memcg: ensure list is empty at rmdir
[linux-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59
60 #include <asm/io.h>
61 #include <asm/pgalloc.h>
62 #include <asm/uaccess.h>
63 #include <asm/tlb.h>
64 #include <asm/tlbflush.h>
65 #include <asm/pgtable.h>
66
67 #include "internal.h"
68
69 #ifndef CONFIG_NEED_MULTIPLE_NODES
70 /* use the per-pgdat data instead for discontigmem - mbligh */
71 unsigned long max_mapnr;
72 struct page *mem_map;
73
74 EXPORT_SYMBOL(max_mapnr);
75 EXPORT_SYMBOL(mem_map);
76 #endif
77
78 unsigned long num_physpages;
79 /*
80  * A number of key systems in x86 including ioremap() rely on the assumption
81  * that high_memory defines the upper bound on direct map memory, then end
82  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
83  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
84  * and ZONE_HIGHMEM.
85  */
86 void * high_memory;
87
88 EXPORT_SYMBOL(num_physpages);
89 EXPORT_SYMBOL(high_memory);
90
91 /*
92  * Randomize the address space (stacks, mmaps, brk, etc.).
93  *
94  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
95  *   as ancient (libc5 based) binaries can segfault. )
96  */
97 int randomize_va_space __read_mostly =
98 #ifdef CONFIG_COMPAT_BRK
99                                         1;
100 #else
101                                         2;
102 #endif
103
104 static int __init disable_randmaps(char *s)
105 {
106         randomize_va_space = 0;
107         return 1;
108 }
109 __setup("norandmaps", disable_randmaps);
110
111 unsigned long zero_pfn __read_mostly;
112 unsigned long highest_memmap_pfn __read_mostly;
113
114 /*
115  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
116  */
117 static int __init init_zero_pfn(void)
118 {
119         zero_pfn = page_to_pfn(ZERO_PAGE(0));
120         return 0;
121 }
122 core_initcall(init_zero_pfn);
123
124 /*
125  * If a p?d_bad entry is found while walking page tables, report
126  * the error, before resetting entry to p?d_none.  Usually (but
127  * very seldom) called out from the p?d_none_or_clear_bad macros.
128  */
129
130 void pgd_clear_bad(pgd_t *pgd)
131 {
132         pgd_ERROR(*pgd);
133         pgd_clear(pgd);
134 }
135
136 void pud_clear_bad(pud_t *pud)
137 {
138         pud_ERROR(*pud);
139         pud_clear(pud);
140 }
141
142 void pmd_clear_bad(pmd_t *pmd)
143 {
144         pmd_ERROR(*pmd);
145         pmd_clear(pmd);
146 }
147
148 /*
149  * Note: this doesn't free the actual pages themselves. That
150  * has been handled earlier when unmapping all the memory regions.
151  */
152 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
153                            unsigned long addr)
154 {
155         pgtable_t token = pmd_pgtable(*pmd);
156         pmd_clear(pmd);
157         pte_free_tlb(tlb, token, addr);
158         tlb->mm->nr_ptes--;
159 }
160
161 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
162                                 unsigned long addr, unsigned long end,
163                                 unsigned long floor, unsigned long ceiling)
164 {
165         pmd_t *pmd;
166         unsigned long next;
167         unsigned long start;
168
169         start = addr;
170         pmd = pmd_offset(pud, addr);
171         do {
172                 next = pmd_addr_end(addr, end);
173                 if (pmd_none_or_clear_bad(pmd))
174                         continue;
175                 free_pte_range(tlb, pmd, addr);
176         } while (pmd++, addr = next, addr != end);
177
178         start &= PUD_MASK;
179         if (start < floor)
180                 return;
181         if (ceiling) {
182                 ceiling &= PUD_MASK;
183                 if (!ceiling)
184                         return;
185         }
186         if (end - 1 > ceiling - 1)
187                 return;
188
189         pmd = pmd_offset(pud, start);
190         pud_clear(pud);
191         pmd_free_tlb(tlb, pmd, start);
192 }
193
194 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
195                                 unsigned long addr, unsigned long end,
196                                 unsigned long floor, unsigned long ceiling)
197 {
198         pud_t *pud;
199         unsigned long next;
200         unsigned long start;
201
202         start = addr;
203         pud = pud_offset(pgd, addr);
204         do {
205                 next = pud_addr_end(addr, end);
206                 if (pud_none_or_clear_bad(pud))
207                         continue;
208                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
209         } while (pud++, addr = next, addr != end);
210
211         start &= PGDIR_MASK;
212         if (start < floor)
213                 return;
214         if (ceiling) {
215                 ceiling &= PGDIR_MASK;
216                 if (!ceiling)
217                         return;
218         }
219         if (end - 1 > ceiling - 1)
220                 return;
221
222         pud = pud_offset(pgd, start);
223         pgd_clear(pgd);
224         pud_free_tlb(tlb, pud, start);
225 }
226
227 /*
228  * This function frees user-level page tables of a process.
229  *
230  * Must be called with pagetable lock held.
231  */
232 void free_pgd_range(struct mmu_gather *tlb,
233                         unsigned long addr, unsigned long end,
234                         unsigned long floor, unsigned long ceiling)
235 {
236         pgd_t *pgd;
237         unsigned long next;
238         unsigned long start;
239
240         /*
241          * The next few lines have given us lots of grief...
242          *
243          * Why are we testing PMD* at this top level?  Because often
244          * there will be no work to do at all, and we'd prefer not to
245          * go all the way down to the bottom just to discover that.
246          *
247          * Why all these "- 1"s?  Because 0 represents both the bottom
248          * of the address space and the top of it (using -1 for the
249          * top wouldn't help much: the masks would do the wrong thing).
250          * The rule is that addr 0 and floor 0 refer to the bottom of
251          * the address space, but end 0 and ceiling 0 refer to the top
252          * Comparisons need to use "end - 1" and "ceiling - 1" (though
253          * that end 0 case should be mythical).
254          *
255          * Wherever addr is brought up or ceiling brought down, we must
256          * be careful to reject "the opposite 0" before it confuses the
257          * subsequent tests.  But what about where end is brought down
258          * by PMD_SIZE below? no, end can't go down to 0 there.
259          *
260          * Whereas we round start (addr) and ceiling down, by different
261          * masks at different levels, in order to test whether a table
262          * now has no other vmas using it, so can be freed, we don't
263          * bother to round floor or end up - the tests don't need that.
264          */
265
266         addr &= PMD_MASK;
267         if (addr < floor) {
268                 addr += PMD_SIZE;
269                 if (!addr)
270                         return;
271         }
272         if (ceiling) {
273                 ceiling &= PMD_MASK;
274                 if (!ceiling)
275                         return;
276         }
277         if (end - 1 > ceiling - 1)
278                 end -= PMD_SIZE;
279         if (addr > end - 1)
280                 return;
281
282         start = addr;
283         pgd = pgd_offset(tlb->mm, addr);
284         do {
285                 next = pgd_addr_end(addr, end);
286                 if (pgd_none_or_clear_bad(pgd))
287                         continue;
288                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
289         } while (pgd++, addr = next, addr != end);
290 }
291
292 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
293                 unsigned long floor, unsigned long ceiling)
294 {
295         while (vma) {
296                 struct vm_area_struct *next = vma->vm_next;
297                 unsigned long addr = vma->vm_start;
298
299                 /*
300                  * Hide vma from rmap and truncate_pagecache before freeing
301                  * pgtables
302                  */
303                 anon_vma_unlink(vma);
304                 unlink_file_vma(vma);
305
306                 if (is_vm_hugetlb_page(vma)) {
307                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
308                                 floor, next? next->vm_start: ceiling);
309                 } else {
310                         /*
311                          * Optimization: gather nearby vmas into one call down
312                          */
313                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
314                                && !is_vm_hugetlb_page(next)) {
315                                 vma = next;
316                                 next = vma->vm_next;
317                                 anon_vma_unlink(vma);
318                                 unlink_file_vma(vma);
319                         }
320                         free_pgd_range(tlb, addr, vma->vm_end,
321                                 floor, next? next->vm_start: ceiling);
322                 }
323                 vma = next;
324         }
325 }
326
327 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
328 {
329         pgtable_t new = pte_alloc_one(mm, address);
330         if (!new)
331                 return -ENOMEM;
332
333         /*
334          * Ensure all pte setup (eg. pte page lock and page clearing) are
335          * visible before the pte is made visible to other CPUs by being
336          * put into page tables.
337          *
338          * The other side of the story is the pointer chasing in the page
339          * table walking code (when walking the page table without locking;
340          * ie. most of the time). Fortunately, these data accesses consist
341          * of a chain of data-dependent loads, meaning most CPUs (alpha
342          * being the notable exception) will already guarantee loads are
343          * seen in-order. See the alpha page table accessors for the
344          * smp_read_barrier_depends() barriers in page table walking code.
345          */
346         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
347
348         spin_lock(&mm->page_table_lock);
349         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
350                 mm->nr_ptes++;
351                 pmd_populate(mm, pmd, new);
352                 new = NULL;
353         }
354         spin_unlock(&mm->page_table_lock);
355         if (new)
356                 pte_free(mm, new);
357         return 0;
358 }
359
360 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
361 {
362         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
363         if (!new)
364                 return -ENOMEM;
365
366         smp_wmb(); /* See comment in __pte_alloc */
367
368         spin_lock(&init_mm.page_table_lock);
369         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
370                 pmd_populate_kernel(&init_mm, pmd, new);
371                 new = NULL;
372         }
373         spin_unlock(&init_mm.page_table_lock);
374         if (new)
375                 pte_free_kernel(&init_mm, new);
376         return 0;
377 }
378
379 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
380 {
381         if (file_rss)
382                 add_mm_counter(mm, file_rss, file_rss);
383         if (anon_rss)
384                 add_mm_counter(mm, anon_rss, anon_rss);
385 }
386
387 /*
388  * This function is called to print an error when a bad pte
389  * is found. For example, we might have a PFN-mapped pte in
390  * a region that doesn't allow it.
391  *
392  * The calling function must still handle the error.
393  */
394 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
395                           pte_t pte, struct page *page)
396 {
397         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
398         pud_t *pud = pud_offset(pgd, addr);
399         pmd_t *pmd = pmd_offset(pud, addr);
400         struct address_space *mapping;
401         pgoff_t index;
402         static unsigned long resume;
403         static unsigned long nr_shown;
404         static unsigned long nr_unshown;
405
406         /*
407          * Allow a burst of 60 reports, then keep quiet for that minute;
408          * or allow a steady drip of one report per second.
409          */
410         if (nr_shown == 60) {
411                 if (time_before(jiffies, resume)) {
412                         nr_unshown++;
413                         return;
414                 }
415                 if (nr_unshown) {
416                         printk(KERN_ALERT
417                                 "BUG: Bad page map: %lu messages suppressed\n",
418                                 nr_unshown);
419                         nr_unshown = 0;
420                 }
421                 nr_shown = 0;
422         }
423         if (nr_shown++ == 0)
424                 resume = jiffies + 60 * HZ;
425
426         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
427         index = linear_page_index(vma, addr);
428
429         printk(KERN_ALERT
430                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
431                 current->comm,
432                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
433         if (page) {
434                 printk(KERN_ALERT
435                 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
436                 page, (void *)page->flags, page_count(page),
437                 page_mapcount(page), page->mapping, page->index);
438         }
439         printk(KERN_ALERT
440                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
441                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
442         /*
443          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
444          */
445         if (vma->vm_ops)
446                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
447                                 (unsigned long)vma->vm_ops->fault);
448         if (vma->vm_file && vma->vm_file->f_op)
449                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
450                                 (unsigned long)vma->vm_file->f_op->mmap);
451         dump_stack();
452         add_taint(TAINT_BAD_PAGE);
453 }
454
455 static inline int is_cow_mapping(unsigned int flags)
456 {
457         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
458 }
459
460 #ifndef is_zero_pfn
461 static inline int is_zero_pfn(unsigned long pfn)
462 {
463         return pfn == zero_pfn;
464 }
465 #endif
466
467 #ifndef my_zero_pfn
468 static inline unsigned long my_zero_pfn(unsigned long addr)
469 {
470         return zero_pfn;
471 }
472 #endif
473
474 /*
475  * vm_normal_page -- This function gets the "struct page" associated with a pte.
476  *
477  * "Special" mappings do not wish to be associated with a "struct page" (either
478  * it doesn't exist, or it exists but they don't want to touch it). In this
479  * case, NULL is returned here. "Normal" mappings do have a struct page.
480  *
481  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
482  * pte bit, in which case this function is trivial. Secondly, an architecture
483  * may not have a spare pte bit, which requires a more complicated scheme,
484  * described below.
485  *
486  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
487  * special mapping (even if there are underlying and valid "struct pages").
488  * COWed pages of a VM_PFNMAP are always normal.
489  *
490  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
491  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
492  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
493  * mapping will always honor the rule
494  *
495  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
496  *
497  * And for normal mappings this is false.
498  *
499  * This restricts such mappings to be a linear translation from virtual address
500  * to pfn. To get around this restriction, we allow arbitrary mappings so long
501  * as the vma is not a COW mapping; in that case, we know that all ptes are
502  * special (because none can have been COWed).
503  *
504  *
505  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
506  *
507  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
508  * page" backing, however the difference is that _all_ pages with a struct
509  * page (that is, those where pfn_valid is true) are refcounted and considered
510  * normal pages by the VM. The disadvantage is that pages are refcounted
511  * (which can be slower and simply not an option for some PFNMAP users). The
512  * advantage is that we don't have to follow the strict linearity rule of
513  * PFNMAP mappings in order to support COWable mappings.
