Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
[linux-2.6-microblaze.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61 #include <linux/string.h>
62 #include <linux/dma-debug.h>
63 #include <linux/debugfs.h>
64
65 #include <asm/io.h>
66 #include <asm/pgalloc.h>
67 #include <asm/uaccess.h>
68 #include <asm/tlb.h>
69 #include <asm/tlbflush.h>
70 #include <asm/pgtable.h>
71
72 #include "internal.h"
73
74 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
76 #endif
77
78 #ifndef CONFIG_NEED_MULTIPLE_NODES
79 /* use the per-pgdat data instead for discontigmem - mbligh */
80 unsigned long max_mapnr;
81 struct page *mem_map;
82
83 EXPORT_SYMBOL(max_mapnr);
84 EXPORT_SYMBOL(mem_map);
85 #endif
86
87 /*
88  * A number of key systems in x86 including ioremap() rely on the assumption
89  * that high_memory defines the upper bound on direct map memory, then end
90  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
91  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92  * and ZONE_HIGHMEM.
93  */
94 void * high_memory;
95
96 EXPORT_SYMBOL(high_memory);
97
98 /*
99  * Randomize the address space (stacks, mmaps, brk, etc.).
100  *
101  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102  *   as ancient (libc5 based) binaries can segfault. )
103  */
104 int randomize_va_space __read_mostly =
105 #ifdef CONFIG_COMPAT_BRK
106                                         1;
107 #else
108                                         2;
109 #endif
110
111 static int __init disable_randmaps(char *s)
112 {
113         randomize_va_space = 0;
114         return 1;
115 }
116 __setup("norandmaps", disable_randmaps);
117
118 unsigned long zero_pfn __read_mostly;
119 unsigned long highest_memmap_pfn __read_mostly;
120
121 EXPORT_SYMBOL(zero_pfn);
122
123 /*
124  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
125  */
126 static int __init init_zero_pfn(void)
127 {
128         zero_pfn = page_to_pfn(ZERO_PAGE(0));
129         return 0;
130 }
131 core_initcall(init_zero_pfn);
132
133
134 #if defined(SPLIT_RSS_COUNTING)
135
136 void sync_mm_rss(struct mm_struct *mm)
137 {
138         int i;
139
140         for (i = 0; i < NR_MM_COUNTERS; i++) {
141                 if (current->rss_stat.count[i]) {
142                         add_mm_counter(mm, i, current->rss_stat.count[i]);
143                         current->rss_stat.count[i] = 0;
144                 }
145         }
146         current->rss_stat.events = 0;
147 }
148
149 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
150 {
151         struct task_struct *task = current;
152
153         if (likely(task->mm == mm))
154                 task->rss_stat.count[member] += val;
155         else
156                 add_mm_counter(mm, member, val);
157 }
158 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
159 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
160
161 /* sync counter once per 64 page faults */
162 #define TASK_RSS_EVENTS_THRESH  (64)
163 static void check_sync_rss_stat(struct task_struct *task)
164 {
165         if (unlikely(task != current))
166                 return;
167         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
168                 sync_mm_rss(task->mm);
169 }
170 #else /* SPLIT_RSS_COUNTING */
171
172 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
173 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
174
175 static void check_sync_rss_stat(struct task_struct *task)
176 {
177 }
178
179 #endif /* SPLIT_RSS_COUNTING */
180
181 #ifdef HAVE_GENERIC_MMU_GATHER
182
183 static int tlb_next_batch(struct mmu_gather *tlb)
184 {
185         struct mmu_gather_batch *batch;
186
187         batch = tlb->active;
188         if (batch->next) {
189                 tlb->active = batch->next;
190                 return 1;
191         }
192
193         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
194                 return 0;
195
196         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
197         if (!batch)
198                 return 0;
199
200         tlb->batch_count++;
201         batch->next = NULL;
202         batch->nr   = 0;
203         batch->max  = MAX_GATHER_BATCH;
204
205         tlb->active->next = batch;
206         tlb->active = batch;
207
208         return 1;
209 }
210
211 /* tlb_gather_mmu
212  *      Called to initialize an (on-stack) mmu_gather structure for page-table
213  *      tear-down from @mm. The @fullmm argument is used when @mm is without
214  *      users and we're going to destroy the full address space (exit/execve).
215  */
216 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
217 {
218         tlb->mm = mm;
219
220         /* Is it from 0 to ~0? */
221         tlb->fullmm     = !(start | (end+1));
222         tlb->need_flush_all = 0;
223         tlb->start      = start;
224         tlb->end        = end;
225         tlb->need_flush = 0;
226         tlb->local.next = NULL;
227         tlb->local.nr   = 0;
228         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
229         tlb->active     = &tlb->local;
230         tlb->batch_count = 0;
231
232 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
233         tlb->batch = NULL;
234 #endif
235 }
236
237 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
238 {
239         tlb->need_flush = 0;
240         tlb_flush(tlb);
241 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
242         tlb_table_flush(tlb);
243 #endif
244 }
245
246 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
247 {
248         struct mmu_gather_batch *batch;
249
250         for (batch = &tlb->local; batch; batch = batch->next) {
251                 free_pages_and_swap_cache(batch->pages, batch->nr);
252                 batch->nr = 0;
253         }
254         tlb->active = &tlb->local;
255 }
256
257 void tlb_flush_mmu(struct mmu_gather *tlb)
258 {
259         if (!tlb->need_flush)
260                 return;
261         tlb_flush_mmu_tlbonly(tlb);
262         tlb_flush_mmu_free(tlb);
263 }
264
265 /* tlb_finish_mmu
266  *      Called at the end of the shootdown operation to free up any resources
267  *      that were required.
268  */
269 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
270 {
271         struct mmu_gather_batch *batch, *next;
272
273         tlb_flush_mmu(tlb);
274
275         /* keep the page table cache within bounds */
276         check_pgt_cache();
277
278         for (batch = tlb->local.next; batch; batch = next) {
279                 next = batch->next;
280                 free_pages((unsigned long)batch, 0);
281         }
282         tlb->local.next = NULL;
283 }
284
285 /* __tlb_remove_page
286  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
287  *      handling the additional races in SMP caused by other CPUs caching valid
288  *      mappings in their TLBs. Returns the number of free page slots left.
289  *      When out of page slots we must call tlb_flush_mmu().
290  */
291 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
292 {
293         struct mmu_gather_batch *batch;
294
295         VM_BUG_ON(!tlb->need_flush);
296
297         batch = tlb->active;
298         batch->pages[batch->nr++] = page;
299         if (batch->nr == batch->max) {
300                 if (!tlb_next_batch(tlb))
301                         return 0;
302                 batch = tlb->active;
303         }
304         VM_BUG_ON_PAGE(batch->nr > batch->max, page);
305
306         return batch->max - batch->nr;
307 }
308
309 #endif /* HAVE_GENERIC_MMU_GATHER */
310
311 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
312
313 /*
314  * See the comment near struct mmu_table_batch.
315  */
316
317 static void tlb_remove_table_smp_sync(void *arg)
318 {
319         /* Simply deliver the interrupt */
320 }
321
322 static void tlb_remove_table_one(void *table)
323 {
324         /*
325          * This isn't an RCU grace period and hence the page-tables cannot be
326          * assumed to be actually RCU-freed.
327          *
328          * It is however sufficient for software page-table walkers that rely on
329          * IRQ disabling. See the comment near struct mmu_table_batch.
330          */
331         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
332         __tlb_remove_table(table);
333 }
334
335 static void tlb_remove_table_rcu(struct rcu_head *head)
336 {
337         struct mmu_table_batch *batch;
338         int i;
339
340         batch = container_of(head, struct mmu_table_batch, rcu);
341
342         for (i = 0; i < batch->nr; i++)
343                 __tlb_remove_table(batch->tables[i]);
344
345         free_page((unsigned long)batch);
346 }
347
348 void tlb_table_flush(struct mmu_gather *tlb)
349 {
350         struct mmu_table_batch **batch = &tlb->batch;
351
352         if (*batch) {
353                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
354                 *batch = NULL;
355         }
356 }
357
358 void tlb_remove_table(struct mmu_gather *tlb, void *table)
359 {
360         struct mmu_table_batch **batch = &tlb->batch;
361
362         tlb->need_flush = 1;
363
364         /*
365          * When there's less then two users of this mm there cannot be a
366          * concurrent page-table walk.
367          */
368         if (atomic_read(&tlb->mm->mm_users) < 2) {
369                 __tlb_remove_table(table);
370                 return;
371         }
372
373         if (*batch == NULL) {
374                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
375                 if (*batch == NULL) {
376                         tlb_remove_table_one(table);
377                         return;
378                 }
379                 (*batch)->nr = 0;
380         }
381         (*batch)->tables[(*batch)->nr++] = table;
382         if ((*batch)->nr == MAX_TABLE_BATCH)
383                 tlb_table_flush(tlb);
384 }
385
386 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
387
388 /*
389  * Note: this doesn't free the actual pages themselves. That
390  * has been handled earlier when unmapping all the memory regions.
391  */
392 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
393                            unsigned long addr)
394 {
395         pgtable_t token = pmd_pgtable(*pmd);
396         pmd_clear(pmd);
397         pte_free_tlb(tlb, token, addr);
398         atomic_long_dec(&tlb->mm->nr_ptes);
399 }
400
401 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
402                                 unsigned long addr, unsigned long end,
403                                 unsigned long floor, unsigned long ceiling)
404 {
405         pmd_t *pmd;
406         unsigned long next;
407         unsigned long start;
408
409         start = addr;
410         pmd = pmd_offset(pud, addr);
411         do {
412                 next = pmd_addr_end(addr, end);
413                 if (pmd_none_or_clear_bad(pmd))
414                         continue;
415                 free_pte_range(tlb, pmd, addr);
416         } while (pmd++, addr = next, addr != end);
417
418         start &= PUD_MASK;
419         if (start < floor)
420                 return;
421         if (ceiling) {
422                 ceiling &= PUD_MASK;
423                 if (!ceiling)
424                         return;
425         }
426         if (end - 1 > ceiling - 1)
427                 return;
428
429         pmd = pmd_offset(pud, start);
430         pud_clear(pud);
431         pmd_free_tlb(tlb, pmd, start);
432 }
433
434 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
435                                 unsigned long addr, unsigned long end,
436                                 unsigned long floor, unsigned long ceiling)
437 {
438         pud_t *pud;
439         unsigned long next;
440         unsigned long start;
441
442         start = addr;
443         pud = pud_offset(pgd, addr);
444         do {
445                 next = pud_addr_end(addr, end);
446                 if (pud_none_or_clear_bad(pud))
447                         continue;
448                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
449         } while (pud++, addr = next, addr != end);
450
451         start &= PGDIR_MASK;
452         if (start < floor)
453                 return;
454         if (ceiling) {
455                 ceiling &= PGDIR_MASK;
456                 if (!ceiling)
457                         return;
458         }
459         if (end - 1 > ceiling - 1)
460                 return;
461
462         pud = pud_offset(pgd, start);
463         pgd_clear(pgd);
464         pud_free_tlb(tlb, pud, start);
465 }
466
467 /*
468  * This function frees user-level page tables of a process.
469  */
470 void free_pgd_range(struct mmu_gather *tlb,
471                         unsigned long addr, unsigned long end,
472                         unsigned long floor, unsigned long ceiling)
473 {
474         pgd_t *pgd;
475         unsigned long next;
476
477         /*
478          * The next few lines have given us lots of grief...
479          *
480          * Why are we testing PMD* at this top level?  Because often
481          * there will be no work to do at all, and we'd prefer not to
482          * go all the way down to the bottom just to discover that.
483          *
484          * Why all these "- 1"s?  Because 0 represents both the bottom
485          * of the address space and the top of it (using -1 for the
486          * top wouldn't help much: the masks would do the wrong thing).
487          * The rule is that addr 0 and floor 0 refer to the bottom of
488          * the address space, but end 0 and ceiling 0 refer to the top
489          * Comparisons need to use "end - 1" and "ceiling - 1" (though
490          * that end 0 case should be mythical).
491          *
492          * Wherever addr is brought up or ceiling brought down, we must
493          * be careful to reject "the opposite 0" before it confuses the
494          * subsequent tests.  But what about where end is brought down
495          * by PMD_SIZE below? no, end can't go down to 0 there.
496          *
497          * Whereas we round start (addr) and ceiling down, by different
498          * masks at different levels, in order to test whether a table
499          * now has no other vmas using it, so can be freed, we don't
500          * bother to round floor or end up - the tests don't need that.
501          */
502
503         addr &= PMD_MASK;
504         if (addr < floor) {
505                 addr += PMD_SIZE;
506                 if (!addr)
507                         return;
508         }
509         if (ceiling) {
510                 ceiling &= PMD_MASK;
511                 if (!ceiling)
512                         return;
513         }
514         if (end - 1 > ceiling - 1)
515                 end -= PMD_SIZE;
516         if (addr > end - 1)
517                 return;
518
519         pgd = pgd_offset(tlb->mm, addr);
520         do {
521                 next = pgd_addr_end(addr, end);
522                 if (pgd_none_or_clear_bad(pgd))
523                         continue;
524                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
525         } while (pgd++, addr = next, addr != end);
526 }
527
528 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
529                 unsigned long floor, unsigned long ceiling)
530 {
531         while (vma) {
532                 struct vm_area_struct *next = vma->vm_next;
533                 unsigned long addr = vma->vm_start;
534
535                 /*
536                  * Hide vma from rmap and truncate_pagecache before freeing
537                  * pgtables
538                  */
539                 unlink_anon_vmas(vma);
540                 unlink_file_vma(vma);
541
542                 if (is_vm_hugetlb_page(vma)) {
543                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
544                                 floor, next? next->vm_start: ceiling);
545                 } else {
546                         /*
547                          * Optimization: gather nearby vmas into one call down
548                          */
549                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
550                                && !is_vm_hugetlb_page(next)) {
551                                 vma = next;
552                                 next = vma->vm_next;
553                                 unlink_anon_vmas(vma);
554                                 unlink_file_vma(vma);
555                         }
556                         free_pgd_range(tlb, addr, vma->vm_end,
557                                 floor, next? next->vm_start: ceiling);
558                 }
559                 vma = next;
560         }
561 }
562
563 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
564                 pmd_t *pmd, unsigned long address)
565 {
566         spinlock_t *ptl;
567         pgtable_t new = pte_alloc_one(mm, address);
568         int wait_split_huge_page;
569         if (!new)
570                 return -ENOMEM;
571
572         /*
573          * Ensure all pte setup (eg. pte page lock and page clearing) are
574          * visible before the pte is made visible to other CPUs by being
575          * put into page tables.
