Merge branch 'for-next' of git://git.samba.org/sfrench/cifs-2.6
[firefly-linux-kernel-4.4.55.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 /*
122  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123  */
124 static int __init init_zero_pfn(void)
125 {
126         zero_pfn = page_to_pfn(ZERO_PAGE(0));
127         return 0;
128 }
129 core_initcall(init_zero_pfn);
130
131
132 #if defined(SPLIT_RSS_COUNTING)
133
134 void sync_mm_rss(struct mm_struct *mm)
135 {
136         int i;
137
138         for (i = 0; i < NR_MM_COUNTERS; i++) {
139                 if (current->rss_stat.count[i]) {
140                         add_mm_counter(mm, i, current->rss_stat.count[i]);
141                         current->rss_stat.count[i] = 0;
142                 }
143         }
144         current->rss_stat.events = 0;
145 }
146
147 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148 {
149         struct task_struct *task = current;
150
151         if (likely(task->mm == mm))
152                 task->rss_stat.count[member] += val;
153         else
154                 add_mm_counter(mm, member, val);
155 }
156 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159 /* sync counter once per 64 page faults */
160 #define TASK_RSS_EVENTS_THRESH  (64)
161 static void check_sync_rss_stat(struct task_struct *task)
162 {
163         if (unlikely(task != current))
164                 return;
165         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166                 sync_mm_rss(task->mm);
167 }
168 #else /* SPLIT_RSS_COUNTING */
169
170 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173 static void check_sync_rss_stat(struct task_struct *task)
174 {
175 }
176
177 #endif /* SPLIT_RSS_COUNTING */
178
179 #ifdef HAVE_GENERIC_MMU_GATHER
180
181 static int tlb_next_batch(struct mmu_gather *tlb)
182 {
183         struct mmu_gather_batch *batch;
184
185         batch = tlb->active;
186         if (batch->next) {
187                 tlb->active = batch->next;
188                 return 1;
189         }
190
191         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192                 return 0;
193
194         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195         if (!batch)
196                 return 0;
197
198         tlb->batch_count++;
199         batch->next = NULL;
200         batch->nr   = 0;
201         batch->max  = MAX_GATHER_BATCH;
202
203         tlb->active->next = batch;
204         tlb->active = batch;
205
206         return 1;
207 }
208
209 /* tlb_gather_mmu
210  *      Called to initialize an (on-stack) mmu_gather structure for page-table
211  *      tear-down from @mm. The @fullmm argument is used when @mm is without
212  *      users and we're going to destroy the full address space (exit/execve).
213  */
214 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
215 {
216         tlb->mm = mm;
217
218         /* Is it from 0 to ~0? */
219         tlb->fullmm     = !(start | (end+1));
220         tlb->need_flush_all = 0;
221         tlb->start      = start;
222         tlb->end        = end;
223         tlb->need_flush = 0;
224         tlb->local.next = NULL;
225         tlb->local.nr   = 0;
226         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
227         tlb->active     = &tlb->local;
228         tlb->batch_count = 0;
229
230 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
231         tlb->batch = NULL;
232 #endif
233 }
234
235 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
236 {
237         tlb->need_flush = 0;
238         tlb_flush(tlb);
239 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
240         tlb_table_flush(tlb);
241 #endif
242 }
243
244 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
245 {
246         struct mmu_gather_batch *batch;
247
248         for (batch = &tlb->local; batch; batch = batch->next) {
249                 free_pages_and_swap_cache(batch->pages, batch->nr);
250                 batch->nr = 0;
251         }
252         tlb->active = &tlb->local;
253 }
254
255 void tlb_flush_mmu(struct mmu_gather *tlb)
256 {
257         if (!tlb->need_flush)
258                 return;
259         tlb_flush_mmu_tlbonly(tlb);
260         tlb_flush_mmu_free(tlb);
261 }
262
263 /* tlb_finish_mmu
264  *      Called at the end of the shootdown operation to free up any resources
265  *      that were required.
266  */
267 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
268 {
269         struct mmu_gather_batch *batch, *next;
270
271         tlb_flush_mmu(tlb);
272
273         /* keep the page table cache within bounds */
274         check_pgt_cache();
275
276         for (batch = tlb->local.next; batch; batch = next) {
277                 next = batch->next;
278                 free_pages((unsigned long)batch, 0);
279         }
280         tlb->local.next = NULL;
281 }
282
283 /* __tlb_remove_page
284  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
285  *      handling the additional races in SMP caused by other CPUs caching valid
286  *      mappings in their TLBs. Returns the number of free page slots left.
287  *      When out of page slots we must call tlb_flush_mmu().
288  */
289 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
290 {
291         struct mmu_gather_batch *batch;
292
293         VM_BUG_ON(!tlb->need_flush);
294
295         batch = tlb->active;
296         batch->pages[batch->nr++] = page;
297         if (batch->nr == batch->max) {
298                 if (!tlb_next_batch(tlb))
299                         return 0;
300                 batch = tlb->active;
301         }
302         VM_BUG_ON_PAGE(batch->nr > batch->max, page);
303
304         return batch->max - batch->nr;
305 }
306
307 #endif /* HAVE_GENERIC_MMU_GATHER */
308
309 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
310
311 /*
312  * See the comment near struct mmu_table_batch.
313  */
314
315 static void tlb_remove_table_smp_sync(void *arg)
316 {
317         /* Simply deliver the interrupt */
318 }
319
320 static void tlb_remove_table_one(void *table)
321 {
322         /*
323          * This isn't an RCU grace period and hence the page-tables cannot be
324          * assumed to be actually RCU-freed.
325          *
326          * It is however sufficient for software page-table walkers that rely on
327          * IRQ disabling. See the comment near struct mmu_table_batch.
328          */
329         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
330         __tlb_remove_table(table);
331 }
332
333 static void tlb_remove_table_rcu(struct rcu_head *head)
334 {
335         struct mmu_table_batch *batch;
336         int i;
337
338         batch = container_of(head, struct mmu_table_batch, rcu);
339
340         for (i = 0; i < batch->nr; i++)
341                 __tlb_remove_table(batch->tables[i]);
342
343         free_page((unsigned long)batch);
344 }
345
346 void tlb_table_flush(struct mmu_gather *tlb)
347 {
348         struct mmu_table_batch **batch = &tlb->batch;
349
350         if (*batch) {
351                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
352                 *batch = NULL;
353         }
354 }
355
356 void tlb_remove_table(struct mmu_gather *tlb, void *table)
357 {
358         struct mmu_table_batch **batch = &tlb->batch;
359
360         tlb->need_flush = 1;
361
362         /*
363          * When there's less then two users of this mm there cannot be a
364          * concurrent page-table walk.
365          */
366         if (atomic_read(&tlb->mm->mm_users) < 2) {
367                 __tlb_remove_table(table);
368                 return;
369         }
370
371         if (*batch == NULL) {
372                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
373                 if (*batch == NULL) {
374                         tlb_remove_table_one(table);
375                         return;
376                 }
377                 (*batch)->nr = 0;
378         }
379         (*batch)->tables[(*batch)->nr++] = table;
380         if ((*batch)->nr == MAX_TABLE_BATCH)
381                 tlb_table_flush(tlb);
382 }
383
384 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
385
386 /*
387  * Note: this doesn't free the actual pages themselves. That
388  * has been handled earlier when unmapping all the memory regions.
389  */
390 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
391                            unsigned long addr)
392 {
393         pgtable_t token = pmd_pgtable(*pmd);
394         pmd_clear(pmd);
395         pte_free_tlb(tlb, token, addr);
396         atomic_long_dec(&tlb->mm->nr_ptes);
397 }
398
399 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
400                                 unsigned long addr, unsigned long end,
401                                 unsigned long floor, unsigned long ceiling)
402 {
403         pmd_t *pmd;
404         unsigned long next;
405         unsigned long start;
406
407         start = addr;
408         pmd = pmd_offset(pud, addr);
409         do {
410                 next = pmd_addr_end(addr, end);
411                 if (pmd_none_or_clear_bad(pmd))
412                         continue;
413                 free_pte_range(tlb, pmd, addr);
414         } while (pmd++, addr = next, addr != end);
415
416         start &= PUD_MASK;
417         if (start < floor)
418                 return;
419         if (ceiling) {
420                 ceiling &= PUD_MASK;
421                 if (!ceiling)
422                         return;
423         }
424         if (end - 1 > ceiling - 1)
425                 return;
426
427         pmd = pmd_offset(pud, start);
428         pud_clear(pud);
429         pmd_free_tlb(tlb, pmd, start);
430 }
431
432 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
433                                 unsigned long addr, unsigned long end,
434                                 unsigned long floor, unsigned long ceiling)
435 {
436         pud_t *pud;
437         unsigned long next;
438         unsigned long start;
439
440         start = addr;
441         pud = pud_offset(pgd, addr);
442         do {
443                 next = pud_addr_end(addr, end);
444                 if (pud_none_or_clear_bad(pud))
445                         continue;
446                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
447         } while (pud++, addr = next, addr != end);
448
449         start &= PGDIR_MASK;
450         if (start < floor)
451                 return;
452         if (ceiling) {
453                 ceiling &= PGDIR_MASK;
454                 if (!ceiling)
455                         return;
456         }
457         if (end - 1 > ceiling - 1)
458                 return;
459
460         pud = pud_offset(pgd, start);
461         pgd_clear(pgd);
462         pud_free_tlb(tlb, pud, start);
463 }
464
465 /*
466  * This function frees user-level page tables of a process.
467  */
468 void free_pgd_range(struct mmu_gather *tlb,
469                         unsigned long addr, unsigned long end,
470                         unsigned long floor, unsigned long ceiling)
471 {
472         pgd_t *pgd;
473         unsigned long next;
474
475         /*
476          * The next few lines have given us lots of grief...
477          *
478          * Why are we testing PMD* at this top level?  Because often
479          * there will be no work to do at all, and we'd prefer not to
480          * go all the way down to the bottom just to discover that.
481          *
482          * Why all these "- 1"s?  Because 0 represents both the bottom
483          * of the address space and the top of it (using -1 for the
484          * top wouldn't help much: the masks would do the wrong thing).
485          * The rule is that addr 0 and floor 0 refer to the bottom of
486          * the address space, but end 0 and ceiling 0 refer to the top
487          * Comparisons need to use "end - 1" and "ceiling - 1" (though
488          * that end 0 case should be mythical).
489          *
490          * Wherever addr is brought up or ceiling brought down, we must
491          * be careful to reject "the opposite 0" before it confuses the
492          * subsequent tests.  But what about where end is brought down
493          * by PMD_SIZE below? no, end can't go down to 0 there.
494          *
495          * Whereas we round start (addr) and ceiling down, by different
496          * masks at different levels, in order to test whether a table
497          * now has no other vmas using it, so can be freed, we don't
498          * bother to round floor or end up - the tests don't need that.
499          */
500
501         addr &= PMD_MASK;
502         if (addr < floor) {
503                 addr += PMD_SIZE;
504                 if (!addr)
505                         return;
506         }
507         if (ceiling) {
508                 ceiling &= PMD_MASK;
509                 if (!ceiling)
510                         return;
511         }
512         if (end - 1 > ceiling - 1)
513                 end -= PMD_SIZE;
514         if (addr > end - 1)
515                 return;
516
517         pgd = pgd_offset(tlb->mm, addr);
518         do {
519                 next = pgd_addr_end(addr, end);
520                 if (pgd_none_or_clear_bad(pgd))
521                         continue;
522                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
523         } while (pgd++, addr = next, addr != end);
524 }
525
526 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
527                 unsigned long floor, unsigned long ceiling)
528 {
529         while (vma) {
530                 struct vm_area_struct *next = vma->vm_next;
531                 unsigned long addr = vma->vm_start;
532
533                 /*
534                  * Hide vma from rmap and truncate_pagecache before freeing
535                  * pgtables
536                  */
537                 unlink_anon_vmas(vma);
538                 unlink_file_vma(vma);
539
540                 if (is_vm_hugetlb_page(vma)) {
541                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
542                                 floor, next? next->vm_start: ceiling);
543                 } else {
544                         /*
545                          * Optimization: gather nearby vmas into one call down
546                          */
547                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
548                                && !is_vm_hugetlb_page(next)) {
549                                 vma = next;
550                                 next = vma->vm_next;
551                                 unlink_anon_vmas(vma);
552                                 unlink_file_vma(vma);
553                         }
554                         free_pgd_range(tlb, addr, vma->vm_end,
555                                 floor, next? next->vm_start: ceiling);
556                 }
557                 vma = next;
558         }
559 }
560
561 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
562                 pmd_t *pmd, unsigned long address)
563 {
564         spinlock_t *ptl;
565         pgtable_t new = pte_alloc_one(mm, address);
566         int wait_split_huge_page;
567         if (!new)
568                 return -ENOMEM;
569
570         /*
571          * Ensure all pte setup (eg. pte page lock and page clearing) are
572          * visible before the pte is made visible to other CPUs by being
573          * put into page tables.
574          *
575          * The other side of the story is the pointer chasing in the page
576          * table walking code (when walking the page table without locking;
577          * ie. most of the time). Fortunately, these data accesses consist
578          * of a chain of data-dependent loads, meaning most CPUs (alpha
579          * being the notable exception) will already guarantee loads are
580          * seen in-order. See the alpha page table accessors for the
581          * smp_read_barrier_depends() barriers in page table walking code.
582          */
583         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
584
585         ptl = pmd_lock(mm, pmd);
586         wait_split_huge_page = 0;
587         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
588                 atomic_long_inc(&mm->nr_ptes);
589                 pmd_populate(mm, pmd, new);
590                 new = NULL;
591         } else if (unlikely(pmd_trans_splitting(*pmd)))
592                 wait_split_huge_page = 1;
593         spin_unlock(ptl);
594         if (new)
595                 pte_free(mm, new);
596         if (wait_split_huge_page)
597                 wait_split_huge_page(vma->anon_vma, pmd);
598         return 0;
599 }
600
601 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
602 {
603         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
604         if (!new)
605                 return -ENOMEM;
606
607         smp_wmb(); /* See comment in __pte_alloc */
608
609         spin_lock(&init_mm.page_table_lock);
610         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
611                 pmd_populate_kernel(&init_mm, pmd, new);
612                 new = NULL;
613         } else
614                 VM_BUG_ON(pmd_trans_splitting(*pmd));
615         spin_unlock(&init_mm.page_table_lock);
616         if (new)
617                 pte_free_kernel(&init_mm, new);
618         return 0;
619 }
620
621 static inline void init_rss_vec(int *rss)
622 {
623         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
624 }
625
626 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
627 {
628         int i;
629
630         if (current->mm == mm)
631                 sync_mm_rss(mm);
632         for (i = 0; i < NR_MM_COUNTERS; i++)
633                 if (rss[i])
634                         add_mm_counter(mm, i, rss[i]);
635 }
636
637 /*
638  * This function is called to print an error when a bad pte
639  * is found. For example, we might have a PFN-mapped pte in
640  * a region that doesn't allow it.
