Merge remote-tracking branch 'asoc/fix/wm8962' into asoc-linus
[platform/adaptation/renesas_rcar/renesas_kernel.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.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 #include <linux/migrate.h>
61 #include <linux/string.h>
62
63 #include <asm/io.h>
64 #include <asm/pgalloc.h>
65 #include <asm/uaccess.h>
66 #include <asm/tlb.h>
67 #include <asm/tlbflush.h>
68 #include <asm/pgtable.h>
69
70 #include "internal.h"
71
72 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
73 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
74 #endif
75
76 #ifndef CONFIG_NEED_MULTIPLE_NODES
77 /* use the per-pgdat data instead for discontigmem - mbligh */
78 unsigned long max_mapnr;
79 struct page *mem_map;
80
81 EXPORT_SYMBOL(max_mapnr);
82 EXPORT_SYMBOL(mem_map);
83 #endif
84
85 /*
86  * A number of key systems in x86 including ioremap() rely on the assumption
87  * that high_memory defines the upper bound on direct map memory, then end
88  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
89  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
90  * and ZONE_HIGHMEM.
91  */
92 void * high_memory;
93
94 EXPORT_SYMBOL(high_memory);
95
96 /*
97  * Randomize the address space (stacks, mmaps, brk, etc.).
98  *
99  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
100  *   as ancient (libc5 based) binaries can segfault. )
101  */
102 int randomize_va_space __read_mostly =
103 #ifdef CONFIG_COMPAT_BRK
104                                         1;
105 #else
106                                         2;
107 #endif
108
109 static int __init disable_randmaps(char *s)
110 {
111         randomize_va_space = 0;
112         return 1;
113 }
114 __setup("norandmaps", disable_randmaps);
115
116 unsigned long zero_pfn __read_mostly;
117 unsigned long highest_memmap_pfn __read_mostly;
118
119 /*
120  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
121  */
122 static int __init init_zero_pfn(void)
123 {
124         zero_pfn = page_to_pfn(ZERO_PAGE(0));
125         return 0;
126 }
127 core_initcall(init_zero_pfn);
128
129
130 #if defined(SPLIT_RSS_COUNTING)
131
132 void sync_mm_rss(struct mm_struct *mm)
133 {
134         int i;
135
136         for (i = 0; i < NR_MM_COUNTERS; i++) {
137                 if (current->rss_stat.count[i]) {
138                         add_mm_counter(mm, i, current->rss_stat.count[i]);
139                         current->rss_stat.count[i] = 0;
140                 }
141         }
142         current->rss_stat.events = 0;
143 }
144
145 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
146 {
147         struct task_struct *task = current;
148
149         if (likely(task->mm == mm))
150                 task->rss_stat.count[member] += val;
151         else
152                 add_mm_counter(mm, member, val);
153 }
154 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
155 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
156
157 /* sync counter once per 64 page faults */
158 #define TASK_RSS_EVENTS_THRESH  (64)
159 static void check_sync_rss_stat(struct task_struct *task)
160 {
161         if (unlikely(task != current))
162                 return;
163         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
164                 sync_mm_rss(task->mm);
165 }
166 #else /* SPLIT_RSS_COUNTING */
167
168 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
169 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
170
171 static void check_sync_rss_stat(struct task_struct *task)
172 {
173 }
174
175 #endif /* SPLIT_RSS_COUNTING */
176
177 #ifdef HAVE_GENERIC_MMU_GATHER
178
179 static int tlb_next_batch(struct mmu_gather *tlb)
180 {
181         struct mmu_gather_batch *batch;
182
183         batch = tlb->active;
184         if (batch->next) {
185                 tlb->active = batch->next;
186                 return 1;
187         }
188
189         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
190                 return 0;
191
192         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
193         if (!batch)
194                 return 0;
195
196         tlb->batch_count++;
197         batch->next = NULL;
198         batch->nr   = 0;
199         batch->max  = MAX_GATHER_BATCH;
200
201         tlb->active->next = batch;
202         tlb->active = batch;
203
204         return 1;
205 }
206
207 /* tlb_gather_mmu
208  *      Called to initialize an (on-stack) mmu_gather structure for page-table
209  *      tear-down from @mm. The @fullmm argument is used when @mm is without
210  *      users and we're going to destroy the full address space (exit/execve).
211  */
212 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
213 {
214         tlb->mm = mm;
215
216         /* Is it from 0 to ~0? */
217         tlb->fullmm     = !(start | (end+1));
218         tlb->need_flush_all = 0;
219         tlb->start      = start;
220         tlb->end        = end;
221         tlb->need_flush = 0;
222         tlb->local.next = NULL;
223         tlb->local.nr   = 0;
224         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
225         tlb->active     = &tlb->local;
226         tlb->batch_count = 0;
227
228 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
229         tlb->batch = NULL;
230 #endif
231 }
232
233 void tlb_flush_mmu(struct mmu_gather *tlb)
234 {
235         struct mmu_gather_batch *batch;
236
237         if (!tlb->need_flush)
238                 return;
239         tlb->need_flush = 0;
240         tlb_flush(tlb);
241 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
242         tlb_table_flush(tlb);
243 #endif
244
245         for (batch = &tlb->local; batch; batch = batch->next) {
246                 free_pages_and_swap_cache(batch->pages, batch->nr);
247                 batch->nr = 0;
248         }
249         tlb->active = &tlb->local;
250 }
251
252 /* tlb_finish_mmu
253  *      Called at the end of the shootdown operation to free up any resources
254  *      that were required.
255  */
256 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
257 {
258         struct mmu_gather_batch *batch, *next;
259
260         tlb_flush_mmu(tlb);
261
262         /* keep the page table cache within bounds */
263         check_pgt_cache();
264
265         for (batch = tlb->local.next; batch; batch = next) {
266                 next = batch->next;
267                 free_pages((unsigned long)batch, 0);
268         }
269         tlb->local.next = NULL;
270 }
271
272 /* __tlb_remove_page
273  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
274  *      handling the additional races in SMP caused by other CPUs caching valid
275  *      mappings in their TLBs. Returns the number of free page slots left.
276  *      When out of page slots we must call tlb_flush_mmu().
277  */
278 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
279 {
280         struct mmu_gather_batch *batch;
281
282         VM_BUG_ON(!tlb->need_flush);
283
284         batch = tlb->active;
285         batch->pages[batch->nr++] = page;
286         if (batch->nr == batch->max) {
287                 if (!tlb_next_batch(tlb))
288                         return 0;
289                 batch = tlb->active;
290         }
291         VM_BUG_ON(batch->nr > batch->max);
292
293         return batch->max - batch->nr;
294 }
295
296 #endif /* HAVE_GENERIC_MMU_GATHER */
297
298 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
299
300 /*
301  * See the comment near struct mmu_table_batch.
302  */
303
304 static void tlb_remove_table_smp_sync(void *arg)
305 {
306         /* Simply deliver the interrupt */
307 }
308
309 static void tlb_remove_table_one(void *table)
310 {
311         /*
312          * This isn't an RCU grace period and hence the page-tables cannot be
313          * assumed to be actually RCU-freed.
314          *
315          * It is however sufficient for software page-table walkers that rely on
316          * IRQ disabling. See the comment near struct mmu_table_batch.
317          */
318         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
319         __tlb_remove_table(table);
320 }
321
322 static void tlb_remove_table_rcu(struct rcu_head *head)
323 {
324         struct mmu_table_batch *batch;
325         int i;
326
327         batch = container_of(head, struct mmu_table_batch, rcu);
328
329         for (i = 0; i < batch->nr; i++)
330                 __tlb_remove_table(batch->tables[i]);
331
332         free_page((unsigned long)batch);
333 }
334
335 void tlb_table_flush(struct mmu_gather *tlb)
336 {
337         struct mmu_table_batch **batch = &tlb->batch;
338
339         if (*batch) {
340                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
341                 *batch = NULL;
342         }
343 }
344
345 void tlb_remove_table(struct mmu_gather *tlb, void *table)
346 {
347         struct mmu_table_batch **batch = &tlb->batch;
348
349         tlb->need_flush = 1;
350
351         /*
352          * When there's less then two users of this mm there cannot be a
353          * concurrent page-table walk.
354          */
355         if (atomic_read(&tlb->mm->mm_users) < 2) {
356                 __tlb_remove_table(table);
357                 return;
358         }
359
360         if (*batch == NULL) {
361                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
362                 if (*batch == NULL) {
363                         tlb_remove_table_one(table);
364                         return;
365                 }
366                 (*batch)->nr = 0;
367         }
368         (*batch)->tables[(*batch)->nr++] = table;
369         if ((*batch)->nr == MAX_TABLE_BATCH)
370                 tlb_table_flush(tlb);
371 }
372
373 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
374
375 /*
376  * Note: this doesn't free the actual pages themselves. That
377  * has been handled earlier when unmapping all the memory regions.
378  */
379 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
380                            unsigned long addr)
381 {
382         pgtable_t token = pmd_pgtable(*pmd);
383         pmd_clear(pmd);
384         pte_free_tlb(tlb, token, addr);
385         tlb->mm->nr_ptes--;
386 }
387
388 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
389                                 unsigned long addr, unsigned long end,
390                                 unsigned long floor, unsigned long ceiling)
391 {
392         pmd_t *pmd;
393         unsigned long next;
394         unsigned long start;
395
396         start = addr;
397         pmd = pmd_offset(pud, addr);
398         do {
399                 next = pmd_addr_end(addr, end);
400                 if (pmd_none_or_clear_bad(pmd))
401                         continue;
402                 free_pte_range(tlb, pmd, addr);
403         } while (pmd++, addr = next, addr != end);
404
405         start &= PUD_MASK;
406         if (start < floor)
407                 return;
408         if (ceiling) {
409                 ceiling &= PUD_MASK;
410                 if (!ceiling)
411                         return;
412         }
413         if (end - 1 > ceiling - 1)
414                 return;
415
416         pmd = pmd_offset(pud, start);
417         pud_clear(pud);
418         pmd_free_tlb(tlb, pmd, start);
419 }
420
421 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
422                                 unsigned long addr, unsigned long end,
423                                 unsigned long floor, unsigned long ceiling)
424 {
425         pud_t *pud;
426         unsigned long next;
427         unsigned long start;
428
429         start = addr;
430         pud = pud_offset(pgd, addr);
431         do {
432                 next = pud_addr_end(addr, end);
433                 if (pud_none_or_clear_bad(pud))
434                         continue;
435                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
436         } while (pud++, addr = next, addr != end);
437
438         start &= PGDIR_MASK;
439         if (start < floor)
440                 return;
441         if (ceiling) {
442                 ceiling &= PGDIR_MASK;
443                 if (!ceiling)
444                         return;
445         }
446         if (end - 1 > ceiling - 1)
447                 return;
448
449         pud = pud_offset(pgd, start);
450         pgd_clear(pgd);
451         pud_free_tlb(tlb, pud, start);
452 }
453
454 /*
455  * This function frees user-level page tables of a process.
456  *
457  * Must be called with pagetable lock held.
458  */
459 void free_pgd_range(struct mmu_gather *tlb,
460                         unsigned long addr, unsigned long end,
461                         unsigned long floor, unsigned long ceiling)
462 {
463         pgd_t *pgd;
464         unsigned long next;
465
466         /*
467          * The next few lines have given us lots of grief...
468          *
469          * Why are we testing PMD* at this top level?  Because often
470          * there will be no work to do at all, and we'd prefer not to
471          * go all the way down to the bottom just to discover that.
472          *
473          * Why all these "- 1"s?  Because 0 represents both the bottom
474          * of the address space and the top of it (using -1 for the
475          * top wouldn't help much: the masks would do the wrong thing).
476          * The rule is that addr 0 and floor 0 refer to the bottom of
477          * the address space, but end 0 and ceiling 0 refer to the top
478          * Comparisons need to use "end - 1" and "ceiling - 1" (though
479          * that end 0 case should be mythical).
480          *
481          * Wherever addr is brought up or ceiling brought down, we must
482          * be careful to reject "the opposite 0" before it confuses the
483          * subsequent tests.  But what about where end is brought down
484          * by PMD_SIZE below? no, end can't go down to 0 there.
485          *
486          * Whereas we round start (addr) and ceiling down, by different
487          * masks at different levels, in order to test whether a table
488          * now has no other vmas using it, so can be freed, we don't
489          * bother to round floor or end up - the tests don't need that.
490          */
491
492         addr &= PMD_MASK;
493         if (addr < floor) {
494                 addr += PMD_SIZE;
495                 if (!addr)
496                         return;
497         }
498         if (ceiling) {
499                 ceiling &= PMD_MASK;
500                 if (!ceiling)
501                         return;
502         }
503         if (end - 1 > ceiling - 1)
504                 end -= PMD_SIZE;
505         if (addr > end - 1)
506                 return;
507
508         pgd = pgd_offset(tlb->mm, addr);
509         do {
510                 next = pgd_addr_end(addr, end);
511                 if (pgd_none_or_clear_bad(pgd))
512                         continue;
513                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
514         } while (pgd++, addr = next, addr != end);
515 }
516
517 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
518                 unsigned long floor, unsigned long ceiling)
519 {
520         while (vma) {
521                 struct vm_area_struct *next = vma->vm_next;
522                 unsigned long addr = vma->vm_start;
523
524                 /*
525                  * Hide vma from rmap and truncate_pagecache before freeing
526                  * pgtables
527                  */
528                 unlink_anon_vmas(vma);
529                 unlink_file_vma(vma);
530
531                 if (is_vm_hugetlb_page(vma)) {
532                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
533                                 floor, next? next->vm_start: ceiling);
534                 } else {
535                         /*
536                          * Optimization: gather nearby vmas into one call down
537                          */
538                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
539                                && !is_vm_hugetlb_page(next)) {
540                                 vma = next;
541                                 next = vma->vm_next;
542                                 unlink_anon_vmas(vma);
543                                 unlink_file_vma(vma);
544                         }
545                         free_pgd_range(tlb, addr, vma->vm_end,
546                                 floor, next? next->vm_start: ceiling);
547                 }
548                 vma = next;
549         }
550 }
551
552 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
553                 pmd_t *pmd, unsigned long address)
554 {
555         pgtable_t new = pte_alloc_one(mm, address);
556         int wait_split_huge_page;
557         if (!new)
558                 return -ENOMEM;
559
560         /*
561          * Ensure all pte setup (eg. pte page lock and page clearing) are
562          * visible before the pte is made visible to other CPUs by being
563          * put into page tables.
564          *
565          * The other side of the story is the pointer chasing in the page
566          * table walking code (when walking the page table without locking;
567          * ie. most of the time). Fortunately, these data accesses consist
568          * of a chain of data-dependent loads, meaning most CPUs (alpha
569          * being the notable exception) will already guarantee loads are
570          * seen in-order. See the alpha page table accessors for the
571          * smp_read_barrier_depends() barriers in page table walking code.
572          */
573         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
574
575         spin_lock(&mm->page_table_lock);
576         wait_split_huge_page = 0;
577         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
578                 mm->nr_ptes++;
579                 pmd_populate(mm, pmd, new);
580                 new = NULL;
581         } else if (unlikely(pmd_trans_splitting(*pmd)))
582                 wait_split_huge_page = 1;
583         spin_unlock(&mm->page_table_lock);
584         if (new)
585                 pte_free(mm, new);
586         if (wait_split_huge_page)
587                 wait_split_huge_page(vma->anon_vma, pmd);
588         return 0;
589 }
590
591 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
592 {
593         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
594         if (!new)
595                 return -ENOMEM;
596
597         smp_wmb(); /* See comment in __pte_alloc */
598
599         spin_lock(&init_mm.page_table_lock);
600         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
601                 pmd_populate_kernel(&init_mm, pmd, new);
602                 new = NULL;
603         } else
604                 VM_BUG_ON(pmd_trans_splitting(*pmd));
605         spin_unlock(&init_mm.page_table_lock);
606         if (new)
607                 pte_free_kernel(&init_mm, new);
608         return 0;
609 }
610
611 static inline void init_rss_vec(int *rss)
612 {
613         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
614 }
615
616 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
617 {
618         int i;
619
620         if (current->mm == mm)
621                 sync_mm_rss(mm);
622         for (i = 0; i < NR_MM_COUNTERS; i++)
623                 if (rss[i])
624                         add_mm_counter(mm, i, rss[i]);
625 }
626
627 /*
628  * This function is called to print an error when a bad pte
629  * is found. For example, we might have a PFN-mapped pte in
630  * a region that doesn't allow it.
631  *
632  * The calling function must still handle the error.
633  */
634 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
635                           pte_t pte, struct page *page)
636 {
637         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
638         pud_t *pud = pud_offset(pgd, addr);
639         pmd_t *pmd = pmd_offset(pud, addr);
640         struct address_space *mapping;
641         pgoff_t index;
642         static unsigned long resume;
643         static unsigned long nr_shown;
644         static unsigned long nr_unshown;
645
646         /*
647          * Allow a burst of 60 reports, then keep quiet for that minute;
648          * or allow a steady drip of one report per second.