576          *
577          * The other side of the story is the pointer chasing in the page
578          * table walking code (when walking the page table without locking;
579          * ie. most of the time). Fortunately, these data accesses consist
580          * of a chain of data-dependent loads, meaning most CPUs (alpha
581          * being the notable exception) will already guarantee loads are
582          * seen in-order. See the alpha page table accessors for the
583          * smp_read_barrier_depends() barriers in page table walking code.
584          */
585         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
586
587         ptl = pmd_lock(mm, pmd);
588         wait_split_huge_page = 0;
589         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
590                 atomic_long_inc(&mm->nr_ptes);
591                 pmd_populate(mm, pmd, new);
592                 new = NULL;
593         } else if (unlikely(pmd_trans_splitting(*pmd)))
594                 wait_split_huge_page = 1;
595         spin_unlock(ptl);
596         if (new)
597                 pte_free(mm, new);
598         if (wait_split_huge_page)
599                 wait_split_huge_page(vma->anon_vma, pmd);
600         return 0;
601 }
602
603 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
604 {
605         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
606         if (!new)
607                 return -ENOMEM;
608
609         smp_wmb(); /* See comment in __pte_alloc */
610
611         spin_lock(&init_mm.page_table_lock);
612         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
613                 pmd_populate_kernel(&init_mm, pmd, new);
614                 new = NULL;
615         } else
616                 VM_BUG_ON(pmd_trans_splitting(*pmd));
617         spin_unlock(&init_mm.page_table_lock);
618         if (new)
619                 pte_free_kernel(&init_mm, new);
620         return 0;
621 }
622
623 static inline void init_rss_vec(int *rss)
624 {
625         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
626 }
627
628 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
629 {
630         int i;
631
632         if (current->mm == mm)
633                 sync_mm_rss(mm);
634         for (i = 0; i < NR_MM_COUNTERS; i++)
635                 if (rss[i])
636                         add_mm_counter(mm, i, rss[i]);
637 }
638
639 /*
640  * This function is called to print an error when a bad pte
641  * is found. For example, we might have a PFN-mapped pte in
642  * a region that doesn't allow it.
643  *
644  * The calling function must still handle the error.
645  */
646 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
647                           pte_t pte, struct page *page)
648 {
649         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
650         pud_t *pud = pud_offset(pgd, addr);
651         pmd_t *pmd = pmd_offset(pud, addr);
652         struct address_space *mapping;
653         pgoff_t index;
654         static unsigned long resume;
655         static unsigned long nr_shown;
656         static unsigned long nr_unshown;
657
658         /*
659          * Allow a burst of 60 reports, then keep quiet for that minute;
660          * or allow a steady drip of one report per second.
661          */
662         if (nr_shown == 60) {
663                 if (time_before(jiffies, resume)) {
664                         nr_unshown++;
665                         return;
666                 }
667                 if (nr_unshown) {
668                         printk(KERN_ALERT
669                                 "BUG: Bad page map: %lu messages suppressed\n",
670                                 nr_unshown);
671                         nr_unshown = 0;
672                 }
673                 nr_shown = 0;
674         }
675         if (nr_shown++ == 0)
676                 resume = jiffies + 60 * HZ;
677
678         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
679         index = linear_page_index(vma, addr);
680
681         printk(KERN_ALERT
682                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
683                 current->comm,
684                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
685         if (page)
686                 dump_page(page, "bad pte");
687         printk(KERN_ALERT
688                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
689                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
690         /*
691          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
692          */
693         if (vma->vm_ops)
694                 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
695                        vma->vm_ops->fault);
696         if (vma->vm_file)
697                 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
698                        vma->vm_file->f_op->mmap);
699         dump_stack();
700         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
701 }
702
703 /*
704  * vm_normal_page -- This function gets the "struct page" associated with a pte.
705  *
706  * "Special" mappings do not wish to be associated with a "struct page" (either
707  * it doesn't exist, or it exists but they don't want to touch it). In this
708  * case, NULL is returned here. "Normal" mappings do have a struct page.
709  *
710  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
711  * pte bit, in which case this function is trivial. Secondly, an architecture
712  * may not have a spare pte bit, which requires a more complicated scheme,
713  * described below.
714  *
715  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
716  * special mapping (even if there are underlying and valid "struct pages").
717  * COWed pages of a VM_PFNMAP are always normal.
718  *
719  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
720  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
721  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
722  * mapping will always honor the rule
723  *
724  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
725  *
726  * And for normal mappings this is false.
727  *
728  * This restricts such mappings to be a linear translation from virtual address
729  * to pfn. To get around this restriction, we allow arbitrary mappings so long
730  * as the vma is not a COW mapping; in that case, we know that all ptes are
731  * special (because none can have been COWed).
732  *
733  *
734  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
735  *
736  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
737  * page" backing, however the difference is that _all_ pages with a struct
738  * page (that is, those where pfn_valid is true) are refcounted and considered
739  * normal pages by the VM. The disadvantage is that pages are refcounted
740  * (which can be slower and simply not an option for some PFNMAP users). The
741  * advantage is that we don't have to follow the strict linearity rule of
742  * PFNMAP mappings in order to support COWable mappings.
743  *
744  */
745 #ifdef __HAVE_ARCH_PTE_SPECIAL
746 # define HAVE_PTE_SPECIAL 1
747 #else
748 # define HAVE_PTE_SPECIAL 0
749 #endif
750 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
751                                 pte_t pte)
752 {
753         unsigned long pfn = pte_pfn(pte);
754
755         if (HAVE_PTE_SPECIAL) {
756                 if (likely(!pte_special(pte)))
757                         goto check_pfn;
758                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
759                         return NULL;
760                 if (!is_zero_pfn(pfn))
761                         print_bad_pte(vma, addr, pte, NULL);
762                 return NULL;
763         }
764
765         /* !HAVE_PTE_SPECIAL case follows: */
766
767         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
768                 if (vma->vm_flags & VM_MIXEDMAP) {
769                         if (!pfn_valid(pfn))
770                                 return NULL;
771                         goto out;
772                 } else {
773                         unsigned long off;
774                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
775                         if (pfn == vma->vm_pgoff + off)
776                                 return NULL;
777                         if (!is_cow_mapping(vma->vm_flags))
778                                 return NULL;
779                 }
780         }
781
782         if (is_zero_pfn(pfn))
783                 return NULL;
784 check_pfn:
785         if (unlikely(pfn > highest_memmap_pfn)) {
786                 print_bad_pte(vma, addr, pte, NULL);
787                 return NULL;
788         }
789
790         /*
791          * NOTE! We still have PageReserved() pages in the page tables.
792          * eg. VDSO mappings can cause them to exist.
793          */
794 out:
795         return pfn_to_page(pfn);
796 }
797
798 /*
799  * copy one vm_area from one task to the other. Assumes the page tables
800  * already present in the new task to be cleared in the whole range
801  * covered by this vma.
802  */
803
804 static inline unsigned long
805 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
806                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
807                 unsigned long addr, int *rss)
808 {
809         unsigned long vm_flags = vma->vm_flags;
810         pte_t pte = *src_pte;
811         struct page *page;
812
813         /* pte contains position in swap or file, so copy. */
814         if (unlikely(!pte_present(pte))) {
815                 if (!pte_file(pte)) {
816                         swp_entry_t entry = pte_to_swp_entry(pte);
817
818                         if (swap_duplicate(entry) < 0)
819                                 return entry.val;
820
821                         /* make sure dst_mm is on swapoff's mmlist. */
822                         if (unlikely(list_empty(&dst_mm->mmlist))) {
823                                 spin_lock(&mmlist_lock);
824                                 if (list_empty(&dst_mm->mmlist))
825                                         list_add(&dst_mm->mmlist,
826                                                  &src_mm->mmlist);
827                                 spin_unlock(&mmlist_lock);
828                         }
829                         if (likely(!non_swap_entry(entry)))
830                                 rss[MM_SWAPENTS]++;
831                         else if (is_migration_entry(entry)) {
832                                 page = migration_entry_to_page(entry);
833
834                                 if (PageAnon(page))
835                                         rss[MM_ANONPAGES]++;
836                                 else
837                                         rss[MM_FILEPAGES]++;
838
839                                 if (is_write_migration_entry(entry) &&
840                                     is_cow_mapping(vm_flags)) {
841                                         /*
842                                          * COW mappings require pages in both
843                                          * parent and child to be set to read.
844                                          */
845                                         make_migration_entry_read(&entry);
846                                         pte = swp_entry_to_pte(entry);
847                                         if (pte_swp_soft_dirty(*src_pte))
848                                                 pte = pte_swp_mksoft_dirty(pte);
849                                         set_pte_at(src_mm, addr, src_pte, pte);
850                                 }
851                         }
852                 }
853                 goto out_set_pte;
854         }
855
856         /*
857          * If it's a COW mapping, write protect it both
858          * in the parent and the child
859          */
860         if (is_cow_mapping(vm_flags)) {
861                 ptep_set_wrprotect(src_mm, addr, src_pte);
862                 pte = pte_wrprotect(pte);
863         }
864
865         /*
866          * If it's a shared mapping, mark it clean in
867          * the child
868          */
869         if (vm_flags & VM_SHARED)
870                 pte = pte_mkclean(pte);
871         pte = pte_mkold(pte);
872
873         page = vm_normal_page(vma, addr, pte);
874         if (page) {
875                 get_page(page);
876                 page_dup_rmap(page);
877                 if (PageAnon(page))
878                         rss[MM_ANONPAGES]++;
879                 else
880                         rss[MM_FILEPAGES]++;
881         }
882
883 out_set_pte:
884         set_pte_at(dst_mm, addr, dst_pte, pte);
885         return 0;
886 }
887
888 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
889                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
890                    unsigned long addr, unsigned long end)
891 {
892         pte_t *orig_src_pte, *orig_dst_pte;
893         pte_t *src_pte, *dst_pte;
894         spinlock_t *src_ptl, *dst_ptl;
895         int progress = 0;
896         int rss[NR_MM_COUNTERS];
897         swp_entry_t entry = (swp_entry_t){0};
898
899 again:
900         init_rss_vec(rss);
901
902         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
903         if (!dst_pte)
904                 return -ENOMEM;
905         src_pte = pte_offset_map(src_pmd, addr);
906         src_ptl = pte_lockptr(src_mm, src_pmd);
907         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
908         orig_src_pte = src_pte;
909         orig_dst_pte = dst_pte;
910         arch_enter_lazy_mmu_mode();
911
912         do {
913                 /*
914                  * We are holding two locks at this point - either of them
915                  * could generate latencies in another task on another CPU.
916                  */
917                 if (progress >= 32) {
918                         progress = 0;
919                         if (need_resched() ||
920                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
921                                 break;
922                 }
923                 if (pte_none(*src_pte)) {
924                         progress++;
925                         continue;
926                 }
927                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
928                                                         vma, addr, rss);
929                 if (entry.val)
930                         break;
931                 progress += 8;
932         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
933
934         arch_leave_lazy_mmu_mode();
935         spin_unlock(src_ptl);
936         pte_unmap(orig_src_pte);
937         add_mm_rss_vec(dst_mm, rss);
938         pte_unmap_unlock(orig_dst_pte, dst_ptl);
939         cond_resched();
940
941         if (entry.val) {
942                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
943                         return -ENOMEM;
944                 progress = 0;
945         }
946         if (addr != end)
947                 goto again;
948         return 0;
949 }
950
951 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
952                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
953                 unsigned long addr, unsigned long end)
954 {
955         pmd_t *src_pmd, *dst_pmd;
956         unsigned long next;
957
958         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
959         if (!dst_pmd)
960                 return -ENOMEM;
961         src_pmd = pmd_offset(src_pud, addr);
962         do {
963                 next = pmd_addr_end(addr, end);
964                 if (pmd_trans_huge(*src_pmd)) {
965                         int err;
966                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
967                         err = copy_huge_pmd(dst_mm, src_mm,
968                                             dst_pmd, src_pmd, addr, vma);
969                         if (err == -ENOMEM)
970                                 return -ENOMEM;
971                         if (!err)
972                                 continue;
973                         /* fall through */
974                 }
975                 if (pmd_none_or_clear_bad(src_pmd))
976                         continue;
977                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
978                                                 vma, addr, next))
979                         return -ENOMEM;
980         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
981         return 0;
982 }
983
984 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
985                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
986                 unsigned long addr, unsigned long end)
987 {
988         pud_t *src_pud, *dst_pud;
989         unsigned long next;
990
991         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
992         if (!dst_pud)
993                 return -ENOMEM;
994         src_pud = pud_offset(src_pgd, addr);
995         do {
996                 next = pud_addr_end(addr, end);
997                 if (pud_none_or_clear_bad(src_pud))
998                         continue;
999                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1000                                                 vma, addr, next))
1001                         return -ENOMEM;
1002         } while (dst_pud++, src_pud++, addr = next, addr != end);
1003         return 0;
1004 }
1005
1006 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1007                 struct vm_area_struct *vma)
1008 {
1009         pgd_t *src_pgd, *dst_pgd;
1010         unsigned long next;
1011         unsigned long addr = vma->vm_start;
1012         unsigned long end = vma->vm_end;
1013         unsigned long mmun_start;       /* For mmu_notifiers */
1014         unsigned long mmun_end;         /* For mmu_notifiers */
1015         bool is_cow;
1016         int ret;
1017
1018         /*
1019          * Don't copy ptes where a page fault will fill them correctly.
1020          * Fork becomes much lighter when there are big shared or private
1021          * readonly mappings. The tradeoff is that copy_page_range is more
1022          * efficient than faulting.
1023          */
1024         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1025                                VM_PFNMAP | VM_MIXEDMAP))) {
1026                 if (!vma->anon_vma)
1027                         return 0;
1028         }
1029
1030         if (is_vm_hugetlb_page(vma))
1031                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1032
1033         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1034                 /*
1035                  * We do not free on error cases below as remove_vma
1036                  * gets called on error from higher level routine
1037                  */
1038                 ret = track_pfn_copy(vma);
1039                 if (ret)
1040                         return ret;
1041         }
1042
1043         /*
1044          * We need to invalidate the secondary MMU mappings only when
1045          * there could be a permission downgrade on the ptes of the
1046          * parent mm. And a permission downgrade will only happen if
1047          * is_cow_mapping() returns true.
