drivers/staging/media/atomisp/pci/atomisp2: use set_memory.h
[linux-2.6-microblaze.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/ksm.h>
53 #include <linux/rmap.h>
54 #include <linux/export.h>
55 #include <linux/delayacct.h>
56 #include <linux/init.h>
57 #include <linux/pfn_t.h>
58 #include <linux/writeback.h>
59 #include <linux/memcontrol.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/kallsyms.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71
72 #include <asm/io.h>
73 #include <asm/mmu_context.h>
74 #include <asm/pgalloc.h>
75 #include <linux/uaccess.h>
76 #include <asm/tlb.h>
77 #include <asm/tlbflush.h>
78 #include <asm/pgtable.h>
79
80 #include "internal.h"
81
82 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
83 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
84 #endif
85
86 #ifndef CONFIG_NEED_MULTIPLE_NODES
87 /* use the per-pgdat data instead for discontigmem - mbligh */
88 unsigned long max_mapnr;
89 EXPORT_SYMBOL(max_mapnr);
90
91 struct page *mem_map;
92 EXPORT_SYMBOL(mem_map);
93 #endif
94
95 /*
96  * A number of key systems in x86 including ioremap() rely on the assumption
97  * that high_memory defines the upper bound on direct map memory, then end
98  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
99  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
100  * and ZONE_HIGHMEM.
101  */
102 void *high_memory;
103 EXPORT_SYMBOL(high_memory);
104
105 /*
106  * Randomize the address space (stacks, mmaps, brk, etc.).
107  *
108  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
109  *   as ancient (libc5 based) binaries can segfault. )
110  */
111 int randomize_va_space __read_mostly =
112 #ifdef CONFIG_COMPAT_BRK
113                                         1;
114 #else
115                                         2;
116 #endif
117
118 static int __init disable_randmaps(char *s)
119 {
120         randomize_va_space = 0;
121         return 1;
122 }
123 __setup("norandmaps", disable_randmaps);
124
125 unsigned long zero_pfn __read_mostly;
126 EXPORT_SYMBOL(zero_pfn);
127
128 unsigned long highest_memmap_pfn __read_mostly;
129
130 /*
131  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
132  */
133 static int __init init_zero_pfn(void)
134 {
135         zero_pfn = page_to_pfn(ZERO_PAGE(0));
136         return 0;
137 }
138 core_initcall(init_zero_pfn);
139
140
141 #if defined(SPLIT_RSS_COUNTING)
142
143 void sync_mm_rss(struct mm_struct *mm)
144 {
145         int i;
146
147         for (i = 0; i < NR_MM_COUNTERS; i++) {
148                 if (current->rss_stat.count[i]) {
149                         add_mm_counter(mm, i, current->rss_stat.count[i]);
150                         current->rss_stat.count[i] = 0;
151                 }
152         }
153         current->rss_stat.events = 0;
154 }
155
156 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
157 {
158         struct task_struct *task = current;
159
160         if (likely(task->mm == mm))
161                 task->rss_stat.count[member] += val;
162         else
163                 add_mm_counter(mm, member, val);
164 }
165 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
166 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
167
168 /* sync counter once per 64 page faults */
169 #define TASK_RSS_EVENTS_THRESH  (64)
170 static void check_sync_rss_stat(struct task_struct *task)
171 {
172         if (unlikely(task != current))
173                 return;
174         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
175                 sync_mm_rss(task->mm);
176 }
177 #else /* SPLIT_RSS_COUNTING */
178
179 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
180 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
181
182 static void check_sync_rss_stat(struct task_struct *task)
183 {
184 }
185
186 #endif /* SPLIT_RSS_COUNTING */
187
188 #ifdef HAVE_GENERIC_MMU_GATHER
189
190 static bool tlb_next_batch(struct mmu_gather *tlb)
191 {
192         struct mmu_gather_batch *batch;
193
194         batch = tlb->active;
195         if (batch->next) {
196                 tlb->active = batch->next;
197                 return true;
198         }
199
200         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
201                 return false;
202
203         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
204         if (!batch)
205                 return false;
206
207         tlb->batch_count++;
208         batch->next = NULL;
209         batch->nr   = 0;
210         batch->max  = MAX_GATHER_BATCH;
211
212         tlb->active->next = batch;
213         tlb->active = batch;
214
215         return true;
216 }
217
218 /* tlb_gather_mmu
219  *      Called to initialize an (on-stack) mmu_gather structure for page-table
220  *      tear-down from @mm. The @fullmm argument is used when @mm is without
221  *      users and we're going to destroy the full address space (exit/execve).
222  */
223 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
224 {
225         tlb->mm = mm;
226
227         /* Is it from 0 to ~0? */
228         tlb->fullmm     = !(start | (end+1));
229         tlb->need_flush_all = 0;
230         tlb->local.next = NULL;
231         tlb->local.nr   = 0;
232         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
233         tlb->active     = &tlb->local;
234         tlb->batch_count = 0;
235
236 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
237         tlb->batch = NULL;
238 #endif
239         tlb->page_size = 0;
240
241         __tlb_reset_range(tlb);
242 }
243
244 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
245 {
246         if (!tlb->end)
247                 return;
248
249         tlb_flush(tlb);
250         mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
251 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
252         tlb_table_flush(tlb);
253 #endif
254         __tlb_reset_range(tlb);
255 }
256
257 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
258 {
259         struct mmu_gather_batch *batch;
260
261         for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
262                 free_pages_and_swap_cache(batch->pages, batch->nr);
263                 batch->nr = 0;
264         }
265         tlb->active = &tlb->local;
266 }
267
268 void tlb_flush_mmu(struct mmu_gather *tlb)
269 {
270         tlb_flush_mmu_tlbonly(tlb);
271         tlb_flush_mmu_free(tlb);
272 }
273
274 /* tlb_finish_mmu
275  *      Called at the end of the shootdown operation to free up any resources
276  *      that were required.
277  */
278 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
279 {
280         struct mmu_gather_batch *batch, *next;
281
282         tlb_flush_mmu(tlb);
283
284         /* keep the page table cache within bounds */
285         check_pgt_cache();
286
287         for (batch = tlb->local.next; batch; batch = next) {
288                 next = batch->next;
289                 free_pages((unsigned long)batch, 0);
290         }
291         tlb->local.next = NULL;
292 }
293
294 /* __tlb_remove_page
295  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
296  *      handling the additional races in SMP caused by other CPUs caching valid
297  *      mappings in their TLBs. Returns the number of free page slots left.
298  *      When out of page slots we must call tlb_flush_mmu().
299  *returns true if the caller should flush.
300  */
301 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
302 {
303         struct mmu_gather_batch *batch;
304
305         VM_BUG_ON(!tlb->end);
306         VM_WARN_ON(tlb->page_size != page_size);
307
308         batch = tlb->active;
309         /*
310          * Add the page and check if we are full. If so
311          * force a flush.
312          */
313         batch->pages[batch->nr++] = page;
314         if (batch->nr == batch->max) {
315                 if (!tlb_next_batch(tlb))
316                         return true;
317                 batch = tlb->active;
318         }
319         VM_BUG_ON_PAGE(batch->nr > batch->max, page);
320
321         return false;
322 }
323
324 #endif /* HAVE_GENERIC_MMU_GATHER */
325
326 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
327
328 /*
329  * See the comment near struct mmu_table_batch.
330  */
331
332 static void tlb_remove_table_smp_sync(void *arg)
333 {
334         /* Simply deliver the interrupt */
335 }
336
337 static void tlb_remove_table_one(void *table)
338 {
339         /*
340          * This isn't an RCU grace period and hence the page-tables cannot be
341          * assumed to be actually RCU-freed.
342          *
343          * It is however sufficient for software page-table walkers that rely on
344          * IRQ disabling. See the comment near struct mmu_table_batch.
345          */
346         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
347         __tlb_remove_table(table);
348 }
349
350 static void tlb_remove_table_rcu(struct rcu_head *head)
351 {
352         struct mmu_table_batch *batch;
353         int i;
354
355         batch = container_of(head, struct mmu_table_batch, rcu);
356
357         for (i = 0; i < batch->nr; i++)
358                 __tlb_remove_table(batch->tables[i]);
359
360         free_page((unsigned long)batch);
361 }
362
363 void tlb_table_flush(struct mmu_gather *tlb)
364 {
365         struct mmu_table_batch **batch = &tlb->batch;
366
367         if (*batch) {
368                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
369                 *batch = NULL;
370         }
371 }
372
373 void tlb_remove_table(struct mmu_gather *tlb, void *table)
374 {
375         struct mmu_table_batch **batch = &tlb->batch;
376
377         /*
378          * When there's less then two users of this mm there cannot be a
379          * concurrent page-table walk.
380          */
381         if (atomic_read(&tlb->mm->mm_users) < 2) {
382                 __tlb_remove_table(table);
383                 return;
384         }
385
386         if (*batch == NULL) {
387                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
388                 if (*batch == NULL) {
389                         tlb_remove_table_one(table);
390                         return;
391                 }
392                 (*batch)->nr = 0;
393         }
394         (*batch)->tables[(*batch)->nr++] = table;
395         if ((*batch)->nr == MAX_TABLE_BATCH)
396                 tlb_table_flush(tlb);
397 }
398
399 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
400
401 /*
402  * Note: this doesn't free the actual pages themselves. That
403  * has been handled earlier when unmapping all the memory regions.
404  */
405 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
406                            unsigned long addr)
407 {
408         pgtable_t token = pmd_pgtable(*pmd);
409         pmd_clear(pmd);
410         pte_free_tlb(tlb, token, addr);
411         atomic_long_dec(&tlb->mm->nr_ptes);
412 }
413
414 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
415                                 unsigned long addr, unsigned long end,
416                                 unsigned long floor, unsigned long ceiling)
417 {
418         pmd_t *pmd;
419         unsigned long next;
420         unsigned long start;
421
422         start = addr;
423         pmd = pmd_offset(pud, addr);
424         do {
425                 next = pmd_addr_end(addr, end);
426                 if (pmd_none_or_clear_bad(pmd))
427                         continue;
428                 free_pte_range(tlb, pmd, addr);
429         } while (pmd++, addr = next, addr != end);
430
431         start &= PUD_MASK;
432         if (start < floor)
433                 return;
434         if (ceiling) {
435                 ceiling &= PUD_MASK;
436                 if (!ceiling)
437                         return;
438         }
439         if (end - 1 > ceiling - 1)
440                 return;
441
442         pmd = pmd_offset(pud, start);
443         pud_clear(pud);
444         pmd_free_tlb(tlb, pmd, start);
445         mm_dec_nr_pmds(tlb->mm);
446 }
447
448 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
449                                 unsigned long addr, unsigned long end,
450                                 unsigned long floor, unsigned long ceiling)
451 {
452         pud_t *pud;
453         unsigned long next;
454         unsigned long start;
455
456         start = addr;
457         pud = pud_offset(p4d, addr);
458         do {
459                 next = pud_addr_end(addr, end);
460                 if (pud_none_or_clear_bad(pud))
461                         continue;
462                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
463         } while (pud++, addr = next, addr != end);
464
465         start &= P4D_MASK;
466         if (start < floor)
467                 return;
468         if (ceiling) {
469                 ceiling &= P4D_MASK;
470                 if (!ceiling)
471                         return;
472         }
473         if (end - 1 > ceiling - 1)
474                 return;
475
476         pud = pud_offset(p4d, start);
477         p4d_clear(p4d);
478         pud_free_tlb(tlb, pud, start);
479 }
480
481 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
482                                 unsigned long addr, unsigned long end,
483                                 unsigned long floor, unsigned long ceiling)
484 {
485         p4d_t *p4d;
486         unsigned long next;
487         unsigned long start;
488
489         start = addr;
490         p4d = p4d_offset(pgd, addr);
491         do {
492                 next = p4d_addr_end(addr, end);
493                 if (p4d_none_or_clear_bad(p4d))
494                         continue;
495                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
496         } while (p4d++, addr = next, addr != end);
497
498         start &= PGDIR_MASK;
499         if (start < floor)
500                 return;
501         if (ceiling) {
502                 ceiling &= PGDIR_MASK;
503                 if (!ceiling)
504                         return;
505         }
506         if (end - 1 > ceiling - 1)
507                 return;
508
509         p4d = p4d_offset(pgd, start);
510         pgd_clear(pgd);
511         p4d_free_tlb(tlb, p4d, start);
512 }
513
514 /*
515  * This function frees user-level page tables of a process.
516  */
517 void free_pgd_range(struct mmu_gather *tlb,
518                         unsigned long addr, unsigned long end,
519                         unsigned long floor, unsigned long ceiling)
520 {
521         pgd_t *pgd;
522         unsigned long next;
523
524         /*
525          * The next few lines have given us lots of grief...
526          *
527          * Why are we testing PMD* at this top level?  Because often
528          * there will be no work to do at all, and we'd prefer not to
529          * go all the way down to the bottom just to discover that.
530          *
531          * Why all these "- 1"s?  Because 0 represents both the bottom
532          * of the address space and the top of it (using -1 for the
533          * top wouldn't help much: the masks would do the wrong thing).
534          * The rule is that addr 0 and floor 0 refer to the bottom of
535          * the address space, but end 0 and ceiling 0 refer to the top
536          * Comparisons need to use "end - 1" and "ceiling - 1" (though
537          * that end 0 case should be mythical).
538          *
539          * Wherever addr is brought up or ceiling brought down, we must
540          * be careful to reject "the opposite 0" before it confuses the
541          * subsequent tests.  But what about where end is brought down
542          * by PMD_SIZE below? no, end can't go down to 0 there.
543          *
544          * Whereas we round start (addr) and ceiling down, by different
545          * masks at different levels, in order to test whether a table
546          * now has no other vmas using it, so can be freed, we don't
547          * bother to round floor or end up - the tests don't need that.
548          */
549
550         addr &= PMD_MASK;
551         if (addr < floor) {
552                 addr += PMD_SIZE;
553                 if (!addr)
554                         return;
555         }
556         if (ceiling) {
557                 ceiling &= PMD_MASK;
558                 if (!ceiling)
559                         return;
560         }
561         if (end - 1 > ceiling - 1)
562                 end -= PMD_SIZE;
563         if (addr > end - 1)
564                 return;
565         /*
566          * We add page table cache pages with PAGE_SIZE,
567          * (see pte_free_tlb()), flush the tlb if we need
568          */
569         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
570         pgd = pgd_offset(tlb->mm, addr);
571         do {
572                 next = pgd_addr_end(addr, end);
573                 if (pgd_none_or_clear_bad(pgd))
574                         continue;
575                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
576         } while (pgd++, addr = next, addr != end);
577 }
578
579 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
580                 unsigned long floor, unsigned long ceiling)
581 {
582         while (vma) {
583                 struct vm_area_struct *next = vma->vm_next;
584                 unsigned long addr = vma->vm_start;
585
586                 /*
587                  * Hide vma from rmap and truncate_pagecache before freeing
588                  * pgtables
589                  */
590                 unlink_anon_vmas(vma);
591                 unlink_file_vma(vma);
592
593                 if (is_vm_hugetlb_page(vma)) {
594                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
595                                 floor, next ? next->vm_start : ceiling);
596                 } else {
597                         /*
598                          * Optimization: gather nearby vmas into one call down
599                          */
600                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
601                                && !is_vm_hugetlb_page(next)) {
602                                 vma = next;
603                                 next = vma->vm_next;
604                                 unlink_anon_vmas(vma);
605                                 unlink_file_vma(vma);
606                         }
607                         free_pgd_range(tlb, addr, vma->vm_end,
608                                 floor, next ? next->vm_start : ceiling);
609                 }
610                 vma = next;
611         }
612 }
613
614 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
615 {
616         spinlock_t *ptl;
617         pgtable_t new = pte_alloc_one(mm, address);
618         if (!new)
619                 return -ENOMEM;
620
621         /*
622          * Ensure all pte setup (eg. pte page lock and page clearing) are
623          * visible before the pte is made visible to other CPUs by being
624          * put into page tables.
625          *
626          * The other side of the story is the pointer chasing in the page
627          * table walking code (when walking the page table without locking;
628          * ie. most of the time). Fortunately, these data accesses consist
629          * of a chain of data-dependent loads, meaning most CPUs (alpha
630          * being the notable exception) will already guarantee loads are
631          * seen in-order. See the alpha page table accessors for the
632          * smp_read_barrier_depends() barriers in page table walking code.
