arch, mm: Remove tlb_fast_mode()
[linux-3.10.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 unsigned long num_physpages;
86 /*
87  * A number of key systems in x86 including ioremap() rely on the assumption
88  * that high_memory defines the upper bound on direct map memory, then end
89  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
90  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
91  * and ZONE_HIGHMEM.
92  */
93 void * high_memory;
94
95 EXPORT_SYMBOL(num_physpages);
96 EXPORT_SYMBOL(high_memory);
97
98 /*
99  * Randomize the address space (stacks, mmaps, brk, etc.).
100  *
101  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102  *   as ancient (libc5 based) binaries can segfault. )
103  */
104 int randomize_va_space __read_mostly =
105 #ifdef CONFIG_COMPAT_BRK
106                                         1;
107 #else
108                                         2;
109 #endif
110
111 static int __init disable_randmaps(char *s)
112 {
113         randomize_va_space = 0;
114         return 1;
115 }
116 __setup("norandmaps", disable_randmaps);
117
118 unsigned long zero_pfn __read_mostly;
119 unsigned long highest_memmap_pfn __read_mostly;
120
121 /*
122  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123  */
124 static int __init init_zero_pfn(void)
125 {
126         zero_pfn = page_to_pfn(ZERO_PAGE(0));
127         return 0;
128 }
129 core_initcall(init_zero_pfn);
130
131
132 #if defined(SPLIT_RSS_COUNTING)
133
134 void sync_mm_rss(struct mm_struct *mm)
135 {
136         int i;
137
138         for (i = 0; i < NR_MM_COUNTERS; i++) {
139                 if (current->rss_stat.count[i]) {
140                         add_mm_counter(mm, i, current->rss_stat.count[i]);
141                         current->rss_stat.count[i] = 0;
142                 }
143         }
144         current->rss_stat.events = 0;
145 }
146
147 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148 {
149         struct task_struct *task = current;
150
151         if (likely(task->mm == mm))
152                 task->rss_stat.count[member] += val;
153         else
154                 add_mm_counter(mm, member, val);
155 }
156 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159 /* sync counter once per 64 page faults */
160 #define TASK_RSS_EVENTS_THRESH  (64)
161 static void check_sync_rss_stat(struct task_struct *task)
162 {
163         if (unlikely(task != current))
164                 return;
165         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166                 sync_mm_rss(task->mm);
167 }
168 #else /* SPLIT_RSS_COUNTING */
169
170 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173 static void check_sync_rss_stat(struct task_struct *task)
174 {
175 }
176
177 #endif /* SPLIT_RSS_COUNTING */
178
179 #ifdef HAVE_GENERIC_MMU_GATHER
180
181 static int tlb_next_batch(struct mmu_gather *tlb)
182 {
183         struct mmu_gather_batch *batch;
184
185         batch = tlb->active;
186         if (batch->next) {
187                 tlb->active = batch->next;
188                 return 1;
189         }
190
191         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192                 return 0;
193
194         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195         if (!batch)
196                 return 0;
197
198         tlb->batch_count++;
199         batch->next = NULL;
200         batch->nr   = 0;
201         batch->max  = MAX_GATHER_BATCH;
202
203         tlb->active->next = batch;
204         tlb->active = batch;
205
206         return 1;
207 }
208
209 /* tlb_gather_mmu
210  *      Called to initialize an (on-stack) mmu_gather structure for page-table
211  *      tear-down from @mm. The @fullmm argument is used when @mm is without
212  *      users and we're going to destroy the full address space (exit/execve).
213  */
214 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
215 {
216         tlb->mm = mm;
217
218         tlb->fullmm     = fullmm;
219         tlb->need_flush_all = 0;
220         tlb->start      = -1UL;
221         tlb->end        = 0;
222         tlb->need_flush = 0;
223         tlb->local.next = NULL;
224         tlb->local.nr   = 0;
225         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
226         tlb->active     = &tlb->local;
227         tlb->batch_count = 0;
228
229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
230         tlb->batch = NULL;
231 #endif
232 }
233
234 void tlb_flush_mmu(struct mmu_gather *tlb)
235 {
236         struct mmu_gather_batch *batch;
237
238         if (!tlb->need_flush)
239                 return;
240         tlb->need_flush = 0;
241         tlb_flush(tlb);
242 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
243         tlb_table_flush(tlb);
244 #endif
245
246         for (batch = &tlb->local; batch; batch = batch->next) {
247                 free_pages_and_swap_cache(batch->pages, batch->nr);
248                 batch->nr = 0;
249         }
250         tlb->active = &tlb->local;
251 }
252
253 /* tlb_finish_mmu
254  *      Called at the end of the shootdown operation to free up any resources
255  *      that were required.
256  */
257 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
258 {
259         struct mmu_gather_batch *batch, *next;
260
261         tlb->start = start;
262         tlb->end   = end;
263         tlb_flush_mmu(tlb);
264
265         /* keep the page table cache within bounds */
266         check_pgt_cache();
267
268         for (batch = tlb->local.next; batch; batch = next) {
269                 next = batch->next;
270                 free_pages((unsigned long)batch, 0);
271         }
272         tlb->local.next = NULL;
273 }
274
275 /* __tlb_remove_page
276  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
277  *      handling the additional races in SMP caused by other CPUs caching valid
278  *      mappings in their TLBs. Returns the number of free page slots left.
279  *      When out of page slots we must call tlb_flush_mmu().
280  */
281 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
282 {
283         struct mmu_gather_batch *batch;
284
285         VM_BUG_ON(!tlb->need_flush);
286
287         batch = tlb->active;
288         batch->pages[batch->nr++] = page;
289         if (batch->nr == batch->max) {
290                 if (!tlb_next_batch(tlb))
291                         return 0;
292                 batch = tlb->active;
293         }
294         VM_BUG_ON(batch->nr > batch->max);
295
296         return batch->max - batch->nr;
297 }
298
299 #endif /* HAVE_GENERIC_MMU_GATHER */
300
301 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
302
303 /*
304  * See the comment near struct mmu_table_batch.
305  */
306
307 static void tlb_remove_table_smp_sync(void *arg)
308 {
309         /* Simply deliver the interrupt */
310 }
311
312 static void tlb_remove_table_one(void *table)
313 {
314         /*
315          * This isn't an RCU grace period and hence the page-tables cannot be
316          * assumed to be actually RCU-freed.
317          *
318          * It is however sufficient for software page-table walkers that rely on
319          * IRQ disabling. See the comment near struct mmu_table_batch.
320          */
321         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
322         __tlb_remove_table(table);
323 }
324
325 static void tlb_remove_table_rcu(struct rcu_head *head)
326 {
327         struct mmu_table_batch *batch;
328         int i;
329
330         batch = container_of(head, struct mmu_table_batch, rcu);
331
332         for (i = 0; i < batch->nr; i++)
333                 __tlb_remove_table(batch->tables[i]);
334
335         free_page((unsigned long)batch);
336 }
337
338 void tlb_table_flush(struct mmu_gather *tlb)
339 {
340         struct mmu_table_batch **batch = &tlb->batch;
341
342         if (*batch) {
343                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
344                 *batch = NULL;
345         }
346 }
347
348 void tlb_remove_table(struct mmu_gather *tlb, void *table)
349 {
350         struct mmu_table_batch **batch = &tlb->batch;
351
352         tlb->need_flush = 1;
353
354         /*
355          * When there's less then two users of this mm there cannot be a
356          * concurrent page-table walk.
357          */
358         if (atomic_read(&tlb->mm->mm_users) < 2) {
359                 __tlb_remove_table(table);
360                 return;
361         }
362
363         if (*batch == NULL) {
364                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
365                 if (*batch == NULL) {
366                         tlb_remove_table_one(table);
367                         return;
368                 }
369                 (*batch)->nr = 0;
370         }
371         (*batch)->tables[(*batch)->nr++] = table;
372         if ((*batch)->nr == MAX_TABLE_BATCH)
373                 tlb_table_flush(tlb);
374 }
375
376 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
377
378 /*
379  * If a p?d_bad entry is found while walking page tables, report
380  * the error, before resetting entry to p?d_none.  Usually (but
381  * very seldom) called out from the p?d_none_or_clear_bad macros.
382  */
383
384 void pgd_clear_bad(pgd_t *pgd)
385 {
386         pgd_ERROR(*pgd);
387         pgd_clear(pgd);
388 }
389
390 void pud_clear_bad(pud_t *pud)
391 {
392         pud_ERROR(*pud);
393         pud_clear(pud);
394 }
395
396 void pmd_clear_bad(pmd_t *pmd)
397 {
398         pmd_ERROR(*pmd);
399         pmd_clear(pmd);
400 }
401
402 /*
403  * Note: this doesn't free the actual pages themselves. That
404  * has been handled earlier when unmapping all the memory regions.
405  */
406 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
407                            unsigned long addr)
408 {
409         pgtable_t token = pmd_pgtable(*pmd);
410         pmd_clear(pmd);
411         pte_free_tlb(tlb, token, addr);
412         tlb->mm->nr_ptes--;
413 }
414
415 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
416                                 unsigned long addr, unsigned long end,
417                                 unsigned long floor, unsigned long ceiling)
418 {
419         pmd_t *pmd;
420         unsigned long next;
421         unsigned long start;
422
423         start = addr;
424         pmd = pmd_offset(pud, addr);
425         do {
426                 next = pmd_addr_end(addr, end);
427                 if (pmd_none_or_clear_bad(pmd))
428                         continue;
429                 free_pte_range(tlb, pmd, addr);
430         } while (pmd++, addr = next, addr != end);
431
432         start &= PUD_MASK;
433         if (start < floor)
434                 return;
435         if (ceiling) {
436                 ceiling &= PUD_MASK;
437                 if (!ceiling)
438                         return;
439         }
440         if (end - 1 > ceiling - 1)
441                 return;
442
443         pmd = pmd_offset(pud, start);
444         pud_clear(pud);
445         pmd_free_tlb(tlb, pmd, start);
446 }
447
448 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
449                                 unsigned long addr, unsigned long end,
450                                 unsigned long floor, unsigned long ceiling)
451 {
452         pud_t *pud;
453         unsigned long next;
454         unsigned long start;
455
456         start = addr;
457         pud = pud_offset(pgd, addr);
458         do {
459                 next = pud_addr_end(addr, end);
460                 if (pud_none_or_clear_bad(pud))
461                         continue;
462                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
463         } while (pud++, addr = next, addr != end);
464
465         start &= PGDIR_MASK;
466         if (start < floor)
467                 return;
468         if (ceiling) {
469                 ceiling &= PGDIR_MASK;
470                 if (!ceiling)
471                         return;
472         }
473         if (end - 1 > ceiling - 1)
474                 return;
475
476         pud = pud_offset(pgd, start);
477         pgd_clear(pgd);
478         pud_free_tlb(tlb, pud, start);
479 }
480
481 /*
482  * This function frees user-level page tables of a process.
483  *
484  * Must be called with pagetable lock held.
485  */
486 void free_pgd_range(struct mmu_gather *tlb,
487                         unsigned long addr, unsigned long end,
488                         unsigned long floor, unsigned long ceiling)
489 {
490         pgd_t *pgd;
491         unsigned long next;
492
493         /*
494          * The next few lines have given us lots of grief...
495          *
496          * Why are we testing PMD* at this top level?  Because often
497          * there will be no work to do at all, and we'd prefer not to
498          * go all the way down to the bottom just to discover that.
499          *
500          * Why all these "- 1"s?  Because 0 represents both the bottom
501          * of the address space and the top of it (using -1 for the
502          * top wouldn't help much: the masks would do the wrong thing).
503          * The rule is that addr 0 and floor 0 refer to the bottom of
504          * the address space, but end 0 and ceiling 0 refer to the top
505          * Comparisons need to use "end - 1" and "ceiling - 1" (though
506          * that end 0 case should be mythical).
507          *
508          * Wherever addr is brought up or ceiling brought down, we must
509          * be careful to reject "the opposite 0" before it confuses the
510          * subsequent tests.  But what about where end is brought down
511          * by PMD_SIZE below? no, end can't go down to 0 there.
512          *
513          * Whereas we round start (addr) and ceiling down, by different
514          * masks at different levels, in order to test whether a table
515          * now has no other vmas using it, so can be freed, we don't
516          * bother to round floor or end up - the tests don't need that.
517          */
518
519         addr &= PMD_MASK;
520         if (addr < floor) {
521                 addr += PMD_SIZE;
522                 if (!addr)
523                         return;
524         }
525         if (ceiling) {
526                 ceiling &= PMD_MASK;
527                 if (!ceiling)
528                         return;
529         }
530         if (end - 1 > ceiling - 1)
531                 end -= PMD_SIZE;
532         if (addr > end - 1)
533                 return;
534
535         pgd = pgd_offset(tlb->mm, addr);
536         do {
537                 next = pgd_addr_end(addr, end);
538                 if (pgd_none_or_clear_bad(pgd))
539                         continue;
540                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
541         } while (pgd++, addr = next, addr != end);
542 }
543
544 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
545                 unsigned long floor, unsigned long ceiling)
546 {
547         while (vma) {
548                 struct vm_area_struct *next = vma->vm_next;
549                 unsigned long addr = vma->vm_start;
550
551                 /*
552                  * Hide vma from rmap and truncate_pagecache before freeing
553                  * pgtables
554                  */
555                 unlink_anon_vmas(vma);
556                 unlink_file_vma(vma);
557
558                 if (is_vm_hugetlb_page(vma)) {
559                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
560                                 floor, next? next->vm_start: ceiling);
561                 } else {
562                         /*
563                          * Optimization: gather nearby vmas into one call down
564                          */
565                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
566                                && !is_vm_hugetlb_page(next)) {
567                                 vma = next;
568                                 next = vma->vm_next;
569                                 unlink_anon_vmas(vma);
570                                 unlink_file_vma(vma);
571                         }
572                         free_pgd_range(tlb, addr, vma->vm_end,
573                                 floor, next? next->vm_start: ceiling);
574                 }
575                 vma = next;
576         }
577 }
578
579 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
580                 pmd_t *pmd, unsigned long address)
581 {
582         pgtable_t new = pte_alloc_one(mm, address);
583         int wait_split_huge_page;
584         if (!new)
585                 return -ENOMEM;
586
587         /*
588          * Ensure all pte setup (eg. pte page lock and page clearing) are
589          * visible before the pte is made visible to other CPUs by being
590          * put into page tables.
591          *
592          * The other side of the story is the pointer chasing in the page
593          * table walking code (when walking the page table without locking;
594          * ie. most of the time). Fortunately, these data accesses consist
595          * of a chain of data-dependent loads, meaning most CPUs (alpha
596          * being the notable exception) will already guarantee loads are
597          * seen in-order. See the alpha page table accessors for the
598          * smp_read_barrier_depends() barriers in page table walking code.
599          */
600         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
601
602         spin_lock(&mm->page_table_lock);
603         wait_split_huge_page = 0;
604         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
605                 mm->nr_ptes++;
606                 pmd_populate(mm, pmd, new);
607                 new = NULL;
608         } else if (unlikely(pmd_trans_splitting(*pmd)))
609                 wait_split_huge_page = 1;
610         spin_unlock(&mm->page_table_lock);
611         if (new)
612                 pte_free(mm, new);
613         if (wait_split_huge_page)
614                 wait_split_huge_page(vma->anon_vma, pmd);
615         return 0;
616 }
617
618 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
619 {
620         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
621         if (!new)
622                 return -ENOMEM;
623
624         smp_wmb(); /* See comment in __pte_alloc */
625
626         spin_lock(&init_mm.page_table_lock);
627         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
628                 pmd_populate_kernel(&init_mm, pmd, new);
629                 new = NULL;
630         } else
631                 VM_BUG_ON(pmd_trans_splitting(*pmd));
632         spin_unlock(&init_mm.page_table_lock);
633         if (new)
634                 pte_free_kernel(&init_mm, new);
635         return 0;
636 }
637
638 static inline void init_rss_vec(int *rss)
639 {
640         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
641 }
642
643 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
644 {
645         int i;
646
647         if (current->mm == mm)
648                 sync_mm_rss(mm);
649         for (i = 0; i < NR_MM_COUNTERS; i++)
650                 if (rss[i])
651                         add_mm_counter(mm, i, rss[i]);
652 }
653
654 /*
655  * This function is called to print an error when a bad pte
656  * is found. For example, we might have a PFN-mapped pte in
657  * a region that doesn't allow it.
658  *
659  * The calling function must still handle the error.
660  */
661 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
662                           pte_t pte, struct page *page)
663 {
664         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
665         pud_t *pud = pud_offset(pgd, addr);
666         pmd_t *pmd = pmd_offset(pud, addr);
667         struct address_space *mapping;
668         pgoff_t index;
669         static unsigned long resume;
670         static unsigned long nr_shown;
671         static unsigned long nr_unshown;
672
673         /*
674          * Allow a burst of 60 reports, then keep quiet for that minute;
675          * or allow a steady drip of one report per second.