1048          */
1049         is_cow = is_cow_mapping(vma->vm_flags);
1050         mmun_start = addr;
1051         mmun_end   = end;
1052         if (is_cow)
1053                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1054                                                     mmun_end);
1055
1056         ret = 0;
1057         dst_pgd = pgd_offset(dst_mm, addr);
1058         src_pgd = pgd_offset(src_mm, addr);
1059         do {
1060                 next = pgd_addr_end(addr, end);
1061                 if (pgd_none_or_clear_bad(src_pgd))
1062                         continue;
1063                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1064                                             vma, addr, next))) {
1065                         ret = -ENOMEM;
1066                         break;
1067                 }
1068         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1069
1070         if (is_cow)
1071                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1072         return ret;
1073 }
1074
1075 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1076                                 struct vm_area_struct *vma, pmd_t *pmd,
1077                                 unsigned long addr, unsigned long end,
1078                                 struct zap_details *details)
1079 {
1080         struct mm_struct *mm = tlb->mm;
1081         int force_flush = 0;
1082         int rss[NR_MM_COUNTERS];
1083         spinlock_t *ptl;
1084         pte_t *start_pte;
1085         pte_t *pte;
1086
1087 again:
1088         init_rss_vec(rss);
1089         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1090         pte = start_pte;
1091         arch_enter_lazy_mmu_mode();
1092         do {
1093                 pte_t ptent = *pte;
1094                 if (pte_none(ptent)) {
1095                         continue;
1096                 }
1097
1098                 if (pte_present(ptent)) {
1099                         struct page *page;
1100
1101                         page = vm_normal_page(vma, addr, ptent);
1102                         if (unlikely(details) && page) {
1103                                 /*
1104                                  * unmap_shared_mapping_pages() wants to
1105                                  * invalidate cache without truncating:
1106                                  * unmap shared but keep private pages.
1107                                  */
1108                                 if (details->check_mapping &&
1109                                     details->check_mapping != page->mapping)
1110                                         continue;
1111                                 /*
1112                                  * Each page->index must be checked when
1113                                  * invalidating or truncating nonlinear.
1114                                  */
1115                                 if (details->nonlinear_vma &&
1116                                     (page->index < details->first_index ||
1117                                      page->index > details->last_index))
1118                                         continue;
1119                         }
1120                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1121                                                         tlb->fullmm);
1122                         tlb_remove_tlb_entry(tlb, pte, addr);
1123                         if (unlikely(!page))
1124                                 continue;
1125                         if (unlikely(details) && details->nonlinear_vma
1126                             && linear_page_index(details->nonlinear_vma,
1127                                                 addr) != page->index) {
1128                                 pte_t ptfile = pgoff_to_pte(page->index);
1129                                 if (pte_soft_dirty(ptent))
1130                                         ptfile = pte_file_mksoft_dirty(ptfile);
1131                                 set_pte_at(mm, addr, pte, ptfile);
1132                         }
1133                         if (PageAnon(page))
1134                                 rss[MM_ANONPAGES]--;
1135                         else {
1136                                 if (pte_dirty(ptent)) {
1137                                         force_flush = 1;
1138                                         set_page_dirty(page);
1139                                 }
1140                                 if (pte_young(ptent) &&
1141                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1142                                         mark_page_accessed(page);
1143                                 rss[MM_FILEPAGES]--;
1144                         }
1145                         page_remove_rmap(page);
1146                         if (unlikely(page_mapcount(page) < 0))
1147                                 print_bad_pte(vma, addr, ptent, page);
1148                         if (unlikely(!__tlb_remove_page(tlb, page))) {
1149                                 force_flush = 1;
1150                                 break;
1151                         }
1152                         continue;
1153                 }
1154                 /*
1155                  * If details->check_mapping, we leave swap entries;
1156                  * if details->nonlinear_vma, we leave file entries.
1157                  */
1158                 if (unlikely(details))
1159                         continue;
1160                 if (pte_file(ptent)) {
1161                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1162                                 print_bad_pte(vma, addr, ptent, NULL);
1163                 } else {
1164                         swp_entry_t entry = pte_to_swp_entry(ptent);
1165
1166                         if (!non_swap_entry(entry))
1167                                 rss[MM_SWAPENTS]--;
1168                         else if (is_migration_entry(entry)) {
1169                                 struct page *page;
1170
1171                                 page = migration_entry_to_page(entry);
1172
1173                                 if (PageAnon(page))
1174                                         rss[MM_ANONPAGES]--;
1175                                 else
1176                                         rss[MM_FILEPAGES]--;
1177                         }
1178                         if (unlikely(!free_swap_and_cache(entry)))
1179                                 print_bad_pte(vma, addr, ptent, NULL);
1180                 }
1181                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1182         } while (pte++, addr += PAGE_SIZE, addr != end);
1183
1184         add_mm_rss_vec(mm, rss);
1185         arch_leave_lazy_mmu_mode();
1186
1187         /* Do the actual TLB flush before dropping ptl */
1188         if (force_flush) {
1189                 unsigned long old_end;
1190
1191                 /*
1192                  * Flush the TLB just for the previous segment,
1193                  * then update the range to be the remaining
1194                  * TLB range.
1195                  */
1196                 old_end = tlb->end;
1197                 tlb->end = addr;
1198                 tlb_flush_mmu_tlbonly(tlb);
1199                 tlb->start = addr;
1200                 tlb->end = old_end;
1201         }
1202         pte_unmap_unlock(start_pte, ptl);
1203
1204         /*
1205          * If we forced a TLB flush (either due to running out of
1206          * batch buffers or because we needed to flush dirty TLB
1207          * entries before releasing the ptl), free the batched
1208          * memory too. Restart if we didn't do everything.
1209          */
1210         if (force_flush) {
1211                 force_flush = 0;
1212                 tlb_flush_mmu_free(tlb);
1213
1214                 if (addr != end)
1215                         goto again;
1216         }
1217
1218         return addr;
1219 }
1220
1221 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1222                                 struct vm_area_struct *vma, pud_t *pud,
1223                                 unsigned long addr, unsigned long end,
1224                                 struct zap_details *details)
1225 {
1226         pmd_t *pmd;
1227         unsigned long next;
1228
1229         pmd = pmd_offset(pud, addr);
1230         do {
1231                 next = pmd_addr_end(addr, end);
1232                 if (pmd_trans_huge(*pmd)) {
1233                         if (next - addr != HPAGE_PMD_SIZE) {
1234 #ifdef CONFIG_DEBUG_VM
1235                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1236                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1237                                                 __func__, addr, end,
1238                                                 vma->vm_start,
1239                                                 vma->vm_end);
1240                                         BUG();
1241                                 }
1242 #endif
1243                                 split_huge_page_pmd(vma, addr, pmd);
1244                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1245                                 goto next;
1246                         /* fall through */
1247                 }
1248                 /*
1249                  * Here there can be other concurrent MADV_DONTNEED or
1250                  * trans huge page faults running, and if the pmd is
1251                  * none or trans huge it can change under us. This is
1252                  * because MADV_DONTNEED holds the mmap_sem in read
1253                  * mode.
1254                  */
1255                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1256                         goto next;
1257                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1258 next:
1259                 cond_resched();
1260         } while (pmd++, addr = next, addr != end);
1261
1262         return addr;
1263 }
1264
1265 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1266                                 struct vm_area_struct *vma, pgd_t *pgd,
1267                                 unsigned long addr, unsigned long end,
1268                                 struct zap_details *details)
1269 {
1270         pud_t *pud;
1271         unsigned long next;
1272
1273         pud = pud_offset(pgd, addr);
1274         do {
1275                 next = pud_addr_end(addr, end);
1276                 if (pud_none_or_clear_bad(pud))
1277                         continue;
1278                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1279         } while (pud++, addr = next, addr != end);
1280
1281         return addr;
1282 }
1283
1284 static void unmap_page_range(struct mmu_gather *tlb,
1285                              struct vm_area_struct *vma,
1286                              unsigned long addr, unsigned long end,
1287                              struct zap_details *details)
1288 {
1289         pgd_t *pgd;
1290         unsigned long next;
1291
1292         if (details && !details->check_mapping && !details->nonlinear_vma)
1293                 details = NULL;
1294
1295         BUG_ON(addr >= end);
1296         tlb_start_vma(tlb, vma);
1297         pgd = pgd_offset(vma->vm_mm, addr);
1298         do {
1299                 next = pgd_addr_end(addr, end);
1300                 if (pgd_none_or_clear_bad(pgd))
1301                         continue;
1302                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1303         } while (pgd++, addr = next, addr != end);
1304         tlb_end_vma(tlb, vma);
1305 }
1306
1307
1308 static void unmap_single_vma(struct mmu_gather *tlb,
1309                 struct vm_area_struct *vma, unsigned long start_addr,
1310                 unsigned long end_addr,
1311                 struct zap_details *details)
1312 {
1313         unsigned long start = max(vma->vm_start, start_addr);
1314         unsigned long end;
1315
1316         if (start >= vma->vm_end)
1317                 return;
1318         end = min(vma->vm_end, end_addr);
1319         if (end <= vma->vm_start)
1320                 return;
1321
1322         if (vma->vm_file)
1323                 uprobe_munmap(vma, start, end);
1324
1325         if (unlikely(vma->vm_flags & VM_PFNMAP))
1326                 untrack_pfn(vma, 0, 0);
1327
1328         if (start != end) {
1329                 if (unlikely(is_vm_hugetlb_page(vma))) {
1330                         /*
1331                          * It is undesirable to test vma->vm_file as it
1332                          * should be non-null for valid hugetlb area.
1333                          * However, vm_file will be NULL in the error
1334                          * cleanup path of mmap_region. When
1335                          * hugetlbfs ->mmap method fails,
1336                          * mmap_region() nullifies vma->vm_file
1337                          * before calling this function to clean up.
1338                          * Since no pte has actually been setup, it is
1339                          * safe to do nothing in this case.
1340                          */
1341                         if (vma->vm_file) {
1342                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1343                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1344                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1345                         }
1346                 } else
1347                         unmap_page_range(tlb, vma, start, end, details);
1348         }
1349 }
1350
1351 /**
1352  * unmap_vmas - unmap a range of memory covered by a list of vma's
1353  * @tlb: address of the caller's struct mmu_gather
1354  * @vma: the starting vma
1355  * @start_addr: virtual address at which to start unmapping
1356  * @end_addr: virtual address at which to end unmapping
1357  *
1358  * Unmap all pages in the vma list.
1359  *
1360  * Only addresses between `start' and `end' will be unmapped.
1361  *
1362  * The VMA list must be sorted in ascending virtual address order.
1363  *
1364  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1365  * range after unmap_vmas() returns.  So the only responsibility here is to
1366  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1367  * drops the lock and schedules.
1368  */
1369 void unmap_vmas(struct mmu_gather *tlb,
1370                 struct vm_area_struct *vma, unsigned long start_addr,
1371                 unsigned long end_addr)
1372 {
1373         struct mm_struct *mm = vma->vm_mm;
1374
1375         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1376         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1377                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1378         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1379 }
1380
1381 /**
1382  * zap_page_range - remove user pages in a given range
1383  * @vma: vm_area_struct holding the applicable pages
1384  * @start: starting address of pages to zap
1385  * @size: number of bytes to zap
1386  * @details: details of nonlinear truncation or shared cache invalidation
1387  *
1388  * Caller must protect the VMA list
1389  */
1390 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1391                 unsigned long size, struct zap_details *details)
1392 {
1393         struct mm_struct *mm = vma->vm_mm;
1394         struct mmu_gather tlb;
1395         unsigned long end = start + size;
1396
1397         lru_add_drain();
1398         tlb_gather_mmu(&tlb, mm, start, end);
1399         update_hiwater_rss(mm);
1400         mmu_notifier_invalidate_range_start(mm, start, end);
1401         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1402                 unmap_single_vma(&tlb, vma, start, end, details);
1403         mmu_notifier_invalidate_range_end(mm, start, end);
1404         tlb_finish_mmu(&tlb, start, end);
1405 }
1406
1407 /**
1408  * zap_page_range_single - remove user pages in a given range
1409  * @vma: vm_area_struct holding the applicable pages
1410  * @address: starting address of pages to zap
1411  * @size: number of bytes to zap
1412  * @details: details of nonlinear truncation or shared cache invalidation
1413  *
1414  * The range must fit into one VMA.
1415  */
1416 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1417                 unsigned long size, struct zap_details *details)
1418 {
1419         struct mm_struct *mm = vma->vm_mm;
1420         struct mmu_gather tlb;
1421         unsigned long end = address + size;
1422
1423         lru_add_drain();
1424         tlb_gather_mmu(&tlb, mm, address, end);
1425         update_hiwater_rss(mm);
1426         mmu_notifier_invalidate_range_start(mm, address, end);
1427         unmap_single_vma(&tlb, vma, address, end, details);
1428         mmu_notifier_invalidate_range_end(mm, address, end);
1429         tlb_finish_mmu(&tlb, address, end);
1430 }
1431
1432 /**
1433  * zap_vma_ptes - remove ptes mapping the vma
1434  * @vma: vm_area_struct holding ptes to be zapped
1435  * @address: starting address of pages to zap
1436  * @size: number of bytes to zap
1437  *
1438  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1439  *
1440  * The entire address range must be fully contained within the vma.
1441  *
1442  * Returns 0 if successful.
1443  */
1444 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1445                 unsigned long size)
1446 {
1447         if (address < vma->vm_start || address + size > vma->vm_end ||
1448                         !(vma->vm_flags & VM_PFNMAP))
1449                 return -1;
1450         zap_page_range_single(vma, address, size, NULL);
1451         return 0;
1452 }
1453 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1454
1455 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1456                         spinlock_t **ptl)
1457 {
1458         pgd_t * pgd = pgd_offset(mm, addr);
1459         pud_t * pud = pud_alloc(mm, pgd, addr);
1460         if (pud) {
1461                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1462                 if (pmd) {
1463                         VM_BUG_ON(pmd_trans_huge(*pmd));
1464                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1465                 }
1466         }
1467         return NULL;
1468 }
1469
1470 /*
1471  * This is the old fallback for page remapping.
1472  *
1473  * For historical reasons, it only allows reserved pages. Only
1474  * old drivers should use this, and they needed to mark their
1475  * pages reserved for the old functions anyway.
1476  */
1477 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1478                         struct page *page, pgprot_t prot)
1479 {
1480         struct mm_struct *mm = vma->vm_mm;
1481         int retval;
1482         pte_t *pte;
1483         spinlock_t *ptl;
1484
1485         retval = -EINVAL;
1486         if (PageAnon(page))
1487                 goto out;
1488         retval = -ENOMEM;
1489         flush_dcache_page(page);
1490         pte = get_locked_pte(mm, addr, &ptl);
1491         if (!pte)
1492                 goto out;
1493         retval = -EBUSY;
1494         if (!pte_none(*pte))
1495                 goto out_unlock;
1496
1497         /* Ok, finally just insert the thing.. */
1498         get_page(page);
1499         inc_mm_counter_fast(mm, MM_FILEPAGES);
1500         page_add_file_rmap(page);
1501         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1502
1503         retval = 0;
1504         pte_unmap_unlock(pte, ptl);
1505         return retval;
1506 out_unlock:
1507         pte_unmap_unlock(pte, ptl);
1508 out:
1509         return retval;
1510 }
1511
1512 /**
1513  * vm_insert_page - insert single page into user vma
1514  * @vma: user vma to map to
1515  * @addr: target user address of this page
1516  * @page: source kernel page
1517  *
1518  * This allows drivers to insert individual pages they've allocated
1519  * into a user vma.