649          */
650         if (nr_shown == 60) {
651                 if (time_before(jiffies, resume)) {
652                         nr_unshown++;
653                         return;
654                 }
655                 if (nr_unshown) {
656                         printk(KERN_ALERT
657                                 "BUG: Bad page map: %lu messages suppressed\n",
658                                 nr_unshown);
659                         nr_unshown = 0;
660                 }
661                 nr_shown = 0;
662         }
663         if (nr_shown++ == 0)
664                 resume = jiffies + 60 * HZ;
665
666         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
667         index = linear_page_index(vma, addr);
668
669         printk(KERN_ALERT
670                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
671                 current->comm,
672                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
673         if (page)
674                 dump_page(page);
675         printk(KERN_ALERT
676                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
677                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
678         /*
679          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
680          */
681         if (vma->vm_ops)
682                 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
683                        vma->vm_ops->fault);
684         if (vma->vm_file && vma->vm_file->f_op)
685                 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
686                        vma->vm_file->f_op->mmap);
687         dump_stack();
688         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
689 }
690
691 static inline bool is_cow_mapping(vm_flags_t flags)
692 {
693         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
694 }
695
696 /*
697  * vm_normal_page -- This function gets the "struct page" associated with a pte.
698  *
699  * "Special" mappings do not wish to be associated with a "struct page" (either
700  * it doesn't exist, or it exists but they don't want to touch it). In this
701  * case, NULL is returned here. "Normal" mappings do have a struct page.
702  *
703  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
704  * pte bit, in which case this function is trivial. Secondly, an architecture
705  * may not have a spare pte bit, which requires a more complicated scheme,
706  * described below.
707  *
708  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
709  * special mapping (even if there are underlying and valid "struct pages").
710  * COWed pages of a VM_PFNMAP are always normal.
711  *
712  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
713  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
714  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
715  * mapping will always honor the rule
716  *
717  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
718  *
719  * And for normal mappings this is false.
720  *
721  * This restricts such mappings to be a linear translation from virtual address
722  * to pfn. To get around this restriction, we allow arbitrary mappings so long
723  * as the vma is not a COW mapping; in that case, we know that all ptes are
724  * special (because none can have been COWed).
725  *
726  *
727  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
728  *
729  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
730  * page" backing, however the difference is that _all_ pages with a struct
731  * page (that is, those where pfn_valid is true) are refcounted and considered
732  * normal pages by the VM. The disadvantage is that pages are refcounted
733  * (which can be slower and simply not an option for some PFNMAP users). The
734  * advantage is that we don't have to follow the strict linearity rule of
735  * PFNMAP mappings in order to support COWable mappings.
736  *
737  */
738 #ifdef __HAVE_ARCH_PTE_SPECIAL
739 # define HAVE_PTE_SPECIAL 1
740 #else
741 # define HAVE_PTE_SPECIAL 0
742 #endif
743 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
744                                 pte_t pte)
745 {
746         unsigned long pfn = pte_pfn(pte);
747
748         if (HAVE_PTE_SPECIAL) {
749                 if (likely(!pte_special(pte)))
750                         goto check_pfn;
751                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
752                         return NULL;
753                 if (!is_zero_pfn(pfn))
754                         print_bad_pte(vma, addr, pte, NULL);
755                 return NULL;
756         }
757
758         /* !HAVE_PTE_SPECIAL case follows: */
759
760         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
761                 if (vma->vm_flags & VM_MIXEDMAP) {
762                         if (!pfn_valid(pfn))
763                                 return NULL;
764                         goto out;
765                 } else {
766                         unsigned long off;
767                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
768                         if (pfn == vma->vm_pgoff + off)
769                                 return NULL;
770                         if (!is_cow_mapping(vma->vm_flags))
771                                 return NULL;
772                 }
773         }
774
775         if (is_zero_pfn(pfn))
776                 return NULL;
777 check_pfn:
778         if (unlikely(pfn > highest_memmap_pfn)) {
779                 print_bad_pte(vma, addr, pte, NULL);
780                 return NULL;
781         }
782
783         /*
784          * NOTE! We still have PageReserved() pages in the page tables.
785          * eg. VDSO mappings can cause them to exist.
786          */
787 out:
788         return pfn_to_page(pfn);
789 }
790
791 /*
792  * copy one vm_area from one task to the other. Assumes the page tables
793  * already present in the new task to be cleared in the whole range
794  * covered by this vma.
795  */
796
797 static inline unsigned long
798 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
799                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
800                 unsigned long addr, int *rss)
801 {
802         unsigned long vm_flags = vma->vm_flags;
803         pte_t pte = *src_pte;
804         struct page *page;
805
806         /* pte contains position in swap or file, so copy. */
807         if (unlikely(!pte_present(pte))) {
808                 if (!pte_file(pte)) {
809                         swp_entry_t entry = pte_to_swp_entry(pte);
810
811                         if (swap_duplicate(entry) < 0)
812                                 return entry.val;
813
814                         /* make sure dst_mm is on swapoff's mmlist. */
815                         if (unlikely(list_empty(&dst_mm->mmlist))) {
816                                 spin_lock(&mmlist_lock);
817                                 if (list_empty(&dst_mm->mmlist))
818                                         list_add(&dst_mm->mmlist,
819                                                  &src_mm->mmlist);
820                                 spin_unlock(&mmlist_lock);
821                         }
822                         if (likely(!non_swap_entry(entry)))
823                                 rss[MM_SWAPENTS]++;
824                         else if (is_migration_entry(entry)) {
825                                 page = migration_entry_to_page(entry);
826
827                                 if (PageAnon(page))
828                                         rss[MM_ANONPAGES]++;
829                                 else
830                                         rss[MM_FILEPAGES]++;
831
832                                 if (is_write_migration_entry(entry) &&
833                                     is_cow_mapping(vm_flags)) {
834                                         /*
835                                          * COW mappings require pages in both
836                                          * parent and child to be set to read.
837                                          */
838                                         make_migration_entry_read(&entry);
839                                         pte = swp_entry_to_pte(entry);
840                                         if (pte_swp_soft_dirty(*src_pte))
841                                                 pte = pte_swp_mksoft_dirty(pte);
842                                         set_pte_at(src_mm, addr, src_pte, pte);
843                                 }
844                         }
845                 }
846                 goto out_set_pte;
847         }
848
849         /*
850          * If it's a COW mapping, write protect it both
851          * in the parent and the child
852          */
853         if (is_cow_mapping(vm_flags)) {
854                 ptep_set_wrprotect(src_mm, addr, src_pte);
855                 pte = pte_wrprotect(pte);
856         }
857
858         /*
859          * If it's a shared mapping, mark it clean in
860          * the child
861          */
862         if (vm_flags & VM_SHARED)
863                 pte = pte_mkclean(pte);
864         pte = pte_mkold(pte);
865
866         page = vm_normal_page(vma, addr, pte);
867         if (page) {
868                 get_page(page);
869                 page_dup_rmap(page);
870                 if (PageAnon(page))
871                         rss[MM_ANONPAGES]++;
872                 else
873                         rss[MM_FILEPAGES]++;
874         }
875
876 out_set_pte:
877         set_pte_at(dst_mm, addr, dst_pte, pte);
878         return 0;
879 }
880
881 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
882                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
883                    unsigned long addr, unsigned long end)
884 {
885         pte_t *orig_src_pte, *orig_dst_pte;
886         pte_t *src_pte, *dst_pte;
887         spinlock_t *src_ptl, *dst_ptl;
888         int progress = 0;
889         int rss[NR_MM_COUNTERS];
890         swp_entry_t entry = (swp_entry_t){0};
891
892 again:
893         init_rss_vec(rss);
894
895         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
896         if (!dst_pte)
897                 return -ENOMEM;
898         src_pte = pte_offset_map(src_pmd, addr);
899         src_ptl = pte_lockptr(src_mm, src_pmd);
900         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
901         orig_src_pte = src_pte;
902         orig_dst_pte = dst_pte;
903         arch_enter_lazy_mmu_mode();
904
905         do {
906                 /*
907                  * We are holding two locks at this point - either of them
908                  * could generate latencies in another task on another CPU.
909                  */
910                 if (progress >= 32) {
911                         progress = 0;
912                         if (need_resched() ||
913                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
914                                 break;
915                 }
916                 if (pte_none(*src_pte)) {
917                         progress++;
918                         continue;
919                 }
920                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
921                                                         vma, addr, rss);
922                 if (entry.val)
923                         break;
924                 progress += 8;
925         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
926
927         arch_leave_lazy_mmu_mode();
928         spin_unlock(src_ptl);
929         pte_unmap(orig_src_pte);
930         add_mm_rss_vec(dst_mm, rss);
931         pte_unmap_unlock(orig_dst_pte, dst_ptl);
932         cond_resched();
933
934         if (entry.val) {
935                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
936                         return -ENOMEM;
937                 progress = 0;
938         }
939         if (addr != end)
940                 goto again;
941         return 0;
942 }
943
944 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
945                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
946                 unsigned long addr, unsigned long end)
947 {
948         pmd_t *src_pmd, *dst_pmd;
949         unsigned long next;
950
951         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
952         if (!dst_pmd)
953                 return -ENOMEM;
954         src_pmd = pmd_offset(src_pud, addr);
955         do {
956                 next = pmd_addr_end(addr, end);
957                 if (pmd_trans_huge(*src_pmd)) {
958                         int err;
959                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
960                         err = copy_huge_pmd(dst_mm, src_mm,
961                                             dst_pmd, src_pmd, addr, vma);
962                         if (err == -ENOMEM)
963                                 return -ENOMEM;
964                         if (!err)
965                                 continue;
966                         /* fall through */
967                 }
968                 if (pmd_none_or_clear_bad(src_pmd))
969                         continue;
970                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
971                                                 vma, addr, next))
972                         return -ENOMEM;
973         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
974         return 0;
975 }
976
977 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
978                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
979                 unsigned long addr, unsigned long end)
980 {
981         pud_t *src_pud, *dst_pud;
982         unsigned long next;
983
984         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
985         if (!dst_pud)
986                 return -ENOMEM;
987         src_pud = pud_offset(src_pgd, addr);
988         do {
989                 next = pud_addr_end(addr, end);
990                 if (pud_none_or_clear_bad(src_pud))
991                         continue;
992                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
993                                                 vma, addr, next))
994                         return -ENOMEM;
995         } while (dst_pud++, src_pud++, addr = next, addr != end);
996         return 0;
997 }
998
999 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1000                 struct vm_area_struct *vma)
1001 {
1002         pgd_t *src_pgd, *dst_pgd;
1003         unsigned long next;
1004         unsigned long addr = vma->vm_start;
1005         unsigned long end = vma->vm_end;
1006         unsigned long mmun_start;       /* For mmu_notifiers */
1007         unsigned long mmun_end;         /* For mmu_notifiers */
1008         bool is_cow;
1009         int ret;
1010
1011         /*
1012          * Don't copy ptes where a page fault will fill them correctly.
1013          * Fork becomes much lighter when there are big shared or private
1014          * readonly mappings. The tradeoff is that copy_page_range is more
1015          * efficient than faulting.
1016          */
1017         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1018                                VM_PFNMAP | VM_MIXEDMAP))) {
1019                 if (!vma->anon_vma)
1020                         return 0;
1021         }
1022
1023         if (is_vm_hugetlb_page(vma))
1024                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1025
1026         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1027                 /*
1028                  * We do not free on error cases below as remove_vma
1029                  * gets called on error from higher level routine
1030                  */
1031                 ret = track_pfn_copy(vma);
1032                 if (ret)
1033                         return ret;
1034         }
1035
1036         /*
1037          * We need to invalidate the secondary MMU mappings only when
1038          * there could be a permission downgrade on the ptes of the
1039          * parent mm. And a permission downgrade will only happen if
1040          * is_cow_mapping() returns true.
1041          */
1042         is_cow = is_cow_mapping(vma->vm_flags);
1043         mmun_start = addr;
1044         mmun_end   = end;
1045         if (is_cow)
1046                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1047                                                     mmun_end);
1048
1049         ret = 0;
1050         dst_pgd = pgd_offset(dst_mm, addr);
1051         src_pgd = pgd_offset(src_mm, addr);
1052         do {
1053                 next = pgd_addr_end(addr, end);
1054                 if (pgd_none_or_clear_bad(src_pgd))
1055                         continue;
1056                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1057                                             vma, addr, next))) {
1058                         ret = -ENOMEM;
1059                         break;
1060                 }
1061         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1062
1063         if (is_cow)
1064                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1065         return ret;
1066 }
1067
1068 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1069                                 struct vm_area_struct *vma, pmd_t *pmd,
1070                                 unsigned long addr, unsigned long end,
1071                                 struct zap_details *details)
1072 {
1073         struct mm_struct *mm = tlb->mm;
1074         int force_flush = 0;
1075         int rss[NR_MM_COUNTERS];
1076         spinlock_t *ptl;
1077         pte_t *start_pte;
1078         pte_t *pte;
1079
1080 again:
1081         init_rss_vec(rss);
1082         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1083         pte = start_pte;
1084         arch_enter_lazy_mmu_mode();
1085         do {
1086                 pte_t ptent = *pte;
1087                 if (pte_none(ptent)) {
1088                         continue;
1089                 }
1090
1091                 if (pte_present(ptent)) {
1092                         struct page *page;
1093
1094                         page = vm_normal_page(vma, addr, ptent);
1095                         if (unlikely(details) && page) {
1096                                 /*
1097                                  * unmap_shared_mapping_pages() wants to
1098                                  * invalidate cache without truncating:
1099                                  * unmap shared but keep private pages.
1100                                  */
1101                                 if (details->check_mapping &&
1102                                     details->check_mapping != page->mapping)
1103                                         continue;
1104                                 /*
1105                                  * Each page->index must be checked when
1106                                  * invalidating or truncating nonlinear.
1107                                  */
1108                                 if (details->nonlinear_vma &&
1109                                     (page->index < details->first_index ||
1110                                      page->index > details->last_index))
1111                                         continue;
1112                         }
1113                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1114                                                         tlb->fullmm);
1115                         tlb_remove_tlb_entry(tlb, pte, addr);
1116                         if (unlikely(!page))
1117                                 continue;
1118                         if (unlikely(details) && details->nonlinear_vma
1119                             && linear_page_index(details->nonlinear_vma,
1120                                                 addr) != page->index) {
1121                                 pte_t ptfile = pgoff_to_pte(page->index);
1122                                 if (pte_soft_dirty(ptent))
1123                                         pte_file_mksoft_dirty(ptfile);
1124                                 set_pte_at(mm, addr, pte, ptfile);
1125                         }
1126                         if (PageAnon(page))
1127                                 rss[MM_ANONPAGES]--;
1128                         else {
1129                                 if (pte_dirty(ptent))
1130                                         set_page_dirty(page);
1131                                 if (pte_young(ptent) &&
1132                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1133                                         mark_page_accessed(page);
1134                                 rss[MM_FILEPAGES]--;
1135                         }
1136                         page_remove_rmap(page);
1137                         if (unlikely(page_mapcount(page) < 0))
1138                                 print_bad_pte(vma, addr, ptent, page);
1139                         force_flush = !__tlb_remove_page(tlb, page);
1140                         if (force_flush)
1141                                 break;
1142                         continue;
1143                 }
1144                 /*
1145                  * If details->check_mapping, we leave swap entries;
1146                  * if details->nonlinear_vma, we leave file entries.
1147                  */
1148                 if (unlikely(details))
1149                         continue;
1150                 if (pte_file(ptent)) {
1151                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1152                                 print_bad_pte(vma, addr, ptent, NULL);
1153                 } else {
1154                         swp_entry_t entry = pte_to_swp_entry(ptent);
1155
1156                         if (!non_swap_entry(entry))
1157                                 rss[MM_SWAPENTS]--;
1158                         else if (is_migration_entry(entry)) {
1159                                 struct page *page;
1160
1161                                 page = migration_entry_to_page(entry);
1162
1163                                 if (PageAnon(page))
1164                                         rss[MM_ANONPAGES]--;
1165                                 else
1166                                         rss[MM_FILEPAGES]--;
1167                         }
1168                         if (unlikely(!free_swap_and_cache(entry)))
1169                                 print_bad_pte(vma, addr, ptent, NULL);
1170                 }
1171                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1172         } while (pte++, addr += PAGE_SIZE, addr != end);
1173
1174         add_mm_rss_vec(mm, rss);
1175         arch_leave_lazy_mmu_mode();
1176         pte_unmap_unlock(start_pte, ptl);
1177
1178         /*
1179          * mmu_gather ran out of room to batch pages, we break out of
1180          * the PTE lock to avoid doing the potential expensive TLB invalidate
1181          * and page-free while holding it.
1182          */
1183         if (force_flush) {
1184                 unsigned long old_end;
1185
1186                 force_flush = 0;
1187
1188                 /*
1189                  * Flush the TLB just for the previous segment,
1190                  * then update the range to be the remaining
1191                  * TLB range.
