53373b7a15126287462c0b93d5189d0d9a687bf5
[muen/linux.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/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 /**
404  * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
405  * @tlb: the mmu_gather structure to initialize
406  * @mm: the mm_struct of the target address space
407  * @start: start of the region that will be removed from the page-table
408  * @end: end of the region that will be removed from the page-table
409  *
410  * Called to initialize an (on-stack) mmu_gather structure for page-table
411  * tear-down from @mm. The @start and @end are set to 0 and -1
412  * respectively when @mm is without users and we're going to destroy
413  * the full address space (exit/execve).
414  */
415 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
416                         unsigned long start, unsigned long end)
417 {
418         arch_tlb_gather_mmu(tlb, mm, start, end);
419         inc_tlb_flush_pending(tlb->mm);
420 }
421
422 void tlb_finish_mmu(struct mmu_gather *tlb,
423                 unsigned long start, unsigned long end)
424 {
425         /*
426          * If there are parallel threads are doing PTE changes on same range
427          * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
428          * flush by batching, a thread has stable TLB entry can fail to flush
429          * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
430          * forcefully if we detect parallel PTE batching threads.
431          */
432         bool force = mm_tlb_flush_nested(tlb->mm);
433
434         arch_tlb_finish_mmu(tlb, start, end, force);
435         dec_tlb_flush_pending(tlb->mm);
436 }
437
438 /*
439  * Note: this doesn't free the actual pages themselves. That
440  * has been handled earlier when unmapping all the memory regions.
441  */
442 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
443                            unsigned long addr)
444 {
445         pgtable_t token = pmd_pgtable(*pmd);
446         pmd_clear(pmd);
447         pte_free_tlb(tlb, token, addr);
448         mm_dec_nr_ptes(tlb->mm);
449 }
450
451 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
452                                 unsigned long addr, unsigned long end,
453                                 unsigned long floor, unsigned long ceiling)
454 {
455         pmd_t *pmd;
456         unsigned long next;
457         unsigned long start;
458
459         start = addr;
460         pmd = pmd_offset(pud, addr);
461         do {
462                 next = pmd_addr_end(addr, end);
463                 if (pmd_none_or_clear_bad(pmd))
464                         continue;
465                 free_pte_range(tlb, pmd, addr);
466         } while (pmd++, addr = next, addr != end);
467
468         start &= PUD_MASK;
469         if (start < floor)
470                 return;
471         if (ceiling) {
472                 ceiling &= PUD_MASK;
473                 if (!ceiling)
474                         return;
475         }
476         if (end - 1 > ceiling - 1)
477                 return;
478
479         pmd = pmd_offset(pud, start);
480         pud_clear(pud);
481         pmd_free_tlb(tlb, pmd, start);
482         mm_dec_nr_pmds(tlb->mm);
483 }
484
485 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
486                                 unsigned long addr, unsigned long end,
487                                 unsigned long floor, unsigned long ceiling)
488 {
489         pud_t *pud;
490         unsigned long next;
491         unsigned long start;
492
493         start = addr;
494         pud = pud_offset(p4d, addr);
495         do {
496                 next = pud_addr_end(addr, end);
497                 if (pud_none_or_clear_bad(pud))
498                         continue;
499                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
500         } while (pud++, addr = next, addr != end);
501
502         start &= P4D_MASK;
503         if (start < floor)
504                 return;
505         if (ceiling) {
506                 ceiling &= P4D_MASK;
507                 if (!ceiling)
508                         return;
509         }
510         if (end - 1 > ceiling - 1)
511                 return;
512
513         pud = pud_offset(p4d, start);
514         p4d_clear(p4d);
515         pud_free_tlb(tlb, pud, start);
516         mm_dec_nr_puds(tlb->mm);
517 }
518
519 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
520                                 unsigned long addr, unsigned long end,
521                                 unsigned long floor, unsigned long ceiling)
522 {
523         p4d_t *p4d;
524         unsigned long next;
525         unsigned long start;
526
527         start = addr;
528         p4d = p4d_offset(pgd, addr);
529         do {
530                 next = p4d_addr_end(addr, end);
531                 if (p4d_none_or_clear_bad(p4d))
532                         continue;
533                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
534         } while (p4d++, addr = next, addr != end);
535
536         start &= PGDIR_MASK;
537         if (start < floor)
538                 return;
539         if (ceiling) {
540                 ceiling &= PGDIR_MASK;
541                 if (!ceiling)
542                         return;
543         }
544         if (end - 1 > ceiling - 1)
545                 return;
546
547         p4d = p4d_offset(pgd, start);
548         pgd_clear(pgd);
549         p4d_free_tlb(tlb, p4d, start);
550 }
551
552 /*
553  * This function frees user-level page tables of a process.
554  */
555 void free_pgd_range(struct mmu_gather *tlb,
556                         unsigned long addr, unsigned long end,
557                         unsigned long floor, unsigned long ceiling)
558 {
559         pgd_t *pgd;
560         unsigned long next;
561
562         /*
563          * The next few lines have given us lots of grief...
564          *
565          * Why are we testing PMD* at this top level?  Because often
566          * there will be no work to do at all, and we'd prefer not to
567          * go all the way down to the bottom just to discover that.
568          *
569          * Why all these "- 1"s?  Because 0 represents both the bottom
570          * of the address space and the top of it (using -1 for the
571          * top wouldn't help much: the masks would do the wrong thing).
572          * The rule is that addr 0 and floor 0 refer to the bottom of
573          * the address space, but end 0 and ceiling 0 refer to the top
574          * Comparisons need to use "end - 1" and "ceiling - 1" (though
575          * that end 0 case should be mythical).
576          *
577          * Wherever addr is brought up or ceiling brought down, we must
578          * be careful to reject "the opposite 0" before it confuses the
579          * subsequent tests.  But what about where end is brought down
580          * by PMD_SIZE below? no, end can't go down to 0 there.
581          *
582          * Whereas we round start (addr) and ceiling down, by different
583          * masks at different levels, in order to test whether a table
584          * now has no other vmas using it, so can be freed, we don't
585          * bother to round floor or end up - the tests don't need that.
586          */
587
588         addr &= PMD_MASK;
589         if (addr < floor) {
590                 addr += PMD_SIZE;
591                 if (!addr)
592                         return;
593         }
594         if (ceiling) {
595                 ceiling &= PMD_MASK;
596                 if (!ceiling)
597                         return;
598         }
599         if (end - 1 > ceiling - 1)
600                 end -= PMD_SIZE;
601         if (addr > end - 1)
602                 return;
603         /*
604          * We add page table cache pages with PAGE_SIZE,
605          * (see pte_free_tlb()), flush the tlb if we need
606          */
607         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
608         pgd = pgd_offset(tlb->mm, addr);
609         do {
610                 next = pgd_addr_end(addr, end);
611                 if (pgd_none_or_clear_bad(pgd))
612                         continue;
613                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
614         } while (pgd++, addr = next, addr != end);
615 }
616
617 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
618                 unsigned long floor, unsigned long ceiling)
619 {
620         while (vma) {
621                 struct vm_area_struct *next = vma->vm_next;
622                 unsigned long addr = vma->vm_start;
623
624                 /*
625                  * Hide vma from rmap and truncate_pagecache before freeing
626                  * pgtables
627                  */
628                 unlink_anon_vmas(vma);
629                 unlink_file_vma(vma);
630
631                 if (is_vm_hugetlb_page(vma)) {
632                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
633                                 floor, next ? next->vm_start : ceiling);
634                 } else {
635                         /*
636                          * Optimization: gather nearby vmas into one call down
637                          */
638                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
639                                && !is_vm_hugetlb_page(next)) {
640                                 vma = next;
641                                 next = vma->vm_next;
642                                 unlink_anon_vmas(vma);
643                                 unlink_file_vma(vma);
644                         }
645                         free_pgd_range(tlb, addr, vma->vm_end,
646                                 floor, next ? next->vm_start : ceiling);
647                 }
648                 vma = next;
649         }
650 }
651
652 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
653 {
654         spinlock_t *ptl;
655         pgtable_t new = pte_alloc_one(mm, address);
656         if (!new)
657                 return -ENOMEM;
658
659         /*
660          * Ensure all pte setup (eg. pte page lock and page clearing) are
661          * visible before the pte is made visible to other CPUs by being
662          * put into page tables.
663          *
664          * The other side of the story is the pointer chasing in the page
665          * table walking code (when walking the page table without locking;
666          * ie. most of the time). Fortunately, these data accesses consist
667          * of a chain of data-dependent loads, meaning most CPUs (alpha
668          * being the notable exception) will already guarantee loads are
669          * seen in-order. See the alpha page table accessors for the
670          * smp_read_barrier_depends() barriers in page table walking code.
671          */
672         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
673
674         ptl = pmd_lock(mm, pmd);
675         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
676                 mm_inc_nr_ptes(mm);
677                 pmd_populate(mm, pmd, new);
678                 new = NULL;
679         }
680         spin_unlock(ptl);
681         if (new)
682                 pte_free(mm, new);
683         return 0;
684 }
685
686 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
687 {
688         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
689         if (!new)
690                 return -ENOMEM;
691
692         smp_wmb(); /* See comment in __pte_alloc */
693
694         spin_lock(&init_mm.page_table_lock);
695         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
696                 pmd_populate_kernel(&init_mm, pmd, new);
697                 new = NULL;
698         }
699         spin_unlock(&init_mm.page_table_lock);
700         if (new)
701                 pte_free_kernel(&init_mm, new);
702         return 0;
703 }
704
705 static inline void init_rss_vec(int *rss)
706 {
707         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
708 }
709
710 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
711 {
712         int i;
713
714         if (current->mm == mm)
715                 sync_mm_rss(mm);
716         for (i = 0; i < NR_MM_COUNTERS; i++)
717                 if (rss[i])
718                         add_mm_counter(mm, i, rss[i]);
719 }
720
721 /*
722  * This function is called to print an error when a bad pte
723  * is found. For example, we might have a PFN-mapped pte in
724  * a region that doesn't allow it.
725  *
726  * The calling function must still handle the error.
727  */
728 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
729                           pte_t pte, struct page *page)
730 {
731         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
732         p4d_t *p4d = p4d_offset(pgd, addr);
733         pud_t *pud = pud_offset(p4d, addr);
734         pmd_t *pmd = pmd_offset(pud, addr);
735         struct address_space *mapping;
736         pgoff_t index;
737         static unsigned long resume;
738         static unsigned long nr_shown;
739         static unsigned long nr_unshown;
740
741         /*
742          * Allow a burst of 60 reports, then keep quiet for that minute;
743          * or allow a steady drip of one report per second.
744          */
745         if (nr_shown == 60) {
746                 if (time_before(jiffies, resume)) {
747                         nr_unshown++;
748                         return;
749                 }
750                 if (nr_unshown) {
751                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
752                                  nr_unshown);
753                         nr_unshown = 0;
754                 }
755                 nr_shown = 0;
756         }
757         if (nr_shown++ == 0)
758                 resume = jiffies + 60 * HZ;
759
760         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
761         index = linear_page_index(vma, addr);
762
763         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
764                  current->comm,
765                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
766         if (page)
767                 dump_page(page, "bad pte");
768         pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
769                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
770         /*
771          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
772          */
773         pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
774                  vma->vm_file,
775                  vma->vm_ops ? vma->vm_ops->fault : NULL,
776                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
777                  mapping ? mapping->a_ops->readpage : NULL);
778         dump_stack();
779         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
780 }
781
782 /*
783  * vm_normal_page -- This function gets the "struct page" associated with a pte.
784  *
785  * "Special" mappings do not wish to be associated with a "struct page" (either
786  * it doesn't exist, or it exists but they don't want to touch it). In this
787  * case, NULL is returned here. "Normal" mappings do have a struct page.
788  *
789  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
790  * pte bit, in which case this function is trivial. Secondly, an architecture
791  * may not have a spare pte bit, which requires a more complicated scheme,
792  * described below.
793  *
794  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
795  * special mapping (even if there are underlying and valid "struct pages").
796  * COWed pages of a VM_PFNMAP are always normal.
797  *
798  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
799  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
800  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
801  * mapping will always honor the rule
802  *
803  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
804  *
805  * And for normal mappings this is false.
806  *
807  * This restricts such mappings to be a linear translation from virtual address
808  * to pfn. To get around this restriction, we allow arbitrary mappings so long
809  * as the vma is not a COW mapping; in that case, we know that all ptes are
810  * special (because none can have been COWed).
811  *
812  *
813  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
814  *
815  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
816  * page" backing, however the difference is that _all_ pages with a struct
817  * page (that is, those where pfn_valid is true) are refcounted and considered
818  * normal pages by the VM. The disadvantage is that pages are refcounted
819  * (which can be slower and simply not an option for some PFNMAP users). The
820  * advantage is that we don't have to follow the strict linearity rule of
821  * PFNMAP mappings in order to support COWable mappings.
822  *
823  */
824 #ifdef __HAVE_ARCH_PTE_SPECIAL
825 # define HAVE_PTE_SPECIAL 1
826 #else
827 # define HAVE_PTE_SPECIAL 0
828 #endif
829 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
830                              pte_t pte, bool with_public_device)
831 {
832         unsigned long pfn = pte_pfn(pte);
833
834         if (HAVE_PTE_SPECIAL) {
835                 if (likely(!pte_special(pte)))
836                         goto check_pfn;
837                 if (vma->vm_ops && vma->vm_ops->find_special_page)
838                         return vma->vm_ops->find_special_page(vma, addr);
839                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
840                         return NULL;
841                 if (is_zero_pfn(pfn))
842                         return NULL;
843
844                 /*
845                  * Device public pages are special pages (they are ZONE_DEVICE
846                  * pages but different from persistent memory). They behave
847                  * allmost like normal pages. The difference is that they are
848                  * not on the lru and thus should never be involve with any-
849                  * thing that involve lru manipulation (mlock, numa balancing,
850                  * ...).
851                  *
852                  * This is why we still want to return NULL for such page from
853                  * vm_normal_page() so that we do not have to special case all
854                  * call site of vm_normal_page().