1520  *
1521  * The page has to be a nice clean _individual_ kernel allocation.
1522  * If you allocate a compound page, you need to have marked it as
1523  * such (__GFP_COMP), or manually just split the page up yourself
1524  * (see split_page()).
1525  *
1526  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1527  * took an arbitrary page protection parameter. This doesn't allow
1528  * that. Your vma protection will have to be set up correctly, which
1529  * means that if you want a shared writable mapping, you'd better
1530  * ask for a shared writable mapping!
1531  *
1532  * The page does not need to be reserved.
1533  *
1534  * Usually this function is called from f_op->mmap() handler
1535  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1536  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1537  * function from other places, for example from page-fault handler.
1538  */
1539 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1540                         struct page *page)
1541 {
1542         if (addr < vma->vm_start || addr >= vma->vm_end)
1543                 return -EFAULT;
1544         if (!page_count(page))
1545                 return -EINVAL;
1546         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1547                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1548                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1549                 vma->vm_flags |= VM_MIXEDMAP;
1550         }
1551         return insert_page(vma, addr, page, vma->vm_page_prot);
1552 }
1553 EXPORT_SYMBOL(vm_insert_page);
1554
1555 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1556                         unsigned long pfn, pgprot_t prot)
1557 {
1558         struct mm_struct *mm = vma->vm_mm;
1559         int retval;
1560         pte_t *pte, entry;
1561         spinlock_t *ptl;
1562
1563         retval = -ENOMEM;
1564         pte = get_locked_pte(mm, addr, &ptl);
1565         if (!pte)
1566                 goto out;
1567         retval = -EBUSY;
1568         if (!pte_none(*pte))
1569                 goto out_unlock;
1570
1571         /* Ok, finally just insert the thing.. */
1572         entry = pte_mkspecial(pfn_pte(pfn, prot));
1573         set_pte_at(mm, addr, pte, entry);
1574         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1575
1576         retval = 0;
1577 out_unlock:
1578         pte_unmap_unlock(pte, ptl);
1579 out:
1580         return retval;
1581 }
1582
1583 /**
1584  * vm_insert_pfn - insert single pfn into user vma
1585  * @vma: user vma to map to
1586  * @addr: target user address of this page
1587  * @pfn: source kernel pfn
1588  *
1589  * Similar to vm_insert_page, this allows drivers to insert individual pages
1590  * they've allocated into a user vma. Same comments apply.
1591  *
1592  * This function should only be called from a vm_ops->fault handler, and
1593  * in that case the handler should return NULL.
1594  *
1595  * vma cannot be a COW mapping.
1596  *
1597  * As this is called only for pages that do not currently exist, we
1598  * do not need to flush old virtual caches or the TLB.
1599  */
1600 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1601                         unsigned long pfn)
1602 {
1603         int ret;
1604         pgprot_t pgprot = vma->vm_page_prot;
1605         /*
1606          * Technically, architectures with pte_special can avoid all these
1607          * restrictions (same for remap_pfn_range).  However we would like
1608          * consistency in testing and feature parity among all, so we should
1609          * try to keep these invariants in place for everybody.
1610          */
1611         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1612         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1613                                                 (VM_PFNMAP|VM_MIXEDMAP));
1614         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1615         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1616
1617         if (addr < vma->vm_start || addr >= vma->vm_end)
1618                 return -EFAULT;
1619         if (track_pfn_insert(vma, &pgprot, pfn))
1620                 return -EINVAL;
1621
1622         ret = insert_pfn(vma, addr, pfn, pgprot);
1623
1624         return ret;
1625 }
1626 EXPORT_SYMBOL(vm_insert_pfn);
1627
1628 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1629                         unsigned long pfn)
1630 {
1631         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1632
1633         if (addr < vma->vm_start || addr >= vma->vm_end)
1634                 return -EFAULT;
1635
1636         /*
1637          * If we don't have pte special, then we have to use the pfn_valid()
1638          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1639          * refcount the page if pfn_valid is true (hence insert_page rather
1640          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1641          * without pte special, it would there be refcounted as a normal page.
1642          */
1643         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1644                 struct page *page;
1645
1646                 page = pfn_to_page(pfn);
1647                 return insert_page(vma, addr, page, vma->vm_page_prot);
1648         }
1649         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1650 }
1651 EXPORT_SYMBOL(vm_insert_mixed);
1652
1653 /*
1654  * maps a range of physical memory into the requested pages. the old
1655  * mappings are removed. any references to nonexistent pages results
1656  * in null mappings (currently treated as "copy-on-access")
1657  */
1658 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1659                         unsigned long addr, unsigned long end,
1660                         unsigned long pfn, pgprot_t prot)
1661 {
1662         pte_t *pte;
1663         spinlock_t *ptl;
1664
1665         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1666         if (!pte)
1667                 return -ENOMEM;
1668         arch_enter_lazy_mmu_mode();
1669         do {
1670                 BUG_ON(!pte_none(*pte));
1671                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1672                 pfn++;
1673         } while (pte++, addr += PAGE_SIZE, addr != end);
1674         arch_leave_lazy_mmu_mode();
1675         pte_unmap_unlock(pte - 1, ptl);
1676         return 0;
1677 }
1678
1679 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1680                         unsigned long addr, unsigned long end,
1681                         unsigned long pfn, pgprot_t prot)
1682 {
1683         pmd_t *pmd;
1684         unsigned long next;
1685
1686         pfn -= addr >> PAGE_SHIFT;
1687         pmd = pmd_alloc(mm, pud, addr);
1688         if (!pmd)
1689                 return -ENOMEM;
1690         VM_BUG_ON(pmd_trans_huge(*pmd));
1691         do {
1692                 next = pmd_addr_end(addr, end);
1693                 if (remap_pte_range(mm, pmd, addr, next,
1694                                 pfn + (addr >> PAGE_SHIFT), prot))
1695                         return -ENOMEM;
1696         } while (pmd++, addr = next, addr != end);
1697         return 0;
1698 }
1699
1700 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1701                         unsigned long addr, unsigned long end,
1702                         unsigned long pfn, pgprot_t prot)
1703 {
1704         pud_t *pud;
1705         unsigned long next;
1706
1707         pfn -= addr >> PAGE_SHIFT;
1708         pud = pud_alloc(mm, pgd, addr);
1709         if (!pud)
1710                 return -ENOMEM;
1711         do {
1712                 next = pud_addr_end(addr, end);
1713                 if (remap_pmd_range(mm, pud, addr, next,
1714                                 pfn + (addr >> PAGE_SHIFT), prot))
1715                         return -ENOMEM;
1716         } while (pud++, addr = next, addr != end);
1717         return 0;
1718 }
1719
1720 /**
1721  * remap_pfn_range - remap kernel memory to userspace
1722  * @vma: user vma to map to
1723  * @addr: target user address to start at
1724  * @pfn: physical address of kernel memory
1725  * @size: size of map area
1726  * @prot: page protection flags for this mapping
1727  *
1728  *  Note: this is only safe if the mm semaphore is held when called.
1729  */
1730 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1731                     unsigned long pfn, unsigned long size, pgprot_t prot)
1732 {
1733         pgd_t *pgd;
1734         unsigned long next;
1735         unsigned long end = addr + PAGE_ALIGN(size);
1736         struct mm_struct *mm = vma->vm_mm;
1737         int err;
1738
1739         /*
1740          * Physically remapped pages are special. Tell the
1741          * rest of the world about it:
1742          *   VM_IO tells people not to look at these pages
1743          *      (accesses can have side effects).
1744          *   VM_PFNMAP tells the core MM that the base pages are just
1745          *      raw PFN mappings, and do not have a "struct page" associated
1746          *      with them.
1747          *   VM_DONTEXPAND
1748          *      Disable vma merging and expanding with mremap().
1749          *   VM_DONTDUMP
1750          *      Omit vma from core dump, even when VM_IO turned off.
1751          *
1752          * There's a horrible special case to handle copy-on-write
1753          * behaviour that some programs depend on. We mark the "original"
1754          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1755          * See vm_normal_page() for details.
1756          */
1757         if (is_cow_mapping(vma->vm_flags)) {
1758                 if (addr != vma->vm_start || end != vma->vm_end)
1759                         return -EINVAL;
1760                 vma->vm_pgoff = pfn;
1761         }
1762
1763         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1764         if (err)
1765                 return -EINVAL;
1766
1767         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1768
1769         BUG_ON(addr >= end);
1770         pfn -= addr >> PAGE_SHIFT;
1771         pgd = pgd_offset(mm, addr);
1772         flush_cache_range(vma, addr, end);
1773         do {
1774                 next = pgd_addr_end(addr, end);
1775                 err = remap_pud_range(mm, pgd, addr, next,
1776                                 pfn + (addr >> PAGE_SHIFT), prot);
1777                 if (err)
1778                         break;
1779         } while (pgd++, addr = next, addr != end);
1780
1781         if (err)
1782                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1783
1784         return err;
1785 }
1786 EXPORT_SYMBOL(remap_pfn_range);
1787
1788 /**
1789  * vm_iomap_memory - remap memory to userspace
1790  * @vma: user vma to map to
1791  * @start: start of area
1792  * @len: size of area
1793  *
1794  * This is a simplified io_remap_pfn_range() for common driver use. The
1795  * driver just needs to give us the physical memory range to be mapped,
1796  * we'll figure out the rest from the vma information.
1797  *
1798  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1799  * whatever write-combining details or similar.
1800  */
1801 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1802 {
1803         unsigned long vm_len, pfn, pages;
1804
1805         /* Check that the physical memory area passed in looks valid */
1806         if (start + len < start)
1807                 return -EINVAL;
1808         /*
1809          * You *really* shouldn't map things that aren't page-aligned,
1810          * but we've historically allowed it because IO memory might
1811          * just have smaller alignment.
1812          */
1813         len += start & ~PAGE_MASK;
1814         pfn = start >> PAGE_SHIFT;
1815         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1816         if (pfn + pages < pfn)
1817                 return -EINVAL;
1818
1819         /* We start the mapping 'vm_pgoff' pages into the area */
1820         if (vma->vm_pgoff > pages)
1821                 return -EINVAL;
1822         pfn += vma->vm_pgoff;
1823         pages -= vma->vm_pgoff;
1824
1825         /* Can we fit all of the mapping? */
1826         vm_len = vma->vm_end - vma->vm_start;
1827         if (vm_len >> PAGE_SHIFT > pages)
1828                 return -EINVAL;
1829
1830         /* Ok, let it rip */
1831         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1832 }
1833 EXPORT_SYMBOL(vm_iomap_memory);
1834
1835 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1836                                      unsigned long addr, unsigned long end,
1837                                      pte_fn_t fn, void *data)
1838 {
1839         pte_t *pte;
1840         int err;
1841         pgtable_t token;
1842         spinlock_t *uninitialized_var(ptl);
1843
1844         pte = (mm == &init_mm) ?
1845                 pte_alloc_kernel(pmd, addr) :
1846                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1847         if (!pte)
1848                 return -ENOMEM;
1849
1850         BUG_ON(pmd_huge(*pmd));
1851
1852         arch_enter_lazy_mmu_mode();
1853
1854         token = pmd_pgtable(*pmd);
1855
1856         do {
1857                 err = fn(pte++, token, addr, data);
1858                 if (err)
1859                         break;
1860         } while (addr += PAGE_SIZE, addr != end);
1861
1862         arch_leave_lazy_mmu_mode();
1863
1864         if (mm != &init_mm)
1865                 pte_unmap_unlock(pte-1, ptl);
1866         return err;
1867 }
1868
1869 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1870                                      unsigned long addr, unsigned long end,
1871                                      pte_fn_t fn, void *data)
1872 {
1873         pmd_t *pmd;
1874         unsigned long next;
1875         int err;
1876
1877         BUG_ON(pud_huge(*pud));
1878
1879         pmd = pmd_alloc(mm, pud, addr);
1880         if (!pmd)
1881                 return -ENOMEM;
1882         do {
1883                 next = pmd_addr_end(addr, end);
1884                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1885                 if (err)
1886                         break;
1887         } while (pmd++, addr = next, addr != end);
1888         return err;
1889 }
1890
1891 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1892                                      unsigned long addr, unsigned long end,
1893                                      pte_fn_t fn, void *data)
1894 {
1895         pud_t *pud;
1896         unsigned long next;
1897         int err;
1898
1899         pud = pud_alloc(mm, pgd, addr);
1900         if (!pud)
1901                 return -ENOMEM;
1902         do {
1903                 next = pud_addr_end(addr, end);
1904                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1905                 if (err)
1906                         break;
1907         } while (pud++, addr = next, addr != end);
1908         return err;
1909 }
1910
1911 /*
1912  * Scan a region of virtual memory, filling in page tables as necessary
1913  * and calling a provided function on each leaf page table.
1914  */
1915 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1916                         unsigned long size, pte_fn_t fn, void *data)
1917 {
1918         pgd_t *pgd;
1919         unsigned long next;
1920         unsigned long end = addr + size;
1921         int err;
1922
1923         BUG_ON(addr >= end);
1924         pgd = pgd_offset(mm, addr);
1925         do {
1926                 next = pgd_addr_end(addr, end);
1927                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1928                 if (err)
1929                         break;
1930         } while (pgd++, addr = next, addr != end);
1931
1932         return err;
1933 }
1934 EXPORT_SYMBOL_GPL(apply_to_page_range);
1935
1936 /*
1937  * handle_pte_fault chooses page fault handler according to an entry
1938  * which was read non-atomically.  Before making any commitment, on
1939  * those architectures or configurations (e.g. i386 with PAE) which
1940  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
1941  * must check under lock before unmapping the pte and proceeding
1942  * (but do_wp_page is only called after already making such a check;
1943  * and do_anonymous_page can safely check later on).
1944  */
1945 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1946                                 pte_t *page_table, pte_t orig_pte)
1947 {
1948         int same = 1;
1949 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1950         if (sizeof(pte_t) > sizeof(unsigned long)) {
1951                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1952                 spin_lock(ptl);
1953                 same = pte_same(*page_table, orig_pte);
1954                 spin_unlock(ptl);
1955         }
1956 #endif
1957         pte_unmap(page_table);
1958         return same;
1959 }
1960
1961 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1962 {
1963         debug_dma_assert_idle(src);
1964
1965         /*
1966          * If the source page was a PFN mapping, we don't have
1967          * a "struct page" for it. We do a best-effort copy by
1968          * just copying from the original user address. If that
1969          * fails, we just zero-fill it. Live with it.