633          */
634         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
635
636         ptl = pmd_lock(mm, pmd);
637         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
638                 atomic_long_inc(&mm->nr_ptes);
639                 pmd_populate(mm, pmd, new);
640                 new = NULL;
641         }
642         spin_unlock(ptl);
643         if (new)
644                 pte_free(mm, new);
645         return 0;
646 }
647
648 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
649 {
650         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
651         if (!new)
652                 return -ENOMEM;
653
654         smp_wmb(); /* See comment in __pte_alloc */
655
656         spin_lock(&init_mm.page_table_lock);
657         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
658                 pmd_populate_kernel(&init_mm, pmd, new);
659                 new = NULL;
660         }
661         spin_unlock(&init_mm.page_table_lock);
662         if (new)
663                 pte_free_kernel(&init_mm, new);
664         return 0;
665 }
666
667 static inline void init_rss_vec(int *rss)
668 {
669         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
670 }
671
672 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
673 {
674         int i;
675
676         if (current->mm == mm)
677                 sync_mm_rss(mm);
678         for (i = 0; i < NR_MM_COUNTERS; i++)
679                 if (rss[i])
680                         add_mm_counter(mm, i, rss[i]);
681 }
682
683 /*
684  * This function is called to print an error when a bad pte
685  * is found. For example, we might have a PFN-mapped pte in
686  * a region that doesn't allow it.
687  *
688  * The calling function must still handle the error.
689  */
690 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
691                           pte_t pte, struct page *page)
692 {
693         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
694         p4d_t *p4d = p4d_offset(pgd, addr);
695         pud_t *pud = pud_offset(p4d, addr);
696         pmd_t *pmd = pmd_offset(pud, addr);
697         struct address_space *mapping;
698         pgoff_t index;
699         static unsigned long resume;
700         static unsigned long nr_shown;
701         static unsigned long nr_unshown;
702
703         /*
704          * Allow a burst of 60 reports, then keep quiet for that minute;
705          * or allow a steady drip of one report per second.
706          */
707         if (nr_shown == 60) {
708                 if (time_before(jiffies, resume)) {
709                         nr_unshown++;
710                         return;
711                 }
712                 if (nr_unshown) {
713                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
714                                  nr_unshown);
715                         nr_unshown = 0;
716                 }
717                 nr_shown = 0;
718         }
719         if (nr_shown++ == 0)
720                 resume = jiffies + 60 * HZ;
721
722         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
723         index = linear_page_index(vma, addr);
724
725         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
726                  current->comm,
727                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
728         if (page)
729                 dump_page(page, "bad pte");
730         pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
731                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
732         /*
733          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
734          */
735         pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
736                  vma->vm_file,
737                  vma->vm_ops ? vma->vm_ops->fault : NULL,
738                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
739                  mapping ? mapping->a_ops->readpage : NULL);
740         dump_stack();
741         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
742 }
743
744 /*
745  * vm_normal_page -- This function gets the "struct page" associated with a pte.
746  *
747  * "Special" mappings do not wish to be associated with a "struct page" (either
748  * it doesn't exist, or it exists but they don't want to touch it). In this
749  * case, NULL is returned here. "Normal" mappings do have a struct page.
750  *
751  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
752  * pte bit, in which case this function is trivial. Secondly, an architecture
753  * may not have a spare pte bit, which requires a more complicated scheme,
754  * described below.
755  *
756  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
757  * special mapping (even if there are underlying and valid "struct pages").
758  * COWed pages of a VM_PFNMAP are always normal.
759  *
760  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
761  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
762  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
763  * mapping will always honor the rule
764  *
765  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
766  *
767  * And for normal mappings this is false.
768  *
769  * This restricts such mappings to be a linear translation from virtual address
770  * to pfn. To get around this restriction, we allow arbitrary mappings so long
771  * as the vma is not a COW mapping; in that case, we know that all ptes are
772  * special (because none can have been COWed).
773  *
774  *
775  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
776  *
777  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
778  * page" backing, however the difference is that _all_ pages with a struct
779  * page (that is, those where pfn_valid is true) are refcounted and considered
780  * normal pages by the VM. The disadvantage is that pages are refcounted
781  * (which can be slower and simply not an option for some PFNMAP users). The
782  * advantage is that we don't have to follow the strict linearity rule of
783  * PFNMAP mappings in order to support COWable mappings.
784  *
785  */
786 #ifdef __HAVE_ARCH_PTE_SPECIAL
787 # define HAVE_PTE_SPECIAL 1
788 #else
789 # define HAVE_PTE_SPECIAL 0
790 #endif
791 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
792                                 pte_t pte)
793 {
794         unsigned long pfn = pte_pfn(pte);
795
796         if (HAVE_PTE_SPECIAL) {
797                 if (likely(!pte_special(pte)))
798                         goto check_pfn;
799                 if (vma->vm_ops && vma->vm_ops->find_special_page)
800                         return vma->vm_ops->find_special_page(vma, addr);
801                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
802                         return NULL;
803                 if (!is_zero_pfn(pfn))
804                         print_bad_pte(vma, addr, pte, NULL);
805                 return NULL;
806         }
807
808         /* !HAVE_PTE_SPECIAL case follows: */
809
810         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
811                 if (vma->vm_flags & VM_MIXEDMAP) {
812                         if (!pfn_valid(pfn))
813                                 return NULL;
814                         goto out;
815                 } else {
816                         unsigned long off;
817                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
818                         if (pfn == vma->vm_pgoff + off)
819                                 return NULL;
820                         if (!is_cow_mapping(vma->vm_flags))
821                                 return NULL;
822                 }
823         }
824
825         if (is_zero_pfn(pfn))
826                 return NULL;
827 check_pfn:
828         if (unlikely(pfn > highest_memmap_pfn)) {
829                 print_bad_pte(vma, addr, pte, NULL);
830                 return NULL;
831         }
832
833         /*
834          * NOTE! We still have PageReserved() pages in the page tables.
835          * eg. VDSO mappings can cause them to exist.
836          */
837 out:
838         return pfn_to_page(pfn);
839 }
840
841 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
842 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
843                                 pmd_t pmd)
844 {
845         unsigned long pfn = pmd_pfn(pmd);
846
847         /*
848          * There is no pmd_special() but there may be special pmds, e.g.
849          * in a direct-access (dax) mapping, so let's just replicate the
850          * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
851          */
852         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
853                 if (vma->vm_flags & VM_MIXEDMAP) {
854                         if (!pfn_valid(pfn))
855                                 return NULL;
856                         goto out;
857                 } else {
858                         unsigned long off;
859                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
860                         if (pfn == vma->vm_pgoff + off)
861                                 return NULL;
862                         if (!is_cow_mapping(vma->vm_flags))
863                                 return NULL;
864                 }
865         }
866
867         if (is_zero_pfn(pfn))
868                 return NULL;
869         if (unlikely(pfn > highest_memmap_pfn))
870                 return NULL;
871
872         /*
873          * NOTE! We still have PageReserved() pages in the page tables.
874          * eg. VDSO mappings can cause them to exist.
875          */
876 out:
877         return pfn_to_page(pfn);
878 }
879 #endif
880
881 /*
882  * copy one vm_area from one task to the other. Assumes the page tables
883  * already present in the new task to be cleared in the whole range
884  * covered by this vma.
885  */
886
887 static inline unsigned long
888 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
889                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
890                 unsigned long addr, int *rss)
891 {
892         unsigned long vm_flags = vma->vm_flags;
893         pte_t pte = *src_pte;
894         struct page *page;
895
896         /* pte contains position in swap or file, so copy. */
897         if (unlikely(!pte_present(pte))) {
898                 swp_entry_t entry = pte_to_swp_entry(pte);
899
900                 if (likely(!non_swap_entry(entry))) {
901                         if (swap_duplicate(entry) < 0)
902                                 return entry.val;
903
904                         /* make sure dst_mm is on swapoff's mmlist. */
905                         if (unlikely(list_empty(&dst_mm->mmlist))) {
906                                 spin_lock(&mmlist_lock);
907                                 if (list_empty(&dst_mm->mmlist))
908                                         list_add(&dst_mm->mmlist,
909                                                         &src_mm->mmlist);
910                                 spin_unlock(&mmlist_lock);
911                         }
912                         rss[MM_SWAPENTS]++;
913                 } else if (is_migration_entry(entry)) {
914                         page = migration_entry_to_page(entry);
915
916                         rss[mm_counter(page)]++;
917
918                         if (is_write_migration_entry(entry) &&
919                                         is_cow_mapping(vm_flags)) {
920                                 /*
921                                  * COW mappings require pages in both
922                                  * parent and child to be set to read.
923                                  */
924                                 make_migration_entry_read(&entry);
925                                 pte = swp_entry_to_pte(entry);
926                                 if (pte_swp_soft_dirty(*src_pte))
927                                         pte = pte_swp_mksoft_dirty(pte);
928                                 set_pte_at(src_mm, addr, src_pte, pte);
929                         }
930                 }
931                 goto out_set_pte;
932         }
933
934         /*
935          * If it's a COW mapping, write protect it both
936          * in the parent and the child
937          */
938         if (is_cow_mapping(vm_flags)) {
939                 ptep_set_wrprotect(src_mm, addr, src_pte);
940                 pte = pte_wrprotect(pte);
941         }
942
943         /*
944          * If it's a shared mapping, mark it clean in
945          * the child
946          */
947         if (vm_flags & VM_SHARED)
948                 pte = pte_mkclean(pte);
949         pte = pte_mkold(pte);
950
951         page = vm_normal_page(vma, addr, pte);
952         if (page) {
953                 get_page(page);
954                 page_dup_rmap(page, false);
955                 rss[mm_counter(page)]++;
956         }
957
958 out_set_pte:
959         set_pte_at(dst_mm, addr, dst_pte, pte);
960         return 0;
961 }
962
963 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
964                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
965                    unsigned long addr, unsigned long end)
966 {
967         pte_t *orig_src_pte, *orig_dst_pte;
968         pte_t *src_pte, *dst_pte;
969         spinlock_t *src_ptl, *dst_ptl;
970         int progress = 0;
971         int rss[NR_MM_COUNTERS];
972         swp_entry_t entry = (swp_entry_t){0};
973
974 again:
975         init_rss_vec(rss);
976
977         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
978         if (!dst_pte)
979                 return -ENOMEM;
980         src_pte = pte_offset_map(src_pmd, addr);
981         src_ptl = pte_lockptr(src_mm, src_pmd);
982         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
983         orig_src_pte = src_pte;
984         orig_dst_pte = dst_pte;
985         arch_enter_lazy_mmu_mode();
986
987         do {
988                 /*
989                  * We are holding two locks at this point - either of them
990                  * could generate latencies in another task on another CPU.
991                  */
992                 if (progress >= 32) {
993                         progress = 0;
994                         if (need_resched() ||
995                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
996                                 break;
997                 }
998                 if (pte_none(*src_pte)) {
999                         progress++;
1000                         continue;
1001                 }
1002                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1003                                                         vma, addr, rss);
1004                 if (entry.val)
1005                         break;
1006                 progress += 8;
1007         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1008
1009         arch_leave_lazy_mmu_mode();
1010         spin_unlock(src_ptl);
1011         pte_unmap(orig_src_pte);
1012         add_mm_rss_vec(dst_mm, rss);
1013         pte_unmap_unlock(orig_dst_pte, dst_ptl);
1014         cond_resched();
1015
1016         if (entry.val) {
1017                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1018                         return -ENOMEM;
1019                 progress = 0;
1020         }
1021         if (addr != end)
1022                 goto again;
1023         return 0;
1024 }
1025
1026 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1027                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1028                 unsigned long addr, unsigned long end)
1029 {
1030         pmd_t *src_pmd, *dst_pmd;
1031         unsigned long next;
1032
1033         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1034         if (!dst_pmd)
1035                 return -ENOMEM;
1036         src_pmd = pmd_offset(src_pud, addr);
1037         do {
1038                 next = pmd_addr_end(addr, end);
1039                 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
1040                         int err;
1041                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1042                         err = copy_huge_pmd(dst_mm, src_mm,
1043                                             dst_pmd, src_pmd, addr, vma);
1044                         if (err == -ENOMEM)
1045                                 return -ENOMEM;
1046                         if (!err)
1047                                 continue;
1048                         /* fall through */
1049                 }
1050                 if (pmd_none_or_clear_bad(src_pmd))
1051                         continue;
1052                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1053                                                 vma, addr, next))
1054                         return -ENOMEM;
1055         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1056         return 0;
1057 }
1058
1059 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1060                 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1061                 unsigned long addr, unsigned long end)
1062 {
1063         pud_t *src_pud, *dst_pud;
1064         unsigned long next;
1065
1066         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1067         if (!dst_pud)
1068                 return -ENOMEM;
1069         src_pud = pud_offset(src_p4d, addr);
1070         do {
1071                 next = pud_addr_end(addr, end);
1072                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1073                         int err;
1074
1075                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1076                         err = copy_huge_pud(dst_mm, src_mm,
1077                                             dst_pud, src_pud, addr, vma);
1078                         if (err == -ENOMEM)
1079                                 return -ENOMEM;
1080                         if (!err)
1081                                 continue;
1082                         /* fall through */
1083                 }
1084                 if (pud_none_or_clear_bad(src_pud))
1085                         continue;
1086                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1087                                                 vma, addr, next))
1088                         return -ENOMEM;
1089         } while (dst_pud++, src_pud++, addr = next, addr != end);
1090         return 0;
1091 }
1092
1093 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1094                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1095                 unsigned long addr, unsigned long end)
1096 {
1097         p4d_t *src_p4d, *dst_p4d;
1098         unsigned long next;
1099
1100         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1101         if (!dst_p4d)
1102                 return -ENOMEM;
1103         src_p4d = p4d_offset(src_pgd, addr);
1104         do {
1105                 next = p4d_addr_end(addr, end);
1106                 if (p4d_none_or_clear_bad(src_p4d))
1107                         continue;
1108                 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1109                                                 vma, addr, next))
1110                         return -ENOMEM;
1111         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1112         return 0;
1113 }
1114
1115 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1116                 struct vm_area_struct *vma)
1117 {
1118         pgd_t *src_pgd, *dst_pgd;
1119         unsigned long next;
1120         unsigned long addr = vma->vm_start;
1121         unsigned long end = vma->vm_end;
1122         unsigned long mmun_start;       /* For mmu_notifiers */
1123         unsigned long mmun_end;         /* For mmu_notifiers */
1124         bool is_cow;
1125         int ret;
1126
1127         /*
1128          * Don't copy ptes where a page fault will fill them correctly.
1129          * Fork becomes much lighter when there are big shared or private
1130          * readonly mappings. The tradeoff is that copy_page_range is more
1131          * efficient than faulting.
1132          */
1133         if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1134                         !vma->anon_vma)
1135                 return 0;
1136
1137         if (is_vm_hugetlb_page(vma))
1138                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1139
1140         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1141                 /*
1142                  * We do not free on error cases below as remove_vma
1143                  * gets called on error from higher level routine
1144                  */
1145                 ret = track_pfn_copy(vma);
1146                 if (ret)
1147                         return ret;
1148         }
1149
1150         /*
1151          * We need to invalidate the secondary MMU mappings only when
1152          * there could be a permission downgrade on the ptes of the
1153          * parent mm. And a permission downgrade will only happen if
1154          * is_cow_mapping() returns true.
1155          */
1156         is_cow = is_cow_mapping(vma->vm_flags);
1157         mmun_start = addr;
1158         mmun_end   = end;
1159         if (is_cow)
1160                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1161                                                     mmun_end);
1162
1163         ret = 0;
1164         dst_pgd = pgd_offset(dst_mm, addr);
1165         src_pgd = pgd_offset(src_mm, addr);
1166         do {
1167                 next = pgd_addr_end(addr, end);
1168                 if (pgd_none_or_clear_bad(src_pgd))
1169                         continue;
1170                 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1171                                             vma, addr, next))) {
1172                         ret = -ENOMEM;
1173                         break;
1174                 }
1175         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1176
1177         if (is_cow)
1178                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1179         return ret;
1180 }
1181
1182 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1183                                 struct vm_area_struct *vma, pmd_t *pmd,
1184                                 unsigned long addr, unsigned long end,
1185                                 struct zap_details *details)
1186 {
1187         struct mm_struct *mm = tlb->mm;
1188         int force_flush = 0;
1189         int rss[NR_MM_COUNTERS];
1190         spinlock_t *ptl;
1191         pte_t *start_pte;
1192         pte_t *pte;
1193         swp_entry_t entry;
1194
1195         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1196 again:
1197         init_rss_vec(rss);
1198         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1199         pte = start_pte;
1200         arch_enter_lazy_mmu_mode();
1201         do {
1202                 pte_t ptent = *pte;
1203                 if (pte_none(ptent))
1204                         continue;
1205
1206                 if (pte_present(ptent)) {
1207                         struct page *page;
1208
1209                         page = vm_normal_page(vma, addr, ptent);
1210                         if (unlikely(details) && page) {
1211                                 /*
1212                                  * unmap_shared_mapping_pages() wants to
1213                                  * invalidate cache without truncating:
1214                                  * unmap shared but keep private pages.