514  *
515  */
516 #ifdef __HAVE_ARCH_PTE_SPECIAL
517 # define HAVE_PTE_SPECIAL 1
518 #else
519 # define HAVE_PTE_SPECIAL 0
520 #endif
521 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
522                                 pte_t pte)
523 {
524         unsigned long pfn = pte_pfn(pte);
525
526         if (HAVE_PTE_SPECIAL) {
527                 if (likely(!pte_special(pte)))
528                         goto check_pfn;
529                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
530                         return NULL;
531                 if (!is_zero_pfn(pfn))
532                         print_bad_pte(vma, addr, pte, NULL);
533                 return NULL;
534         }
535
536         /* !HAVE_PTE_SPECIAL case follows: */
537
538         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
539                 if (vma->vm_flags & VM_MIXEDMAP) {
540                         if (!pfn_valid(pfn))
541                                 return NULL;
542                         goto out;
543                 } else {
544                         unsigned long off;
545                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
546                         if (pfn == vma->vm_pgoff + off)
547                                 return NULL;
548                         if (!is_cow_mapping(vma->vm_flags))
549                                 return NULL;
550                 }
551         }
552
553         if (is_zero_pfn(pfn))
554                 return NULL;
555 check_pfn:
556         if (unlikely(pfn > highest_memmap_pfn)) {
557                 print_bad_pte(vma, addr, pte, NULL);
558                 return NULL;
559         }
560
561         /*
562          * NOTE! We still have PageReserved() pages in the page tables.
563          * eg. VDSO mappings can cause them to exist.
564          */
565 out:
566         return pfn_to_page(pfn);
567 }
568
569 /*
570  * copy one vm_area from one task to the other. Assumes the page tables
571  * already present in the new task to be cleared in the whole range
572  * covered by this vma.
573  */
574
575 static inline unsigned long
576 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
577                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
578                 unsigned long addr, int *rss)
579 {
580         unsigned long vm_flags = vma->vm_flags;
581         pte_t pte = *src_pte;
582         struct page *page;
583
584         /* pte contains position in swap or file, so copy. */
585         if (unlikely(!pte_present(pte))) {
586                 if (!pte_file(pte)) {
587                         swp_entry_t entry = pte_to_swp_entry(pte);
588
589                         if (swap_duplicate(entry) < 0)
590                                 return entry.val;
591
592                         /* make sure dst_mm is on swapoff's mmlist. */
593                         if (unlikely(list_empty(&dst_mm->mmlist))) {
594                                 spin_lock(&mmlist_lock);
595                                 if (list_empty(&dst_mm->mmlist))
596                                         list_add(&dst_mm->mmlist,
597                                                  &src_mm->mmlist);
598                                 spin_unlock(&mmlist_lock);
599                         }
600                         if (is_write_migration_entry(entry) &&
601                                         is_cow_mapping(vm_flags)) {
602                                 /*
603                                  * COW mappings require pages in both parent
604                                  * and child to be set to read.
605                                  */
606                                 make_migration_entry_read(&entry);
607                                 pte = swp_entry_to_pte(entry);
608                                 set_pte_at(src_mm, addr, src_pte, pte);
609                         }
610                 }
611                 goto out_set_pte;
612         }
613
614         /*
615          * If it's a COW mapping, write protect it both
616          * in the parent and the child
617          */
618         if (is_cow_mapping(vm_flags)) {
619                 ptep_set_wrprotect(src_mm, addr, src_pte);
620                 pte = pte_wrprotect(pte);
621         }
622
623         /*
624          * If it's a shared mapping, mark it clean in
625          * the child
626          */
627         if (vm_flags & VM_SHARED)
628                 pte = pte_mkclean(pte);
629         pte = pte_mkold(pte);
630
631         page = vm_normal_page(vma, addr, pte);
632         if (page) {
633                 get_page(page);
634                 page_dup_rmap(page);
635                 rss[PageAnon(page)]++;
636         }
637
638 out_set_pte:
639         set_pte_at(dst_mm, addr, dst_pte, pte);
640         return 0;
641 }
642
643 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
644                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
645                 unsigned long addr, unsigned long end)
646 {
647         pte_t *orig_src_pte, *orig_dst_pte;
648         pte_t *src_pte, *dst_pte;
649         spinlock_t *src_ptl, *dst_ptl;
650         int progress = 0;
651         int rss[2];
652         swp_entry_t entry = (swp_entry_t){0};
653
654 again:
655         rss[1] = rss[0] = 0;
656         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
657         if (!dst_pte)
658                 return -ENOMEM;
659         src_pte = pte_offset_map_nested(src_pmd, addr);
660         src_ptl = pte_lockptr(src_mm, src_pmd);
661         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
662         orig_src_pte = src_pte;
663         orig_dst_pte = dst_pte;
664         arch_enter_lazy_mmu_mode();
665
666         do {
667                 /*
668                  * We are holding two locks at this point - either of them
669                  * could generate latencies in another task on another CPU.
670                  */
671                 if (progress >= 32) {
672                         progress = 0;
673                         if (need_resched() ||
674                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
675                                 break;
676                 }
677                 if (pte_none(*src_pte)) {
678                         progress++;
679                         continue;
680                 }
681                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
682                                                         vma, addr, rss);
683                 if (entry.val)
684                         break;
685                 progress += 8;
686         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
687
688         arch_leave_lazy_mmu_mode();
689         spin_unlock(src_ptl);
690         pte_unmap_nested(orig_src_pte);
691         add_mm_rss(dst_mm, rss[0], rss[1]);
692         pte_unmap_unlock(orig_dst_pte, dst_ptl);
693         cond_resched();
694
695         if (entry.val) {
696                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
697                         return -ENOMEM;
698                 progress = 0;
699         }
700         if (addr != end)
701                 goto again;
702         return 0;
703 }
704
705 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
706                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
707                 unsigned long addr, unsigned long end)
708 {
709         pmd_t *src_pmd, *dst_pmd;
710         unsigned long next;
711
712         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
713         if (!dst_pmd)
714                 return -ENOMEM;
715         src_pmd = pmd_offset(src_pud, addr);
716         do {
717                 next = pmd_addr_end(addr, end);
718                 if (pmd_none_or_clear_bad(src_pmd))
719                         continue;
720                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
721                                                 vma, addr, next))
722                         return -ENOMEM;
723         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
724         return 0;
725 }
726
727 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
728                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
729                 unsigned long addr, unsigned long end)
730 {
731         pud_t *src_pud, *dst_pud;
732         unsigned long next;
733
734         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
735         if (!dst_pud)
736                 return -ENOMEM;
737         src_pud = pud_offset(src_pgd, addr);
738         do {
739                 next = pud_addr_end(addr, end);
740                 if (pud_none_or_clear_bad(src_pud))
741                         continue;
742                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
743                                                 vma, addr, next))
744                         return -ENOMEM;
745         } while (dst_pud++, src_pud++, addr = next, addr != end);
746         return 0;
747 }
748
749 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
750                 struct vm_area_struct *vma)
751 {
752         pgd_t *src_pgd, *dst_pgd;
753         unsigned long next;
754         unsigned long addr = vma->vm_start;
755         unsigned long end = vma->vm_end;
756         int ret;
757
758         /*
759          * Don't copy ptes where a page fault will fill them correctly.
760          * Fork becomes much lighter when there are big shared or private
761          * readonly mappings. The tradeoff is that copy_page_range is more
762          * efficient than faulting.
763          */
764         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
765                 if (!vma->anon_vma)
766                         return 0;
767         }
768
769         if (is_vm_hugetlb_page(vma))
770                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
771
772         if (unlikely(is_pfn_mapping(vma))) {
773                 /*
774                  * We do not free on error cases below as remove_vma
775                  * gets called on error from higher level routine
776                  */
777                 ret = track_pfn_vma_copy(vma);
778                 if (ret)
779                         return ret;
780         }
781
782         /*
783          * We need to invalidate the secondary MMU mappings only when
784          * there could be a permission downgrade on the ptes of the
785          * parent mm. And a permission downgrade will only happen if
786          * is_cow_mapping() returns true.
787          */
788         if (is_cow_mapping(vma->vm_flags))
789                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
790
791         ret = 0;
792         dst_pgd = pgd_offset(dst_mm, addr);
793         src_pgd = pgd_offset(src_mm, addr);
794         do {
795                 next = pgd_addr_end(addr, end);
796                 if (pgd_none_or_clear_bad(src_pgd))
797                         continue;
798                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
799                                             vma, addr, next))) {
800                         ret = -ENOMEM;
801                         break;
802                 }
803         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
804
805         if (is_cow_mapping(vma->vm_flags))
806                 mmu_notifier_invalidate_range_end(src_mm,
807                                                   vma->vm_start, end);
808         return ret;
809 }
810
811 static unsigned long zap_pte_range(struct mmu_gather *tlb,
812                                 struct vm_area_struct *vma, pmd_t *pmd,
813                                 unsigned long addr, unsigned long end,
814                                 long *zap_work, struct zap_details *details)
815 {
816         struct mm_struct *mm = tlb->mm;
817         pte_t *pte;
818         spinlock_t *ptl;
819         int file_rss = 0;
820         int anon_rss = 0;
821
822         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
823         arch_enter_lazy_mmu_mode();
824         do {
825                 pte_t ptent = *pte;
826                 if (pte_none(ptent)) {
827                         (*zap_work)--;
828                         continue;
829                 }
830
831                 (*zap_work) -= PAGE_SIZE;
832
833                 if (pte_present(ptent)) {
834                         struct page *page;
835
836                         page = vm_normal_page(vma, addr, ptent);
837                         if (unlikely(details) && page) {
838                                 /*
839                                  * unmap_shared_mapping_pages() wants to
840                                  * invalidate cache without truncating:
841                                  * unmap shared but keep private pages.
842                                  */
843                                 if (details->check_mapping &&
844                                     details->check_mapping != page->mapping)
845                                         continue;
846                                 /*
847                                  * Each page->index must be checked when
848                                  * invalidating or truncating nonlinear.
849                                  */
850                                 if (details->nonlinear_vma &&
851                                     (page->index < details->first_index ||
852                                      page->index > details->last_index))
853                                         continue;
854                         }
855                         ptent = ptep_get_and_clear_full(mm, addr, pte,
856                                                         tlb->fullmm);
857                         tlb_remove_tlb_entry(tlb, pte, addr);
858                         if (unlikely(!page))
859                                 continue;
860                         if (unlikely(details) && details->nonlinear_vma
861                             && linear_page_index(details->nonlinear_vma,
862                                                 addr) != page->index)
863                                 set_pte_at(mm, addr, pte,
864                                            pgoff_to_pte(page->index));
865                         if (PageAnon(page))
866                                 anon_rss--;
867                         else {
868                                 if (pte_dirty(ptent))
869                                         set_page_dirty(page);
870                                 if (pte_young(ptent) &&
871                                     likely(!VM_SequentialReadHint(vma)))
872                                         mark_page_accessed(page);
873                                 file_rss--;
874                         }
875                         page_remove_rmap(page);
876                         if (unlikely(page_mapcount(page) < 0))
877                                 print_bad_pte(vma, addr, ptent, page);
878                         tlb_remove_page(tlb, page);
879                         continue;
880                 }
881                 /*
882                  * If details->check_mapping, we leave swap entries;
883                  * if details->nonlinear_vma, we leave file entries.
884                  */
885                 if (unlikely(details))
886                         continue;
887                 if (pte_file(ptent)) {
888                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
889                                 print_bad_pte(vma, addr, ptent, NULL);
890                 } else if
891                   (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
892                         print_bad_pte(vma, addr, ptent, NULL);
893                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
894         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
895
896         add_mm_rss(mm, file_rss, anon_rss);
897         arch_leave_lazy_mmu_mode();
898         pte_unmap_unlock(pte - 1, ptl);
899
900         return addr;
901 }
902
903 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
904                                 struct vm_area_struct *vma, pud_t *pud,
905                                 unsigned long addr, unsigned long end,
906                                 long *zap_work, struct zap_details *details)
907 {
908         pmd_t *pmd;
909         unsigned long next;
910
911         pmd = pmd_offset(pud, addr);
912         do {
913                 next = pmd_addr_end(addr, end);
914                 if (pmd_none_or_clear_bad(pmd)) {
915                         (*zap_work)--;
916                         continue;
917                 }
918                 next = zap_pte_range(tlb, vma, pmd, addr, next,
919                                                 zap_work, details);
920         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
921
922         return addr;
923 }
924
925 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
926                                 struct vm_area_struct *vma, pgd_t *pgd,
927                                 unsigned long addr, unsigned long end,
928                                 long *zap_work, struct zap_details *details)
929 {
930         pud_t *pud;
931         unsigned long next;
932
933         pud = pud_offset(pgd, addr);
934         do {
935                 next = pud_addr_end(addr, end);
936                 if (pud_none_or_clear_bad(pud)) {
937                         (*zap_work)--;
938                         continue;
939                 }
940                 next = zap_pmd_range(tlb, vma, pud, addr, next,
941                                                 zap_work, details);
942         } while (pud++, addr = next, (addr != end && *zap_work > 0));
943
944         return addr;
945 }
946
947 static unsigned long unmap_page_range(struct mmu_gather *tlb,
948                                 struct vm_area_struct *vma,
949                                 unsigned long addr, unsigned long end,
950                                 long *zap_work, struct zap_details *details)
951 {
952         pgd_t *pgd;
953         unsigned long next;
954
955         if (details && !details->check_mapping && !details->nonlinear_vma)
956                 details = NULL;
957
958         BUG_ON(addr >= end);
959         mem_cgroup_uncharge_start();
960         tlb_start_vma(tlb, vma);
961         pgd = pgd_offset(vma->vm_mm, addr);
962         do {
963                 next = pgd_addr_end(addr, end);
964                 if (pgd_none_or_clear_bad(pgd)) {
965                         (*zap_work)--;
966                         continue;
967                 }
968                 next = zap_pud_range(tlb, vma, pgd, addr, next,
969                                                 zap_work, details);
970         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
971         tlb_end_vma(tlb, vma);
972         mem_cgroup_uncharge_end();
973
974         return addr;
975 }
976
977 #ifdef CONFIG_PREEMPT
978 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
979 #else
980 /* No preempt: go for improved straight-line efficiency */
981 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
982 #endif
983
984 /**
985  * unmap_vmas - unmap a range of memory covered by a list of vma's
986  * @tlbp: address of the caller's struct mmu_gather
987  * @vma: the starting vma
988  * @start_addr: virtual address at which to start unmapping
989  * @end_addr: virtual address at which to end unmapping
990  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
991  * @details: details of nonlinear truncation or shared cache invalidation
992  *
993  * Returns the end address of the unmapping (restart addr if interrupted).