641  *
642  * The calling function must still handle the error.
643  */
644 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
645                           pte_t pte, struct page *page)
646 {
647         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
648         pud_t *pud = pud_offset(pgd, addr);
649         pmd_t *pmd = pmd_offset(pud, addr);
650         struct address_space *mapping;
651         pgoff_t index;
652         static unsigned long resume;
653         static unsigned long nr_shown;
654         static unsigned long nr_unshown;
655
656         /*
657          * Allow a burst of 60 reports, then keep quiet for that minute;
658          * or allow a steady drip of one report per second.
659          */
660         if (nr_shown == 60) {
661                 if (time_before(jiffies, resume)) {
662                         nr_unshown++;
663                         return;
664                 }
665                 if (nr_unshown) {
666                         printk(KERN_ALERT
667                                 "BUG: Bad page map: %lu messages suppressed\n",
668                                 nr_unshown);
669                         nr_unshown = 0;
670                 }
671                 nr_shown = 0;
672         }
673         if (nr_shown++ == 0)
674                 resume = jiffies + 60 * HZ;
675
676         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
677         index = linear_page_index(vma, addr);
678
679         printk(KERN_ALERT
680                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
681                 current->comm,
682                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
683         if (page)
684                 dump_page(page, "bad pte");
685         printk(KERN_ALERT
686                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
687                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
688         /*
689          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
690          */
691         if (vma->vm_ops)
692                 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
693                        vma->vm_ops->fault);
694         if (vma->vm_file)
695                 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
696                        vma->vm_file->f_op->mmap);
697         dump_stack();
698         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
699 }
700
701 /*
702  * vm_normal_page -- This function gets the "struct page" associated with a pte.
703  *
704  * "Special" mappings do not wish to be associated with a "struct page" (either
705  * it doesn't exist, or it exists but they don't want to touch it). In this
706  * case, NULL is returned here. "Normal" mappings do have a struct page.
707  *
708  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
709  * pte bit, in which case this function is trivial. Secondly, an architecture
710  * may not have a spare pte bit, which requires a more complicated scheme,
711  * described below.
712  *
713  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
714  * special mapping (even if there are underlying and valid "struct pages").
715  * COWed pages of a VM_PFNMAP are always normal.
716  *
717  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
718  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
719  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
720  * mapping will always honor the rule
721  *
722  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
723  *
724  * And for normal mappings this is false.
725  *
726  * This restricts such mappings to be a linear translation from virtual address
727  * to pfn. To get around this restriction, we allow arbitrary mappings so long
728  * as the vma is not a COW mapping; in that case, we know that all ptes are
729  * special (because none can have been COWed).
730  *
731  *
732  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
733  *
734  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
735  * page" backing, however the difference is that _all_ pages with a struct
736  * page (that is, those where pfn_valid is true) are refcounted and considered
737  * normal pages by the VM. The disadvantage is that pages are refcounted
738  * (which can be slower and simply not an option for some PFNMAP users). The
739  * advantage is that we don't have to follow the strict linearity rule of
740  * PFNMAP mappings in order to support COWable mappings.
741  *
742  */
743 #ifdef __HAVE_ARCH_PTE_SPECIAL
744 # define HAVE_PTE_SPECIAL 1
745 #else
746 # define HAVE_PTE_SPECIAL 0
747 #endif
748 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
749                                 pte_t pte)
750 {
751         unsigned long pfn = pte_pfn(pte);
752
753         if (HAVE_PTE_SPECIAL) {
754                 if (likely(!pte_special(pte) || pte_numa(pte)))
755                         goto check_pfn;
756                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
757                         return NULL;
758                 if (!is_zero_pfn(pfn))
759                         print_bad_pte(vma, addr, pte, NULL);
760                 return NULL;
761         }
762
763         /* !HAVE_PTE_SPECIAL case follows: */
764
765         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
766                 if (vma->vm_flags & VM_MIXEDMAP) {
767                         if (!pfn_valid(pfn))
768                                 return NULL;
769                         goto out;
770                 } else {
771                         unsigned long off;
772                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
773                         if (pfn == vma->vm_pgoff + off)
774                                 return NULL;
775                         if (!is_cow_mapping(vma->vm_flags))
776                                 return NULL;
777                 }
778         }
779
780 check_pfn:
781         if (unlikely(pfn > highest_memmap_pfn)) {
782                 print_bad_pte(vma, addr, pte, NULL);
783                 return NULL;
784         }
785
786         if (is_zero_pfn(pfn))
787                 return NULL;
788
789         /*
790          * NOTE! We still have PageReserved() pages in the page tables.
791          * eg. VDSO mappings can cause them to exist.
792          */
793 out:
794         return pfn_to_page(pfn);
795 }
796
797 /*
798  * copy one vm_area from one task to the other. Assumes the page tables
799  * already present in the new task to be cleared in the whole range
800  * covered by this vma.
801  */
802
803 static inline unsigned long
804 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
805                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
806                 unsigned long addr, int *rss)
807 {
808         unsigned long vm_flags = vma->vm_flags;
809         pte_t pte = *src_pte;
810         struct page *page;
811
812         /* pte contains position in swap or file, so copy. */
813         if (unlikely(!pte_present(pte))) {
814                 if (!pte_file(pte)) {
815                         swp_entry_t entry = pte_to_swp_entry(pte);
816
817                         if (swap_duplicate(entry) < 0)
818                                 return entry.val;
819
820                         /* make sure dst_mm is on swapoff's mmlist. */
821                         if (unlikely(list_empty(&dst_mm->mmlist))) {
822                                 spin_lock(&mmlist_lock);
823                                 if (list_empty(&dst_mm->mmlist))
824                                         list_add(&dst_mm->mmlist,
825                                                  &src_mm->mmlist);
826                                 spin_unlock(&mmlist_lock);
827                         }
828                         if (likely(!non_swap_entry(entry)))
829                                 rss[MM_SWAPENTS]++;
830                         else if (is_migration_entry(entry)) {
831                                 page = migration_entry_to_page(entry);
832
833                                 if (PageAnon(page))
834                                         rss[MM_ANONPAGES]++;
835                                 else
836                                         rss[MM_FILEPAGES]++;
837
838                                 if (is_write_migration_entry(entry) &&
839                                     is_cow_mapping(vm_flags)) {
840                                         /*
841                                          * COW mappings require pages in both
842                                          * parent and child to be set to read.
843                                          */
844                                         make_migration_entry_read(&entry);
845                                         pte = swp_entry_to_pte(entry);
846                                         if (pte_swp_soft_dirty(*src_pte))
847                                                 pte = pte_swp_mksoft_dirty(pte);
848                                         set_pte_at(src_mm, addr, src_pte, pte);
849                                 }
850                         }
851                 }
852                 goto out_set_pte;
853         }
854
855         /*
856          * If it's a COW mapping, write protect it both
857          * in the parent and the child
858          */
859         if (is_cow_mapping(vm_flags)) {
860                 ptep_set_wrprotect(src_mm, addr, src_pte);
861                 pte = pte_wrprotect(pte);
862         }
863
864         /*
865          * If it's a shared mapping, mark it clean in
866          * the child
867          */
868         if (vm_flags & VM_SHARED)
869                 pte = pte_mkclean(pte);
870         pte = pte_mkold(pte);
871
872         page = vm_normal_page(vma, addr, pte);
873         if (page) {
874                 get_page(page);
875                 page_dup_rmap(page);
876                 if (PageAnon(page))
877                         rss[MM_ANONPAGES]++;
878                 else
879                         rss[MM_FILEPAGES]++;
880         }
881
882 out_set_pte:
883         set_pte_at(dst_mm, addr, dst_pte, pte);
884         return 0;
885 }
886
887 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
888                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
889                    unsigned long addr, unsigned long end)
890 {
891         pte_t *orig_src_pte, *orig_dst_pte;
892         pte_t *src_pte, *dst_pte;
893         spinlock_t *src_ptl, *dst_ptl;
894         int progress = 0;
895         int rss[NR_MM_COUNTERS];
896         swp_entry_t entry = (swp_entry_t){0};
897
898 again:
899         init_rss_vec(rss);
900
901         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
902         if (!dst_pte)
903                 return -ENOMEM;
904         src_pte = pte_offset_map(src_pmd, addr);
905         src_ptl = pte_lockptr(src_mm, src_pmd);
906         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
907         orig_src_pte = src_pte;
908         orig_dst_pte = dst_pte;
909         arch_enter_lazy_mmu_mode();
910
911         do {
912                 /*
913                  * We are holding two locks at this point - either of them
914                  * could generate latencies in another task on another CPU.
915                  */
916                 if (progress >= 32) {
917                         progress = 0;
918                         if (need_resched() ||
919                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
920                                 break;
921                 }
922                 if (pte_none(*src_pte)) {
923                         progress++;
924                         continue;
925                 }
926                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
927                                                         vma, addr, rss);
928                 if (entry.val)
929                         break;
930                 progress += 8;
931         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
932
933         arch_leave_lazy_mmu_mode();
934         spin_unlock(src_ptl);
935         pte_unmap(orig_src_pte);
936         add_mm_rss_vec(dst_mm, rss);
937         pte_unmap_unlock(orig_dst_pte, dst_ptl);
938         cond_resched();
939
940         if (entry.val) {
941                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
942                         return -ENOMEM;
943                 progress = 0;
944         }
945         if (addr != end)
946                 goto again;
947         return 0;
948 }
949
950 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
951                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
952                 unsigned long addr, unsigned long end)
953 {
954         pmd_t *src_pmd, *dst_pmd;
955         unsigned long next;
956
957         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
958         if (!dst_pmd)
959                 return -ENOMEM;
960         src_pmd = pmd_offset(src_pud, addr);
961         do {
962                 next = pmd_addr_end(addr, end);
963                 if (pmd_trans_huge(*src_pmd)) {
964                         int err;
965                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
966                         err = copy_huge_pmd(dst_mm, src_mm,
967                                             dst_pmd, src_pmd, addr, vma);
968                         if (err == -ENOMEM)
969                                 return -ENOMEM;
970                         if (!err)
971                                 continue;
972                         /* fall through */
973                 }
974                 if (pmd_none_or_clear_bad(src_pmd))
975                         continue;
976                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
977                                                 vma, addr, next))
978                         return -ENOMEM;
979         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
980         return 0;
981 }
982
983 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
984                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
985                 unsigned long addr, unsigned long end)
986 {
987         pud_t *src_pud, *dst_pud;
988         unsigned long next;
989
990         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
991         if (!dst_pud)
992                 return -ENOMEM;
993         src_pud = pud_offset(src_pgd, addr);
994         do {
995                 next = pud_addr_end(addr, end);
996                 if (pud_none_or_clear_bad(src_pud))
997                         continue;
998                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
999                                                 vma, addr, next))
1000                         return -ENOMEM;
1001         } while (dst_pud++, src_pud++, addr = next, addr != end);
1002         return 0;
1003 }
1004
1005 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1006                 struct vm_area_struct *vma)
1007 {
1008         pgd_t *src_pgd, *dst_pgd;
1009         unsigned long next;
1010         unsigned long addr = vma->vm_start;
1011         unsigned long end = vma->vm_end;
1012         unsigned long mmun_start;       /* For mmu_notifiers */
1013         unsigned long mmun_end;         /* For mmu_notifiers */
1014         bool is_cow;
1015         int ret;
1016
1017         /*
1018          * Don't copy ptes where a page fault will fill them correctly.
1019          * Fork becomes much lighter when there are big shared or private
1020          * readonly mappings. The tradeoff is that copy_page_range is more
1021          * efficient than faulting.
1022          */
1023         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1024                                VM_PFNMAP | VM_MIXEDMAP))) {
1025                 if (!vma->anon_vma)
1026                         return 0;
1027         }
1028
1029         if (is_vm_hugetlb_page(vma))
1030                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1031
1032         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1033                 /*
1034                  * We do not free on error cases below as remove_vma
1035                  * gets called on error from higher level routine
1036                  */
1037                 ret = track_pfn_copy(vma);
1038                 if (ret)
1039                         return ret;
1040         }
1041
1042         /*
1043          * We need to invalidate the secondary MMU mappings only when
1044          * there could be a permission downgrade on the ptes of the
1045          * parent mm. And a permission downgrade will only happen if
1046          * is_cow_mapping() returns true.
1047          */
1048         is_cow = is_cow_mapping(vma->vm_flags);
1049         mmun_start = addr;
1050         mmun_end   = end;
1051         if (is_cow)
1052                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1053                                                     mmun_end);
1054
1055         ret = 0;
1056         dst_pgd = pgd_offset(dst_mm, addr);
1057         src_pgd = pgd_offset(src_mm, addr);
1058         do {
1059                 next = pgd_addr_end(addr, end);
1060                 if (pgd_none_or_clear_bad(src_pgd))
1061                         continue;
1062                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1063                                             vma, addr, next))) {
1064                         ret = -ENOMEM;
1065                         break;
1066                 }
1067         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1068
1069         if (is_cow)
1070                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1071         return ret;
1072 }
1073
1074 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1075                                 struct vm_area_struct *vma, pmd_t *pmd,
1076                                 unsigned long addr, unsigned long end,
1077                                 struct zap_details *details)
1078 {
1079         struct mm_struct *mm = tlb->mm;
1080         int force_flush = 0;
1081         int rss[NR_MM_COUNTERS];
1082         spinlock_t *ptl;
1083         pte_t *start_pte;
1084         pte_t *pte;
1085
1086 again:
1087         init_rss_vec(rss);
1088         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1089         pte = start_pte;
1090         arch_enter_lazy_mmu_mode();
1091         do {
1092                 pte_t ptent = *pte;
1093                 if (pte_none(ptent)) {
1094                         continue;
1095                 }
1096
1097                 if (pte_present(ptent)) {
1098                         struct page *page;
1099
1100                         page = vm_normal_page(vma, addr, ptent);
1101                         if (unlikely(details) && page) {
1102                                 /*
1103                                  * unmap_shared_mapping_pages() wants to
1104                                  * invalidate cache without truncating:
1105                                  * unmap shared but keep private pages.
1106                                  */
1107                                 if (details->check_mapping &&
1108                                     details->check_mapping != page->mapping)
1109                                         continue;
1110                                 /*
1111                                  * Each page->index must be checked when
1112                                  * invalidating or truncating nonlinear.