855                  */
856                 if (likely(pfn <= highest_memmap_pfn)) {
857                         struct page *page = pfn_to_page(pfn);
858
859                         if (is_device_public_page(page)) {
860                                 if (with_public_device)
861                                         return page;
862                                 return NULL;
863                         }
864                 }
865                 print_bad_pte(vma, addr, pte, NULL);
866                 return NULL;
867         }
868
869         /* !HAVE_PTE_SPECIAL case follows: */
870
871         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
872                 if (vma->vm_flags & VM_MIXEDMAP) {
873                         if (!pfn_valid(pfn))
874                                 return NULL;
875                         goto out;
876                 } else {
877                         unsigned long off;
878                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
879                         if (pfn == vma->vm_pgoff + off)
880                                 return NULL;
881                         if (!is_cow_mapping(vma->vm_flags))
882                                 return NULL;
883                 }
884         }
885
886         if (is_zero_pfn(pfn))
887                 return NULL;
888 check_pfn:
889         if (unlikely(pfn > highest_memmap_pfn)) {
890                 print_bad_pte(vma, addr, pte, NULL);
891                 return NULL;
892         }
893
894         /*
895          * NOTE! We still have PageReserved() pages in the page tables.
896          * eg. VDSO mappings can cause them to exist.
897          */
898 out:
899         return pfn_to_page(pfn);
900 }
901
902 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
903 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
904                                 pmd_t pmd)
905 {
906         unsigned long pfn = pmd_pfn(pmd);
907
908         /*
909          * There is no pmd_special() but there may be special pmds, e.g.
910          * in a direct-access (dax) mapping, so let's just replicate the
911          * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
912          */
913         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
914                 if (vma->vm_flags & VM_MIXEDMAP) {
915                         if (!pfn_valid(pfn))
916                                 return NULL;
917                         goto out;
918                 } else {
919                         unsigned long off;
920                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
921                         if (pfn == vma->vm_pgoff + off)
922                                 return NULL;
923                         if (!is_cow_mapping(vma->vm_flags))
924                                 return NULL;
925                 }
926         }
927
928         if (is_zero_pfn(pfn))
929                 return NULL;
930         if (unlikely(pfn > highest_memmap_pfn))
931                 return NULL;
932
933         /*
934          * NOTE! We still have PageReserved() pages in the page tables.
935          * eg. VDSO mappings can cause them to exist.
936          */
937 out:
938         return pfn_to_page(pfn);
939 }
940 #endif
941
942 /*
943  * copy one vm_area from one task to the other. Assumes the page tables
944  * already present in the new task to be cleared in the whole range
945  * covered by this vma.
946  */
947
948 static inline unsigned long
949 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
950                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
951                 unsigned long addr, int *rss)
952 {
953         unsigned long vm_flags = vma->vm_flags;
954         pte_t pte = *src_pte;
955         struct page *page;
956
957         /* pte contains position in swap or file, so copy. */
958         if (unlikely(!pte_present(pte))) {
959                 swp_entry_t entry = pte_to_swp_entry(pte);
960
961                 if (likely(!non_swap_entry(entry))) {
962                         if (swap_duplicate(entry) < 0)
963                                 return entry.val;
964
965                         /* make sure dst_mm is on swapoff's mmlist. */
966                         if (unlikely(list_empty(&dst_mm->mmlist))) {
967                                 spin_lock(&mmlist_lock);
968                                 if (list_empty(&dst_mm->mmlist))
969                                         list_add(&dst_mm->mmlist,
970                                                         &src_mm->mmlist);
971                                 spin_unlock(&mmlist_lock);
972                         }
973                         rss[MM_SWAPENTS]++;
974                 } else if (is_migration_entry(entry)) {
975                         page = migration_entry_to_page(entry);
976
977                         rss[mm_counter(page)]++;
978
979                         if (is_write_migration_entry(entry) &&
980                                         is_cow_mapping(vm_flags)) {
981                                 /*
982                                  * COW mappings require pages in both
983                                  * parent and child to be set to read.
984                                  */
985                                 make_migration_entry_read(&entry);
986                                 pte = swp_entry_to_pte(entry);
987                                 if (pte_swp_soft_dirty(*src_pte))
988                                         pte = pte_swp_mksoft_dirty(pte);
989                                 set_pte_at(src_mm, addr, src_pte, pte);
990                         }
991                 } else if (is_device_private_entry(entry)) {
992                         page = device_private_entry_to_page(entry);
993
994                         /*
995                          * Update rss count even for unaddressable pages, as
996                          * they should treated just like normal pages in this
997                          * respect.
998                          *
999                          * We will likely want to have some new rss counters
1000                          * for unaddressable pages, at some point. But for now
1001                          * keep things as they are.
1002                          */
1003                         get_page(page);
1004                         rss[mm_counter(page)]++;
1005                         page_dup_rmap(page, false);
1006
1007                         /*
1008                          * We do not preserve soft-dirty information, because so
1009                          * far, checkpoint/restore is the only feature that
1010                          * requires that. And checkpoint/restore does not work
1011                          * when a device driver is involved (you cannot easily
1012                          * save and restore device driver state).
1013                          */
1014                         if (is_write_device_private_entry(entry) &&
1015                             is_cow_mapping(vm_flags)) {
1016                                 make_device_private_entry_read(&entry);
1017                                 pte = swp_entry_to_pte(entry);
1018                                 set_pte_at(src_mm, addr, src_pte, pte);
1019                         }
1020                 }
1021                 goto out_set_pte;
1022         }
1023
1024         /*
1025          * If it's a COW mapping, write protect it both
1026          * in the parent and the child
1027          */
1028         if (is_cow_mapping(vm_flags)) {
1029                 ptep_set_wrprotect(src_mm, addr, src_pte);
1030                 pte = pte_wrprotect(pte);
1031         }
1032
1033         /*
1034          * If it's a shared mapping, mark it clean in
1035          * the child
1036          */
1037         if (vm_flags & VM_SHARED)
1038                 pte = pte_mkclean(pte);
1039         pte = pte_mkold(pte);
1040
1041         page = vm_normal_page(vma, addr, pte);
1042         if (page) {
1043                 get_page(page);
1044                 page_dup_rmap(page, false);
1045                 rss[mm_counter(page)]++;
1046         } else if (pte_devmap(pte)) {
1047                 page = pte_page(pte);
1048
1049                 /*
1050                  * Cache coherent device memory behave like regular page and
1051                  * not like persistent memory page. For more informations see
1052                  * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1053                  */
1054                 if (is_device_public_page(page)) {
1055                         get_page(page);
1056                         page_dup_rmap(page, false);
1057                         rss[mm_counter(page)]++;
1058                 }
1059         }
1060
1061 out_set_pte:
1062         set_pte_at(dst_mm, addr, dst_pte, pte);
1063         return 0;
1064 }
1065
1066 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1067                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1068                    unsigned long addr, unsigned long end)
1069 {
1070         pte_t *orig_src_pte, *orig_dst_pte;
1071         pte_t *src_pte, *dst_pte;
1072         spinlock_t *src_ptl, *dst_ptl;
1073         int progress = 0;
1074         int rss[NR_MM_COUNTERS];
1075         swp_entry_t entry = (swp_entry_t){0};
1076
1077 again:
1078         init_rss_vec(rss);
1079
1080         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1081         if (!dst_pte)
1082                 return -ENOMEM;
1083         src_pte = pte_offset_map(src_pmd, addr);
1084         src_ptl = pte_lockptr(src_mm, src_pmd);
1085         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1086         orig_src_pte = src_pte;
1087         orig_dst_pte = dst_pte;
1088         arch_enter_lazy_mmu_mode();
1089
1090         do {
1091                 /*
1092                  * We are holding two locks at this point - either of them
1093                  * could generate latencies in another task on another CPU.
1094                  */
1095                 if (progress >= 32) {
1096                         progress = 0;
1097                         if (need_resched() ||
1098                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1099                                 break;
1100                 }
1101                 if (pte_none(*src_pte)) {
1102                         progress++;
1103                         continue;
1104                 }
1105                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1106                                                         vma, addr, rss);
1107                 if (entry.val)
1108                         break;
1109                 progress += 8;
1110         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1111
1112         arch_leave_lazy_mmu_mode();
1113         spin_unlock(src_ptl);
1114         pte_unmap(orig_src_pte);
1115         add_mm_rss_vec(dst_mm, rss);
1116         pte_unmap_unlock(orig_dst_pte, dst_ptl);
1117         cond_resched();
1118
1119         if (entry.val) {
1120                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1121                         return -ENOMEM;
1122                 progress = 0;
1123         }
1124         if (addr != end)
1125                 goto again;
1126         return 0;
1127 }
1128
1129 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1130                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1131                 unsigned long addr, unsigned long end)
1132 {
1133         pmd_t *src_pmd, *dst_pmd;
1134         unsigned long next;
1135
1136         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1137         if (!dst_pmd)
1138                 return -ENOMEM;
1139         src_pmd = pmd_offset(src_pud, addr);
1140         do {
1141                 next = pmd_addr_end(addr, end);
1142                 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1143                         || pmd_devmap(*src_pmd)) {
1144                         int err;
1145                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1146                         err = copy_huge_pmd(dst_mm, src_mm,
1147                                             dst_pmd, src_pmd, addr, vma);
1148                         if (err == -ENOMEM)
1149                                 return -ENOMEM;
1150                         if (!err)
1151                                 continue;
1152                         /* fall through */
1153                 }
1154                 if (pmd_none_or_clear_bad(src_pmd))
1155                         continue;
1156                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1157                                                 vma, addr, next))
1158                         return -ENOMEM;
1159         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1160         return 0;
1161 }
1162
1163 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1164                 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1165                 unsigned long addr, unsigned long end)
1166 {
1167         pud_t *src_pud, *dst_pud;
1168         unsigned long next;
1169
1170         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1171         if (!dst_pud)
1172                 return -ENOMEM;
1173         src_pud = pud_offset(src_p4d, addr);
1174         do {
1175                 next = pud_addr_end(addr, end);
1176                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1177                         int err;
1178
1179                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1180                         err = copy_huge_pud(dst_mm, src_mm,
1181                                             dst_pud, src_pud, addr, vma);
1182                         if (err == -ENOMEM)
1183                                 return -ENOMEM;
1184                         if (!err)
1185                                 continue;
1186                         /* fall through */
1187                 }
1188                 if (pud_none_or_clear_bad(src_pud))
1189                         continue;
1190                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1191                                                 vma, addr, next))
1192                         return -ENOMEM;
1193         } while (dst_pud++, src_pud++, addr = next, addr != end);
1194         return 0;
1195 }
1196
1197 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1198                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1199                 unsigned long addr, unsigned long end)
1200 {
1201         p4d_t *src_p4d, *dst_p4d;
1202         unsigned long next;
1203
1204         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1205         if (!dst_p4d)
1206                 return -ENOMEM;
1207         src_p4d = p4d_offset(src_pgd, addr);
1208         do {
1209                 next = p4d_addr_end(addr, end);
1210                 if (p4d_none_or_clear_bad(src_p4d))
1211                         continue;
1212                 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1213                                                 vma, addr, next))
1214                         return -ENOMEM;
1215         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1216         return 0;
1217 }
1218
1219 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1220                 struct vm_area_struct *vma)
1221 {
1222         pgd_t *src_pgd, *dst_pgd;
1223         unsigned long next;
1224         unsigned long addr = vma->vm_start;
1225         unsigned long end = vma->vm_end;
1226         unsigned long mmun_start;       /* For mmu_notifiers */
1227         unsigned long mmun_end;         /* For mmu_notifiers */
1228         bool is_cow;
1229         int ret;
1230
1231         /*
1232          * Don't copy ptes where a page fault will fill them correctly.
1233          * Fork becomes much lighter when there are big shared or private
1234          * readonly mappings. The tradeoff is that copy_page_range is more
1235          * efficient than faulting.
1236          */
1237         if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1238                         !vma->anon_vma)
1239                 return 0;
1240
1241         if (is_vm_hugetlb_page(vma))
1242                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1243
1244         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1245                 /*
1246                  * We do not free on error cases below as remove_vma
1247                  * gets called on error from higher level routine
1248                  */
1249                 ret = track_pfn_copy(vma);
1250                 if (ret)
1251                         return ret;
1252         }
1253
1254         /*
1255          * We need to invalidate the secondary MMU mappings only when
1256          * there could be a permission downgrade on the ptes of the
1257          * parent mm. And a permission downgrade will only happen if
1258          * is_cow_mapping() returns true.
1259          */
1260         is_cow = is_cow_mapping(vma->vm_flags);
1261         mmun_start = addr;
1262         mmun_end   = end;
1263         if (is_cow)
1264                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1265                                                     mmun_end);
1266
1267         ret = 0;
1268         dst_pgd = pgd_offset(dst_mm, addr);
1269         src_pgd = pgd_offset(src_mm, addr);
1270         do {
1271                 next = pgd_addr_end(addr, end);
1272                 if (pgd_none_or_clear_bad(src_pgd))
1273                         continue;
1274                 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1275                                             vma, addr, next))) {
1276                         ret = -ENOMEM;
1277                         break;
1278                 }
1279         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1280
1281         if (is_cow)
1282                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1283         return ret;
1284 }
1285
1286 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1287                                 struct vm_area_struct *vma, pmd_t *pmd,
1288                                 unsigned long addr, unsigned long end,
1289                                 struct zap_details *details)
1290 {
1291         struct mm_struct *mm = tlb->mm;
1292         int force_flush = 0;
1293         int rss[NR_MM_COUNTERS];
1294         spinlock_t *ptl;
1295         pte_t *start_pte;
1296         pte_t *pte;
1297         swp_entry_t entry;
1298
1299         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1300 again:
1301         init_rss_vec(rss);
1302         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1303         pte = start_pte;
1304         flush_tlb_batched_pending(mm);
1305         arch_enter_lazy_mmu_mode();
1306         do {
1307                 pte_t ptent = *pte;
1308                 if (pte_none(ptent))
1309                         continue;
1310
1311                 if (pte_present(ptent)) {
1312                         struct page *page;
1313
1314                         page = _vm_normal_page(vma, addr, ptent, true);
1315                         if (unlikely(details) && page) {
1316                                 /*
1317                                  * unmap_shared_mapping_pages() wants to
1318                                  * invalidate cache without truncating:
1319                                  * unmap shared but keep private pages.
1320                                  */
1321                                 if (details->check_mapping &&
1322                                     details->check_mapping != page_rmapping(page))
1323                                         continue;
1324                         }
1325                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1326                                                         tlb->fullmm);
1327                         tlb_remove_tlb_entry(tlb, pte, addr);
1328                         if (unlikely(!page))
1329                                 continue;
1330
1331                         if (!PageAnon(page)) {
1332                                 if (pte_dirty(ptent)) {
1333                                         force_flush = 1;
1334                                         set_page_dirty(page);
1335                                 }
1336                                 if (pte_young(ptent) &&
1337                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1338                                         mark_page_accessed(page);
1339                         }
1340                         rss[mm_counter(page)]--;
1341                         page_remove_rmap(page, false);
1342                         if (unlikely(page_mapcount(page) < 0))
1343                                 print_bad_pte(vma, addr, ptent, page);
1344                         if (unlikely(__tlb_remove_page(tlb, page))) {
1345                                 force_flush = 1;
1346                                 addr += PAGE_SIZE;
1347                                 break;
1348                         }
1349                         continue;
1350                 }
1351
1352                 entry = pte_to_swp_entry(ptent);
1353                 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1354                         struct page *page = device_private_entry_to_page(entry);
1355
1356                         if (unlikely(details && details->check_mapping)) {
1357                                 /*
1358                                  * unmap_shared_mapping_pages() wants to
1359                                  * invalidate cache without truncating:
1360                                  * unmap shared but keep private pages.
