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