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