994  *
995  * Unmap all pages in the vma list.
996  *
997  * We aim to not hold locks for too long (for scheduling latency reasons).
998  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
999  * return the ending mmu_gather to the caller.
1000  *
1001  * Only addresses between `start' and `end' will be unmapped.
1002  *
1003  * The VMA list must be sorted in ascending virtual address order.
1004  *
1005  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1006  * range after unmap_vmas() returns.  So the only responsibility here is to
1007  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1008  * drops the lock and schedules.
1009  */
1010 unsigned long unmap_vmas(struct mmu_gather **tlbp,
1011                 struct vm_area_struct *vma, unsigned long start_addr,
1012                 unsigned long end_addr, unsigned long *nr_accounted,
1013                 struct zap_details *details)
1014 {
1015         long zap_work = ZAP_BLOCK_SIZE;
1016         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
1017         int tlb_start_valid = 0;
1018         unsigned long start = start_addr;
1019         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1020         int fullmm = (*tlbp)->fullmm;
1021         struct mm_struct *mm = vma->vm_mm;
1022
1023         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1024         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1025                 unsigned long end;
1026
1027                 start = max(vma->vm_start, start_addr);
1028                 if (start >= vma->vm_end)
1029                         continue;
1030                 end = min(vma->vm_end, end_addr);
1031                 if (end <= vma->vm_start)
1032                         continue;
1033
1034                 if (vma->vm_flags & VM_ACCOUNT)
1035                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1036
1037                 if (unlikely(is_pfn_mapping(vma)))
1038                         untrack_pfn_vma(vma, 0, 0);
1039
1040                 while (start != end) {
1041                         if (!tlb_start_valid) {
1042                                 tlb_start = start;
1043                                 tlb_start_valid = 1;
1044                         }
1045
1046                         if (unlikely(is_vm_hugetlb_page(vma))) {
1047                                 /*
1048                                  * It is undesirable to test vma->vm_file as it
1049                                  * should be non-null for valid hugetlb area.
1050                                  * However, vm_file will be NULL in the error
1051                                  * cleanup path of do_mmap_pgoff. When
1052                                  * hugetlbfs ->mmap method fails,
1053                                  * do_mmap_pgoff() nullifies vma->vm_file
1054                                  * before calling this function to clean up.
1055                                  * Since no pte has actually been setup, it is
1056                                  * safe to do nothing in this case.
1057                                  */
1058                                 if (vma->vm_file) {
1059                                         unmap_hugepage_range(vma, start, end, NULL);
1060                                         zap_work -= (end - start) /
1061                                         pages_per_huge_page(hstate_vma(vma));
1062                                 }
1063
1064                                 start = end;
1065                         } else
1066                                 start = unmap_page_range(*tlbp, vma,
1067                                                 start, end, &zap_work, details);
1068
1069                         if (zap_work > 0) {
1070                                 BUG_ON(start != end);
1071                                 break;
1072                         }
1073
1074                         tlb_finish_mmu(*tlbp, tlb_start, start);
1075
1076                         if (need_resched() ||
1077                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1078                                 if (i_mmap_lock) {
1079                                         *tlbp = NULL;
1080                                         goto out;
1081                                 }
1082                                 cond_resched();
1083                         }
1084
1085                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1086                         tlb_start_valid = 0;
1087                         zap_work = ZAP_BLOCK_SIZE;
1088                 }
1089         }
1090 out:
1091         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1092         return start;   /* which is now the end (or restart) address */
1093 }
1094
1095 /**
1096  * zap_page_range - remove user pages in a given range
1097  * @vma: vm_area_struct holding the applicable pages
1098  * @address: starting address of pages to zap
1099  * @size: number of bytes to zap
1100  * @details: details of nonlinear truncation or shared cache invalidation
1101  */
1102 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1103                 unsigned long size, struct zap_details *details)
1104 {
1105         struct mm_struct *mm = vma->vm_mm;
1106         struct mmu_gather *tlb;
1107         unsigned long end = address + size;
1108         unsigned long nr_accounted = 0;
1109
1110         lru_add_drain();
1111         tlb = tlb_gather_mmu(mm, 0);
1112         update_hiwater_rss(mm);
1113         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1114         if (tlb)
1115                 tlb_finish_mmu(tlb, address, end);
1116         return end;
1117 }
1118
1119 /**
1120  * zap_vma_ptes - remove ptes mapping the vma
1121  * @vma: vm_area_struct holding ptes to be zapped
1122  * @address: starting address of pages to zap
1123  * @size: number of bytes to zap
1124  *
1125  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1126  *
1127  * The entire address range must be fully contained within the vma.
1128  *
1129  * Returns 0 if successful.
1130  */
1131 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1132                 unsigned long size)
1133 {
1134         if (address < vma->vm_start || address + size > vma->vm_end ||
1135                         !(vma->vm_flags & VM_PFNMAP))
1136                 return -1;
1137         zap_page_range(vma, address, size, NULL);
1138         return 0;
1139 }
1140 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1141
1142 /*
1143  * Do a quick page-table lookup for a single page.
1144  */
1145 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1146                         unsigned int flags)
1147 {
1148         pgd_t *pgd;
1149         pud_t *pud;
1150         pmd_t *pmd;
1151         pte_t *ptep, pte;
1152         spinlock_t *ptl;
1153         struct page *page;
1154         struct mm_struct *mm = vma->vm_mm;
1155
1156         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1157         if (!IS_ERR(page)) {
1158                 BUG_ON(flags & FOLL_GET);
1159                 goto out;
1160         }
1161
1162         page = NULL;
1163         pgd = pgd_offset(mm, address);
1164         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1165                 goto no_page_table;
1166
1167         pud = pud_offset(pgd, address);
1168         if (pud_none(*pud))
1169                 goto no_page_table;
1170         if (pud_huge(*pud)) {
1171                 BUG_ON(flags & FOLL_GET);
1172                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1173                 goto out;
1174         }
1175         if (unlikely(pud_bad(*pud)))
1176                 goto no_page_table;
1177
1178         pmd = pmd_offset(pud, address);
1179         if (pmd_none(*pmd))
1180                 goto no_page_table;
1181         if (pmd_huge(*pmd)) {
1182                 BUG_ON(flags & FOLL_GET);
1183                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1184                 goto out;
1185         }
1186         if (unlikely(pmd_bad(*pmd)))
1187                 goto no_page_table;
1188
1189         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1190
1191         pte = *ptep;
1192         if (!pte_present(pte))
1193                 goto no_page;
1194         if ((flags & FOLL_WRITE) && !pte_write(pte))
1195                 goto unlock;
1196
1197         page = vm_normal_page(vma, address, pte);
1198         if (unlikely(!page)) {
1199                 if ((flags & FOLL_DUMP) ||
1200                     !is_zero_pfn(pte_pfn(pte)))
1201                         goto bad_page;
1202                 page = pte_page(pte);
1203         }
1204
1205         if (flags & FOLL_GET)
1206                 get_page(page);
1207         if (flags & FOLL_TOUCH) {
1208                 if ((flags & FOLL_WRITE) &&
1209                     !pte_dirty(pte) && !PageDirty(page))
1210                         set_page_dirty(page);
1211                 /*
1212                  * pte_mkyoung() would be more correct here, but atomic care
1213                  * is needed to avoid losing the dirty bit: it is easier to use
1214                  * mark_page_accessed().
1215                  */
1216                 mark_page_accessed(page);
1217         }
1218 unlock:
1219         pte_unmap_unlock(ptep, ptl);
1220 out:
1221         return page;
1222
1223 bad_page:
1224         pte_unmap_unlock(ptep, ptl);
1225         return ERR_PTR(-EFAULT);
1226
1227 no_page:
1228         pte_unmap_unlock(ptep, ptl);
1229         if (!pte_none(pte))
1230                 return page;
1231
1232 no_page_table:
1233         /*
1234          * When core dumping an enormous anonymous area that nobody
1235          * has touched so far, we don't want to allocate unnecessary pages or
1236          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1237          * then get_dump_page() will return NULL to leave a hole in the dump.
1238          * But we can only make this optimization where a hole would surely
1239          * be zero-filled if handle_mm_fault() actually did handle it.
1240          */
1241         if ((flags & FOLL_DUMP) &&
1242             (!vma->vm_ops || !vma->vm_ops->fault))
1243                 return ERR_PTR(-EFAULT);
1244         return page;
1245 }
1246
1247 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1248                      unsigned long start, int nr_pages, unsigned int gup_flags,
1249                      struct page **pages, struct vm_area_struct **vmas)
1250 {
1251         int i;
1252         unsigned long vm_flags;
1253
1254         if (nr_pages <= 0)
1255                 return 0;
1256
1257         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1258
1259         /* 
1260          * Require read or write permissions.
1261          * If FOLL_FORCE is set, we only require the "MAY" flags.
1262          */
1263         vm_flags  = (gup_flags & FOLL_WRITE) ?
1264                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1265         vm_flags &= (gup_flags & FOLL_FORCE) ?
1266                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1267         i = 0;
1268
1269         do {
1270                 struct vm_area_struct *vma;
1271
1272                 vma = find_extend_vma(mm, start);
1273                 if (!vma && in_gate_area(tsk, start)) {
1274                         unsigned long pg = start & PAGE_MASK;
1275                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1276                         pgd_t *pgd;
1277                         pud_t *pud;
1278                         pmd_t *pmd;
1279                         pte_t *pte;
1280
1281                         /* user gate pages are read-only */
1282                         if (gup_flags & FOLL_WRITE)
1283                                 return i ? : -EFAULT;
1284                         if (pg > TASK_SIZE)
1285                                 pgd = pgd_offset_k(pg);
1286                         else
1287                                 pgd = pgd_offset_gate(mm, pg);
1288                         BUG_ON(pgd_none(*pgd));
1289                         pud = pud_offset(pgd, pg);
1290                         BUG_ON(pud_none(*pud));
1291                         pmd = pmd_offset(pud, pg);
1292                         if (pmd_none(*pmd))
1293                                 return i ? : -EFAULT;
1294                         pte = pte_offset_map(pmd, pg);
1295                         if (pte_none(*pte)) {
1296                                 pte_unmap(pte);
1297                                 return i ? : -EFAULT;
1298                         }
1299                         if (pages) {
1300                                 struct page *page = vm_normal_page(gate_vma, start, *pte);
1301                                 pages[i] = page;
1302                                 if (page)
1303                                         get_page(page);
1304                         }
1305                         pte_unmap(pte);
1306                         if (vmas)
1307                                 vmas[i] = gate_vma;
1308                         i++;
1309                         start += PAGE_SIZE;
1310                         nr_pages--;
1311                         continue;
1312                 }
1313
1314                 if (!vma ||
1315                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1316                     !(vm_flags & vma->vm_flags))
1317                         return i ? : -EFAULT;
1318
1319                 if (is_vm_hugetlb_page(vma)) {
1320                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1321                                         &start, &nr_pages, i, gup_flags);
1322                         continue;
1323                 }
1324
1325                 do {
1326                         struct page *page;
1327                         unsigned int foll_flags = gup_flags;
1328
1329                         /*
1330                          * If we have a pending SIGKILL, don't keep faulting
1331                          * pages and potentially allocating memory.
1332                          */
1333                         if (unlikely(fatal_signal_pending(current)))
1334                                 return i ? i : -ERESTARTSYS;
1335
1336                         cond_resched();
1337                         while (!(page = follow_page(vma, start, foll_flags))) {
1338                                 int ret;
1339
1340                                 ret = handle_mm_fault(mm, vma, start,
1341                                         (foll_flags & FOLL_WRITE) ?
1342                                         FAULT_FLAG_WRITE : 0);
1343
1344                                 if (ret & VM_FAULT_ERROR) {
1345                                         if (ret & VM_FAULT_OOM)
1346                                                 return i ? i : -ENOMEM;
1347                                         if (ret &
1348                                             (VM_FAULT_HWPOISON|VM_FAULT_SIGBUS))
1349                                                 return i ? i : -EFAULT;
1350                                         BUG();
1351                                 }
1352                                 if (ret & VM_FAULT_MAJOR)
1353                                         tsk->maj_flt++;
1354                                 else
1355                                         tsk->min_flt++;
1356
1357                                 /*
1358                                  * The VM_FAULT_WRITE bit tells us that
1359                                  * do_wp_page has broken COW when necessary,
1360                                  * even if maybe_mkwrite decided not to set
1361                                  * pte_write. We can thus safely do subsequent
1362                                  * page lookups as if they were reads. But only
1363                                  * do so when looping for pte_write is futile:
1364                                  * in some cases userspace may also be wanting
1365                                  * to write to the gotten user page, which a
1366                                  * read fault here might prevent (a readonly
1367                                  * page might get reCOWed by userspace write).
