7c521a6ec7c685d548b6498edbf6c68eab6c11ca
[sfrench/cifs-2.6.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/kallsyms.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
73
74 #include <asm/io.h>
75 #include <asm/mmu_context.h>
76 #include <asm/pgalloc.h>
77 #include <linux/uaccess.h>
78 #include <asm/tlb.h>
79 #include <asm/tlbflush.h>
80 #include <asm/pgtable.h>
81
82 #include "internal.h"
83
84 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
85 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
86 #endif
87
88 #ifndef CONFIG_NEED_MULTIPLE_NODES
89 /* use the per-pgdat data instead for discontigmem - mbligh */
90 unsigned long max_mapnr;
91 EXPORT_SYMBOL(max_mapnr);
92
93 struct page *mem_map;
94 EXPORT_SYMBOL(mem_map);
95 #endif
96
97 /*
98  * A number of key systems in x86 including ioremap() rely on the assumption
99  * that high_memory defines the upper bound on direct map memory, then end
100  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
101  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
102  * and ZONE_HIGHMEM.
103  */
104 void *high_memory;
105 EXPORT_SYMBOL(high_memory);
106
107 /*
108  * Randomize the address space (stacks, mmaps, brk, etc.).
109  *
110  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
111  *   as ancient (libc5 based) binaries can segfault. )
112  */
113 int randomize_va_space __read_mostly =
114 #ifdef CONFIG_COMPAT_BRK
115                                         1;
116 #else
117                                         2;
118 #endif
119
120 static int __init disable_randmaps(char *s)
121 {
122         randomize_va_space = 0;
123         return 1;
124 }
125 __setup("norandmaps", disable_randmaps);
126
127 unsigned long zero_pfn __read_mostly;
128 EXPORT_SYMBOL(zero_pfn);
129
130 unsigned long highest_memmap_pfn __read_mostly;
131
132 /*
133  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
134  */
135 static int __init init_zero_pfn(void)
136 {
137         zero_pfn = page_to_pfn(ZERO_PAGE(0));
138         return 0;
139 }
140 core_initcall(init_zero_pfn);
141
142
143 #if defined(SPLIT_RSS_COUNTING)
144
145 void sync_mm_rss(struct mm_struct *mm)
146 {
147         int i;
148
149         for (i = 0; i < NR_MM_COUNTERS; i++) {
150                 if (current->rss_stat.count[i]) {
151                         add_mm_counter(mm, i, current->rss_stat.count[i]);
152                         current->rss_stat.count[i] = 0;
153                 }
154         }
155         current->rss_stat.events = 0;
156 }
157
158 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
159 {
160         struct task_struct *task = current;
161
162         if (likely(task->mm == mm))
163                 task->rss_stat.count[member] += val;
164         else
165                 add_mm_counter(mm, member, val);
166 }
167 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
168 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
169
170 /* sync counter once per 64 page faults */
171 #define TASK_RSS_EVENTS_THRESH  (64)
172 static void check_sync_rss_stat(struct task_struct *task)
173 {
174         if (unlikely(task != current))
175                 return;
176         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
177                 sync_mm_rss(task->mm);
178 }
179 #else /* SPLIT_RSS_COUNTING */
180
181 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
182 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
183
184 static void check_sync_rss_stat(struct task_struct *task)
185 {
186 }
187
188 #endif /* SPLIT_RSS_COUNTING */
189
190 #ifdef HAVE_GENERIC_MMU_GATHER
191
192 static bool tlb_next_batch(struct mmu_gather *tlb)
193 {
194         struct mmu_gather_batch *batch;
195
196         batch = tlb->active;
197         if (batch->next) {
198                 tlb->active = batch->next;
199                 return true;
200         }
201
202         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
203                 return false;
204
205         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
206         if (!batch)
207                 return false;
208
209         tlb->batch_count++;
210         batch->next = NULL;
211         batch->nr   = 0;
212         batch->max  = MAX_GATHER_BATCH;
213
214         tlb->active->next = batch;
215         tlb->active = batch;
216
217         return true;
218 }
219
220 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
221                                 unsigned long start, unsigned long end)
222 {
223         tlb->mm = mm;
224
225         /* Is it from 0 to ~0? */
226         tlb->fullmm     = !(start | (end+1));
227         tlb->need_flush_all = 0;
228         tlb->local.next = NULL;
229         tlb->local.nr   = 0;
230         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
231         tlb->active     = &tlb->local;
232         tlb->batch_count = 0;
233
234 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
235         tlb->batch = NULL;
236 #endif
237         tlb->page_size = 0;
238
239         __tlb_reset_range(tlb);
240 }
241
242 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
243 {
244         if (!tlb->end)
245                 return;
246
247         tlb_flush(tlb);
248         mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
249 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
250         tlb_table_flush(tlb);
251 #endif
252         __tlb_reset_range(tlb);
253 }
254
255 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
256 {
257         struct mmu_gather_batch *batch;
258
259         for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
260                 free_pages_and_swap_cache(batch->pages, batch->nr);
261                 batch->nr = 0;
262         }
263         tlb->active = &tlb->local;
264 }
265
266 void tlb_flush_mmu(struct mmu_gather *tlb)
267 {
268         tlb_flush_mmu_tlbonly(tlb);
269         tlb_flush_mmu_free(tlb);
270 }
271
272 /* tlb_finish_mmu
273  *      Called at the end of the shootdown operation to free up any resources
274  *      that were required.
275  */
276 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
277                 unsigned long start, unsigned long end, bool force)
278 {
279         struct mmu_gather_batch *batch, *next;
280
281         if (force)
282                 __tlb_adjust_range(tlb, start, end - start);
283
284         tlb_flush_mmu(tlb);
285
286         /* keep the page table cache within bounds */
287         check_pgt_cache();
288
289         for (batch = tlb->local.next; batch; batch = next) {
290                 next = batch->next;
291                 free_pages((unsigned long)batch, 0);
292         }
293         tlb->local.next = NULL;
294 }
295
296 /* __tlb_remove_page
297  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
298  *      handling the additional races in SMP caused by other CPUs caching valid
299  *      mappings in their TLBs. Returns the number of free page slots left.
300  *      When out of page slots we must call tlb_flush_mmu().
301  *returns true if the caller should flush.
302  */
303 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
304 {
305         struct mmu_gather_batch *batch;
306
307         VM_BUG_ON(!tlb->end);
308         VM_WARN_ON(tlb->page_size != page_size);
309
310         batch = tlb->active;
311         /*
312          * Add the page and check if we are full. If so
313          * force a flush.
314          */
315         batch->pages[batch->nr++] = page;
316         if (batch->nr == batch->max) {
317                 if (!tlb_next_batch(tlb))
318                         return true;
319                 batch = tlb->active;
320         }
321         VM_BUG_ON_PAGE(batch->nr > batch->max, page);
322
323         return false;
324 }
325
326 #endif /* HAVE_GENERIC_MMU_GATHER */
327
328 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
329
330 /*
331  * See the comment near struct mmu_table_batch.
332  */
333
334 static void tlb_remove_table_smp_sync(void *arg)
335 {
336         /* Simply deliver the interrupt */
337 }
338
339 static void tlb_remove_table_one(void *table)
340 {
341         /*
342          * This isn't an RCU grace period and hence the page-tables cannot be
343          * assumed to be actually RCU-freed.
344          *
345          * It is however sufficient for software page-table walkers that rely on
346          * IRQ disabling. See the comment near struct mmu_table_batch.
347          */
348         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
349         __tlb_remove_table(table);
350 }
351
352 static void tlb_remove_table_rcu(struct rcu_head *head)
353 {
354         struct mmu_table_batch *batch;
355         int i;
356
357         batch = container_of(head, struct mmu_table_batch, rcu);
358
359         for (i = 0; i < batch->nr; i++)
360                 __tlb_remove_table(batch->tables[i]);
361
362         free_page((unsigned long)batch);
363 }
364
365 void tlb_table_flush(struct mmu_gather *tlb)
366 {
367         struct mmu_table_batch **batch = &tlb->batch;
368
369         if (*batch) {
370                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
371                 *batch = NULL;
372         }
373 }
374
375 void tlb_remove_table(struct mmu_gather *tlb, void *table)
376 {
377         struct mmu_table_batch **batch = &tlb->batch;
378
379         /*
380          * When there's less then two users of this mm there cannot be a
381          * concurrent page-table walk.
382          */
383         if (atomic_read(&tlb->mm->mm_users) < 2) {
384                 __tlb_remove_table(table);
385                 return;
386         }
387
388         if (*batch == NULL) {
389                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
390                 if (*batch == NULL) {
391                         tlb_remove_table_one(table);
392                         return;
393                 }
394                 (*batch)->nr = 0;
395         }
396         (*batch)->tables[(*batch)->nr++] = table;
397         if ((*batch)->nr == MAX_TABLE_BATCH)
398                 tlb_table_flush(tlb);
399 }
400
401 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
402
403 /* tlb_gather_mmu
404  *      Called to initialize an (on-stack) mmu_gather structure for page-table
405  *      tear-down from @mm. The @fullmm argument is used when @mm is without
406  *      users and we're going to destroy the full address space (exit/execve).
407  */
408 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
409                         unsigned long start, unsigned long end)
410 {
411         arch_tlb_gather_mmu(tlb, mm, start, end);
412         inc_tlb_flush_pending(tlb->mm);
413 }
414
415 void tlb_finish_mmu(struct mmu_gather *tlb,
416                 unsigned long start, unsigned long end)
417 {
418         /*
419          * If there are parallel threads are doing PTE changes on same range
420          * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
421          * flush by batching, a thread has stable TLB entry can fail to flush
422          * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
423          * forcefully if we detect parallel PTE batching threads.
424          */
425         bool force = mm_tlb_flush_nested(tlb->mm);
426
427         arch_tlb_finish_mmu(tlb, start, end, force);
428         dec_tlb_flush_pending(tlb->mm);
429 }
430
431 /*
432  * Note: this doesn't free the actual pages themselves. That
433  * has been handled earlier when unmapping all the memory regions.
434  */
435 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
436                            unsigned long addr)
437 {
438         pgtable_t token = pmd_pgtable(*pmd);
439         pmd_clear(pmd);
440         pte_free_tlb(tlb, token, addr);
441         atomic_long_dec(&tlb->mm->nr_ptes);
442 }
443
444 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
445                                 unsigned long addr, unsigned long end,
446                                 unsigned long floor, unsigned long ceiling)
447 {
448         pmd_t *pmd;
449         unsigned long next;
450         unsigned long start;
451
452         start = addr;
453         pmd = pmd_offset(pud, addr);
454         do {
455                 next = pmd_addr_end(addr, end);
456                 if (pmd_none_or_clear_bad(pmd))
457                         continue;
458                 free_pte_range(tlb, pmd, addr);
459         } while (pmd++, addr = next, addr != end);
460
461         start &= PUD_MASK;
462         if (start < floor)
463                 return;
464         if (ceiling) {
465                 ceiling &= PUD_MASK;
466                 if (!ceiling)
467                         return;
468         }
469         if (end - 1 > ceiling - 1)
470                 return;
471
472         pmd = pmd_offset(pud, start);
473         pud_clear(pud);
474         pmd_free_tlb(tlb, pmd, start);
475         mm_dec_nr_pmds(tlb->mm);
476 }
477
478 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
479                                 unsigned long addr, unsigned long end,
480                                 unsigned long floor, unsigned long ceiling)
481 {
482         pud_t *pud;
483         unsigned long next;
484         unsigned long start;
485
486         start = addr;
487         pud = pud_offset(p4d, addr);
488         do {
489                 next = pud_addr_end(addr, end);
490                 if (pud_none_or_clear_bad(pud))
491                         continue;
492                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
493         } while (pud++, addr = next, addr != end);
494
495         start &= P4D_MASK;
496         if (start < floor)
497                 return;
498         if (ceiling) {
499                 ceiling &= P4D_MASK;
500                 if (!ceiling)
501                         return;
502         }
503         if (end - 1 > ceiling - 1)
504                 return;
505
506         pud = pud_offset(p4d, start);
507         p4d_clear(p4d);
508         pud_free_tlb(tlb, pud, start);
509 }
510
511 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
512                                 unsigned long addr, unsigned long end,
513                                 unsigned long floor, unsigned long ceiling)
514 {
515         p4d_t *p4d;
516         unsigned long next;
517         unsigned long start;
518
519         start = addr;
520         p4d = p4d_offset(pgd, addr);
521         do {
522                 next = p4d_addr_end(addr, end);
523                 if (p4d_none_or_clear_bad(p4d))
524                         continue;
525                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
526         } while (p4d++, addr = next, addr != end);
527
528         start &= PGDIR_MASK;
529         if (start < floor)
530                 return;
531         if (ceiling) {
532                 ceiling &= PGDIR_MASK;
533                 if (!ceiling)
534                         return;
535         }
536         if (end - 1 > ceiling - 1)
537                 return;
538
539         p4d = p4d_offset(pgd, start);
540         pgd_clear(pgd);
541         p4d_free_tlb(tlb, p4d, start);
542 }
543
544 /*
545  * This function frees user-level page tables of a process.