1215                                  */
1216                                 if (details->check_mapping &&
1217                                     details->check_mapping != page_rmapping(page))
1218                                         continue;
1219                         }
1220                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1221                                                         tlb->fullmm);
1222                         tlb_remove_tlb_entry(tlb, pte, addr);
1223                         if (unlikely(!page))
1224                                 continue;
1225
1226                         if (!PageAnon(page)) {
1227                                 if (pte_dirty(ptent)) {
1228                                         force_flush = 1;
1229                                         set_page_dirty(page);
1230                                 }
1231                                 if (pte_young(ptent) &&
1232                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1233                                         mark_page_accessed(page);
1234                         }
1235                         rss[mm_counter(page)]--;
1236                         page_remove_rmap(page, false);
1237                         if (unlikely(page_mapcount(page) < 0))
1238                                 print_bad_pte(vma, addr, ptent, page);
1239                         if (unlikely(__tlb_remove_page(tlb, page))) {
1240                                 force_flush = 1;
1241                                 addr += PAGE_SIZE;
1242                                 break;
1243                         }
1244                         continue;
1245                 }
1246                 /* If details->check_mapping, we leave swap entries. */
1247                 if (unlikely(details))
1248                         continue;
1249
1250                 entry = pte_to_swp_entry(ptent);
1251                 if (!non_swap_entry(entry))
1252                         rss[MM_SWAPENTS]--;
1253                 else if (is_migration_entry(entry)) {
1254                         struct page *page;
1255
1256                         page = migration_entry_to_page(entry);
1257                         rss[mm_counter(page)]--;
1258                 }
1259                 if (unlikely(!free_swap_and_cache(entry)))
1260                         print_bad_pte(vma, addr, ptent, NULL);
1261                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1262         } while (pte++, addr += PAGE_SIZE, addr != end);
1263
1264         add_mm_rss_vec(mm, rss);
1265         arch_leave_lazy_mmu_mode();
1266
1267         /* Do the actual TLB flush before dropping ptl */
1268         if (force_flush)
1269                 tlb_flush_mmu_tlbonly(tlb);
1270         pte_unmap_unlock(start_pte, ptl);
1271
1272         /*
1273          * If we forced a TLB flush (either due to running out of
1274          * batch buffers or because we needed to flush dirty TLB
1275          * entries before releasing the ptl), free the batched
1276          * memory too. Restart if we didn't do everything.
1277          */
1278         if (force_flush) {
1279                 force_flush = 0;
1280                 tlb_flush_mmu_free(tlb);
1281                 if (addr != end)
1282                         goto again;
1283         }
1284
1285         return addr;
1286 }
1287
1288 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1289                                 struct vm_area_struct *vma, pud_t *pud,
1290                                 unsigned long addr, unsigned long end,
1291                                 struct zap_details *details)
1292 {
1293         pmd_t *pmd;
1294         unsigned long next;
1295
1296         pmd = pmd_offset(pud, addr);
1297         do {
1298                 next = pmd_addr_end(addr, end);
1299                 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1300                         if (next - addr != HPAGE_PMD_SIZE) {
1301                                 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1302                                     !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1303                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1304                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1305                                 goto next;
1306                         /* fall through */
1307                 }
1308                 /*
1309                  * Here there can be other concurrent MADV_DONTNEED or
1310                  * trans huge page faults running, and if the pmd is
1311                  * none or trans huge it can change under us. This is
1312                  * because MADV_DONTNEED holds the mmap_sem in read
1313                  * mode.
1314                  */
1315                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1316                         goto next;
1317                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1318 next:
1319                 cond_resched();
1320         } while (pmd++, addr = next, addr != end);
1321
1322         return addr;
1323 }
1324
1325 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1326                                 struct vm_area_struct *vma, p4d_t *p4d,
1327                                 unsigned long addr, unsigned long end,
1328                                 struct zap_details *details)
1329 {
1330         pud_t *pud;
1331         unsigned long next;
1332
1333         pud = pud_offset(p4d, addr);
1334         do {
1335                 next = pud_addr_end(addr, end);
1336                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1337                         if (next - addr != HPAGE_PUD_SIZE) {
1338                                 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1339                                 split_huge_pud(vma, pud, addr);
1340                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1341                                 goto next;
1342                         /* fall through */
1343                 }
1344                 if (pud_none_or_clear_bad(pud))
1345                         continue;
1346                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1347 next:
1348                 cond_resched();
1349         } while (pud++, addr = next, addr != end);
1350
1351         return addr;
1352 }
1353
1354 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1355                                 struct vm_area_struct *vma, pgd_t *pgd,
1356                                 unsigned long addr, unsigned long end,
1357                                 struct zap_details *details)
1358 {
1359         p4d_t *p4d;
1360         unsigned long next;
1361
1362         p4d = p4d_offset(pgd, addr);
1363         do {
1364                 next = p4d_addr_end(addr, end);
1365                 if (p4d_none_or_clear_bad(p4d))
1366                         continue;
1367                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1368         } while (p4d++, addr = next, addr != end);
1369
1370         return addr;
1371 }
1372
1373 void unmap_page_range(struct mmu_gather *tlb,
1374                              struct vm_area_struct *vma,
1375                              unsigned long addr, unsigned long end,
1376                              struct zap_details *details)
1377 {
1378         pgd_t *pgd;
1379         unsigned long next;
1380
1381         BUG_ON(addr >= end);
1382         tlb_start_vma(tlb, vma);
1383         pgd = pgd_offset(vma->vm_mm, addr);
1384         do {
1385                 next = pgd_addr_end(addr, end);
1386                 if (pgd_none_or_clear_bad(pgd))
1387                         continue;
1388                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1389         } while (pgd++, addr = next, addr != end);
1390         tlb_end_vma(tlb, vma);
1391 }
1392
1393
1394 static void unmap_single_vma(struct mmu_gather *tlb,
1395                 struct vm_area_struct *vma, unsigned long start_addr,
1396                 unsigned long end_addr,
1397                 struct zap_details *details)
1398 {
1399         unsigned long start = max(vma->vm_start, start_addr);
1400         unsigned long end;
1401
1402         if (start >= vma->vm_end)
1403                 return;
1404         end = min(vma->vm_end, end_addr);
1405         if (end <= vma->vm_start)
1406                 return;
1407
1408         if (vma->vm_file)
1409                 uprobe_munmap(vma, start, end);
1410
1411         if (unlikely(vma->vm_flags & VM_PFNMAP))
1412                 untrack_pfn(vma, 0, 0);
1413
1414         if (start != end) {
1415                 if (unlikely(is_vm_hugetlb_page(vma))) {
1416                         /*
1417                          * It is undesirable to test vma->vm_file as it
1418                          * should be non-null for valid hugetlb area.
1419                          * However, vm_file will be NULL in the error
1420                          * cleanup path of mmap_region. When
1421                          * hugetlbfs ->mmap method fails,
1422                          * mmap_region() nullifies vma->vm_file
1423                          * before calling this function to clean up.
1424                          * Since no pte has actually been setup, it is
1425                          * safe to do nothing in this case.
1426                          */
1427                         if (vma->vm_file) {
1428                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1429                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1430                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1431                         }
1432                 } else
1433                         unmap_page_range(tlb, vma, start, end, details);
1434         }
1435 }
1436
1437 /**
1438  * unmap_vmas - unmap a range of memory covered by a list of vma's
1439  * @tlb: address of the caller's struct mmu_gather
1440  * @vma: the starting vma
1441  * @start_addr: virtual address at which to start unmapping
1442  * @end_addr: virtual address at which to end unmapping
1443  *
1444  * Unmap all pages in the vma list.
1445  *
1446  * Only addresses between `start' and `end' will be unmapped.
1447  *
1448  * The VMA list must be sorted in ascending virtual address order.
1449  *
1450  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1451  * range after unmap_vmas() returns.  So the only responsibility here is to
1452  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1453  * drops the lock and schedules.
1454  */
1455 void unmap_vmas(struct mmu_gather *tlb,
1456                 struct vm_area_struct *vma, unsigned long start_addr,
1457                 unsigned long end_addr)
1458 {
1459         struct mm_struct *mm = vma->vm_mm;
1460
1461         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1462         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1463                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1464         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1465 }
1466
1467 /**
1468  * zap_page_range - remove user pages in a given range
1469  * @vma: vm_area_struct holding the applicable pages
1470  * @start: starting address of pages to zap
1471  * @size: number of bytes to zap
1472  *
1473  * Caller must protect the VMA list
1474  */
1475 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1476                 unsigned long size)
1477 {
1478         struct mm_struct *mm = vma->vm_mm;
1479         struct mmu_gather tlb;
1480         unsigned long end = start + size;
1481
1482         lru_add_drain();
1483         tlb_gather_mmu(&tlb, mm, start, end);
1484         update_hiwater_rss(mm);
1485         mmu_notifier_invalidate_range_start(mm, start, end);
1486         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1487                 unmap_single_vma(&tlb, vma, start, end, NULL);
1488         mmu_notifier_invalidate_range_end(mm, start, end);
1489         tlb_finish_mmu(&tlb, start, end);
1490 }
1491
1492 /**
1493  * zap_page_range_single - remove user pages in a given range
1494  * @vma: vm_area_struct holding the applicable pages
1495  * @address: starting address of pages to zap
1496  * @size: number of bytes to zap
1497  * @details: details of shared cache invalidation
1498  *
1499  * The range must fit into one VMA.
1500  */
1501 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1502                 unsigned long size, struct zap_details *details)
1503 {
1504         struct mm_struct *mm = vma->vm_mm;
1505         struct mmu_gather tlb;
1506         unsigned long end = address + size;
1507
1508         lru_add_drain();
1509         tlb_gather_mmu(&tlb, mm, address, end);
1510         update_hiwater_rss(mm);
1511         mmu_notifier_invalidate_range_start(mm, address, end);
1512         unmap_single_vma(&tlb, vma, address, end, details);
1513         mmu_notifier_invalidate_range_end(mm, address, end);
1514         tlb_finish_mmu(&tlb, address, end);
1515 }
1516
1517 /**
1518  * zap_vma_ptes - remove ptes mapping the vma
1519  * @vma: vm_area_struct holding ptes to be zapped
1520  * @address: starting address of pages to zap
1521  * @size: number of bytes to zap
1522  *
1523  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1524  *
1525  * The entire address range must be fully contained within the vma.
1526  *
1527  * Returns 0 if successful.
1528  */
1529 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1530                 unsigned long size)
1531 {
1532         if (address < vma->vm_start || address + size > vma->vm_end ||
1533                         !(vma->vm_flags & VM_PFNMAP))
1534                 return -1;
1535         zap_page_range_single(vma, address, size, NULL);
1536         return 0;
1537 }
1538 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1539
1540 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1541                         spinlock_t **ptl)
1542 {
1543         pgd_t *pgd;
1544         p4d_t *p4d;
1545         pud_t *pud;
1546         pmd_t *pmd;
1547
1548         pgd = pgd_offset(mm, addr);
1549         p4d = p4d_alloc(mm, pgd, addr);
1550         if (!p4d)
1551                 return NULL;
1552         pud = pud_alloc(mm, p4d, addr);
1553         if (!pud)
1554                 return NULL;
1555         pmd = pmd_alloc(mm, pud, addr);
1556         if (!pmd)
1557                 return NULL;
1558
1559         VM_BUG_ON(pmd_trans_huge(*pmd));
1560         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1561 }
1562
1563 /*
1564  * This is the old fallback for page remapping.
1565  *
1566  * For historical reasons, it only allows reserved pages. Only
1567  * old drivers should use this, and they needed to mark their
1568  * pages reserved for the old functions anyway.
1569  */
1570 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1571                         struct page *page, pgprot_t prot)
1572 {
1573         struct mm_struct *mm = vma->vm_mm;
1574         int retval;
1575         pte_t *pte;
1576         spinlock_t *ptl;
1577
1578         retval = -EINVAL;
1579         if (PageAnon(page))
1580                 goto out;
1581         retval = -ENOMEM;
1582         flush_dcache_page(page);
1583         pte = get_locked_pte(mm, addr, &ptl);
1584         if (!pte)
1585                 goto out;
1586         retval = -EBUSY;
1587         if (!pte_none(*pte))
1588                 goto out_unlock;
1589
1590         /* Ok, finally just insert the thing.. */
1591         get_page(page);
1592         inc_mm_counter_fast(mm, mm_counter_file(page));
1593         page_add_file_rmap(page, false);
1594         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1595
1596         retval = 0;
1597         pte_unmap_unlock(pte, ptl);
1598         return retval;
1599 out_unlock:
1600         pte_unmap_unlock(pte, ptl);
1601 out:
1602         return retval;
1603 }
1604
1605 /**
1606  * vm_insert_page - insert single page into user vma
1607  * @vma: user vma to map to
1608  * @addr: target user address of this page
1609  * @page: source kernel page
1610  *
1611  * This allows drivers to insert individual pages they've allocated
1612  * into a user vma.
1613  *
1614  * The page has to be a nice clean _individual_ kernel allocation.
1615  * If you allocate a compound page, you need to have marked it as
1616  * such (__GFP_COMP), or manually just split the page up yourself
1617  * (see split_page()).
1618  *
1619  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1620  * took an arbitrary page protection parameter. This doesn't allow
1621  * that. Your vma protection will have to be set up correctly, which
1622  * means that if you want a shared writable mapping, you'd better
1623  * ask for a shared writable mapping!
1624  *
1625  * The page does not need to be reserved.
1626  *
1627  * Usually this function is called from f_op->mmap() handler
1628  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1629  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1630  * function from other places, for example from page-fault handler.
1631  */
1632 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1633                         struct page *page)
1634 {
1635         if (addr < vma->vm_start || addr >= vma->vm_end)
1636                 return -EFAULT;
1637         if (!page_count(page))
1638                 return -EINVAL;
1639         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1640                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1641                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1642                 vma->vm_flags |= VM_MIXEDMAP;
1643         }
1644         return insert_page(vma, addr, page, vma->vm_page_prot);
1645 }
1646 EXPORT_SYMBOL(vm_insert_page);
1647
1648 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1649                         pfn_t pfn, pgprot_t prot)
1650 {
1651         struct mm_struct *mm = vma->vm_mm;
1652         int retval;
1653         pte_t *pte, entry;
1654         spinlock_t *ptl;
1655
1656         retval = -ENOMEM;
1657         pte = get_locked_pte(mm, addr, &ptl);
1658         if (!pte)
1659                 goto out;
1660         retval = -EBUSY;
1661         if (!pte_none(*pte))
1662                 goto out_unlock;
1663
1664         /* Ok, finally just insert the thing.. */
1665         if (pfn_t_devmap(pfn))
1666                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1667         else
1668                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1669         set_pte_at(mm, addr, pte, entry);
1670         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1671
1672         retval = 0;
1673 out_unlock:
1674         pte_unmap_unlock(pte, ptl);
1675 out:
1676         return retval;
1677 }
1678
1679 /**
1680  * vm_insert_pfn - insert single pfn into user vma
1681  * @vma: user vma to map to
1682  * @addr: target user address of this page
1683  * @pfn: source kernel pfn
1684  *
1685  * Similar to vm_insert_page, this allows drivers to insert individual pages
1686  * they've allocated into a user vma. Same comments apply.
1687  *
1688  * This function should only be called from a vm_ops->fault handler, and
1689  * in that case the handler should return NULL.
1690  *
1691  * vma cannot be a COW mapping.
1692  *
1693  * As this is called only for pages that do not currently exist, we
1694  * do not need to flush old virtual caches or the TLB.
1695  */
1696 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1697                         unsigned long pfn)
1698 {
1699         return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1700 }
1701 EXPORT_SYMBOL(vm_insert_pfn);
1702
1703 /**
1704  * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1705  * @vma: user vma to map to
1706  * @addr: target user address of this page
1707  * @pfn: source kernel pfn
1708  * @pgprot: pgprot flags for the inserted page
1709  *
1710  * This is exactly like vm_insert_pfn, except that it allows drivers to
1711  * to override pgprot on a per-page basis.
1712  *
1713  * This only makes sense for IO mappings, and it makes no sense for
1714  * cow mappings.  In general, using multiple vmas is preferable;
1715  * vm_insert_pfn_prot should only be used if using multiple VMAs is
1716  * impractical.
1717  */
1718 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1719                         unsigned long pfn, pgprot_t pgprot)
1720 {
1721         int ret;
1722         /*
1723          * Technically, architectures with pte_special can avoid all these
1724          * restrictions (same for remap_pfn_range).  However we would like
1725          * consistency in testing and feature parity among all, so we should
1726          * try to keep these invariants in place for everybody.