1113                                  */
1114                                 if (details->nonlinear_vma &&
1115                                     (page->index < details->first_index ||
1116                                      page->index > details->last_index))
1117                                         continue;
1118                         }
1119                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1120                                                         tlb->fullmm);
1121                         tlb_remove_tlb_entry(tlb, pte, addr);
1122                         if (unlikely(!page))
1123                                 continue;
1124                         if (unlikely(details) && details->nonlinear_vma
1125                             && linear_page_index(details->nonlinear_vma,
1126                                                 addr) != page->index) {
1127                                 pte_t ptfile = pgoff_to_pte(page->index);
1128                                 if (pte_soft_dirty(ptent))
1129                                         pte_file_mksoft_dirty(ptfile);
1130                                 set_pte_at(mm, addr, pte, ptfile);
1131                         }
1132                         if (PageAnon(page))
1133                                 rss[MM_ANONPAGES]--;
1134                         else {
1135                                 if (pte_dirty(ptent)) {
1136                                         force_flush = 1;
1137                                         set_page_dirty(page);
1138                                 }
1139                                 if (pte_young(ptent) &&
1140                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1141                                         mark_page_accessed(page);
1142                                 rss[MM_FILEPAGES]--;
1143                         }
1144                         page_remove_rmap(page);
1145                         if (unlikely(page_mapcount(page) < 0))
1146                                 print_bad_pte(vma, addr, ptent, page);
1147                         if (unlikely(!__tlb_remove_page(tlb, page))) {
1148                                 force_flush = 1;
1149                                 break;
1150                         }
1151                         continue;
1152                 }
1153                 /*
1154                  * If details->check_mapping, we leave swap entries;
1155                  * if details->nonlinear_vma, we leave file entries.
1156                  */
1157                 if (unlikely(details))
1158                         continue;
1159                 if (pte_file(ptent)) {
1160                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1161                                 print_bad_pte(vma, addr, ptent, NULL);
1162                 } else {
1163                         swp_entry_t entry = pte_to_swp_entry(ptent);
1164
1165                         if (!non_swap_entry(entry))
1166                                 rss[MM_SWAPENTS]--;
1167                         else if (is_migration_entry(entry)) {
1168                                 struct page *page;
1169
1170                                 page = migration_entry_to_page(entry);
1171
1172                                 if (PageAnon(page))
1173                                         rss[MM_ANONPAGES]--;
1174                                 else
1175                                         rss[MM_FILEPAGES]--;
1176                         }
1177                         if (unlikely(!free_swap_and_cache(entry)))
1178                                 print_bad_pte(vma, addr, ptent, NULL);
1179                 }
1180                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1181         } while (pte++, addr += PAGE_SIZE, addr != end);
1182
1183         add_mm_rss_vec(mm, rss);
1184         arch_leave_lazy_mmu_mode();
1185
1186         /* Do the actual TLB flush before dropping ptl */
1187         if (force_flush) {
1188                 unsigned long old_end;
1189
1190                 /*
1191                  * Flush the TLB just for the previous segment,
1192                  * then update the range to be the remaining
1193                  * TLB range.
1194                  */
1195                 old_end = tlb->end;
1196                 tlb->end = addr;
1197                 tlb_flush_mmu_tlbonly(tlb);
1198                 tlb->start = addr;
1199                 tlb->end = old_end;
1200         }
1201         pte_unmap_unlock(start_pte, ptl);
1202
1203         /*
1204          * If we forced a TLB flush (either due to running out of
1205          * batch buffers or because we needed to flush dirty TLB
1206          * entries before releasing the ptl), free the batched
1207          * memory too. Restart if we didn't do everything.
1208          */
1209         if (force_flush) {
1210                 force_flush = 0;
1211                 tlb_flush_mmu_free(tlb);
1212
1213                 if (addr != end)
1214                         goto again;
1215         }
1216
1217         return addr;
1218 }
1219
1220 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1221                                 struct vm_area_struct *vma, pud_t *pud,
1222                                 unsigned long addr, unsigned long end,
1223                                 struct zap_details *details)
1224 {
1225         pmd_t *pmd;
1226         unsigned long next;
1227
1228         pmd = pmd_offset(pud, addr);
1229         do {
1230                 next = pmd_addr_end(addr, end);
1231                 if (pmd_trans_huge(*pmd)) {
1232                         if (next - addr != HPAGE_PMD_SIZE) {
1233 #ifdef CONFIG_DEBUG_VM
1234                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1235                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1236                                                 __func__, addr, end,
1237                                                 vma->vm_start,
1238                                                 vma->vm_end);
1239                                         BUG();
1240                                 }
1241 #endif
1242                                 split_huge_page_pmd(vma, addr, pmd);
1243                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1244                                 goto next;
1245                         /* fall through */
1246                 }
1247                 /*
1248                  * Here there can be other concurrent MADV_DONTNEED or
1249                  * trans huge page faults running, and if the pmd is
1250                  * none or trans huge it can change under us. This is
1251                  * because MADV_DONTNEED holds the mmap_sem in read
1252                  * mode.
1253                  */
1254                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1255                         goto next;
1256                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1257 next:
1258                 cond_resched();
1259         } while (pmd++, addr = next, addr != end);
1260
1261         return addr;
1262 }
1263
1264 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1265                                 struct vm_area_struct *vma, pgd_t *pgd,
1266                                 unsigned long addr, unsigned long end,
1267                                 struct zap_details *details)
1268 {
1269         pud_t *pud;
1270         unsigned long next;
1271
1272         pud = pud_offset(pgd, addr);
1273         do {
1274                 next = pud_addr_end(addr, end);
1275                 if (pud_none_or_clear_bad(pud))
1276                         continue;
1277                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1278         } while (pud++, addr = next, addr != end);
1279
1280         return addr;
1281 }
1282
1283 static void unmap_page_range(struct mmu_gather *tlb,
1284                              struct vm_area_struct *vma,
1285                              unsigned long addr, unsigned long end,
1286                              struct zap_details *details)
1287 {
1288         pgd_t *pgd;
1289         unsigned long next;
1290
1291         if (details && !details->check_mapping && !details->nonlinear_vma)
1292                 details = NULL;
1293
1294         BUG_ON(addr >= end);
1295         tlb_start_vma(tlb, vma);
1296         pgd = pgd_offset(vma->vm_mm, addr);
1297         do {
1298                 next = pgd_addr_end(addr, end);
1299                 if (pgd_none_or_clear_bad(pgd))
1300                         continue;
1301                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1302         } while (pgd++, addr = next, addr != end);
1303         tlb_end_vma(tlb, vma);
1304 }
1305
1306
1307 static void unmap_single_vma(struct mmu_gather *tlb,
1308                 struct vm_area_struct *vma, unsigned long start_addr,
1309                 unsigned long end_addr,
1310                 struct zap_details *details)
1311 {
1312         unsigned long start = max(vma->vm_start, start_addr);
1313         unsigned long end;
1314
1315         if (start >= vma->vm_end)
1316                 return;
1317         end = min(vma->vm_end, end_addr);
1318         if (end <= vma->vm_start)
1319                 return;
1320
1321         if (vma->vm_file)
1322                 uprobe_munmap(vma, start, end);
1323
1324         if (unlikely(vma->vm_flags & VM_PFNMAP))
1325                 untrack_pfn(vma, 0, 0);
1326
1327         if (start != end) {
1328                 if (unlikely(is_vm_hugetlb_page(vma))) {
1329                         /*
1330                          * It is undesirable to test vma->vm_file as it
1331                          * should be non-null for valid hugetlb area.
1332                          * However, vm_file will be NULL in the error
1333                          * cleanup path of mmap_region. When
1334                          * hugetlbfs ->mmap method fails,
1335                          * mmap_region() nullifies vma->vm_file
1336                          * before calling this function to clean up.
1337                          * Since no pte has actually been setup, it is
1338                          * safe to do nothing in this case.
1339                          */
1340                         if (vma->vm_file) {
1341                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1342                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1343                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1344                         }
1345                 } else
1346                         unmap_page_range(tlb, vma, start, end, details);
1347         }
1348 }
1349
1350 /**
1351  * unmap_vmas - unmap a range of memory covered by a list of vma's
1352  * @tlb: address of the caller's struct mmu_gather
1353  * @vma: the starting vma
1354  * @start_addr: virtual address at which to start unmapping
1355  * @end_addr: virtual address at which to end unmapping
1356  *
1357  * Unmap all pages in the vma list.
1358  *
1359  * Only addresses between `start' and `end' will be unmapped.
1360  *
1361  * The VMA list must be sorted in ascending virtual address order.
1362  *
1363  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1364  * range after unmap_vmas() returns.  So the only responsibility here is to
1365  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1366  * drops the lock and schedules.
1367  */
1368 void unmap_vmas(struct mmu_gather *tlb,
1369                 struct vm_area_struct *vma, unsigned long start_addr,
1370                 unsigned long end_addr)
1371 {
1372         struct mm_struct *mm = vma->vm_mm;
1373
1374         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1375         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1376                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1377         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1378 }
1379
1380 /**
1381  * zap_page_range - remove user pages in a given range
1382  * @vma: vm_area_struct holding the applicable pages
1383  * @start: starting address of pages to zap
1384  * @size: number of bytes to zap
1385  * @details: details of nonlinear truncation or shared cache invalidation
1386  *
1387  * Caller must protect the VMA list
1388  */
1389 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1390                 unsigned long size, struct zap_details *details)
1391 {
1392         struct mm_struct *mm = vma->vm_mm;
1393         struct mmu_gather tlb;
1394         unsigned long end = start + size;
1395
1396         lru_add_drain();
1397         tlb_gather_mmu(&tlb, mm, start, end);
1398         update_hiwater_rss(mm);
1399         mmu_notifier_invalidate_range_start(mm, start, end);
1400         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1401                 unmap_single_vma(&tlb, vma, start, end, details);
1402         mmu_notifier_invalidate_range_end(mm, start, end);
1403         tlb_finish_mmu(&tlb, start, end);
1404 }
1405
1406 /**
1407  * zap_page_range_single - remove user pages in a given range
1408  * @vma: vm_area_struct holding the applicable pages
1409  * @address: starting address of pages to zap
1410  * @size: number of bytes to zap
1411  * @details: details of nonlinear truncation or shared cache invalidation
1412  *
1413  * The range must fit into one VMA.
1414  */
1415 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1416                 unsigned long size, struct zap_details *details)
1417 {
1418         struct mm_struct *mm = vma->vm_mm;
1419         struct mmu_gather tlb;
1420         unsigned long end = address + size;
1421
1422         lru_add_drain();
1423         tlb_gather_mmu(&tlb, mm, address, end);
1424         update_hiwater_rss(mm);
1425         mmu_notifier_invalidate_range_start(mm, address, end);
1426         unmap_single_vma(&tlb, vma, address, end, details);
1427         mmu_notifier_invalidate_range_end(mm, address, end);
1428         tlb_finish_mmu(&tlb, address, end);
1429 }
1430
1431 /**
1432  * zap_vma_ptes - remove ptes mapping the vma
1433  * @vma: vm_area_struct holding ptes to be zapped
1434  * @address: starting address of pages to zap
1435  * @size: number of bytes to zap
1436  *
1437  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1438  *
1439  * The entire address range must be fully contained within the vma.
1440  *
1441  * Returns 0 if successful.
1442  */
1443 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1444                 unsigned long size)
1445 {
1446         if (address < vma->vm_start || address + size > vma->vm_end ||
1447                         !(vma->vm_flags & VM_PFNMAP))
1448                 return -1;
1449         zap_page_range_single(vma, address, size, NULL);
1450         return 0;
1451 }
1452 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1453
1454 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1455                         spinlock_t **ptl)
1456 {
1457         pgd_t * pgd = pgd_offset(mm, addr);
1458         pud_t * pud = pud_alloc(mm, pgd, addr);
1459         if (pud) {
1460                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1461                 if (pmd) {
1462                         VM_BUG_ON(pmd_trans_huge(*pmd));
1463                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1464                 }
1465         }
1466         return NULL;
1467 }
1468
1469 /*
1470  * This is the old fallback for page remapping.
1471  *
1472  * For historical reasons, it only allows reserved pages. Only
1473  * old drivers should use this, and they needed to mark their
1474  * pages reserved for the old functions anyway.
1475  */
1476 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1477                         struct page *page, pgprot_t prot)
1478 {
1479         struct mm_struct *mm = vma->vm_mm;
1480         int retval;
1481         pte_t *pte;
1482         spinlock_t *ptl;
1483
1484         retval = -EINVAL;
1485         if (PageAnon(page))
1486                 goto out;
1487         retval = -ENOMEM;
1488         flush_dcache_page(page);
1489         pte = get_locked_pte(mm, addr, &ptl);
1490         if (!pte)
1491                 goto out;
1492         retval = -EBUSY;
1493         if (!pte_none(*pte))
1494                 goto out_unlock;
1495
1496         /* Ok, finally just insert the thing.. */
1497         get_page(page);
1498         inc_mm_counter_fast(mm, MM_FILEPAGES);
1499         page_add_file_rmap(page);
1500         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1501
1502         retval = 0;
1503         pte_unmap_unlock(pte, ptl);
1504         return retval;
1505 out_unlock:
1506         pte_unmap_unlock(pte, ptl);
1507 out:
1508         return retval;
1509 }
1510
1511 /**
1512  * vm_insert_page - insert single page into user vma
1513  * @vma: user vma to map to
1514  * @addr: target user address of this page
1515  * @page: source kernel page
1516  *
1517  * This allows drivers to insert individual pages they've allocated
1518  * into a user vma.
1519  *
1520  * The page has to be a nice clean _individual_ kernel allocation.
1521  * If you allocate a compound page, you need to have marked it as
1522  * such (__GFP_COMP), or manually just split the page up yourself
1523  * (see split_page()).
1524  *
1525  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1526  * took an arbitrary page protection parameter. This doesn't allow
1527  * that. Your vma protection will have to be set up correctly, which
1528  * means that if you want a shared writable mapping, you'd better
1529  * ask for a shared writable mapping!
1530  *
1531  * The page does not need to be reserved.
1532  *
1533  * Usually this function is called from f_op->mmap() handler
1534  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1535  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1536  * function from other places, for example from page-fault handler.
