219                                 pgd = pgd_offset_gate(mm, pg);
1220                         BUG_ON(pgd_none(*pgd));
1221                         pud = pud_offset(pgd, pg);
1222                         BUG_ON(pud_none(*pud));
1223                         pmd = pmd_offset(pud, pg);
1224                         if (pmd_none(*pmd))
1225                                 return i ? : -EFAULT;
1226                         pte = pte_offset_map(pmd, pg);
1227                         if (pte_none(*pte)) {
1228                                 pte_unmap(pte);
1229                                 return i ? : -EFAULT;
1230                         }
1231                         if (pages) {
1232                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1233                                 pages[i] = page;
1234                                 if (page)
1235                                         get_page(page);
1236                         }
1237                         pte_unmap(pte);
1238                         if (vmas)
1239                                 vmas[i] = gate_vma;
1240                         i++;
1241                         start += PAGE_SIZE;
1242                         len--;
1243                         continue;
1244                 }
1245
1246                 if (!vma ||
1247                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1248                     (!ignore && !(vm_flags & vma->vm_flags)))
1249                         return i ? : -EFAULT;
1250
1251                 if (is_vm_hugetlb_page(vma)) {
1252                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1253                                                 &start, &len, i, write);
1254                         continue;
1255                 }
1256
1257                 foll_flags = FOLL_TOUCH;
1258                 if (pages)
1259                         foll_flags |= FOLL_GET;
1260                 if (!write && use_zero_page(vma))
1261                         foll_flags |= FOLL_ANON;
1262
1263                 do {
1264                         struct page *page;
1265
1266                         /*
1267                          * If tsk is ooming, cut off its access to large memory
1268                          * allocations. It has a pending SIGKILL, but it can't
1269                          * be processed until returning to user space.
1270                          */
1271                         if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1272                                 return i ? i : -ENOMEM;
1273
1274                         if (write)
1275                                 foll_flags |= FOLL_WRITE;
1276
1277                         cond_resched();
1278                         while (!(page = follow_page(vma, start, foll_flags))) {
1279                                 int ret;
1280                                 ret = handle_mm_fault(mm, vma, start,
1281                                                 foll_flags & FOLL_WRITE);
1282                                 if (ret & VM_FAULT_ERROR) {
1283                                         if (ret & VM_FAULT_OOM)
1284                                                 return i ? i : -ENOMEM;
1285                                         else if (ret & VM_FAULT_SIGBUS)
1286                                                 return i ? i : -EFAULT;
1287                                         BUG();
1288                                 }
1289                                 if (ret & VM_FAULT_MAJOR)
1290                                         tsk->maj_flt++;
1291                                 else
1292                                         tsk->min_flt++;
1293
1294                                 /*
1295                                  * The VM_FAULT_WRITE bit tells us that
1296                                  * do_wp_page has broken COW when necessary,
1297                                  * even if maybe_mkwrite decided not to set
1298                                  * pte_write. We can thus safely do subsequent
1299                                  * page lookups as if they were reads. But only
1300                                  * do so when looping for pte_write is futile:
1301                                  * in some cases userspace may also be wanting
1302                                  * to write to the gotten user page, which a
1303                                  * read fault here might prevent (a readonly
1304                                  * page might get reCOWed by userspace write).
1305                                  */
1306                                 if ((ret & VM_FAULT_WRITE) &&
1307                                     !(vma->vm_flags & VM_WRITE))
1308                                         foll_flags &= ~FOLL_WRITE;
1309
1310                                 cond_resched();
1311                         }
1312                         if (IS_ERR(page))
1313                                 return i ? i : PTR_ERR(page);
1314                         if (pages) {
1315                                 pages[i] = page;
1316
1317                                 flush_anon_page(vma, page, start);
1318                                 flush_dcache_page(page);
1319                         }
1320                         if (vmas)
1321                                 vmas[i] = vma;
1322                         i++;
1323                         start += PAGE_SIZE;
1324                         len--;
1325                 } while (len && start < vma->vm_end);
1326         } while (len);
1327         return i;
1328 }
1329
1330 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1331                 unsigned long start, int len, int write, int force,
1332                 struct page **pages, struct vm_area_struct **vmas)
1333 {
1334         int flags = 0;
1335
1336         if (write)
1337                 flags |= GUP_FLAGS_WRITE;
1338         if (force)
1339                 flags |= GUP_FLAGS_FORCE;
1340
1341         return __get_user_pages(tsk, mm,
1342                                 start, len, flags,
1343                                 pages, vmas);
1344 }
1345
1346 EXPORT_SYMBOL(get_user_pages);
1347
1348 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1349                         spinlock_t **ptl)
1350 {
1351         pgd_t * pgd = pgd_offset(mm, addr);
1352         pud_t * pud = pud_alloc(mm, pgd, addr);
1353         if (pud) {
1354                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1355                 if (pmd)
1356                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1357         }
1358         return NULL;
1359 }
1360
1361 /*
1362  * This is the old fallback for page remapping.
1363  *
1364  * For historical reasons, it only allows reserved pages. Only
1365  * old drivers should use this, and they needed to mark their
1366  * pages reserved for the old functions anyway.
1367  */
1368 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1369                         struct page *page, pgprot_t prot)
1370 {
1371         struct mm_struct *mm = vma->vm_mm;
1372         int retval;
1373         pte_t *pte;
1374         spinlock_t *ptl;
1375
1376         retval = -EINVAL;
1377         if (PageAnon(page))
1378                 goto out;
1379         retval = -ENOMEM;
1380         flush_dcache_page(page);
1381         pte = get_locked_pte(mm, addr, &ptl);
1382         if (!pte)
1383                 goto out;
1384         retval = -EBUSY;
1385         if (!pte_none(*pte))
1386                 goto out_unlock;
1387
1388         /* Ok, finally just insert the thing.. */
1389         get_page(page);
1390         inc_mm_counter(mm, file_rss);
1391         page_add_file_rmap(page);
1392         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1393
1394         retval = 0;
1395         pte_unmap_unlock(pte, ptl);
1396         return retval;
1397 out_unlock:
1398         pte_unmap_unlock(pte, ptl);
1399 out:
1400         return retval;
1401 }
1402
1403 /**
1404  * vm_insert_page - insert single page into user vma
1405  * @vma: user vma to map to
1406  * @addr: target user address of this page
1407  * @page: source kernel page
1408  *
1409  * This allows drivers to insert individual pages they've allocated
1410  * into a user vma.
1411  *
1412  * The page has to be a nice clean _individual_ kernel allocation.
1413  * If you allocate a compound page, you need to have marked it as
1414  * such (__GFP_COMP), or manually just split the page up yourself
1415  * (see split_page()).
1416  *
1417  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1418  * took an arbitrary page protection parameter. This doesn't allow
1419  * that. Your vma protection will have to be set up correctly, which
1420  * means that if you want a shared writable mapping, you'd better
1421  * ask for a shared writable mapping!
1422  *
1423  * The page does not need to be reserved.
1424  */
1425 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1426                         struct page *page)
1427 {
1428         if (addr < vma->vm_start || addr >= vma->vm_end)
1429                 return -EFAULT;
1430         if (!page_count(page))
1431                 return -EINVAL;
1432         vma->vm_flags |= VM_INSERTPAGE;
1433         return insert_page(vma, addr, page, vma->vm_page_prot);
1434 }
1435 EXPORT_SYMBOL(vm_insert_page);
1436
1437 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1438                         unsigned long pfn, pgprot_t prot)
1439 {
1440         struct mm_struct *mm = vma->vm_mm;
1441         int retval;
1442         pte_t *pte, entry;
1443         spinlock_t *ptl;
1444
1445         retval = -ENOMEM;
1446         pte = get_locked_pte(mm, addr, &ptl);
1447         if (!pte)
1448                 goto out;
1449         retval = -EBUSY;
1450         if (!pte_none(*pte))
1451                 goto out_unlock;
1452
1453         /* Ok, finally just insert the thing.. */
1454         entry = pte_mkspecial(pfn_pte(pfn, prot));
1455         set_pte_at(mm, addr, pte, entry);
1456         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1457
1458         retval = 0;
1459 out_unlock:
1460         pte_unmap_unlock(pte, ptl);
1461 out:
1462         return retval;
1463 }
1464
1465 /**
1466  * vm_insert_pfn - insert single pfn into user vma
1467  * @vma: user vma to map to
1468  * @addr: target user address of this page
1469  * @pfn: source kernel pfn
1470  *
1471  * Similar to vm_inert_page, this allows drivers to insert individual pages
1472  * they've allocated into a user vma. Same comments apply.