1361                                  */
1362                                 if (details->check_mapping !=
1363                                     page_rmapping(page))
1364                                         continue;
1365                         }
1366
1367                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1368                         rss[mm_counter(page)]--;
1369                         page_remove_rmap(page, false);
1370                         put_page(page);
1371                         continue;
1372                 }
1373
1374                 /* If details->check_mapping, we leave swap entries. */
1375                 if (unlikely(details))
1376                         continue;
1377
1378                 entry = pte_to_swp_entry(ptent);
1379                 if (!non_swap_entry(entry))
1380                         rss[MM_SWAPENTS]--;
1381                 else if (is_migration_entry(entry)) {
1382                         struct page *page;
1383
1384                         page = migration_entry_to_page(entry);
1385                         rss[mm_counter(page)]--;
1386                 }
1387                 if (unlikely(!free_swap_and_cache(entry)))
1388                         print_bad_pte(vma, addr, ptent, NULL);
1389                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1390         } while (pte++, addr += PAGE_SIZE, addr != end);
1391
1392         add_mm_rss_vec(mm, rss);
1393         arch_leave_lazy_mmu_mode();
1394
1395         /* Do the actual TLB flush before dropping ptl */
1396         if (force_flush)
1397                 tlb_flush_mmu_tlbonly(tlb);
1398         pte_unmap_unlock(start_pte, ptl);
1399
1400         /*
1401          * If we forced a TLB flush (either due to running out of
1402          * batch buffers or because we needed to flush dirty TLB
1403          * entries before releasing the ptl), free the batched
1404          * memory too. Restart if we didn't do everything.
1405          */
1406         if (force_flush) {
1407                 force_flush = 0;
1408                 tlb_flush_mmu_free(tlb);
1409                 if (addr != end)
1410                         goto again;
1411         }
1412
1413         return addr;
1414 }
1415
1416 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1417                                 struct vm_area_struct *vma, pud_t *pud,
1418                                 unsigned long addr, unsigned long end,
1419                                 struct zap_details *details)
1420 {
1421         pmd_t *pmd;
1422         unsigned long next;
1423
1424         pmd = pmd_offset(pud, addr);
1425         do {
1426                 next = pmd_addr_end(addr, end);
1427                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1428                         if (next - addr != HPAGE_PMD_SIZE) {
1429                                 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1430                                     !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1431                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1432                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1433                                 goto next;
1434                         /* fall through */
1435                 }
1436                 /*
1437                  * Here there can be other concurrent MADV_DONTNEED or
1438                  * trans huge page faults running, and if the pmd is
1439                  * none or trans huge it can change under us. This is
1440                  * because MADV_DONTNEED holds the mmap_sem in read
1441                  * mode.
1442                  */
1443                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1444                         goto next;
1445                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1446 next:
1447                 cond_resched();
1448         } while (pmd++, addr = next, addr != end);
1449
1450         return addr;
1451 }
1452
1453 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1454                                 struct vm_area_struct *vma, p4d_t *p4d,
1455                                 unsigned long addr, unsigned long end,
1456                                 struct zap_details *details)
1457 {
1458         pud_t *pud;
1459         unsigned long next;
1460
1461         pud = pud_offset(p4d, addr);
1462         do {
1463                 next = pud_addr_end(addr, end);
1464                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1465                         if (next - addr != HPAGE_PUD_SIZE) {
1466                                 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1467                                 split_huge_pud(vma, pud, addr);
1468                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1469                                 goto next;
1470                         /* fall through */
1471                 }
1472                 if (pud_none_or_clear_bad(pud))
1473                         continue;
1474                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1475 next:
1476                 cond_resched();
1477         } while (pud++, addr = next, addr != end);
1478
1479         return addr;
1480 }
1481
1482 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1483                                 struct vm_area_struct *vma, pgd_t *pgd,
1484                                 unsigned long addr, unsigned long end,
1485                                 struct zap_details *details)
1486 {
1487         p4d_t *p4d;
1488         unsigned long next;
1489
1490         p4d = p4d_offset(pgd, addr);
1491         do {
1492                 next = p4d_addr_end(addr, end);
1493                 if (p4d_none_or_clear_bad(p4d))
1494                         continue;
1495                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1496         } while (p4d++, addr = next, addr != end);
1497
1498         return addr;
1499 }
1500
1501 void unmap_page_range(struct mmu_gather *tlb,
1502                              struct vm_area_struct *vma,
1503                              unsigned long addr, unsigned long end,
1504                              struct zap_details *details)
1505 {
1506         pgd_t *pgd;
1507         unsigned long next;
1508
1509         BUG_ON(addr >= end);
1510         tlb_start_vma(tlb, vma);
1511         pgd = pgd_offset(vma->vm_mm, addr);
1512         do {
1513                 next = pgd_addr_end(addr, end);
1514                 if (pgd_none_or_clear_bad(pgd))
1515                         continue;
1516                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1517         } while (pgd++, addr = next, addr != end);
1518         tlb_end_vma(tlb, vma);
1519 }
1520
1521
1522 static void unmap_single_vma(struct mmu_gather *tlb,
1523                 struct vm_area_struct *vma, unsigned long start_addr,
1524                 unsigned long end_addr,
1525                 struct zap_details *details)
1526 {
1527         unsigned long start = max(vma->vm_start, start_addr);
1528         unsigned long end;
1529
1530         if (start >= vma->vm_end)
1531                 return;
1532         end = min(vma->vm_end, end_addr);
1533         if (end <= vma->vm_start)
1534                 return;
1535
1536         if (vma->vm_file)
1537                 uprobe_munmap(vma, start, end);
1538
1539         if (unlikely(vma->vm_flags & VM_PFNMAP))
1540                 untrack_pfn(vma, 0, 0);
1541
1542         if (start != end) {
1543                 if (unlikely(is_vm_hugetlb_page(vma))) {
1544                         /*
1545                          * It is undesirable to test vma->vm_file as it
1546                          * should be non-null for valid hugetlb area.
1547                          * However, vm_file will be NULL in the error
1548                          * cleanup path of mmap_region. When
1549                          * hugetlbfs ->mmap method fails,
1550                          * mmap_region() nullifies vma->vm_file
1551                          * before calling this function to clean up.
1552                          * Since no pte has actually been setup, it is
1553                          * safe to do nothing in this case.
1554                          */
1555                         if (vma->vm_file) {
1556                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1557                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1558                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1559                         }
1560                 } else
1561                         unmap_page_range(tlb, vma, start, end, details);
1562         }
1563 }
1564
1565 /**
1566  * unmap_vmas - unmap a range of memory covered by a list of vma's
1567  * @tlb: address of the caller's struct mmu_gather
1568  * @vma: the starting vma
1569  * @start_addr: virtual address at which to start unmapping
1570  * @end_addr: virtual address at which to end unmapping
1571  *
1572  * Unmap all pages in the vma list.
1573  *
1574  * Only addresses between `start' and `end' will be unmapped.
1575  *
1576  * The VMA list must be sorted in ascending virtual address order.
1577  *
1578  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1579  * range after unmap_vmas() returns.  So the only responsibility here is to
1580  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1581  * drops the lock and schedules.
1582  */
1583 void unmap_vmas(struct mmu_gather *tlb,
1584                 struct vm_area_struct *vma, unsigned long start_addr,
1585                 unsigned long end_addr)
1586 {
1587         struct mm_struct *mm = vma->vm_mm;
1588
1589         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1590         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1591                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1592         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1593 }
1594
1595 /**
1596  * zap_page_range - remove user pages in a given range
1597  * @vma: vm_area_struct holding the applicable pages
1598  * @start: starting address of pages to zap
1599  * @size: number of bytes to zap
1600  *
1601  * Caller must protect the VMA list
1602  */
1603 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1604                 unsigned long size)
1605 {
1606         struct mm_struct *mm = vma->vm_mm;
1607         struct mmu_gather tlb;
1608         unsigned long end = start + size;
1609
1610         lru_add_drain();
1611         tlb_gather_mmu(&tlb, mm, start, end);
1612         update_hiwater_rss(mm);
1613         mmu_notifier_invalidate_range_start(mm, start, end);
1614         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1615                 unmap_single_vma(&tlb, vma, start, end, NULL);
1616
1617                 /*
1618                  * zap_page_range does not specify whether mmap_sem should be
1619                  * held for read or write. That allows parallel zap_page_range
1620                  * operations to unmap a PTE and defer a flush meaning that
1621                  * this call observes pte_none and fails to flush the TLB.
1622                  * Rather than adding a complex API, ensure that no stale
1623                  * TLB entries exist when this call returns.
1624                  */
1625                 flush_tlb_range(vma, start, end);
1626         }
1627
1628         mmu_notifier_invalidate_range_end(mm, start, end);
1629         tlb_finish_mmu(&tlb, start, end);
1630 }
1631
1632 /**
1633  * zap_page_range_single - remove user pages in a given range
1634  * @vma: vm_area_struct holding the applicable pages
1635  * @address: starting address of pages to zap
1636  * @size: number of bytes to zap
1637  * @details: details of shared cache invalidation
1638  *
1639  * The range must fit into one VMA.
1640  */
1641 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1642                 unsigned long size, struct zap_details *details)
1643 {
1644         struct mm_struct *mm = vma->vm_mm;
1645         struct mmu_gather tlb;
1646         unsigned long end = address + size;
1647
1648         lru_add_drain();
1649         tlb_gather_mmu(&tlb, mm, address, end);
1650         update_hiwater_rss(mm);
1651         mmu_notifier_invalidate_range_start(mm, address, end);
1652         unmap_single_vma(&tlb, vma, address, end, details);
1653         mmu_notifier_invalidate_range_end(mm, address, end);
1654         tlb_finish_mmu(&tlb, address, end);
1655 }
1656
1657 /**
1658  * zap_vma_ptes - remove ptes mapping the vma
1659  * @vma: vm_area_struct holding ptes to be zapped
1660  * @address: starting address of pages to zap
1661  * @size: number of bytes to zap
1662  *
1663  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1664  *
1665  * The entire address range must be fully contained within the vma.
1666  *
1667  * Returns 0 if successful.
1668  */
1669 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1670                 unsigned long size)
1671 {
1672         if (address < vma->vm_start || address + size > vma->vm_end ||
1673                         !(vma->vm_flags & VM_PFNMAP))
1674                 return -1;
1675         zap_page_range_single(vma, address, size, NULL);
1676         return 0;
1677 }
1678 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1679
1680 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1681                         spinlock_t **ptl)
1682 {
1683         pgd_t *pgd;
1684         p4d_t *p4d;
1685         pud_t *pud;
1686         pmd_t *pmd;
1687
1688         pgd = pgd_offset(mm, addr);
1689         p4d = p4d_alloc(mm, pgd, addr);
1690         if (!p4d)
1691                 return NULL;
1692         pud = pud_alloc(mm, p4d, addr);
1693         if (!pud)
1694                 return NULL;
1695         pmd = pmd_alloc(mm, pud, addr);
1696         if (!pmd)
1697                 return NULL;
1698
1699         VM_BUG_ON(pmd_trans_huge(*pmd));
1700         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1701 }
1702
1703 /*
1704  * This is the old fallback for page remapping.
1705  *
1706  * For historical reasons, it only allows reserved pages. Only
1707  * old drivers should use this, and they needed to mark their
1708  * pages reserved for the old functions anyway.
1709  */
1710 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1711                         struct page *page, pgprot_t prot)
1712 {
1713         struct mm_struct *mm = vma->vm_mm;
1714         int retval;
1715         pte_t *pte;
1716         spinlock_t *ptl;
1717
1718         retval = -EINVAL;
1719         if (PageAnon(page))
1720                 goto out;
1721         retval = -ENOMEM;
1722         flush_dcache_page(page);
1723         pte = get_locked_pte(mm, addr, &ptl);
1724         if (!pte)
1725                 goto out;
1726         retval = -EBUSY;
1727         if (!pte_none(*pte))
1728                 goto out_unlock;
1729
1730         /* Ok, finally just insert the thing.. */
1731         get_page(page);
1732         inc_mm_counter_fast(mm, mm_counter_file(page));
1733         page_add_file_rmap(page, false);
1734         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1735
1736         retval = 0;
1737         pte_unmap_unlock(pte, ptl);
1738         return retval;
1739 out_unlock:
1740         pte_unmap_unlock(pte, ptl);
1741 out:
1742         return retval;
1743 }
1744
1745 /**
1746  * vm_insert_page - insert single page into user vma
1747  * @vma: user vma to map to
1748  * @addr: target user address of this page
1749  * @page: source kernel page
1750  *
1751  * This allows drivers to insert individual pages they've allocated
1752  * into a user vma.
1753  *
1754  * The page has to be a nice clean _individual_ kernel allocation.
1755  * If you allocate a compound page, you need to have marked it as
1756  * such (__GFP_COMP), or manually just split the page up yourself
1757  * (see split_page()).
1758  *
1759  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1760  * took an arbitrary page protection parameter. This doesn't allow
1761  * that. Your vma protection will have to be set up correctly, which
1762  * means that if you want a shared writable mapping, you'd better
1763  * ask for a shared writable mapping!
1764  *
1765  * The page does not need to be reserved.
1766  *
1767  * Usually this function is called from f_op->mmap() handler
1768  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1769  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1770  * function from other places, for example from page-fault handler.
1771  */
1772 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1773                         struct page *page)
1774 {
1775         if (addr < vma->vm_start || addr >= vma->vm_end)
1776                 return -EFAULT;
1777         if (!page_count(page))
1778                 return -EINVAL;
1779         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1780                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1781                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1782                 vma->vm_flags |= VM_MIXEDMAP;
1783         }
1784         return insert_page(vma, addr, page, vma->vm_page_prot);
1785 }
1786 EXPORT_SYMBOL(vm_insert_page);
1787
1788 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1789                         pfn_t pfn, pgprot_t prot, bool mkwrite)
1790 {
1791         struct mm_struct *mm = vma->vm_mm;
1792         int retval;
1793         pte_t *pte, entry;
1794         spinlock_t *ptl;
1795
1796         retval = -ENOMEM;
1797         pte = get_locked_pte(mm, addr, &ptl);
1798         if (!pte)
1799                 goto out;
1800         retval = -EBUSY;
1801         if (!pte_none(*pte)) {
1802                 if (mkwrite) {
1803                         /*
1804                          * For read faults on private mappings the PFN passed
1805                          * in may not match the PFN we have mapped if the
1806                          * mapped PFN is a writeable COW page.  In the mkwrite
1807                          * case we are creating a writable PTE for a shared
1808                          * mapping and we expect the PFNs to match.