546  */
547 void free_pgd_range(struct mmu_gather *tlb,
548                         unsigned long addr, unsigned long end,
549                         unsigned long floor, unsigned long ceiling)
550 {
551         pgd_t *pgd;
552         unsigned long next;
553
554         /*
555          * The next few lines have given us lots of grief...
556          *
557          * Why are we testing PMD* at this top level?  Because often
558          * there will be no work to do at all, and we'd prefer not to
559          * go all the way down to the bottom just to discover that.
560          *
561          * Why all these "- 1"s?  Because 0 represents both the bottom
562          * of the address space and the top of it (using -1 for the
563          * top wouldn't help much: the masks would do the wrong thing).
564          * The rule is that addr 0 and floor 0 refer to the bottom of
565          * the address space, but end 0 and ceiling 0 refer to the top
566          * Comparisons need to use "end - 1" and "ceiling - 1" (though
567          * that end 0 case should be mythical).
568          *
569          * Wherever addr is brought up or ceiling brought down, we must
570          * be careful to reject "the opposite 0" before it confuses the
571          * subsequent tests.  But what about where end is brought down
572          * by PMD_SIZE below? no, end can't go down to 0 there.
573          *
574          * Whereas we round start (addr) and ceiling down, by different
575          * masks at different levels, in order to test whether a table
576          * now has no other vmas using it, so can be freed, we don't
577          * bother to round floor or end up - the tests don't need that.
578          */
579
580         addr &= PMD_MASK;
581         if (addr < floor) {
582                 addr += PMD_SIZE;
583                 if (!addr)
584                         return;
585         }
586         if (ceiling) {
587                 ceiling &= PMD_MASK;
588                 if (!ceiling)
589                         return;
590         }
591         if (end - 1 > ceiling - 1)
592                 end -= PMD_SIZE;
593         if (addr > end - 1)
594                 return;
595         /*
596          * We add page table cache pages with PAGE_SIZE,
597          * (see pte_free_tlb()), flush the tlb if we need
598          */
599         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
600         pgd = pgd_offset(tlb->mm, addr);
601         do {
602                 next = pgd_addr_end(addr, end);
603                 if (pgd_none_or_clear_bad(pgd))
604                         continue;
605                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
606         } while (pgd++, addr = next, addr != end);
607 }
608
609 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
610                 unsigned long floor, unsigned long ceiling)
611 {
612         while (vma) {
613                 struct vm_area_struct *next = vma->vm_next;
614                 unsigned long addr = vma->vm_start;
615
616                 /*
617                  * Hide vma from rmap and truncate_pagecache before freeing
618                  * pgtables
619                  */
620                 unlink_anon_vmas(vma);
621                 unlink_file_vma(vma);
622
623                 if (is_vm_hugetlb_page(vma)) {
624                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
625                                 floor, next ? next->vm_start : ceiling);
626                 } else {
627                         /*
628                          * Optimization: gather nearby vmas into one call down
629                          */
630                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
631                                && !is_vm_hugetlb_page(next)) {
632                                 vma = next;
633                                 next = vma->vm_next;
634                                 unlink_anon_vmas(vma);
635                                 unlink_file_vma(vma);
636                         }
637                         free_pgd_range(tlb, addr, vma->vm_end,
638                                 floor, next ? next->vm_start : ceiling);
639                 }
640                 vma = next;
641         }
642 }
643
644 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
645 {
646         spinlock_t *ptl;
647         pgtable_t new = pte_alloc_one(mm, address);
648         if (!new)
649                 return -ENOMEM;
650
651         /*
652          * Ensure all pte setup (eg. pte page lock and page clearing) are
653          * visible before the pte is made visible to other CPUs by being
654          * put into page tables.
655          *
656          * The other side of the story is the pointer chasing in the page
657          * table walking code (when walking the page table without locking;
658          * ie. most of the time). Fortunately, these data accesses consist
659          * of a chain of data-dependent loads, meaning most CPUs (alpha
660          * being the notable exception) will already guarantee loads are
661          * seen in-order. See the alpha page table accessors for the
662          * smp_read_barrier_depends() barriers in page table walking code.
663          */
664         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
665
666         ptl = pmd_lock(mm, pmd);
667         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
668                 atomic_long_inc(&mm->nr_ptes);
669                 pmd_populate(mm, pmd, new);
670                 new = NULL;
671         }
672         spin_unlock(ptl);
673         if (new)
674                 pte_free(mm, new);
675         return 0;
676 }
677
678 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
679 {
680         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
681         if (!new)
682                 return -ENOMEM;
683
684         smp_wmb(); /* See comment in __pte_alloc */
685
686         spin_lock(&init_mm.page_table_lock);
687         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
688                 pmd_populate_kernel(&init_mm, pmd, new);
689                 new = NULL;
690         }
691         spin_unlock(&init_mm.page_table_lock);
692         if (new)
693                 pte_free_kernel(&init_mm, new);
694         return 0;
695 }
696
697 static inline void init_rss_vec(int *rss)
698 {
699         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
700 }
701
702 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
703 {
704         int i;
705
706         if (current->mm == mm)
707                 sync_mm_rss(mm);
708         for (i = 0; i < NR_MM_COUNTERS; i++)
709                 if (rss[i])
710                         add_mm_counter(mm, i, rss[i]);
711 }
712
713 /*
714  * This function is called to print an error when a bad pte
715  * is found. For example, we might have a PFN-mapped pte in
716  * a region that doesn't allow it.
717  *
718  * The calling function must still handle the error.
719  */
720 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
721                           pte_t pte, struct page *page)
722 {
723         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
724         p4d_t *p4d = p4d_offset(pgd, addr);
725         pud_t *pud = pud_offset(p4d, addr);
726         pmd_t *pmd = pmd_offset(pud, addr);
727         struct address_space *mapping;
728         pgoff_t index;
729         static unsigned long resume;
730         static unsigned long nr_shown;
731         static unsigned long nr_unshown;
732
733         /*
734          * Allow a burst of 60 reports, then keep quiet for that minute;
735          * or allow a steady drip of one report per second.
736          */
737         if (nr_shown == 60) {
738                 if (time_before(jiffies, resume)) {
739                         nr_unshown++;
740                         return;
741                 }
742                 if (nr_unshown) {
743                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
744                                  nr_unshown);
745                         nr_unshown = 0;
746                 }
747                 nr_shown = 0;
748         }
749         if (nr_shown++ == 0)
750                 resume = jiffies + 60 * HZ;
751
752         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
753         index = linear_page_index(vma, addr);
754
755         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
756                  current->comm,
757                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
758         if (page)
759                 dump_page(page, "bad pte");
760         pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
761                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
762         /*
763          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
764          */
765         pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
766                  vma->vm_file,
767                  vma->vm_ops ? vma->vm_ops->fault : NULL,
768                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
769                  mapping ? mapping->a_ops->readpage : NULL);
770         dump_stack();
771         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
772 }
773
774 /*
775  * vm_normal_page -- This function gets the "struct page" associated with a pte.
776  *
777  * "Special" mappings do not wish to be associated with a "struct page" (either
778  * it doesn't exist, or it exists but they don't want to touch it). In this
779  * case, NULL is returned here. "Normal" mappings do have a struct page.
780  *
781  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
782  * pte bit, in which case this function is trivial. Secondly, an architecture
783  * may not have a spare pte bit, which requires a more complicated scheme,
784  * described below.
785  *
786  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
787  * special mapping (even if there are underlying and valid "struct pages").
788  * COWed pages of a VM_PFNMAP are always normal.
789  *
790  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
791  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
792  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
793  * mapping will always honor the rule
794  *
795  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
796  *
797  * And for normal mappings this is false.
798  *
799  * This restricts such mappings to be a linear translation from virtual address
800  * to pfn. To get around this restriction, we allow arbitrary mappings so long
801  * as the vma is not a COW mapping; in that case, we know that all ptes are
802  * special (because none can have been COWed).
803  *
804  *
805  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
806  *
807  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
808  * page" backing, however the difference is that _all_ pages with a struct
809  * page (that is, those where pfn_valid is true) are refcounted and considered
810  * normal pages by the VM. The disadvantage is that pages are refcounted
811  * (which can be slower and simply not an option for some PFNMAP users). The
812  * advantage is that we don't have to follow the strict linearity rule of
813  * PFNMAP mappings in order to support COWable mappings.
814  *
815  */
816 #ifdef __HAVE_ARCH_PTE_SPECIAL
817 # define HAVE_PTE_SPECIAL 1
818 #else
819 # define HAVE_PTE_SPECIAL 0
820 #endif
821 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
822                              pte_t pte, bool with_public_device)
823 {
824         unsigned long pfn = pte_pfn(pte);
825
826         if (HAVE_PTE_SPECIAL) {
827                 if (likely(!pte_special(pte)))
828                         goto check_pfn;
829                 if (vma->vm_ops && vma->vm_ops->find_special_page)
830                         return vma->vm_ops->find_special_page(vma, addr);
831                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
832                         return NULL;
833                 if (is_zero_pfn(pfn))
834                         return NULL;
835
836                 /*
837                  * Device public pages are special pages (they are ZONE_DEVICE
838                  * pages but different from persistent memory). They behave
839                  * allmost like normal pages. The difference is that they are
840                  * not on the lru and thus should never be involve with any-
841                  * thing that involve lru manipulation (mlock, numa balancing,
842                  * ...).
843                  *
844                  * This is why we still want to return NULL for such page from
845                  * vm_normal_page() so that we do not have to special case all
846                  * call site of vm_normal_page().
847                  */
848                 if (likely(pfn < highest_memmap_pfn)) {
849                         struct page *page = pfn_to_page(pfn);
850
851                         if (is_device_public_page(page)) {
852                                 if (with_public_device)
853                                         return page;
854                                 return NULL;
855                         }
856                 }
857                 print_bad_pte(vma, addr, pte, NULL);
858                 return NULL;
859         }
860
861         /* !HAVE_PTE_SPECIAL case follows: */
862
863         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
864                 if (vma->vm_flags & VM_MIXEDMAP) {
865                         if (!pfn_valid(pfn))
866                                 return NULL;
867                         goto out;
868                 } else {
869                         unsigned long off;
870                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
871                         if (pfn == vma->vm_pgoff + off)
872                                 return NULL;
873                         if (!is_cow_mapping(vma->vm_flags))
874                                 return NULL;
875                 }
876         }
877
878         if (is_zero_pfn(pfn))
879                 return NULL;
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          * !HAVE_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  * Returns 0 if successful.