1473  *
1474  * This function should only be called from a vm_ops->fault handler, and
1475  * in that case the handler should return NULL.
1476  *
1477  * vma cannot be a COW mapping.
1478  *
1479  * As this is called only for pages that do not currently exist, we
1480  * do not need to flush old virtual caches or the TLB.
1481  */
1482 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1483                         unsigned long pfn)
1484 {
1485         int ret;
1486         /*
1487          * Technically, architectures with pte_special can avoid all these
1488          * restrictions (same for remap_pfn_range).  However we would like
1489          * consistency in testing and feature parity among all, so we should
1490          * try to keep these invariants in place for everybody.
1491          */
1492         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1493         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1494                                                 (VM_PFNMAP|VM_MIXEDMAP));
1495         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1496         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1497
1498         if (addr < vma->vm_start || addr >= vma->vm_end)
1499                 return -EFAULT;
1500         if (track_pfn_vma_new(vma, vma->vm_page_prot, pfn, PAGE_SIZE))
1501                 return -EINVAL;
1502
1503         ret = insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1504
1505         if (ret)
1506                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1507
1508         return ret;
1509 }
1510 EXPORT_SYMBOL(vm_insert_pfn);
1511
1512 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1513                         unsigned long pfn)
1514 {
1515         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1516
1517         if (addr < vma->vm_start || addr >= vma->vm_end)
1518                 return -EFAULT;
1519
1520         /*
1521          * If we don't have pte special, then we have to use the pfn_valid()
1522          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1523          * refcount the page if pfn_valid is true (hence insert_page rather
1524          * than insert_pfn).
1525          */
1526         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1527                 struct page *page;
1528
1529                 page = pfn_to_page(pfn);
1530                 return insert_page(vma, addr, page, vma->vm_page_prot);
1531         }
1532         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1533 }
1534 EXPORT_SYMBOL(vm_insert_mixed);
1535
1536 /*
1537  * maps a range of physical memory into the requested pages. the old
1538  * mappings are removed. any references to nonexistent pages results
1539  * in null mappings (currently treated as "copy-on-access")
1540  */
1541 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1542                         unsigned long addr, unsigned long end,
1543                         unsigned long pfn, pgprot_t prot)
1544 {
1545         pte_t *pte;
1546         spinlock_t *ptl;
1547
1548         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1549         if (!pte)
1550                 return -ENOMEM;
1551         arch_enter_lazy_mmu_mode();
1552         do {
1553                 BUG_ON(!pte_none(*pte));
1554                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1555                 pfn++;
1556         } while (pte++, addr += PAGE_SIZE, addr != end);
1557         arch_leave_lazy_mmu_mode();
1558         pte_unmap_unlock(pte - 1, ptl);
1559         return 0;
1560 }
1561
1562 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1563                         unsigned long addr, unsigned long end,
1564                         unsigned long pfn, pgprot_t prot)
1565 {
1566         pmd_t *pmd;
1567         unsigned long next;
1568
1569         pfn -= addr >> PAGE_SHIFT;
1570         pmd = pmd_alloc(mm, pud, addr);
1571         if (!pmd)
1572                 return -ENOMEM;
1573         do {
1574                 next = pmd_addr_end(addr, end);
1575                 if (remap_pte_range(mm, pmd, addr, next,
1576                                 pfn + (addr >> PAGE_SHIFT), prot))
1577                         return -ENOMEM;
1578         } while (pmd++, addr = next, addr != end);
1579         return 0;
1580 }
1581
1582 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1583                         unsigned long addr, unsigned long end,
1584                         unsigned long pfn, pgprot_t prot)
1585 {
1586         pud_t *pud;
1587         unsigned long next;
1588
1589         pfn -= addr >> PAGE_SHIFT;
1590         pud = pud_alloc(mm, pgd, addr);
1591         if (!pud)
1592                 return -ENOMEM;
1593         do {
1594                 next = pud_addr_end(addr, end);
1595                 if (remap_pmd_range(mm, pud, addr, next,
1596                                 pfn + (addr >> PAGE_SHIFT), prot))
1597                         return -ENOMEM;
1598         } while (pud++, addr = next, addr != end);
1599         return 0;
1600 }
1601
1602 /**
1603  * remap_pfn_range - remap kernel memory to userspace
1604  * @vma: user vma to map to
1605  * @addr: target user address to start at
1606  * @pfn: physical address of kernel memory
1607  * @size: size of map area
1608  * @prot: page protection flags for this mapping
1609  *
1610  *  Note: this is only safe if the mm semaphore is held when called.
1611  */
1612 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1613                     unsigned long pfn, unsigned long size, pgprot_t prot)
1614 {
1615         pgd_t *pgd;
1616         unsigned long next;
1617         unsigned long end = addr + PAGE_ALIGN(size);
1618         struct mm_struct *mm = vma->vm_mm;
1619         int err;
1620
1621         /*
1622          * Physically remapped pages are special. Tell the
1623          * rest of the world about it:
1624          *   VM_IO tells people not to look at these pages
1625          *      (accesses can have side effects).
1626          *   VM_RESERVED is specified all over the place, because
1627          *      in 2.4 it kept swapout's vma scan off this vma; but
1628          *      in 2.6 the LRU scan won't even find its pages, so this
1629          *      flag means no more than count its pages in reserved_vm,
1630          *      and omit it from core dump, even when VM_IO turned off.
1631          *   VM_PFNMAP tells the core MM that the base pages are just
1632          *      raw PFN mappings, and do not have a "struct page" associated
1633          *      with them.
1634          *
1635          * There's a horrible special case to handle copy-on-write
1636          * behaviour that some programs depend on. We mark the "original"
1637          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1638          */
1639         if (addr == vma->vm_start && end == vma->vm_end)
1640                 vma->vm_pgoff = pfn;
1641         else if (is_cow_mapping(vma->vm_flags))
1642                 return -EINVAL;
1643
1644         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1645
1646         err = track_pfn_vma_new(vma, prot, pfn, PAGE_ALIGN(size));
1647         if (err)
1648                 return -EINVAL;
1649
1650         BUG_ON(addr >= end);
1651         pfn -= addr >> PAGE_SHIFT;
1652         pgd = pgd_offset(mm, addr);
1653         flush_cache_range(vma, addr, end);
1654         do {
1655                 next = pgd_addr_end(addr, end);
1656                 err = remap_pud_range(mm, pgd, addr, next,
1657                                 pfn + (addr >> PAGE_SHIFT), prot);
1658                 if (err)
1659                         break;
1660         } while (pgd++, addr = next, addr != end);
1661
1662         if (err)
1663                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1664
1665         return err;
1666 }
1667 EXPORT_SYMBOL(remap_pfn_range);
1668
1669 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1670                                      unsigned long addr, unsigned long end,
1671                                      pte_fn_t fn, void *data)
1672 {
1673         pte_t *pte;
1674         int err;
1675         pgtable_t token;
1676         spinlock_t *uninitialized_var(ptl);
1677
1678         pte = (mm == &init_mm) ?