1809                          */
1810                         if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1811                                 goto out_unlock;
1812                         entry = *pte;
1813                         goto out_mkwrite;
1814                 } else
1815                         goto out_unlock;
1816         }
1817
1818         /* Ok, finally just insert the thing.. */
1819         if (pfn_t_devmap(pfn))
1820                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1821         else
1822                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1823
1824 out_mkwrite:
1825         if (mkwrite) {
1826                 entry = pte_mkyoung(entry);
1827                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1828         }
1829
1830         set_pte_at(mm, addr, pte, entry);
1831         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1832
1833         retval = 0;
1834 out_unlock:
1835         pte_unmap_unlock(pte, ptl);
1836 out:
1837         return retval;
1838 }
1839
1840 /**
1841  * vm_insert_pfn - insert single pfn into user vma
1842  * @vma: user vma to map to
1843  * @addr: target user address of this page
1844  * @pfn: source kernel pfn
1845  *
1846  * Similar to vm_insert_page, this allows drivers to insert individual pages
1847  * they've allocated into a user vma. Same comments apply.
1848  *
1849  * This function should only be called from a vm_ops->fault handler, and
1850  * in that case the handler should return NULL.
1851  *
1852  * vma cannot be a COW mapping.
1853  *
1854  * As this is called only for pages that do not currently exist, we
1855  * do not need to flush old virtual caches or the TLB.
1856  */
1857 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1858                         unsigned long pfn)
1859 {
1860         return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1861 }
1862 EXPORT_SYMBOL(vm_insert_pfn);
1863
1864 /**
1865  * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1866  * @vma: user vma to map to
1867  * @addr: target user address of this page
1868  * @pfn: source kernel pfn
1869  * @pgprot: pgprot flags for the inserted page
1870  *
1871  * This is exactly like vm_insert_pfn, except that it allows drivers to
1872  * to override pgprot on a per-page basis.
1873  *
1874  * This only makes sense for IO mappings, and it makes no sense for
1875  * cow mappings.  In general, using multiple vmas is preferable;
1876  * vm_insert_pfn_prot should only be used if using multiple VMAs is
1877  * impractical.
1878  */
1879 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1880                         unsigned long pfn, pgprot_t pgprot)
1881 {
1882         int ret;
1883         /*
1884          * Technically, architectures with pte_special can avoid all these
1885          * restrictions (same for remap_pfn_range).  However we would like
1886          * consistency in testing and feature parity among all, so we should
1887          * try to keep these invariants in place for everybody.
1888          */
1889         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1890         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1891                                                 (VM_PFNMAP|VM_MIXEDMAP));
1892         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1893         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1894
1895         if (addr < vma->vm_start || addr >= vma->vm_end)
1896                 return -EFAULT;
1897
1898         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1899
1900         ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1901                         false);
1902
1903         return ret;
1904 }
1905 EXPORT_SYMBOL(vm_insert_pfn_prot);
1906
1907 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1908                         pfn_t pfn, bool mkwrite)
1909 {
1910         pgprot_t pgprot = vma->vm_page_prot;
1911
1912         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1913
1914         if (addr < vma->vm_start || addr >= vma->vm_end)
1915                 return -EFAULT;
1916
1917         track_pfn_insert(vma, &pgprot, pfn);
1918
1919         /*
1920          * If we don't have pte special, then we have to use the pfn_valid()
1921          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1922          * refcount the page if pfn_valid is true (hence insert_page rather
1923          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1924          * without pte special, it would there be refcounted as a normal page.
1925          */
1926         if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1927                 struct page *page;
1928
1929                 /*
1930                  * At this point we are committed to insert_page()
1931                  * regardless of whether the caller specified flags that
1932                  * result in pfn_t_has_page() == false.
1933                  */
1934                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1935                 return insert_page(vma, addr, page, pgprot);
1936         }
1937         return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1938 }
1939
1940 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1941                         pfn_t pfn)
1942 {
1943         return __vm_insert_mixed(vma, addr, pfn, false);
1944
1945 }
1946 EXPORT_SYMBOL(vm_insert_mixed);
1947
1948 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1949                         pfn_t pfn)
1950 {
1951         return __vm_insert_mixed(vma, addr, pfn, true);
1952 }
1953 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1954
1955 /*
1956  * maps a range of physical memory into the requested pages. the old
1957  * mappings are removed. any references to nonexistent pages results
1958  * in null mappings (currently treated as "copy-on-access")
1959  */
1960 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1961                         unsigned long addr, unsigned long end,
1962                         unsigned long pfn, pgprot_t prot)
1963 {
1964         pte_t *pte;
1965         spinlock_t *ptl;
1966
1967         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1968         if (!pte)
1969                 return -ENOMEM;
1970         arch_enter_lazy_mmu_mode();
1971         do {
1972                 BUG_ON(!pte_none(*pte));
1973                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1974                 pfn++;
1975         } while (pte++, addr += PAGE_SIZE, addr != end);
1976         arch_leave_lazy_mmu_mode();
1977         pte_unmap_unlock(pte - 1, ptl);
1978         return 0;
1979 }
1980
1981 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1982                         unsigned long addr, unsigned long end,
1983                         unsigned long pfn, pgprot_t prot)
1984 {
1985         pmd_t *pmd;
1986         unsigned long next;
1987
1988         pfn -= addr >> PAGE_SHIFT;
1989         pmd = pmd_alloc(mm, pud, addr);
1990         if (!pmd)
1991                 return -ENOMEM;
1992         VM_BUG_ON(pmd_trans_huge(*pmd));
1993         do {
1994                 next = pmd_addr_end(addr, end);
1995                 if (remap_pte_range(mm, pmd, addr, next,
1996                                 pfn + (addr >> PAGE_SHIFT), prot))
1997                         return -ENOMEM;
1998         } while (pmd++, addr = next, addr != end);
1999         return 0;
2000 }
2001
2002 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2003                         unsigned long addr, unsigned long end,
2004                         unsigned long pfn, pgprot_t prot)
2005 {
2006         pud_t *pud;
2007         unsigned long next;
2008
2009         pfn -= addr >> PAGE_SHIFT;
2010         pud = pud_alloc(mm, p4d, addr);
2011         if (!pud)
2012                 return -ENOMEM;
2013         do {
2014                 next = pud_addr_end(addr, end);
2015                 if (remap_pmd_range(mm, pud, addr, next,
2016                                 pfn + (addr >> PAGE_SHIFT), prot))
2017                         return -ENOMEM;
2018         } while (pud++, addr = next, addr != end);
2019         return 0;
2020 }
2021
2022 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2023                         unsigned long addr, unsigned long end,
2024                         unsigned long pfn, pgprot_t prot)
2025 {
2026         p4d_t *p4d;
2027         unsigned long next;
2028
2029         pfn -= addr >> PAGE_SHIFT;
2030         p4d = p4d_alloc(mm, pgd, addr);
2031         if (!p4d)
2032                 return -ENOMEM;
2033         do {
2034                 next = p4d_addr_end(addr, end);
2035                 if (remap_pud_range(mm, p4d, addr, next,
2036                                 pfn + (addr >> PAGE_SHIFT), prot))
2037                         return -ENOMEM;
2038         } while (p4d++, addr = next, addr != end);
2039         return 0;
2040 }
2041
2042 /**
2043  * remap_pfn_range - remap kernel memory to userspace
2044  * @vma: user vma to map to
2045  * @addr: target user address to start at
2046  * @pfn: physical address of kernel memory
2047  * @size: size of map area
2048  * @prot: page protection flags for this mapping
2049  *
2050  *  Note: this is only safe if the mm semaphore is held when called.
2051  */
2052 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2053                     unsigned long pfn, unsigned long size, pgprot_t prot)
2054 {
2055         pgd_t *pgd;
2056         unsigned long next;
2057         unsigned long end = addr + PAGE_ALIGN(size);
2058         struct mm_struct *mm = vma->vm_mm;
2059         unsigned long remap_pfn = pfn;
2060         int err;
2061
2062         /*
2063          * Physically remapped pages are special. Tell the
2064          * rest of the world about it:
2065          *   VM_IO tells people not to look at these pages
2066          *      (accesses can have side effects).
2067          *   VM_PFNMAP tells the core MM that the base pages are just
2068          *      raw PFN mappings, and do not have a "struct page" associated
2069          *      with them.
2070          *   VM_DONTEXPAND
2071          *      Disable vma merging and expanding with mremap().
2072          *   VM_DONTDUMP
2073          *      Omit vma from core dump, even when VM_IO turned off.
2074          *
2075          * There's a horrible special case to handle copy-on-write
2076          * behaviour that some programs depend on. We mark the "original"
2077          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2078          * See vm_normal_page() for details.
2079          */
2080         if (is_cow_mapping(vma->vm_flags)) {
2081                 if (addr != vma->vm_start || end != vma->vm_end)
2082                         return -EINVAL;
2083                 vma->vm_pgoff = pfn;
2084         }
2085
2086         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2087         if (err)
2088                 return -EINVAL;
2089
2090         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2091
2092         BUG_ON(addr >= end);
2093         pfn -= addr >> PAGE_SHIFT;
2094         pgd = pgd_offset(mm, addr);
2095         flush_cache_range(vma, addr, end);
2096         do {
2097                 next = pgd_addr_end(addr, end);
2098                 err = remap_p4d_range(mm, pgd, addr, next,
2099                                 pfn + (addr >> PAGE_SHIFT), prot);
2100                 if (err)
2101                         break;
2102         } while (pgd++, addr = next, addr != end);
2103
2104         if (err)
2105                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2106
2107         return err;
2108 }
2109 EXPORT_SYMBOL(remap_pfn_range);
2110
2111 /**
2112  * vm_iomap_memory - remap memory to userspace
2113  * @vma: user vma to map to
2114  * @start: start of area
2115  * @len: size of area
2116  *
2117  * This is a simplified io_remap_pfn_range() for common driver use. The
2118  * driver just needs to give us the physical memory range to be mapped,
2119  * we'll figure out the rest from the vma information.
2120  *
2121  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2122  * whatever write-combining details or similar.
2123  */
2124 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2125 {
2126         unsigned long vm_len, pfn, pages;
2127
2128         /* Check that the physical memory area passed in looks valid */
2129         if (start + len < start)
2130                 return -EINVAL;
2131         /*
2132          * You *really* shouldn't map things that aren't page-aligned,
2133          * but we've historically allowed it because IO memory might
2134          * just have smaller alignment.
2135          */
2136         len += start & ~PAGE_MASK;
2137         pfn = start >> PAGE_SHIFT;
2138         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2139         if (pfn + pages < pfn)
2140                 return -EINVAL;
2141
2142         /* We start the mapping 'vm_pgoff' pages into the area */
2143         if (vma->vm_pgoff > pages)
2144                 return -EINVAL;
2145         pfn += vma->vm_pgoff;
2146         pages -= vma->vm_pgoff;
2147
2148         /* Can we fit all of the mapping? */
2149         vm_len = vma->vm_end - vma->vm_start;
2150         if (vm_len >> PAGE_SHIFT > pages)
2151                 return -EINVAL;
2152
2153         /* Ok, let it rip */
2154         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2155 }
2156 EXPORT_SYMBOL(vm_iomap_memory);
2157
2158 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2159                                      unsigned long addr, unsigned long end,
2160                                      pte_fn_t fn, void *data)
2161 {
2162         pte_t *pte;
2163         int err;
2164         pgtable_t token;
2165         spinlock_t *uninitialized_var(ptl);
2166
2167         pte = (mm == &init_mm) ?
2168                 pte_alloc_kernel(pmd, addr) :
2169                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2170         if (!pte)
2171                 return -ENOMEM;
2172
2173         BUG_ON(pmd_huge(*pmd));
2174
2175         arch_enter_lazy_mmu_mode();
2176
2177         token = pmd_pgtable(*pmd);
2178
2179         do {
2180                 err = fn(pte++, token, addr, data);
2181                 if (err)
2182                         break;
2183         } while (addr += PAGE_SIZE, addr != end);
2184
2185         arch_leave_lazy_mmu_mode();
2186
2187         if (mm != &init_mm)
2188                 pte_unmap_unlock(pte-1, ptl);
2189         return err;
2190 }
2191
2192 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2193                                      unsigned long addr, unsigned long end,
2194                                      pte_fn_t fn, void *data)
2195 {
2196         pmd_t *pmd;
2197         unsigned long next;
2198         int err;
2199
2200         BUG_ON(pud_huge(*pud));
2201
2202         pmd = pmd_alloc(mm, pud, addr);
2203         if (!pmd)
2204                 return -ENOMEM;
2205         do {
2206                 next = pmd_addr_end(addr, end);
2207                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2208                 if (err)
2209                         break;
2210         } while (pmd++, addr = next, addr != end);
2211         return err;
2212 }
2213
2214 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2215                                      unsigned long addr, unsigned long end,
2216                                      pte_fn_t fn, void *data)
2217 {
2218         pud_t *pud;
2219         unsigned long next;
2220         int err;
2221
2222         pud = pud_alloc(mm, p4d, addr);
2223         if (!pud)
2224                 return -ENOMEM;
2225         do {
2226                 next = pud_addr_end(addr, end);
2227                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2228                 if (err)
2229                         break;
2230         } while (pud++, addr = next, addr != end);
2231         return err;
2232 }
2233
2234 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2235                                      unsigned long addr, unsigned long end,
2236                                      pte_fn_t fn, void *data)
2237 {
2238         p4d_t *p4d;
2239         unsigned long next;
2240         int err;
2241
2242         p4d = p4d_alloc(mm, pgd, addr);
2243         if (!p4d)
2244                 return -ENOMEM;
2245         do {
2246                 next = p4d_addr_end(addr, end);
2247                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2248                 if (err)
2249                         break;
2250         } while (p4d++, addr = next, addr != end);
2251         return err;
2252 }
2253
2254 /*
2255  * Scan a region of virtual memory, filling in page tables as necessary
2256  * and calling a provided function on each leaf page table.
