mlock: do not hold mmap_sem for extended periods of time
[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 #include <linux/gfp.h>
60
61 #include <asm/io.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
64 #include <asm/tlb.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
67
68 #include "internal.h"
69
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
73 struct page *mem_map;
74
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
77 #endif
78
79 unsigned long num_physpages;
80 /*
81  * A number of key systems in x86 including ioremap() rely on the assumption
82  * that high_memory defines the upper bound on direct map memory, then end
83  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
84  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85  * and ZONE_HIGHMEM.
86  */
87 void * high_memory;
88
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
91
92 /*
93  * Randomize the address space (stacks, mmaps, brk, etc.).
94  *
95  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96  *   as ancient (libc5 based) binaries can segfault. )
97  */
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
100                                         1;
101 #else
102                                         2;
103 #endif
104
105 static int __init disable_randmaps(char *s)
106 {
107         randomize_va_space = 0;
108         return 1;
109 }
110 __setup("norandmaps", disable_randmaps);
111
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
114
115 /*
116  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117  */
118 static int __init init_zero_pfn(void)
119 {
120         zero_pfn = page_to_pfn(ZERO_PAGE(0));
121         return 0;
122 }
123 core_initcall(init_zero_pfn);
124
125
126 #if defined(SPLIT_RSS_COUNTING)
127
128 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
129 {
130         int i;
131
132         for (i = 0; i < NR_MM_COUNTERS; i++) {
133                 if (task->rss_stat.count[i]) {
134                         add_mm_counter(mm, i, task->rss_stat.count[i]);
135                         task->rss_stat.count[i] = 0;
136                 }
137         }
138         task->rss_stat.events = 0;
139 }
140
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142 {
143         struct task_struct *task = current;
144
145         if (likely(task->mm == mm))
146                 task->rss_stat.count[member] += val;
147         else
148                 add_mm_counter(mm, member, val);
149 }
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH  (64)
155 static void check_sync_rss_stat(struct task_struct *task)
156 {
157         if (unlikely(task != current))
158                 return;
159         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160                 __sync_task_rss_stat(task, task->mm);
161 }
162
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
164 {
165         long val = 0;
166
167         /*
168          * Don't use task->mm here...for avoiding to use task_get_mm()..
169          * The caller must guarantee task->mm is not invalid.
170          */
171         val = atomic_long_read(&mm->rss_stat.count[member]);
172         /*
173          * counter is updated in asynchronous manner and may go to minus.
174          * But it's never be expected number for users.
175          */
176         if (val < 0)
177                 return 0;
178         return (unsigned long)val;
179 }
180
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182 {
183         __sync_task_rss_stat(task, mm);
184 }
185 #else
186
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
189
190 static void check_sync_rss_stat(struct task_struct *task)
191 {
192 }
193
194 #endif
195
196 /*
197  * If a p?d_bad entry is found while walking page tables, report
198  * the error, before resetting entry to p?d_none.  Usually (but
199  * very seldom) called out from the p?d_none_or_clear_bad macros.
200  */
201
202 void pgd_clear_bad(pgd_t *pgd)
203 {
204         pgd_ERROR(*pgd);
205         pgd_clear(pgd);
206 }
207
208 void pud_clear_bad(pud_t *pud)
209 {
210         pud_ERROR(*pud);
211         pud_clear(pud);
212 }
213
214 void pmd_clear_bad(pmd_t *pmd)
215 {
216         pmd_ERROR(*pmd);
217         pmd_clear(pmd);
218 }
219
220 /*
221  * Note: this doesn't free the actual pages themselves. That
222  * has been handled earlier when unmapping all the memory regions.
223  */
224 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
225                            unsigned long addr)
226 {
227         pgtable_t token = pmd_pgtable(*pmd);
228         pmd_clear(pmd);
229         pte_free_tlb(tlb, token, addr);
230         tlb->mm->nr_ptes--;
231 }
232
233 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
234                                 unsigned long addr, unsigned long end,
235                                 unsigned long floor, unsigned long ceiling)
236 {
237         pmd_t *pmd;
238         unsigned long next;
239         unsigned long start;
240
241         start = addr;
242         pmd = pmd_offset(pud, addr);
243         do {
244                 next = pmd_addr_end(addr, end);
245                 if (pmd_none_or_clear_bad(pmd))
246                         continue;
247                 free_pte_range(tlb, pmd, addr);
248         } while (pmd++, addr = next, addr != end);
249
250         start &= PUD_MASK;
251         if (start < floor)
252                 return;
253         if (ceiling) {
254                 ceiling &= PUD_MASK;
255                 if (!ceiling)
256                         return;
257         }
258         if (end - 1 > ceiling - 1)
259                 return;
260
261         pmd = pmd_offset(pud, start);
262         pud_clear(pud);
263         pmd_free_tlb(tlb, pmd, start);
264 }
265
266 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
267                                 unsigned long addr, unsigned long end,
268                                 unsigned long floor, unsigned long ceiling)
269 {
270         pud_t *pud;
271         unsigned long next;
272         unsigned long start;
273
274         start = addr;
275         pud = pud_offset(pgd, addr);
276         do {
277                 next = pud_addr_end(addr, end);
278                 if (pud_none_or_clear_bad(pud))
279                         continue;
280                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
281         } while (pud++, addr = next, addr != end);
282
283         start &= PGDIR_MASK;
284         if (start < floor)
285                 return;
286         if (ceiling) {
287                 ceiling &= PGDIR_MASK;
288                 if (!ceiling)
289                         return;
290         }
291         if (end - 1 > ceiling - 1)
292                 return;
293
294         pud = pud_offset(pgd, start);
295         pgd_clear(pgd);
296         pud_free_tlb(tlb, pud, start);
297 }
298
299 /*
300  * This function frees user-level page tables of a process.
301  *
302  * Must be called with pagetable lock held.
303  */
304 void free_pgd_range(struct mmu_gather *tlb,
305                         unsigned long addr, unsigned long end,
306                         unsigned long floor, unsigned long ceiling)
307 {
308         pgd_t *pgd;
309         unsigned long next;
310
311         /*
312          * The next few lines have given us lots of grief...
313          *
314          * Why are we testing PMD* at this top level?  Because often
315          * there will be no work to do at all, and we'd prefer not to
316          * go all the way down to the bottom just to discover that.
317          *
318          * Why all these "- 1"s?  Because 0 represents both the bottom
319          * of the address space and the top of it (using -1 for the
320          * top wouldn't help much: the masks would do the wrong thing).
321          * The rule is that addr 0 and floor 0 refer to the bottom of
322          * the address space, but end 0 and ceiling 0 refer to the top
323          * Comparisons need to use "end - 1" and "ceiling - 1" (though
324          * that end 0 case should be mythical).
325          *
326          * Wherever addr is brought up or ceiling brought down, we must
327          * be careful to reject "the opposite 0" before it confuses the
328          * subsequent tests.  But what about where end is brought down
329          * by PMD_SIZE below? no, end can't go down to 0 there.
330          *
331          * Whereas we round start (addr) and ceiling down, by different
332          * masks at different levels, in order to test whether a table
333          * now has no other vmas using it, so can be freed, we don't
334          * bother to round floor or end up - the tests don't need that.
335          */
336
337         addr &= PMD_MASK;
338         if (addr < floor) {
339                 addr += PMD_SIZE;
340                 if (!addr)
341                         return;
342         }
343         if (ceiling) {
344                 ceiling &= PMD_MASK;
345                 if (!ceiling)
346                         return;
347         }
348         if (end - 1 > ceiling - 1)
349                 end -= PMD_SIZE;
350         if (addr > end - 1)
351                 return;
352
353         pgd = pgd_offset(tlb->mm, addr);
354         do {
355                 next = pgd_addr_end(addr, end);
356                 if (pgd_none_or_clear_bad(pgd))
357                         continue;
358                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
359         } while (pgd++, addr = next, addr != end);
360 }
361
362 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
363                 unsigned long floor, unsigned long ceiling)
364 {
365         while (vma) {
366                 struct vm_area_struct *next = vma->vm_next;
367                 unsigned long addr = vma->vm_start;
368
369                 /*
370                  * Hide vma from rmap and truncate_pagecache before freeing
371                  * pgtables
372                  */
373                 unlink_anon_vmas(vma);
374                 unlink_file_vma(vma);
375
376                 if (is_vm_hugetlb_page(vma)) {
377                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
378                                 floor, next? next->vm_start: ceiling);
379                 } else {
380                         /*
381                          * Optimization: gather nearby vmas into one call down
382                          */
383                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
384                                && !is_vm_hugetlb_page(next)) {
385                                 vma = next;
386                                 next = vma->vm_next;
387                                 unlink_anon_vmas(vma);
388                                 unlink_file_vma(vma);
389                         }
390                         free_pgd_range(tlb, addr, vma->vm_end,
391                                 floor, next? next->vm_start: ceiling);
392                 }
393                 vma = next;
394         }
395 }
396
397 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
398 {
399         pgtable_t new = pte_alloc_one(mm, address);
400         if (!new)
401                 return -ENOMEM;
402
403         /*
404          * Ensure all pte setup (eg. pte page lock and page clearing) are
405          * visible before the pte is made visible to other CPUs by being
406          * put into page tables.
407          *
408          * The other side of the story is the pointer chasing in the page
409          * table walking code (when walking the page table without locking;
410          * ie. most of the time). Fortunately, these data accesses consist
411          * of a chain of data-dependent loads, meaning most CPUs (alpha
412          * being the notable exception) will already guarantee loads are
413          * seen in-order. See the alpha page table accessors for the
414          * smp_read_barrier_depends() barriers in page table walking code.
415          */
416         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
417
418         spin_lock(&mm->page_table_lock);
419         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
420                 mm->nr_ptes++;
421                 pmd_populate(mm, pmd, new);
422                 new = NULL;
423         }
424         spin_unlock(&mm->page_table_lock);
425         if (new)
426                 pte_free(mm, new);
427         return 0;
428 }
429
430 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
431 {
432         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
433         if (!new)
434                 return -ENOMEM;
435
436         smp_wmb(); /* See comment in __pte_alloc */
437
438         spin_lock(&init_mm.page_table_lock);
439         if (!pmd_present(*pmd)) {       /* Has another populated it ? */
440                 pmd_populate_kernel(&init_mm, pmd, new);
441                 new = NULL;
442         }
443         spin_unlock(&init_mm.page_table_lock);
444         if (new)
445                 pte_free_kernel(&init_mm, new);
446         return 0;
447 }
448
449 static inline void init_rss_vec(int *rss)
450 {
451         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
452 }
453
454 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
455 {
456         int i;
457
458         if (current->mm == mm)
459                 sync_mm_rss(current, mm);
460         for (i = 0; i < NR_MM_COUNTERS; i++)
461                 if (rss[i])
462                         add_mm_counter(mm, i, rss[i]);
463 }
464
465 /*
466  * This function is called to print an error when a bad pte
467  * is found. For example, we might have a PFN-mapped pte in
468  * a region that doesn't allow it.
469  *
470  * The calling function must still handle the error.
471  */
472 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
473                           pte_t pte, struct page *page)
474 {
475         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
476         pud_t *pud = pud_offset(pgd, addr);
477         pmd_t *pmd = pmd_offset(pud, addr);
478         struct address_space *mapping;
479         pgoff_t index;
480         static unsigned long resume;
481         static unsigned long nr_shown;
482         static unsigned long nr_unshown;
483
484         /*
485          * Allow a burst of 60 reports, then keep quiet for that minute;
486          * or allow a steady drip of one report per second.
487          */
488         if (nr_shown == 60) {
489                 if (time_before(jiffies, resume)) {
490                         nr_unshown++;
491                         return;
492                 }
493                 if (nr_unshown) {
494                         printk(KERN_ALERT
495                                 "BUG: Bad page map: %lu messages suppressed\n",
496                                 nr_unshown);
497                         nr_unshown = 0;
498                 }
499                 nr_shown = 0;
500         }
501         if (nr_shown++ == 0)
502                 resume = jiffies + 60 * HZ;
503
504         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
505         index = linear_page_index(vma, addr);
506
507         printk(KERN_ALERT
508                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
509                 current->comm,
510                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
511         if (page)
512                 dump_page(page);
513         printk(KERN_ALERT
514                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
515                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
516         /*
517          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
518          */
519         if (vma->vm_ops)
520                 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
521                                 (unsigned long)vma->vm_ops->fault);
522         if (vma->vm_file && vma->vm_file->f_op)
523                 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
524                                 (unsigned long)vma->vm_file->f_op->mmap);
525         dump_stack();
526         add_taint(TAINT_BAD_PAGE);
527 }
528
529 static inline int is_cow_mapping(unsigned int flags)
530 {
531         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
532 }
533
534 #ifndef is_zero_pfn
535 static inline int is_zero_pfn(unsigned long pfn)
536 {
537         return pfn == zero_pfn;
538 }
539 #endif
540
541 #ifndef my_zero_pfn
542 static inline unsigned long my_zero_pfn(unsigned long addr)
543 {
544         return zero_pfn;
545 }
546 #endif
547
548 /*
549  * vm_normal_page -- This function gets the "struct page" associated with a pte.
550  *
551  * "Special" mappings do not wish to be associated with a "struct page" (either
552  * it doesn't exist, or it exists but they don't want to touch it). In this
553  * case, NULL is returned here. "Normal" mappings do have a struct page.
554  *
555  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
556  * pte bit, in which case this function is trivial. Secondly, an architecture
557  * may not have a spare pte bit, which requires a more complicated scheme,
558  * described below.
559  *
560  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
561  * special mapping (even if there are underlying and valid "struct pages").
562  * COWed pages of a VM_PFNMAP are always normal.
563  *
564  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
565  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
566  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
567  * mapping will always honor the rule
568  *
569  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
570  *
571  * And for normal mappings this is false.
572  *
573  * This restricts such mappings to be a linear translation from virtual address
574  * to pfn. To get around this restriction, we allow arbitrary mappings so long
575  * as the vma is not a COW mapping; in that case, we know that all ptes are
576  * special (because none can have been COWed).
577  *
578  *
579  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
580  *
581  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
582  * page" backing, however the difference is that _all_ pages with a struct
583  * page (that is, those where pfn_valid is true) are refcounted and considered
584  * normal pages by the VM. The disadvantage is that pages are refcounted
585  * (which can be slower and simply not an option for some PFNMAP users). The
586  * advantage is that we don't have to follow the strict linearity rule of
587  * PFNMAP mappings in order to support COWable mappings.