1368                                  */
1369                                 if ((ret & VM_FAULT_WRITE) &&
1370                                     !(vma->vm_flags & VM_WRITE))
1371                                         foll_flags &= ~FOLL_WRITE;
1372
1373                                 cond_resched();
1374                         }
1375                         if (IS_ERR(page))
1376                                 return i ? i : PTR_ERR(page);
1377                         if (pages) {
1378                                 pages[i] = page;
1379
1380                                 flush_anon_page(vma, page, start);
1381                                 flush_dcache_page(page);
1382                         }
1383                         if (vmas)
1384                                 vmas[i] = vma;
1385                         i++;
1386                         start += PAGE_SIZE;
1387                         nr_pages--;
1388                 } while (nr_pages && start < vma->vm_end);
1389         } while (nr_pages);
1390         return i;
1391 }
1392
1393 /**
1394  * get_user_pages() - pin user pages in memory
1395  * @tsk:        task_struct of target task
1396  * @mm:         mm_struct of target mm
1397  * @start:      starting user address
1398  * @nr_pages:   number of pages from start to pin
1399  * @write:      whether pages will be written to by the caller
1400  * @force:      whether to force write access even if user mapping is
1401  *              readonly. This will result in the page being COWed even
1402  *              in MAP_SHARED mappings. You do not want this.
1403  * @pages:      array that receives pointers to the pages pinned.
1404  *              Should be at least nr_pages long. Or NULL, if caller
1405  *              only intends to ensure the pages are faulted in.
1406  * @vmas:       array of pointers to vmas corresponding to each page.
1407  *              Or NULL if the caller does not require them.
1408  *
1409  * Returns number of pages pinned. This may be fewer than the number
1410  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1411  * were pinned, returns -errno. Each page returned must be released
1412  * with a put_page() call when it is finished with. vmas will only
1413  * remain valid while mmap_sem is held.
1414  *
1415  * Must be called with mmap_sem held for read or write.
1416  *
1417  * get_user_pages walks a process's page tables and takes a reference to
1418  * each struct page that each user address corresponds to at a given
1419  * instant. That is, it takes the page that would be accessed if a user
1420  * thread accesses the given user virtual address at that instant.
1421  *
1422  * This does not guarantee that the page exists in the user mappings when
1423  * get_user_pages returns, and there may even be a completely different
1424  * page there in some cases (eg. if mmapped pagecache has been invalidated
1425  * and subsequently re faulted). However it does guarantee that the page
1426  * won't be freed completely. And mostly callers simply care that the page
1427  * contains data that was valid *at some point in time*. Typically, an IO
1428  * or similar operation cannot guarantee anything stronger anyway because
1429  * locks can't be held over the syscall boundary.
1430  *
1431  * If write=0, the page must not be written to. If the page is written to,
1432  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1433  * after the page is finished with, and before put_page is called.
1434  *
1435  * get_user_pages is typically used for fewer-copy IO operations, to get a
1436  * handle on the memory by some means other than accesses via the user virtual
1437  * addresses. The pages may be submitted for DMA to devices or accessed via
1438  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1439  * use the correct cache flushing APIs.
1440  *
1441  * See also get_user_pages_fast, for performance critical applications.
1442  */
1443 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1444                 unsigned long start, int nr_pages, int write, int force,
1445                 struct page **pages, struct vm_area_struct **vmas)
1446 {
1447         int flags = FOLL_TOUCH;
1448
1449         if (pages)
1450                 flags |= FOLL_GET;
1451         if (write)
1452                 flags |= FOLL_WRITE;
1453         if (force)
1454                 flags |= FOLL_FORCE;
1455
1456         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
1457 }
1458 EXPORT_SYMBOL(get_user_pages);
1459
1460 /**
1461  * get_dump_page() - pin user page in memory while writing it to core dump
1462  * @addr: user address
1463  *
1464  * Returns struct page pointer of user page pinned for dump,
1465  * to be freed afterwards by page_cache_release() or put_page().
1466  *
1467  * Returns NULL on any kind of failure - a hole must then be inserted into
1468  * the corefile, to preserve alignment with its headers; and also returns
1469  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1470  * allowing a hole to be left in the corefile to save diskspace.
1471  *
1472  * Called without mmap_sem, but after all other threads have been killed.
1473  */
1474 #ifdef CONFIG_ELF_CORE
1475 struct page *get_dump_page(unsigned long addr)
1476 {
1477         struct vm_area_struct *vma;
1478         struct page *page;
1479
1480         if (__get_user_pages(current, current->mm, addr, 1,
1481                         FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma) < 1)
1482                 return NULL;
1483         flush_cache_page(vma, addr, page_to_pfn(page));
1484         return page;
1485 }
1486 #endif /* CONFIG_ELF_CORE */
1487
1488 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1489                         spinlock_t **ptl)
1490 {
1491         pgd_t * pgd = pgd_offset(mm, addr);
1492         pud_t * pud = pud_alloc(mm, pgd, addr);
1493         if (pud) {
1494                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1495                 if (pmd)
1496                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1497         }
1498         return NULL;
1499 }
1500
1501 /*
1502  * This is the old fallback for page remapping.
1503  *
1504  * For historical reasons, it only allows reserved pages. Only
1505  * old drivers should use this, and they needed to mark their
1506  * pages reserved for the old functions anyway.
1507  */
1508 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1509                         struct page *page, pgprot_t prot)
1510 {
1511         struct mm_struct *mm = vma->vm_mm;
1512         int retval;
1513         pte_t *pte;
1514         spinlock_t *ptl;
1515
1516         retval = -EINVAL;
1517         if (PageAnon(page))
1518                 goto out;
1519         retval = -ENOMEM;
1520         flush_dcache_page(page);
1521         pte = get_locked_pte(mm, addr, &ptl);
1522         if (!pte)
1523                 goto out;
1524         retval = -EBUSY;
1525         if (!pte_none(*pte))
1526                 goto out_unlock;
1527
1528         /* Ok, finally just insert the thing.. */
1529         get_page(page);
1530         inc_mm_counter(mm, file_rss);
1531         page_add_file_rmap(page);
1532         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1533
1534         retval = 0;
1535         pte_unmap_unlock(pte, ptl);
1536         return retval;
1537 out_unlock:
1538         pte_unmap_unlock(pte, ptl);
1539 out:
1540         return retval;
1541 }
1542
1543 /**
1544  * vm_insert_page - insert single page into user vma
1545  * @vma: user vma to map to
1546  * @addr: target user address of this page
1547  * @page: source kernel page
1548  *
1549  * This allows drivers to insert individual pages they've allocated
1550  * into a user vma.
1551  *
1552  * The page has to be a nice clean _individual_ kernel allocation.
1553  * If you allocate a compound page, you need to have marked it as
1554  * such (__GFP_COMP), or manually just split the page up yourself
1555  * (see split_page()).
1556  *
1557  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1558  * took an arbitrary page protection parameter. This doesn't allow
1559  * that. Your vma protection will have to be set up correctly, which
1560  * means that if you want a shared writable mapping, you'd better
1561  * ask for a shared writable mapping!
1562  *
1563  * The page does not need to be reserved.
1564  */
1565 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1566                         struct page *page)
1567 {
1568         if (addr < vma->vm_start || addr >= vma->vm_end)
1569                 return -EFAULT;
1570         if (!page_count(page))
1571                 return -EINVAL;
1572         vma->vm_flags |= VM_INSERTPAGE;
1573         return insert_page(vma, addr, page, vma->vm_page_prot);
1574 }
1575 EXPORT_SYMBOL(vm_insert_page);
1576
1577 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1578                         unsigned long pfn, pgprot_t prot)
1579 {
1580         struct mm_struct *mm = vma->vm_mm;
1581         int retval;
1582         pte_t *pte, entry;
1583         spinlock_t *ptl;
1584
1585         retval = -ENOMEM;
1586         pte = get_locked_pte(mm, addr, &ptl);
1587         if (!pte)
1588                 goto out;
1589         retval = -EBUSY;
1590         if (!pte_none(*pte))
1591                 goto out_unlock;
1592
1593         /* Ok, finally just insert the thing.. */
1594         entry = pte_mkspecial(pfn_pte(pfn, prot));
1595         set_pte_at(mm, addr, pte, entry);
1596         update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1597
1598         retval = 0;
1599 out_unlock:
1600         pte_unmap_unlock(pte, ptl);
1601 out:
1602         return retval;
1603 }
1604
1605 /**
1606  * vm_insert_pfn - insert single pfn into user vma
1607  * @vma: user vma to map to
1608  * @addr: target user address of this page
1609  * @pfn: source kernel pfn
1610  *
1611  * Similar to vm_inert_page, this allows drivers to insert individual pages
1612  * they've allocated into a user vma. Same comments apply.
1613  *
1614  * This function should only be called from a vm_ops->fault handler, and
1615  * in that case the handler should return NULL.
1616  *
1617  * vma cannot be a COW mapping.
1618  *
1619  * As this is called only for pages that do not currently exist, we
1620  * do not need to flush old virtual caches or the TLB.
1621  */
1622 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1623                         unsigned long pfn)
1624 {
1625         int ret;
1626         pgprot_t pgprot = vma->vm_page_prot;
1627         /*
1628          * Technically, architectures with pte_special can avoid all these
1629          * restrictions (same for remap_pfn_range).  However we would like
1630          * consistency in testing and feature parity among all, so we should
1631          * try to keep these invariants in place for everybody.
1632          */
1633         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1634         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1635                                                 (VM_PFNMAP|VM_MIXEDMAP));
1636         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1637         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1638
1639         if (addr < vma->vm_start || addr >= vma->vm_end)
1640                 return -EFAULT;
1641         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1642                 return -EINVAL;
1643
1644         ret = insert_pfn(vma, addr, pfn, pgprot);
1645
1646         if (ret)
1647                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1648
1649         return ret;
1650 }
1651 EXPORT_SYMBOL(vm_insert_pfn);
1652
1653 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1654                         unsigned long pfn)
1655 {
1656         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1657
1658         if (addr < vma->vm_start || addr >= vma->vm_end)
1659                 return -EFAULT;
1660
1661         /*
1662          * If we don't have pte special, then we have to use the pfn_valid()
1663          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1664          * refcount the page if pfn_valid is true (hence insert_page rather
1665          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1666          * without pte special, it would there be refcounted as a normal page.
1667          */
1668         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1669                 struct page *page;
1670
1671                 page = pfn_to_page(pfn);
1672                 return insert_page(vma, addr, page, vma->vm_page_prot);
1673         }
1674         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1675 }
1676 EXPORT_SYMBOL(vm_insert_mixed);
1677
1678 /*
1679  * maps a range of physical memory into the requested pages. the old
1680  * mappings are removed. any references to nonexistent pages results
1681  * in null mappings (currently treated as "copy-on-access")
1682  */
1683 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1684                         unsigned long addr, unsigned long end,
1685                         unsigned long pfn, pgprot_t prot)
1686 {
1687         pte_t *pte;
1688         spinlock_t *ptl;
1689
1690         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1691         if (!pte)
1692                 return -ENOMEM;
1693         arch_enter_lazy_mmu_mode();
1694         do {
1695                 BUG_ON(!pte_none(*pte));
1696                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1697                 pfn++;
1698         } while (pte++, addr += PAGE_SIZE, addr != end);
1699         arch_leave_lazy_mmu_mode();
1700         pte_unmap_unlock(pte - 1, ptl);
1701         return 0;
1702 }
1703
1704 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1705                         unsigned long addr, unsigned long end,
1706                         unsigned long pfn, pgprot_t prot)
1707 {
1708         pmd_t *pmd;
1709         unsigned long next;
1710
1711         pfn -= addr >> PAGE_SHIFT;
1712         pmd = pmd_alloc(mm, pud, addr);
1713         if (!pmd)
1714                 return -ENOMEM;
1715         do {
1716                 next = pmd_addr_end(addr, end);
1717                 if (remap_pte_range(mm, pmd, addr, next,
1718                                 pfn + (addr >> PAGE_SHIFT), prot))
1719                         return -ENOMEM;
1720         } while (pmd++, addr = next, addr != end);
1721         return 0;
1722 }
1723
1724 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1725                         unsigned long addr, unsigned long end,
1726                         unsigned long pfn, pgprot_t prot)
1727 {
1728         pud_t *pud;
1729         unsigned long next;
1730
1731         pfn -= addr >> PAGE_SHIFT;
1732         pud = pud_alloc(mm, pgd, addr);
1733         if (!pud)
1734                 return -ENOMEM;
1735         do {
1736                 next = pud_addr_end(addr, end);
1737                 if (remap_pmd_range(mm, pud, addr, next,
1738                                 pfn + (addr >> PAGE_SHIFT), prot))
1739                         return -ENOMEM;
1740         } while (pud++, addr = next, addr != end);
1741         return 0;
1742 }
1743
1744 /**
1745  * remap_pfn_range - remap kernel memory to userspace
1746  * @vma: user vma to map to
1747  * @addr: target user address to start at
1748  * @pfn: physical address of kernel memory
1749  * @size: size of map area
1750  * @prot: page protection flags for this mapping
1751  *
1752  *  Note: this is only safe if the mm semaphore is held when called.
