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