588  *
589  */
590 #ifdef __HAVE_ARCH_PTE_SPECIAL
591 # define HAVE_PTE_SPECIAL 1
592 #else
593 # define HAVE_PTE_SPECIAL 0
594 #endif
595 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
596                                 pte_t pte)
597 {
598         unsigned long pfn = pte_pfn(pte);
599
600         if (HAVE_PTE_SPECIAL) {
601                 if (likely(!pte_special(pte)))
602                         goto check_pfn;
603                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
604                         return NULL;
605                 if (!is_zero_pfn(pfn))
606                         print_bad_pte(vma, addr, pte, NULL);
607                 return NULL;
608         }
609
610         /* !HAVE_PTE_SPECIAL case follows: */
611
612         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
613                 if (vma->vm_flags & VM_MIXEDMAP) {
614                         if (!pfn_valid(pfn))
615                                 return NULL;
616                         goto out;
617                 } else {
618                         unsigned long off;
619                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
620                         if (pfn == vma->vm_pgoff + off)
621                                 return NULL;
622                         if (!is_cow_mapping(vma->vm_flags))
623                                 return NULL;
624                 }
625         }
626
627         if (is_zero_pfn(pfn))
628                 return NULL;
629 check_pfn:
630         if (unlikely(pfn > highest_memmap_pfn)) {
631                 print_bad_pte(vma, addr, pte, NULL);
632                 return NULL;
633         }
634
635         /*
636          * NOTE! We still have PageReserved() pages in the page tables.
637          * eg. VDSO mappings can cause them to exist.
638          */
639 out:
640         return pfn_to_page(pfn);
641 }
642
643 /*
644  * copy one vm_area from one task to the other. Assumes the page tables
645  * already present in the new task to be cleared in the whole range
646  * covered by this vma.
647  */
648
649 static inline unsigned long
650 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
651                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
652                 unsigned long addr, int *rss)
653 {
654         unsigned long vm_flags = vma->vm_flags;
655         pte_t pte = *src_pte;
656         struct page *page;
657
658         /* pte contains position in swap or file, so copy. */
659         if (unlikely(!pte_present(pte))) {
660                 if (!pte_file(pte)) {
661                         swp_entry_t entry = pte_to_swp_entry(pte);
662
663                         if (swap_duplicate(entry) < 0)
664                                 return entry.val;
665
666                         /* make sure dst_mm is on swapoff's mmlist. */
667                         if (unlikely(list_empty(&dst_mm->mmlist))) {
668                                 spin_lock(&mmlist_lock);
669                                 if (list_empty(&dst_mm->mmlist))
670                                         list_add(&dst_mm->mmlist,
671                                                  &src_mm->mmlist);
672                                 spin_unlock(&mmlist_lock);
673                         }
674                         if (likely(!non_swap_entry(entry)))
675                                 rss[MM_SWAPENTS]++;
676                         else if (is_write_migration_entry(entry) &&
677                                         is_cow_mapping(vm_flags)) {
678                                 /*
679                                  * COW mappings require pages in both parent
680                                  * and child to be set to read.
681                                  */
682                                 make_migration_entry_read(&entry);
683                                 pte = swp_entry_to_pte(entry);
684                                 set_pte_at(src_mm, addr, src_pte, pte);
685                         }
686                 }
687                 goto out_set_pte;
688         }
689
690         /*
691          * If it's a COW mapping, write protect it both
692          * in the parent and the child
693          */
694         if (is_cow_mapping(vm_flags)) {
695                 ptep_set_wrprotect(src_mm, addr, src_pte);
696                 pte = pte_wrprotect(pte);
697         }
698
699         /*
700          * If it's a shared mapping, mark it clean in
701          * the child
702          */
703         if (vm_flags & VM_SHARED)
704                 pte = pte_mkclean(pte);
705         pte = pte_mkold(pte);
706
707         page = vm_normal_page(vma, addr, pte);
708         if (page) {
709                 get_page(page);
710                 page_dup_rmap(page);
711                 if (PageAnon(page))
712                         rss[MM_ANONPAGES]++;
713                 else
714                         rss[MM_FILEPAGES]++;
715         }
716
717 out_set_pte:
718         set_pte_at(dst_mm, addr, dst_pte, pte);
719         return 0;
720 }
721
722 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
723                 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
724                 unsigned long addr, unsigned long end)
725 {
726         pte_t *orig_src_pte, *orig_dst_pte;
727         pte_t *src_pte, *dst_pte;
728         spinlock_t *src_ptl, *dst_ptl;
729         int progress = 0;
730         int rss[NR_MM_COUNTERS];
731         swp_entry_t entry = (swp_entry_t){0};
732
733 again:
734         init_rss_vec(rss);
735
736         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
737         if (!dst_pte)
738                 return -ENOMEM;
739         src_pte = pte_offset_map(src_pmd, addr);
740         src_ptl = pte_lockptr(src_mm, src_pmd);
741         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
742         orig_src_pte = src_pte;
743         orig_dst_pte = dst_pte;
744         arch_enter_lazy_mmu_mode();
745
746         do {
747                 /*
748                  * We are holding two locks at this point - either of them
749                  * could generate latencies in another task on another CPU.
750                  */
751                 if (progress >= 32) {
752                         progress = 0;
753                         if (need_resched() ||
754                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
755                                 break;
756                 }
757                 if (pte_none(*src_pte)) {
758                         progress++;
759                         continue;
760                 }
761                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
762                                                         vma, addr, rss);
763                 if (entry.val)
764                         break;
765                 progress += 8;
766         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
767
768         arch_leave_lazy_mmu_mode();
769         spin_unlock(src_ptl);
770         pte_unmap(orig_src_pte);
771         add_mm_rss_vec(dst_mm, rss);
772         pte_unmap_unlock(orig_dst_pte, dst_ptl);
773         cond_resched();
774
775         if (entry.val) {
776                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
777                         return -ENOMEM;
778                 progress = 0;
779         }
780         if (addr != end)
781                 goto again;
782         return 0;
783 }
784
785 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
786                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
787                 unsigned long addr, unsigned long end)
788 {
789         pmd_t *src_pmd, *dst_pmd;
790         unsigned long next;
791
792         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
793         if (!dst_pmd)
794                 return -ENOMEM;
795         src_pmd = pmd_offset(src_pud, addr);
796         do {
797                 next = pmd_addr_end(addr, end);
798                 if (pmd_none_or_clear_bad(src_pmd))
799                         continue;
800                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
801                                                 vma, addr, next))
802                         return -ENOMEM;
803         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
804         return 0;
805 }
806
807 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
808                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
809                 unsigned long addr, unsigned long end)
810 {
811         pud_t *src_pud, *dst_pud;
812         unsigned long next;
813
814         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
815         if (!dst_pud)
816                 return -ENOMEM;
817         src_pud = pud_offset(src_pgd, addr);
818         do {
819                 next = pud_addr_end(addr, end);
820                 if (pud_none_or_clear_bad(src_pud))
821                         continue;
822                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
823                                                 vma, addr, next))
824                         return -ENOMEM;
825         } while (dst_pud++, src_pud++, addr = next, addr != end);
826         return 0;
827 }
828
829 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
830                 struct vm_area_struct *vma)
831 {
832         pgd_t *src_pgd, *dst_pgd;
833         unsigned long next;
834         unsigned long addr = vma->vm_start;
835         unsigned long end = vma->vm_end;
836         int ret;
837
838         /*
839          * Don't copy ptes where a page fault will fill them correctly.
840          * Fork becomes much lighter when there are big shared or private
841          * readonly mappings. The tradeoff is that copy_page_range is more
842          * efficient than faulting.
843          */
844         if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
845                 if (!vma->anon_vma)
846                         return 0;
847         }
848
849         if (is_vm_hugetlb_page(vma))
850                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
851
852         if (unlikely(is_pfn_mapping(vma))) {
853                 /*
854                  * We do not free on error cases below as remove_vma
855                  * gets called on error from higher level routine
856                  */
857                 ret = track_pfn_vma_copy(vma);
858                 if (ret)
859                         return ret;
860         }
861
862         /*
863          * We need to invalidate the secondary MMU mappings only when
864          * there could be a permission downgrade on the ptes of the
865          * parent mm. And a permission downgrade will only happen if
866          * is_cow_mapping() returns true.
867          */
868         if (is_cow_mapping(vma->vm_flags))
869                 mmu_notifier_invalidate_range_start(src_mm, addr, end);
870
871         ret = 0;
872         dst_pgd = pgd_offset(dst_mm, addr);
873         src_pgd = pgd_offset(src_mm, addr);
874         do {
875                 next = pgd_addr_end(addr, end);
876                 if (pgd_none_or_clear_bad(src_pgd))
877                         continue;
878                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
879                                             vma, addr, next))) {
880                         ret = -ENOMEM;
881                         break;
882                 }
883         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
884
885         if (is_cow_mapping(vma->vm_flags))
886                 mmu_notifier_invalidate_range_end(src_mm,
887                                                   vma->vm_start, end);
888         return ret;
889 }
890
891 static unsigned long zap_pte_range(struct mmu_gather *tlb,
892                                 struct vm_area_struct *vma, pmd_t *pmd,
893                                 unsigned long addr, unsigned long end,
894                                 long *zap_work, struct zap_details *details)
895 {
896         struct mm_struct *mm = tlb->mm;
897         pte_t *pte;
898         spinlock_t *ptl;
899         int rss[NR_MM_COUNTERS];
900
901         init_rss_vec(rss);
902
903         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
904         arch_enter_lazy_mmu_mode();
905         do {
906                 pte_t ptent = *pte;
907                 if (pte_none(ptent)) {
908                         (*zap_work)--;
909                         continue;
910                 }
911
912                 (*zap_work) -= PAGE_SIZE;
913
914                 if (pte_present(ptent)) {
915                         struct page *page;
916
917                         page = vm_normal_page(vma, addr, ptent);
918                         if (unlikely(details) && page) {
919                                 /*
920                                  * unmap_shared_mapping_pages() wants to
921                                  * invalidate cache without truncating:
922                                  * unmap shared but keep private pages.
923                                  */
924                                 if (details->check_mapping &&
925                                     details->check_mapping != page->mapping)
926                                         continue;
927                                 /*
928                                  * Each page->index must be checked when
929                                  * invalidating or truncating nonlinear.
930                                  */
931                                 if (details->nonlinear_vma &&
932                                     (page->index < details->first_index ||
933                                      page->index > details->last_index))
934                                         continue;
935                         }
936                         ptent = ptep_get_and_clear_full(mm, addr, pte,
937                                                         tlb->fullmm);
938                         tlb_remove_tlb_entry(tlb, pte, addr);
939                         if (unlikely(!page))
940                                 continue;
941                         if (unlikely(details) && details->nonlinear_vma
942                             && linear_page_index(details->nonlinear_vma,
943                                                 addr) != page->index)
944                                 set_pte_at(mm, addr, pte,
945                                            pgoff_to_pte(page->index));
946                         if (PageAnon(page))
947                                 rss[MM_ANONPAGES]--;
948                         else {
949                                 if (pte_dirty(ptent))
950                                         set_page_dirty(page);
951                                 if (pte_young(ptent) &&
952                                     likely(!VM_SequentialReadHint(vma)))
953                                         mark_page_accessed(page);
954                                 rss[MM_FILEPAGES]--;
955                         }
956                         page_remove_rmap(page);
957                         if (unlikely(page_mapcount(page) < 0))
958                                 print_bad_pte(vma, addr, ptent, page);
959                         tlb_remove_page(tlb, page);
960                         continue;
961                 }
962                 /*
963                  * If details->check_mapping, we leave swap entries;
964                  * if details->nonlinear_vma, we leave file entries.
965                  */
966                 if (unlikely(details))
967                         continue;
968                 if (pte_file(ptent)) {
969                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
970                                 print_bad_pte(vma, addr, ptent, NULL);
971                 } else {
972                         swp_entry_t entry = pte_to_swp_entry(ptent);
973
974                         if (!non_swap_entry(entry))
975                                 rss[MM_SWAPENTS]--;
976                         if (unlikely(!free_swap_and_cache(entry)))
977                                 print_bad_pte(vma, addr, ptent, NULL);
978                 }
979                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
980         } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
981
982         add_mm_rss_vec(mm, rss);
983         arch_leave_lazy_mmu_mode();
984         pte_unmap_unlock(pte - 1, ptl);
985
986         return addr;
987 }
988
989 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
990                                 struct vm_area_struct *vma, pud_t *pud,
991                                 unsigned long addr, unsigned long end,
992                                 long *zap_work, struct zap_details *details)
993 {
994         pmd_t *pmd;
995         unsigned long next;
996
997         pmd = pmd_offset(pud, addr);
998         do {
999                 next = pmd_addr_end(addr, end);
1000                 if (pmd_none_or_clear_bad(pmd)) {
1001                         (*zap_work)--;
1002                         continue;
1003                 }
1004                 next = zap_pte_range(tlb, vma, pmd, addr, next,
1005                                                 zap_work, details);
1006         } while (pmd++, addr = next, (addr != end && *zap_work > 0));
1007
1008         return addr;
1009 }
1010
1011 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1012                                 struct vm_area_struct *vma, pgd_t *pgd,
1013                                 unsigned long addr, unsigned long end,
1014                                 long *zap_work, struct zap_details *details)
1015 {
1016         pud_t *pud;
1017         unsigned long next;
1018
1019         pud = pud_offset(pgd, addr);
1020         do {
1021                 next = pud_addr_end(addr, end);
1022                 if (pud_none_or_clear_bad(pud)) {
1023                         (*zap_work)--;
1024                         continue;
1025                 }
1026                 next = zap_pmd_range(tlb, vma, pud, addr, next,
1027                                                 zap_work, details);
1028         } while (pud++, addr = next, (addr != end && *zap_work > 0));
1029
1030         return addr;
1031 }
1032
1033 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1034                                 struct vm_area_struct *vma,
1035                                 unsigned long addr, unsigned long end,
1036                                 long *zap_work, struct zap_details *details)
1037 {
1038         pgd_t *pgd;
1039         unsigned long next;
1040
1041         if (details && !details->check_mapping && !details->nonlinear_vma)
1042                 details = NULL;
1043
1044         BUG_ON(addr >= end);
1045         mem_cgroup_uncharge_start();
1046         tlb_start_vma(tlb, vma);
1047         pgd = pgd_offset(vma->vm_mm, addr);
1048         do {
1049                 next = pgd_addr_end(addr, end);
1050                 if (pgd_none_or_clear_bad(pgd)) {
1051                         (*zap_work)--;
1052                         continue;
1053                 }
1054                 next = zap_pud_range(tlb, vma, pgd, addr, next,
1055                                                 zap_work, details);
1056         } while (pgd++, addr = next, (addr != end && *zap_work > 0));
1057         tlb_end_vma(tlb, vma);
1058         mem_cgroup_uncharge_end();
1059
1060         return addr;
1061 }
1062
1063 #ifdef CONFIG_PREEMPT
1064 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1065 #else
1066 /* No preempt: go for improved straight-line efficiency */
1067 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1068 #endif
1069
1070 /**
1071  * unmap_vmas - unmap a range of memory covered by a list of vma's
1072  * @tlbp: address of the caller's struct mmu_gather
1073  * @vma: the starting vma
1074  * @start_addr: virtual address at which to start unmapping
1075  * @end_addr: virtual address at which to end unmapping
1076  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1077  * @details: details of nonlinear truncation or shared cache invalidation
1078  *
1079  * Returns the end address of the unmapping (restart addr if interrupted).