676          */
677         if (nr_shown == 60) {
678                 if (time_before(jiffies, resume)) {
679                         nr_unshown++;
680                         return;
681                 }
682                 if (nr_unshown) {
683                         printk(KERN_ALERT
684                                 "BUG: Bad page map: %lu messages suppressed\n",
685                                 nr_unshown);
686                         nr_unshown = 0;
687                 }
688                 nr_shown = 0;
689         }
690         if (nr_shown++ == 0)
691                 resume = jiffies + 60 * HZ;
692
693         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
694         index = linear_page_index(vma, addr);
695
696         printk(KERN_ALERT
697                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
698                 current->comm,
699                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
700         if (page)
701                 dump_page(page);
702         printk(KERN_ALERT
703                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
704                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
705         /*
706          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
707          */
708         if (vma->vm_ops)
709                 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
710                        vma->vm_ops->fault);
711         if (vma->vm_file && vma->vm_file->f_op)
712                 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
713                        vma->vm_file->f_op->mmap);
714         dump_stack();
715         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
716 }
717
718 static inline bool is_cow_mapping(vm_flags_t flags)
719 {
720         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
721 }
722
723 /*
724  * vm_normal_page -- This function gets the "struct page" associated with a pte.
725  *
726  * "Special" mappings do not wish to be associated with a "struct page" (either
727  * it doesn't exist, or it exists but they don't want to touch it). In this
728  * case, NULL is returned here. "Normal" mappings do have a struct page.
729  *
730  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
731  * pte bit, in which case this function is trivial. Secondly, an architecture
732  * may not have a spare pte bit, which requires a more complicated scheme,
733  * described below.
734  *
735  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
736  * special mapping (even if there are underlying and valid "struct pages").
737  * COWed pages of a VM_PFNMAP are always normal.
738  *
739  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
740  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
741  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
742  * mapping will always honor the rule
743  *
744  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
745  *
746  * And for normal mappings this is false.
747  *
748  * This restricts such mappings to be a linear translation from virtual address
749  * to pfn. To get around this restriction, we allow arbitrary mappings so long
750  * as the vma is not a COW mapping; in that case, we know that all ptes are
751  * special (because none can have been COWed).
752  *
753  *
754  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
755  *
756  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
757  * page" backing, however the difference is that _all_ pages with a struct
758  * page (that is, those where pfn_valid is true) are refcounted and considered
759  * normal pages by the VM. The disadvantage is that pages are refcounted
760  * (which can be slower and simply not an option for some PFNMAP users). The
761  * advantage is that we don't have to follow the strict linearity rule of
762  * PFNMAP mappings in order to support COWable mappings.
763  *
764  */
765 #ifdef __HAVE_ARCH_PTE_SPECIAL
766 # define HAVE_PTE_SPECIAL 1
767 #else
768 # define HAVE_PTE_SPECIAL 0
769 #endif
770 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
771                                 pte_t pte)
772 {
773         unsigned long pfn = pte_pfn(pte);
774
775         if (HAVE_PTE_SPECIAL) {
776                 if (likely(!pte_special(pte)))
777                         goto check_pfn;
778                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
779                         return NULL;
780                 if (!is_zero_pfn(pfn))
781                         print_bad_pte(vma, addr, pte, NULL);
782                 return NULL;
783         }
784
785         /* !HAVE_PTE_SPECIAL case follows: */
786
787         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
788                 if (vma->vm_flags & VM_MIXEDMAP) {
789                         if (!pfn_valid(pfn))
790                                 return NULL;
791                         goto out;
792                 } else {
793                         unsigned long off;
794                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
795                         if (pfn == vma->vm_pgoff + off)
796                                 return NULL;
797                         if (!is_cow_mapping(vma->vm_flags))
798                                 return NULL;
799                 }
800         }
801
802         if (is_zero_pfn(pfn))
803                 return NULL;
804 check_pfn:
805         if (unlikely(pfn > highest_memmap_pfn)) {
806                 print_bad_pte(vma, addr, pte, NULL);
807                 return NULL;
808         }
809
810         /*
811          * NOTE! We still have PageReserved() pages in the page tables.
812          * eg. VDSO mappings can cause them to exist.
813          */
814 out:
815         return pfn_to_page(pfn);
816 }
817
818 /*
819  * copy one vm_area from one task to the other. Assumes the page tables
820  * already present in the new task to be cleared in the whole range
821  * covered by this vma.
822  */
823
824 static inline unsigned long
825 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
826                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
827                 unsigned long addr, int *rss)
828 {
829         unsigned long vm_flags = vma->vm_flags;
830         pte_t pte = *src_pte;
831         struct page *page;
832
833         /* pte contains position in swap or file, so copy. */
834         if (unlikely(!pte_present(pte))) {
835                 if (!pte_file(pte)) {
836                         swp_entry_t entry = pte_to_swp_entry(pte);
837
838                         if (swap_duplicate(entry) < 0)
839                                 return entry.val;
840
841                         /* make sure dst_mm is on swapoff's mmlist. */
842                         if (unlikely(list_empty(&dst_mm->mmlist))) {
843                                 spin_lock(&mmlist_lock);
844                                 if (list_empty(&dst_mm->mmlist))
845                                         list_add(&dst_mm->mmlist,
846                                                  &src_mm->mmlist);
847                                 spin_unlock(&mmlist_lock);
848                         }
849                         if (likely(!non_swap_entry(entry)))
850                                 rss[MM_SWAPENTS]++;
851                         else if (is_migration_entry(entry)) {
852                                 page = migration_entry_to_page(entry);
853
854                                 if (PageAnon(page))
855                                         rss[MM_ANONPAGES]++;
856                                 else
857                                         rss[MM_FILEPAGES]++;
858
859                                 if (is_write_migration_entry(entry) &&
860                                     is_cow_mapping(vm_flags)) {
861                                         /*
862                                          * COW mappings require pages in both
863                                          * parent and child to be set to read.
864                                          */
865                                         make_migration_entry_read(&entry);
866                                         pte = swp_entry_to_pte(entry);
867                                         set_pte_at(src_mm, addr, src_pte, pte);
868                                 }
869                         }
870                 }
871                 goto out_set_pte;
872         }
873
874         /*
875          * If it's a COW mapping, write protect it both
876          * in the parent and the child
877          */
878         if (is_cow_mapping(vm_flags)) {
879                 ptep_set_wrprotect(src_mm, addr, src_pte);
880                 pte = pte_wrprotect(pte);
881         }
882
883         /*
884          * If it's a shared mapping, mark it clean in
885          * the child
886          */
887         if (vm_flags & VM_SHARED)
888                 pte = pte_mkclean(pte);
889         pte = pte_mkold(pte);
890
891         page = vm_normal_page(vma, addr, pte);
892         if (page) {
893                 get_page(page);
894                 page_dup_rmap(page);
895                 if (PageAnon(page))
896                         rss[MM_ANONPAGES]++;
897                 else
898                         rss[MM_FILEPAGES]++;
899         }
900
901 out_set_pte:
902         set_pte_at(dst_mm, addr, dst_pte, pte);
903         return 0;
904 }
905
906 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
907                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
908                    unsigned long addr, unsigned long end)
909 {
910         pte_t *orig_src_pte, *orig_dst_pte;
911         pte_t *src_pte, *dst_pte;
912         spinlock_t *src_ptl, *dst_ptl;
913         int progress = 0;
914         int rss[NR_MM_COUNTERS];
915         swp_entry_t entry = (swp_entry_t){0};
916
917 again:
918         init_rss_vec(rss);
919
920         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
921         if (!dst_pte)
922                 return -ENOMEM;
923         src_pte = pte_offset_map(src_pmd, addr);
924         src_ptl = pte_lockptr(src_mm, src_pmd);
925         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
926         orig_src_pte = src_pte;
927         orig_dst_pte = dst_pte;
928         arch_enter_lazy_mmu_mode();
929
930         do {
931                 /*
932                  * We are holding two locks at this point - either of them
933                  * could generate latencies in another task on another CPU.
934                  */
935                 if (progress >= 32) {
936                         progress = 0;
937                         if (need_resched() ||
938                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
939                                 break;
940                 }
941                 if (pte_none(*src_pte)) {
942                         progress++;
943                         continue;
944                 }
945                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
946                                                         vma, addr, rss);
947                 if (entry.val)
948                         break;
949                 progress += 8;
950         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
951
952         arch_leave_lazy_mmu_mode();
953         spin_unlock(src_ptl);
954         pte_unmap(orig_src_pte);
955         add_mm_rss_vec(dst_mm, rss);
956         pte_unmap_unlock(orig_dst_pte, dst_ptl);
957         cond_resched();
958
959         if (entry.val) {
960                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
961                         return -ENOMEM;
962                 progress = 0;
963         }
964         if (addr != end)
965                 goto again;
966         return 0;
967 }
968
969 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
970                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
971                 unsigned long addr, unsigned long end)
972 {
973         pmd_t *src_pmd, *dst_pmd;
974         unsigned long next;
975
976         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
977         if (!dst_pmd)
978                 return -ENOMEM;
979         src_pmd = pmd_offset(src_pud, addr);
980         do {
981                 next = pmd_addr_end(addr, end);
982                 if (pmd_trans_huge(*src_pmd)) {
983                         int err;
984                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
985                         err = copy_huge_pmd(dst_mm, src_mm,
986                                             dst_pmd, src_pmd, addr, vma);
987                         if (err == -ENOMEM)
988                                 return -ENOMEM;
989                         if (!err)
990                                 continue;
991                         /* fall through */
992                 }
993                 if (pmd_none_or_clear_bad(src_pmd))
994                         continue;
995                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
996                                                 vma, addr, next))
997                         return -ENOMEM;
998         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
999         return 0;
1000 }
1001
1002 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1003                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1004                 unsigned long addr, unsigned long end)
1005 {
1006         pud_t *src_pud, *dst_pud;
1007         unsigned long next;
1008
1009         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1010         if (!dst_pud)
1011                 return -ENOMEM;
1012         src_pud = pud_offset(src_pgd, addr);
1013         do {
1014                 next = pud_addr_end(addr, end);
1015                 if (pud_none_or_clear_bad(src_pud))
1016                         continue;
1017                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1018                                                 vma, addr, next))
1019                         return -ENOMEM;
1020         } while (dst_pud++, src_pud++, addr = next, addr != end);
1021         return 0;
1022 }
1023
1024 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1025                 struct vm_area_struct *vma)
1026 {
1027         pgd_t *src_pgd, *dst_pgd;
1028         unsigned long next;
1029         unsigned long addr = vma->vm_start;
1030         unsigned long end = vma->vm_end;
1031         unsigned long mmun_start;       /* For mmu_notifiers */
1032         unsigned long mmun_end;         /* For mmu_notifiers */
1033         bool is_cow;
1034         int ret;
1035
1036         /*
1037          * Don't copy ptes where a page fault will fill them correctly.
1038          * Fork becomes much lighter when there are big shared or private
1039          * readonly mappings. The tradeoff is that copy_page_range is more
1040          * efficient than faulting.
1041          */
1042         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1043                                VM_PFNMAP | VM_MIXEDMAP))) {
1044                 if (!vma->anon_vma)
1045                         return 0;
1046         }
1047
1048         if (is_vm_hugetlb_page(vma))
1049                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1050
1051         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1052                 /*
1053                  * We do not free on error cases below as remove_vma
1054                  * gets called on error from higher level routine
1055                  */
1056                 ret = track_pfn_copy(vma);
1057                 if (ret)
1058                         return ret;
1059         }
1060
1061         /*
1062          * We need to invalidate the secondary MMU mappings only when
1063          * there could be a permission downgrade on the ptes of the
1064          * parent mm. And a permission downgrade will only happen if
1065          * is_cow_mapping() returns true.
1066          */
1067         is_cow = is_cow_mapping(vma->vm_flags);
1068         mmun_start = addr;
1069         mmun_end   = end;
1070         if (is_cow)
1071                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1072                                                     mmun_end);
1073
1074         ret = 0;
1075         dst_pgd = pgd_offset(dst_mm, addr);
1076         src_pgd = pgd_offset(src_mm, addr);
1077         do {
1078                 next = pgd_addr_end(addr, end);
1079                 if (pgd_none_or_clear_bad(src_pgd))
1080                         continue;
1081                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1082                                             vma, addr, next))) {
1083                         ret = -ENOMEM;
1084                         break;
1085                 }
1086         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1087
1088         if (is_cow)
1089                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1090         return ret;
1091 }
1092
1093 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1094                                 struct vm_area_struct *vma, pmd_t *pmd,
1095                                 unsigned long addr, unsigned long end,
1096                                 struct zap_details *details)
1097 {
1098         struct mm_struct *mm = tlb->mm;
1099         int force_flush = 0;
1100         int rss[NR_MM_COUNTERS];
1101         spinlock_t *ptl;
1102         pte_t *start_pte;
1103         pte_t *pte;
1104
1105 again:
1106         init_rss_vec(rss);
1107         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1108         pte = start_pte;
1109         arch_enter_lazy_mmu_mode();
1110         do {
1111                 pte_t ptent = *pte;
1112                 if (pte_none(ptent)) {
1113                         continue;
1114                 }
1115
1116                 if (pte_present(ptent)) {
1117                         struct page *page;
1118
1119                         page = vm_normal_page(vma, addr, ptent);
1120                         if (unlikely(details) && page) {
1121                                 /*
1122                                  * unmap_shared_mapping_pages() wants to
1123                                  * invalidate cache without truncating:
1124                                  * unmap shared but keep private pages.
1125                                  */
1126                                 if (details->check_mapping &&
1127                                     details->check_mapping != page->mapping)
1128                                         continue;
1129                                 /*
1130                                  * Each page->index must be checked when
1131                                  * invalidating or truncating nonlinear.
1132                                  */
1133                                 if (details->nonlinear_vma &&
1134                                     (page->index < details->first_index ||
1135                                      page->index > details->last_index))
1136                                         continue;
1137                         }
1138                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1139                                                         tlb->fullmm);
1140                         tlb_remove_tlb_entry(tlb, pte, addr);
1141                         if (unlikely(!page))
1142                                 continue;
1143                         if (unlikely(details) && details->nonlinear_vma
1144                             && linear_page_index(details->nonlinear_vma,
1145                                                 addr) != page->index)
1146                                 set_pte_at(mm, addr, pte,
1147                                            pgoff_to_pte(page->index));
1148                         if (PageAnon(page))
1149                                 rss[MM_ANONPAGES]--;
1150                         else {
1151                                 if (pte_dirty(ptent))
1152                                         set_page_dirty(page);
1153                                 if (pte_young(ptent) &&
1154                                     likely(!VM_SequentialReadHint(vma)))
1155                                         mark_page_accessed(page);
1156                                 rss[MM_FILEPAGES]--;
1157                         }
1158                         page_remove_rmap(page);
1159                         if (unlikely(page_mapcount(page) < 0))
1160                                 print_bad_pte(vma, addr, ptent, page);
1161                         force_flush = !__tlb_remove_page(tlb, page);
1162                         if (force_flush)
1163                                 break;
1164                         continue;
1165                 }
1166                 /*
1167                  * If details->check_mapping, we leave swap entries;
1168                  * if details->nonlinear_vma, we leave file entries.
1169                  */
1170                 if (unlikely(details))
1171                         continue;
1172                 if (pte_file(ptent)) {
1173                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1174                                 print_bad_pte(vma, addr, ptent, NULL);
1175                 } else {
1176                         swp_entry_t entry = pte_to_swp_entry(ptent);
1177
1178                         if (!non_swap_entry(entry))
1179                                 rss[MM_SWAPENTS]--;
1180                         else if (is_migration_entry(entry)) {
1181                                 struct page *page;
1182
1183                                 page = migration_entry_to_page(entry);
1184
1185                                 if (PageAnon(page))
1186                                         rss[MM_ANONPAGES]--;
1187                                 else
1188                                         rss[MM_FILEPAGES]--;
1189                         }
1190                         if (unlikely(!free_swap_and_cache(entry)))
1191                                 print_bad_pte(vma, addr, ptent, NULL);
1192                 }
1193                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1194         } while (pte++, addr += PAGE_SIZE, addr != end);
1195
1196         add_mm_rss_vec(mm, rss);
1197         arch_leave_lazy_mmu_mode();
1198         pte_unmap_unlock(start_pte, ptl);
1199
1200         /*
1201          * mmu_gather ran out of room to batch pages, we break out of
1202          * the PTE lock to avoid doing the potential expensive TLB invalidate
1203          * and page-free while holding it.