1192                  */
1193                 old_end = tlb->end;
1194                 tlb->end = addr;
1195
1196                 tlb_flush_mmu(tlb);
1197
1198                 tlb->start = addr;
1199                 tlb->end = old_end;
1200
1201                 if (addr != end)
1202                         goto again;
1203         }
1204
1205         return addr;
1206 }
1207
1208 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1209                                 struct vm_area_struct *vma, pud_t *pud,
1210                                 unsigned long addr, unsigned long end,
1211                                 struct zap_details *details)
1212 {
1213         pmd_t *pmd;
1214         unsigned long next;
1215
1216         pmd = pmd_offset(pud, addr);
1217         do {
1218                 next = pmd_addr_end(addr, end);
1219                 if (pmd_trans_huge(*pmd)) {
1220                         if (next - addr != HPAGE_PMD_SIZE) {
1221 #ifdef CONFIG_DEBUG_VM
1222                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1223                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1224                                                 __func__, addr, end,
1225                                                 vma->vm_start,
1226                                                 vma->vm_end);
1227                                         BUG();
1228                                 }
1229 #endif
1230                                 split_huge_page_pmd(vma, addr, pmd);
1231                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1232                                 goto next;
1233                         /* fall through */
1234                 }
1235                 /*
1236                  * Here there can be other concurrent MADV_DONTNEED or
1237                  * trans huge page faults running, and if the pmd is
1238                  * none or trans huge it can change under us. This is
1239                  * because MADV_DONTNEED holds the mmap_sem in read
1240                  * mode.
1241                  */
1242                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1243                         goto next;
1244                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1245 next:
1246                 cond_resched();
1247         } while (pmd++, addr = next, addr != end);
1248
1249         return addr;
1250 }
1251
1252 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1253                                 struct vm_area_struct *vma, pgd_t *pgd,
1254                                 unsigned long addr, unsigned long end,
1255                                 struct zap_details *details)
1256 {
1257         pud_t *pud;
1258         unsigned long next;
1259
1260         pud = pud_offset(pgd, addr);
1261         do {
1262                 next = pud_addr_end(addr, end);
1263                 if (pud_none_or_clear_bad(pud))
1264                         continue;
1265                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1266         } while (pud++, addr = next, addr != end);
1267
1268         return addr;
1269 }
1270
1271 static void unmap_page_range(struct mmu_gather *tlb,
1272                              struct vm_area_struct *vma,
1273                              unsigned long addr, unsigned long end,
1274                              struct zap_details *details)
1275 {
1276         pgd_t *pgd;
1277         unsigned long next;
1278
1279         if (details && !details->check_mapping && !details->nonlinear_vma)
1280                 details = NULL;
1281
1282         BUG_ON(addr >= end);
1283         mem_cgroup_uncharge_start();
1284         tlb_start_vma(tlb, vma);
1285         pgd = pgd_offset(vma->vm_mm, addr);
1286         do {
1287                 next = pgd_addr_end(addr, end);
1288                 if (pgd_none_or_clear_bad(pgd))
1289                         continue;
1290                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1291         } while (pgd++, addr = next, addr != end);
1292         tlb_end_vma(tlb, vma);
1293         mem_cgroup_uncharge_end();
1294 }
1295
1296
1297 static void unmap_single_vma(struct mmu_gather *tlb,
1298                 struct vm_area_struct *vma, unsigned long start_addr,
1299                 unsigned long end_addr,
1300                 struct zap_details *details)
1301 {
1302         unsigned long start = max(vma->vm_start, start_addr);
1303         unsigned long end;
1304
1305         if (start >= vma->vm_end)
1306                 return;
1307         end = min(vma->vm_end, end_addr);
1308         if (end <= vma->vm_start)
1309                 return;
1310
1311         if (vma->vm_file)
1312                 uprobe_munmap(vma, start, end);
1313
1314         if (unlikely(vma->vm_flags & VM_PFNMAP))
1315                 untrack_pfn(vma, 0, 0);
1316
1317         if (start != end) {
1318                 if (unlikely(is_vm_hugetlb_page(vma))) {
1319                         /*
1320                          * It is undesirable to test vma->vm_file as it
1321                          * should be non-null for valid hugetlb area.
1322                          * However, vm_file will be NULL in the error
1323                          * cleanup path of do_mmap_pgoff. When
1324                          * hugetlbfs ->mmap method fails,
1325                          * do_mmap_pgoff() nullifies vma->vm_file
1326                          * before calling this function to clean up.
1327                          * Since no pte has actually been setup, it is
1328                          * safe to do nothing in this case.
1329                          */
1330                         if (vma->vm_file) {
1331                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1332                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1333                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1334                         }
1335                 } else
1336                         unmap_page_range(tlb, vma, start, end, details);
1337         }
1338 }
1339
1340 /**
1341  * unmap_vmas - unmap a range of memory covered by a list of vma's
1342  * @tlb: address of the caller's struct mmu_gather
1343  * @vma: the starting vma
1344  * @start_addr: virtual address at which to start unmapping
1345  * @end_addr: virtual address at which to end unmapping
1346  *
1347  * Unmap all pages in the vma list.
1348  *
1349  * Only addresses between `start' and `end' will be unmapped.
1350  *
1351  * The VMA list must be sorted in ascending virtual address order.
1352  *
1353  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1354  * range after unmap_vmas() returns.  So the only responsibility here is to
1355  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1356  * drops the lock and schedules.
1357  */
1358 void unmap_vmas(struct mmu_gather *tlb,
1359                 struct vm_area_struct *vma, unsigned long start_addr,
1360                 unsigned long end_addr)
1361 {
1362         struct mm_struct *mm = vma->vm_mm;
1363
1364         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1365         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1366                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1367         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1368 }
1369
1370 /**
1371  * zap_page_range - remove user pages in a given range
1372  * @vma: vm_area_struct holding the applicable pages
1373  * @start: starting address of pages to zap
1374  * @size: number of bytes to zap
1375  * @details: details of nonlinear truncation or shared cache invalidation
1376  *
1377  * Caller must protect the VMA list
1378  */
1379 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1380                 unsigned long size, struct zap_details *details)
1381 {
1382         struct mm_struct *mm = vma->vm_mm;
1383         struct mmu_gather tlb;
1384         unsigned long end = start + size;
1385
1386         lru_add_drain();
1387         tlb_gather_mmu(&tlb, mm, start, end);
1388         update_hiwater_rss(mm);
1389         mmu_notifier_invalidate_range_start(mm, start, end);
1390         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1391                 unmap_single_vma(&tlb, vma, start, end, details);
1392         mmu_notifier_invalidate_range_end(mm, start, end);
1393         tlb_finish_mmu(&tlb, start, end);
1394 }
1395
1396 /**
1397  * zap_page_range_single - remove user pages in a given range
1398  * @vma: vm_area_struct holding the applicable pages
1399  * @address: starting address of pages to zap
1400  * @size: number of bytes to zap
1401  * @details: details of nonlinear truncation or shared cache invalidation
1402  *
1403  * The range must fit into one VMA.
1404  */
1405 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1406                 unsigned long size, struct zap_details *details)
1407 {
1408         struct mm_struct *mm = vma->vm_mm;
1409         struct mmu_gather tlb;
1410         unsigned long end = address + size;
1411
1412         lru_add_drain();
1413         tlb_gather_mmu(&tlb, mm, address, end);
1414         update_hiwater_rss(mm);
1415         mmu_notifier_invalidate_range_start(mm, address, end);
1416         unmap_single_vma(&tlb, vma, address, end, details);
1417         mmu_notifier_invalidate_range_end(mm, address, end);
1418         tlb_finish_mmu(&tlb, address, end);
1419 }
1420
1421 /**
1422  * zap_vma_ptes - remove ptes mapping the vma
1423  * @vma: vm_area_struct holding ptes to be zapped
1424  * @address: starting address of pages to zap
1425  * @size: number of bytes to zap
1426  *
1427  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1428  *
1429  * The entire address range must be fully contained within the vma.
1430  *
1431  * Returns 0 if successful.
1432  */
1433 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1434                 unsigned long size)
1435 {
1436         if (address < vma->vm_start || address + size > vma->vm_end ||
1437                         !(vma->vm_flags & VM_PFNMAP))
1438                 return -1;
1439         zap_page_range_single(vma, address, size, NULL);
1440         return 0;
1441 }
1442 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1443
1444 /**
1445  * follow_page_mask - look up a page descriptor from a user-virtual address
1446  * @vma: vm_area_struct mapping @address
1447  * @address: virtual address to look up
1448  * @flags: flags modifying lookup behaviour
1449  * @page_mask: on output, *page_mask is set according to the size of the page
1450  *
1451  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1452  *
1453  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1454  * an error pointer if there is a mapping to something not represented
1455  * by a page descriptor (see also vm_normal_page()).
1456  */
1457 struct page *follow_page_mask(struct vm_area_struct *vma,
1458                               unsigned long address, unsigned int flags,
1459                               unsigned int *page_mask)
1460 {
1461         pgd_t *pgd;
1462         pud_t *pud;
1463         pmd_t *pmd;
1464         pte_t *ptep, pte;
1465         spinlock_t *ptl;
1466         struct page *page;
1467         struct mm_struct *mm = vma->vm_mm;
1468
1469         *page_mask = 0;
1470
1471         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1472         if (!IS_ERR(page)) {
1473                 BUG_ON(flags & FOLL_GET);
1474                 goto out;
1475         }
1476
1477         page = NULL;
1478         pgd = pgd_offset(mm, address);
1479         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1480                 goto no_page_table;
1481
1482         pud = pud_offset(pgd, address);
1483         if (pud_none(*pud))
1484                 goto no_page_table;
1485         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1486                 if (flags & FOLL_GET)
1487                         goto out;
1488                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1489                 goto out;
1490         }
1491         if (unlikely(pud_bad(*pud)))
1492                 goto no_page_table;
1493
1494         pmd = pmd_offset(pud, address);
1495         if (pmd_none(*pmd))
1496                 goto no_page_table;
1497         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1498                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1499                 if (flags & FOLL_GET) {
1500                         /*
1501                          * Refcount on tail pages are not well-defined and
1502                          * shouldn't be taken. The caller should handle a NULL
1503                          * return when trying to follow tail pages.
1504                          */
1505                         if (PageHead(page))
1506                                 get_page(page);
1507                         else {
1508                                 page = NULL;
1509                                 goto out;
1510                         }
1511                 }
1512                 goto out;
1513         }
1514         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1515                 goto no_page_table;
1516         if (pmd_trans_huge(*pmd)) {
1517                 if (flags & FOLL_SPLIT) {
1518                         split_huge_page_pmd(vma, address, pmd);
1519                         goto split_fallthrough;
1520                 }
1521                 spin_lock(&mm->page_table_lock);
1522                 if (likely(pmd_trans_huge(*pmd))) {
1523                         if (unlikely(pmd_trans_splitting(*pmd))) {
1524                                 spin_unlock(&mm->page_table_lock);
1525                                 wait_split_huge_page(vma->anon_vma, pmd);
1526                         } else {
1527                                 page = follow_trans_huge_pmd(vma, address,
1528                                                              pmd, flags);
1529                                 spin_unlock(&mm->page_table_lock);
1530                                 *page_mask = HPAGE_PMD_NR - 1;
1531                                 goto out;
1532                         }
1533                 } else
1534                         spin_unlock(&mm->page_table_lock);
1535                 /* fall through */
1536         }
1537 split_fallthrough:
1538         if (unlikely(pmd_bad(*pmd)))
1539                 goto no_page_table;
1540
1541         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1542
1543         pte = *ptep;
1544         if (!pte_present(pte)) {
1545                 swp_entry_t entry;
1546                 /*
1547                  * KSM's break_ksm() relies upon recognizing a ksm page
1548                  * even while it is being migrated, so for that case we
1549                  * need migration_entry_wait().
1550                  */
1551                 if (likely(!(flags & FOLL_MIGRATION)))
1552                         goto no_page;
1553                 if (pte_none(pte) || pte_file(pte))
1554                         goto no_page;
1555                 entry = pte_to_swp_entry(pte);
1556                 if (!is_migration_entry(entry))
1557                         goto no_page;
1558                 pte_unmap_unlock(ptep, ptl);
1559                 migration_entry_wait(mm, pmd, address);
1560                 goto split_fallthrough;
1561         }
1562         if ((flags & FOLL_NUMA) && pte_numa(pte))
1563                 goto no_page;
1564         if ((flags & FOLL_WRITE) && !pte_write(pte))
1565                 goto unlock;
1566
1567         page = vm_normal_page(vma, address, pte);
1568         if (unlikely(!page)) {
1569                 if ((flags & FOLL_DUMP) ||
1570                     !is_zero_pfn(pte_pfn(pte)))
1571                         goto bad_page;
1572                 page = pte_page(pte);
1573         }
1574
1575         if (flags & FOLL_GET)
1576                 get_page_foll(page);
1577         if (flags & FOLL_TOUCH) {
1578                 if ((flags & FOLL_WRITE) &&
1579                     !pte_dirty(pte) && !PageDirty(page))
1580                         set_page_dirty(page);
1581                 /*
1582                  * pte_mkyoung() would be more correct here, but atomic care
1583                  * is needed to avoid losing the dirty bit: it is easier to use
1584                  * mark_page_accessed().
1585                  */
1586                 mark_page_accessed(page);
1587         }
1588         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1589                 /*
1590                  * The preliminary mapping check is mainly to avoid the
1591                  * pointless overhead of lock_page on the ZERO_PAGE
1592                  * which might bounce very badly if there is contention.
1593                  *
1594                  * If the page is already locked, we don't need to
1595                  * handle it now - vmscan will handle it later if and
1596                  * when it attempts to reclaim the page.
1597                  */
1598                 if (page->mapping && trylock_page(page)) {
1599                         lru_add_drain();  /* push cached pages to LRU */
1600                         /*
1601                          * Because we lock page here, and migration is
1602                          * blocked by the pte's page reference, and we
1603                          * know the page is still mapped, we don't even
1604                          * need to check for file-cache page truncation.
1605                          */
1606                         mlock_vma_page(page);
1607                         unlock_page(page);
1608                 }
1609         }
1610 unlock:
1611         pte_unmap_unlock(ptep, ptl);
1612 out:
1613         return page;
1614
1615 bad_page:
1616         pte_unmap_unlock(ptep, ptl);
1617         return ERR_PTR(-EFAULT);
1618
1619 no_page:
1620         pte_unmap_unlock(ptep, ptl);
1621         if (!pte_none(pte))
1622                 return page;
1623
1624 no_page_table:
1625         /*
1626          * When core dumping an enormous anonymous area that nobody
1627          * has touched so far, we don't want to allocate unnecessary pages or
1628          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1629          * then get_dump_page() will return NULL to leave a hole in the dump.
1630          * But we can only make this optimization where a hole would surely
1631          * be zero-filled if handle_mm_fault() actually did handle it.
1632          */
1633         if ((flags & FOLL_DUMP) &&
1634             (!vma->vm_ops || !vma->vm_ops->fault))
1635                 return ERR_PTR(-EFAULT);
1636         return page;
1637 }
1638
1639 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1640 {
1641         return stack_guard_page_start(vma, addr) ||
1642                stack_guard_page_end(vma, addr+PAGE_SIZE);
1643 }
1644
1645 /**
1646  * __get_user_pages() - pin user pages in memory
1647  * @tsk:        task_struct of target task
1648  * @mm:         mm_struct of target mm
1649  * @start:      starting user address
1650  * @nr_pages:   number of pages from start to pin
1651  * @gup_flags:  flags modifying pin behaviour
1652  * @pages:      array that receives pointers to the pages pinned.
1653  *              Should be at least nr_pages long. Or NULL, if caller
1654  *              only intends to ensure the pages are faulted in.
1655  * @vmas:       array of pointers to vmas corresponding to each page.
1656  *              Or NULL if the caller does not require them.
1657  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1658  *
1659  * Returns number of pages pinned. This may be fewer than the number
1660  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1661  * were pinned, returns -errno. Each page returned must be released
1662  * with a put_page() call when it is finished with. vmas will only
1663  * remain valid while mmap_sem is held.
1664  *
1665  * Must be called with mmap_sem held for read or write.
1666  *
1667  * __get_user_pages walks a process's page tables and takes a reference to
1668  * each struct page that each user address corresponds to at a given
1669  * instant. That is, it takes the page that would be accessed if a user
1670  * thread accesses the given user virtual address at that instant.
1671  *
1672  * This does not guarantee that the page exists in the user mappings when
1673  * __get_user_pages returns, and there may even be a completely different
1674  * page there in some cases (eg. if mmapped pagecache has been invalidated
1675  * and subsequently re faulted). However it does guarantee that the page
1676  * won't be freed completely. And mostly callers simply care that the page
1677  * contains data that was valid *at some point in time*. Typically, an IO
1678  * or similar operation cannot guarantee anything stronger anyway because
1679  * locks can't be held over the syscall boundary.
1680  *
1681  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1682  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1683  * appropriate) must be called after the page is finished with, and
1684  * before put_page is called.
1685  *
1686  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1687  * or mmap_sem contention, and if waiting is needed to pin all pages,
1688  * *@nonblocking will be set to 0.
1689  *
1690  * In most cases, get_user_pages or get_user_pages_fast should be used
1691  * instead of __get_user_pages. __get_user_pages should be used only if
1692  * you need some special @gup_flags.
1693  */
1694 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1695                 unsigned long start, unsigned long nr_pages,
1696                 unsigned int gup_flags, struct page **pages,
1697                 struct vm_area_struct **vmas, int *nonblocking)
1698 {
1699         long i;
1700         unsigned long vm_flags;
1701         unsigned int page_mask;
1702
1703         if (!nr_pages)
1704                 return 0;
1705
1706         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1707
1708         /* 
1709          * Require read or write permissions.