1679                 pte_alloc_kernel(pmd, addr) :
1680                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1681         if (!pte)
1682                 return -ENOMEM;
1683
1684         BUG_ON(pmd_huge(*pmd));
1685
1686         arch_enter_lazy_mmu_mode();
1687
1688         token = pmd_pgtable(*pmd);
1689
1690         do {
1691                 err = fn(pte, token, addr, data);
1692                 if (err)
1693                         break;
1694         } while (pte++, addr += PAGE_SIZE, addr != end);
1695
1696         arch_leave_lazy_mmu_mode();
1697
1698         if (mm != &init_mm)
1699                 pte_unmap_unlock(pte-1, ptl);
1700         return err;
1701 }
1702
1703 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1704                                      unsigned long addr, unsigned long end,
1705                                      pte_fn_t fn, void *data)
1706 {
1707         pmd_t *pmd;
1708         unsigned long next;
1709         int err;
1710
1711         BUG_ON(pud_huge(*pud));
1712
1713         pmd = pmd_alloc(mm, pud, addr);
1714         if (!pmd)
1715                 return -ENOMEM;
1716         do {
1717                 next = pmd_addr_end(addr, end);
1718                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1719                 if (err)
1720                         break;
1721         } while (pmd++, addr = next, addr != end);
1722         return err;
1723 }
1724
1725 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1726                                      unsigned long addr, unsigned long end,
1727                                      pte_fn_t fn, void *data)
1728 {
1729         pud_t *pud;
1730         unsigned long next;
1731         int err;
1732
1733         pud = pud_alloc(mm, pgd, addr);
1734         if (!pud)
1735                 return -ENOMEM;
1736         do {
1737                 next = pud_addr_end(addr, end);
1738                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1739                 if (err)
1740                         break;
1741         } while (pud++, addr = next, addr != end);
1742         return err;
1743 }
1744
1745 /*
1746  * Scan a region of virtual memory, filling in page tables as necessary
1747  * and calling a provided function on each leaf page table.
1748  */
1749 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1750                         unsigned long size, pte_fn_t fn, void *data)
1751 {
1752         pgd_t *pgd;
1753         unsigned long next;
1754         unsigned long start = addr, end = addr + size;
1755         int err;
1756
1757         BUG_ON(addr >= end);
1758         mmu_notifier_invalidate_range_start(mm, start, end);
1759         pgd = pgd_offset(mm, addr);
1760         do {
1761                 next = pgd_addr_end(addr, end);
1762                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1763                 if (err)
1764                         break;
1765         } while (pgd++, addr = next, addr != end);
1766         mmu_notifier_invalidate_range_end(mm, start, end);
1767         return err;
1768 }
1769 EXPORT_SYMBOL_GPL(apply_to_page_range);
1770
1771 /*
1772  * handle_pte_fault chooses page fault handler according to an entry
1773  * which was read non-atomically.  Before making any commitment, on
1774  * those architectures or configurations (e.g. i386 with PAE) which
1775  * might give a mix of unmatched parts, do_swap_page and do_file_page
1776  * must check under lock before unmapping the pte and proceeding
1777  * (but do_wp_page is only called after already making such a check;
1778  * and do_anonymous_page and do_no_page can safely check later on).
1779  */
1780 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1781                                 pte_t *page_table, pte_t orig_pte)
1782 {
1783         int same = 1;
1784 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1785         if (sizeof(pte_t) > sizeof(unsigned long)) {
1786                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1787                 spin_lock(ptl);
1788                 same = pte_same(*page_table, orig_pte);
1789                 spin_unlock(ptl);
1790         }
1791 #endif
1792         pte_unmap(page_table);
1793         return same;
1794 }
1795
1796 /*
1797  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1798  * servicing faults for write access.  In the normal case, do always want
1799  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1800  * that do not have writing enabled, when used by access_process_vm.
1801  */
1802 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1803 {
1804         if (likely(vma->vm_flags & VM_WRITE))
1805                 pte = pte_mkwrite(pte);
1806         return pte;
1807 }
1808
1809 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1810 {
1811         /*
1812          * If the source page was a PFN mapping, we don't have
1813          * a "struct page" for it. We do a best-effort copy by
1814          * just copying from the original user address. If that
1815          * fails, we just zero-fill it. Live with it.
1816          */
1817         if (unlikely(!src)) {
1818                 void *kaddr = kmap_atomic(dst, KM_USER0);
1819                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1820
1821                 /*
1822                  * This really shouldn't fail, because the page is there
1823                  * in the page tables. But it might just be unreadable,
1824                  * in which case we just give up and fill the result with
1825                  * zeroes.
1826                  */
1827                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1828                         memset(kaddr, 0, PAGE_SIZE);
1829                 kunmap_atomic(kaddr, KM_USER0);
1830                 flush_dcache_page(dst);
1831         } else
1832                 copy_user_highpage(dst, src, va, vma);
1833 }
1834
1835 /*
1836  * This routine handles present pages, when users try to write
1837  * to a shared page. It is done by copying the page to a new address
1838  * and decrementing the shared-page counter for the old page.
1839  *
1840  * Note that this routine assumes that the protection checks have been
1841  * done by the caller (the low-level page fault routine in most cases).
1842  * Thus we can safely just mark it writable once we've done any necessary
1843  * COW.
1844  *
1845  * We also mark the page dirty at this point even though the page will
1846  * change only once the write actually happens. This avoids a few races,
1847  * and potentially makes it more efficient.
1848  *
1849  * We enter with non-exclusive mmap_sem (to exclude vma changes,
1850  * but allow concurrent faults), with pte both mapped and locked.
1851  * We return with mmap_sem still held, but pte unmapped and unlocked.
1852  */
1853 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1854                 unsigned long address, pte_t *page_table, pmd_t *pmd,
1855                 spinlock_t *ptl, pte_t orig_pte)
1856 {
1857         struct page *old_page, *new_page;
1858         pte_t entry;
1859         int reuse = 0, ret = 0;
1860         int page_mkwrite = 0;
1861         struct page *dirty_page = NULL;
1862
1863         old_page = vm_normal_page(vma, address, orig_pte);
1864         if (!old_page) {
1865                 /*
1866                  * VM_MIXEDMAP !pfn_valid() case
1867                  *
1868                  * We should not cow pages in a shared writeable mapping.
1869                  * Just mark the pages writable as we can't do any dirty
1870                  * accounting on raw pfn maps.
1871                  */
1872                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1873                                      (VM_WRITE|VM_SHARED))
1874                         goto reuse;
1875                 goto gotten;
1876         }
1877
1878         /*
1879          * Take out anonymous pages first, anonymous shared vmas are
1880          * not dirty accountable.
1881          */
1882         if (PageAnon(old_page)) {
1883                 if (!trylock_page(old_page)) {
1884                         page_cache_get(old_page);
1885                         pte_unmap_unlock(page_table, ptl);
1886                         lock_page(old_page);
1887                         page_table = pte_offset_map_lock(mm, pmd, address,
1888                                                          &ptl);
1889                         if (!pte_same(*page_table, orig_pte)) {
1890                                 unlock_page(old_page);
1891                                 page_cache_release(old_page);
1892                                 goto unlock;
1893                         }
1894                         page_cache_release(old_page);
1895                 }
1896                 reuse = reuse_swap_page(old_page);
1897                 unlock_page(old_page);
1898         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1899                                         (VM_WRITE|VM_SHARED))) {
1900                 /*
1901                  * Only catch write-faults on shared writable pages,
1902                  * read-only shared pages can get COWed by
1903                  * get_user_pages(.write=1, .force=1).
1904                  */
1905                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1906                         /*
1907                          * Notify the address space that the page is about to
1908                          * become writable so that it can prohibit this or wait
1909                          * for the page to get into an appropriate state.
1910                          *
1911                          * We do this without the lock held, so that it can
1912                          * sleep if it needs to.
1913                          */
1914                         page_cache_get(old_page);
1915                         pte_unmap_unlock(page_table, ptl);
1916
1917                         if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1918                                 goto unwritable_page;
1919
1920                         /*
1921                          * Since we dropped the lock we need to revalidate
1922                          * the PTE as someone else may have changed it.  If
1923                          * they did, we just return, as we can count on the
1924                          * MMU to tell us if they didn't also make it writable.