1970          */
1971         if (unlikely(!src)) {
1972                 void *kaddr = kmap_atomic(dst);
1973                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1974
1975                 /*
1976                  * This really shouldn't fail, because the page is there
1977                  * in the page tables. But it might just be unreadable,
1978                  * in which case we just give up and fill the result with
1979                  * zeroes.
1980                  */
1981                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1982                         clear_page(kaddr);
1983                 kunmap_atomic(kaddr);
1984                 flush_dcache_page(dst);
1985         } else
1986                 copy_user_highpage(dst, src, va, vma);
1987 }
1988
1989 /*
1990  * Notify the address space that the page is about to become writable so that
1991  * it can prohibit this or wait for the page to get into an appropriate state.
1992  *
1993  * We do this without the lock held, so that it can sleep if it needs to.
1994  */
1995 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1996                unsigned long address)
1997 {
1998         struct vm_fault vmf;
1999         int ret;
2000
2001         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2002         vmf.pgoff = page->index;
2003         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2004         vmf.page = page;
2005
2006         ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2007         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2008                 return ret;
2009         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2010                 lock_page(page);
2011                 if (!page->mapping) {
2012                         unlock_page(page);
2013                         return 0; /* retry */
2014                 }
2015                 ret |= VM_FAULT_LOCKED;
2016         } else
2017                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2018         return ret;
2019 }
2020
2021 /*
2022  * This routine handles present pages, when users try to write
2023  * to a shared page. It is done by copying the page to a new address
2024  * and decrementing the shared-page counter for the old page.
2025  *
2026  * Note that this routine assumes that the protection checks have been
2027  * done by the caller (the low-level page fault routine in most cases).
2028  * Thus we can safely just mark it writable once we've done any necessary
2029  * COW.
2030  *
2031  * We also mark the page dirty at this point even though the page will
2032  * change only once the write actually happens. This avoids a few races,
2033  * and potentially makes it more efficient.
2034  *
2035  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2036  * but allow concurrent faults), with pte both mapped and locked.
2037  * We return with mmap_sem still held, but pte unmapped and unlocked.
2038  */
2039 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2040                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2041                 spinlock_t *ptl, pte_t orig_pte)
2042         __releases(ptl)
2043 {
2044         struct page *old_page, *new_page = NULL;
2045         pte_t entry;
2046         int ret = 0;
2047         int page_mkwrite = 0;
2048         struct page *dirty_page = NULL;
2049         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2050         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2051         struct mem_cgroup *memcg;
2052
2053         old_page = vm_normal_page(vma, address, orig_pte);
2054         if (!old_page) {
2055                 /*
2056                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2057                  * VM_PFNMAP VMA.
2058                  *
2059                  * We should not cow pages in a shared writeable mapping.
2060                  * Just mark the pages writable as we can't do any dirty
2061                  * accounting on raw pfn maps.
2062                  */
2063                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2064                                      (VM_WRITE|VM_SHARED))
2065                         goto reuse;
2066                 goto gotten;
2067         }
2068
2069         /*
2070          * Take out anonymous pages first, anonymous shared vmas are
2071          * not dirty accountable.
2072          */
2073         if (PageAnon(old_page) && !PageKsm(old_page)) {
2074                 if (!trylock_page(old_page)) {
2075                         page_cache_get(old_page);
2076                         pte_unmap_unlock(page_table, ptl);
2077                         lock_page(old_page);
2078                         page_table = pte_offset_map_lock(mm, pmd, address,
2079                                                          &ptl);
2080                         if (!pte_same(*page_table, orig_pte)) {
2081                                 unlock_page(old_page);
2082                                 goto unlock;
2083                         }
2084                         page_cache_release(old_page);
2085                 }
2086                 if (reuse_swap_page(old_page)) {
2087                         /*
2088                          * The page is all ours.  Move it to our anon_vma so
2089                          * the rmap code will not search our parent or siblings.
2090                          * Protected against the rmap code by the page lock.
2091                          */
2092                         page_move_anon_rmap(old_page, vma, address);
2093                         unlock_page(old_page);
2094                         goto reuse;
2095                 }
2096                 unlock_page(old_page);
2097         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2098                                         (VM_WRITE|VM_SHARED))) {
2099                 /*
2100                  * Only catch write-faults on shared writable pages,
2101                  * read-only shared pages can get COWed by
2102                  * get_user_pages(.write=1, .force=1).
2103                  */
2104                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2105                         int tmp;
2106                         page_cache_get(old_page);
2107                         pte_unmap_unlock(page_table, ptl);
2108                         tmp = do_page_mkwrite(vma, old_page, address);
2109                         if (unlikely(!tmp || (tmp &
2110                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2111                                 page_cache_release(old_page);
2112                                 return tmp;
2113                         }
2114                         /*
2115                          * Since we dropped the lock we need to revalidate
2116                          * the PTE as someone else may have changed it.  If
2117                          * they did, we just return, as we can count on the
2118                          * MMU to tell us if they didn't also make it writable.
2119                          */
2120                         page_table = pte_offset_map_lock(mm, pmd, address,
2121                                                          &ptl);
2122                         if (!pte_same(*page_table, orig_pte)) {
2123                                 unlock_page(old_page);
2124                                 goto unlock;
2125                         }
2126
2127                         page_mkwrite = 1;
2128                 }
2129                 dirty_page = old_page;
2130                 get_page(dirty_page);
2131
2132 reuse:
2133                 /*
2134                  * Clear the pages cpupid information as the existing
2135                  * information potentially belongs to a now completely
2136                  * unrelated process.
2137                  */
2138                 if (old_page)
2139                         page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2140
2141                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2142                 entry = pte_mkyoung(orig_pte);
2143                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2144                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2145                         update_mmu_cache(vma, address, page_table);
2146                 pte_unmap_unlock(page_table, ptl);
2147                 ret |= VM_FAULT_WRITE;
2148
2149                 if (!dirty_page)
2150                         return ret;
2151
2152                 /*
2153                  * Yes, Virginia, this is actually required to prevent a race
2154                  * with clear_page_dirty_for_io() from clearing the page dirty
2155                  * bit after it clear all dirty ptes, but before a racing
2156                  * do_wp_page installs a dirty pte.
2157                  *
2158                  * do_shared_fault is protected similarly.
2159                  */
2160                 if (!page_mkwrite) {
2161                         wait_on_page_locked(dirty_page);
2162                         set_page_dirty_balance(dirty_page);
2163                         /* file_update_time outside page_lock */
2164                         if (vma->vm_file)
2165                                 file_update_time(vma->vm_file);
2166                 }
2167                 put_page(dirty_page);
2168                 if (page_mkwrite) {
2169                         struct address_space *mapping = dirty_page->mapping;
2170
2171                         set_page_dirty(dirty_page);
2172                         unlock_page(dirty_page);
2173                         page_cache_release(dirty_page);
2174                         if (mapping)    {
2175                                 /*
2176                                  * Some device drivers do not set page.mapping
2177                                  * but still dirty their pages
2178                                  */
2179                                 balance_dirty_pages_ratelimited(mapping);
2180                         }
2181                 }
2182
2183                 return ret;
2184         }
2185
2186         /*
2187          * Ok, we need to copy. Oh, well..
2188          */
2189         page_cache_get(old_page);
2190 gotten:
2191         pte_unmap_unlock(page_table, ptl);
2192
2193         if (unlikely(anon_vma_prepare(vma)))
2194                 goto oom;
2195
2196         if (is_zero_pfn(pte_pfn(orig_pte))) {
2197                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2198                 if (!new_page)
2199                         goto oom;
2200         } else {
2201                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2202                 if (!new_page)
2203                         goto oom;
2204                 cow_user_page(new_page, old_page, address, vma);
2205         }
2206         __SetPageUptodate(new_page);
2207
2208         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
2209                 goto oom_free_new;
2210
2211         mmun_start  = address & PAGE_MASK;
2212         mmun_end    = mmun_start + PAGE_SIZE;
2213         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2214
2215         /*
2216          * Re-check the pte - we dropped the lock
2217          */
2218         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2219         if (likely(pte_same(*page_table, orig_pte))) {
2220                 if (old_page) {
2221                         if (!PageAnon(old_page)) {
2222                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2223                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2224                         }
2225                 } else
2226                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2227                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2228                 entry = mk_pte(new_page, vma->vm_page_prot);
2229                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2230                 /*
2231                  * Clear the pte entry and flush it first, before updating the
2232                  * pte with the new entry. This will avoid a race condition
2233                  * seen in the presence of one thread doing SMC and another
2234                  * thread doing COW.
2235                  */
2236                 ptep_clear_flush(vma, address, page_table);
2237                 page_add_new_anon_rmap(new_page, vma, address);
2238                 mem_cgroup_commit_charge(new_page, memcg, false);
2239                 lru_cache_add_active_or_unevictable(new_page, vma);
2240                 /*
2241                  * We call the notify macro here because, when using secondary
2242                  * mmu page tables (such as kvm shadow page tables), we want the
2243                  * new page to be mapped directly into the secondary page table.
2244                  */
2245                 set_pte_at_notify(mm, address, page_table, entry);
2246                 update_mmu_cache(vma, address, page_table);
2247                 if (old_page) {
2248                         /*
2249                          * Only after switching the pte to the new page may
2250                          * we remove the mapcount here. Otherwise another
2251                          * process may come and find the rmap count decremented
2252                          * before the pte is switched to the new page, and
2253                          * "reuse" the old page writing into it while our pte
2254                          * here still points into it and can be read by other
2255                          * threads.
2256                          *
2257                          * The critical issue is to order this
2258                          * page_remove_rmap with the ptp_clear_flush above.
2259                          * Those stores are ordered by (if nothing else,)
2260                          * the barrier present in the atomic_add_negative
2261                          * in page_remove_rmap.
2262                          *
2263                          * Then the TLB flush in ptep_clear_flush ensures that
2264                          * no process can access the old page before the
2265                          * decremented mapcount is visible. And the old page
2266                          * cannot be reused until after the decremented
2267                          * mapcount is visible. So transitively, TLBs to
2268                          * old page will be flushed before it can be reused.
2269                          */
2270                         page_remove_rmap(old_page);
2271                 }
2272
2273                 /* Free the old page.. */
2274                 new_page = old_page;
2275                 ret |= VM_FAULT_WRITE;
2276         } else
2277                 mem_cgroup_cancel_charge(new_page, memcg);
2278
2279         if (new_page)
2280                 page_cache_release(new_page);
2281 unlock:
2282         pte_unmap_unlock(page_table, ptl);
2283         if (mmun_end > mmun_start)
2284                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2285         if (old_page) {
2286                 /*
2287                  * Don't let another task, with possibly unlocked vma,
2288                  * keep the mlocked page.
2289                  */
2290                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2291                         lock_page(old_page);    /* LRU manipulation */
2292                         munlock_vma_page(old_page);
2293                         unlock_page(old_page);
2294                 }
2295                 page_cache_release(old_page);
2296         }
2297         return ret;
2298 oom_free_new:
2299         page_cache_release(new_page);
2300 oom:
2301         if (old_page)
2302                 page_cache_release(old_page);
2303         return VM_FAULT_OOM;
2304 }
2305
2306 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2307                 unsigned long start_addr, unsigned long end_addr,
2308                 struct zap_details *details)
2309 {
2310         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2311 }
2312
2313 static inline void unmap_mapping_range_tree(struct rb_root *root,
2314                                             struct zap_details *details)
2315 {
2316         struct vm_area_struct *vma;
2317         pgoff_t vba, vea, zba, zea;
2318
2319         vma_interval_tree_foreach(vma, root,
2320                         details->first_index, details->last_index) {
2321
2322                 vba = vma->vm_pgoff;
2323                 vea = vba + vma_pages(vma) - 1;
2324                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2325                 zba = details->first_index;
2326                 if (zba < vba)
2327                         zba = vba;
2328                 zea = details->last_index;
2329                 if (zea > vea)
2330                         zea = vea;
2331
2332                 unmap_mapping_range_vma(vma,
2333                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2334                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2335                                 details);
2336         }
2337 }
2338
2339 static inline void unmap_mapping_range_list(struct list_head *head,
2340                                             struct zap_details *details)
2341 {
2342         struct vm_area_struct *vma;
2343
2344         /*
2345          * In nonlinear VMAs there is no correspondence between virtual address
2346          * offset and file offset.  So we must perform an exhaustive search
2347          * across *all* the pages in each nonlinear VMA, not just the pages
2348          * whose virtual address lies outside the file truncation point.
2349          */
2350         list_for_each_entry(vma, head, shared.nonlinear) {
2351                 details->nonlinear_vma = vma;
2352                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2353         }
2354 }
2355
2356 /**
2357  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2358  * @mapping: the address space containing mmaps to be unmapped.
2359  * @holebegin: byte in first page to unmap, relative to the start of
2360  * the underlying file.  This will be rounded down to a PAGE_SIZE
2361  * boundary.  Note that this is different from truncate_pagecache(), which
2362  * must keep the partial page.  In contrast, we must get rid of
2363  * partial pages.
2364  * @holelen: size of prospective hole in bytes.  This will be rounded
2365  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2366  * end of the file.
2367  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2368  * but 0 when invalidating pagecache, don't throw away private data.
2369  */
2370 void unmap_mapping_range(struct address_space *mapping,
2371                 loff_t const holebegin, loff_t const holelen, int even_cows)
2372 {
2373         struct zap_details details;
2374         pgoff_t hba = holebegin >> PAGE_SHIFT;
2375         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2376
2377         /* Check for overflow. */
2378         if (sizeof(holelen) > sizeof(hlen)) {
2379                 long long holeend =
2380                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2381                 if (holeend & ~(long long)ULONG_MAX)
2382                         hlen = ULONG_MAX - hba + 1;
2383         }
2384
2385         details.check_mapping = even_cows? NULL: mapping;
2386         details.nonlinear_vma = NULL;
2387         details.first_index = hba;
2388         details.last_index = hba + hlen - 1;
2389         if (details.last_index < details.first_index)
2390                 details.last_index = ULONG_MAX;
2391
2392
2393         mutex_lock(&mapping->i_mmap_mutex);
2394         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2395                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2396         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2397                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2398         mutex_unlock(&mapping->i_mmap_mutex);
2399 }
2400 EXPORT_SYMBOL(unmap_mapping_range);
2401
2402 /*
2403  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2404  * but allow concurrent faults), and pte mapped but not yet locked.
2405  * We return with pte unmapped and unlocked.
2406  *
2407  * We return with the mmap_sem locked or unlocked in the same cases
2408  * as does filemap_fault().