1710          * If FOLL_FORCE is set, we only require the "MAY" flags.
1711          */
1712         vm_flags  = (gup_flags & FOLL_WRITE) ?
1713                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1714         vm_flags &= (gup_flags & FOLL_FORCE) ?
1715                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1716
1717         /*
1718          * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1719          * would be called on PROT_NONE ranges. We must never invoke
1720          * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1721          * page faults would unprotect the PROT_NONE ranges if
1722          * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1723          * bitflag. So to avoid that, don't set FOLL_NUMA if
1724          * FOLL_FORCE is set.
1725          */
1726         if (!(gup_flags & FOLL_FORCE))
1727                 gup_flags |= FOLL_NUMA;
1728
1729         i = 0;
1730
1731         do {
1732                 struct vm_area_struct *vma;
1733
1734                 vma = find_extend_vma(mm, start);
1735                 if (!vma && in_gate_area(mm, start)) {
1736                         unsigned long pg = start & PAGE_MASK;
1737                         pgd_t *pgd;
1738                         pud_t *pud;
1739                         pmd_t *pmd;
1740                         pte_t *pte;
1741
1742                         /* user gate pages are read-only */
1743                         if (gup_flags & FOLL_WRITE)
1744                                 return i ? : -EFAULT;
1745                         if (pg > TASK_SIZE)
1746                                 pgd = pgd_offset_k(pg);
1747                         else
1748                                 pgd = pgd_offset_gate(mm, pg);
1749                         BUG_ON(pgd_none(*pgd));
1750                         pud = pud_offset(pgd, pg);
1751                         BUG_ON(pud_none(*pud));
1752                         pmd = pmd_offset(pud, pg);
1753                         if (pmd_none(*pmd))
1754                                 return i ? : -EFAULT;
1755                         VM_BUG_ON(pmd_trans_huge(*pmd));
1756                         pte = pte_offset_map(pmd, pg);
1757                         if (pte_none(*pte)) {
1758                                 pte_unmap(pte);
1759                                 return i ? : -EFAULT;
1760                         }
1761                         vma = get_gate_vma(mm);
1762                         if (pages) {
1763                                 struct page *page;
1764
1765                                 page = vm_normal_page(vma, start, *pte);
1766                                 if (!page) {
1767                                         if (!(gup_flags & FOLL_DUMP) &&
1768                                              is_zero_pfn(pte_pfn(*pte)))
1769                                                 page = pte_page(*pte);
1770                                         else {
1771                                                 pte_unmap(pte);
1772                                                 return i ? : -EFAULT;
1773                                         }
1774                                 }
1775                                 pages[i] = page;
1776                                 get_page(page);
1777                         }
1778                         pte_unmap(pte);
1779                         page_mask = 0;
1780                         goto next_page;
1781                 }
1782
1783                 if (!vma ||
1784                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1785                     !(vm_flags & vma->vm_flags))
1786                         return i ? : -EFAULT;
1787
1788                 if (is_vm_hugetlb_page(vma)) {
1789                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1790                                         &start, &nr_pages, i, gup_flags);
1791                         continue;
1792                 }
1793
1794                 do {
1795                         struct page *page;
1796                         unsigned int foll_flags = gup_flags;
1797                         unsigned int page_increm;
1798
1799                         /*
1800                          * If we have a pending SIGKILL, don't keep faulting
1801                          * pages and potentially allocating memory.
1802                          */
1803                         if (unlikely(fatal_signal_pending(current)))
1804                                 return i ? i : -ERESTARTSYS;
1805
1806                         cond_resched();
1807                         while (!(page = follow_page_mask(vma, start,
1808                                                 foll_flags, &page_mask))) {
1809                                 int ret;
1810                                 unsigned int fault_flags = 0;
1811
1812                                 /* For mlock, just skip the stack guard page. */
1813                                 if (foll_flags & FOLL_MLOCK) {
1814                                         if (stack_guard_page(vma, start))
1815                                                 goto next_page;
1816                                 }
1817                                 if (foll_flags & FOLL_WRITE)
1818                                         fault_flags |= FAULT_FLAG_WRITE;
1819                                 if (nonblocking)
1820                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1821                                 if (foll_flags & FOLL_NOWAIT)
1822                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1823
1824                                 ret = handle_mm_fault(mm, vma, start,
1825                                                         fault_flags);
1826
1827                                 if (ret & VM_FAULT_ERROR) {
1828                                         if (ret & VM_FAULT_OOM)
1829                                                 return i ? i : -ENOMEM;
1830                                         if (ret & (VM_FAULT_HWPOISON |
1831                                                    VM_FAULT_HWPOISON_LARGE)) {
1832                                                 if (i)
1833                                                         return i;
1834                                                 else if (gup_flags & FOLL_HWPOISON)
1835                                                         return -EHWPOISON;
1836                                                 else
1837                                                         return -EFAULT;
1838                                         }
1839                                         if (ret & VM_FAULT_SIGBUS)
1840                                                 return i ? i : -EFAULT;
1841                                         BUG();
1842                                 }
1843
1844                                 if (tsk) {
1845                                         if (ret & VM_FAULT_MAJOR)
1846                                                 tsk->maj_flt++;
1847                                         else
1848                                                 tsk->min_flt++;
1849                                 }
1850
1851                                 if (ret & VM_FAULT_RETRY) {
1852                                         if (nonblocking)
1853                                                 *nonblocking = 0;
1854                                         return i;
1855                                 }
1856
1857                                 /*
1858                                  * The VM_FAULT_WRITE bit tells us that
1859                                  * do_wp_page has broken COW when necessary,
1860                                  * even if maybe_mkwrite decided not to set
1861                                  * pte_write. We can thus safely do subsequent
1862                                  * page lookups as if they were reads. But only
1863                                  * do so when looping for pte_write is futile:
1864                                  * in some cases userspace may also be wanting
1865                                  * to write to the gotten user page, which a
1866                                  * read fault here might prevent (a readonly
1867                                  * page might get reCOWed by userspace write).
1868                                  */
1869                                 if ((ret & VM_FAULT_WRITE) &&
1870                                     !(vma->vm_flags & VM_WRITE))
1871                                         foll_flags &= ~FOLL_WRITE;
1872
1873                                 cond_resched();
1874                         }
1875                         if (IS_ERR(page))
1876                                 return i ? i : PTR_ERR(page);
1877                         if (pages) {
1878                                 pages[i] = page;
1879
1880                                 flush_anon_page(vma, page, start);
1881                                 flush_dcache_page(page);
1882                                 page_mask = 0;
1883                         }
1884 next_page:
1885                         if (vmas) {
1886                                 vmas[i] = vma;
1887                                 page_mask = 0;
1888                         }
1889                         page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1890                         if (page_increm > nr_pages)
1891                                 page_increm = nr_pages;
1892                         i += page_increm;
1893                         start += page_increm * PAGE_SIZE;
1894                         nr_pages -= page_increm;
1895                 } while (nr_pages && start < vma->vm_end);
1896         } while (nr_pages);
1897         return i;
1898 }
1899 EXPORT_SYMBOL(__get_user_pages);
1900
1901 /*
1902  * fixup_user_fault() - manually resolve a user page fault
1903  * @tsk:        the task_struct to use for page fault accounting, or
1904  *              NULL if faults are not to be recorded.
1905  * @mm:         mm_struct of target mm
1906  * @address:    user address
1907  * @fault_flags:flags to pass down to handle_mm_fault()
1908  *
1909  * This is meant to be called in the specific scenario where for locking reasons
1910  * we try to access user memory in atomic context (within a pagefault_disable()
1911  * section), this returns -EFAULT, and we want to resolve the user fault before
1912  * trying again.
1913  *
1914  * Typically this is meant to be used by the futex code.
1915  *
1916  * The main difference with get_user_pages() is that this function will
1917  * unconditionally call handle_mm_fault() which will in turn perform all the
1918  * necessary SW fixup of the dirty and young bits in the PTE, while
1919  * handle_mm_fault() only guarantees to update these in the struct page.
1920  *
1921  * This is important for some architectures where those bits also gate the
1922  * access permission to the page because they are maintained in software.  On
1923  * such architectures, gup() will not be enough to make a subsequent access
1924  * succeed.
1925  *
1926  * This should be called with the mm_sem held for read.
1927  */
1928 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1929                      unsigned long address, unsigned int fault_flags)
1930 {
1931         struct vm_area_struct *vma;
1932         int ret;
1933
1934         vma = find_extend_vma(mm, address);
1935         if (!vma || address < vma->vm_start)
1936                 return -EFAULT;
1937
1938         ret = handle_mm_fault(mm, vma, address, fault_flags);
1939         if (ret & VM_FAULT_ERROR) {
1940                 if (ret & VM_FAULT_OOM)
1941                         return -ENOMEM;
1942                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1943                         return -EHWPOISON;
1944                 if (ret & VM_FAULT_SIGBUS)
1945                         return -EFAULT;
1946                 BUG();
1947         }
1948         if (tsk) {
1949                 if (ret & VM_FAULT_MAJOR)
1950                         tsk->maj_flt++;
1951                 else
1952                         tsk->min_flt++;
1953         }
1954         return 0;
1955 }
1956
1957 /*
1958  * get_user_pages() - pin user pages in memory
1959  * @tsk:        the task_struct to use for page fault accounting, or
1960  *              NULL if faults are not to be recorded.
1961  * @mm:         mm_struct of target mm
1962  * @start:      starting user address
1963  * @nr_pages:   number of pages from start to pin
1964  * @write:      whether pages will be written to by the caller
1965  * @force:      whether to force write access even if user mapping is
1966  *              readonly. This will result in the page being COWed even
1967  *              in MAP_SHARED mappings. You do not want this.
1968  * @pages:      array that receives pointers to the pages pinned.
1969  *              Should be at least nr_pages long. Or NULL, if caller
1970  *              only intends to ensure the pages are faulted in.
1971  * @vmas:       array of pointers to vmas corresponding to each page.
1972  *              Or NULL if the caller does not require them.
1973  *
1974  * Returns number of pages pinned. This may be fewer than the number
1975  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1976  * were pinned, returns -errno. Each page returned must be released
1977  * with a put_page() call when it is finished with. vmas will only
1978  * remain valid while mmap_sem is held.
1979  *
1980  * Must be called with mmap_sem held for read or write.
1981  *
1982  * get_user_pages walks a process's page tables and takes a reference to
1983  * each struct page that each user address corresponds to at a given
1984  * instant. That is, it takes the page that would be accessed if a user
1985  * thread accesses the given user virtual address at that instant.
1986  *
1987  * This does not guarantee that the page exists in the user mappings when
1988  * get_user_pages returns, and there may even be a completely different
1989  * page there in some cases (eg. if mmapped pagecache has been invalidated
1990  * and subsequently re faulted). However it does guarantee that the page
1991  * won't be freed completely. And mostly callers simply care that the page
1992  * contains data that was valid *at some point in time*. Typically, an IO
1993  * or similar operation cannot guarantee anything stronger anyway because
1994  * locks can't be held over the syscall boundary.
1995  *
1996  * If write=0, the page must not be written to. If the page is written to,
1997  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1998  * after the page is finished with, and before put_page is called.
1999  *
2000  * get_user_pages is typically used for fewer-copy IO operations, to get a
2001  * handle on the memory by some means other than accesses via the user virtual
2002  * addresses. The pages may be submitted for DMA to devices or accessed via
2003  * their kernel linear mapping (via the kmap APIs). Care should be taken to
2004  * use the correct cache flushing APIs.
2005  *
2006  * See also get_user_pages_fast, for performance critical applications.
2007  */
2008 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2009                 unsigned long start, unsigned long nr_pages, int write,
2010                 int force, struct page **pages, struct vm_area_struct **vmas)
2011 {
2012         int flags = FOLL_TOUCH;
2013
2014         if (pages)
2015                 flags |= FOLL_GET;
2016         if (write)
2017                 flags |= FOLL_WRITE;
2018         if (force)
2019                 flags |= FOLL_FORCE;
2020
2021         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2022                                 NULL);
2023 }
2024 EXPORT_SYMBOL(get_user_pages);
2025
2026 /**
2027  * get_dump_page() - pin user page in memory while writing it to core dump
2028  * @addr: user address
2029  *
2030  * Returns struct page pointer of user page pinned for dump,
2031  * to be freed afterwards by page_cache_release() or put_page().
2032  *
2033  * Returns NULL on any kind of failure - a hole must then be inserted into
2034  * the corefile, to preserve alignment with its headers; and also returns
2035  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2036  * allowing a hole to be left in the corefile to save diskspace.
2037  *
2038  * Called without mmap_sem, but after all other threads have been killed.
2039  */
2040 #ifdef CONFIG_ELF_CORE
2041 struct page *get_dump_page(unsigned long addr)
2042 {
2043         struct vm_area_struct *vma;
2044         struct page *page;
2045
2046         if (__get_user_pages(current, current->mm, addr, 1,
2047                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2048                              NULL) < 1)
2049                 return NULL;
2050         flush_cache_page(vma, addr, page_to_pfn(page));
2051         return page;
2052 }
2053 #endif /* CONFIG_ELF_CORE */
2054
2055 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2056                         spinlock_t **ptl)
2057 {
2058         pgd_t * pgd = pgd_offset(mm, addr);
2059         pud_t * pud = pud_alloc(mm, pgd, addr);
2060         if (pud) {
2061                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2062                 if (pmd) {
2063                         VM_BUG_ON(pmd_trans_huge(*pmd));
2064                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2065                 }
2066         }
2067         return NULL;
2068 }
2069
2070 /*
2071  * This is the old fallback for page remapping.
2072  *
2073  * For historical reasons, it only allows reserved pages. Only
2074  * old drivers should use this, and they needed to mark their
2075  * pages reserved for the old functions anyway.
2076  */
2077 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2078                         struct page *page, pgprot_t prot)
2079 {
2080         struct mm_struct *mm = vma->vm_mm;
2081         int retval;
2082         pte_t *pte;
2083         spinlock_t *ptl;
2084
2085         retval = -EINVAL;
2086         if (PageAnon(page))
2087                 goto out;
2088         retval = -ENOMEM;
2089         flush_dcache_page(page);
2090         pte = get_locked_pte(mm, addr, &ptl);
2091         if (!pte)
2092                 goto out;
2093         retval = -EBUSY;
2094         if (!pte_none(*pte))
2095                 goto out_unlock;
2096
2097         /* Ok, finally just insert the thing.. */
2098         get_page(page);
2099         inc_mm_counter_fast(mm, MM_FILEPAGES);
2100         page_add_file_rmap(page);
2101         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2102
2103         retval = 0;
2104         pte_unmap_unlock(pte, ptl);
2105         return retval;
2106 out_unlock:
2107         pte_unmap_unlock(pte, ptl);
2108 out:
2109         return retval;
2110 }
2111
2112 /**
2113  * vm_insert_page - insert single page into user vma
2114  * @vma: user vma to map to
2115  * @addr: target user address of this page
2116  * @page: source kernel page
2117  *
2118  * This allows drivers to insert individual pages they've allocated
2119  * into a user vma.
2120  *
2121  * The page has to be a nice clean _individual_ kernel allocation.
2122  * If you allocate a compound page, you need to have marked it as
2123  * such (__GFP_COMP), or manually just split the page up yourself
2124  * (see split_page()).
2125  *
2126  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2127  * took an arbitrary page protection parameter. This doesn't allow
2128  * that. Your vma protection will have to be set up correctly, which
2129  * means that if you want a shared writable mapping, you'd better
2130  * ask for a shared writable mapping!
2131  *
2132  * The page does not need to be reserved.
2133  *
2134  * Usually this function is called from f_op->mmap() handler
2135  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2136  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2137  * function from other places, for example from page-fault handler.
2138  */
2139 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2140                         struct page *page)
2141 {
2142         if (addr < vma->vm_start || addr >= vma->vm_end)
2143                 return -EFAULT;
2144         if (!page_count(page))
2145                 return -EINVAL;
2146         if (!(vma->vm_flags & VM_MIXEDMAP)) {
2147                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2148                 BUG_ON(vma->vm_flags & VM_PFNMAP);
2149                 vma->vm_flags |= VM_MIXEDMAP;
2150         }
2151         return insert_page(vma, addr, page, vma->vm_page_prot);
2152 }
2153 EXPORT_SYMBOL(vm_insert_page);
2154
2155 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2156                         unsigned long pfn, pgprot_t prot)
2157 {
2158         struct mm_struct *mm = vma->vm_mm;
2159         int retval;
2160         pte_t *pte, entry;
2161         spinlock_t *ptl;
2162
2163         retval = -ENOMEM;
2164         pte = get_locked_pte(mm, addr, &ptl);
2165         if (!pte)
2166                 goto out;
2167         retval = -EBUSY;
2168         if (!pte_none(*pte))
2169                 goto out_unlock;
2170
2171         /* Ok, finally just insert the thing.. */
2172         entry = pte_mkspecial(pfn_pte(pfn, prot));
2173         set_pte_at(mm, addr, pte, entry);
2174         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2175
2176         retval = 0;
2177 out_unlock:
2178         pte_unmap_unlock(pte, ptl);
2179 out:
2180         return retval;
2181 }
2182
2183 /**
2184  * vm_insert_pfn - insert single pfn into user vma
2185  * @vma: user vma to map to
2186  * @addr: target user address of this page
2187  * @pfn: source kernel pfn
2188  *
2189  * Similar to vm_insert_page, this allows drivers to insert individual pages
2190  * they've allocated into a user vma. Same comments apply.