1204          */
1205         if (force_flush) {
1206                 force_flush = 0;
1207
1208 #ifdef HAVE_GENERIC_MMU_GATHER
1209                 tlb->start = addr;
1210                 tlb->end = end;
1211 #endif
1212                 tlb_flush_mmu(tlb);
1213                 if (addr != end)
1214                         goto again;
1215         }
1216
1217         return addr;
1218 }
1219
1220 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1221                                 struct vm_area_struct *vma, pud_t *pud,
1222                                 unsigned long addr, unsigned long end,
1223                                 struct zap_details *details)
1224 {
1225         pmd_t *pmd;
1226         unsigned long next;
1227
1228         pmd = pmd_offset(pud, addr);
1229         do {
1230                 next = pmd_addr_end(addr, end);
1231                 if (pmd_trans_huge(*pmd)) {
1232                         if (next - addr != HPAGE_PMD_SIZE) {
1233 #ifdef CONFIG_DEBUG_VM
1234                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1235                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1236                                                 __func__, addr, end,
1237                                                 vma->vm_start,
1238                                                 vma->vm_end);
1239                                         BUG();
1240                                 }
1241 #endif
1242                                 split_huge_page_pmd(vma, addr, pmd);
1243                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1244                                 goto next;
1245                         /* fall through */
1246                 }
1247                 /*
1248                  * Here there can be other concurrent MADV_DONTNEED or
1249                  * trans huge page faults running, and if the pmd is
1250                  * none or trans huge it can change under us. This is
1251                  * because MADV_DONTNEED holds the mmap_sem in read
1252                  * mode.
1253                  */
1254                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1255                         goto next;
1256                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1257 next:
1258                 cond_resched();
1259         } while (pmd++, addr = next, addr != end);
1260
1261         return addr;
1262 }
1263
1264 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1265                                 struct vm_area_struct *vma, pgd_t *pgd,
1266                                 unsigned long addr, unsigned long end,
1267                                 struct zap_details *details)
1268 {
1269         pud_t *pud;
1270         unsigned long next;
1271
1272         pud = pud_offset(pgd, addr);
1273         do {
1274                 next = pud_addr_end(addr, end);
1275                 if (pud_none_or_clear_bad(pud))
1276                         continue;
1277                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1278         } while (pud++, addr = next, addr != end);
1279
1280         return addr;
1281 }
1282
1283 static void unmap_page_range(struct mmu_gather *tlb,
1284                              struct vm_area_struct *vma,
1285                              unsigned long addr, unsigned long end,
1286                              struct zap_details *details)
1287 {
1288         pgd_t *pgd;
1289         unsigned long next;
1290
1291         if (details && !details->check_mapping && !details->nonlinear_vma)
1292                 details = NULL;
1293
1294         BUG_ON(addr >= end);
1295         mem_cgroup_uncharge_start();
1296         tlb_start_vma(tlb, vma);
1297         pgd = pgd_offset(vma->vm_mm, addr);
1298         do {
1299                 next = pgd_addr_end(addr, end);
1300                 if (pgd_none_or_clear_bad(pgd))
1301                         continue;
1302                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1303         } while (pgd++, addr = next, addr != end);
1304         tlb_end_vma(tlb, vma);
1305         mem_cgroup_uncharge_end();
1306 }
1307
1308
1309 static void unmap_single_vma(struct mmu_gather *tlb,
1310                 struct vm_area_struct *vma, unsigned long start_addr,
1311                 unsigned long end_addr,
1312                 struct zap_details *details)
1313 {
1314         unsigned long start = max(vma->vm_start, start_addr);
1315         unsigned long end;
1316
1317         if (start >= vma->vm_end)
1318                 return;
1319         end = min(vma->vm_end, end_addr);
1320         if (end <= vma->vm_start)
1321                 return;
1322
1323         if (vma->vm_file)
1324                 uprobe_munmap(vma, start, end);
1325
1326         if (unlikely(vma->vm_flags & VM_PFNMAP))
1327                 untrack_pfn(vma, 0, 0);
1328
1329         if (start != end) {
1330                 if (unlikely(is_vm_hugetlb_page(vma))) {
1331                         /*
1332                          * It is undesirable to test vma->vm_file as it
1333                          * should be non-null for valid hugetlb area.
1334                          * However, vm_file will be NULL in the error
1335                          * cleanup path of do_mmap_pgoff. When
1336                          * hugetlbfs ->mmap method fails,
1337                          * do_mmap_pgoff() nullifies vma->vm_file
1338                          * before calling this function to clean up.
1339                          * Since no pte has actually been setup, it is
1340                          * safe to do nothing in this case.
1341                          */
1342                         if (vma->vm_file) {
1343                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1344                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1345                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1346                         }
1347                 } else
1348                         unmap_page_range(tlb, vma, start, end, details);
1349         }
1350 }
1351
1352 /**
1353  * unmap_vmas - unmap a range of memory covered by a list of vma's
1354  * @tlb: address of the caller's struct mmu_gather
1355  * @vma: the starting vma
1356  * @start_addr: virtual address at which to start unmapping
1357  * @end_addr: virtual address at which to end unmapping
1358  *
1359  * Unmap all pages in the vma list.
1360  *
1361  * Only addresses between `start' and `end' will be unmapped.
1362  *
1363  * The VMA list must be sorted in ascending virtual address order.
1364  *
1365  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1366  * range after unmap_vmas() returns.  So the only responsibility here is to
1367  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1368  * drops the lock and schedules.
1369  */
1370 void unmap_vmas(struct mmu_gather *tlb,
1371                 struct vm_area_struct *vma, unsigned long start_addr,
1372                 unsigned long end_addr)
1373 {
1374         struct mm_struct *mm = vma->vm_mm;
1375
1376         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1377         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1378                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1379         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1380 }
1381
1382 /**
1383  * zap_page_range - remove user pages in a given range
1384  * @vma: vm_area_struct holding the applicable pages
1385  * @start: starting address of pages to zap
1386  * @size: number of bytes to zap
1387  * @details: details of nonlinear truncation or shared cache invalidation
1388  *
1389  * Caller must protect the VMA list
1390  */
1391 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1392                 unsigned long size, struct zap_details *details)
1393 {
1394         struct mm_struct *mm = vma->vm_mm;
1395         struct mmu_gather tlb;
1396         unsigned long end = start + size;
1397
1398         lru_add_drain();
1399         tlb_gather_mmu(&tlb, mm, 0);
1400         update_hiwater_rss(mm);
1401         mmu_notifier_invalidate_range_start(mm, start, end);
1402         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1403                 unmap_single_vma(&tlb, vma, start, end, details);
1404         mmu_notifier_invalidate_range_end(mm, start, end);
1405         tlb_finish_mmu(&tlb, start, end);
1406 }
1407
1408 /**
1409  * zap_page_range_single - remove user pages in a given range
1410  * @vma: vm_area_struct holding the applicable pages
1411  * @address: starting address of pages to zap
1412  * @size: number of bytes to zap
1413  * @details: details of nonlinear truncation or shared cache invalidation
1414  *
1415  * The range must fit into one VMA.
1416  */
1417 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1418                 unsigned long size, struct zap_details *details)
1419 {
1420         struct mm_struct *mm = vma->vm_mm;
1421         struct mmu_gather tlb;
1422         unsigned long end = address + size;
1423
1424         lru_add_drain();
1425         tlb_gather_mmu(&tlb, mm, 0);
1426         update_hiwater_rss(mm);
1427         mmu_notifier_invalidate_range_start(mm, address, end);
1428         unmap_single_vma(&tlb, vma, address, end, details);
1429         mmu_notifier_invalidate_range_end(mm, address, end);
1430         tlb_finish_mmu(&tlb, address, end);
1431 }
1432
1433 /**
1434  * zap_vma_ptes - remove ptes mapping the vma
1435  * @vma: vm_area_struct holding ptes to be zapped
1436  * @address: starting address of pages to zap
1437  * @size: number of bytes to zap
1438  *
1439  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1440  *
1441  * The entire address range must be fully contained within the vma.
1442  *
1443  * Returns 0 if successful.
1444  */
1445 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1446                 unsigned long size)
1447 {
1448         if (address < vma->vm_start || address + size > vma->vm_end ||
1449                         !(vma->vm_flags & VM_PFNMAP))
1450                 return -1;
1451         zap_page_range_single(vma, address, size, NULL);
1452         return 0;
1453 }
1454 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1455
1456 /**
1457  * follow_page_mask - look up a page descriptor from a user-virtual address
1458  * @vma: vm_area_struct mapping @address
1459  * @address: virtual address to look up
1460  * @flags: flags modifying lookup behaviour
1461  * @page_mask: on output, *page_mask is set according to the size of the page
1462  *
1463  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1464  *
1465  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1466  * an error pointer if there is a mapping to something not represented
1467  * by a page descriptor (see also vm_normal_page()).
1468  */
1469 struct page *follow_page_mask(struct vm_area_struct *vma,
1470                               unsigned long address, unsigned int flags,
1471                               unsigned int *page_mask)
1472 {
1473         pgd_t *pgd;
1474         pud_t *pud;
1475         pmd_t *pmd;
1476         pte_t *ptep, pte;
1477         spinlock_t *ptl;
1478         struct page *page;
1479         struct mm_struct *mm = vma->vm_mm;
1480
1481         *page_mask = 0;
1482
1483         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1484         if (!IS_ERR(page)) {
1485                 BUG_ON(flags & FOLL_GET);
1486                 goto out;
1487         }
1488
1489         page = NULL;
1490         pgd = pgd_offset(mm, address);
1491         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1492                 goto no_page_table;
1493
1494         pud = pud_offset(pgd, address);
1495         if (pud_none(*pud))
1496                 goto no_page_table;
1497         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1498                 BUG_ON(flags & FOLL_GET);
1499                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1500                 goto out;
1501         }
1502         if (unlikely(pud_bad(*pud)))
1503                 goto no_page_table;
1504
1505         pmd = pmd_offset(pud, address);
1506         if (pmd_none(*pmd))
1507                 goto no_page_table;
1508         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1509                 BUG_ON(flags & FOLL_GET);
1510                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1511                 goto out;
1512         }
1513         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1514                 goto no_page_table;
1515         if (pmd_trans_huge(*pmd)) {
1516                 if (flags & FOLL_SPLIT) {
1517                         split_huge_page_pmd(vma, address, pmd);
1518                         goto split_fallthrough;
1519                 }
1520                 spin_lock(&mm->page_table_lock);
1521                 if (likely(pmd_trans_huge(*pmd))) {
1522                         if (unlikely(pmd_trans_splitting(*pmd))) {
1523                                 spin_unlock(&mm->page_table_lock);
1524                                 wait_split_huge_page(vma->anon_vma, pmd);
1525                         } else {
1526                                 page = follow_trans_huge_pmd(vma, address,
1527                                                              pmd, flags);
1528                                 spin_unlock(&mm->page_table_lock);
1529                                 *page_mask = HPAGE_PMD_NR - 1;
1530                                 goto out;
1531                         }
1532                 } else
1533                         spin_unlock(&mm->page_table_lock);
1534                 /* fall through */
1535         }
1536 split_fallthrough:
1537         if (unlikely(pmd_bad(*pmd)))
1538                 goto no_page_table;
1539
1540         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1541
1542         pte = *ptep;
1543         if (!pte_present(pte)) {
1544                 swp_entry_t entry;
1545                 /*
1546                  * KSM's break_ksm() relies upon recognizing a ksm page
1547                  * even while it is being migrated, so for that case we
1548                  * need migration_entry_wait().
1549                  */
1550                 if (likely(!(flags & FOLL_MIGRATION)))
1551                         goto no_page;
1552                 if (pte_none(pte) || pte_file(pte))
1553                         goto no_page;
1554                 entry = pte_to_swp_entry(pte);
1555                 if (!is_migration_entry(entry))
1556                         goto no_page;
1557                 pte_unmap_unlock(ptep, ptl);
1558                 migration_entry_wait(mm, pmd, address);
1559                 goto split_fallthrough;
1560         }
1561         if ((flags & FOLL_NUMA) && pte_numa(pte))
1562                 goto no_page;
1563         if ((flags & FOLL_WRITE) && !pte_write(pte))
1564                 goto unlock;
1565
1566         page = vm_normal_page(vma, address, pte);
1567         if (unlikely(!page)) {
1568                 if ((flags & FOLL_DUMP) ||
1569                     !is_zero_pfn(pte_pfn(pte)))
1570                         goto bad_page;
1571                 page = pte_page(pte);
1572         }
1573
1574         if (flags & FOLL_GET)
1575                 get_page_foll(page);
1576         if (flags & FOLL_TOUCH) {
1577                 if ((flags & FOLL_WRITE) &&
1578                     !pte_dirty(pte) && !PageDirty(page))
1579                         set_page_dirty(page);
1580                 /*
1581                  * pte_mkyoung() would be more correct here, but atomic care
1582                  * is needed to avoid losing the dirty bit: it is easier to use
1583                  * mark_page_accessed().
1584                  */
1585                 mark_page_accessed(page);
1586         }
1587         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1588                 /*
1589                  * The preliminary mapping check is mainly to avoid the
1590                  * pointless overhead of lock_page on the ZERO_PAGE
1591                  * which might bounce very badly if there is contention.
1592                  *
1593                  * If the page is already locked, we don't need to
1594                  * handle it now - vmscan will handle it later if and
1595                  * when it attempts to reclaim the page.
1596                  */
1597                 if (page->mapping && trylock_page(page)) {
1598                         lru_add_drain();  /* push cached pages to LRU */
1599                         /*
1600                          * Because we lock page here, and migration is
1601                          * blocked by the pte's page reference, and we
1602                          * know the page is still mapped, we don't even
1603                          * need to check for file-cache page truncation.
1604                          */
1605                         mlock_vma_page(page);
1606                         unlock_page(page);
1607                 }
1608         }
1609 unlock:
1610         pte_unmap_unlock(ptep, ptl);
1611 out:
1612         return page;
1613
1614 bad_page:
1615         pte_unmap_unlock(ptep, ptl);
1616         return ERR_PTR(-EFAULT);
1617
1618 no_page:
1619         pte_unmap_unlock(ptep, ptl);
1620         if (!pte_none(pte))
1621                 return page;
1622
1623 no_page_table:
1624         /*
1625          * When core dumping an enormous anonymous area that nobody
1626          * has touched so far, we don't want to allocate unnecessary pages or
1627          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1628          * then get_dump_page() will return NULL to leave a hole in the dump.
1629          * But we can only make this optimization where a hole would surely
1630          * be zero-filled if handle_mm_fault() actually did handle it.
1631          */
1632         if ((flags & FOLL_DUMP) &&
1633             (!vma->vm_ops || !vma->vm_ops->fault))
1634                 return ERR_PTR(-EFAULT);
1635         return page;
1636 }
1637
1638 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1639 {
1640         return stack_guard_page_start(vma, addr) ||
1641                stack_guard_page_end(vma, addr+PAGE_SIZE);
1642 }
1643
1644 /**
1645  * __get_user_pages() - pin user pages in memory
1646  * @tsk:        task_struct of target task
1647  * @mm:         mm_struct of target mm
1648  * @start:      starting user address
1649  * @nr_pages:   number of pages from start to pin
1650  * @gup_flags:  flags modifying pin behaviour
1651  * @pages:      array that receives pointers to the pages pinned.
1652  *              Should be at least nr_pages long. Or NULL, if caller
1653  *              only intends to ensure the pages are faulted in.
1654  * @vmas:       array of pointers to vmas corresponding to each page.
1655  *              Or NULL if the caller does not require them.
1656  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1657  *
1658  * Returns number of pages pinned. This may be fewer than the number
1659  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1660  * were pinned, returns -errno. Each page returned must be released
1661  * with a put_page() call when it is finished with. vmas will only
1662  * remain valid while mmap_sem is held.
1663  *
1664  * Must be called with mmap_sem held for read or write.
1665  *
1666  * __get_user_pages walks a process's page tables and takes a reference to
1667  * each struct page that each user address corresponds to at a given
1668  * instant. That is, it takes the page that would be accessed if a user
1669  * thread accesses the given user virtual address at that instant.
1670  *
1671  * This does not guarantee that the page exists in the user mappings when
1672  * __get_user_pages returns, and there may even be a completely different
1673  * page there in some cases (eg. if mmapped pagecache has been invalidated
1674  * and subsequently re faulted). However it does guarantee that the page
1675  * won't be freed completely. And mostly callers simply care that the page
1676  * contains data that was valid *at some point in time*. Typically, an IO
1677  * or similar operation cannot guarantee anything stronger anyway because
1678  * locks can't be held over the syscall boundary.
1679  *
1680  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1681  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1682  * appropriate) must be called after the page is finished with, and
1683  * before put_page is called.
1684  *
1685  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1686  * or mmap_sem contention, and if waiting is needed to pin all pages,
1687  * *@nonblocking will be set to 0.