1660  */
1661 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1662                 unsigned long size)
1663 {
1664         if (address < vma->vm_start || address + size > vma->vm_end ||
1665                         !(vma->vm_flags & VM_PFNMAP))
1666                 return -1;
1667         zap_page_range_single(vma, address, size, NULL);
1668         return 0;
1669 }
1670 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1671
1672 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1673                         spinlock_t **ptl)
1674 {
1675         pgd_t *pgd;
1676         p4d_t *p4d;
1677         pud_t *pud;
1678         pmd_t *pmd;
1679
1680         pgd = pgd_offset(mm, addr);
1681         p4d = p4d_alloc(mm, pgd, addr);
1682         if (!p4d)
1683                 return NULL;
1684         pud = pud_alloc(mm, p4d, addr);
1685         if (!pud)
1686                 return NULL;
1687         pmd = pmd_alloc(mm, pud, addr);
1688         if (!pmd)
1689                 return NULL;
1690
1691         VM_BUG_ON(pmd_trans_huge(*pmd));
1692         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1693 }
1694
1695 /*
1696  * This is the old fallback for page remapping.
1697  *
1698  * For historical reasons, it only allows reserved pages. Only
1699  * old drivers should use this, and they needed to mark their
1700  * pages reserved for the old functions anyway.
1701  */
1702 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1703                         struct page *page, pgprot_t prot)
1704 {
1705         struct mm_struct *mm = vma->vm_mm;
1706         int retval;
1707         pte_t *pte;
1708         spinlock_t *ptl;
1709
1710         retval = -EINVAL;
1711         if (PageAnon(page))
1712                 goto out;
1713         retval = -ENOMEM;
1714         flush_dcache_page(page);
1715         pte = get_locked_pte(mm, addr, &ptl);
1716         if (!pte)
1717                 goto out;
1718         retval = -EBUSY;
1719         if (!pte_none(*pte))
1720                 goto out_unlock;
1721
1722         /* Ok, finally just insert the thing.. */
1723         get_page(page);
1724         inc_mm_counter_fast(mm, mm_counter_file(page));
1725         page_add_file_rmap(page, false);
1726         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1727
1728         retval = 0;
1729         pte_unmap_unlock(pte, ptl);
1730         return retval;
1731 out_unlock:
1732         pte_unmap_unlock(pte, ptl);
1733 out:
1734         return retval;
1735 }
1736
1737 /**
1738  * vm_insert_page - insert single page into user vma
1739  * @vma: user vma to map to
1740  * @addr: target user address of this page
1741  * @page: source kernel page
1742  *
1743  * This allows drivers to insert individual pages they've allocated
1744  * into a user vma.
1745  *
1746  * The page has to be a nice clean _individual_ kernel allocation.
1747  * If you allocate a compound page, you need to have marked it as
1748  * such (__GFP_COMP), or manually just split the page up yourself
1749  * (see split_page()).
1750  *
1751  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1752  * took an arbitrary page protection parameter. This doesn't allow
1753  * that. Your vma protection will have to be set up correctly, which
1754  * means that if you want a shared writable mapping, you'd better
1755  * ask for a shared writable mapping!
1756  *
1757  * The page does not need to be reserved.
1758  *
1759  * Usually this function is called from f_op->mmap() handler
1760  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1761  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1762  * function from other places, for example from page-fault handler.
1763  */
1764 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1765                         struct page *page)
1766 {
1767         if (addr < vma->vm_start || addr >= vma->vm_end)
1768                 return -EFAULT;
1769         if (!page_count(page))
1770                 return -EINVAL;
1771         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1772                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1773                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1774                 vma->vm_flags |= VM_MIXEDMAP;
1775         }
1776         return insert_page(vma, addr, page, vma->vm_page_prot);
1777 }
1778 EXPORT_SYMBOL(vm_insert_page);
1779
1780 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1781                         pfn_t pfn, pgprot_t prot, bool mkwrite)
1782 {
1783         struct mm_struct *mm = vma->vm_mm;
1784         int retval;
1785         pte_t *pte, entry;
1786         spinlock_t *ptl;
1787
1788         retval = -ENOMEM;
1789         pte = get_locked_pte(mm, addr, &ptl);
1790         if (!pte)
1791                 goto out;
1792         retval = -EBUSY;
1793         if (!pte_none(*pte)) {
1794                 if (mkwrite) {
1795                         /*
1796                          * For read faults on private mappings the PFN passed
1797                          * in may not match the PFN we have mapped if the
1798                          * mapped PFN is a writeable COW page.  In the mkwrite
1799                          * case we are creating a writable PTE for a shared
1800                          * mapping and we expect the PFNs to match.
1801                          */
1802                         if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1803                                 goto out_unlock;
1804                         entry = *pte;
1805                         goto out_mkwrite;
1806                 } else
1807                         goto out_unlock;
1808         }
1809
1810         /* Ok, finally just insert the thing.. */
1811         if (pfn_t_devmap(pfn))
1812                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1813         else
1814                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1815
1816 out_mkwrite:
1817         if (mkwrite) {
1818                 entry = pte_mkyoung(entry);
1819                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1820         }
1821
1822         set_pte_at(mm, addr, pte, entry);
1823         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1824
1825         retval = 0;
1826 out_unlock:
1827         pte_unmap_unlock(pte, ptl);
1828 out:
1829         return retval;
1830 }
1831
1832 /**
1833  * vm_insert_pfn - insert single pfn into user vma
1834  * @vma: user vma to map to
1835  * @addr: target user address of this page
1836  * @pfn: source kernel pfn
1837  *
1838  * Similar to vm_insert_page, this allows drivers to insert individual pages
1839  * they've allocated into a user vma. Same comments apply.
1840  *
1841  * This function should only be called from a vm_ops->fault handler, and
1842  * in that case the handler should return NULL.
1843  *
1844  * vma cannot be a COW mapping.
1845  *
1846  * As this is called only for pages that do not currently exist, we
1847  * do not need to flush old virtual caches or the TLB.
1848  */
1849 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1850                         unsigned long pfn)
1851 {
1852         return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1853 }
1854 EXPORT_SYMBOL(vm_insert_pfn);
1855
1856 /**
1857  * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1858  * @vma: user vma to map to
1859  * @addr: target user address of this page
1860  * @pfn: source kernel pfn
1861  * @pgprot: pgprot flags for the inserted page
1862  *
1863  * This is exactly like vm_insert_pfn, except that it allows drivers to
1864  * to override pgprot on a per-page basis.
1865  *
1866  * This only makes sense for IO mappings, and it makes no sense for
1867  * cow mappings.  In general, using multiple vmas is preferable;
1868  * vm_insert_pfn_prot should only be used if using multiple VMAs is
1869  * impractical.
1870  */
1871 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1872                         unsigned long pfn, pgprot_t pgprot)
1873 {
1874         int ret;
1875         /*
1876          * Technically, architectures with pte_special can avoid all these
1877          * restrictions (same for remap_pfn_range).  However we would like
1878          * consistency in testing and feature parity among all, so we should
1879          * try to keep these invariants in place for everybody.
1880          */
1881         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1882         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1883                                                 (VM_PFNMAP|VM_MIXEDMAP));
1884         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1885         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1886
1887         if (addr < vma->vm_start || addr >= vma->vm_end)
1888                 return -EFAULT;
1889
1890         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1891
1892         ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1893                         false);
1894
1895         return ret;
1896 }
1897 EXPORT_SYMBOL(vm_insert_pfn_prot);
1898
1899 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1900                         pfn_t pfn, bool mkwrite)
1901 {
1902         pgprot_t pgprot = vma->vm_page_prot;
1903
1904         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1905
1906         if (addr < vma->vm_start || addr >= vma->vm_end)
1907                 return -EFAULT;
1908
1909         track_pfn_insert(vma, &pgprot, pfn);
1910
1911         /*
1912          * If we don't have pte special, then we have to use the pfn_valid()
1913          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1914          * refcount the page if pfn_valid is true (hence insert_page rather
1915          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1916          * without pte special, it would there be refcounted as a normal page.
1917          */
1918         if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1919                 struct page *page;
1920
1921                 /*
1922                  * At this point we are committed to insert_page()
1923                  * regardless of whether the caller specified flags that
1924                  * result in pfn_t_has_page() == false.
1925                  */
1926                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1927                 return insert_page(vma, addr, page, pgprot);
1928         }
1929         return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1930 }
1931
1932 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1933                         pfn_t pfn)
1934 {
1935         return __vm_insert_mixed(vma, addr, pfn, false);
1936
1937 }
1938 EXPORT_SYMBOL(vm_insert_mixed);
1939
1940 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1941                         pfn_t pfn)
1942 {
1943         return __vm_insert_mixed(vma, addr, pfn, true);
1944 }
1945 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1946
1947 /*
1948  * maps a range of physical memory into the requested pages. the old
1949  * mappings are removed. any references to nonexistent pages results
1950  * in null mappings (currently treated as "copy-on-access")
1951  */
1952 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1953                         unsigned long addr, unsigned long end,
1954                         unsigned long pfn, pgprot_t prot)
1955 {
1956         pte_t *pte;
1957         spinlock_t *ptl;
1958
1959         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1960         if (!pte)
1961                 return -ENOMEM;
1962         arch_enter_lazy_mmu_mode();
1963         do {
1964                 BUG_ON(!pte_none(*pte));
1965                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1966                 pfn++;
1967         } while (pte++, addr += PAGE_SIZE, addr != end);
1968         arch_leave_lazy_mmu_mode();
1969         pte_unmap_unlock(pte - 1, ptl);
1970         return 0;
1971 }
1972
1973 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1974                         unsigned long addr, unsigned long end,
1975                         unsigned long pfn, pgprot_t prot)
1976 {
1977         pmd_t *pmd;
1978         unsigned long next;
1979
1980         pfn -= addr >> PAGE_SHIFT;
1981         pmd = pmd_alloc(mm, pud, addr);
1982         if (!pmd)
1983                 return -ENOMEM;
1984         VM_BUG_ON(pmd_trans_huge(*pmd));
1985         do {
1986                 next = pmd_addr_end(addr, end);
1987                 if (remap_pte_range(mm, pmd, addr, next,
1988                                 pfn + (addr >> PAGE_SHIFT), prot))
1989                         return -ENOMEM;
1990         } while (pmd++, addr = next, addr != end);
1991         return 0;
1992 }
1993
1994 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1995                         unsigned long addr, unsigned long end,
1996                         unsigned long pfn, pgprot_t prot)
1997 {
1998         pud_t *pud;
1999         unsigned long next;
2000
2001         pfn -= addr >> PAGE_SHIFT;
2002         pud = pud_alloc(mm, p4d, addr);
2003         if (!pud)
2004                 return -ENOMEM;
2005         do {
2006                 next = pud_addr_end(addr, end);
2007                 if (remap_pmd_range(mm, pud, addr, next,
2008                                 pfn + (addr >> PAGE_SHIFT), prot))
2009                         return -ENOMEM;
2010         } while (pud++, addr = next, addr != end);
2011         return 0;
2012 }
2013
2014 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2015                         unsigned long addr, unsigned long end,
2016                         unsigned long pfn, pgprot_t prot)
2017 {
2018         p4d_t *p4d;
2019         unsigned long next;
2020
2021         pfn -= addr >> PAGE_SHIFT;
2022         p4d = p4d_alloc(mm, pgd, addr);
2023         if (!p4d)
2024                 return -ENOMEM;
2025         do {
2026                 next = p4d_addr_end(addr, end);
2027                 if (remap_pud_range(mm, p4d, addr, next,
2028                                 pfn + (addr >> PAGE_SHIFT), prot))
2029                         return -ENOMEM;
2030         } while (p4d++, addr = next, addr != end);
2031         return 0;
2032 }
2033
2034 /**
2035  * remap_pfn_range - remap kernel memory to userspace
2036  * @vma: user vma to map to
2037  * @addr: target user address to start at
2038  * @pfn: physical address of kernel memory
2039  * @size: size of map area
2040  * @prot: page protection flags for this mapping
2041  *
2042  *  Note: this is only safe if the mm semaphore is held when called.