1080  *
1081  * Unmap all pages in the vma list.
1082  *
1083  * We aim to not hold locks for too long (for scheduling latency reasons).
1084  * So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
1085  * return the ending mmu_gather to the caller.
1086  *
1087  * Only addresses between `start' and `end' will be unmapped.
1088  *
1089  * The VMA list must be sorted in ascending virtual address order.
1090  *
1091  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1092  * range after unmap_vmas() returns.  So the only responsibility here is to
1093  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1094  * drops the lock and schedules.
1095  */
1096 unsigned long unmap_vmas(struct mmu_gather **tlbp,
1097                 struct vm_area_struct *vma, unsigned long start_addr,
1098                 unsigned long end_addr, unsigned long *nr_accounted,
1099                 struct zap_details *details)
1100 {
1101         long zap_work = ZAP_BLOCK_SIZE;
1102         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
1103         int tlb_start_valid = 0;
1104         unsigned long start = start_addr;
1105         spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
1106         int fullmm = (*tlbp)->fullmm;
1107         struct mm_struct *mm = vma->vm_mm;
1108
1109         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1110         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1111                 unsigned long end;
1112
1113                 start = max(vma->vm_start, start_addr);
1114                 if (start >= vma->vm_end)
1115                         continue;
1116                 end = min(vma->vm_end, end_addr);
1117                 if (end <= vma->vm_start)
1118                         continue;
1119
1120                 if (vma->vm_flags & VM_ACCOUNT)
1121                         *nr_accounted += (end - start) >> PAGE_SHIFT;
1122
1123                 if (unlikely(is_pfn_mapping(vma)))
1124                         untrack_pfn_vma(vma, 0, 0);
1125
1126                 while (start != end) {
1127                         if (!tlb_start_valid) {
1128                                 tlb_start = start;
1129                                 tlb_start_valid = 1;
1130                         }
1131
1132                         if (unlikely(is_vm_hugetlb_page(vma))) {
1133                                 /*
1134                                  * It is undesirable to test vma->vm_file as it
1135                                  * should be non-null for valid hugetlb area.
1136                                  * However, vm_file will be NULL in the error
1137                                  * cleanup path of do_mmap_pgoff. When
1138                                  * hugetlbfs ->mmap method fails,
1139                                  * do_mmap_pgoff() nullifies vma->vm_file
1140                                  * before calling this function to clean up.
1141                                  * Since no pte has actually been setup, it is
1142                                  * safe to do nothing in this case.
1143                                  */
1144                                 if (vma->vm_file) {
1145                                         unmap_hugepage_range(vma, start, end, NULL);
1146                                         zap_work -= (end - start) /
1147                                         pages_per_huge_page(hstate_vma(vma));
1148                                 }
1149
1150                                 start = end;
1151                         } else
1152                                 start = unmap_page_range(*tlbp, vma,
1153                                                 start, end, &zap_work, details);
1154
1155                         if (zap_work > 0) {
1156                                 BUG_ON(start != end);
1157                                 break;
1158                         }
1159
1160                         tlb_finish_mmu(*tlbp, tlb_start, start);
1161
1162                         if (need_resched() ||
1163                                 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1164                                 if (i_mmap_lock) {
1165                                         *tlbp = NULL;
1166                                         goto out;
1167                                 }
1168                                 cond_resched();
1169                         }
1170
1171                         *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1172                         tlb_start_valid = 0;
1173                         zap_work = ZAP_BLOCK_SIZE;
1174                 }
1175         }
1176 out:
1177         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1178         return start;   /* which is now the end (or restart) address */
1179 }
1180
1181 /**
1182  * zap_page_range - remove user pages in a given range
1183  * @vma: vm_area_struct holding the applicable pages
1184  * @address: starting address of pages to zap
1185  * @size: number of bytes to zap
1186  * @details: details of nonlinear truncation or shared cache invalidation
1187  */
1188 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1189                 unsigned long size, struct zap_details *details)
1190 {
1191         struct mm_struct *mm = vma->vm_mm;
1192         struct mmu_gather *tlb;
1193         unsigned long end = address + size;
1194         unsigned long nr_accounted = 0;
1195
1196         lru_add_drain();
1197         tlb = tlb_gather_mmu(mm, 0);
1198         update_hiwater_rss(mm);
1199         end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1200         if (tlb)
1201                 tlb_finish_mmu(tlb, address, end);
1202         return end;
1203 }
1204
1205 /**
1206  * zap_vma_ptes - remove ptes mapping the vma
1207  * @vma: vm_area_struct holding ptes to be zapped
1208  * @address: starting address of pages to zap
1209  * @size: number of bytes to zap
1210  *
1211  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1212  *
1213  * The entire address range must be fully contained within the vma.
1214  *
1215  * Returns 0 if successful.
1216  */
1217 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1218                 unsigned long size)
1219 {
1220         if (address < vma->vm_start || address + size > vma->vm_end ||
1221                         !(vma->vm_flags & VM_PFNMAP))
1222                 return -1;
1223         zap_page_range(vma, address, size, NULL);
1224         return 0;
1225 }
1226 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1227
1228 /**
1229  * follow_page - look up a page descriptor from a user-virtual address
1230  * @vma: vm_area_struct mapping @address
1231  * @address: virtual address to look up
1232  * @flags: flags modifying lookup behaviour
1233  *
1234  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1235  *
1236  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1237  * an error pointer if there is a mapping to something not represented
1238  * by a page descriptor (see also vm_normal_page()).
1239  */
1240 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1241                         unsigned int flags)
1242 {
1243         pgd_t *pgd;
1244         pud_t *pud;
1245         pmd_t *pmd;
1246         pte_t *ptep, pte;
1247         spinlock_t *ptl;
1248         struct page *page;
1249         struct mm_struct *mm = vma->vm_mm;
1250
1251         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1252         if (!IS_ERR(page)) {
1253                 BUG_ON(flags & FOLL_GET);
1254                 goto out;
1255         }
1256
1257         page = NULL;
1258         pgd = pgd_offset(mm, address);
1259         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1260                 goto no_page_table;
1261
1262         pud = pud_offset(pgd, address);
1263         if (pud_none(*pud))
1264                 goto no_page_table;
1265         if (pud_huge(*pud)) {
1266                 BUG_ON(flags & FOLL_GET);
1267                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1268                 goto out;
1269         }
1270         if (unlikely(pud_bad(*pud)))
1271                 goto no_page_table;
1272
1273         pmd = pmd_offset(pud, address);
1274         if (pmd_none(*pmd))
1275                 goto no_page_table;
1276         if (pmd_huge(*pmd)) {
1277                 BUG_ON(flags & FOLL_GET);
1278                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1279                 goto out;
1280         }
1281         if (unlikely(pmd_bad(*pmd)))
1282                 goto no_page_table;
1283
1284         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1285
1286         pte = *ptep;
1287         if (!pte_present(pte))
1288                 goto no_page;
1289         if ((flags & FOLL_WRITE) && !pte_write(pte))
1290                 goto unlock;
1291
1292         page = vm_normal_page(vma, address, pte);
1293         if (unlikely(!page)) {
1294                 if ((flags & FOLL_DUMP) ||
1295                     !is_zero_pfn(pte_pfn(pte)))
1296                         goto bad_page;
1297                 page = pte_page(pte);
1298         }
1299
1300         if (flags & FOLL_GET)
1301                 get_page(page);
1302         if (flags & FOLL_TOUCH) {
1303                 if ((flags & FOLL_WRITE) &&
1304                     !pte_dirty(pte) && !PageDirty(page))
1305                         set_page_dirty(page);
1306                 /*
1307                  * pte_mkyoung() would be more correct here, but atomic care
1308                  * is needed to avoid losing the dirty bit: it is easier to use
1309                  * mark_page_accessed().
1310                  */
1311                 mark_page_accessed(page);
1312         }
1313         if (flags & FOLL_MLOCK) {
1314                 /*
1315                  * The preliminary mapping check is mainly to avoid the
1316                  * pointless overhead of lock_page on the ZERO_PAGE
1317                  * which might bounce very badly if there is contention.
1318                  *
1319                  * If the page is already locked, we don't need to
1320                  * handle it now - vmscan will handle it later if and
1321                  * when it attempts to reclaim the page.
1322                  */
1323                 if (page->mapping && trylock_page(page)) {
1324                         lru_add_drain();  /* push cached pages to LRU */
1325                         /*
1326                          * Because we lock page here and migration is
1327                          * blocked by the pte's page reference, we need
1328                          * only check for file-cache page truncation.
1329                          */
1330                         if (page->mapping)
1331                                 mlock_vma_page(page);
1332                         unlock_page(page);
1333                 }
1334         }
1335 unlock:
1336         pte_unmap_unlock(ptep, ptl);
1337 out:
1338         return page;
1339
1340 bad_page:
1341         pte_unmap_unlock(ptep, ptl);
1342         return ERR_PTR(-EFAULT);
1343
1344 no_page:
1345         pte_unmap_unlock(ptep, ptl);
1346         if (!pte_none(pte))
1347                 return page;
1348
1349 no_page_table:
1350         /*
1351          * When core dumping an enormous anonymous area that nobody
1352          * has touched so far, we don't want to allocate unnecessary pages or
1353          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1354          * then get_dump_page() will return NULL to leave a hole in the dump.
1355          * But we can only make this optimization where a hole would surely
1356          * be zero-filled if handle_mm_fault() actually did handle it.
1357          */
1358         if ((flags & FOLL_DUMP) &&
1359             (!vma->vm_ops || !vma->vm_ops->fault))
1360                 return ERR_PTR(-EFAULT);
1361         return page;
1362 }
1363
1364 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1365                      unsigned long start, int nr_pages, unsigned int gup_flags,
1366                      struct page **pages, struct vm_area_struct **vmas,
1367                      int *nonblocking)
1368 {
1369         int i;
1370         unsigned long vm_flags;
1371
1372         if (nr_pages <= 0)
1373                 return 0;
1374
1375         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1376
1377         /* 
1378          * Require read or write permissions.
1379          * If FOLL_FORCE is set, we only require the "MAY" flags.
1380          */
1381         vm_flags  = (gup_flags & FOLL_WRITE) ?
1382                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1383         vm_flags &= (gup_flags & FOLL_FORCE) ?
1384                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1385         i = 0;
1386
1387         do {
1388                 struct vm_area_struct *vma;
1389
1390                 vma = find_extend_vma(mm, start);
1391                 if (!vma && in_gate_area(tsk, start)) {
1392                         unsigned long pg = start & PAGE_MASK;
1393                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1394                         pgd_t *pgd;
1395                         pud_t *pud;
1396                         pmd_t *pmd;
1397                         pte_t *pte;
1398
1399                         /* user gate pages are read-only */
1400                         if (gup_flags & FOLL_WRITE)
1401                                 return i ? : -EFAULT;
1402                         if (pg > TASK_SIZE)
1403                                 pgd = pgd_offset_k(pg);
1404                         else
1405                                 pgd = pgd_offset_gate(mm, pg);
1406                         BUG_ON(pgd_none(*pgd));
1407                         pud = pud_offset(pgd, pg);
1408                         BUG_ON(pud_none(*pud));
1409                         pmd = pmd_offset(pud, pg);
1410                         if (pmd_none(*pmd))
1411                                 return i ? : -EFAULT;
1412                         pte = pte_offset_map(pmd, pg);
1413                         if (pte_none(*pte)) {
1414                                 pte_unmap(pte);
1415                                 return i ? : -EFAULT;
1416                         }
1417                         if (pages) {
1418                                 struct page *page;
1419
1420                                 page = vm_normal_page(gate_vma, start, *pte);
1421                                 if (!page) {
1422                                         if (!(gup_flags & FOLL_DUMP) &&
1423                                              is_zero_pfn(pte_pfn(*pte)))
1424                                                 page = pte_page(*pte);
1425                                         else {
1426                                                 pte_unmap(pte);
1427                                                 return i ? : -EFAULT;
1428                                         }
1429                                 }
1430                                 pages[i] = page;
1431                                 get_page(page);
1432                         }
1433                         pte_unmap(pte);
1434                         if (vmas)
1435                                 vmas[i] = gate_vma;
1436                         i++;
1437                         start += PAGE_SIZE;
1438                         nr_pages--;
1439                         continue;
1440                 }
1441
1442                 if (!vma ||
1443                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1444                     !(vm_flags & vma->vm_flags))
1445                         return i ? : -EFAULT;
1446
1447                 if (is_vm_hugetlb_page(vma)) {
1448                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1449                                         &start, &nr_pages, i, gup_flags);
1450                         continue;
1451                 }
1452
1453                 do {
1454                         struct page *page;
1455                         unsigned int foll_flags = gup_flags;
1456
1457                         /*
1458                          * If we have a pending SIGKILL, don't keep faulting
1459                          * pages and potentially allocating memory.