2409  */
2410 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2411                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2412                 unsigned int flags, pte_t orig_pte)
2413 {
2414         spinlock_t *ptl;
2415         struct page *page, *swapcache;
2416         struct mem_cgroup *memcg;
2417         swp_entry_t entry;
2418         pte_t pte;
2419         int locked;
2420         int exclusive = 0;
2421         int ret = 0;
2422
2423         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2424                 goto out;
2425
2426         entry = pte_to_swp_entry(orig_pte);
2427         if (unlikely(non_swap_entry(entry))) {
2428                 if (is_migration_entry(entry)) {
2429                         migration_entry_wait(mm, pmd, address);
2430                 } else if (is_hwpoison_entry(entry)) {
2431                         ret = VM_FAULT_HWPOISON;
2432                 } else {
2433                         print_bad_pte(vma, address, orig_pte, NULL);
2434                         ret = VM_FAULT_SIGBUS;
2435                 }
2436                 goto out;
2437         }
2438         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2439         page = lookup_swap_cache(entry);
2440         if (!page) {
2441                 page = swapin_readahead(entry,
2442                                         GFP_HIGHUSER_MOVABLE, vma, address);
2443                 if (!page) {
2444                         /*
2445                          * Back out if somebody else faulted in this pte
2446                          * while we released the pte lock.
2447                          */
2448                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2449                         if (likely(pte_same(*page_table, orig_pte)))
2450                                 ret = VM_FAULT_OOM;
2451                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2452                         goto unlock;
2453                 }
2454
2455                 /* Had to read the page from swap area: Major fault */
2456                 ret = VM_FAULT_MAJOR;
2457                 count_vm_event(PGMAJFAULT);
2458                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2459         } else if (PageHWPoison(page)) {
2460                 /*
2461                  * hwpoisoned dirty swapcache pages are kept for killing
2462                  * owner processes (which may be unknown at hwpoison time)
2463                  */
2464                 ret = VM_FAULT_HWPOISON;
2465                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2466                 swapcache = page;
2467                 goto out_release;
2468         }
2469
2470         swapcache = page;
2471         locked = lock_page_or_retry(page, mm, flags);
2472
2473         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2474         if (!locked) {
2475                 ret |= VM_FAULT_RETRY;
2476                 goto out_release;
2477         }
2478
2479         /*
2480          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2481          * release the swapcache from under us.  The page pin, and pte_same
2482          * test below, are not enough to exclude that.  Even if it is still
2483          * swapcache, we need to check that the page's swap has not changed.
2484          */
2485         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2486                 goto out_page;
2487
2488         page = ksm_might_need_to_copy(page, vma, address);
2489         if (unlikely(!page)) {
2490                 ret = VM_FAULT_OOM;
2491                 page = swapcache;
2492                 goto out_page;
2493         }
2494
2495         if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
2496                 ret = VM_FAULT_OOM;
2497                 goto out_page;
2498         }
2499
2500         /*
2501          * Back out if somebody else already faulted in this pte.
2502          */
2503         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2504         if (unlikely(!pte_same(*page_table, orig_pte)))
2505                 goto out_nomap;
2506
2507         if (unlikely(!PageUptodate(page))) {
2508                 ret = VM_FAULT_SIGBUS;
2509                 goto out_nomap;
2510         }
2511
2512         /*
2513          * The page isn't present yet, go ahead with the fault.
2514          *
2515          * Be careful about the sequence of operations here.
2516          * To get its accounting right, reuse_swap_page() must be called
2517          * while the page is counted on swap but not yet in mapcount i.e.
2518          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2519          * must be called after the swap_free(), or it will never succeed.
2520          */
2521
2522         inc_mm_counter_fast(mm, MM_ANONPAGES);
2523         dec_mm_counter_fast(mm, MM_SWAPENTS);
2524         pte = mk_pte(page, vma->vm_page_prot);
2525         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2526                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2527                 flags &= ~FAULT_FLAG_WRITE;
2528                 ret |= VM_FAULT_WRITE;
2529                 exclusive = 1;
2530         }
2531         flush_icache_page(vma, page);
2532         if (pte_swp_soft_dirty(orig_pte))
2533                 pte = pte_mksoft_dirty(pte);
2534         set_pte_at(mm, address, page_table, pte);
2535         if (page == swapcache) {
2536                 do_page_add_anon_rmap(page, vma, address, exclusive);
2537                 mem_cgroup_commit_charge(page, memcg, true);
2538         } else { /* ksm created a completely new copy */
2539                 page_add_new_anon_rmap(page, vma, address);
2540                 mem_cgroup_commit_charge(page, memcg, false);
2541                 lru_cache_add_active_or_unevictable(page, vma);
2542         }
2543
2544         swap_free(entry);
2545         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2546                 try_to_free_swap(page);
2547         unlock_page(page);
2548         if (page != swapcache) {
2549                 /*
2550                  * Hold the lock to avoid the swap entry to be reused
2551                  * until we take the PT lock for the pte_same() check
2552                  * (to avoid false positives from pte_same). For
2553                  * further safety release the lock after the swap_free
2554                  * so that the swap count won't change under a
2555                  * parallel locked swapcache.
2556                  */
2557                 unlock_page(swapcache);
2558                 page_cache_release(swapcache);
2559         }
2560
2561         if (flags & FAULT_FLAG_WRITE) {
2562                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2563                 if (ret & VM_FAULT_ERROR)
2564                         ret &= VM_FAULT_ERROR;
2565                 goto out;
2566         }
2567
2568         /* No need to invalidate - it was non-present before */
2569         update_mmu_cache(vma, address, page_table);
2570 unlock:
2571         pte_unmap_unlock(page_table, ptl);
2572 out:
2573         return ret;
2574 out_nomap:
2575         mem_cgroup_cancel_charge(page, memcg);
2576         pte_unmap_unlock(page_table, ptl);
2577 out_page:
2578         unlock_page(page);
2579 out_release:
2580         page_cache_release(page);
2581         if (page != swapcache) {
2582                 unlock_page(swapcache);
2583                 page_cache_release(swapcache);
2584         }
2585         return ret;
2586 }
2587
2588 /*
2589  * This is like a special single-page "expand_{down|up}wards()",
2590  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2591  * doesn't hit another vma.
2592  */
2593 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2594 {
2595         address &= PAGE_MASK;
2596         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2597                 struct vm_area_struct *prev = vma->vm_prev;
2598
2599                 /*
2600                  * Is there a mapping abutting this one below?
2601                  *
2602                  * That's only ok if it's the same stack mapping
2603                  * that has gotten split..
2604                  */
2605                 if (prev && prev->vm_end == address)
2606                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2607
2608                 expand_downwards(vma, address - PAGE_SIZE);
2609         }
2610         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2611                 struct vm_area_struct *next = vma->vm_next;
2612
2613                 /* As VM_GROWSDOWN but s/below/above/ */
2614                 if (next && next->vm_start == address + PAGE_SIZE)
2615                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2616
2617                 expand_upwards(vma, address + PAGE_SIZE);
2618         }
2619         return 0;
2620 }
2621
2622 /*
2623  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2624  * but allow concurrent faults), and pte mapped but not yet locked.
2625  * We return with mmap_sem still held, but pte unmapped and unlocked.
2626  */
2627 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2628                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2629                 unsigned int flags)
2630 {
2631         struct mem_cgroup *memcg;
2632         struct page *page;
2633         spinlock_t *ptl;
2634         pte_t entry;
2635
2636         pte_unmap(page_table);
2637
2638         /* Check if we need to add a guard page to the stack */
2639         if (check_stack_guard_page(vma, address) < 0)
2640                 return VM_FAULT_SIGBUS;
2641
2642         /* Use the zero-page for reads */
2643         if (!(flags & FAULT_FLAG_WRITE)) {
2644                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2645                                                 vma->vm_page_prot));
2646                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2647                 if (!pte_none(*page_table))
2648                         goto unlock;
2649                 goto setpte;
2650         }
2651
2652         /* Allocate our own private page. */
2653         if (unlikely(anon_vma_prepare(vma)))
2654                 goto oom;
2655         page = alloc_zeroed_user_highpage_movable(vma, address);
2656         if (!page)
2657                 goto oom;
2658         /*
2659          * The memory barrier inside __SetPageUptodate makes sure that
2660          * preceeding stores to the page contents become visible before
2661          * the set_pte_at() write.
2662          */
2663         __SetPageUptodate(page);
2664
2665         if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
2666                 goto oom_free_page;
2667
2668         entry = mk_pte(page, vma->vm_page_prot);
2669         if (vma->vm_flags & VM_WRITE)
2670                 entry = pte_mkwrite(pte_mkdirty(entry));
2671
2672         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2673         if (!pte_none(*page_table))
2674                 goto release;
2675
2676         inc_mm_counter_fast(mm, MM_ANONPAGES);
2677         page_add_new_anon_rmap(page, vma, address);
2678         mem_cgroup_commit_charge(page, memcg, false);
2679         lru_cache_add_active_or_unevictable(page, vma);
2680 setpte:
2681         set_pte_at(mm, address, page_table, entry);
2682
2683         /* No need to invalidate - it was non-present before */
2684         update_mmu_cache(vma, address, page_table);
2685 unlock:
2686         pte_unmap_unlock(page_table, ptl);
2687         return 0;
2688 release:
2689         mem_cgroup_cancel_charge(page, memcg);
2690         page_cache_release(page);
2691         goto unlock;
2692 oom_free_page:
2693         page_cache_release(page);
2694 oom:
2695         return VM_FAULT_OOM;
2696 }
2697
2698 /*
2699  * The mmap_sem must have been held on entry, and may have been
2700  * released depending on flags and vma->vm_ops->fault() return value.
2701  * See filemap_fault() and __lock_page_retry().
2702  */
2703 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2704                 pgoff_t pgoff, unsigned int flags, struct page **page)
2705 {
2706         struct vm_fault vmf;
2707         int ret;
2708
2709         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2710         vmf.pgoff = pgoff;
2711         vmf.flags = flags;
2712         vmf.page = NULL;
2713
2714         ret = vma->vm_ops->fault(vma, &vmf);
2715         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2716                 return ret;
2717
2718         if (unlikely(PageHWPoison(vmf.page))) {
2719                 if (ret & VM_FAULT_LOCKED)
2720                         unlock_page(vmf.page);
2721                 page_cache_release(vmf.page);
2722                 return VM_FAULT_HWPOISON;
2723         }
2724
2725         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2726                 lock_page(vmf.page);
2727         else
2728                 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2729
2730         *page = vmf.page;
2731         return ret;
2732 }
2733
2734 /**
2735  * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2736  *
2737  * @vma: virtual memory area
2738  * @address: user virtual address
2739  * @page: page to map
2740  * @pte: pointer to target page table entry
2741  * @write: true, if new entry is writable
2742  * @anon: true, if it's anonymous page
2743  *
2744  * Caller must hold page table lock relevant for @pte.
2745  *
2746  * Target users are page handler itself and implementations of
2747  * vm_ops->map_pages.
2748  */
2749 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2750                 struct page *page, pte_t *pte, bool write, bool anon)
2751 {
2752         pte_t entry;
2753
2754         flush_icache_page(vma, page);
2755         entry = mk_pte(page, vma->vm_page_prot);
2756         if (write)
2757                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2758         else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
2759                 entry = pte_mksoft_dirty(entry);
2760         if (anon) {
2761                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2762                 page_add_new_anon_rmap(page, vma, address);
2763         } else {
2764                 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2765                 page_add_file_rmap(page);
2766         }
2767         set_pte_at(vma->vm_mm, address, pte, entry);
2768
2769         /* no need to invalidate: a not-present page won't be cached */
2770         update_mmu_cache(vma, address, pte);
2771 }
2772
2773 static unsigned long fault_around_bytes __read_mostly =
2774         rounddown_pow_of_two(65536);
2775
2776 #ifdef CONFIG_DEBUG_FS
2777 static int fault_around_bytes_get(void *data, u64 *val)
2778 {
2779         *val = fault_around_bytes;
2780         return 0;
2781 }
2782
2783 /*
2784  * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2785  * rounded down to nearest page order. It's what do_fault_around() expects to
2786  * see.
2787  */
2788 static int fault_around_bytes_set(void *data, u64 val)
2789 {
2790         if (val / PAGE_SIZE > PTRS_PER_PTE)
2791                 return -EINVAL;
2792         if (val > PAGE_SIZE)
2793                 fault_around_bytes = rounddown_pow_of_two(val);
2794         else
2795                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2796         return 0;
2797 }
2798 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2799                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2800
2801 static int __init fault_around_debugfs(void)
2802 {
2803         void *ret;
2804
2805         ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2806                         &fault_around_bytes_fops);
2807         if (!ret)
2808                 pr_warn("Failed to create fault_around_bytes in debugfs");
2809         return 0;
2810 }
2811 late_initcall(fault_around_debugfs);
2812 #endif
2813
2814 /*
2815  * do_fault_around() tries to map few pages around the fault address. The hope
2816  * is that the pages will be needed soon and this will lower the number of
2817  * faults to handle.
2818  *
2819  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2820  * not ready to be mapped: not up-to-date, locked, etc.
2821  *
2822  * This function is called with the page table lock taken. In the split ptlock
2823  * case the page table lock only protects only those entries which belong to
2824  * the page table corresponding to the fault address.
2825  *
2826  * This function doesn't cross the VMA boundaries, in order to call map_pages()
2827  * only once.
2828  *
2829  * fault_around_pages() defines how many pages we'll try to map.
2830  * do_fault_around() expects it to return a power of two less than or equal to
2831  * PTRS_PER_PTE.
2832  *
2833  * The virtual address of the area that we map is naturally aligned to the
2834  * fault_around_pages() value (and therefore to page order).  This way it's
2835  * easier to guarantee that we don't cross page table boundaries.
2836  */
2837 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2838                 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2839 {
2840         unsigned long start_addr, nr_pages, mask;
2841         pgoff_t max_pgoff;
2842         struct vm_fault vmf;
2843         int off;
2844
2845         nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2846         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2847
2848         start_addr = max(address & mask, vma->vm_start);
2849         off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2850         pte -= off;
2851         pgoff -= off;
2852
2853         /*
2854          *  max_pgoff is either end of page table or end of vma
2855          *  or fault_around_pages() from pgoff, depending what is nearest.
2856          */
2857         max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2858                 PTRS_PER_PTE - 1;
2859         max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2860                         pgoff + nr_pages - 1);
2861
2862         /* Check if it makes any sense to call ->map_pages */
2863         while (!pte_none(*pte)) {
2864                 if (++pgoff > max_pgoff)
2865                         return;
2866                 start_addr += PAGE_SIZE;
2867                 if (start_addr >= vma->vm_end)
2868                         return;
2869                 pte++;
2870         }
2871
2872         vmf.virtual_address = (void __user *) start_addr;
2873         vmf.pte = pte;
2874         vmf.pgoff = pgoff;
2875         vmf.max_pgoff = max_pgoff;
2876         vmf.flags = flags;
2877         vma->vm_ops->map_pages(vma, &vmf);
2878 }
2879
2880 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2881                 unsigned long address, pmd_t *pmd,
2882                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2883 {
2884         struct page *fault_page;
2885         spinlock_t *ptl;
2886         pte_t *pte;
2887         int ret = 0;
2888
2889         /*
2890          * Let's call ->map_pages() first and use ->fault() as fallback
2891          * if page by the offset is not ready to be mapped (cold cache or
2892          * something).