1727          */
1728         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1729         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1730                                                 (VM_PFNMAP|VM_MIXEDMAP));
1731         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1732         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1733
1734         if (addr < vma->vm_start || addr >= vma->vm_end)
1735                 return -EFAULT;
1736
1737         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1738
1739         ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1740
1741         return ret;
1742 }
1743 EXPORT_SYMBOL(vm_insert_pfn_prot);
1744
1745 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1746                         pfn_t pfn)
1747 {
1748         pgprot_t pgprot = vma->vm_page_prot;
1749
1750         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1751
1752         if (addr < vma->vm_start || addr >= vma->vm_end)
1753                 return -EFAULT;
1754
1755         track_pfn_insert(vma, &pgprot, pfn);
1756
1757         /*
1758          * If we don't have pte special, then we have to use the pfn_valid()
1759          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1760          * refcount the page if pfn_valid is true (hence insert_page rather
1761          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1762          * without pte special, it would there be refcounted as a normal page.
1763          */
1764         if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1765                 struct page *page;
1766
1767                 /*
1768                  * At this point we are committed to insert_page()
1769                  * regardless of whether the caller specified flags that
1770                  * result in pfn_t_has_page() == false.
1771                  */
1772                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1773                 return insert_page(vma, addr, page, pgprot);
1774         }
1775         return insert_pfn(vma, addr, pfn, pgprot);
1776 }
1777 EXPORT_SYMBOL(vm_insert_mixed);
1778
1779 /*
1780  * maps a range of physical memory into the requested pages. the old
1781  * mappings are removed. any references to nonexistent pages results
1782  * in null mappings (currently treated as "copy-on-access")
1783  */
1784 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1785                         unsigned long addr, unsigned long end,
1786                         unsigned long pfn, pgprot_t prot)
1787 {
1788         pte_t *pte;
1789         spinlock_t *ptl;
1790
1791         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1792         if (!pte)
1793                 return -ENOMEM;
1794         arch_enter_lazy_mmu_mode();
1795         do {
1796                 BUG_ON(!pte_none(*pte));
1797                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1798                 pfn++;
1799         } while (pte++, addr += PAGE_SIZE, addr != end);
1800         arch_leave_lazy_mmu_mode();
1801         pte_unmap_unlock(pte - 1, ptl);
1802         return 0;
1803 }
1804
1805 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1806                         unsigned long addr, unsigned long end,
1807                         unsigned long pfn, pgprot_t prot)
1808 {
1809         pmd_t *pmd;
1810         unsigned long next;
1811
1812         pfn -= addr >> PAGE_SHIFT;
1813         pmd = pmd_alloc(mm, pud, addr);
1814         if (!pmd)
1815                 return -ENOMEM;
1816         VM_BUG_ON(pmd_trans_huge(*pmd));
1817         do {
1818                 next = pmd_addr_end(addr, end);
1819                 if (remap_pte_range(mm, pmd, addr, next,
1820                                 pfn + (addr >> PAGE_SHIFT), prot))
1821                         return -ENOMEM;
1822         } while (pmd++, addr = next, addr != end);
1823         return 0;
1824 }
1825
1826 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1827                         unsigned long addr, unsigned long end,
1828                         unsigned long pfn, pgprot_t prot)
1829 {
1830         pud_t *pud;
1831         unsigned long next;
1832
1833         pfn -= addr >> PAGE_SHIFT;
1834         pud = pud_alloc(mm, p4d, addr);
1835         if (!pud)
1836                 return -ENOMEM;
1837         do {
1838                 next = pud_addr_end(addr, end);
1839                 if (remap_pmd_range(mm, pud, addr, next,
1840                                 pfn + (addr >> PAGE_SHIFT), prot))
1841                         return -ENOMEM;
1842         } while (pud++, addr = next, addr != end);
1843         return 0;
1844 }
1845
1846 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1847                         unsigned long addr, unsigned long end,
1848                         unsigned long pfn, pgprot_t prot)
1849 {
1850         p4d_t *p4d;
1851         unsigned long next;
1852
1853         pfn -= addr >> PAGE_SHIFT;
1854         p4d = p4d_alloc(mm, pgd, addr);
1855         if (!p4d)
1856                 return -ENOMEM;
1857         do {
1858                 next = p4d_addr_end(addr, end);
1859                 if (remap_pud_range(mm, p4d, addr, next,
1860                                 pfn + (addr >> PAGE_SHIFT), prot))
1861                         return -ENOMEM;
1862         } while (p4d++, addr = next, addr != end);
1863         return 0;
1864 }
1865
1866 /**
1867  * remap_pfn_range - remap kernel memory to userspace
1868  * @vma: user vma to map to
1869  * @addr: target user address to start at
1870  * @pfn: physical address of kernel memory
1871  * @size: size of map area
1872  * @prot: page protection flags for this mapping
1873  *
1874  *  Note: this is only safe if the mm semaphore is held when called.
1875  */
1876 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1877                     unsigned long pfn, unsigned long size, pgprot_t prot)
1878 {
1879         pgd_t *pgd;
1880         unsigned long next;
1881         unsigned long end = addr + PAGE_ALIGN(size);
1882         struct mm_struct *mm = vma->vm_mm;
1883         unsigned long remap_pfn = pfn;
1884         int err;
1885
1886         /*
1887          * Physically remapped pages are special. Tell the
1888          * rest of the world about it:
1889          *   VM_IO tells people not to look at these pages
1890          *      (accesses can have side effects).
1891          *   VM_PFNMAP tells the core MM that the base pages are just
1892          *      raw PFN mappings, and do not have a "struct page" associated
1893          *      with them.
1894          *   VM_DONTEXPAND
1895          *      Disable vma merging and expanding with mremap().
1896          *   VM_DONTDUMP
1897          *      Omit vma from core dump, even when VM_IO turned off.
1898          *
1899          * There's a horrible special case to handle copy-on-write
1900          * behaviour that some programs depend on. We mark the "original"
1901          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1902          * See vm_normal_page() for details.
1903          */
1904         if (is_cow_mapping(vma->vm_flags)) {
1905                 if (addr != vma->vm_start || end != vma->vm_end)
1906                         return -EINVAL;
1907                 vma->vm_pgoff = pfn;
1908         }
1909
1910         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1911         if (err)
1912                 return -EINVAL;
1913
1914         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1915
1916         BUG_ON(addr >= end);
1917         pfn -= addr >> PAGE_SHIFT;
1918         pgd = pgd_offset(mm, addr);
1919         flush_cache_range(vma, addr, end);
1920         do {
1921                 next = pgd_addr_end(addr, end);
1922                 err = remap_p4d_range(mm, pgd, addr, next,
1923                                 pfn + (addr >> PAGE_SHIFT), prot);
1924                 if (err)
1925                         break;
1926         } while (pgd++, addr = next, addr != end);
1927
1928         if (err)
1929                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1930
1931         return err;
1932 }
1933 EXPORT_SYMBOL(remap_pfn_range);
1934
1935 /**
1936  * vm_iomap_memory - remap memory to userspace
1937  * @vma: user vma to map to
1938  * @start: start of area
1939  * @len: size of area
1940  *
1941  * This is a simplified io_remap_pfn_range() for common driver use. The
1942  * driver just needs to give us the physical memory range to be mapped,
1943  * we'll figure out the rest from the vma information.
1944  *
1945  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1946  * whatever write-combining details or similar.
1947  */
1948 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1949 {
1950         unsigned long vm_len, pfn, pages;
1951
1952         /* Check that the physical memory area passed in looks valid */
1953         if (start + len < start)
1954                 return -EINVAL;
1955         /*
1956          * You *really* shouldn't map things that aren't page-aligned,
1957          * but we've historically allowed it because IO memory might
1958          * just have smaller alignment.
1959          */
1960         len += start & ~PAGE_MASK;
1961         pfn = start >> PAGE_SHIFT;
1962         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1963         if (pfn + pages < pfn)
1964                 return -EINVAL;
1965
1966         /* We start the mapping 'vm_pgoff' pages into the area */
1967         if (vma->vm_pgoff > pages)
1968                 return -EINVAL;
1969         pfn += vma->vm_pgoff;
1970         pages -= vma->vm_pgoff;
1971
1972         /* Can we fit all of the mapping? */
1973         vm_len = vma->vm_end - vma->vm_start;
1974         if (vm_len >> PAGE_SHIFT > pages)
1975                 return -EINVAL;
1976
1977         /* Ok, let it rip */
1978         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1979 }
1980 EXPORT_SYMBOL(vm_iomap_memory);
1981
1982 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1983                                      unsigned long addr, unsigned long end,
1984                                      pte_fn_t fn, void *data)
1985 {
1986         pte_t *pte;
1987         int err;
1988         pgtable_t token;
1989         spinlock_t *uninitialized_var(ptl);
1990
1991         pte = (mm == &init_mm) ?
1992                 pte_alloc_kernel(pmd, addr) :
1993                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1994         if (!pte)
1995                 return -ENOMEM;
1996
1997         BUG_ON(pmd_huge(*pmd));
1998
1999         arch_enter_lazy_mmu_mode();
2000
2001         token = pmd_pgtable(*pmd);
2002
2003         do {
2004                 err = fn(pte++, token, addr, data);
2005                 if (err)
2006                         break;
2007         } while (addr += PAGE_SIZE, addr != end);
2008
2009         arch_leave_lazy_mmu_mode();
2010
2011         if (mm != &init_mm)
2012                 pte_unmap_unlock(pte-1, ptl);
2013         return err;
2014 }
2015
2016 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2017                                      unsigned long addr, unsigned long end,
2018                                      pte_fn_t fn, void *data)
2019 {
2020         pmd_t *pmd;
2021         unsigned long next;
2022         int err;
2023
2024         BUG_ON(pud_huge(*pud));
2025
2026         pmd = pmd_alloc(mm, pud, addr);
2027         if (!pmd)
2028                 return -ENOMEM;
2029         do {
2030                 next = pmd_addr_end(addr, end);
2031                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2032                 if (err)
2033                         break;
2034         } while (pmd++, addr = next, addr != end);
2035         return err;
2036 }
2037
2038 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2039                                      unsigned long addr, unsigned long end,
2040                                      pte_fn_t fn, void *data)
2041 {
2042         pud_t *pud;
2043         unsigned long next;
2044         int err;
2045
2046         pud = pud_alloc(mm, p4d, addr);
2047         if (!pud)
2048                 return -ENOMEM;
2049         do {
2050                 next = pud_addr_end(addr, end);
2051                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2052                 if (err)
2053                         break;
2054         } while (pud++, addr = next, addr != end);
2055         return err;
2056 }
2057
2058 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2059                                      unsigned long addr, unsigned long end,
2060                                      pte_fn_t fn, void *data)
2061 {
2062         p4d_t *p4d;
2063         unsigned long next;
2064         int err;
2065
2066         p4d = p4d_alloc(mm, pgd, addr);
2067         if (!p4d)
2068                 return -ENOMEM;
2069         do {
2070                 next = p4d_addr_end(addr, end);
2071                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2072                 if (err)
2073                         break;
2074         } while (p4d++, addr = next, addr != end);
2075         return err;
2076 }
2077
2078 /*
2079  * Scan a region of virtual memory, filling in page tables as necessary
2080  * and calling a provided function on each leaf page table.
2081  */
2082 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2083                         unsigned long size, pte_fn_t fn, void *data)
2084 {
2085         pgd_t *pgd;
2086         unsigned long next;
2087         unsigned long end = addr + size;
2088         int err;
2089
2090         if (WARN_ON(addr >= end))
2091                 return -EINVAL;
2092
2093         pgd = pgd_offset(mm, addr);
2094         do {
2095                 next = pgd_addr_end(addr, end);
2096                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2097                 if (err)
2098                         break;
2099         } while (pgd++, addr = next, addr != end);
2100
2101         return err;
2102 }
2103 EXPORT_SYMBOL_GPL(apply_to_page_range);
2104
2105 /*
2106  * handle_pte_fault chooses page fault handler according to an entry which was
2107  * read non-atomically.  Before making any commitment, on those architectures
2108  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2109  * parts, do_swap_page must check under lock before unmapping the pte and
2110  * proceeding (but do_wp_page is only called after already making such a check;
2111  * and do_anonymous_page can safely check later on).
2112  */
2113 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2114                                 pte_t *page_table, pte_t orig_pte)
2115 {
2116         int same = 1;
2117 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2118         if (sizeof(pte_t) > sizeof(unsigned long)) {
2119                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2120                 spin_lock(ptl);
2121                 same = pte_same(*page_table, orig_pte);
2122                 spin_unlock(ptl);
2123         }
2124 #endif
2125         pte_unmap(page_table);
2126         return same;
2127 }
2128
2129 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2130 {
2131         debug_dma_assert_idle(src);
2132
2133         /*
2134          * If the source page was a PFN mapping, we don't have
2135          * a "struct page" for it. We do a best-effort copy by
2136          * just copying from the original user address. If that
2137          * fails, we just zero-fill it. Live with it.
2138          */
2139         if (unlikely(!src)) {
2140                 void *kaddr = kmap_atomic(dst);
2141                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2142
2143                 /*
2144                  * This really shouldn't fail, because the page is there
2145                  * in the page tables. But it might just be unreadable,
2146                  * in which case we just give up and fill the result with
2147                  * zeroes.
2148                  */
2149                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2150                         clear_page(kaddr);
2151                 kunmap_atomic(kaddr);
2152                 flush_dcache_page(dst);
2153         } else
2154                 copy_user_highpage(dst, src, va, vma);
2155 }
2156
2157 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2158 {
2159         struct file *vm_file = vma->vm_file;
2160
2161         if (vm_file)
2162                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2163
2164         /*
2165          * Special mappings (e.g. VDSO) do not have any file so fake
2166          * a default GFP_KERNEL for them.
2167          */
2168         return GFP_KERNEL;
2169 }
2170
2171 /*
2172  * Notify the address space that the page is about to become writable so that
2173  * it can prohibit this or wait for the page to get into an appropriate state.
2174  *
2175  * We do this without the lock held, so that it can sleep if it needs to.
2176  */
2177 static int do_page_mkwrite(struct vm_fault *vmf)
2178 {
2179         int ret;
2180         struct page *page = vmf->page;
2181         unsigned int old_flags = vmf->flags;
2182
2183         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2184
2185         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2186         /* Restore original flags so that caller is not surprised */
2187         vmf->flags = old_flags;
2188         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2189                 return ret;
2190         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2191                 lock_page(page);
2192                 if (!page->mapping) {
2193                         unlock_page(page);
2194                         return 0; /* retry */
2195                 }
2196                 ret |= VM_FAULT_LOCKED;
2197         } else
2198                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2199         return ret;
2200 }
2201
2202 /*
2203  * Handle dirtying of a page in shared file mapping on a write fault.
2204  *
2205  * The function expects the page to be locked and unlocks it.
2206  */
2207 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2208                                     struct page *page)
2209 {
2210         struct address_space *mapping;
2211         bool dirtied;
2212         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2213
2214         dirtied = set_page_dirty(page);
2215         VM_BUG_ON_PAGE(PageAnon(page), page);
2216         /*
2217          * Take a local copy of the address_space - page.mapping may be zeroed
2218          * by truncate after unlock_page().   The address_space itself remains
2219          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2220          * release semantics to prevent the compiler from undoing this copying.
2221          */
2222         mapping = page_rmapping(page);
2223         unlock_page(page);
2224
2225         if ((dirtied || page_mkwrite) && mapping) {
2226                 /*
2227                  * Some device drivers do not set page.mapping
2228                  * but still dirty their pages
2229                  */
2230                 balance_dirty_pages_ratelimited(mapping);
2231         }
2232
2233         if (!page_mkwrite)
2234                 file_update_time(vma->vm_file);
2235 }
2236
2237 /*
2238  * Handle write page faults for pages that can be reused in the current vma
2239  *
2240  * This can happen either due to the mapping being with the VM_SHARED flag,
2241  * or due to us being the last reference standing to the page. In either
2242  * case, all we need to do here is to mark the page as writable and update
2243  * any related book-keeping.
2244  */
2245 static inline void wp_page_reuse(struct vm_fault *vmf)
2246         __releases(vmf->ptl)
2247 {
2248         struct vm_area_struct *vma = vmf->vma;
2249         struct page *page = vmf->page;
2250         pte_t entry;
2251         /*
2252          * Clear the pages cpupid information as the existing
2253          * information potentially belongs to a now completely
2254          * unrelated process.
2255          */
2256         if (page)
2257                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2258
2259         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2260         entry = pte_mkyoung(vmf->orig_pte);
2261         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2262         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2263                 update_mmu_cache(vma, vmf->address, vmf->pte);
2264         pte_unmap_unlock(vmf->pte, vmf->ptl);
2265 }
2266
2267 /*
2268  * Handle the case of a page which we actually need to copy to a new page.
2269  *
2270  * Called with mmap_sem locked and the old page referenced, but
2271  * without the ptl held.
2272  *
2273  * High level logic flow:
2274  *
2275  * - Allocate a page, copy the content of the old page to the new one.
2276  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2277  * - Take the PTL. If the pte changed, bail out and release the allocated page
2278  * - If the pte is still the way we remember it, update the page table and all
2279  *   relevant references. This includes dropping the reference the page-table
2280  *   held to the old page, as well as updating the rmap.