1460                          */
1461                         if (unlikely(fatal_signal_pending(current)))
1462                                 return i ? i : -ERESTARTSYS;
1463
1464                         cond_resched();
1465                         while (!(page = follow_page(vma, start, foll_flags))) {
1466                                 int ret;
1467                                 unsigned int fault_flags = 0;
1468
1469                                 if (foll_flags & FOLL_WRITE)
1470                                         fault_flags |= FAULT_FLAG_WRITE;
1471                                 if (nonblocking)
1472                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1473
1474                                 ret = handle_mm_fault(mm, vma, start,
1475                                                         fault_flags);
1476
1477                                 if (ret & VM_FAULT_ERROR) {
1478                                         if (ret & VM_FAULT_OOM)
1479                                                 return i ? i : -ENOMEM;
1480                                         if (ret &
1481                                             (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE|
1482                                              VM_FAULT_SIGBUS))
1483                                                 return i ? i : -EFAULT;
1484                                         BUG();
1485                                 }
1486                                 if (ret & VM_FAULT_MAJOR)
1487                                         tsk->maj_flt++;
1488                                 else
1489                                         tsk->min_flt++;
1490
1491                                 if (ret & VM_FAULT_RETRY) {
1492                                         *nonblocking = 0;
1493                                         return i;
1494                                 }
1495
1496                                 /*
1497                                  * The VM_FAULT_WRITE bit tells us that
1498                                  * do_wp_page has broken COW when necessary,
1499                                  * even if maybe_mkwrite decided not to set
1500                                  * pte_write. We can thus safely do subsequent
1501                                  * page lookups as if they were reads. But only
1502                                  * do so when looping for pte_write is futile:
1503                                  * in some cases userspace may also be wanting
1504                                  * to write to the gotten user page, which a
1505                                  * read fault here might prevent (a readonly
1506                                  * page might get reCOWed by userspace write).
1507                                  */
1508                                 if ((ret & VM_FAULT_WRITE) &&
1509                                     !(vma->vm_flags & VM_WRITE))
1510                                         foll_flags &= ~FOLL_WRITE;
1511
1512                                 cond_resched();
1513                         }
1514                         if (IS_ERR(page))
1515                                 return i ? i : PTR_ERR(page);
1516                         if (pages) {
1517                                 pages[i] = page;
1518
1519                                 flush_anon_page(vma, page, start);
1520                                 flush_dcache_page(page);
1521                         }
1522                         if (vmas)
1523                                 vmas[i] = vma;
1524                         i++;
1525                         start += PAGE_SIZE;
1526                         nr_pages--;
1527                 } while (nr_pages && start < vma->vm_end);
1528         } while (nr_pages);
1529         return i;
1530 }
1531
1532 /**
1533  * get_user_pages() - pin user pages in memory
1534  * @tsk:        task_struct of target task
1535  * @mm:         mm_struct of target mm
1536  * @start:      starting user address
1537  * @nr_pages:   number of pages from start to pin
1538  * @write:      whether pages will be written to by the caller
1539  * @force:      whether to force write access even if user mapping is
1540  *              readonly. This will result in the page being COWed even
1541  *              in MAP_SHARED mappings. You do not want this.
1542  * @pages:      array that receives pointers to the pages pinned.
1543  *              Should be at least nr_pages long. Or NULL, if caller
1544  *              only intends to ensure the pages are faulted in.
1545  * @vmas:       array of pointers to vmas corresponding to each page.
1546  *              Or NULL if the caller does not require them.
1547  *
1548  * Returns number of pages pinned. This may be fewer than the number
1549  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1550  * were pinned, returns -errno. Each page returned must be released
1551  * with a put_page() call when it is finished with. vmas will only
1552  * remain valid while mmap_sem is held.
1553  *
1554  * Must be called with mmap_sem held for read or write.
1555  *
1556  * get_user_pages walks a process's page tables and takes a reference to
1557  * each struct page that each user address corresponds to at a given
1558  * instant. That is, it takes the page that would be accessed if a user
1559  * thread accesses the given user virtual address at that instant.
1560  *
1561  * This does not guarantee that the page exists in the user mappings when
1562  * get_user_pages returns, and there may even be a completely different
1563  * page there in some cases (eg. if mmapped pagecache has been invalidated
1564  * and subsequently re faulted). However it does guarantee that the page
1565  * won't be freed completely. And mostly callers simply care that the page
1566  * contains data that was valid *at some point in time*. Typically, an IO
1567  * or similar operation cannot guarantee anything stronger anyway because
1568  * locks can't be held over the syscall boundary.
1569  *
1570  * If write=0, the page must not be written to. If the page is written to,
1571  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1572  * after the page is finished with, and before put_page is called.
1573  *
1574  * get_user_pages is typically used for fewer-copy IO operations, to get a
1575  * handle on the memory by some means other than accesses via the user virtual
1576  * addresses. The pages may be submitted for DMA to devices or accessed via
1577  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1578  * use the correct cache flushing APIs.
1579  *
1580  * See also get_user_pages_fast, for performance critical applications.
1581  */
1582 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1583                 unsigned long start, int nr_pages, int write, int force,
1584                 struct page **pages, struct vm_area_struct **vmas)
1585 {
1586         int flags = FOLL_TOUCH;
1587
1588         if (pages)
1589                 flags |= FOLL_GET;
1590         if (write)
1591                 flags |= FOLL_WRITE;
1592         if (force)
1593                 flags |= FOLL_FORCE;
1594
1595         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1596                                 NULL);
1597 }
1598 EXPORT_SYMBOL(get_user_pages);
1599
1600 /**
1601  * get_dump_page() - pin user page in memory while writing it to core dump
1602  * @addr: user address
1603  *
1604  * Returns struct page pointer of user page pinned for dump,
1605  * to be freed afterwards by page_cache_release() or put_page().
1606  *
1607  * Returns NULL on any kind of failure - a hole must then be inserted into
1608  * the corefile, to preserve alignment with its headers; and also returns
1609  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1610  * allowing a hole to be left in the corefile to save diskspace.
1611  *
1612  * Called without mmap_sem, but after all other threads have been killed.
1613  */
1614 #ifdef CONFIG_ELF_CORE
1615 struct page *get_dump_page(unsigned long addr)
1616 {
1617         struct vm_area_struct *vma;
1618         struct page *page;
1619
1620         if (__get_user_pages(current, current->mm, addr, 1,
1621                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1622                              NULL) < 1)
1623                 return NULL;
1624         flush_cache_page(vma, addr, page_to_pfn(page));
1625         return page;
1626 }
1627 #endif /* CONFIG_ELF_CORE */
1628
1629 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1630                         spinlock_t **ptl)
1631 {
1632         pgd_t * pgd = pgd_offset(mm, addr);
1633         pud_t * pud = pud_alloc(mm, pgd, addr);
1634         if (pud) {
1635                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1636                 if (pmd)
1637                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1638         }
1639         return NULL;
1640 }
1641
1642 /*
1643  * This is the old fallback for page remapping.
1644  *
1645  * For historical reasons, it only allows reserved pages. Only
1646  * old drivers should use this, and they needed to mark their
1647  * pages reserved for the old functions anyway.
1648  */
1649 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1650                         struct page *page, pgprot_t prot)
1651 {
1652         struct mm_struct *mm = vma->vm_mm;
1653         int retval;
1654         pte_t *pte;
1655         spinlock_t *ptl;
1656
1657         retval = -EINVAL;
1658         if (PageAnon(page))
1659                 goto out;
1660         retval = -ENOMEM;
1661         flush_dcache_page(page);
1662         pte = get_locked_pte(mm, addr, &ptl);
1663         if (!pte)
1664                 goto out;
1665         retval = -EBUSY;
1666         if (!pte_none(*pte))
1667                 goto out_unlock;
1668
1669         /* Ok, finally just insert the thing.. */
1670         get_page(page);
1671         inc_mm_counter_fast(mm, MM_FILEPAGES);
1672         page_add_file_rmap(page);
1673         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1674
1675         retval = 0;
1676         pte_unmap_unlock(pte, ptl);
1677         return retval;
1678 out_unlock:
1679         pte_unmap_unlock(pte, ptl);
1680 out:
1681         return retval;
1682 }
1683
1684 /**
1685  * vm_insert_page - insert single page into user vma
1686  * @vma: user vma to map to
1687  * @addr: target user address of this page
1688  * @page: source kernel page
1689  *
1690  * This allows drivers to insert individual pages they've allocated
1691  * into a user vma.
1692  *
1693  * The page has to be a nice clean _individual_ kernel allocation.
1694  * If you allocate a compound page, you need to have marked it as
1695  * such (__GFP_COMP), or manually just split the page up yourself
1696  * (see split_page()).
1697  *
1698  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1699  * took an arbitrary page protection parameter. This doesn't allow
1700  * that. Your vma protection will have to be set up correctly, which
1701  * means that if you want a shared writable mapping, you'd better
1702  * ask for a shared writable mapping!
1703  *
1704  * The page does not need to be reserved.
1705  */
1706 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1707                         struct page *page)
1708 {
1709         if (addr < vma->vm_start || addr >= vma->vm_end)
1710                 return -EFAULT;
1711         if (!page_count(page))
1712                 return -EINVAL;
1713         vma->vm_flags |= VM_INSERTPAGE;
1714         return insert_page(vma, addr, page, vma->vm_page_prot);
1715 }
1716 EXPORT_SYMBOL(vm_insert_page);
1717
1718 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1719                         unsigned long pfn, pgprot_t prot)
1720 {
1721         struct mm_struct *mm = vma->vm_mm;
1722         int retval;
1723         pte_t *pte, entry;
1724         spinlock_t *ptl;
1725
1726         retval = -ENOMEM;
1727         pte = get_locked_pte(mm, addr, &ptl);
1728         if (!pte)
1729                 goto out;
1730         retval = -EBUSY;
1731         if (!pte_none(*pte))
1732                 goto out_unlock;
1733
1734         /* Ok, finally just insert the thing.. */
1735         entry = pte_mkspecial(pfn_pte(pfn, prot));
1736         set_pte_at(mm, addr, pte, entry);
1737         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1738
1739         retval = 0;
1740 out_unlock:
1741         pte_unmap_unlock(pte, ptl);
1742 out:
1743         return retval;
1744 }
1745
1746 /**
1747  * vm_insert_pfn - insert single pfn into user vma
1748  * @vma: user vma to map to
1749  * @addr: target user address of this page
1750  * @pfn: source kernel pfn
1751  *
1752  * Similar to vm_inert_page, this allows drivers to insert individual pages
1753  * they've allocated into a user vma. Same comments apply.
1754  *
1755  * This function should only be called from a vm_ops->fault handler, and
1756  * in that case the handler should return NULL.
1757  *
1758  * vma cannot be a COW mapping.
1759  *
1760  * As this is called only for pages that do not currently exist, we
1761  * do not need to flush old virtual caches or the TLB.
1762  */
1763 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1764                         unsigned long pfn)
1765 {
1766         int ret;
1767         pgprot_t pgprot = vma->vm_page_prot;
1768         /*
1769          * Technically, architectures with pte_special can avoid all these
1770          * restrictions (same for remap_pfn_range).  However we would like
1771          * consistency in testing and feature parity among all, so we should
1772          * try to keep these invariants in place for everybody.
1773          */
1774         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1775         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1776                                                 (VM_PFNMAP|VM_MIXEDMAP));
1777         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1778         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1779
1780         if (addr < vma->vm_start || addr >= vma->vm_end)
1781                 return -EFAULT;
1782         if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1783                 return -EINVAL;
1784
1785         ret = insert_pfn(vma, addr, pfn, pgprot);
1786
1787         if (ret)
1788                 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1789
1790         return ret;
1791 }
1792 EXPORT_SYMBOL(vm_insert_pfn);
1793
1794 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1795                         unsigned long pfn)
1796 {
1797         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1798
1799         if (addr < vma->vm_start || addr >= vma->vm_end)
1800                 return -EFAULT;
1801
1802         /*
1803          * If we don't have pte special, then we have to use the pfn_valid()
1804          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1805          * refcount the page if pfn_valid is true (hence insert_page rather
1806          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1807          * without pte special, it would there be refcounted as a normal page.
1808          */
1809         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1810                 struct page *page;
1811
1812                 page = pfn_to_page(pfn);
1813                 return insert_page(vma, addr, page, vma->vm_page_prot);
1814         }
1815         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1816 }
1817 EXPORT_SYMBOL(vm_insert_mixed);
1818
1819 /*
1820  * maps a range of physical memory into the requested pages. the old
1821  * mappings are removed. any references to nonexistent pages results
1822  * in null mappings (currently treated as "copy-on-access")
1823  */
1824 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1825                         unsigned long addr, unsigned long end,
1826                         unsigned long pfn, pgprot_t prot)
1827 {
1828         pte_t *pte;
1829         spinlock_t *ptl;
1830
1831         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1832         if (!pte)
1833                 return -ENOMEM;
1834         arch_enter_lazy_mmu_mode();
1835         do {
1836                 BUG_ON(!pte_none(*pte));
1837                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1838                 pfn++;
1839         } while (pte++, addr += PAGE_SIZE, addr != end);
1840         arch_leave_lazy_mmu_mode();
1841         pte_unmap_unlock(pte - 1, ptl);
1842         return 0;
1843 }
1844
1845 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1846                         unsigned long addr, unsigned long end,
1847                         unsigned long pfn, pgprot_t prot)
1848 {
1849         pmd_t *pmd;
1850         unsigned long next;
1851
1852         pfn -= addr >> PAGE_SHIFT;
1853         pmd = pmd_alloc(mm, pud, addr);
1854         if (!pmd)
1855                 return -ENOMEM;
1856         do {
1857                 next = pmd_addr_end(addr, end);
1858                 if (remap_pte_range(mm, pmd, addr, next,
1859                                 pfn + (addr >> PAGE_SHIFT), prot))
1860                         return -ENOMEM;
1861         } while (pmd++, addr = next, addr != end);
1862         return 0;
1863 }
1864
1865 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1866                         unsigned long addr, unsigned long end,
1867                         unsigned long pfn, pgprot_t prot)
1868 {
1869         pud_t *pud;
1870         unsigned long next;
1871
1872         pfn -= addr >> PAGE_SHIFT;
1873         pud = pud_alloc(mm, pgd, addr);
1874         if (!pud)
1875                 return -ENOMEM;
1876         do {
1877                 next = pud_addr_end(addr, end);
1878                 if (remap_pmd_range(mm, pud, addr, next,
1879                                 pfn + (addr >> PAGE_SHIFT), prot))
1880                         return -ENOMEM;
1881         } while (pud++, addr = next, addr != end);
1882         return 0;
1883 }
1884
1885 /**
1886  * remap_pfn_range - remap kernel memory to userspace
1887  * @vma: user vma to map to
1888  * @addr: target user address to start at
1889  * @pfn: physical address of kernel memory
1890  * @size: size of map area
1891  * @prot: page protection flags for this mapping
1892  *
1893  *  Note: this is only safe if the mm semaphore is held when called.