1688  *
1689  * In most cases, get_user_pages or get_user_pages_fast should be used
1690  * instead of __get_user_pages. __get_user_pages should be used only if
1691  * you need some special @gup_flags.
1692  */
1693 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1694                 unsigned long start, unsigned long nr_pages,
1695                 unsigned int gup_flags, struct page **pages,
1696                 struct vm_area_struct **vmas, int *nonblocking)
1697 {
1698         long i;
1699         unsigned long vm_flags;
1700         unsigned int page_mask;
1701
1702         if (!nr_pages)
1703                 return 0;
1704
1705         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1706
1707         /* 
1708          * Require read or write permissions.
1709          * If FOLL_FORCE is set, we only require the "MAY" flags.
1710          */
1711         vm_flags  = (gup_flags & FOLL_WRITE) ?
1712                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1713         vm_flags &= (gup_flags & FOLL_FORCE) ?
1714                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1715
1716         /*
1717          * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1718          * would be called on PROT_NONE ranges. We must never invoke
1719          * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1720          * page faults would unprotect the PROT_NONE ranges if
1721          * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1722          * bitflag. So to avoid that, don't set FOLL_NUMA if
1723          * FOLL_FORCE is set.
1724          */
1725         if (!(gup_flags & FOLL_FORCE))
1726                 gup_flags |= FOLL_NUMA;
1727
1728         i = 0;
1729
1730         do {
1731                 struct vm_area_struct *vma;
1732
1733                 vma = find_extend_vma(mm, start);
1734                 if (!vma && in_gate_area(mm, start)) {
1735                         unsigned long pg = start & PAGE_MASK;
1736                         pgd_t *pgd;
1737                         pud_t *pud;
1738                         pmd_t *pmd;
1739                         pte_t *pte;
1740
1741                         /* user gate pages are read-only */
1742                         if (gup_flags & FOLL_WRITE)
1743                                 return i ? : -EFAULT;
1744                         if (pg > TASK_SIZE)
1745                                 pgd = pgd_offset_k(pg);
1746                         else
1747                                 pgd = pgd_offset_gate(mm, pg);
1748                         BUG_ON(pgd_none(*pgd));
1749                         pud = pud_offset(pgd, pg);
1750                         BUG_ON(pud_none(*pud));
1751                         pmd = pmd_offset(pud, pg);
1752                         if (pmd_none(*pmd))
1753                                 return i ? : -EFAULT;
1754                         VM_BUG_ON(pmd_trans_huge(*pmd));
1755                         pte = pte_offset_map(pmd, pg);
1756                         if (pte_none(*pte)) {
1757                                 pte_unmap(pte);
1758                                 return i ? : -EFAULT;
1759                         }
1760                         vma = get_gate_vma(mm);
1761                         if (pages) {
1762                                 struct page *page;
1763
1764                                 page = vm_normal_page(vma, start, *pte);
1765                                 if (!page) {
1766                                         if (!(gup_flags & FOLL_DUMP) &&
1767                                              is_zero_pfn(pte_pfn(*pte)))
1768                                                 page = pte_page(*pte);
1769                                         else {
1770                                                 pte_unmap(pte);
1771                                                 return i ? : -EFAULT;
1772                                         }
1773                                 }
1774                                 pages[i] = page;
1775                                 get_page(page);
1776                         }
1777                         pte_unmap(pte);
1778                         page_mask = 0;
1779                         goto next_page;
1780                 }
1781
1782                 if (!vma ||
1783                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1784                     !(vm_flags & vma->vm_flags))
1785                         return i ? : -EFAULT;
1786
1787                 if (is_vm_hugetlb_page(vma)) {
1788                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1789                                         &start, &nr_pages, i, gup_flags);
1790                         continue;
1791                 }
1792
1793                 do {
1794                         struct page *page;
1795                         unsigned int foll_flags = gup_flags;
1796                         unsigned int page_increm;
1797
1798                         /*
1799                          * If we have a pending SIGKILL, don't keep faulting
1800                          * pages and potentially allocating memory.
1801                          */
1802                         if (unlikely(fatal_signal_pending(current)))
1803                                 return i ? i : -ERESTARTSYS;
1804
1805                         cond_resched();
1806                         while (!(page = follow_page_mask(vma, start,
1807                                                 foll_flags, &page_mask))) {
1808                                 int ret;
1809                                 unsigned int fault_flags = 0;
1810
1811                                 /* For mlock, just skip the stack guard page. */
1812                                 if (foll_flags & FOLL_MLOCK) {
1813                                         if (stack_guard_page(vma, start))
1814                                                 goto next_page;
1815                                 }
1816                                 if (foll_flags & FOLL_WRITE)
1817                                         fault_flags |= FAULT_FLAG_WRITE;
1818                                 if (nonblocking)
1819                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1820                                 if (foll_flags & FOLL_NOWAIT)
1821                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1822
1823                                 ret = handle_mm_fault(mm, vma, start,
1824                                                         fault_flags);
1825
1826                                 if (ret & VM_FAULT_ERROR) {
1827                                         if (ret & VM_FAULT_OOM)
1828                                                 return i ? i : -ENOMEM;
1829                                         if (ret & (VM_FAULT_HWPOISON |
1830                                                    VM_FAULT_HWPOISON_LARGE)) {
1831                                                 if (i)
1832                                                         return i;
1833                                                 else if (gup_flags & FOLL_HWPOISON)
1834                                                         return -EHWPOISON;
1835                                                 else
1836                                                         return -EFAULT;
1837                                         }
1838                                         if (ret & VM_FAULT_SIGBUS)
1839                                                 return i ? i : -EFAULT;
1840                                         BUG();
1841                                 }
1842
1843                                 if (tsk) {
1844                                         if (ret & VM_FAULT_MAJOR)
1845                                                 tsk->maj_flt++;
1846                                         else
1847                                                 tsk->min_flt++;
1848                                 }
1849
1850                                 if (ret & VM_FAULT_RETRY) {
1851                                         if (nonblocking)
1852                                                 *nonblocking = 0;
1853                                         return i;
1854                                 }
1855
1856                                 /*
1857                                  * The VM_FAULT_WRITE bit tells us that
1858                                  * do_wp_page has broken COW when necessary,
1859                                  * even if maybe_mkwrite decided not to set
1860                                  * pte_write. We can thus safely do subsequent
1861                                  * page lookups as if they were reads. But only
1862                                  * do so when looping for pte_write is futile:
1863                                  * in some cases userspace may also be wanting
1864                                  * to write to the gotten user page, which a
1865                                  * read fault here might prevent (a readonly
1866                                  * page might get reCOWed by userspace write).
1867                                  */
1868                                 if ((ret & VM_FAULT_WRITE) &&
1869                                     !(vma->vm_flags & VM_WRITE))
1870                                         foll_flags &= ~FOLL_WRITE;
1871
1872                                 cond_resched();
1873                         }
1874                         if (IS_ERR(page))
1875                                 return i ? i : PTR_ERR(page);
1876                         if (pages) {
1877                                 pages[i] = page;
1878
1879                                 flush_anon_page(vma, page, start);
1880                                 flush_dcache_page(page);
1881                                 page_mask = 0;
1882                         }
1883 next_page:
1884                         if (vmas) {
1885                                 vmas[i] = vma;
1886                                 page_mask = 0;
1887                         }
1888                         page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1889                         if (page_increm > nr_pages)
1890                                 page_increm = nr_pages;
1891                         i += page_increm;
1892                         start += page_increm * PAGE_SIZE;
1893                         nr_pages -= page_increm;
1894                 } while (nr_pages && start < vma->vm_end);
1895         } while (nr_pages);
1896         return i;
1897 }
1898 EXPORT_SYMBOL(__get_user_pages);
1899
1900 /*
1901  * fixup_user_fault() - manually resolve a user page fault
1902  * @tsk:        the task_struct to use for page fault accounting, or
1903  *              NULL if faults are not to be recorded.
1904  * @mm:         mm_struct of target mm
1905  * @address:    user address
1906  * @fault_flags:flags to pass down to handle_mm_fault()
1907  *
1908  * This is meant to be called in the specific scenario where for locking reasons
1909  * we try to access user memory in atomic context (within a pagefault_disable()
1910  * section), this returns -EFAULT, and we want to resolve the user fault before
1911  * trying again.
1912  *
1913  * Typically this is meant to be used by the futex code.
1914  *
1915  * The main difference with get_user_pages() is that this function will
1916  * unconditionally call handle_mm_fault() which will in turn perform all the
1917  * necessary SW fixup of the dirty and young bits in the PTE, while
1918  * handle_mm_fault() only guarantees to update these in the struct page.
1919  *
1920  * This is important for some architectures where those bits also gate the
1921  * access permission to the page because they are maintained in software.  On
1922  * such architectures, gup() will not be enough to make a subsequent access
1923  * succeed.
1924  *
1925  * This should be called with the mm_sem held for read.
1926  */
1927 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1928                      unsigned long address, unsigned int fault_flags)
1929 {
1930         struct vm_area_struct *vma;
1931         int ret;
1932
1933         vma = find_extend_vma(mm, address);
1934         if (!vma || address < vma->vm_start)
1935                 return -EFAULT;
1936
1937         ret = handle_mm_fault(mm, vma, address, fault_flags);
1938         if (ret & VM_FAULT_ERROR) {
1939                 if (ret & VM_FAULT_OOM)
1940                         return -ENOMEM;
1941                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1942                         return -EHWPOISON;
1943                 if (ret & VM_FAULT_SIGBUS)
1944                         return -EFAULT;
1945                 BUG();
1946         }
1947         if (tsk) {
1948                 if (ret & VM_FAULT_MAJOR)
1949                         tsk->maj_flt++;
1950                 else
1951                         tsk->min_flt++;
1952         }
1953         return 0;
1954 }
1955
1956 /*
1957  * get_user_pages() - pin user pages in memory
1958  * @tsk:        the task_struct to use for page fault accounting, or
1959  *              NULL if faults are not to be recorded.
1960  * @mm:         mm_struct of target mm
1961  * @start:      starting user address
1962  * @nr_pages:   number of pages from start to pin
1963  * @write:      whether pages will be written to by the caller
1964  * @force:      whether to force write access even if user mapping is
1965  *              readonly. This will result in the page being COWed even
1966  *              in MAP_SHARED mappings. You do not want this.
1967  * @pages:      array that receives pointers to the pages pinned.
1968  *              Should be at least nr_pages long. Or NULL, if caller
1969  *              only intends to ensure the pages are faulted in.
1970  * @vmas:       array of pointers to vmas corresponding to each page.
1971  *              Or NULL if the caller does not require them.
1972  *
1973  * Returns number of pages pinned. This may be fewer than the number
1974  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1975  * were pinned, returns -errno. Each page returned must be released
1976  * with a put_page() call when it is finished with. vmas will only
1977  * remain valid while mmap_sem is held.
1978  *
1979  * Must be called with mmap_sem held for read or write.
1980  *
1981  * get_user_pages walks a process's page tables and takes a reference to
1982  * each struct page that each user address corresponds to at a given
1983  * instant. That is, it takes the page that would be accessed if a user
1984  * thread accesses the given user virtual address at that instant.
1985  *
1986  * This does not guarantee that the page exists in the user mappings when
1987  * get_user_pages returns, and there may even be a completely different
1988  * page there in some cases (eg. if mmapped pagecache has been invalidated
1989  * and subsequently re faulted). However it does guarantee that the page
1990  * won't be freed completely. And mostly callers simply care that the page
1991  * contains data that was valid *at some point in time*. Typically, an IO
1992  * or similar operation cannot guarantee anything stronger anyway because
1993  * locks can't be held over the syscall boundary.
1994  *
1995  * If write=0, the page must not be written to. If the page is written to,
1996  * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1997  * after the page is finished with, and before put_page is called.
1998  *
1999  * get_user_pages is typically used for fewer-copy IO operations, to get a
2000  * handle on the memory by some means other than accesses via the user virtual
2001  * addresses. The pages may be submitted for DMA to devices or accessed via
2002  * their kernel linear mapping (via the kmap APIs). Care should be taken to
2003  * use the correct cache flushing APIs.
2004  *
2005  * See also get_user_pages_fast, for performance critical applications.
2006  */
2007 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2008                 unsigned long start, unsigned long nr_pages, int write,
2009                 int force, struct page **pages, struct vm_area_struct **vmas)
2010 {
2011         int flags = FOLL_TOUCH;
2012
2013         if (pages)
2014                 flags |= FOLL_GET;
2015         if (write)
2016                 flags |= FOLL_WRITE;
2017         if (force)
2018                 flags |= FOLL_FORCE;
2019
2020         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2021                                 NULL);
2022 }
2023 EXPORT_SYMBOL(get_user_pages);
2024
2025 /**
2026  * get_dump_page() - pin user page in memory while writing it to core dump
2027  * @addr: user address
2028  *
2029  * Returns struct page pointer of user page pinned for dump,
2030  * to be freed afterwards by page_cache_release() or put_page().
2031  *
2032  * Returns NULL on any kind of failure - a hole must then be inserted into
2033  * the corefile, to preserve alignment with its headers; and also returns
2034  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2035  * allowing a hole to be left in the corefile to save diskspace.
2036  *
2037  * Called without mmap_sem, but after all other threads have been killed.
2038  */
2039 #ifdef CONFIG_ELF_CORE
2040 struct page *get_dump_page(unsigned long addr)
2041 {
2042         struct vm_area_struct *vma;
2043         struct page *page;
2044
2045         if (__get_user_pages(current, current->mm, addr, 1,
2046                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2047                              NULL) < 1)
2048                 return NULL;
2049         flush_cache_page(vma, addr, page_to_pfn(page));
2050         return page;
2051 }
2052 #endif /* CONFIG_ELF_CORE */
2053
2054 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2055                         spinlock_t **ptl)
2056 {
2057         pgd_t * pgd = pgd_offset(mm, addr);
2058         pud_t * pud = pud_alloc(mm, pgd, addr);
2059         if (pud) {
2060                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2061                 if (pmd) {
2062                         VM_BUG_ON(pmd_trans_huge(*pmd));
2063                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2064                 }
2065         }
2066         return NULL;
2067 }
2068
2069 /*
2070  * This is the old fallback for page remapping.
2071  *
2072  * For historical reasons, it only allows reserved pages. Only
2073  * old drivers should use this, and they needed to mark their
2074  * pages reserved for the old functions anyway.
2075  */
2076 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2077                         struct page *page, pgprot_t prot)
2078 {
2079         struct mm_struct *mm = vma->vm_mm;
2080         int retval;
2081         pte_t *pte;
2082         spinlock_t *ptl;
2083
2084         retval = -EINVAL;
2085         if (PageAnon(page))
2086                 goto out;
2087         retval = -ENOMEM;
2088         flush_dcache_page(page);
2089         pte = get_locked_pte(mm, addr, &ptl);
2090         if (!pte)
2091                 goto out;
2092         retval = -EBUSY;
2093         if (!pte_none(*pte))
2094                 goto out_unlock;
2095
2096         /* Ok, finally just insert the thing.. */
2097         get_page(page);
2098         inc_mm_counter_fast(mm, MM_FILEPAGES);
2099         page_add_file_rmap(page);
2100         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2101
2102         retval = 0;
2103         pte_unmap_unlock(pte, ptl);
2104         return retval;
2105 out_unlock:
2106         pte_unmap_unlock(pte, ptl);
2107 out:
2108         return retval;
2109 }
2110
2111 /**
2112  * vm_insert_page - insert single page into user vma
2113  * @vma: user vma to map to
2114  * @addr: target user address of this page
2115  * @page: source kernel page
2116  *
2117  * This allows drivers to insert individual pages they've allocated
2118  * into a user vma.
2119  *
2120  * The page has to be a nice clean _individual_ kernel allocation.
2121  * If you allocate a compound page, you need to have marked it as
2122  * such (__GFP_COMP), or manually just split the page up yourself
2123  * (see split_page()).
2124  *
2125  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2126  * took an arbitrary page protection parameter. This doesn't allow
2127  * that. Your vma protection will have to be set up correctly, which
2128  * means that if you want a shared writable mapping, you'd better
2129  * ask for a shared writable mapping!
2130  *
2131  * The page does not need to be reserved.
2132  *
2133  * Usually this function is called from f_op->mmap() handler
2134  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2135  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2136  * function from other places, for example from page-fault handler.