2281  * - In any case, unlock the PTL and drop the reference we took to the old page.
2282  */
2283 static int wp_page_copy(struct vm_fault *vmf)
2284 {
2285         struct vm_area_struct *vma = vmf->vma;
2286         struct mm_struct *mm = vma->vm_mm;
2287         struct page *old_page = vmf->page;
2288         struct page *new_page = NULL;
2289         pte_t entry;
2290         int page_copied = 0;
2291         const unsigned long mmun_start = vmf->address & PAGE_MASK;
2292         const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2293         struct mem_cgroup *memcg;
2294
2295         if (unlikely(anon_vma_prepare(vma)))
2296                 goto oom;
2297
2298         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2299                 new_page = alloc_zeroed_user_highpage_movable(vma,
2300                                                               vmf->address);
2301                 if (!new_page)
2302                         goto oom;
2303         } else {
2304                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2305                                 vmf->address);
2306                 if (!new_page)
2307                         goto oom;
2308                 cow_user_page(new_page, old_page, vmf->address, vma);
2309         }
2310
2311         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2312                 goto oom_free_new;
2313
2314         __SetPageUptodate(new_page);
2315
2316         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2317
2318         /*
2319          * Re-check the pte - we dropped the lock
2320          */
2321         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2322         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2323                 if (old_page) {
2324                         if (!PageAnon(old_page)) {
2325                                 dec_mm_counter_fast(mm,
2326                                                 mm_counter_file(old_page));
2327                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2328                         }
2329                 } else {
2330                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2331                 }
2332                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2333                 entry = mk_pte(new_page, vma->vm_page_prot);
2334                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2335                 /*
2336                  * Clear the pte entry and flush it first, before updating the
2337                  * pte with the new entry. This will avoid a race condition
2338                  * seen in the presence of one thread doing SMC and another
2339                  * thread doing COW.
2340                  */
2341                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2342                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2343                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2344                 lru_cache_add_active_or_unevictable(new_page, vma);
2345                 /*
2346                  * We call the notify macro here because, when using secondary
2347                  * mmu page tables (such as kvm shadow page tables), we want the
2348                  * new page to be mapped directly into the secondary page table.
2349                  */
2350                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2351                 update_mmu_cache(vma, vmf->address, vmf->pte);
2352                 if (old_page) {
2353                         /*
2354                          * Only after switching the pte to the new page may
2355                          * we remove the mapcount here. Otherwise another
2356                          * process may come and find the rmap count decremented
2357                          * before the pte is switched to the new page, and
2358                          * "reuse" the old page writing into it while our pte
2359                          * here still points into it and can be read by other
2360                          * threads.
2361                          *
2362                          * The critical issue is to order this
2363                          * page_remove_rmap with the ptp_clear_flush above.
2364                          * Those stores are ordered by (if nothing else,)
2365                          * the barrier present in the atomic_add_negative
2366                          * in page_remove_rmap.
2367                          *
2368                          * Then the TLB flush in ptep_clear_flush ensures that
2369                          * no process can access the old page before the
2370                          * decremented mapcount is visible. And the old page
2371                          * cannot be reused until after the decremented
2372                          * mapcount is visible. So transitively, TLBs to
2373                          * old page will be flushed before it can be reused.
2374                          */
2375                         page_remove_rmap(old_page, false);
2376                 }
2377
2378                 /* Free the old page.. */
2379                 new_page = old_page;
2380                 page_copied = 1;
2381         } else {
2382                 mem_cgroup_cancel_charge(new_page, memcg, false);
2383         }
2384
2385         if (new_page)
2386                 put_page(new_page);
2387
2388         pte_unmap_unlock(vmf->pte, vmf->ptl);
2389         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2390         if (old_page) {
2391                 /*
2392                  * Don't let another task, with possibly unlocked vma,
2393                  * keep the mlocked page.
2394                  */
2395                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2396                         lock_page(old_page);    /* LRU manipulation */
2397                         if (PageMlocked(old_page))
2398                                 munlock_vma_page(old_page);
2399                         unlock_page(old_page);
2400                 }
2401                 put_page(old_page);
2402         }
2403         return page_copied ? VM_FAULT_WRITE : 0;
2404 oom_free_new:
2405         put_page(new_page);
2406 oom:
2407         if (old_page)
2408                 put_page(old_page);
2409         return VM_FAULT_OOM;
2410 }
2411
2412 /**
2413  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2414  *                        writeable once the page is prepared
2415  *
2416  * @vmf: structure describing the fault
2417  *
2418  * This function handles all that is needed to finish a write page fault in a
2419  * shared mapping due to PTE being read-only once the mapped page is prepared.
2420  * It handles locking of PTE and modifying it. The function returns
2421  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2422  * lock.
2423  *
2424  * The function expects the page to be locked or other protection against
2425  * concurrent faults / writeback (such as DAX radix tree locks).
2426  */
2427 int finish_mkwrite_fault(struct vm_fault *vmf)
2428 {
2429         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2430         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2431                                        &vmf->ptl);
2432         /*
2433          * We might have raced with another page fault while we released the
2434          * pte_offset_map_lock.
2435          */
2436         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2437                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2438                 return VM_FAULT_NOPAGE;
2439         }
2440         wp_page_reuse(vmf);
2441         return 0;
2442 }
2443
2444 /*
2445  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2446  * mapping
2447  */
2448 static int wp_pfn_shared(struct vm_fault *vmf)
2449 {
2450         struct vm_area_struct *vma = vmf->vma;
2451
2452         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2453                 int ret;
2454
2455                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2456                 vmf->flags |= FAULT_FLAG_MKWRITE;
2457                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2458                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2459                         return ret;
2460                 return finish_mkwrite_fault(vmf);
2461         }
2462         wp_page_reuse(vmf);
2463         return VM_FAULT_WRITE;
2464 }
2465
2466 static int wp_page_shared(struct vm_fault *vmf)
2467         __releases(vmf->ptl)
2468 {
2469         struct vm_area_struct *vma = vmf->vma;
2470
2471         get_page(vmf->page);
2472
2473         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2474                 int tmp;
2475
2476                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2477                 tmp = do_page_mkwrite(vmf);
2478                 if (unlikely(!tmp || (tmp &
2479                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2480                         put_page(vmf->page);
2481                         return tmp;
2482                 }
2483                 tmp = finish_mkwrite_fault(vmf);
2484                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2485                         unlock_page(vmf->page);
2486                         put_page(vmf->page);
2487                         return tmp;
2488                 }
2489         } else {
2490                 wp_page_reuse(vmf);
2491                 lock_page(vmf->page);
2492         }
2493         fault_dirty_shared_page(vma, vmf->page);
2494         put_page(vmf->page);
2495
2496         return VM_FAULT_WRITE;
2497 }
2498
2499 /*
2500  * This routine handles present pages, when users try to write
2501  * to a shared page. It is done by copying the page to a new address
2502  * and decrementing the shared-page counter for the old page.
2503  *
2504  * Note that this routine assumes that the protection checks have been
2505  * done by the caller (the low-level page fault routine in most cases).
2506  * Thus we can safely just mark it writable once we've done any necessary
2507  * COW.
2508  *
2509  * We also mark the page dirty at this point even though the page will
2510  * change only once the write actually happens. This avoids a few races,
2511  * and potentially makes it more efficient.
2512  *
2513  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2514  * but allow concurrent faults), with pte both mapped and locked.
2515  * We return with mmap_sem still held, but pte unmapped and unlocked.
2516  */
2517 static int do_wp_page(struct vm_fault *vmf)
2518         __releases(vmf->ptl)
2519 {
2520         struct vm_area_struct *vma = vmf->vma;
2521
2522         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2523         if (!vmf->page) {
2524                 /*
2525                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2526                  * VM_PFNMAP VMA.
2527                  *
2528                  * We should not cow pages in a shared writeable mapping.
2529                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2530                  */
2531                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2532                                      (VM_WRITE|VM_SHARED))
2533                         return wp_pfn_shared(vmf);
2534
2535                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2536                 return wp_page_copy(vmf);
2537         }
2538
2539         /*
2540          * Take out anonymous pages first, anonymous shared vmas are
2541          * not dirty accountable.
2542          */
2543         if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2544                 int total_mapcount;
2545                 if (!trylock_page(vmf->page)) {
2546                         get_page(vmf->page);
2547                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2548                         lock_page(vmf->page);
2549                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2550                                         vmf->address, &vmf->ptl);
2551                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2552                                 unlock_page(vmf->page);
2553                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2554                                 put_page(vmf->page);
2555                                 return 0;
2556                         }
2557                         put_page(vmf->page);
2558                 }
2559                 if (reuse_swap_page(vmf->page, &total_mapcount)) {
2560                         if (total_mapcount == 1) {
2561                                 /*
2562                                  * The page is all ours. Move it to
2563                                  * our anon_vma so the rmap code will
2564                                  * not search our parent or siblings.
2565                                  * Protected against the rmap code by
2566                                  * the page lock.
2567                                  */
2568                                 page_move_anon_rmap(vmf->page, vma);
2569                         }
2570                         unlock_page(vmf->page);
2571                         wp_page_reuse(vmf);
2572                         return VM_FAULT_WRITE;
2573                 }
2574                 unlock_page(vmf->page);
2575         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2576                                         (VM_WRITE|VM_SHARED))) {
2577                 return wp_page_shared(vmf);
2578         }
2579
2580         /*
2581          * Ok, we need to copy. Oh, well..
2582          */
2583         get_page(vmf->page);
2584
2585         pte_unmap_unlock(vmf->pte, vmf->ptl);
2586         return wp_page_copy(vmf);
2587 }
2588
2589 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2590                 unsigned long start_addr, unsigned long end_addr,
2591                 struct zap_details *details)
2592 {
2593         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2594 }
2595
2596 static inline void unmap_mapping_range_tree(struct rb_root *root,
2597                                             struct zap_details *details)
2598 {
2599         struct vm_area_struct *vma;
2600         pgoff_t vba, vea, zba, zea;
2601
2602         vma_interval_tree_foreach(vma, root,
2603                         details->first_index, details->last_index) {
2604
2605                 vba = vma->vm_pgoff;
2606                 vea = vba + vma_pages(vma) - 1;
2607                 zba = details->first_index;
2608                 if (zba < vba)
2609                         zba = vba;
2610                 zea = details->last_index;
2611                 if (zea > vea)
2612                         zea = vea;
2613
2614                 unmap_mapping_range_vma(vma,
2615                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2616                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2617                                 details);
2618         }
2619 }
2620
2621 /**
2622  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2623  * address_space corresponding to the specified page range in the underlying
2624  * file.
2625  *
2626  * @mapping: the address space containing mmaps to be unmapped.
2627  * @holebegin: byte in first page to unmap, relative to the start of
2628  * the underlying file.  This will be rounded down to a PAGE_SIZE
2629  * boundary.  Note that this is different from truncate_pagecache(), which
2630  * must keep the partial page.  In contrast, we must get rid of
2631  * partial pages.
2632  * @holelen: size of prospective hole in bytes.  This will be rounded
2633  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2634  * end of the file.
2635  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2636  * but 0 when invalidating pagecache, don't throw away private data.
2637  */
2638 void unmap_mapping_range(struct address_space *mapping,
2639                 loff_t const holebegin, loff_t const holelen, int even_cows)
2640 {
2641         struct zap_details details = { };
2642         pgoff_t hba = holebegin >> PAGE_SHIFT;
2643         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2644
2645         /* Check for overflow. */
2646         if (sizeof(holelen) > sizeof(hlen)) {
2647                 long long holeend =
2648                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2649                 if (holeend & ~(long long)ULONG_MAX)
2650                         hlen = ULONG_MAX - hba + 1;
2651         }
2652
2653         details.check_mapping = even_cows ? NULL : mapping;
2654         details.first_index = hba;
2655         details.last_index = hba + hlen - 1;
2656         if (details.last_index < details.first_index)
2657                 details.last_index = ULONG_MAX;
2658
2659         i_mmap_lock_write(mapping);
2660         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2661                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2662         i_mmap_unlock_write(mapping);
2663 }
2664 EXPORT_SYMBOL(unmap_mapping_range);
2665
2666 /*
2667  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2668  * but allow concurrent faults), and pte mapped but not yet locked.
2669  * We return with pte unmapped and unlocked.
2670  *
2671  * We return with the mmap_sem locked or unlocked in the same cases
2672  * as does filemap_fault().
2673  */
2674 int do_swap_page(struct vm_fault *vmf)
2675 {
2676         struct vm_area_struct *vma = vmf->vma;
2677         struct page *page, *swapcache;
2678         struct mem_cgroup *memcg;
2679         swp_entry_t entry;
2680         pte_t pte;
2681         int locked;
2682         int exclusive = 0;
2683         int ret = 0;
2684
2685         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2686                 goto out;
2687
2688         entry = pte_to_swp_entry(vmf->orig_pte);
2689         if (unlikely(non_swap_entry(entry))) {
2690                 if (is_migration_entry(entry)) {
2691                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2692                                              vmf->address);
2693                 } else if (is_hwpoison_entry(entry)) {
2694                         ret = VM_FAULT_HWPOISON;
2695                 } else {
2696                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2697                         ret = VM_FAULT_SIGBUS;
2698                 }
2699                 goto out;
2700         }
2701         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2702         page = lookup_swap_cache(entry);
2703         if (!page) {
2704                 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, vma,
2705                                         vmf->address);
2706                 if (!page) {
2707                         /*
2708                          * Back out if somebody else faulted in this pte
2709                          * while we released the pte lock.
2710                          */
2711                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2712                                         vmf->address, &vmf->ptl);
2713                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2714                                 ret = VM_FAULT_OOM;
2715                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2716                         goto unlock;
2717                 }
2718
2719                 /* Had to read the page from swap area: Major fault */
2720                 ret = VM_FAULT_MAJOR;
2721                 count_vm_event(PGMAJFAULT);
2722                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
2723         } else if (PageHWPoison(page)) {
2724                 /*
2725                  * hwpoisoned dirty swapcache pages are kept for killing
2726                  * owner processes (which may be unknown at hwpoison time)
2727                  */
2728                 ret = VM_FAULT_HWPOISON;
2729                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2730                 swapcache = page;
2731                 goto out_release;
2732         }
2733
2734         swapcache = page;
2735         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2736
2737         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2738         if (!locked) {
2739                 ret |= VM_FAULT_RETRY;
2740                 goto out_release;
2741         }
2742
2743         /*
2744          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2745          * release the swapcache from under us.  The page pin, and pte_same
2746          * test below, are not enough to exclude that.  Even if it is still
2747          * swapcache, we need to check that the page's swap has not changed.
2748          */
2749         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2750                 goto out_page;
2751
2752         page = ksm_might_need_to_copy(page, vma, vmf->address);
2753         if (unlikely(!page)) {
2754                 ret = VM_FAULT_OOM;
2755                 page = swapcache;
2756                 goto out_page;
2757         }
2758
2759         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2760                                 &memcg, false)) {
2761                 ret = VM_FAULT_OOM;
2762                 goto out_page;
2763         }
2764
2765         /*
2766          * Back out if somebody else already faulted in this pte.
2767          */
2768         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2769                         &vmf->ptl);
2770         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2771                 goto out_nomap;
2772
2773         if (unlikely(!PageUptodate(page))) {
2774                 ret = VM_FAULT_SIGBUS;
2775                 goto out_nomap;
2776         }
2777
2778         /*
2779          * The page isn't present yet, go ahead with the fault.
2780          *
2781          * Be careful about the sequence of operations here.
2782          * To get its accounting right, reuse_swap_page() must be called
2783          * while the page is counted on swap but not yet in mapcount i.e.
2784          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2785          * must be called after the swap_free(), or it will never succeed.
2786          */
2787
2788         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2789         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2790         pte = mk_pte(page, vma->vm_page_prot);
2791         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2792                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2793                 vmf->flags &= ~FAULT_FLAG_WRITE;
2794                 ret |= VM_FAULT_WRITE;
2795                 exclusive = RMAP_EXCLUSIVE;
2796         }
2797         flush_icache_page(vma, page);
2798         if (pte_swp_soft_dirty(vmf->orig_pte))
2799                 pte = pte_mksoft_dirty(pte);
2800         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2801         vmf->orig_pte = pte;
2802         if (page == swapcache) {
2803                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2804                 mem_cgroup_commit_charge(page, memcg, true, false);
2805                 activate_page(page);
2806         } else { /* ksm created a completely new copy */
2807                 page_add_new_anon_rmap(page, vma, vmf->address, false);
2808                 mem_cgroup_commit_charge(page, memcg, false, false);
2809                 lru_cache_add_active_or_unevictable(page, vma);
2810         }
2811
2812         swap_free(entry);
2813         if (mem_cgroup_swap_full(page) ||
2814             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2815                 try_to_free_swap(page);
2816         unlock_page(page);
2817         if (page != swapcache) {
2818                 /*
2819                  * Hold the lock to avoid the swap entry to be reused
2820                  * until we take the PT lock for the pte_same() check
2821                  * (to avoid false positives from pte_same). For
2822                  * further safety release the lock after the swap_free
2823                  * so that the swap count won't change under a
2824                  * parallel locked swapcache.