1537  */
1538 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1539                         struct page *page)
1540 {
1541         if (addr < vma->vm_start || addr >= vma->vm_end)
1542                 return -EFAULT;
1543         if (!page_count(page))
1544                 return -EINVAL;
1545         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1546                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1547                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1548                 vma->vm_flags |= VM_MIXEDMAP;
1549         }
1550         return insert_page(vma, addr, page, vma->vm_page_prot);
1551 }
1552 EXPORT_SYMBOL(vm_insert_page);
1553
1554 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1555                         unsigned long pfn, pgprot_t prot)
1556 {
1557         struct mm_struct *mm = vma->vm_mm;
1558         int retval;
1559         pte_t *pte, entry;
1560         spinlock_t *ptl;
1561
1562         retval = -ENOMEM;
1563         pte = get_locked_pte(mm, addr, &ptl);
1564         if (!pte)
1565                 goto out;
1566         retval = -EBUSY;
1567         if (!pte_none(*pte))
1568                 goto out_unlock;
1569
1570         /* Ok, finally just insert the thing.. */
1571         entry = pte_mkspecial(pfn_pte(pfn, prot));
1572         set_pte_at(mm, addr, pte, entry);
1573         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1574
1575         retval = 0;
1576 out_unlock:
1577         pte_unmap_unlock(pte, ptl);
1578 out:
1579         return retval;
1580 }
1581
1582 /**
1583  * vm_insert_pfn - insert single pfn into user vma
1584  * @vma: user vma to map to
1585  * @addr: target user address of this page
1586  * @pfn: source kernel pfn
1587  *
1588  * Similar to vm_insert_page, this allows drivers to insert individual pages
1589  * they've allocated into a user vma. Same comments apply.
1590  *
1591  * This function should only be called from a vm_ops->fault handler, and
1592  * in that case the handler should return NULL.
1593  *
1594  * vma cannot be a COW mapping.
1595  *
1596  * As this is called only for pages that do not currently exist, we
1597  * do not need to flush old virtual caches or the TLB.
1598  */
1599 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1600                         unsigned long pfn)
1601 {
1602         int ret;
1603         pgprot_t pgprot = vma->vm_page_prot;
1604         /*
1605          * Technically, architectures with pte_special can avoid all these
1606          * restrictions (same for remap_pfn_range).  However we would like
1607          * consistency in testing and feature parity among all, so we should
1608          * try to keep these invariants in place for everybody.
1609          */
1610         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1611         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1612                                                 (VM_PFNMAP|VM_MIXEDMAP));
1613         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1614         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1615
1616         if (addr < vma->vm_start || addr >= vma->vm_end)
1617                 return -EFAULT;
1618         if (track_pfn_insert(vma, &pgprot, pfn))
1619                 return -EINVAL;
1620
1621         ret = insert_pfn(vma, addr, pfn, pgprot);
1622
1623         return ret;
1624 }
1625 EXPORT_SYMBOL(vm_insert_pfn);
1626
1627 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1628                         unsigned long pfn)
1629 {
1630         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1631
1632         if (addr < vma->vm_start || addr >= vma->vm_end)
1633                 return -EFAULT;
1634
1635         /*
1636          * If we don't have pte special, then we have to use the pfn_valid()
1637          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1638          * refcount the page if pfn_valid is true (hence insert_page rather
1639          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1640          * without pte special, it would there be refcounted as a normal page.
1641          */
1642         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1643                 struct page *page;
1644
1645                 page = pfn_to_page(pfn);
1646                 return insert_page(vma, addr, page, vma->vm_page_prot);
1647         }
1648         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1649 }
1650 EXPORT_SYMBOL(vm_insert_mixed);
1651
1652 /*
1653  * maps a range of physical memory into the requested pages. the old
1654  * mappings are removed. any references to nonexistent pages results
1655  * in null mappings (currently treated as "copy-on-access")
1656  */
1657 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1658                         unsigned long addr, unsigned long end,
1659                         unsigned long pfn, pgprot_t prot)
1660 {
1661         pte_t *pte;
1662         spinlock_t *ptl;
1663
1664         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1665         if (!pte)
1666                 return -ENOMEM;
1667         arch_enter_lazy_mmu_mode();
1668         do {
1669                 BUG_ON(!pte_none(*pte));
1670                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1671                 pfn++;
1672         } while (pte++, addr += PAGE_SIZE, addr != end);
1673         arch_leave_lazy_mmu_mode();
1674         pte_unmap_unlock(pte - 1, ptl);
1675         return 0;
1676 }
1677
1678 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1679                         unsigned long addr, unsigned long end,
1680                         unsigned long pfn, pgprot_t prot)
1681 {
1682         pmd_t *pmd;
1683         unsigned long next;
1684
1685         pfn -= addr >> PAGE_SHIFT;
1686         pmd = pmd_alloc(mm, pud, addr);
1687         if (!pmd)
1688                 return -ENOMEM;
1689         VM_BUG_ON(pmd_trans_huge(*pmd));
1690         do {
1691                 next = pmd_addr_end(addr, end);
1692                 if (remap_pte_range(mm, pmd, addr, next,
1693                                 pfn + (addr >> PAGE_SHIFT), prot))
1694                         return -ENOMEM;
1695         } while (pmd++, addr = next, addr != end);
1696         return 0;
1697 }
1698
1699 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1700                         unsigned long addr, unsigned long end,
1701                         unsigned long pfn, pgprot_t prot)
1702 {
1703         pud_t *pud;
1704         unsigned long next;
1705
1706         pfn -= addr >> PAGE_SHIFT;
1707         pud = pud_alloc(mm, pgd, addr);
1708         if (!pud)
1709                 return -ENOMEM;
1710         do {
1711                 next = pud_addr_end(addr, end);
1712                 if (remap_pmd_range(mm, pud, addr, next,
1713                                 pfn + (addr >> PAGE_SHIFT), prot))
1714                         return -ENOMEM;
1715         } while (pud++, addr = next, addr != end);
1716         return 0;
1717 }
1718
1719 /**
1720  * remap_pfn_range - remap kernel memory to userspace
1721  * @vma: user vma to map to
1722  * @addr: target user address to start at
1723  * @pfn: physical address of kernel memory
1724  * @size: size of map area
1725  * @prot: page protection flags for this mapping
1726  *
1727  *  Note: this is only safe if the mm semaphore is held when called.
1728  */
1729 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1730                     unsigned long pfn, unsigned long size, pgprot_t prot)
1731 {
1732         pgd_t *pgd;
1733         unsigned long next;
1734         unsigned long end = addr + PAGE_ALIGN(size);
1735         struct mm_struct *mm = vma->vm_mm;
1736         int err;
1737
1738         /*
1739          * Physically remapped pages are special. Tell the
1740          * rest of the world about it:
1741          *   VM_IO tells people not to look at these pages
1742          *      (accesses can have side effects).
1743          *   VM_PFNMAP tells the core MM that the base pages are just
1744          *      raw PFN mappings, and do not have a "struct page" associated
1745          *      with them.
1746          *   VM_DONTEXPAND
1747          *      Disable vma merging and expanding with mremap().
1748          *   VM_DONTDUMP
1749          *      Omit vma from core dump, even when VM_IO turned off.
1750          *
1751          * There's a horrible special case to handle copy-on-write
1752          * behaviour that some programs depend on. We mark the "original"
1753          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1754          * See vm_normal_page() for details.
1755          */
1756         if (is_cow_mapping(vma->vm_flags)) {
1757                 if (addr != vma->vm_start || end != vma->vm_end)
1758                         return -EINVAL;
1759                 vma->vm_pgoff = pfn;
1760         }
1761
1762         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1763         if (err)
1764                 return -EINVAL;
1765
1766         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1767
1768         BUG_ON(addr >= end);
1769         pfn -= addr >> PAGE_SHIFT;
1770         pgd = pgd_offset(mm, addr);
1771         flush_cache_range(vma, addr, end);
1772         do {
1773                 next = pgd_addr_end(addr, end);
1774                 err = remap_pud_range(mm, pgd, addr, next,
1775                                 pfn + (addr >> PAGE_SHIFT), prot);
1776                 if (err)
1777                         break;
1778         } while (pgd++, addr = next, addr != end);
1779
1780         if (err)
1781                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1782
1783         return err;
1784 }
1785 EXPORT_SYMBOL(remap_pfn_range);
1786
1787 /**
1788  * vm_iomap_memory - remap memory to userspace
1789  * @vma: user vma to map to
1790  * @start: start of area
1791  * @len: size of area
1792  *
1793  * This is a simplified io_remap_pfn_range() for common driver use. The
1794  * driver just needs to give us the physical memory range to be mapped,
1795  * we'll figure out the rest from the vma information.
1796  *
1797  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1798  * whatever write-combining details or similar.
1799  */
1800 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1801 {
1802         unsigned long vm_len, pfn, pages;
1803
1804         /* Check that the physical memory area passed in looks valid */
1805         if (start + len < start)
1806                 return -EINVAL;
1807         /*
1808          * You *really* shouldn't map things that aren't page-aligned,
1809          * but we've historically allowed it because IO memory might
1810          * just have smaller alignment.
1811          */
1812         len += start & ~PAGE_MASK;
1813         pfn = start >> PAGE_SHIFT;
1814         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1815         if (pfn + pages < pfn)
1816                 return -EINVAL;
1817
1818         /* We start the mapping 'vm_pgoff' pages into the area */
1819         if (vma->vm_pgoff > pages)
1820                 return -EINVAL;
1821         pfn += vma->vm_pgoff;
1822         pages -= vma->vm_pgoff;
1823
1824         /* Can we fit all of the mapping? */
1825         vm_len = vma->vm_end - vma->vm_start;
1826         if (vm_len >> PAGE_SHIFT > pages)
1827                 return -EINVAL;
1828
1829         /* Ok, let it rip */
1830         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1831 }
1832 EXPORT_SYMBOL(vm_iomap_memory);
1833
1834 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1835                                      unsigned long addr, unsigned long end,
1836                                      pte_fn_t fn, void *data)
1837 {
1838         pte_t *pte;
1839         int err;
1840         pgtable_t token;
1841         spinlock_t *uninitialized_var(ptl);
1842
1843         pte = (mm == &init_mm) ?
1844                 pte_alloc_kernel(pmd, addr) :
1845                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1846         if (!pte)
1847                 return -ENOMEM;
1848
1849         BUG_ON(pmd_huge(*pmd));
1850
1851         arch_enter_lazy_mmu_mode();
1852
1853         token = pmd_pgtable(*pmd);
1854
1855         do {
1856                 err = fn(pte++, token, addr, data);
1857                 if (err)
1858                         break;
1859         } while (addr += PAGE_SIZE, addr != end);
1860
1861         arch_leave_lazy_mmu_mode();
1862
1863         if (mm != &init_mm)
1864                 pte_unmap_unlock(pte-1, ptl);
1865         return err;
1866 }
1867
1868 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1869                                      unsigned long addr, unsigned long end,
1870                                      pte_fn_t fn, void *data)
1871 {
1872         pmd_t *pmd;
1873         unsigned long next;
1874         int err;
1875
1876         BUG_ON(pud_huge(*pud));
1877
1878         pmd = pmd_alloc(mm, pud, addr);
1879         if (!pmd)
1880                 return -ENOMEM;
1881         do {
1882                 next = pmd_addr_end(addr, end);
1883                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1884                 if (err)
1885                         break;
1886         } while (pmd++, addr = next, addr != end);
1887         return err;
1888 }
1889
1890 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1891                                      unsigned long addr, unsigned long end,
1892                                      pte_fn_t fn, void *data)
1893 {
1894         pud_t *pud;
1895         unsigned long next;
1896         int err;
1897
1898         pud = pud_alloc(mm, pgd, addr);
1899         if (!pud)
1900                 return -ENOMEM;
1901         do {
1902                 next = pud_addr_end(addr, end);
1903                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1904                 if (err)
1905                         break;
1906         } while (pud++, addr = next, addr != end);
1907         return err;
1908 }
1909
1910 /*
1911  * Scan a region of virtual memory, filling in page tables as necessary
1912  * and calling a provided function on each leaf page table.
1913  */
1914 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1915                         unsigned long size, pte_fn_t fn, void *data)
1916 {
1917         pgd_t *pgd;
1918         unsigned long next;
1919         unsigned long end = addr + size;
1920         int err;
1921
1922         BUG_ON(addr >= end);
1923         pgd = pgd_offset(mm, addr);
1924         do {
1925                 next = pgd_addr_end(addr, end);
1926                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1927                 if (err)
1928                         break;
1929         } while (pgd++, addr = next, addr != end);
1930
1931         return err;
1932 }
1933 EXPORT_SYMBOL_GPL(apply_to_page_range);
1934
1935 /*
1936  * handle_pte_fault chooses page fault handler according to an entry
1937  * which was read non-atomically.  Before making any commitment, on
1938  * those architectures or configurations (e.g. i386 with PAE) which
1939  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
1940  * must check under lock before unmapping the pte and proceeding
1941  * (but do_wp_page is only called after already making such a check;
1942  * and do_anonymous_page can safely check later on).
1943  */
1944 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1945                                 pte_t *page_table, pte_t orig_pte)
1946 {
1947         int same = 1;
1948 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1949         if (sizeof(pte_t) > sizeof(unsigned long)) {
1950                 spinlock_t *ptl = pte_lockptr(mm, pmd);
1951                 spin_lock(ptl);
1952                 same = pte_same(*page_table, orig_pte);
1953                 spin_unlock(ptl);
1954         }
1955 #endif
1956         pte_unmap(page_table);
1957         return same;
1958 }
1959
1960 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1961 {
1962         debug_dma_assert_idle(src);
1963
1964         /*
1965          * If the source page was a PFN mapping, we don't have
1966          * a "struct page" for it. We do a best-effort copy by
1967          * just copying from the original user address. If that
1968          * fails, we just zero-fill it. Live with it.
1969          */
1970         if (unlikely(!src)) {
1971                 void *kaddr = kmap_atomic(dst);
1972                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1973
1974                 /*
1975                  * This really shouldn't fail, because the page is there
1976                  * in the page tables. But it might just be unreadable,
1977                  * in which case we just give up and fill the result with
1978                  * zeroes.
1979                  */
1980                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1981                         clear_page(kaddr);
1982                 kunmap_atomic(kaddr);
1983                 flush_dcache_page(dst);
1984         } else
1985                 copy_user_highpage(dst, src, va, vma);
1986 }
1987
1988 /*
1989  * Notify the address space that the page is about to become writable so that
1990  * it can prohibit this or wait for the page to get into an appropriate state.
1991  *
1992  * We do this without the lock held, so that it can sleep if it needs to.