2137  */
2138 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2139                         struct page *page)
2140 {
2141         if (addr < vma->vm_start || addr >= vma->vm_end)
2142                 return -EFAULT;
2143         if (!page_count(page))
2144                 return -EINVAL;
2145         if (!(vma->vm_flags & VM_MIXEDMAP)) {
2146                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2147                 BUG_ON(vma->vm_flags & VM_PFNMAP);
2148                 vma->vm_flags |= VM_MIXEDMAP;
2149         }
2150         return insert_page(vma, addr, page, vma->vm_page_prot);
2151 }
2152 EXPORT_SYMBOL(vm_insert_page);
2153
2154 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2155                         unsigned long pfn, pgprot_t prot)
2156 {
2157         struct mm_struct *mm = vma->vm_mm;
2158         int retval;
2159         pte_t *pte, entry;
2160         spinlock_t *ptl;
2161
2162         retval = -ENOMEM;
2163         pte = get_locked_pte(mm, addr, &ptl);
2164         if (!pte)
2165                 goto out;
2166         retval = -EBUSY;
2167         if (!pte_none(*pte))
2168                 goto out_unlock;
2169
2170         /* Ok, finally just insert the thing.. */
2171         entry = pte_mkspecial(pfn_pte(pfn, prot));
2172         set_pte_at(mm, addr, pte, entry);
2173         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2174
2175         retval = 0;
2176 out_unlock:
2177         pte_unmap_unlock(pte, ptl);
2178 out:
2179         return retval;
2180 }
2181
2182 /**
2183  * vm_insert_pfn - insert single pfn into user vma
2184  * @vma: user vma to map to
2185  * @addr: target user address of this page
2186  * @pfn: source kernel pfn
2187  *
2188  * Similar to vm_insert_page, this allows drivers to insert individual pages
2189  * they've allocated into a user vma. Same comments apply.
2190  *
2191  * This function should only be called from a vm_ops->fault handler, and
2192  * in that case the handler should return NULL.
2193  *
2194  * vma cannot be a COW mapping.
2195  *
2196  * As this is called only for pages that do not currently exist, we
2197  * do not need to flush old virtual caches or the TLB.
2198  */
2199 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2200                         unsigned long pfn)
2201 {
2202         int ret;
2203         pgprot_t pgprot = vma->vm_page_prot;
2204         /*
2205          * Technically, architectures with pte_special can avoid all these
2206          * restrictions (same for remap_pfn_range).  However we would like
2207          * consistency in testing and feature parity among all, so we should
2208          * try to keep these invariants in place for everybody.
2209          */
2210         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2211         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2212                                                 (VM_PFNMAP|VM_MIXEDMAP));
2213         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2214         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2215
2216         if (addr < vma->vm_start || addr >= vma->vm_end)
2217                 return -EFAULT;
2218         if (track_pfn_insert(vma, &pgprot, pfn))
2219                 return -EINVAL;
2220
2221         ret = insert_pfn(vma, addr, pfn, pgprot);
2222
2223         return ret;
2224 }
2225 EXPORT_SYMBOL(vm_insert_pfn);
2226
2227 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2228                         unsigned long pfn)
2229 {
2230         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2231
2232         if (addr < vma->vm_start || addr >= vma->vm_end)
2233                 return -EFAULT;
2234
2235         /*
2236          * If we don't have pte special, then we have to use the pfn_valid()
2237          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2238          * refcount the page if pfn_valid is true (hence insert_page rather
2239          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2240          * without pte special, it would there be refcounted as a normal page.
2241          */
2242         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2243                 struct page *page;
2244
2245                 page = pfn_to_page(pfn);
2246                 return insert_page(vma, addr, page, vma->vm_page_prot);
2247         }
2248         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2249 }
2250 EXPORT_SYMBOL(vm_insert_mixed);
2251
2252 /*
2253  * maps a range of physical memory into the requested pages. the old
2254  * mappings are removed. any references to nonexistent pages results
2255  * in null mappings (currently treated as "copy-on-access")
2256  */
2257 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2258                         unsigned long addr, unsigned long end,
2259                         unsigned long pfn, pgprot_t prot)
2260 {
2261         pte_t *pte;
2262         spinlock_t *ptl;
2263
2264         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2265         if (!pte)
2266                 return -ENOMEM;
2267         arch_enter_lazy_mmu_mode();
2268         do {
2269                 BUG_ON(!pte_none(*pte));
2270                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2271                 pfn++;
2272         } while (pte++, addr += PAGE_SIZE, addr != end);
2273         arch_leave_lazy_mmu_mode();
2274         pte_unmap_unlock(pte - 1, ptl);
2275         return 0;
2276 }
2277
2278 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2279                         unsigned long addr, unsigned long end,
2280                         unsigned long pfn, pgprot_t prot)
2281 {
2282         pmd_t *pmd;
2283         unsigned long next;
2284
2285         pfn -= addr >> PAGE_SHIFT;
2286         pmd = pmd_alloc(mm, pud, addr);
2287         if (!pmd)
2288                 return -ENOMEM;
2289         VM_BUG_ON(pmd_trans_huge(*pmd));
2290         do {
2291                 next = pmd_addr_end(addr, end);
2292                 if (remap_pte_range(mm, pmd, addr, next,
2293                                 pfn + (addr >> PAGE_SHIFT), prot))
2294                         return -ENOMEM;
2295         } while (pmd++, addr = next, addr != end);
2296         return 0;
2297 }
2298
2299 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2300                         unsigned long addr, unsigned long end,
2301                         unsigned long pfn, pgprot_t prot)
2302 {
2303         pud_t *pud;
2304         unsigned long next;
2305
2306         pfn -= addr >> PAGE_SHIFT;
2307         pud = pud_alloc(mm, pgd, addr);
2308         if (!pud)
2309                 return -ENOMEM;
2310         do {
2311                 next = pud_addr_end(addr, end);
2312                 if (remap_pmd_range(mm, pud, addr, next,
2313                                 pfn + (addr >> PAGE_SHIFT), prot))
2314                         return -ENOMEM;
2315         } while (pud++, addr = next, addr != end);
2316         return 0;
2317 }
2318
2319 /**
2320  * remap_pfn_range - remap kernel memory to userspace
2321  * @vma: user vma to map to
2322  * @addr: target user address to start at
2323  * @pfn: physical address of kernel memory
2324  * @size: size of map area
2325  * @prot: page protection flags for this mapping
2326  *
2327  *  Note: this is only safe if the mm semaphore is held when called.
2328  */
2329 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2330                     unsigned long pfn, unsigned long size, pgprot_t prot)
2331 {
2332         pgd_t *pgd;
2333         unsigned long next;
2334         unsigned long end = addr + PAGE_ALIGN(size);
2335         struct mm_struct *mm = vma->vm_mm;
2336         int err;
2337
2338         /*
2339          * Physically remapped pages are special. Tell the
2340          * rest of the world about it:
2341          *   VM_IO tells people not to look at these pages
2342          *      (accesses can have side effects).
2343          *   VM_PFNMAP tells the core MM that the base pages are just
2344          *      raw PFN mappings, and do not have a "struct page" associated
2345          *      with them.
2346          *   VM_DONTEXPAND
2347          *      Disable vma merging and expanding with mremap().
2348          *   VM_DONTDUMP
2349          *      Omit vma from core dump, even when VM_IO turned off.
2350          *
2351          * There's a horrible special case to handle copy-on-write
2352          * behaviour that some programs depend on. We mark the "original"
2353          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2354          * See vm_normal_page() for details.
2355          */
2356         if (is_cow_mapping(vma->vm_flags)) {
2357                 if (addr != vma->vm_start || end != vma->vm_end)
2358                         return -EINVAL;
2359                 vma->vm_pgoff = pfn;
2360         }
2361
2362         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2363         if (err)
2364                 return -EINVAL;
2365
2366         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2367
2368         BUG_ON(addr >= end);
2369         pfn -= addr >> PAGE_SHIFT;
2370         pgd = pgd_offset(mm, addr);
2371         flush_cache_range(vma, addr, end);
2372         do {
2373                 next = pgd_addr_end(addr, end);
2374                 err = remap_pud_range(mm, pgd, addr, next,
2375                                 pfn + (addr >> PAGE_SHIFT), prot);
2376                 if (err)
2377                         break;
2378         } while (pgd++, addr = next, addr != end);
2379
2380         if (err)
2381                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2382
2383         return err;
2384 }
2385 EXPORT_SYMBOL(remap_pfn_range);
2386
2387 /**
2388  * vm_iomap_memory - remap memory to userspace
2389  * @vma: user vma to map to
2390  * @start: start of area
2391  * @len: size of area
2392  *
2393  * This is a simplified io_remap_pfn_range() for common driver use. The
2394  * driver just needs to give us the physical memory range to be mapped,
2395  * we'll figure out the rest from the vma information.
2396  *
2397  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2398  * whatever write-combining details or similar.
2399  */
2400 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2401 {
2402         unsigned long vm_len, pfn, pages;
2403
2404         /* Check that the physical memory area passed in looks valid */
2405         if (start + len < start)
2406                 return -EINVAL;
2407         /*
2408          * You *really* shouldn't map things that aren't page-aligned,
2409          * but we've historically allowed it because IO memory might
2410          * just have smaller alignment.
2411          */
2412         len += start & ~PAGE_MASK;
2413         pfn = start >> PAGE_SHIFT;
2414         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2415         if (pfn + pages < pfn)
2416                 return -EINVAL;
2417
2418         /* We start the mapping 'vm_pgoff' pages into the area */
2419         if (vma->vm_pgoff > pages)
2420                 return -EINVAL;
2421         pfn += vma->vm_pgoff;
2422         pages -= vma->vm_pgoff;
2423
2424         /* Can we fit all of the mapping? */
2425         vm_len = vma->vm_end - vma->vm_start;
2426         if (vm_len >> PAGE_SHIFT > pages)
2427                 return -EINVAL;
2428
2429         /* Ok, let it rip */
2430         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2431 }
2432 EXPORT_SYMBOL(vm_iomap_memory);
2433
2434 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2435                                      unsigned long addr, unsigned long end,
2436                                      pte_fn_t fn, void *data)
2437 {
2438         pte_t *pte;
2439         int err;
2440         pgtable_t token;
2441         spinlock_t *uninitialized_var(ptl);
2442
2443         pte = (mm == &init_mm) ?
2444                 pte_alloc_kernel(pmd, addr) :
2445                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2446         if (!pte)
2447                 return -ENOMEM;
2448
2449         BUG_ON(pmd_huge(*pmd));
2450
2451         arch_enter_lazy_mmu_mode();
2452
2453         token = pmd_pgtable(*pmd);
2454
2455         do {
2456                 err = fn(pte++, token, addr, data);
2457                 if (err)
2458                         break;
2459         } while (addr += PAGE_SIZE, addr != end);
2460
2461         arch_leave_lazy_mmu_mode();
2462
2463         if (mm != &init_mm)
2464                 pte_unmap_unlock(pte-1, ptl);
2465         return err;
2466 }
2467
2468 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2469                                      unsigned long addr, unsigned long end,
2470                                      pte_fn_t fn, void *data)
2471 {
2472         pmd_t *pmd;
2473         unsigned long next;
2474         int err;
2475
2476         BUG_ON(pud_huge(*pud));
2477
2478         pmd = pmd_alloc(mm, pud, addr);
2479         if (!pmd)
2480                 return -ENOMEM;
2481         do {
2482                 next = pmd_addr_end(addr, end);
2483                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2484                 if (err)
2485                         break;
2486         } while (pmd++, addr = next, addr != end);
2487         return err;
2488 }
2489
2490 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2491                                      unsigned long addr, unsigned long end,
2492                                      pte_fn_t fn, void *data)
2493 {
2494         pud_t *pud;
2495         unsigned long next;
2496         int err;
2497
2498         pud = pud_alloc(mm, pgd, addr);
2499         if (!pud)
2500                 return -ENOMEM;
2501         do {
2502                 next = pud_addr_end(addr, end);
2503                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2504                 if (err)
2505                         break;
2506         } while (pud++, addr = next, addr != end);
2507         return err;
2508 }
2509
2510 /*
2511  * Scan a region of virtual memory, filling in page tables as necessary
2512  * and calling a provided function on each leaf page table.
2513  */
2514 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2515                         unsigned long size, pte_fn_t fn, void *data)
2516 {
2517         pgd_t *pgd;
2518         unsigned long next;
2519         unsigned long end = addr + size;
2520         int err;
2521
2522         BUG_ON(addr >= end);
2523         pgd = pgd_offset(mm, addr);
2524         do {
2525                 next = pgd_addr_end(addr, end);
2526                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2527                 if (err)
2528                         break;
2529         } while (pgd++, addr = next, addr != end);
2530
2531         return err;
2532 }
2533 EXPORT_SYMBOL_GPL(apply_to_page_range);
2534
2535 /*
2536  * handle_pte_fault chooses page fault handler according to an entry
2537  * which was read non-atomically.  Before making any commitment, on
2538  * those architectures or configurations (e.g. i386 with PAE) which
2539  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2540  * must check under lock before unmapping the pte and proceeding
2541  * (but do_wp_page is only called after already making such a check;
2542  * and do_anonymous_page can safely check later on).
2543  */
2544 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2545                                 pte_t *page_table, pte_t orig_pte)
2546 {
2547         int same = 1;
2548 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2549         if (sizeof(pte_t) > sizeof(unsigned long)) {
2550                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2551                 spin_lock(ptl);
2552                 same = pte_same(*page_table, orig_pte);
2553                 spin_unlock(ptl);
2554         }
2555 #endif
2556         pte_unmap(page_table);
2557         return same;
2558 }
2559
2560 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2561 {
2562         /*
2563          * If the source page was a PFN mapping, we don't have
2564          * a "struct page" for it. We do a best-effort copy by
2565          * just copying from the original user address. If that
2566          * fails, we just zero-fill it. Live with it.
2567          */
2568         if (unlikely(!src)) {
2569                 void *kaddr = kmap_atomic(dst);
2570                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2571
2572                 /*
2573                  * This really shouldn't fail, because the page is there
2574                  * in the page tables. But it might just be unreadable,
2575                  * in which case we just give up and fill the result with
2576                  * zeroes.
2577                  */
2578                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2579                         clear_page(kaddr);
2580                 kunmap_atomic(kaddr);
2581                 flush_dcache_page(dst);
2582         } else
2583                 copy_user_highpage(dst, src, va, vma);
2584 }
2585
2586 /*
2587  * This routine handles present pages, when users try to write
2588  * to a shared page. It is done by copying the page to a new address
2589  * and decrementing the shared-page counter for the old page.
2590  *
2591  * Note that this routine assumes that the protection checks have been
2592  * done by the caller (the low-level page fault routine in most cases).
2593  * Thus we can safely just mark it writable once we've done any necessary
2594  * COW.
2595  *
2596  * We also mark the page dirty at this point even though the page will
2597  * change only once the write actually happens. This avoids a few races,
2598  * and potentially makes it more efficient.
2599  *
2600  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2601  * but allow concurrent faults), with pte both mapped and locked.
2602  * We return with mmap_sem still held, but pte unmapped and unlocked.
2603  */
2604 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2605                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2606                 spinlock_t *ptl, pte_t orig_pte)
2607         __releases(ptl)
2608 {
2609         struct page *old_page, *new_page = NULL;
2610         pte_t entry;
2611         int ret = 0;
2612         int page_mkwrite = 0;
2613         struct page *dirty_page = NULL;
2614         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2615         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2616
2617         old_page = vm_normal_page(vma, address, orig_pte);
2618         if (!old_page) {
2619                 /*
2620                  * VM_MIXEDMAP !pfn_valid() case
2621                  *
2622                  * We should not cow pages in a shared writeable mapping.
2623                  * Just mark the pages writable as we can't do any dirty
2624                  * accounting on raw pfn maps.
2625                  */
2626                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2627                                      (VM_WRITE|VM_SHARED))
2628                         goto reuse;
2629                 goto gotten;
2630         }
2631
2632         /*
2633          * Take out anonymous pages first, anonymous shared vmas are
2634          * not dirty accountable.
2635          */
2636         if (PageAnon(old_page) && !PageKsm(old_page)) {
2637                 if (!trylock_page(old_page)) {
2638                         page_cache_get(old_page);
2639                         pte_unmap_unlock(page_table, ptl);
2640                         lock_page(old_page);
2641                         page_table = pte_offset_map_lock(mm, pmd, address,
2642                                                          &ptl);
2643                         if (!pte_same(*page_table, orig_pte)) {
2644                                 unlock_page(old_page);
2645                                 goto unlock;
2646                         }
2647                         page_cache_release(old_page);
2648                 }
2649                 if (reuse_swap_page(old_page)) {
2650                         /*
2651                          * The page is all ours.  Move it to our anon_vma so
2652                          * the rmap code will not search our parent or siblings.
2653                          * Protected against the rmap code by the page lock.
2654                          */
2655                         page_move_anon_rmap(old_page, vma, address);
2656                         unlock_page(old_page);
2657                         goto reuse;
2658                 }
2659                 unlock_page(old_page);
2660         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2661                                         (VM_WRITE|VM_SHARED))) {
2662                 /*
2663                  * Only catch write-faults on shared writable pages,
2664                  * read-only shared pages can get COWed by
2665                  * get_user_pages(.write=1, .force=1).
