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