1894  */
1895 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1896                     unsigned long pfn, unsigned long size, pgprot_t prot)
1897 {
1898         pgd_t *pgd;
1899         unsigned long next;
1900         unsigned long end = addr + PAGE_ALIGN(size);
1901         struct mm_struct *mm = vma->vm_mm;
1902         int err;
1903
1904         /*
1905          * Physically remapped pages are special. Tell the
1906          * rest of the world about it:
1907          *   VM_IO tells people not to look at these pages
1908          *      (accesses can have side effects).
1909          *   VM_RESERVED is specified all over the place, because
1910          *      in 2.4 it kept swapout's vma scan off this vma; but
1911          *      in 2.6 the LRU scan won't even find its pages, so this
1912          *      flag means no more than count its pages in reserved_vm,
1913          *      and omit it from core dump, even when VM_IO turned off.
1914          *   VM_PFNMAP tells the core MM that the base pages are just
1915          *      raw PFN mappings, and do not have a "struct page" associated
1916          *      with them.
1917          *
1918          * There's a horrible special case to handle copy-on-write
1919          * behaviour that some programs depend on. We mark the "original"
1920          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1921          */
1922         if (addr == vma->vm_start && end == vma->vm_end) {
1923                 vma->vm_pgoff = pfn;
1924                 vma->vm_flags |= VM_PFN_AT_MMAP;
1925         } else if (is_cow_mapping(vma->vm_flags))
1926                 return -EINVAL;
1927
1928         vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1929
1930         err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1931         if (err) {
1932                 /*
1933                  * To indicate that track_pfn related cleanup is not
1934                  * needed from higher level routine calling unmap_vmas
1935                  */
1936                 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1937                 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1938                 return -EINVAL;
1939         }
1940
1941         BUG_ON(addr >= end);
1942         pfn -= addr >> PAGE_SHIFT;
1943         pgd = pgd_offset(mm, addr);
1944         flush_cache_range(vma, addr, end);
1945         do {
1946                 next = pgd_addr_end(addr, end);
1947                 err = remap_pud_range(mm, pgd, addr, next,
1948                                 pfn + (addr >> PAGE_SHIFT), prot);
1949                 if (err)
1950                         break;
1951         } while (pgd++, addr = next, addr != end);
1952
1953         if (err)
1954                 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1955
1956         return err;
1957 }
1958 EXPORT_SYMBOL(remap_pfn_range);
1959
1960 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1961                                      unsigned long addr, unsigned long end,
1962                                      pte_fn_t fn, void *data)
1963 {
1964         pte_t *pte;
1965         int err;
1966         pgtable_t token;
1967         spinlock_t *uninitialized_var(ptl);
1968
1969         pte = (mm == &init_mm) ?
1970                 pte_alloc_kernel(pmd, addr) :
1971                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1972         if (!pte)
1973                 return -ENOMEM;
1974
1975         BUG_ON(pmd_huge(*pmd));
1976
1977         arch_enter_lazy_mmu_mode();
1978
1979         token = pmd_pgtable(*pmd);
1980
1981         do {
1982                 err = fn(pte++, token, addr, data);
1983                 if (err)
1984                         break;
1985         } while (addr += PAGE_SIZE, addr != end);
1986
1987         arch_leave_lazy_mmu_mode();
1988
1989         if (mm != &init_mm)
1990                 pte_unmap_unlock(pte-1, ptl);
1991         return err;
1992 }
1993
1994 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1995                                      unsigned long addr, unsigned long end,
1996                                      pte_fn_t fn, void *data)
1997 {
1998         pmd_t *pmd;
1999         unsigned long next;
2000         int err;
2001
2002         BUG_ON(pud_huge(*pud));
2003
2004         pmd = pmd_alloc(mm, pud, addr);
2005         if (!pmd)
2006                 return -ENOMEM;
2007         do {
2008                 next = pmd_addr_end(addr, end);
2009                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2010                 if (err)
2011                         break;
2012         } while (pmd++, addr = next, addr != end);
2013         return err;
2014 }
2015
2016 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2017                                      unsigned long addr, unsigned long end,
2018                                      pte_fn_t fn, void *data)
2019 {
2020         pud_t *pud;
2021         unsigned long next;
2022         int err;
2023
2024         pud = pud_alloc(mm, pgd, addr);
2025         if (!pud)
2026                 return -ENOMEM;
2027         do {
2028                 next = pud_addr_end(addr, end);
2029                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2030                 if (err)
2031                         break;
2032         } while (pud++, addr = next, addr != end);
2033         return err;
2034 }
2035
2036 /*
2037  * Scan a region of virtual memory, filling in page tables as necessary
2038  * and calling a provided function on each leaf page table.
2039  */
2040 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2041                         unsigned long size, pte_fn_t fn, void *data)
2042 {
2043         pgd_t *pgd;
2044         unsigned long next;
2045         unsigned long end = addr + size;
2046         int err;
2047
2048         BUG_ON(addr >= end);
2049         pgd = pgd_offset(mm, addr);
2050         do {
2051                 next = pgd_addr_end(addr, end);
2052                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2053                 if (err)
2054                         break;
2055         } while (pgd++, addr = next, addr != end);
2056
2057         return err;
2058 }
2059 EXPORT_SYMBOL_GPL(apply_to_page_range);
2060
2061 /*
2062  * handle_pte_fault chooses page fault handler according to an entry
2063  * which was read non-atomically.  Before making any commitment, on
2064  * those architectures or configurations (e.g. i386 with PAE) which
2065  * might give a mix of unmatched parts, do_swap_page and do_file_page
2066  * must check under lock before unmapping the pte and proceeding
2067  * (but do_wp_page is only called after already making such a check;
2068  * and do_anonymous_page and do_no_page can safely check later on).
2069  */
2070 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2071                                 pte_t *page_table, pte_t orig_pte)
2072 {
2073         int same = 1;
2074 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2075         if (sizeof(pte_t) > sizeof(unsigned long)) {
2076                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2077                 spin_lock(ptl);
2078                 same = pte_same(*page_table, orig_pte);
2079                 spin_unlock(ptl);
2080         }
2081 #endif
2082         pte_unmap(page_table);
2083         return same;
2084 }
2085
2086 /*
2087  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
2088  * servicing faults for write access.  In the normal case, do always want
2089  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
2090  * that do not have writing enabled, when used by access_process_vm.
2091  */
2092 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
2093 {
2094         if (likely(vma->vm_flags & VM_WRITE))
2095                 pte = pte_mkwrite(pte);
2096         return pte;
2097 }
2098
2099 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2100 {
2101         /*
2102          * If the source page was a PFN mapping, we don't have
2103          * a "struct page" for it. We do a best-effort copy by
2104          * just copying from the original user address. If that
2105          * fails, we just zero-fill it. Live with it.
2106          */
2107         if (unlikely(!src)) {
2108                 void *kaddr = kmap_atomic(dst, KM_USER0);
2109                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2110
2111                 /*
2112                  * This really shouldn't fail, because the page is there
2113                  * in the page tables. But it might just be unreadable,
2114                  * in which case we just give up and fill the result with
2115                  * zeroes.
2116                  */
2117                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2118                         clear_page(kaddr);
2119                 kunmap_atomic(kaddr, KM_USER0);
2120                 flush_dcache_page(dst);
2121         } else
2122                 copy_user_highpage(dst, src, va, vma);
2123 }
2124
2125 /*
2126  * This routine handles present pages, when users try to write
2127  * to a shared page. It is done by copying the page to a new address
2128  * and decrementing the shared-page counter for the old page.
2129  *
2130  * Note that this routine assumes that the protection checks have been
2131  * done by the caller (the low-level page fault routine in most cases).
2132  * Thus we can safely just mark it writable once we've done any necessary
2133  * COW.
2134  *
2135  * We also mark the page dirty at this point even though the page will
2136  * change only once the write actually happens. This avoids a few races,
2137  * and potentially makes it more efficient.
2138  *
2139  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2140  * but allow concurrent faults), with pte both mapped and locked.
2141  * We return with mmap_sem still held, but pte unmapped and unlocked.
2142  */
2143 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2144                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2145                 spinlock_t *ptl, pte_t orig_pte)
2146         __releases(ptl)
2147 {
2148         struct page *old_page, *new_page;
2149         pte_t entry;
2150         int ret = 0;
2151         int page_mkwrite = 0;
2152         struct page *dirty_page = NULL;
2153
2154         old_page = vm_normal_page(vma, address, orig_pte);
2155         if (!old_page) {
2156                 /*
2157                  * VM_MIXEDMAP !pfn_valid() case
2158                  *
2159                  * We should not cow pages in a shared writeable mapping.
2160                  * Just mark the pages writable as we can't do any dirty
2161                  * accounting on raw pfn maps.
2162                  */
2163                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2164                                      (VM_WRITE|VM_SHARED))
2165                         goto reuse;
2166                 goto gotten;
2167         }
2168
2169         /*
2170          * Take out anonymous pages first, anonymous shared vmas are
2171          * not dirty accountable.
2172          */
2173         if (PageAnon(old_page) && !PageKsm(old_page)) {
2174                 if (!trylock_page(old_page)) {
2175                         page_cache_get(old_page);
2176                         pte_unmap_unlock(page_table, ptl);
2177                         lock_page(old_page);
2178                         page_table = pte_offset_map_lock(mm, pmd, address,
2179                                                          &ptl);
2180                         if (!pte_same(*page_table, orig_pte)) {
2181                                 unlock_page(old_page);
2182                                 page_cache_release(old_page);
2183                                 goto unlock;
2184                         }
2185                         page_cache_release(old_page);
2186                 }
2187                 if (reuse_swap_page(old_page)) {
2188                         /*
2189                          * The page is all ours.  Move it to our anon_vma so
2190                          * the rmap code will not search our parent or siblings.
2191                          * Protected against the rmap code by the page lock.
2192                          */
2193                         page_move_anon_rmap(old_page, vma, address);
2194                         unlock_page(old_page);
2195                         goto reuse;
2196                 }
2197                 unlock_page(old_page);
2198         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2199                                         (VM_WRITE|VM_SHARED))) {
2200                 /*
2201                  * Only catch write-faults on shared writable pages,
2202                  * read-only shared pages can get COWed by
2203                  * get_user_pages(.write=1, .force=1).
2204                  */
2205                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2206                         struct vm_fault vmf;
2207                         int tmp;
2208
2209                         vmf.virtual_address = (void __user *)(address &
2210                                                                 PAGE_MASK);
2211                         vmf.pgoff = old_page->index;
2212                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2213                         vmf.page = old_page;
2214
2215                         /*
2216                          * Notify the address space that the page is about to
2217                          * become writable so that it can prohibit this or wait
2218                          * for the page to get into an appropriate state.
2219                          *
2220                          * We do this without the lock held, so that it can
2221                          * sleep if it needs to.
2222                          */
2223                         page_cache_get(old_page);
2224                         pte_unmap_unlock(page_table, ptl);
2225
2226                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2227                         if (unlikely(tmp &
2228                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2229                                 ret = tmp;
2230                                 goto unwritable_page;
2231                         }
2232                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2233                                 lock_page(old_page);
2234                                 if (!old_page->mapping) {
2235                                         ret = 0; /* retry the fault */
2236                                         unlock_page(old_page);
2237                                         goto unwritable_page;
2238                                 }
2239                         } else
2240                                 VM_BUG_ON(!PageLocked(old_page));
2241
2242                         /*
2243                          * Since we dropped the lock we need to revalidate
2244                          * the PTE as someone else may have changed it.  If
2245                          * they did, we just return, as we can count on the
2246                          * MMU to tell us if they didn't also make it writable.
2247                          */
2248                         page_table = pte_offset_map_lock(mm, pmd, address,
2249                                                          &ptl);
2250                         if (!pte_same(*page_table, orig_pte)) {
2251                                 unlock_page(old_page);
2252                                 page_cache_release(old_page);
2253                                 goto unlock;
2254                         }
2255
2256                         page_mkwrite = 1;
2257                 }
2258                 dirty_page = old_page;
2259                 get_page(dirty_page);
2260
2261 reuse:
2262                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2263                 entry = pte_mkyoung(orig_pte);
2264                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2265                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2266                         update_mmu_cache(vma, address, page_table);
2267                 pte_unmap_unlock(page_table, ptl);
2268                 ret |= VM_FAULT_WRITE;
2269
2270                 if (!dirty_page)
2271                         return ret;
2272
2273                 /*
2274                  * Yes, Virginia, this is actually required to prevent a race
2275                  * with clear_page_dirty_for_io() from clearing the page dirty
2276                  * bit after it clear all dirty ptes, but before a racing
2277                  * do_wp_page installs a dirty pte.
2278                  *
2279                  * do_no_page is protected similarly.
2280                  */
2281                 if (!page_mkwrite) {
2282                         wait_on_page_locked(dirty_page);
2283                         set_page_dirty_balance(dirty_page, page_mkwrite);
2284                 }
2285                 put_page(dirty_page);
2286                 if (page_mkwrite) {
2287                         struct address_space *mapping = dirty_page->mapping;
2288
2289                         set_page_dirty(dirty_page);
2290                         unlock_page(dirty_page);
2291                         page_cache_release(dirty_page);
2292                         if (mapping)    {
2293                                 /*
2294                                  * Some device drivers do not set page.mapping
2295                                  * but still dirty their pages
2296                                  */
2297                                 balance_dirty_pages_ratelimited(mapping);
2298                         }
2299                 }
2300
2301                 /* file_update_time outside page_lock */
2302                 if (vma->vm_file)
2303                         file_update_time(vma->vm_file);
2304
2305                 return ret;
2306         }
2307
2308         /*
2309          * Ok, we need to copy. Oh, well..
2310          */
2311         page_cache_get(old_page);
2312 gotten:
2313         pte_unmap_unlock(page_table, ptl);
2314
2315         if (unlikely(anon_vma_prepare(vma)))
2316                 goto oom;
2317
2318         if (is_zero_pfn(pte_pfn(orig_pte))) {
2319                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2320                 if (!new_page)
2321                         goto oom;
2322         } else {
2323                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2324                 if (!new_page)
2325                         goto oom;
2326                 cow_user_page(new_page, old_page, address, vma);
2327         }
2328         __SetPageUptodate(new_page);
2329
2330         /*
2331          * Don't let another task, with possibly unlocked vma,
2332          * keep the mlocked page.