2191  *
2192  * This function should only be called from a vm_ops->fault handler, and
2193  * in that case the handler should return NULL.
2194  *
2195  * vma cannot be a COW mapping.
2196  *
2197  * As this is called only for pages that do not currently exist, we
2198  * do not need to flush old virtual caches or the TLB.
2199  */
2200 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2201                         unsigned long pfn)
2202 {
2203         int ret;
2204         pgprot_t pgprot = vma->vm_page_prot;
2205         /*
2206          * Technically, architectures with pte_special can avoid all these
2207          * restrictions (same for remap_pfn_range).  However we would like
2208          * consistency in testing and feature parity among all, so we should
2209          * try to keep these invariants in place for everybody.
2210          */
2211         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2212         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2213                                                 (VM_PFNMAP|VM_MIXEDMAP));
2214         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2215         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2216
2217         if (addr < vma->vm_start || addr >= vma->vm_end)
2218                 return -EFAULT;
2219         if (track_pfn_insert(vma, &pgprot, pfn))
2220                 return -EINVAL;
2221
2222         ret = insert_pfn(vma, addr, pfn, pgprot);
2223
2224         return ret;
2225 }
2226 EXPORT_SYMBOL(vm_insert_pfn);
2227
2228 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2229                         unsigned long pfn)
2230 {
2231         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2232
2233         if (addr < vma->vm_start || addr >= vma->vm_end)
2234                 return -EFAULT;
2235
2236         /*
2237          * If we don't have pte special, then we have to use the pfn_valid()
2238          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2239          * refcount the page if pfn_valid is true (hence insert_page rather
2240          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2241          * without pte special, it would there be refcounted as a normal page.
2242          */
2243         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2244                 struct page *page;
2245
2246                 page = pfn_to_page(pfn);
2247                 return insert_page(vma, addr, page, vma->vm_page_prot);
2248         }
2249         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2250 }
2251 EXPORT_SYMBOL(vm_insert_mixed);
2252
2253 /*
2254  * maps a range of physical memory into the requested pages. the old
2255  * mappings are removed. any references to nonexistent pages results
2256  * in null mappings (currently treated as "copy-on-access")
2257  */
2258 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2259                         unsigned long addr, unsigned long end,
2260                         unsigned long pfn, pgprot_t prot)
2261 {
2262         pte_t *pte;
2263         spinlock_t *ptl;
2264
2265         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2266         if (!pte)
2267                 return -ENOMEM;
2268         arch_enter_lazy_mmu_mode();
2269         do {
2270                 BUG_ON(!pte_none(*pte));
2271                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2272                 pfn++;
2273         } while (pte++, addr += PAGE_SIZE, addr != end);
2274         arch_leave_lazy_mmu_mode();
2275         pte_unmap_unlock(pte - 1, ptl);
2276         return 0;
2277 }
2278
2279 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2280                         unsigned long addr, unsigned long end,
2281                         unsigned long pfn, pgprot_t prot)
2282 {
2283         pmd_t *pmd;
2284         unsigned long next;
2285
2286         pfn -= addr >> PAGE_SHIFT;
2287         pmd = pmd_alloc(mm, pud, addr);
2288         if (!pmd)
2289                 return -ENOMEM;
2290         VM_BUG_ON(pmd_trans_huge(*pmd));
2291         do {
2292                 next = pmd_addr_end(addr, end);
2293                 if (remap_pte_range(mm, pmd, addr, next,
2294                                 pfn + (addr >> PAGE_SHIFT), prot))
2295                         return -ENOMEM;
2296         } while (pmd++, addr = next, addr != end);
2297         return 0;
2298 }
2299
2300 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2301                         unsigned long addr, unsigned long end,
2302                         unsigned long pfn, pgprot_t prot)
2303 {
2304         pud_t *pud;
2305         unsigned long next;
2306
2307         pfn -= addr >> PAGE_SHIFT;
2308         pud = pud_alloc(mm, pgd, addr);
2309         if (!pud)
2310                 return -ENOMEM;
2311         do {
2312                 next = pud_addr_end(addr, end);
2313                 if (remap_pmd_range(mm, pud, addr, next,
2314                                 pfn + (addr >> PAGE_SHIFT), prot))
2315                         return -ENOMEM;
2316         } while (pud++, addr = next, addr != end);
2317         return 0;
2318 }
2319
2320 /**
2321  * remap_pfn_range - remap kernel memory to userspace
2322  * @vma: user vma to map to
2323  * @addr: target user address to start at
2324  * @pfn: physical address of kernel memory
2325  * @size: size of map area
2326  * @prot: page protection flags for this mapping
2327  *
2328  *  Note: this is only safe if the mm semaphore is held when called.
2329  */
2330 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2331                     unsigned long pfn, unsigned long size, pgprot_t prot)
2332 {
2333         pgd_t *pgd;
2334         unsigned long next;
2335         unsigned long end = addr + PAGE_ALIGN(size);
2336         struct mm_struct *mm = vma->vm_mm;
2337         int err;
2338
2339         /*
2340          * Physically remapped pages are special. Tell the
2341          * rest of the world about it:
2342          *   VM_IO tells people not to look at these pages
2343          *      (accesses can have side effects).
2344          *   VM_PFNMAP tells the core MM that the base pages are just
2345          *      raw PFN mappings, and do not have a "struct page" associated
2346          *      with them.
2347          *   VM_DONTEXPAND
2348          *      Disable vma merging and expanding with mremap().
2349          *   VM_DONTDUMP
2350          *      Omit vma from core dump, even when VM_IO turned off.
2351          *
2352          * There's a horrible special case to handle copy-on-write
2353          * behaviour that some programs depend on. We mark the "original"
2354          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2355          * See vm_normal_page() for details.
2356          */
2357         if (is_cow_mapping(vma->vm_flags)) {
2358                 if (addr != vma->vm_start || end != vma->vm_end)
2359                         return -EINVAL;
2360                 vma->vm_pgoff = pfn;
2361         }
2362
2363         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2364         if (err)
2365                 return -EINVAL;
2366
2367         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2368
2369         BUG_ON(addr >= end);
2370         pfn -= addr >> PAGE_SHIFT;
2371         pgd = pgd_offset(mm, addr);
2372         flush_cache_range(vma, addr, end);
2373         do {
2374                 next = pgd_addr_end(addr, end);
2375                 err = remap_pud_range(mm, pgd, addr, next,
2376                                 pfn + (addr >> PAGE_SHIFT), prot);
2377                 if (err)
2378                         break;
2379         } while (pgd++, addr = next, addr != end);
2380
2381         if (err)
2382                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2383
2384         return err;
2385 }
2386 EXPORT_SYMBOL(remap_pfn_range);
2387
2388 /**
2389  * vm_iomap_memory - remap memory to userspace
2390  * @vma: user vma to map to
2391  * @start: start of area
2392  * @len: size of area
2393  *
2394  * This is a simplified io_remap_pfn_range() for common driver use. The
2395  * driver just needs to give us the physical memory range to be mapped,
2396  * we'll figure out the rest from the vma information.
2397  *
2398  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2399  * whatever write-combining details or similar.
2400  */
2401 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2402 {
2403         unsigned long vm_len, pfn, pages;
2404
2405         /* Check that the physical memory area passed in looks valid */
2406         if (start + len < start)
2407                 return -EINVAL;
2408         /*
2409          * You *really* shouldn't map things that aren't page-aligned,
2410          * but we've historically allowed it because IO memory might
2411          * just have smaller alignment.
2412          */
2413         len += start & ~PAGE_MASK;
2414         pfn = start >> PAGE_SHIFT;
2415         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2416         if (pfn + pages < pfn)
2417                 return -EINVAL;
2418
2419         /* We start the mapping 'vm_pgoff' pages into the area */
2420         if (vma->vm_pgoff > pages)
2421                 return -EINVAL;
2422         pfn += vma->vm_pgoff;
2423         pages -= vma->vm_pgoff;
2424
2425         /* Can we fit all of the mapping? */
2426         vm_len = vma->vm_end - vma->vm_start;
2427         if (vm_len >> PAGE_SHIFT > pages)
2428                 return -EINVAL;
2429
2430         /* Ok, let it rip */
2431         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2432 }
2433 EXPORT_SYMBOL(vm_iomap_memory);
2434
2435 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2436                                      unsigned long addr, unsigned long end,
2437                                      pte_fn_t fn, void *data)
2438 {
2439         pte_t *pte;
2440         int err;
2441         pgtable_t token;
2442         spinlock_t *uninitialized_var(ptl);
2443
2444         pte = (mm == &init_mm) ?
2445                 pte_alloc_kernel(pmd, addr) :
2446                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2447         if (!pte)
2448                 return -ENOMEM;
2449
2450         BUG_ON(pmd_huge(*pmd));
2451
2452         arch_enter_lazy_mmu_mode();
2453
2454         token = pmd_pgtable(*pmd);
2455
2456         do {
2457                 err = fn(pte++, token, addr, data);
2458                 if (err)
2459                         break;
2460         } while (addr += PAGE_SIZE, addr != end);
2461
2462         arch_leave_lazy_mmu_mode();
2463
2464         if (mm != &init_mm)
2465                 pte_unmap_unlock(pte-1, ptl);
2466         return err;
2467 }
2468
2469 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2470                                      unsigned long addr, unsigned long end,
2471                                      pte_fn_t fn, void *data)
2472 {
2473         pmd_t *pmd;
2474         unsigned long next;
2475         int err;
2476
2477         BUG_ON(pud_huge(*pud));
2478
2479         pmd = pmd_alloc(mm, pud, addr);
2480         if (!pmd)
2481                 return -ENOMEM;
2482         do {
2483                 next = pmd_addr_end(addr, end);
2484                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2485                 if (err)
2486                         break;
2487         } while (pmd++, addr = next, addr != end);
2488         return err;
2489 }
2490
2491 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2492                                      unsigned long addr, unsigned long end,
2493                                      pte_fn_t fn, void *data)
2494 {
2495         pud_t *pud;
2496         unsigned long next;
2497         int err;
2498
2499         pud = pud_alloc(mm, pgd, addr);
2500         if (!pud)
2501                 return -ENOMEM;
2502         do {
2503                 next = pud_addr_end(addr, end);
2504                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2505                 if (err)
2506                         break;
2507         } while (pud++, addr = next, addr != end);
2508         return err;
2509 }
2510
2511 /*
2512  * Scan a region of virtual memory, filling in page tables as necessary
2513  * and calling a provided function on each leaf page table.
2514  */
2515 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2516                         unsigned long size, pte_fn_t fn, void *data)
2517 {
2518         pgd_t *pgd;
2519         unsigned long next;
2520         unsigned long end = addr + size;
2521         int err;
2522
2523         BUG_ON(addr >= end);
2524         pgd = pgd_offset(mm, addr);
2525         do {
2526                 next = pgd_addr_end(addr, end);
2527                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2528                 if (err)
2529                         break;
2530         } while (pgd++, addr = next, addr != end);
2531
2532         return err;
2533 }
2534 EXPORT_SYMBOL_GPL(apply_to_page_range);
2535
2536 /*
2537  * handle_pte_fault chooses page fault handler according to an entry
2538  * which was read non-atomically.  Before making any commitment, on
2539  * those architectures or configurations (e.g. i386 with PAE) which
2540  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2541  * must check under lock before unmapping the pte and proceeding
2542  * (but do_wp_page is only called after already making such a check;
2543  * and do_anonymous_page can safely check later on).
2544  */
2545 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2546                                 pte_t *page_table, pte_t orig_pte)
2547 {
2548         int same = 1;
2549 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2550         if (sizeof(pte_t) > sizeof(unsigned long)) {
2551                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2552                 spin_lock(ptl);
2553                 same = pte_same(*page_table, orig_pte);
2554                 spin_unlock(ptl);
2555         }
2556 #endif
2557         pte_unmap(page_table);
2558         return same;
2559 }
2560
2561 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2562 {
2563         /*
2564          * If the source page was a PFN mapping, we don't have
2565          * a "struct page" for it. We do a best-effort copy by
2566          * just copying from the original user address. If that
2567          * fails, we just zero-fill it. Live with it.
2568          */
2569         if (unlikely(!src)) {
2570                 void *kaddr = kmap_atomic(dst);
2571                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2572
2573                 /*
2574                  * This really shouldn't fail, because the page is there
2575                  * in the page tables. But it might just be unreadable,
2576                  * in which case we just give up and fill the result with
2577                  * zeroes.
2578                  */
2579                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2580                         clear_page(kaddr);
2581                 kunmap_atomic(kaddr);
2582                 flush_dcache_page(dst);
2583         } else
2584                 copy_user_highpage(dst, src, va, vma);
2585 }
2586
2587 /*
2588  * This routine handles present pages, when users try to write
2589  * to a shared page. It is done by copying the page to a new address
2590  * and decrementing the shared-page counter for the old page.
2591  *
2592  * Note that this routine assumes that the protection checks have been
2593  * done by the caller (the low-level page fault routine in most cases).
2594  * Thus we can safely just mark it writable once we've done any necessary
2595  * COW.
2596  *
2597  * We also mark the page dirty at this point even though the page will
2598  * change only once the write actually happens. This avoids a few races,
2599  * and potentially makes it more efficient.
2600  *
2601  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2602  * but allow concurrent faults), with pte both mapped and locked.
2603  * We return with mmap_sem still held, but pte unmapped and unlocked.
2604  */
2605 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2606                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2607                 spinlock_t *ptl, pte_t orig_pte)
2608         __releases(ptl)
2609 {
2610         struct page *old_page, *new_page = NULL;
2611         pte_t entry;
2612         int ret = 0;
2613         int page_mkwrite = 0;
2614         struct page *dirty_page = NULL;
2615         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2616         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2617
2618         old_page = vm_normal_page(vma, address, orig_pte);
2619         if (!old_page) {
2620                 /*
2621                  * VM_MIXEDMAP !pfn_valid() case
2622                  *
2623                  * We should not cow pages in a shared writeable mapping.
2624                  * Just mark the pages writable as we can't do any dirty
2625                  * accounting on raw pfn maps.
2626                  */
2627                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2628                                      (VM_WRITE|VM_SHARED))
2629                         goto reuse;
2630                 goto gotten;
2631         }
2632
2633         /*
2634          * Take out anonymous pages first, anonymous shared vmas are
2635          * not dirty accountable.
2636          */
2637         if (PageAnon(old_page) && !PageKsm(old_page)) {
2638                 if (!trylock_page(old_page)) {
2639                         page_cache_get(old_page);
2640                         pte_unmap_unlock(page_table, ptl);
2641                         lock_page(old_page);
2642                         page_table = pte_offset_map_lock(mm, pmd, address,
2643                                                          &ptl);
2644                         if (!pte_same(*page_table, orig_pte)) {
2645                                 unlock_page(old_page);
2646                                 goto unlock;
2647                         }
2648                         page_cache_release(old_page);
2649                 }
2650                 if (reuse_swap_page(old_page)) {
2651                         /*
2652                          * The page is all ours.  Move it to our anon_vma so
2653                          * the rmap code will not search our parent or siblings.
2654                          * Protected against the rmap code by the page lock.
2655                          */
2656                         page_move_anon_rmap(old_page, vma, address);
2657                         unlock_page(old_page);
2658                         goto reuse;
2659                 }
2660                 unlock_page(old_page);
2661         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2662                                         (VM_WRITE|VM_SHARED))) {
2663                 /*
2664                  * Only catch write-faults on shared writable pages,
2665                  * read-only shared pages can get COWed by
2666                  * get_user_pages(.write=1, .force=1).
2667                  */
2668                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2669                         struct vm_fault vmf;
2670                         int tmp;
2671
2672                         vmf.virtual_address = (void __user *)(address &
2673                                                                 PAGE_MASK);
2674                         vmf.pgoff = old_page->index;
2675                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2676                         vmf.page = old_page;
2677
2678                         /*
2679                          * Notify the address space that the page is about to
2680                          * become writable so that it can prohibit this or wait
2681                          * for the page to get into an appropriate state.
2682                          *
2683                          * We do this without the lock held, so that it can
2684                          * sleep if it needs to.
2685                          */
2686                         page_cache_get(old_page);
2687                         pte_unmap_unlock(page_table, ptl);
2688
2689                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2690                         if (unlikely(tmp &
2691                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2692                                 ret = tmp;
2693                                 goto unwritable_page;
2694                         }
2695                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2696                                 lock_page(old_page);
2697                                 if (!old_page->mapping) {
2698                                         ret = 0; /* retry the fault */
2699                                         unlock_page(old_page);
2700                                         goto unwritable_page;
2701                                 }
2702                         } else
2703                                 VM_BUG_ON(!PageLocked(old_page));
2704
2705                         /*
2706                          * Since we dropped the lock we need to revalidate
2707                          * the PTE as someone else may have changed it.  If
2708                          * they did, we just return, as we can count on the
2709                          * MMU to tell us if they didn't also make it writable.