1925                          */
1926                         page_table = pte_offset_map_lock(mm, pmd, address,
1927                                                          &ptl);
1928                         page_cache_release(old_page);
1929                         if (!pte_same(*page_table, orig_pte))
1930                                 goto unlock;
1931
1932                         page_mkwrite = 1;
1933                 }
1934                 dirty_page = old_page;
1935                 get_page(dirty_page);
1936                 reuse = 1;
1937         }
1938
1939         if (reuse) {
1940 reuse:
1941                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1942                 entry = pte_mkyoung(orig_pte);
1943                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1944                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
1945                         update_mmu_cache(vma, address, entry);
1946                 ret |= VM_FAULT_WRITE;
1947                 goto unlock;
1948         }
1949
1950         /*
1951          * Ok, we need to copy. Oh, well..
1952          */
1953         page_cache_get(old_page);
1954 gotten:
1955         pte_unmap_unlock(page_table, ptl);
1956
1957         if (unlikely(anon_vma_prepare(vma)))
1958                 goto oom;
1959         VM_BUG_ON(old_page == ZERO_PAGE(0));
1960         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1961         if (!new_page)
1962                 goto oom;
1963         /*
1964          * Don't let another task, with possibly unlocked vma,
1965          * keep the mlocked page.
1966          */
1967         if (vma->vm_flags & VM_LOCKED) {
1968                 lock_page(old_page);    /* for LRU manipulation */
1969                 clear_page_mlock(old_page);
1970                 unlock_page(old_page);
1971         }
1972         cow_user_page(new_page, old_page, address, vma);
1973         __SetPageUptodate(new_page);
1974
1975         if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
1976                 goto oom_free_new;
1977
1978         /*
1979          * Re-check the pte - we dropped the lock
1980          */
1981         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1982         if (likely(pte_same(*page_table, orig_pte))) {
1983                 if (old_page) {
1984                         if (!PageAnon(old_page)) {
1985                                 dec_mm_counter(mm, file_rss);
1986                                 inc_mm_counter(mm, anon_rss);
1987                         }
1988                 } else
1989                         inc_mm_counter(mm, anon_rss);
1990                 flush_cache_page(vma, address, pte_pfn(orig_pte));
1991                 entry = mk_pte(new_page, vma->vm_page_prot);
1992                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1993                 /*
1994                  * Clear the pte entry and flush it first, before updating the
1995                  * pte with the new entry. This will avoid a race condition
1996                  * seen in the presence of one thread doing SMC and another
1997                  * thread doing COW.
1998                  */
1999                 ptep_clear_flush_notify(vma, address, page_table);
2000                 page_add_new_anon_rmap(new_page, vma, address);
2001                 set_pte_at(mm, address, page_table, entry);
2002                 update_mmu_cache(vma, address, entry);
2003                 if (old_page) {
2004                         /*
2005                          * Only after switching the pte to the new page may
2006                          * we remove the mapcount here. Otherwise another
2007                          * process may come and find the rmap count decremented
2008                          * before the pte is switched to the new page, and
2009                          * "reuse" the old page writing into it while our pte
2010                          * here still points into it and can be read by other
2011                          * threads.
2012                          *
2013                          * The critical issue is to order this
2014                          * page_remove_rmap with the ptp_clear_flush above.
2015                          * Those stores are ordered by (if nothing else,)
2016                          * the barrier present in the atomic_add_negative
2017                          * in page_remove_rmap.
2018                          *
2019                          * Then the TLB flush in ptep_clear_flush ensures that
2020                          * no process can access the old page before the
2021                          * decremented mapcount is visible. And the old page
2022                          * cannot be reused until after the decremented
2023                          * mapcount is visible. So transitively, TLBs to
2024                          * old page will be flushed before it can be reused.
2025                          */
2026                         page_remove_rmap(old_page);
2027                 }
2028
2029                 /* Free the old page.. */
2030                 new_page = old_page;
2031                 ret |= VM_FAULT_WRITE;
2032         } else
2033                 mem_cgroup_uncharge_page(new_page);
2034
2035         if (new_page)
2036                 page_cache_release(new_page);
2037         if (old_page)
2038                 page_cache_release(old_page);
2039 unlock:
2040         pte_unmap_unlock(page_table, ptl);
2041         if (dirty_page) {
2042                 if (vma->vm_file)
2043                         file_update_time(vma->vm_file);
2044
2045                 /*
2046                  * Yes, Virginia, this is actually required to prevent a race
2047                  * with clear_page_dirty_for_io() from clearing the page dirty
2048                  * bit after it clear all dirty ptes, but before a racing
2049                  * do_wp_page installs a dirty pte.
2050                  *
2051                  * do_no_page is protected similarly.
2052                  */
2053                 wait_on_page_locked(dirty_page);
2054                 set_page_dirty_balance(dirty_page, page_mkwrite);
2055                 put_page(dirty_page);
2056         }
2057         return ret;
2058 oom_free_new:
2059         page_cache_release(new_page);
2060 oom:
2061         if (old_page)
2062                 page_cache_release(old_page);
2063         return VM_FAULT_OOM;
2064
2065 unwritable_page:
2066         page_cache_release(old_page);
2067         return VM_FAULT_SIGBUS;
2068 }
2069
2070 /*
2071  * Helper functions for unmap_mapping_range().
2072  *
2073  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2074  *
2075  * We have to restart searching the prio_tree whenever we drop the lock,
2076  * since the iterator is only valid while the lock is held, and anyway
2077  * a later vma might be split and reinserted earlier while lock dropped.
2078  *
2079  * The list of nonlinear vmas could be handled more efficiently, using
2080  * a placeholder, but handle it in the same way until a need is shown.
2081  * It is important to search the prio_tree before nonlinear list: a vma
2082  * may become nonlinear and be shifted from prio_tree to nonlinear list
2083  * while the lock is dropped; but never shifted from list to prio_tree.
2084  *
2085  * In order to make forward progress despite restarting the search,
2086  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2087  * quickly skip it next time around.  Since the prio_tree search only
2088  * shows us those vmas affected by unmapping the range in question, we
2089  * can't efficiently keep all vmas in step with mapping->truncate_count:
2090  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2091  * mapping->truncate_count and vma->vm_truncate_count are protected by
2092  * i_mmap_lock.
2093  *
2094  * In order to make forward progress despite repeatedly restarting some
2095  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2096  * and restart from that address when we reach that vma again.  It might
2097  * have been split or merged, shrunk or extended, but never shifted: so
2098  * restart_addr remains valid so long as it remains in the vma's range.
2099  * unmap_mapping_range forces truncate_count to leap over page-aligned
2100  * values so we can save vma's restart_addr in its truncate_count field.
2101  */
2102 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2103
2104 static void reset_vma_truncate_counts(struct address_space *mapping)
2105 {
2106         struct vm_area_struct *vma;
2107         struct prio_tree_iter iter;
2108
2109         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2110                 vma->vm_truncate_count = 0;
2111         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2112                 vma->vm_truncate_count = 0;
2113 }
2114
2115 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2116                 unsigned long start_addr, unsigned long end_addr,
2117                 struct zap_details *details)
2118 {
2119         unsigned long restart_addr;
2120         int need_break;
2121
2122         /*
2123          * files that support invalidating or truncating portions of the
2124          * file from under mmaped areas must have their ->fault function
2125          * return a locked page (and set VM_FAULT_LOCKED in the return).
2126          * This provides synchronisation against concurrent unmapping here.