2825                  */
2826                 unlock_page(swapcache);
2827                 put_page(swapcache);
2828         }
2829
2830         if (vmf->flags & FAULT_FLAG_WRITE) {
2831                 ret |= do_wp_page(vmf);
2832                 if (ret & VM_FAULT_ERROR)
2833                         ret &= VM_FAULT_ERROR;
2834                 goto out;
2835         }
2836
2837         /* No need to invalidate - it was non-present before */
2838         update_mmu_cache(vma, vmf->address, vmf->pte);
2839 unlock:
2840         pte_unmap_unlock(vmf->pte, vmf->ptl);
2841 out:
2842         return ret;
2843 out_nomap:
2844         mem_cgroup_cancel_charge(page, memcg, false);
2845         pte_unmap_unlock(vmf->pte, vmf->ptl);
2846 out_page:
2847         unlock_page(page);
2848 out_release:
2849         put_page(page);
2850         if (page != swapcache) {
2851                 unlock_page(swapcache);
2852                 put_page(swapcache);
2853         }
2854         return ret;
2855 }
2856
2857 /*
2858  * This is like a special single-page "expand_{down|up}wards()",
2859  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2860  * doesn't hit another vma.
2861  */
2862 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2863 {
2864         address &= PAGE_MASK;
2865         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2866                 struct vm_area_struct *prev = vma->vm_prev;
2867
2868                 /*
2869                  * Is there a mapping abutting this one below?
2870                  *
2871                  * That's only ok if it's the same stack mapping
2872                  * that has gotten split..
2873                  */
2874                 if (prev && prev->vm_end == address)
2875                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2876
2877                 return expand_downwards(vma, address - PAGE_SIZE);
2878         }
2879         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2880                 struct vm_area_struct *next = vma->vm_next;
2881
2882                 /* As VM_GROWSDOWN but s/below/above/ */
2883                 if (next && next->vm_start == address + PAGE_SIZE)
2884                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2885
2886                 return expand_upwards(vma, address + PAGE_SIZE);
2887         }
2888         return 0;
2889 }
2890
2891 /*
2892  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2893  * but allow concurrent faults), and pte mapped but not yet locked.
2894  * We return with mmap_sem still held, but pte unmapped and unlocked.
2895  */
2896 static int do_anonymous_page(struct vm_fault *vmf)
2897 {
2898         struct vm_area_struct *vma = vmf->vma;
2899         struct mem_cgroup *memcg;
2900         struct page *page;
2901         pte_t entry;
2902
2903         /* File mapping without ->vm_ops ? */
2904         if (vma->vm_flags & VM_SHARED)
2905                 return VM_FAULT_SIGBUS;
2906
2907         /* Check if we need to add a guard page to the stack */
2908         if (check_stack_guard_page(vma, vmf->address) < 0)
2909                 return VM_FAULT_SIGSEGV;
2910
2911         /*
2912          * Use pte_alloc() instead of pte_alloc_map().  We can't run
2913          * pte_offset_map() on pmds where a huge pmd might be created
2914          * from a different thread.
2915          *
2916          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2917          * parallel threads are excluded by other means.
2918          *
2919          * Here we only have down_read(mmap_sem).
2920          */
2921         if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2922                 return VM_FAULT_OOM;
2923
2924         /* See the comment in pte_alloc_one_map() */
2925         if (unlikely(pmd_trans_unstable(vmf->pmd)))
2926                 return 0;
2927
2928         /* Use the zero-page for reads */
2929         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2930                         !mm_forbids_zeropage(vma->vm_mm)) {
2931                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2932                                                 vma->vm_page_prot));
2933                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2934                                 vmf->address, &vmf->ptl);
2935                 if (!pte_none(*vmf->pte))
2936                         goto unlock;
2937                 /* Deliver the page fault to userland, check inside PT lock */
2938                 if (userfaultfd_missing(vma)) {
2939                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2940                         return handle_userfault(vmf, VM_UFFD_MISSING);
2941                 }
2942                 goto setpte;
2943         }
2944
2945         /* Allocate our own private page. */
2946         if (unlikely(anon_vma_prepare(vma)))
2947                 goto oom;
2948         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2949         if (!page)
2950                 goto oom;
2951
2952         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2953                 goto oom_free_page;
2954
2955         /*
2956          * The memory barrier inside __SetPageUptodate makes sure that
2957          * preceeding stores to the page contents become visible before
2958          * the set_pte_at() write.
2959          */
2960         __SetPageUptodate(page);
2961
2962         entry = mk_pte(page, vma->vm_page_prot);
2963         if (vma->vm_flags & VM_WRITE)
2964                 entry = pte_mkwrite(pte_mkdirty(entry));
2965
2966         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2967                         &vmf->ptl);
2968         if (!pte_none(*vmf->pte))
2969                 goto release;
2970
2971         /* Deliver the page fault to userland, check inside PT lock */
2972         if (userfaultfd_missing(vma)) {
2973                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2974                 mem_cgroup_cancel_charge(page, memcg, false);
2975                 put_page(page);
2976                 return handle_userfault(vmf, VM_UFFD_MISSING);
2977         }
2978
2979         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2980         page_add_new_anon_rmap(page, vma, vmf->address, false);
2981         mem_cgroup_commit_charge(page, memcg, false, false);
2982         lru_cache_add_active_or_unevictable(page, vma);
2983 setpte:
2984         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2985
2986         /* No need to invalidate - it was non-present before */
2987         update_mmu_cache(vma, vmf->address, vmf->pte);
2988 unlock:
2989         pte_unmap_unlock(vmf->pte, vmf->ptl);
2990         return 0;
2991 release:
2992         mem_cgroup_cancel_charge(page, memcg, false);
2993         put_page(page);
2994         goto unlock;
2995 oom_free_page:
2996         put_page(page);
2997 oom:
2998         return VM_FAULT_OOM;
2999 }
3000
3001 /*
3002  * The mmap_sem must have been held on entry, and may have been
3003  * released depending on flags and vma->vm_ops->fault() return value.
3004  * See filemap_fault() and __lock_page_retry().
3005  */
3006 static int __do_fault(struct vm_fault *vmf)
3007 {
3008         struct vm_area_struct *vma = vmf->vma;
3009         int ret;
3010
3011         ret = vma->vm_ops->fault(vmf);
3012         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3013                             VM_FAULT_DONE_COW)))
3014                 return ret;
3015
3016         if (unlikely(PageHWPoison(vmf->page))) {
3017                 if (ret & VM_FAULT_LOCKED)
3018                         unlock_page(vmf->page);
3019                 put_page(vmf->page);
3020                 vmf->page = NULL;
3021                 return VM_FAULT_HWPOISON;
3022         }
3023
3024         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3025                 lock_page(vmf->page);
3026         else
3027                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3028
3029         return ret;
3030 }
3031
3032 static int pte_alloc_one_map(struct vm_fault *vmf)
3033 {
3034         struct vm_area_struct *vma = vmf->vma;
3035
3036         if (!pmd_none(*vmf->pmd))
3037                 goto map_pte;
3038         if (vmf->prealloc_pte) {
3039                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3040                 if (unlikely(!pmd_none(*vmf->pmd))) {
3041                         spin_unlock(vmf->ptl);
3042                         goto map_pte;
3043                 }
3044
3045                 atomic_long_inc(&vma->vm_mm->nr_ptes);
3046                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3047                 spin_unlock(vmf->ptl);
3048                 vmf->prealloc_pte = NULL;
3049         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3050                 return VM_FAULT_OOM;
3051         }
3052 map_pte:
3053         /*
3054          * If a huge pmd materialized under us just retry later.  Use
3055          * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3056          * didn't become pmd_trans_huge under us and then back to pmd_none, as
3057          * a result of MADV_DONTNEED running immediately after a huge pmd fault
3058          * in a different thread of this mm, in turn leading to a misleading
3059          * pmd_trans_huge() retval.  All we have to ensure is that it is a
3060          * regular pmd that we can walk with pte_offset_map() and we can do that
3061          * through an atomic read in C, which is what pmd_trans_unstable()
3062          * provides.
3063          */
3064         if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
3065                 return VM_FAULT_NOPAGE;
3066
3067         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3068                         &vmf->ptl);
3069         return 0;
3070 }
3071
3072 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3073
3074 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3075 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3076                 unsigned long haddr)
3077 {
3078         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3079                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3080                 return false;
3081         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3082                 return false;
3083         return true;
3084 }
3085
3086 static void deposit_prealloc_pte(struct vm_fault *vmf)
3087 {
3088         struct vm_area_struct *vma = vmf->vma;
3089
3090         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3091         /*
3092          * We are going to consume the prealloc table,
3093          * count that as nr_ptes.
3094          */
3095         atomic_long_inc(&vma->vm_mm->nr_ptes);
3096         vmf->prealloc_pte = NULL;
3097 }
3098
3099 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3100 {
3101         struct vm_area_struct *vma = vmf->vma;
3102         bool write = vmf->flags & FAULT_FLAG_WRITE;
3103         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3104         pmd_t entry;
3105         int i, ret;
3106
3107         if (!transhuge_vma_suitable(vma, haddr))
3108                 return VM_FAULT_FALLBACK;
3109
3110         ret = VM_FAULT_FALLBACK;
3111         page = compound_head(page);
3112
3113         /*
3114          * Archs like ppc64 need additonal space to store information
3115          * related to pte entry. Use the preallocated table for that.
3116          */
3117         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3118                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3119                 if (!vmf->prealloc_pte)
3120                         return VM_FAULT_OOM;
3121                 smp_wmb(); /* See comment in __pte_alloc() */
3122         }
3123
3124         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3125         if (unlikely(!pmd_none(*vmf->pmd)))
3126                 goto out;
3127
3128         for (i = 0; i < HPAGE_PMD_NR; i++)
3129                 flush_icache_page(vma, page + i);
3130
3131         entry = mk_huge_pmd(page, vma->vm_page_prot);
3132         if (write)
3133                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3134
3135         add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3136         page_add_file_rmap(page, true);
3137         /*
3138          * deposit and withdraw with pmd lock held
3139          */
3140         if (arch_needs_pgtable_deposit())
3141                 deposit_prealloc_pte(vmf);
3142
3143         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3144
3145         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3146
3147         /* fault is handled */
3148         ret = 0;
3149         count_vm_event(THP_FILE_MAPPED);
3150 out:
3151         spin_unlock(vmf->ptl);
3152         return ret;
3153 }
3154 #else
3155 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3156 {
3157         BUILD_BUG();
3158         return 0;
3159 }
3160 #endif
3161
3162 /**
3163  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3164  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3165  *
3166  * @vmf: fault environment
3167  * @memcg: memcg to charge page (only for private mappings)
3168  * @page: page to map
3169  *
3170  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3171  * return.
3172  *
3173  * Target users are page handler itself and implementations of
3174  * vm_ops->map_pages.
3175  */
3176 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3177                 struct page *page)
3178 {
3179         struct vm_area_struct *vma = vmf->vma;
3180         bool write = vmf->flags & FAULT_FLAG_WRITE;
3181         pte_t entry;
3182         int ret;
3183
3184         if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3185                         IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3186                 /* THP on COW? */
3187                 VM_BUG_ON_PAGE(memcg, page);
3188
3189                 ret = do_set_pmd(vmf, page);
3190                 if (ret != VM_FAULT_FALLBACK)
3191                         return ret;
3192         }
3193
3194         if (!vmf->pte) {
3195                 ret = pte_alloc_one_map(vmf);
3196                 if (ret)
3197                         return ret;
3198         }
3199
3200         /* Re-check under ptl */
3201         if (unlikely(!pte_none(*vmf->pte)))
3202                 return VM_FAULT_NOPAGE;
3203
3204         flush_icache_page(vma, page);
3205         entry = mk_pte(page, vma->vm_page_prot);
3206         if (write)
3207                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3208         /* copy-on-write page */
3209         if (write && !(vma->vm_flags & VM_SHARED)) {
3210                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3211                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3212                 mem_cgroup_commit_charge(page, memcg, false, false);
3213                 lru_cache_add_active_or_unevictable(page, vma);
3214         } else {
3215                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3216                 page_add_file_rmap(page, false);
3217         }
3218         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3219
3220         /* no need to invalidate: a not-present page won't be cached */
3221         update_mmu_cache(vma, vmf->address, vmf->pte);
3222
3223         return 0;
3224 }
3225
3226
3227 /**
3228  * finish_fault - finish page fault once we have prepared the page to fault
3229  *
3230  * @vmf: structure describing the fault
3231  *
3232  * This function handles all that is needed to finish a page fault once the
3233  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3234  * given page, adds reverse page mapping, handles memcg charges and LRU
3235  * addition. The function returns 0 on success, VM_FAULT_ code in case of
3236  * error.
3237  *
3238  * The function expects the page to be locked and on success it consumes a
3239  * reference of a page being mapped (for the PTE which maps it).
3240  */
3241 int finish_fault(struct vm_fault *vmf)
3242 {
3243         struct page *page;
3244         int ret;
3245
3246         /* Did we COW the page? */
3247         if ((vmf->flags & FAULT_FLAG_WRITE) &&
3248             !(vmf->vma->vm_flags & VM_SHARED))
3249                 page = vmf->cow_page;
3250         else
3251                 page = vmf->page;
3252         ret = alloc_set_pte(vmf, vmf->memcg, page);
3253         if (vmf->pte)
3254                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3255         return ret;
3256 }
3257
3258 static unsigned long fault_around_bytes __read_mostly =
3259         rounddown_pow_of_two(65536);
3260
3261 #ifdef CONFIG_DEBUG_FS
3262 static int fault_around_bytes_get(void *data, u64 *val)
3263 {
3264         *val = fault_around_bytes;
3265         return 0;
3266 }
3267
3268 /*
3269  * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3270  * rounded down to nearest page order. It's what do_fault_around() expects to
3271  * see.
3272  */
3273 static int fault_around_bytes_set(void *data, u64 val)
3274 {
3275         if (val / PAGE_SIZE > PTRS_PER_PTE)
3276                 return -EINVAL;
3277         if (val > PAGE_SIZE)
3278                 fault_around_bytes = rounddown_pow_of_two(val);
3279         else
3280                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3281         return 0;
3282 }
3283 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3284                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3285
3286 static int __init fault_around_debugfs(void)
3287 {
3288         void *ret;
3289
3290         ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3291                         &fault_around_bytes_fops);
3292         if (!ret)
3293                 pr_warn("Failed to create fault_around_bytes in debugfs");
3294         return 0;
3295 }
3296 late_initcall(fault_around_debugfs);
3297 #endif
3298
3299 /*
3300  * do_fault_around() tries to map few pages around the fault address. The hope
3301  * is that the pages will be needed soon and this will lower the number of
3302  * faults to handle.
3303  *
3304  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3305  * not ready to be mapped: not up-to-date, locked, etc.
3306  *
3307  * This function is called with the page table lock taken. In the split ptlock
3308  * case the page table lock only protects only those entries which belong to
3309  * the page table corresponding to the fault address.
3310  *
3311  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3312  * only once.
3313  *
3314  * fault_around_pages() defines how many pages we'll try to map.
3315  * do_fault_around() expects it to return a power of two less than or equal to
3316  * PTRS_PER_PTE.
3317  *
3318  * The virtual address of the area that we map is naturally aligned to the
3319  * fault_around_pages() value (and therefore to page order).  This way it's
3320  * easier to guarantee that we don't cross page table boundaries.
3321  */
3322 static int do_fault_around(struct vm_fault *vmf)
3323 {
3324         unsigned long address = vmf->address, nr_pages, mask;
3325         pgoff_t start_pgoff = vmf->pgoff;
3326         pgoff_t end_pgoff;
3327         int off, ret = 0;
3328
3329         nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3330         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3331
3332         vmf->address = max(address & mask, vmf->vma->vm_start);
3333         off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3334         start_pgoff -= off;
3335
3336         /*
3337          *  end_pgoff is either end of page table or end of vma
3338          *  or fault_around_pages() from start_pgoff, depending what is nearest.