1993  */
1994 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1995                unsigned long address)
1996 {
1997         struct vm_fault vmf;
1998         int ret;
1999
2000         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2001         vmf.pgoff = page->index;
2002         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2003         vmf.page = page;
2004
2005         ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2006         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2007                 return ret;
2008         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2009                 lock_page(page);
2010                 if (!page->mapping) {
2011                         unlock_page(page);
2012                         return 0; /* retry */
2013                 }
2014                 ret |= VM_FAULT_LOCKED;
2015         } else
2016                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2017         return ret;
2018 }
2019
2020 /*
2021  * This routine handles present pages, when users try to write
2022  * to a shared page. It is done by copying the page to a new address
2023  * and decrementing the shared-page counter for the old page.
2024  *
2025  * Note that this routine assumes that the protection checks have been
2026  * done by the caller (the low-level page fault routine in most cases).
2027  * Thus we can safely just mark it writable once we've done any necessary
2028  * COW.
2029  *
2030  * We also mark the page dirty at this point even though the page will
2031  * change only once the write actually happens. This avoids a few races,
2032  * and potentially makes it more efficient.
2033  *
2034  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2035  * but allow concurrent faults), with pte both mapped and locked.
2036  * We return with mmap_sem still held, but pte unmapped and unlocked.
2037  */
2038 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2039                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2040                 spinlock_t *ptl, pte_t orig_pte)
2041         __releases(ptl)
2042 {
2043         struct page *old_page, *new_page = NULL;
2044         pte_t entry;
2045         int ret = 0;
2046         int page_mkwrite = 0;
2047         struct page *dirty_page = NULL;
2048         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2049         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2050         struct mem_cgroup *memcg;
2051
2052         old_page = vm_normal_page(vma, address, orig_pte);
2053         if (!old_page) {
2054                 /*
2055                  * VM_MIXEDMAP !pfn_valid() case
2056                  *
2057                  * We should not cow pages in a shared writeable mapping.
2058                  * Just mark the pages writable as we can't do any dirty
2059                  * accounting on raw pfn maps.
2060                  */
2061                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2062                                      (VM_WRITE|VM_SHARED))
2063                         goto reuse;
2064                 goto gotten;
2065         }
2066
2067         /*
2068          * Take out anonymous pages first, anonymous shared vmas are
2069          * not dirty accountable.
2070          */
2071         if (PageAnon(old_page) && !PageKsm(old_page)) {
2072                 if (!trylock_page(old_page)) {
2073                         page_cache_get(old_page);
2074                         pte_unmap_unlock(page_table, ptl);
2075                         lock_page(old_page);
2076                         page_table = pte_offset_map_lock(mm, pmd, address,
2077                                                          &ptl);
2078                         if (!pte_same(*page_table, orig_pte)) {
2079                                 unlock_page(old_page);
2080                                 goto unlock;
2081                         }
2082                         page_cache_release(old_page);
2083                 }
2084                 if (reuse_swap_page(old_page)) {
2085                         /*
2086                          * The page is all ours.  Move it to our anon_vma so
2087                          * the rmap code will not search our parent or siblings.
2088                          * Protected against the rmap code by the page lock.
2089                          */
2090                         page_move_anon_rmap(old_page, vma, address);
2091                         unlock_page(old_page);
2092                         goto reuse;
2093                 }
2094                 unlock_page(old_page);
2095         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2096                                         (VM_WRITE|VM_SHARED))) {
2097                 /*
2098                  * Only catch write-faults on shared writable pages,
2099                  * read-only shared pages can get COWed by
2100                  * get_user_pages(.write=1, .force=1).
2101                  */
2102                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2103                         int tmp;
2104                         page_cache_get(old_page);
2105                         pte_unmap_unlock(page_table, ptl);
2106                         tmp = do_page_mkwrite(vma, old_page, address);
2107                         if (unlikely(!tmp || (tmp &
2108                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2109                                 page_cache_release(old_page);
2110                                 return tmp;
2111                         }
2112                         /*
2113                          * Since we dropped the lock we need to revalidate
2114                          * the PTE as someone else may have changed it.  If
2115                          * they did, we just return, as we can count on the
2116                          * MMU to tell us if they didn't also make it writable.
2117                          */
2118                         page_table = pte_offset_map_lock(mm, pmd, address,
2119                                                          &ptl);
2120                         if (!pte_same(*page_table, orig_pte)) {
2121                                 unlock_page(old_page);
2122                                 goto unlock;
2123                         }
2124
2125                         page_mkwrite = 1;
2126                 }
2127                 dirty_page = old_page;
2128                 get_page(dirty_page);
2129
2130 reuse:
2131                 /*
2132                  * Clear the pages cpupid information as the existing
2133                  * information potentially belongs to a now completely
2134                  * unrelated process.
2135                  */
2136                 if (old_page)
2137                         page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2138
2139                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2140                 entry = pte_mkyoung(orig_pte);
2141                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2142                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2143                         update_mmu_cache(vma, address, page_table);
2144                 pte_unmap_unlock(page_table, ptl);
2145                 ret |= VM_FAULT_WRITE;
2146
2147                 if (!dirty_page)
2148                         return ret;
2149
2150                 /*
2151                  * Yes, Virginia, this is actually required to prevent a race
2152                  * with clear_page_dirty_for_io() from clearing the page dirty
2153                  * bit after it clear all dirty ptes, but before a racing
2154                  * do_wp_page installs a dirty pte.
2155                  *
2156                  * do_shared_fault is protected similarly.
2157                  */
2158                 if (!page_mkwrite) {
2159                         wait_on_page_locked(dirty_page);
2160                         set_page_dirty_balance(dirty_page);
2161                         /* file_update_time outside page_lock */
2162                         if (vma->vm_file)
2163                                 file_update_time(vma->vm_file);
2164                 }
2165                 put_page(dirty_page);
2166                 if (page_mkwrite) {
2167                         struct address_space *mapping = dirty_page->mapping;
2168
2169                         set_page_dirty(dirty_page);
2170                         unlock_page(dirty_page);
2171                         page_cache_release(dirty_page);
2172                         if (mapping)    {
2173                                 /*
2174                                  * Some device drivers do not set page.mapping
2175                                  * but still dirty their pages
2176                                  */
2177                                 balance_dirty_pages_ratelimited(mapping);
2178                         }
2179                 }
2180
2181                 return ret;
2182         }
2183
2184         /*
2185          * Ok, we need to copy. Oh, well..
2186          */
2187         page_cache_get(old_page);
2188 gotten:
2189         pte_unmap_unlock(page_table, ptl);
2190
2191         if (unlikely(anon_vma_prepare(vma)))
2192                 goto oom;
2193
2194         if (is_zero_pfn(pte_pfn(orig_pte))) {
2195                 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2196                 if (!new_page)
2197                         goto oom;
2198         } else {
2199                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2200                 if (!new_page)
2201                         goto oom;
2202                 cow_user_page(new_page, old_page, address, vma);
2203         }
2204         __SetPageUptodate(new_page);
2205
2206         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
2207                 goto oom_free_new;
2208
2209         mmun_start  = address & PAGE_MASK;
2210         mmun_end    = mmun_start + PAGE_SIZE;
2211         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2212
2213         /*
2214          * Re-check the pte - we dropped the lock
2215          */
2216         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2217         if (likely(pte_same(*page_table, orig_pte))) {
2218                 if (old_page) {
2219                         if (!PageAnon(old_page)) {
2220                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2221                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2222                         }
2223                 } else
2224                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2225                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2226                 entry = mk_pte(new_page, vma->vm_page_prot);
2227                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2228                 /*
2229                  * Clear the pte entry and flush it first, before updating the
2230                  * pte with the new entry. This will avoid a race condition
2231                  * seen in the presence of one thread doing SMC and another
2232                  * thread doing COW.
2233                  */
2234                 ptep_clear_flush(vma, address, page_table);
2235                 page_add_new_anon_rmap(new_page, vma, address);
2236                 mem_cgroup_commit_charge(new_page, memcg, false);
2237                 lru_cache_add_active_or_unevictable(new_page, vma);
2238                 /*
2239                  * We call the notify macro here because, when using secondary
2240                  * mmu page tables (such as kvm shadow page tables), we want the
2241                  * new page to be mapped directly into the secondary page table.
2242                  */
2243                 set_pte_at_notify(mm, address, page_table, entry);
2244                 update_mmu_cache(vma, address, page_table);
2245                 if (old_page) {
2246                         /*
2247                          * Only after switching the pte to the new page may
2248                          * we remove the mapcount here. Otherwise another
2249                          * process may come and find the rmap count decremented
2250                          * before the pte is switched to the new page, and
2251                          * "reuse" the old page writing into it while our pte
2252                          * here still points into it and can be read by other
2253                          * threads.
2254                          *
2255                          * The critical issue is to order this
2256                          * page_remove_rmap with the ptp_clear_flush above.
2257                          * Those stores are ordered by (if nothing else,)
2258                          * the barrier present in the atomic_add_negative
2259                          * in page_remove_rmap.
2260                          *
2261                          * Then the TLB flush in ptep_clear_flush ensures that
2262                          * no process can access the old page before the
2263                          * decremented mapcount is visible. And the old page
2264                          * cannot be reused until after the decremented
2265                          * mapcount is visible. So transitively, TLBs to
2266                          * old page will be flushed before it can be reused.
2267                          */
2268                         page_remove_rmap(old_page);
2269                 }
2270
2271                 /* Free the old page.. */
2272                 new_page = old_page;
2273                 ret |= VM_FAULT_WRITE;
2274         } else
2275                 mem_cgroup_cancel_charge(new_page, memcg);
2276
2277         if (new_page)
2278                 page_cache_release(new_page);
2279 unlock:
2280         pte_unmap_unlock(page_table, ptl);
2281         if (mmun_end > mmun_start)
2282                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2283         if (old_page) {
2284                 /*
2285                  * Don't let another task, with possibly unlocked vma,
2286                  * keep the mlocked page.
2287                  */
2288                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2289                         lock_page(old_page);    /* LRU manipulation */
2290                         munlock_vma_page(old_page);
2291                         unlock_page(old_page);
2292                 }
2293                 page_cache_release(old_page);
2294         }
2295         return ret;
2296 oom_free_new:
2297         page_cache_release(new_page);
2298 oom:
2299         if (old_page)
2300                 page_cache_release(old_page);
2301         return VM_FAULT_OOM;
2302 }
2303
2304 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2305                 unsigned long start_addr, unsigned long end_addr,
2306                 struct zap_details *details)
2307 {
2308         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2309 }
2310
2311 static inline void unmap_mapping_range_tree(struct rb_root *root,
2312                                             struct zap_details *details)
2313 {
2314         struct vm_area_struct *vma;
2315         pgoff_t vba, vea, zba, zea;
2316
2317         vma_interval_tree_foreach(vma, root,
2318                         details->first_index, details->last_index) {
2319
2320                 vba = vma->vm_pgoff;
2321                 vea = vba + vma_pages(vma) - 1;
2322                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2323                 zba = details->first_index;
2324                 if (zba < vba)
2325                         zba = vba;
2326                 zea = details->last_index;
2327                 if (zea > vea)
2328                         zea = vea;
2329
2330                 unmap_mapping_range_vma(vma,
2331                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2332                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2333                                 details);
2334         }
2335 }
2336
2337 static inline void unmap_mapping_range_list(struct list_head *head,
2338                                             struct zap_details *details)
2339 {
2340         struct vm_area_struct *vma;
2341
2342         /*
2343          * In nonlinear VMAs there is no correspondence between virtual address
2344          * offset and file offset.  So we must perform an exhaustive search
2345          * across *all* the pages in each nonlinear VMA, not just the pages
2346          * whose virtual address lies outside the file truncation point.
2347          */
2348         list_for_each_entry(vma, head, shared.nonlinear) {
2349                 details->nonlinear_vma = vma;
2350                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2351         }
2352 }
2353
2354 /**
2355  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2356  * @mapping: the address space containing mmaps to be unmapped.
2357  * @holebegin: byte in first page to unmap, relative to the start of
2358  * the underlying file.  This will be rounded down to a PAGE_SIZE
2359  * boundary.  Note that this is different from truncate_pagecache(), which
2360  * must keep the partial page.  In contrast, we must get rid of
2361  * partial pages.
2362  * @holelen: size of prospective hole in bytes.  This will be rounded
2363  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2364  * end of the file.
2365  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2366  * but 0 when invalidating pagecache, don't throw away private data.
2367  */
2368 void unmap_mapping_range(struct address_space *mapping,
2369                 loff_t const holebegin, loff_t const holelen, int even_cows)
2370 {
2371         struct zap_details details;
2372         pgoff_t hba = holebegin >> PAGE_SHIFT;
2373         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2374
2375         /* Check for overflow. */
2376         if (sizeof(holelen) > sizeof(hlen)) {
2377                 long long holeend =
2378                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2379                 if (holeend & ~(long long)ULONG_MAX)
2380                         hlen = ULONG_MAX - hba + 1;
2381         }
2382
2383         details.check_mapping = even_cows? NULL: mapping;
2384         details.nonlinear_vma = NULL;
2385         details.first_index = hba;
2386         details.last_index = hba + hlen - 1;
2387         if (details.last_index < details.first_index)
2388                 details.last_index = ULONG_MAX;
2389
2390
2391         mutex_lock(&mapping->i_mmap_mutex);
2392         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2393                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2394         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2395                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2396         mutex_unlock(&mapping->i_mmap_mutex);
2397 }
2398 EXPORT_SYMBOL(unmap_mapping_range);
2399
2400 /*
2401  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2402  * but allow concurrent faults), and pte mapped but not yet locked.
2403  * We return with pte unmapped and unlocked.
2404  *
2405  * We return with the mmap_sem locked or unlocked in the same cases
2406  * as does filemap_fault().
2407  */
2408 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2409                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2410                 unsigned int flags, pte_t orig_pte)
2411 {
2412         spinlock_t *ptl;
2413         struct page *page, *swapcache;
2414         struct mem_cgroup *memcg;
2415         swp_entry_t entry;
2416         pte_t pte;
2417         int locked;
2418         int exclusive = 0;
2419         int ret = 0;
2420
2421         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2422                 goto out;
2423
2424         entry = pte_to_swp_entry(orig_pte);
2425         if (unlikely(non_swap_entry(entry))) {
2426                 if (is_migration_entry(entry)) {
2427                         migration_entry_wait(mm, pmd, address);
2428                 } else if (is_hwpoison_entry(entry)) {
2429                         ret = VM_FAULT_HWPOISON;
2430                 } else {
2431                         print_bad_pte(vma, address, orig_pte, NULL);
2432                         ret = VM_FAULT_SIGBUS;
2433                 }
2434                 goto out;
2435         }
2436         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2437         page = lookup_swap_cache(entry);
2438         if (!page) {
2439                 page = swapin_readahead(entry,
2440                                         GFP_HIGHUSER_MOVABLE, vma, address);
2441                 if (!page) {
2442                         /*
2443                          * Back out if somebody else faulted in this pte
2444                          * while we released the pte lock.
