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