2333          */
2334         if ((vma->vm_flags & VM_LOCKED) && old_page) {
2335                 lock_page(old_page);    /* for LRU manipulation */
2336                 clear_page_mlock(old_page);
2337                 unlock_page(old_page);
2338         }
2339
2340         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2341                 goto oom_free_new;
2342
2343         /*
2344          * Re-check the pte - we dropped the lock
2345          */
2346         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2347         if (likely(pte_same(*page_table, orig_pte))) {
2348                 if (old_page) {
2349                         if (!PageAnon(old_page)) {
2350                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2351                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2352                         }
2353                 } else
2354                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2355                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2356                 entry = mk_pte(new_page, vma->vm_page_prot);
2357                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2358                 /*
2359                  * Clear the pte entry and flush it first, before updating the
2360                  * pte with the new entry. This will avoid a race condition
2361                  * seen in the presence of one thread doing SMC and another
2362                  * thread doing COW.
2363                  */
2364                 ptep_clear_flush(vma, address, page_table);
2365                 page_add_new_anon_rmap(new_page, vma, address);
2366                 /*
2367                  * We call the notify macro here because, when using secondary
2368                  * mmu page tables (such as kvm shadow page tables), we want the
2369                  * new page to be mapped directly into the secondary page table.
2370                  */
2371                 set_pte_at_notify(mm, address, page_table, entry);
2372                 update_mmu_cache(vma, address, page_table);
2373                 if (old_page) {
2374                         /*
2375                          * Only after switching the pte to the new page may
2376                          * we remove the mapcount here. Otherwise another
2377                          * process may come and find the rmap count decremented
2378                          * before the pte is switched to the new page, and
2379                          * "reuse" the old page writing into it while our pte
2380                          * here still points into it and can be read by other
2381                          * threads.
2382                          *
2383                          * The critical issue is to order this
2384                          * page_remove_rmap with the ptp_clear_flush above.
2385                          * Those stores are ordered by (if nothing else,)
2386                          * the barrier present in the atomic_add_negative
2387                          * in page_remove_rmap.
2388                          *
2389                          * Then the TLB flush in ptep_clear_flush ensures that
2390                          * no process can access the old page before the
2391                          * decremented mapcount is visible. And the old page
2392                          * cannot be reused until after the decremented
2393                          * mapcount is visible. So transitively, TLBs to
2394                          * old page will be flushed before it can be reused.
2395                          */
2396                         page_remove_rmap(old_page);
2397                 }
2398
2399                 /* Free the old page.. */
2400                 new_page = old_page;
2401                 ret |= VM_FAULT_WRITE;
2402         } else
2403                 mem_cgroup_uncharge_page(new_page);
2404
2405         if (new_page)
2406                 page_cache_release(new_page);
2407         if (old_page)
2408                 page_cache_release(old_page);
2409 unlock:
2410         pte_unmap_unlock(page_table, ptl);
2411         return ret;
2412 oom_free_new:
2413         page_cache_release(new_page);
2414 oom:
2415         if (old_page) {
2416                 if (page_mkwrite) {
2417                         unlock_page(old_page);
2418                         page_cache_release(old_page);
2419                 }
2420                 page_cache_release(old_page);
2421         }
2422         return VM_FAULT_OOM;
2423
2424 unwritable_page:
2425         page_cache_release(old_page);
2426         return ret;
2427 }
2428
2429 /*
2430  * Helper functions for unmap_mapping_range().
2431  *
2432  * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2433  *
2434  * We have to restart searching the prio_tree whenever we drop the lock,
2435  * since the iterator is only valid while the lock is held, and anyway
2436  * a later vma might be split and reinserted earlier while lock dropped.
2437  *
2438  * The list of nonlinear vmas could be handled more efficiently, using
2439  * a placeholder, but handle it in the same way until a need is shown.
2440  * It is important to search the prio_tree before nonlinear list: a vma
2441  * may become nonlinear and be shifted from prio_tree to nonlinear list
2442  * while the lock is dropped; but never shifted from list to prio_tree.
2443  *
2444  * In order to make forward progress despite restarting the search,
2445  * vm_truncate_count is used to mark a vma as now dealt with, so we can
2446  * quickly skip it next time around.  Since the prio_tree search only
2447  * shows us those vmas affected by unmapping the range in question, we
2448  * can't efficiently keep all vmas in step with mapping->truncate_count:
2449  * so instead reset them all whenever it wraps back to 0 (then go to 1).
2450  * mapping->truncate_count and vma->vm_truncate_count are protected by
2451  * i_mmap_lock.
2452  *
2453  * In order to make forward progress despite repeatedly restarting some
2454  * large vma, note the restart_addr from unmap_vmas when it breaks out:
2455  * and restart from that address when we reach that vma again.  It might
2456  * have been split or merged, shrunk or extended, but never shifted: so
2457  * restart_addr remains valid so long as it remains in the vma's range.
2458  * unmap_mapping_range forces truncate_count to leap over page-aligned
2459  * values so we can save vma's restart_addr in its truncate_count field.
2460  */
2461 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2462
2463 static void reset_vma_truncate_counts(struct address_space *mapping)
2464 {
2465         struct vm_area_struct *vma;
2466         struct prio_tree_iter iter;
2467
2468         vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2469                 vma->vm_truncate_count = 0;
2470         list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2471                 vma->vm_truncate_count = 0;
2472 }
2473
2474 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2475                 unsigned long start_addr, unsigned long end_addr,
2476                 struct zap_details *details)
2477 {
2478         unsigned long restart_addr;
2479         int need_break;
2480
2481         /*
2482          * files that support invalidating or truncating portions of the
2483          * file from under mmaped areas must have their ->fault function
2484          * return a locked page (and set VM_FAULT_LOCKED in the return).
2485          * This provides synchronisation against concurrent unmapping here.
2486          */
2487
2488 again:
2489         restart_addr = vma->vm_truncate_count;
2490         if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2491                 start_addr = restart_addr;
2492                 if (start_addr >= end_addr) {
2493                         /* Top of vma has been split off since last time */
2494                         vma->vm_truncate_count = details->truncate_count;
2495                         return 0;
2496                 }
2497         }
2498
2499         restart_addr = zap_page_range(vma, start_addr,
2500                                         end_addr - start_addr, details);
2501         need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2502
2503         if (restart_addr >= end_addr) {
2504                 /* We have now completed this vma: mark it so */
2505                 vma->vm_truncate_count = details->truncate_count;
2506                 if (!need_break)
2507                         return 0;
2508         } else {
2509                 /* Note restart_addr in vma's truncate_count field */
2510                 vma->vm_truncate_count = restart_addr;
2511                 if (!need_break)
2512                         goto again;
2513         }
2514
2515         spin_unlock(details->i_mmap_lock);
2516         cond_resched();
2517         spin_lock(details->i_mmap_lock);
2518         return -EINTR;
2519 }
2520
2521 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2522                                             struct zap_details *details)
2523 {
2524         struct vm_area_struct *vma;
2525         struct prio_tree_iter iter;
2526         pgoff_t vba, vea, zba, zea;
2527
2528 restart:
2529         vma_prio_tree_foreach(vma, &iter, root,
2530                         details->first_index, details->last_index) {
2531                 /* Skip quickly over those we have already dealt with */
2532                 if (vma->vm_truncate_count == details->truncate_count)
2533                         continue;
2534
2535                 vba = vma->vm_pgoff;
2536                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2537                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2538                 zba = details->first_index;
2539                 if (zba < vba)
2540                         zba = vba;
2541                 zea = details->last_index;
2542                 if (zea > vea)
2543                         zea = vea;
2544
2545                 if (unmap_mapping_range_vma(vma,
2546                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2547                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2548                                 details) < 0)
2549                         goto restart;
2550         }
2551 }
2552
2553 static inline void unmap_mapping_range_list(struct list_head *head,
2554                                             struct zap_details *details)
2555 {
2556         struct vm_area_struct *vma;
2557
2558         /*
2559          * In nonlinear VMAs there is no correspondence between virtual address
2560          * offset and file offset.  So we must perform an exhaustive search
2561          * across *all* the pages in each nonlinear VMA, not just the pages
2562          * whose virtual address lies outside the file truncation point.
2563          */
2564 restart:
2565         list_for_each_entry(vma, head, shared.vm_set.list) {
2566                 /* Skip quickly over those we have already dealt with */
2567                 if (vma->vm_truncate_count == details->truncate_count)
2568                         continue;
2569                 details->nonlinear_vma = vma;
2570                 if (unmap_mapping_range_vma(vma, vma->vm_start,
2571                                         vma->vm_end, details) < 0)
2572                         goto restart;
2573         }
2574 }
2575
2576 /**
2577  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2578  * @mapping: the address space containing mmaps to be unmapped.
2579  * @holebegin: byte in first page to unmap, relative to the start of
2580  * the underlying file.  This will be rounded down to a PAGE_SIZE
2581  * boundary.  Note that this is different from truncate_pagecache(), which
2582  * must keep the partial page.  In contrast, we must get rid of
2583  * partial pages.
2584  * @holelen: size of prospective hole in bytes.  This will be rounded
2585  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2586  * end of the file.
2587  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2588  * but 0 when invalidating pagecache, don't throw away private data.
2589  */
2590 void unmap_mapping_range(struct address_space *mapping,
2591                 loff_t const holebegin, loff_t const holelen, int even_cows)
2592 {
2593         struct zap_details details;
2594         pgoff_t hba = holebegin >> PAGE_SHIFT;
2595         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2596
2597         /* Check for overflow. */
2598         if (sizeof(holelen) > sizeof(hlen)) {
2599                 long long holeend =
2600                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2601                 if (holeend & ~(long long)ULONG_MAX)
2602                         hlen = ULONG_MAX - hba + 1;
2603         }
2604
2605         details.check_mapping = even_cows? NULL: mapping;
2606         details.nonlinear_vma = NULL;
2607         details.first_index = hba;
2608         details.last_index = hba + hlen - 1;
2609         if (details.last_index < details.first_index)
2610                 details.last_index = ULONG_MAX;
2611         details.i_mmap_lock = &mapping->i_mmap_lock;
2612
2613         spin_lock(&mapping->i_mmap_lock);
2614
2615         /* Protect against endless unmapping loops */
2616         mapping->truncate_count++;
2617         if (unlikely(is_restart_addr(mapping->truncate_count))) {
2618                 if (mapping->truncate_count == 0)
2619                         reset_vma_truncate_counts(mapping);
2620                 mapping->truncate_count++;
2621         }
2622         details.truncate_count = mapping->truncate_count;
2623
2624         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2625                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2626         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2627                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2628         spin_unlock(&mapping->i_mmap_lock);
2629 }
2630 EXPORT_SYMBOL(unmap_mapping_range);
2631
2632 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2633 {
2634         struct address_space *mapping = inode->i_mapping;
2635
2636         /*
2637          * If the underlying filesystem is not going to provide
2638          * a way to truncate a range of blocks (punch a hole) -
2639          * we should return failure right now.
2640          */
2641         if (!inode->i_op->truncate_range)
2642                 return -ENOSYS;
2643
2644         mutex_lock(&inode->i_mutex);
2645         down_write(&inode->i_alloc_sem);
2646         unmap_mapping_range(mapping, offset, (end - offset), 1);
2647         truncate_inode_pages_range(mapping, offset, end);
2648         unmap_mapping_range(mapping, offset, (end - offset), 1);
2649         inode->i_op->truncate_range(inode, offset, end);
2650         up_write(&inode->i_alloc_sem);
2651         mutex_unlock(&inode->i_mutex);
2652
2653         return 0;
2654 }
2655
2656 /*
2657  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2658  * but allow concurrent faults), and pte mapped but not yet locked.
2659  * We return with mmap_sem still held, but pte unmapped and unlocked.
2660  */
2661 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2662                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2663                 unsigned int flags, pte_t orig_pte)
2664 {
2665         spinlock_t *ptl;
2666         struct page *page, *swapcache = NULL;
2667         swp_entry_t entry;
2668         pte_t pte;
2669         int locked;
2670         struct mem_cgroup *ptr = NULL;
2671         int exclusive = 0;
2672         int ret = 0;
2673
2674         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2675                 goto out;
2676
2677         entry = pte_to_swp_entry(orig_pte);
2678         if (unlikely(non_swap_entry(entry))) {
2679                 if (is_migration_entry(entry)) {
2680                         migration_entry_wait(mm, pmd, address);
2681                 } else if (is_hwpoison_entry(entry)) {
2682                         ret = VM_FAULT_HWPOISON;
2683                 } else {
2684                         print_bad_pte(vma, address, orig_pte, NULL);
2685                         ret = VM_FAULT_SIGBUS;
2686                 }
2687                 goto out;
2688         }
2689         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2690         page = lookup_swap_cache(entry);
2691         if (!page) {
2692                 grab_swap_token(mm); /* Contend for token _before_ read-in */
2693                 page = swapin_readahead(entry,
2694                                         GFP_HIGHUSER_MOVABLE, vma, address);
2695                 if (!page) {
2696                         /*
2697                          * Back out if somebody else faulted in this pte
2698                          * while we released the pte lock.
2699                          */
2700                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2701                         if (likely(pte_same(*page_table, orig_pte)))
2702                                 ret = VM_FAULT_OOM;
2703                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2704                         goto unlock;
2705                 }
2706
2707                 /* Had to read the page from swap area: Major fault */
2708                 ret = VM_FAULT_MAJOR;
2709                 count_vm_event(PGMAJFAULT);
2710         } else if (PageHWPoison(page)) {
2711                 /*
2712                  * hwpoisoned dirty swapcache pages are kept for killing
2713                  * owner processes (which may be unknown at hwpoison time)
2714                  */
2715                 ret = VM_FAULT_HWPOISON;
2716                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2717                 goto out_release;
2718         }
2719
2720         locked = lock_page_or_retry(page, mm, flags);
2721         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2722         if (!locked) {
2723                 ret |= VM_FAULT_RETRY;
2724                 goto out_release;
2725         }
2726
2727         /*
2728          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2729          * release the swapcache from under us.  The page pin, and pte_same
2730          * test below, are not enough to exclude that.  Even if it is still
2731          * swapcache, we need to check that the page's swap has not changed.