1753  */
1754 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1755                     unsigned long pfn, unsigned long size, pgprot_t prot)
1756 {
1757         pgd_t *pgd;
1758         unsigned long next;
1759         unsigned long end = addr + PAGE_ALIGN(size);
1760         struct mm_struct *mm = vma->vm_mm;
1761         int err;
1762
1763         /*
1764          * Physically remapped pages are special. Tell the
1765          * rest of the world about it:
1766          *   VM_IO tells people not to look at these pages
1767          *      (accesses can have side effects).
1768          *   VM_RESERVED is specified all over the place, because
1769          *      in 2.4 it kept swapout's vma scan off this vma; but
1770          *      in 2.6 the LRU scan won't even find its pages, so this
1771          *      flag means no more than count its pages in reserved_vm,
1772          *      and omit it from core dump, even when VM_IO turned off.
1773          *   VM_PFNMAP tells the core MM that the base pages are just
1774          *      raw PFN mappings, and do not have a "struct page" associated
1775          *      with them.
1776          *
1777          * There's a horrible special case to handle copy-on-write
1778          * behaviour that some programs depend on. We mark the "original"
1779          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1780          */
1781         if (addr == vma->vm_start && end == vma->vm_end) {
1782                 vma->vm_pgoff = pfn;
1783                 vma->vm_flags |= VM_PFN_AT_MMAP;
1784         } else if (is_cow_mapping(vma->vm_flags))
1785                 return -EINVAL;
1786
1787         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1788
1789         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1790         if (err) {
1791                 /*
1792                  * To indicate that track_pfn related cleanup is not
1793                  * needed from higher level routine calling unmap_vmas
1794                  */
1795                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1796                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1797                 return -EINVAL;
1798         }
1799
1800         BUG_ON(addr >= end);
1801         pfn -= addr >> PAGE_SHIFT;
1802         pgd = pgd_offset(mm, addr);
1803         flush_cache_range(vma, addr, end);
1804         do {
1805                 next = pgd_addr_end(addr, end);
1806                 err = remap_pud_range(mm, pgd, addr, next,
1807                                 pfn + (addr >> PAGE_SHIFT), prot);
1808                 if (err)
1809                         break;
1810         } while (pgd++, addr = next, addr != end);
1811
1812         if (err)
1813                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1814
1815         return err;
1816 }
1817 EXPORT_SYMBOL(remap_pfn_range);
1818
1819 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1820                                      unsigned long addr, unsigned long end,
1821                                      pte_fn_t fn, void *data)
1822 {
1823         pte_t *pte;
1824         int err;
1825         pgtable_t token;
1826         spinlock_t *uninitialized_var(ptl);
1827
1828         pte = (mm == &init_mm) ?
1829                 pte_alloc_kernel(pmd, addr) :
1830                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1831         if (!pte)
1832                 return -ENOMEM;
1833
1834         BUG_ON(pmd_huge(*pmd));
1835
1836         arch_enter_lazy_mmu_mode();
1837
1838         token = pmd_pgtable(*pmd);
1839
1840         do {
1841                 err = fn(pte++, token, addr, data);
1842                 if (err)
1843                         break;
1844         } while (addr += PAGE_SIZE, addr != end);
1845
1846         arch_leave_lazy_mmu_mode();
1847
1848         if (mm != &init_mm)
1849                 pte_unmap_unlock(pte-1, ptl);
1850         return err;
1851 }
1852
1853 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1854                                      unsigned long addr, unsigned long end,
1855                                      pte_fn_t fn, void *data)
1856 {
1857         pmd_t *pmd;
1858         unsigned long next;
1859         int err;
1860
1861         BUG_ON(pud_huge(*pud));
1862
1863         pmd = pmd_alloc(mm, pud, addr);
1864         if (!pmd)
1865                 return -ENOMEM;
1866         do {
1867                 next = pmd_addr_end(addr, end);
1868                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1869                 if (err)
1870                         break;
1871         } while (pmd++, addr = next, addr != end);
1872         return err;
1873 }
1874
1875 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1876                                      unsigned long addr, unsigned long end,
1877                                      pte_fn_t fn, void *data)
1878 {
1879         pud_t *pud;
1880         unsigned long next;
1881         int err;
1882
1883         pud = pud_alloc(mm, pgd, addr);
1884         if (!pud)
1885                 return -ENOMEM;
1886         do {
1887                 next = pud_addr_end(addr, end);
1888                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1889                 if (err)
1890                         break;
1891         } while (pud++, addr = next, addr != end);
1892         return err;
1893 }
1894
1895 /*
1896  * Scan a region of virtual memory, filling in page tables as necessary
1897  * and calling a provided function on each leaf page table.
1898  */
1899 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1900                         unsigned long size, pte_fn_t fn, void *data)
1901 {
1902         pgd_t *pgd;
1903         unsigned long next;
1904         unsigned long start = addr, end = addr + size;
1905         int err;
1906
1907         BUG_ON(addr >= end);
1908         mmu_notifier_invalidate_range_start(mm, start, end);
1909         pgd = pgd_offset(mm, addr);
1910         do {
1911                 next = pgd_addr_end(addr, end);
1912                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1913                 if (err)
1914                         break;
1915         } while (pgd++, addr = next, addr != end);
1916         mmu_notifier_invalidate_range_end(mm, start, end);
1917         return err;
1918 }
1919 EXPORT_SYMBOL_GPL(apply_to_page_range);
1920
1921 /*
1922  * handle_pte_fault chooses page fault handler according to an entry
1923  * which was read non-atomically.  Before making any commitment, on
1924  * those architectures or configurations (e.g. i386 with PAE) which
1925  * might give a mix of unmatched parts, do_swap_page and do_file_page
1926  * must check under lock before unmapping the pte and proceeding
1927  * (but do_wp_page is only called after already making such a check;
1928  * and do_anonymous_page and do_no_page can safely check later on).
1929  */
1930 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1931                                 pte_t *page_table, pte_t orig_pte)
1932 {
1933         int same = 1;
1934 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1935         if (sizeof(pte_t) > sizeof(unsigned long)) {
1936                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1937                 spin_lock(ptl);
1938                 same = pte_same(*page_table, orig_pte);
1939                 spin_unlock(ptl);
1940         }
1941 #endif
1942         pte_unmap(page_table);
1943         return same;
1944 }
1945
1946 /*
1947  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1948  * servicing faults for write access.  In the normal case, do always want
1949  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1950  * that do not have writing enabled, when used by access_process_vm.
1951  */
1952 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1953 {
1954         if (likely(vma->vm_flags & VM_WRITE))
1955                 pte = pte_mkwrite(pte);
1956         return pte;
1957 }
1958
1959 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1960 {
1961         /*
1962          * If the source page was a PFN mapping, we don't have
1963          * a "struct page" for it. We do a best-effort copy by
1964          * just copying from the original user address. If that
1965          * fails, we just zero-fill it. Live with it.
1966          */
1967         if (unlikely(!src)) {
1968                 void *kaddr = kmap_atomic(dst, KM_USER0);
1969                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1970
1971                 /*
1972                  * This really shouldn't fail, because the page is there
1973                  * in the page tables. But it might just be unreadable,
1974                  * in which case we just give up and fill the result with
1975                  * zeroes.
1976                  */
1977                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1978                         memset(kaddr, 0, PAGE_SIZE);
1979                 kunmap_atomic(kaddr, KM_USER0);
1980                 flush_dcache_page(dst);
1981         } else
1982                 copy_user_highpage(dst, src, va, vma);
1983 }
1984
1985 /*
1986  * This routine handles present pages, when users try to write
1987  * to a shared page. It is done by copying the page to a new address
1988  * and decrementing the shared-page counter for the old page.
1989  *
1990  * Note that this routine assumes that the protection checks have been
1991  * done by the caller (the low-level page fault routine in most cases).
1992  * Thus we can safely just mark it writable once we've done any necessary
1993  * COW.
1994  *
1995  * We also mark the page dirty at this point even though the page will
1996  * change only once the write actually happens. This avoids a few races,
1997  * and potentially makes it more efficient.
1998  *
1999  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2000  * but allow concurrent faults), with pte both mapped and locked.
2001  * We return with mmap_sem still held, but pte unmapped and unlocked.
2002  */
2003 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2004                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2005                 spinlock_t *ptl, pte_t orig_pte)
2006 {
2007         struct page *old_page, *new_page;
2008         pte_t entry;
2009         int reuse = 0, ret = 0;
2010         int page_mkwrite = 0;
2011         struct page *dirty_page = NULL;
2012
2013         old_page = vm_normal_page(vma, address, orig_pte);
2014         if (!old_page) {
2015                 /*
2016                  * VM_MIXEDMAP !pfn_valid() case
2017                  *
2018                  * We should not cow pages in a shared writeable mapping.
2019                  * Just mark the pages writable as we can't do any dirty
2020                  * accounting on raw pfn maps.
2021                  */
2022                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2023                                      (VM_WRITE|VM_SHARED))
2024                         goto reuse;
2025                 goto gotten;
2026         }
2027
2028         /*
2029          * Take out anonymous pages first, anonymous shared vmas are
2030          * not dirty accountable.
2031          */
2032         if (PageAnon(old_page) && !PageKsm(old_page)) {
2033                 if (!trylock_page(old_page)) {
2034                         page_cache_get(old_page);
2035                         pte_unmap_unlock(page_table, ptl);
2036                         lock_page(old_page);
2037                         page_table = pte_offset_map_lock(mm, pmd, address,
2038                                                          &ptl);
2039                         if (!pte_same(*page_table, orig_pte)) {
2040                                 unlock_page(old_page);
2041                                 page_cache_release(old_page);
2042                                 goto unlock;
2043                         }
2044                         page_cache_release(old_page);
2045                 }
2046                 reuse = reuse_swap_page(old_page);
2047                 unlock_page(old_page);
2048         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2049                                         (VM_WRITE|VM_SHARED))) {
2050                 /*
2051                  * Only catch write-faults on shared writable pages,
2052                  * read-only shared pages can get COWed by
2053                  * get_user_pages(.write=1, .force=1).
2054                  */
2055                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2056                         struct vm_fault vmf;
2057                         int tmp;
2058
2059                         vmf.virtual_address = (void __user *)(address &
2060                                                                 PAGE_MASK);
2061                         vmf.pgoff = old_page->index;
2062                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2063                         vmf.page = old_page;
2064
2065                         /*
2066                          * Notify the address space that the page is about to
2067                          * become writable so that it can prohibit this or wait
2068                          * for the page to get into an appropriate state.
2069                          *
2070                          * We do this without the lock held, so that it can
2071                          * sleep if it needs to.
2072                          */
2073                         page_cache_get(old_page);
2074                         pte_unmap_unlock(page_table, ptl);
2075
2076                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2077                         if (unlikely(tmp &
2078                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2079                                 ret = tmp;
2080                                 goto unwritable_page;
2081                         }
2082                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2083                                 lock_page(old_page);
2084                                 if (!old_page->mapping) {
2085                                         ret = 0; /* retry the fault */
2086                                         unlock_page(old_page);
2087                                         goto unwritable_page;
2088                                 }
2089                         } else
2090                                 VM_BUG_ON(!PageLocked(old_page));
2091
2092                         /*
2093                          * Since we dropped the lock we need to revalidate
2094                          * the PTE as someone else may have changed it.  If
2095                          * they did, we just return, as we can count on the
2096                          * MMU to tell us if they didn't also make it writable.
2097                          */
2098                         page_table = pte_offset_map_lock(mm, pmd, address,
2099                                                          &ptl);
2100                         if (!pte_same(*page_table, orig_pte)) {
2101                                 unlock_page(old_page);
2102                                 page_cache_release(old_page);
2103                                 goto unlock;
2104                         }
2105
2106                         page_mkwrite = 1;
2107                 }
2108                 dirty_page = old_page;
2109                 get_page(dirty_page);
2110                 reuse = 1;
2111         }
2112
2113         if (reuse) {
2114 reuse:
2115                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2116                 entry = pte_mkyoung(orig_pte);
2117                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2118                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2119                         update_mmu_cache(vma, address, entry);
2120                 ret |= VM_FAULT_WRITE;
2121                 goto unlock;
2122         }
2123
2124         /*
2125          * Ok, we need to copy. Oh, well..
2126          */
2127         page_cache_get(old_page);
2128 gotten:
2129         pte_unmap_unlock(page_table, ptl);
2130
2131         if (unlikely(anon_vma_prepare(vma)))
2132                 goto oom;
2133
2134         if (is_zero_pfn(pte_pfn(orig_pte))) {
2135                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2136                 if (!new_page)
2137                         goto oom;
2138         } else {
2139                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2140                 if (!new_page)
2141                         goto oom;
2142                 cow_user_page(new_page, old_page, address, vma);
2143         }
2144         __SetPageUptodate(new_page);
2145
2146         /*
2147          * Don't let another task, with possibly unlocked vma,
2148          * keep the mlocked page.
