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