2127          */
2128
2129 again:
2130         restart_addr = vma->vm_truncate_count;
2131         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2132                 start_addr = restart_addr;
2133                 if (start_addr >= end_addr) {
2134                         /* Top of vma has been split off since last time */
2135                         vma->vm_truncate_count = details->truncate_count;
2136                         return 0;
2137                 }
2138         }
2139
2140         restart_addr = zap_page_range(vma, start_addr,
2141                                         end_addr - start_addr, details);
2142         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2143
2144         if (restart_addr >= end_addr) {
2145                 /* We have now completed this vma: mark it so */
2146                 vma->vm_truncate_count = details->truncate_count;
2147                 if (!need_break)
2148                         return 0;
2149         } else {
2150                 /* Note restart_addr in vma's truncate_count field */
2151                 vma->vm_truncate_count = restart_addr;
2152                 if (!need_break)
2153                         goto again;
2154         }
2155
2156         spin_unlock(details->i_mmap_lock);
2157         cond_resched();
2158         spin_lock(details->i_mmap_lock);
2159         return -EINTR;
2160 }
2161
2162 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2163                                             struct zap_details *details)
2164 {
2165         struct vm_area_struct *vma;
2166         struct prio_tree_iter iter;
2167         pgoff_t vba, vea, zba, zea;
2168
2169 restart:
2170         vma_prio_tree_foreach(vma, &iter, root,
2171                         details->first_index, details->last_index) {
2172                 /* Skip quickly over those we have already dealt with */
2173                 if (vma->vm_truncate_count == details->truncate_count)
2174                         continue;
2175
2176                 vba = vma->vm_pgoff;
2177                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2178                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2179                 zba = details->first_index;
2180                 if (zba < vba)
2181                         zba = vba;
2182                 zea = details->last_index;
2183                 if (zea > vea)
2184                         zea = vea;
2185
2186                 if (unmap_mapping_range_vma(vma,
2187                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2188                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2189                                 details) < 0)
2190                         goto restart;
2191         }
2192 }
2193
2194 static inline void unmap_mapping_range_list(struct list_head *head,
2195                                             struct zap_details *details)
2196 {
2197         struct vm_area_struct *vma;
2198
2199         /*
2200          * In nonlinear VMAs there is no correspondence between virtual address
2201          * offset and file offset.  So we must perform an exhaustive search
2202          * across *all* the pages in each nonlinear VMA, not just the pages
2203          * whose virtual address lies outside the file truncation point.
2204          */
2205 restart:
2206         list_for_each_entry(vma, head, shared.vm_set.list) {
2207                 /* Skip quickly over those we have already dealt with */
2208                 if (vma->vm_truncate_count == details->truncate_count)
2209                         continue;
2210                 details->nonlinear_vma = vma;
2211                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2212                                         vma->vm_end, details) < 0)
2213                         goto restart;
2214         }
2215 }
2216
2217 /**
2218  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2219  * @mapping: the address space containing mmaps to be unmapped.
2220  * @holebegin: byte in first page to unmap, relative to the start of
2221  * the underlying file.  This will be rounded down to a PAGE_SIZE
2222  * boundary.  Note that this is different from vmtruncate(), which
2223  * must keep the partial page.  In contrast, we must get rid of
2224  * partial pages.
2225  * @holelen: size of prospective hole in bytes.  This will be rounded
2226  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2227  * end of the file.
2228  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2229  * but 0 when invalidating pagecache, don't throw away private data.
2230  */
2231 void unmap_mapping_range(struct address_space *mapping,
2232                 loff_t const holebegin, loff_t const holelen, int even_cows)
2233 {
2234         struct zap_details details;
2235         pgoff_t hba = holebegin >> PAGE_SHIFT;
2236         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2237
2238         /* Check for overflow. */
2239         if (sizeof(holelen) > sizeof(hlen)) {
2240                 long long holeend =
2241                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2242                 if (holeend & ~(long long)ULONG_MAX)
2243                         hlen = ULONG_MAX - hba + 1;
2244         }
2245
2246         details.check_mapping = even_cows? NULL: mapping;
2247         details.nonlinear_vma = NULL;
2248         details.first_index = hba;
2249         details.last_index = hba + hlen - 1;
2250         if (details.last_index < details.first_index)
2251                 details.last_index = ULONG_MAX;
2252         details.i_mmap_lock = &mapping->i_mmap_lock;
2253
2254         spin_lock(&mapping->i_mmap_lock);
2255
2256         /* Protect against endless unmapping loops */
2257         mapping->truncate_count++;
2258         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2259                 if (mapping->truncate_count == 0)
2260                         reset_vma_truncate_counts(mapping);
2261                 mapping->truncate_count++;
2262         }
2263         details.truncate_count = mapping->truncate_count;
2264
2265         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2266                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2267         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2268                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2269         spin_unlock(&mapping->i_mmap_lock);
2270 }
2271 EXPORT_SYMBOL(unmap_mapping_range);
2272
2273 /**
2274  * vmtruncate - unmap mappings "freed" by truncate() syscall
2275  * @inode: inode of the file used
2276  * @offset: file offset to start truncating
2277  *
2278  * NOTE! We have to be ready to update the memory sharing
2279  * between the file and the memory map for a potential last
2280  * incomplete page.  Ugly, but necessary.
2281  */
2282 int vmtruncate(struct inode * inode, loff_t offset)
2283 {
2284         if (inode->i_size < offset) {
2285                 unsigned long limit;
2286
2287                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2288                 if (limit != RLIM_INFINITY && offset > limit)
2289                         goto out_sig;
2290                 if (offset > inode->i_sb->s_maxbytes)
2291                         goto out_big;
2292                 i_size_write(inode, offset);
2293         } else {
2294                 struct address_space *mapping = inode->i_mapping;
2295
2296                 /*
2297                  * truncation of in-use swapfiles is disallowed - it would
2298                  * cause subsequent swapout to scribble on the now-freed
2299                  * blocks.
2300                  */
2301                 if (IS_SWAPFILE(inode))
2302                         return -ETXTBSY;
2303                 i_size_write(inode, offset);
2304
2305                 /*
2306                  * unmap_mapping_range is called twice, first simply for
2307                  * efficiency so that truncate_inode_pages does fewer
2308                  * single-page unmaps.  However after this first call, and
2309                  * before truncate_inode_pages finishes, it is possible for
2310                  * private pages to be COWed, which remain after
2311                  * truncate_inode_pages finishes, hence the second
2312                  * unmap_mapping_range call must be made for correctness.
2313                  */
2314                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2315                 truncate_inode_pages(mapping, offset);
2316                 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2317         }
2318
2319         if (inode->i_op->truncate)
2320                 inode->i_op->truncate(inode);
2321         return 0;
2322
2323 out_sig:
2324         send_sig(SIGXFSZ, current, 0);
2325 out_big:
2326         return -EFBIG;
2327 }
2328 EXPORT_SYMBOL(vmtruncate);
2329
2330 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2331 {
2332         struct address_space *mapping = inode->i_mapping;
2333
2334         /*
2335          * If the underlying filesystem is not going to provide
2336          * a way to truncate a range of blocks (punch a hole) -
2337          * we should return failure right now.
2338          */
2339         if (!inode->i_op->truncate_range)
2340                 return -ENOSYS;
2341
2342         mutex_lock(&inode->i_mutex);
2343         down_write(&inode->i_alloc_sem);
2344         unmap_mapping_range(mapping, offset, (end - offset), 1);
2345         truncate_inode_pages_range(mapping, offset, end);
2346         unmap_mapping_range(mapping, offset, (end - offset), 1);
2347         inode->i_op->truncate_range(inode, offset, end);
2348         up_write(&inode->i_alloc_sem);
2349         mutex_unlock(&inode->i_mutex);
2350
2351         return 0;
2352 }
2353
2354 /*
2355  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2356  * but allow concurrent faults), and pte mapped but not yet locked.
2357  * We return with mmap_sem still held, but pte unmapped and unlocked.
2358  */
2359 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2360                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2361                 int write_access, pte_t orig_pte)
2362 {
2363         spinlock_t *ptl;
2364         struct page *page;
2365         swp_entry_t entry;
2366         pte_t pte;
2367         int ret = 0;
2368
2369         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2370                 goto out;
2371
2372         entry = pte_to_swp_entry(orig_pte);
2373         if (is_migration_entry(entry)) {
2374                 migration_entry_wait(mm, pmd, address);
2375                 goto out;
2376         }
2377         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2378         page = lookup_swap_cache(entry);
2379         if (!page) {
2380                 grab_swap_token(); /* Contend for token _before_ read-in */
2381                 page = swapin_readahead(entry,
2382                                         GFP_HIGHUSER_MOVABLE, vma, address);
2383                 if (!page) {
2384                         /*
2385                          * Back out if somebody else faulted in this pte
2386                          * while we released the pte lock.