2257  */
2258 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2259                         unsigned long size, pte_fn_t fn, void *data)
2260 {
2261         pgd_t *pgd;
2262         unsigned long next;
2263         unsigned long end = addr + size;
2264         int err;
2265
2266         if (WARN_ON(addr >= end))
2267                 return -EINVAL;
2268
2269         pgd = pgd_offset(mm, addr);
2270         do {
2271                 next = pgd_addr_end(addr, end);
2272                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2273                 if (err)
2274                         break;
2275         } while (pgd++, addr = next, addr != end);
2276
2277         return err;
2278 }
2279 EXPORT_SYMBOL_GPL(apply_to_page_range);
2280
2281 /*
2282  * handle_pte_fault chooses page fault handler according to an entry which was
2283  * read non-atomically.  Before making any commitment, on those architectures
2284  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2285  * parts, do_swap_page must check under lock before unmapping the pte and
2286  * proceeding (but do_wp_page is only called after already making such a check;
2287  * and do_anonymous_page can safely check later on).
2288  */
2289 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2290                                 pte_t *page_table, pte_t orig_pte)
2291 {
2292         int same = 1;
2293 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2294         if (sizeof(pte_t) > sizeof(unsigned long)) {
2295                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2296                 spin_lock(ptl);
2297                 same = pte_same(*page_table, orig_pte);
2298                 spin_unlock(ptl);
2299         }
2300 #endif
2301         pte_unmap(page_table);
2302         return same;
2303 }
2304
2305 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2306 {
2307         debug_dma_assert_idle(src);
2308
2309         /*
2310          * If the source page was a PFN mapping, we don't have
2311          * a "struct page" for it. We do a best-effort copy by
2312          * just copying from the original user address. If that
2313          * fails, we just zero-fill it. Live with it.
2314          */
2315         if (unlikely(!src)) {
2316                 void *kaddr = kmap_atomic(dst);
2317                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2318
2319                 /*
2320                  * This really shouldn't fail, because the page is there
2321                  * in the page tables. But it might just be unreadable,
2322                  * in which case we just give up and fill the result with
2323                  * zeroes.
2324                  */
2325                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2326                         clear_page(kaddr);
2327                 kunmap_atomic(kaddr);
2328                 flush_dcache_page(dst);
2329         } else
2330                 copy_user_highpage(dst, src, va, vma);
2331 }
2332
2333 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2334 {
2335         struct file *vm_file = vma->vm_file;
2336
2337         if (vm_file)
2338                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2339
2340         /*
2341          * Special mappings (e.g. VDSO) do not have any file so fake
2342          * a default GFP_KERNEL for them.
2343          */
2344         return GFP_KERNEL;
2345 }
2346
2347 /*
2348  * Notify the address space that the page is about to become writable so that
2349  * it can prohibit this or wait for the page to get into an appropriate state.
2350  *
2351  * We do this without the lock held, so that it can sleep if it needs to.
2352  */
2353 static int do_page_mkwrite(struct vm_fault *vmf)
2354 {
2355         int ret;
2356         struct page *page = vmf->page;
2357         unsigned int old_flags = vmf->flags;
2358
2359         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2360
2361         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2362         /* Restore original flags so that caller is not surprised */
2363         vmf->flags = old_flags;
2364         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2365                 return ret;
2366         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2367                 lock_page(page);
2368                 if (!page->mapping) {
2369                         unlock_page(page);
2370                         return 0; /* retry */
2371                 }
2372                 ret |= VM_FAULT_LOCKED;
2373         } else
2374                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2375         return ret;
2376 }
2377
2378 /*
2379  * Handle dirtying of a page in shared file mapping on a write fault.
2380  *
2381  * The function expects the page to be locked and unlocks it.
2382  */
2383 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2384                                     struct page *page)
2385 {
2386         struct address_space *mapping;
2387         bool dirtied;
2388         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2389
2390         dirtied = set_page_dirty(page);
2391         VM_BUG_ON_PAGE(PageAnon(page), page);
2392         /*
2393          * Take a local copy of the address_space - page.mapping may be zeroed
2394          * by truncate after unlock_page().   The address_space itself remains
2395          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2396          * release semantics to prevent the compiler from undoing this copying.
2397          */
2398         mapping = page_rmapping(page);
2399         unlock_page(page);
2400
2401         if ((dirtied || page_mkwrite) && mapping) {
2402                 /*
2403                  * Some device drivers do not set page.mapping
2404                  * but still dirty their pages
2405                  */
2406                 balance_dirty_pages_ratelimited(mapping);
2407         }
2408
2409         if (!page_mkwrite)
2410                 file_update_time(vma->vm_file);
2411 }
2412
2413 /*
2414  * Handle write page faults for pages that can be reused in the current vma
2415  *
2416  * This can happen either due to the mapping being with the VM_SHARED flag,
2417  * or due to us being the last reference standing to the page. In either
2418  * case, all we need to do here is to mark the page as writable and update
2419  * any related book-keeping.
2420  */
2421 static inline void wp_page_reuse(struct vm_fault *vmf)
2422         __releases(vmf->ptl)
2423 {
2424         struct vm_area_struct *vma = vmf->vma;
2425         struct page *page = vmf->page;
2426         pte_t entry;
2427         /*
2428          * Clear the pages cpupid information as the existing
2429          * information potentially belongs to a now completely
2430          * unrelated process.
2431          */
2432         if (page)
2433                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2434
2435         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2436         entry = pte_mkyoung(vmf->orig_pte);
2437         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2438         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2439                 update_mmu_cache(vma, vmf->address, vmf->pte);
2440         pte_unmap_unlock(vmf->pte, vmf->ptl);
2441 }
2442
2443 /*
2444  * Handle the case of a page which we actually need to copy to a new page.
2445  *
2446  * Called with mmap_sem locked and the old page referenced, but
2447  * without the ptl held.
2448  *
2449  * High level logic flow:
2450  *
2451  * - Allocate a page, copy the content of the old page to the new one.
2452  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2453  * - Take the PTL. If the pte changed, bail out and release the allocated page
2454  * - If the pte is still the way we remember it, update the page table and all
2455  *   relevant references. This includes dropping the reference the page-table
2456  *   held to the old page, as well as updating the rmap.
2457  * - In any case, unlock the PTL and drop the reference we took to the old page.
2458  */
2459 static int wp_page_copy(struct vm_fault *vmf)
2460 {
2461         struct vm_area_struct *vma = vmf->vma;
2462         struct mm_struct *mm = vma->vm_mm;
2463         struct page *old_page = vmf->page;
2464         struct page *new_page = NULL;
2465         pte_t entry;
2466         int page_copied = 0;
2467         const unsigned long mmun_start = vmf->address & PAGE_MASK;
2468         const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2469         struct mem_cgroup *memcg;
2470
2471         if (unlikely(anon_vma_prepare(vma)))
2472                 goto oom;
2473
2474         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2475                 new_page = alloc_zeroed_user_highpage_movable(vma,
2476                                                               vmf->address);
2477                 if (!new_page)
2478                         goto oom;
2479         } else {
2480                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2481                                 vmf->address);
2482                 if (!new_page)
2483                         goto oom;
2484                 cow_user_page(new_page, old_page, vmf->address, vma);
2485         }
2486
2487         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2488                 goto oom_free_new;
2489
2490         __SetPageUptodate(new_page);
2491
2492         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2493
2494         /*
2495          * Re-check the pte - we dropped the lock
2496          */
2497         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2498         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2499                 if (old_page) {
2500                         if (!PageAnon(old_page)) {
2501                                 dec_mm_counter_fast(mm,
2502                                                 mm_counter_file(old_page));
2503                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2504                         }
2505                 } else {
2506                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2507                 }
2508                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2509                 entry = mk_pte(new_page, vma->vm_page_prot);
2510                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2511                 /*
2512                  * Clear the pte entry and flush it first, before updating the
2513                  * pte with the new entry. This will avoid a race condition
2514                  * seen in the presence of one thread doing SMC and another
2515                  * thread doing COW.
2516                  */
2517                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2518                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2519                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2520                 lru_cache_add_active_or_unevictable(new_page, vma);
2521                 /*
2522                  * We call the notify macro here because, when using secondary
2523                  * mmu page tables (such as kvm shadow page tables), we want the
2524                  * new page to be mapped directly into the secondary page table.
2525                  */
2526                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2527                 update_mmu_cache(vma, vmf->address, vmf->pte);
2528                 if (old_page) {
2529                         /*
2530                          * Only after switching the pte to the new page may
2531                          * we remove the mapcount here. Otherwise another
2532                          * process may come and find the rmap count decremented
2533                          * before the pte is switched to the new page, and
2534                          * "reuse" the old page writing into it while our pte
2535                          * here still points into it and can be read by other
2536                          * threads.
2537                          *
2538                          * The critical issue is to order this
2539                          * page_remove_rmap with the ptp_clear_flush above.
2540                          * Those stores are ordered by (if nothing else,)
2541                          * the barrier present in the atomic_add_negative
2542                          * in page_remove_rmap.
2543                          *
2544                          * Then the TLB flush in ptep_clear_flush ensures that
2545                          * no process can access the old page before the
2546                          * decremented mapcount is visible. And the old page
2547                          * cannot be reused until after the decremented
2548                          * mapcount is visible. So transitively, TLBs to
2549                          * old page will be flushed before it can be reused.
2550                          */
2551                         page_remove_rmap(old_page, false);
2552                 }
2553
2554                 /* Free the old page.. */
2555                 new_page = old_page;
2556                 page_copied = 1;
2557         } else {
2558                 mem_cgroup_cancel_charge(new_page, memcg, false);
2559         }
2560
2561         if (new_page)
2562                 put_page(new_page);
2563
2564         pte_unmap_unlock(vmf->pte, vmf->ptl);
2565         /*
2566          * No need to double call mmu_notifier->invalidate_range() callback as
2567          * the above ptep_clear_flush_notify() did already call it.
2568          */
2569         mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2570         if (old_page) {
2571                 /*
2572                  * Don't let another task, with possibly unlocked vma,
2573                  * keep the mlocked page.
2574                  */
2575                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2576                         lock_page(old_page);    /* LRU manipulation */
2577                         if (PageMlocked(old_page))
2578                                 munlock_vma_page(old_page);
2579                         unlock_page(old_page);
2580                 }
2581                 put_page(old_page);
2582         }
2583         return page_copied ? VM_FAULT_WRITE : 0;
2584 oom_free_new:
2585         put_page(new_page);
2586 oom:
2587         if (old_page)
2588                 put_page(old_page);
2589         return VM_FAULT_OOM;
2590 }
2591
2592 /**
2593  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2594  *                        writeable once the page is prepared
2595  *
2596  * @vmf: structure describing the fault
2597  *
2598  * This function handles all that is needed to finish a write page fault in a
2599  * shared mapping due to PTE being read-only once the mapped page is prepared.
2600  * It handles locking of PTE and modifying it. The function returns
2601  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2602  * lock.
2603  *
2604  * The function expects the page to be locked or other protection against
2605  * concurrent faults / writeback (such as DAX radix tree locks).
2606  */
2607 int finish_mkwrite_fault(struct vm_fault *vmf)
2608 {
2609         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2610         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2611                                        &vmf->ptl);
2612         /*
2613          * We might have raced with another page fault while we released the
2614          * pte_offset_map_lock.
2615          */
2616         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2617                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2618                 return VM_FAULT_NOPAGE;
2619         }
2620         wp_page_reuse(vmf);
2621         return 0;
2622 }
2623
2624 /*
2625  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2626  * mapping
2627  */
2628 static int wp_pfn_shared(struct vm_fault *vmf)
2629 {
2630         struct vm_area_struct *vma = vmf->vma;
2631
2632         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2633                 int ret;
2634
2635                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2636                 vmf->flags |= FAULT_FLAG_MKWRITE;
2637                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2638                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2639                         return ret;
2640                 return finish_mkwrite_fault(vmf);
2641         }
2642         wp_page_reuse(vmf);
2643         return VM_FAULT_WRITE;
2644 }
2645
2646 static int wp_page_shared(struct vm_fault *vmf)
2647         __releases(vmf->ptl)
2648 {
2649         struct vm_area_struct *vma = vmf->vma;
2650
2651         get_page(vmf->page);
2652
2653         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2654                 int tmp;
2655
2656                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2657                 tmp = do_page_mkwrite(vmf);
2658                 if (unlikely(!tmp || (tmp &
2659                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2660                         put_page(vmf->page);
2661                         return tmp;
2662                 }
2663                 tmp = finish_mkwrite_fault(vmf);
2664                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2665                         unlock_page(vmf->page);
2666                         put_page(vmf->page);
2667                         return tmp;
2668                 }
2669         } else {
2670                 wp_page_reuse(vmf);
2671                 lock_page(vmf->page);
2672         }
2673         fault_dirty_shared_page(vma, vmf->page);
2674         put_page(vmf->page);
2675
2676         return VM_FAULT_WRITE;
2677 }
2678
2679 /*
2680  * This routine handles present pages, when users try to write
2681  * to a shared page. It is done by copying the page to a new address
2682  * and decrementing the shared-page counter for the old page.
2683  *
2684  * Note that this routine assumes that the protection checks have been
2685  * done by the caller (the low-level page fault routine in most cases).
2686  * Thus we can safely just mark it writable once we've done any necessary
2687  * COW.
2688  *
2689  * We also mark the page dirty at this point even though the page will
2690  * change only once the write actually happens. This avoids a few races,
2691  * and potentially makes it more efficient.
2692  *
2693  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2694  * but allow concurrent faults), with pte both mapped and locked.
2695  * We return with mmap_sem still held, but pte unmapped and unlocked.
2696  */
2697 static int do_wp_page(struct vm_fault *vmf)
2698         __releases(vmf->ptl)
2699 {
2700         struct vm_area_struct *vma = vmf->vma;
2701
2702         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2703         if (!vmf->page) {
2704                 /*
2705                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2706                  * VM_PFNMAP VMA.
2707                  *
2708                  * We should not cow pages in a shared writeable mapping.
2709                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2710                  */
2711                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2712                                      (VM_WRITE|VM_SHARED))
2713                         return wp_pfn_shared(vmf);
2714
2715                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2716                 return wp_page_copy(vmf);
2717         }
2718
2719         /*
2720          * Take out anonymous pages first, anonymous shared vmas are
2721          * not dirty accountable.