2445                          */
2446                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2447                         if (likely(pte_same(*page_table, orig_pte)))
2448                                 ret = VM_FAULT_OOM;
2449                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2450                         goto unlock;
2451                 }
2452
2453                 /* Had to read the page from swap area: Major fault */
2454                 ret = VM_FAULT_MAJOR;
2455                 count_vm_event(PGMAJFAULT);
2456                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2457         } else if (PageHWPoison(page)) {
2458                 /*
2459                  * hwpoisoned dirty swapcache pages are kept for killing
2460                  * owner processes (which may be unknown at hwpoison time)
2461                  */
2462                 ret = VM_FAULT_HWPOISON;
2463                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2464                 swapcache = page;
2465                 goto out_release;
2466         }
2467
2468         swapcache = page;
2469         locked = lock_page_or_retry(page, mm, flags);
2470
2471         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2472         if (!locked) {
2473                 ret |= VM_FAULT_RETRY;
2474                 goto out_release;
2475         }
2476
2477         /*
2478          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2479          * release the swapcache from under us.  The page pin, and pte_same
2480          * test below, are not enough to exclude that.  Even if it is still
2481          * swapcache, we need to check that the page's swap has not changed.
2482          */
2483         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2484                 goto out_page;
2485
2486         page = ksm_might_need_to_copy(page, vma, address);
2487         if (unlikely(!page)) {
2488                 ret = VM_FAULT_OOM;
2489                 page = swapcache;
2490                 goto out_page;
2491         }
2492
2493         if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
2494                 ret = VM_FAULT_OOM;
2495                 goto out_page;
2496         }
2497
2498         /*
2499          * Back out if somebody else already faulted in this pte.
2500          */
2501         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2502         if (unlikely(!pte_same(*page_table, orig_pte)))
2503                 goto out_nomap;
2504
2505         if (unlikely(!PageUptodate(page))) {
2506                 ret = VM_FAULT_SIGBUS;
2507                 goto out_nomap;
2508         }
2509
2510         /*
2511          * The page isn't present yet, go ahead with the fault.
2512          *
2513          * Be careful about the sequence of operations here.
2514          * To get its accounting right, reuse_swap_page() must be called
2515          * while the page is counted on swap but not yet in mapcount i.e.
2516          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2517          * must be called after the swap_free(), or it will never succeed.
2518          */
2519
2520         inc_mm_counter_fast(mm, MM_ANONPAGES);
2521         dec_mm_counter_fast(mm, MM_SWAPENTS);
2522         pte = mk_pte(page, vma->vm_page_prot);
2523         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2524                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2525                 flags &= ~FAULT_FLAG_WRITE;
2526                 ret |= VM_FAULT_WRITE;
2527                 exclusive = 1;
2528         }
2529         flush_icache_page(vma, page);
2530         if (pte_swp_soft_dirty(orig_pte))
2531                 pte = pte_mksoft_dirty(pte);
2532         set_pte_at(mm, address, page_table, pte);
2533         if (page == swapcache) {
2534                 do_page_add_anon_rmap(page, vma, address, exclusive);
2535                 mem_cgroup_commit_charge(page, memcg, true);
2536         } else { /* ksm created a completely new copy */
2537                 page_add_new_anon_rmap(page, vma, address);
2538                 mem_cgroup_commit_charge(page, memcg, false);
2539                 lru_cache_add_active_or_unevictable(page, vma);
2540         }
2541
2542         swap_free(entry);
2543         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2544                 try_to_free_swap(page);
2545         unlock_page(page);
2546         if (page != swapcache) {
2547                 /*
2548                  * Hold the lock to avoid the swap entry to be reused
2549                  * until we take the PT lock for the pte_same() check
2550                  * (to avoid false positives from pte_same). For
2551                  * further safety release the lock after the swap_free
2552                  * so that the swap count won't change under a
2553                  * parallel locked swapcache.
2554                  */
2555                 unlock_page(swapcache);
2556                 page_cache_release(swapcache);
2557         }
2558
2559         if (flags & FAULT_FLAG_WRITE) {
2560                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2561                 if (ret & VM_FAULT_ERROR)
2562                         ret &= VM_FAULT_ERROR;
2563                 goto out;
2564         }
2565
2566         /* No need to invalidate - it was non-present before */
2567         update_mmu_cache(vma, address, page_table);
2568 unlock:
2569         pte_unmap_unlock(page_table, ptl);
2570 out:
2571         return ret;
2572 out_nomap:
2573         mem_cgroup_cancel_charge(page, memcg);
2574         pte_unmap_unlock(page_table, ptl);
2575 out_page:
2576         unlock_page(page);
2577 out_release:
2578         page_cache_release(page);
2579         if (page != swapcache) {
2580                 unlock_page(swapcache);
2581                 page_cache_release(swapcache);
2582         }
2583         return ret;
2584 }
2585
2586 /*
2587  * This is like a special single-page "expand_{down|up}wards()",
2588  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2589  * doesn't hit another vma.
2590  */
2591 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2592 {
2593         address &= PAGE_MASK;
2594         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2595                 struct vm_area_struct *prev = vma->vm_prev;
2596
2597                 /*
2598                  * Is there a mapping abutting this one below?
2599                  *
2600                  * That's only ok if it's the same stack mapping
2601                  * that has gotten split..
2602                  */
2603                 if (prev && prev->vm_end == address)
2604                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2605
2606                 expand_downwards(vma, address - PAGE_SIZE);
2607         }
2608         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2609                 struct vm_area_struct *next = vma->vm_next;
2610
2611                 /* As VM_GROWSDOWN but s/below/above/ */
2612                 if (next && next->vm_start == address + PAGE_SIZE)
2613                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2614
2615                 expand_upwards(vma, address + PAGE_SIZE);
2616         }
2617         return 0;
2618 }
2619
2620 /*
2621  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2622  * but allow concurrent faults), and pte mapped but not yet locked.
2623  * We return with mmap_sem still held, but pte unmapped and unlocked.
2624  */
2625 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2626                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2627                 unsigned int flags)
2628 {
2629         struct mem_cgroup *memcg;
2630         struct page *page;
2631         spinlock_t *ptl;
2632         pte_t entry;
2633
2634         pte_unmap(page_table);
2635
2636         /* Check if we need to add a guard page to the stack */
2637         if (check_stack_guard_page(vma, address) < 0)
2638                 return VM_FAULT_SIGBUS;
2639
2640         /* Use the zero-page for reads */
2641         if (!(flags & FAULT_FLAG_WRITE)) {
2642                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2643                                                 vma->vm_page_prot));
2644                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2645                 if (!pte_none(*page_table))
2646                         goto unlock;
2647                 goto setpte;
2648         }
2649
2650         /* Allocate our own private page. */
2651         if (unlikely(anon_vma_prepare(vma)))
2652                 goto oom;
2653         page = alloc_zeroed_user_highpage_movable(vma, address);
2654         if (!page)
2655                 goto oom;
2656         /*
2657          * The memory barrier inside __SetPageUptodate makes sure that
2658          * preceeding stores to the page contents become visible before
2659          * the set_pte_at() write.
2660          */
2661         __SetPageUptodate(page);
2662
2663         if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
2664                 goto oom_free_page;
2665
2666         entry = mk_pte(page, vma->vm_page_prot);
2667         if (vma->vm_flags & VM_WRITE)
2668                 entry = pte_mkwrite(pte_mkdirty(entry));
2669
2670         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2671         if (!pte_none(*page_table))
2672                 goto release;
2673
2674         inc_mm_counter_fast(mm, MM_ANONPAGES);
2675         page_add_new_anon_rmap(page, vma, address);
2676         mem_cgroup_commit_charge(page, memcg, false);
2677         lru_cache_add_active_or_unevictable(page, vma);
2678 setpte:
2679         set_pte_at(mm, address, page_table, entry);
2680
2681         /* No need to invalidate - it was non-present before */
2682         update_mmu_cache(vma, address, page_table);
2683 unlock:
2684         pte_unmap_unlock(page_table, ptl);
2685         return 0;
2686 release:
2687         mem_cgroup_cancel_charge(page, memcg);
2688         page_cache_release(page);
2689         goto unlock;
2690 oom_free_page:
2691         page_cache_release(page);
2692 oom:
2693         return VM_FAULT_OOM;
2694 }
2695
2696 /*
2697  * The mmap_sem must have been held on entry, and may have been
2698  * released depending on flags and vma->vm_ops->fault() return value.
2699  * See filemap_fault() and __lock_page_retry().
2700  */
2701 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2702                 pgoff_t pgoff, unsigned int flags, struct page **page)
2703 {
2704         struct vm_fault vmf;
2705         int ret;
2706
2707         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2708         vmf.pgoff = pgoff;
2709         vmf.flags = flags;
2710         vmf.page = NULL;
2711
2712         ret = vma->vm_ops->fault(vma, &vmf);
2713         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2714                 return ret;
2715
2716         if (unlikely(PageHWPoison(vmf.page))) {
2717                 if (ret & VM_FAULT_LOCKED)
2718                         unlock_page(vmf.page);
2719                 page_cache_release(vmf.page);
2720                 return VM_FAULT_HWPOISON;
2721         }
2722
2723         if (unlikely(!(ret & VM_FAULT_LOCKED)))
2724                 lock_page(vmf.page);
2725         else
2726                 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2727
2728         *page = vmf.page;
2729         return ret;
2730 }
2731
2732 /**
2733  * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2734  *
2735  * @vma: virtual memory area
2736  * @address: user virtual address
2737  * @page: page to map
2738  * @pte: pointer to target page table entry
2739  * @write: true, if new entry is writable
2740  * @anon: true, if it's anonymous page
2741  *
2742  * Caller must hold page table lock relevant for @pte.
2743  *
2744  * Target users are page handler itself and implementations of
2745  * vm_ops->map_pages.
2746  */
2747 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2748                 struct page *page, pte_t *pte, bool write, bool anon)
2749 {
2750         pte_t entry;
2751
2752         flush_icache_page(vma, page);
2753         entry = mk_pte(page, vma->vm_page_prot);
2754         if (write)
2755                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2756         else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
2757                 entry = pte_mksoft_dirty(entry);
2758         if (anon) {
2759                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2760                 page_add_new_anon_rmap(page, vma, address);
2761         } else {
2762                 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2763                 page_add_file_rmap(page);
2764         }
2765         set_pte_at(vma->vm_mm, address, pte, entry);
2766
2767         /* no need to invalidate: a not-present page won't be cached */
2768         update_mmu_cache(vma, address, pte);
2769 }
2770
2771 static unsigned long fault_around_bytes __read_mostly =
2772         rounddown_pow_of_two(65536);
2773
2774 #ifdef CONFIG_DEBUG_FS
2775 static int fault_around_bytes_get(void *data, u64 *val)
2776 {
2777         *val = fault_around_bytes;
2778         return 0;
2779 }
2780
2781 /*
2782  * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2783  * rounded down to nearest page order. It's what do_fault_around() expects to
2784  * see.
2785  */
2786 static int fault_around_bytes_set(void *data, u64 val)
2787 {
2788         if (val / PAGE_SIZE > PTRS_PER_PTE)
2789                 return -EINVAL;
2790         if (val > PAGE_SIZE)
2791                 fault_around_bytes = rounddown_pow_of_two(val);
2792         else
2793                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2794         return 0;
2795 }
2796 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2797                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2798
2799 static int __init fault_around_debugfs(void)
2800 {
2801         void *ret;
2802
2803         ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2804                         &fault_around_bytes_fops);
2805         if (!ret)
2806                 pr_warn("Failed to create fault_around_bytes in debugfs");
2807         return 0;
2808 }
2809 late_initcall(fault_around_debugfs);
2810 #endif
2811
2812 /*
2813  * do_fault_around() tries to map few pages around the fault address. The hope
2814  * is that the pages will be needed soon and this will lower the number of
2815  * faults to handle.
2816  *
2817  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2818  * not ready to be mapped: not up-to-date, locked, etc.
2819  *
2820  * This function is called with the page table lock taken. In the split ptlock
2821  * case the page table lock only protects only those entries which belong to
2822  * the page table corresponding to the fault address.
2823  *
2824  * This function doesn't cross the VMA boundaries, in order to call map_pages()
2825  * only once.
2826  *
2827  * fault_around_pages() defines how many pages we'll try to map.
2828  * do_fault_around() expects it to return a power of two less than or equal to
2829  * PTRS_PER_PTE.
2830  *
2831  * The virtual address of the area that we map is naturally aligned to the
2832  * fault_around_pages() value (and therefore to page order).  This way it's
2833  * easier to guarantee that we don't cross page table boundaries.
2834  */
2835 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2836                 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2837 {
2838         unsigned long start_addr, nr_pages, mask;
2839         pgoff_t max_pgoff;
2840         struct vm_fault vmf;
2841         int off;
2842
2843         nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2844         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2845
2846         start_addr = max(address & mask, vma->vm_start);
2847         off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2848         pte -= off;
2849         pgoff -= off;
2850
2851         /*
2852          *  max_pgoff is either end of page table or end of vma
2853          *  or fault_around_pages() from pgoff, depending what is nearest.
2854          */
2855         max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2856                 PTRS_PER_PTE - 1;
2857         max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2858                         pgoff + nr_pages - 1);
2859
2860         /* Check if it makes any sense to call ->map_pages */
2861         while (!pte_none(*pte)) {
2862                 if (++pgoff > max_pgoff)
2863                         return;
2864                 start_addr += PAGE_SIZE;
2865                 if (start_addr >= vma->vm_end)
2866                         return;
2867                 pte++;
2868         }
2869
2870         vmf.virtual_address = (void __user *) start_addr;
2871         vmf.pte = pte;
2872         vmf.pgoff = pgoff;
2873         vmf.max_pgoff = max_pgoff;
2874         vmf.flags = flags;
2875         vma->vm_ops->map_pages(vma, &vmf);
2876 }
2877
2878 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2879                 unsigned long address, pmd_t *pmd,
2880                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2881 {
2882         struct page *fault_page;
2883         spinlock_t *ptl;
2884         pte_t *pte;
2885         int ret = 0;
2886
2887         /*
2888          * Let's call ->map_pages() first and use ->fault() as fallback
2889          * if page by the offset is not ready to be mapped (cold cache or
2890          * something).