2387                          */
2388                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2389                         if (likely(pte_same(*page_table, orig_pte)))
2390                                 ret = VM_FAULT_OOM;
2391                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2392                         goto unlock;
2393                 }
2394
2395                 /* Had to read the page from swap area: Major fault */
2396                 ret = VM_FAULT_MAJOR;
2397                 count_vm_event(PGMAJFAULT);
2398         }
2399
2400         mark_page_accessed(page);
2401
2402         lock_page(page);
2403         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2404
2405         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2406                 ret = VM_FAULT_OOM;
2407                 unlock_page(page);
2408                 goto out;
2409         }
2410
2411         /*
2412          * Back out if somebody else already faulted in this pte.
2413          */
2414         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2415         if (unlikely(!pte_same(*page_table, orig_pte)))
2416                 goto out_nomap;
2417
2418         if (unlikely(!PageUptodate(page))) {
2419                 ret = VM_FAULT_SIGBUS;
2420                 goto out_nomap;
2421         }
2422
2423         /* The page isn't present yet, go ahead with the fault. */
2424
2425         inc_mm_counter(mm, anon_rss);
2426         pte = mk_pte(page, vma->vm_page_prot);
2427         if (write_access && reuse_swap_page(page)) {
2428                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2429                 write_access = 0;
2430         }
2431
2432         flush_icache_page(vma, page);
2433         set_pte_at(mm, address, page_table, pte);
2434         page_add_anon_rmap(page, vma, address);
2435
2436         swap_free(entry);
2437         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2438                 try_to_free_swap(page);
2439         unlock_page(page);
2440
2441         if (write_access) {
2442                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2443                 if (ret & VM_FAULT_ERROR)
2444                         ret &= VM_FAULT_ERROR;
2445                 goto out;
2446         }
2447
2448         /* No need to invalidate - it was non-present before */
2449         update_mmu_cache(vma, address, pte);
2450 unlock:
2451         pte_unmap_unlock(page_table, ptl);
2452 out:
2453         return ret;
2454 out_nomap:
2455         mem_cgroup_uncharge_page(page);
2456         pte_unmap_unlock(page_table, ptl);
2457         unlock_page(page);
2458         page_cache_release(page);
2459         return ret;
2460 }
2461
2462 /*
2463  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2464  * but allow concurrent faults), and pte mapped but not yet locked.
2465  * We return with mmap_sem still held, but pte unmapped and unlocked.
2466  */
2467 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2468                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2469                 int write_access)
2470 {
2471         struct page *page;
2472         spinlock_t *ptl;
2473         pte_t entry;
2474
2475         /* Allocate our own private page. */
2476         pte_unmap(page_table);
2477
2478         if (unlikely(anon_vma_prepare(vma)))
2479                 goto oom;
2480         page = alloc_zeroed_user_highpage_movable(vma, address);
2481         if (!page)
2482                 goto oom;
2483         __SetPageUptodate(page);
2484
2485         if (mem_cgroup_charge(page, mm, GFP_KERNEL))
2486                 goto oom_free_page;
2487
2488         entry = mk_pte(page, vma->vm_page_prot);
2489         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2490
2491         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2492         if (!pte_none(*page_table))
2493                 goto release;
2494         inc_mm_counter(mm, anon_rss);
2495         page_add_new_anon_rmap(page, vma, address);
2496         set_pte_at(mm, address, page_table, entry);
2497
2498         /* No need to invalidate - it was non-present before */
2499         update_mmu_cache(vma, address, entry);
2500 unlock:
2501         pte_unmap_unlock(page_table, ptl);
2502         return 0;
2503 release:
2504         mem_cgroup_uncharge_page(page);
2505         page_cache_release(page);
2506         goto unlock;
2507 oom_free_page:
2508         page_cache_release(page);
2509 oom:
2510         return VM_FAULT_OOM;
2511 }
2512
2513 /*
2514  * __do_fault() tries to create a new page mapping. It aggressively
2515  * tries to share with existing pages, but makes a separate copy if
2516  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2517  * the next page fault.
2518  *
2519  * As this is called only for pages that do not currently exist, we
2520  * do not need to flush old virtual caches or the TLB.
2521  *
2522  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2523  * but allow concurrent faults), and pte neither mapped nor locked.
2524  * We return with mmap_sem still held, but pte unmapped and unlocked.
2525  */
2526 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2527                 unsigned long address, pmd_t *pmd,
2528                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2529 {
2530         pte_t *page_table;
2531         spinlock_t *ptl;
2532         struct page *page;
2533         pte_t entry;
2534         int anon = 0;
2535         int charged = 0;
2536         struct page *dirty_page = NULL;
2537         struct vm_fault vmf;
2538         int ret;
2539         int page_mkwrite = 0;
2540
2541         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2542         vmf.pgoff = pgoff;
2543         vmf.flags = flags;
2544         vmf.page = NULL;
2545
2546         ret = vma->vm_ops->fault(vma, &vmf);
2547         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2548                 return ret;
2549
2550         /*
2551          * For consistency in subsequent calls, make the faulted page always
2552          * locked.
2553          */
2554         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2555                 lock_page(vmf.page);
2556         else
2557                 VM_BUG_ON(!PageLocked(vmf.page));
2558
2559         /*
2560          * Should we do an early C-O-W break?
2561          */
2562         page = vmf.page;
2563         if (flags & FAULT_FLAG_WRITE) {
2564                 if (!(vma->vm_flags & VM_SHARED)) {
2565                         anon = 1;
2566                         if (unlikely(anon_vma_prepare(vma))) {
2567                                 ret = VM_FAULT_OOM;
2568                                 goto out;
2569                         }
2570                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2571                                                 vma, address);
2572                         if (!page) {
2573                                 ret = VM_FAULT_OOM;
2574                                 goto out;
2575                         }
2576                         if (mem_cgroup_charge(page, mm, GFP_KERNEL)) {
2577                                 ret = VM_FAULT_OOM;
2578                                 page_cache_release(page);
2579                                 goto out;
2580                         }
2581                         charged = 1;
2582                         /*
2583                          * Don't let another task, with possibly unlocked vma,
2584                          * keep the mlocked page.
2585                          */
2586                         if (vma->vm_flags & VM_LOCKED)
2587                                 clear_page_mlock(vmf.page);
2588                         copy_user_highpage(page, vmf.page, address, vma);
2589                         __SetPageUptodate(page);
2590                 } else {
2591                         /*
2592                          * If the page will be shareable, see if the backing
2593                          * address space wants to know that the page is about
2594                          * to become writable
2595                          */
2596                         if (vma->vm_ops->page_mkwrite) {
2597                                 unlock_page(page);
2598                                 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2599                                         ret = VM_FAULT_SIGBUS;
2600                                         anon = 1; /* no anon but release vmf.page */
2601                                         goto out_unlocked;
2602                                 }
2603                                 lock_page(page);
2604                                 /*
2605                                  * XXX: this is not quite right (racy vs
2606                                  * invalidate) to unlock and relock the page
2607                                  * like this, however a better fix requires
2608                                  * reworking page_mkwrite locking API, which
2609                                  * is better done later.
2610                                  */
2611                                 if (!page->mapping) {
2612                                         ret = 0;
2613                                         anon = 1; /* no anon but release vmf.page */
2614                                         goto out;
2615                                 }
2616                                 page_mkwrite = 1;
2617                         }
2618                 }
2619
2620         }
2621
2622         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2623
2624         /*
2625          * This silly early PAGE_DIRTY setting removes a race
2626          * due to the bad i386 page protection. But it's valid
2627          * for other architectures too.
2628          *
2629          * Note that if write_access is true, we either now have
2630          * an exclusive copy of the page, or this is a shared mapping,
2631          * so we can make it writable and dirty to avoid having to
2632          * handle that later.