2893          */
2894         if (vma->vm_ops->map_pages && !(flags & FAULT_FLAG_NONLINEAR) &&
2895             fault_around_bytes >> PAGE_SHIFT > 1) {
2896                 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2897                 do_fault_around(vma, address, pte, pgoff, flags);
2898                 if (!pte_same(*pte, orig_pte))
2899                         goto unlock_out;
2900                 pte_unmap_unlock(pte, ptl);
2901         }
2902
2903         ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2904         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2905                 return ret;
2906
2907         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2908         if (unlikely(!pte_same(*pte, orig_pte))) {
2909                 pte_unmap_unlock(pte, ptl);
2910                 unlock_page(fault_page);
2911                 page_cache_release(fault_page);
2912                 return ret;
2913         }
2914         do_set_pte(vma, address, fault_page, pte, false, false);
2915         unlock_page(fault_page);
2916 unlock_out:
2917         pte_unmap_unlock(pte, ptl);
2918         return ret;
2919 }
2920
2921 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2922                 unsigned long address, pmd_t *pmd,
2923                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2924 {
2925         struct page *fault_page, *new_page;
2926         struct mem_cgroup *memcg;
2927         spinlock_t *ptl;
2928         pte_t *pte;
2929         int ret;
2930
2931         if (unlikely(anon_vma_prepare(vma)))
2932                 return VM_FAULT_OOM;
2933
2934         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2935         if (!new_page)
2936                 return VM_FAULT_OOM;
2937
2938         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
2939                 page_cache_release(new_page);
2940                 return VM_FAULT_OOM;
2941         }
2942
2943         ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2944         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2945                 goto uncharge_out;
2946
2947         copy_user_highpage(new_page, fault_page, address, vma);
2948         __SetPageUptodate(new_page);
2949
2950         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2951         if (unlikely(!pte_same(*pte, orig_pte))) {
2952                 pte_unmap_unlock(pte, ptl);
2953                 unlock_page(fault_page);
2954                 page_cache_release(fault_page);
2955                 goto uncharge_out;
2956         }
2957         do_set_pte(vma, address, new_page, pte, true, true);
2958         mem_cgroup_commit_charge(new_page, memcg, false);
2959         lru_cache_add_active_or_unevictable(new_page, vma);
2960         pte_unmap_unlock(pte, ptl);
2961         unlock_page(fault_page);
2962         page_cache_release(fault_page);
2963         return ret;
2964 uncharge_out:
2965         mem_cgroup_cancel_charge(new_page, memcg);
2966         page_cache_release(new_page);
2967         return ret;
2968 }
2969
2970 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2971                 unsigned long address, pmd_t *pmd,
2972                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2973 {
2974         struct page *fault_page;
2975         struct address_space *mapping;
2976         spinlock_t *ptl;
2977         pte_t *pte;
2978         int dirtied = 0;
2979         int ret, tmp;
2980
2981         ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2982         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2983                 return ret;
2984
2985         /*
2986          * Check if the backing address space wants to know that the page is
2987          * about to become writable
2988          */
2989         if (vma->vm_ops->page_mkwrite) {
2990                 unlock_page(fault_page);
2991                 tmp = do_page_mkwrite(vma, fault_page, address);
2992                 if (unlikely(!tmp ||
2993                                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2994                         page_cache_release(fault_page);
2995                         return tmp;
2996                 }
2997         }
2998
2999         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3000         if (unlikely(!pte_same(*pte, orig_pte))) {
3001                 pte_unmap_unlock(pte, ptl);
3002                 unlock_page(fault_page);
3003                 page_cache_release(fault_page);
3004                 return ret;
3005         }
3006         do_set_pte(vma, address, fault_page, pte, true, false);
3007         pte_unmap_unlock(pte, ptl);
3008
3009         if (set_page_dirty(fault_page))
3010                 dirtied = 1;
3011         mapping = fault_page->mapping;
3012         unlock_page(fault_page);
3013         if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3014                 /*
3015                  * Some device drivers do not set page.mapping but still
3016                  * dirty their pages
3017                  */
3018                 balance_dirty_pages_ratelimited(mapping);
3019         }
3020
3021         /* file_update_time outside page_lock */
3022         if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3023                 file_update_time(vma->vm_file);
3024
3025         return ret;
3026 }
3027
3028 /*
3029  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3030  * but allow concurrent faults).
3031  * The mmap_sem may have been released depending on flags and our
3032  * return value.  See filemap_fault() and __lock_page_or_retry().
3033  */
3034 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3035                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3036                 unsigned int flags, pte_t orig_pte)
3037 {
3038         pgoff_t pgoff = (((address & PAGE_MASK)
3039                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3040
3041         pte_unmap(page_table);
3042         if (!(flags & FAULT_FLAG_WRITE))
3043                 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3044                                 orig_pte);
3045         if (!(vma->vm_flags & VM_SHARED))
3046                 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3047                                 orig_pte);
3048         return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3049 }
3050
3051 /*
3052  * Fault of a previously existing named mapping. Repopulate the pte
3053  * from the encoded file_pte if possible. This enables swappable
3054  * nonlinear vmas.
3055  *
3056  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3057  * but allow concurrent faults), and pte mapped but not yet locked.
3058  * We return with pte unmapped and unlocked.
3059  * The mmap_sem may have been released depending on flags and our
3060  * return value.  See filemap_fault() and __lock_page_or_retry().
3061  */
3062 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3063                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3064                 unsigned int flags, pte_t orig_pte)
3065 {
3066         pgoff_t pgoff;
3067
3068         flags |= FAULT_FLAG_NONLINEAR;
3069
3070         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3071                 return 0;
3072
3073         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3074                 /*
3075                  * Page table corrupted: show pte and kill process.
3076                  */
3077                 print_bad_pte(vma, address, orig_pte, NULL);
3078                 return VM_FAULT_SIGBUS;
3079         }
3080
3081         pgoff = pte_to_pgoff(orig_pte);
3082         if (!(flags & FAULT_FLAG_WRITE))
3083                 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3084                                 orig_pte);
3085         if (!(vma->vm_flags & VM_SHARED))
3086                 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3087                                 orig_pte);
3088         return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3089 }
3090
3091 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3092                                 unsigned long addr, int page_nid,
3093                                 int *flags)
3094 {
3095         get_page(page);
3096
3097         count_vm_numa_event(NUMA_HINT_FAULTS);
3098         if (page_nid == numa_node_id()) {
3099                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3100                 *flags |= TNF_FAULT_LOCAL;
3101         }
3102
3103         return mpol_misplaced(page, vma, addr);
3104 }
3105
3106 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3107                    unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3108 {
3109         struct page *page = NULL;
3110         spinlock_t *ptl;
3111         int page_nid = -1;
3112         int last_cpupid;
3113         int target_nid;
3114         bool migrated = false;
3115         int flags = 0;
3116
3117         /*
3118         * The "pte" at this point cannot be used safely without
3119         * validation through pte_unmap_same(). It's of NUMA type but
3120         * the pfn may be screwed if the read is non atomic.
3121         *
3122         * ptep_modify_prot_start is not called as this is clearing
3123         * the _PAGE_NUMA bit and it is not really expected that there
3124         * would be concurrent hardware modifications to the PTE.
3125         */
3126         ptl = pte_lockptr(mm, pmd);
3127         spin_lock(ptl);
3128         if (unlikely(!pte_same(*ptep, pte))) {
3129                 pte_unmap_unlock(ptep, ptl);
3130                 goto out;
3131         }
3132
3133         pte = pte_mknonnuma(pte);
3134         set_pte_at(mm, addr, ptep, pte);
3135         update_mmu_cache(vma, addr, ptep);
3136
3137         page = vm_normal_page(vma, addr, pte);
3138         if (!page) {
3139                 pte_unmap_unlock(ptep, ptl);
3140                 return 0;
3141         }
3142         BUG_ON(is_zero_pfn(page_to_pfn(page)));
3143
3144         /*
3145          * Avoid grouping on DSO/COW pages in specific and RO pages
3146          * in general, RO pages shouldn't hurt as much anyway since
3147          * they can be in shared cache state.
3148          */
3149         if (!pte_write(pte))
3150                 flags |= TNF_NO_GROUP;
3151
3152         /*
3153          * Flag if the page is shared between multiple address spaces. This
3154          * is later used when determining whether to group tasks together
3155          */
3156         if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3157                 flags |= TNF_SHARED;
3158
3159         last_cpupid = page_cpupid_last(page);
3160         page_nid = page_to_nid(page);
3161         target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3162         pte_unmap_unlock(ptep, ptl);
3163         if (target_nid == -1) {
3164                 put_page(page);
3165                 goto out;
3166         }
3167
3168         /* Migrate to the requested node */
3169         migrated = migrate_misplaced_page(page, vma, target_nid);
3170         if (migrated) {
3171                 page_nid = target_nid;
3172                 flags |= TNF_MIGRATED;
3173         }
3174
3175 out:
3176         if (page_nid != -1)
3177                 task_numa_fault(last_cpupid, page_nid, 1, flags);
3178         return 0;
3179 }
3180
3181 /*
3182  * These routines also need to handle stuff like marking pages dirty
3183  * and/or accessed for architectures that don't do it in hardware (most
3184  * RISC architectures).  The early dirtying is also good on the i386.
3185  *
3186  * There is also a hook called "update_mmu_cache()" that architectures
3187  * with external mmu caches can use to update those (ie the Sparc or
3188  * PowerPC hashed page tables that act as extended TLBs).
3189  *
3190  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3191  * but allow concurrent faults), and pte mapped but not yet locked.
3192  * We return with pte unmapped and unlocked.
3193  *
3194  * The mmap_sem may have been released depending on flags and our
3195  * return value.  See filemap_fault() and __lock_page_or_retry().
3196  */
3197 static int handle_pte_fault(struct mm_struct *mm,
3198                      struct vm_area_struct *vma, unsigned long address,
3199                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3200 {
3201         pte_t entry;
3202         spinlock_t *ptl;
3203
3204         entry = ACCESS_ONCE(*pte);
3205         if (!pte_present(entry)) {
3206                 if (pte_none(entry)) {
3207                         if (vma->vm_ops) {
3208                                 if (likely(vma->vm_ops->fault))
3209                                         return do_linear_fault(mm, vma, address,
3210                                                 pte, pmd, flags, entry);
3211                         }
3212                         return do_anonymous_page(mm, vma, address,
3213                                                  pte, pmd, flags);
3214                 }
3215                 if (pte_file(entry))
3216                         return do_nonlinear_fault(mm, vma, address,
3217                                         pte, pmd, flags, entry);
3218                 return do_swap_page(mm, vma, address,
3219                                         pte, pmd, flags, entry);
3220         }
3221
3222         if (pte_numa(entry))
3223                 return do_numa_page(mm, vma, address, entry, pte, pmd);
3224
3225         ptl = pte_lockptr(mm, pmd);
3226         spin_lock(ptl);
3227         if (unlikely(!pte_same(*pte, entry)))
3228                 goto unlock;
3229         if (flags & FAULT_FLAG_WRITE) {
3230                 if (!pte_write(entry))
3231                         return do_wp_page(mm, vma, address,
3232                                         pte, pmd, ptl, entry);
3233                 entry = pte_mkdirty(entry);
3234         }
3235         entry = pte_mkyoung(entry);
3236         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3237                 update_mmu_cache(vma, address, pte);
3238         } else {
3239                 /*
3240                  * This is needed only for protection faults but the arch code
3241                  * is not yet telling us if this is a protection fault or not.
3242                  * This still avoids useless tlb flushes for .text page faults
3243                  * with threads.
3244                  */
3245                 if (flags & FAULT_FLAG_WRITE)
3246                         flush_tlb_fix_spurious_fault(vma, address);
3247         }
3248 unlock:
3249         pte_unmap_unlock(pte, ptl);
3250         return 0;
3251 }
3252
3253 /*
3254  * By the time we get here, we already hold the mm semaphore
3255  *
3256  * The mmap_sem may have been released depending on flags and our
3257  * return value.  See filemap_fault() and __lock_page_or_retry().
3258  */
3259 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3260                              unsigned long address, unsigned int flags)
3261 {
3262         pgd_t *pgd;
3263         pud_t *pud;
3264         pmd_t *pmd;
3265         pte_t *pte;
3266
3267         if (unlikely(is_vm_hugetlb_page(vma)))
3268                 return hugetlb_fault(mm, vma, address, flags);
3269
3270         pgd = pgd_offset(mm, address);
3271         pud = pud_alloc(mm, pgd, address);
3272         if (!pud)
3273                 return VM_FAULT_OOM;
3274         pmd = pmd_alloc(mm, pud, address);
3275         if (!pmd)
3276                 return VM_FAULT_OOM;
3277         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3278                 int ret = VM_FAULT_FALLBACK;
3279                 if (!vma->vm_ops)
3280                         ret = do_huge_pmd_anonymous_page(mm, vma, address,
3281                                         pmd, flags);
3282                 if (!(ret & VM_FAULT_FALLBACK))
3283                         return ret;
3284         } else {
3285                 pmd_t orig_pmd = *pmd;
3286                 int ret;
3287
3288                 barrier();
3289                 if (pmd_trans_huge(orig_pmd)) {
3290                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3291
3292                         /*
3293                          * If the pmd is splitting, return and retry the
3294                          * the fault.  Alternative: wait until the split
3295                          * is done, and goto retry.
3296                          */
3297                         if (pmd_trans_splitting(orig_pmd))
3298                                 return 0;
3299
3300                         if (pmd_numa(orig_pmd))
3301                                 return do_huge_pmd_numa_page(mm, vma, address,
3302                                                              orig_pmd, pmd);
3303
3304                         if (dirty && !pmd_write(orig_pmd)) {
3305                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3306                                                           orig_pmd);
3307                                 if (!(ret & VM_FAULT_FALLBACK))
3308                                         return ret;
3309                         } else {
3310                                 huge_pmd_set_accessed(mm, vma, address, pmd,
3311                                                       orig_pmd, dirty);
3312                                 return 0;
3313                         }
3314                 }
3315         }
3316
3317         /*
3318          * Use __pte_alloc instead of pte_alloc_map, because we can't
3319          * run pte_offset_map on the pmd, if an huge pmd could
3320          * materialize from under us from a different thread.