2722          */
2723         if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2724                 int total_map_swapcount;
2725                 if (!trylock_page(vmf->page)) {
2726                         get_page(vmf->page);
2727                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2728                         lock_page(vmf->page);
2729                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2730                                         vmf->address, &vmf->ptl);
2731                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2732                                 unlock_page(vmf->page);
2733                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2734                                 put_page(vmf->page);
2735                                 return 0;
2736                         }
2737                         put_page(vmf->page);
2738                 }
2739                 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2740                         if (total_map_swapcount == 1) {
2741                                 /*
2742                                  * The page is all ours. Move it to
2743                                  * our anon_vma so the rmap code will
2744                                  * not search our parent or siblings.
2745                                  * Protected against the rmap code by
2746                                  * the page lock.
2747                                  */
2748                                 page_move_anon_rmap(vmf->page, vma);
2749                         }
2750                         unlock_page(vmf->page);
2751                         wp_page_reuse(vmf);
2752                         return VM_FAULT_WRITE;
2753                 }
2754                 unlock_page(vmf->page);
2755         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2756                                         (VM_WRITE|VM_SHARED))) {
2757                 return wp_page_shared(vmf);
2758         }
2759
2760         /*
2761          * Ok, we need to copy. Oh, well..
2762          */
2763         get_page(vmf->page);
2764
2765         pte_unmap_unlock(vmf->pte, vmf->ptl);
2766         return wp_page_copy(vmf);
2767 }
2768
2769 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2770                 unsigned long start_addr, unsigned long end_addr,
2771                 struct zap_details *details)
2772 {
2773         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2774 }
2775
2776 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2777                                             struct zap_details *details)
2778 {
2779         struct vm_area_struct *vma;
2780         pgoff_t vba, vea, zba, zea;
2781
2782         vma_interval_tree_foreach(vma, root,
2783                         details->first_index, details->last_index) {
2784
2785                 vba = vma->vm_pgoff;
2786                 vea = vba + vma_pages(vma) - 1;
2787                 zba = details->first_index;
2788                 if (zba < vba)
2789                         zba = vba;
2790                 zea = details->last_index;
2791                 if (zea > vea)
2792                         zea = vea;
2793
2794                 unmap_mapping_range_vma(vma,
2795                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2796                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2797                                 details);
2798         }
2799 }
2800
2801 /**
2802  * unmap_mapping_pages() - Unmap pages from processes.
2803  * @mapping: The address space containing pages to be unmapped.
2804  * @start: Index of first page to be unmapped.
2805  * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
2806  * @even_cows: Whether to unmap even private COWed pages.
2807  *
2808  * Unmap the pages in this address space from any userspace process which
2809  * has them mmaped.  Generally, you want to remove COWed pages as well when
2810  * a file is being truncated, but not when invalidating pages from the page
2811  * cache.
2812  */
2813 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2814                 pgoff_t nr, bool even_cows)
2815 {
2816         struct zap_details details = { };
2817
2818         details.check_mapping = even_cows ? NULL : mapping;
2819         details.first_index = start;
2820         details.last_index = start + nr - 1;
2821         if (details.last_index < details.first_index)
2822                 details.last_index = ULONG_MAX;
2823
2824         i_mmap_lock_write(mapping);
2825         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2826                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2827         i_mmap_unlock_write(mapping);
2828 }
2829
2830 /**
2831  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2832  * address_space corresponding to the specified byte range in the underlying
2833  * file.
2834  *
2835  * @mapping: the address space containing mmaps to be unmapped.
2836  * @holebegin: byte in first page to unmap, relative to the start of
2837  * the underlying file.  This will be rounded down to a PAGE_SIZE
2838  * boundary.  Note that this is different from truncate_pagecache(), which
2839  * must keep the partial page.  In contrast, we must get rid of
2840  * partial pages.
2841  * @holelen: size of prospective hole in bytes.  This will be rounded
2842  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2843  * end of the file.
2844  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2845  * but 0 when invalidating pagecache, don't throw away private data.
2846  */
2847 void unmap_mapping_range(struct address_space *mapping,
2848                 loff_t const holebegin, loff_t const holelen, int even_cows)
2849 {
2850         pgoff_t hba = holebegin >> PAGE_SHIFT;
2851         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2852
2853         /* Check for overflow. */
2854         if (sizeof(holelen) > sizeof(hlen)) {
2855                 long long holeend =
2856                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2857                 if (holeend & ~(long long)ULONG_MAX)
2858                         hlen = ULONG_MAX - hba + 1;
2859         }
2860
2861         unmap_mapping_pages(mapping, hba, hlen, even_cows);
2862 }
2863 EXPORT_SYMBOL(unmap_mapping_range);
2864
2865 /*
2866  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2867  * but allow concurrent faults), and pte mapped but not yet locked.
2868  * We return with pte unmapped and unlocked.
2869  *
2870  * We return with the mmap_sem locked or unlocked in the same cases
2871  * as does filemap_fault().
2872  */
2873 int do_swap_page(struct vm_fault *vmf)
2874 {
2875         struct vm_area_struct *vma = vmf->vma;
2876         struct page *page = NULL, *swapcache = NULL;
2877         struct mem_cgroup *memcg;
2878         struct vma_swap_readahead swap_ra;
2879         swp_entry_t entry;
2880         pte_t pte;
2881         int locked;
2882         int exclusive = 0;
2883         int ret = 0;
2884         bool vma_readahead = swap_use_vma_readahead();
2885
2886         if (vma_readahead) {
2887                 page = swap_readahead_detect(vmf, &swap_ra);
2888                 swapcache = page;
2889         }
2890
2891         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
2892                 if (page)
2893                         put_page(page);
2894                 goto out;
2895         }
2896
2897         entry = pte_to_swp_entry(vmf->orig_pte);
2898         if (unlikely(non_swap_entry(entry))) {
2899                 if (is_migration_entry(entry)) {
2900                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2901                                              vmf->address);
2902                 } else if (is_device_private_entry(entry)) {
2903                         /*
2904                          * For un-addressable device memory we call the pgmap
2905                          * fault handler callback. The callback must migrate
2906                          * the page back to some CPU accessible page.
2907                          */
2908                         ret = device_private_entry_fault(vma, vmf->address, entry,
2909                                                  vmf->flags, vmf->pmd);
2910                 } else if (is_hwpoison_entry(entry)) {
2911                         ret = VM_FAULT_HWPOISON;
2912                 } else {
2913                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2914                         ret = VM_FAULT_SIGBUS;
2915                 }
2916                 goto out;
2917         }
2918
2919
2920         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2921         if (!page) {
2922                 page = lookup_swap_cache(entry, vma_readahead ? vma : NULL,
2923                                          vmf->address);
2924                 swapcache = page;
2925         }
2926
2927         if (!page) {
2928                 struct swap_info_struct *si = swp_swap_info(entry);
2929
2930                 if (si->flags & SWP_SYNCHRONOUS_IO &&
2931                                 __swap_count(si, entry) == 1) {
2932                         /* skip swapcache */
2933                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2934                         if (page) {
2935                                 __SetPageLocked(page);
2936                                 __SetPageSwapBacked(page);
2937                                 set_page_private(page, entry.val);
2938                                 lru_cache_add_anon(page);
2939                                 swap_readpage(page, true);
2940                         }
2941                 } else {
2942                         if (vma_readahead)
2943                                 page = do_swap_page_readahead(entry,
2944                                         GFP_HIGHUSER_MOVABLE, vmf, &swap_ra);
2945                         else
2946                                 page = swapin_readahead(entry,
2947                                        GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2948                         swapcache = page;
2949                 }
2950
2951                 if (!page) {
2952                         /*
2953                          * Back out if somebody else faulted in this pte
2954                          * while we released the pte lock.
2955                          */
2956                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2957                                         vmf->address, &vmf->ptl);
2958                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2959                                 ret = VM_FAULT_OOM;
2960                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2961                         goto unlock;
2962                 }
2963
2964                 /* Had to read the page from swap area: Major fault */
2965                 ret = VM_FAULT_MAJOR;
2966                 count_vm_event(PGMAJFAULT);
2967                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2968         } else if (PageHWPoison(page)) {
2969                 /*
2970                  * hwpoisoned dirty swapcache pages are kept for killing
2971                  * owner processes (which may be unknown at hwpoison time)
2972                  */
2973                 ret = VM_FAULT_HWPOISON;
2974                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2975                 swapcache = page;
2976                 goto out_release;
2977         }
2978
2979         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2980
2981         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2982         if (!locked) {
2983                 ret |= VM_FAULT_RETRY;
2984                 goto out_release;
2985         }
2986
2987         /*
2988          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2989          * release the swapcache from under us.  The page pin, and pte_same
2990          * test below, are not enough to exclude that.  Even if it is still
2991          * swapcache, we need to check that the page's swap has not changed.
2992          */
2993         if (unlikely((!PageSwapCache(page) ||
2994                         page_private(page) != entry.val)) && swapcache)
2995                 goto out_page;
2996
2997         page = ksm_might_need_to_copy(page, vma, vmf->address);
2998         if (unlikely(!page)) {
2999                 ret = VM_FAULT_OOM;
3000                 page = swapcache;
3001                 goto out_page;
3002         }
3003
3004         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
3005                                 &memcg, false)) {
3006                 ret = VM_FAULT_OOM;
3007                 goto out_page;
3008         }
3009
3010         /*
3011          * Back out if somebody else already faulted in this pte.
3012          */
3013         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3014                         &vmf->ptl);
3015         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3016                 goto out_nomap;
3017
3018         if (unlikely(!PageUptodate(page))) {
3019                 ret = VM_FAULT_SIGBUS;
3020                 goto out_nomap;
3021         }
3022
3023         /*
3024          * The page isn't present yet, go ahead with the fault.
3025          *
3026          * Be careful about the sequence of operations here.
3027          * To get its accounting right, reuse_swap_page() must be called
3028          * while the page is counted on swap but not yet in mapcount i.e.
3029          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3030          * must be called after the swap_free(), or it will never succeed.
3031          */
3032
3033         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3034         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3035         pte = mk_pte(page, vma->vm_page_prot);
3036         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3037                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3038                 vmf->flags &= ~FAULT_FLAG_WRITE;
3039                 ret |= VM_FAULT_WRITE;
3040                 exclusive = RMAP_EXCLUSIVE;
3041         }
3042         flush_icache_page(vma, page);
3043         if (pte_swp_soft_dirty(vmf->orig_pte))
3044                 pte = pte_mksoft_dirty(pte);
3045         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3046         vmf->orig_pte = pte;
3047
3048         /* ksm created a completely new copy */
3049         if (unlikely(page != swapcache && swapcache)) {
3050                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3051                 mem_cgroup_commit_charge(page, memcg, false, false);
3052                 lru_cache_add_active_or_unevictable(page, vma);
3053         } else {
3054                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3055                 mem_cgroup_commit_charge(page, memcg, true, false);
3056                 activate_page(page);
3057         }
3058
3059         swap_free(entry);
3060         if (mem_cgroup_swap_full(page) ||
3061             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3062                 try_to_free_swap(page);
3063         unlock_page(page);
3064         if (page != swapcache && swapcache) {
3065                 /*
3066                  * Hold the lock to avoid the swap entry to be reused
3067                  * until we take the PT lock for the pte_same() check
3068                  * (to avoid false positives from pte_same). For
3069                  * further safety release the lock after the swap_free
3070                  * so that the swap count won't change under a
3071                  * parallel locked swapcache.
3072                  */
3073                 unlock_page(swapcache);
3074                 put_page(swapcache);
3075         }
3076
3077         if (vmf->flags & FAULT_FLAG_WRITE) {
3078                 ret |= do_wp_page(vmf);
3079                 if (ret & VM_FAULT_ERROR)
3080                         ret &= VM_FAULT_ERROR;
3081                 goto out;
3082         }
3083
3084         /* No need to invalidate - it was non-present before */
3085         update_mmu_cache(vma, vmf->address, vmf->pte);
3086 unlock:
3087         pte_unmap_unlock(vmf->pte, vmf->ptl);
3088 out:
3089         return ret;
3090 out_nomap:
3091         mem_cgroup_cancel_charge(page, memcg, false);
3092         pte_unmap_unlock(vmf->pte, vmf->ptl);
3093 out_page:
3094         unlock_page(page);
3095 out_release:
3096         put_page(page);
3097         if (page != swapcache && swapcache) {
3098                 unlock_page(swapcache);
3099                 put_page(swapcache);
3100         }
3101         return ret;
3102 }
3103
3104 /*
3105  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3106  * but allow concurrent faults), and pte mapped but not yet locked.
3107  * We return with mmap_sem still held, but pte unmapped and unlocked.
3108  */
3109 static int do_anonymous_page(struct vm_fault *vmf)
3110 {
3111         struct vm_area_struct *vma = vmf->vma;
3112         struct mem_cgroup *memcg;
3113         struct page *page;
3114         int ret = 0;
3115         pte_t entry;
3116
3117         /* File mapping without ->vm_ops ? */
3118         if (vma->vm_flags & VM_SHARED)
3119                 return VM_FAULT_SIGBUS;
3120
3121         /*
3122          * Use pte_alloc() instead of pte_alloc_map().  We can't run
3123          * pte_offset_map() on pmds where a huge pmd might be created
3124          * from a different thread.
3125          *
3126          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3127          * parallel threads are excluded by other means.
3128          *
3129          * Here we only have down_read(mmap_sem).
3130          */
3131         if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3132                 return VM_FAULT_OOM;
3133
3134         /* See the comment in pte_alloc_one_map() */
3135         if (unlikely(pmd_trans_unstable(vmf->pmd)))
3136                 return 0;
3137
3138         /* Use the zero-page for reads */
3139         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3140                         !mm_forbids_zeropage(vma->vm_mm)) {
3141                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3142                                                 vma->vm_page_prot));
3143                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3144                                 vmf->address, &vmf->ptl);
3145                 if (!pte_none(*vmf->pte))
3146                         goto unlock;
3147                 ret = check_stable_address_space(vma->vm_mm);
3148                 if (ret)
3149                         goto unlock;
3150                 /* Deliver the page fault to userland, check inside PT lock */
3151                 if (userfaultfd_missing(vma)) {
3152                         pte_unmap_unlock(vmf->pte, vmf->ptl);
3153                         return handle_userfault(vmf, VM_UFFD_MISSING);
3154                 }
3155                 goto setpte;
3156         }
3157
3158         /* Allocate our own private page. */
3159         if (unlikely(anon_vma_prepare(vma)))
3160                 goto oom;
3161         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3162         if (!page)
3163                 goto oom;
3164
3165         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3166                 goto oom_free_page;
3167
3168         /*
3169          * The memory barrier inside __SetPageUptodate makes sure that
3170          * preceeding stores to the page contents become visible before
3171          * the set_pte_at() write.