2891          */
2892         if (vma->vm_ops->map_pages && !(flags & FAULT_FLAG_NONLINEAR) &&
2893             fault_around_bytes >> PAGE_SHIFT > 1) {
2894                 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2895                 do_fault_around(vma, address, pte, pgoff, flags);
2896                 if (!pte_same(*pte, orig_pte))
2897                         goto unlock_out;
2898                 pte_unmap_unlock(pte, ptl);
2899         }
2900
2901         ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2902         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2903                 return ret;
2904
2905         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2906         if (unlikely(!pte_same(*pte, orig_pte))) {
2907                 pte_unmap_unlock(pte, ptl);
2908                 unlock_page(fault_page);
2909                 page_cache_release(fault_page);
2910                 return ret;
2911         }
2912         do_set_pte(vma, address, fault_page, pte, false, false);
2913         unlock_page(fault_page);
2914 unlock_out:
2915         pte_unmap_unlock(pte, ptl);
2916         return ret;
2917 }
2918
2919 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2920                 unsigned long address, pmd_t *pmd,
2921                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2922 {
2923         struct page *fault_page, *new_page;
2924         struct mem_cgroup *memcg;
2925         spinlock_t *ptl;
2926         pte_t *pte;
2927         int ret;
2928
2929         if (unlikely(anon_vma_prepare(vma)))
2930                 return VM_FAULT_OOM;
2931
2932         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2933         if (!new_page)
2934                 return VM_FAULT_OOM;
2935
2936         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
2937                 page_cache_release(new_page);
2938                 return VM_FAULT_OOM;
2939         }
2940
2941         ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2942         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2943                 goto uncharge_out;
2944
2945         copy_user_highpage(new_page, fault_page, address, vma);
2946         __SetPageUptodate(new_page);
2947
2948         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2949         if (unlikely(!pte_same(*pte, orig_pte))) {
2950                 pte_unmap_unlock(pte, ptl);
2951                 unlock_page(fault_page);
2952                 page_cache_release(fault_page);
2953                 goto uncharge_out;
2954         }
2955         do_set_pte(vma, address, new_page, pte, true, true);
2956         mem_cgroup_commit_charge(new_page, memcg, false);
2957         lru_cache_add_active_or_unevictable(new_page, vma);
2958         pte_unmap_unlock(pte, ptl);
2959         unlock_page(fault_page);
2960         page_cache_release(fault_page);
2961         return ret;
2962 uncharge_out:
2963         mem_cgroup_cancel_charge(new_page, memcg);
2964         page_cache_release(new_page);
2965         return ret;
2966 }
2967
2968 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2969                 unsigned long address, pmd_t *pmd,
2970                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2971 {
2972         struct page *fault_page;
2973         struct address_space *mapping;
2974         spinlock_t *ptl;
2975         pte_t *pte;
2976         int dirtied = 0;
2977         int ret, tmp;
2978
2979         ret = __do_fault(vma, address, pgoff, flags, &fault_page);
2980         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2981                 return ret;
2982
2983         /*
2984          * Check if the backing address space wants to know that the page is
2985          * about to become writable
2986          */
2987         if (vma->vm_ops->page_mkwrite) {
2988                 unlock_page(fault_page);
2989                 tmp = do_page_mkwrite(vma, fault_page, address);
2990                 if (unlikely(!tmp ||
2991                                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2992                         page_cache_release(fault_page);
2993                         return tmp;
2994                 }
2995         }
2996
2997         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2998         if (unlikely(!pte_same(*pte, orig_pte))) {
2999                 pte_unmap_unlock(pte, ptl);
3000                 unlock_page(fault_page);
3001                 page_cache_release(fault_page);
3002                 return ret;
3003         }
3004         do_set_pte(vma, address, fault_page, pte, true, false);
3005         pte_unmap_unlock(pte, ptl);
3006
3007         if (set_page_dirty(fault_page))
3008                 dirtied = 1;
3009         mapping = fault_page->mapping;
3010         unlock_page(fault_page);
3011         if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3012                 /*
3013                  * Some device drivers do not set page.mapping but still
3014                  * dirty their pages
3015                  */
3016                 balance_dirty_pages_ratelimited(mapping);
3017         }
3018
3019         /* file_update_time outside page_lock */
3020         if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3021                 file_update_time(vma->vm_file);
3022
3023         return ret;
3024 }
3025
3026 /*
3027  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3028  * but allow concurrent faults).
3029  * The mmap_sem may have been released depending on flags and our
3030  * return value.  See filemap_fault() and __lock_page_or_retry().
3031  */
3032 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3033                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3034                 unsigned int flags, pte_t orig_pte)
3035 {
3036         pgoff_t pgoff = (((address & PAGE_MASK)
3037                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3038
3039         pte_unmap(page_table);
3040         if (!(flags & FAULT_FLAG_WRITE))
3041                 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3042                                 orig_pte);
3043         if (!(vma->vm_flags & VM_SHARED))
3044                 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3045                                 orig_pte);
3046         return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3047 }
3048
3049 /*
3050  * Fault of a previously existing named mapping. Repopulate the pte
3051  * from the encoded file_pte if possible. This enables swappable
3052  * nonlinear vmas.
3053  *
3054  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3055  * but allow concurrent faults), and pte mapped but not yet locked.
3056  * We return with pte unmapped and unlocked.
3057  * The mmap_sem may have been released depending on flags and our
3058  * return value.  See filemap_fault() and __lock_page_or_retry().
3059  */
3060 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3061                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3062                 unsigned int flags, pte_t orig_pte)
3063 {
3064         pgoff_t pgoff;
3065
3066         flags |= FAULT_FLAG_NONLINEAR;
3067
3068         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3069                 return 0;
3070
3071         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3072                 /*
3073                  * Page table corrupted: show pte and kill process.
3074                  */
3075                 print_bad_pte(vma, address, orig_pte, NULL);
3076                 return VM_FAULT_SIGBUS;
3077         }
3078
3079         pgoff = pte_to_pgoff(orig_pte);
3080         if (!(flags & FAULT_FLAG_WRITE))
3081                 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3082                                 orig_pte);
3083         if (!(vma->vm_flags & VM_SHARED))
3084                 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3085                                 orig_pte);
3086         return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3087 }
3088
3089 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3090                                 unsigned long addr, int page_nid,
3091                                 int *flags)
3092 {
3093         get_page(page);
3094
3095         count_vm_numa_event(NUMA_HINT_FAULTS);
3096         if (page_nid == numa_node_id()) {
3097                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3098                 *flags |= TNF_FAULT_LOCAL;
3099         }
3100
3101         return mpol_misplaced(page, vma, addr);
3102 }
3103
3104 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3105                    unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3106 {
3107         struct page *page = NULL;
3108         spinlock_t *ptl;
3109         int page_nid = -1;
3110         int last_cpupid;
3111         int target_nid;
3112         bool migrated = false;
3113         int flags = 0;
3114
3115         /*
3116         * The "pte" at this point cannot be used safely without
3117         * validation through pte_unmap_same(). It's of NUMA type but
3118         * the pfn may be screwed if the read is non atomic.
3119         *
3120         * ptep_modify_prot_start is not called as this is clearing
3121         * the _PAGE_NUMA bit and it is not really expected that there
3122         * would be concurrent hardware modifications to the PTE.
3123         */
3124         ptl = pte_lockptr(mm, pmd);
3125         spin_lock(ptl);
3126         if (unlikely(!pte_same(*ptep, pte))) {
3127                 pte_unmap_unlock(ptep, ptl);
3128                 goto out;
3129         }
3130
3131         pte = pte_mknonnuma(pte);
3132         set_pte_at(mm, addr, ptep, pte);
3133         update_mmu_cache(vma, addr, ptep);
3134
3135         page = vm_normal_page(vma, addr, pte);
3136         if (!page) {
3137                 pte_unmap_unlock(ptep, ptl);
3138                 return 0;
3139         }
3140         BUG_ON(is_zero_pfn(page_to_pfn(page)));
3141
3142         /*
3143          * Avoid grouping on DSO/COW pages in specific and RO pages
3144          * in general, RO pages shouldn't hurt as much anyway since
3145          * they can be in shared cache state.
3146          */
3147         if (!pte_write(pte))
3148                 flags |= TNF_NO_GROUP;
3149
3150         /*
3151          * Flag if the page is shared between multiple address spaces. This
3152          * is later used when determining whether to group tasks together
3153          */
3154         if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3155                 flags |= TNF_SHARED;
3156
3157         last_cpupid = page_cpupid_last(page);
3158         page_nid = page_to_nid(page);
3159         target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3160         pte_unmap_unlock(ptep, ptl);
3161         if (target_nid == -1) {
3162                 put_page(page);
3163                 goto out;
3164         }
3165
3166         /* Migrate to the requested node */
3167         migrated = migrate_misplaced_page(page, vma, target_nid);
3168         if (migrated) {
3169                 page_nid = target_nid;
3170                 flags |= TNF_MIGRATED;
3171         }
3172
3173 out:
3174         if (page_nid != -1)
3175                 task_numa_fault(last_cpupid, page_nid, 1, flags);
3176         return 0;
3177 }
3178
3179 /*
3180  * These routines also need to handle stuff like marking pages dirty
3181  * and/or accessed for architectures that don't do it in hardware (most
3182  * RISC architectures).  The early dirtying is also good on the i386.
3183  *
3184  * There is also a hook called "update_mmu_cache()" that architectures
3185  * with external mmu caches can use to update those (ie the Sparc or
3186  * PowerPC hashed page tables that act as extended TLBs).
3187  *
3188  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3189  * but allow concurrent faults), and pte mapped but not yet locked.
3190  * We return with pte unmapped and unlocked.
3191  *
3192  * The mmap_sem may have been released depending on flags and our
3193  * return value.  See filemap_fault() and __lock_page_or_retry().
3194  */
3195 static int handle_pte_fault(struct mm_struct *mm,
3196                      struct vm_area_struct *vma, unsigned long address,
3197                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3198 {
3199         pte_t entry;
3200         spinlock_t *ptl;
3201
3202         entry = ACCESS_ONCE(*pte);
3203         if (!pte_present(entry)) {
3204                 if (pte_none(entry)) {
3205                         if (vma->vm_ops) {
3206                                 if (likely(vma->vm_ops->fault))
3207                                         return do_linear_fault(mm, vma, address,
3208                                                 pte, pmd, flags, entry);
3209                         }
3210                         return do_anonymous_page(mm, vma, address,
3211                                                  pte, pmd, flags);
3212                 }
3213                 if (pte_file(entry))
3214                         return do_nonlinear_fault(mm, vma, address,
3215                                         pte, pmd, flags, entry);
3216                 return do_swap_page(mm, vma, address,
3217                                         pte, pmd, flags, entry);
3218         }
3219
3220         if (pte_numa(entry))
3221                 return do_numa_page(mm, vma, address, entry, pte, pmd);
3222
3223         ptl = pte_lockptr(mm, pmd);
3224         spin_lock(ptl);
3225         if (unlikely(!pte_same(*pte, entry)))
3226                 goto unlock;
3227         if (flags & FAULT_FLAG_WRITE) {
3228                 if (!pte_write(entry))
3229                         return do_wp_page(mm, vma, address,
3230                                         pte, pmd, ptl, entry);
3231                 entry = pte_mkdirty(entry);
3232         }
3233         entry = pte_mkyoung(entry);
3234         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3235                 update_mmu_cache(vma, address, pte);
3236         } else {
3237                 /*
3238                  * This is needed only for protection faults but the arch code
3239                  * is not yet telling us if this is a protection fault or not.
3240                  * This still avoids useless tlb flushes for .text page faults
3241                  * with threads.
3242                  */
3243                 if (flags & FAULT_FLAG_WRITE)
3244                         flush_tlb_fix_spurious_fault(vma, address);
3245         }
3246 unlock:
3247         pte_unmap_unlock(pte, ptl);
3248         return 0;
3249 }
3250
3251 /*
3252  * By the time we get here, we already hold the mm semaphore
3253  *
3254  * The mmap_sem may have been released depending on flags and our
3255  * return value.  See filemap_fault() and __lock_page_or_retry().
3256  */
3257 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3258                              unsigned long address, unsigned int flags)
3259 {
3260         pgd_t *pgd;
3261         pud_t *pud;
3262         pmd_t *pmd;
3263         pte_t *pte;
3264
3265         if (unlikely(is_vm_hugetlb_page(vma)))
3266                 return hugetlb_fault(mm, vma, address, flags);
3267
3268         pgd = pgd_offset(mm, address);
3269         pud = pud_alloc(mm, pgd, address);
3270         if (!pud)
3271                 return VM_FAULT_OOM;
3272         pmd = pmd_alloc(mm, pud, address);
3273         if (!pmd)
3274                 return VM_FAULT_OOM;
3275         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3276                 int ret = VM_FAULT_FALLBACK;
3277                 if (!vma->vm_ops)
3278                         ret = do_huge_pmd_anonymous_page(mm, vma, address,
3279                                         pmd, flags);
3280                 if (!(ret & VM_FAULT_FALLBACK))
3281                         return ret;
3282         } else {
3283                 pmd_t orig_pmd = *pmd;
3284                 int ret;
3285
3286                 barrier();
3287                 if (pmd_trans_huge(orig_pmd)) {
3288                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3289
3290                         /*
3291                          * If the pmd is splitting, return and retry the
3292                          * the fault.  Alternative: wait until the split
3293                          * is done, and goto retry.
3294                          */
3295                         if (pmd_trans_splitting(orig_pmd))
3296                                 return 0;
3297
3298                         if (pmd_numa(orig_pmd))
3299                                 return do_huge_pmd_numa_page(mm, vma, address,
3300                                                              orig_pmd, pmd);
3301
3302                         if (dirty && !pmd_write(orig_pmd)) {
3303                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3304                                                           orig_pmd);
3305                                 if (!(ret & VM_FAULT_FALLBACK))
3306                                         return ret;
3307                         } else {
3308                                 huge_pmd_set_accessed(mm, vma, address, pmd,
3309                                                       orig_pmd, dirty);
3310                                 return 0;
3311                         }
3312                 }
3313         }
3314
3315         /*
3316          * Use __pte_alloc instead of pte_alloc_map, because we can't
3317          * run pte_offset_map on the pmd, if an huge pmd could
3318          * materialize from under us from a different thread.
3319          */
3320         if (unlikely(pmd_none(*pmd)) &&
3321             unlikely(__pte_alloc(mm, vma, pmd, address)))
3322                 return VM_FAULT_OOM;
3323         /* if an huge pmd materialized from under us just retry later */
3324         if (unlikely(pmd_trans_huge(*pmd)))
3325                 return 0;
3326         /*
3327          * A regular pmd is established and it can't morph into a huge pmd
3328          * from under us anymore at this point because we hold the mmap_sem
3329          * read mode and khugepaged takes it in write mode. So now it's
3330          * safe to run pte_offset_map().