2043  */
2044 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2045                     unsigned long pfn, unsigned long size, pgprot_t prot)
2046 {
2047         pgd_t *pgd;
2048         unsigned long next;
2049         unsigned long end = addr + PAGE_ALIGN(size);
2050         struct mm_struct *mm = vma->vm_mm;
2051         unsigned long remap_pfn = pfn;
2052         int err;
2053
2054         /*
2055          * Physically remapped pages are special. Tell the
2056          * rest of the world about it:
2057          *   VM_IO tells people not to look at these pages
2058          *      (accesses can have side effects).
2059          *   VM_PFNMAP tells the core MM that the base pages are just
2060          *      raw PFN mappings, and do not have a "struct page" associated
2061          *      with them.
2062          *   VM_DONTEXPAND
2063          *      Disable vma merging and expanding with mremap().
2064          *   VM_DONTDUMP
2065          *      Omit vma from core dump, even when VM_IO turned off.
2066          *
2067          * There's a horrible special case to handle copy-on-write
2068          * behaviour that some programs depend on. We mark the "original"
2069          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2070          * See vm_normal_page() for details.
2071          */
2072         if (is_cow_mapping(vma->vm_flags)) {
2073                 if (addr != vma->vm_start || end != vma->vm_end)
2074                         return -EINVAL;
2075                 vma->vm_pgoff = pfn;
2076         }
2077
2078         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2079         if (err)
2080                 return -EINVAL;
2081
2082         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2083
2084         BUG_ON(addr >= end);
2085         pfn -= addr >> PAGE_SHIFT;
2086         pgd = pgd_offset(mm, addr);
2087         flush_cache_range(vma, addr, end);
2088         do {
2089                 next = pgd_addr_end(addr, end);
2090                 err = remap_p4d_range(mm, pgd, addr, next,
2091                                 pfn + (addr >> PAGE_SHIFT), prot);
2092                 if (err)
2093                         break;
2094         } while (pgd++, addr = next, addr != end);
2095
2096         if (err)
2097                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2098
2099         return err;
2100 }
2101 EXPORT_SYMBOL(remap_pfn_range);
2102
2103 /**
2104  * vm_iomap_memory - remap memory to userspace
2105  * @vma: user vma to map to
2106  * @start: start of area
2107  * @len: size of area
2108  *
2109  * This is a simplified io_remap_pfn_range() for common driver use. The
2110  * driver just needs to give us the physical memory range to be mapped,
2111  * we'll figure out the rest from the vma information.
2112  *
2113  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2114  * whatever write-combining details or similar.
2115  */
2116 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2117 {
2118         unsigned long vm_len, pfn, pages;
2119
2120         /* Check that the physical memory area passed in looks valid */
2121         if (start + len < start)
2122                 return -EINVAL;
2123         /*
2124          * You *really* shouldn't map things that aren't page-aligned,
2125          * but we've historically allowed it because IO memory might
2126          * just have smaller alignment.
2127          */
2128         len += start & ~PAGE_MASK;
2129         pfn = start >> PAGE_SHIFT;
2130         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2131         if (pfn + pages < pfn)
2132                 return -EINVAL;
2133
2134         /* We start the mapping 'vm_pgoff' pages into the area */
2135         if (vma->vm_pgoff > pages)
2136                 return -EINVAL;
2137         pfn += vma->vm_pgoff;
2138         pages -= vma->vm_pgoff;
2139
2140         /* Can we fit all of the mapping? */
2141         vm_len = vma->vm_end - vma->vm_start;
2142         if (vm_len >> PAGE_SHIFT > pages)
2143                 return -EINVAL;
2144
2145         /* Ok, let it rip */
2146         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2147 }
2148 EXPORT_SYMBOL(vm_iomap_memory);
2149
2150 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2151                                      unsigned long addr, unsigned long end,
2152                                      pte_fn_t fn, void *data)
2153 {
2154         pte_t *pte;
2155         int err;
2156         pgtable_t token;
2157         spinlock_t *uninitialized_var(ptl);
2158
2159         pte = (mm == &init_mm) ?
2160                 pte_alloc_kernel(pmd, addr) :
2161                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2162         if (!pte)
2163                 return -ENOMEM;
2164
2165         BUG_ON(pmd_huge(*pmd));
2166
2167         arch_enter_lazy_mmu_mode();
2168
2169         token = pmd_pgtable(*pmd);
2170
2171         do {
2172                 err = fn(pte++, token, addr, data);
2173                 if (err)
2174                         break;
2175         } while (addr += PAGE_SIZE, addr != end);
2176
2177         arch_leave_lazy_mmu_mode();
2178
2179         if (mm != &init_mm)
2180                 pte_unmap_unlock(pte-1, ptl);
2181         return err;
2182 }
2183
2184 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2185                                      unsigned long addr, unsigned long end,
2186                                      pte_fn_t fn, void *data)
2187 {
2188         pmd_t *pmd;
2189         unsigned long next;
2190         int err;
2191
2192         BUG_ON(pud_huge(*pud));
2193
2194         pmd = pmd_alloc(mm, pud, addr);
2195         if (!pmd)
2196                 return -ENOMEM;
2197         do {
2198                 next = pmd_addr_end(addr, end);
2199                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2200                 if (err)
2201                         break;
2202         } while (pmd++, addr = next, addr != end);
2203         return err;
2204 }
2205
2206 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2207                                      unsigned long addr, unsigned long end,
2208                                      pte_fn_t fn, void *data)
2209 {
2210         pud_t *pud;
2211         unsigned long next;
2212         int err;
2213
2214         pud = pud_alloc(mm, p4d, addr);
2215         if (!pud)
2216                 return -ENOMEM;
2217         do {
2218                 next = pud_addr_end(addr, end);
2219                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2220                 if (err)
2221                         break;
2222         } while (pud++, addr = next, addr != end);
2223         return err;
2224 }
2225
2226 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2227                                      unsigned long addr, unsigned long end,
2228                                      pte_fn_t fn, void *data)
2229 {
2230         p4d_t *p4d;
2231         unsigned long next;
2232         int err;
2233
2234         p4d = p4d_alloc(mm, pgd, addr);
2235         if (!p4d)
2236                 return -ENOMEM;
2237         do {
2238                 next = p4d_addr_end(addr, end);
2239                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2240                 if (err)
2241                         break;
2242         } while (p4d++, addr = next, addr != end);
2243         return err;
2244 }
2245
2246 /*
2247  * Scan a region of virtual memory, filling in page tables as necessary
2248  * and calling a provided function on each leaf page table.
2249  */
2250 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2251                         unsigned long size, pte_fn_t fn, void *data)
2252 {
2253         pgd_t *pgd;
2254         unsigned long next;
2255         unsigned long end = addr + size;
2256         int err;
2257
2258         if (WARN_ON(addr >= end))
2259                 return -EINVAL;
2260
2261         pgd = pgd_offset(mm, addr);
2262         do {
2263                 next = pgd_addr_end(addr, end);
2264                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2265                 if (err)
2266                         break;
2267         } while (pgd++, addr = next, addr != end);
2268
2269         return err;
2270 }
2271 EXPORT_SYMBOL_GPL(apply_to_page_range);
2272
2273 /*
2274  * handle_pte_fault chooses page fault handler according to an entry which was
2275  * read non-atomically.  Before making any commitment, on those architectures
2276  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2277  * parts, do_swap_page must check under lock before unmapping the pte and
2278  * proceeding (but do_wp_page is only called after already making such a check;
2279  * and do_anonymous_page can safely check later on).
2280  */
2281 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2282                                 pte_t *page_table, pte_t orig_pte)
2283 {
2284         int same = 1;
2285 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2286         if (sizeof(pte_t) > sizeof(unsigned long)) {
2287                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2288                 spin_lock(ptl);
2289                 same = pte_same(*page_table, orig_pte);
2290                 spin_unlock(ptl);
2291         }
2292 #endif
2293         pte_unmap(page_table);
2294         return same;
2295 }
2296
2297 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2298 {
2299         debug_dma_assert_idle(src);
2300
2301         /*
2302          * If the source page was a PFN mapping, we don't have
2303          * a "struct page" for it. We do a best-effort copy by
2304          * just copying from the original user address. If that
2305          * fails, we just zero-fill it. Live with it.
2306          */
2307         if (unlikely(!src)) {
2308                 void *kaddr = kmap_atomic(dst);
2309                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2310
2311                 /*
2312                  * This really shouldn't fail, because the page is there
2313                  * in the page tables. But it might just be unreadable,
2314                  * in which case we just give up and fill the result with
2315                  * zeroes.
2316                  */
2317                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2318                         clear_page(kaddr);
2319                 kunmap_atomic(kaddr);
2320                 flush_dcache_page(dst);
2321         } else
2322                 copy_user_highpage(dst, src, va, vma);
2323 }
2324
2325 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2326 {
2327         struct file *vm_file = vma->vm_file;
2328
2329         if (vm_file)
2330                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2331
2332         /*
2333          * Special mappings (e.g. VDSO) do not have any file so fake
2334          * a default GFP_KERNEL for them.
2335          */
2336         return GFP_KERNEL;
2337 }
2338
2339 /*
2340  * Notify the address space that the page is about to become writable so that
2341  * it can prohibit this or wait for the page to get into an appropriate state.
2342  *
2343  * We do this without the lock held, so that it can sleep if it needs to.
2344  */
2345 static int do_page_mkwrite(struct vm_fault *vmf)
2346 {
2347         int ret;
2348         struct page *page = vmf->page;
2349         unsigned int old_flags = vmf->flags;
2350
2351         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2352
2353         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2354         /* Restore original flags so that caller is not surprised */
2355         vmf->flags = old_flags;
2356         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2357                 return ret;
2358         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2359                 lock_page(page);
2360                 if (!page->mapping) {
2361                         unlock_page(page);
2362                         return 0; /* retry */
2363                 }
2364                 ret |= VM_FAULT_LOCKED;
2365         } else
2366                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2367         return ret;
2368 }
2369
2370 /*
2371  * Handle dirtying of a page in shared file mapping on a write fault.
2372  *
2373  * The function expects the page to be locked and unlocks it.
2374  */
2375 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2376                                     struct page *page)
2377 {
2378         struct address_space *mapping;
2379         bool dirtied;
2380         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2381
2382         dirtied = set_page_dirty(page);
2383         VM_BUG_ON_PAGE(PageAnon(page), page);
2384         /*
2385          * Take a local copy of the address_space - page.mapping may be zeroed
2386          * by truncate after unlock_page().   The address_space itself remains
2387          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2388          * release semantics to prevent the compiler from undoing this copying.
