mm: swap: clean up swap readahead
[muen/linux.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/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.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 #include <linux/oom.h>
72
73 #include <asm/io.h>
74 #include <asm/mmu_context.h>
75 #include <asm/pgalloc.h>
76 #include <linux/uaccess.h>
77 #include <asm/tlb.h>
78 #include <asm/tlbflush.h>
79 #include <asm/pgtable.h>
80
81 #include "internal.h"
82
83 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
84 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
85 #endif
86
87 #ifndef CONFIG_NEED_MULTIPLE_NODES
88 /* use the per-pgdat data instead for discontigmem - mbligh */
89 unsigned long max_mapnr;
90 EXPORT_SYMBOL(max_mapnr);
91
92 struct page *mem_map;
93 EXPORT_SYMBOL(mem_map);
94 #endif
95
96 /*
97  * A number of key systems in x86 including ioremap() rely on the assumption
98  * that high_memory defines the upper bound on direct map memory, then end
99  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
100  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
101  * and ZONE_HIGHMEM.
102  */
103 void *high_memory;
104 EXPORT_SYMBOL(high_memory);
105
106 /*
107  * Randomize the address space (stacks, mmaps, brk, etc.).
108  *
109  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110  *   as ancient (libc5 based) binaries can segfault. )
111  */
112 int randomize_va_space __read_mostly =
113 #ifdef CONFIG_COMPAT_BRK
114                                         1;
115 #else
116                                         2;
117 #endif
118
119 static int __init disable_randmaps(char *s)
120 {
121         randomize_va_space = 0;
122         return 1;
123 }
124 __setup("norandmaps", disable_randmaps);
125
126 unsigned long zero_pfn __read_mostly;
127 EXPORT_SYMBOL(zero_pfn);
128
129 unsigned long highest_memmap_pfn __read_mostly;
130
131 /*
132  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
133  */
134 static int __init init_zero_pfn(void)
135 {
136         zero_pfn = page_to_pfn(ZERO_PAGE(0));
137         return 0;
138 }
139 core_initcall(init_zero_pfn);
140
141
142 #if defined(SPLIT_RSS_COUNTING)
143
144 void sync_mm_rss(struct mm_struct *mm)
145 {
146         int i;
147
148         for (i = 0; i < NR_MM_COUNTERS; i++) {
149                 if (current->rss_stat.count[i]) {
150                         add_mm_counter(mm, i, current->rss_stat.count[i]);
151                         current->rss_stat.count[i] = 0;
152                 }
153         }
154         current->rss_stat.events = 0;
155 }
156
157 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
158 {
159         struct task_struct *task = current;
160
161         if (likely(task->mm == mm))
162                 task->rss_stat.count[member] += val;
163         else
164                 add_mm_counter(mm, member, val);
165 }
166 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
167 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
168
169 /* sync counter once per 64 page faults */
170 #define TASK_RSS_EVENTS_THRESH  (64)
171 static void check_sync_rss_stat(struct task_struct *task)
172 {
173         if (unlikely(task != current))
174                 return;
175         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
176                 sync_mm_rss(task->mm);
177 }
178 #else /* SPLIT_RSS_COUNTING */
179
180 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
181 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
182
183 static void check_sync_rss_stat(struct task_struct *task)
184 {
185 }
186
187 #endif /* SPLIT_RSS_COUNTING */
188
189 #ifdef HAVE_GENERIC_MMU_GATHER
190
191 static bool tlb_next_batch(struct mmu_gather *tlb)
192 {
193         struct mmu_gather_batch *batch;
194
195         batch = tlb->active;
196         if (batch->next) {
197                 tlb->active = batch->next;
198                 return true;
199         }
200
201         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
202                 return false;
203
204         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
205         if (!batch)
206                 return false;
207
208         tlb->batch_count++;
209         batch->next = NULL;
210         batch->nr   = 0;
211         batch->max  = MAX_GATHER_BATCH;
212
213         tlb->active->next = batch;
214         tlb->active = batch;
215
216         return true;
217 }
218
219 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
220                                 unsigned long start, unsigned long end)
221 {
222         tlb->mm = mm;
223
224         /* Is it from 0 to ~0? */
225         tlb->fullmm     = !(start | (end+1));
226         tlb->need_flush_all = 0;
227         tlb->local.next = NULL;
228         tlb->local.nr   = 0;
229         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
230         tlb->active     = &tlb->local;
231         tlb->batch_count = 0;
232
233 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
234         tlb->batch = NULL;
235 #endif
236         tlb->page_size = 0;
237
238         __tlb_reset_range(tlb);
239 }
240
241 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
242 {
243         if (!tlb->end)
244                 return;
245
246         tlb_flush(tlb);
247         mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
248 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
249         tlb_table_flush(tlb);
250 #endif
251         __tlb_reset_range(tlb);
252 }
253
254 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
255 {
256         struct mmu_gather_batch *batch;
257
258         for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
259                 free_pages_and_swap_cache(batch->pages, batch->nr);
260                 batch->nr = 0;
261         }
262         tlb->active = &tlb->local;
263 }
264
265 void tlb_flush_mmu(struct mmu_gather *tlb)
266 {
267         tlb_flush_mmu_tlbonly(tlb);
268         tlb_flush_mmu_free(tlb);
269 }
270
271 /* tlb_finish_mmu
272  *      Called at the end of the shootdown operation to free up any resources
273  *      that were required.
274  */
275 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
276                 unsigned long start, unsigned long end, bool force)
277 {
278         struct mmu_gather_batch *batch, *next;
279
280         if (force)
281                 __tlb_adjust_range(tlb, start, end - start);
282
283         tlb_flush_mmu(tlb);
284
285         /* keep the page table cache within bounds */
286         check_pgt_cache();
287
288         for (batch = tlb->local.next; batch; batch = next) {
289                 next = batch->next;
290                 free_pages((unsigned long)batch, 0);
291         }
292         tlb->local.next = NULL;
293 }
294
295 /* __tlb_remove_page
296  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
297  *      handling the additional races in SMP caused by other CPUs caching valid
298  *      mappings in their TLBs. Returns the number of free page slots left.
299  *      When out of page slots we must call tlb_flush_mmu().
300  *returns true if the caller should flush.
301  */
302 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
303 {
304         struct mmu_gather_batch *batch;
305
306         VM_BUG_ON(!tlb->end);
307         VM_WARN_ON(tlb->page_size != page_size);
308
309         batch = tlb->active;
310         /*
311          * Add the page and check if we are full. If so
312          * force a flush.
313          */
314         batch->pages[batch->nr++] = page;
315         if (batch->nr == batch->max) {
316                 if (!tlb_next_batch(tlb))
317                         return true;
318                 batch = tlb->active;
319         }
320         VM_BUG_ON_PAGE(batch->nr > batch->max, page);
321
322         return false;
323 }
324
325 #endif /* HAVE_GENERIC_MMU_GATHER */
326
327 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
328
329 /*
330  * See the comment near struct mmu_table_batch.
331  */
332
333 static void tlb_remove_table_smp_sync(void *arg)
334 {
335         /* Simply deliver the interrupt */
336 }
337
338 static void tlb_remove_table_one(void *table)
339 {
340         /*
341          * This isn't an RCU grace period and hence the page-tables cannot be
342          * assumed to be actually RCU-freed.
343          *
344          * It is however sufficient for software page-table walkers that rely on
345          * IRQ disabling. See the comment near struct mmu_table_batch.
346          */
347         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
348         __tlb_remove_table(table);
349 }
350
351 static void tlb_remove_table_rcu(struct rcu_head *head)
352 {
353         struct mmu_table_batch *batch;
354         int i;
355
356         batch = container_of(head, struct mmu_table_batch, rcu);
357
358         for (i = 0; i < batch->nr; i++)
359                 __tlb_remove_table(batch->tables[i]);
360
361         free_page((unsigned long)batch);
362 }
363
364 void tlb_table_flush(struct mmu_gather *tlb)
365 {
366         struct mmu_table_batch **batch = &tlb->batch;
367
368         if (*batch) {
369                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
370                 *batch = NULL;
371         }
372 }
373
374 void tlb_remove_table(struct mmu_gather *tlb, void *table)
375 {
376         struct mmu_table_batch **batch = &tlb->batch;
377
378         /*
379          * When there's less then two users of this mm there cannot be a
380          * concurrent page-table walk.
381          */
382         if (atomic_read(&tlb->mm->mm_users) < 2) {
383                 __tlb_remove_table(table);
384                 return;
385         }
386
387         if (*batch == NULL) {
388                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
389                 if (*batch == NULL) {
390                         tlb_remove_table_one(table);
391                         return;
392                 }
393                 (*batch)->nr = 0;
394         }
395         (*batch)->tables[(*batch)->nr++] = table;
396         if ((*batch)->nr == MAX_TABLE_BATCH)
397                 tlb_table_flush(tlb);
398 }
399
400 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
401
402 /**
403  * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
404  * @tlb: the mmu_gather structure to initialize
405  * @mm: the mm_struct of the target address space
406  * @start: start of the region that will be removed from the page-table
407  * @end: end of the region that will be removed from the page-table
408  *
409  * Called to initialize an (on-stack) mmu_gather structure for page-table
410  * tear-down from @mm. The @start and @end are set to 0 and -1
411  * respectively when @mm is without users and we're going to destroy
412  * the full address space (exit/execve).
413  */
414 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
415                         unsigned long start, unsigned long end)
416 {
417         arch_tlb_gather_mmu(tlb, mm, start, end);
418         inc_tlb_flush_pending(tlb->mm);
419 }
420
421 void tlb_finish_mmu(struct mmu_gather *tlb,
422                 unsigned long start, unsigned long end)
423 {
424         /*
425          * If there are parallel threads are doing PTE changes on same range
426          * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
427          * flush by batching, a thread has stable TLB entry can fail to flush
428          * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
429          * forcefully if we detect parallel PTE batching threads.
430          */
431         bool force = mm_tlb_flush_nested(tlb->mm);
432
433         arch_tlb_finish_mmu(tlb, start, end, force);
434         dec_tlb_flush_pending(tlb->mm);
435 }
436
437 /*
438  * Note: this doesn't free the actual pages themselves. That
439  * has been handled earlier when unmapping all the memory regions.
440  */
441 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
442                            unsigned long addr)
443 {
444         pgtable_t token = pmd_pgtable(*pmd);
445         pmd_clear(pmd);
446         pte_free_tlb(tlb, token, addr);
447         mm_dec_nr_ptes(tlb->mm);
448 }
449
450 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
451                                 unsigned long addr, unsigned long end,
452                                 unsigned long floor, unsigned long ceiling)
453 {
454         pmd_t *pmd;
455         unsigned long next;
456         unsigned long start;
457
458         start = addr;
459         pmd = pmd_offset(pud, addr);
460         do {
461                 next = pmd_addr_end(addr, end);
462                 if (pmd_none_or_clear_bad(pmd))
463                         continue;
464                 free_pte_range(tlb, pmd, addr);
465         } while (pmd++, addr = next, addr != end);
466
467         start &= PUD_MASK;
468         if (start < floor)
469                 return;
470         if (ceiling) {
471                 ceiling &= PUD_MASK;
472                 if (!ceiling)
473                         return;
474         }
475         if (end - 1 > ceiling - 1)
476                 return;
477
478         pmd = pmd_offset(pud, start);
479         pud_clear(pud);
480         pmd_free_tlb(tlb, pmd, start);
481         mm_dec_nr_pmds(tlb->mm);
482 }
483
484 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
485                                 unsigned long addr, unsigned long end,
486                                 unsigned long floor, unsigned long ceiling)
487 {
488         pud_t *pud;
489         unsigned long next;
490         unsigned long start;
491
492         start = addr;
493         pud = pud_offset(p4d, addr);
494         do {
495                 next = pud_addr_end(addr, end);
496                 if (pud_none_or_clear_bad(pud))
497                         continue;
498                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
499         } while (pud++, addr = next, addr != end);
500
501         start &= P4D_MASK;
502         if (start < floor)
503                 return;
504         if (ceiling) {
505                 ceiling &= P4D_MASK;
506                 if (!ceiling)
507                         return;
508         }
509         if (end - 1 > ceiling - 1)
510                 return;
511
512         pud = pud_offset(p4d, start);
513         p4d_clear(p4d);
514         pud_free_tlb(tlb, pud, start);
515         mm_dec_nr_puds(tlb->mm);
516 }
517
518 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
519                                 unsigned long addr, unsigned long end,
520                                 unsigned long floor, unsigned long ceiling)
521 {
522         p4d_t *p4d;
523         unsigned long next;
524         unsigned long start;
525
526         start = addr;
527         p4d = p4d_offset(pgd, addr);
528         do {
529                 next = p4d_addr_end(addr, end);
530                 if (p4d_none_or_clear_bad(p4d))
531                         continue;
532                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
533         } while (p4d++, addr = next, addr != end);
534
535         start &= PGDIR_MASK;
536         if (start < floor)
537                 return;
538         if (ceiling) {
539                 ceiling &= PGDIR_MASK;
540                 if (!ceiling)
541                         return;
542         }
543         if (end - 1 > ceiling - 1)
544                 return;
545
546         p4d = p4d_offset(pgd, start);
547         pgd_clear(pgd);
548         p4d_free_tlb(tlb, p4d, start);
549 }
550
551 /*
552  * This function frees user-level page tables of a process.
553  */
554 void free_pgd_range(struct mmu_gather *tlb,
555                         unsigned long addr, unsigned long end,
556                         unsigned long floor, unsigned long ceiling)
557 {
558         pgd_t *pgd;
559         unsigned long next;
560
561         /*
562          * The next few lines have given us lots of grief...
563          *
564          * Why are we testing PMD* at this top level?  Because often
565          * there will be no work to do at all, and we'd prefer not to
566          * go all the way down to the bottom just to discover that.
567          *
568          * Why all these "- 1"s?  Because 0 represents both the bottom
569          * of the address space and the top of it (using -1 for the
570          * top wouldn't help much: the masks would do the wrong thing).
571          * The rule is that addr 0 and floor 0 refer to the bottom of
572          * the address space, but end 0 and ceiling 0 refer to the top
573          * Comparisons need to use "end - 1" and "ceiling - 1" (though
574          * that end 0 case should be mythical).
575          *
576          * Wherever addr is brought up or ceiling brought down, we must
577          * be careful to reject "the opposite 0" before it confuses the
578          * subsequent tests.  But what about where end is brought down
579          * by PMD_SIZE below? no, end can't go down to 0 there.
580          *
581          * Whereas we round start (addr) and ceiling down, by different
582          * masks at different levels, in order to test whether a table
583          * now has no other vmas using it, so can be freed, we don't
584          * bother to round floor or end up - the tests don't need that.
585          */
586
587         addr &= PMD_MASK;
588         if (addr < floor) {
589                 addr += PMD_SIZE;
590                 if (!addr)
591                         return;
592         }
593         if (ceiling) {
594                 ceiling &= PMD_MASK;
595                 if (!ceiling)
596                         return;
597         }
598         if (end - 1 > ceiling - 1)
599                 end -= PMD_SIZE;
600         if (addr > end - 1)
601                 return;
602         /*
603          * We add page table cache pages with PAGE_SIZE,
604          * (see pte_free_tlb()), flush the tlb if we need
605          */
606         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
607         pgd = pgd_offset(tlb->mm, addr);
608         do {
609                 next = pgd_addr_end(addr, end);
610                 if (pgd_none_or_clear_bad(pgd))
611                         continue;
612                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
613         } while (pgd++, addr = next, addr != end);
614 }
615
616 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
617                 unsigned long floor, unsigned long ceiling)
618 {
619         while (vma) {
620                 struct vm_area_struct *next = vma->vm_next;
621                 unsigned long addr = vma->vm_start;
622
623                 /*
624                  * Hide vma from rmap and truncate_pagecache before freeing
625                  * pgtables
626                  */
627                 unlink_anon_vmas(vma);
628                 unlink_file_vma(vma);
629
630                 if (is_vm_hugetlb_page(vma)) {
631                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
632                                 floor, next ? next->vm_start : ceiling);
633                 } else {
634                         /*
635                          * Optimization: gather nearby vmas into one call down
636                          */
637                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
638                                && !is_vm_hugetlb_page(next)) {
639                                 vma = next;
640                                 next = vma->vm_next;
641                                 unlink_anon_vmas(vma);
642                                 unlink_file_vma(vma);
643                         }
644                         free_pgd_range(tlb, addr, vma->vm_end,
645                                 floor, next ? next->vm_start : ceiling);
646                 }
647                 vma = next;
648         }
649 }
650
651 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
652 {
653         spinlock_t *ptl;
654         pgtable_t new = pte_alloc_one(mm, address);
655         if (!new)
656                 return -ENOMEM;
657
658         /*
659          * Ensure all pte setup (eg. pte page lock and page clearing) are
660          * visible before the pte is made visible to other CPUs by being
661          * put into page tables.
662          *
663          * The other side of the story is the pointer chasing in the page
664          * table walking code (when walking the page table without locking;
665          * ie. most of the time). Fortunately, these data accesses consist
666          * of a chain of data-dependent loads, meaning most CPUs (alpha
667          * being the notable exception) will already guarantee loads are
668          * seen in-order. See the alpha page table accessors for the
669          * smp_read_barrier_depends() barriers in page table walking code.
670          */
671         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
672
673         ptl = pmd_lock(mm, pmd);
674         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
675                 mm_inc_nr_ptes(mm);
676                 pmd_populate(mm, pmd, new);
677                 new = NULL;
678         }
679         spin_unlock(ptl);
680         if (new)
681                 pte_free(mm, new);
682         return 0;
683 }
684
685 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
686 {
687         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
688         if (!new)
689                 return -ENOMEM;
690
691         smp_wmb(); /* See comment in __pte_alloc */
692
693         spin_lock(&init_mm.page_table_lock);
694         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
695                 pmd_populate_kernel(&init_mm, pmd, new);
696                 new = NULL;
697         }
698         spin_unlock(&init_mm.page_table_lock);
699         if (new)
700                 pte_free_kernel(&init_mm, new);
701         return 0;
702 }
703
704 static inline void init_rss_vec(int *rss)
705 {
706         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
707 }
708
709 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
710 {
711         int i;
712
713         if (current->mm == mm)
714                 sync_mm_rss(mm);
715         for (i = 0; i < NR_MM_COUNTERS; i++)
716                 if (rss[i])
717                         add_mm_counter(mm, i, rss[i]);
718 }
719
720 /*
721  * This function is called to print an error when a bad pte
722  * is found. For example, we might have a PFN-mapped pte in
723  * a region that doesn't allow it.
724  *
725  * The calling function must still handle the error.
726  */
727 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
728                           pte_t pte, struct page *page)
729 {
730         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
731         p4d_t *p4d = p4d_offset(pgd, addr);
732         pud_t *pud = pud_offset(p4d, addr);
733         pmd_t *pmd = pmd_offset(pud, addr);
734         struct address_space *mapping;
735         pgoff_t index;
736         static unsigned long resume;
737         static unsigned long nr_shown;
738         static unsigned long nr_unshown;
739
740         /*
741          * Allow a burst of 60 reports, then keep quiet for that minute;
742          * or allow a steady drip of one report per second.
743          */
744         if (nr_shown == 60) {
745                 if (time_before(jiffies, resume)) {
746                         nr_unshown++;
747                         return;
748                 }
749                 if (nr_unshown) {
750                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
751                                  nr_unshown);
752                         nr_unshown = 0;
753                 }
754                 nr_shown = 0;
755         }
756         if (nr_shown++ == 0)
757                 resume = jiffies + 60 * HZ;
758
759         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
760         index = linear_page_index(vma, addr);
761
762         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
763                  current->comm,
764                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
765         if (page)
766                 dump_page(page, "bad pte");
767         pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
768                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
769         pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
770                  vma->vm_file,
771                  vma->vm_ops ? vma->vm_ops->fault : NULL,
772                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
773                  mapping ? mapping->a_ops->readpage : NULL);
774         dump_stack();
775         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
776 }
777
778 /*
779  * vm_normal_page -- This function gets the "struct page" associated with a pte.
780  *
781  * "Special" mappings do not wish to be associated with a "struct page" (either
782  * it doesn't exist, or it exists but they don't want to touch it). In this
783  * case, NULL is returned here. "Normal" mappings do have a struct page.
784  *
785  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
786  * pte bit, in which case this function is trivial. Secondly, an architecture
787  * may not have a spare pte bit, which requires a more complicated scheme,
788  * described below.
789  *
790  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
791  * special mapping (even if there are underlying and valid "struct pages").
792  * COWed pages of a VM_PFNMAP are always normal.
793  *
794  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
795  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
796  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
797  * mapping will always honor the rule
798  *
799  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
800  *
801  * And for normal mappings this is false.
802  *
803  * This restricts such mappings to be a linear translation from virtual address
804  * to pfn. To get around this restriction, we allow arbitrary mappings so long
805  * as the vma is not a COW mapping; in that case, we know that all ptes are
806  * special (because none can have been COWed).
807  *
808  *
809  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
810  *
811  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
812  * page" backing, however the difference is that _all_ pages with a struct
813  * page (that is, those where pfn_valid is true) are refcounted and considered
814  * normal pages by the VM. The disadvantage is that pages are refcounted
815  * (which can be slower and simply not an option for some PFNMAP users). The
816  * advantage is that we don't have to follow the strict linearity rule of
817  * PFNMAP mappings in order to support COWable mappings.
818  *
819  */
820 #ifdef __HAVE_ARCH_PTE_SPECIAL
821 # define HAVE_PTE_SPECIAL 1
822 #else
823 # define HAVE_PTE_SPECIAL 0
824 #endif
825 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
826                              pte_t pte, bool with_public_device)
827 {
828         unsigned long pfn = pte_pfn(pte);
829
830         if (HAVE_PTE_SPECIAL) {
831                 if (likely(!pte_special(pte)))
832                         goto check_pfn;
833                 if (vma->vm_ops && vma->vm_ops->find_special_page)
834                         return vma->vm_ops->find_special_page(vma, addr);
835                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
836                         return NULL;
837                 if (is_zero_pfn(pfn))
838                         return NULL;
839
840                 /*
841                  * Device public pages are special pages (they are ZONE_DEVICE
842                  * pages but different from persistent memory). They behave
843                  * allmost like normal pages. The difference is that they are
844                  * not on the lru and thus should never be involve with any-
845                  * thing that involve lru manipulation (mlock, numa balancing,
846                  * ...).
847                  *
848                  * This is why we still want to return NULL for such page from
849                  * vm_normal_page() so that we do not have to special case all
850                  * call site of vm_normal_page().
851                  */
852                 if (likely(pfn <= highest_memmap_pfn)) {
853                         struct page *page = pfn_to_page(pfn);
854
855                         if (is_device_public_page(page)) {
856                                 if (with_public_device)
857                                         return page;
858                                 return NULL;
859                         }
860                 }
861                 print_bad_pte(vma, addr, pte, NULL);
862                 return NULL;
863         }
864
865         /* !HAVE_PTE_SPECIAL case follows: */
866
867         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
868                 if (vma->vm_flags & VM_MIXEDMAP) {
869                         if (!pfn_valid(pfn))
870                                 return NULL;
871                         goto out;
872                 } else {
873                         unsigned long off;
874                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
875                         if (pfn == vma->vm_pgoff + off)
876                                 return NULL;
877                         if (!is_cow_mapping(vma->vm_flags))
878                                 return NULL;
879                 }
880         }
881
882         if (is_zero_pfn(pfn))
883                 return NULL;
884 check_pfn:
885         if (unlikely(pfn > highest_memmap_pfn)) {
886                 print_bad_pte(vma, addr, pte, NULL);
887                 return NULL;
888         }
889
890         /*
891          * NOTE! We still have PageReserved() pages in the page tables.
892          * eg. VDSO mappings can cause them to exist.
893          */
894 out:
895         return pfn_to_page(pfn);
896 }
897
898 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
899 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
900                                 pmd_t pmd)
901 {
902         unsigned long pfn = pmd_pfn(pmd);
903
904         /*
905          * There is no pmd_special() but there may be special pmds, e.g.
906          * in a direct-access (dax) mapping, so let's just replicate the
907          * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
908          */
909         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
910                 if (vma->vm_flags & VM_MIXEDMAP) {
911                         if (!pfn_valid(pfn))
912                                 return NULL;
913                         goto out;
914                 } else {
915                         unsigned long off;
916                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
917                         if (pfn == vma->vm_pgoff + off)
918                                 return NULL;
919                         if (!is_cow_mapping(vma->vm_flags))
920                                 return NULL;
921                 }
922         }
923
924         if (is_zero_pfn(pfn))
925                 return NULL;
926         if (unlikely(pfn > highest_memmap_pfn))
927                 return NULL;
928
929         /*
930          * NOTE! We still have PageReserved() pages in the page tables.
931          * eg. VDSO mappings can cause them to exist.
932          */
933 out:
934         return pfn_to_page(pfn);
935 }
936 #endif
937
938 /*
939  * copy one vm_area from one task to the other. Assumes the page tables
940  * already present in the new task to be cleared in the whole range
941  * covered by this vma.
942  */
943
944 static inline unsigned long
945 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
946                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
947                 unsigned long addr, int *rss)
948 {
949         unsigned long vm_flags = vma->vm_flags;
950         pte_t pte = *src_pte;
951         struct page *page;
952
953         /* pte contains position in swap or file, so copy. */
954         if (unlikely(!pte_present(pte))) {
955                 swp_entry_t entry = pte_to_swp_entry(pte);
956
957                 if (likely(!non_swap_entry(entry))) {
958                         if (swap_duplicate(entry) < 0)
959                                 return entry.val;
960
961                         /* make sure dst_mm is on swapoff's mmlist. */
962                         if (unlikely(list_empty(&dst_mm->mmlist))) {
963                                 spin_lock(&mmlist_lock);
964                                 if (list_empty(&dst_mm->mmlist))
965                                         list_add(&dst_mm->mmlist,
966                                                         &src_mm->mmlist);
967                                 spin_unlock(&mmlist_lock);
968                         }
969                         rss[MM_SWAPENTS]++;
970                 } else if (is_migration_entry(entry)) {
971                         page = migration_entry_to_page(entry);
972
973                         rss[mm_counter(page)]++;
974
975                         if (is_write_migration_entry(entry) &&
976                                         is_cow_mapping(vm_flags)) {
977                                 /*
978                                  * COW mappings require pages in both
979                                  * parent and child to be set to read.
980                                  */
981                                 make_migration_entry_read(&entry);
982                                 pte = swp_entry_to_pte(entry);
983                                 if (pte_swp_soft_dirty(*src_pte))
984                                         pte = pte_swp_mksoft_dirty(pte);
985                                 set_pte_at(src_mm, addr, src_pte, pte);
986                         }
987                 } else if (is_device_private_entry(entry)) {
988                         page = device_private_entry_to_page(entry);
989
990                         /*
991                          * Update rss count even for unaddressable pages, as
992                          * they should treated just like normal pages in this
993                          * respect.
994                          *
995                          * We will likely want to have some new rss counters
996                          * for unaddressable pages, at some point. But for now
997                          * keep things as they are.
998                          */
999                         get_page(page);
1000                         rss[mm_counter(page)]++;
1001                         page_dup_rmap(page, false);
1002
1003                         /*
1004                          * We do not preserve soft-dirty information, because so
1005                          * far, checkpoint/restore is the only feature that
1006                          * requires that. And checkpoint/restore does not work
1007                          * when a device driver is involved (you cannot easily
1008                          * save and restore device driver state).
1009                          */
1010                         if (is_write_device_private_entry(entry) &&
1011                             is_cow_mapping(vm_flags)) {
1012                                 make_device_private_entry_read(&entry);
1013                                 pte = swp_entry_to_pte(entry);
1014                                 set_pte_at(src_mm, addr, src_pte, pte);
1015                         }
1016                 }
1017                 goto out_set_pte;
1018         }
1019
1020         /*
1021          * If it's a COW mapping, write protect it both
1022          * in the parent and the child
1023          */
1024         if (is_cow_mapping(vm_flags)) {
1025                 ptep_set_wrprotect(src_mm, addr, src_pte);
1026                 pte = pte_wrprotect(pte);
1027         }
1028
1029         /*
1030          * If it's a shared mapping, mark it clean in
1031          * the child
1032          */
1033         if (vm_flags & VM_SHARED)
1034                 pte = pte_mkclean(pte);
1035         pte = pte_mkold(pte);
1036
1037         page = vm_normal_page(vma, addr, pte);
1038         if (page) {
1039                 get_page(page);
1040                 page_dup_rmap(page, false);
1041                 rss[mm_counter(page)]++;
1042         } else if (pte_devmap(pte)) {
1043                 page = pte_page(pte);
1044
1045                 /*
1046                  * Cache coherent device memory behave like regular page and
1047                  * not like persistent memory page. For more informations see
1048                  * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1049                  */
1050                 if (is_device_public_page(page)) {
1051                         get_page(page);
1052                         page_dup_rmap(page, false);
1053                         rss[mm_counter(page)]++;
1054                 }
1055         }
1056
1057 out_set_pte:
1058         set_pte_at(dst_mm, addr, dst_pte, pte);
1059         return 0;
1060 }
1061
1062 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1063                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1064                    unsigned long addr, unsigned long end)
1065 {
1066         pte_t *orig_src_pte, *orig_dst_pte;
1067         pte_t *src_pte, *dst_pte;
1068         spinlock_t *src_ptl, *dst_ptl;
1069         int progress = 0;
1070         int rss[NR_MM_COUNTERS];
1071         swp_entry_t entry = (swp_entry_t){0};
1072
1073 again:
1074         init_rss_vec(rss);
1075
1076         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1077         if (!dst_pte)
1078                 return -ENOMEM;
1079         src_pte = pte_offset_map(src_pmd, addr);
1080         src_ptl = pte_lockptr(src_mm, src_pmd);
1081         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1082         orig_src_pte = src_pte;
1083         orig_dst_pte = dst_pte;
1084         arch_enter_lazy_mmu_mode();
1085
1086         do {
1087                 /*
1088                  * We are holding two locks at this point - either of them
1089                  * could generate latencies in another task on another CPU.
1090                  */
1091                 if (progress >= 32) {
1092                         progress = 0;
1093                         if (need_resched() ||
1094                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1095                                 break;
1096                 }
1097                 if (pte_none(*src_pte)) {
1098                         progress++;
1099                         continue;
1100                 }
1101                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1102                                                         vma, addr, rss);
1103                 if (entry.val)
1104                         break;
1105                 progress += 8;
1106         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1107
1108         arch_leave_lazy_mmu_mode();
1109         spin_unlock(src_ptl);
1110         pte_unmap(orig_src_pte);
1111         add_mm_rss_vec(dst_mm, rss);
1112         pte_unmap_unlock(orig_dst_pte, dst_ptl);
1113         cond_resched();
1114
1115         if (entry.val) {
1116                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1117                         return -ENOMEM;
1118                 progress = 0;
1119         }
1120         if (addr != end)
1121                 goto again;
1122         return 0;
1123 }
1124
1125 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1126                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1127                 unsigned long addr, unsigned long end)
1128 {
1129         pmd_t *src_pmd, *dst_pmd;
1130         unsigned long next;
1131
1132         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1133         if (!dst_pmd)
1134                 return -ENOMEM;
1135         src_pmd = pmd_offset(src_pud, addr);
1136         do {
1137                 next = pmd_addr_end(addr, end);
1138                 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1139                         || pmd_devmap(*src_pmd)) {
1140                         int err;
1141                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1142                         err = copy_huge_pmd(dst_mm, src_mm,
1143                                             dst_pmd, src_pmd, addr, vma);
1144                         if (err == -ENOMEM)
1145                                 return -ENOMEM;
1146                         if (!err)
1147                                 continue;
1148                         /* fall through */
1149                 }
1150                 if (pmd_none_or_clear_bad(src_pmd))
1151                         continue;
1152                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1153                                                 vma, addr, next))
1154                         return -ENOMEM;
1155         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1156         return 0;
1157 }
1158
1159 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1160                 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1161                 unsigned long addr, unsigned long end)
1162 {
1163         pud_t *src_pud, *dst_pud;
1164         unsigned long next;
1165
1166         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1167         if (!dst_pud)
1168                 return -ENOMEM;
1169         src_pud = pud_offset(src_p4d, addr);
1170         do {
1171                 next = pud_addr_end(addr, end);
1172                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1173                         int err;
1174
1175                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1176                         err = copy_huge_pud(dst_mm, src_mm,
1177                                             dst_pud, src_pud, addr, vma);
1178                         if (err == -ENOMEM)
1179                                 return -ENOMEM;
1180                         if (!err)
1181                                 continue;
1182                         /* fall through */
1183                 }
1184                 if (pud_none_or_clear_bad(src_pud))
1185                         continue;
1186                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1187                                                 vma, addr, next))
1188                         return -ENOMEM;
1189         } while (dst_pud++, src_pud++, addr = next, addr != end);
1190         return 0;
1191 }
1192
1193 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1194                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1195                 unsigned long addr, unsigned long end)
1196 {
1197         p4d_t *src_p4d, *dst_p4d;
1198         unsigned long next;
1199
1200         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1201         if (!dst_p4d)
1202                 return -ENOMEM;
1203         src_p4d = p4d_offset(src_pgd, addr);
1204         do {
1205                 next = p4d_addr_end(addr, end);
1206                 if (p4d_none_or_clear_bad(src_p4d))
1207                         continue;
1208                 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1209                                                 vma, addr, next))
1210                         return -ENOMEM;
1211         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1212         return 0;
1213 }
1214
1215 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1216                 struct vm_area_struct *vma)
1217 {
1218         pgd_t *src_pgd, *dst_pgd;
1219         unsigned long next;
1220         unsigned long addr = vma->vm_start;
1221         unsigned long end = vma->vm_end;
1222         unsigned long mmun_start;       /* For mmu_notifiers */
1223         unsigned long mmun_end;         /* For mmu_notifiers */
1224         bool is_cow;
1225         int ret;
1226
1227         /*
1228          * Don't copy ptes where a page fault will fill them correctly.
1229          * Fork becomes much lighter when there are big shared or private
1230          * readonly mappings. The tradeoff is that copy_page_range is more
1231          * efficient than faulting.
1232          */
1233         if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1234                         !vma->anon_vma)
1235                 return 0;
1236
1237         if (is_vm_hugetlb_page(vma))
1238                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1239
1240         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1241                 /*
1242                  * We do not free on error cases below as remove_vma
1243                  * gets called on error from higher level routine
1244                  */
1245                 ret = track_pfn_copy(vma);
1246                 if (ret)
1247                         return ret;
1248         }
1249
1250         /*
1251          * We need to invalidate the secondary MMU mappings only when
1252          * there could be a permission downgrade on the ptes of the
1253          * parent mm. And a permission downgrade will only happen if
1254          * is_cow_mapping() returns true.
1255          */
1256         is_cow = is_cow_mapping(vma->vm_flags);
1257         mmun_start = addr;
1258         mmun_end   = end;
1259         if (is_cow)
1260                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1261                                                     mmun_end);
1262
1263         ret = 0;
1264         dst_pgd = pgd_offset(dst_mm, addr);
1265         src_pgd = pgd_offset(src_mm, addr);
1266         do {
1267                 next = pgd_addr_end(addr, end);
1268                 if (pgd_none_or_clear_bad(src_pgd))
1269                         continue;
1270                 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1271                                             vma, addr, next))) {
1272                         ret = -ENOMEM;
1273                         break;
1274                 }
1275         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1276
1277         if (is_cow)
1278                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1279         return ret;
1280 }
1281
1282 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1283                                 struct vm_area_struct *vma, pmd_t *pmd,
1284                                 unsigned long addr, unsigned long end,
1285                                 struct zap_details *details)
1286 {
1287         struct mm_struct *mm = tlb->mm;
1288         int force_flush = 0;
1289         int rss[NR_MM_COUNTERS];
1290         spinlock_t *ptl;
1291         pte_t *start_pte;
1292         pte_t *pte;
1293         swp_entry_t entry;
1294
1295         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1296 again:
1297         init_rss_vec(rss);
1298         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1299         pte = start_pte;
1300         flush_tlb_batched_pending(mm);
1301         arch_enter_lazy_mmu_mode();
1302         do {
1303                 pte_t ptent = *pte;
1304                 if (pte_none(ptent))
1305                         continue;
1306
1307                 if (pte_present(ptent)) {
1308                         struct page *page;
1309
1310                         page = _vm_normal_page(vma, addr, ptent, true);
1311                         if (unlikely(details) && page) {
1312                                 /*
1313                                  * unmap_shared_mapping_pages() wants to
1314                                  * invalidate cache without truncating:
1315                                  * unmap shared but keep private pages.
1316                                  */
1317                                 if (details->check_mapping &&
1318                                     details->check_mapping != page_rmapping(page))
1319                                         continue;
1320                         }
1321                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1322                                                         tlb->fullmm);
1323                         tlb_remove_tlb_entry(tlb, pte, addr);
1324                         if (unlikely(!page))
1325                                 continue;
1326
1327                         if (!PageAnon(page)) {
1328                                 if (pte_dirty(ptent)) {
1329                                         force_flush = 1;
1330                                         set_page_dirty(page);
1331                                 }
1332                                 if (pte_young(ptent) &&
1333                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1334                                         mark_page_accessed(page);
1335                         }
1336                         rss[mm_counter(page)]--;
1337                         page_remove_rmap(page, false);
1338                         if (unlikely(page_mapcount(page) < 0))
1339                                 print_bad_pte(vma, addr, ptent, page);
1340                         if (unlikely(__tlb_remove_page(tlb, page))) {
1341                                 force_flush = 1;
1342                                 addr += PAGE_SIZE;
1343                                 break;
1344                         }
1345                         continue;
1346                 }
1347
1348                 entry = pte_to_swp_entry(ptent);
1349                 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1350                         struct page *page = device_private_entry_to_page(entry);
1351
1352                         if (unlikely(details && details->check_mapping)) {
1353                                 /*
1354                                  * unmap_shared_mapping_pages() wants to
1355                                  * invalidate cache without truncating:
1356                                  * unmap shared but keep private pages.
1357                                  */
1358                                 if (details->check_mapping !=
1359                                     page_rmapping(page))
1360                                         continue;
1361                         }
1362
1363                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1364                         rss[mm_counter(page)]--;
1365                         page_remove_rmap(page, false);
1366                         put_page(page);
1367                         continue;
1368                 }
1369
1370                 /* If details->check_mapping, we leave swap entries. */
1371                 if (unlikely(details))
1372                         continue;
1373
1374                 entry = pte_to_swp_entry(ptent);
1375                 if (!non_swap_entry(entry))
1376                         rss[MM_SWAPENTS]--;
1377                 else if (is_migration_entry(entry)) {
1378                         struct page *page;
1379
1380                         page = migration_entry_to_page(entry);
1381                         rss[mm_counter(page)]--;
1382                 }
1383                 if (unlikely(!free_swap_and_cache(entry)))
1384                         print_bad_pte(vma, addr, ptent, NULL);
1385                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1386         } while (pte++, addr += PAGE_SIZE, addr != end);
1387
1388         add_mm_rss_vec(mm, rss);
1389         arch_leave_lazy_mmu_mode();
1390
1391         /* Do the actual TLB flush before dropping ptl */
1392         if (force_flush)
1393                 tlb_flush_mmu_tlbonly(tlb);
1394         pte_unmap_unlock(start_pte, ptl);
1395
1396         /*
1397          * If we forced a TLB flush (either due to running out of
1398          * batch buffers or because we needed to flush dirty TLB
1399          * entries before releasing the ptl), free the batched
1400          * memory too. Restart if we didn't do everything.
1401          */
1402         if (force_flush) {
1403                 force_flush = 0;
1404                 tlb_flush_mmu_free(tlb);
1405                 if (addr != end)
1406                         goto again;
1407         }
1408
1409         return addr;
1410 }
1411
1412 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1413                                 struct vm_area_struct *vma, pud_t *pud,
1414                                 unsigned long addr, unsigned long end,
1415                                 struct zap_details *details)
1416 {
1417         pmd_t *pmd;
1418         unsigned long next;
1419
1420         pmd = pmd_offset(pud, addr);
1421         do {
1422                 next = pmd_addr_end(addr, end);
1423                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1424                         if (next - addr != HPAGE_PMD_SIZE) {
1425                                 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1426                                     !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1427                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1428                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1429                                 goto next;
1430                         /* fall through */
1431                 }
1432                 /*
1433                  * Here there can be other concurrent MADV_DONTNEED or
1434                  * trans huge page faults running, and if the pmd is
1435                  * none or trans huge it can change under us. This is
1436                  * because MADV_DONTNEED holds the mmap_sem in read
1437                  * mode.
1438                  */
1439                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1440                         goto next;
1441                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1442 next:
1443                 cond_resched();
1444         } while (pmd++, addr = next, addr != end);
1445
1446         return addr;
1447 }
1448
1449 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1450                                 struct vm_area_struct *vma, p4d_t *p4d,
1451                                 unsigned long addr, unsigned long end,
1452                                 struct zap_details *details)
1453 {
1454         pud_t *pud;
1455         unsigned long next;
1456
1457         pud = pud_offset(p4d, addr);
1458         do {
1459                 next = pud_addr_end(addr, end);
1460                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1461                         if (next - addr != HPAGE_PUD_SIZE) {
1462                                 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1463                                 split_huge_pud(vma, pud, addr);
1464                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1465                                 goto next;
1466                         /* fall through */
1467                 }
1468                 if (pud_none_or_clear_bad(pud))
1469                         continue;
1470                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1471 next:
1472                 cond_resched();
1473         } while (pud++, addr = next, addr != end);
1474
1475         return addr;
1476 }
1477
1478 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1479                                 struct vm_area_struct *vma, pgd_t *pgd,
1480                                 unsigned long addr, unsigned long end,
1481                                 struct zap_details *details)
1482 {
1483         p4d_t *p4d;
1484         unsigned long next;
1485
1486         p4d = p4d_offset(pgd, addr);
1487         do {
1488                 next = p4d_addr_end(addr, end);
1489                 if (p4d_none_or_clear_bad(p4d))
1490                         continue;
1491                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1492         } while (p4d++, addr = next, addr != end);
1493
1494         return addr;
1495 }
1496
1497 void unmap_page_range(struct mmu_gather *tlb,
1498                              struct vm_area_struct *vma,
1499                              unsigned long addr, unsigned long end,
1500                              struct zap_details *details)
1501 {
1502         pgd_t *pgd;
1503         unsigned long next;
1504
1505         BUG_ON(addr >= end);
1506         tlb_start_vma(tlb, vma);
1507         pgd = pgd_offset(vma->vm_mm, addr);
1508         do {
1509                 next = pgd_addr_end(addr, end);
1510                 if (pgd_none_or_clear_bad(pgd))
1511                         continue;
1512                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1513         } while (pgd++, addr = next, addr != end);
1514         tlb_end_vma(tlb, vma);
1515 }
1516
1517
1518 static void unmap_single_vma(struct mmu_gather *tlb,
1519                 struct vm_area_struct *vma, unsigned long start_addr,
1520                 unsigned long end_addr,
1521                 struct zap_details *details)
1522 {
1523         unsigned long start = max(vma->vm_start, start_addr);
1524         unsigned long end;
1525
1526         if (start >= vma->vm_end)
1527                 return;
1528         end = min(vma->vm_end, end_addr);
1529         if (end <= vma->vm_start)
1530                 return;
1531
1532         if (vma->vm_file)
1533                 uprobe_munmap(vma, start, end);
1534
1535         if (unlikely(vma->vm_flags & VM_PFNMAP))
1536                 untrack_pfn(vma, 0, 0);
1537
1538         if (start != end) {
1539                 if (unlikely(is_vm_hugetlb_page(vma))) {
1540                         /*
1541                          * It is undesirable to test vma->vm_file as it
1542                          * should be non-null for valid hugetlb area.
1543                          * However, vm_file will be NULL in the error
1544                          * cleanup path of mmap_region. When
1545                          * hugetlbfs ->mmap method fails,
1546                          * mmap_region() nullifies vma->vm_file
1547                          * before calling this function to clean up.
1548                          * Since no pte has actually been setup, it is
1549                          * safe to do nothing in this case.
1550                          */
1551                         if (vma->vm_file) {
1552                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1553                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1554                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1555                         }
1556                 } else
1557                         unmap_page_range(tlb, vma, start, end, details);
1558         }
1559 }
1560
1561 /**
1562  * unmap_vmas - unmap a range of memory covered by a list of vma's
1563  * @tlb: address of the caller's struct mmu_gather
1564  * @vma: the starting vma
1565  * @start_addr: virtual address at which to start unmapping
1566  * @end_addr: virtual address at which to end unmapping
1567  *
1568  * Unmap all pages in the vma list.
1569  *
1570  * Only addresses between `start' and `end' will be unmapped.
1571  *
1572  * The VMA list must be sorted in ascending virtual address order.
1573  *
1574  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1575  * range after unmap_vmas() returns.  So the only responsibility here is to
1576  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1577  * drops the lock and schedules.
1578  */
1579 void unmap_vmas(struct mmu_gather *tlb,
1580                 struct vm_area_struct *vma, unsigned long start_addr,
1581                 unsigned long end_addr)
1582 {
1583         struct mm_struct *mm = vma->vm_mm;
1584
1585         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1586         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1587                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1588         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1589 }
1590
1591 /**
1592  * zap_page_range - remove user pages in a given range
1593  * @vma: vm_area_struct holding the applicable pages
1594  * @start: starting address of pages to zap
1595  * @size: number of bytes to zap
1596  *
1597  * Caller must protect the VMA list
1598  */
1599 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1600                 unsigned long size)
1601 {
1602         struct mm_struct *mm = vma->vm_mm;
1603         struct mmu_gather tlb;
1604         unsigned long end = start + size;
1605
1606         lru_add_drain();
1607         tlb_gather_mmu(&tlb, mm, start, end);
1608         update_hiwater_rss(mm);
1609         mmu_notifier_invalidate_range_start(mm, start, end);
1610         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1611                 unmap_single_vma(&tlb, vma, start, end, NULL);
1612
1613                 /*
1614                  * zap_page_range does not specify whether mmap_sem should be
1615                  * held for read or write. That allows parallel zap_page_range
1616                  * operations to unmap a PTE and defer a flush meaning that
1617                  * this call observes pte_none and fails to flush the TLB.
1618                  * Rather than adding a complex API, ensure that no stale
1619                  * TLB entries exist when this call returns.
1620                  */
1621                 flush_tlb_range(vma, start, end);
1622         }
1623
1624         mmu_notifier_invalidate_range_end(mm, start, end);
1625         tlb_finish_mmu(&tlb, start, end);
1626 }
1627
1628 /**
1629  * zap_page_range_single - remove user pages in a given range
1630  * @vma: vm_area_struct holding the applicable pages
1631  * @address: starting address of pages to zap
1632  * @size: number of bytes to zap
1633  * @details: details of shared cache invalidation
1634  *
1635  * The range must fit into one VMA.
1636  */
1637 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1638                 unsigned long size, struct zap_details *details)
1639 {
1640         struct mm_struct *mm = vma->vm_mm;
1641         struct mmu_gather tlb;
1642         unsigned long end = address + size;
1643
1644         lru_add_drain();
1645         tlb_gather_mmu(&tlb, mm, address, end);
1646         update_hiwater_rss(mm);
1647         mmu_notifier_invalidate_range_start(mm, address, end);
1648         unmap_single_vma(&tlb, vma, address, end, details);
1649         mmu_notifier_invalidate_range_end(mm, address, end);
1650         tlb_finish_mmu(&tlb, address, end);
1651 }
1652
1653 /**
1654  * zap_vma_ptes - remove ptes mapping the vma
1655  * @vma: vm_area_struct holding ptes to be zapped
1656  * @address: starting address of pages to zap
1657  * @size: number of bytes to zap
1658  *
1659  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1660  *
1661  * The entire address range must be fully contained within the vma.
1662  *
1663  * Returns 0 if successful.
1664  */
1665 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1666                 unsigned long size)
1667 {
1668         if (address < vma->vm_start || address + size > vma->vm_end ||
1669                         !(vma->vm_flags & VM_PFNMAP))
1670                 return -1;
1671         zap_page_range_single(vma, address, size, NULL);
1672         return 0;
1673 }
1674 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1675
1676 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1677                         spinlock_t **ptl)
1678 {
1679         pgd_t *pgd;
1680         p4d_t *p4d;
1681         pud_t *pud;
1682         pmd_t *pmd;
1683
1684         pgd = pgd_offset(mm, addr);
1685         p4d = p4d_alloc(mm, pgd, addr);
1686         if (!p4d)
1687                 return NULL;
1688         pud = pud_alloc(mm, p4d, addr);
1689         if (!pud)
1690                 return NULL;
1691         pmd = pmd_alloc(mm, pud, addr);
1692         if (!pmd)
1693                 return NULL;
1694
1695         VM_BUG_ON(pmd_trans_huge(*pmd));
1696         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1697 }
1698
1699 /*
1700  * This is the old fallback for page remapping.
1701  *
1702  * For historical reasons, it only allows reserved pages. Only
1703  * old drivers should use this, and they needed to mark their
1704  * pages reserved for the old functions anyway.
1705  */
1706 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1707                         struct page *page, pgprot_t prot)
1708 {
1709         struct mm_struct *mm = vma->vm_mm;
1710         int retval;
1711         pte_t *pte;
1712         spinlock_t *ptl;
1713
1714         retval = -EINVAL;
1715         if (PageAnon(page))
1716                 goto out;
1717         retval = -ENOMEM;
1718         flush_dcache_page(page);
1719         pte = get_locked_pte(mm, addr, &ptl);
1720         if (!pte)
1721                 goto out;
1722         retval = -EBUSY;
1723         if (!pte_none(*pte))
1724                 goto out_unlock;
1725
1726         /* Ok, finally just insert the thing.. */
1727         get_page(page);
1728         inc_mm_counter_fast(mm, mm_counter_file(page));
1729         page_add_file_rmap(page, false);
1730         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1731
1732         retval = 0;
1733         pte_unmap_unlock(pte, ptl);
1734         return retval;
1735 out_unlock:
1736         pte_unmap_unlock(pte, ptl);
1737 out:
1738         return retval;
1739 }
1740
1741 /**
1742  * vm_insert_page - insert single page into user vma
1743  * @vma: user vma to map to
1744  * @addr: target user address of this page
1745  * @page: source kernel page
1746  *
1747  * This allows drivers to insert individual pages they've allocated
1748  * into a user vma.
1749  *
1750  * The page has to be a nice clean _individual_ kernel allocation.
1751  * If you allocate a compound page, you need to have marked it as
1752  * such (__GFP_COMP), or manually just split the page up yourself
1753  * (see split_page()).
1754  *
1755  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1756  * took an arbitrary page protection parameter. This doesn't allow
1757  * that. Your vma protection will have to be set up correctly, which
1758  * means that if you want a shared writable mapping, you'd better
1759  * ask for a shared writable mapping!
1760  *
1761  * The page does not need to be reserved.
1762  *
1763  * Usually this function is called from f_op->mmap() handler
1764  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1765  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1766  * function from other places, for example from page-fault handler.
1767  */
1768 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1769                         struct page *page)
1770 {
1771         if (addr < vma->vm_start || addr >= vma->vm_end)
1772                 return -EFAULT;
1773         if (!page_count(page))
1774                 return -EINVAL;
1775         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1776                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1777                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1778                 vma->vm_flags |= VM_MIXEDMAP;
1779         }
1780         return insert_page(vma, addr, page, vma->vm_page_prot);
1781 }
1782 EXPORT_SYMBOL(vm_insert_page);
1783
1784 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1785                         pfn_t pfn, pgprot_t prot, bool mkwrite)
1786 {
1787         struct mm_struct *mm = vma->vm_mm;
1788         int retval;
1789         pte_t *pte, entry;
1790         spinlock_t *ptl;
1791
1792         retval = -ENOMEM;
1793         pte = get_locked_pte(mm, addr, &ptl);
1794         if (!pte)
1795                 goto out;
1796         retval = -EBUSY;
1797         if (!pte_none(*pte)) {
1798                 if (mkwrite) {
1799                         /*
1800                          * For read faults on private mappings the PFN passed
1801                          * in may not match the PFN we have mapped if the
1802                          * mapped PFN is a writeable COW page.  In the mkwrite
1803                          * case we are creating a writable PTE for a shared
1804                          * mapping and we expect the PFNs to match.
1805                          */
1806                         if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1807                                 goto out_unlock;
1808                         entry = *pte;
1809                         goto out_mkwrite;
1810                 } else
1811                         goto out_unlock;
1812         }
1813
1814         /* Ok, finally just insert the thing.. */
1815         if (pfn_t_devmap(pfn))
1816                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1817         else
1818                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1819
1820 out_mkwrite:
1821         if (mkwrite) {
1822                 entry = pte_mkyoung(entry);
1823                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1824         }
1825
1826         set_pte_at(mm, addr, pte, entry);
1827         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1828
1829         retval = 0;
1830 out_unlock:
1831         pte_unmap_unlock(pte, ptl);
1832 out:
1833         return retval;
1834 }
1835
1836 /**
1837  * vm_insert_pfn - insert single pfn into user vma
1838  * @vma: user vma to map to
1839  * @addr: target user address of this page
1840  * @pfn: source kernel pfn
1841  *
1842  * Similar to vm_insert_page, this allows drivers to insert individual pages
1843  * they've allocated into a user vma. Same comments apply.
1844  *
1845  * This function should only be called from a vm_ops->fault handler, and
1846  * in that case the handler should return NULL.
1847  *
1848  * vma cannot be a COW mapping.
1849  *
1850  * As this is called only for pages that do not currently exist, we
1851  * do not need to flush old virtual caches or the TLB.
1852  */
1853 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1854                         unsigned long pfn)
1855 {
1856         return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1857 }
1858 EXPORT_SYMBOL(vm_insert_pfn);
1859
1860 /**
1861  * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1862  * @vma: user vma to map to
1863  * @addr: target user address of this page
1864  * @pfn: source kernel pfn
1865  * @pgprot: pgprot flags for the inserted page
1866  *
1867  * This is exactly like vm_insert_pfn, except that it allows drivers to
1868  * to override pgprot on a per-page basis.
1869  *
1870  * This only makes sense for IO mappings, and it makes no sense for
1871  * cow mappings.  In general, using multiple vmas is preferable;
1872  * vm_insert_pfn_prot should only be used if using multiple VMAs is
1873  * impractical.
1874  */
1875 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1876                         unsigned long pfn, pgprot_t pgprot)
1877 {
1878         int ret;
1879         /*
1880          * Technically, architectures with pte_special can avoid all these
1881          * restrictions (same for remap_pfn_range).  However we would like
1882          * consistency in testing and feature parity among all, so we should
1883          * try to keep these invariants in place for everybody.
1884          */
1885         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1886         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1887                                                 (VM_PFNMAP|VM_MIXEDMAP));
1888         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1889         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1890
1891         if (addr < vma->vm_start || addr >= vma->vm_end)
1892                 return -EFAULT;
1893
1894         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1895
1896         ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1897                         false);
1898
1899         return ret;
1900 }
1901 EXPORT_SYMBOL(vm_insert_pfn_prot);
1902
1903 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1904 {
1905         /* these checks mirror the abort conditions in vm_normal_page */
1906         if (vma->vm_flags & VM_MIXEDMAP)
1907                 return true;
1908         if (pfn_t_devmap(pfn))
1909                 return true;
1910         if (pfn_t_special(pfn))
1911                 return true;
1912         if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1913                 return true;
1914         return false;
1915 }
1916
1917 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1918                         pfn_t pfn, bool mkwrite)
1919 {
1920         pgprot_t pgprot = vma->vm_page_prot;
1921
1922         BUG_ON(!vm_mixed_ok(vma, pfn));
1923
1924         if (addr < vma->vm_start || addr >= vma->vm_end)
1925                 return -EFAULT;
1926
1927         track_pfn_insert(vma, &pgprot, pfn);
1928
1929         /*
1930          * If we don't have pte special, then we have to use the pfn_valid()
1931          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1932          * refcount the page if pfn_valid is true (hence insert_page rather
1933          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1934          * without pte special, it would there be refcounted as a normal page.
1935          */
1936         if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1937                 struct page *page;
1938
1939                 /*
1940                  * At this point we are committed to insert_page()
1941                  * regardless of whether the caller specified flags that
1942                  * result in pfn_t_has_page() == false.
1943                  */
1944                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1945                 return insert_page(vma, addr, page, pgprot);
1946         }
1947         return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1948 }
1949
1950 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1951                         pfn_t pfn)
1952 {
1953         return __vm_insert_mixed(vma, addr, pfn, false);
1954
1955 }
1956 EXPORT_SYMBOL(vm_insert_mixed);
1957
1958 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1959                         pfn_t pfn)
1960 {
1961         return __vm_insert_mixed(vma, addr, pfn, true);
1962 }
1963 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1964
1965 /*
1966  * maps a range of physical memory into the requested pages. the old
1967  * mappings are removed. any references to nonexistent pages results
1968  * in null mappings (currently treated as "copy-on-access")
1969  */
1970 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1971                         unsigned long addr, unsigned long end,
1972                         unsigned long pfn, pgprot_t prot)
1973 {
1974         pte_t *pte;
1975         spinlock_t *ptl;
1976
1977         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1978         if (!pte)
1979                 return -ENOMEM;
1980         arch_enter_lazy_mmu_mode();
1981         do {
1982                 BUG_ON(!pte_none(*pte));
1983                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1984                 pfn++;
1985         } while (pte++, addr += PAGE_SIZE, addr != end);
1986         arch_leave_lazy_mmu_mode();
1987         pte_unmap_unlock(pte - 1, ptl);
1988         return 0;
1989 }
1990
1991 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1992                         unsigned long addr, unsigned long end,
1993                         unsigned long pfn, pgprot_t prot)
1994 {
1995         pmd_t *pmd;
1996         unsigned long next;
1997
1998         pfn -= addr >> PAGE_SHIFT;
1999         pmd = pmd_alloc(mm, pud, addr);
2000         if (!pmd)
2001                 return -ENOMEM;
2002         VM_BUG_ON(pmd_trans_huge(*pmd));
2003         do {
2004                 next = pmd_addr_end(addr, end);
2005                 if (remap_pte_range(mm, pmd, addr, next,
2006                                 pfn + (addr >> PAGE_SHIFT), prot))
2007                         return -ENOMEM;
2008         } while (pmd++, addr = next, addr != end);
2009         return 0;
2010 }
2011
2012 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2013                         unsigned long addr, unsigned long end,
2014                         unsigned long pfn, pgprot_t prot)
2015 {
2016         pud_t *pud;
2017         unsigned long next;
2018
2019         pfn -= addr >> PAGE_SHIFT;
2020         pud = pud_alloc(mm, p4d, addr);
2021         if (!pud)
2022                 return -ENOMEM;
2023         do {
2024                 next = pud_addr_end(addr, end);
2025                 if (remap_pmd_range(mm, pud, addr, next,
2026                                 pfn + (addr >> PAGE_SHIFT), prot))
2027                         return -ENOMEM;
2028         } while (pud++, addr = next, addr != end);
2029         return 0;
2030 }
2031
2032 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2033                         unsigned long addr, unsigned long end,
2034                         unsigned long pfn, pgprot_t prot)
2035 {
2036         p4d_t *p4d;
2037         unsigned long next;
2038
2039         pfn -= addr >> PAGE_SHIFT;
2040         p4d = p4d_alloc(mm, pgd, addr);
2041         if (!p4d)
2042                 return -ENOMEM;
2043         do {
2044                 next = p4d_addr_end(addr, end);
2045                 if (remap_pud_range(mm, p4d, addr, next,
2046                                 pfn + (addr >> PAGE_SHIFT), prot))
2047                         return -ENOMEM;
2048         } while (p4d++, addr = next, addr != end);
2049         return 0;
2050 }
2051
2052 /**
2053  * remap_pfn_range - remap kernel memory to userspace
2054  * @vma: user vma to map to
2055  * @addr: target user address to start at
2056  * @pfn: physical address of kernel memory
2057  * @size: size of map area
2058  * @prot: page protection flags for this mapping
2059  *
2060  *  Note: this is only safe if the mm semaphore is held when called.
2061  */
2062 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2063                     unsigned long pfn, unsigned long size, pgprot_t prot)
2064 {
2065         pgd_t *pgd;
2066         unsigned long next;
2067         unsigned long end = addr + PAGE_ALIGN(size);
2068         struct mm_struct *mm = vma->vm_mm;
2069         unsigned long remap_pfn = pfn;
2070         int err;
2071
2072         /*
2073          * Physically remapped pages are special. Tell the
2074          * rest of the world about it:
2075          *   VM_IO tells people not to look at these pages
2076          *      (accesses can have side effects).
2077          *   VM_PFNMAP tells the core MM that the base pages are just
2078          *      raw PFN mappings, and do not have a "struct page" associated
2079          *      with them.
2080          *   VM_DONTEXPAND
2081          *      Disable vma merging and expanding with mremap().
2082          *   VM_DONTDUMP
2083          *      Omit vma from core dump, even when VM_IO turned off.
2084          *
2085          * There's a horrible special case to handle copy-on-write
2086          * behaviour that some programs depend on. We mark the "original"
2087          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2088          * See vm_normal_page() for details.
2089          */
2090         if (is_cow_mapping(vma->vm_flags)) {
2091                 if (addr != vma->vm_start || end != vma->vm_end)
2092                         return -EINVAL;
2093                 vma->vm_pgoff = pfn;
2094         }
2095
2096         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2097         if (err)
2098                 return -EINVAL;
2099
2100         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2101
2102         BUG_ON(addr >= end);
2103         pfn -= addr >> PAGE_SHIFT;
2104         pgd = pgd_offset(mm, addr);
2105         flush_cache_range(vma, addr, end);
2106         do {
2107                 next = pgd_addr_end(addr, end);
2108                 err = remap_p4d_range(mm, pgd, addr, next,
2109                                 pfn + (addr >> PAGE_SHIFT), prot);
2110                 if (err)
2111                         break;
2112         } while (pgd++, addr = next, addr != end);
2113
2114         if (err)
2115                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2116
2117         return err;
2118 }
2119 EXPORT_SYMBOL(remap_pfn_range);
2120
2121 /**
2122  * vm_iomap_memory - remap memory to userspace
2123  * @vma: user vma to map to
2124  * @start: start of area
2125  * @len: size of area
2126  *
2127  * This is a simplified io_remap_pfn_range() for common driver use. The
2128  * driver just needs to give us the physical memory range to be mapped,
2129  * we'll figure out the rest from the vma information.
2130  *
2131  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2132  * whatever write-combining details or similar.
2133  */
2134 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2135 {
2136         unsigned long vm_len, pfn, pages;
2137
2138         /* Check that the physical memory area passed in looks valid */
2139         if (start + len < start)
2140                 return -EINVAL;
2141         /*
2142          * You *really* shouldn't map things that aren't page-aligned,
2143          * but we've historically allowed it because IO memory might
2144          * just have smaller alignment.
2145          */
2146         len += start & ~PAGE_MASK;
2147         pfn = start >> PAGE_SHIFT;
2148         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2149         if (pfn + pages < pfn)
2150                 return -EINVAL;
2151
2152         /* We start the mapping 'vm_pgoff' pages into the area */
2153         if (vma->vm_pgoff > pages)
2154                 return -EINVAL;
2155         pfn += vma->vm_pgoff;
2156         pages -= vma->vm_pgoff;
2157
2158         /* Can we fit all of the mapping? */
2159         vm_len = vma->vm_end - vma->vm_start;
2160         if (vm_len >> PAGE_SHIFT > pages)
2161                 return -EINVAL;
2162
2163         /* Ok, let it rip */
2164         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2165 }
2166 EXPORT_SYMBOL(vm_iomap_memory);
2167
2168 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2169                                      unsigned long addr, unsigned long end,
2170                                      pte_fn_t fn, void *data)
2171 {
2172         pte_t *pte;
2173         int err;
2174         pgtable_t token;
2175         spinlock_t *uninitialized_var(ptl);
2176
2177         pte = (mm == &init_mm) ?
2178                 pte_alloc_kernel(pmd, addr) :
2179                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2180         if (!pte)
2181                 return -ENOMEM;
2182
2183         BUG_ON(pmd_huge(*pmd));
2184
2185         arch_enter_lazy_mmu_mode();
2186
2187         token = pmd_pgtable(*pmd);
2188
2189         do {
2190                 err = fn(pte++, token, addr, data);
2191                 if (err)
2192                         break;
2193         } while (addr += PAGE_SIZE, addr != end);
2194
2195         arch_leave_lazy_mmu_mode();
2196
2197         if (mm != &init_mm)
2198                 pte_unmap_unlock(pte-1, ptl);
2199         return err;
2200 }
2201
2202 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2203                                      unsigned long addr, unsigned long end,
2204                                      pte_fn_t fn, void *data)
2205 {
2206         pmd_t *pmd;
2207         unsigned long next;
2208         int err;
2209
2210         BUG_ON(pud_huge(*pud));
2211
2212         pmd = pmd_alloc(mm, pud, addr);
2213         if (!pmd)
2214                 return -ENOMEM;
2215         do {
2216                 next = pmd_addr_end(addr, end);
2217                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2218                 if (err)
2219                         break;
2220         } while (pmd++, addr = next, addr != end);
2221         return err;
2222 }
2223
2224 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2225                                      unsigned long addr, unsigned long end,
2226                                      pte_fn_t fn, void *data)
2227 {
2228         pud_t *pud;
2229         unsigned long next;
2230         int err;
2231
2232         pud = pud_alloc(mm, p4d, addr);
2233         if (!pud)
2234                 return -ENOMEM;
2235         do {
2236                 next = pud_addr_end(addr, end);
2237                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2238                 if (err)
2239                         break;
2240         } while (pud++, addr = next, addr != end);
2241         return err;
2242 }
2243
2244 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2245                                      unsigned long addr, unsigned long end,
2246                                      pte_fn_t fn, void *data)
2247 {
2248         p4d_t *p4d;
2249         unsigned long next;
2250         int err;
2251
2252         p4d = p4d_alloc(mm, pgd, addr);
2253         if (!p4d)
2254                 return -ENOMEM;
2255         do {
2256                 next = p4d_addr_end(addr, end);
2257                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2258                 if (err)
2259                         break;
2260         } while (p4d++, addr = next, addr != end);
2261         return err;
2262 }
2263
2264 /*
2265  * Scan a region of virtual memory, filling in page tables as necessary
2266  * and calling a provided function on each leaf page table.
2267  */
2268 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2269                         unsigned long size, pte_fn_t fn, void *data)
2270 {
2271         pgd_t *pgd;
2272         unsigned long next;
2273         unsigned long end = addr + size;
2274         int err;
2275
2276         if (WARN_ON(addr >= end))
2277                 return -EINVAL;
2278
2279         pgd = pgd_offset(mm, addr);
2280         do {
2281                 next = pgd_addr_end(addr, end);
2282                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2283                 if (err)
2284                         break;
2285         } while (pgd++, addr = next, addr != end);
2286
2287         return err;
2288 }
2289 EXPORT_SYMBOL_GPL(apply_to_page_range);
2290
2291 /*
2292  * handle_pte_fault chooses page fault handler according to an entry which was
2293  * read non-atomically.  Before making any commitment, on those architectures
2294  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2295  * parts, do_swap_page must check under lock before unmapping the pte and
2296  * proceeding (but do_wp_page is only called after already making such a check;
2297  * and do_anonymous_page can safely check later on).
2298  */
2299 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2300                                 pte_t *page_table, pte_t orig_pte)
2301 {
2302         int same = 1;
2303 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2304         if (sizeof(pte_t) > sizeof(unsigned long)) {
2305                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2306                 spin_lock(ptl);
2307                 same = pte_same(*page_table, orig_pte);
2308                 spin_unlock(ptl);
2309         }
2310 #endif
2311         pte_unmap(page_table);
2312         return same;
2313 }
2314
2315 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2316 {
2317         debug_dma_assert_idle(src);
2318
2319         /*
2320          * If the source page was a PFN mapping, we don't have
2321          * a "struct page" for it. We do a best-effort copy by
2322          * just copying from the original user address. If that
2323          * fails, we just zero-fill it. Live with it.
2324          */
2325         if (unlikely(!src)) {
2326                 void *kaddr = kmap_atomic(dst);
2327                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2328
2329                 /*
2330                  * This really shouldn't fail, because the page is there
2331                  * in the page tables. But it might just be unreadable,
2332                  * in which case we just give up and fill the result with
2333                  * zeroes.
2334                  */
2335                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2336                         clear_page(kaddr);
2337                 kunmap_atomic(kaddr);
2338                 flush_dcache_page(dst);
2339         } else
2340                 copy_user_highpage(dst, src, va, vma);
2341 }
2342
2343 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2344 {
2345         struct file *vm_file = vma->vm_file;
2346
2347         if (vm_file)
2348                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2349
2350         /*
2351          * Special mappings (e.g. VDSO) do not have any file so fake
2352          * a default GFP_KERNEL for them.
2353          */
2354         return GFP_KERNEL;
2355 }
2356
2357 /*
2358  * Notify the address space that the page is about to become writable so that
2359  * it can prohibit this or wait for the page to get into an appropriate state.
2360  *
2361  * We do this without the lock held, so that it can sleep if it needs to.
2362  */
2363 static int do_page_mkwrite(struct vm_fault *vmf)
2364 {
2365         int ret;
2366         struct page *page = vmf->page;
2367         unsigned int old_flags = vmf->flags;
2368
2369         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2370
2371         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2372         /* Restore original flags so that caller is not surprised */
2373         vmf->flags = old_flags;
2374         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2375                 return ret;
2376         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2377                 lock_page(page);
2378                 if (!page->mapping) {
2379                         unlock_page(page);
2380                         return 0; /* retry */
2381                 }
2382                 ret |= VM_FAULT_LOCKED;
2383         } else
2384                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2385         return ret;
2386 }
2387
2388 /*
2389  * Handle dirtying of a page in shared file mapping on a write fault.
2390  *
2391  * The function expects the page to be locked and unlocks it.
2392  */
2393 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2394                                     struct page *page)
2395 {
2396         struct address_space *mapping;
2397         bool dirtied;
2398         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2399
2400         dirtied = set_page_dirty(page);
2401         VM_BUG_ON_PAGE(PageAnon(page), page);
2402         /*
2403          * Take a local copy of the address_space - page.mapping may be zeroed
2404          * by truncate after unlock_page().   The address_space itself remains
2405          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2406          * release semantics to prevent the compiler from undoing this copying.
2407          */
2408         mapping = page_rmapping(page);
2409         unlock_page(page);
2410
2411         if ((dirtied || page_mkwrite) && mapping) {
2412                 /*
2413                  * Some device drivers do not set page.mapping
2414                  * but still dirty their pages
2415                  */
2416                 balance_dirty_pages_ratelimited(mapping);
2417         }
2418
2419         if (!page_mkwrite)
2420                 file_update_time(vma->vm_file);
2421 }
2422
2423 /*
2424  * Handle write page faults for pages that can be reused in the current vma
2425  *
2426  * This can happen either due to the mapping being with the VM_SHARED flag,
2427  * or due to us being the last reference standing to the page. In either
2428  * case, all we need to do here is to mark the page as writable and update
2429  * any related book-keeping.
2430  */
2431 static inline void wp_page_reuse(struct vm_fault *vmf)
2432         __releases(vmf->ptl)
2433 {
2434         struct vm_area_struct *vma = vmf->vma;
2435         struct page *page = vmf->page;
2436         pte_t entry;
2437         /*
2438          * Clear the pages cpupid information as the existing
2439          * information potentially belongs to a now completely
2440          * unrelated process.
2441          */
2442         if (page)
2443                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2444
2445         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2446         entry = pte_mkyoung(vmf->orig_pte);
2447         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2448         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2449                 update_mmu_cache(vma, vmf->address, vmf->pte);
2450         pte_unmap_unlock(vmf->pte, vmf->ptl);
2451 }
2452
2453 /*
2454  * Handle the case of a page which we actually need to copy to a new page.
2455  *
2456  * Called with mmap_sem locked and the old page referenced, but
2457  * without the ptl held.
2458  *
2459  * High level logic flow:
2460  *
2461  * - Allocate a page, copy the content of the old page to the new one.
2462  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2463  * - Take the PTL. If the pte changed, bail out and release the allocated page
2464  * - If the pte is still the way we remember it, update the page table and all
2465  *   relevant references. This includes dropping the reference the page-table
2466  *   held to the old page, as well as updating the rmap.
2467  * - In any case, unlock the PTL and drop the reference we took to the old page.
2468  */
2469 static int wp_page_copy(struct vm_fault *vmf)
2470 {
2471         struct vm_area_struct *vma = vmf->vma;
2472         struct mm_struct *mm = vma->vm_mm;
2473         struct page *old_page = vmf->page;
2474         struct page *new_page = NULL;
2475         pte_t entry;
2476         int page_copied = 0;
2477         const unsigned long mmun_start = vmf->address & PAGE_MASK;
2478         const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2479         struct mem_cgroup *memcg;
2480
2481         if (unlikely(anon_vma_prepare(vma)))
2482                 goto oom;
2483
2484         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2485                 new_page = alloc_zeroed_user_highpage_movable(vma,
2486                                                               vmf->address);
2487                 if (!new_page)
2488                         goto oom;
2489         } else {
2490                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2491                                 vmf->address);
2492                 if (!new_page)
2493                         goto oom;
2494                 cow_user_page(new_page, old_page, vmf->address, vma);
2495         }
2496
2497         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2498                 goto oom_free_new;
2499
2500         __SetPageUptodate(new_page);
2501
2502         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2503
2504         /*
2505          * Re-check the pte - we dropped the lock
2506          */
2507         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2508         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2509                 if (old_page) {
2510                         if (!PageAnon(old_page)) {
2511                                 dec_mm_counter_fast(mm,
2512                                                 mm_counter_file(old_page));
2513                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2514                         }
2515                 } else {
2516                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2517                 }
2518                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2519                 entry = mk_pte(new_page, vma->vm_page_prot);
2520                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2521                 /*
2522                  * Clear the pte entry and flush it first, before updating the
2523                  * pte with the new entry. This will avoid a race condition
2524                  * seen in the presence of one thread doing SMC and another
2525                  * thread doing COW.
2526                  */
2527                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2528                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2529                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2530                 lru_cache_add_active_or_unevictable(new_page, vma);
2531                 /*
2532                  * We call the notify macro here because, when using secondary
2533                  * mmu page tables (such as kvm shadow page tables), we want the
2534                  * new page to be mapped directly into the secondary page table.
2535                  */
2536                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2537                 update_mmu_cache(vma, vmf->address, vmf->pte);
2538                 if (old_page) {
2539                         /*
2540                          * Only after switching the pte to the new page may
2541                          * we remove the mapcount here. Otherwise another
2542                          * process may come and find the rmap count decremented
2543                          * before the pte is switched to the new page, and
2544                          * "reuse" the old page writing into it while our pte
2545                          * here still points into it and can be read by other
2546                          * threads.
2547                          *
2548                          * The critical issue is to order this
2549                          * page_remove_rmap with the ptp_clear_flush above.
2550                          * Those stores are ordered by (if nothing else,)
2551                          * the barrier present in the atomic_add_negative
2552                          * in page_remove_rmap.
2553                          *
2554                          * Then the TLB flush in ptep_clear_flush ensures that
2555                          * no process can access the old page before the
2556                          * decremented mapcount is visible. And the old page
2557                          * cannot be reused until after the decremented
2558                          * mapcount is visible. So transitively, TLBs to
2559                          * old page will be flushed before it can be reused.
2560                          */
2561                         page_remove_rmap(old_page, false);
2562                 }
2563
2564                 /* Free the old page.. */
2565                 new_page = old_page;
2566                 page_copied = 1;
2567         } else {
2568                 mem_cgroup_cancel_charge(new_page, memcg, false);
2569         }
2570
2571         if (new_page)
2572                 put_page(new_page);
2573
2574         pte_unmap_unlock(vmf->pte, vmf->ptl);
2575         /*
2576          * No need to double call mmu_notifier->invalidate_range() callback as
2577          * the above ptep_clear_flush_notify() did already call it.
2578          */
2579         mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2580         if (old_page) {
2581                 /*
2582                  * Don't let another task, with possibly unlocked vma,
2583                  * keep the mlocked page.
2584                  */
2585                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2586                         lock_page(old_page);    /* LRU manipulation */
2587                         if (PageMlocked(old_page))
2588                                 munlock_vma_page(old_page);
2589                         unlock_page(old_page);
2590                 }
2591                 put_page(old_page);
2592         }
2593         return page_copied ? VM_FAULT_WRITE : 0;
2594 oom_free_new:
2595         put_page(new_page);
2596 oom:
2597         if (old_page)
2598                 put_page(old_page);
2599         return VM_FAULT_OOM;
2600 }
2601
2602 /**
2603  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2604  *                        writeable once the page is prepared
2605  *
2606  * @vmf: structure describing the fault
2607  *
2608  * This function handles all that is needed to finish a write page fault in a
2609  * shared mapping due to PTE being read-only once the mapped page is prepared.
2610  * It handles locking of PTE and modifying it. The function returns
2611  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2612  * lock.
2613  *
2614  * The function expects the page to be locked or other protection against
2615  * concurrent faults / writeback (such as DAX radix tree locks).
2616  */
2617 int finish_mkwrite_fault(struct vm_fault *vmf)
2618 {
2619         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2620         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2621                                        &vmf->ptl);
2622         /*
2623          * We might have raced with another page fault while we released the
2624          * pte_offset_map_lock.
2625          */
2626         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2627                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2628                 return VM_FAULT_NOPAGE;
2629         }
2630         wp_page_reuse(vmf);
2631         return 0;
2632 }
2633
2634 /*
2635  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2636  * mapping
2637  */
2638 static int wp_pfn_shared(struct vm_fault *vmf)
2639 {
2640         struct vm_area_struct *vma = vmf->vma;
2641
2642         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2643                 int ret;
2644
2645                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2646                 vmf->flags |= FAULT_FLAG_MKWRITE;
2647                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2648                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2649                         return ret;
2650                 return finish_mkwrite_fault(vmf);
2651         }
2652         wp_page_reuse(vmf);
2653         return VM_FAULT_WRITE;
2654 }
2655
2656 static int wp_page_shared(struct vm_fault *vmf)
2657         __releases(vmf->ptl)
2658 {
2659         struct vm_area_struct *vma = vmf->vma;
2660
2661         get_page(vmf->page);
2662
2663         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2664                 int tmp;
2665
2666                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2667                 tmp = do_page_mkwrite(vmf);
2668                 if (unlikely(!tmp || (tmp &
2669                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2670                         put_page(vmf->page);
2671                         return tmp;
2672                 }
2673                 tmp = finish_mkwrite_fault(vmf);
2674                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2675                         unlock_page(vmf->page);
2676                         put_page(vmf->page);
2677                         return tmp;
2678                 }
2679         } else {
2680                 wp_page_reuse(vmf);
2681                 lock_page(vmf->page);
2682         }
2683         fault_dirty_shared_page(vma, vmf->page);
2684         put_page(vmf->page);
2685
2686         return VM_FAULT_WRITE;
2687 }
2688
2689 /*
2690  * This routine handles present pages, when users try to write
2691  * to a shared page. It is done by copying the page to a new address
2692  * and decrementing the shared-page counter for the old page.
2693  *
2694  * Note that this routine assumes that the protection checks have been
2695  * done by the caller (the low-level page fault routine in most cases).
2696  * Thus we can safely just mark it writable once we've done any necessary
2697  * COW.
2698  *
2699  * We also mark the page dirty at this point even though the page will
2700  * change only once the write actually happens. This avoids a few races,
2701  * and potentially makes it more efficient.
2702  *
2703  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2704  * but allow concurrent faults), with pte both mapped and locked.
2705  * We return with mmap_sem still held, but pte unmapped and unlocked.
2706  */
2707 static int do_wp_page(struct vm_fault *vmf)
2708         __releases(vmf->ptl)
2709 {
2710         struct vm_area_struct *vma = vmf->vma;
2711
2712         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2713         if (!vmf->page) {
2714                 /*
2715                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2716                  * VM_PFNMAP VMA.
2717                  *
2718                  * We should not cow pages in a shared writeable mapping.
2719                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2720                  */
2721                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2722                                      (VM_WRITE|VM_SHARED))
2723                         return wp_pfn_shared(vmf);
2724
2725                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2726                 return wp_page_copy(vmf);
2727         }
2728
2729         /*
2730          * Take out anonymous pages first, anonymous shared vmas are
2731          * not dirty accountable.
2732          */
2733         if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2734                 int total_map_swapcount;
2735                 if (!trylock_page(vmf->page)) {
2736                         get_page(vmf->page);
2737                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2738                         lock_page(vmf->page);
2739                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2740                                         vmf->address, &vmf->ptl);
2741                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2742                                 unlock_page(vmf->page);
2743                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2744                                 put_page(vmf->page);
2745                                 return 0;
2746                         }
2747                         put_page(vmf->page);
2748                 }
2749                 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2750                         if (total_map_swapcount == 1) {
2751                                 /*
2752                                  * The page is all ours. Move it to
2753                                  * our anon_vma so the rmap code will
2754                                  * not search our parent or siblings.
2755                                  * Protected against the rmap code by
2756                                  * the page lock.
2757                                  */
2758                                 page_move_anon_rmap(vmf->page, vma);
2759                         }
2760                         unlock_page(vmf->page);
2761                         wp_page_reuse(vmf);
2762                         return VM_FAULT_WRITE;
2763                 }
2764                 unlock_page(vmf->page);
2765         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2766                                         (VM_WRITE|VM_SHARED))) {
2767                 return wp_page_shared(vmf);
2768         }
2769
2770         /*
2771          * Ok, we need to copy. Oh, well..
2772          */
2773         get_page(vmf->page);
2774
2775         pte_unmap_unlock(vmf->pte, vmf->ptl);
2776         return wp_page_copy(vmf);
2777 }
2778
2779 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2780                 unsigned long start_addr, unsigned long end_addr,
2781                 struct zap_details *details)
2782 {
2783         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2784 }
2785
2786 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2787                                             struct zap_details *details)
2788 {
2789         struct vm_area_struct *vma;
2790         pgoff_t vba, vea, zba, zea;
2791
2792         vma_interval_tree_foreach(vma, root,
2793                         details->first_index, details->last_index) {
2794
2795                 vba = vma->vm_pgoff;
2796                 vea = vba + vma_pages(vma) - 1;
2797                 zba = details->first_index;
2798                 if (zba < vba)
2799                         zba = vba;
2800                 zea = details->last_index;
2801                 if (zea > vea)
2802                         zea = vea;
2803
2804                 unmap_mapping_range_vma(vma,
2805                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2806                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2807                                 details);
2808         }
2809 }
2810
2811 /**
2812  * unmap_mapping_pages() - Unmap pages from processes.
2813  * @mapping: The address space containing pages to be unmapped.
2814  * @start: Index of first page to be unmapped.
2815  * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
2816  * @even_cows: Whether to unmap even private COWed pages.
2817  *
2818  * Unmap the pages in this address space from any userspace process which
2819  * has them mmaped.  Generally, you want to remove COWed pages as well when
2820  * a file is being truncated, but not when invalidating pages from the page
2821  * cache.
2822  */
2823 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2824                 pgoff_t nr, bool even_cows)
2825 {
2826         struct zap_details details = { };
2827
2828         details.check_mapping = even_cows ? NULL : mapping;
2829         details.first_index = start;
2830         details.last_index = start + nr - 1;
2831         if (details.last_index < details.first_index)
2832                 details.last_index = ULONG_MAX;
2833
2834         i_mmap_lock_write(mapping);
2835         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2836                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2837         i_mmap_unlock_write(mapping);
2838 }
2839
2840 /**
2841  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2842  * address_space corresponding to the specified byte range in the underlying
2843  * file.
2844  *
2845  * @mapping: the address space containing mmaps to be unmapped.
2846  * @holebegin: byte in first page to unmap, relative to the start of
2847  * the underlying file.  This will be rounded down to a PAGE_SIZE
2848  * boundary.  Note that this is different from truncate_pagecache(), which
2849  * must keep the partial page.  In contrast, we must get rid of
2850  * partial pages.
2851  * @holelen: size of prospective hole in bytes.  This will be rounded
2852  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2853  * end of the file.
2854  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2855  * but 0 when invalidating pagecache, don't throw away private data.
2856  */
2857 void unmap_mapping_range(struct address_space *mapping,
2858                 loff_t const holebegin, loff_t const holelen, int even_cows)
2859 {
2860         pgoff_t hba = holebegin >> PAGE_SHIFT;
2861         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2862
2863         /* Check for overflow. */
2864         if (sizeof(holelen) > sizeof(hlen)) {
2865                 long long holeend =
2866                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2867                 if (holeend & ~(long long)ULONG_MAX)
2868                         hlen = ULONG_MAX - hba + 1;
2869         }
2870
2871         unmap_mapping_pages(mapping, hba, hlen, even_cows);
2872 }
2873 EXPORT_SYMBOL(unmap_mapping_range);
2874
2875 /*
2876  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2877  * but allow concurrent faults), and pte mapped but not yet locked.
2878  * We return with pte unmapped and unlocked.
2879  *
2880  * We return with the mmap_sem locked or unlocked in the same cases
2881  * as does filemap_fault().
2882  */
2883 int do_swap_page(struct vm_fault *vmf)
2884 {
2885         struct vm_area_struct *vma = vmf->vma;
2886         struct page *page = NULL, *swapcache;
2887         struct mem_cgroup *memcg;
2888         swp_entry_t entry;
2889         pte_t pte;
2890         int locked;
2891         int exclusive = 0;
2892         int ret = 0;
2893
2894         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2895                 goto out;
2896
2897         entry = pte_to_swp_entry(vmf->orig_pte);
2898         if (unlikely(non_swap_entry(entry))) {
2899                 if (is_migration_entry(entry)) {
2900                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2901                                              vmf->address);
2902                 } else if (is_device_private_entry(entry)) {
2903                         /*
2904                          * For un-addressable device memory we call the pgmap
2905                          * fault handler callback. The callback must migrate
2906                          * the page back to some CPU accessible page.
2907                          */
2908                         ret = device_private_entry_fault(vma, vmf->address, entry,
2909                                                  vmf->flags, vmf->pmd);
2910                 } else if (is_hwpoison_entry(entry)) {
2911                         ret = VM_FAULT_HWPOISON;
2912                 } else {
2913                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2914                         ret = VM_FAULT_SIGBUS;
2915                 }
2916                 goto out;
2917         }
2918
2919
2920         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2921         page = lookup_swap_cache(entry, vma, vmf->address);
2922         swapcache = page;
2923
2924         if (!page) {
2925                 struct swap_info_struct *si = swp_swap_info(entry);
2926
2927                 if (si->flags & SWP_SYNCHRONOUS_IO &&
2928                                 __swap_count(si, entry) == 1) {
2929                         /* skip swapcache */
2930                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2931                         if (page) {
2932                                 __SetPageLocked(page);
2933                                 __SetPageSwapBacked(page);
2934                                 set_page_private(page, entry.val);
2935                                 lru_cache_add_anon(page);
2936                                 swap_readpage(page, true);
2937                         }
2938                 } else {
2939                         if (swap_use_vma_readahead())
2940                                 page = do_swap_page_readahead(entry,
2941                                         GFP_HIGHUSER_MOVABLE, vmf);
2942                         else
2943                                 page = swapin_readahead(entry,
2944                                        GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2945                         swapcache = page;
2946                 }
2947
2948                 if (!page) {
2949                         /*
2950                          * Back out if somebody else faulted in this pte
2951                          * while we released the pte lock.
2952                          */
2953                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2954                                         vmf->address, &vmf->ptl);
2955                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2956                                 ret = VM_FAULT_OOM;
2957                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2958                         goto unlock;
2959                 }
2960
2961                 /* Had to read the page from swap area: Major fault */
2962                 ret = VM_FAULT_MAJOR;
2963                 count_vm_event(PGMAJFAULT);
2964                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2965         } else if (PageHWPoison(page)) {
2966                 /*
2967                  * hwpoisoned dirty swapcache pages are kept for killing
2968                  * owner processes (which may be unknown at hwpoison time)
2969                  */
2970                 ret = VM_FAULT_HWPOISON;
2971                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2972                 goto out_release;
2973         }
2974
2975         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2976
2977         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2978         if (!locked) {
2979                 ret |= VM_FAULT_RETRY;
2980                 goto out_release;
2981         }
2982
2983         /*
2984          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2985          * release the swapcache from under us.  The page pin, and pte_same
2986          * test below, are not enough to exclude that.  Even if it is still
2987          * swapcache, we need to check that the page's swap has not changed.
2988          */
2989         if (unlikely((!PageSwapCache(page) ||
2990                         page_private(page) != entry.val)) && swapcache)
2991                 goto out_page;
2992
2993         page = ksm_might_need_to_copy(page, vma, vmf->address);
2994         if (unlikely(!page)) {
2995                 ret = VM_FAULT_OOM;
2996                 page = swapcache;
2997                 goto out_page;
2998         }
2999
3000         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
3001                                 &memcg, false)) {
3002                 ret = VM_FAULT_OOM;
3003                 goto out_page;
3004         }
3005
3006         /*
3007          * Back out if somebody else already faulted in this pte.
3008          */
3009         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3010                         &vmf->ptl);
3011         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3012                 goto out_nomap;
3013
3014         if (unlikely(!PageUptodate(page))) {
3015                 ret = VM_FAULT_SIGBUS;
3016                 goto out_nomap;
3017         }
3018
3019         /*
3020          * The page isn't present yet, go ahead with the fault.
3021          *
3022          * Be careful about the sequence of operations here.
3023          * To get its accounting right, reuse_swap_page() must be called
3024          * while the page is counted on swap but not yet in mapcount i.e.
3025          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3026          * must be called after the swap_free(), or it will never succeed.
3027          */
3028
3029         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3030         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3031         pte = mk_pte(page, vma->vm_page_prot);
3032         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3033                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3034                 vmf->flags &= ~FAULT_FLAG_WRITE;
3035                 ret |= VM_FAULT_WRITE;
3036                 exclusive = RMAP_EXCLUSIVE;
3037         }
3038         flush_icache_page(vma, page);
3039         if (pte_swp_soft_dirty(vmf->orig_pte))
3040                 pte = pte_mksoft_dirty(pte);
3041         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3042         arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3043         vmf->orig_pte = pte;
3044
3045         /* ksm created a completely new copy */
3046         if (unlikely(page != swapcache && swapcache)) {
3047                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3048                 mem_cgroup_commit_charge(page, memcg, false, false);
3049                 lru_cache_add_active_or_unevictable(page, vma);
3050         } else {
3051                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3052                 mem_cgroup_commit_charge(page, memcg, true, false);
3053                 activate_page(page);
3054         }
3055
3056         swap_free(entry);
3057         if (mem_cgroup_swap_full(page) ||
3058             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3059                 try_to_free_swap(page);
3060         unlock_page(page);
3061         if (page != swapcache && swapcache) {
3062                 /*
3063                  * Hold the lock to avoid the swap entry to be reused
3064                  * until we take the PT lock for the pte_same() check
3065                  * (to avoid false positives from pte_same). For
3066                  * further safety release the lock after the swap_free
3067                  * so that the swap count won't change under a
3068                  * parallel locked swapcache.