2633          */
2634         /* Only go through if we didn't race with anybody else... */
2635         if (likely(pte_same(*page_table, orig_pte))) {
2636                 flush_icache_page(vma, page);
2637                 entry = mk_pte(page, vma->vm_page_prot);
2638                 if (flags & FAULT_FLAG_WRITE)
2639                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2640                 if (anon) {
2641                         inc_mm_counter(mm, anon_rss);
2642                         page_add_new_anon_rmap(page, vma, address);
2643                 } else {
2644                         inc_mm_counter(mm, file_rss);
2645                         page_add_file_rmap(page);
2646                         if (flags & FAULT_FLAG_WRITE) {
2647                                 dirty_page = page;
2648                                 get_page(dirty_page);
2649                         }
2650                 }
2651                 set_pte_at(mm, address, page_table, entry);
2652
2653                 /* no need to invalidate: a not-present page won't be cached */
2654                 update_mmu_cache(vma, address, entry);
2655         } else {
2656                 if (charged)
2657                         mem_cgroup_uncharge_page(page);
2658                 if (anon)
2659                         page_cache_release(page);
2660                 else
2661                         anon = 1; /* no anon but release faulted_page */
2662         }
2663
2664         pte_unmap_unlock(page_table, ptl);
2665
2666 out:
2667         unlock_page(vmf.page);
2668 out_unlocked:
2669         if (anon)
2670                 page_cache_release(vmf.page);
2671         else if (dirty_page) {
2672                 if (vma->vm_file)
2673                         file_update_time(vma->vm_file);
2674
2675                 set_page_dirty_balance(dirty_page, page_mkwrite);
2676                 put_page(dirty_page);
2677         }
2678
2679         return ret;
2680 }
2681
2682 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2683                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2684                 int write_access, pte_t orig_pte)
2685 {
2686         pgoff_t pgoff = (((address & PAGE_MASK)
2687                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2688         unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2689
2690         pte_unmap(page_table);
2691         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2692 }
2693
2694 /*
2695  * Fault of a previously existing named mapping. Repopulate the pte
2696  * from the encoded file_pte if possible. This enables swappable
2697  * nonlinear vmas.
2698  *
2699  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2700  * but allow concurrent faults), and pte mapped but not yet locked.
2701  * We return with mmap_sem still held, but pte unmapped and unlocked.
2702  */
2703 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2704                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2705                 int write_access, pte_t orig_pte)
2706 {
2707         unsigned int flags = FAULT_FLAG_NONLINEAR |
2708                                 (write_access ? FAULT_FLAG_WRITE : 0);
2709         pgoff_t pgoff;
2710
2711         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2712                 return 0;
2713
2714         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2715                 /*
2716                  * Page table corrupted: show pte and kill process.
2717                  */
2718                 print_bad_pte(vma, address, orig_pte, NULL);
2719                 return VM_FAULT_OOM;
2720         }
2721
2722         pgoff = pte_to_pgoff(orig_pte);
2723         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2724 }
2725
2726 /*
2727  * These routines also need to handle stuff like marking pages dirty
2728  * and/or accessed for architectures that don't do it in hardware (most
2729  * RISC architectures).  The early dirtying is also good on the i386.
2730  *
2731  * There is also a hook called "update_mmu_cache()" that architectures
2732  * with external mmu caches can use to update those (ie the Sparc or
2733  * PowerPC hashed page tables that act as extended TLBs).
2734  *
2735  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2736  * but allow concurrent faults), and pte mapped but not yet locked.
2737  * We return with mmap_sem still held, but pte unmapped and unlocked.
2738  */
2739 static inline int handle_pte_fault(struct mm_struct *mm,
2740                 struct vm_area_struct *vma, unsigned long address,
2741                 pte_t *pte, pmd_t *pmd, int write_access)
2742 {
2743         pte_t entry;
2744         spinlock_t *ptl;
2745
2746         entry = *pte;
2747         if (!pte_present(entry)) {
2748                 if (pte_none(entry)) {
2749                         if (vma->vm_ops) {
2750                                 if (likely(vma->vm_ops->fault))
2751                                         return do_linear_fault(mm, vma, address,
2752                                                 pte, pmd, write_access, entry);
2753                         }
2754                         return do_anonymous_page(mm, vma, address,
2755                                                  pte, pmd, write_access);
2756                 }
2757                 if (pte_file(entry))
2758                         return do_nonlinear_fault(mm, vma, address,
2759                                         pte, pmd, write_access, entry);
2760                 return do_swap_page(mm, vma, address,
2761                                         pte, pmd, write_access, entry);
2762         }
2763
2764         ptl = pte_lockptr(mm, pmd);
2765         spin_lock(ptl);
2766         if (unlikely(!pte_same(*pte, entry)))
2767                 goto unlock;
2768         if (write_access) {
2769                 if (!pte_write(entry))
2770                         return do_wp_page(mm, vma, address,
2771                                         pte, pmd, ptl, entry);
2772                 entry = pte_mkdirty(entry);
2773         }
2774         entry = pte_mkyoung(entry);
2775         if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2776                 update_mmu_cache(vma, address, entry);
2777         } else {
2778                 /*
2779                  * This is needed only for protection faults but the arch code
2780                  * is not yet telling us if this is a protection fault or not.
2781                  * This still avoids useless tlb flushes for .text page faults
2782                  * with threads.
2783                  */
2784                 if (write_access)
2785                         flush_tlb_page(vma, address);
2786         }
2787 unlock:
2788         pte_unmap_unlock(pte, ptl);
2789         return 0;
2790 }
2791
2792 /*
2793  * By the time we get here, we already hold the mm semaphore
2794  */
2795 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2796                 unsigned long address, int write_access)
2797 {
2798         pgd_t *pgd;
2799         pud_t *pud;
2800         pmd_t *pmd;
2801         pte_t *pte;
2802
2803         __set_current_state(TASK_RUNNING);
2804
2805         count_vm_event(PGFAULT);
2806
2807         if (unlikely(is_vm_hugetlb_page(vma)))
2808                 return hugetlb_fault(mm, vma, address, write_access);
2809
2810         pgd = pgd_offset(mm, address);
2811         pud = pud_alloc(mm, pgd, address);
2812         if (!pud)
2813                 return VM_FAULT_OOM;
2814         pmd = pmd_alloc(mm, pud, address);
2815         if (!pmd)
2816                 return VM_FAULT_OOM;
2817         pte = pte_alloc_map(mm, pmd, address);
2818         if (!pte)
2819                 return VM_FAULT_OOM;
2820
2821         return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2822 }
2823
2824 #ifndef __PAGETABLE_PUD_FOLDED
2825 /*
2826  * Allocate page upper directory.
2827  * We've already handled the fast-path in-line.
2828  */
2829 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2830 {
2831         pud_t *new = pud_alloc_one(mm, address);
2832         if (!new)
2833                 return -ENOMEM;
2834
2835         smp_wmb(); /* See comment in __pte_alloc */
2836
2837         spin_lock(&mm->page_table_lock);
2838         if (pgd_present(*pgd))          /* Another has populated it */
2839                 pud_free(mm, new);
2840         else
2841                 pgd_populate(mm, pgd, new);
2842         spin_unlock(&mm->page_table_lock);
2843         return 0;
2844 }
2845 #endif /* __PAGETABLE_PUD_FOLDED */
2846
2847 #ifndef __PAGETABLE_PMD_FOLDED
2848 /*
2849  * Allocate page middle directory.
2850  * We've already handled the fast-path in-line.