2666                  */
2667                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2668                         struct vm_fault vmf;
2669                         int tmp;
2670
2671                         vmf.virtual_address = (void __user *)(address &
2672                                                                 PAGE_MASK);
2673                         vmf.pgoff = old_page->index;
2674                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2675                         vmf.page = old_page;
2676
2677                         /*
2678                          * Notify the address space that the page is about to
2679                          * become writable so that it can prohibit this or wait
2680                          * for the page to get into an appropriate state.
2681                          *
2682                          * We do this without the lock held, so that it can
2683                          * sleep if it needs to.
2684                          */
2685                         page_cache_get(old_page);
2686                         pte_unmap_unlock(page_table, ptl);
2687
2688                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2689                         if (unlikely(tmp &
2690                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2691                                 ret = tmp;
2692                                 goto unwritable_page;
2693                         }
2694                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2695                                 lock_page(old_page);
2696                                 if (!old_page->mapping) {
2697                                         ret = 0; /* retry the fault */
2698                                         unlock_page(old_page);
2699                                         goto unwritable_page;
2700                                 }
2701                         } else
2702                                 VM_BUG_ON(!PageLocked(old_page));
2703
2704                         /*
2705                          * Since we dropped the lock we need to revalidate
2706                          * the PTE as someone else may have changed it.  If
2707                          * they did, we just return, as we can count on the
2708                          * MMU to tell us if they didn't also make it writable.
2709                          */
2710                         page_table = pte_offset_map_lock(mm, pmd, address,
2711                                                          &ptl);
2712                         if (!pte_same(*page_table, orig_pte)) {
2713                                 unlock_page(old_page);
2714                                 goto unlock;
2715                         }
2716
2717                         page_mkwrite = 1;
2718                 }
2719                 dirty_page = old_page;
2720                 get_page(dirty_page);
2721
2722 reuse:
2723                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2724                 entry = pte_mkyoung(orig_pte);
2725                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2726                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2727                         update_mmu_cache(vma, address, page_table);
2728                 pte_unmap_unlock(page_table, ptl);
2729                 ret |= VM_FAULT_WRITE;
2730
2731                 if (!dirty_page)
2732                         return ret;
2733
2734                 /*
2735                  * Yes, Virginia, this is actually required to prevent a race
2736                  * with clear_page_dirty_for_io() from clearing the page dirty
2737                  * bit after it clear all dirty ptes, but before a racing
2738                  * do_wp_page installs a dirty pte.
2739                  *
2740                  * __do_fault is protected similarly.
2741                  */
2742                 if (!page_mkwrite) {
2743                         wait_on_page_locked(dirty_page);
2744                         set_page_dirty_balance(dirty_page, page_mkwrite);
2745                         /* file_update_time outside page_lock */
2746                         if (vma->vm_file)
2747                                 file_update_time(vma->vm_file);
2748                 }
2749                 put_page(dirty_page);
2750                 if (page_mkwrite) {
2751                         struct address_space *mapping = dirty_page->mapping;
2752
2753                         set_page_dirty(dirty_page);
2754                         unlock_page(dirty_page);
2755                         page_cache_release(dirty_page);
2756                         if (mapping)    {
2757                                 /*
2758                                  * Some device drivers do not set page.mapping
2759                                  * but still dirty their pages
2760                                  */
2761                                 balance_dirty_pages_ratelimited(mapping);
2762                         }
2763                 }
2764
2765                 return ret;
2766         }
2767
2768         /*
2769          * Ok, we need to copy. Oh, well..
2770          */
2771         page_cache_get(old_page);
2772 gotten:
2773         pte_unmap_unlock(page_table, ptl);
2774
2775         if (unlikely(anon_vma_prepare(vma)))
2776                 goto oom;
2777
2778         if (is_zero_pfn(pte_pfn(orig_pte))) {
2779                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2780                 if (!new_page)
2781                         goto oom;
2782         } else {
2783                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2784                 if (!new_page)
2785                         goto oom;
2786                 cow_user_page(new_page, old_page, address, vma);
2787         }
2788         __SetPageUptodate(new_page);
2789
2790         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2791                 goto oom_free_new;
2792
2793         mmun_start  = address & PAGE_MASK;
2794         mmun_end    = mmun_start + PAGE_SIZE;
2795         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2796
2797         /*
2798          * Re-check the pte - we dropped the lock
2799          */
2800         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2801         if (likely(pte_same(*page_table, orig_pte))) {
2802                 if (old_page) {
2803                         if (!PageAnon(old_page)) {
2804                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2805                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2806                         }
2807                 } else
2808                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2809                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2810                 entry = mk_pte(new_page, vma->vm_page_prot);
2811                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2812                 /*
2813                  * Clear the pte entry and flush it first, before updating the
2814                  * pte with the new entry. This will avoid a race condition
2815                  * seen in the presence of one thread doing SMC and another
2816                  * thread doing COW.
2817                  */
2818                 ptep_clear_flush(vma, address, page_table);
2819                 page_add_new_anon_rmap(new_page, vma, address);
2820                 /*
2821                  * We call the notify macro here because, when using secondary
2822                  * mmu page tables (such as kvm shadow page tables), we want the
2823                  * new page to be mapped directly into the secondary page table.
2824                  */
2825                 set_pte_at_notify(mm, address, page_table, entry);
2826                 update_mmu_cache(vma, address, page_table);
2827                 if (old_page) {
2828                         /*
2829                          * Only after switching the pte to the new page may
2830                          * we remove the mapcount here. Otherwise another
2831                          * process may come and find the rmap count decremented
2832                          * before the pte is switched to the new page, and
2833                          * "reuse" the old page writing into it while our pte
2834                          * here still points into it and can be read by other
2835                          * threads.
2836                          *
2837                          * The critical issue is to order this
2838                          * page_remove_rmap with the ptp_clear_flush above.
2839                          * Those stores are ordered by (if nothing else,)
2840                          * the barrier present in the atomic_add_negative
2841                          * in page_remove_rmap.
2842                          *
2843                          * Then the TLB flush in ptep_clear_flush ensures that
2844                          * no process can access the old page before the
2845                          * decremented mapcount is visible. And the old page
2846                          * cannot be reused until after the decremented
2847                          * mapcount is visible. So transitively, TLBs to
2848                          * old page will be flushed before it can be reused.
2849                          */
2850                         page_remove_rmap(old_page);
2851                 }
2852
2853                 /* Free the old page.. */
2854                 new_page = old_page;
2855                 ret |= VM_FAULT_WRITE;
2856         } else
2857                 mem_cgroup_uncharge_page(new_page);
2858
2859         if (new_page)
2860                 page_cache_release(new_page);
2861 unlock:
2862         pte_unmap_unlock(page_table, ptl);
2863         if (mmun_end > mmun_start)
2864                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2865         if (old_page) {
2866                 /*
2867                  * Don't let another task, with possibly unlocked vma,
2868                  * keep the mlocked page.
2869                  */
2870                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2871                         lock_page(old_page);    /* LRU manipulation */
2872                         munlock_vma_page(old_page);
2873                         unlock_page(old_page);
2874                 }
2875                 page_cache_release(old_page);
2876         }
2877         return ret;
2878 oom_free_new:
2879         page_cache_release(new_page);
2880 oom:
2881         if (old_page)
2882                 page_cache_release(old_page);
2883         return VM_FAULT_OOM;
2884
2885 unwritable_page:
2886         page_cache_release(old_page);
2887         return ret;
2888 }
2889
2890 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2891                 unsigned long start_addr, unsigned long end_addr,
2892                 struct zap_details *details)
2893 {
2894         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2895 }
2896
2897 static inline void unmap_mapping_range_tree(struct rb_root *root,
2898                                             struct zap_details *details)
2899 {
2900         struct vm_area_struct *vma;
2901         pgoff_t vba, vea, zba, zea;
2902
2903         vma_interval_tree_foreach(vma, root,
2904                         details->first_index, details->last_index) {
2905
2906                 vba = vma->vm_pgoff;
2907                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2908                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2909                 zba = details->first_index;
2910                 if (zba < vba)
2911                         zba = vba;
2912                 zea = details->last_index;
2913                 if (zea > vea)
2914                         zea = vea;
2915
2916                 unmap_mapping_range_vma(vma,
2917                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2918                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2919                                 details);
2920         }
2921 }
2922
2923 static inline void unmap_mapping_range_list(struct list_head *head,
2924                                             struct zap_details *details)
2925 {
2926         struct vm_area_struct *vma;
2927
2928         /*
2929          * In nonlinear VMAs there is no correspondence between virtual address
2930          * offset and file offset.  So we must perform an exhaustive search
2931          * across *all* the pages in each nonlinear VMA, not just the pages
2932          * whose virtual address lies outside the file truncation point.
2933          */
2934         list_for_each_entry(vma, head, shared.nonlinear) {
2935                 details->nonlinear_vma = vma;
2936                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2937         }
2938 }
2939
2940 /**
2941  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2942  * @mapping: the address space containing mmaps to be unmapped.
2943  * @holebegin: byte in first page to unmap, relative to the start of
2944  * the underlying file.  This will be rounded down to a PAGE_SIZE
2945  * boundary.  Note that this is different from truncate_pagecache(), which
2946  * must keep the partial page.  In contrast, we must get rid of
2947  * partial pages.
2948  * @holelen: size of prospective hole in bytes.  This will be rounded
2949  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2950  * end of the file.
2951  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2952  * but 0 when invalidating pagecache, don't throw away private data.
2953  */
2954 void unmap_mapping_range(struct address_space *mapping,
2955                 loff_t const holebegin, loff_t const holelen, int even_cows)
2956 {
2957         struct zap_details details;
2958         pgoff_t hba = holebegin >> PAGE_SHIFT;
2959         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2960
2961         /* Check for overflow. */
2962         if (sizeof(holelen) > sizeof(hlen)) {
2963                 long long holeend =
2964                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2965                 if (holeend & ~(long long)ULONG_MAX)
2966                         hlen = ULONG_MAX - hba + 1;
2967         }
2968
2969         details.check_mapping = even_cows? NULL: mapping;
2970         details.nonlinear_vma = NULL;
2971         details.first_index = hba;
2972         details.last_index = hba + hlen - 1;
2973         if (details.last_index < details.first_index)
2974                 details.last_index = ULONG_MAX;
2975
2976
2977         mutex_lock(&mapping->i_mmap_mutex);
2978         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2979                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2980         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2981                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2982         mutex_unlock(&mapping->i_mmap_mutex);
2983 }
2984 EXPORT_SYMBOL(unmap_mapping_range);
2985
2986 /*
2987  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2988  * but allow concurrent faults), and pte mapped but not yet locked.
2989  * We return with mmap_sem still held, but pte unmapped and unlocked.
2990  */
2991 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2992                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2993                 unsigned int flags, pte_t orig_pte)
2994 {
2995         spinlock_t *ptl;
2996         struct page *page, *swapcache;
2997         swp_entry_t entry;
2998         pte_t pte;
2999         int locked;
3000         struct mem_cgroup *ptr;
3001         int exclusive = 0;
3002         int ret = 0;
3003
3004         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3005                 goto out;
3006
3007         entry = pte_to_swp_entry(orig_pte);
3008         if (unlikely(non_swap_entry(entry))) {
3009                 if (is_migration_entry(entry)) {
3010                         migration_entry_wait(mm, pmd, address);
3011                 } else if (is_hwpoison_entry(entry)) {
3012                         ret = VM_FAULT_HWPOISON;
3013                 } else {
3014                         print_bad_pte(vma, address, orig_pte, NULL);
3015                         ret = VM_FAULT_SIGBUS;
3016                 }
3017                 goto out;
3018         }
3019         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3020         page = lookup_swap_cache(entry);
3021         if (!page) {
3022                 page = swapin_readahead(entry,
3023                                         GFP_HIGHUSER_MOVABLE, vma, address);
3024                 if (!page) {
3025                         /*
3026                          * Back out if somebody else faulted in this pte
3027                          * while we released the pte lock.
3028                          */
3029                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3030                         if (likely(pte_same(*page_table, orig_pte)))
3031                                 ret = VM_FAULT_OOM;
3032                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3033                         goto unlock;
3034                 }
3035
3036                 /* Had to read the page from swap area: Major fault */
3037                 ret = VM_FAULT_MAJOR;
3038                 count_vm_event(PGMAJFAULT);
3039                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3040         } else if (PageHWPoison(page)) {
3041                 /*
3042                  * hwpoisoned dirty swapcache pages are kept for killing
3043                  * owner processes (which may be unknown at hwpoison time)
3044                  */
3045                 ret = VM_FAULT_HWPOISON;
3046                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3047                 swapcache = page;
3048                 goto out_release;
3049         }
3050
3051         swapcache = page;
3052         locked = lock_page_or_retry(page, mm, flags);
3053
3054         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3055         if (!locked) {
3056                 ret |= VM_FAULT_RETRY;
3057                 goto out_release;
3058         }
3059
3060         /*
3061          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3062          * release the swapcache from under us.  The page pin, and pte_same
3063          * test below, are not enough to exclude that.  Even if it is still
3064          * swapcache, we need to check that the page's swap has not changed.
3065          */
3066         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3067                 goto out_page;
3068
3069         page = ksm_might_need_to_copy(page, vma, address);
3070         if (unlikely(!page)) {
3071                 ret = VM_FAULT_OOM;
3072                 page = swapcache;
3073                 goto out_page;
3074         }
3075
3076         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3077                 ret = VM_FAULT_OOM;
3078                 goto out_page;
3079         }
3080
3081         /*
3082          * Back out if somebody else already faulted in this pte.
3083          */
3084         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3085         if (unlikely(!pte_same(*page_table, orig_pte)))
3086                 goto out_nomap;
3087
3088         if (unlikely(!PageUptodate(page))) {
3089                 ret = VM_FAULT_SIGBUS;
3090                 goto out_nomap;
3091         }
3092
3093         /*
3094          * The page isn't present yet, go ahead with the fault.
3095          *
3096          * Be careful about the sequence of operations here.
3097          * To get its accounting right, reuse_swap_page() must be called
3098          * while the page is counted on swap but not yet in mapcount i.e.
3099          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3100          * must be called after the swap_free(), or it will never succeed.
3101          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3102          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3103          * in page->private. In this case, a record in swap_cgroup  is silently
3104          * discarded at swap_free().
3105          */
3106
3107         inc_mm_counter_fast(mm, MM_ANONPAGES);
3108         dec_mm_counter_fast(mm, MM_SWAPENTS);
3109         pte = mk_pte(page, vma->vm_page_prot);
3110         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3111                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3112                 flags &= ~FAULT_FLAG_WRITE;
3113                 ret |= VM_FAULT_WRITE;
3114                 exclusive = 1;
3115         }
3116         flush_icache_page(vma, page);
3117         set_pte_at(mm, address, page_table, pte);
3118         if (page == swapcache)
3119                 do_page_add_anon_rmap(page, vma, address, exclusive);
3120         else /* ksm created a completely new copy */
3121                 page_add_new_anon_rmap(page, vma, address);
3122         /* It's better to call commit-charge after rmap is established */
3123         mem_cgroup_commit_charge_swapin(page, ptr);
3124
3125         swap_free(entry);
3126         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3127                 try_to_free_swap(page);
3128         unlock_page(page);
3129         if (page != swapcache) {
3130                 /*
3131                  * Hold the lock to avoid the swap entry to be reused
3132                  * until we take the PT lock for the pte_same() check
3133                  * (to avoid false positives from pte_same). For
3134                  * further safety release the lock after the swap_free
3135                  * so that the swap count won't change under a
3136                  * parallel locked swapcache.
3137                  */
3138                 unlock_page(swapcache);
3139                 page_cache_release(swapcache);
3140         }
3141
3142         if (flags & FAULT_FLAG_WRITE) {
3143                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3144                 if (ret & VM_FAULT_ERROR)
3145                         ret &= VM_FAULT_ERROR;
3146                 goto out;
3147         }
3148
3149         /* No need to invalidate - it was non-present before */
3150         update_mmu_cache(vma, address, page_table);
3151 unlock:
3152         pte_unmap_unlock(page_table, ptl);
3153 out:
3154         return ret;
3155 out_nomap:
3156         mem_cgroup_cancel_charge_swapin(ptr);
3157         pte_unmap_unlock(page_table, ptl);
3158 out_page:
3159         unlock_page(page);
3160 out_release:
3161         page_cache_release(page);
3162         if (page != swapcache) {
3163                 unlock_page(swapcache);
3164                 page_cache_release(swapcache);
3165         }
3166         return ret;
3167 }
3168
3169 /*
3170  * This is like a special single-page "expand_{down|up}wards()",
3171  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3172  * doesn't hit another vma.
3173  */
3174 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3175 {
3176         address &= PAGE_MASK;
3177         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3178                 struct vm_area_struct *prev = vma->vm_prev;
3179
3180                 /*
3181                  * Is there a mapping abutting this one below?