2149          */
2150         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2151                 lock_page(old_page);    /* for LRU manipulation */
2152                 clear_page_mlock(old_page);
2153                 unlock_page(old_page);
2154         }
2155
2156         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2157                 goto oom_free_new;
2158
2159         /*
2160          * Re-check the pte - we dropped the lock
2161          */
2162         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2163         if (likely(pte_same(*page_table, orig_pte))) {
2164                 if (old_page) {
2165                         if (!PageAnon(old_page)) {
2166                                 dec_mm_counter(mm, file_rss);
2167                                 inc_mm_counter(mm, anon_rss);
2168                         }
2169                 } else
2170                         inc_mm_counter(mm, anon_rss);
2171                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2172                 entry = mk_pte(new_page, vma->vm_page_prot);
2173                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2174                 /*
2175                  * Clear the pte entry and flush it first, before updating the
2176                  * pte with the new entry. This will avoid a race condition
2177                  * seen in the presence of one thread doing SMC and another
2178                  * thread doing COW.
2179                  */
2180                 ptep_clear_flush(vma, address, page_table);
2181                 page_add_new_anon_rmap(new_page, vma, address);
2182                 /*
2183                  * We call the notify macro here because, when using secondary
2184                  * mmu page tables (such as kvm shadow page tables), we want the
2185                  * new page to be mapped directly into the secondary page table.
2186                  */
2187                 set_pte_at_notify(mm, address, page_table, entry);
2188                 update_mmu_cache(vma, address, entry);
2189                 if (old_page) {
2190                         /*
2191                          * Only after switching the pte to the new page may
2192                          * we remove the mapcount here. Otherwise another
2193                          * process may come and find the rmap count decremented
2194                          * before the pte is switched to the new page, and
2195                          * "reuse" the old page writing into it while our pte
2196                          * here still points into it and can be read by other
2197                          * threads.
2198                          *
2199                          * The critical issue is to order this
2200                          * page_remove_rmap with the ptp_clear_flush above.
2201                          * Those stores are ordered by (if nothing else,)
2202                          * the barrier present in the atomic_add_negative
2203                          * in page_remove_rmap.
2204                          *
2205                          * Then the TLB flush in ptep_clear_flush ensures that
2206                          * no process can access the old page before the
2207                          * decremented mapcount is visible. And the old page
2208                          * cannot be reused until after the decremented
2209                          * mapcount is visible. So transitively, TLBs to
2210                          * old page will be flushed before it can be reused.
2211                          */
2212                         page_remove_rmap(old_page);
2213                 }
2214
2215                 /* Free the old page.. */
2216                 new_page = old_page;
2217                 ret |= VM_FAULT_WRITE;
2218         } else
2219                 mem_cgroup_uncharge_page(new_page);
2220
2221         if (new_page)
2222                 page_cache_release(new_page);
2223         if (old_page)
2224                 page_cache_release(old_page);
2225 unlock:
2226         pte_unmap_unlock(page_table, ptl);
2227         if (dirty_page) {
2228                 /*
2229                  * Yes, Virginia, this is actually required to prevent a race
2230                  * with clear_page_dirty_for_io() from clearing the page dirty
2231                  * bit after it clear all dirty ptes, but before a racing
2232                  * do_wp_page installs a dirty pte.
2233                  *
2234                  * do_no_page is protected similarly.
2235                  */
2236                 if (!page_mkwrite) {
2237                         wait_on_page_locked(dirty_page);
2238                         set_page_dirty_balance(dirty_page, page_mkwrite);
2239                 }
2240                 put_page(dirty_page);
2241                 if (page_mkwrite) {
2242                         struct address_space *mapping = dirty_page->mapping;
2243
2244                         set_page_dirty(dirty_page);
2245                         unlock_page(dirty_page);
2246                         page_cache_release(dirty_page);
2247                         if (mapping)    {
2248                                 /*
2249                                  * Some device drivers do not set page.mapping
2250                                  * but still dirty their pages
2251                                  */
2252                                 balance_dirty_pages_ratelimited(mapping);
2253                         }
2254                 }
2255
2256                 /* file_update_time outside page_lock */
2257                 if (vma->vm_file)
2258                         file_update_time(vma->vm_file);
2259         }
2260         return ret;
2261 oom_free_new:
2262         page_cache_release(new_page);
2263 oom:
2264         if (old_page) {
2265                 if (page_mkwrite) {
2266                         unlock_page(old_page);
2267                         page_cache_release(old_page);
2268                 }
2269                 page_cache_release(old_page);
2270         }
2271         return VM_FAULT_OOM;
2272
2273 unwritable_page:
2274         page_cache_release(old_page);
2275         return ret;
2276 }
2277
2278 /*
2279  * Helper functions for unmap_mapping_range().
2280  *
2281  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2282  *
2283  * We have to restart searching the prio_tree whenever we drop the lock,
2284  * since the iterator is only valid while the lock is held, and anyway
2285  * a later vma might be split and reinserted earlier while lock dropped.
2286  *
2287  * The list of nonlinear vmas could be handled more efficiently, using
2288  * a placeholder, but handle it in the same way until a need is shown.
2289  * It is important to search the prio_tree before nonlinear list: a vma
2290  * may become nonlinear and be shifted from prio_tree to nonlinear list
2291  * while the lock is dropped; but never shifted from list to prio_tree.
2292  *
2293  * In order to make forward progress despite restarting the search,
2294  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2295  * quickly skip it next time around.  Since the prio_tree search only
2296  * shows us those vmas affected by unmapping the range in question, we
2297  * can't efficiently keep all vmas in step with mapping->truncate_count:
2298  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2299  * mapping->truncate_count and vma->vm_truncate_count are protected by
2300  * i_mmap_lock.
2301  *
2302  * In order to make forward progress despite repeatedly restarting some
2303  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2304  * and restart from that address when we reach that vma again.  It might
2305  * have been split or merged, shrunk or extended, but never shifted: so
2306  * restart_addr remains valid so long as it remains in the vma's range.
2307  * unmap_mapping_range forces truncate_count to leap over page-aligned
2308  * values so we can save vma's restart_addr in its truncate_count field.
2309  */
2310 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2311
2312 static void reset_vma_truncate_counts(struct address_space *mapping)
2313 {
2314         struct vm_area_struct *vma;
2315         struct prio_tree_iter iter;
2316
2317         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2318                 vma->vm_truncate_count = 0;
2319         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2320                 vma->vm_truncate_count = 0;
2321 }
2322
2323 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2324                 unsigned long start_addr, unsigned long end_addr,
2325                 struct zap_details *details)
2326 {
2327         unsigned long restart_addr;
2328         int need_break;
2329
2330         /*
2331          * files that support invalidating or truncating portions of the
2332          * file from under mmaped areas must have their ->fault function
2333          * return a locked page (and set VM_FAULT_LOCKED in the return).
2334          * This provides synchronisation against concurrent unmapping here.
2335          */
2336
2337 again:
2338         restart_addr = vma->vm_truncate_count;
2339         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2340                 start_addr = restart_addr;
2341                 if (start_addr >= end_addr) {
2342                         /* Top of vma has been split off since last time */
2343                         vma->vm_truncate_count = details->truncate_count;
2344                         return 0;
2345                 }
2346         }
2347
2348         restart_addr = zap_page_range(vma, start_addr,
2349                                         end_addr - start_addr, details);
2350         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2351
2352         if (restart_addr >= end_addr) {
2353                 /* We have now completed this vma: mark it so */
2354                 vma->vm_truncate_count = details->truncate_count;
2355                 if (!need_break)
2356                         return 0;
2357         } else {
2358                 /* Note restart_addr in vma's truncate_count field */
2359                 vma->vm_truncate_count = restart_addr;
2360                 if (!need_break)
2361                         goto again;
2362         }
2363
2364         spin_unlock(details->i_mmap_lock);
2365         cond_resched();
2366         spin_lock(details->i_mmap_lock);
2367         return -EINTR;
2368 }
2369
2370 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2371                                             struct zap_details *details)
2372 {
2373         struct vm_area_struct *vma;
2374         struct prio_tree_iter iter;
2375         pgoff_t vba, vea, zba, zea;
2376
2377 restart:
2378         vma_prio_tree_foreach(vma, &iter, root,
2379                         details->first_index, details->last_index) {
2380                 /* Skip quickly over those we have already dealt with */
2381                 if (vma->vm_truncate_count == details->truncate_count)
2382                         continue;
2383
2384                 vba = vma->vm_pgoff;
2385                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2386                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2387                 zba = details->first_index;
2388                 if (zba < vba)
2389                         zba = vba;
2390                 zea = details->last_index;
2391                 if (zea > vea)
2392                         zea = vea;
2393
2394                 if (unmap_mapping_range_vma(vma,
2395                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2396                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2397                                 details) < 0)
2398                         goto restart;
2399         }
2400 }
2401
2402 static inline void unmap_mapping_range_list(struct list_head *head,
2403                                             struct zap_details *details)
2404 {
2405         struct vm_area_struct *vma;
2406
2407         /*
2408          * In nonlinear VMAs there is no correspondence between virtual address
2409          * offset and file offset.  So we must perform an exhaustive search
2410          * across *all* the pages in each nonlinear VMA, not just the pages
2411          * whose virtual address lies outside the file truncation point.
2412          */
2413 restart:
2414         list_for_each_entry(vma, head, shared.vm_set.list) {
2415                 /* Skip quickly over those we have already dealt with */
2416                 if (vma->vm_truncate_count == details->truncate_count)
2417                         continue;
2418                 details->nonlinear_vma = vma;
2419                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2420                                         vma->vm_end, details) < 0)
2421                         goto restart;
2422         }
2423 }
2424
2425 /**
2426  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2427  * @mapping: the address space containing mmaps to be unmapped.
2428  * @holebegin: byte in first page to unmap, relative to the start of
2429  * the underlying file.  This will be rounded down to a PAGE_SIZE
2430  * boundary.  Note that this is different from truncate_pagecache(), which
2431  * must keep the partial page.  In contrast, we must get rid of
2432  * partial pages.
2433  * @holelen: size of prospective hole in bytes.  This will be rounded
2434  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2435  * end of the file.
2436  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2437  * but 0 when invalidating pagecache, don't throw away private data.
2438  */
2439 void unmap_mapping_range(struct address_space *mapping,
2440                 loff_t const holebegin, loff_t const holelen, int even_cows)
2441 {
2442         struct zap_details details;
2443         pgoff_t hba = holebegin >> PAGE_SHIFT;
2444         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2445
2446         /* Check for overflow. */
2447         if (sizeof(holelen) > sizeof(hlen)) {
2448                 long long holeend =
2449                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2450                 if (holeend & ~(long long)ULONG_MAX)
2451                         hlen = ULONG_MAX - hba + 1;
2452         }
2453
2454         details.check_mapping = even_cows? NULL: mapping;
2455         details.nonlinear_vma = NULL;
2456         details.first_index = hba;
2457         details.last_index = hba + hlen - 1;
2458         if (details.last_index < details.first_index)
2459                 details.last_index = ULONG_MAX;
2460         details.i_mmap_lock = &mapping->i_mmap_lock;
2461
2462         spin_lock(&mapping->i_mmap_lock);
2463
2464         /* Protect against endless unmapping loops */
2465         mapping->truncate_count++;
2466         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2467                 if (mapping->truncate_count == 0)
2468                         reset_vma_truncate_counts(mapping);
2469                 mapping->truncate_count++;
2470         }
2471         details.truncate_count = mapping->truncate_count;
2472
2473         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2474                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2475         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2476                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2477         spin_unlock(&mapping->i_mmap_lock);
2478 }
2479 EXPORT_SYMBOL(unmap_mapping_range);
2480
2481 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2482 {
2483         struct address_space *mapping = inode->i_mapping;
2484
2485         /*
2486          * If the underlying filesystem is not going to provide
2487          * a way to truncate a range of blocks (punch a hole) -
2488          * we should return failure right now.
2489          */
2490         if (!inode->i_op->truncate_range)
2491                 return -ENOSYS;
2492
2493         mutex_lock(&inode->i_mutex);
2494         down_write(&inode->i_alloc_sem);
2495         unmap_mapping_range(mapping, offset, (end - offset), 1);
2496         truncate_inode_pages_range(mapping, offset, end);
2497         unmap_mapping_range(mapping, offset, (end - offset), 1);
2498         inode->i_op->truncate_range(inode, offset, end);
2499         up_write(&inode->i_alloc_sem);
2500         mutex_unlock(&inode->i_mutex);
2501
2502         return 0;
2503 }
2504
2505 /*
2506  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2507  * but allow concurrent faults), and pte mapped but not yet locked.
2508  * We return with mmap_sem still held, but pte unmapped and unlocked.
2509  */
2510 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2511                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2512                 unsigned int flags, pte_t orig_pte)
2513 {
2514         spinlock_t *ptl;
2515         struct page *page;
2516         swp_entry_t entry;
2517         pte_t pte;
2518         struct mem_cgroup *ptr = NULL;
2519         int ret = 0;
2520
2521         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2522                 goto out;
2523
2524         entry = pte_to_swp_entry(orig_pte);
2525         if (unlikely(non_swap_entry(entry))) {
2526                 if (is_migration_entry(entry)) {
2527                         migration_entry_wait(mm, pmd, address);
2528                 } else if (is_hwpoison_entry(entry)) {
2529                         ret = VM_FAULT_HWPOISON;
2530                 } else {
2531                         print_bad_pte(vma, address, orig_pte, NULL);
2532                         ret = VM_FAULT_SIGBUS;
2533                 }
2534                 goto out;
2535         }
2536         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2537         page = lookup_swap_cache(entry);
2538         if (!page) {
2539                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2540                 page = swapin_readahead(entry,
2541                                         GFP_HIGHUSER_MOVABLE, vma, address);
2542                 if (!page) {
2543                         /*
2544                          * Back out if somebody else faulted in this pte
2545                          * while we released the pte lock.