2710                          */
2711                         page_table = pte_offset_map_lock(mm, pmd, address,
2712                                                          &ptl);
2713                         if (!pte_same(*page_table, orig_pte)) {
2714                                 unlock_page(old_page);
2715                                 goto unlock;
2716                         }
2717
2718                         page_mkwrite = 1;
2719                 }
2720                 dirty_page = old_page;
2721                 get_page(dirty_page);
2722
2723 reuse:
2724                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2725                 entry = pte_mkyoung(orig_pte);
2726                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2727                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2728                         update_mmu_cache(vma, address, page_table);
2729                 pte_unmap_unlock(page_table, ptl);
2730                 ret |= VM_FAULT_WRITE;
2731
2732                 if (!dirty_page)
2733                         return ret;
2734
2735                 /*
2736                  * Yes, Virginia, this is actually required to prevent a race
2737                  * with clear_page_dirty_for_io() from clearing the page dirty
2738                  * bit after it clear all dirty ptes, but before a racing
2739                  * do_wp_page installs a dirty pte.
2740                  *
2741                  * __do_fault is protected similarly.
2742                  */
2743                 if (!page_mkwrite) {
2744                         wait_on_page_locked(dirty_page);
2745                         set_page_dirty_balance(dirty_page, page_mkwrite);
2746                         /* file_update_time outside page_lock */
2747                         if (vma->vm_file)
2748                                 file_update_time(vma->vm_file);
2749                 }
2750                 put_page(dirty_page);
2751                 if (page_mkwrite) {
2752                         struct address_space *mapping = dirty_page->mapping;
2753
2754                         set_page_dirty(dirty_page);
2755                         unlock_page(dirty_page);
2756                         page_cache_release(dirty_page);
2757                         if (mapping)    {
2758                                 /*
2759                                  * Some device drivers do not set page.mapping
2760                                  * but still dirty their pages
2761                                  */
2762                                 balance_dirty_pages_ratelimited(mapping);
2763                         }
2764                 }
2765
2766                 return ret;
2767         }
2768
2769         /*
2770          * Ok, we need to copy. Oh, well..
2771          */
2772         page_cache_get(old_page);
2773 gotten:
2774         pte_unmap_unlock(page_table, ptl);
2775
2776         if (unlikely(anon_vma_prepare(vma)))
2777                 goto oom;
2778
2779         if (is_zero_pfn(pte_pfn(orig_pte))) {
2780                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2781                 if (!new_page)
2782                         goto oom;
2783         } else {
2784                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2785                 if (!new_page)
2786                         goto oom;
2787                 cow_user_page(new_page, old_page, address, vma);
2788         }
2789         __SetPageUptodate(new_page);
2790
2791         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2792                 goto oom_free_new;
2793
2794         mmun_start  = address & PAGE_MASK;
2795         mmun_end    = mmun_start + PAGE_SIZE;
2796         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2797
2798         /*
2799          * Re-check the pte - we dropped the lock
2800          */
2801         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2802         if (likely(pte_same(*page_table, orig_pte))) {
2803                 if (old_page) {
2804                         if (!PageAnon(old_page)) {
2805                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2806                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2807                         }
2808                 } else
2809                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2810                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2811                 entry = mk_pte(new_page, vma->vm_page_prot);
2812                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2813                 /*
2814                  * Clear the pte entry and flush it first, before updating the
2815                  * pte with the new entry. This will avoid a race condition
2816                  * seen in the presence of one thread doing SMC and another
2817                  * thread doing COW.
2818                  */
2819                 ptep_clear_flush(vma, address, page_table);
2820                 page_add_new_anon_rmap(new_page, vma, address);
2821                 /*
2822                  * We call the notify macro here because, when using secondary
2823                  * mmu page tables (such as kvm shadow page tables), we want the
2824                  * new page to be mapped directly into the secondary page table.
2825                  */
2826                 set_pte_at_notify(mm, address, page_table, entry);
2827                 update_mmu_cache(vma, address, page_table);
2828                 if (old_page) {
2829                         /*
2830                          * Only after switching the pte to the new page may
2831                          * we remove the mapcount here. Otherwise another
2832                          * process may come and find the rmap count decremented
2833                          * before the pte is switched to the new page, and
2834                          * "reuse" the old page writing into it while our pte
2835                          * here still points into it and can be read by other
2836                          * threads.
2837                          *
2838                          * The critical issue is to order this
2839                          * page_remove_rmap with the ptp_clear_flush above.
2840                          * Those stores are ordered by (if nothing else,)
2841                          * the barrier present in the atomic_add_negative
2842                          * in page_remove_rmap.
2843                          *
2844                          * Then the TLB flush in ptep_clear_flush ensures that
2845                          * no process can access the old page before the
2846                          * decremented mapcount is visible. And the old page
2847                          * cannot be reused until after the decremented
2848                          * mapcount is visible. So transitively, TLBs to
2849                          * old page will be flushed before it can be reused.
2850                          */
2851                         page_remove_rmap(old_page);
2852                 }
2853
2854                 /* Free the old page.. */
2855                 new_page = old_page;
2856                 ret |= VM_FAULT_WRITE;
2857         } else
2858                 mem_cgroup_uncharge_page(new_page);
2859
2860         if (new_page)
2861                 page_cache_release(new_page);
2862 unlock:
2863         pte_unmap_unlock(page_table, ptl);
2864         if (mmun_end > mmun_start)
2865                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2866         if (old_page) {
2867                 /*
2868                  * Don't let another task, with possibly unlocked vma,
2869                  * keep the mlocked page.
2870                  */
2871                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2872                         lock_page(old_page);    /* LRU manipulation */
2873                         munlock_vma_page(old_page);
2874                         unlock_page(old_page);
2875                 }
2876                 page_cache_release(old_page);
2877         }
2878         return ret;
2879 oom_free_new:
2880         page_cache_release(new_page);
2881 oom:
2882         if (old_page)
2883                 page_cache_release(old_page);
2884         return VM_FAULT_OOM;
2885
2886 unwritable_page:
2887         page_cache_release(old_page);
2888         return ret;
2889 }
2890
2891 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2892                 unsigned long start_addr, unsigned long end_addr,
2893                 struct zap_details *details)
2894 {
2895         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2896 }
2897
2898 static inline void unmap_mapping_range_tree(struct rb_root *root,
2899                                             struct zap_details *details)
2900 {
2901         struct vm_area_struct *vma;
2902         pgoff_t vba, vea, zba, zea;
2903
2904         vma_interval_tree_foreach(vma, root,
2905                         details->first_index, details->last_index) {
2906
2907                 vba = vma->vm_pgoff;
2908                 vea = vba + vma_pages(vma) - 1;
2909                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2910                 zba = details->first_index;
2911                 if (zba < vba)
2912                         zba = vba;
2913                 zea = details->last_index;
2914                 if (zea > vea)
2915                         zea = vea;
2916
2917                 unmap_mapping_range_vma(vma,
2918                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2919                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2920                                 details);
2921         }
2922 }
2923
2924 static inline void unmap_mapping_range_list(struct list_head *head,
2925                                             struct zap_details *details)
2926 {
2927         struct vm_area_struct *vma;
2928
2929         /*
2930          * In nonlinear VMAs there is no correspondence between virtual address
2931          * offset and file offset.  So we must perform an exhaustive search
2932          * across *all* the pages in each nonlinear VMA, not just the pages
2933          * whose virtual address lies outside the file truncation point.
2934          */
2935         list_for_each_entry(vma, head, shared.nonlinear) {
2936                 details->nonlinear_vma = vma;
2937                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2938         }
2939 }
2940
2941 /**
2942  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2943  * @mapping: the address space containing mmaps to be unmapped.
2944  * @holebegin: byte in first page to unmap, relative to the start of
2945  * the underlying file.  This will be rounded down to a PAGE_SIZE
2946  * boundary.  Note that this is different from truncate_pagecache(), which
2947  * must keep the partial page.  In contrast, we must get rid of
2948  * partial pages.
2949  * @holelen: size of prospective hole in bytes.  This will be rounded
2950  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2951  * end of the file.
2952  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2953  * but 0 when invalidating pagecache, don't throw away private data.
2954  */
2955 void unmap_mapping_range(struct address_space *mapping,
2956                 loff_t const holebegin, loff_t const holelen, int even_cows)
2957 {
2958         struct zap_details details;
2959         pgoff_t hba = holebegin >> PAGE_SHIFT;
2960         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2961
2962         /* Check for overflow. */
2963         if (sizeof(holelen) > sizeof(hlen)) {
2964                 long long holeend =
2965                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2966                 if (holeend & ~(long long)ULONG_MAX)
2967                         hlen = ULONG_MAX - hba + 1;
2968         }
2969
2970         details.check_mapping = even_cows? NULL: mapping;
2971         details.nonlinear_vma = NULL;
2972         details.first_index = hba;
2973         details.last_index = hba + hlen - 1;
2974         if (details.last_index < details.first_index)
2975                 details.last_index = ULONG_MAX;
2976
2977
2978         mutex_lock(&mapping->i_mmap_mutex);
2979         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2980                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2981         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2982                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2983         mutex_unlock(&mapping->i_mmap_mutex);
2984 }
2985 EXPORT_SYMBOL(unmap_mapping_range);
2986
2987 /*
2988  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2989  * but allow concurrent faults), and pte mapped but not yet locked.
2990  * We return with mmap_sem still held, but pte unmapped and unlocked.
2991  */
2992 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2993                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2994                 unsigned int flags, pte_t orig_pte)
2995 {
2996         spinlock_t *ptl;
2997         struct page *page, *swapcache;
2998         swp_entry_t entry;
2999         pte_t pte;
3000         int locked;
3001         struct mem_cgroup *ptr;
3002         int exclusive = 0;
3003         int ret = 0;
3004
3005         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3006                 goto out;
3007
3008         entry = pte_to_swp_entry(orig_pte);
3009         if (unlikely(non_swap_entry(entry))) {
3010                 if (is_migration_entry(entry)) {
3011                         migration_entry_wait(mm, pmd, address);
3012                 } else if (is_hwpoison_entry(entry)) {
3013                         ret = VM_FAULT_HWPOISON;
3014                 } else {
3015                         print_bad_pte(vma, address, orig_pte, NULL);
3016                         ret = VM_FAULT_SIGBUS;
3017                 }
3018                 goto out;
3019         }
3020         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3021         page = lookup_swap_cache(entry);
3022         if (!page) {
3023                 page = swapin_readahead(entry,
3024                                         GFP_HIGHUSER_MOVABLE, vma, address);
3025                 if (!page) {
3026                         /*
3027                          * Back out if somebody else faulted in this pte
3028                          * while we released the pte lock.
3029                          */
3030                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3031                         if (likely(pte_same(*page_table, orig_pte)))
3032                                 ret = VM_FAULT_OOM;
3033                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3034                         goto unlock;
3035                 }
3036
3037                 /* Had to read the page from swap area: Major fault */
3038                 ret = VM_FAULT_MAJOR;
3039                 count_vm_event(PGMAJFAULT);
3040                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3041         } else if (PageHWPoison(page)) {
3042                 /*
3043                  * hwpoisoned dirty swapcache pages are kept for killing
3044                  * owner processes (which may be unknown at hwpoison time)
3045                  */
3046                 ret = VM_FAULT_HWPOISON;
3047                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3048                 swapcache = page;
3049                 goto out_release;
3050         }
3051
3052         swapcache = page;
3053         locked = lock_page_or_retry(page, mm, flags);
3054
3055         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3056         if (!locked) {
3057                 ret |= VM_FAULT_RETRY;
3058                 goto out_release;
3059         }
3060
3061         /*
3062          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3063          * release the swapcache from under us.  The page pin, and pte_same
3064          * test below, are not enough to exclude that.  Even if it is still
3065          * swapcache, we need to check that the page's swap has not changed.
3066          */
3067         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3068                 goto out_page;
3069
3070         page = ksm_might_need_to_copy(page, vma, address);
3071         if (unlikely(!page)) {
3072                 ret = VM_FAULT_OOM;
3073                 page = swapcache;
3074                 goto out_page;
3075         }
3076
3077         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3078                 ret = VM_FAULT_OOM;
3079                 goto out_page;
3080         }
3081
3082         /*
3083          * Back out if somebody else already faulted in this pte.
3084          */
3085         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3086         if (unlikely(!pte_same(*page_table, orig_pte)))
3087                 goto out_nomap;
3088
3089         if (unlikely(!PageUptodate(page))) {
3090                 ret = VM_FAULT_SIGBUS;
3091                 goto out_nomap;
3092         }
3093
3094         /*
3095          * The page isn't present yet, go ahead with the fault.
3096          *
3097          * Be careful about the sequence of operations here.
3098          * To get its accounting right, reuse_swap_page() must be called
3099          * while the page is counted on swap but not yet in mapcount i.e.
3100          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3101          * must be called after the swap_free(), or it will never succeed.
3102          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3103          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3104          * in page->private. In this case, a record in swap_cgroup  is silently
3105          * discarded at swap_free().
3106          */
3107
3108         inc_mm_counter_fast(mm, MM_ANONPAGES);
3109         dec_mm_counter_fast(mm, MM_SWAPENTS);
3110         pte = mk_pte(page, vma->vm_page_prot);
3111         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3112                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3113                 flags &= ~FAULT_FLAG_WRITE;
3114                 ret |= VM_FAULT_WRITE;
3115                 exclusive = 1;
3116         }
3117         flush_icache_page(vma, page);
3118         if (pte_swp_soft_dirty(orig_pte))
3119                 pte = pte_mksoft_dirty(pte);
3120         set_pte_at(mm, address, page_table, pte);
3121         if (page == swapcache)
3122                 do_page_add_anon_rmap(page, vma, address, exclusive);
3123         else /* ksm created a completely new copy */
3124                 page_add_new_anon_rmap(page, vma, address);
3125         /* It's better to call commit-charge after rmap is established */
3126         mem_cgroup_commit_charge_swapin(page, ptr);
3127
3128         swap_free(entry);
3129         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3130                 try_to_free_swap(page);
3131         unlock_page(page);
3132         if (page != swapcache) {
3133                 /*
3134                  * Hold the lock to avoid the swap entry to be reused
3135                  * until we take the PT lock for the pte_same() check
3136                  * (to avoid false positives from pte_same). For
3137                  * further safety release the lock after the swap_free
3138                  * so that the swap count won't change under a
3139                  * parallel locked swapcache.
3140                  */
3141                 unlock_page(swapcache);
3142                 page_cache_release(swapcache);
3143         }
3144
3145         if (flags & FAULT_FLAG_WRITE) {
3146                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3147                 if (ret & VM_FAULT_ERROR)
3148                         ret &= VM_FAULT_ERROR;
3149                 goto out;
3150         }
3151
3152         /* No need to invalidate - it was non-present before */
3153         update_mmu_cache(vma, address, page_table);
3154 unlock:
3155         pte_unmap_unlock(page_table, ptl);
3156 out:
3157         return ret;
3158 out_nomap:
3159         mem_cgroup_cancel_charge_swapin(ptr);
3160         pte_unmap_unlock(page_table, ptl);
3161 out_page:
3162         unlock_page(page);
3163 out_release:
3164         page_cache_release(page);
3165         if (page != swapcache) {
3166                 unlock_page(swapcache);
3167                 page_cache_release(swapcache);
3168         }
3169         return ret;
3170 }
3171
3172 /*
3173  * This is like a special single-page "expand_{down|up}wards()",
3174  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3175  * doesn't hit another vma.
3176  */
3177 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3178 {
3179         address &= PAGE_MASK;
3180         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3181                 struct vm_area_struct *prev = vma->vm_prev;
3182
3183                 /*
3184                  * Is there a mapping abutting this one below?
3185                  *
3186                  * That's only ok if it's the same stack mapping
3187                  * that has gotten split..
3188                  */
3189                 if (prev && prev->vm_end == address)
3190                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3191
3192                 expand_downwards(vma, address - PAGE_SIZE);
3193         }
3194         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3195                 struct vm_area_struct *next = vma->vm_next;
3196
3197                 /* As VM_GROWSDOWN but s/below/above/ */
3198                 if (next && next->vm_start == address + PAGE_SIZE)
3199                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3200
3201                 expand_upwards(vma, address + PAGE_SIZE);
3202         }
3203         return 0;
3204 }
3205
3206 /*
3207  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3208  * but allow concurrent faults), and pte mapped but not yet locked.
3209  * We return with mmap_sem still held, but pte unmapped and unlocked.
3210  */
3211 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3212                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3213                 unsigned int flags)
3214 {
3215         struct page *page;
3216         spinlock_t *ptl;
3217         pte_t entry;
3218
3219         pte_unmap(page_table);
3220
3221         /* Check if we need to add a guard page to the stack */
3222         if (check_stack_guard_page(vma, address) < 0)
3223                 return VM_FAULT_SIGBUS;
3224
3225         /* Use the zero-page for reads */
3226         if (!(flags & FAULT_FLAG_WRITE)) {
3227                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3228                                                 vma->vm_page_prot));
3229                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3230                 if (!pte_none(*page_table))
3231                         goto unlock;
3232                 goto setpte;
3233         }
3234
3235         /* Allocate our own private page. */
3236         if (unlikely(anon_vma_prepare(vma)))
3237                 goto oom;
3238         page = alloc_zeroed_user_highpage_movable(vma, address);
3239         if (!page)
3240                 goto oom;
3241         /*
3242          * The memory barrier inside __SetPageUptodate makes sure that
3243          * preceeding stores to the page contents become visible before
3244          * the set_pte_at() write.