2732          */
2733         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2734                 goto out_page;
2735
2736         if (ksm_might_need_to_copy(page, vma, address)) {
2737                 swapcache = page;
2738                 page = ksm_does_need_to_copy(page, vma, address);
2739
2740                 if (unlikely(!page)) {
2741                         ret = VM_FAULT_OOM;
2742                         page = swapcache;
2743                         swapcache = NULL;
2744                         goto out_page;
2745                 }
2746         }
2747
2748         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2749                 ret = VM_FAULT_OOM;
2750                 goto out_page;
2751         }
2752
2753         /*
2754          * Back out if somebody else already faulted in this pte.
2755          */
2756         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2757         if (unlikely(!pte_same(*page_table, orig_pte)))
2758                 goto out_nomap;
2759
2760         if (unlikely(!PageUptodate(page))) {
2761                 ret = VM_FAULT_SIGBUS;
2762                 goto out_nomap;
2763         }
2764
2765         /*
2766          * The page isn't present yet, go ahead with the fault.
2767          *
2768          * Be careful about the sequence of operations here.
2769          * To get its accounting right, reuse_swap_page() must be called
2770          * while the page is counted on swap but not yet in mapcount i.e.
2771          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2772          * must be called after the swap_free(), or it will never succeed.
2773          * Because delete_from_swap_page() may be called by reuse_swap_page(),
2774          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2775          * in page->private. In this case, a record in swap_cgroup  is silently
2776          * discarded at swap_free().
2777          */
2778
2779         inc_mm_counter_fast(mm, MM_ANONPAGES);
2780         dec_mm_counter_fast(mm, MM_SWAPENTS);
2781         pte = mk_pte(page, vma->vm_page_prot);
2782         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2783                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2784                 flags &= ~FAULT_FLAG_WRITE;
2785                 ret |= VM_FAULT_WRITE;
2786                 exclusive = 1;
2787         }
2788         flush_icache_page(vma, page);
2789         set_pte_at(mm, address, page_table, pte);
2790         do_page_add_anon_rmap(page, vma, address, exclusive);
2791         /* It's better to call commit-charge after rmap is established */
2792         mem_cgroup_commit_charge_swapin(page, ptr);
2793
2794         swap_free(entry);
2795         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2796                 try_to_free_swap(page);
2797         unlock_page(page);
2798         if (swapcache) {
2799                 /*
2800                  * Hold the lock to avoid the swap entry to be reused
2801                  * until we take the PT lock for the pte_same() check
2802                  * (to avoid false positives from pte_same). For
2803                  * further safety release the lock after the swap_free
2804                  * so that the swap count won't change under a
2805                  * parallel locked swapcache.
2806                  */
2807                 unlock_page(swapcache);
2808                 page_cache_release(swapcache);
2809         }
2810
2811         if (flags & FAULT_FLAG_WRITE) {
2812                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2813                 if (ret & VM_FAULT_ERROR)
2814                         ret &= VM_FAULT_ERROR;
2815                 goto out;
2816         }
2817
2818         /* No need to invalidate - it was non-present before */
2819         update_mmu_cache(vma, address, page_table);
2820 unlock:
2821         pte_unmap_unlock(page_table, ptl);
2822 out:
2823         return ret;
2824 out_nomap:
2825         mem_cgroup_cancel_charge_swapin(ptr);
2826         pte_unmap_unlock(page_table, ptl);
2827 out_page:
2828         unlock_page(page);
2829 out_release:
2830         page_cache_release(page);
2831         if (swapcache) {
2832                 unlock_page(swapcache);
2833                 page_cache_release(swapcache);
2834         }
2835         return ret;
2836 }
2837
2838 /*
2839  * This is like a special single-page "expand_{down|up}wards()",
2840  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2841  * doesn't hit another vma.
2842  */
2843 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2844 {
2845         address &= PAGE_MASK;
2846         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2847                 struct vm_area_struct *prev = vma->vm_prev;
2848
2849                 /*
2850                  * Is there a mapping abutting this one below?
2851                  *
2852                  * That's only ok if it's the same stack mapping
2853                  * that has gotten split..
2854                  */
2855                 if (prev && prev->vm_end == address)
2856                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2857
2858                 expand_stack(vma, address - PAGE_SIZE);
2859         }
2860         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2861                 struct vm_area_struct *next = vma->vm_next;
2862
2863                 /* As VM_GROWSDOWN but s/below/above/ */
2864                 if (next && next->vm_start == address + PAGE_SIZE)
2865                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2866
2867                 expand_upwards(vma, address + PAGE_SIZE);
2868         }
2869         return 0;
2870 }
2871
2872 /*
2873  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2874  * but allow concurrent faults), and pte mapped but not yet locked.
2875  * We return with mmap_sem still held, but pte unmapped and unlocked.
2876  */
2877 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2878                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2879                 unsigned int flags)
2880 {
2881         struct page *page;
2882         spinlock_t *ptl;
2883         pte_t entry;
2884
2885         pte_unmap(page_table);
2886
2887         /* Check if we need to add a guard page to the stack */
2888         if (check_stack_guard_page(vma, address) < 0)
2889                 return VM_FAULT_SIGBUS;
2890
2891         /* Use the zero-page for reads */
2892         if (!(flags & FAULT_FLAG_WRITE)) {
2893                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2894                                                 vma->vm_page_prot));
2895                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2896                 if (!pte_none(*page_table))
2897                         goto unlock;
2898                 goto setpte;
2899         }
2900
2901         /* Allocate our own private page. */
2902         if (unlikely(anon_vma_prepare(vma)))
2903                 goto oom;
2904         page = alloc_zeroed_user_highpage_movable(vma, address);
2905         if (!page)
2906                 goto oom;
2907         __SetPageUptodate(page);
2908
2909         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2910                 goto oom_free_page;
2911
2912         entry = mk_pte(page, vma->vm_page_prot);
2913         if (vma->vm_flags & VM_WRITE)
2914                 entry = pte_mkwrite(pte_mkdirty(entry));
2915
2916         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2917         if (!pte_none(*page_table))
2918                 goto release;
2919
2920         inc_mm_counter_fast(mm, MM_ANONPAGES);
2921         page_add_new_anon_rmap(page, vma, address);
2922 setpte:
2923         set_pte_at(mm, address, page_table, entry);
2924
2925         /* No need to invalidate - it was non-present before */
2926         update_mmu_cache(vma, address, page_table);
2927 unlock:
2928         pte_unmap_unlock(page_table, ptl);
2929         return 0;
2930 release:
2931         mem_cgroup_uncharge_page(page);
2932         page_cache_release(page);
2933         goto unlock;
2934 oom_free_page:
2935         page_cache_release(page);
2936 oom:
2937         return VM_FAULT_OOM;
2938 }
2939
2940 /*
2941  * __do_fault() tries to create a new page mapping. It aggressively
2942  * tries to share with existing pages, but makes a separate copy if
2943  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2944  * the next page fault.
2945  *
2946  * As this is called only for pages that do not currently exist, we
2947  * do not need to flush old virtual caches or the TLB.
2948  *
2949  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2950  * but allow concurrent faults), and pte neither mapped nor locked.
2951  * We return with mmap_sem still held, but pte unmapped and unlocked.
2952  */
2953 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2954                 unsigned long address, pmd_t *pmd,
2955                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2956 {
2957         pte_t *page_table;
2958         spinlock_t *ptl;
2959         struct page *page;
2960         pte_t entry;
2961         int anon = 0;
2962         int charged = 0;
2963         struct page *dirty_page = NULL;
2964         struct vm_fault vmf;
2965         int ret;
2966         int page_mkwrite = 0;
2967
2968         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2969         vmf.pgoff = pgoff;
2970         vmf.flags = flags;
2971         vmf.page = NULL;
2972
2973         ret = vma->vm_ops->fault(vma, &vmf);
2974         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
2975                             VM_FAULT_RETRY)))
2976                 return ret;
2977
2978         if (unlikely(PageHWPoison(vmf.page))) {
2979                 if (ret & VM_FAULT_LOCKED)
2980                         unlock_page(vmf.page);
2981                 return VM_FAULT_HWPOISON;
2982         }
2983
2984         /*
2985          * For consistency in subsequent calls, make the faulted page always
2986          * locked.
2987          */
2988         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2989                 lock_page(vmf.page);
2990         else
2991                 VM_BUG_ON(!PageLocked(vmf.page));
2992
2993         /*
2994          * Should we do an early C-O-W break?
2995          */
2996         page = vmf.page;
2997         if (flags & FAULT_FLAG_WRITE) {
2998                 if (!(vma->vm_flags & VM_SHARED)) {
2999                         anon = 1;
3000                         if (unlikely(anon_vma_prepare(vma))) {
3001                                 ret = VM_FAULT_OOM;
3002                                 goto out;
3003                         }
3004                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3005                                                 vma, address);
3006                         if (!page) {
3007                                 ret = VM_FAULT_OOM;
3008                                 goto out;
3009                         }
3010                         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3011                                 ret = VM_FAULT_OOM;
3012                                 page_cache_release(page);
3013                                 goto out;
3014                         }
3015                         charged = 1;
3016                         /*
3017                          * Don't let another task, with possibly unlocked vma,
3018                          * keep the mlocked page.
3019                          */
3020                         if (vma->vm_flags & VM_LOCKED)
3021                                 clear_page_mlock(vmf.page);
3022                         copy_user_highpage(page, vmf.page, address, vma);
3023                         __SetPageUptodate(page);
3024                 } else {
3025                         /*
3026                          * If the page will be shareable, see if the backing
3027                          * address space wants to know that the page is about
3028                          * to become writable
3029                          */
3030                         if (vma->vm_ops->page_mkwrite) {
3031                                 int tmp;
3032
3033                                 unlock_page(page);
3034                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3035                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3036                                 if (unlikely(tmp &
3037                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3038                                         ret = tmp;
3039                                         goto unwritable_page;
3040                                 }
3041                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3042                                         lock_page(page);
3043                                         if (!page->mapping) {
3044                                                 ret = 0; /* retry the fault */
3045                                                 unlock_page(page);
3046                                                 goto unwritable_page;
3047                                         }
3048                                 } else
3049                                         VM_BUG_ON(!PageLocked(page));
3050                                 page_mkwrite = 1;
3051                         }
3052                 }
3053
3054         }
3055
3056         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3057
3058         /*
3059          * This silly early PAGE_DIRTY setting removes a race
3060          * due to the bad i386 page protection. But it's valid
3061          * for other architectures too.
3062          *
3063          * Note that if FAULT_FLAG_WRITE is set, we either now have
3064          * an exclusive copy of the page, or this is a shared mapping,
3065          * so we can make it writable and dirty to avoid having to
3066          * handle that later.
3067          */
3068         /* Only go through if we didn't race with anybody else... */
3069         if (likely(pte_same(*page_table, orig_pte))) {
3070                 flush_icache_page(vma, page);
3071                 entry = mk_pte(page, vma->vm_page_prot);
3072                 if (flags & FAULT_FLAG_WRITE)
3073                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3074                 if (anon) {
3075                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3076                         page_add_new_anon_rmap(page, vma, address);
3077                 } else {
3078                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3079                         page_add_file_rmap(page);
3080                         if (flags & FAULT_FLAG_WRITE) {
3081                                 dirty_page = page;
3082                                 get_page(dirty_page);
3083                         }
3084                 }
3085                 set_pte_at(mm, address, page_table, entry);
3086
3087                 /* no need to invalidate: a not-present page won't be cached */
3088                 update_mmu_cache(vma, address, page_table);
3089         } else {
3090                 if (charged)
3091                         mem_cgroup_uncharge_page(page);
3092                 if (anon)
3093                         page_cache_release(page);
3094                 else
3095                         anon = 1; /* no anon but release faulted_page */
3096         }
3097
3098         pte_unmap_unlock(page_table, ptl);
3099
3100 out:
3101         if (dirty_page) {
3102                 struct address_space *mapping = page->mapping;
3103
3104                 if (set_page_dirty(dirty_page))
3105                         page_mkwrite = 1;
3106                 unlock_page(dirty_page);
3107                 put_page(dirty_page);
3108                 if (page_mkwrite && mapping) {
3109                         /*
3110                          * Some device drivers do not set page.mapping but still
3111                          * dirty their pages
3112                          */
3113                         balance_dirty_pages_ratelimited(mapping);
3114                 }
3115
3116                 /* file_update_time outside page_lock */
3117                 if (vma->vm_file)
3118                         file_update_time(vma->vm_file);
3119         } else {
3120                 unlock_page(vmf.page);
3121                 if (anon)
3122                         page_cache_release(vmf.page);
3123         }
3124
3125         return ret;
3126
3127 unwritable_page:
3128         page_cache_release(page);
3129         return ret;
3130 }
3131
3132 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3133                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3134                 unsigned int flags, pte_t orig_pte)
3135 {
3136         pgoff_t pgoff = (((address & PAGE_MASK)
3137                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3138
3139         pte_unmap(page_table);
3140         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3141 }
3142
3143 /*
3144  * Fault of a previously existing named mapping. Repopulate the pte
3145  * from the encoded file_pte if possible. This enables swappable
3146  * nonlinear vmas.
3147  *
3148  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3149  * but allow concurrent faults), and pte mapped but not yet locked.
3150  * We return with mmap_sem still held, but pte unmapped and unlocked.
3151  */
3152 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3153                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3154                 unsigned int flags, pte_t orig_pte)
3155 {
3156         pgoff_t pgoff;
3157
3158         flags |= FAULT_FLAG_NONLINEAR;
3159
3160         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3161                 return 0;
3162
3163         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3164                 /*
3165                  * Page table corrupted: show pte and kill process.
3166                  */
3167                 print_bad_pte(vma, address, orig_pte, NULL);
3168                 return VM_FAULT_SIGBUS;
3169         }
3170
3171         pgoff = pte_to_pgoff(orig_pte);
3172         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3173 }
3174
3175 /*
3176  * These routines also need to handle stuff like marking pages dirty
3177  * and/or accessed for architectures that don't do it in hardware (most
3178  * RISC architectures).  The early dirtying is also good on the i386.
3179  *
3180  * There is also a hook called "update_mmu_cache()" that architectures
3181  * with external mmu caches can use to update those (ie the Sparc or
3182  * PowerPC hashed page tables that act as extended TLBs).
3183  *
3184  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3185  * but allow concurrent faults), and pte mapped but not yet locked.
3186  * We return with mmap_sem still held, but pte unmapped and unlocked.