3339          */
3340         end_pgoff = start_pgoff -
3341                 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3342                 PTRS_PER_PTE - 1;
3343         end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3344                         start_pgoff + nr_pages - 1);
3345
3346         if (pmd_none(*vmf->pmd)) {
3347                 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3348                                                   vmf->address);
3349                 if (!vmf->prealloc_pte)
3350                         goto out;
3351                 smp_wmb(); /* See comment in __pte_alloc() */
3352         }
3353
3354         vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3355
3356         /* Huge page is mapped? Page fault is solved */
3357         if (pmd_trans_huge(*vmf->pmd)) {
3358                 ret = VM_FAULT_NOPAGE;
3359                 goto out;
3360         }
3361
3362         /* ->map_pages() haven't done anything useful. Cold page cache? */
3363         if (!vmf->pte)
3364                 goto out;
3365
3366         /* check if the page fault is solved */
3367         vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3368         if (!pte_none(*vmf->pte))
3369                 ret = VM_FAULT_NOPAGE;
3370         pte_unmap_unlock(vmf->pte, vmf->ptl);
3371 out:
3372         vmf->address = address;
3373         vmf->pte = NULL;
3374         return ret;
3375 }
3376
3377 static int do_read_fault(struct vm_fault *vmf)
3378 {
3379         struct vm_area_struct *vma = vmf->vma;
3380         int ret = 0;
3381
3382         /*
3383          * Let's call ->map_pages() first and use ->fault() as fallback
3384          * if page by the offset is not ready to be mapped (cold cache or
3385          * something).
3386          */
3387         if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3388                 ret = do_fault_around(vmf);
3389                 if (ret)
3390                         return ret;
3391         }
3392
3393         ret = __do_fault(vmf);
3394         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3395                 return ret;
3396
3397         ret |= finish_fault(vmf);
3398         unlock_page(vmf->page);
3399         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3400                 put_page(vmf->page);
3401         return ret;
3402 }
3403
3404 static int do_cow_fault(struct vm_fault *vmf)
3405 {
3406         struct vm_area_struct *vma = vmf->vma;
3407         int ret;
3408
3409         if (unlikely(anon_vma_prepare(vma)))
3410                 return VM_FAULT_OOM;
3411
3412         vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3413         if (!vmf->cow_page)
3414                 return VM_FAULT_OOM;
3415
3416         if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3417                                 &vmf->memcg, false)) {
3418                 put_page(vmf->cow_page);
3419                 return VM_FAULT_OOM;
3420         }
3421
3422         ret = __do_fault(vmf);
3423         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3424                 goto uncharge_out;
3425         if (ret & VM_FAULT_DONE_COW)
3426                 return ret;
3427
3428         copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3429         __SetPageUptodate(vmf->cow_page);
3430
3431         ret |= finish_fault(vmf);
3432         unlock_page(vmf->page);
3433         put_page(vmf->page);
3434         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3435                 goto uncharge_out;
3436         return ret;
3437 uncharge_out:
3438         mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3439         put_page(vmf->cow_page);
3440         return ret;
3441 }
3442
3443 static int do_shared_fault(struct vm_fault *vmf)
3444 {
3445         struct vm_area_struct *vma = vmf->vma;
3446         int ret, tmp;
3447
3448         ret = __do_fault(vmf);
3449         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3450                 return ret;
3451
3452         /*
3453          * Check if the backing address space wants to know that the page is
3454          * about to become writable
3455          */
3456         if (vma->vm_ops->page_mkwrite) {
3457                 unlock_page(vmf->page);
3458                 tmp = do_page_mkwrite(vmf);
3459                 if (unlikely(!tmp ||
3460                                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3461                         put_page(vmf->page);
3462                         return tmp;
3463                 }
3464         }
3465
3466         ret |= finish_fault(vmf);
3467         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3468                                         VM_FAULT_RETRY))) {
3469                 unlock_page(vmf->page);
3470                 put_page(vmf->page);
3471                 return ret;
3472         }
3473
3474         fault_dirty_shared_page(vma, vmf->page);
3475         return ret;
3476 }
3477
3478 /*
3479  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3480  * but allow concurrent faults).
3481  * The mmap_sem may have been released depending on flags and our
3482  * return value.  See filemap_fault() and __lock_page_or_retry().
3483  */
3484 static int do_fault(struct vm_fault *vmf)
3485 {
3486         struct vm_area_struct *vma = vmf->vma;
3487         int ret;
3488
3489         /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3490         if (!vma->vm_ops->fault)
3491                 ret = VM_FAULT_SIGBUS;
3492         else if (!(vmf->flags & FAULT_FLAG_WRITE))
3493                 ret = do_read_fault(vmf);
3494         else if (!(vma->vm_flags & VM_SHARED))
3495                 ret = do_cow_fault(vmf);
3496         else
3497                 ret = do_shared_fault(vmf);
3498
3499         /* preallocated pagetable is unused: free it */
3500         if (vmf->prealloc_pte) {
3501                 pte_free(vma->vm_mm, vmf->prealloc_pte);
3502                 vmf->prealloc_pte = NULL;
3503         }
3504         return ret;
3505 }
3506
3507 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3508                                 unsigned long addr, int page_nid,
3509                                 int *flags)
3510 {
3511         get_page(page);
3512
3513         count_vm_numa_event(NUMA_HINT_FAULTS);
3514         if (page_nid == numa_node_id()) {
3515                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3516                 *flags |= TNF_FAULT_LOCAL;
3517         }
3518
3519         return mpol_misplaced(page, vma, addr);
3520 }
3521
3522 static int do_numa_page(struct vm_fault *vmf)
3523 {
3524         struct vm_area_struct *vma = vmf->vma;
3525         struct page *page = NULL;
3526         int page_nid = -1;
3527         int last_cpupid;
3528         int target_nid;
3529         bool migrated = false;
3530         pte_t pte;
3531         bool was_writable = pte_savedwrite(vmf->orig_pte);
3532         int flags = 0;
3533
3534         /*
3535          * The "pte" at this point cannot be used safely without
3536          * validation through pte_unmap_same(). It's of NUMA type but
3537          * the pfn may be screwed if the read is non atomic.
3538          */
3539         vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3540         spin_lock(vmf->ptl);
3541         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3542                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3543                 goto out;
3544         }
3545
3546         /*
3547          * Make it present again, Depending on how arch implementes non
3548          * accessible ptes, some can allow access by kernel mode.
3549          */
3550         pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3551         pte = pte_modify(pte, vma->vm_page_prot);
3552         pte = pte_mkyoung(pte);
3553         if (was_writable)
3554                 pte = pte_mkwrite(pte);
3555         ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3556         update_mmu_cache(vma, vmf->address, vmf->pte);
3557
3558         page = vm_normal_page(vma, vmf->address, pte);
3559         if (!page) {
3560                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3561                 return 0;
3562         }
3563
3564         /* TODO: handle PTE-mapped THP */
3565         if (PageCompound(page)) {
3566                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3567                 return 0;
3568         }
3569
3570         /*
3571          * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3572          * much anyway since they can be in shared cache state. This misses
3573          * the case where a mapping is writable but the process never writes
3574          * to it but pte_write gets cleared during protection updates and
3575          * pte_dirty has unpredictable behaviour between PTE scan updates,
3576          * background writeback, dirty balancing and application behaviour.
3577          */
3578         if (!pte_write(pte))
3579                 flags |= TNF_NO_GROUP;
3580
3581         /*
3582          * Flag if the page is shared between multiple address spaces. This
3583          * is later used when determining whether to group tasks together
3584          */
3585         if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3586                 flags |= TNF_SHARED;
3587
3588         last_cpupid = page_cpupid_last(page);
3589         page_nid = page_to_nid(page);
3590         target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3591                         &flags);
3592         pte_unmap_unlock(vmf->pte, vmf->ptl);
3593         if (target_nid == -1) {
3594                 put_page(page);
3595                 goto out;
3596         }
3597
3598         /* Migrate to the requested node */
3599         migrated = migrate_misplaced_page(page, vma, target_nid);
3600         if (migrated) {
3601                 page_nid = target_nid;
3602                 flags |= TNF_MIGRATED;
3603         } else
3604                 flags |= TNF_MIGRATE_FAIL;
3605
3606 out:
3607         if (page_nid != -1)
3608                 task_numa_fault(last_cpupid, page_nid, 1, flags);
3609         return 0;
3610 }
3611
3612 static int create_huge_pmd(struct vm_fault *vmf)
3613 {
3614         if (vma_is_anonymous(vmf->vma))
3615                 return do_huge_pmd_anonymous_page(vmf);
3616         if (vmf->vma->vm_ops->huge_fault)
3617                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3618         return VM_FAULT_FALLBACK;
3619 }
3620
3621 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3622 {
3623         if (vma_is_anonymous(vmf->vma))
3624                 return do_huge_pmd_wp_page(vmf, orig_pmd);
3625         if (vmf->vma->vm_ops->huge_fault)
3626                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3627
3628         /* COW handled on pte level: split pmd */
3629         VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3630         __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3631
3632         return VM_FAULT_FALLBACK;
3633 }
3634
3635 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3636 {
3637         return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3638 }
3639
3640 static int create_huge_pud(struct vm_fault *vmf)
3641 {
3642 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3643         /* No support for anonymous transparent PUD pages yet */
3644         if (vma_is_anonymous(vmf->vma))
3645                 return VM_FAULT_FALLBACK;
3646         if (vmf->vma->vm_ops->huge_fault)
3647                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3648 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3649         return VM_FAULT_FALLBACK;
3650 }
3651
3652 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3653 {
3654 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3655         /* No support for anonymous transparent PUD pages yet */
3656         if (vma_is_anonymous(vmf->vma))
3657                 return VM_FAULT_FALLBACK;
3658         if (vmf->vma->vm_ops->huge_fault)
3659                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3660 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3661         return VM_FAULT_FALLBACK;
3662 }
3663
3664 /*
3665  * These routines also need to handle stuff like marking pages dirty
3666  * and/or accessed for architectures that don't do it in hardware (most
3667  * RISC architectures).  The early dirtying is also good on the i386.
3668  *
3669  * There is also a hook called "update_mmu_cache()" that architectures
3670  * with external mmu caches can use to update those (ie the Sparc or
3671  * PowerPC hashed page tables that act as extended TLBs).
3672  *
3673  * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3674  * concurrent faults).
3675  *
3676  * The mmap_sem may have been released depending on flags and our return value.
3677  * See filemap_fault() and __lock_page_or_retry().
3678  */
3679 static int handle_pte_fault(struct vm_fault *vmf)
3680 {
3681         pte_t entry;
3682
3683         if (unlikely(pmd_none(*vmf->pmd))) {
3684                 /*
3685                  * Leave __pte_alloc() until later: because vm_ops->fault may
3686                  * want to allocate huge page, and if we expose page table
3687                  * for an instant, it will be difficult to retract from
3688                  * concurrent faults and from rmap lookups.
3689                  */
3690                 vmf->pte = NULL;
3691         } else {
3692                 /* See comment in pte_alloc_one_map() */
3693                 if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
3694                         return 0;
3695                 /*
3696                  * A regular pmd is established and it can't morph into a huge
3697                  * pmd from under us anymore at this point because we hold the
3698                  * mmap_sem read mode and khugepaged takes it in write mode.
3699                  * So now it's safe to run pte_offset_map().
3700                  */
3701                 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3702                 vmf->orig_pte = *vmf->pte;
3703
3704                 /*
3705                  * some architectures can have larger ptes than wordsize,
3706                  * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3707                  * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3708                  * atomic accesses.  The code below just needs a consistent
3709                  * view for the ifs and we later double check anyway with the
3710                  * ptl lock held. So here a barrier will do.
3711                  */
3712                 barrier();
3713                 if (pte_none(vmf->orig_pte)) {
3714                         pte_unmap(vmf->pte);
3715                         vmf->pte = NULL;
3716                 }
3717         }
3718
3719         if (!vmf->pte) {
3720                 if (vma_is_anonymous(vmf->vma))
3721                         return do_anonymous_page(vmf);
3722                 else
3723                         return do_fault(vmf);
3724         }
3725
3726         if (!pte_present(vmf->orig_pte))
3727                 return do_swap_page(vmf);
3728
3729         if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3730                 return do_numa_page(vmf);
3731
3732         vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3733         spin_lock(vmf->ptl);
3734         entry = vmf->orig_pte;
3735         if (unlikely(!pte_same(*vmf->pte, entry)))
3736                 goto unlock;
3737         if (vmf->flags & FAULT_FLAG_WRITE) {
3738                 if (!pte_write(entry))
3739                         return do_wp_page(vmf);
3740                 entry = pte_mkdirty(entry);
3741         }
3742         entry = pte_mkyoung(entry);
3743         if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3744                                 vmf->flags & FAULT_FLAG_WRITE)) {
3745                 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3746         } else {
3747                 /*
3748                  * This is needed only for protection faults but the arch code
3749                  * is not yet telling us if this is a protection fault or not.
3750                  * This still avoids useless tlb flushes for .text page faults
3751                  * with threads.
3752                  */
3753                 if (vmf->flags & FAULT_FLAG_WRITE)
3754                         flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3755         }
3756 unlock:
3757         pte_unmap_unlock(vmf->pte, vmf->ptl);
3758         return 0;
3759 }
3760
3761 /*
3762  * By the time we get here, we already hold the mm semaphore
3763  *
3764  * The mmap_sem may have been released depending on flags and our
3765  * return value.  See filemap_fault() and __lock_page_or_retry().
3766  */
3767 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3768                 unsigned int flags)
3769 {
3770         struct vm_fault vmf = {
3771                 .vma = vma,
3772                 .address = address & PAGE_MASK,
3773                 .flags = flags,
3774                 .pgoff = linear_page_index(vma, address),
3775                 .gfp_mask = __get_fault_gfp_mask(vma),
3776         };
3777         struct mm_struct *mm = vma->vm_mm;
3778         pgd_t *pgd;
3779         p4d_t *p4d;
3780         int ret;
3781
3782         pgd = pgd_offset(mm, address);
3783         p4d = p4d_alloc(mm, pgd, address);
3784         if (!p4d)
3785                 return VM_FAULT_OOM;
3786
3787         vmf.pud = pud_alloc(mm, p4d, address);
3788         if (!vmf.pud)
3789                 return VM_FAULT_OOM;
3790         if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
3791                 ret = create_huge_pud(&vmf);
3792                 if (!(ret & VM_FAULT_FALLBACK))
3793                         return ret;
3794         } else {
3795                 pud_t orig_pud = *vmf.pud;
3796
3797                 barrier();
3798                 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3799                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3800
3801                         /* NUMA case for anonymous PUDs would go here */
3802
3803                         if (dirty && !pud_write(orig_pud)) {
3804                                 ret = wp_huge_pud(&vmf, orig_pud);
3805                                 if (!(ret & VM_FAULT_FALLBACK))
3806                                         return ret;
3807                         } else {
3808                                 huge_pud_set_accessed(&vmf, orig_pud);
3809                                 return 0;
3810                         }
3811                 }
3812         }
3813
3814         vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3815         if (!vmf.pmd)
3816                 return VM_FAULT_OOM;
3817         if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3818                 ret = create_huge_pmd(&vmf);
3819                 if (!(ret & VM_FAULT_FALLBACK))
3820                         return ret;
3821         } else {
3822                 pmd_t orig_pmd = *vmf.pmd;
3823
3824                 barrier();
3825                 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3826                         if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3827                                 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3828
3829                         if ((vmf.flags & FAULT_FLAG_WRITE) &&
3830                                         !pmd_write(orig_pmd)) {
3831                                 ret = wp_huge_pmd(&vmf, orig_pmd);
3832                                 if (!(ret & VM_FAULT_FALLBACK))
3833                                         return ret;
3834                         } else {
3835                                 huge_pmd_set_accessed(&vmf, orig_pmd);
3836                                 return 0;
3837                         }
3838                 }
3839         }
3840
3841         return handle_pte_fault(&vmf);
3842 }
3843
3844 /*
3845  * By the time we get here, we already hold the mm semaphore
3846  *
3847  * The mmap_sem may have been released depending on flags and our
3848  * return value.  See filemap_fault() and __lock_page_or_retry().
3849  */
3850 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3851                 unsigned int flags)
3852 {
3853         int ret;
3854
3855         __set_current_state(TASK_RUNNING);
3856
3857         count_vm_event(PGFAULT);
3858         mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
3859
3860         /* do counter updates before entering really critical section. */
3861         check_sync_rss_stat(current);
3862
3863         /*
3864          * Enable the memcg OOM handling for faults triggered in user
3865          * space.  Kernel faults are handled more gracefully.
3866          */
3867         if (flags & FAULT_FLAG_USER)
3868                 mem_cgroup_oom_enable();
3869
3870         if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3871                                             flags & FAULT_FLAG_INSTRUCTION,
3872                                             flags & FAULT_FLAG_REMOTE))
3873                 return VM_FAULT_SIGSEGV;
3874
3875         if (unlikely(is_vm_hugetlb_page(vma)))
3876                 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3877         else
3878                 ret = __handle_mm_fault(vma, address, flags);
3879
3880         if (flags & FAULT_FLAG_USER) {
3881                 mem_cgroup_oom_disable();
3882                 /*
3883                  * The task may have entered a memcg OOM situation but
3884                  * if the allocation error was handled gracefully (no
3885                  * VM_FAULT_OOM), there is no need to kill anything.