3182                  *
3183                  * That's only ok if it's the same stack mapping
3184                  * that has gotten split..
3185                  */
3186                 if (prev && prev->vm_end == address)
3187                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3188
3189                 expand_downwards(vma, address - PAGE_SIZE);
3190         }
3191         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3192                 struct vm_area_struct *next = vma->vm_next;
3193
3194                 /* As VM_GROWSDOWN but s/below/above/ */
3195                 if (next && next->vm_start == address + PAGE_SIZE)
3196                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3197
3198                 expand_upwards(vma, address + PAGE_SIZE);
3199         }
3200         return 0;
3201 }
3202
3203 /*
3204  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3205  * but allow concurrent faults), and pte mapped but not yet locked.
3206  * We return with mmap_sem still held, but pte unmapped and unlocked.
3207  */
3208 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3209                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3210                 unsigned int flags)
3211 {
3212         struct page *page;
3213         spinlock_t *ptl;
3214         pte_t entry;
3215
3216         pte_unmap(page_table);
3217
3218         /* Check if we need to add a guard page to the stack */
3219         if (check_stack_guard_page(vma, address) < 0)
3220                 return VM_FAULT_SIGBUS;
3221
3222         /* Use the zero-page for reads */
3223         if (!(flags & FAULT_FLAG_WRITE)) {
3224                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3225                                                 vma->vm_page_prot));
3226                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3227                 if (!pte_none(*page_table))
3228                         goto unlock;
3229                 goto setpte;
3230         }
3231
3232         /* Allocate our own private page. */
3233         if (unlikely(anon_vma_prepare(vma)))
3234                 goto oom;
3235         page = alloc_zeroed_user_highpage_movable(vma, address);
3236         if (!page)
3237                 goto oom;
3238         /*
3239          * The memory barrier inside __SetPageUptodate makes sure that
3240          * preceeding stores to the page contents become visible before
3241          * the set_pte_at() write.
3242          */
3243         __SetPageUptodate(page);
3244
3245         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3246                 goto oom_free_page;
3247
3248         entry = mk_pte(page, vma->vm_page_prot);
3249         if (vma->vm_flags & VM_WRITE)
3250                 entry = pte_mkwrite(pte_mkdirty(entry));
3251
3252         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3253         if (!pte_none(*page_table))
3254                 goto release;
3255
3256         inc_mm_counter_fast(mm, MM_ANONPAGES);
3257         page_add_new_anon_rmap(page, vma, address);
3258 setpte:
3259         set_pte_at(mm, address, page_table, entry);
3260
3261         /* No need to invalidate - it was non-present before */
3262         update_mmu_cache(vma, address, page_table);
3263 unlock:
3264         pte_unmap_unlock(page_table, ptl);
3265         return 0;
3266 release:
3267         mem_cgroup_uncharge_page(page);
3268         page_cache_release(page);
3269         goto unlock;
3270 oom_free_page:
3271         page_cache_release(page);
3272 oom:
3273         return VM_FAULT_OOM;
3274 }
3275
3276 /*
3277  * __do_fault() tries to create a new page mapping. It aggressively
3278  * tries to share with existing pages, but makes a separate copy if
3279  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3280  * the next page fault.
3281  *
3282  * As this is called only for pages that do not currently exist, we
3283  * do not need to flush old virtual caches or the TLB.
3284  *
3285  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3286  * but allow concurrent faults), and pte neither mapped nor locked.
3287  * We return with mmap_sem still held, but pte unmapped and unlocked.
3288  */
3289 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3290                 unsigned long address, pmd_t *pmd,
3291                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3292 {
3293         pte_t *page_table;
3294         spinlock_t *ptl;
3295         struct page *page;
3296         struct page *cow_page;
3297         pte_t entry;
3298         int anon = 0;
3299         struct page *dirty_page = NULL;
3300         struct vm_fault vmf;
3301         int ret;
3302         int page_mkwrite = 0;
3303
3304         /*
3305          * If we do COW later, allocate page befor taking lock_page()
3306          * on the file cache page. This will reduce lock holding time.
3307          */
3308         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3309
3310                 if (unlikely(anon_vma_prepare(vma)))
3311                         return VM_FAULT_OOM;
3312
3313                 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3314                 if (!cow_page)
3315                         return VM_FAULT_OOM;
3316
3317                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3318                         page_cache_release(cow_page);
3319                         return VM_FAULT_OOM;
3320                 }
3321         } else
3322                 cow_page = NULL;
3323
3324         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3325         vmf.pgoff = pgoff;
3326         vmf.flags = flags;
3327         vmf.page = NULL;
3328
3329         ret = vma->vm_ops->fault(vma, &vmf);
3330         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3331                             VM_FAULT_RETRY)))
3332                 goto uncharge_out;
3333
3334         if (unlikely(PageHWPoison(vmf.page))) {
3335                 if (ret & VM_FAULT_LOCKED)
3336                         unlock_page(vmf.page);
3337                 ret = VM_FAULT_HWPOISON;
3338                 goto uncharge_out;
3339         }
3340
3341         /*
3342          * For consistency in subsequent calls, make the faulted page always
3343          * locked.
3344          */
3345         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3346                 lock_page(vmf.page);
3347         else
3348                 VM_BUG_ON(!PageLocked(vmf.page));
3349
3350         /*
3351          * Should we do an early C-O-W break?
3352          */
3353         page = vmf.page;
3354         if (flags & FAULT_FLAG_WRITE) {
3355                 if (!(vma->vm_flags & VM_SHARED)) {
3356                         page = cow_page;
3357                         anon = 1;
3358                         copy_user_highpage(page, vmf.page, address, vma);
3359                         __SetPageUptodate(page);
3360                 } else {
3361                         /*
3362                          * If the page will be shareable, see if the backing
3363                          * address space wants to know that the page is about
3364                          * to become writable
3365                          */
3366                         if (vma->vm_ops->page_mkwrite) {
3367                                 int tmp;
3368
3369                                 unlock_page(page);
3370                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3371                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3372                                 if (unlikely(tmp &
3373                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3374                                         ret = tmp;
3375                                         goto unwritable_page;
3376                                 }
3377                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3378                                         lock_page(page);
3379                                         if (!page->mapping) {
3380                                                 ret = 0; /* retry the fault */
3381                                                 unlock_page(page);
3382                                                 goto unwritable_page;
3383                                         }
3384                                 } else
3385                                         VM_BUG_ON(!PageLocked(page));
3386                                 page_mkwrite = 1;
3387                         }
3388                 }
3389
3390         }
3391
3392         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3393
3394         /*
3395          * This silly early PAGE_DIRTY setting removes a race
3396          * due to the bad i386 page protection. But it's valid
3397          * for other architectures too.
3398          *
3399          * Note that if FAULT_FLAG_WRITE is set, we either now have
3400          * an exclusive copy of the page, or this is a shared mapping,
3401          * so we can make it writable and dirty to avoid having to
3402          * handle that later.
3403          */
3404         /* Only go through if we didn't race with anybody else... */
3405         if (likely(pte_same(*page_table, orig_pte))) {
3406                 flush_icache_page(vma, page);
3407                 entry = mk_pte(page, vma->vm_page_prot);
3408                 if (flags & FAULT_FLAG_WRITE)
3409                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3410                 if (anon) {
3411                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3412                         page_add_new_anon_rmap(page, vma, address);
3413                 } else {
3414                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3415                         page_add_file_rmap(page);
3416                         if (flags & FAULT_FLAG_WRITE) {
3417                                 dirty_page = page;
3418                                 get_page(dirty_page);
3419                         }
3420                 }
3421                 set_pte_at(mm, address, page_table, entry);
3422
3423                 /* no need to invalidate: a not-present page won't be cached */
3424                 update_mmu_cache(vma, address, page_table);
3425         } else {
3426                 if (cow_page)
3427                         mem_cgroup_uncharge_page(cow_page);
3428                 if (anon)
3429                         page_cache_release(page);
3430                 else
3431                         anon = 1; /* no anon but release faulted_page */
3432         }
3433
3434         pte_unmap_unlock(page_table, ptl);
3435
3436         if (dirty_page) {
3437                 struct address_space *mapping = page->mapping;
3438                 int dirtied = 0;
3439
3440                 if (set_page_dirty(dirty_page))
3441                         dirtied = 1;
3442                 unlock_page(dirty_page);
3443                 put_page(dirty_page);
3444                 if ((dirtied || page_mkwrite) && mapping) {
3445                         /*
3446                          * Some device drivers do not set page.mapping but still
3447                          * dirty their pages
3448                          */
3449                         balance_dirty_pages_ratelimited(mapping);
3450                 }
3451
3452                 /* file_update_time outside page_lock */
3453                 if (vma->vm_file && !page_mkwrite)
3454                         file_update_time(vma->vm_file);
3455         } else {
3456                 unlock_page(vmf.page);
3457                 if (anon)
3458                         page_cache_release(vmf.page);
3459         }
3460
3461         return ret;
3462
3463 unwritable_page:
3464         page_cache_release(page);
3465         return ret;
3466 uncharge_out:
3467         /* fs's fault handler get error */
3468         if (cow_page) {
3469                 mem_cgroup_uncharge_page(cow_page);
3470                 page_cache_release(cow_page);
3471         }
3472         return ret;
3473 }
3474
3475 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3476                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3477                 unsigned int flags, pte_t orig_pte)
3478 {
3479         pgoff_t pgoff = (((address & PAGE_MASK)
3480                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3481
3482         pte_unmap(page_table);
3483         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3484 }
3485
3486 /*
3487  * Fault of a previously existing named mapping. Repopulate the pte
3488  * from the encoded file_pte if possible. This enables swappable
3489  * nonlinear vmas.
3490  *
3491  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3492  * but allow concurrent faults), and pte mapped but not yet locked.
3493  * We return with mmap_sem still held, but pte unmapped and unlocked.
3494  */
3495 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3496                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3497                 unsigned int flags, pte_t orig_pte)
3498 {
3499         pgoff_t pgoff;
3500
3501         flags |= FAULT_FLAG_NONLINEAR;
3502
3503         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3504                 return 0;
3505
3506         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3507                 /*
3508                  * Page table corrupted: show pte and kill process.
3509                  */
3510                 print_bad_pte(vma, address, orig_pte, NULL);
3511                 return VM_FAULT_SIGBUS;
3512         }
3513
3514         pgoff = pte_to_pgoff(orig_pte);
3515         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3516 }
3517
3518 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3519                                 unsigned long addr, int current_nid)
3520 {
3521         get_page(page);
3522
3523         count_vm_numa_event(NUMA_HINT_FAULTS);
3524         if (current_nid == numa_node_id())
3525                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3526
3527         return mpol_misplaced(page, vma, addr);
3528 }
3529
3530 int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3531                    unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3532 {
3533         struct page *page = NULL;
3534         spinlock_t *ptl;
3535         int current_nid = -1;
3536         int target_nid;
3537         bool migrated = false;
3538
3539         /*
3540         * The "pte" at this point cannot be used safely without
3541         * validation through pte_unmap_same(). It's of NUMA type but
3542         * the pfn may be screwed if the read is non atomic.
3543         *
3544         * ptep_modify_prot_start is not called as this is clearing
3545         * the _PAGE_NUMA bit and it is not really expected that there
3546         * would be concurrent hardware modifications to the PTE.
3547         */
3548         ptl = pte_lockptr(mm, pmd);
3549         spin_lock(ptl);
3550         if (unlikely(!pte_same(*ptep, pte))) {
3551                 pte_unmap_unlock(ptep, ptl);
3552                 goto out;
3553         }
3554
3555         pte = pte_mknonnuma(pte);
3556         set_pte_at(mm, addr, ptep, pte);
3557         update_mmu_cache(vma, addr, ptep);
3558
3559         page = vm_normal_page(vma, addr, pte);
3560         if (!page) {
3561                 pte_unmap_unlock(ptep, ptl);
3562                 return 0;
3563         }
3564
3565         current_nid = page_to_nid(page);
3566         target_nid = numa_migrate_prep(page, vma, addr, current_nid);
3567         pte_unmap_unlock(ptep, ptl);
3568         if (target_nid == -1) {
3569                 /*
3570                  * Account for the fault against the current node if it not
3571                  * being replaced regardless of where the page is located.
3572                  */
3573                 current_nid = numa_node_id();
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                 current_nid = target_nid;
3582
3583 out:
3584         if (current_nid != -1)
3585                 task_numa_fault(current_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         int local_nid = numa_node_id();
3601
3602         spin_lock(&mm->page_table_lock);
3603         pmd = *pmdp;
3604         if (pmd_numa(pmd)) {
3605                 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3606                 numa = true;
3607         }
3608         spin_unlock(&mm->page_table_lock);
3609
3610         if (!numa)
3611                 return 0;
3612
3613         /* we're in a page fault so some vma must be in the range */
3614         BUG_ON(!vma);
3615         BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3616         offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3617         VM_BUG_ON(offset >= PMD_SIZE);
3618         orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3619         pte += offset >> PAGE_SHIFT;
3620         for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3621                 pte_t pteval = *pte;
3622                 struct page *page;
3623                 int curr_nid = local_nid;
3624                 int target_nid;
3625                 bool migrated;
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                 /*
3648                  * Note that the NUMA fault is later accounted to either
3649                  * the node that is currently running or where the page is
3650                  * migrated to.
3651                  */
3652                 curr_nid = local_nid;
3653                 target_nid = numa_migrate_prep(page, vma, addr,
3654                                                page_to_nid(page));
3655                 if (target_nid == -1) {
3656                         put_page(page);
3657                         continue;
3658                 }
3659
3660                 /* Migrate to the requested node */
3661                 pte_unmap_unlock(pte, ptl);
3662                 migrated = migrate_misplaced_page(page, target_nid);
3663                 if (migrated)
3664                         curr_nid = target_nid;
3665                 task_numa_fault(curr_nid, 1, migrated);
3666
3667                 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
3668         }
3669         pte_unmap_unlock(orig_pte, ptl);
3670
3671         return 0;
3672 }
3673 #else
3674 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3675                      unsigned long addr, pmd_t *pmdp)
3676 {
3677         BUG();
3678         return 0;
3679 }
3680 #endif /* CONFIG_NUMA_BALANCING */
3681
3682 /*
3683  * These routines also need to handle stuff like marking pages dirty
3684  * and/or accessed for architectures that don't do it in hardware (most
3685  * RISC architectures).  The early dirtying is also good on the i386.
3686  *
3687  * There is also a hook called "update_mmu_cache()" that architectures
3688  * with external mmu caches can use to update those (ie the Sparc or
3689  * PowerPC hashed page tables that act as extended TLBs).
3690  *
3691  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3692  * but allow concurrent faults), and pte mapped but not yet locked.
3693  * We return with mmap_sem still held, but pte unmapped and unlocked.
3694  */
3695 int handle_pte_fault(struct mm_struct *mm,
3696                      struct vm_area_struct *vma, unsigned long address,
3697                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3698 {
3699         pte_t entry;
3700         spinlock_t *ptl;
3701
3702         entry = *pte;
3703         if (!pte_present(entry)) {
3704                 if (pte_none(entry)) {
3705                         if (vma->vm_ops) {
3706                                 if (likely(vma->vm_ops->fault))
3707                                         return do_linear_fault(mm, vma, address,
3708                                                 pte, pmd, flags, entry);
3709                         }
3710                         return do_anonymous_page(mm, vma, address,
3711                                                  pte, pmd, flags);
3712                 }
3713                 if (pte_file(entry))
3714                         return do_nonlinear_fault(mm, vma, address,
3715                                         pte, pmd, flags, entry);
3716                 return do_swap_page(mm, vma, address,
3717                                         pte, pmd, flags, entry);
3718         }
3719
3720         if (pte_numa(entry))
3721                 return do_numa_page(mm, vma, address, entry, pte, pmd);
3722
3723         ptl = pte_lockptr(mm, pmd);
3724         spin_lock(ptl);
3725         if (unlikely(!pte_same(*pte, entry)))
3726                 goto unlock;
3727         if (flags & FAULT_FLAG_WRITE) {
3728                 if (!pte_write(entry))
3729                         return do_wp_page(mm, vma, address,
3730                                         pte, pmd, ptl, entry);
3731                 entry = pte_mkdirty(entry);
3732         }
3733         entry = pte_mkyoung(entry);
3734         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3735                 update_mmu_cache(vma, address, pte);
3736         } else {
3737                 /*
3738                  * This is needed only for protection faults but the arch code
3739                  * is not yet telling us if this is a protection fault or not.
3740                  * This still avoids useless tlb flushes for .text page faults
3741                  * with threads.