3187  */
3188 static inline int handle_pte_fault(struct mm_struct *mm,
3189                 struct vm_area_struct *vma, unsigned long address,
3190                 pte_t *pte, pmd_t *pmd, unsigned int flags)
3191 {
3192         pte_t entry;
3193         spinlock_t *ptl;
3194
3195         entry = *pte;
3196         if (!pte_present(entry)) {
3197                 if (pte_none(entry)) {
3198                         if (vma->vm_ops) {
3199                                 if (likely(vma->vm_ops->fault))
3200                                         return do_linear_fault(mm, vma, address,
3201                                                 pte, pmd, flags, entry);
3202                         }
3203                         return do_anonymous_page(mm, vma, address,
3204                                                  pte, pmd, flags);
3205                 }
3206                 if (pte_file(entry))
3207                         return do_nonlinear_fault(mm, vma, address,
3208                                         pte, pmd, flags, entry);
3209                 return do_swap_page(mm, vma, address,
3210                                         pte, pmd, flags, entry);
3211         }
3212
3213         ptl = pte_lockptr(mm, pmd);
3214         spin_lock(ptl);
3215         if (unlikely(!pte_same(*pte, entry)))
3216                 goto unlock;
3217         if (flags & FAULT_FLAG_WRITE) {
3218                 if (!pte_write(entry))
3219                         return do_wp_page(mm, vma, address,
3220                                         pte, pmd, ptl, entry);
3221                 entry = pte_mkdirty(entry);
3222         }
3223         entry = pte_mkyoung(entry);
3224         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3225                 update_mmu_cache(vma, address, pte);
3226         } else {
3227                 /*
3228                  * This is needed only for protection faults but the arch code
3229                  * is not yet telling us if this is a protection fault or not.
3230                  * This still avoids useless tlb flushes for .text page faults
3231                  * with threads.
3232                  */
3233                 if (flags & FAULT_FLAG_WRITE)
3234                         flush_tlb_fix_spurious_fault(vma, address);
3235         }
3236 unlock:
3237         pte_unmap_unlock(pte, ptl);
3238         return 0;
3239 }
3240
3241 /*
3242  * By the time we get here, we already hold the mm semaphore
3243  */
3244 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3245                 unsigned long address, unsigned int flags)
3246 {
3247         pgd_t *pgd;
3248         pud_t *pud;
3249         pmd_t *pmd;
3250         pte_t *pte;
3251
3252         __set_current_state(TASK_RUNNING);
3253
3254         count_vm_event(PGFAULT);
3255
3256         /* do counter updates before entering really critical section. */
3257         check_sync_rss_stat(current);
3258
3259         if (unlikely(is_vm_hugetlb_page(vma)))
3260                 return hugetlb_fault(mm, vma, address, flags);
3261
3262         pgd = pgd_offset(mm, address);
3263         pud = pud_alloc(mm, pgd, address);
3264         if (!pud)
3265                 return VM_FAULT_OOM;
3266         pmd = pmd_alloc(mm, pud, address);
3267         if (!pmd)
3268                 return VM_FAULT_OOM;
3269         pte = pte_alloc_map(mm, pmd, address);
3270         if (!pte)
3271                 return VM_FAULT_OOM;
3272
3273         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3274 }
3275
3276 #ifndef __PAGETABLE_PUD_FOLDED
3277 /*
3278  * Allocate page upper directory.
3279  * We've already handled the fast-path in-line.
3280  */
3281 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3282 {
3283         pud_t *new = pud_alloc_one(mm, address);
3284         if (!new)
3285                 return -ENOMEM;
3286
3287         smp_wmb(); /* See comment in __pte_alloc */
3288
3289         spin_lock(&mm->page_table_lock);
3290         if (pgd_present(*pgd))          /* Another has populated it */
3291                 pud_free(mm, new);
3292         else
3293                 pgd_populate(mm, pgd, new);
3294         spin_unlock(&mm->page_table_lock);
3295         return 0;
3296 }
3297 #endif /* __PAGETABLE_PUD_FOLDED */
3298
3299 #ifndef __PAGETABLE_PMD_FOLDED
3300 /*
3301  * Allocate page middle directory.
3302  * We've already handled the fast-path in-line.
3303  */
3304 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3305 {
3306         pmd_t *new = pmd_alloc_one(mm, address);
3307         if (!new)
3308                 return -ENOMEM;
3309
3310         smp_wmb(); /* See comment in __pte_alloc */
3311
3312         spin_lock(&mm->page_table_lock);
3313 #ifndef __ARCH_HAS_4LEVEL_HACK
3314         if (pud_present(*pud))          /* Another has populated it */
3315                 pmd_free(mm, new);
3316         else
3317                 pud_populate(mm, pud, new);
3318 #else
3319         if (pgd_present(*pud))          /* Another has populated it */
3320                 pmd_free(mm, new);
3321         else
3322                 pgd_populate(mm, pud, new);
3323 #endif /* __ARCH_HAS_4LEVEL_HACK */
3324         spin_unlock(&mm->page_table_lock);
3325         return 0;
3326 }
3327 #endif /* __PAGETABLE_PMD_FOLDED */
3328
3329 int make_pages_present(unsigned long addr, unsigned long end)
3330 {
3331         int ret, len, write;
3332         struct vm_area_struct * vma;
3333
3334         vma = find_vma(current->mm, addr);
3335         if (!vma)
3336                 return -ENOMEM;
3337         /*
3338          * We want to touch writable mappings with a write fault in order
3339          * to break COW, except for shared mappings because these don't COW
3340          * and we would not want to dirty them for nothing.
3341          */
3342         write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3343         BUG_ON(addr >= end);
3344         BUG_ON(end > vma->vm_end);
3345         len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3346         ret = get_user_pages(current, current->mm, addr,
3347                         len, write, 0, NULL, NULL);
3348         if (ret < 0)
3349                 return ret;
3350         return ret == len ? 0 : -EFAULT;
3351 }
3352
3353 #if !defined(__HAVE_ARCH_GATE_AREA)
3354
3355 #if defined(AT_SYSINFO_EHDR)
3356 static struct vm_area_struct gate_vma;
3357
3358 static int __init gate_vma_init(void)
3359 {
3360         gate_vma.vm_mm = NULL;
3361         gate_vma.vm_start = FIXADDR_USER_START;
3362         gate_vma.vm_end = FIXADDR_USER_END;
3363         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3364         gate_vma.vm_page_prot = __P101;
3365         /*
3366          * Make sure the vDSO gets into every core dump.
3367          * Dumping its contents makes post-mortem fully interpretable later
3368          * without matching up the same kernel and hardware config to see
3369          * what PC values meant.
3370          */
3371         gate_vma.vm_flags |= VM_ALWAYSDUMP;
3372         return 0;
3373 }
3374 __initcall(gate_vma_init);
3375 #endif
3376
3377 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3378 {
3379 #ifdef AT_SYSINFO_EHDR
3380         return &gate_vma;
3381 #else
3382         return NULL;
3383 #endif
3384 }
3385
3386 int in_gate_area_no_task(unsigned long addr)
3387 {
3388 #ifdef AT_SYSINFO_EHDR
3389         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3390                 return 1;
3391 #endif
3392         return 0;
3393 }
3394
3395 #endif  /* __HAVE_ARCH_GATE_AREA */
3396
3397 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3398                 pte_t **ptepp, spinlock_t **ptlp)
3399 {
3400         pgd_t *pgd;
3401         pud_t *pud;
3402         pmd_t *pmd;
3403         pte_t *ptep;
3404
3405         pgd = pgd_offset(mm, address);
3406         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3407                 goto out;
3408
3409         pud = pud_offset(pgd, address);
3410         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3411                 goto out;
3412
3413         pmd = pmd_offset(pud, address);
3414         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3415                 goto out;
3416
3417         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3418         if (pmd_huge(*pmd))
3419                 goto out;
3420
3421         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3422         if (!ptep)
3423                 goto out;
3424         if (!pte_present(*ptep))
3425                 goto unlock;
3426         *ptepp = ptep;
3427         return 0;
3428 unlock:
3429         pte_unmap_unlock(ptep, *ptlp);
3430 out:
3431         return -EINVAL;
3432 }
3433
3434 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3435                              pte_t **ptepp, spinlock_t **ptlp)
3436 {
3437         int res;
3438
3439         /* (void) is needed to make gcc happy */
3440         (void) __cond_lock(*ptlp,
3441                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3442         return res;
3443 }
3444
3445 /**
3446  * follow_pfn - look up PFN at a user virtual address
3447  * @vma: memory mapping
3448  * @address: user virtual address
3449  * @pfn: location to store found PFN
3450  *
3451  * Only IO mappings and raw PFN mappings are allowed.
3452  *
3453  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3454  */
3455 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3456         unsigned long *pfn)
3457 {
3458         int ret = -EINVAL;
3459         spinlock_t *ptl;
3460         pte_t *ptep;
3461
3462         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3463                 return ret;
3464
3465         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3466         if (ret)
3467                 return ret;
3468         *pfn = pte_pfn(*ptep);
3469         pte_unmap_unlock(ptep, ptl);
3470         return 0;
3471 }
3472 EXPORT_SYMBOL(follow_pfn);
3473
3474 #ifdef CONFIG_HAVE_IOREMAP_PROT
3475 int follow_phys(struct vm_area_struct *vma,
3476                 unsigned long address, unsigned int flags,
3477                 unsigned long *prot, resource_size_t *phys)
3478 {
3479         int ret = -EINVAL;
3480         pte_t *ptep, pte;
3481         spinlock_t *ptl;
3482
3483         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3484                 goto out;
3485
3486         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3487                 goto out;
3488         pte = *ptep;
3489
3490         if ((flags & FOLL_WRITE) && !pte_write(pte))
3491                 goto unlock;
3492
3493         *prot = pgprot_val(pte_pgprot(pte));
3494         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3495
3496         ret = 0;
3497 unlock:
3498         pte_unmap_unlock(ptep, ptl);
3499 out:
3500         return ret;
3501 }
3502
3503 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3504                         void *buf, int len, int write)
3505 {
3506         resource_size_t phys_addr;
3507         unsigned long prot = 0;
3508         void __iomem *maddr;
3509         int offset = addr & (PAGE_SIZE-1);
3510
3511         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3512                 return -EINVAL;
3513
3514         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3515         if (write)
3516                 memcpy_toio(maddr + offset, buf, len);
3517         else
3518                 memcpy_fromio(buf, maddr + offset, len);
3519         iounmap(maddr);
3520
3521         return len;
3522 }
3523 #endif
3524
3525 /*
3526  * Access another process' address space.
3527  * Source/target buffer must be kernel space,
3528  * Do not walk the page table directly, use get_user_pages
3529  */
3530 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3531 {
3532         struct mm_struct *mm;
3533         struct vm_area_struct *vma;
3534         void *old_buf = buf;
3535
3536         mm = get_task_mm(tsk);
3537         if (!mm)
3538                 return 0;
3539
3540         down_read(&mm->mmap_sem);
3541         /* ignore errors, just check how much was successfully transferred */
3542         while (len) {
3543                 int bytes, ret, offset;
3544                 void *maddr;
3545                 struct page *page = NULL;
3546
3547                 ret = get_user_pages(tsk, mm, addr, 1,
3548                                 write, 1, &page, &vma);
3549                 if (ret <= 0) {
3550                         /*
3551                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3552                          * we can access using slightly different code.
3553                          */
3554 #ifdef CONFIG_HAVE_IOREMAP_PROT
3555                         vma = find_vma(mm, addr);
3556                         if (!vma)
3557                                 break;
3558                         if (vma->vm_ops && vma->vm_ops->access)
3559                                 ret = vma->vm_ops->access(vma, addr, buf,
3560                                                           len, write);
3561                         if (ret <= 0)
3562 #endif
3563                                 break;
3564                         bytes = ret;
3565                 } else {
3566                         bytes = len;
3567                         offset = addr & (PAGE_SIZE-1);
3568                         if (bytes > PAGE_SIZE-offset)
3569                                 bytes = PAGE_SIZE-offset;
3570
3571                         maddr = kmap(page);
3572                         if (write) {
3573                                 copy_to_user_page(vma, page, addr,
3574                                                   maddr + offset, buf, bytes);
3575                                 set_page_dirty_lock(page);
3576                         } else {
3577                                 copy_from_user_page(vma, page, addr,
3578                                                     buf, maddr + offset, bytes);
3579                         }
3580                         kunmap(page);
3581                         page_cache_release(page);
3582                 }
3583                 len -= bytes;
3584                 buf += bytes;
3585                 addr += bytes;
3586         }
3587         up_read(&mm->mmap_sem);
3588         mmput(mm);
3589
3590         return buf - old_buf;
3591 }
3592
3593 /*
3594  * Print the name of a VMA.
3595  */
3596 void print_vma_addr(char *prefix, unsigned long ip)
3597 {
3598         struct mm_struct *mm = current->mm;
3599         struct vm_area_struct *vma;
3600
3601         /*
3602          * Do not print if we are in atomic
3603          * contexts (in exception stacks, etc.):
3604          */
3605         if (preempt_count())
3606                 return;
3607
3608         down_read(&mm->mmap_sem);
3609         vma = find_vma(mm, ip);
3610         if (vma && vma->vm_file) {
3611                 struct file *f = vma->vm_file;
3612                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3613                 if (buf) {
3614                         char *p, *s;
3615
3616                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3617                         if (IS_ERR(p))
3618                                 p = "?";
3619                         s = strrchr(p, '/');
3620                         if (s)
3621                                 p = s+1;
3622                         printk("%s%s[%lx+%lx]", prefix, p,
3623                                         vma->vm_start,
3624                                         vma->vm_end - vma->vm_start);
3625                         free_page((unsigned long)buf);
3626                 }
3627         }
3628         up_read(&current->mm->mmap_sem);
3629 }
3630
3631 #ifdef CONFIG_PROVE_LOCKING
3632 void might_fault(void)
3633 {
3634         /*
3635          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3636          * holding the mmap_sem, this is safe because kernel memory doesn't
3637          * get paged out, therefore we'll never actually fault, and the
3638          * below annotations will generate false positives.
3639          */
3640         if (segment_eq(get_fs(), KERNEL_DS))
3641                 return;
3642
3643         might_sleep();
3644         /*
3645          * it would be nicer only to annotate paths which are not under
3646          * pagefault_disable, however that requires a larger audit and
3647          * providing helpers like get_user_atomic.
3648          */
3649         if (!in_atomic() && current->mm)
3650                 might_lock_read(&current->mm->mmap_sem);
3651 }
3652 EXPORT_SYMBOL(might_fault);
3653 #endif