3321          */
3322         if (unlikely(pmd_none(*pmd)) &&
3323             unlikely(__pte_alloc(mm, vma, pmd, address)))
3324                 return VM_FAULT_OOM;
3325         /* if an huge pmd materialized from under us just retry later */
3326         if (unlikely(pmd_trans_huge(*pmd)))
3327                 return 0;
3328         /*
3329          * A regular pmd is established and it can't morph into a huge pmd
3330          * from under us anymore at this point because we hold the mmap_sem
3331          * read mode and khugepaged takes it in write mode. So now it's
3332          * safe to run pte_offset_map().
3333          */
3334         pte = pte_offset_map(pmd, address);
3335
3336         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3337 }
3338
3339 /*
3340  * By the time we get here, we already hold the mm semaphore
3341  *
3342  * The mmap_sem may have been released depending on flags and our
3343  * return value.  See filemap_fault() and __lock_page_or_retry().
3344  */
3345 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3346                     unsigned long address, unsigned int flags)
3347 {
3348         int ret;
3349
3350         __set_current_state(TASK_RUNNING);
3351
3352         count_vm_event(PGFAULT);
3353         mem_cgroup_count_vm_event(mm, PGFAULT);
3354
3355         /* do counter updates before entering really critical section. */
3356         check_sync_rss_stat(current);
3357
3358         /*
3359          * Enable the memcg OOM handling for faults triggered in user
3360          * space.  Kernel faults are handled more gracefully.
3361          */
3362         if (flags & FAULT_FLAG_USER)
3363                 mem_cgroup_oom_enable();
3364
3365         ret = __handle_mm_fault(mm, vma, address, flags);
3366
3367         if (flags & FAULT_FLAG_USER) {
3368                 mem_cgroup_oom_disable();
3369                 /*
3370                  * The task may have entered a memcg OOM situation but
3371                  * if the allocation error was handled gracefully (no
3372                  * VM_FAULT_OOM), there is no need to kill anything.
3373                  * Just clean up the OOM state peacefully.
3374                  */
3375                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3376                         mem_cgroup_oom_synchronize(false);
3377         }
3378
3379         return ret;
3380 }
3381
3382 #ifndef __PAGETABLE_PUD_FOLDED
3383 /*
3384  * Allocate page upper directory.
3385  * We've already handled the fast-path in-line.
3386  */
3387 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3388 {
3389         pud_t *new = pud_alloc_one(mm, address);
3390         if (!new)
3391                 return -ENOMEM;
3392
3393         smp_wmb(); /* See comment in __pte_alloc */
3394
3395         spin_lock(&mm->page_table_lock);
3396         if (pgd_present(*pgd))          /* Another has populated it */
3397                 pud_free(mm, new);
3398         else
3399                 pgd_populate(mm, pgd, new);
3400         spin_unlock(&mm->page_table_lock);
3401         return 0;
3402 }
3403 #endif /* __PAGETABLE_PUD_FOLDED */
3404
3405 #ifndef __PAGETABLE_PMD_FOLDED
3406 /*
3407  * Allocate page middle directory.
3408  * We've already handled the fast-path in-line.
3409  */
3410 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3411 {
3412         pmd_t *new = pmd_alloc_one(mm, address);
3413         if (!new)
3414                 return -ENOMEM;
3415
3416         smp_wmb(); /* See comment in __pte_alloc */
3417
3418         spin_lock(&mm->page_table_lock);
3419 #ifndef __ARCH_HAS_4LEVEL_HACK
3420         if (pud_present(*pud))          /* Another has populated it */
3421                 pmd_free(mm, new);
3422         else
3423                 pud_populate(mm, pud, new);
3424 #else
3425         if (pgd_present(*pud))          /* Another has populated it */
3426                 pmd_free(mm, new);
3427         else
3428                 pgd_populate(mm, pud, new);
3429 #endif /* __ARCH_HAS_4LEVEL_HACK */
3430         spin_unlock(&mm->page_table_lock);
3431         return 0;
3432 }
3433 #endif /* __PAGETABLE_PMD_FOLDED */
3434
3435 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3436                 pte_t **ptepp, spinlock_t **ptlp)
3437 {
3438         pgd_t *pgd;
3439         pud_t *pud;
3440         pmd_t *pmd;
3441         pte_t *ptep;
3442
3443         pgd = pgd_offset(mm, address);
3444         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3445                 goto out;
3446
3447         pud = pud_offset(pgd, address);
3448         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3449                 goto out;
3450
3451         pmd = pmd_offset(pud, address);
3452         VM_BUG_ON(pmd_trans_huge(*pmd));
3453         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3454                 goto out;
3455
3456         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3457         if (pmd_huge(*pmd))
3458                 goto out;
3459
3460         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3461         if (!ptep)
3462                 goto out;
3463         if (!pte_present(*ptep))
3464                 goto unlock;
3465         *ptepp = ptep;
3466         return 0;
3467 unlock:
3468         pte_unmap_unlock(ptep, *ptlp);
3469 out:
3470         return -EINVAL;
3471 }
3472
3473 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3474                              pte_t **ptepp, spinlock_t **ptlp)
3475 {
3476         int res;
3477
3478         /* (void) is needed to make gcc happy */
3479         (void) __cond_lock(*ptlp,
3480                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3481         return res;
3482 }
3483
3484 /**
3485  * follow_pfn - look up PFN at a user virtual address
3486  * @vma: memory mapping
3487  * @address: user virtual address
3488  * @pfn: location to store found PFN
3489  *
3490  * Only IO mappings and raw PFN mappings are allowed.
3491  *
3492  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3493  */
3494 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3495         unsigned long *pfn)
3496 {
3497         int ret = -EINVAL;
3498         spinlock_t *ptl;
3499         pte_t *ptep;
3500
3501         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3502                 return ret;
3503
3504         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3505         if (ret)
3506                 return ret;
3507         *pfn = pte_pfn(*ptep);
3508         pte_unmap_unlock(ptep, ptl);
3509         return 0;
3510 }
3511 EXPORT_SYMBOL(follow_pfn);
3512
3513 #ifdef CONFIG_HAVE_IOREMAP_PROT
3514 int follow_phys(struct vm_area_struct *vma,
3515                 unsigned long address, unsigned int flags,
3516                 unsigned long *prot, resource_size_t *phys)
3517 {
3518         int ret = -EINVAL;
3519         pte_t *ptep, pte;
3520         spinlock_t *ptl;
3521
3522         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3523                 goto out;
3524
3525         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3526                 goto out;
3527         pte = *ptep;
3528
3529         if ((flags & FOLL_WRITE) && !pte_write(pte))
3530                 goto unlock;
3531
3532         *prot = pgprot_val(pte_pgprot(pte));
3533         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3534
3535         ret = 0;
3536 unlock:
3537         pte_unmap_unlock(ptep, ptl);
3538 out:
3539         return ret;
3540 }
3541
3542 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3543                         void *buf, int len, int write)
3544 {
3545         resource_size_t phys_addr;
3546         unsigned long prot = 0;
3547         void __iomem *maddr;
3548         int offset = addr & (PAGE_SIZE-1);
3549
3550         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3551                 return -EINVAL;
3552
3553         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3554         if (write)
3555                 memcpy_toio(maddr + offset, buf, len);
3556         else
3557                 memcpy_fromio(buf, maddr + offset, len);
3558         iounmap(maddr);
3559
3560         return len;
3561 }
3562 EXPORT_SYMBOL_GPL(generic_access_phys);
3563 #endif
3564
3565 /*
3566  * Access another process' address space as given in mm.  If non-NULL, use the
3567  * given task for page fault accounting.
3568  */
3569 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3570                 unsigned long addr, void *buf, int len, int write)
3571 {
3572         struct vm_area_struct *vma;
3573         void *old_buf = buf;
3574
3575         down_read(&mm->mmap_sem);
3576         /* ignore errors, just check how much was successfully transferred */
3577         while (len) {
3578                 int bytes, ret, offset;
3579                 void *maddr;
3580                 struct page *page = NULL;
3581
3582                 ret = get_user_pages(tsk, mm, addr, 1,
3583                                 write, 1, &page, &vma);
3584                 if (ret <= 0) {
3585 #ifndef CONFIG_HAVE_IOREMAP_PROT
3586                         break;
3587 #else
3588                         /*
3589                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3590                          * we can access using slightly different code.
3591                          */
3592                         vma = find_vma(mm, addr);
3593                         if (!vma || vma->vm_start > addr)
3594                                 break;
3595                         if (vma->vm_ops && vma->vm_ops->access)
3596                                 ret = vma->vm_ops->access(vma, addr, buf,
3597                                                           len, write);
3598                         if (ret <= 0)
3599                                 break;
3600                         bytes = ret;
3601 #endif
3602                 } else {
3603                         bytes = len;
3604                         offset = addr & (PAGE_SIZE-1);
3605                         if (bytes > PAGE_SIZE-offset)
3606                                 bytes = PAGE_SIZE-offset;
3607
3608                         maddr = kmap(page);
3609                         if (write) {
3610                                 copy_to_user_page(vma, page, addr,
3611                                                   maddr + offset, buf, bytes);
3612                                 set_page_dirty_lock(page);
3613                         } else {
3614                                 copy_from_user_page(vma, page, addr,
3615                                                     buf, maddr + offset, bytes);
3616                         }
3617                         kunmap(page);
3618                         page_cache_release(page);
3619                 }
3620                 len -= bytes;
3621                 buf += bytes;
3622                 addr += bytes;
3623         }
3624         up_read(&mm->mmap_sem);
3625
3626         return buf - old_buf;
3627 }
3628
3629 /**
3630  * access_remote_vm - access another process' address space
3631  * @mm:         the mm_struct of the target address space
3632  * @addr:       start address to access
3633  * @buf:        source or destination buffer
3634  * @len:        number of bytes to transfer
3635  * @write:      whether the access is a write
3636  *
3637  * The caller must hold a reference on @mm.
3638  */
3639 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3640                 void *buf, int len, int write)
3641 {
3642         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3643 }
3644
3645 /*
3646  * Access another process' address space.
3647  * Source/target buffer must be kernel space,
3648  * Do not walk the page table directly, use get_user_pages
3649  */
3650 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3651                 void *buf, int len, int write)
3652 {
3653         struct mm_struct *mm;
3654         int ret;
3655
3656         mm = get_task_mm(tsk);
3657         if (!mm)
3658                 return 0;
3659
3660         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3661         mmput(mm);
3662
3663         return ret;
3664 }
3665
3666 /*
3667  * Print the name of a VMA.
3668  */
3669 void print_vma_addr(char *prefix, unsigned long ip)
3670 {
3671         struct mm_struct *mm = current->mm;
3672         struct vm_area_struct *vma;
3673
3674         /*
3675          * Do not print if we are in atomic
3676          * contexts (in exception stacks, etc.):
3677          */
3678         if (preempt_count())
3679                 return;
3680
3681         down_read(&mm->mmap_sem);
3682         vma = find_vma(mm, ip);
3683         if (vma && vma->vm_file) {
3684                 struct file *f = vma->vm_file;
3685                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3686                 if (buf) {
3687                         char *p;
3688
3689                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3690                         if (IS_ERR(p))
3691                                 p = "?";
3692                         printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3693                                         vma->vm_start,
3694                                         vma->vm_end - vma->vm_start);
3695                         free_page((unsigned long)buf);
3696                 }
3697         }
3698         up_read(&mm->mmap_sem);
3699 }
3700
3701 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3702 void might_fault(void)
3703 {
3704         /*
3705          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3706          * holding the mmap_sem, this is safe because kernel memory doesn't
3707          * get paged out, therefore we'll never actually fault, and the
3708          * below annotations will generate false positives.
3709          */
3710         if (segment_eq(get_fs(), KERNEL_DS))
3711                 return;
3712
3713         /*
3714          * it would be nicer only to annotate paths which are not under
3715          * pagefault_disable, however that requires a larger audit and
3716          * providing helpers like get_user_atomic.
3717          */
3718         if (in_atomic())
3719                 return;
3720
3721         __might_sleep(__FILE__, __LINE__, 0);
3722
3723         if (current->mm)
3724                 might_lock_read(&current->mm->mmap_sem);
3725 }
3726 EXPORT_SYMBOL(might_fault);
3727 #endif
3728
3729 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3730 static void clear_gigantic_page(struct page *page,
3731                                 unsigned long addr,
3732                                 unsigned int pages_per_huge_page)
3733 {
3734         int i;
3735         struct page *p = page;
3736
3737         might_sleep();
3738         for (i = 0; i < pages_per_huge_page;
3739              i++, p = mem_map_next(p, page, i)) {
3740                 cond_resched();
3741                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3742         }
3743 }
3744 void clear_huge_page(struct page *page,
3745                      unsigned long addr, unsigned int pages_per_huge_page)
3746 {
3747         int i;
3748
3749         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3750                 clear_gigantic_page(page, addr, pages_per_huge_page);
3751                 return;
3752         }
3753
3754         might_sleep();
3755         for (i = 0; i < pages_per_huge_page; i++) {
3756                 cond_resched();
3757                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3758         }
3759 }
3760
3761 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3762                                     unsigned long addr,
3763                                     struct vm_area_struct *vma,
3764                                     unsigned int pages_per_huge_page)
3765 {
3766         int i;
3767         struct page *dst_base = dst;
3768         struct page *src_base = src;
3769
3770         for (i = 0; i < pages_per_huge_page; ) {
3771                 cond_resched();
3772                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3773
3774                 i++;
3775                 dst = mem_map_next(dst, dst_base, i);
3776                 src = mem_map_next(src, src_base, i);
3777         }
3778 }
3779
3780 void copy_user_huge_page(struct page *dst, struct page *src,
3781                          unsigned long addr, struct vm_area_struct *vma,
3782                          unsigned int pages_per_huge_page)
3783 {
3784         int i;
3785
3786         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3787                 copy_user_gigantic_page(dst, src, addr, vma,
3788                                         pages_per_huge_page);
3789                 return;
3790         }
3791
3792         might_sleep();
3793         for (i = 0; i < pages_per_huge_page; i++) {
3794                 cond_resched();
3795                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3796         }
3797 }
3798 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3799
3800 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3801
3802 static struct kmem_cache *page_ptl_cachep;
3803
3804 void __init ptlock_cache_init(void)
3805 {
3806         page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3807                         SLAB_PANIC, NULL);
3808 }
3809
3810 bool ptlock_alloc(struct page *page)
3811 {
3812         spinlock_t *ptl;
3813
3814         ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3815         if (!ptl)
3816                 return false;
3817         page->ptl = ptl;
3818         return true;
3819 }
3820
3821 void ptlock_free(struct page *page)
3822 {
3823         kmem_cache_free(page_ptl_cachep, page->ptl);
3824 }
3825 #endif