2851  */
2852 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2853 {
2854         pmd_t *new = pmd_alloc_one(mm, address);
2855         if (!new)
2856                 return -ENOMEM;
2857
2858         smp_wmb(); /* See comment in __pte_alloc */
2859
2860         spin_lock(&mm->page_table_lock);
2861 #ifndef __ARCH_HAS_4LEVEL_HACK
2862         if (pud_present(*pud))          /* Another has populated it */
2863                 pmd_free(mm, new);
2864         else
2865                 pud_populate(mm, pud, new);
2866 #else
2867         if (pgd_present(*pud))          /* Another has populated it */
2868                 pmd_free(mm, new);
2869         else
2870                 pgd_populate(mm, pud, new);
2871 #endif /* __ARCH_HAS_4LEVEL_HACK */
2872         spin_unlock(&mm->page_table_lock);
2873         return 0;
2874 }
2875 #endif /* __PAGETABLE_PMD_FOLDED */
2876
2877 int make_pages_present(unsigned long addr, unsigned long end)
2878 {
2879         int ret, len, write;
2880         struct vm_area_struct * vma;
2881
2882         vma = find_vma(current->mm, addr);
2883         if (!vma)
2884                 return -ENOMEM;
2885         write = (vma->vm_flags & VM_WRITE) != 0;
2886         BUG_ON(addr >= end);
2887         BUG_ON(end > vma->vm_end);
2888         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2889         ret = get_user_pages(current, current->mm, addr,
2890                         len, write, 0, NULL, NULL);
2891         if (ret < 0)
2892                 return ret;
2893         return ret == len ? 0 : -EFAULT;
2894 }
2895
2896 #if !defined(__HAVE_ARCH_GATE_AREA)
2897
2898 #if defined(AT_SYSINFO_EHDR)
2899 static struct vm_area_struct gate_vma;
2900
2901 static int __init gate_vma_init(void)
2902 {
2903         gate_vma.vm_mm = NULL;
2904         gate_vma.vm_start = FIXADDR_USER_START;
2905         gate_vma.vm_end = FIXADDR_USER_END;
2906         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2907         gate_vma.vm_page_prot = __P101;
2908         /*
2909          * Make sure the vDSO gets into every core dump.
2910          * Dumping its contents makes post-mortem fully interpretable later
2911          * without matching up the same kernel and hardware config to see
2912          * what PC values meant.
2913          */
2914         gate_vma.vm_flags |= VM_ALWAYSDUMP;
2915         return 0;
2916 }
2917 __initcall(gate_vma_init);
2918 #endif
2919
2920 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2921 {
2922 #ifdef AT_SYSINFO_EHDR
2923         return &gate_vma;
2924 #else
2925         return NULL;
2926 #endif
2927 }
2928
2929 int in_gate_area_no_task(unsigned long addr)
2930 {
2931 #ifdef AT_SYSINFO_EHDR
2932         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2933                 return 1;
2934 #endif
2935         return 0;
2936 }
2937
2938 #endif  /* __HAVE_ARCH_GATE_AREA */
2939
2940 #ifdef CONFIG_HAVE_IOREMAP_PROT
2941 int follow_phys(struct vm_area_struct *vma,
2942                 unsigned long address, unsigned int flags,
2943                 unsigned long *prot, resource_size_t *phys)
2944 {
2945         pgd_t *pgd;
2946         pud_t *pud;
2947         pmd_t *pmd;
2948         pte_t *ptep, pte;
2949         spinlock_t *ptl;
2950         resource_size_t phys_addr = 0;
2951         struct mm_struct *mm = vma->vm_mm;
2952         int ret = -EINVAL;
2953
2954         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
2955                 goto out;
2956
2957         pgd = pgd_offset(mm, address);
2958         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
2959                 goto out;
2960
2961         pud = pud_offset(pgd, address);
2962         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
2963                 goto out;
2964
2965         pmd = pmd_offset(pud, address);
2966         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
2967                 goto out;
2968
2969         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
2970         if (pmd_huge(*pmd))
2971                 goto out;
2972
2973         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
2974         if (!ptep)
2975                 goto out;
2976
2977         pte = *ptep;
2978         if (!pte_present(pte))
2979                 goto unlock;
2980         if ((flags & FOLL_WRITE) && !pte_write(pte))
2981                 goto unlock;
2982         phys_addr = pte_pfn(pte);
2983         phys_addr <<= PAGE_SHIFT; /* Shift here to avoid overflow on PAE */
2984
2985         *prot = pgprot_val(pte_pgprot(pte));
2986         *phys = phys_addr;
2987         ret = 0;
2988
2989 unlock:
2990         pte_unmap_unlock(ptep, ptl);
2991 out:
2992         return ret;
2993 }
2994
2995 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
2996                         void *buf, int len, int write)
2997 {
2998         resource_size_t phys_addr;
2999         unsigned long prot = 0;
3000         void __iomem *maddr;
3001         int offset = addr & (PAGE_SIZE-1);
3002
3003         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3004                 return -EINVAL;
3005
3006         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3007         if (write)
3008                 memcpy_toio(maddr + offset, buf, len);
3009         else
3010                 memcpy_fromio(buf, maddr + offset, len);
3011         iounmap(maddr);
3012
3013         return len;
3014 }
3015 #endif
3016
3017 /*
3018  * Access another process' address space.
3019  * Source/target buffer must be kernel space,
3020  * Do not walk the page table directly, use get_user_pages
3021  */
3022 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3023 {
3024         struct mm_struct *mm;
3025         struct vm_area_struct *vma;
3026         void *old_buf = buf;
3027
3028         mm = get_task_mm(tsk);
3029         if (!mm)
3030                 return 0;
3031
3032         down_read(&mm->mmap_sem);
3033         /* ignore errors, just check how much was successfully transferred */
3034         while (len) {
3035                 int bytes, ret, offset;
3036                 void *maddr;
3037                 struct page *page = NULL;
3038
3039                 ret = get_user_pages(tsk, mm, addr, 1,
3040                                 write, 1, &page, &vma);
3041                 if (ret <= 0) {
3042                         /*
3043                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3044                          * we can access using slightly different code.
3045                          */
3046 #ifdef CONFIG_HAVE_IOREMAP_PROT
3047                         vma = find_vma(mm, addr);
3048                         if (!vma)
3049                                 break;
3050                         if (vma->vm_ops && vma->vm_ops->access)
3051                                 ret = vma->vm_ops->access(vma, addr, buf,
3052                                                           len, write);
3053                         if (ret <= 0)
3054 #endif
3055                                 break;
3056                         bytes = ret;
3057                 } else {
3058                         bytes = len;
3059                         offset = addr & (PAGE_SIZE-1);
3060                         if (bytes > PAGE_SIZE-offset)
3061                                 bytes = PAGE_SIZE-offset;
3062
3063                         maddr = kmap(page);
3064                         if (write) {
3065                                 copy_to_user_page(vma, page, addr,
3066                                                   maddr + offset, buf, bytes);
3067                                 set_page_dirty_lock(page);
3068                         } else {
3069                                 copy_from_user_page(vma, page, addr,
3070                                                     buf, maddr + offset, bytes);
3071                         }
3072                         kunmap(page);
3073                         page_cache_release(page);
3074                 }
3075                 len -= bytes;
3076                 buf += bytes;
3077                 addr += bytes;
3078         }
3079         up_read(&mm->mmap_sem);
3080         mmput(mm);
3081
3082         return buf - old_buf;
3083 }
3084
3085 /*
3086  * Print the name of a VMA.
3087  */
3088 void print_vma_addr(char *prefix, unsigned long ip)
3089 {
3090         struct mm_struct *mm = current->mm;
3091         struct vm_area_struct *vma;
3092
3093         /*
3094          * Do not print if we are in atomic
3095          * contexts (in exception stacks, etc.):
3096          */
3097         if (preempt_count())
3098                 return;
3099
3100         down_read(&mm->mmap_sem);
3101         vma = find_vma(mm, ip);
3102         if (vma && vma->vm_file) {
3103                 struct file *f = vma->vm_file;
3104                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3105                 if (buf) {
3106                         char *p, *s;
3107
3108                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3109                         if (IS_ERR(p))
3110                                 p = "?";
3111                         s = strrchr(p, '/');
3112                         if (s)
3113                                 p = s+1;
3114                         printk("%s%s[%lx+%lx]", prefix, p,
3115                                         vma->vm_start,
3116                                         vma->vm_end - vma->vm_start);
3117                         free_page((unsigned long)buf);
3118                 }
3119         }
3120         up_read(&current->mm->mmap_sem);
3121 }
3122
3123 #ifdef CONFIG_PROVE_LOCKING
3124 void might_fault(void)
3125 {
3126         might_sleep();
3127         /*
3128          * it would be nicer only to annotate paths which are not under
3129          * pagefault_disable, however that requires a larger audit and
3130          * providing helpers like get_user_atomic.
3131          */
3132         if (!in_atomic() && current->mm)
3133                 might_lock_read(&current->mm->mmap_sem);
3134 }
3135 EXPORT_SYMBOL(might_fault);
3136 #endif