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