3245          */
3246         __SetPageUptodate(page);
3247
3248         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3249                 goto oom_free_page;
3250
3251         entry = mk_pte(page, vma->vm_page_prot);
3252         if (vma->vm_flags & VM_WRITE)
3253                 entry = pte_mkwrite(pte_mkdirty(entry));
3254
3255         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3256         if (!pte_none(*page_table))
3257                 goto release;
3258
3259         inc_mm_counter_fast(mm, MM_ANONPAGES);
3260         page_add_new_anon_rmap(page, vma, address);
3261 setpte:
3262         set_pte_at(mm, address, page_table, entry);
3263
3264         /* No need to invalidate - it was non-present before */
3265         update_mmu_cache(vma, address, page_table);
3266 unlock:
3267         pte_unmap_unlock(page_table, ptl);
3268         return 0;
3269 release:
3270         mem_cgroup_uncharge_page(page);
3271         page_cache_release(page);
3272         goto unlock;
3273 oom_free_page:
3274         page_cache_release(page);
3275 oom:
3276         return VM_FAULT_OOM;
3277 }
3278
3279 /*
3280  * __do_fault() tries to create a new page mapping. It aggressively
3281  * tries to share with existing pages, but makes a separate copy if
3282  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3283  * the next page fault.
3284  *
3285  * As this is called only for pages that do not currently exist, we
3286  * do not need to flush old virtual caches or the TLB.
3287  *
3288  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3289  * but allow concurrent faults), and pte neither mapped nor locked.
3290  * We return with mmap_sem still held, but pte unmapped and unlocked.
3291  */
3292 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3293                 unsigned long address, pmd_t *pmd,
3294                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3295 {
3296         pte_t *page_table;
3297         spinlock_t *ptl;
3298         struct page *page;
3299         struct page *cow_page;
3300         pte_t entry;
3301         int anon = 0;
3302         struct page *dirty_page = NULL;
3303         struct vm_fault vmf;
3304         int ret;
3305         int page_mkwrite = 0;
3306
3307         /*
3308          * If we do COW later, allocate page befor taking lock_page()
3309          * on the file cache page. This will reduce lock holding time.
3310          */
3311         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3312
3313                 if (unlikely(anon_vma_prepare(vma)))
3314                         return VM_FAULT_OOM;
3315
3316                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3317                 if (!cow_page)
3318                         return VM_FAULT_OOM;
3319
3320                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3321                         page_cache_release(cow_page);
3322                         return VM_FAULT_OOM;
3323                 }
3324         } else
3325                 cow_page = NULL;
3326
3327         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3328         vmf.pgoff = pgoff;
3329         vmf.flags = flags;
3330         vmf.page = NULL;
3331
3332         ret = vma->vm_ops->fault(vma, &vmf);
3333         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3334                             VM_FAULT_RETRY)))
3335                 goto uncharge_out;
3336
3337         if (unlikely(PageHWPoison(vmf.page))) {
3338                 if (ret & VM_FAULT_LOCKED)
3339                         unlock_page(vmf.page);
3340                 ret = VM_FAULT_HWPOISON;
3341                 goto uncharge_out;
3342         }
3343
3344         /*
3345          * For consistency in subsequent calls, make the faulted page always
3346          * locked.
3347          */
3348         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3349                 lock_page(vmf.page);
3350         else
3351                 VM_BUG_ON(!PageLocked(vmf.page));
3352
3353         /*
3354          * Should we do an early C-O-W break?
3355          */
3356         page = vmf.page;
3357         if (flags & FAULT_FLAG_WRITE) {
3358                 if (!(vma->vm_flags & VM_SHARED)) {
3359                         page = cow_page;
3360                         anon = 1;
3361                         copy_user_highpage(page, vmf.page, address, vma);
3362                         __SetPageUptodate(page);
3363                 } else {
3364                         /*
3365                          * If the page will be shareable, see if the backing
3366                          * address space wants to know that the page is about
3367                          * to become writable
3368                          */
3369                         if (vma->vm_ops->page_mkwrite) {
3370                                 int tmp;
3371
3372                                 unlock_page(page);
3373                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3374                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3375                                 if (unlikely(tmp &
3376                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3377                                         ret = tmp;
3378                                         goto unwritable_page;
3379                                 }
3380                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3381                                         lock_page(page);
3382                                         if (!page->mapping) {
3383                                                 ret = 0; /* retry the fault */
3384                                                 unlock_page(page);
3385                                                 goto unwritable_page;
3386                                         }
3387                                 } else
3388                                         VM_BUG_ON(!PageLocked(page));
3389                                 page_mkwrite = 1;
3390                         }
3391                 }
3392
3393         }
3394
3395         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3396
3397         /*
3398          * This silly early PAGE_DIRTY setting removes a race
3399          * due to the bad i386 page protection. But it's valid
3400          * for other architectures too.
3401          *
3402          * Note that if FAULT_FLAG_WRITE is set, we either now have
3403          * an exclusive copy of the page, or this is a shared mapping,
3404          * so we can make it writable and dirty to avoid having to
3405          * handle that later.
3406          */
3407         /* Only go through if we didn't race with anybody else... */
3408         if (likely(pte_same(*page_table, orig_pte))) {
3409                 flush_icache_page(vma, page);
3410                 entry = mk_pte(page, vma->vm_page_prot);
3411                 if (flags & FAULT_FLAG_WRITE)
3412                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3413                 else if (pte_file(orig_pte) && pte_file_soft_dirty(orig_pte))
3414                         pte_mksoft_dirty(entry);
3415                 if (anon) {
3416                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3417                         page_add_new_anon_rmap(page, vma, address);
3418                 } else {
3419                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3420                         page_add_file_rmap(page);
3421                         if (flags & FAULT_FLAG_WRITE) {
3422                                 dirty_page = page;
3423                                 get_page(dirty_page);
3424                         }
3425                 }
3426                 set_pte_at(mm, address, page_table, entry);
3427
3428                 /* no need to invalidate: a not-present page won't be cached */
3429                 update_mmu_cache(vma, address, page_table);
3430         } else {
3431                 if (cow_page)
3432                         mem_cgroup_uncharge_page(cow_page);
3433                 if (anon)
3434                         page_cache_release(page);
3435                 else
3436                         anon = 1; /* no anon but release faulted_page */
3437         }
3438
3439         pte_unmap_unlock(page_table, ptl);
3440
3441         if (dirty_page) {
3442                 struct address_space *mapping = page->mapping;
3443                 int dirtied = 0;
3444
3445                 if (set_page_dirty(dirty_page))
3446                         dirtied = 1;
3447                 unlock_page(dirty_page);
3448                 put_page(dirty_page);
3449                 if ((dirtied || page_mkwrite) && mapping) {
3450                         /*
3451                          * Some device drivers do not set page.mapping but still
3452                          * dirty their pages
3453                          */
3454                         balance_dirty_pages_ratelimited(mapping);
3455                 }
3456
3457                 /* file_update_time outside page_lock */
3458                 if (vma->vm_file && !page_mkwrite)
3459                         file_update_time(vma->vm_file);
3460         } else {
3461                 unlock_page(vmf.page);
3462                 if (anon)
3463                         page_cache_release(vmf.page);
3464         }
3465
3466         return ret;
3467
3468 unwritable_page:
3469         page_cache_release(page);
3470         return ret;
3471 uncharge_out:
3472         /* fs's fault handler get error */
3473         if (cow_page) {
3474                 mem_cgroup_uncharge_page(cow_page);
3475                 page_cache_release(cow_page);
3476         }
3477         return ret;
3478 }
3479
3480 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3481                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3482                 unsigned int flags, pte_t orig_pte)
3483 {
3484         pgoff_t pgoff = (((address & PAGE_MASK)
3485                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3486
3487         pte_unmap(page_table);
3488         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3489 }
3490
3491 /*
3492  * Fault of a previously existing named mapping. Repopulate the pte
3493  * from the encoded file_pte if possible. This enables swappable
3494  * nonlinear vmas.
3495  *
3496  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3497  * but allow concurrent faults), and pte mapped but not yet locked.
3498  * We return with mmap_sem still held, but pte unmapped and unlocked.
3499  */
3500 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3501                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3502                 unsigned int flags, pte_t orig_pte)
3503 {
3504         pgoff_t pgoff;
3505
3506         flags |= FAULT_FLAG_NONLINEAR;
3507
3508         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3509                 return 0;
3510
3511         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3512                 /*
3513                  * Page table corrupted: show pte and kill process.
3514                  */
3515                 print_bad_pte(vma, address, orig_pte, NULL);
3516                 return VM_FAULT_SIGBUS;
3517         }
3518
3519         pgoff = pte_to_pgoff(orig_pte);
3520         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3521 }
3522
3523 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3524                                 unsigned long addr, int page_nid)
3525 {
3526         get_page(page);
3527
3528         count_vm_numa_event(NUMA_HINT_FAULTS);
3529         if (page_nid == numa_node_id())
3530                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3531
3532         return mpol_misplaced(page, vma, addr);
3533 }
3534
3535 int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3536                    unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3537 {
3538         struct page *page = NULL;
3539         spinlock_t *ptl;
3540         int page_nid = -1;
3541         int target_nid;
3542         bool migrated = false;
3543
3544         /*
3545         * The "pte" at this point cannot be used safely without
3546         * validation through pte_unmap_same(). It's of NUMA type but
3547         * the pfn may be screwed if the read is non atomic.
3548         *
3549         * ptep_modify_prot_start is not called as this is clearing
3550         * the _PAGE_NUMA bit and it is not really expected that there
3551         * would be concurrent hardware modifications to the PTE.
3552         */
3553         ptl = pte_lockptr(mm, pmd);
3554         spin_lock(ptl);
3555         if (unlikely(!pte_same(*ptep, pte))) {
3556                 pte_unmap_unlock(ptep, ptl);
3557                 goto out;
3558         }
3559
3560         pte = pte_mknonnuma(pte);
3561         set_pte_at(mm, addr, ptep, pte);
3562         update_mmu_cache(vma, addr, ptep);
3563
3564         page = vm_normal_page(vma, addr, pte);
3565         if (!page) {
3566                 pte_unmap_unlock(ptep, ptl);
3567                 return 0;
3568         }
3569
3570         page_nid = page_to_nid(page);
3571         target_nid = numa_migrate_prep(page, vma, addr, page_nid);
3572         pte_unmap_unlock(ptep, ptl);
3573         if (target_nid == -1) {
3574                 put_page(page);
3575                 goto out;
3576         }
3577
3578         /* Migrate to the requested node */
3579         migrated = migrate_misplaced_page(page, target_nid);
3580         if (migrated)
3581                 page_nid = target_nid;
3582
3583 out:
3584         if (page_nid != -1)
3585                 task_numa_fault(page_nid, 1, migrated);
3586         return 0;
3587 }
3588
3589 /* NUMA hinting page fault entry point for regular pmds */
3590 #ifdef CONFIG_NUMA_BALANCING
3591 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3592                      unsigned long addr, pmd_t *pmdp)
3593 {
3594         pmd_t pmd;
3595         pte_t *pte, *orig_pte;
3596         unsigned long _addr = addr & PMD_MASK;
3597         unsigned long offset;
3598         spinlock_t *ptl;
3599         bool numa = false;
3600
3601         spin_lock(&mm->page_table_lock);
3602         pmd = *pmdp;
3603         if (pmd_numa(pmd)) {
3604                 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3605                 numa = true;
3606         }
3607         spin_unlock(&mm->page_table_lock);
3608
3609         if (!numa)
3610                 return 0;
3611
3612         /* we're in a page fault so some vma must be in the range */
3613         BUG_ON(!vma);
3614         BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3615         offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3616         VM_BUG_ON(offset >= PMD_SIZE);
3617         orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3618         pte += offset >> PAGE_SHIFT;
3619         for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3620                 pte_t pteval = *pte;
3621                 struct page *page;
3622                 int page_nid = -1;
3623                 int target_nid;
3624                 bool migrated = false;
3625
3626                 if (!pte_present(pteval))
3627                         continue;
3628                 if (!pte_numa(pteval))
3629                         continue;
3630                 if (addr >= vma->vm_end) {
3631                         vma = find_vma(mm, addr);
3632                         /* there's a pte present so there must be a vma */
3633                         BUG_ON(!vma);
3634                         BUG_ON(addr < vma->vm_start);
3635                 }
3636                 if (pte_numa(pteval)) {
3637                         pteval = pte_mknonnuma(pteval);
3638                         set_pte_at(mm, addr, pte, pteval);
3639                 }
3640                 page = vm_normal_page(vma, addr, pteval);
3641                 if (unlikely(!page))
3642                         continue;
3643                 /* only check non-shared pages */
3644                 if (unlikely(page_mapcount(page) != 1))
3645                         continue;
3646
3647                 page_nid = page_to_nid(page);
3648                 target_nid = numa_migrate_prep(page, vma, addr, page_nid);
3649                 pte_unmap_unlock(pte, ptl);
3650                 if (target_nid != -1) {
3651                         migrated = migrate_misplaced_page(page, target_nid);
3652                         if (migrated)
3653                                 page_nid = target_nid;
3654                 } else {
3655                         put_page(page);
3656                 }
3657
3658                 if (page_nid != -1)
3659                         task_numa_fault(page_nid, 1, migrated);
3660
3661                 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
3662         }
3663         pte_unmap_unlock(orig_pte, ptl);
3664
3665         return 0;
3666 }
3667 #else
3668 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3669                      unsigned long addr, pmd_t *pmdp)
3670 {
3671         BUG();
3672         return 0;
3673 }
3674 #endif /* CONFIG_NUMA_BALANCING */
3675
3676 /*
3677  * These routines also need to handle stuff like marking pages dirty
3678  * and/or accessed for architectures that don't do it in hardware (most
3679  * RISC architectures).  The early dirtying is also good on the i386.
3680  *
3681  * There is also a hook called "update_mmu_cache()" that architectures
3682  * with external mmu caches can use to update those (ie the Sparc or
3683  * PowerPC hashed page tables that act as extended TLBs).
3684  *
3685  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3686  * but allow concurrent faults), and pte mapped but not yet locked.
3687  * We return with mmap_sem still held, but pte unmapped and unlocked.
3688  */
3689 static int handle_pte_fault(struct mm_struct *mm,
3690                      struct vm_area_struct *vma, unsigned long address,
3691                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3692 {
3693         pte_t entry;
3694         spinlock_t *ptl;
3695
3696         entry = *pte;
3697         if (!pte_present(entry)) {
3698                 if (pte_none(entry)) {
3699                         if (vma->vm_ops) {
3700                                 if (likely(vma->vm_ops->fault))
3701                                         return do_linear_fault(mm, vma, address,
3702                                                 pte, pmd, flags, entry);
3703                         }
3704                         return do_anonymous_page(mm, vma, address,
3705                                                  pte, pmd, flags);
3706                 }
3707                 if (pte_file(entry))
3708                         return do_nonlinear_fault(mm, vma, address,
3709                                         pte, pmd, flags, entry);
3710                 return do_swap_page(mm, vma, address,
3711                                         pte, pmd, flags, entry);
3712         }
3713
3714         if (pte_numa(entry))
3715                 return do_numa_page(mm, vma, address, entry, pte, pmd);
3716
3717         ptl = pte_lockptr(mm, pmd);
3718         spin_lock(ptl);
3719         if (unlikely(!pte_same(*pte, entry)))
3720                 goto unlock;
3721         if (flags & FAULT_FLAG_WRITE) {
3722                 if (!pte_write(entry))
3723                         return do_wp_page(mm, vma, address,
3724                                         pte, pmd, ptl, entry);
3725                 entry = pte_mkdirty(entry);
3726         }
3727         entry = pte_mkyoung(entry);
3728         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3729                 update_mmu_cache(vma, address, pte);
3730         } else {
3731                 /*
3732                  * This is needed only for protection faults but the arch code
3733                  * is not yet telling us if this is a protection fault or not.
3734                  * This still avoids useless tlb flushes for .text page faults
3735                  * with threads.
3736                  */
3737                 if (flags & FAULT_FLAG_WRITE)
3738                         flush_tlb_fix_spurious_fault(vma, address);
3739         }
3740 unlock:
3741         pte_unmap_unlock(pte, ptl);
3742         return 0;
3743 }
3744
3745 /*
3746  * By the time we get here, we already hold the mm semaphore
3747  */
3748 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3749                              unsigned long address, unsigned int flags)
3750 {
3751         pgd_t *pgd;
3752         pud_t *pud;
3753         pmd_t *pmd;
3754         pte_t *pte;
3755
3756         if (unlikely(is_vm_hugetlb_page(vma)))
3757                 return hugetlb_fault(mm, vma, address, flags);
3758
3759 retry:
3760         pgd = pgd_offset(mm, address);
3761         pud = pud_alloc(mm, pgd, address);
3762         if (!pud)
3763                 return VM_FAULT_OOM;
3764         pmd = pmd_alloc(mm, pud, address);
3765         if (!pmd)
3766                 return VM_FAULT_OOM;
3767         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3768                 int ret = VM_FAULT_FALLBACK;
3769                 if (!vma->vm_ops)
3770                         ret = do_huge_pmd_anonymous_page(mm, vma, address,
3771                                         pmd, flags);
3772                 if (!(ret & VM_FAULT_FALLBACK))
3773                         return ret;
3774         } else {
3775                 pmd_t orig_pmd = *pmd;
3776                 int ret;
3777
3778                 barrier();
3779                 if (pmd_trans_huge(orig_pmd)) {
3780                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3781
3782                         /*
3783                          * If the pmd is splitting, return and retry the
3784                          * the fault.  Alternative: wait until the split
3785                          * is done, and goto retry.
