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