3742                  */
3743                 if (flags & FAULT_FLAG_WRITE)
3744                         flush_tlb_fix_spurious_fault(vma, address);
3745         }
3746 unlock:
3747         pte_unmap_unlock(pte, ptl);
3748         return 0;
3749 }
3750
3751 /*
3752  * By the time we get here, we already hold the mm semaphore
3753  */
3754 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3755                 unsigned long address, unsigned int flags)
3756 {
3757         pgd_t *pgd;
3758         pud_t *pud;
3759         pmd_t *pmd;
3760         pte_t *pte;
3761
3762         __set_current_state(TASK_RUNNING);
3763
3764         count_vm_event(PGFAULT);
3765         mem_cgroup_count_vm_event(mm, PGFAULT);
3766
3767         /* do counter updates before entering really critical section. */
3768         check_sync_rss_stat(current);
3769
3770         if (unlikely(is_vm_hugetlb_page(vma)))
3771                 return hugetlb_fault(mm, vma, address, flags);
3772
3773 retry:
3774         pgd = pgd_offset(mm, address);
3775         pud = pud_alloc(mm, pgd, address);
3776         if (!pud)
3777                 return VM_FAULT_OOM;
3778         pmd = pmd_alloc(mm, pud, address);
3779         if (!pmd)
3780                 return VM_FAULT_OOM;
3781         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3782                 if (!vma->vm_ops)
3783                         return do_huge_pmd_anonymous_page(mm, vma, address,
3784                                                           pmd, flags);
3785         } else {
3786                 pmd_t orig_pmd = *pmd;
3787                 int ret;
3788
3789                 barrier();
3790                 if (pmd_trans_huge(orig_pmd)) {
3791                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3792
3793                         /*
3794                          * If the pmd is splitting, return and retry the
3795                          * the fault.  Alternative: wait until the split
3796                          * is done, and goto retry.
3797                          */
3798                         if (pmd_trans_splitting(orig_pmd))
3799                                 return 0;
3800
3801                         if (pmd_numa(orig_pmd))
3802                                 return do_huge_pmd_numa_page(mm, vma, address,
3803                                                              orig_pmd, pmd);
3804
3805                         if (dirty && !pmd_write(orig_pmd)) {
3806                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3807                                                           orig_pmd);
3808                                 /*
3809                                  * If COW results in an oom, the huge pmd will
3810                                  * have been split, so retry the fault on the
3811                                  * pte for a smaller charge.
3812                                  */
3813                                 if (unlikely(ret & VM_FAULT_OOM))
3814                                         goto retry;
3815                                 return ret;
3816                         } else {
3817                                 huge_pmd_set_accessed(mm, vma, address, pmd,
3818                                                       orig_pmd, dirty);
3819                         }
3820
3821                         return 0;
3822                 }
3823         }
3824
3825         if (pmd_numa(*pmd))
3826                 return do_pmd_numa_page(mm, vma, address, pmd);
3827
3828         /*
3829          * Use __pte_alloc instead of pte_alloc_map, because we can't
3830          * run pte_offset_map on the pmd, if an huge pmd could
3831          * materialize from under us from a different thread.
3832          */
3833         if (unlikely(pmd_none(*pmd)) &&
3834             unlikely(__pte_alloc(mm, vma, pmd, address)))
3835                 return VM_FAULT_OOM;
3836         /* if an huge pmd materialized from under us just retry later */
3837         if (unlikely(pmd_trans_huge(*pmd)))
3838                 return 0;
3839         /*
3840          * A regular pmd is established and it can't morph into a huge pmd
3841          * from under us anymore at this point because we hold the mmap_sem
3842          * read mode and khugepaged takes it in write mode. So now it's
3843          * safe to run pte_offset_map().
3844          */
3845         pte = pte_offset_map(pmd, address);
3846
3847         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3848 }
3849
3850 #ifndef __PAGETABLE_PUD_FOLDED
3851 /*
3852  * Allocate page upper directory.
3853  * We've already handled the fast-path in-line.
3854  */
3855 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3856 {
3857         pud_t *new = pud_alloc_one(mm, address);
3858         if (!new)
3859                 return -ENOMEM;
3860
3861         smp_wmb(); /* See comment in __pte_alloc */
3862
3863         spin_lock(&mm->page_table_lock);
3864         if (pgd_present(*pgd))          /* Another has populated it */
3865                 pud_free(mm, new);
3866         else
3867                 pgd_populate(mm, pgd, new);
3868         spin_unlock(&mm->page_table_lock);
3869         return 0;
3870 }
3871 #endif /* __PAGETABLE_PUD_FOLDED */
3872
3873 #ifndef __PAGETABLE_PMD_FOLDED
3874 /*
3875  * Allocate page middle directory.
3876  * We've already handled the fast-path in-line.
3877  */
3878 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3879 {
3880         pmd_t *new = pmd_alloc_one(mm, address);
3881         if (!new)
3882                 return -ENOMEM;
3883
3884         smp_wmb(); /* See comment in __pte_alloc */
3885
3886         spin_lock(&mm->page_table_lock);
3887 #ifndef __ARCH_HAS_4LEVEL_HACK
3888         if (pud_present(*pud))          /* Another has populated it */
3889                 pmd_free(mm, new);
3890         else
3891                 pud_populate(mm, pud, new);
3892 #else
3893         if (pgd_present(*pud))          /* Another has populated it */
3894                 pmd_free(mm, new);
3895         else
3896                 pgd_populate(mm, pud, new);
3897 #endif /* __ARCH_HAS_4LEVEL_HACK */
3898         spin_unlock(&mm->page_table_lock);
3899         return 0;
3900 }
3901 #endif /* __PAGETABLE_PMD_FOLDED */
3902
3903 #if !defined(__HAVE_ARCH_GATE_AREA)
3904
3905 #if defined(AT_SYSINFO_EHDR)
3906 static struct vm_area_struct gate_vma;
3907
3908 static int __init gate_vma_init(void)
3909 {
3910         gate_vma.vm_mm = NULL;
3911         gate_vma.vm_start = FIXADDR_USER_START;
3912         gate_vma.vm_end = FIXADDR_USER_END;
3913         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3914         gate_vma.vm_page_prot = __P101;
3915
3916         return 0;
3917 }
3918 __initcall(gate_vma_init);
3919 #endif
3920
3921 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3922 {
3923 #ifdef AT_SYSINFO_EHDR
3924         return &gate_vma;
3925 #else
3926         return NULL;
3927 #endif
3928 }
3929
3930 int in_gate_area_no_mm(unsigned long addr)
3931 {
3932 #ifdef AT_SYSINFO_EHDR
3933         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3934                 return 1;
3935 #endif
3936         return 0;
3937 }
3938
3939 #endif  /* __HAVE_ARCH_GATE_AREA */
3940
3941 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3942                 pte_t **ptepp, spinlock_t **ptlp)
3943 {
3944         pgd_t *pgd;
3945         pud_t *pud;
3946         pmd_t *pmd;
3947         pte_t *ptep;
3948
3949         pgd = pgd_offset(mm, address);
3950         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3951                 goto out;
3952
3953         pud = pud_offset(pgd, address);
3954         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3955                 goto out;
3956
3957         pmd = pmd_offset(pud, address);
3958         VM_BUG_ON(pmd_trans_huge(*pmd));
3959         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3960                 goto out;
3961
3962         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3963         if (pmd_huge(*pmd))
3964                 goto out;
3965
3966         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3967         if (!ptep)
3968                 goto out;
3969         if (!pte_present(*ptep))
3970                 goto unlock;
3971         *ptepp = ptep;
3972         return 0;
3973 unlock:
3974         pte_unmap_unlock(ptep, *ptlp);
3975 out:
3976         return -EINVAL;
3977 }
3978
3979 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3980                              pte_t **ptepp, spinlock_t **ptlp)
3981 {
3982         int res;
3983
3984         /* (void) is needed to make gcc happy */
3985         (void) __cond_lock(*ptlp,
3986                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3987         return res;
3988 }
3989
3990 /**
3991  * follow_pfn - look up PFN at a user virtual address
3992  * @vma: memory mapping
3993  * @address: user virtual address
3994  * @pfn: location to store found PFN
3995  *
3996  * Only IO mappings and raw PFN mappings are allowed.
3997  *
3998  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3999  */
4000 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4001         unsigned long *pfn)
4002 {
4003         int ret = -EINVAL;
4004         spinlock_t *ptl;
4005         pte_t *ptep;
4006
4007         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4008                 return ret;
4009
4010         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4011         if (ret)
4012                 return ret;
4013         *pfn = pte_pfn(*ptep);
4014         pte_unmap_unlock(ptep, ptl);
4015         return 0;
4016 }
4017 EXPORT_SYMBOL(follow_pfn);
4018
4019 #ifdef CONFIG_HAVE_IOREMAP_PROT
4020 int follow_phys(struct vm_area_struct *vma,
4021                 unsigned long address, unsigned int flags,
4022                 unsigned long *prot, resource_size_t *phys)
4023 {
4024         int ret = -EINVAL;
4025         pte_t *ptep, pte;
4026         spinlock_t *ptl;
4027
4028         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4029                 goto out;
4030
4031         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4032                 goto out;
4033         pte = *ptep;
4034
4035         if ((flags & FOLL_WRITE) && !pte_write(pte))
4036                 goto unlock;
4037
4038         *prot = pgprot_val(pte_pgprot(pte));
4039         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4040
4041         ret = 0;
4042 unlock:
4043         pte_unmap_unlock(ptep, ptl);
4044 out:
4045         return ret;
4046 }
4047
4048 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4049                         void *buf, int len, int write)
4050 {
4051         resource_size_t phys_addr;
4052         unsigned long prot = 0;
4053         void __iomem *maddr;
4054         int offset = addr & (PAGE_SIZE-1);
4055
4056         if (follow_phys(vma, addr, write, &prot, &phys_addr))
4057                 return -EINVAL;
4058
4059         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4060         if (write)
4061                 memcpy_toio(maddr + offset, buf, len);
4062         else
4063                 memcpy_fromio(buf, maddr + offset, len);
4064         iounmap(maddr);
4065
4066         return len;
4067 }
4068 #endif
4069
4070 /*
4071  * Access another process' address space as given in mm.  If non-NULL, use the
4072  * given task for page fault accounting.
4073  */
4074 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4075                 unsigned long addr, void *buf, int len, int write)
4076 {
4077         struct vm_area_struct *vma;
4078         void *old_buf = buf;
4079
4080         down_read(&mm->mmap_sem);
4081         /* ignore errors, just check how much was successfully transferred */
4082         while (len) {
4083                 int bytes, ret, offset;
4084                 void *maddr;
4085                 struct page *page = NULL;
4086
4087                 ret = get_user_pages(tsk, mm, addr, 1,
4088                                 write, 1, &page, &vma);
4089                 if (ret <= 0) {
4090                         /*
4091                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
4092                          * we can access using slightly different code.
4093                          */
4094 #ifdef CONFIG_HAVE_IOREMAP_PROT
4095                         vma = find_vma(mm, addr);
4096                         if (!vma || vma->vm_start > addr)
4097                                 break;
4098                         if (vma->vm_ops && vma->vm_ops->access)
4099                                 ret = vma->vm_ops->access(vma, addr, buf,
4100                                                           len, write);
4101                         if (ret <= 0)
4102 #endif
4103                                 break;
4104                         bytes = ret;
4105                 } else {
4106                         bytes = len;
4107                         offset = addr & (PAGE_SIZE-1);
4108                         if (bytes > PAGE_SIZE-offset)
4109                                 bytes = PAGE_SIZE-offset;
4110
4111                         maddr = kmap(page);
4112                         if (write) {
4113                                 copy_to_user_page(vma, page, addr,
4114                                                   maddr + offset, buf, bytes);
4115                                 set_page_dirty_lock(page);
4116                         } else {
4117                                 copy_from_user_page(vma, page, addr,
4118                                                     buf, maddr + offset, bytes);
4119                         }
4120                         kunmap(page);
4121                         page_cache_release(page);
4122                 }
4123                 len -= bytes;
4124                 buf += bytes;
4125                 addr += bytes;
4126         }
4127         up_read(&mm->mmap_sem);
4128
4129         return buf - old_buf;
4130 }
4131
4132 /**
4133  * access_remote_vm - access another process' address space
4134  * @mm:         the mm_struct of the target address space
4135  * @addr:       start address to access
4136  * @buf:        source or destination buffer
4137  * @len:        number of bytes to transfer
4138  * @write:      whether the access is a write
4139  *
4140  * The caller must hold a reference on @mm.
4141  */
4142 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4143                 void *buf, int len, int write)
4144 {
4145         return __access_remote_vm(NULL, mm, addr, buf, len, write);
4146 }
4147
4148 /*
4149  * Access another process' address space.
4150  * Source/target buffer must be kernel space,
4151  * Do not walk the page table directly, use get_user_pages
4152  */
4153 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4154                 void *buf, int len, int write)
4155 {
4156         struct mm_struct *mm;
4157         int ret;
4158
4159         mm = get_task_mm(tsk);
4160         if (!mm)
4161                 return 0;
4162
4163         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4164         mmput(mm);
4165
4166         return ret;
4167 }
4168
4169 /*
4170  * Print the name of a VMA.
4171  */
4172 void print_vma_addr(char *prefix, unsigned long ip)
4173 {
4174         struct mm_struct *mm = current->mm;
4175         struct vm_area_struct *vma;
4176
4177         /*
4178          * Do not print if we are in atomic
4179          * contexts (in exception stacks, etc.):
4180          */
4181         if (preempt_count())
4182                 return;
4183
4184         down_read(&mm->mmap_sem);
4185         vma = find_vma(mm, ip);
4186         if (vma && vma->vm_file) {
4187                 struct file *f = vma->vm_file;
4188                 char *buf = (char *)__get_free_page(GFP_KERNEL);
4189                 if (buf) {
4190                         char *p;
4191
4192                         p = d_path(&f->f_path, buf, PAGE_SIZE);
4193                         if (IS_ERR(p))
4194                                 p = "?";
4195                         printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4196                                         vma->vm_start,
4197                                         vma->vm_end - vma->vm_start);
4198                         free_page((unsigned long)buf);
4199                 }
4200         }
4201         up_read(&mm->mmap_sem);
4202 }
4203
4204 #ifdef CONFIG_PROVE_LOCKING
4205 void might_fault(void)
4206 {
4207         /*
4208          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4209          * holding the mmap_sem, this is safe because kernel memory doesn't
4210          * get paged out, therefore we'll never actually fault, and the
4211          * below annotations will generate false positives.
4212          */
4213         if (segment_eq(get_fs(), KERNEL_DS))
4214                 return;
4215
4216         might_sleep();
4217         /*
4218          * it would be nicer only to annotate paths which are not under
4219          * pagefault_disable, however that requires a larger audit and
4220          * providing helpers like get_user_atomic.
4221          */
4222         if (!in_atomic() && current->mm)
4223                 might_lock_read(&current->mm->mmap_sem);
4224 }
4225 EXPORT_SYMBOL(might_fault);
4226 #endif
4227
4228 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4229 static void clear_gigantic_page(struct page *page,
4230                                 unsigned long addr,
4231                                 unsigned int pages_per_huge_page)
4232 {
4233         int i;
4234         struct page *p = page;
4235
4236         might_sleep();
4237         for (i = 0; i < pages_per_huge_page;
4238              i++, p = mem_map_next(p, page, i)) {
4239                 cond_resched();
4240                 clear_user_highpage(p, addr + i * PAGE_SIZE);
4241         }
4242 }
4243 void clear_huge_page(struct page *page,
4244                      unsigned long addr, unsigned int pages_per_huge_page)
4245 {
4246         int i;
4247
4248         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4249                 clear_gigantic_page(page, addr, pages_per_huge_page);
4250                 return;
4251         }
4252
4253         might_sleep();
4254         for (i = 0; i < pages_per_huge_page; i++) {
4255                 cond_resched();
4256                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4257         }
4258 }
4259
4260 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4261                                     unsigned long addr,
4262                                     struct vm_area_struct *vma,
4263                                     unsigned int pages_per_huge_page)
4264 {
4265         int i;
4266         struct page *dst_base = dst;
4267         struct page *src_base = src;
4268
4269         for (i = 0; i < pages_per_huge_page; ) {
4270                 cond_resched();
4271                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4272
4273                 i++;
4274                 dst = mem_map_next(dst, dst_base, i);
4275                 src = mem_map_next(src, src_base, i);
4276         }
4277 }
4278
4279 void copy_user_huge_page(struct page *dst, struct page *src,
4280                          unsigned long addr, struct vm_area_struct *vma,
4281                          unsigned int pages_per_huge_page)
4282 {
4283         int i;
4284
4285         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4286                 copy_user_gigantic_page(dst, src, addr, vma,
4287                                         pages_per_huge_page);
4288                 return;
4289         }
4290
4291         might_sleep();
4292         for (i = 0; i < pages_per_huge_page; i++) {
4293                 cond_resched();
4294                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4295         }
4296 }
4297 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */