Merge branch 'akpm' (patches from Andrew)
[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 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
821                              pte_t pte, bool with_public_device)
822 {
823         unsigned long pfn = pte_pfn(pte);
824
825         if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
826                 if (likely(!pte_special(pte)))
827                         goto check_pfn;
828                 if (vma->vm_ops && vma->vm_ops->find_special_page)
829                         return vma->vm_ops->find_special_page(vma, addr);
830                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
831                         return NULL;
832                 if (is_zero_pfn(pfn))
833                         return NULL;
834
835                 /*
836                  * Device public pages are special pages (they are ZONE_DEVICE
837                  * pages but different from persistent memory). They behave
838                  * allmost like normal pages. The difference is that they are
839                  * not on the lru and thus should never be involve with any-
840                  * thing that involve lru manipulation (mlock, numa balancing,
841                  * ...).
842                  *
843                  * This is why we still want to return NULL for such page from
844                  * vm_normal_page() so that we do not have to special case all
845                  * call site of vm_normal_page().
846                  */
847                 if (likely(pfn <= highest_memmap_pfn)) {
848                         struct page *page = pfn_to_page(pfn);
849
850                         if (is_device_public_page(page)) {
851                                 if (with_public_device)
852                                         return page;
853                                 return NULL;
854                         }
855                 }
856                 print_bad_pte(vma, addr, pte, NULL);
857                 return NULL;
858         }
859
860         /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
861
862         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
863                 if (vma->vm_flags & VM_MIXEDMAP) {
864                         if (!pfn_valid(pfn))
865                                 return NULL;
866                         goto out;
867                 } else {
868                         unsigned long off;
869                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
870                         if (pfn == vma->vm_pgoff + off)
871                                 return NULL;
872                         if (!is_cow_mapping(vma->vm_flags))
873                                 return NULL;
874                 }
875         }
876
877         if (is_zero_pfn(pfn))
878                 return NULL;
879
880 check_pfn:
881         if (unlikely(pfn > highest_memmap_pfn)) {
882                 print_bad_pte(vma, addr, pte, NULL);
883                 return NULL;
884         }
885
886         /*
887          * NOTE! We still have PageReserved() pages in the page tables.
888          * eg. VDSO mappings can cause them to exist.
889          */
890 out:
891         return pfn_to_page(pfn);
892 }
893
894 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
895 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
896                                 pmd_t pmd)
897 {
898         unsigned long pfn = pmd_pfn(pmd);
899
900         /*
901          * There is no pmd_special() but there may be special pmds, e.g.
902          * in a direct-access (dax) mapping, so let's just replicate the
903          * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
904          */
905         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
906                 if (vma->vm_flags & VM_MIXEDMAP) {
907                         if (!pfn_valid(pfn))
908                                 return NULL;
909                         goto out;
910                 } else {
911                         unsigned long off;
912                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
913                         if (pfn == vma->vm_pgoff + off)
914                                 return NULL;
915                         if (!is_cow_mapping(vma->vm_flags))
916                                 return NULL;
917                 }
918         }
919
920         if (is_zero_pfn(pfn))
921                 return NULL;
922         if (unlikely(pfn > highest_memmap_pfn))
923                 return NULL;
924
925         /*
926          * NOTE! We still have PageReserved() pages in the page tables.
927          * eg. VDSO mappings can cause them to exist.
928          */
929 out:
930         return pfn_to_page(pfn);
931 }
932 #endif
933
934 /*
935  * copy one vm_area from one task to the other. Assumes the page tables
936  * already present in the new task to be cleared in the whole range
937  * covered by this vma.
938  */
939
940 static inline unsigned long
941 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
942                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
943                 unsigned long addr, int *rss)
944 {
945         unsigned long vm_flags = vma->vm_flags;
946         pte_t pte = *src_pte;
947         struct page *page;
948
949         /* pte contains position in swap or file, so copy. */
950         if (unlikely(!pte_present(pte))) {
951                 swp_entry_t entry = pte_to_swp_entry(pte);
952
953                 if (likely(!non_swap_entry(entry))) {
954                         if (swap_duplicate(entry) < 0)
955                                 return entry.val;
956
957                         /* make sure dst_mm is on swapoff's mmlist. */
958                         if (unlikely(list_empty(&dst_mm->mmlist))) {
959                                 spin_lock(&mmlist_lock);
960                                 if (list_empty(&dst_mm->mmlist))
961                                         list_add(&dst_mm->mmlist,
962                                                         &src_mm->mmlist);
963                                 spin_unlock(&mmlist_lock);
964                         }
965                         rss[MM_SWAPENTS]++;
966                 } else if (is_migration_entry(entry)) {
967                         page = migration_entry_to_page(entry);
968
969                         rss[mm_counter(page)]++;
970
971                         if (is_write_migration_entry(entry) &&
972                                         is_cow_mapping(vm_flags)) {
973                                 /*
974                                  * COW mappings require pages in both
975                                  * parent and child to be set to read.
976                                  */
977                                 make_migration_entry_read(&entry);
978                                 pte = swp_entry_to_pte(entry);
979                                 if (pte_swp_soft_dirty(*src_pte))
980                                         pte = pte_swp_mksoft_dirty(pte);
981                                 set_pte_at(src_mm, addr, src_pte, pte);
982                         }
983                 } else if (is_device_private_entry(entry)) {
984                         page = device_private_entry_to_page(entry);
985
986                         /*
987                          * Update rss count even for unaddressable pages, as
988                          * they should treated just like normal pages in this
989                          * respect.
990                          *
991                          * We will likely want to have some new rss counters
992                          * for unaddressable pages, at some point. But for now
993                          * keep things as they are.
994                          */
995                         get_page(page);
996                         rss[mm_counter(page)]++;
997                         page_dup_rmap(page, false);
998
999                         /*
1000                          * We do not preserve soft-dirty information, because so
1001                          * far, checkpoint/restore is the only feature that
1002                          * requires that. And checkpoint/restore does not work
1003                          * when a device driver is involved (you cannot easily
1004                          * save and restore device driver state).
1005                          */
1006                         if (is_write_device_private_entry(entry) &&
1007                             is_cow_mapping(vm_flags)) {
1008                                 make_device_private_entry_read(&entry);
1009                                 pte = swp_entry_to_pte(entry);
1010                                 set_pte_at(src_mm, addr, src_pte, pte);
1011                         }
1012                 }
1013                 goto out_set_pte;
1014         }
1015
1016         /*
1017          * If it's a COW mapping, write protect it both
1018          * in the parent and the child
1019          */
1020         if (is_cow_mapping(vm_flags)) {
1021                 ptep_set_wrprotect(src_mm, addr, src_pte);
1022                 pte = pte_wrprotect(pte);
1023         }
1024
1025         /*
1026          * If it's a shared mapping, mark it clean in
1027          * the child
1028          */
1029         if (vm_flags & VM_SHARED)
1030                 pte = pte_mkclean(pte);
1031         pte = pte_mkold(pte);
1032
1033         page = vm_normal_page(vma, addr, pte);
1034         if (page) {
1035                 get_page(page);
1036                 page_dup_rmap(page, false);
1037                 rss[mm_counter(page)]++;
1038         } else if (pte_devmap(pte)) {
1039                 page = pte_page(pte);
1040
1041                 /*
1042                  * Cache coherent device memory behave like regular page and
1043                  * not like persistent memory page. For more informations see
1044                  * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1045                  */
1046                 if (is_device_public_page(page)) {
1047                         get_page(page);
1048                         page_dup_rmap(page, false);
1049                         rss[mm_counter(page)]++;
1050                 }
1051         }
1052
1053 out_set_pte:
1054         set_pte_at(dst_mm, addr, dst_pte, pte);
1055         return 0;
1056 }
1057
1058 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1059                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1060                    unsigned long addr, unsigned long end)
1061 {
1062         pte_t *orig_src_pte, *orig_dst_pte;
1063         pte_t *src_pte, *dst_pte;
1064         spinlock_t *src_ptl, *dst_ptl;
1065         int progress = 0;
1066         int rss[NR_MM_COUNTERS];
1067         swp_entry_t entry = (swp_entry_t){0};
1068
1069 again:
1070         init_rss_vec(rss);
1071
1072         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1073         if (!dst_pte)
1074                 return -ENOMEM;
1075         src_pte = pte_offset_map(src_pmd, addr);
1076         src_ptl = pte_lockptr(src_mm, src_pmd);
1077         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1078         orig_src_pte = src_pte;
1079         orig_dst_pte = dst_pte;
1080         arch_enter_lazy_mmu_mode();
1081
1082         do {
1083                 /*
1084                  * We are holding two locks at this point - either of them
1085                  * could generate latencies in another task on another CPU.
1086                  */
1087                 if (progress >= 32) {
1088                         progress = 0;
1089                         if (need_resched() ||
1090                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1091                                 break;
1092                 }
1093                 if (pte_none(*src_pte)) {
1094                         progress++;
1095                         continue;
1096                 }
1097                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1098                                                         vma, addr, rss);
1099                 if (entry.val)
1100                         break;
1101                 progress += 8;
1102         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1103
1104         arch_leave_lazy_mmu_mode();
1105         spin_unlock(src_ptl);
1106         pte_unmap(orig_src_pte);
1107         add_mm_rss_vec(dst_mm, rss);
1108         pte_unmap_unlock(orig_dst_pte, dst_ptl);
1109         cond_resched();
1110
1111         if (entry.val) {
1112                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1113                         return -ENOMEM;
1114                 progress = 0;
1115         }
1116         if (addr != end)
1117                 goto again;
1118         return 0;
1119 }
1120
1121 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1122                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1123                 unsigned long addr, unsigned long end)
1124 {
1125         pmd_t *src_pmd, *dst_pmd;
1126         unsigned long next;
1127
1128         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1129         if (!dst_pmd)
1130                 return -ENOMEM;
1131         src_pmd = pmd_offset(src_pud, addr);
1132         do {
1133                 next = pmd_addr_end(addr, end);
1134                 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1135                         || pmd_devmap(*src_pmd)) {
1136                         int err;
1137                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1138                         err = copy_huge_pmd(dst_mm, src_mm,
1139                                             dst_pmd, src_pmd, addr, vma);
1140                         if (err == -ENOMEM)
1141                                 return -ENOMEM;
1142                         if (!err)
1143                                 continue;
1144                         /* fall through */
1145                 }
1146                 if (pmd_none_or_clear_bad(src_pmd))
1147                         continue;
1148                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1149                                                 vma, addr, next))
1150                         return -ENOMEM;
1151         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1152         return 0;
1153 }
1154
1155 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1156                 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1157                 unsigned long addr, unsigned long end)
1158 {
1159         pud_t *src_pud, *dst_pud;
1160         unsigned long next;
1161
1162         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1163         if (!dst_pud)
1164                 return -ENOMEM;
1165         src_pud = pud_offset(src_p4d, addr);
1166         do {
1167                 next = pud_addr_end(addr, end);
1168                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1169                         int err;
1170
1171                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1172                         err = copy_huge_pud(dst_mm, src_mm,
1173                                             dst_pud, src_pud, addr, vma);
1174                         if (err == -ENOMEM)
1175                                 return -ENOMEM;
1176                         if (!err)
1177                                 continue;
1178                         /* fall through */
1179                 }
1180                 if (pud_none_or_clear_bad(src_pud))
1181                         continue;
1182                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1183                                                 vma, addr, next))
1184                         return -ENOMEM;
1185         } while (dst_pud++, src_pud++, addr = next, addr != end);
1186         return 0;
1187 }
1188
1189 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1190                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1191                 unsigned long addr, unsigned long end)
1192 {
1193         p4d_t *src_p4d, *dst_p4d;
1194         unsigned long next;
1195
1196         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1197         if (!dst_p4d)
1198                 return -ENOMEM;
1199         src_p4d = p4d_offset(src_pgd, addr);
1200         do {
1201                 next = p4d_addr_end(addr, end);
1202                 if (p4d_none_or_clear_bad(src_p4d))
1203                         continue;
1204                 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1205                                                 vma, addr, next))
1206                         return -ENOMEM;
1207         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1208         return 0;
1209 }
1210
1211 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1212                 struct vm_area_struct *vma)
1213 {
1214         pgd_t *src_pgd, *dst_pgd;
1215         unsigned long next;
1216         unsigned long addr = vma->vm_start;
1217         unsigned long end = vma->vm_end;
1218         unsigned long mmun_start;       /* For mmu_notifiers */
1219         unsigned long mmun_end;         /* For mmu_notifiers */
1220         bool is_cow;
1221         int ret;
1222
1223         /*
1224          * Don't copy ptes where a page fault will fill them correctly.
1225          * Fork becomes much lighter when there are big shared or private
1226          * readonly mappings. The tradeoff is that copy_page_range is more
1227          * efficient than faulting.
1228          */
1229         if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1230                         !vma->anon_vma)
1231                 return 0;
1232
1233         if (is_vm_hugetlb_page(vma))
1234                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1235
1236         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1237                 /*
1238                  * We do not free on error cases below as remove_vma
1239                  * gets called on error from higher level routine
1240                  */
1241                 ret = track_pfn_copy(vma);
1242                 if (ret)
1243                         return ret;
1244         }
1245
1246         /*
1247          * We need to invalidate the secondary MMU mappings only when
1248          * there could be a permission downgrade on the ptes of the
1249          * parent mm. And a permission downgrade will only happen if
1250          * is_cow_mapping() returns true.
1251          */
1252         is_cow = is_cow_mapping(vma->vm_flags);
1253         mmun_start = addr;
1254         mmun_end   = end;
1255         if (is_cow)
1256                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1257                                                     mmun_end);
1258
1259         ret = 0;
1260         dst_pgd = pgd_offset(dst_mm, addr);
1261         src_pgd = pgd_offset(src_mm, addr);
1262         do {
1263                 next = pgd_addr_end(addr, end);
1264                 if (pgd_none_or_clear_bad(src_pgd))
1265                         continue;
1266                 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1267                                             vma, addr, next))) {
1268                         ret = -ENOMEM;
1269                         break;
1270                 }
1271         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1272
1273         if (is_cow)
1274                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1275         return ret;
1276 }
1277
1278 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1279                                 struct vm_area_struct *vma, pmd_t *pmd,
1280                                 unsigned long addr, unsigned long end,
1281                                 struct zap_details *details)
1282 {
1283         struct mm_struct *mm = tlb->mm;
1284         int force_flush = 0;
1285         int rss[NR_MM_COUNTERS];
1286         spinlock_t *ptl;
1287         pte_t *start_pte;
1288         pte_t *pte;
1289         swp_entry_t entry;
1290
1291         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1292 again:
1293         init_rss_vec(rss);
1294         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1295         pte = start_pte;
1296         flush_tlb_batched_pending(mm);
1297         arch_enter_lazy_mmu_mode();
1298         do {
1299                 pte_t ptent = *pte;
1300                 if (pte_none(ptent))
1301                         continue;
1302
1303                 if (pte_present(ptent)) {
1304                         struct page *page;
1305
1306                         page = _vm_normal_page(vma, addr, ptent, true);
1307                         if (unlikely(details) && page) {
1308                                 /*
1309                                  * unmap_shared_mapping_pages() wants to
1310                                  * invalidate cache without truncating:
1311                                  * unmap shared but keep private pages.
1312                                  */
1313                                 if (details->check_mapping &&
1314                                     details->check_mapping != page_rmapping(page))
1315                                         continue;
1316                         }
1317                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1318                                                         tlb->fullmm);
1319                         tlb_remove_tlb_entry(tlb, pte, addr);
1320                         if (unlikely(!page))
1321                                 continue;
1322
1323                         if (!PageAnon(page)) {
1324                                 if (pte_dirty(ptent)) {
1325                                         force_flush = 1;
1326                                         set_page_dirty(page);
1327                                 }
1328                                 if (pte_young(ptent) &&
1329                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1330                                         mark_page_accessed(page);
1331                         }
1332                         rss[mm_counter(page)]--;
1333                         page_remove_rmap(page, false);
1334                         if (unlikely(page_mapcount(page) < 0))
1335                                 print_bad_pte(vma, addr, ptent, page);
1336                         if (unlikely(__tlb_remove_page(tlb, page))) {
1337                                 force_flush = 1;
1338                                 addr += PAGE_SIZE;
1339                                 break;
1340                         }
1341                         continue;
1342                 }
1343
1344                 entry = pte_to_swp_entry(ptent);
1345                 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1346                         struct page *page = device_private_entry_to_page(entry);
1347
1348                         if (unlikely(details && details->check_mapping)) {
1349                                 /*
1350                                  * unmap_shared_mapping_pages() wants to
1351                                  * invalidate cache without truncating:
1352                                  * unmap shared but keep private pages.
1353                                  */
1354                                 if (details->check_mapping !=
1355                                     page_rmapping(page))
1356                                         continue;
1357                         }
1358
1359                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1360                         rss[mm_counter(page)]--;
1361                         page_remove_rmap(page, false);
1362                         put_page(page);
1363                         continue;
1364                 }
1365
1366                 /* If details->check_mapping, we leave swap entries. */
1367                 if (unlikely(details))
1368                         continue;
1369
1370                 entry = pte_to_swp_entry(ptent);
1371                 if (!non_swap_entry(entry))
1372                         rss[MM_SWAPENTS]--;
1373                 else if (is_migration_entry(entry)) {
1374                         struct page *page;
1375
1376                         page = migration_entry_to_page(entry);
1377                         rss[mm_counter(page)]--;
1378                 }
1379                 if (unlikely(!free_swap_and_cache(entry)))
1380                         print_bad_pte(vma, addr, ptent, NULL);
1381                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1382         } while (pte++, addr += PAGE_SIZE, addr != end);
1383
1384         add_mm_rss_vec(mm, rss);
1385         arch_leave_lazy_mmu_mode();
1386
1387         /* Do the actual TLB flush before dropping ptl */
1388         if (force_flush)
1389                 tlb_flush_mmu_tlbonly(tlb);
1390         pte_unmap_unlock(start_pte, ptl);
1391
1392         /*
1393          * If we forced a TLB flush (either due to running out of
1394          * batch buffers or because we needed to flush dirty TLB
1395          * entries before releasing the ptl), free the batched
1396          * memory too. Restart if we didn't do everything.
1397          */
1398         if (force_flush) {
1399                 force_flush = 0;
1400                 tlb_flush_mmu_free(tlb);
1401                 if (addr != end)
1402                         goto again;
1403         }
1404
1405         return addr;
1406 }
1407
1408 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1409                                 struct vm_area_struct *vma, pud_t *pud,
1410                                 unsigned long addr, unsigned long end,
1411                                 struct zap_details *details)
1412 {
1413         pmd_t *pmd;
1414         unsigned long next;
1415
1416         pmd = pmd_offset(pud, addr);
1417         do {
1418                 next = pmd_addr_end(addr, end);
1419                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1420                         if (next - addr != HPAGE_PMD_SIZE) {
1421                                 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1422                                     !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1423                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1424                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1425                                 goto next;
1426                         /* fall through */
1427                 }
1428                 /*
1429                  * Here there can be other concurrent MADV_DONTNEED or
1430                  * trans huge page faults running, and if the pmd is
1431                  * none or trans huge it can change under us. This is
1432                  * because MADV_DONTNEED holds the mmap_sem in read
1433                  * mode.
1434                  */
1435                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1436                         goto next;
1437                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1438 next:
1439                 cond_resched();
1440         } while (pmd++, addr = next, addr != end);
1441
1442         return addr;
1443 }
1444
1445 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1446                                 struct vm_area_struct *vma, p4d_t *p4d,
1447                                 unsigned long addr, unsigned long end,
1448                                 struct zap_details *details)
1449 {
1450         pud_t *pud;
1451         unsigned long next;
1452
1453         pud = pud_offset(p4d, addr);
1454         do {
1455                 next = pud_addr_end(addr, end);
1456                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1457                         if (next - addr != HPAGE_PUD_SIZE) {
1458                                 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1459                                 split_huge_pud(vma, pud, addr);
1460                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1461                                 goto next;
1462                         /* fall through */
1463                 }
1464                 if (pud_none_or_clear_bad(pud))
1465                         continue;
1466                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1467 next:
1468                 cond_resched();
1469         } while (pud++, addr = next, addr != end);
1470
1471         return addr;
1472 }
1473
1474 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1475                                 struct vm_area_struct *vma, pgd_t *pgd,
1476                                 unsigned long addr, unsigned long end,
1477                                 struct zap_details *details)
1478 {
1479         p4d_t *p4d;
1480         unsigned long next;
1481
1482         p4d = p4d_offset(pgd, addr);
1483         do {
1484                 next = p4d_addr_end(addr, end);
1485                 if (p4d_none_or_clear_bad(p4d))
1486                         continue;
1487                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1488         } while (p4d++, addr = next, addr != end);
1489
1490         return addr;
1491 }
1492
1493 void unmap_page_range(struct mmu_gather *tlb,
1494                              struct vm_area_struct *vma,
1495                              unsigned long addr, unsigned long end,
1496                              struct zap_details *details)
1497 {
1498         pgd_t *pgd;
1499         unsigned long next;
1500
1501         BUG_ON(addr >= end);
1502         tlb_start_vma(tlb, vma);
1503         pgd = pgd_offset(vma->vm_mm, addr);
1504         do {
1505                 next = pgd_addr_end(addr, end);
1506                 if (pgd_none_or_clear_bad(pgd))
1507                         continue;
1508                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1509         } while (pgd++, addr = next, addr != end);
1510         tlb_end_vma(tlb, vma);
1511 }
1512
1513
1514 static void unmap_single_vma(struct mmu_gather *tlb,
1515                 struct vm_area_struct *vma, unsigned long start_addr,
1516                 unsigned long end_addr,
1517                 struct zap_details *details)
1518 {
1519         unsigned long start = max(vma->vm_start, start_addr);
1520         unsigned long end;
1521
1522         if (start >= vma->vm_end)
1523                 return;
1524         end = min(vma->vm_end, end_addr);
1525         if (end <= vma->vm_start)
1526                 return;
1527
1528         if (vma->vm_file)
1529                 uprobe_munmap(vma, start, end);
1530
1531         if (unlikely(vma->vm_flags & VM_PFNMAP))
1532                 untrack_pfn(vma, 0, 0);
1533
1534         if (start != end) {
1535                 if (unlikely(is_vm_hugetlb_page(vma))) {
1536                         /*
1537                          * It is undesirable to test vma->vm_file as it
1538                          * should be non-null for valid hugetlb area.
1539                          * However, vm_file will be NULL in the error
1540                          * cleanup path of mmap_region. When
1541                          * hugetlbfs ->mmap method fails,
1542                          * mmap_region() nullifies vma->vm_file
1543                          * before calling this function to clean up.
1544                          * Since no pte has actually been setup, it is
1545                          * safe to do nothing in this case.
1546                          */
1547                         if (vma->vm_file) {
1548                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1549                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1550                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1551                         }
1552                 } else
1553                         unmap_page_range(tlb, vma, start, end, details);
1554         }
1555 }
1556
1557 /**
1558  * unmap_vmas - unmap a range of memory covered by a list of vma's
1559  * @tlb: address of the caller's struct mmu_gather
1560  * @vma: the starting vma
1561  * @start_addr: virtual address at which to start unmapping
1562  * @end_addr: virtual address at which to end unmapping
1563  *
1564  * Unmap all pages in the vma list.
1565  *
1566  * Only addresses between `start' and `end' will be unmapped.
1567  *
1568  * The VMA list must be sorted in ascending virtual address order.
1569  *
1570  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1571  * range after unmap_vmas() returns.  So the only responsibility here is to
1572  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1573  * drops the lock and schedules.
1574  */
1575 void unmap_vmas(struct mmu_gather *tlb,
1576                 struct vm_area_struct *vma, unsigned long start_addr,
1577                 unsigned long end_addr)
1578 {
1579         struct mm_struct *mm = vma->vm_mm;
1580
1581         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1582         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1583                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1584         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1585 }
1586
1587 /**
1588  * zap_page_range - remove user pages in a given range
1589  * @vma: vm_area_struct holding the applicable pages
1590  * @start: starting address of pages to zap
1591  * @size: number of bytes to zap
1592  *
1593  * Caller must protect the VMA list
1594  */
1595 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1596                 unsigned long size)
1597 {
1598         struct mm_struct *mm = vma->vm_mm;
1599         struct mmu_gather tlb;
1600         unsigned long end = start + size;
1601
1602         lru_add_drain();
1603         tlb_gather_mmu(&tlb, mm, start, end);
1604         update_hiwater_rss(mm);
1605         mmu_notifier_invalidate_range_start(mm, start, end);
1606         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1607                 unmap_single_vma(&tlb, vma, start, end, NULL);
1608
1609                 /*
1610                  * zap_page_range does not specify whether mmap_sem should be
1611                  * held for read or write. That allows parallel zap_page_range
1612                  * operations to unmap a PTE and defer a flush meaning that
1613                  * this call observes pte_none and fails to flush the TLB.
1614                  * Rather than adding a complex API, ensure that no stale
1615                  * TLB entries exist when this call returns.
1616                  */
1617                 flush_tlb_range(vma, start, end);
1618         }
1619
1620         mmu_notifier_invalidate_range_end(mm, start, end);
1621         tlb_finish_mmu(&tlb, start, end);
1622 }
1623
1624 /**
1625  * zap_page_range_single - remove user pages in a given range
1626  * @vma: vm_area_struct holding the applicable pages
1627  * @address: starting address of pages to zap
1628  * @size: number of bytes to zap
1629  * @details: details of shared cache invalidation
1630  *
1631  * The range must fit into one VMA.
1632  */
1633 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1634                 unsigned long size, struct zap_details *details)
1635 {
1636         struct mm_struct *mm = vma->vm_mm;
1637         struct mmu_gather tlb;
1638         unsigned long end = address + size;
1639
1640         lru_add_drain();
1641         tlb_gather_mmu(&tlb, mm, address, end);
1642         update_hiwater_rss(mm);
1643         mmu_notifier_invalidate_range_start(mm, address, end);
1644         unmap_single_vma(&tlb, vma, address, end, details);
1645         mmu_notifier_invalidate_range_end(mm, address, end);
1646         tlb_finish_mmu(&tlb, address, end);
1647 }
1648
1649 /**
1650  * zap_vma_ptes - remove ptes mapping the vma
1651  * @vma: vm_area_struct holding ptes to be zapped
1652  * @address: starting address of pages to zap
1653  * @size: number of bytes to zap
1654  *
1655  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1656  *
1657  * The entire address range must be fully contained within the vma.
1658  *
1659  */
1660 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1661                 unsigned long size)
1662 {
1663         if (address < vma->vm_start || address + size > vma->vm_end ||
1664                         !(vma->vm_flags & VM_PFNMAP))
1665                 return;
1666
1667         zap_page_range_single(vma, address, size, NULL);
1668 }
1669 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1670
1671 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1672                         spinlock_t **ptl)
1673 {
1674         pgd_t *pgd;
1675         p4d_t *p4d;
1676         pud_t *pud;
1677         pmd_t *pmd;
1678
1679         pgd = pgd_offset(mm, addr);
1680         p4d = p4d_alloc(mm, pgd, addr);
1681         if (!p4d)
1682                 return NULL;
1683         pud = pud_alloc(mm, p4d, addr);
1684         if (!pud)
1685                 return NULL;
1686         pmd = pmd_alloc(mm, pud, addr);
1687         if (!pmd)
1688                 return NULL;
1689
1690         VM_BUG_ON(pmd_trans_huge(*pmd));
1691         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1692 }
1693
1694 /*
1695  * This is the old fallback for page remapping.
1696  *
1697  * For historical reasons, it only allows reserved pages. Only
1698  * old drivers should use this, and they needed to mark their
1699  * pages reserved for the old functions anyway.
1700  */
1701 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1702                         struct page *page, pgprot_t prot)
1703 {
1704         struct mm_struct *mm = vma->vm_mm;
1705         int retval;
1706         pte_t *pte;
1707         spinlock_t *ptl;
1708
1709         retval = -EINVAL;
1710         if (PageAnon(page))
1711                 goto out;
1712         retval = -ENOMEM;
1713         flush_dcache_page(page);
1714         pte = get_locked_pte(mm, addr, &ptl);
1715         if (!pte)
1716                 goto out;
1717         retval = -EBUSY;
1718         if (!pte_none(*pte))
1719                 goto out_unlock;
1720
1721         /* Ok, finally just insert the thing.. */
1722         get_page(page);
1723         inc_mm_counter_fast(mm, mm_counter_file(page));
1724         page_add_file_rmap(page, false);
1725         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1726
1727         retval = 0;
1728         pte_unmap_unlock(pte, ptl);
1729         return retval;
1730 out_unlock:
1731         pte_unmap_unlock(pte, ptl);
1732 out:
1733         return retval;
1734 }
1735
1736 /**
1737  * vm_insert_page - insert single page into user vma
1738  * @vma: user vma to map to
1739  * @addr: target user address of this page
1740  * @page: source kernel page
1741  *
1742  * This allows drivers to insert individual pages they've allocated
1743  * into a user vma.
1744  *
1745  * The page has to be a nice clean _individual_ kernel allocation.
1746  * If you allocate a compound page, you need to have marked it as
1747  * such (__GFP_COMP), or manually just split the page up yourself
1748  * (see split_page()).
1749  *
1750  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1751  * took an arbitrary page protection parameter. This doesn't allow
1752  * that. Your vma protection will have to be set up correctly, which
1753  * means that if you want a shared writable mapping, you'd better
1754  * ask for a shared writable mapping!
1755  *
1756  * The page does not need to be reserved.
1757  *
1758  * Usually this function is called from f_op->mmap() handler
1759  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1760  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1761  * function from other places, for example from page-fault handler.
1762  */
1763 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1764                         struct page *page)
1765 {
1766         if (addr < vma->vm_start || addr >= vma->vm_end)
1767                 return -EFAULT;
1768         if (!page_count(page))
1769                 return -EINVAL;
1770         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1771                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1772                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1773                 vma->vm_flags |= VM_MIXEDMAP;
1774         }
1775         return insert_page(vma, addr, page, vma->vm_page_prot);
1776 }
1777 EXPORT_SYMBOL(vm_insert_page);
1778
1779 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1780                         pfn_t pfn, pgprot_t prot, bool mkwrite)
1781 {
1782         struct mm_struct *mm = vma->vm_mm;
1783         int retval;
1784         pte_t *pte, entry;
1785         spinlock_t *ptl;
1786
1787         retval = -ENOMEM;
1788         pte = get_locked_pte(mm, addr, &ptl);
1789         if (!pte)
1790                 goto out;
1791         retval = -EBUSY;
1792         if (!pte_none(*pte)) {
1793                 if (mkwrite) {
1794                         /*
1795                          * For read faults on private mappings the PFN passed
1796                          * in may not match the PFN we have mapped if the
1797                          * mapped PFN is a writeable COW page.  In the mkwrite
1798                          * case we are creating a writable PTE for a shared
1799                          * mapping and we expect the PFNs to match.
1800                          */
1801                         if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1802                                 goto out_unlock;
1803                         entry = *pte;
1804                         goto out_mkwrite;
1805                 } else
1806                         goto out_unlock;
1807         }
1808
1809         /* Ok, finally just insert the thing.. */
1810         if (pfn_t_devmap(pfn))
1811                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1812         else
1813                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1814
1815 out_mkwrite:
1816         if (mkwrite) {
1817                 entry = pte_mkyoung(entry);
1818                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1819         }
1820
1821         set_pte_at(mm, addr, pte, entry);
1822         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1823
1824         retval = 0;
1825 out_unlock:
1826         pte_unmap_unlock(pte, ptl);
1827 out:
1828         return retval;
1829 }
1830
1831 /**
1832  * vm_insert_pfn - insert single pfn into user vma
1833  * @vma: user vma to map to
1834  * @addr: target user address of this page
1835  * @pfn: source kernel pfn
1836  *
1837  * Similar to vm_insert_page, this allows drivers to insert individual pages
1838  * they've allocated into a user vma. Same comments apply.
1839  *
1840  * This function should only be called from a vm_ops->fault handler, and
1841  * in that case the handler should return NULL.
1842  *
1843  * vma cannot be a COW mapping.
1844  *
1845  * As this is called only for pages that do not currently exist, we
1846  * do not need to flush old virtual caches or the TLB.
1847  */
1848 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1849                         unsigned long pfn)
1850 {
1851         return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1852 }
1853 EXPORT_SYMBOL(vm_insert_pfn);
1854
1855 /**
1856  * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1857  * @vma: user vma to map to
1858  * @addr: target user address of this page
1859  * @pfn: source kernel pfn
1860  * @pgprot: pgprot flags for the inserted page
1861  *
1862  * This is exactly like vm_insert_pfn, except that it allows drivers to
1863  * to override pgprot on a per-page basis.
1864  *
1865  * This only makes sense for IO mappings, and it makes no sense for
1866  * cow mappings.  In general, using multiple vmas is preferable;
1867  * vm_insert_pfn_prot should only be used if using multiple VMAs is
1868  * impractical.
1869  */
1870 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1871                         unsigned long pfn, pgprot_t pgprot)
1872 {
1873         int ret;
1874         /*
1875          * Technically, architectures with pte_special can avoid all these
1876          * restrictions (same for remap_pfn_range).  However we would like
1877          * consistency in testing and feature parity among all, so we should
1878          * try to keep these invariants in place for everybody.
1879          */
1880         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1881         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1882                                                 (VM_PFNMAP|VM_MIXEDMAP));
1883         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1884         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1885
1886         if (addr < vma->vm_start || addr >= vma->vm_end)
1887                 return -EFAULT;
1888
1889         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1890
1891         ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1892                         false);
1893
1894         return ret;
1895 }
1896 EXPORT_SYMBOL(vm_insert_pfn_prot);
1897
1898 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1899 {
1900         /* these checks mirror the abort conditions in vm_normal_page */
1901         if (vma->vm_flags & VM_MIXEDMAP)
1902                 return true;
1903         if (pfn_t_devmap(pfn))
1904                 return true;
1905         if (pfn_t_special(pfn))
1906                 return true;
1907         if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1908                 return true;
1909         return false;
1910 }
1911
1912 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1913                         pfn_t pfn, bool mkwrite)
1914 {
1915         pgprot_t pgprot = vma->vm_page_prot;
1916
1917         BUG_ON(!vm_mixed_ok(vma, pfn));
1918
1919         if (addr < vma->vm_start || addr >= vma->vm_end)
1920                 return -EFAULT;
1921
1922         track_pfn_insert(vma, &pgprot, pfn);
1923
1924         /*
1925          * If we don't have pte special, then we have to use the pfn_valid()
1926          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1927          * refcount the page if pfn_valid is true (hence insert_page rather
1928          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1929          * without pte special, it would there be refcounted as a normal page.
1930          */
1931         if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1932             !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1933                 struct page *page;
1934
1935                 /*
1936                  * At this point we are committed to insert_page()
1937                  * regardless of whether the caller specified flags that
1938                  * result in pfn_t_has_page() == false.
1939                  */
1940                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1941                 return insert_page(vma, addr, page, pgprot);
1942         }
1943         return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1944 }
1945
1946 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1947                         pfn_t pfn)
1948 {
1949         return __vm_insert_mixed(vma, addr, pfn, false);
1950
1951 }
1952 EXPORT_SYMBOL(vm_insert_mixed);
1953
1954 /*
1955  *  If the insertion of PTE failed because someone else already added a
1956  *  different entry in the mean time, we treat that as success as we assume
1957  *  the same entry was actually inserted.
1958  */
1959
1960 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1961                 unsigned long addr, pfn_t pfn)
1962 {
1963         int err;
1964
1965         err =  __vm_insert_mixed(vma, addr, pfn, true);
1966         if (err == -ENOMEM)
1967                 return VM_FAULT_OOM;
1968         if (err < 0 && err != -EBUSY)
1969                 return VM_FAULT_SIGBUS;
1970         return VM_FAULT_NOPAGE;
1971 }
1972 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1973
1974 /*
1975  * maps a range of physical memory into the requested pages. the old
1976  * mappings are removed. any references to nonexistent pages results
1977  * in null mappings (currently treated as "copy-on-access")
1978  */
1979 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1980                         unsigned long addr, unsigned long end,
1981                         unsigned long pfn, pgprot_t prot)
1982 {
1983         pte_t *pte;
1984         spinlock_t *ptl;
1985
1986         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1987         if (!pte)
1988                 return -ENOMEM;
1989         arch_enter_lazy_mmu_mode();
1990         do {
1991                 BUG_ON(!pte_none(*pte));
1992                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1993                 pfn++;
1994         } while (pte++, addr += PAGE_SIZE, addr != end);
1995         arch_leave_lazy_mmu_mode();
1996         pte_unmap_unlock(pte - 1, ptl);
1997         return 0;
1998 }
1999
2000 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2001                         unsigned long addr, unsigned long end,
2002                         unsigned long pfn, pgprot_t prot)
2003 {
2004         pmd_t *pmd;
2005         unsigned long next;
2006
2007         pfn -= addr >> PAGE_SHIFT;
2008         pmd = pmd_alloc(mm, pud, addr);
2009         if (!pmd)
2010                 return -ENOMEM;
2011         VM_BUG_ON(pmd_trans_huge(*pmd));
2012         do {
2013                 next = pmd_addr_end(addr, end);
2014                 if (remap_pte_range(mm, pmd, addr, next,
2015                                 pfn + (addr >> PAGE_SHIFT), prot))
2016                         return -ENOMEM;
2017         } while (pmd++, addr = next, addr != end);
2018         return 0;
2019 }
2020
2021 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2022                         unsigned long addr, unsigned long end,
2023                         unsigned long pfn, pgprot_t prot)
2024 {
2025         pud_t *pud;
2026         unsigned long next;
2027
2028         pfn -= addr >> PAGE_SHIFT;
2029         pud = pud_alloc(mm, p4d, addr);
2030         if (!pud)
2031                 return -ENOMEM;
2032         do {
2033                 next = pud_addr_end(addr, end);
2034                 if (remap_pmd_range(mm, pud, addr, next,
2035                                 pfn + (addr >> PAGE_SHIFT), prot))
2036                         return -ENOMEM;
2037         } while (pud++, addr = next, addr != end);
2038         return 0;
2039 }
2040
2041 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2042                         unsigned long addr, unsigned long end,
2043                         unsigned long pfn, pgprot_t prot)
2044 {
2045         p4d_t *p4d;
2046         unsigned long next;
2047
2048         pfn -= addr >> PAGE_SHIFT;
2049         p4d = p4d_alloc(mm, pgd, addr);
2050         if (!p4d)
2051                 return -ENOMEM;
2052         do {
2053                 next = p4d_addr_end(addr, end);
2054                 if (remap_pud_range(mm, p4d, addr, next,
2055                                 pfn + (addr >> PAGE_SHIFT), prot))
2056                         return -ENOMEM;
2057         } while (p4d++, addr = next, addr != end);
2058         return 0;
2059 }
2060
2061 /**
2062  * remap_pfn_range - remap kernel memory to userspace
2063  * @vma: user vma to map to
2064  * @addr: target user address to start at
2065  * @pfn: physical address of kernel memory
2066  * @size: size of map area
2067  * @prot: page protection flags for this mapping
2068  *
2069  *  Note: this is only safe if the mm semaphore is held when called.
2070  */
2071 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2072                     unsigned long pfn, unsigned long size, pgprot_t prot)
2073 {
2074         pgd_t *pgd;
2075         unsigned long next;
2076         unsigned long end = addr + PAGE_ALIGN(size);
2077         struct mm_struct *mm = vma->vm_mm;
2078         unsigned long remap_pfn = pfn;
2079         int err;
2080
2081         /*
2082          * Physically remapped pages are special. Tell the
2083          * rest of the world about it:
2084          *   VM_IO tells people not to look at these pages
2085          *      (accesses can have side effects).
2086          *   VM_PFNMAP tells the core MM that the base pages are just
2087          *      raw PFN mappings, and do not have a "struct page" associated
2088          *      with them.
2089          *   VM_DONTEXPAND
2090          *      Disable vma merging and expanding with mremap().
2091          *   VM_DONTDUMP
2092          *      Omit vma from core dump, even when VM_IO turned off.
2093          *
2094          * There's a horrible special case to handle copy-on-write
2095          * behaviour that some programs depend on. We mark the "original"
2096          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2097          * See vm_normal_page() for details.
2098          */
2099         if (is_cow_mapping(vma->vm_flags)) {
2100                 if (addr != vma->vm_start || end != vma->vm_end)
2101                         return -EINVAL;
2102                 vma->vm_pgoff = pfn;
2103         }
2104
2105         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2106         if (err)
2107                 return -EINVAL;
2108
2109         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2110
2111         BUG_ON(addr >= end);
2112         pfn -= addr >> PAGE_SHIFT;
2113         pgd = pgd_offset(mm, addr);
2114         flush_cache_range(vma, addr, end);
2115         do {
2116                 next = pgd_addr_end(addr, end);
2117                 err = remap_p4d_range(mm, pgd, addr, next,
2118                                 pfn + (addr >> PAGE_SHIFT), prot);
2119                 if (err)
2120                         break;
2121         } while (pgd++, addr = next, addr != end);
2122
2123         if (err)
2124                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2125
2126         return err;
2127 }
2128 EXPORT_SYMBOL(remap_pfn_range);
2129
2130 /**
2131  * vm_iomap_memory - remap memory to userspace
2132  * @vma: user vma to map to
2133  * @start: start of area
2134  * @len: size of area
2135  *
2136  * This is a simplified io_remap_pfn_range() for common driver use. The
2137  * driver just needs to give us the physical memory range to be mapped,
2138  * we'll figure out the rest from the vma information.
2139  *
2140  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2141  * whatever write-combining details or similar.
2142  */
2143 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2144 {
2145         unsigned long vm_len, pfn, pages;
2146
2147         /* Check that the physical memory area passed in looks valid */
2148         if (start + len < start)
2149                 return -EINVAL;
2150         /*
2151          * You *really* shouldn't map things that aren't page-aligned,
2152          * but we've historically allowed it because IO memory might
2153          * just have smaller alignment.
2154          */
2155         len += start & ~PAGE_MASK;
2156         pfn = start >> PAGE_SHIFT;
2157         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2158         if (pfn + pages < pfn)
2159                 return -EINVAL;
2160
2161         /* We start the mapping 'vm_pgoff' pages into the area */
2162         if (vma->vm_pgoff > pages)
2163                 return -EINVAL;
2164         pfn += vma->vm_pgoff;
2165         pages -= vma->vm_pgoff;
2166
2167         /* Can we fit all of the mapping? */
2168         vm_len = vma->vm_end - vma->vm_start;
2169         if (vm_len >> PAGE_SHIFT > pages)
2170                 return -EINVAL;
2171
2172         /* Ok, let it rip */
2173         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2174 }
2175 EXPORT_SYMBOL(vm_iomap_memory);
2176
2177 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2178                                      unsigned long addr, unsigned long end,
2179                                      pte_fn_t fn, void *data)
2180 {
2181         pte_t *pte;
2182         int err;
2183         pgtable_t token;
2184         spinlock_t *uninitialized_var(ptl);
2185
2186         pte = (mm == &init_mm) ?
2187                 pte_alloc_kernel(pmd, addr) :
2188                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2189         if (!pte)
2190                 return -ENOMEM;
2191
2192         BUG_ON(pmd_huge(*pmd));
2193
2194         arch_enter_lazy_mmu_mode();
2195
2196         token = pmd_pgtable(*pmd);
2197
2198         do {
2199                 err = fn(pte++, token, addr, data);
2200                 if (err)
2201                         break;
2202         } while (addr += PAGE_SIZE, addr != end);
2203
2204         arch_leave_lazy_mmu_mode();
2205
2206         if (mm != &init_mm)
2207                 pte_unmap_unlock(pte-1, ptl);
2208         return err;
2209 }
2210
2211 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2212                                      unsigned long addr, unsigned long end,
2213                                      pte_fn_t fn, void *data)
2214 {
2215         pmd_t *pmd;
2216         unsigned long next;
2217         int err;
2218
2219         BUG_ON(pud_huge(*pud));
2220
2221         pmd = pmd_alloc(mm, pud, addr);
2222         if (!pmd)
2223                 return -ENOMEM;
2224         do {
2225                 next = pmd_addr_end(addr, end);
2226                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2227                 if (err)
2228                         break;
2229         } while (pmd++, addr = next, addr != end);
2230         return err;
2231 }
2232
2233 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2234                                      unsigned long addr, unsigned long end,
2235                                      pte_fn_t fn, void *data)
2236 {
2237         pud_t *pud;
2238         unsigned long next;
2239         int err;
2240
2241         pud = pud_alloc(mm, p4d, addr);
2242         if (!pud)
2243                 return -ENOMEM;
2244         do {
2245                 next = pud_addr_end(addr, end);
2246                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2247                 if (err)
2248                         break;
2249         } while (pud++, addr = next, addr != end);
2250         return err;
2251 }
2252
2253 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2254                                      unsigned long addr, unsigned long end,
2255                                      pte_fn_t fn, void *data)
2256 {
2257         p4d_t *p4d;
2258         unsigned long next;
2259         int err;
2260
2261         p4d = p4d_alloc(mm, pgd, addr);
2262         if (!p4d)
2263                 return -ENOMEM;
2264         do {
2265                 next = p4d_addr_end(addr, end);
2266                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2267                 if (err)
2268                         break;
2269         } while (p4d++, addr = next, addr != end);
2270         return err;
2271 }
2272
2273 /*
2274  * Scan a region of virtual memory, filling in page tables as necessary
2275  * and calling a provided function on each leaf page table.
2276  */
2277 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2278                         unsigned long size, pte_fn_t fn, void *data)
2279 {
2280         pgd_t *pgd;
2281         unsigned long next;
2282         unsigned long end = addr + size;
2283         int err;
2284
2285         if (WARN_ON(addr >= end))
2286                 return -EINVAL;
2287
2288         pgd = pgd_offset(mm, addr);
2289         do {
2290                 next = pgd_addr_end(addr, end);
2291                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2292                 if (err)
2293                         break;
2294         } while (pgd++, addr = next, addr != end);
2295
2296         return err;
2297 }
2298 EXPORT_SYMBOL_GPL(apply_to_page_range);
2299
2300 /*
2301  * handle_pte_fault chooses page fault handler according to an entry which was
2302  * read non-atomically.  Before making any commitment, on those architectures
2303  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2304  * parts, do_swap_page must check under lock before unmapping the pte and
2305  * proceeding (but do_wp_page is only called after already making such a check;
2306  * and do_anonymous_page can safely check later on).
2307  */
2308 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2309                                 pte_t *page_table, pte_t orig_pte)
2310 {
2311         int same = 1;
2312 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2313         if (sizeof(pte_t) > sizeof(unsigned long)) {
2314                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2315                 spin_lock(ptl);
2316                 same = pte_same(*page_table, orig_pte);
2317                 spin_unlock(ptl);
2318         }
2319 #endif
2320         pte_unmap(page_table);
2321         return same;
2322 }
2323
2324 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2325 {
2326         debug_dma_assert_idle(src);
2327
2328         /*
2329          * If the source page was a PFN mapping, we don't have
2330          * a "struct page" for it. We do a best-effort copy by
2331          * just copying from the original user address. If that
2332          * fails, we just zero-fill it. Live with it.
2333          */
2334         if (unlikely(!src)) {
2335                 void *kaddr = kmap_atomic(dst);
2336                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2337
2338                 /*
2339                  * This really shouldn't fail, because the page is there
2340                  * in the page tables. But it might just be unreadable,
2341                  * in which case we just give up and fill the result with
2342                  * zeroes.
2343                  */
2344                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2345                         clear_page(kaddr);
2346                 kunmap_atomic(kaddr);
2347                 flush_dcache_page(dst);
2348         } else
2349                 copy_user_highpage(dst, src, va, vma);
2350 }
2351
2352 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2353 {
2354         struct file *vm_file = vma->vm_file;
2355
2356         if (vm_file)
2357                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2358
2359         /*
2360          * Special mappings (e.g. VDSO) do not have any file so fake
2361          * a default GFP_KERNEL for them.
2362          */
2363         return GFP_KERNEL;
2364 }
2365
2366 /*
2367  * Notify the address space that the page is about to become writable so that
2368  * it can prohibit this or wait for the page to get into an appropriate state.
2369  *
2370  * We do this without the lock held, so that it can sleep if it needs to.
2371  */
2372 static int do_page_mkwrite(struct vm_fault *vmf)
2373 {
2374         int ret;
2375         struct page *page = vmf->page;
2376         unsigned int old_flags = vmf->flags;
2377
2378         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2379
2380         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2381         /* Restore original flags so that caller is not surprised */
2382         vmf->flags = old_flags;
2383         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2384                 return ret;
2385         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2386                 lock_page(page);
2387                 if (!page->mapping) {
2388                         unlock_page(page);
2389                         return 0; /* retry */
2390                 }
2391                 ret |= VM_FAULT_LOCKED;
2392         } else
2393                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2394         return ret;
2395 }
2396
2397 /*
2398  * Handle dirtying of a page in shared file mapping on a write fault.
2399  *
2400  * The function expects the page to be locked and unlocks it.
2401  */
2402 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2403                                     struct page *page)
2404 {
2405         struct address_space *mapping;
2406         bool dirtied;
2407         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2408
2409         dirtied = set_page_dirty(page);
2410         VM_BUG_ON_PAGE(PageAnon(page), page);
2411         /*
2412          * Take a local copy of the address_space - page.mapping may be zeroed
2413          * by truncate after unlock_page().   The address_space itself remains
2414          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2415          * release semantics to prevent the compiler from undoing this copying.
2416          */
2417         mapping = page_rmapping(page);
2418         unlock_page(page);
2419
2420         if ((dirtied || page_mkwrite) && mapping) {
2421                 /*
2422                  * Some device drivers do not set page.mapping
2423                  * but still dirty their pages
2424                  */
2425                 balance_dirty_pages_ratelimited(mapping);
2426         }
2427
2428         if (!page_mkwrite)
2429                 file_update_time(vma->vm_file);
2430 }
2431
2432 /*
2433  * Handle write page faults for pages that can be reused in the current vma
2434  *
2435  * This can happen either due to the mapping being with the VM_SHARED flag,
2436  * or due to us being the last reference standing to the page. In either
2437  * case, all we need to do here is to mark the page as writable and update
2438  * any related book-keeping.
2439  */
2440 static inline void wp_page_reuse(struct vm_fault *vmf)
2441         __releases(vmf->ptl)
2442 {
2443         struct vm_area_struct *vma = vmf->vma;
2444         struct page *page = vmf->page;
2445         pte_t entry;
2446         /*
2447          * Clear the pages cpupid information as the existing
2448          * information potentially belongs to a now completely
2449          * unrelated process.
2450          */
2451         if (page)
2452                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2453
2454         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2455         entry = pte_mkyoung(vmf->orig_pte);
2456         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2457         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2458                 update_mmu_cache(vma, vmf->address, vmf->pte);
2459         pte_unmap_unlock(vmf->pte, vmf->ptl);
2460 }
2461
2462 /*
2463  * Handle the case of a page which we actually need to copy to a new page.
2464  *
2465  * Called with mmap_sem locked and the old page referenced, but
2466  * without the ptl held.
2467  *
2468  * High level logic flow:
2469  *
2470  * - Allocate a page, copy the content of the old page to the new one.
2471  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2472  * - Take the PTL. If the pte changed, bail out and release the allocated page
2473  * - If the pte is still the way we remember it, update the page table and all
2474  *   relevant references. This includes dropping the reference the page-table
2475  *   held to the old page, as well as updating the rmap.
2476  * - In any case, unlock the PTL and drop the reference we took to the old page.
2477  */
2478 static int wp_page_copy(struct vm_fault *vmf)
2479 {
2480         struct vm_area_struct *vma = vmf->vma;
2481         struct mm_struct *mm = vma->vm_mm;
2482         struct page *old_page = vmf->page;
2483         struct page *new_page = NULL;
2484         pte_t entry;
2485         int page_copied = 0;
2486         const unsigned long mmun_start = vmf->address & PAGE_MASK;
2487         const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2488         struct mem_cgroup *memcg;
2489
2490         if (unlikely(anon_vma_prepare(vma)))
2491                 goto oom;
2492
2493         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2494                 new_page = alloc_zeroed_user_highpage_movable(vma,
2495                                                               vmf->address);
2496                 if (!new_page)
2497                         goto oom;
2498         } else {
2499                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2500                                 vmf->address);
2501                 if (!new_page)
2502                         goto oom;
2503                 cow_user_page(new_page, old_page, vmf->address, vma);
2504         }
2505
2506         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2507                 goto oom_free_new;
2508
2509         __SetPageUptodate(new_page);
2510
2511         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2512
2513         /*
2514          * Re-check the pte - we dropped the lock
2515          */
2516         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2517         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2518                 if (old_page) {
2519                         if (!PageAnon(old_page)) {
2520                                 dec_mm_counter_fast(mm,
2521                                                 mm_counter_file(old_page));
2522                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2523                         }
2524                 } else {
2525                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2526                 }
2527                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2528                 entry = mk_pte(new_page, vma->vm_page_prot);
2529                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2530                 /*
2531                  * Clear the pte entry and flush it first, before updating the
2532                  * pte with the new entry. This will avoid a race condition
2533                  * seen in the presence of one thread doing SMC and another
2534                  * thread doing COW.
2535                  */
2536                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2537                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2538                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2539                 lru_cache_add_active_or_unevictable(new_page, vma);
2540                 /*
2541                  * We call the notify macro here because, when using secondary
2542                  * mmu page tables (such as kvm shadow page tables), we want the
2543                  * new page to be mapped directly into the secondary page table.
2544                  */
2545                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2546                 update_mmu_cache(vma, vmf->address, vmf->pte);
2547                 if (old_page) {
2548                         /*
2549                          * Only after switching the pte to the new page may
2550                          * we remove the mapcount here. Otherwise another
2551                          * process may come and find the rmap count decremented
2552                          * before the pte is switched to the new page, and
2553                          * "reuse" the old page writing into it while our pte
2554                          * here still points into it and can be read by other
2555                          * threads.
2556                          *
2557                          * The critical issue is to order this
2558                          * page_remove_rmap with the ptp_clear_flush above.
2559                          * Those stores are ordered by (if nothing else,)
2560                          * the barrier present in the atomic_add_negative
2561                          * in page_remove_rmap.
2562                          *
2563                          * Then the TLB flush in ptep_clear_flush ensures that
2564                          * no process can access the old page before the
2565                          * decremented mapcount is visible. And the old page
2566                          * cannot be reused until after the decremented
2567                          * mapcount is visible. So transitively, TLBs to
2568                          * old page will be flushed before it can be reused.
2569                          */
2570                         page_remove_rmap(old_page, false);
2571                 }
2572
2573                 /* Free the old page.. */
2574                 new_page = old_page;
2575                 page_copied = 1;
2576         } else {
2577                 mem_cgroup_cancel_charge(new_page, memcg, false);
2578         }
2579
2580         if (new_page)
2581                 put_page(new_page);
2582
2583         pte_unmap_unlock(vmf->pte, vmf->ptl);
2584         /*
2585          * No need to double call mmu_notifier->invalidate_range() callback as
2586          * the above ptep_clear_flush_notify() did already call it.
2587          */
2588         mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2589         if (old_page) {
2590                 /*
2591                  * Don't let another task, with possibly unlocked vma,
2592                  * keep the mlocked page.
2593                  */
2594                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2595                         lock_page(old_page);    /* LRU manipulation */
2596                         if (PageMlocked(old_page))
2597                                 munlock_vma_page(old_page);
2598                         unlock_page(old_page);
2599                 }
2600                 put_page(old_page);
2601         }
2602         return page_copied ? VM_FAULT_WRITE : 0;
2603 oom_free_new:
2604         put_page(new_page);
2605 oom:
2606         if (old_page)
2607                 put_page(old_page);
2608         return VM_FAULT_OOM;
2609 }
2610
2611 /**
2612  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2613  *                        writeable once the page is prepared
2614  *
2615  * @vmf: structure describing the fault
2616  *
2617  * This function handles all that is needed to finish a write page fault in a
2618  * shared mapping due to PTE being read-only once the mapped page is prepared.
2619  * It handles locking of PTE and modifying it. The function returns
2620  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2621  * lock.
2622  *
2623  * The function expects the page to be locked or other protection against
2624  * concurrent faults / writeback (such as DAX radix tree locks).
2625  */
2626 int finish_mkwrite_fault(struct vm_fault *vmf)
2627 {
2628         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2629         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2630                                        &vmf->ptl);
2631         /*
2632          * We might have raced with another page fault while we released the
2633          * pte_offset_map_lock.
2634          */
2635         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2636                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2637                 return VM_FAULT_NOPAGE;
2638         }
2639         wp_page_reuse(vmf);
2640         return 0;
2641 }
2642
2643 /*
2644  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2645  * mapping
2646  */
2647 static int wp_pfn_shared(struct vm_fault *vmf)
2648 {
2649         struct vm_area_struct *vma = vmf->vma;
2650
2651         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2652                 int ret;
2653
2654                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2655                 vmf->flags |= FAULT_FLAG_MKWRITE;
2656                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2657                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2658                         return ret;
2659                 return finish_mkwrite_fault(vmf);
2660         }
2661         wp_page_reuse(vmf);
2662         return VM_FAULT_WRITE;
2663 }
2664
2665 static int wp_page_shared(struct vm_fault *vmf)
2666         __releases(vmf->ptl)
2667 {
2668         struct vm_area_struct *vma = vmf->vma;
2669
2670         get_page(vmf->page);
2671
2672         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2673                 int tmp;
2674
2675                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2676                 tmp = do_page_mkwrite(vmf);
2677                 if (unlikely(!tmp || (tmp &
2678                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2679                         put_page(vmf->page);
2680                         return tmp;
2681                 }
2682                 tmp = finish_mkwrite_fault(vmf);
2683                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2684                         unlock_page(vmf->page);
2685                         put_page(vmf->page);
2686                         return tmp;
2687                 }
2688         } else {
2689                 wp_page_reuse(vmf);
2690                 lock_page(vmf->page);
2691         }
2692         fault_dirty_shared_page(vma, vmf->page);
2693         put_page(vmf->page);
2694
2695         return VM_FAULT_WRITE;
2696 }
2697
2698 /*
2699  * This routine handles present pages, when users try to write
2700  * to a shared page. It is done by copying the page to a new address
2701  * and decrementing the shared-page counter for the old page.
2702  *
2703  * Note that this routine assumes that the protection checks have been
2704  * done by the caller (the low-level page fault routine in most cases).
2705  * Thus we can safely just mark it writable once we've done any necessary
2706  * COW.
2707  *
2708  * We also mark the page dirty at this point even though the page will
2709  * change only once the write actually happens. This avoids a few races,
2710  * and potentially makes it more efficient.
2711  *
2712  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2713  * but allow concurrent faults), with pte both mapped and locked.
2714  * We return with mmap_sem still held, but pte unmapped and unlocked.
2715  */
2716 static int do_wp_page(struct vm_fault *vmf)
2717         __releases(vmf->ptl)
2718 {
2719         struct vm_area_struct *vma = vmf->vma;
2720
2721         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2722         if (!vmf->page) {
2723                 /*
2724                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2725                  * VM_PFNMAP VMA.
2726                  *
2727                  * We should not cow pages in a shared writeable mapping.
2728                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2729                  */
2730                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2731                                      (VM_WRITE|VM_SHARED))
2732                         return wp_pfn_shared(vmf);
2733
2734                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2735                 return wp_page_copy(vmf);
2736         }
2737
2738         /*
2739          * Take out anonymous pages first, anonymous shared vmas are
2740          * not dirty accountable.
2741          */
2742         if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2743                 int total_map_swapcount;
2744                 if (!trylock_page(vmf->page)) {
2745                         get_page(vmf->page);
2746                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2747                         lock_page(vmf->page);
2748                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2749                                         vmf->address, &vmf->ptl);
2750                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2751                                 unlock_page(vmf->page);
2752                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2753                                 put_page(vmf->page);
2754                                 return 0;
2755                         }
2756                         put_page(vmf->page);
2757                 }
2758                 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2759                         if (total_map_swapcount == 1) {
2760                                 /*
2761                                  * The page is all ours. Move it to
2762                                  * our anon_vma so the rmap code will
2763                                  * not search our parent or siblings.
2764                                  * Protected against the rmap code by
2765                                  * the page lock.
2766                                  */
2767                                 page_move_anon_rmap(vmf->page, vma);
2768                         }
2769                         unlock_page(vmf->page);
2770                         wp_page_reuse(vmf);
2771                         return VM_FAULT_WRITE;
2772                 }
2773                 unlock_page(vmf->page);
2774         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2775                                         (VM_WRITE|VM_SHARED))) {
2776                 return wp_page_shared(vmf);
2777         }
2778
2779         /*
2780          * Ok, we need to copy. Oh, well..
2781          */
2782         get_page(vmf->page);
2783
2784         pte_unmap_unlock(vmf->pte, vmf->ptl);
2785         return wp_page_copy(vmf);
2786 }
2787
2788 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2789                 unsigned long start_addr, unsigned long end_addr,
2790                 struct zap_details *details)
2791 {
2792         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2793 }
2794
2795 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2796                                             struct zap_details *details)
2797 {
2798         struct vm_area_struct *vma;
2799         pgoff_t vba, vea, zba, zea;
2800
2801         vma_interval_tree_foreach(vma, root,
2802                         details->first_index, details->last_index) {
2803
2804                 vba = vma->vm_pgoff;
2805                 vea = vba + vma_pages(vma) - 1;
2806                 zba = details->first_index;
2807                 if (zba < vba)
2808                         zba = vba;
2809                 zea = details->last_index;
2810                 if (zea > vea)
2811                         zea = vea;
2812
2813                 unmap_mapping_range_vma(vma,
2814                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2815                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2816                                 details);
2817         }
2818 }
2819
2820 /**
2821  * unmap_mapping_pages() - Unmap pages from processes.
2822  * @mapping: The address space containing pages to be unmapped.
2823  * @start: Index of first page to be unmapped.
2824  * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
2825  * @even_cows: Whether to unmap even private COWed pages.
2826  *
2827  * Unmap the pages in this address space from any userspace process which
2828  * has them mmaped.  Generally, you want to remove COWed pages as well when
2829  * a file is being truncated, but not when invalidating pages from the page
2830  * cache.
2831  */
2832 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2833                 pgoff_t nr, bool even_cows)
2834 {
2835         struct zap_details details = { };
2836
2837         details.check_mapping = even_cows ? NULL : mapping;
2838         details.first_index = start;
2839         details.last_index = start + nr - 1;
2840         if (details.last_index < details.first_index)
2841                 details.last_index = ULONG_MAX;
2842
2843         i_mmap_lock_write(mapping);
2844         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2845                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2846         i_mmap_unlock_write(mapping);
2847 }
2848
2849 /**
2850  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2851  * address_space corresponding to the specified byte range in the underlying
2852  * file.
2853  *
2854  * @mapping: the address space containing mmaps to be unmapped.
2855  * @holebegin: byte in first page to unmap, relative to the start of
2856  * the underlying file.  This will be rounded down to a PAGE_SIZE
2857  * boundary.  Note that this is different from truncate_pagecache(), which
2858  * must keep the partial page.  In contrast, we must get rid of
2859  * partial pages.
2860  * @holelen: size of prospective hole in bytes.  This will be rounded
2861  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2862  * end of the file.
2863  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2864  * but 0 when invalidating pagecache, don't throw away private data.
2865  */
2866 void unmap_mapping_range(struct address_space *mapping,
2867                 loff_t const holebegin, loff_t const holelen, int even_cows)
2868 {
2869         pgoff_t hba = holebegin >> PAGE_SHIFT;
2870         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2871
2872         /* Check for overflow. */
2873         if (sizeof(holelen) > sizeof(hlen)) {
2874                 long long holeend =
2875                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2876                 if (holeend & ~(long long)ULONG_MAX)
2877                         hlen = ULONG_MAX - hba + 1;
2878         }
2879
2880         unmap_mapping_pages(mapping, hba, hlen, even_cows);
2881 }
2882 EXPORT_SYMBOL(unmap_mapping_range);
2883
2884 /*
2885  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2886  * but allow concurrent faults), and pte mapped but not yet locked.
2887  * We return with pte unmapped and unlocked.
2888  *
2889  * We return with the mmap_sem locked or unlocked in the same cases
2890  * as does filemap_fault().
2891  */
2892 int do_swap_page(struct vm_fault *vmf)
2893 {
2894         struct vm_area_struct *vma = vmf->vma;
2895         struct page *page = NULL, *swapcache;
2896         struct mem_cgroup *memcg;
2897         swp_entry_t entry;
2898         pte_t pte;
2899         int locked;
2900         int exclusive = 0;
2901         int ret = 0;
2902
2903         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2904                 goto out;
2905
2906         entry = pte_to_swp_entry(vmf->orig_pte);
2907         if (unlikely(non_swap_entry(entry))) {
2908                 if (is_migration_entry(entry)) {
2909                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2910                                              vmf->address);
2911                 } else if (is_device_private_entry(entry)) {
2912                         /*
2913                          * For un-addressable device memory we call the pgmap
2914                          * fault handler callback. The callback must migrate
2915                          * the page back to some CPU accessible page.
2916                          */
2917                         ret = device_private_entry_fault(vma, vmf->address, entry,
2918                                                  vmf->flags, vmf->pmd);
2919                 } else if (is_hwpoison_entry(entry)) {
2920                         ret = VM_FAULT_HWPOISON;
2921                 } else {
2922                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2923                         ret = VM_FAULT_SIGBUS;
2924                 }
2925                 goto out;
2926         }
2927
2928
2929         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2930         page = lookup_swap_cache(entry, vma, vmf->address);
2931         swapcache = page;
2932
2933         if (!page) {
2934                 struct swap_info_struct *si = swp_swap_info(entry);
2935
2936                 if (si->flags & SWP_SYNCHRONOUS_IO &&
2937                                 __swap_count(si, entry) == 1) {
2938                         /* skip swapcache */
2939                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2940                                                         vmf->address);
2941                         if (page) {
2942                                 __SetPageLocked(page);
2943                                 __SetPageSwapBacked(page);
2944                                 set_page_private(page, entry.val);
2945                                 lru_cache_add_anon(page);
2946                                 swap_readpage(page, true);
2947                         }
2948                 } else {
2949                         page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2950                                                 vmf);
2951                         swapcache = page;
2952                 }
2953
2954                 if (!page) {
2955                         /*
2956                          * Back out if somebody else faulted in this pte
2957                          * while we released the pte lock.
2958                          */
2959                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2960                                         vmf->address, &vmf->ptl);
2961                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2962                                 ret = VM_FAULT_OOM;
2963                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2964                         goto unlock;
2965                 }
2966
2967                 /* Had to read the page from swap area: Major fault */
2968                 ret = VM_FAULT_MAJOR;
2969                 count_vm_event(PGMAJFAULT);
2970                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2971         } else if (PageHWPoison(page)) {
2972                 /*
2973                  * hwpoisoned dirty swapcache pages are kept for killing
2974                  * owner processes (which may be unknown at hwpoison time)
2975                  */
2976                 ret = VM_FAULT_HWPOISON;
2977                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2978                 goto out_release;
2979         }
2980
2981         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2982
2983         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2984         if (!locked) {
2985                 ret |= VM_FAULT_RETRY;
2986                 goto out_release;
2987         }
2988
2989         /*
2990          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2991          * release the swapcache from under us.  The page pin, and pte_same
2992          * test below, are not enough to exclude that.  Even if it is still
2993          * swapcache, we need to check that the page's swap has not changed.
2994          */
2995         if (unlikely((!PageSwapCache(page) ||
2996                         page_private(page) != entry.val)) && swapcache)
2997                 goto out_page;
2998
2999         page = ksm_might_need_to_copy(page, vma, vmf->address);
3000         if (unlikely(!page)) {
3001                 ret = VM_FAULT_OOM;
3002                 page = swapcache;
3003                 goto out_page;
3004         }
3005
3006         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
3007                                 &memcg, false)) {
3008                 ret = VM_FAULT_OOM;
3009                 goto out_page;
3010         }
3011
3012         /*
3013          * Back out if somebody else already faulted in this pte.
3014          */
3015         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3016                         &vmf->ptl);
3017         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3018                 goto out_nomap;
3019
3020         if (unlikely(!PageUptodate(page))) {
3021                 ret = VM_FAULT_SIGBUS;
3022                 goto out_nomap;
3023         }
3024
3025         /*
3026          * The page isn't present yet, go ahead with the fault.
3027          *
3028          * Be careful about the sequence of operations here.
3029          * To get its accounting right, reuse_swap_page() must be called
3030          * while the page is counted on swap but not yet in mapcount i.e.
3031          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3032          * must be called after the swap_free(), or it will never succeed.
3033          */
3034
3035         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3036         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3037         pte = mk_pte(page, vma->vm_page_prot);
3038         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3039                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3040                 vmf->flags &= ~FAULT_FLAG_WRITE;
3041                 ret |= VM_FAULT_WRITE;
3042                 exclusive = RMAP_EXCLUSIVE;
3043         }
3044         flush_icache_page(vma, page);
3045         if (pte_swp_soft_dirty(vmf->orig_pte))
3046                 pte = pte_mksoft_dirty(pte);
3047         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3048         arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3049         vmf->orig_pte = pte;
3050
3051         /* ksm created a completely new copy */
3052         if (unlikely(page != swapcache && swapcache)) {
3053                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3054                 mem_cgroup_commit_charge(page, memcg, false, false);
3055                 lru_cache_add_active_or_unevictable(page, vma);
3056         } else {
3057                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3058                 mem_cgroup_commit_charge(page, memcg, true, false);
3059                 activate_page(page);
3060         }
3061
3062         swap_free(entry);
3063         if (mem_cgroup_swap_full(page) ||
3064             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3065                 try_to_free_swap(page);
3066         unlock_page(page);
3067         if (page != swapcache && swapcache) {
3068                 /*
3069                  * Hold the lock to avoid the swap entry to be reused
3070                  * until we take the PT lock for the pte_same() check
3071                  * (to avoid false positives from pte_same). For
3072                  * further safety release the lock after the swap_free
3073                  * so that the swap count won't change under a
3074                  * parallel locked swapcache.
3075                  */
3076                 unlock_page(swapcache);
3077                 put_page(swapcache);
3078         }
3079
3080         if (vmf->flags & FAULT_FLAG_WRITE) {
3081                 ret |= do_wp_page(vmf);
3082                 if (ret & VM_FAULT_ERROR)
3083                         ret &= VM_FAULT_ERROR;
3084                 goto out;
3085         }
3086
3087         /* No need to invalidate - it was non-present before */
3088         update_mmu_cache(vma, vmf->address, vmf->pte);
3089 unlock:
3090         pte_unmap_unlock(vmf->pte, vmf->ptl);
3091 out:
3092         return ret;
3093 out_nomap:
3094         mem_cgroup_cancel_charge(page, memcg, false);
3095         pte_unmap_unlock(vmf->pte, vmf->ptl);
3096 out_page:
3097         unlock_page(page);
3098 out_release:
3099         put_page(page);
3100         if (page != swapcache && swapcache) {
3101                 unlock_page(swapcache);
3102                 put_page(swapcache);
3103         }
3104         return ret;
3105 }
3106
3107 /*
3108  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3109  * but allow concurrent faults), and pte mapped but not yet locked.
3110  * We return with mmap_sem still held, but pte unmapped and unlocked.
3111  */
3112 static int do_anonymous_page(struct vm_fault *vmf)
3113 {
3114         struct vm_area_struct *vma = vmf->vma;
3115         struct mem_cgroup *memcg;
3116         struct page *page;
3117         int ret = 0;
3118         pte_t entry;
3119
3120         /* File mapping without ->vm_ops ? */
3121         if (vma->vm_flags & VM_SHARED)
3122                 return VM_FAULT_SIGBUS;
3123
3124         /*
3125          * Use pte_alloc() instead of pte_alloc_map().  We can't run
3126          * pte_offset_map() on pmds where a huge pmd might be created
3127          * from a different thread.
3128          *
3129          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3130          * parallel threads are excluded by other means.
3131          *
3132          * Here we only have down_read(mmap_sem).
3133          */
3134         if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3135                 return VM_FAULT_OOM;
3136
3137         /* See the comment in pte_alloc_one_map() */
3138         if (unlikely(pmd_trans_unstable(vmf->pmd)))
3139                 return 0;
3140
3141         /* Use the zero-page for reads */
3142         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3143                         !mm_forbids_zeropage(vma->vm_mm)) {
3144                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3145                                                 vma->vm_page_prot));
3146                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3147                                 vmf->address, &vmf->ptl);
3148                 if (!pte_none(*vmf->pte))
3149                         goto unlock;
3150                 ret = check_stable_address_space(vma->vm_mm);
3151                 if (ret)
3152                         goto unlock;
3153                 /* Deliver the page fault to userland, check inside PT lock */
3154                 if (userfaultfd_missing(vma)) {
3155                         pte_unmap_unlock(vmf->pte, vmf->ptl);
3156                         return handle_userfault(vmf, VM_UFFD_MISSING);
3157                 }
3158                 goto setpte;
3159         }
3160
3161         /* Allocate our own private page. */
3162         if (unlikely(anon_vma_prepare(vma)))
3163                 goto oom;
3164         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3165         if (!page)
3166                 goto oom;
3167
3168         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3169                 goto oom_free_page;
3170
3171         /*
3172          * The memory barrier inside __SetPageUptodate makes sure that
3173          * preceeding stores to the page contents become visible before
3174          * the set_pte_at() write.
3175          */
3176         __SetPageUptodate(page);
3177
3178         entry = mk_pte(page, vma->vm_page_prot);
3179         if (vma->vm_flags & VM_WRITE)
3180                 entry = pte_mkwrite(pte_mkdirty(entry));
3181
3182         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3183                         &vmf->ptl);
3184         if (!pte_none(*vmf->pte))
3185                 goto release;
3186
3187         ret = check_stable_address_space(vma->vm_mm);
3188         if (ret)
3189                 goto release;
3190
3191         /* Deliver the page fault to userland, check inside PT lock */
3192         if (userfaultfd_missing(vma)) {
3193                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3194                 mem_cgroup_cancel_charge(page, memcg, false);
3195                 put_page(page);
3196                 return handle_userfault(vmf, VM_UFFD_MISSING);
3197         }
3198
3199         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3200         page_add_new_anon_rmap(page, vma, vmf->address, false);
3201         mem_cgroup_commit_charge(page, memcg, false, false);
3202         lru_cache_add_active_or_unevictable(page, vma);
3203 setpte:
3204         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3205
3206         /* No need to invalidate - it was non-present before */
3207         update_mmu_cache(vma, vmf->address, vmf->pte);
3208 unlock:
3209         pte_unmap_unlock(vmf->pte, vmf->ptl);
3210         return ret;
3211 release:
3212         mem_cgroup_cancel_charge(page, memcg, false);
3213         put_page(page);
3214         goto unlock;
3215 oom_free_page:
3216         put_page(page);
3217 oom:
3218         return VM_FAULT_OOM;
3219 }
3220
3221 /*
3222  * The mmap_sem must have been held on entry, and may have been
3223  * released depending on flags and vma->vm_ops->fault() return value.
3224  * See filemap_fault() and __lock_page_retry().
3225  */
3226 static int __do_fault(struct vm_fault *vmf)
3227 {
3228         struct vm_area_struct *vma = vmf->vma;
3229         int ret;
3230
3231         ret = vma->vm_ops->fault(vmf);
3232         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3233                             VM_FAULT_DONE_COW)))
3234                 return ret;
3235
3236         if (unlikely(PageHWPoison(vmf->page))) {
3237                 if (ret & VM_FAULT_LOCKED)
3238                         unlock_page(vmf->page);
3239                 put_page(vmf->page);
3240                 vmf->page = NULL;
3241                 return VM_FAULT_HWPOISON;
3242         }
3243
3244         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3245                 lock_page(vmf->page);
3246         else
3247                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3248
3249         return ret;
3250 }
3251
3252 /*
3253  * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3254  * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3255  * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3256  * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3257  */
3258 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3259 {
3260         return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3261 }
3262
3263 static int pte_alloc_one_map(struct vm_fault *vmf)
3264 {
3265         struct vm_area_struct *vma = vmf->vma;
3266
3267         if (!pmd_none(*vmf->pmd))
3268                 goto map_pte;
3269         if (vmf->prealloc_pte) {
3270                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3271                 if (unlikely(!pmd_none(*vmf->pmd))) {
3272                         spin_unlock(vmf->ptl);
3273                         goto map_pte;
3274                 }
3275
3276                 mm_inc_nr_ptes(vma->vm_mm);
3277                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3278                 spin_unlock(vmf->ptl);
3279                 vmf->prealloc_pte = NULL;
3280         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3281                 return VM_FAULT_OOM;
3282         }
3283 map_pte:
3284         /*
3285          * If a huge pmd materialized under us just retry later.  Use
3286          * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3287          * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3288          * under us and then back to pmd_none, as a result of MADV_DONTNEED
3289          * running immediately after a huge pmd fault in a different thread of
3290          * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3291          * All we have to ensure is that it is a regular pmd that we can walk
3292          * with pte_offset_map() and we can do that through an atomic read in
3293          * C, which is what pmd_trans_unstable() provides.
3294          */
3295         if (pmd_devmap_trans_unstable(vmf->pmd))
3296                 return VM_FAULT_NOPAGE;
3297
3298         /*
3299          * At this point we know that our vmf->pmd points to a page of ptes
3300          * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3301          * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3302          * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3303          * be valid and we will re-check to make sure the vmf->pte isn't
3304          * pte_none() under vmf->ptl protection when we return to
3305          * alloc_set_pte().
3306          */
3307         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3308                         &vmf->ptl);
3309         return 0;
3310 }
3311
3312 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3313
3314 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3315 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3316                 unsigned long haddr)
3317 {
3318         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3319                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3320                 return false;
3321         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3322                 return false;
3323         return true;
3324 }
3325
3326 static void deposit_prealloc_pte(struct vm_fault *vmf)
3327 {
3328         struct vm_area_struct *vma = vmf->vma;
3329
3330         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3331         /*
3332          * We are going to consume the prealloc table,
3333          * count that as nr_ptes.
3334          */
3335         mm_inc_nr_ptes(vma->vm_mm);
3336         vmf->prealloc_pte = NULL;
3337 }
3338
3339 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3340 {
3341         struct vm_area_struct *vma = vmf->vma;
3342         bool write = vmf->flags & FAULT_FLAG_WRITE;
3343         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3344         pmd_t entry;
3345         int i, ret;
3346
3347         if (!transhuge_vma_suitable(vma, haddr))
3348                 return VM_FAULT_FALLBACK;
3349
3350         ret = VM_FAULT_FALLBACK;
3351         page = compound_head(page);
3352
3353         /*
3354          * Archs like ppc64 need additonal space to store information
3355          * related to pte entry. Use the preallocated table for that.
3356          */
3357         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3358                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3359                 if (!vmf->prealloc_pte)
3360                         return VM_FAULT_OOM;
3361                 smp_wmb(); /* See comment in __pte_alloc() */
3362         }
3363
3364         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3365         if (unlikely(!pmd_none(*vmf->pmd)))
3366                 goto out;
3367
3368         for (i = 0; i < HPAGE_PMD_NR; i++)
3369                 flush_icache_page(vma, page + i);
3370
3371         entry = mk_huge_pmd(page, vma->vm_page_prot);
3372         if (write)
3373                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3374
3375         add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3376         page_add_file_rmap(page, true);
3377         /*
3378          * deposit and withdraw with pmd lock held
3379          */
3380         if (arch_needs_pgtable_deposit())
3381                 deposit_prealloc_pte(vmf);
3382
3383         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3384
3385         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3386
3387         /* fault is handled */
3388         ret = 0;
3389         count_vm_event(THP_FILE_MAPPED);
3390 out:
3391         spin_unlock(vmf->ptl);
3392         return ret;
3393 }
3394 #else
3395 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3396 {
3397         BUILD_BUG();
3398         return 0;
3399 }
3400 #endif
3401
3402 /**
3403  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3404  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3405  *
3406  * @vmf: fault environment
3407  * @memcg: memcg to charge page (only for private mappings)
3408  * @page: page to map
3409  *
3410  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3411  * return.
3412  *
3413  * Target users are page handler itself and implementations of
3414  * vm_ops->map_pages.
3415  */
3416 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3417                 struct page *page)
3418 {
3419         struct vm_area_struct *vma = vmf->vma;
3420         bool write = vmf->flags & FAULT_FLAG_WRITE;
3421         pte_t entry;
3422         int ret;
3423
3424         if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3425                         IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3426                 /* THP on COW? */
3427                 VM_BUG_ON_PAGE(memcg, page);
3428
3429                 ret = do_set_pmd(vmf, page);
3430                 if (ret != VM_FAULT_FALLBACK)
3431                         return ret;
3432         }
3433
3434         if (!vmf->pte) {
3435                 ret = pte_alloc_one_map(vmf);
3436                 if (ret)
3437                         return ret;
3438         }
3439
3440         /* Re-check under ptl */
3441         if (unlikely(!pte_none(*vmf->pte)))
3442                 return VM_FAULT_NOPAGE;
3443
3444         flush_icache_page(vma, page);
3445         entry = mk_pte(page, vma->vm_page_prot);
3446         if (write)
3447                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3448         /* copy-on-write page */
3449         if (write && !(vma->vm_flags & VM_SHARED)) {
3450                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3451                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3452                 mem_cgroup_commit_charge(page, memcg, false, false);
3453                 lru_cache_add_active_or_unevictable(page, vma);
3454         } else {
3455                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3456                 page_add_file_rmap(page, false);
3457         }
3458         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3459
3460         /* no need to invalidate: a not-present page won't be cached */
3461         update_mmu_cache(vma, vmf->address, vmf->pte);
3462
3463         return 0;
3464 }
3465
3466
3467 /**
3468  * finish_fault - finish page fault once we have prepared the page to fault
3469  *
3470  * @vmf: structure describing the fault
3471  *
3472  * This function handles all that is needed to finish a page fault once the
3473  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3474  * given page, adds reverse page mapping, handles memcg charges and LRU
3475  * addition. The function returns 0 on success, VM_FAULT_ code in case of
3476  * error.
3477  *
3478  * The function expects the page to be locked and on success it consumes a
3479  * reference of a page being mapped (for the PTE which maps it).
3480  */
3481 int finish_fault(struct vm_fault *vmf)
3482 {
3483         struct page *page;
3484         int ret = 0;
3485
3486         /* Did we COW the page? */
3487         if ((vmf->flags & FAULT_FLAG_WRITE) &&
3488             !(vmf->vma->vm_flags & VM_SHARED))
3489                 page = vmf->cow_page;
3490         else