3172          */
3173         __SetPageUptodate(page);
3174
3175         entry = mk_pte(page, vma->vm_page_prot);
3176         if (vma->vm_flags & VM_WRITE)
3177                 entry = pte_mkwrite(pte_mkdirty(entry));
3178
3179         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3180                         &vmf->ptl);
3181         if (!pte_none(*vmf->pte))
3182                 goto release;
3183
3184         ret = check_stable_address_space(vma->vm_mm);
3185         if (ret)
3186                 goto release;
3187
3188         /* Deliver the page fault to userland, check inside PT lock */
3189         if (userfaultfd_missing(vma)) {
3190                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3191                 mem_cgroup_cancel_charge(page, memcg, false);
3192                 put_page(page);
3193                 return handle_userfault(vmf, VM_UFFD_MISSING);
3194         }
3195
3196         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3197         page_add_new_anon_rmap(page, vma, vmf->address, false);
3198         mem_cgroup_commit_charge(page, memcg, false, false);
3199         lru_cache_add_active_or_unevictable(page, vma);
3200 setpte:
3201         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3202
3203         /* No need to invalidate - it was non-present before */
3204         update_mmu_cache(vma, vmf->address, vmf->pte);
3205 unlock:
3206         pte_unmap_unlock(vmf->pte, vmf->ptl);
3207         return ret;
3208 release:
3209         mem_cgroup_cancel_charge(page, memcg, false);
3210         put_page(page);
3211         goto unlock;
3212 oom_free_page:
3213         put_page(page);
3214 oom:
3215         return VM_FAULT_OOM;
3216 }
3217
3218 /*
3219  * The mmap_sem must have been held on entry, and may have been
3220  * released depending on flags and vma->vm_ops->fault() return value.
3221  * See filemap_fault() and __lock_page_retry().
3222  */
3223 static int __do_fault(struct vm_fault *vmf)
3224 {
3225         struct vm_area_struct *vma = vmf->vma;
3226         int ret;
3227
3228         ret = vma->vm_ops->fault(vmf);
3229         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3230                             VM_FAULT_DONE_COW)))
3231                 return ret;
3232
3233         if (unlikely(PageHWPoison(vmf->page))) {
3234                 if (ret & VM_FAULT_LOCKED)
3235                         unlock_page(vmf->page);
3236                 put_page(vmf->page);
3237                 vmf->page = NULL;
3238                 return VM_FAULT_HWPOISON;
3239         }
3240
3241         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3242                 lock_page(vmf->page);
3243         else
3244                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3245
3246         return ret;
3247 }
3248
3249 /*
3250  * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3251  * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3252  * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3253  * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3254  */
3255 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3256 {
3257         return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3258 }
3259
3260 static int pte_alloc_one_map(struct vm_fault *vmf)
3261 {
3262         struct vm_area_struct *vma = vmf->vma;
3263
3264         if (!pmd_none(*vmf->pmd))
3265                 goto map_pte;
3266         if (vmf->prealloc_pte) {
3267                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3268                 if (unlikely(!pmd_none(*vmf->pmd))) {
3269                         spin_unlock(vmf->ptl);
3270                         goto map_pte;
3271                 }
3272
3273                 mm_inc_nr_ptes(vma->vm_mm);
3274                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3275                 spin_unlock(vmf->ptl);
3276                 vmf->prealloc_pte = NULL;
3277         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3278                 return VM_FAULT_OOM;
3279         }
3280 map_pte:
3281         /*
3282          * If a huge pmd materialized under us just retry later.  Use
3283          * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3284          * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3285          * under us and then back to pmd_none, as a result of MADV_DONTNEED
3286          * running immediately after a huge pmd fault in a different thread of
3287          * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3288          * All we have to ensure is that it is a regular pmd that we can walk
3289          * with pte_offset_map() and we can do that through an atomic read in
3290          * C, which is what pmd_trans_unstable() provides.
3291          */
3292         if (pmd_devmap_trans_unstable(vmf->pmd))
3293                 return VM_FAULT_NOPAGE;
3294
3295         /*
3296          * At this point we know that our vmf->pmd points to a page of ptes
3297          * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3298          * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3299          * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3300          * be valid and we will re-check to make sure the vmf->pte isn't
3301          * pte_none() under vmf->ptl protection when we return to
3302          * alloc_set_pte().
3303          */
3304         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3305                         &vmf->ptl);
3306         return 0;
3307 }
3308
3309 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3310
3311 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3312 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3313                 unsigned long haddr)
3314 {
3315         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3316                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3317                 return false;
3318         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3319                 return false;
3320         return true;
3321 }
3322
3323 static void deposit_prealloc_pte(struct vm_fault *vmf)
3324 {
3325         struct vm_area_struct *vma = vmf->vma;
3326
3327         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3328         /*
3329          * We are going to consume the prealloc table,
3330          * count that as nr_ptes.
3331          */
3332         mm_inc_nr_ptes(vma->vm_mm);
3333         vmf->prealloc_pte = NULL;
3334 }
3335
3336 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3337 {
3338         struct vm_area_struct *vma = vmf->vma;
3339         bool write = vmf->flags & FAULT_FLAG_WRITE;
3340         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3341         pmd_t entry;
3342         int i, ret;
3343
3344         if (!transhuge_vma_suitable(vma, haddr))
3345                 return VM_FAULT_FALLBACK;
3346
3347         ret = VM_FAULT_FALLBACK;
3348         page = compound_head(page);
3349
3350         /*
3351          * Archs like ppc64 need additonal space to store information
3352          * related to pte entry. Use the preallocated table for that.
3353          */
3354         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3355                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3356                 if (!vmf->prealloc_pte)
3357                         return VM_FAULT_OOM;
3358                 smp_wmb(); /* See comment in __pte_alloc() */
3359         }
3360
3361         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3362         if (unlikely(!pmd_none(*vmf->pmd)))
3363                 goto out;
3364
3365         for (i = 0; i < HPAGE_PMD_NR; i++)
3366                 flush_icache_page(vma, page + i);
3367
3368         entry = mk_huge_pmd(page, vma->vm_page_prot);
3369         if (write)
3370                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3371
3372         add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3373         page_add_file_rmap(page, true);
3374         /*
3375          * deposit and withdraw with pmd lock held
3376          */
3377         if (arch_needs_pgtable_deposit())
3378                 deposit_prealloc_pte(vmf);
3379
3380         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3381
3382         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3383
3384         /* fault is handled */
3385         ret = 0;
3386         count_vm_event(THP_FILE_MAPPED);
3387 out:
3388         spin_unlock(vmf->ptl);
3389         return ret;
3390 }
3391 #else
3392 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3393 {
3394         BUILD_BUG();
3395         return 0;
3396 }
3397 #endif
3398
3399 /**
3400  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3401  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3402  *
3403  * @vmf: fault environment
3404  * @memcg: memcg to charge page (only for private mappings)
3405  * @page: page to map
3406  *
3407  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3408  * return.
3409  *
3410  * Target users are page handler itself and implementations of
3411  * vm_ops->map_pages.
3412  */
3413 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3414                 struct page *page)
3415 {
3416         struct vm_area_struct *vma = vmf->vma;
3417         bool write = vmf->flags & FAULT_FLAG_WRITE;
3418         pte_t entry;
3419         int ret;
3420
3421         if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3422                         IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3423                 /* THP on COW? */
3424                 VM_BUG_ON_PAGE(memcg, page);
3425
3426                 ret = do_set_pmd(vmf, page);
3427                 if (ret != VM_FAULT_FALLBACK)
3428                         return ret;
3429         }
3430
3431         if (!vmf->pte) {
3432                 ret = pte_alloc_one_map(vmf);
3433                 if (ret)
3434                         return ret;
3435         }
3436
3437         /* Re-check under ptl */
3438         if (unlikely(!pte_none(*vmf->pte)))
3439                 return VM_FAULT_NOPAGE;
3440
3441         flush_icache_page(vma, page);
3442         entry = mk_pte(page, vma->vm_page_prot);
3443         if (write)
3444                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3445         /* copy-on-write page */
3446         if (write && !(vma->vm_flags & VM_SHARED)) {
3447                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3448                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3449                 mem_cgroup_commit_charge(page, memcg, false, false);
3450                 lru_cache_add_active_or_unevictable(page, vma);
3451         } else {
3452                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3453                 page_add_file_rmap(page, false);
3454         }
3455         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3456
3457         /* no need to invalidate: a not-present page won't be cached */
3458         update_mmu_cache(vma, vmf->address, vmf->pte);
3459
3460         return 0;
3461 }
3462
3463
3464 /**
3465  * finish_fault - finish page fault once we have prepared the page to fault
3466  *
3467  * @vmf: structure describing the fault
3468  *
3469  * This function handles all that is needed to finish a page fault once the
3470  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3471  * given page, adds reverse page mapping, handles memcg charges and LRU
3472  * addition. The function returns 0 on success, VM_FAULT_ code in case of
3473  * error.
3474  *
3475  * The function expects the page to be locked and on success it consumes a
3476  * reference of a page being mapped (for the PTE which maps it).
3477  */
3478 int finish_fault(struct vm_fault *vmf)
3479 {
3480         struct page *page;
3481         int ret = 0;
3482
3483         /* Did we COW the page? */
3484         if ((vmf->flags & FAULT_FLAG_WRITE) &&
3485             !(vmf->vma->vm_flags & VM_SHARED))
3486                 page = vmf->cow_page;
3487         else
3488                 page = vmf->page;
3489
3490         /*
3491          * check even for read faults because we might have lost our CoWed
3492          * page
3493          */
3494         if (!(vmf->vma->vm_flags & VM_SHARED))
3495                 ret = check_stable_address_space(vmf->vma->vm_mm);
3496         if (!ret)
3497                 ret = alloc_set_pte(vmf, vmf->memcg, page);
3498         if (vmf->pte)
3499                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3500         return ret;
3501 }
3502
3503 static unsigned long fault_around_bytes __read_mostly =
3504         rounddown_pow_of_two(65536);
3505
3506 #ifdef CONFIG_DEBUG_FS
3507 static int fault_around_bytes_get(void *data, u64 *val)
3508 {
3509         *val = fault_around_bytes;
3510         return 0;
3511 }
3512
3513 /*
3514  * fault_around_bytes must be rounded down to the nearest page order as it's
3515  * what do_fault_around() expects to see.
3516  */
3517 static int fault_around_bytes_set(void *data, u64 val)
3518 {
3519         if (val / PAGE_SIZE > PTRS_PER_PTE)
3520                 return -EINVAL;
3521         if (val > PAGE_SIZE)
3522                 fault_around_bytes = rounddown_pow_of_two(val);
3523         else
3524                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3525         return 0;
3526 }
3527 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3528                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3529
3530 static int __init fault_around_debugfs(void)
3531 {
3532         void *ret;
3533
3534         ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3535                         &fault_around_bytes_fops);
3536         if (!ret)
3537                 pr_warn("Failed to create fault_around_bytes in debugfs");
3538         return 0;
3539 }
3540 late_initcall(fault_around_debugfs);
3541 #endif
3542
3543 /*
3544  * do_fault_around() tries to map few pages around the fault address. The hope
3545  * is that the pages will be needed soon and this will lower the number of
3546  * faults to handle.
3547  *
3548  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3549  * not ready to be mapped: not up-to-date, locked, etc.
3550  *
3551  * This function is called with the page table lock taken. In the split ptlock
3552  * case the page table lock only protects only those entries which belong to
3553  * the page table corresponding to the fault address.
3554  *
3555  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3556  * only once.
3557  *
3558  * fault_around_bytes defines how many bytes we'll try to map.
3559  * do_fault_around() expects it to be set to a power of two less than or equal
3560  * to PTRS_PER_PTE.
3561  *
3562  * The virtual address of the area that we map is naturally aligned to
3563  * fault_around_bytes rounded down to the machine page size
3564  * (and therefore to page order).  This way it's easier to guarantee
3565  * that we don't cross page table boundaries.
3566  */
3567 static int do_fault_around(struct vm_fault *vmf)
3568 {
3569         unsigned long address = vmf->address, nr_pages, mask;
3570         pgoff_t start_pgoff = vmf->pgoff;
3571         pgoff_t end_pgoff;
3572         int off, ret = 0;
3573
3574         nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3575         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3576
3577         vmf->address = max(address & mask, vmf->vma->vm_start);
3578         off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3579         start_pgoff -= off;
3580
3581         /*
3582          *  end_pgoff is either the end of the page table, the end of
3583          *  the vma or nr_pages from start_pgoff, depending what is nearest.
3584          */
3585         end_pgoff = start_pgoff -
3586                 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3587                 PTRS_PER_PTE - 1;
3588         end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3589                         start_pgoff + nr_pages - 1);
3590
3591         if (pmd_none(*vmf->pmd)) {
3592                 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3593                                                   vmf->address);
3594                 if (!vmf->prealloc_pte)
3595                         goto out;
3596                 smp_wmb(); /* See comment in __pte_alloc() */
3597         }
3598
3599         vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3600
3601         /* Huge page is mapped? Page fault is solved */
3602         if (pmd_trans_huge(*vmf->pmd)) {
3603                 ret = VM_FAULT_NOPAGE;
3604                 goto out;
3605         }
3606
3607         /* ->map_pages() haven't done anything useful. Cold page cache? */
3608         if (!vmf->pte)
3609                 goto out;
3610
3611         /* check if the page fault is solved */
3612         vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3613         if (!pte_none(*vmf->pte))
3614                 ret = VM_FAULT_NOPAGE;
3615         pte_unmap_unlock(vmf->pte, vmf->ptl);
3616 out:
3617         vmf->address = address;
3618         vmf->pte = NULL;
3619         return ret;
3620 }
3621
3622 static int do_read_fault(struct vm_fault *vmf)
3623 {
3624         struct vm_area_struct *vma = vmf->vma;
3625         int ret = 0;
3626
3627         /*
3628          * Let's call ->map_pages() first and use ->fault() as fallback
3629          * if page by the offset is not ready to be mapped (cold cache or
3630          * something).