2389          */
2390         mapping = page_rmapping(page);
2391         unlock_page(page);
2392
2393         if ((dirtied || page_mkwrite) && mapping) {
2394                 /*
2395                  * Some device drivers do not set page.mapping
2396                  * but still dirty their pages
2397                  */
2398                 balance_dirty_pages_ratelimited(mapping);
2399         }
2400
2401         if (!page_mkwrite)
2402                 file_update_time(vma->vm_file);
2403 }
2404
2405 /*
2406  * Handle write page faults for pages that can be reused in the current vma
2407  *
2408  * This can happen either due to the mapping being with the VM_SHARED flag,
2409  * or due to us being the last reference standing to the page. In either
2410  * case, all we need to do here is to mark the page as writable and update
2411  * any related book-keeping.
2412  */
2413 static inline void wp_page_reuse(struct vm_fault *vmf)
2414         __releases(vmf->ptl)
2415 {
2416         struct vm_area_struct *vma = vmf->vma;
2417         struct page *page = vmf->page;
2418         pte_t entry;
2419         /*
2420          * Clear the pages cpupid information as the existing
2421          * information potentially belongs to a now completely
2422          * unrelated process.
2423          */
2424         if (page)
2425                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2426
2427         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2428         entry = pte_mkyoung(vmf->orig_pte);
2429         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2430         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2431                 update_mmu_cache(vma, vmf->address, vmf->pte);
2432         pte_unmap_unlock(vmf->pte, vmf->ptl);
2433 }
2434
2435 /*
2436  * Handle the case of a page which we actually need to copy to a new page.
2437  *
2438  * Called with mmap_sem locked and the old page referenced, but
2439  * without the ptl held.
2440  *
2441  * High level logic flow:
2442  *
2443  * - Allocate a page, copy the content of the old page to the new one.
2444  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2445  * - Take the PTL. If the pte changed, bail out and release the allocated page
2446  * - If the pte is still the way we remember it, update the page table and all
2447  *   relevant references. This includes dropping the reference the page-table
2448  *   held to the old page, as well as updating the rmap.
2449  * - In any case, unlock the PTL and drop the reference we took to the old page.
2450  */
2451 static int wp_page_copy(struct vm_fault *vmf)
2452 {
2453         struct vm_area_struct *vma = vmf->vma;
2454         struct mm_struct *mm = vma->vm_mm;
2455         struct page *old_page = vmf->page;
2456         struct page *new_page = NULL;
2457         pte_t entry;
2458         int page_copied = 0;
2459         const unsigned long mmun_start = vmf->address & PAGE_MASK;
2460         const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2461         struct mem_cgroup *memcg;
2462
2463         if (unlikely(anon_vma_prepare(vma)))
2464                 goto oom;
2465
2466         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2467                 new_page = alloc_zeroed_user_highpage_movable(vma,
2468                                                               vmf->address);
2469                 if (!new_page)
2470                         goto oom;
2471         } else {
2472                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2473                                 vmf->address);
2474                 if (!new_page)
2475                         goto oom;
2476                 cow_user_page(new_page, old_page, vmf->address, vma);
2477         }
2478
2479         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2480                 goto oom_free_new;
2481
2482         __SetPageUptodate(new_page);
2483
2484         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2485
2486         /*
2487          * Re-check the pte - we dropped the lock
2488          */
2489         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2490         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2491                 if (old_page) {
2492                         if (!PageAnon(old_page)) {
2493                                 dec_mm_counter_fast(mm,
2494                                                 mm_counter_file(old_page));
2495                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2496                         }
2497                 } else {
2498                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2499                 }
2500                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2501                 entry = mk_pte(new_page, vma->vm_page_prot);
2502                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2503                 /*
2504                  * Clear the pte entry and flush it first, before updating the
2505                  * pte with the new entry. This will avoid a race condition
2506                  * seen in the presence of one thread doing SMC and another
2507                  * thread doing COW.
2508                  */
2509                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2510                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2511                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2512                 lru_cache_add_active_or_unevictable(new_page, vma);
2513                 /*
2514                  * We call the notify macro here because, when using secondary
2515                  * mmu page tables (such as kvm shadow page tables), we want the
2516                  * new page to be mapped directly into the secondary page table.
2517                  */
2518                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2519                 update_mmu_cache(vma, vmf->address, vmf->pte);
2520                 if (old_page) {
2521                         /*
2522                          * Only after switching the pte to the new page may
2523                          * we remove the mapcount here. Otherwise another
2524                          * process may come and find the rmap count decremented
2525                          * before the pte is switched to the new page, and
2526                          * "reuse" the old page writing into it while our pte
2527                          * here still points into it and can be read by other
2528                          * threads.
2529                          *
2530                          * The critical issue is to order this
2531                          * page_remove_rmap with the ptp_clear_flush above.
2532                          * Those stores are ordered by (if nothing else,)
2533                          * the barrier present in the atomic_add_negative
2534                          * in page_remove_rmap.
2535                          *
2536                          * Then the TLB flush in ptep_clear_flush ensures that
2537                          * no process can access the old page before the
2538                          * decremented mapcount is visible. And the old page
2539                          * cannot be reused until after the decremented
2540                          * mapcount is visible. So transitively, TLBs to
2541                          * old page will be flushed before it can be reused.
2542                          */
2543                         page_remove_rmap(old_page, false);
2544                 }
2545
2546                 /* Free the old page.. */
2547                 new_page = old_page;
2548                 page_copied = 1;
2549         } else {
2550                 mem_cgroup_cancel_charge(new_page, memcg, false);
2551         }
2552
2553         if (new_page)
2554                 put_page(new_page);
2555
2556         pte_unmap_unlock(vmf->pte, vmf->ptl);
2557         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2558         if (old_page) {
2559                 /*
2560                  * Don't let another task, with possibly unlocked vma,
2561                  * keep the mlocked page.
2562                  */
2563                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2564                         lock_page(old_page);    /* LRU manipulation */
2565                         if (PageMlocked(old_page))
2566                                 munlock_vma_page(old_page);
2567                         unlock_page(old_page);
2568                 }
2569                 put_page(old_page);
2570         }
2571         return page_copied ? VM_FAULT_WRITE : 0;
2572 oom_free_new:
2573         put_page(new_page);
2574 oom:
2575         if (old_page)
2576                 put_page(old_page);
2577         return VM_FAULT_OOM;
2578 }
2579
2580 /**
2581  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2582  *                        writeable once the page is prepared
2583  *
2584  * @vmf: structure describing the fault
2585  *
2586  * This function handles all that is needed to finish a write page fault in a
2587  * shared mapping due to PTE being read-only once the mapped page is prepared.
2588  * It handles locking of PTE and modifying it. The function returns
2589  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2590  * lock.
2591  *
2592  * The function expects the page to be locked or other protection against
2593  * concurrent faults / writeback (such as DAX radix tree locks).
2594  */
2595 int finish_mkwrite_fault(struct vm_fault *vmf)
2596 {
2597         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2598         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2599                                        &vmf->ptl);
2600         /*
2601          * We might have raced with another page fault while we released the
2602          * pte_offset_map_lock.
2603          */
2604         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2605                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2606                 return VM_FAULT_NOPAGE;
2607         }
2608         wp_page_reuse(vmf);
2609         return 0;
2610 }
2611
2612 /*
2613  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2614  * mapping
2615  */
2616 static int wp_pfn_shared(struct vm_fault *vmf)
2617 {
2618         struct vm_area_struct *vma = vmf->vma;
2619
2620         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2621                 int ret;
2622
2623                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2624                 vmf->flags |= FAULT_FLAG_MKWRITE;
2625                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2626                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2627                         return ret;
2628                 return finish_mkwrite_fault(vmf);
2629         }
2630         wp_page_reuse(vmf);
2631         return VM_FAULT_WRITE;
2632 }
2633
2634 static int wp_page_shared(struct vm_fault *vmf)
2635         __releases(vmf->ptl)
2636 {
2637         struct vm_area_struct *vma = vmf->vma;
2638
2639         get_page(vmf->page);
2640
2641         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2642                 int tmp;
2643
2644                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2645                 tmp = do_page_mkwrite(vmf);
2646                 if (unlikely(!tmp || (tmp &
2647                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2648                         put_page(vmf->page);
2649                         return tmp;
2650                 }
2651                 tmp = finish_mkwrite_fault(vmf);
2652                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2653                         unlock_page(vmf->page);
2654                         put_page(vmf->page);
2655                         return tmp;
2656                 }
2657         } else {
2658                 wp_page_reuse(vmf);
2659                 lock_page(vmf->page);
2660         }
2661         fault_dirty_shared_page(vma, vmf->page);
2662         put_page(vmf->page);
2663
2664         return VM_FAULT_WRITE;
2665 }
2666
2667 /*
2668  * This routine handles present pages, when users try to write
2669  * to a shared page. It is done by copying the page to a new address
2670  * and decrementing the shared-page counter for the old page.
2671  *
2672  * Note that this routine assumes that the protection checks have been
2673  * done by the caller (the low-level page fault routine in most cases).
2674  * Thus we can safely just mark it writable once we've done any necessary
2675  * COW.
2676  *
2677  * We also mark the page dirty at this point even though the page will
2678  * change only once the write actually happens. This avoids a few races,
2679  * and potentially makes it more efficient.
2680  *
2681  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2682  * but allow concurrent faults), with pte both mapped and locked.
2683  * We return with mmap_sem still held, but pte unmapped and unlocked.
2684  */
2685 static int do_wp_page(struct vm_fault *vmf)
2686         __releases(vmf->ptl)
2687 {
2688         struct vm_area_struct *vma = vmf->vma;
2689
2690         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2691         if (!vmf->page) {
2692                 /*
2693                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2694                  * VM_PFNMAP VMA.
2695                  *
2696                  * We should not cow pages in a shared writeable mapping.
2697                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2698                  */
2699                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2700                                      (VM_WRITE|VM_SHARED))
2701                         return wp_pfn_shared(vmf);
2702
2703                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2704                 return wp_page_copy(vmf);
2705         }
2706
2707         /*
2708          * Take out anonymous pages first, anonymous shared vmas are
2709          * not dirty accountable.
2710          */
2711         if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2712                 int total_map_swapcount;
2713                 if (!trylock_page(vmf->page)) {
2714                         get_page(vmf->page);
2715                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2716                         lock_page(vmf->page);
2717                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2718                                         vmf->address, &vmf->ptl);
2719                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2720                                 unlock_page(vmf->page);
2721                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2722                                 put_page(vmf->page);
2723                                 return 0;
2724                         }
2725                         put_page(vmf->page);
2726                 }
2727                 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2728                         if (total_map_swapcount == 1) {
2729                                 /*
2730                                  * The page is all ours. Move it to
2731                                  * our anon_vma so the rmap code will
2732                                  * not search our parent or siblings.
2733                                  * Protected against the rmap code by
2734                                  * the page lock.
2735                                  */
2736                                 page_move_anon_rmap(vmf->page, vma);
2737                         }
2738                         unlock_page(vmf->page);
2739                         wp_page_reuse(vmf);
2740                         return VM_FAULT_WRITE;
2741                 }
2742                 unlock_page(vmf->page);
2743         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2744                                         (VM_WRITE|VM_SHARED))) {
2745                 return wp_page_shared(vmf);
2746         }
2747
2748         /*
2749          * Ok, we need to copy. Oh, well..
2750          */
2751         get_page(vmf->page);
2752
2753         pte_unmap_unlock(vmf->pte, vmf->ptl);
2754         return wp_page_copy(vmf);
2755 }
2756
2757 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2758                 unsigned long start_addr, unsigned long end_addr,
2759                 struct zap_details *details)
2760 {
2761         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2762 }
2763
2764 static inline void unmap_mapping_range_tree(struct rb_root *root,
2765                                             struct zap_details *details)
2766 {
2767         struct vm_area_struct *vma;
2768         pgoff_t vba, vea, zba, zea;
2769
2770         vma_interval_tree_foreach(vma, root,
2771                         details->first_index, details->last_index) {
2772
2773                 vba = vma->vm_pgoff;
2774                 vea = vba + vma_pages(vma) - 1;
2775                 zba = details->first_index;
2776                 if (zba < vba)
2777                         zba = vba;
2778                 zea = details->last_index;
2779                 if (zea > vea)
2780                         zea = vea;
2781
2782                 unmap_mapping_range_vma(vma,
2783                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2784                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2785                                 details);
2786         }
2787 }
2788
2789 /**
2790  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2791  * address_space corresponding to the specified page range in the underlying
2792  * file.