3069                  */
3070                 unlock_page(swapcache);
3071                 put_page(swapcache);
3072         }
3073
3074         if (vmf->flags & FAULT_FLAG_WRITE) {
3075                 ret |= do_wp_page(vmf);
3076                 if (ret & VM_FAULT_ERROR)
3077                         ret &= VM_FAULT_ERROR;
3078                 goto out;
3079         }
3080
3081         /* No need to invalidate - it was non-present before */
3082         update_mmu_cache(vma, vmf->address, vmf->pte);
3083 unlock:
3084         pte_unmap_unlock(vmf->pte, vmf->ptl);
3085 out:
3086         return ret;
3087 out_nomap:
3088         mem_cgroup_cancel_charge(page, memcg, false);
3089         pte_unmap_unlock(vmf->pte, vmf->ptl);
3090 out_page:
3091         unlock_page(page);
3092 out_release:
3093         put_page(page);
3094         if (page != swapcache && swapcache) {
3095                 unlock_page(swapcache);
3096                 put_page(swapcache);
3097         }
3098         return ret;
3099 }
3100
3101 /*
3102  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3103  * but allow concurrent faults), and pte mapped but not yet locked.
3104  * We return with mmap_sem still held, but pte unmapped and unlocked.
3105  */
3106 static int do_anonymous_page(struct vm_fault *vmf)
3107 {
3108         struct vm_area_struct *vma = vmf->vma;
3109         struct mem_cgroup *memcg;
3110         struct page *page;
3111         int ret = 0;
3112         pte_t entry;
3113
3114         /* File mapping without ->vm_ops ? */
3115         if (vma->vm_flags & VM_SHARED)
3116                 return VM_FAULT_SIGBUS;
3117
3118         /*
3119          * Use pte_alloc() instead of pte_alloc_map().  We can't run
3120          * pte_offset_map() on pmds where a huge pmd might be created
3121          * from a different thread.
3122          *
3123          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3124          * parallel threads are excluded by other means.
3125          *
3126          * Here we only have down_read(mmap_sem).
3127          */
3128         if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3129                 return VM_FAULT_OOM;
3130
3131         /* See the comment in pte_alloc_one_map() */
3132         if (unlikely(pmd_trans_unstable(vmf->pmd)))
3133                 return 0;
3134
3135         /* Use the zero-page for reads */
3136         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3137                         !mm_forbids_zeropage(vma->vm_mm)) {
3138                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3139                                                 vma->vm_page_prot));
3140                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3141                                 vmf->address, &vmf->ptl);
3142                 if (!pte_none(*vmf->pte))
3143                         goto unlock;
3144                 ret = check_stable_address_space(vma->vm_mm);
3145                 if (ret)
3146                         goto unlock;
3147                 /* Deliver the page fault to userland, check inside PT lock */
3148                 if (userfaultfd_missing(vma)) {
3149                         pte_unmap_unlock(vmf->pte, vmf->ptl);
3150                         return handle_userfault(vmf, VM_UFFD_MISSING);
3151                 }
3152                 goto setpte;
3153         }
3154
3155         /* Allocate our own private page. */
3156         if (unlikely(anon_vma_prepare(vma)))
3157                 goto oom;
3158         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3159         if (!page)
3160                 goto oom;
3161
3162         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3163                 goto oom_free_page;
3164
3165         /*
3166          * The memory barrier inside __SetPageUptodate makes sure that
3167          * preceeding stores to the page contents become visible before
3168          * the set_pte_at() write.
3169          */
3170         __SetPageUptodate(page);
3171
3172         entry = mk_pte(page, vma->vm_page_prot);
3173         if (vma->vm_flags & VM_WRITE)
3174                 entry = pte_mkwrite(pte_mkdirty(entry));
3175
3176         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3177                         &vmf->ptl);
3178         if (!pte_none(*vmf->pte))
3179                 goto release;
3180
3181         ret = check_stable_address_space(vma->vm_mm);
3182         if (ret)
3183                 goto release;
3184
3185         /* Deliver the page fault to userland, check inside PT lock */
3186         if (userfaultfd_missing(vma)) {
3187                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3188                 mem_cgroup_cancel_charge(page, memcg, false);
3189                 put_page(page);
3190                 return handle_userfault(vmf, VM_UFFD_MISSING);
3191         }
3192
3193         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3194         page_add_new_anon_rmap(page, vma, vmf->address, false);
3195         mem_cgroup_commit_charge(page, memcg, false, false);
3196         lru_cache_add_active_or_unevictable(page, vma);
3197 setpte:
3198         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3199
3200         /* No need to invalidate - it was non-present before */
3201         update_mmu_cache(vma, vmf->address, vmf->pte);
3202 unlock:
3203         pte_unmap_unlock(vmf->pte, vmf->ptl);
3204         return ret;
3205 release:
3206         mem_cgroup_cancel_charge(page, memcg, false);
3207         put_page(page);
3208         goto unlock;
3209 oom_free_page:
3210         put_page(page);
3211 oom:
3212         return VM_FAULT_OOM;
3213 }
3214
3215 /*
3216  * The mmap_sem must have been held on entry, and may have been
3217  * released depending on flags and vma->vm_ops->fault() return value.
3218  * See filemap_fault() and __lock_page_retry().
3219  */
3220 static int __do_fault(struct vm_fault *vmf)
3221 {
3222         struct vm_area_struct *vma = vmf->vma;
3223         int ret;
3224
3225         ret = vma->vm_ops->fault(vmf);
3226         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3227                             VM_FAULT_DONE_COW)))
3228                 return ret;
3229
3230         if (unlikely(PageHWPoison(vmf->page))) {
3231                 if (ret & VM_FAULT_LOCKED)
3232                         unlock_page(vmf->page);
3233                 put_page(vmf->page);
3234                 vmf->page = NULL;
3235                 return VM_FAULT_HWPOISON;
3236         }
3237
3238         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3239                 lock_page(vmf->page);
3240         else
3241                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3242
3243         return ret;
3244 }
3245
3246 /*
3247  * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3248  * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3249  * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3250  * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3251  */
3252 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3253 {
3254         return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3255 }
3256
3257 static int pte_alloc_one_map(struct vm_fault *vmf)
3258 {
3259         struct vm_area_struct *vma = vmf->vma;
3260
3261         if (!pmd_none(*vmf->pmd))
3262                 goto map_pte;
3263         if (vmf->prealloc_pte) {
3264                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3265                 if (unlikely(!pmd_none(*vmf->pmd))) {
3266                         spin_unlock(vmf->ptl);
3267                         goto map_pte;
3268                 }
3269
3270                 mm_inc_nr_ptes(vma->vm_mm);
3271                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3272                 spin_unlock(vmf->ptl);
3273                 vmf->prealloc_pte = NULL;
3274         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3275                 return VM_FAULT_OOM;
3276         }
3277 map_pte:
3278         /*
3279          * If a huge pmd materialized under us just retry later.  Use
3280          * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3281          * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3282          * under us and then back to pmd_none, as a result of MADV_DONTNEED
3283          * running immediately after a huge pmd fault in a different thread of
3284          * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3285          * All we have to ensure is that it is a regular pmd that we can walk
3286          * with pte_offset_map() and we can do that through an atomic read in
3287          * C, which is what pmd_trans_unstable() provides.
3288          */
3289         if (pmd_devmap_trans_unstable(vmf->pmd))
3290                 return VM_FAULT_NOPAGE;
3291
3292         /*
3293          * At this point we know that our vmf->pmd points to a page of ptes
3294          * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3295          * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3296          * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3297          * be valid and we will re-check to make sure the vmf->pte isn't
3298          * pte_none() under vmf->ptl protection when we return to
3299          * alloc_set_pte().
3300          */
3301         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3302                         &vmf->ptl);
3303         return 0;
3304 }
3305
3306 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3307
3308 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3309 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3310                 unsigned long haddr)
3311 {
3312         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3313                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3314                 return false;
3315         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3316                 return false;
3317         return true;
3318 }
3319
3320 static void deposit_prealloc_pte(struct vm_fault *vmf)
3321 {
3322         struct vm_area_struct *vma = vmf->vma;
3323
3324         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3325         /*
3326          * We are going to consume the prealloc table,
3327          * count that as nr_ptes.
3328          */
3329         mm_inc_nr_ptes(vma->vm_mm);
3330         vmf->prealloc_pte = NULL;
3331 }
3332
3333 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3334 {
3335         struct vm_area_struct *vma = vmf->vma;
3336         bool write = vmf->flags & FAULT_FLAG_WRITE;
3337         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3338         pmd_t entry;
3339         int i, ret;
3340
3341         if (!transhuge_vma_suitable(vma, haddr))
3342                 return VM_FAULT_FALLBACK;
3343
3344         ret = VM_FAULT_FALLBACK;
3345         page = compound_head(page);
3346
3347         /*
3348          * Archs like ppc64 need additonal space to store information
3349          * related to pte entry. Use the preallocated table for that.
3350          */
3351         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3352                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3353                 if (!vmf->prealloc_pte)
3354                         return VM_FAULT_OOM;
3355                 smp_wmb(); /* See comment in __pte_alloc() */
3356         }
3357
3358         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3359         if (unlikely(!pmd_none(*vmf->pmd)))
3360                 goto out;
3361
3362         for (i = 0; i < HPAGE_PMD_NR; i++)
3363                 flush_icache_page(vma, page + i);
3364
3365         entry = mk_huge_pmd(page, vma->vm_page_prot);
3366         if (write)
3367                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3368
3369         add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3370         page_add_file_rmap(page, true);
3371         /*
3372          * deposit and withdraw with pmd lock held
3373          */
3374         if (arch_needs_pgtable_deposit())
3375                 deposit_prealloc_pte(vmf);
3376
3377         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3378
3379         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3380
3381         /* fault is handled */
3382         ret = 0;
3383         count_vm_event(THP_FILE_MAPPED);
3384 out:
3385         spin_unlock(vmf->ptl);
3386         return ret;
3387 }
3388 #else
3389 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3390 {
3391         BUILD_BUG();
3392         return 0;
3393 }
3394 #endif
3395
3396 /**
3397  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3398  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3399  *
3400  * @vmf: fault environment
3401  * @memcg: memcg to charge page (only for private mappings)
3402  * @page: page to map
3403  *
3404  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3405  * return.
3406  *
3407  * Target users are page handler itself and implementations of
3408  * vm_ops->map_pages.
3409  */
3410 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3411                 struct page *page)
3412 {
3413         struct vm_area_struct *vma = vmf->vma;
3414         bool write = vmf->flags & FAULT_FLAG_WRITE;
3415         pte_t entry;
3416         int ret;
3417
3418         if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3419                         IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3420                 /* THP on COW? */
3421                 VM_BUG_ON_PAGE(memcg, page);
3422
3423                 ret = do_set_pmd(vmf, page);
3424                 if (ret != VM_FAULT_FALLBACK)
3425                         return ret;
3426         }
3427
3428         if (!vmf->pte) {
3429                 ret = pte_alloc_one_map(vmf);
3430                 if (ret)
3431                         return ret;
3432         }
3433
3434         /* Re-check under ptl */
3435         if (unlikely(!pte_none(*vmf->pte)))
3436                 return VM_FAULT_NOPAGE;
3437
3438         flush_icache_page(vma, page);
3439         entry = mk_pte(page, vma->vm_page_prot);
3440         if (write)
3441                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3442         /* copy-on-write page */
3443         if (write && !(vma->vm_flags & VM_SHARED)) {
3444                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3445                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3446                 mem_cgroup_commit_charge(page, memcg, false, false);
3447                 lru_cache_add_active_or_unevictable(page, vma);
3448         } else {
3449                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3450                 page_add_file_rmap(page, false);
3451         }
3452         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3453
3454         /* no need to invalidate: a not-present page won't be cached */
3455         update_mmu_cache(vma, vmf->address, vmf->pte);
3456
3457         return 0;
3458 }
3459
3460
3461 /**
3462  * finish_fault - finish page fault once we have prepared the page to fault
3463  *
3464  * @vmf: structure describing the fault
3465  *
3466  * This function handles all that is needed to finish a page fault once the
3467  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3468  * given page, adds reverse page mapping, handles memcg charges and LRU
3469  * addition. The function returns 0 on success, VM_FAULT_ code in case of
3470  * error.
3471  *
3472  * The function expects the page to be locked and on success it consumes a
3473  * reference of a page being mapped (for the PTE which maps it).
3474  */
3475 int finish_fault(struct vm_fault *vmf)
3476 {
3477         struct page *page;
3478         int ret = 0;
3479
3480         /* Did we COW the page? */
3481         if ((vmf->flags & FAULT_FLAG_WRITE) &&
3482             !(vmf->vma->vm_flags & VM_SHARED))
3483                 page = vmf->cow_page;
3484         else
3485                 page = vmf->page;
3486
3487         /*
3488          * check even for read faults because we might have lost our CoWed
3489          * page
3490          */