3331          */
3332         pte = pte_offset_map(pmd, address);
3333
3334         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3335 }
3336
3337 /*
3338  * By the time we get here, we already hold the mm semaphore
3339  *
3340  * The mmap_sem may have been released depending on flags and our
3341  * return value.  See filemap_fault() and __lock_page_or_retry().
3342  */
3343 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3344                     unsigned long address, unsigned int flags)
3345 {
3346         int ret;
3347
3348         __set_current_state(TASK_RUNNING);
3349
3350         count_vm_event(PGFAULT);
3351         mem_cgroup_count_vm_event(mm, PGFAULT);
3352
3353         /* do counter updates before entering really critical section. */
3354         check_sync_rss_stat(current);
3355
3356         /*
3357          * Enable the memcg OOM handling for faults triggered in user
3358          * space.  Kernel faults are handled more gracefully.
3359          */
3360         if (flags & FAULT_FLAG_USER)
3361                 mem_cgroup_oom_enable();
3362
3363         ret = __handle_mm_fault(mm, vma, address, flags);
3364
3365         if (flags & FAULT_FLAG_USER) {
3366                 mem_cgroup_oom_disable();
3367                 /*
3368                  * The task may have entered a memcg OOM situation but
3369                  * if the allocation error was handled gracefully (no
3370                  * VM_FAULT_OOM), there is no need to kill anything.
3371                  * Just clean up the OOM state peacefully.
3372                  */
3373                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3374                         mem_cgroup_oom_synchronize(false);
3375         }
3376
3377         return ret;
3378 }
3379
3380 #ifndef __PAGETABLE_PUD_FOLDED
3381 /*
3382  * Allocate page upper directory.
3383  * We've already handled the fast-path in-line.
3384  */
3385 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3386 {
3387         pud_t *new = pud_alloc_one(mm, address);
3388         if (!new)
3389                 return -ENOMEM;
3390
3391         smp_wmb(); /* See comment in __pte_alloc */
3392
3393         spin_lock(&mm->page_table_lock);
3394         if (pgd_present(*pgd))          /* Another has populated it */
3395                 pud_free(mm, new);
3396         else
3397                 pgd_populate(mm, pgd, new);
3398         spin_unlock(&mm->page_table_lock);
3399         return 0;
3400 }
3401 #endif /* __PAGETABLE_PUD_FOLDED */
3402
3403 #ifndef __PAGETABLE_PMD_FOLDED
3404 /*
3405  * Allocate page middle directory.
3406  * We've already handled the fast-path in-line.
3407  */
3408 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3409 {
3410         pmd_t *new = pmd_alloc_one(mm, address);
3411         if (!new)
3412                 return -ENOMEM;
3413
3414         smp_wmb(); /* See comment in __pte_alloc */
3415
3416         spin_lock(&mm->page_table_lock);
3417 #ifndef __ARCH_HAS_4LEVEL_HACK
3418         if (pud_present(*pud))          /* Another has populated it */
3419                 pmd_free(mm, new);
3420         else
3421                 pud_populate(mm, pud, new);
3422 #else
3423         if (pgd_present(*pud))          /* Another has populated it */
3424                 pmd_free(mm, new);
3425         else
3426                 pgd_populate(mm, pud, new);
3427 #endif /* __ARCH_HAS_4LEVEL_HACK */
3428         spin_unlock(&mm->page_table_lock);
3429         return 0;
3430 }
3431 #endif /* __PAGETABLE_PMD_FOLDED */
3432
3433 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3434                 pte_t **ptepp, spinlock_t **ptlp)
3435 {
3436         pgd_t *pgd;
3437         pud_t *pud;
3438         pmd_t *pmd;
3439         pte_t *ptep;
3440
3441         pgd = pgd_offset(mm, address);
3442         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3443                 goto out;
3444
3445         pud = pud_offset(pgd, address);
3446         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3447                 goto out;
3448
3449         pmd = pmd_offset(pud, address);
3450         VM_BUG_ON(pmd_trans_huge(*pmd));
3451         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3452                 goto out;
3453
3454         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3455         if (pmd_huge(*pmd))
3456                 goto out;
3457
3458         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3459         if (!ptep)
3460                 goto out;
3461         if (!pte_present(*ptep))
3462                 goto unlock;
3463         *ptepp = ptep;
3464         return 0;
3465 unlock:
3466         pte_unmap_unlock(ptep, *ptlp);
3467 out:
3468         return -EINVAL;
3469 }
3470
3471 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3472                              pte_t **ptepp, spinlock_t **ptlp)
3473 {
3474         int res;
3475
3476         /* (void) is needed to make gcc happy */
3477         (void) __cond_lock(*ptlp,
3478                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
3479         return res;
3480 }
3481
3482 /**
3483  * follow_pfn - look up PFN at a user virtual address
3484  * @vma: memory mapping
3485  * @address: user virtual address
3486  * @pfn: location to store found PFN
3487  *
3488  * Only IO mappings and raw PFN mappings are allowed.
3489  *
3490  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3491  */
3492 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3493         unsigned long *pfn)
3494 {
3495         int ret = -EINVAL;
3496         spinlock_t *ptl;
3497         pte_t *ptep;
3498
3499         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3500                 return ret;
3501
3502         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3503         if (ret)
3504                 return ret;
3505         *pfn = pte_pfn(*ptep);
3506         pte_unmap_unlock(ptep, ptl);
3507         return 0;
3508 }
3509 EXPORT_SYMBOL(follow_pfn);
3510
3511 #ifdef CONFIG_HAVE_IOREMAP_PROT
3512 int follow_phys(struct vm_area_struct *vma,
3513                 unsigned long address, unsigned int flags,
3514                 unsigned long *prot, resource_size_t *phys)
3515 {
3516         int ret = -EINVAL;
3517         pte_t *ptep, pte;
3518         spinlock_t *ptl;
3519
3520         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3521                 goto out;
3522
3523         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3524                 goto out;
3525         pte = *ptep;
3526
3527         if ((flags & FOLL_WRITE) && !pte_write(pte))
3528                 goto unlock;
3529
3530         *prot = pgprot_val(pte_pgprot(pte));
3531         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3532
3533         ret = 0;
3534 unlock:
3535         pte_unmap_unlock(ptep, ptl);
3536 out:
3537         return ret;
3538 }
3539
3540 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3541                         void *buf, int len, int write)
3542 {
3543         resource_size_t phys_addr;
3544         unsigned long prot = 0;
3545         void __iomem *maddr;
3546         int offset = addr & (PAGE_SIZE-1);
3547
3548         if (follow_phys(vma, addr, write, &prot, &phys_addr))
3549                 return -EINVAL;
3550
3551         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3552         if (write)
3553                 memcpy_toio(maddr + offset, buf, len);
3554         else
3555                 memcpy_fromio(buf, maddr + offset, len);
3556         iounmap(maddr);
3557
3558         return len;
3559 }
3560 EXPORT_SYMBOL_GPL(generic_access_phys);
3561 #endif
3562
3563 /*
3564  * Access another process' address space as given in mm.  If non-NULL, use the
3565  * given task for page fault accounting.
3566  */
3567 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3568                 unsigned long addr, void *buf, int len, int write)
3569 {
3570         struct vm_area_struct *vma;
3571         void *old_buf = buf;
3572
3573         down_read(&mm->mmap_sem);
3574         /* ignore errors, just check how much was successfully transferred */
3575         while (len) {
3576                 int bytes, ret, offset;
3577                 void *maddr;
3578                 struct page *page = NULL;
3579
3580                 ret = get_user_pages(tsk, mm, addr, 1,
3581                                 write, 1, &page, &vma);
3582                 if (ret <= 0) {
3583 #ifndef CONFIG_HAVE_IOREMAP_PROT
3584                         break;
3585 #else
3586                         /*
3587                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
3588                          * we can access using slightly different code.
3589                          */
3590                         vma = find_vma(mm, addr);
3591                         if (!vma || vma->vm_start > addr)
3592                                 break;
3593                         if (vma->vm_ops && vma->vm_ops->access)
3594                                 ret = vma->vm_ops->access(vma, addr, buf,
3595                                                           len, write);
3596                         if (ret <= 0)
3597                                 break;
3598                         bytes = ret;
3599 #endif
3600                 } else {
3601                         bytes = len;
3602                         offset = addr & (PAGE_SIZE-1);
3603                         if (bytes > PAGE_SIZE-offset)
3604                                 bytes = PAGE_SIZE-offset;
3605
3606                         maddr = kmap(page);
3607                         if (write) {
3608                                 copy_to_user_page(vma, page, addr,
3609                                                   maddr + offset, buf, bytes);
3610                                 set_page_dirty_lock(page);
3611                         } else {
3612                                 copy_from_user_page(vma, page, addr,
3613                                                     buf, maddr + offset, bytes);
3614                         }
3615                         kunmap(page);
3616                         page_cache_release(page);
3617                 }
3618                 len -= bytes;
3619                 buf += bytes;
3620                 addr += bytes;
3621         }
3622         up_read(&mm->mmap_sem);
3623
3624         return buf - old_buf;
3625 }
3626
3627 /**
3628  * access_remote_vm - access another process' address space
3629  * @mm:         the mm_struct of the target address space
3630  * @addr:       start address to access
3631  * @buf:        source or destination buffer
3632  * @len:        number of bytes to transfer
3633  * @write:      whether the access is a write
3634  *
3635  * The caller must hold a reference on @mm.
3636  */
3637 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3638                 void *buf, int len, int write)
3639 {
3640         return __access_remote_vm(NULL, mm, addr, buf, len, write);
3641 }
3642
3643 /*
3644  * Access another process' address space.
3645  * Source/target buffer must be kernel space,
3646  * Do not walk the page table directly, use get_user_pages
3647  */
3648 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3649                 void *buf, int len, int write)
3650 {
3651         struct mm_struct *mm;
3652         int ret;
3653
3654         mm = get_task_mm(tsk);
3655         if (!mm)
3656                 return 0;
3657
3658         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3659         mmput(mm);
3660
3661         return ret;
3662 }
3663
3664 /*
3665  * Print the name of a VMA.
3666  */
3667 void print_vma_addr(char *prefix, unsigned long ip)
3668 {
3669         struct mm_struct *mm = current->mm;
3670         struct vm_area_struct *vma;
3671
3672         /*
3673          * Do not print if we are in atomic
3674          * contexts (in exception stacks, etc.):
3675          */
3676         if (preempt_count())
3677                 return;
3678
3679         down_read(&mm->mmap_sem);
3680         vma = find_vma(mm, ip);
3681         if (vma && vma->vm_file) {
3682                 struct file *f = vma->vm_file;
3683                 char *buf = (char *)__get_free_page(GFP_KERNEL);
3684                 if (buf) {
3685                         char *p;
3686
3687                         p = d_path(&f->f_path, buf, PAGE_SIZE);
3688                         if (IS_ERR(p))
3689                                 p = "?";
3690                         printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3691                                         vma->vm_start,
3692                                         vma->vm_end - vma->vm_start);
3693                         free_page((unsigned long)buf);
3694                 }
3695         }
3696         up_read(&mm->mmap_sem);
3697 }
3698
3699 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3700 void might_fault(void)
3701 {
3702         /*
3703          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3704          * holding the mmap_sem, this is safe because kernel memory doesn't
3705          * get paged out, therefore we'll never actually fault, and the
3706          * below annotations will generate false positives.
3707          */
3708         if (segment_eq(get_fs(), KERNEL_DS))
3709                 return;
3710
3711         /*
3712          * it would be nicer only to annotate paths which are not under
3713          * pagefault_disable, however that requires a larger audit and
3714          * providing helpers like get_user_atomic.
3715          */
3716         if (in_atomic())
3717                 return;
3718
3719         __might_sleep(__FILE__, __LINE__, 0);
3720
3721         if (current->mm)
3722                 might_lock_read(&current->mm->mmap_sem);
3723 }
3724 EXPORT_SYMBOL(might_fault);
3725 #endif
3726
3727 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3728 static void clear_gigantic_page(struct page *page,
3729                                 unsigned long addr,
3730                                 unsigned int pages_per_huge_page)
3731 {
3732         int i;
3733         struct page *p = page;
3734
3735         might_sleep();
3736         for (i = 0; i < pages_per_huge_page;
3737              i++, p = mem_map_next(p, page, i)) {
3738                 cond_resched();
3739                 clear_user_highpage(p, addr + i * PAGE_SIZE);
3740         }
3741 }
3742 void clear_huge_page(struct page *page,
3743                      unsigned long addr, unsigned int pages_per_huge_page)
3744 {
3745         int i;
3746
3747         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3748                 clear_gigantic_page(page, addr, pages_per_huge_page);
3749                 return;
3750         }
3751
3752         might_sleep();
3753         for (i = 0; i < pages_per_huge_page; i++) {
3754                 cond_resched();
3755                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3756         }
3757 }
3758
3759 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3760                                     unsigned long addr,
3761                                     struct vm_area_struct *vma,
3762                                     unsigned int pages_per_huge_page)
3763 {
3764         int i;
3765         struct page *dst_base = dst;
3766         struct page *src_base = src;
3767
3768         for (i = 0; i < pages_per_huge_page; ) {
3769                 cond_resched();
3770                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3771
3772                 i++;
3773                 dst = mem_map_next(dst, dst_base, i);
3774                 src = mem_map_next(src, src_base, i);
3775         }
3776 }
3777
3778 void copy_user_huge_page(struct page *dst, struct page *src,
3779                          unsigned long addr, struct vm_area_struct *vma,
3780                          unsigned int pages_per_huge_page)
3781 {
3782         int i;
3783
3784         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3785                 copy_user_gigantic_page(dst, src, addr, vma,
3786                                         pages_per_huge_page);
3787                 return;
3788         }
3789
3790         might_sleep();
3791         for (i = 0; i < pages_per_huge_page; i++) {
3792                 cond_resched();
3793                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3794         }
3795 }
3796 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3797
3798 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3799
3800 static struct kmem_cache *page_ptl_cachep;
3801
3802 void __init ptlock_cache_init(void)
3803 {
3804         page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3805                         SLAB_PANIC, NULL);
3806 }
3807
3808 bool ptlock_alloc(struct page *page)
3809 {
3810         spinlock_t *ptl;
3811
3812         ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3813         if (!ptl)
3814                 return false;
3815         page->ptl = ptl;
3816         return true;
3817 }
3818
3819 void ptlock_free(struct page *page)
3820 {
3821         kmem_cache_free(page_ptl_cachep, page->ptl);
3822 }
3823 #endif