2793  *
2794  * @mapping: the address space containing mmaps to be unmapped.
2795  * @holebegin: byte in first page to unmap, relative to the start of
2796  * the underlying file.  This will be rounded down to a PAGE_SIZE
2797  * boundary.  Note that this is different from truncate_pagecache(), which
2798  * must keep the partial page.  In contrast, we must get rid of
2799  * partial pages.
2800  * @holelen: size of prospective hole in bytes.  This will be rounded
2801  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2802  * end of the file.
2803  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2804  * but 0 when invalidating pagecache, don't throw away private data.
2805  */
2806 void unmap_mapping_range(struct address_space *mapping,
2807                 loff_t const holebegin, loff_t const holelen, int even_cows)
2808 {
2809         struct zap_details details = { };
2810         pgoff_t hba = holebegin >> PAGE_SHIFT;
2811         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2812
2813         /* Check for overflow. */
2814         if (sizeof(holelen) > sizeof(hlen)) {
2815                 long long holeend =
2816                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2817                 if (holeend & ~(long long)ULONG_MAX)
2818                         hlen = ULONG_MAX - hba + 1;
2819         }
2820
2821         details.check_mapping = even_cows ? NULL : mapping;
2822         details.first_index = hba;
2823         details.last_index = hba + hlen - 1;
2824         if (details.last_index < details.first_index)
2825                 details.last_index = ULONG_MAX;
2826
2827         i_mmap_lock_write(mapping);
2828         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2829                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2830         i_mmap_unlock_write(mapping);
2831 }
2832 EXPORT_SYMBOL(unmap_mapping_range);
2833
2834 /*
2835  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2836  * but allow concurrent faults), and pte mapped but not yet locked.
2837  * We return with pte unmapped and unlocked.
2838  *
2839  * We return with the mmap_sem locked or unlocked in the same cases
2840  * as does filemap_fault().
2841  */
2842 int do_swap_page(struct vm_fault *vmf)
2843 {
2844         struct vm_area_struct *vma = vmf->vma;
2845         struct page *page = NULL, *swapcache;
2846         struct mem_cgroup *memcg;
2847         struct vma_swap_readahead swap_ra;
2848         swp_entry_t entry;
2849         pte_t pte;
2850         int locked;
2851         int exclusive = 0;
2852         int ret = 0;
2853         bool vma_readahead = swap_use_vma_readahead();
2854
2855         if (vma_readahead)
2856                 page = swap_readahead_detect(vmf, &swap_ra);
2857         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
2858                 if (page)
2859                         put_page(page);
2860                 goto out;
2861         }
2862
2863         entry = pte_to_swp_entry(vmf->orig_pte);
2864         if (unlikely(non_swap_entry(entry))) {
2865                 if (is_migration_entry(entry)) {
2866                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2867                                              vmf->address);
2868                 } else if (is_device_private_entry(entry)) {
2869                         /*
2870                          * For un-addressable device memory we call the pgmap
2871                          * fault handler callback. The callback must migrate
2872                          * the page back to some CPU accessible page.
2873                          */
2874                         ret = device_private_entry_fault(vma, vmf->address, entry,
2875                                                  vmf->flags, vmf->pmd);
2876                 } else if (is_hwpoison_entry(entry)) {
2877                         ret = VM_FAULT_HWPOISON;
2878                 } else {
2879                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2880                         ret = VM_FAULT_SIGBUS;
2881                 }
2882                 goto out;
2883         }
2884         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2885         if (!page)
2886                 page = lookup_swap_cache(entry, vma_readahead ? vma : NULL,
2887                                          vmf->address);
2888         if (!page) {
2889                 if (vma_readahead)
2890                         page = do_swap_page_readahead(entry,
2891                                 GFP_HIGHUSER_MOVABLE, vmf, &swap_ra);
2892                 else
2893                         page = swapin_readahead(entry,
2894                                 GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2895                 if (!page) {
2896                         /*
2897                          * Back out if somebody else faulted in this pte
2898                          * while we released the pte lock.
2899                          */
2900                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2901                                         vmf->address, &vmf->ptl);
2902                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2903                                 ret = VM_FAULT_OOM;
2904                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2905                         goto unlock;
2906                 }
2907
2908                 /* Had to read the page from swap area: Major fault */
2909                 ret = VM_FAULT_MAJOR;
2910                 count_vm_event(PGMAJFAULT);
2911                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2912         } else if (PageHWPoison(page)) {
2913                 /*
2914                  * hwpoisoned dirty swapcache pages are kept for killing
2915                  * owner processes (which may be unknown at hwpoison time)
2916                  */
2917                 ret = VM_FAULT_HWPOISON;
2918                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2919                 swapcache = page;
2920                 goto out_release;
2921         }
2922
2923         swapcache = page;
2924         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2925
2926         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2927         if (!locked) {
2928                 ret |= VM_FAULT_RETRY;
2929                 goto out_release;
2930         }
2931
2932         /*
2933          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2934          * release the swapcache from under us.  The page pin, and pte_same
2935          * test below, are not enough to exclude that.  Even if it is still
2936          * swapcache, we need to check that the page's swap has not changed.
2937          */
2938         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2939                 goto out_page;
2940
2941         page = ksm_might_need_to_copy(page, vma, vmf->address);
2942         if (unlikely(!page)) {
2943                 ret = VM_FAULT_OOM;
2944                 page = swapcache;
2945                 goto out_page;
2946         }
2947
2948         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2949                                 &memcg, false)) {
2950                 ret = VM_FAULT_OOM;
2951                 goto out_page;
2952         }
2953
2954         /*
2955          * Back out if somebody else already faulted in this pte.
2956          */
2957         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2958                         &vmf->ptl);
2959         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2960                 goto out_nomap;
2961
2962         if (unlikely(!PageUptodate(page))) {
2963                 ret = VM_FAULT_SIGBUS;
2964                 goto out_nomap;
2965         }
2966
2967         /*
2968          * The page isn't present yet, go ahead with the fault.
2969          *
2970          * Be careful about the sequence of operations here.
2971          * To get its accounting right, reuse_swap_page() must be called
2972          * while the page is counted on swap but not yet in mapcount i.e.
2973          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2974          * must be called after the swap_free(), or it will never succeed.
2975          */
2976
2977         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2978         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2979         pte = mk_pte(page, vma->vm_page_prot);
2980         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2981                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2982                 vmf->flags &= ~FAULT_FLAG_WRITE;
2983                 ret |= VM_FAULT_WRITE;
2984                 exclusive = RMAP_EXCLUSIVE;
2985         }
2986         flush_icache_page(vma, page);
2987         if (pte_swp_soft_dirty(vmf->orig_pte))
2988                 pte = pte_mksoft_dirty(pte);
2989         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2990         vmf->orig_pte = pte;
2991         if (page == swapcache) {
2992                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2993                 mem_cgroup_commit_charge(page, memcg, true, false);
2994                 activate_page(page);
2995         } else { /* ksm created a completely new copy */
2996                 page_add_new_anon_rmap(page, vma, vmf->address, false);
2997                 mem_cgroup_commit_charge(page, memcg, false, false);
2998                 lru_cache_add_active_or_unevictable(page, vma);
2999         }
3000
3001         swap_free(entry);
3002         if (mem_cgroup_swap_full(page) ||
3003             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3004                 try_to_free_swap(page);
3005         unlock_page(page);
3006         if (page != swapcache) {
3007                 /*
3008                  * Hold the lock to avoid the swap entry to be reused
3009                  * until we take the PT lock for the pte_same() check
3010                  * (to avoid false positives from pte_same). For
3011                  * further safety release the lock after the swap_free
3012                  * so that the swap count won't change under a
3013                  * parallel locked swapcache.
3014                  */
3015                 unlock_page(swapcache);
3016                 put_page(swapcache);
3017         }
3018
3019         if (vmf->flags & FAULT_FLAG_WRITE) {
3020                 ret |= do_wp_page(vmf);
3021                 if (ret & VM_FAULT_ERROR)
3022                         ret &= VM_FAULT_ERROR;
3023                 goto out;
3024         }
3025
3026         /* No need to invalidate - it was non-present before */
3027         update_mmu_cache(vma, vmf->address, vmf->pte);
3028 unlock:
3029         pte_unmap_unlock(vmf->pte, vmf->ptl);
3030 out:
3031         return ret;
3032 out_nomap:
3033         mem_cgroup_cancel_charge(page, memcg, false);
3034         pte_unmap_unlock(vmf->pte, vmf->ptl);
3035 out_page:
3036         unlock_page(page);
3037 out_release:
3038         put_page(page);
3039         if (page != swapcache) {
3040                 unlock_page(swapcache);
3041                 put_page(swapcache);
3042         }
3043         return ret;
3044 }
3045
3046 /*
3047  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3048  * but allow concurrent faults), and pte mapped but not yet locked.
3049  * We return with mmap_sem still held, but pte unmapped and unlocked.
3050  */
3051 static int do_anonymous_page(struct vm_fault *vmf)
3052 {
3053         struct vm_area_struct *vma = vmf->vma;
3054         struct mem_cgroup *memcg;
3055         struct page *page;
3056         int ret = 0;
3057         pte_t entry;
3058
3059         /* File mapping without ->vm_ops ? */
3060         if (vma->vm_flags & VM_SHARED)
3061                 return VM_FAULT_SIGBUS;
3062
3063         /*
3064          * Use pte_alloc() instead of pte_alloc_map().  We can't run
3065          * pte_offset_map() on pmds where a huge pmd might be created
3066          * from a different thread.
3067          *
3068          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3069          * parallel threads are excluded by other means.
3070          *
3071          * Here we only have down_read(mmap_sem).
3072          */
3073         if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3074                 return VM_FAULT_OOM;
3075
3076         /* See the comment in pte_alloc_one_map() */
3077         if (unlikely(pmd_trans_unstable(vmf->pmd)))
3078                 return 0;
3079
3080         /* Use the zero-page for reads */
3081         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3082                         !mm_forbids_zeropage(vma->vm_mm)) {
3083                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3084                                                 vma->vm_page_prot));
3085                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3086                                 vmf->address, &vmf->ptl);
3087                 if (!pte_none(*vmf->pte))
3088                         goto unlock;
3089                 ret = check_stable_address_space(vma->vm_mm);
3090                 if (ret)
3091                         goto unlock;
3092                 /* Deliver the page fault to userland, check inside PT lock */
3093                 if (userfaultfd_missing(vma)) {
3094                         pte_unmap_unlock(vmf->pte, vmf->ptl);
3095                         return handle_userfault(vmf, VM_UFFD_MISSING);
3096                 }
3097                 goto setpte;
3098         }
3099
3100         /* Allocate our own private page. */
3101         if (unlikely(anon_vma_prepare(vma)))
3102                 goto oom;
3103         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3104         if (!page)
3105                 goto oom;
3106
3107         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3108                 goto oom_free_page;
3109
3110         /*
3111          * The memory barrier inside __SetPageUptodate makes sure that
3112          * preceeding stores to the page contents become visible before
3113          * the set_pte_at() write.