2546                          */
2547                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2548                         if (likely(pte_same(*page_table, orig_pte)))
2549                                 ret = VM_FAULT_OOM;
2550                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2551                         goto unlock;
2552                 }
2553
2554                 /* Had to read the page from swap area: Major fault */
2555                 ret = VM_FAULT_MAJOR;
2556                 count_vm_event(PGMAJFAULT);
2557         } else if (PageHWPoison(page)) {
2558                 /*
2559                  * hwpoisoned dirty swapcache pages are kept for killing
2560                  * owner processes (which may be unknown at hwpoison time)
2561                  */
2562                 ret = VM_FAULT_HWPOISON;
2563                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2564                 goto out_release;
2565         }
2566
2567         lock_page(page);
2568         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2569
2570         page = ksm_might_need_to_copy(page, vma, address);
2571         if (!page) {
2572                 ret = VM_FAULT_OOM;
2573                 goto out;
2574         }
2575
2576         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2577                 ret = VM_FAULT_OOM;
2578                 goto out_page;
2579         }
2580
2581         /*
2582          * Back out if somebody else already faulted in this pte.
2583          */
2584         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2585         if (unlikely(!pte_same(*page_table, orig_pte)))
2586                 goto out_nomap;
2587
2588         if (unlikely(!PageUptodate(page))) {
2589                 ret = VM_FAULT_SIGBUS;
2590                 goto out_nomap;
2591         }
2592
2593         /*
2594          * The page isn't present yet, go ahead with the fault.
2595          *
2596          * Be careful about the sequence of operations here.
2597          * To get its accounting right, reuse_swap_page() must be called
2598          * while the page is counted on swap but not yet in mapcount i.e.
2599          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2600          * must be called after the swap_free(), or it will never succeed.
2601          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2602          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2603          * in page->private. In this case, a record in swap_cgroup  is silently
2604          * discarded at swap_free().
2605          */
2606
2607         inc_mm_counter(mm, anon_rss);
2608         pte = mk_pte(page, vma->vm_page_prot);
2609         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2610                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2611                 flags &= ~FAULT_FLAG_WRITE;
2612         }
2613         flush_icache_page(vma, page);
2614         set_pte_at(mm, address, page_table, pte);
2615         page_add_anon_rmap(page, vma, address);
2616         /* It's better to call commit-charge after rmap is established */
2617         mem_cgroup_commit_charge_swapin(page, ptr);
2618
2619         swap_free(entry);
2620         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2621                 try_to_free_swap(page);
2622         unlock_page(page);
2623
2624         if (flags & FAULT_FLAG_WRITE) {
2625                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2626                 if (ret & VM_FAULT_ERROR)
2627                         ret &= VM_FAULT_ERROR;
2628                 goto out;
2629         }
2630
2631         /* No need to invalidate - it was non-present before */
2632         update_mmu_cache(vma, address, pte);
2633 unlock:
2634         pte_unmap_unlock(page_table, ptl);
2635 out:
2636         return ret;
2637 out_nomap:
2638         mem_cgroup_cancel_charge_swapin(ptr);
2639         pte_unmap_unlock(page_table, ptl);
2640 out_page:
2641         unlock_page(page);
2642 out_release:
2643         page_cache_release(page);
2644         return ret;
2645 }
2646
2647 /*
2648  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2649  * but allow concurrent faults), and pte mapped but not yet locked.
2650  * We return with mmap_sem still held, but pte unmapped and unlocked.
2651  */
2652 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2653                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2654                 unsigned int flags)
2655 {
2656         struct page *page;
2657         spinlock_t *ptl;
2658         pte_t entry;
2659
2660         if (!(flags & FAULT_FLAG_WRITE)) {
2661                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2662                                                 vma->vm_page_prot));
2663                 ptl = pte_lockptr(mm, pmd);
2664                 spin_lock(ptl);
2665                 if (!pte_none(*page_table))
2666                         goto unlock;
2667                 goto setpte;
2668         }
2669
2670         /* Allocate our own private page. */
2671         pte_unmap(page_table);
2672
2673         if (unlikely(anon_vma_prepare(vma)))
2674                 goto oom;
2675         page = alloc_zeroed_user_highpage_movable(vma, address);
2676         if (!page)
2677                 goto oom;
2678         __SetPageUptodate(page);
2679
2680         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2681                 goto oom_free_page;
2682
2683         entry = mk_pte(page, vma->vm_page_prot);
2684         if (vma->vm_flags & VM_WRITE)
2685                 entry = pte_mkwrite(pte_mkdirty(entry));
2686
2687         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2688         if (!pte_none(*page_table))
2689                 goto release;
2690
2691         inc_mm_counter(mm, anon_rss);
2692         page_add_new_anon_rmap(page, vma, address);
2693 setpte:
2694         set_pte_at(mm, address, page_table, entry);
2695
2696         /* No need to invalidate - it was non-present before */
2697         update_mmu_cache(vma, address, entry);
2698 unlock:
2699         pte_unmap_unlock(page_table, ptl);
2700         return 0;
2701 release:
2702         mem_cgroup_uncharge_page(page);
2703         page_cache_release(page);
2704         goto unlock;
2705 oom_free_page:
2706         page_cache_release(page);
2707 oom:
2708         return VM_FAULT_OOM;
2709 }
2710
2711 /*
2712  * __do_fault() tries to create a new page mapping. It aggressively
2713  * tries to share with existing pages, but makes a separate copy if
2714  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2715  * the next page fault.
2716  *
2717  * As this is called only for pages that do not currently exist, we
2718  * do not need to flush old virtual caches or the TLB.
2719  *
2720  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2721  * but allow concurrent faults), and pte neither mapped nor locked.
2722  * We return with mmap_sem still held, but pte unmapped and unlocked.
2723  */
2724 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2725                 unsigned long address, pmd_t *pmd,
2726                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2727 {
2728         pte_t *page_table;
2729         spinlock_t *ptl;
2730         struct page *page;
2731         pte_t entry;
2732         int anon = 0;
2733         int charged = 0;
2734         struct page *dirty_page = NULL;
2735         struct vm_fault vmf;
2736         int ret;
2737         int page_mkwrite = 0;
2738
2739         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2740         vmf.pgoff = pgoff;
2741         vmf.flags = flags;
2742         vmf.page = NULL;
2743
2744         ret = vma->vm_ops->fault(vma, &vmf);
2745         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2746                 return ret;
2747
2748         if (unlikely(PageHWPoison(vmf.page))) {
2749                 if (ret & VM_FAULT_LOCKED)
2750                         unlock_page(vmf.page);
2751                 return VM_FAULT_HWPOISON;
2752         }
2753
2754         /*
2755          * For consistency in subsequent calls, make the faulted page always
2756          * locked.
2757          */
2758         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2759                 lock_page(vmf.page);
2760         else
2761                 VM_BUG_ON(!PageLocked(vmf.page));
2762
2763         /*
2764          * Should we do an early C-O-W break?
2765          */
2766         page = vmf.page;
2767         if (flags & FAULT_FLAG_WRITE) {
2768                 if (!(vma->vm_flags & VM_SHARED)) {
2769                         anon = 1;
2770                         if (unlikely(anon_vma_prepare(vma))) {
2771                                 ret = VM_FAULT_OOM;
2772                                 goto out;
2773                         }
2774                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2775                                                 vma, address);
2776                         if (!page) {
2777                                 ret = VM_FAULT_OOM;
2778                                 goto out;
2779                         }
2780                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2781                                 ret = VM_FAULT_OOM;
2782                                 page_cache_release(page);
2783                                 goto out;
2784                         }
2785                         charged = 1;
2786                         /*
2787                          * Don't let another task, with possibly unlocked vma,
2788                          * keep the mlocked page.
2789                          */
2790                         if (vma->vm_flags & VM_LOCKED)
2791                                 clear_page_mlock(vmf.page);
2792                         copy_user_highpage(page, vmf.page, address, vma);
2793                         __SetPageUptodate(page);
2794                 } else {
2795                         /*
2796                          * If the page will be shareable, see if the backing
2797                          * address space wants to know that the page is about
2798                          * to become writable
2799                          */
2800                         if (vma->vm_ops->page_mkwrite) {
2801                                 int tmp;
2802
2803                                 unlock_page(page);
2804                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2805                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2806                                 if (unlikely(tmp &
2807                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2808                                         ret = tmp;
2809                                         goto unwritable_page;
2810                                 }
2811                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2812                                         lock_page(page);
2813                                         if (!page->mapping) {
2814                                                 ret = 0; /* retry the fault */
2815                                                 unlock_page(page);
2816                                                 goto unwritable_page;
2817                                         }
2818                                 } else
2819                                         VM_BUG_ON(!PageLocked(page));
2820                                 page_mkwrite = 1;
2821                         }
2822                 }
2823
2824         }
2825
2826         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2827
2828         /*
2829          * This silly early PAGE_DIRTY setting removes a race
2830          * due to the bad i386 page protection. But it's valid
2831          * for other architectures too.
2832          *
2833          * Note that if FAULT_FLAG_WRITE is set, we either now have
2834          * an exclusive copy of the page, or this is a shared mapping,
2835          * so we can make it writable and dirty to avoid having to
2836          * handle that later.
2837          */
2838         /* Only go through if we didn't race with anybody else... */
2839         if (likely(pte_same(*page_table, orig_pte))) {
2840                 flush_icache_page(vma, page);
2841                 entry = mk_pte(page, vma->vm_page_prot);
2842                 if (flags & FAULT_FLAG_WRITE)
2843                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2844                 if (anon) {
2845                         inc_mm_counter(mm, anon_rss);
2846                         page_add_new_anon_rmap(page, vma, address);
2847                 } else {
2848                         inc_mm_counter(mm, file_rss);
2849                         page_add_file_rmap(page);
2850                         if (flags & FAULT_FLAG_WRITE) {
2851                                 dirty_page = page;
2852                                 get_page(dirty_page);
2853                         }
2854                 }
2855                 set_pte_at(mm, address, page_table, entry);
2856
2857                 /* no need to invalidate: a not-present page won't be cached */
2858                 update_mmu_cache(vma, address, entry);
2859         } else {
2860                 if (charged)
2861                         mem_cgroup_uncharge_page(page);
2862                 if (anon)
2863                         page_cache_release(page);
2864                 else
2865                         anon = 1; /* no anon but release faulted_page */
2866         }
2867
2868         pte_unmap_unlock(page_table, ptl);
2869
2870 out:
2871         if (dirty_page) {
2872                 struct address_space *mapping = page->mapping;
2873
2874                 if (set_page_dirty(dirty_page))
2875                         page_mkwrite = 1;
2876                 unlock_page(dirty_page);
2877                 put_page(dirty_page);
2878                 if (page_mkwrite && mapping) {
2879                         /*
2880                          * Some device drivers do not set page.mapping but still
2881                          * dirty their pages
2882                          */
2883                         balance_dirty_pages_ratelimited(mapping);
2884                 }
2885
2886                 /* file_update_time outside page_lock */
2887                 if (vma->vm_file)
2888                         file_update_time(vma->vm_file);
2889         } else {
2890                 unlock_page(vmf.page);
2891                 if (anon)
2892                         page_cache_release(vmf.page);
2893         }
2894
2895         return ret;
2896
2897 unwritable_page:
2898         page_cache_release(page);
2899         return ret;
2900 }
2901
2902 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2903                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2904                 unsigned int flags, pte_t orig_pte)
2905 {
2906         pgoff_t pgoff = (((address & PAGE_MASK)
2907                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2908
2909         pte_unmap(page_table);
2910         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2911 }
2912
2913 /*
2914  * Fault of a previously existing named mapping. Repopulate the pte
2915  * from the encoded file_pte if possible. This enables swappable
2916  * nonlinear vmas.
2917  *
2918  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2919  * but allow concurrent faults), and pte mapped but not yet locked.
2920  * We return with mmap_sem still held, but pte unmapped and unlocked.
2921  */
2922 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2923                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2924                 unsigned int flags, pte_t orig_pte)
2925 {
2926         pgoff_t pgoff;
2927
2928         flags |= FAULT_FLAG_NONLINEAR;
2929
2930         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2931                 return 0;
2932
2933         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2934                 /*
2935                  * Page table corrupted: show pte and kill process.
2936                  */
2937                 print_bad_pte(vma, address, orig_pte, NULL);
2938                 return VM_FAULT_SIGBUS;
2939         }
2940
2941         pgoff = pte_to_pgoff(orig_pte);
2942         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2943 }
2944
2945 /*
2946  * These routines also need to handle stuff like marking pages dirty
2947  * and/or accessed for architectures that don't do it in hardware (most
2948  * RISC architectures).  The early dirtying is also good on the i386.
2949  *
2950  * There is also a hook called "update_mmu_cache()" that architectures
2951  * with external mmu caches can use to update those (ie the Sparc or
2952  * PowerPC hashed page tables that act as extended TLBs).
2953  *
2954  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2955  * but allow concurrent faults), and pte mapped but not yet locked.
2956  * We return with mmap_sem still held, but pte unmapped and unlocked.