3886                  * Just clean up the OOM state peacefully.
3887                  */
3888                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3889                         mem_cgroup_oom_synchronize(false);
3890         }
3891
3892         /*
3893          * This mm has been already reaped by the oom reaper and so the
3894          * refault cannot be trusted in general. Anonymous refaults would
3895          * lose data and give a zero page instead e.g. This is especially
3896          * problem for use_mm() because regular tasks will just die and
3897          * the corrupted data will not be visible anywhere while kthread
3898          * will outlive the oom victim and potentially propagate the date
3899          * further.
3900          */
3901         if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3902                                 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3903                 ret = VM_FAULT_SIGBUS;
3904
3905         return ret;
3906 }
3907 EXPORT_SYMBOL_GPL(handle_mm_fault);
3908
3909 #ifndef __PAGETABLE_P4D_FOLDED
3910 /*
3911  * Allocate p4d page table.
3912  * We've already handled the fast-path in-line.
3913  */
3914 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3915 {
3916         p4d_t *new = p4d_alloc_one(mm, address);
3917         if (!new)
3918                 return -ENOMEM;
3919
3920         smp_wmb(); /* See comment in __pte_alloc */
3921
3922         spin_lock(&mm->page_table_lock);
3923         if (pgd_present(*pgd))          /* Another has populated it */
3924                 p4d_free(mm, new);
3925         else
3926                 pgd_populate(mm, pgd, new);
3927         spin_unlock(&mm->page_table_lock);
3928         return 0;
3929 }
3930 #endif /* __PAGETABLE_P4D_FOLDED */
3931
3932 #ifndef __PAGETABLE_PUD_FOLDED
3933 /*
3934  * Allocate page upper directory.
3935  * We've already handled the fast-path in-line.
3936  */
3937 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
3938 {
3939         pud_t *new = pud_alloc_one(mm, address);
3940         if (!new)
3941                 return -ENOMEM;
3942
3943         smp_wmb(); /* See comment in __pte_alloc */
3944
3945         spin_lock(&mm->page_table_lock);
3946 #ifndef __ARCH_HAS_5LEVEL_HACK
3947         if (p4d_present(*p4d))          /* Another has populated it */
3948                 pud_free(mm, new);
3949         else
3950                 p4d_populate(mm, p4d, new);
3951 #else
3952         if (pgd_present(*p4d))          /* Another has populated it */
3953                 pud_free(mm, new);
3954         else
3955                 pgd_populate(mm, p4d, new);
3956 #endif /* __ARCH_HAS_5LEVEL_HACK */
3957         spin_unlock(&mm->page_table_lock);
3958         return 0;
3959 }
3960 #endif /* __PAGETABLE_PUD_FOLDED */
3961
3962 #ifndef __PAGETABLE_PMD_FOLDED
3963 /*
3964  * Allocate page middle directory.
3965  * We've already handled the fast-path in-line.
3966  */
3967 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3968 {
3969         spinlock_t *ptl;
3970         pmd_t *new = pmd_alloc_one(mm, address);
3971         if (!new)
3972                 return -ENOMEM;
3973
3974         smp_wmb(); /* See comment in __pte_alloc */
3975
3976         ptl = pud_lock(mm, pud);
3977 #ifndef __ARCH_HAS_4LEVEL_HACK
3978         if (!pud_present(*pud)) {
3979                 mm_inc_nr_pmds(mm);
3980                 pud_populate(mm, pud, new);
3981         } else  /* Another has populated it */
3982                 pmd_free(mm, new);
3983 #else
3984         if (!pgd_present(*pud)) {
3985                 mm_inc_nr_pmds(mm);
3986                 pgd_populate(mm, pud, new);
3987         } else /* Another has populated it */
3988                 pmd_free(mm, new);
3989 #endif /* __ARCH_HAS_4LEVEL_HACK */
3990         spin_unlock(ptl);
3991         return 0;
3992 }
3993 #endif /* __PAGETABLE_PMD_FOLDED */
3994
3995 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3996                 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3997 {
3998         pgd_t *pgd;
3999         p4d_t *p4d;
4000         pud_t *pud;
4001         pmd_t *pmd;
4002         pte_t *ptep;
4003
4004         pgd = pgd_offset(mm, address);
4005         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4006                 goto out;
4007
4008         p4d = p4d_offset(pgd, address);
4009         if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4010                 goto out;
4011
4012         pud = pud_offset(p4d, address);
4013         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4014                 goto out;
4015
4016         pmd = pmd_offset(pud, address);
4017         VM_BUG_ON(pmd_trans_huge(*pmd));
4018
4019         if (pmd_huge(*pmd)) {
4020                 if (!pmdpp)
4021                         goto out;
4022
4023                 *ptlp = pmd_lock(mm, pmd);
4024                 if (pmd_huge(*pmd)) {
4025                         *pmdpp = pmd;
4026                         return 0;
4027                 }
4028                 spin_unlock(*ptlp);
4029         }
4030
4031         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4032                 goto out;
4033
4034         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4035         if (!ptep)
4036                 goto out;
4037         if (!pte_present(*ptep))
4038                 goto unlock;
4039         *ptepp = ptep;
4040         return 0;
4041 unlock:
4042         pte_unmap_unlock(ptep, *ptlp);
4043 out:
4044         return -EINVAL;
4045 }
4046
4047 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4048                              pte_t **ptepp, spinlock_t **ptlp)
4049 {
4050         int res;
4051
4052         /* (void) is needed to make gcc happy */
4053         (void) __cond_lock(*ptlp,
4054                            !(res = __follow_pte_pmd(mm, address, ptepp, NULL,
4055                                            ptlp)));
4056         return res;
4057 }
4058
4059 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4060                              pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4061 {
4062         int res;
4063
4064         /* (void) is needed to make gcc happy */
4065         (void) __cond_lock(*ptlp,
4066                            !(res = __follow_pte_pmd(mm, address, ptepp, pmdpp,
4067                                            ptlp)));
4068         return res;
4069 }
4070 EXPORT_SYMBOL(follow_pte_pmd);
4071
4072 /**
4073  * follow_pfn - look up PFN at a user virtual address
4074  * @vma: memory mapping
4075  * @address: user virtual address
4076  * @pfn: location to store found PFN
4077  *
4078  * Only IO mappings and raw PFN mappings are allowed.
4079  *
4080  * Returns zero and the pfn at @pfn on success, -ve otherwise.
4081  */
4082 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4083         unsigned long *pfn)
4084 {
4085         int ret = -EINVAL;
4086         spinlock_t *ptl;
4087         pte_t *ptep;
4088
4089         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4090                 return ret;
4091
4092         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4093         if (ret)
4094                 return ret;
4095         *pfn = pte_pfn(*ptep);
4096         pte_unmap_unlock(ptep, ptl);
4097         return 0;
4098 }
4099 EXPORT_SYMBOL(follow_pfn);
4100
4101 #ifdef CONFIG_HAVE_IOREMAP_PROT
4102 int follow_phys(struct vm_area_struct *vma,
4103                 unsigned long address, unsigned int flags,
4104                 unsigned long *prot, resource_size_t *phys)
4105 {
4106         int ret = -EINVAL;
4107         pte_t *ptep, pte;
4108         spinlock_t *ptl;
4109
4110         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4111                 goto out;
4112
4113         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4114                 goto out;
4115         pte = *ptep;
4116
4117         if ((flags & FOLL_WRITE) && !pte_write(pte))
4118                 goto unlock;
4119
4120         *prot = pgprot_val(pte_pgprot(pte));
4121         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4122
4123         ret = 0;
4124 unlock:
4125         pte_unmap_unlock(ptep, ptl);
4126 out:
4127         return ret;
4128 }
4129
4130 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4131                         void *buf, int len, int write)
4132 {
4133         resource_size_t phys_addr;
4134         unsigned long prot = 0;
4135         void __iomem *maddr;
4136         int offset = addr & (PAGE_SIZE-1);
4137
4138         if (follow_phys(vma, addr, write, &prot, &phys_addr))
4139                 return -EINVAL;
4140
4141         maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4142         if (write)
4143                 memcpy_toio(maddr + offset, buf, len);
4144         else
4145                 memcpy_fromio(buf, maddr + offset, len);
4146         iounmap(maddr);
4147
4148         return len;
4149 }
4150 EXPORT_SYMBOL_GPL(generic_access_phys);
4151 #endif
4152
4153 /*
4154  * Access another process' address space as given in mm.  If non-NULL, use the
4155  * given task for page fault accounting.
4156  */
4157 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4158                 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4159 {
4160         struct vm_area_struct *vma;
4161         void *old_buf = buf;
4162         int write = gup_flags & FOLL_WRITE;
4163
4164         down_read(&mm->mmap_sem);
4165         /* ignore errors, just check how much was successfully transferred */
4166         while (len) {
4167                 int bytes, ret, offset;
4168                 void *maddr;
4169                 struct page *page = NULL;
4170
4171                 ret = get_user_pages_remote(tsk, mm, addr, 1,
4172                                 gup_flags, &page, &vma, NULL);
4173                 if (ret <= 0) {
4174 #ifndef CONFIG_HAVE_IOREMAP_PROT
4175                         break;
4176 #else
4177                         /*
4178                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
4179                          * we can access using slightly different code.
4180                          */
4181                         vma = find_vma(mm, addr);
4182                         if (!vma || vma->vm_start > addr)
4183                                 break;
4184                         if (vma->vm_ops && vma->vm_ops->access)
4185                                 ret = vma->vm_ops->access(vma, addr, buf,
4186                                                           len, write);
4187                         if (ret <= 0)
4188                                 break;
4189                         bytes = ret;
4190 #endif
4191                 } else {
4192                         bytes = len;
4193                         offset = addr & (PAGE_SIZE-1);
4194                         if (bytes > PAGE_SIZE-offset)
4195                                 bytes = PAGE_SIZE-offset;
4196
4197                         maddr = kmap(page);
4198                         if (write) {
4199                                 copy_to_user_page(vma, page, addr,
4200                                                   maddr + offset, buf, bytes);
4201                                 set_page_dirty_lock(page);
4202                         } else {
4203                                 copy_from_user_page(vma, page, addr,
4204                                                     buf, maddr + offset, bytes);
4205                         }
4206                         kunmap(page);
4207                         put_page(page);
4208                 }
4209                 len -= bytes;
4210                 buf += bytes;
4211                 addr += bytes;
4212         }
4213         up_read(&mm->mmap_sem);
4214
4215         return buf - old_buf;
4216 }
4217
4218 /**
4219  * access_remote_vm - access another process' address space
4220  * @mm:         the mm_struct of the target address space
4221  * @addr:       start address to access
4222  * @buf:        source or destination buffer
4223  * @len:        number of bytes to transfer
4224  * @gup_flags:  flags modifying lookup behaviour
4225  *
4226  * The caller must hold a reference on @mm.
4227  */
4228 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4229                 void *buf, int len, unsigned int gup_flags)
4230 {
4231         return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4232 }
4233
4234 /*
4235  * Access another process' address space.
4236  * Source/target buffer must be kernel space,
4237  * Do not walk the page table directly, use get_user_pages
4238  */
4239 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4240                 void *buf, int len, unsigned int gup_flags)
4241 {
4242         struct mm_struct *mm;
4243         int ret;
4244
4245         mm = get_task_mm(tsk);
4246         if (!mm)
4247                 return 0;
4248
4249         ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4250
4251         mmput(mm);
4252
4253         return ret;
4254 }
4255 EXPORT_SYMBOL_GPL(access_process_vm);
4256
4257 /*
4258  * Print the name of a VMA.
4259  */
4260 void print_vma_addr(char *prefix, unsigned long ip)
4261 {
4262         struct mm_struct *mm = current->mm;
4263         struct vm_area_struct *vma;
4264
4265         /*
4266          * Do not print if we are in atomic
4267          * contexts (in exception stacks, etc.):
4268          */
4269         if (preempt_count())
4270                 return;
4271
4272         down_read(&mm->mmap_sem);
4273         vma = find_vma(mm, ip);
4274         if (vma && vma->vm_file) {
4275                 struct file *f = vma->vm_file;
4276                 char *buf = (char *)__get_free_page(GFP_KERNEL);
4277                 if (buf) {
4278                         char *p;
4279
4280                         p = file_path(f, buf, PAGE_SIZE);
4281                         if (IS_ERR(p))
4282                                 p = "?";
4283                         printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4284                                         vma->vm_start,
4285                                         vma->vm_end - vma->vm_start);
4286                         free_page((unsigned long)buf);
4287                 }
4288         }
4289         up_read(&mm->mmap_sem);
4290 }
4291
4292 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4293 void __might_fault(const char *file, int line)
4294 {
4295         /*
4296          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4297          * holding the mmap_sem, this is safe because kernel memory doesn't
4298          * get paged out, therefore we'll never actually fault, and the
4299          * below annotations will generate false positives.
4300          */
4301         if (uaccess_kernel())
4302                 return;
4303         if (pagefault_disabled())
4304                 return;
4305         __might_sleep(file, line, 0);
4306 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4307         if (current->mm)
4308                 might_lock_read(&current->mm->mmap_sem);
4309 #endif
4310 }
4311 EXPORT_SYMBOL(__might_fault);
4312 #endif
4313
4314 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4315 static void clear_gigantic_page(struct page *page,
4316                                 unsigned long addr,
4317                                 unsigned int pages_per_huge_page)
4318 {
4319         int i;
4320         struct page *p = page;
4321
4322         might_sleep();
4323         for (i = 0; i < pages_per_huge_page;
4324              i++, p = mem_map_next(p, page, i)) {
4325                 cond_resched();
4326                 clear_user_highpage(p, addr + i * PAGE_SIZE);
4327         }
4328 }
4329 void clear_huge_page(struct page *page,
4330                      unsigned long addr, unsigned int pages_per_huge_page)
4331 {
4332         int i;
4333
4334         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4335                 clear_gigantic_page(page, addr, pages_per_huge_page);
4336                 return;
4337         }
4338
4339         might_sleep();
4340         for (i = 0; i < pages_per_huge_page; i++) {
4341                 cond_resched();
4342                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4343         }
4344 }
4345
4346 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4347                                     unsigned long addr,
4348                                     struct vm_area_struct *vma,
4349                                     unsigned int pages_per_huge_page)
4350 {
4351         int i;
4352         struct page *dst_base = dst;
4353         struct page *src_base = src;
4354
4355         for (i = 0; i < pages_per_huge_page; ) {
4356                 cond_resched();
4357                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4358
4359                 i++;
4360                 dst = mem_map_next(dst, dst_base, i);
4361                 src = mem_map_next(src, src_base, i);
4362         }
4363 }
4364
4365 void copy_user_huge_page(struct page *dst, struct page *src,
4366                          unsigned long addr, struct vm_area_struct *vma,
4367                          unsigned int pages_per_huge_page)
4368 {
4369         int i;
4370
4371         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4372                 copy_user_gigantic_page(dst, src, addr, vma,
4373                                         pages_per_huge_page);
4374                 return;
4375         }
4376
4377         might_sleep();
4378         for (i = 0; i < pages_per_huge_page; i++) {
4379                 cond_resched();
4380                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4381         }
4382 }
4383
4384 long copy_huge_page_from_user(struct page *dst_page,
4385                                 const void __user *usr_src,
4386                                 unsigned int pages_per_huge_page,
4387                                 bool allow_pagefault)
4388 {
4389         void *src = (void *)usr_src;
4390         void *page_kaddr;
4391         unsigned long i, rc = 0;
4392         unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4393
4394         for (i = 0; i < pages_per_huge_page; i++) {
4395                 if (allow_pagefault)
4396                         page_kaddr = kmap(dst_page + i);
4397                 else
4398                         page_kaddr = kmap_atomic(dst_page + i);
4399                 rc = copy_from_user(page_kaddr,
4400                                 (const void __user *)(src + i * PAGE_SIZE),
4401                                 PAGE_SIZE);
4402                 if (allow_pagefault)
4403                         kunmap(dst_page + i);
4404                 else
4405                         kunmap_atomic(page_kaddr);
4406
4407                 ret_val -= (PAGE_SIZE - rc);
4408                 if (rc)
4409                         break;
4410
4411                 cond_resched();
4412         }
4413         return ret_val;
4414 }
4415 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4416
4417 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4418
4419 static struct kmem_cache *page_ptl_cachep;
4420
4421 void __init ptlock_cache_init(void)
4422 {
4423         page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4424                         SLAB_PANIC, NULL);
4425 }
4426
4427 bool ptlock_alloc(struct page *page)
4428 {
4429         spinlock_t *ptl;
4430
4431         ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4432         if (!ptl)
4433                 return false;
4434         page->ptl = ptl;
4435         return true;
4436 }
4437
4438 void ptlock_free(struct page *page)
4439 {
4440         kmem_cache_free(page_ptl_cachep, page->ptl);
4441 }
4442 #endif