3114          */
3115         __SetPageUptodate(page);
3116
3117         entry = mk_pte(page, vma->vm_page_prot);
3118         if (vma->vm_flags & VM_WRITE)
3119                 entry = pte_mkwrite(pte_mkdirty(entry));
3120
3121         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3122                         &vmf->ptl);
3123         if (!pte_none(*vmf->pte))
3124                 goto release;
3125
3126         ret = check_stable_address_space(vma->vm_mm);
3127         if (ret)
3128                 goto release;
3129
3130         /* Deliver the page fault to userland, check inside PT lock */
3131         if (userfaultfd_missing(vma)) {
3132                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3133                 mem_cgroup_cancel_charge(page, memcg, false);
3134                 put_page(page);
3135                 return handle_userfault(vmf, VM_UFFD_MISSING);
3136         }
3137
3138         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3139         page_add_new_anon_rmap(page, vma, vmf->address, false);
3140         mem_cgroup_commit_charge(page, memcg, false, false);
3141         lru_cache_add_active_or_unevictable(page, vma);
3142 setpte:
3143         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3144
3145         /* No need to invalidate - it was non-present before */
3146         update_mmu_cache(vma, vmf->address, vmf->pte);
3147 unlock:
3148         pte_unmap_unlock(vmf->pte, vmf->ptl);
3149         return ret;
3150 release:
3151         mem_cgroup_cancel_charge(page, memcg, false);
3152         put_page(page);
3153         goto unlock;
3154 oom_free_page:
3155         put_page(page);
3156 oom:
3157         return VM_FAULT_OOM;
3158 }
3159
3160 /*
3161  * The mmap_sem must have been held on entry, and may have been
3162  * released depending on flags and vma->vm_ops->fault() return value.
3163  * See filemap_fault() and __lock_page_retry().
3164  */
3165 static int __do_fault(struct vm_fault *vmf)
3166 {
3167         struct vm_area_struct *vma = vmf->vma;
3168         int ret;
3169
3170         ret = vma->vm_ops->fault(vmf);
3171         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3172                             VM_FAULT_DONE_COW)))
3173                 return ret;
3174
3175         if (unlikely(PageHWPoison(vmf->page))) {
3176                 if (ret & VM_FAULT_LOCKED)
3177                         unlock_page(vmf->page);
3178                 put_page(vmf->page);
3179                 vmf->page = NULL;
3180                 return VM_FAULT_HWPOISON;
3181         }
3182
3183         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3184                 lock_page(vmf->page);
3185         else
3186                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3187
3188         return ret;
3189 }
3190
3191 /*
3192  * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3193  * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3194  * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3195  * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3196  */
3197 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3198 {
3199         return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3200 }
3201
3202 static int pte_alloc_one_map(struct vm_fault *vmf)
3203 {
3204         struct vm_area_struct *vma = vmf->vma;
3205
3206         if (!pmd_none(*vmf->pmd))
3207                 goto map_pte;
3208         if (vmf->prealloc_pte) {
3209                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3210                 if (unlikely(!pmd_none(*vmf->pmd))) {
3211                         spin_unlock(vmf->ptl);
3212                         goto map_pte;
3213                 }
3214
3215                 atomic_long_inc(&vma->vm_mm->nr_ptes);
3216                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3217                 spin_unlock(vmf->ptl);
3218                 vmf->prealloc_pte = NULL;
3219         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3220                 return VM_FAULT_OOM;
3221         }
3222 map_pte:
3223         /*
3224          * If a huge pmd materialized under us just retry later.  Use
3225          * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3226          * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3227          * under us and then back to pmd_none, as a result of MADV_DONTNEED
3228          * running immediately after a huge pmd fault in a different thread of
3229          * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3230          * All we have to ensure is that it is a regular pmd that we can walk
3231          * with pte_offset_map() and we can do that through an atomic read in
3232          * C, which is what pmd_trans_unstable() provides.
3233          */
3234         if (pmd_devmap_trans_unstable(vmf->pmd))
3235                 return VM_FAULT_NOPAGE;
3236
3237         /*
3238          * At this point we know that our vmf->pmd points to a page of ptes
3239          * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3240          * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3241          * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3242          * be valid and we will re-check to make sure the vmf->pte isn't
3243          * pte_none() under vmf->ptl protection when we return to
3244          * alloc_set_pte().
3245          */
3246         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3247                         &vmf->ptl);
3248         return 0;
3249 }
3250
3251 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3252
3253 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3254 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3255                 unsigned long haddr)
3256 {
3257         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3258                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3259                 return false;
3260         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3261                 return false;
3262         return true;
3263 }
3264
3265 static void deposit_prealloc_pte(struct vm_fault *vmf)
3266 {
3267         struct vm_area_struct *vma = vmf->vma;
3268
3269         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3270         /*
3271          * We are going to consume the prealloc table,
3272          * count that as nr_ptes.
3273          */
3274         atomic_long_inc(&vma->vm_mm->nr_ptes);
3275         vmf->prealloc_pte = NULL;
3276 }
3277
3278 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3279 {
3280         struct vm_area_struct *vma = vmf->vma;
3281         bool write = vmf->flags & FAULT_FLAG_WRITE;
3282         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3283         pmd_t entry;
3284         int i, ret;
3285
3286         if (!transhuge_vma_suitable(vma, haddr))
3287                 return VM_FAULT_FALLBACK;
3288
3289         ret = VM_FAULT_FALLBACK;
3290         page = compound_head(page);
3291
3292         /*
3293          * Archs like ppc64 need additonal space to store information
3294          * related to pte entry. Use the preallocated table for that.
3295          */
3296         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3297                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3298                 if (!vmf->prealloc_pte)
3299                         return VM_FAULT_OOM;
3300                 smp_wmb(); /* See comment in __pte_alloc() */
3301         }
3302
3303         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3304         if (unlikely(!pmd_none(*vmf->pmd)))
3305                 goto out;
3306
3307         for (i = 0; i < HPAGE_PMD_NR; i++)
3308                 flush_icache_page(vma, page + i);
3309
3310         entry = mk_huge_pmd(page, vma->vm_page_prot);
3311         if (write)
3312                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3313
3314         add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3315         page_add_file_rmap(page, true);
3316         /*
3317          * deposit and withdraw with pmd lock held
3318          */
3319         if (arch_needs_pgtable_deposit())
3320                 deposit_prealloc_pte(vmf);
3321
3322         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3323
3324         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3325
3326         /* fault is handled */
3327         ret = 0;
3328         count_vm_event(THP_FILE_MAPPED);
3329 out:
3330         spin_unlock(vmf->ptl);
3331         return ret;
3332 }
3333 #else
3334 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3335 {
3336         BUILD_BUG();
3337         return 0;
3338 }
3339 #endif
3340
3341 /**
3342  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3343  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3344  *
3345  * @vmf: fault environment
3346  * @memcg: memcg to charge page (only for private mappings)
3347  * @page: page to map
3348  *
3349  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3350  * return.
3351  *
3352  * Target users are page handler itself and implementations of
3353  * vm_ops->map_pages.
3354  */
3355 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3356                 struct page *page)
3357 {
3358         struct vm_area_struct *vma = vmf->vma;
3359         bool write = vmf->flags & FAULT_FLAG_WRITE;
3360         pte_t entry;
3361         int ret;
3362
3363         if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3364                         IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3365                 /* THP on COW? */
3366                 VM_BUG_ON_PAGE(memcg, page);
3367
3368                 ret = do_set_pmd(vmf, page);
3369                 if (ret != VM_FAULT_FALLBACK)
3370                         return ret;
3371         }
3372
3373         if (!vmf->pte) {
3374                 ret = pte_alloc_one_map(vmf);
3375                 if (ret)
3376                         return ret;
3377         }
3378
3379         /* Re-check under ptl */
3380         if (unlikely(!pte_none(*vmf->pte)))
3381                 return VM_FAULT_NOPAGE;
3382
3383         flush_icache_page(vma, page);
3384         entry = mk_pte(page, vma->vm_page_prot);
3385         if (write)
3386                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3387         /* copy-on-write page */
3388         if (write && !(vma->vm_flags & VM_SHARED)) {
3389                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3390                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3391                 mem_cgroup_commit_charge(page, memcg, false, false);
3392                 lru_cache_add_active_or_unevictable(page, vma);
3393         } else {
3394                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3395                 page_add_file_rmap(page, false);
3396         }
3397         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3398
3399         /* no need to invalidate: a not-present page won't be cached */
3400         update_mmu_cache(vma, vmf->address, vmf->pte);
3401
3402         return 0;
3403 }
3404
3405
3406 /**
3407  * finish_fault - finish page fault once we have prepared the page to fault
3408  *
3409  * @vmf: structure describing the fault
3410  *
3411  * This function handles all that is needed to finish a page fault once the
3412  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3413  * given page, adds reverse page mapping, handles memcg charges and LRU
3414  * addition. The function returns 0 on success, VM_FAULT_ code in case of
3415  * error.
3416  *
3417  * The function expects the page to be locked and on success it consumes a
3418  * reference of a page being mapped (for the PTE which maps it).
3419  */
3420 int finish_fault(struct vm_fault *vmf)
3421 {
3422         struct page *page;
3423         int ret = 0;
3424
3425         /* Did we COW the page? */
3426         if ((vmf->flags & FAULT_FLAG_WRITE) &&
3427             !(vmf->vma->vm_flags & VM_SHARED))
3428                 page = vmf->cow_page;
3429         else
3430                 page = vmf->page;
3431
3432         /*
3433          * check even for read faults because we might have lost our CoWed
3434          * page
3435          */
3436         if (!(vmf->vma->vm_flags & VM_SHARED))
3437                 ret = check_stable_address_space(vmf->vma->vm_mm);
3438         if (!ret)
3439                 ret = alloc_set_pte(vmf, vmf->memcg, page);
3440         if (vmf->pte)
3441                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3442         return ret;
3443 }
3444
3445 static unsigned long fault_around_bytes __read_mostly =
3446         rounddown_pow_of_two(65536);
3447
3448 #ifdef CONFIG_DEBUG_FS
3449 static int fault_around_bytes_get(void *data, u64 *val)
3450 {
3451         *val = fault_around_bytes;
3452         return 0;
3453 }
3454
3455 /*
3456  * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3457  * rounded down to nearest page order. It's what do_fault_around() expects to
3458  * see.
3459  */
3460 static int fault_around_bytes_set(void *data, u64 val)
3461 {
3462         if (val / PAGE_SIZE > PTRS_PER_PTE)
3463                 return -EINVAL;
3464         if (val > PAGE_SIZE)
3465                 fault_around_bytes = rounddown_pow_of_two(val);
3466         else
3467                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3468         return 0;
3469 }
3470 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3471                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3472
3473 static int __init fault_around_debugfs(void)
3474 {
3475         void *ret;
3476
3477         ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3478                         &fault_around_bytes_fops);
3479         if (!ret)
3480                 pr_warn("Failed to create fault_around_bytes in debugfs");
3481         return 0;
3482 }
3483 late_initcall(fault_around_debugfs);
3484 #endif
3485
3486 /*
3487  * do_fault_around() tries to map few pages around the fault address. The hope
3488  * is that the pages will be needed soon and this will lower the number of
3489  * faults to handle.
3490  *
3491  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3492  * not ready to be mapped: not up-to-date, locked, etc.
3493  *
3494  * This function is called with the page table lock taken. In the split ptlock
3495  * case the page table lock only protects only those entries which belong to
3496  * the page table corresponding to the fault address.
3497  *
3498  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3499  * only once.
3500  *
3501  * fault_around_pages() defines how many pages we'll try to map.
3502  * do_fault_around() expects it to return a power of two less than or equal to
3503  * PTRS_PER_PTE.
3504  *
3505  * The virtual address of the area that we map is naturally aligned to the
3506  * fault_around_pages() value (and therefore to page order).  This way it's
3507  * easier to guarantee that we don't cross page table boundaries.
3508  */
3509 static int do_fault_around(struct vm_fault *vmf)
3510 {
3511         unsigned long address = vmf->address, nr_pages, mask;
3512         pgoff_t start_pgoff = vmf->pgoff;
3513         pgoff_t end_pgoff;
3514         int off, ret = 0;
3515
3516         nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3517         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3518
3519         vmf->address = max(address & mask, vmf->vma->vm_start);
3520         off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3521         start_pgoff -= off;
3522
3523         /*
3524          *  end_pgoff is either end of page table or end of vma
3525          *  or fault_around_pages() from start_pgoff, depending what is nearest.
3526          */
3527         end_pgoff = start_pgoff -
3528                 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3529                 PTRS_PER_PTE - 1;
3530         end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3531                         start_pgoff + nr_pages - 1);
3532
3533         if (pmd_none(*vmf->pmd)) {
3534                 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3535                                                   vmf->address);
3536                 if (!vmf->prealloc_pte)
3537                         goto out;
3538                 smp_wmb(); /* See comment in __pte_alloc() */
3539         }
3540
3541         vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3542
3543         /* Huge page is mapped? Page fault is solved */
3544         if (pmd_trans_huge(*vmf->pmd)) {
3545                 ret = VM_FAULT_NOPAGE;
3546                 goto out;
3547         }
3548
3549         /* ->map_pages() haven't done anything useful. Cold page cache? */
3550         if (!vmf->pte)
3551                 goto out;
3552
3553         /* check if the page fault is solved */
3554         vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3555         if (!pte_none(*vmf->pte))
3556                 ret = VM_FAULT_NOPAGE;
3557         pte_unmap_unlock(vmf->pte, vmf->ptl);
3558 out:
3559         vmf->address = address;
3560         vmf->pte = NULL;
3561         return ret;
3562 }
3563
3564 static int do_read_fault(struct vm_fault *vmf)
3565 {
3566         struct vm_area_struct *vma = vmf->vma;
3567         int ret = 0;
3568
3569         /*
3570          * Let's call ->map_pages() first and use ->fault() as fallback
3571          * if page by the offset is not ready to be mapped (cold cache or
3572          * something).
3573          */
3574         if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3575                 ret = do_fault_around(vmf);
3576                 if (ret)
3577                         return ret;
3578         }
3579
3580         ret = __do_fault(vmf);
3581         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3582                 return ret;
3583
3584         ret |= finish_fault(vmf);
3585         unlock_page(vmf->page);
3586         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3587                 put_page(vmf->page);
3588         return ret;
3589 }
3590
3591 static int do_cow_fault(struct vm_fault *vmf)
3592 {
3593         struct vm_area_struct *vma = vmf->vma;
3594         int ret;
3595
3596         if (unlikely(anon_vma_prepare(vma)))
3597                 return VM_FAULT_OOM;
3598
3599         vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3600         if (!vmf->cow_page)
3601                 return VM_FAULT_OOM;
3602
3603         if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3604                                 &vmf->memcg, false)) {
3605                 put_page(vmf->cow_page);
3606                 return VM_FAULT_OOM;
3607         }
3608
3609         ret = __do_fault(vmf);
3610         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3611                 goto uncharge_out;
3612         if (ret & VM_FAULT_DONE_COW)
3613                 return ret;
3614
3615         copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3616         __SetPageUptodate(vmf->cow_page);
3617
3618         ret |= finish_fault(vmf);
3619         unlock_page(vmf->page);
3620         put_page(vmf->page);
3621         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3622                 goto uncharge_out;
3623         return ret;
3624 uncharge_out:
3625         mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3626         put_page(vmf->cow_page);
3627         return ret;
3628 }
3629
3630 static int do_shared_fault(struct vm_fault *vmf)
3631 {
3632         struct vm_area_struct *vma = vmf->vma;
3633         int ret, tmp;
3634
3635         ret = __do_fault(vmf);
3636         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3637                 return ret;
3638
3639         /*
3640          * Check if the backing address space wants to know that the page is
3641          * about to become writable
3642          */
3643         if (vma->vm_ops->page_mkwrite) {
3644                 unlock_page(vmf->page);
3645                 tmp = do_page_mkwrite(vmf);
3646                 if (unlikely(!tmp ||
3647                                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3648                         put_page(vmf->page);
3649                         return tmp;
3650                 }
3651         }
3652
3653         ret |= finish_fault(vmf);
3654         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3655                                         VM_FAULT_RETRY))) {
3656                 unlock_page(vmf->page);
3657                 put_page(vmf->page);
3658                 return ret;
3659         }
3660
3661         fault_dirty_shared_page(vma, vmf->page);
3662         return ret;
3663 }
3664
3665 /*
3666  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3667  * but allow concurrent faults).
3668  * The mmap_sem may have been released depending on flags and our
3669  * return value.  See filemap_fault() and __lock_page_or_retry().
3670  */
3671 static int do_fault(struct vm_fault *vmf)
3672 {
3673         struct vm_area_struct *vma = vmf->vma;
3674         int ret;
3675
3676         /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3677         if (!vma->vm_ops->fault)
3678                 ret = VM_FAULT_SIGBUS;
3679         else if (!(vmf->flags & FAULT_FLAG_WRITE))
3680                 ret = do_read_fault(vmf);
3681         else if (!(vma->vm_flags & VM_SHARED))
3682                 ret = do_cow_fault(vmf);
3683         else
3684                 ret = do_shared_fault(vmf);
3685
3686         /* preallocated pagetable is unused: free it */
3687         if (vmf->prealloc_pte) {
3688                 pte_free(vma->vm_mm, vmf->prealloc_pte);
3689                 vmf->prealloc_pte = NULL;
3690         }
3691         return ret;
3692 }
3693
3694 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3695                                 unsigned long addr, int page_nid,
3696                                 int *flags)
3697 {
3698         get_page(page);
3699
3700         count_vm_numa_event(NUMA_HINT_FAULTS);
3701         if (page_nid == numa_node_id()) {
3702                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3703                 *flags |= TNF_FAULT_LOCAL;
3704         }
3705
3706         return mpol_misplaced(page, vma, addr);
3707 }
3708
3709 static int do_numa_page(struct vm_fault *vmf)
3710 {
3711         struct vm_area_struct *vma = vmf->vma;
3712         struct page *page = NULL;
3713         int page_nid = -1;
3714         int last_cpupid;
3715         int target_nid;
3716         bool migrated = false;
3717         pte_t pte;
3718         bool was_writable = pte_savedwrite(vmf->orig_pte);
3719         int flags = 0;
3720
3721         /*
3722          * The "pte" at this point cannot be used safely without
3723          * validation through pte_unmap_same(). It's of NUMA type but
3724          * the pfn may be screwed if the read is non atomic.
3725          */
3726         vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3727         spin_lock(vmf->ptl);
3728         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3729                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3730                 goto out;
3731         }
3732
3733         /*
3734          * Make it present again, Depending on how arch implementes non
3735          * accessible ptes, some can allow access by kernel mode.
3736          */
3737         pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3738         pte = pte_modify(pte, vma->vm_page_prot);
3739         pte = pte_mkyoung(pte);
3740         if (was_writable)
3741                 pte = pte_mkwrite(pte);
3742         ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3743         update_mmu_cache(vma, vmf->address, vmf->pte);
3744
3745         page = vm_normal_page(vma, vmf->address, pte);
3746         if (!page) {
3747                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3748                 return 0;
3749         }
3750
3751         /* TODO: handle PTE-mapped THP */
3752         if (PageCompound(page)) {
3753                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3754                 return 0;
3755         }
3756
3757         /*
3758          * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3759          * much anyway since they can be in shared cache state. This misses
3760          * the case where a mapping is writable but the process never writes
3761          * to it but pte_write gets cleared during protection updates and
3762          * pte_dirty has unpredictable behaviour between PTE scan updates,
3763          * background writeback, dirty balancing and application behaviour.
3764          */
3765         if (!pte_write(pte))
3766                 flags |= TNF_NO_GROUP;
3767
3768         /*
3769          * Flag if the page is shared between multiple address spaces. This
3770          * is later used when determining whether to group tasks together
3771          */
3772         if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3773                 flags |= TNF_SHARED;
3774
3775         last_cpupid = page_cpupid_last(page);
3776         page_nid = page_to_nid(page);
3777         target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3778                         &flags);
3779         pte_unmap_unlock(vmf->pte, vmf->ptl);
3780         if (target_nid == -1) {
3781                 put_page(page);
3782                 goto out;
3783         }
3784
3785         /* Migrate to the requested node */
3786         migrated = migrate_misplaced_page(page, vma, target_nid);
3787         if (migrated) {
3788                 page_nid = target_nid;
3789                 flags |= TNF_MIGRATED;
3790         } else
3791                 flags |= TNF_MIGRATE_FAIL;
3792
3793 out:
3794         if (page_nid != -1)
3795                 task_numa_fault(last_cpupid, page_nid, 1, flags);
3796         return 0;
3797 }
3798
3799 static inline int create_huge_pmd(struct vm_fault *vmf)
3800 {
3801         if (vma_is_anonymous(vmf->vma))
3802                 return do_huge_pmd_anonymous_page(vmf);
3803         if (vmf->vma->vm_ops->huge_fault)
3804                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3805         return VM_FAULT_FALLBACK;
3806 }
3807
3808 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3809 {
3810         if (vma_is_anonymous(vmf->vma))
3811                 return do_huge_pmd_wp_page(vmf, orig_pmd);
3812         if (vmf->vma->vm_ops->huge_fault)
3813                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3814
3815         /* COW handled on pte level: split pmd */
3816         VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3817         __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3818
3819         return VM_FAULT_FALLBACK;
3820 }
3821
3822 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3823 {
3824         return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3825 }
3826
3827 static int create_huge_pud(struct vm_fault *vmf)
3828 {
3829 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3830         /* No support for anonymous transparent PUD pages yet */
3831         if (vma_is_anonymous(vmf->vma))
3832                 return VM_FAULT_FALLBACK;
3833         if (vmf->vma->vm_ops->huge_fault)
3834                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3835 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3836         return VM_FAULT_FALLBACK;
3837 }
3838
3839 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3840 {
3841 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3842         /* No support for anonymous transparent PUD pages yet */
3843         if (vma_is_anonymous(vmf->vma))
3844                 return VM_FAULT_FALLBACK;
3845         if (vmf->vma->vm_ops->huge_fault)
3846                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3847 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3848         return VM_FAULT_FALLBACK;
3849 }
3850
3851 /*
3852  * These routines also need to handle stuff like marking pages dirty
3853  * and/or accessed for architectures that don't do it in hardware (most
3854  * RISC architectures).  The early dirtying is also good on the i386.
3855  *
3856  * There is also a hook called "update_mmu_cache()" that architectures
3857  * with external mmu caches can use to update those (ie the Sparc or
3858  * PowerPC hashed page tables that act as extended TLBs).
3859  *
3860  * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3861  * concurrent faults).
3862  *
3863  * The mmap_sem may have been released depending on flags and our return value.
3864  * See filemap_fault() and __lock_page_or_retry().
3865  */
3866 static int handle_pte_fault(struct vm_fault *vmf)
3867 {
3868         pte_t entry;
3869
3870         if (unlikely(pmd_none(*vmf->pmd))) {
3871                 /*
3872                  * Leave __pte_alloc() until later: because vm_ops->fault may
3873                  * want to allocate huge page, and if we expose page table
3874                  * for an instant, it will be difficult to retract from
3875                  * concurrent fault