2957  */
2958 static inline int handle_pte_fault(struct mm_struct *mm,
2959                 struct vm_area_struct *vma, unsigned long address,
2960                 pte_t *pte, pmd_t *pmd, unsigned int flags)
2961 {
2962         pte_t entry;
2963         spinlock_t *ptl;
2964
2965         entry = *pte;
2966         if (!pte_present(entry)) {
2967                 if (pte_none(entry)) {
2968                         if (vma->vm_ops) {
2969                                 if (likely(vma->vm_ops->fault))
2970                                         return do_linear_fault(mm, vma, address,
2971                                                 pte, pmd, flags, entry);
2972                         }
2973                         return do_anonymous_page(mm, vma, address,
2974                                                  pte, pmd, flags);
2975                 }
2976                 if (pte_file(entry))
2977                         return do_nonlinear_fault(mm, vma, address,
2978                                         pte, pmd, flags, entry);
2979                 return do_swap_page(mm, vma, address,
2980                                         pte, pmd, flags, entry);
2981         }
2982
2983         ptl = pte_lockptr(mm, pmd);
2984         spin_lock(ptl);
2985         if (unlikely(!pte_same(*pte, entry)))
2986                 goto unlock;
2987         if (flags & FAULT_FLAG_WRITE) {
2988                 if (!pte_write(entry))
2989                         return do_wp_page(mm, vma, address,
2990                                         pte, pmd, ptl, entry);
2991                 entry = pte_mkdirty(entry);
2992         }
2993         entry = pte_mkyoung(entry);
2994         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
2995                 update_mmu_cache(vma, address, entry);
2996         } else {
2997                 /*
2998                  * This is needed only for protection faults but the arch code
2999                  * is not yet telling us if this is a protection fault or not.
3000                  * This still avoids useless tlb flushes for .text page faults
3001                  * with threads.
3002                  */
3003                 if (flags & FAULT_FLAG_WRITE)
3004                         flush_tlb_page(vma, address);
3005         }
3006 unlock:
3007         pte_unmap_unlock(pte, ptl);
3008         return 0;
3009 }
3010
3011 /*
3012  * By the time we get here, we already hold the mm semaphore
3013  */
3014 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3015                 unsigned long address, unsigned int flags)
3016 {
3017         pgd_t *pgd;
3018         pud_t *pud;
3019         pmd_t *pmd;
3020         pte_t *pte;
3021
3022         __set_current_state(TASK_RUNNING);
3023
3024         count_vm_event(PGFAULT);
3025
3026         if (unlikely(is_vm_hugetlb_page(vma)))
3027                 return hugetlb_fault(mm, vma, address, flags);
3028
3029         pgd = pgd_offset(mm, address);
3030         pud = pud_alloc(mm, pgd, address);
3031         if (!pud)
3032                 return VM_FAULT_OOM;
3033         pmd = pmd_alloc(mm, pud, address);
3034         if (!pmd)
3035                 return VM_FAULT_OOM;
3036         pte = pte_alloc_map(mm, pmd, address);
3037         if (!pte)
3038                 return VM_FAULT_OOM;
3039
3040         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3041 }
3042
3043 #ifndef __PAGETABLE_PUD_FOLDED
3044 /*
3045  * Allocate page upper directory.
3046  * We've already handled the fast-path in-line.
3047  */
3048 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3049 {
3050         pud_t *new = pud_alloc_one(mm, address);
3051         if (!new)
3052                 return -ENOMEM;
3053
3054         smp_wmb(); /* See comment in __pte_alloc */
3055
3056         spin_lock(&mm->page_table_lock);
3057         if (pgd_present(*pgd))          /* Another has populated it */
3058                 pud_free(mm, new);
3059         else
3060                 pgd_populate(mm, pgd, new);
3061         spin_unlock(&mm->page_table_lock);
3062         return 0;
3063 }
3064 #endif /* __PAGETABLE_PUD_FOLDED */
3065
3066 #ifndef __PAGETABLE_PMD_FOLDED
3067 /*
3068  * Allocate page middle directory.
3069  * We've already handled the fast-path in-line.
3070  */
3071 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3072 {
3073         pmd_t *new = pmd_alloc_one(mm, address);
3074         if (!new)
3075                 return -ENOMEM;
3076
3077         smp_wmb(); /* See comment in __pte_alloc */
3078
3079         spin_lock(&mm->page_table_lock);
3080 #ifndef __ARCH_HAS_4LEVEL_HACK
3081         if (pud_present(*pud))          /* Another has populated it */
3082                 pmd_free(mm, new);
3083         else
3084                 pud_populate(mm, pud, new);
3085 #else
3086         if (pgd_present(*pud))          /* Another has populated it */
3087                 pmd_free(mm, new);
3088         else
3089                 pgd_populate(mm, pud, new);
3090 #endif /* __ARCH_HAS_4LEVEL_HACK */
3091         spin_unlock(&mm->page_table_lock);
3092         return 0;
3093 }
3094 #endif /* __PAGETABLE_PMD_FOLDED */
3095
3096 int make_pages_present(unsigned long addr, unsigned long end)
3097 {
3098         int ret, len, write;
3099         struct vm_area_struct * vma;
3100
3101         vma = find_vma(current->mm, addr);
3102         if (!vma)
3103                 return -ENOMEM;
3104         write = (vma->vm_flags & VM_WRITE) != 0;
3105         BUG_ON(addr >= end);
3106         BUG_ON(end > vma->vm_end);
3107         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3108         ret = get_user_pages(current, current->mm, addr,
3109                         len, write, 0, NULL, NULL);
3110         if (ret < 0)
3111                 return ret;
3112         return ret == len ? 0 : -EFAULT;
3113 }
3114
3115 #if !defined(__HAVE_ARCH_GATE_AREA)
3116
3117 #if defined(AT_SYSINFO_EHDR)
3118 static struct vm_area_struct gate_vma;
3119
3120 static int __init gate_vma_init(void)
3121 {
3122         gate_vma.vm_mm = NULL;
3123         gate_vma.vm_start = FIXADDR_USER_START;
3124         gate_vma.vm_end = FIXADDR_USER_END;
3125         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3126         gate_vma.vm_page_prot = __P101;
3127         /*
3128          * Make sure the vDSO gets into every core dump.
3129          * Dumping its contents makes post-mortem fully interpretable later
3130          * without matching up the same kernel and hardware config to see
3131          * what PC values meant.
3132          */
3133         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3134         return 0;
3135 }
3136 __initcall(gate_vma_init);
3137 #endif
3138
3139 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3140 {
3141 #ifdef AT_SYSINFO_EHDR
3142         return &gate_vma;
3143 #else
3144         return NULL;
3145 #endif
3146 }
3147
3148 int in_gate_area_no_task(unsigned long addr)
3149 {
3150 #ifdef AT_SYSINFO_EHDR
3151         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3152                 return 1;
3153 #endif
3154         return 0;
3155 }
3156
3157 #endif  /* __HAVE_ARCH_GATE_AREA */
3158
3159 static int follow_pte(struct mm_struct *mm, unsigned long address,
3160                 pte_t **ptepp, spinlock_t **ptlp)
3161 {
3162         pgd_t *pgd;
3163         pud_t *pud;
3164         pmd_t *pmd;
3165         pte_t *ptep;
3166
3167         pgd = pgd_offset(mm, address);
3168         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3169                 goto out;
3170
3171         pud = pud_offset(pgd, address);
3172         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3173                 goto out;
3174
3175         pmd = pmd_offset(pud, address);
3176         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3177                 goto out;
3178
3179         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3180         if (pmd_huge(*pmd))
3181                 goto out;
3182
3183         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3184         if (!ptep)
3185                 goto out;
3186         if (!pte_present(*ptep))
3187                 goto unlock;
3188         *ptepp = ptep;
3189         return 0;
3190 unlock:
3191         pte_unmap_unlock(ptep, *ptlp);
3192 out:
3193         return -EINVAL;
3194 }
3195
3196 /**
3197  * follow_pfn - look up PFN at a user virtual address
3198  * @vma: memory mapping
3199  * @address: user virtual address
3200  * @pfn: location to store found PFN
3201  *
3202  * Only IO mappings and raw PFN mappings are allowed.
3203  *
3204  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3205  */
3206 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3207         unsigned long *pfn)
3208 {
3209         int ret = -EINVAL;
3210         spinlock_t *ptl;
3211         pte_t *ptep;
3212
3213         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3214                 return ret;
3215
3216         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3217         if (ret)
3218                 return ret;
3219         *pfn = pte_pfn(*ptep);
3220         pte_unmap_unlock(ptep, ptl);
3221         return 0;
3222 }
3223 EXPORT_SYMBOL(follow_pfn);
3224
3225 #ifdef CONFIG_HAVE_IOREMAP_PROT
3226 int follow_phys(struct vm_area_struct *vma,
3227                 unsigned long address, unsigned int flags,
3228                 unsigned long *prot, resource_size_t *phys)
3229 {
3230         int ret = -EINVAL;
3231         pte_t *ptep, pte;
3232         spinlock_t *ptl;
3233
3234         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3235                 goto out;
3236
3237         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3238                 goto out;
3239         pte = *ptep;
3240
3241         if ((flags & FOLL_WRITE) && !pte_write(pte))
3242                 goto unlock;
3243
3244         *prot = pgprot_val(pte_pgprot(pte));
3245         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3246
3247         ret = 0;
3248 unlock:
3249         pte_unmap_unlock(ptep, ptl);
3250 out:
3251         return ret;
3252 }
3253
3254 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3255                         void *buf, int len, int write)
3256 {
3257         resource_size_t phys_addr;
3258         unsigned long prot = 0;
3259         void __iomem *maddr;
3260         int offset = addr & (PAGE_SIZE-1);
3261
3262         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3263                 return -EINVAL;
3264
3265         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3266         if (write)
3267                 memcpy_toio(maddr + offset, buf, len);
3268         else
3269                 memcpy_fromio(buf, maddr + offset, len);
3270         iounmap(maddr);
3271
3272         return len;
3273 }
3274 #endif
3275
3276 /*
3277  * Access another process' address space.
3278  * Source/target buffer must be kernel space,
3279  * Do not walk the page table directly, use get_user_pages
3280  */
3281 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3282 {
3283         struct mm_struct *mm;
3284         struct vm_area_struct *vma;
3285         void *old_buf = buf;
3286
3287         mm = get_task_mm(tsk);
3288         if (!mm)
3289                 return 0;
3290
3291         down_read(&mm->mmap_sem);
3292         /* ignore errors, just check how much was successfully transferred */
3293         while (len) {
3294                 int bytes, ret, offset;
3295                 void *maddr;
3296                 struct page *page = NULL;
3297
3298                 ret = get_user_pages(tsk, mm, addr, 1,
3299                                 write, 1, &page, &vma);
3300                 if (ret <= 0) {
3301                         /*
3302                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3303                          * we can access using slightly different code.
3304                          */
3305 #ifdef CONFIG_HAVE_IOREMAP_PROT
3306                         vma = find_vma(mm, addr);
3307                         if (!vma)
3308                                 break;
3309                         if (vma->vm_ops && vma->vm_ops->access)
3310                                 ret = vma->vm_ops->access(vma, addr, buf,
3311                                                           len, write);
3312                         if (ret <= 0)
3313 #endif
3314                                 break;
3315                         bytes = ret;
3316                 } else {
3317                         bytes = len;
3318                         offset = addr & (PAGE_SIZE-1);
3319                         if (bytes > PAGE_SIZE-offset)
3320                                 bytes = PAGE_SIZE-offset;
3321
3322                         maddr = kmap(page);
3323                         if (write) {
3324                                 copy_to_user_page(vma, page, addr,
3325                                                   maddr + offset, buf, bytes);
3326                                 set_page_dirty_lock(page);
3327                         } else {
3328                                 copy_from_user_page(vma, page, addr,
3329                                                     buf, maddr + offset, bytes);
3330                         }
3331                         kunmap(page);
3332                         page_cache_release(page);
3333                 }
3334                 len -= bytes;
3335                 buf += bytes;
3336                 addr += bytes;
3337         }
3338         up_read(&mm->mmap_sem);
3339         mmput(mm);
3340
3341         return buf - old_buf;
3342 }
3343
3344 /*
3345  * Print the name of a VMA.
3346  */
3347 void print_vma_addr(char *prefix, unsigned long ip)
3348 {
3349         struct mm_struct *mm = current->mm;
3350         struct vm_area_struct *vma;
3351
3352         /*
3353          * Do not print if we are in atomic
3354          * contexts (in exception stacks, etc.):
3355          */
3356         if (preempt_count())
3357                 return;
3358
3359         down_read(&mm->mmap_sem);
3360         vma = find_vma(mm, ip);
3361         if (vma && vma->vm_file) {
3362                 struct file *f = vma->vm_file;
3363                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3364                 if (buf) {
3365                         char *p, *s;
3366
3367                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3368                         if (IS_ERR(p))
3369                                 p = "?";
3370                         s = strrchr(p, '/');
3371                         if (s)
3372                                 p = s+1;
3373                         printk("%s%s[%lx+%lx]", prefix, p,
3374                                         vma->vm_start,
3375                                         vma->vm_end - vma->vm_start);
3376                         free_page((unsigned long)buf);
3377                 }
3378         }
3379         up_read(&current->mm->mmap_sem);
3380 }
3381
3382 #ifdef CONFIG_PROVE_LOCKING
3383 void might_fault(void)
3384 {
3385         /*
3386          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3387          * holding the mmap_sem, this is safe because kernel memory doesn't
3388          * get paged out, therefore we'll never actually fault, and the
3389          * below annotations will generate false positives.
3390          */
3391         if (segment_eq(get_fs(), KERNEL_DS))
3392                 return;
3393
3394         might_sleep();
3395         /*
3396          * it would be nicer only to annotate paths which are not under
3397          * pagefault_disable, however that requires a larger audit and
3398          * providing helpers like get_user_atomic.
3399          */
3400         if (!in_atomic() && current->mm)
3401                 might_lock_read(&current->mm->mmap_sem);
3402 }
3403 EXPORT_SYMBOL(might_fault);
3404 #endif