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