3631          */
3632         if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3633                 ret = do_fault_around(vmf);
3634                 if (ret)
3635                         return ret;
3636         }
3637
3638         ret = __do_fault(vmf);
3639         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3640                 return ret;
3641
3642         ret |= finish_fault(vmf);
3643         unlock_page(vmf->page);
3644         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3645                 put_page(vmf->page);
3646         return ret;
3647 }
3648
3649 static int do_cow_fault(struct vm_fault *vmf)
3650 {
3651         struct vm_area_struct *vma = vmf->vma;
3652         int ret;
3653
3654         if (unlikely(anon_vma_prepare(vma)))
3655                 return VM_FAULT_OOM;
3656
3657         vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3658         if (!vmf->cow_page)
3659                 return VM_FAULT_OOM;
3660
3661         if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3662                                 &vmf->memcg, false)) {
3663                 put_page(vmf->cow_page);
3664                 return VM_FAULT_OOM;
3665         }
3666
3667         ret = __do_fault(vmf);
3668         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3669                 goto uncharge_out;
3670         if (ret & VM_FAULT_DONE_COW)
3671                 return ret;
3672
3673         copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3674         __SetPageUptodate(vmf->cow_page);
3675
3676         ret |= finish_fault(vmf);
3677         unlock_page(vmf->page);
3678         put_page(vmf->page);
3679         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3680                 goto uncharge_out;
3681         return ret;
3682 uncharge_out:
3683         mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3684         put_page(vmf->cow_page);
3685         return ret;
3686 }
3687
3688 static int do_shared_fault(struct vm_fault *vmf)
3689 {
3690         struct vm_area_struct *vma = vmf->vma;
3691         int ret, tmp;
3692
3693         ret = __do_fault(vmf);
3694         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3695                 return ret;
3696
3697         /*
3698          * Check if the backing address space wants to know that the page is
3699          * about to become writable
3700          */
3701         if (vma->vm_ops->page_mkwrite) {
3702                 unlock_page(vmf->page);
3703                 tmp = do_page_mkwrite(vmf);
3704                 if (unlikely(!tmp ||
3705                                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3706                         put_page(vmf->page);
3707                         return tmp;
3708                 }
3709         }
3710
3711         ret |= finish_fault(vmf);
3712         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3713                                         VM_FAULT_RETRY))) {
3714                 unlock_page(vmf->page);
3715                 put_page(vmf->page);
3716                 return ret;
3717         }
3718
3719         fault_dirty_shared_page(vma, vmf->page);
3720         return ret;
3721 }
3722
3723 /*
3724  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3725  * but allow concurrent faults).
3726  * The mmap_sem may have been released depending on flags and our
3727  * return value.  See filemap_fault() and __lock_page_or_retry().
3728  */
3729 static int do_fault(struct vm_fault *vmf)
3730 {
3731         struct vm_area_struct *vma = vmf->vma;
3732         int ret;
3733
3734         /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3735         if (!vma->vm_ops->fault)
3736                 ret = VM_FAULT_SIGBUS;
3737         else if (!(vmf->flags & FAULT_FLAG_WRITE))
3738                 ret = do_read_fault(vmf);
3739         else if (!(vma->vm_flags & VM_SHARED))
3740                 ret = do_cow_fault(vmf);
3741         else
3742                 ret = do_shared_fault(vmf);
3743
3744         /* preallocated pagetable is unused: free it */
3745         if (vmf->prealloc_pte) {
3746                 pte_free(vma->vm_mm, vmf->prealloc_pte);
3747                 vmf->prealloc_pte = NULL;
3748         }
3749         return ret;
3750 }
3751
3752 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3753                                 unsigned long addr, int page_nid,
3754                                 int *flags)
3755 {
3756         get_page(page);
3757
3758         count_vm_numa_event(NUMA_HINT_FAULTS);
3759         if (page_nid == numa_node_id()) {
3760                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3761                 *flags |= TNF_FAULT_LOCAL;
3762         }
3763
3764         return mpol_misplaced(page, vma, addr);
3765 }
3766
3767 static int do_numa_page(struct vm_fault *vmf)
3768 {
3769         struct vm_area_struct *vma = vmf->vma;
3770         struct page *page = NULL;
3771         int page_nid = -1;
3772         int last_cpupid;
3773         int target_nid;
3774         bool migrated = false;
3775         pte_t pte;
3776         bool was_writable = pte_savedwrite(vmf->orig_pte);
3777         int flags = 0;
3778
3779         /*
3780          * The "pte" at this point cannot be used safely without
3781          * validation through pte_unmap_same(). It's of NUMA type but
3782          * the pfn may be screwed if the read is non atomic.
3783          */
3784         vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3785         spin_lock(vmf->ptl);
3786         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3787                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3788                 goto out;
3789         }
3790
3791         /*
3792          * Make it present again, Depending on how arch implementes non
3793          * accessible ptes, some can allow access by kernel mode.
3794          */
3795         pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3796         pte = pte_modify(pte, vma->vm_page_prot);
3797         pte = pte_mkyoung(pte);
3798         if (was_writable)
3799                 pte = pte_mkwrite(pte);
3800         ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3801         update_mmu_cache(vma, vmf->address, vmf->pte);
3802
3803         page = vm_normal_page(vma, vmf->address, pte);
3804         if (!page) {
3805                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3806                 return 0;
3807         }
3808
3809         /* TODO: handle PTE-mapped THP */
3810         if (PageCompound(page)) {
3811                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3812                 return 0;
3813         }
3814
3815         /*
3816          * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3817          * much anyway since they can be in shared cache state. This misses
3818          * the case where a mapping is writable but the process never writes
3819          * to it but pte_write gets cleared during protection updates and
3820          * pte_dirty has unpredictable behaviour between PTE scan updates,
3821          * background writeback, dirty balancing and application behaviour.
3822          */
3823         if (!pte_write(pte))
3824                 flags |= TNF_NO_GROUP;
3825
3826         /*
3827          * Flag if the page is shared between multiple address spaces. This
3828          * is later used when determining whether to group tasks together
3829          */
3830         if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3831                 flags |= TNF_SHARED;
3832
3833         last_cpupid = page_cpupid_last(page);
3834         page_nid = page_to_nid(page);
3835         target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3836                         &flags);
3837         pte_unmap_unlock(vmf->pte, vmf->ptl);
3838         if (target_nid == -1) {
3839                 put_page(page);
3840                 goto out;
3841         }
3842
3843         /* Migrate to the requested node */
3844         migrated = migrate_misplaced_page(page, vma, target_nid);
3845         if (migrated) {
3846                 page_nid = target_nid;
3847                 flags |= TNF_MIGRATED;
3848         } else
3849                 flags |= TNF_MIGRATE_FAIL;
3850
3851 out:
3852         if (page_nid != -1)
3853                 task_numa_fault(last_cpupid, page_nid, 1, flags);
3854         return 0;
3855 }
3856
3857 static inline int create_huge_pmd(struct vm_fault *vmf)
3858 {
3859         if (vma_is_anonymous(vmf->vma))
3860                 return do_huge_pmd_anonymous_page(vmf);
3861         if (vmf->vma->vm_ops->huge_fault)
3862                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3863         return VM_FAULT_FALLBACK;
3864 }
3865
3866 /* `inline' is required to avoid gcc 4.1.2 build error */
3867 static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3868 {
3869         if (vma_is_anonymous(vmf->vma))
3870                 return do_huge_pmd_wp_page(vmf, orig_pmd);
3871         if (vmf->vma->vm_ops->huge_fault)
3872                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3873
3874         /* COW handled on pte level: split pmd */
3875         VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3876         __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3877
3878         return VM_FAULT_FALLBACK;
3879 }
3880
3881 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3882 {
3883         return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3884 }
3885
3886 static int create_huge_pud(struct vm_fault *vmf)
3887 {
3888 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3889         /* No support for anonymous transparent PUD pages yet */
3890         if (vma_is_anonymous(vmf->vma))
3891                 return VM_FAULT_FALLBACK;
3892         if (vmf->vma->vm_ops->huge_fault)
3893                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3894 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3895         return VM_FAULT_FALLBACK;
3896 }
3897
3898 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3899 {
3900 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3901         /* No support for anonymous transparent PUD pages yet */
3902         if (vma_is_anonymous(vmf->vma))
3903                 return VM_FAULT_FALLBACK;
3904         if (vmf->vma->vm_ops->huge_fault)
3905                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3906 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3907         return VM_FAULT_FALLBACK;
3908 }
3909
3910 /*
3911  * These routines also need to handle stuff like marking pages dirty
3912  * and/or accessed for architectures that don't do it in hardware (most
3913  * RISC architectures).  The early dirtying is also good on the i386.
3914  *
3915  * There is also a hook called "update_mmu_cache()" that architectures
3916  * with external mmu caches can use to update those (ie the Sparc or
3917  * PowerPC hashed page tables that act as extended TLBs).
3918  *
3919  * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3920  * concurrent faults).
3921  *
3922  * The mmap_sem may have been released depending on flags and our return value.
3923  * See filemap_fault() and __lock_page_or_retry().
3924  */
3925 static int handle_pte_fault(struct vm_fault *vmf)
3926 {
3927         pte_t entry;
3928
3929         if (unlikely(pmd_none(*vmf->pmd))) {
3930                 /*
3931                  * Leave __pte_alloc() until later: because vm_ops->fault may
3932                  * want to allocate huge page, and if we expose page table
3933                  * for an instant, it will be difficult to retract from
3934                  * concurrent faults and from rmap lookups.
3935                  */
3936                 vmf->pte = NULL;
3937         } else {
3938                 /* See comment in pte_alloc_one_map() */
3939                 if (pmd_devmap_trans_unstable(vmf->pmd))
3940                         return 0;
3941                 /*
3942                  * A regular pmd is established and it can't morph into a huge
3943                  * pmd from under us anymore at this point because we hold the
3944                  * mmap_sem read mode and khugepaged takes it in write mode.
3945                  * So now it's safe to run pte_offset_map().
3946                  */
3947                 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3948                 vmf->orig_pte = *vmf->pte;
3949
3950                 /*
3951                  * some architectures can have larger ptes than wordsize,
3952                  * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3953                  * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3954                  * accesses.  The code below just needs a consistent view
3955                  * for the ifs and we later double check anyway with the
3956                  * ptl lock held. So here a barrier will do.
3957                  */
3958                 barrier();
3959                 if (pte_none(vmf->orig_pte)) {
3960                         pte_unmap(vmf->pte);
3961                         vmf->pte = NULL;
3962                 }
3963         }
3964
3965         if (!vmf->pte) {
3966                 if (vma_is_anonymous(vmf->vma))
3967                         return do_anonymous_page(vmf);
3968                 else
3969                         return do_fault(vmf);
3970         }
3971
3972         if (!pte_present(vmf->orig_pte))
3973                 return do_swap_page(vmf);
3974
3975         if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3976                 return do_numa_page(vmf);
3977
3978         vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3979         spin_lock(vmf->ptl);
3980         entry = vmf->orig_pte;
3981         if (unlikely(!pte_same(*vmf->pte, entry)))
3982                 goto unlock;
3983         if (vmf->flags & FAULT_FLAG_WRITE) {
3984                 if (!pte_write(entry))
3985                         return do_wp_page(vmf);
3986                 entry = pte_mkdirty(entry);
3987         }
3988         entry = pte_mkyoung(entry);
3989         if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3990                                 vmf->flags & FAULT_FLAG_WRITE)) {
3991                 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3992         } else {
3993                 /*
3994                  * This is needed only for protection faults but the arch code
3995                  * is not yet telling us if this is a protection fault or not.
3996                  * This still avoids useless tlb flushes for .text page faults
3997                  * with threads.
3998                  */
3999                 if (vmf->flags & FAULT_FLAG_WRITE)
4000                         flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4001         }
4002 unlock:
4003         pte_unmap_unlock(vmf->pte, vmf->ptl);
4004         return 0;
4005 }
4006
4007 /*
4008  * By the time we get here, we already hold the mm semaphore
4009  *
4010  * The mmap_sem may have been released depending on flags and our
4011  * return value.  See filemap_fault() and __lock_page_or_retry().
4012  */
4013 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4014                 unsigned int flags)
4015 {
4016         struct vm_fault vmf = {
4017                 .vma = vma,
4018                 .address = address & PAGE_MASK,
4019                 .flags = flags,
4020                 .pgoff = linear_page_index(vma, address),
4021                 .gfp_mask = __get_fault_gfp_mask(vma),
4022         };
4023         unsigned int dirty = flags & FAULT_FLAG_WRITE;
4024         struct mm_struct *mm = vma->vm_mm;
4025         pgd_t *pgd;
4026         p4d_t *p4d;
4027         int ret;
4028
4029         pgd = pgd_offset(mm, address);
4030         p4d = p4d_alloc(mm, pgd, address);
4031         if (!p4d)
4032                 return VM_FAULT_OOM;
4033
4034         vmf.pud = pud_alloc(mm, p4d, address);
4035         if (!vmf.pud)
4036                 return VM_FAULT_OOM;
4037         if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4038                 ret = create_huge_pud(&vmf);
4039                 if (!(ret & VM_FAULT_FALLBACK))
4040                         return ret;
4041         } else {
4042                 pud_t orig_pud = *vmf.pud;
4043
4044                 barrier();
4045                 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4046
4047                         /* NUMA case for anonymous PUDs would go here */
4048
4049                         if (dirty && !pud_write(orig_pud)) {
4050                                 ret = wp_huge_pud(&vmf, orig_pud);
4051                                 if (!(ret & VM_FAULT_FALLBACK))
4052                                         return ret;
4053                         } else {
4054                                 huge_pud_set_accessed(&vmf, orig_pud);
4055                                 return 0;
4056                         }
4057                 }
4058         }
4059
4060         vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4061         if (!vmf.pmd)
4062                 return VM_FAULT_OOM;
4063         if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4064                 ret = create_huge_pmd(&vmf);
4065                 if (!(ret & VM_FAULT_FALLBACK))
4066                         return ret;
4067         } else {
4068                 pmd_t orig_pmd = *vmf.pmd;
4069
4070                 barrier();
4071                 if (unlikely(is_swap_pmd(orig_pmd))) {
4072                         VM_BUG_ON(thp_migration_supported() &&
4073                                           !is_pmd_migration_entry(orig_pmd));
4074                         if (is_pmd_migration_entry(orig_pmd))
4075                                 pmd_migration_entry_wait(mm, vmf.pmd);
4076                         return 0;
4077                 }
4078                 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4079                         if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4080                                 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4081
4082                         if (dirty && !pmd_write(orig_pmd)) {
4083                                 ret = wp_huge_pmd(&vmf, orig_pmd);
4084                                 if (!(ret & VM_FAULT_FALLBACK))
4085                                         return ret;
4086                         } else {
4087                                 huge_pmd_set_accessed(&vmf, orig_pmd);
4088                                 return 0;
4089                         }
4090                 }
4091         }
4092
4093         return handle_pte_fault(&vmf);
4094 }
4095
4096 /*
4097  * By the time we get here, we already hold the mm semaphore
4098  *
4099  * The mmap_sem may have been released depending on flags and our
4100  * return value.  See filemap_fault() and __lock_page_or_retry().
4101  */
4102 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4103               &n