3786                          */
3787                         if (pmd_trans_splitting(orig_pmd))
3788                                 return 0;
3789
3790                         if (pmd_numa(orig_pmd))
3791                                 return do_huge_pmd_numa_page(mm, vma, address,
3792                                                              orig_pmd, pmd);
3793
3794                         if (dirty && !pmd_write(orig_pmd)) {
3795                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3796                                                           orig_pmd);
3797                                 /*
3798                                  * If COW results in an oom, the huge pmd will
3799                                  * have been split, so retry the fault on the
3800                                  * pte for a smaller charge.
3801                                  */
3802                                 if (unlikely(ret & VM_FAULT_OOM))
3803                                         goto retry;
3804                                 return ret;
3805                         } else {
3806                                 huge_pmd_set_accessed(mm, vma, address, pmd,
3807                                                       orig_pmd, dirty);
3808                         }
3809
3810                         return 0;
3811                 }
3812         }
3813
3814         if (pmd_numa(*pmd))
3815                 return do_pmd_numa_page(mm, vma, address, pmd);
3816
3817         /*
3818          * Use __pte_alloc instead of pte_alloc_map, because we can't
3819          * run pte_offset_map on the pmd, if an huge pmd could
3820          * materialize from under us from a different thread.
3821          */
3822         if (unlikely(pmd_none(*pmd)) &&
3823             unlikely(__pte_alloc(mm, vma, pmd, address)))
3824                 return VM_FAULT_OOM;
3825         /* if an huge pmd materialized from under us just retry later */
3826         if (unlikely(pmd_trans_huge(*pmd)))
3827                 return 0;
3828         /*
3829          * A regular pmd is established and it can't morph into a huge pmd
3830          * from under us anymore at this point because we hold the mmap_sem
3831          * read mode and khugepaged takes it in write mode. So now it's
3832          * safe to run pte_offset_map().
3833          */
3834         pte = pte_offset_map(pmd, address);
3835
3836         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3837 }
3838
3839 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3840                     unsigned long address, unsigned int flags)
3841 {
3842         int ret;
3843
3844         __set_current_state(TASK_RUNNING);
3845
3846         count_vm_event(PGFAULT);
3847         mem_cgroup_count_vm_event(mm, PGFAULT);
3848
3849         /* do counter updates before entering really critical section. */
3850         check_sync_rss_stat(current);
3851
3852         /*
3853          * Enable the memcg OOM handling for faults triggered in user
3854          * space.  Kernel faults are handled more gracefully.
3855          */
3856         if (flags & FAULT_FLAG_USER)
3857                 mem_cgroup_oom_enable();
3858
3859         ret = __handle_mm_fault(mm, vma, address, flags);
3860
3861         if (flags & FAULT_FLAG_USER) {
3862                 mem_cgroup_oom_disable();
3863                 /*
3864                  * The task may have entered a memcg OOM situation but
3865                  * if the allocation error was handled gracefully (no
3866                  * VM_FAULT_OOM), there is no need to kill anything.
3867                  * Just clean up the OOM state peacefully.
3868                  */
3869                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3870                         mem_cgroup_oom_synchronize(false);
3871         }
3872
3873         return ret;
3874 }
3875
3876 #ifndef __PAGETABLE_PUD_FOLDED
3877 /*
3878  * Allocate page upper directory.
3879  * We've already handled the fast-path in-line.
3880  */
3881 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3882 {
3883         pud_t *new = pud_alloc_one(mm, address);
3884         if (!new)
3885                 return -ENOMEM;
3886
3887         smp_wmb(); /* See comment in __pte_alloc */
3888
3889         spin_lock(&mm->page_table_lock);
3890         if (pgd_present(*pgd))          /* Another has populated it */
3891                 pud_free(mm, new);
3892         else
3893                 pgd_populate(mm, pgd, new);
3894         spin_unlock(&mm->page_table_lock);
3895         return 0;
3896 }
3897 #endif /* __PAGETABLE_PUD_FOLDED */
3898
3899 #ifndef __PAGETABLE_PMD_FOLDED
3900 /*
3901  * Allocate page middle directory.
3902  * We've already handled the fast-path in-line.
3903  */
3904 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3905 {
3906         pmd_t *new = pmd_alloc_one(mm, address);
3907         if (!new)
3908                 return -ENOMEM;
3909
3910         smp_wmb(); /* See comment in __pte_alloc */
3911
3912         spin_lock(&mm->page_table_lock);
3913 #ifndef __ARCH_HAS_4LEVEL_HACK
3914         if (pud_present(*pud))          /* Another has populated it */
3915                 pmd_free(mm, new);
3916         else
3917                 pud_populate(mm, pud, new);
3918 #else
3919         if (pgd_present(*pud))          /* Another has populated it */
3920                 pmd_free(mm, new);
3921         else
3922                 pgd_populate(mm, pud, new);
3923 #endif /* __ARCH_HAS_4LEVEL_HACK */
3924         spin_unlock(&mm->page_table_lock);
3925         return 0;
3926 }
3927 #endif /* __PAGETABLE_PMD_FOLDED */
3928
3929 #if !defined(__HAVE_ARCH_GATE_AREA)
3930
3931 #if defined(AT_SYSINFO_EHDR)
3932 static struct vm_area_struct gate_vma;
3933
3934 static int __init gate_vma_init(void)
3935 {
3936         gate_vma.vm_mm = NULL;
3937         gate_vma.vm_start = FIXADDR_USER_START;
3938         gate_vma.vm_end = FIXADDR_USER_END;
3939         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3940         gate_vma.vm_page_prot = __P101;
3941
3942         return 0;
3943 }
3944 __initcall(gate_vma_init);
3945 #endif
3946
3947 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3948 {
3949 #ifdef AT_SYSINFO_EHDR
3950         return &gate_vma;
3951 #else
3952         return NULL;
3953 #endif
3954 }
3955
3956 int in_gate_area_no_mm(unsigned long addr)
3957 {
3958 #ifdef AT_SYSINFO_EHDR
3959         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3960                 return 1;
3961 #endif
3962         return 0;
3963 }
3964
3965 #endif  /* __HAVE_ARCH_GATE_AREA */
3966
3967 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3968                 pte_t **ptepp, spinlock_t **ptlp)
3969 {
3970         pgd_t *pgd;
3971         pud_t *pud;
3972         pmd_t *pmd;
3973         pte_t *ptep;
3974
3975         pgd = pgd_offset(mm, address);
3976         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3977                 goto out;
3978
3979         pud = pud_offset(pgd, address);
3980         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3981                 goto out;
3982
3983         pmd = pmd_offset(pud, address);
3984         VM_BUG_ON(pmd_trans_huge(*pmd));
3985         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3986                 goto out;
3987
3988         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3989         if (pmd_huge(*pmd))
3990                 goto out;
3991
3992         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3993         if (!ptep)
3994                 goto out;
3995         if (!pte_present(*ptep))
3996                 goto unlock;
3997         *ptepp = ptep;
3998         return 0;
3999 unlock:
4000         pte_unmap_unlock(ptep, *ptlp);
4001 out:
4002         return -EINVAL;
4003 }
4004
4005 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4006                              pte_t **ptepp, spinlock_t **ptlp)
4007 {
4008         int res;
4009
4010         /* (void) is needed to make gcc happy */
4011         (void) __cond_lock(*ptlp,
4012                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
4013         return res;
4014 }
4015
4016 /**
4017  * follow_pfn - look up PFN at a user virtual address
4018  * @vma: memory mapping
4019  * @address: user virtual address
4020  * @pfn: location to store found PFN
4021  *
4022  * Only IO mappings and raw PFN mappings are allowed.
4023  *
4024  * Returns zero and the pfn at @pfn on success, -ve otherwise.
4025  */
4026 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4027         unsigned long *pfn)
4028 {
4029         int ret = -EINVAL;
4030         spinlock_t *ptl;
4031         pte_t *ptep;
4032
4033         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4034                 return ret;
4035
4036         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4037         if (ret)
4038                 return ret;
4039         *pfn = pte_pfn(*ptep);
4040         pte_unmap_unlock(ptep, ptl);
4041         return 0;
4042 }
4043 EXPORT_SYMBOL(follow_pfn);
4044
4045 #ifdef CONFIG_HAVE_IOREMAP_PROT
4046 int follow_phys(struct vm_area_struct *vma,
4047                 unsigned long address, unsigned int flags,
4048                 unsigned long *prot, resource_size_t *phys)
4049 {
4050         int ret = -EINVAL;
4051         pte_t *ptep, pte;
4052         spinlock_t *ptl;
4053
4054         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4055                 goto out;
4056
4057         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4058                 goto out;
4059         pte = *ptep;
4060
4061         if ((flags & FOLL_WRITE) && !pte_write(pte))
4062                 goto unlock;
4063
4064         *prot = pgprot_val(pte_pgprot(pte));
4065         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4066
4067         ret = 0;
4068 unlock:
4069         pte_unmap_unlock(ptep, ptl);
4070 out:
4071         return ret;
4072 }
4073
4074 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4075                         void *buf, int len, int write)
4076 {
4077         resource_size_t phys_addr;
4078         unsigned long prot = 0;
4079         void __iomem *maddr;
4080         int offset = addr & (PAGE_SIZE-1);
4081
4082         if (follow_phys(vma, addr, write, &prot, &phys_addr))
4083                 return -EINVAL;
4084
4085         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4086         if (write)
4087                 memcpy_toio(maddr + offset, buf, len);
4088         else
4089                 memcpy_fromio(buf, maddr + offset, len);
4090         iounmap(maddr);
4091
4092         return len;
4093 }
4094 EXPORT_SYMBOL_GPL(generic_access_phys);
4095 #endif
4096
4097 /*
4098  * Access another process' address space as given in mm.  If non-NULL, use the
4099  * given task for page fault accounting.
4100  */
4101 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4102                 unsigned long addr, void *buf, int len, int write)
4103 {
4104         struct vm_area_struct *vma;
4105         void *old_buf = buf;
4106
4107         down_read(&mm->mmap_sem);
4108         /* ignore errors, just check how much was successfully transferred */
4109         while (len) {
4110                 int bytes, ret, offset;
4111                 void *maddr;
4112                 struct page *page = NULL;
4113
4114                 ret = get_user_pages(tsk, mm, addr, 1,
4115                                 write, 1, &page, &vma);
4116                 if (ret <= 0) {
4117                         /*
4118                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
4119                          * we can access using slightly different code.
4120                          */
4121 #ifdef CONFIG_HAVE_IOREMAP_PROT
4122                         vma = find_vma(mm, addr);
4123                         if (!vma || vma->vm_start > addr)
4124                                 break;
4125                         if (vma->vm_ops && vma->vm_ops->access)
4126                                 ret = vma->vm_ops->access(vma, addr, buf,
4127                                                           len, write);
4128                         if (ret <= 0)
4129 #endif
4130                                 break;
4131                         bytes = ret;
4132                 } else {
4133                         bytes = len;
4134                         offset = addr & (PAGE_SIZE-1);
4135                         if (bytes > PAGE_SIZE-offset)
4136                                 bytes = PAGE_SIZE-offset;
4137
4138                         maddr = kmap(page);
4139                         if (write) {
4140                                 copy_to_user_page(vma, page, addr,
4141                                                   maddr + offset, buf, bytes);
4142                                 set_page_dirty_lock(page);
4143                         } else {
4144                                 copy_from_user_page(vma, page, addr,
4145                                                     buf, maddr + offset, bytes);
4146                         }
4147                         kunmap(page);
4148                         page_cache_release(page);
4149                 }
4150                 len -= bytes;
4151                 buf += bytes;
4152                 addr += bytes;
4153         }
4154         up_read(&mm->mmap_sem);
4155
4156         return buf - old_buf;
4157 }
4158
4159 /**
4160  * access_remote_vm - access another process' address space
4161  * @mm:         the mm_struct of the target address space
4162  * @addr:       start address to access
4163  * @buf:        source or destination buffer
4164  * @len:        number of bytes to transfer
4165  * @write:      whether the access is a write
4166  *
4167  * The caller must hold a reference on @mm.
4168  */
4169 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4170                 void *buf, int len, int write)
4171 {
4172         return __access_remote_vm(NULL, mm, addr, buf, len, write);
4173 }
4174
4175 /*
4176  * Access another process' address space.
4177  * Source/target buffer must be kernel space,
4178  * Do not walk the page table directly, use get_user_pages
4179  */
4180 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4181                 void *buf, int len, int write)
4182 {
4183         struct mm_struct *mm;
4184         int ret;
4185
4186         mm = get_task_mm(tsk);
4187         if (!mm)
4188                 return 0;
4189
4190         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4191         mmput(mm);
4192
4193         return ret;
4194 }
4195
4196 /*
4197  * Print the name of a VMA.
4198  */
4199 void print_vma_addr(char *prefix, unsigned long ip)
4200 {
4201         struct mm_struct *mm = current->mm;
4202         struct vm_area_struct *vma;
4203
4204         /*
4205          * Do not print if we are in atomic
4206          * contexts (in exception stacks, etc.):
4207          */
4208         if (preempt_count())
4209                 return;
4210
4211         down_read(&mm->mmap_sem);
4212         vma = find_vma(mm, ip);
4213         if (vma && vma->vm_file) {
4214                 struct file *f = vma->vm_file;
4215                 char *buf = (char *)__get_free_page(GFP_KERNEL);
4216                 if (buf) {
4217                         char *p;
4218
4219                         p = d_path(&f->f_path, buf, PAGE_SIZE);
4220                         if (IS_ERR(p))
4221                                 p = "?";
4222                         printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4223                                         vma->vm_start,
4224                                         vma->vm_end - vma->vm_start);
4225                         free_page((unsigned long)buf);
4226                 }
4227         }
4228         up_read(&mm->mmap_sem);
4229 }
4230
4231 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4232 void might_fault(void)
4233 {
4234         /*
4235          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4236          * holding the mmap_sem, this is safe because kernel memory doesn't
4237          * get paged out, therefore we'll never actually fault, and the
4238          * below annotations will generate false positives.
4239          */
4240         if (segment_eq(get_fs(), KERNEL_DS))
4241                 return;
4242
4243         /*
4244          * it would be nicer only to annotate paths which are not under
4245          * pagefault_disable, however that requires a larger audit and
4246          * providing helpers like get_user_atomic.
4247          */
4248         if (in_atomic())
4249                 return;
4250
4251         __might_sleep(__FILE__, __LINE__, 0);
4252
4253         if (current->mm)
4254                 might_lock_read(&current->mm->mmap_sem);
4255 }
4256 EXPORT_SYMBOL(might_fault);
4257 #endif
4258
4259 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4260 static void clear_gigantic_page(struct page *page,
4261                                 unsigned long addr,
4262                                 unsigned int pages_per_huge_page)
4263 {
4264         int i;
4265         struct page *p = page;
4266
4267         might_sleep();
4268         for (i = 0; i < pages_per_huge_page;
4269              i++, p = mem_map_next(p, page, i)) {
4270                 cond_resched();
4271                 clear_user_highpage(p, addr + i * PAGE_SIZE);
4272         }
4273 }
4274 void clear_huge_page(struct page *page,
4275                      unsigned long addr, unsigned int pages_per_huge_page)
4276 {
4277         int i;
4278
4279         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4280                 clear_gigantic_page(page, addr, pages_per_huge_page);
4281                 return;
4282         }
4283
4284         might_sleep();
4285         for (i = 0; i < pages_per_huge_page; i++) {
4286                 cond_resched();
4287                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4288         }
4289 }
4290
4291 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4292                                     unsigned long addr,
4293                                     struct vm_area_struct *vma,
4294                                     unsigned int pages_per_huge_page)
4295 {
4296         int i;
4297         struct page *dst_base = dst;
4298         struct page *src_base = src;
4299
4300         for (i = 0; i < pages_per_huge_page; ) {
4301                 cond_resched();
4302                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4303
4304                 i++;
4305                 dst = mem_map_next(dst, dst_base, i);
4306                 src = mem_map_next(src, src_base, i);
4307         }
4308 }
4309
4310 void copy_user_huge_page(struct page *dst, struct page *src,
4311                          unsigned long addr, struct vm_area_struct *vma,
4312                          unsigned int pages_per_huge_page)
4313 {
4314         int i;
4315
4316         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4317                 copy_user_gigantic_page(dst, src, addr, vma,
4318                                         pages_per_huge_page);
4319                 return;
4320         }
4321
4322         might_sleep();
4323         for (i = 0; i < pages_per_huge_page; i++) {
4324                 cond_resched();
4325                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4326         }
4327 }
4328 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */