mm: do not rely on preempt_count in print_vma_addr
[sfrench/cifs-2.6.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/kallsyms.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
73
74 #include <asm/io.h>
75 #include <asm/mmu_context.h>
76 #include <asm/pgalloc.h>
77 #include <linux/uaccess.h>
78 #include <asm/tlb.h>
79 #include <asm/tlbflush.h>
80 #include <asm/pgtable.h>
81
82 #include "internal.h"
83
84 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
85 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
86 #endif
87
88 #ifndef CONFIG_NEED_MULTIPLE_NODES
89 /* use the per-pgdat data instead for discontigmem - mbligh */
90 unsigned long max_mapnr;
91 EXPORT_SYMBOL(max_mapnr);
92
93 struct page *mem_map;
94 EXPORT_SYMBOL(mem_map);
95 #endif
96
97 /*
98  * A number of key systems in x86 including ioremap() rely on the assumption
99  * that high_memory defines the upper bound on direct map memory, then end
100  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
101  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
102  * and ZONE_HIGHMEM.
103  */
104 void *high_memory;
105 EXPORT_SYMBOL(high_memory);
106
107 /*
108  * Randomize the address space (stacks, mmaps, brk, etc.).
109  *
110  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
111  *   as ancient (libc5 based) binaries can segfault. )
112  */
113 int randomize_va_space __read_mostly =
114 #ifdef CONFIG_COMPAT_BRK
115                                         1;
116 #else
117                                         2;
118 #endif
119
120 static int __init disable_randmaps(char *s)
121 {
122         randomize_va_space = 0;
123         return 1;
124 }
125 __setup("norandmaps", disable_randmaps);
126
127 unsigned long zero_pfn __read_mostly;
128 EXPORT_SYMBOL(zero_pfn);
129
130 unsigned long highest_memmap_pfn __read_mostly;
131
132 /*
133  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
134  */
135 static int __init init_zero_pfn(void)
136 {
137         zero_pfn = page_to_pfn(ZERO_PAGE(0));
138         return 0;
139 }
140 core_initcall(init_zero_pfn);
141
142
143 #if defined(SPLIT_RSS_COUNTING)
144
145 void sync_mm_rss(struct mm_struct *mm)
146 {
147         int i;
148
149         for (i = 0; i < NR_MM_COUNTERS; i++) {
150                 if (current->rss_stat.count[i]) {
151                         add_mm_counter(mm, i, current->rss_stat.count[i]);
152                         current->rss_stat.count[i] = 0;
153                 }
154         }
155         current->rss_stat.events = 0;
156 }
157
158 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
159 {
160         struct task_struct *task = current;
161
162         if (likely(task->mm == mm))
163                 task->rss_stat.count[member] += val;
164         else
165                 add_mm_counter(mm, member, val);
166 }
167 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
168 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
169
170 /* sync counter once per 64 page faults */
171 #define TASK_RSS_EVENTS_THRESH  (64)
172 static void check_sync_rss_stat(struct task_struct *task)
173 {
174         if (unlikely(task != current))
175                 return;
176         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
177                 sync_mm_rss(task->mm);
178 }
179 #else /* SPLIT_RSS_COUNTING */
180
181 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
182 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
183
184 static void check_sync_rss_stat(struct task_struct *task)
185 {
186 }
187
188 #endif /* SPLIT_RSS_COUNTING */
189
190 #ifdef HAVE_GENERIC_MMU_GATHER
191
192 static bool tlb_next_batch(struct mmu_gather *tlb)
193 {
194         struct mmu_gather_batch *batch;
195
196         batch = tlb->active;
197         if (batch->next) {
198                 tlb->active = batch->next;
199                 return true;
200         }
201
202         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
203                 return false;
204
205         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
206         if (!batch)
207                 return false;
208
209         tlb->batch_count++;
210         batch->next = NULL;
211         batch->nr   = 0;
212         batch->max  = MAX_GATHER_BATCH;
213
214         tlb->active->next = batch;
215         tlb->active = batch;
216
217         return true;
218 }
219
220 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
221                                 unsigned long start, unsigned long end)
222 {
223         tlb->mm = mm;
224
225         /* Is it from 0 to ~0? */
226         tlb->fullmm     = !(start | (end+1));
227         tlb->need_flush_all = 0;
228         tlb->local.next = NULL;
229         tlb->local.nr   = 0;
230         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
231         tlb->active     = &tlb->local;
232         tlb->batch_count = 0;
233
234 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
235         tlb->batch = NULL;
236 #endif
237         tlb->page_size = 0;
238
239         __tlb_reset_range(tlb);
240 }
241
242 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
243 {
244         if (!tlb->end)
245                 return;
246
247         tlb_flush(tlb);
248         mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
249 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
250         tlb_table_flush(tlb);
251 #endif
252         __tlb_reset_range(tlb);
253 }
254
255 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
256 {
257         struct mmu_gather_batch *batch;
258
259         for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
260                 free_pages_and_swap_cache(batch->pages, batch->nr);
261                 batch->nr = 0;
262         }
263         tlb->active = &tlb->local;
264 }
265
266 void tlb_flush_mmu(struct mmu_gather *tlb)
267 {
268         tlb_flush_mmu_tlbonly(tlb);
269         tlb_flush_mmu_free(tlb);
270 }
271
272 /* tlb_finish_mmu
273  *      Called at the end of the shootdown operation to free up any resources
274  *      that were required.
275  */
276 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
277                 unsigned long start, unsigned long end, bool force)
278 {
279         struct mmu_gather_batch *batch, *next;
280
281         if (force)
282                 __tlb_adjust_range(tlb, start, end - start);
283
284         tlb_flush_mmu(tlb);
285
286         /* keep the page table cache within bounds */
287         check_pgt_cache();
288
289         for (batch = tlb->local.next; batch; batch = next) {
290                 next = batch->next;
291                 free_pages((unsigned long)batch, 0);
292         }
293         tlb->local.next = NULL;
294 }
295
296 /* __tlb_remove_page
297  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
298  *      handling the additional races in SMP caused by other CPUs caching valid
299  *      mappings in their TLBs. Returns the number of free page slots left.
300  *      When out of page slots we must call tlb_flush_mmu().
301  *returns true if the caller should flush.
302  */
303 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
304 {
305         struct mmu_gather_batch *batch;
306
307         VM_BUG_ON(!tlb->end);
308         VM_WARN_ON(tlb->page_size != page_size);
309
310         batch = tlb->active;
311         /*
312          * Add the page and check if we are full. If so
313          * force a flush.
314          */
315         batch->pages[batch->nr++] = page;
316         if (batch->nr == batch->max) {
317                 if (!tlb_next_batch(tlb))
318                         return true;
319                 batch = tlb->active;
320         }
321         VM_BUG_ON_PAGE(batch->nr > batch->max, page);
322
323         return false;
324 }
325
326 #endif /* HAVE_GENERIC_MMU_GATHER */
327
328 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
329
330 /*
331  * See the comment near struct mmu_table_batch.
332  */
333
334 static void tlb_remove_table_smp_sync(void *arg)
335 {
336         /* Simply deliver the interrupt */
337 }
338
339 static void tlb_remove_table_one(void *table)
340 {
341         /*
342          * This isn't an RCU grace period and hence the page-tables cannot be
343          * assumed to be actually RCU-freed.
344          *
345          * It is however sufficient for software page-table walkers that rely on
346          * IRQ disabling. See the comment near struct mmu_table_batch.
347          */
348         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
349         __tlb_remove_table(table);
350 }
351
352 static void tlb_remove_table_rcu(struct rcu_head *head)
353 {
354         struct mmu_table_batch *batch;
355         int i;
356
357         batch = container_of(head, struct mmu_table_batch, rcu);
358
359         for (i = 0; i < batch->nr; i++)
360                 __tlb_remove_table(batch->tables[i]);
361
362         free_page((unsigned long)batch);
363 }
364
365 void tlb_table_flush(struct mmu_gather *tlb)
366 {
367         struct mmu_table_batch **batch = &tlb->batch;
368
369         if (*batch) {
370                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
371                 *batch = NULL;
372         }
373 }
374
375 void tlb_remove_table(struct mmu_gather *tlb, void *table)
376 {
377         struct mmu_table_batch **batch = &tlb->batch;
378
379         /*
380          * When there's less then two users of this mm there cannot be a
381          * concurrent page-table walk.
382          */
383         if (atomic_read(&tlb->mm->mm_users) < 2) {
384                 __tlb_remove_table(table);
385                 return;
386         }
387
388         if (*batch == NULL) {
389                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
390                 if (*batch == NULL) {
391                         tlb_remove_table_one(table);
392                         return;
393                 }
394                 (*batch)->nr = 0;
395         }
396         (*batch)->tables[(*batch)->nr++] = table;
397         if ((*batch)->nr == MAX_TABLE_BATCH)
398                 tlb_table_flush(tlb);
399 }
400
401 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
402
403 /* tlb_gather_mmu
404  *      Called to initialize an (on-stack) mmu_gather structure for page-table
405  *      tear-down from @mm. The @fullmm argument is used when @mm is without
406  *      users and we're going to destroy the full address space (exit/execve).
407  */
408 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
409                         unsigned long start, unsigned long end)
410 {
411         arch_tlb_gather_mmu(tlb, mm, start, end);
412         inc_tlb_flush_pending(tlb->mm);
413 }
414
415 void tlb_finish_mmu(struct mmu_gather *tlb,
416                 unsigned long start, unsigned long end)
417 {
418         /*
419          * If there are parallel threads are doing PTE changes on same range
420          * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
421          * flush by batching, a thread has stable TLB entry can fail to flush
422          * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
423          * forcefully if we detect parallel PTE batching threads.
424          */
425         bool force = mm_tlb_flush_nested(tlb->mm);
426
427         arch_tlb_finish_mmu(tlb, start, end, force);
428         dec_tlb_flush_pending(tlb->mm);
429 }
430
431 /*
432  * Note: this doesn't free the actual pages themselves. That
433  * has been handled earlier when unmapping all the memory regions.
434  */
435 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
436                            unsigned long addr)
437 {
438         pgtable_t token = pmd_pgtable(*pmd);
439         pmd_clear(pmd);
440         pte_free_tlb(tlb, token, addr);
441         mm_dec_nr_ptes(tlb->mm);
442 }
443
444 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
445                                 unsigned long addr, unsigned long end,
446                                 unsigned long floor, unsigned long ceiling)
447 {
448         pmd_t *pmd;
449         unsigned long next;
450         unsigned long start;
451
452         start = addr;
453         pmd = pmd_offset(pud, addr);
454         do {
455                 next = pmd_addr_end(addr, end);
456                 if (pmd_none_or_clear_bad(pmd))
457                         continue;
458                 free_pte_range(tlb, pmd, addr);
459         } while (pmd++, addr = next, addr != end);
460
461         start &= PUD_MASK;
462         if (start < floor)
463                 return;
464         if (ceiling) {
465                 ceiling &= PUD_MASK;
466                 if (!ceiling)
467                         return;
468         }
469         if (end - 1 > ceiling - 1)
470                 return;
471
472         pmd = pmd_offset(pud, start);
473         pud_clear(pud);
474         pmd_free_tlb(tlb, pmd, start);
475         mm_dec_nr_pmds(tlb->mm);
476 }
477
478 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
479                                 unsigned long addr, unsigned long end,
480                                 unsigned long floor, unsigned long ceiling)
481 {
482         pud_t *pud;
483         unsigned long next;
484         unsigned long start;
485
486         start = addr;
487         pud = pud_offset(p4d, addr);
488         do {
489                 next = pud_addr_end(addr, end);
490                 if (pud_none_or_clear_bad(pud))
491                         continue;
492                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
493         } while (pud++, addr = next, addr != end);
494
495         start &= P4D_MASK;
496         if (start < floor)
497                 return;
498         if (ceiling) {
499                 ceiling &= P4D_MASK;
500                 if (!ceiling)
501                         return;
502         }
503         if (end - 1 > ceiling - 1)
504                 return;
505
506         pud = pud_offset(p4d, start);
507         p4d_clear(p4d);
508         pud_free_tlb(tlb, pud, start);
509         mm_dec_nr_puds(tlb->mm);
510 }
511
512 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
513                                 unsigned long addr, unsigned long end,
514                                 unsigned long floor, unsigned long ceiling)
515 {
516         p4d_t *p4d;
517         unsigned long next;
518         unsigned long start;
519
520         start = addr;
521         p4d = p4d_offset(pgd, addr);
522         do {
523                 next = p4d_addr_end(addr, end);
524                 if (p4d_none_or_clear_bad(p4d))
525                         continue;
526                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
527         } while (p4d++, addr = next, addr != end);
528
529         start &= PGDIR_MASK;
530         if (start < floor)
531                 return;
532         if (ceiling) {
533                 ceiling &= PGDIR_MASK;
534                 if (!ceiling)
535                         return;
536         }
537         if (end - 1 > ceiling - 1)
538                 return;
539
540         p4d = p4d_offset(pgd, start);
541         pgd_clear(pgd);
542         p4d_free_tlb(tlb, p4d, start);
543 }
544
545 /*
546  * This function frees user-level page tables of a process.
547  */
548 void free_pgd_range(struct mmu_gather *tlb,
549                         unsigned long addr, unsigned long end,
550                         unsigned long floor, unsigned long ceiling)
551 {
552         pgd_t *pgd;
553         unsigned long next;
554
555         /*
556          * The next few lines have given us lots of grief...
557          *
558          * Why are we testing PMD* at this top level?  Because often
559          * there will be no work to do at all, and we'd prefer not to
560          * go all the way down to the bottom just to discover that.
561          *
562          * Why all these "- 1"s?  Because 0 represents both the bottom
563          * of the address space and the top of it (using -1 for the
564          * top wouldn't help much: the masks would do the wrong thing).
565          * The rule is that addr 0 and floor 0 refer to the bottom of
566          * the address space, but end 0 and ceiling 0 refer to the top
567          * Comparisons need to use "end - 1" and "ceiling - 1" (though
568          * that end 0 case should be mythical).
569          *
570          * Wherever addr is brought up or ceiling brought down, we must
571          * be careful to reject "the opposite 0" before it confuses the
572          * subsequent tests.  But what about where end is brought down
573          * by PMD_SIZE below? no, end can't go down to 0 there.
574          *
575          * Whereas we round start (addr) and ceiling down, by different
576          * masks at different levels, in order to test whether a table
577          * now has no other vmas using it, so can be freed, we don't
578          * bother to round floor or end up - the tests don't need that.
579          */
580
581         addr &= PMD_MASK;
582         if (addr < floor) {
583                 addr += PMD_SIZE;
584                 if (!addr)
585                         return;
586         }
587         if (ceiling) {
588                 ceiling &= PMD_MASK;
589                 if (!ceiling)
590                         return;
591         }
592         if (end - 1 > ceiling - 1)
593                 end -= PMD_SIZE;
594         if (addr > end - 1)
595                 return;
596         /*
597          * We add page table cache pages with PAGE_SIZE,
598          * (see pte_free_tlb()), flush the tlb if we need
599          */
600         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
601         pgd = pgd_offset(tlb->mm, addr);
602         do {
603                 next = pgd_addr_end(addr, end);
604                 if (pgd_none_or_clear_bad(pgd))
605                         continue;
606                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
607         } while (pgd++, addr = next, addr != end);
608 }
609
610 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
611                 unsigned long floor, unsigned long ceiling)
612 {
613         while (vma) {
614                 struct vm_area_struct *next = vma->vm_next;
615                 unsigned long addr = vma->vm_start;
616
617                 /*
618                  * Hide vma from rmap and truncate_pagecache before freeing
619                  * pgtables
620                  */
621                 unlink_anon_vmas(vma);
622                 unlink_file_vma(vma);
623
624                 if (is_vm_hugetlb_page(vma)) {
625                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
626                                 floor, next ? next->vm_start : ceiling);
627                 } else {
628                         /*
629                          * Optimization: gather nearby vmas into one call down
630                          */
631                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
632                                && !is_vm_hugetlb_page(next)) {
633                                 vma = next;
634                                 next = vma->vm_next;
635                                 unlink_anon_vmas(vma);
636                                 unlink_file_vma(vma);
637                         }
638                         free_pgd_range(tlb, addr, vma->vm_end,
639                                 floor, next ? next->vm_start : ceiling);
640                 }
641                 vma = next;
642         }
643 }
644
645 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
646 {
647         spinlock_t *ptl;
648         pgtable_t new = pte_alloc_one(mm, address);
649         if (!new)
650                 return -ENOMEM;
651
652         /*
653          * Ensure all pte setup (eg. pte page lock and page clearing) are
654          * visible before the pte is made visible to other CPUs by being
655          * put into page tables.
656          *
657          * The other side of the story is the pointer chasing in the page
658          * table walking code (when walking the page table without locking;
659          * ie. most of the time). Fortunately, these data accesses consist
660          * of a chain of data-dependent loads, meaning most CPUs (alpha
661          * being the notable exception) will already guarantee loads are
662          * seen in-order. See the alpha page table accessors for the
663          * smp_read_barrier_depends() barriers in page table walking code.
664          */
665         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
666
667         ptl = pmd_lock(mm, pmd);
668         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
669                 mm_inc_nr_ptes(mm);
670                 pmd_populate(mm, pmd, new);
671                 new = NULL;
672         }
673         spin_unlock(ptl);
674         if (new)
675                 pte_free(mm, new);
676         return 0;
677 }
678
679 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
680 {
681         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
682         if (!new)
683                 return -ENOMEM;
684
685         smp_wmb(); /* See comment in __pte_alloc */
686
687         spin_lock(&init_mm.page_table_lock);
688         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
689                 pmd_populate_kernel(&init_mm, pmd, new);
690                 new = NULL;
691         }
692         spin_unlock(&init_mm.page_table_lock);
693         if (new)
694                 pte_free_kernel(&init_mm, new);
695         return 0;
696 }
697
698 static inline void init_rss_vec(int *rss)
699 {
700         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
701 }
702
703 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
704 {
705         int i;
706
707         if (current->mm == mm)
708                 sync_mm_rss(mm);
709         for (i = 0; i < NR_MM_COUNTERS; i++)
710                 if (rss[i])
711                         add_mm_counter(mm, i, rss[i]);
712 }
713
714 /*
715  * This function is called to print an error when a bad pte
716  * is found. For example, we might have a PFN-mapped pte in
717  * a region that doesn't allow it.
718  *
719  * The calling function must still handle the error.
720  */
721 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
722                           pte_t pte, struct page *page)
723 {
724         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
725         p4d_t *p4d = p4d_offset(pgd, addr);
726         pud_t *pud = pud_offset(p4d, addr);
727         pmd_t *pmd = pmd_offset(pud, addr);
728         struct address_space *mapping;
729         pgoff_t index;
730         static unsigned long resume;
731         static unsigned long nr_shown;
732         static unsigned long nr_unshown;
733
734         /*
735          * Allow a burst of 60 reports, then keep quiet for that minute;
736          * or allow a steady drip of one report per second.
737          */
738         if (nr_shown == 60) {
739                 if (time_before(jiffies, resume)) {
740                         nr_unshown++;
741                         return;
742                 }
743                 if (nr_unshown) {
744                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
745                                  nr_unshown);
746                         nr_unshown = 0;
747                 }
748                 nr_shown = 0;
749         }
750         if (nr_shown++ == 0)
751                 resume = jiffies + 60 * HZ;
752
753         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
754         index = linear_page_index(vma, addr);
755
756         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
757                  current->comm,
758                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
759         if (page)
760                 dump_page(page, "bad pte");
761         pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
762                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
763         /*
764          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
765          */
766         pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
767                  vma->vm_file,
768                  vma->vm_ops ? vma->vm_ops->fault : NULL,
769                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
770                  mapping ? mapping->a_ops->readpage : NULL);
771         dump_stack();
772         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
773 }
774
775 /*
776  * vm_normal_page -- This function gets the "struct page" associated with a pte.
777  *
778  * "Special" mappings do not wish to be associated with a "struct page" (either
779  * it doesn't exist, or it exists but they don't want to touch it). In this
780  * case, NULL is returned here. "Normal" mappings do have a struct page.
781  *
782  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
783  * pte bit, in which case this function is trivial. Secondly, an architecture
784  * may not have a spare pte bit, which requires a more complicated scheme,
785  * described below.
786  *
787  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
788  * special mapping (even if there are underlying and valid "struct pages").
789  * COWed pages of a VM_PFNMAP are always normal.
790  *
791  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
792  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
793  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
794  * mapping will always honor the rule
795  *
796  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
797  *
798  * And for normal mappings this is false.
799  *
800  * This restricts such mappings to be a linear translation from virtual address
801  * to pfn. To get around this restriction, we allow arbitrary mappings so long
802  * as the vma is not a COW mapping; in that case, we know that all ptes are
803  * special (because none can have been COWed).
804  *
805  *
806  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
807  *
808  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
809  * page" backing, however the difference is that _all_ pages with a struct
810  * page (that is, those where pfn_valid is true) are refcounted and considered
811  * normal pages by the VM. The disadvantage is that pages are refcounted
812  * (which can be slower and simply not an option for some PFNMAP users). The
813  * advantage is that we don't have to follow the strict linearity rule of
814  * PFNMAP mappings in order to support COWable mappings.
815  *
816  */
817 #ifdef __HAVE_ARCH_PTE_SPECIAL
818 # define HAVE_PTE_SPECIAL 1
819 #else
820 # define HAVE_PTE_SPECIAL 0
821 #endif
822 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
823                              pte_t pte, bool with_public_device)
824 {
825         unsigned long pfn = pte_pfn(pte);
826
827         if (HAVE_PTE_SPECIAL) {
828                 if (likely(!pte_special(pte)))
829                         goto check_pfn;
830                 if (vma->vm_ops && vma->vm_ops->find_special_page)
831                         return vma->vm_ops->find_special_page(vma, addr);
832                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
833                         return NULL;
834                 if (is_zero_pfn(pfn))
835                         return NULL;
836
837                 /*
838                  * Device public pages are special pages (they are ZONE_DEVICE
839                  * pages but different from persistent memory). They behave
840                  * allmost like normal pages. The difference is that they are
841                  * not on the lru and thus should never be involve with any-
842                  * thing that involve lru manipulation (mlock, numa balancing,
843                  * ...).
844                  *
845                  * This is why we still want to return NULL for such page from
846                  * vm_normal_page() so that we do not have to special case all
847                  * call site of vm_normal_page().
848                  */
849                 if (likely(pfn <= highest_memmap_pfn)) {
850                         struct page *page = pfn_to_page(pfn);
851
852                         if (is_device_public_page(page)) {
853                                 if (with_public_device)
854                                         return page;
855                                 return NULL;
856                         }
857                 }
858                 print_bad_pte(vma, addr, pte, NULL);
859                 return NULL;
860         }
861
862         /* !HAVE_PTE_SPECIAL case follows: */
863
864         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
865                 if (vma->vm_flags & VM_MIXEDMAP) {
866                         if (!pfn_valid(pfn))
867                                 return NULL;
868                         goto out;
869                 } else {
870                         unsigned long off;
871                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
872                         if (pfn == vma->vm_pgoff + off)
873                                 return NULL;
874                         if (!is_cow_mapping(vma->vm_flags))
875                                 return NULL;
876                 }
877         }
878
879         if (is_zero_pfn(pfn))
880                 return NULL;
881 check_pfn:
882         if (unlikely(pfn > highest_memmap_pfn)) {
883                 print_bad_pte(vma, addr, pte, NULL);
884                 return NULL;
885         }
886
887         /*
888          * NOTE! We still have PageReserved() pages in the page tables.
889          * eg. VDSO mappings can cause them to exist.
890          */
891 out:
892         return pfn_to_page(pfn);
893 }
894
895 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
896 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
897                                 pmd_t pmd)
898 {
899         unsigned long pfn = pmd_pfn(pmd);
900
901         /*
902          * There is no pmd_special() but there may be special pmds, e.g.
903          * in a direct-access (dax) mapping, so let's just replicate the
904          * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
905          */
906         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
907                 if (vma->vm_flags & VM_MIXEDMAP) {
908                         if (!pfn_valid(pfn))
909                                 return NULL;
910                         goto out;
911                 } else {
912                         unsigned long off;
913                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
914                         if (pfn == vma->vm_pgoff + off)
915                                 return NULL;
916                         if (!is_cow_mapping(vma->vm_flags))
917                                 return NULL;
918                 }
919         }
920
921         if (is_zero_pfn(pfn))
922                 return NULL;
923         if (unlikely(pfn > highest_memmap_pfn))
924                 return NULL;
925
926         /*
927          * NOTE! We still have PageReserved() pages in the page tables.
928          * eg. VDSO mappings can cause them to exist.
929          */
930 out:
931         return pfn_to_page(pfn);
932 }
933 #endif
934
935 /*
936  * copy one vm_area from one task to the other. Assumes the page tables
937  * already present in the new task to be cleared in the whole range
938  * covered by this vma.
939  */
940
941 static inline unsigned long
942 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
943                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
944                 unsigned long addr, int *rss)
945 {
946         unsigned long vm_flags = vma->vm_flags;
947         pte_t pte = *src_pte;
948         struct page *page;
949
950         /* pte contains position in swap or file, so copy. */
951         if (unlikely(!pte_present(pte))) {
952                 swp_entry_t entry = pte_to_swp_entry(pte);
953
954                 if (likely(!non_swap_entry(entry))) {
955                         if (swap_duplicate(entry) < 0)
956                                 return entry.val;
957
958                         /* make sure dst_mm is on swapoff's mmlist. */
959                         if (unlikely(list_empty(&dst_mm->mmlist))) {
960                                 spin_lock(&mmlist_lock);
961                                 if (list_empty(&dst_mm->mmlist))
962                                         list_add(&dst_mm->mmlist,
963                                                         &src_mm->mmlist);
964                                 spin_unlock(&mmlist_lock);
965                         }
966                         rss[MM_SWAPENTS]++;
967                 } else if (is_migration_entry(entry)) {
968                         page = migration_entry_to_page(entry);
969
970                         rss[mm_counter(page)]++;
971
972                         if (is_write_migration_entry(entry) &&
973                                         is_cow_mapping(vm_flags)) {
974                                 /*
975                                  * COW mappings require pages in both
976                                  * parent and child to be set to read.
977                                  */
978                                 make_migration_entry_read(&entry);
979                                 pte = swp_entry_to_pte(entry);
980                                 if (pte_swp_soft_dirty(*src_pte))
981                                         pte = pte_swp_mksoft_dirty(pte);
982                                 set_pte_at(src_mm, addr, src_pte, pte);
983                         }
984                 } else if (is_device_private_entry(entry)) {
985                         page = device_private_entry_to_page(entry);
986
987                         /*
988                          * Update rss count even for unaddressable pages, as
989                          * they should treated just like normal pages in this
990                          * respect.
991                          *
992                          * We will likely want to have some new rss counters
993                          * for unaddressable pages, at some point. But for now
994                          * keep things as they are.
995                          */
996                         get_page(page);
997                         rss[mm_counter(page)]++;
998                         page_dup_rmap(page, false);
999
1000                         /*
1001                          * We do not preserve soft-dirty information, because so
1002                          * far, checkpoint/restore is the only feature that
1003                          * requires that. And checkpoint/restore does not work
1004                          * when a device driver is involved (you cannot easily
1005                          * save and restore device driver state).
1006                          */
1007                         if (is_write_device_private_entry(entry) &&
1008                             is_cow_mapping(vm_flags)) {
1009                                 make_device_private_entry_read(&entry);
1010                                 pte = swp_entry_to_pte(entry);
1011                                 set_pte_at(src_mm, addr, src_pte, pte);
1012                         }
1013                 }
1014                 goto out_set_pte;
1015         }
1016
1017         /*
1018          * If it's a COW mapping, write protect it both
1019          * in the parent and the child
1020          */
1021         if (is_cow_mapping(vm_flags)) {
1022                 ptep_set_wrprotect(src_mm, addr, src_pte);
1023                 pte = pte_wrprotect(pte);
1024         }
1025
1026         /*
1027          * If it's a shared mapping, mark it clean in
1028          * the child
1029          */
1030         if (vm_flags & VM_SHARED)
1031                 pte = pte_mkclean(pte);
1032         pte = pte_mkold(pte);
1033
1034         page = vm_normal_page(vma, addr, pte);
1035         if (page) {
1036                 get_page(page);
1037                 page_dup_rmap(page, false);
1038                 rss[mm_counter(page)]++;
1039         } else if (pte_devmap(pte)) {
1040                 page = pte_page(pte);
1041
1042                 /*
1043                  * Cache coherent device memory behave like regular page and
1044                  * not like persistent memory page. For more informations see
1045                  * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1046                  */
1047                 if (is_device_public_page(page)) {
1048                         get_page(page);
1049                         page_dup_rmap(page, false);
1050                         rss[mm_counter(page)]++;
1051                 }
1052         }
1053
1054 out_set_pte:
1055         set_pte_at(dst_mm, addr, dst_pte, pte);
1056         return 0;
1057 }
1058
1059 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1060                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1061                    unsigned long addr, unsigned long end)
1062 {
1063         pte_t *orig_src_pte, *orig_dst_pte;
1064         pte_t *src_pte, *dst_pte;
1065         spinlock_t *src_ptl, *dst_ptl;
1066         int progress = 0;
1067         int rss[NR_MM_COUNTERS];
1068         swp_entry_t entry = (swp_entry_t){0};
1069
1070 again:
1071         init_rss_vec(rss);
1072
1073         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1074         if (!dst_pte)
1075                 return -ENOMEM;
1076         src_pte = pte_offset_map(src_pmd, addr);
1077         src_ptl = pte_lockptr(src_mm, src_pmd);
1078         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1079         orig_src_pte = src_pte;
1080         orig_dst_pte = dst_pte;
1081         arch_enter_lazy_mmu_mode();
1082
1083         do {
1084                 /*
1085                  * We are holding two locks at this point - either of them
1086                  * could generate latencies in another task on another CPU.
1087                  */
1088                 if (progress >= 32) {
1089                         progress = 0;
1090                         if (need_resched() ||
1091                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1092                                 break;
1093                 }
1094                 if (pte_none(*src_pte)) {
1095                         progress++;
1096                         continue;
1097                 }
1098                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1099                                                         vma, addr, rss);
1100                 if (entry.val)
1101                         break;
1102                 progress += 8;
1103         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1104
1105         arch_leave_lazy_mmu_mode();
1106         spin_unlock(src_ptl);
1107         pte_unmap(orig_src_pte);
1108         add_mm_rss_vec(dst_mm, rss);
1109         pte_unmap_unlock(orig_dst_pte, dst_ptl);
1110         cond_resched();
1111
1112         if (entry.val) {
1113                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1114                         return -ENOMEM;
1115                 progress = 0;
1116         }
1117         if (addr != end)
1118                 goto again;
1119         return 0;
1120 }
1121
1122 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1123                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1124                 unsigned long addr, unsigned long end)
1125 {
1126         pmd_t *src_pmd, *dst_pmd;
1127         unsigned long next;
1128
1129         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1130         if (!dst_pmd)
1131                 return -ENOMEM;
1132         src_pmd = pmd_offset(src_pud, addr);
1133         do {
1134                 next = pmd_addr_end(addr, end);
1135                 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1136                         || pmd_devmap(*src_pmd)) {
1137                         int err;
1138                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1139                         err = copy_huge_pmd(dst_mm, src_mm,
1140                                             dst_pmd, src_pmd, addr, vma);
1141                         if (err == -ENOMEM)
1142                                 return -ENOMEM;
1143                         if (!err)
1144                                 continue;
1145                         /* fall through */
1146                 }
1147                 if (pmd_none_or_clear_bad(src_pmd))
1148                         continue;
1149                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1150                                                 vma, addr, next))
1151                         return -ENOMEM;
1152         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1153         return 0;
1154 }
1155
1156 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1157                 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1158                 unsigned long addr, unsigned long end)
1159 {
1160         pud_t *src_pud, *dst_pud;
1161         unsigned long next;
1162
1163         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1164         if (!dst_pud)
1165                 return -ENOMEM;
1166         src_pud = pud_offset(src_p4d, addr);
1167         do {
1168                 next = pud_addr_end(addr, end);
1169                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1170                         int err;
1171
1172                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1173                         err = copy_huge_pud(dst_mm, src_mm,
1174                                             dst_pud, src_pud, addr, vma);
1175                         if (err == -ENOMEM)
1176                                 return -ENOMEM;
1177                         if (!err)
1178                                 continue;
1179                         /* fall through */
1180                 }
1181                 if (pud_none_or_clear_bad(src_pud))
1182                         continue;
1183                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1184                                                 vma, addr, next))
1185                         return -ENOMEM;
1186         } while (dst_pud++, src_pud++, addr = next, addr != end);
1187         return 0;
1188 }
1189
1190 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1191                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1192                 unsigned long addr, unsigned long end)
1193 {
1194         p4d_t *src_p4d, *dst_p4d;
1195         unsigned long next;
1196
1197         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1198         if (!dst_p4d)
1199                 return -ENOMEM;
1200         src_p4d = p4d_offset(src_pgd, addr);
1201         do {
1202                 next = p4d_addr_end(addr, end);
1203                 if (p4d_none_or_clear_bad(src_p4d))
1204                         continue;
1205                 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1206                                                 vma, addr, next))
1207                         return -ENOMEM;
1208         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1209         return 0;
1210 }
1211
1212 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1213                 struct vm_area_struct *vma)
1214 {
1215         pgd_t *src_pgd, *dst_pgd;
1216         unsigned long next;
1217         unsigned long addr = vma->vm_start;
1218         unsigned long end = vma->vm_end;
1219         unsigned long mmun_start;       /* For mmu_notifiers */
1220         unsigned long mmun_end;         /* For mmu_notifiers */
1221         bool is_cow;
1222         int ret;
1223
1224         /*
1225          * Don't copy ptes where a page fault will fill them correctly.
1226          * Fork becomes much lighter when there are big shared or private
1227          * readonly mappings. The tradeoff is that copy_page_range is more
1228          * efficient than faulting.
1229          */
1230         if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1231                         !vma->anon_vma)
1232                 return 0;
1233
1234         if (is_vm_hugetlb_page(vma))
1235                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1236
1237         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1238                 /*
1239                  * We do not free on error cases below as remove_vma
1240                  * gets called on error from higher level routine
1241                  */
1242                 ret = track_pfn_copy(vma);
1243                 if (ret)
1244                         return ret;
1245         }
1246
1247         /*
1248          * We need to invalidate the secondary MMU mappings only when
1249          * there could be a permission downgrade on the ptes of the
1250          * parent mm. And a permission downgrade will only happen if
1251          * is_cow_mapping() returns true.
1252          */
1253         is_cow = is_cow_mapping(vma->vm_flags);
1254         mmun_start = addr;
1255         mmun_end   = end;
1256         if (is_cow)
1257                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1258                                                     mmun_end);
1259
1260         ret = 0;
1261         dst_pgd = pgd_offset(dst_mm, addr);
1262         src_pgd = pgd_offset(src_mm, addr);
1263         do {
1264                 next = pgd_addr_end(addr, end);
1265                 if (pgd_none_or_clear_bad(src_pgd))
1266                         continue;
1267                 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1268                                             vma, addr, next))) {
1269                         ret = -ENOMEM;
1270                         break;
1271                 }
1272         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1273
1274         if (is_cow)
1275                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1276         return ret;
1277 }
1278
1279 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1280                                 struct vm_area_struct *vma, pmd_t *pmd,
1281                                 unsigned long addr, unsigned long end,
1282                                 struct zap_details *details)
1283 {
1284         struct mm_struct *mm = tlb->mm;
1285         int force_flush = 0;
1286         int rss[NR_MM_COUNTERS];
1287         spinlock_t *ptl;
1288         pte_t *start_pte;
1289         pte_t *pte;
1290         swp_entry_t entry;
1291
1292         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1293 again:
1294         init_rss_vec(rss);
1295         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1296         pte = start_pte;
1297         flush_tlb_batched_pending(mm);
1298         arch_enter_lazy_mmu_mode();
1299         do {
1300                 pte_t ptent = *pte;
1301                 if (pte_none(ptent))
1302                         continue;
1303
1304                 if (pte_present(ptent)) {
1305                         struct page *page;
1306
1307                         page = _vm_normal_page(vma, addr, ptent, true);
1308                         if (unlikely(details) && page) {
1309                                 /*
1310                                  * unmap_shared_mapping_pages() wants to
1311                                  * invalidate cache without truncating:
1312                                  * unmap shared but keep private pages.
1313                                  */
1314                                 if (details->check_mapping &&
1315                                     details->check_mapping != page_rmapping(page))
1316                                         continue;
1317                         }
1318                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1319                                                         tlb->fullmm);
1320                         tlb_remove_tlb_entry(tlb, pte, addr);
1321                         if (unlikely(!page))
1322                                 continue;
1323
1324                         if (!PageAnon(page)) {
1325                                 if (pte_dirty(ptent)) {
1326                                         force_flush = 1;
1327                                         set_page_dirty(page);
1328                                 }
1329                                 if (pte_young(ptent) &&
1330                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1331                                         mark_page_accessed(page);
1332                         }
1333                         rss[mm_counter(page)]--;
1334                         page_remove_rmap(page, false);
1335                         if (unlikely(page_mapcount(page) < 0))
1336                                 print_bad_pte(vma, addr, ptent, page);
1337                         if (unlikely(__tlb_remove_page(tlb, page))) {
1338                                 force_flush = 1;
1339                                 addr += PAGE_SIZE;
1340                                 break;
1341                         }
1342                         continue;
1343                 }
1344
1345                 entry = pte_to_swp_entry(ptent);
1346                 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1347                         struct page *page = device_private_entry_to_page(entry);
1348
1349                         if (unlikely(details && details->check_mapping)) {
1350                                 /*
1351                                  * unmap_shared_mapping_pages() wants to
1352                                  * invalidate cache without truncating:
1353                                  * unmap shared but keep private pages.
1354                                  */
1355                                 if (details->check_mapping !=
1356                                     page_rmapping(page))
1357                                         continue;
1358                         }
1359
1360                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1361                         rss[mm_counter(page)]--;
1362                         page_remove_rmap(page, false);
1363                         put_page(page);
1364                         continue;
1365                 }
1366
1367                 /* If details->check_mapping, we leave swap entries. */
1368                 if (unlikely(details))
1369                         continue;
1370
1371                 entry = pte_to_swp_entry(ptent);
1372                 if (!non_swap_entry(entry))
1373                         rss[MM_SWAPENTS]--;
1374                 else if (is_migration_entry(entry)) {
1375                         struct page *page;
1376
1377                         page = migration_entry_to_page(entry);
1378                         rss[mm_counter(page)]--;
1379                 }
1380                 if (unlikely(!free_swap_and_cache(entry)))
1381                         print_bad_pte(vma, addr, ptent, NULL);
1382                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1383         } while (pte++, addr += PAGE_SIZE, addr != end);
1384
1385         add_mm_rss_vec(mm, rss);
1386         arch_leave_lazy_mmu_mode();
1387
1388         /* Do the actual TLB flush before dropping ptl */
1389         if (force_flush)
1390                 tlb_flush_mmu_tlbonly(tlb);
1391         pte_unmap_unlock(start_pte, ptl);
1392
1393         /*
1394          * If we forced a TLB flush (either due to running out of
1395          * batch buffers or because we needed to flush dirty TLB
1396          * entries before releasing the ptl), free the batched
1397          * memory too. Restart if we didn't do everything.
1398          */
1399         if (force_flush) {
1400                 force_flush = 0;
1401                 tlb_flush_mmu_free(tlb);
1402                 if (addr != end)
1403                         goto again;
1404         }
1405
1406         return addr;
1407 }
1408
1409 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1410                                 struct vm_area_struct *vma, pud_t *pud,
1411                                 unsigned long addr, unsigned long end,
1412                                 struct zap_details *details)
1413 {
1414         pmd_t *pmd;
1415         unsigned long next;
1416
1417         pmd = pmd_offset(pud, addr);
1418         do {
1419                 next = pmd_addr_end(addr, end);
1420                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1421                         if (next - addr != HPAGE_PMD_SIZE) {
1422                                 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1423                                     !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1424                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1425                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1426                                 goto next;
1427                         /* fall through */
1428                 }
1429                 /*
1430                  * Here there can be other concurrent MADV_DONTNEED or
1431                  * trans huge page faults running, and if the pmd is
1432                  * none or trans huge it can change under us. This is
1433                  * because MADV_DONTNEED holds the mmap_sem in read
1434                  * mode.
1435                  */
1436                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1437                         goto next;
1438                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1439 next:
1440                 cond_resched();
1441         } while (pmd++, addr = next, addr != end);
1442
1443         return addr;
1444 }
1445
1446 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1447                                 struct vm_area_struct *vma, p4d_t *p4d,
1448                                 unsigned long addr, unsigned long end,
1449                                 struct zap_details *details)
1450 {
1451         pud_t *pud;
1452         unsigned long next;
1453
1454         pud = pud_offset(p4d, addr);
1455         do {
1456                 next = pud_addr_end(addr, end);
1457                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1458                         if (next - addr != HPAGE_PUD_SIZE) {
1459                                 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1460                                 split_huge_pud(vma, pud, addr);
1461                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1462                                 goto next;
1463                         /* fall through */
1464                 }
1465                 if (pud_none_or_clear_bad(pud))
1466                         continue;
1467                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1468 next:
1469                 cond_resched();
1470         } while (pud++, addr = next, addr != end);
1471
1472         return addr;
1473 }
1474
1475 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1476                                 struct vm_area_struct *vma, pgd_t *pgd,
1477                                 unsigned long addr, unsigned long end,
1478                                 struct zap_details *details)
1479 {
1480         p4d_t *p4d;
1481         unsigned long next;
1482
1483         p4d = p4d_offset(pgd, addr);
1484         do {
1485                 next = p4d_addr_end(addr, end);
1486                 if (p4d_none_or_clear_bad(p4d))
1487                         continue;
1488                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1489         } while (p4d++, addr = next, addr != end);
1490
1491         return addr;
1492 }
1493
1494 void unmap_page_range(struct mmu_gather *tlb,
1495                              struct vm_area_struct *vma,
1496                              unsigned long addr, unsigned long end,
1497                              struct zap_details *details)
1498 {
1499         pgd_t *pgd;
1500         unsigned long next;
1501
1502         BUG_ON(addr >= end);
1503         tlb_start_vma(tlb, vma);
1504         pgd = pgd_offset(vma->vm_mm, addr);
1505         do {
1506                 next = pgd_addr_end(addr, end);
1507                 if (pgd_none_or_clear_bad(pgd))
1508                         continue;
1509                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1510         } while (pgd++, addr = next, addr != end);
1511         tlb_end_vma(tlb, vma);
1512 }
1513
1514
1515 static void unmap_single_vma(struct mmu_gather *tlb,
1516                 struct vm_area_struct *vma, unsigned long start_addr,
1517                 unsigned long end_addr,
1518                 struct zap_details *details)
1519 {
1520         unsigned long start = max(vma->vm_start, start_addr);
1521         unsigned long end;
1522
1523         if (start >= vma->vm_end)
1524                 return;
1525         end = min(vma->vm_end, end_addr);
1526         if (end <= vma->vm_start)
1527                 return;
1528
1529         if (vma->vm_file)
1530                 uprobe_munmap(vma, start, end);
1531
1532         if (unlikely(vma->vm_flags & VM_PFNMAP))
1533                 untrack_pfn(vma, 0, 0);
1534
1535         if (start != end) {
1536                 if (unlikely(is_vm_hugetlb_page(vma))) {
1537                         /*
1538                          * It is undesirable to test vma->vm_file as it
1539                          * should be non-null for valid hugetlb area.
1540                          * However, vm_file will be NULL in the error
1541                          * cleanup path of mmap_region. When
1542                          * hugetlbfs ->mmap method fails,
1543                          * mmap_region() nullifies vma->vm_file
1544                          * before calling this function to clean up.
1545                          * Since no pte has actually been setup, it is
1546                          * safe to do nothing in this case.
1547                          */
1548                         if (vma->vm_file) {
1549                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1550                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1551                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1552                         }
1553                 } else
1554                         unmap_page_range(tlb, vma, start, end, details);
1555         }
1556 }
1557
1558 /**
1559  * unmap_vmas - unmap a range of memory covered by a list of vma's
1560  * @tlb: address of the caller's struct mmu_gather
1561  * @vma: the starting vma
1562  * @start_addr: virtual address at which to start unmapping
1563  * @end_addr: virtual address at which to end unmapping
1564  *
1565  * Unmap all pages in the vma list.
1566  *
1567  * Only addresses between `start' and `end' will be unmapped.
1568  *
1569  * The VMA list must be sorted in ascending virtual address order.
1570  *
1571  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1572  * range after unmap_vmas() returns.  So the only responsibility here is to
1573  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1574  * drops the lock and schedules.
1575  */
1576 void unmap_vmas(struct mmu_gather *tlb,
1577                 struct vm_area_struct *vma, unsigned long start_addr,
1578                 unsigned long end_addr)
1579 {
1580         struct mm_struct *mm = vma->vm_mm;
1581
1582         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1583         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1584                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1585         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1586 }
1587
1588 /**
1589  * zap_page_range - remove user pages in a given range
1590  * @vma: vm_area_struct holding the applicable pages
1591  * @start: starting address of pages to zap
1592  * @size: number of bytes to zap
1593  *
1594  * Caller must protect the VMA list
1595  */
1596 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1597                 unsigned long size)
1598 {
1599         struct mm_struct *mm = vma->vm_mm;
1600         struct mmu_gather tlb;
1601         unsigned long end = start + size;
1602
1603         lru_add_drain();
1604         tlb_gather_mmu(&tlb, mm, start, end);
1605         update_hiwater_rss(mm);
1606         mmu_notifier_invalidate_range_start(mm, start, end);
1607         for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1608                 unmap_single_vma(&tlb, vma, start, end, NULL);
1609
1610                 /*
1611                  * zap_page_range does not specify whether mmap_sem should be
1612                  * held for read or write. That allows parallel zap_page_range
1613                  * operations to unmap a PTE and defer a flush meaning that
1614                  * this call observes pte_none and fails to flush the TLB.
1615                  * Rather than adding a complex API, ensure that no stale
1616                  * TLB entries exist when this call returns.
1617                  */
1618                 flush_tlb_range(vma, start, end);
1619         }
1620
1621         mmu_notifier_invalidate_range_end(mm, start, end);
1622         tlb_finish_mmu(&tlb, start, end);
1623 }
1624
1625 /**
1626  * zap_page_range_single - remove user pages in a given range
1627  * @vma: vm_area_struct holding the applicable pages
1628  * @address: starting address of pages to zap
1629  * @size: number of bytes to zap
1630  * @details: details of shared cache invalidation
1631  *
1632  * The range must fit into one VMA.
1633  */
1634 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1635                 unsigned long size, struct zap_details *details)
1636 {
1637         struct mm_struct *mm = vma->vm_mm;
1638         struct mmu_gather tlb;
1639         unsigned long end = address + size;
1640
1641         lru_add_drain();
1642         tlb_gather_mmu(&tlb, mm, address, end);
1643         update_hiwater_rss(mm);
1644         mmu_notifier_invalidate_range_start(mm, address, end);
1645         unmap_single_vma(&tlb, vma, address, end, details);
1646         mmu_notifier_invalidate_range_end(mm, address, end);
1647         tlb_finish_mmu(&tlb, address, end);
1648 }
1649
1650 /**
1651  * zap_vma_ptes - remove ptes mapping the vma
1652  * @vma: vm_area_struct holding ptes to be zapped
1653  * @address: starting address of pages to zap
1654  * @size: number of bytes to zap
1655  *
1656  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1657  *
1658  * The entire address range must be fully contained within the vma.
1659  *
1660  * Returns 0 if successful.
1661  */
1662 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1663                 unsigned long size)
1664 {
1665         if (address < vma->vm_start || address + size > vma->vm_end ||
1666                         !(vma->vm_flags & VM_PFNMAP))
1667                 return -1;
1668         zap_page_range_single(vma, address, size, NULL);
1669         return 0;
1670 }
1671 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1672
1673 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1674                         spinlock_t **ptl)
1675 {
1676         pgd_t *pgd;
1677         p4d_t *p4d;
1678         pud_t *pud;
1679         pmd_t *pmd;
1680
1681         pgd = pgd_offset(mm, addr);
1682         p4d = p4d_alloc(mm, pgd, addr);
1683         if (!p4d)
1684                 return NULL;
1685         pud = pud_alloc(mm, p4d, addr);
1686         if (!pud)
1687                 return NULL;
1688         pmd = pmd_alloc(mm, pud, addr);
1689         if (!pmd)
1690                 return NULL;
1691
1692         VM_BUG_ON(pmd_trans_huge(*pmd));
1693         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1694 }
1695
1696 /*
1697  * This is the old fallback for page remapping.
1698  *
1699  * For historical reasons, it only allows reserved pages. Only
1700  * old drivers should use this, and they needed to mark their
1701  * pages reserved for the old functions anyway.
1702  */
1703 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1704                         struct page *page, pgprot_t prot)
1705 {
1706         struct mm_struct *mm = vma->vm_mm;
1707         int retval;
1708         pte_t *pte;
1709         spinlock_t *ptl;
1710
1711         retval = -EINVAL;
1712         if (PageAnon(page))
1713                 goto out;
1714         retval = -ENOMEM;
1715         flush_dcache_page(page);
1716         pte = get_locked_pte(mm, addr, &ptl);
1717         if (!pte)
1718                 goto out;
1719         retval = -EBUSY;
1720         if (!pte_none(*pte))
1721                 goto out_unlock;
1722
1723         /* Ok, finally just insert the thing.. */
1724         get_page(page);
1725         inc_mm_counter_fast(mm, mm_counter_file(page));
1726         page_add_file_rmap(page, false);
1727         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1728
1729         retval = 0;
1730         pte_unmap_unlock(pte, ptl);
1731         return retval;
1732 out_unlock:
1733         pte_unmap_unlock(pte, ptl);
1734 out:
1735         return retval;
1736 }
1737
1738 /**
1739  * vm_insert_page - insert single page into user vma
1740  * @vma: user vma to map to
1741  * @addr: target user address of this page
1742  * @page: source kernel page
1743  *
1744  * This allows drivers to insert individual pages they've allocated
1745  * into a user vma.
1746  *
1747  * The page has to be a nice clean _individual_ kernel allocation.
1748  * If you allocate a compound page, you need to have marked it as
1749  * such (__GFP_COMP), or manually just split the page up yourself
1750  * (see split_page()).
1751  *
1752  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1753  * took an arbitrary page protection parameter. This doesn't allow
1754  * that. Your vma protection will have to be set up correctly, which
1755  * means that if you want a shared writable mapping, you'd better
1756  * ask for a shared writable mapping!
1757  *
1758  * The page does not need to be reserved.
1759  *
1760  * Usually this function is called from f_op->mmap() handler
1761  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1762  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1763  * function from other places, for example from page-fault handler.
1764  */
1765 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1766                         struct page *page)
1767 {
1768         if (addr < vma->vm_start || addr >= vma->vm_end)
1769                 return -EFAULT;
1770         if (!page_count(page))
1771                 return -EINVAL;
1772         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1773                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1774                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1775                 vma->vm_flags |= VM_MIXEDMAP;
1776         }
1777         return insert_page(vma, addr, page, vma->vm_page_prot);
1778 }
1779 EXPORT_SYMBOL(vm_insert_page);
1780
1781 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1782                         pfn_t pfn, pgprot_t prot, bool mkwrite)
1783 {
1784         struct mm_struct *mm = vma->vm_mm;
1785         int retval;
1786         pte_t *pte, entry;
1787         spinlock_t *ptl;
1788
1789         retval = -ENOMEM;
1790         pte = get_locked_pte(mm, addr, &ptl);
1791         if (!pte)
1792                 goto out;
1793         retval = -EBUSY;
1794         if (!pte_none(*pte)) {
1795                 if (mkwrite) {
1796                         /*
1797                          * For read faults on private mappings the PFN passed
1798                          * in may not match the PFN we have mapped if the
1799                          * mapped PFN is a writeable COW page.  In the mkwrite
1800                          * case we are creating a writable PTE for a shared
1801                          * mapping and we expect the PFNs to match.
1802                          */
1803                         if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1804                                 goto out_unlock;
1805                         entry = *pte;
1806                         goto out_mkwrite;
1807                 } else
1808                         goto out_unlock;
1809         }
1810
1811         /* Ok, finally just insert the thing.. */
1812         if (pfn_t_devmap(pfn))
1813                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1814         else
1815                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1816
1817 out_mkwrite:
1818         if (mkwrite) {
1819                 entry = pte_mkyoung(entry);
1820                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1821         }
1822
1823         set_pte_at(mm, addr, pte, entry);
1824         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1825
1826         retval = 0;
1827 out_unlock:
1828         pte_unmap_unlock(pte, ptl);
1829 out:
1830         return retval;
1831 }
1832
1833 /**
1834  * vm_insert_pfn - insert single pfn into user vma
1835  * @vma: user vma to map to
1836  * @addr: target user address of this page
1837  * @pfn: source kernel pfn
1838  *
1839  * Similar to vm_insert_page, this allows drivers to insert individual pages
1840  * they've allocated into a user vma. Same comments apply.
1841  *
1842  * This function should only be called from a vm_ops->fault handler, and
1843  * in that case the handler should return NULL.
1844  *
1845  * vma cannot be a COW mapping.
1846  *
1847  * As this is called only for pages that do not currently exist, we
1848  * do not need to flush old virtual caches or the TLB.
1849  */
1850 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1851                         unsigned long pfn)
1852 {
1853         return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1854 }
1855 EXPORT_SYMBOL(vm_insert_pfn);
1856
1857 /**
1858  * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1859  * @vma: user vma to map to
1860  * @addr: target user address of this page
1861  * @pfn: source kernel pfn
1862  * @pgprot: pgprot flags for the inserted page
1863  *
1864  * This is exactly like vm_insert_pfn, except that it allows drivers to
1865  * to override pgprot on a per-page basis.
1866  *
1867  * This only makes sense for IO mappings, and it makes no sense for
1868  * cow mappings.  In general, using multiple vmas is preferable;
1869  * vm_insert_pfn_prot should only be used if using multiple VMAs is
1870  * impractical.
1871  */
1872 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1873                         unsigned long pfn, pgprot_t pgprot)
1874 {
1875         int ret;
1876         /*
1877          * Technically, architectures with pte_special can avoid all these
1878          * restrictions (same for remap_pfn_range).  However we would like
1879          * consistency in testing and feature parity among all, so we should
1880          * try to keep these invariants in place for everybody.
1881          */
1882         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1883         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1884                                                 (VM_PFNMAP|VM_MIXEDMAP));
1885         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1886         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1887
1888         if (addr < vma->vm_start || addr >= vma->vm_end)
1889                 return -EFAULT;
1890
1891         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1892
1893         ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1894                         false);
1895
1896         return ret;
1897 }
1898 EXPORT_SYMBOL(vm_insert_pfn_prot);
1899
1900 static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1901                         pfn_t pfn, bool mkwrite)
1902 {
1903         pgprot_t pgprot = vma->vm_page_prot;
1904
1905         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1906
1907         if (addr < vma->vm_start || addr >= vma->vm_end)
1908                 return -EFAULT;
1909
1910         track_pfn_insert(vma, &pgprot, pfn);
1911
1912         /*
1913          * If we don't have pte special, then we have to use the pfn_valid()
1914          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1915          * refcount the page if pfn_valid is true (hence insert_page rather
1916          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1917          * without pte special, it would there be refcounted as a normal page.
1918          */
1919         if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1920                 struct page *page;
1921
1922                 /*
1923                  * At this point we are committed to insert_page()
1924                  * regardless of whether the caller specified flags that
1925                  * result in pfn_t_has_page() == false.
1926                  */
1927                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1928                 return insert_page(vma, addr, page, pgprot);
1929         }
1930         return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1931 }
1932
1933 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1934                         pfn_t pfn)
1935 {
1936         return __vm_insert_mixed(vma, addr, pfn, false);
1937
1938 }
1939 EXPORT_SYMBOL(vm_insert_mixed);
1940
1941 int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1942                         pfn_t pfn)
1943 {
1944         return __vm_insert_mixed(vma, addr, pfn, true);
1945 }
1946 EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1947
1948 /*
1949  * maps a range of physical memory into the requested pages. the old
1950  * mappings are removed. any references to nonexistent pages results
1951  * in null mappings (currently treated as "copy-on-access")
1952  */
1953 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1954                         unsigned long addr, unsigned long end,
1955                         unsigned long pfn, pgprot_t prot)
1956 {
1957         pte_t *pte;
1958         spinlock_t *ptl;
1959
1960         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1961         if (!pte)
1962                 return -ENOMEM;
1963         arch_enter_lazy_mmu_mode();
1964         do {
1965                 BUG_ON(!pte_none(*pte));
1966                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1967                 pfn++;
1968         } while (pte++, addr += PAGE_SIZE, addr != end);
1969         arch_leave_lazy_mmu_mode();
1970         pte_unmap_unlock(pte - 1, ptl);
1971         return 0;
1972 }
1973
1974 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1975                         unsigned long addr, unsigned long end,
1976                         unsigned long pfn, pgprot_t prot)
1977 {
1978         pmd_t *pmd;
1979         unsigned long next;
1980
1981         pfn -= addr >> PAGE_SHIFT;
1982         pmd = pmd_alloc(mm, pud, addr);
1983         if (!pmd)
1984                 return -ENOMEM;
1985         VM_BUG_ON(pmd_trans_huge(*pmd));
1986         do {
1987                 next = pmd_addr_end(addr, end);
1988                 if (remap_pte_range(mm, pmd, addr, next,
1989                                 pfn + (addr >> PAGE_SHIFT), prot))
1990                         return -ENOMEM;
1991         } while (pmd++, addr = next, addr != end);
1992         return 0;
1993 }
1994
1995 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1996                         unsigned long addr, unsigned long end,
1997                         unsigned long pfn, pgprot_t prot)
1998 {
1999         pud_t *pud;
2000         unsigned long next;
2001
2002         pfn -= addr >> PAGE_SHIFT;
2003         pud = pud_alloc(mm, p4d, addr);
2004         if (!pud)
2005                 return -ENOMEM;
2006         do {
2007                 next = pud_addr_end(addr, end);
2008                 if (remap_pmd_range(mm, pud, addr, next,
2009                                 pfn + (addr >> PAGE_SHIFT), prot))
2010                         return -ENOMEM;
2011         } while (pud++, addr = next, addr != end);
2012         return 0;
2013 }
2014
2015 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2016                         unsigned long addr, unsigned long end,
2017                         unsigned long pfn, pgprot_t prot)
2018 {
2019         p4d_t *p4d;
2020         unsigned long next;
2021
2022         pfn -= addr >> PAGE_SHIFT;
2023         p4d = p4d_alloc(mm, pgd, addr);
2024         if (!p4d)
2025                 return -ENOMEM;
2026         do {
2027                 next = p4d_addr_end(addr, end);
2028                 if (remap_pud_range(mm, p4d, addr, next,
2029                                 pfn + (addr >> PAGE_SHIFT), prot))
2030                         return -ENOMEM;
2031         } while (p4d++, addr = next, addr != end);
2032         return 0;
2033 }
2034
2035 /**
2036  * remap_pfn_range - remap kernel memory to userspace
2037  * @vma: user vma to map to
2038  * @addr: target user address to start at
2039  * @pfn: physical address of kernel memory
2040  * @size: size of map area
2041  * @prot: page protection flags for this mapping
2042  *
2043  *  Note: this is only safe if the mm semaphore is held when called.
2044  */
2045 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2046                     unsigned long pfn, unsigned long size, pgprot_t prot)
2047 {
2048         pgd_t *pgd;
2049         unsigned long next;
2050         unsigned long end = addr + PAGE_ALIGN(size);
2051         struct mm_struct *mm = vma->vm_mm;
2052         unsigned long remap_pfn = pfn;
2053         int err;
2054
2055         /*
2056          * Physically remapped pages are special. Tell the
2057          * rest of the world about it:
2058          *   VM_IO tells people not to look at these pages
2059          *      (accesses can have side effects).
2060          *   VM_PFNMAP tells the core MM that the base pages are just
2061          *      raw PFN mappings, and do not have a "struct page" associated
2062          *      with them.
2063          *   VM_DONTEXPAND
2064          *      Disable vma merging and expanding with mremap().
2065          *   VM_DONTDUMP
2066          *      Omit vma from core dump, even when VM_IO turned off.
2067          *
2068          * There's a horrible special case to handle copy-on-write
2069          * behaviour that some programs depend on. We mark the "original"
2070          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2071          * See vm_normal_page() for details.
2072          */
2073         if (is_cow_mapping(vma->vm_flags)) {
2074                 if (addr != vma->vm_start || end != vma->vm_end)
2075                         return -EINVAL;
2076                 vma->vm_pgoff = pfn;
2077         }
2078
2079         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2080         if (err)
2081                 return -EINVAL;
2082
2083         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2084
2085         BUG_ON(addr >= end);
2086         pfn -= addr >> PAGE_SHIFT;
2087         pgd = pgd_offset(mm, addr);
2088         flush_cache_range(vma, addr, end);
2089         do {
2090                 next = pgd_addr_end(addr, end);
2091                 err = remap_p4d_range(mm, pgd, addr, next,
2092                                 pfn + (addr >> PAGE_SHIFT), prot);
2093                 if (err)
2094                         break;
2095         } while (pgd++, addr = next, addr != end);
2096
2097         if (err)
2098                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2099
2100         return err;
2101 }
2102 EXPORT_SYMBOL(remap_pfn_range);
2103
2104 /**
2105  * vm_iomap_memory - remap memory to userspace
2106  * @vma: user vma to map to
2107  * @start: start of area
2108  * @len: size of area
2109  *
2110  * This is a simplified io_remap_pfn_range() for common driver use. The
2111  * driver just needs to give us the physical memory range to be mapped,
2112  * we'll figure out the rest from the vma information.
2113  *
2114  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2115  * whatever write-combining details or similar.
2116  */
2117 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2118 {
2119         unsigned long vm_len, pfn, pages;
2120
2121         /* Check that the physical memory area passed in looks valid */
2122         if (start + len < start)
2123                 return -EINVAL;
2124         /*
2125          * You *really* shouldn't map things that aren't page-aligned,
2126          * but we've historically allowed it because IO memory might
2127          * just have smaller alignment.
2128          */
2129         len += start & ~PAGE_MASK;
2130         pfn = start >> PAGE_SHIFT;
2131         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2132         if (pfn + pages < pfn)
2133                 return -EINVAL;
2134
2135         /* We start the mapping 'vm_pgoff' pages into the area */
2136         if (vma->vm_pgoff > pages)
2137                 return -EINVAL;
2138         pfn += vma->vm_pgoff;
2139         pages -= vma->vm_pgoff;
2140
2141         /* Can we fit all of the mapping? */
2142         vm_len = vma->vm_end - vma->vm_start;
2143         if (vm_len >> PAGE_SHIFT > pages)
2144                 return -EINVAL;
2145
2146         /* Ok, let it rip */
2147         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2148 }
2149 EXPORT_SYMBOL(vm_iomap_memory);
2150
2151 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2152                                      unsigned long addr, unsigned long end,
2153                                      pte_fn_t fn, void *data)
2154 {
2155         pte_t *pte;
2156         int err;
2157         pgtable_t token;
2158         spinlock_t *uninitialized_var(ptl);
2159
2160         pte = (mm == &init_mm) ?
2161                 pte_alloc_kernel(pmd, addr) :
2162                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2163         if (!pte)
2164                 return -ENOMEM;
2165
2166         BUG_ON(pmd_huge(*pmd));
2167
2168         arch_enter_lazy_mmu_mode();
2169
2170         token = pmd_pgtable(*pmd);
2171
2172         do {
2173                 err = fn(pte++, token, addr, data);
2174                 if (err)
2175                         break;
2176         } while (addr += PAGE_SIZE, addr != end);
2177
2178         arch_leave_lazy_mmu_mode();
2179
2180         if (mm != &init_mm)
2181                 pte_unmap_unlock(pte-1, ptl);
2182         return err;
2183 }
2184
2185 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2186                                      unsigned long addr, unsigned long end,
2187                                      pte_fn_t fn, void *data)
2188 {
2189         pmd_t *pmd;
2190         unsigned long next;
2191         int err;
2192
2193         BUG_ON(pud_huge(*pud));
2194
2195         pmd = pmd_alloc(mm, pud, addr);
2196         if (!pmd)
2197                 return -ENOMEM;
2198         do {
2199                 next = pmd_addr_end(addr, end);
2200                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2201                 if (err)
2202                         break;
2203         } while (pmd++, addr = next, addr != end);
2204         return err;
2205 }
2206
2207 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2208                                      unsigned long addr, unsigned long end,
2209                                      pte_fn_t fn, void *data)
2210 {
2211         pud_t *pud;
2212         unsigned long next;
2213         int err;
2214
2215         pud = pud_alloc(mm, p4d, addr);
2216         if (!pud)
2217                 return -ENOMEM;
2218         do {
2219                 next = pud_addr_end(addr, end);
2220                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2221                 if (err)
2222                         break;
2223         } while (pud++, addr = next, addr != end);
2224         return err;
2225 }
2226
2227 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2228                                      unsigned long addr, unsigned long end,
2229                                      pte_fn_t fn, void *data)
2230 {
2231         p4d_t *p4d;
2232         unsigned long next;
2233         int err;
2234
2235         p4d = p4d_alloc(mm, pgd, addr);
2236         if (!p4d)
2237                 return -ENOMEM;
2238         do {
2239                 next = p4d_addr_end(addr, end);
2240                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2241                 if (err)
2242                         break;
2243         } while (p4d++, addr = next, addr != end);
2244         return err;
2245 }
2246
2247 /*
2248  * Scan a region of virtual memory, filling in page tables as necessary
2249  * and calling a provided function on each leaf page table.
2250  */
2251 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2252                         unsigned long size, pte_fn_t fn, void *data)
2253 {
2254         pgd_t *pgd;
2255         unsigned long next;
2256         unsigned long end = addr + size;
2257         int err;
2258
2259         if (WARN_ON(addr >= end))
2260                 return -EINVAL;
2261
2262         pgd = pgd_offset(mm, addr);
2263         do {
2264                 next = pgd_addr_end(addr, end);
2265                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2266                 if (err)
2267                         break;
2268         } while (pgd++, addr = next, addr != end);
2269
2270         return err;
2271 }
2272 EXPORT_SYMBOL_GPL(apply_to_page_range);
2273
2274 /*
2275  * handle_pte_fault chooses page fault handler according to an entry which was
2276  * read non-atomically.  Before making any commitment, on those architectures
2277  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2278  * parts, do_swap_page must check under lock before unmapping the pte and
2279  * proceeding (but do_wp_page is only called after already making such a check;
2280  * and do_anonymous_page can safely check later on).
2281  */
2282 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2283                                 pte_t *page_table, pte_t orig_pte)
2284 {
2285         int same = 1;
2286 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2287         if (sizeof(pte_t) > sizeof(unsigned long)) {
2288                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2289                 spin_lock(ptl);
2290                 same = pte_same(*page_table, orig_pte);
2291                 spin_unlock(ptl);
2292         }
2293 #endif
2294         pte_unmap(page_table);
2295         return same;
2296 }
2297
2298 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2299 {
2300         debug_dma_assert_idle(src);
2301
2302         /*
2303          * If the source page was a PFN mapping, we don't have
2304          * a "struct page" for it. We do a best-effort copy by
2305          * just copying from the original user address. If that
2306          * fails, we just zero-fill it. Live with it.
2307          */
2308         if (unlikely(!src)) {
2309                 void *kaddr = kmap_atomic(dst);
2310                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2311
2312                 /*
2313                  * This really shouldn't fail, because the page is there
2314                  * in the page tables. But it might just be unreadable,
2315                  * in which case we just give up and fill the result with
2316                  * zeroes.
2317                  */
2318                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2319                         clear_page(kaddr);
2320                 kunmap_atomic(kaddr);
2321                 flush_dcache_page(dst);
2322         } else
2323                 copy_user_highpage(dst, src, va, vma);
2324 }
2325
2326 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2327 {
2328         struct file *vm_file = vma->vm_file;
2329
2330         if (vm_file)
2331                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2332
2333         /*
2334          * Special mappings (e.g. VDSO) do not have any file so fake
2335          * a default GFP_KERNEL for them.
2336          */
2337         return GFP_KERNEL;
2338 }
2339
2340 /*
2341  * Notify the address space that the page is about to become writable so that
2342  * it can prohibit this or wait for the page to get into an appropriate state.
2343  *
2344  * We do this without the lock held, so that it can sleep if it needs to.
2345  */
2346 static int do_page_mkwrite(struct vm_fault *vmf)
2347 {
2348         int ret;
2349         struct page *page = vmf->page;
2350         unsigned int old_flags = vmf->flags;
2351
2352         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2353
2354         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2355         /* Restore original flags so that caller is not surprised */
2356         vmf->flags = old_flags;
2357         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2358                 return ret;
2359         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2360                 lock_page(page);
2361                 if (!page->mapping) {
2362                         unlock_page(page);
2363                         return 0; /* retry */
2364                 }
2365                 ret |= VM_FAULT_LOCKED;
2366         } else
2367                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2368         return ret;
2369 }
2370
2371 /*
2372  * Handle dirtying of a page in shared file mapping on a write fault.
2373  *
2374  * The function expects the page to be locked and unlocks it.
2375  */
2376 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2377                                     struct page *page)
2378 {
2379         struct address_space *mapping;
2380         bool dirtied;
2381         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2382
2383         dirtied = set_page_dirty(page);
2384         VM_BUG_ON_PAGE(PageAnon(page), page);
2385         /*
2386          * Take a local copy of the address_space - page.mapping may be zeroed
2387          * by truncate after unlock_page().   The address_space itself remains
2388          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2389          * release semantics to prevent the compiler from undoing this copying.
2390          */
2391         mapping = page_rmapping(page);
2392         unlock_page(page);
2393
2394         if ((dirtied || page_mkwrite) && mapping) {
2395                 /*
2396                  * Some device drivers do not set page.mapping
2397                  * but still dirty their pages
2398                  */
2399                 balance_dirty_pages_ratelimited(mapping);
2400         }
2401
2402         if (!page_mkwrite)
2403                 file_update_time(vma->vm_file);
2404 }
2405
2406 /*
2407  * Handle write page faults for pages that can be reused in the current vma
2408  *
2409  * This can happen either due to the mapping being with the VM_SHARED flag,
2410  * or due to us being the last reference standing to the page. In either
2411  * case, all we need to do here is to mark the page as writable and update
2412  * any related book-keeping.
2413  */
2414 static inline void wp_page_reuse(struct vm_fault *vmf)
2415         __releases(vmf->ptl)
2416 {
2417         struct vm_area_struct *vma = vmf->vma;
2418         struct page *page = vmf->page;
2419         pte_t entry;
2420         /*
2421          * Clear the pages cpupid information as the existing
2422          * information potentially belongs to a now completely
2423          * unrelated process.
2424          */
2425         if (page)
2426                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2427
2428         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2429         entry = pte_mkyoung(vmf->orig_pte);
2430         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2431         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2432                 update_mmu_cache(vma, vmf->address, vmf->pte);
2433         pte_unmap_unlock(vmf->pte, vmf->ptl);
2434 }
2435
2436 /*
2437  * Handle the case of a page which we actually need to copy to a new page.
2438  *
2439  * Called with mmap_sem locked and the old page referenced, but
2440  * without the ptl held.
2441  *
2442  * High level logic flow:
2443  *
2444  * - Allocate a page, copy the content of the old page to the new one.
2445  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2446  * - Take the PTL. If the pte changed, bail out and release the allocated page
2447  * - If the pte is still the way we remember it, update the page table and all
2448  *   relevant references. This includes dropping the reference the page-table
2449  *   held to the old page, as well as updating the rmap.
2450  * - In any case, unlock the PTL and drop the reference we took to the old page.
2451  */
2452 static int wp_page_copy(struct vm_fault *vmf)
2453 {
2454         struct vm_area_struct *vma = vmf->vma;
2455         struct mm_struct *mm = vma->vm_mm;
2456         struct page *old_page = vmf->page;
2457         struct page *new_page = NULL;
2458         pte_t entry;
2459         int page_copied = 0;
2460         const unsigned long mmun_start = vmf->address & PAGE_MASK;
2461         const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2462         struct mem_cgroup *memcg;
2463
2464         if (unlikely(anon_vma_prepare(vma)))
2465                 goto oom;
2466
2467         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2468                 new_page = alloc_zeroed_user_highpage_movable(vma,
2469                                                               vmf->address);
2470                 if (!new_page)
2471                         goto oom;
2472         } else {
2473                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2474                                 vmf->address);
2475                 if (!new_page)
2476                         goto oom;
2477                 cow_user_page(new_page, old_page, vmf->address, vma);
2478         }
2479
2480         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2481                 goto oom_free_new;
2482
2483         __SetPageUptodate(new_page);
2484
2485         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2486
2487         /*
2488          * Re-check the pte - we dropped the lock
2489          */
2490         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2491         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2492                 if (old_page) {
2493                         if (!PageAnon(old_page)) {
2494                                 dec_mm_counter_fast(mm,
2495                                                 mm_counter_file(old_page));
2496                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2497                         }
2498                 } else {
2499                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2500                 }
2501                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2502                 entry = mk_pte(new_page, vma->vm_page_prot);
2503                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2504                 /*
2505                  * Clear the pte entry and flush it first, before updating the
2506                  * pte with the new entry. This will avoid a race condition
2507                  * seen in the presence of one thread doing SMC and another
2508                  * thread doing COW.
2509                  */
2510                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2511                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2512                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2513                 lru_cache_add_active_or_unevictable(new_page, vma);
2514                 /*
2515                  * We call the notify macro here because, when using secondary
2516                  * mmu page tables (such as kvm shadow page tables), we want the
2517                  * new page to be mapped directly into the secondary page table.
2518                  */
2519                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2520                 update_mmu_cache(vma, vmf->address, vmf->pte);
2521                 if (old_page) {
2522                         /*
2523                          * Only after switching the pte to the new page may
2524                          * we remove the mapcount here. Otherwise another
2525                          * process may come and find the rmap count decremented
2526                          * before the pte is switched to the new page, and
2527                          * "reuse" the old page writing into it while our pte
2528                          * here still points into it and can be read by other
2529                          * threads.
2530                          *
2531                          * The critical issue is to order this
2532                          * page_remove_rmap with the ptp_clear_flush above.
2533                          * Those stores are ordered by (if nothing else,)
2534                          * the barrier present in the atomic_add_negative
2535                          * in page_remove_rmap.
2536                          *
2537                          * Then the TLB flush in ptep_clear_flush ensures that
2538                          * no process can access the old page before the
2539                          * decremented mapcount is visible. And the old page
2540                          * cannot be reused until after the decremented
2541                          * mapcount is visible. So transitively, TLBs to
2542                          * old page will be flushed before it can be reused.
2543                          */
2544                         page_remove_rmap(old_page, false);
2545                 }
2546
2547                 /* Free the old page.. */
2548                 new_page = old_page;
2549                 page_copied = 1;
2550         } else {
2551                 mem_cgroup_cancel_charge(new_page, memcg, false);
2552         }
2553
2554         if (new_page)
2555                 put_page(new_page);
2556
2557         pte_unmap_unlock(vmf->pte, vmf->ptl);
2558         /*
2559          * No need to double call mmu_notifier->invalidate_range() callback as
2560          * the above ptep_clear_flush_notify() did already call it.
2561          */
2562         mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2563         if (old_page) {
2564                 /*
2565                  * Don't let another task, with possibly unlocked vma,
2566                  * keep the mlocked page.
2567                  */
2568                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2569                         lock_page(old_page);    /* LRU manipulation */
2570                         if (PageMlocked(old_page))
2571                                 munlock_vma_page(old_page);
2572                         unlock_page(old_page);
2573                 }
2574                 put_page(old_page);
2575         }
2576         return page_copied ? VM_FAULT_WRITE : 0;
2577 oom_free_new:
2578         put_page(new_page);
2579 oom:
2580         if (old_page)
2581                 put_page(old_page);
2582         return VM_FAULT_OOM;
2583 }
2584
2585 /**
2586  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2587  *                        writeable once the page is prepared
2588  *
2589  * @vmf: structure describing the fault
2590  *
2591  * This function handles all that is needed to finish a write page fault in a
2592  * shared mapping due to PTE being read-only once the mapped page is prepared.
2593  * It handles locking of PTE and modifying it. The function returns
2594  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2595  * lock.
2596  *
2597  * The function expects the page to be locked or other protection against
2598  * concurrent faults / writeback (such as DAX radix tree locks).
2599  */
2600 int finish_mkwrite_fault(struct vm_fault *vmf)
2601 {
2602         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2603         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2604                                        &vmf->ptl);
2605         /*
2606          * We might have raced with another page fault while we released the
2607          * pte_offset_map_lock.
2608          */
2609         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2610                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2611                 return VM_FAULT_NOPAGE;
2612         }
2613         wp_page_reuse(vmf);
2614         return 0;
2615 }
2616
2617 /*
2618  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2619  * mapping
2620  */
2621 static int wp_pfn_shared(struct vm_fault *vmf)
2622 {
2623         struct vm_area_struct *vma = vmf->vma;
2624
2625         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2626                 int ret;
2627
2628                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2629                 vmf->flags |= FAULT_FLAG_MKWRITE;
2630                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2631                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2632                         return ret;
2633                 return finish_mkwrite_fault(vmf);
2634         }
2635         wp_page_reuse(vmf);
2636         return VM_FAULT_WRITE;
2637 }
2638
2639 static int wp_page_shared(struct vm_fault *vmf)
2640         __releases(vmf->ptl)
2641 {
2642         struct vm_area_struct *vma = vmf->vma;
2643
2644         get_page(vmf->page);
2645
2646         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2647                 int tmp;
2648
2649                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2650                 tmp = do_page_mkwrite(vmf);
2651                 if (unlikely(!tmp || (tmp &
2652                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2653                         put_page(vmf->page);
2654                         return tmp;
2655                 }
2656                 tmp = finish_mkwrite_fault(vmf);
2657                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2658                         unlock_page(vmf->page);
2659                         put_page(vmf->page);
2660                         return tmp;
2661                 }
2662         } else {
2663                 wp_page_reuse(vmf);
2664                 lock_page(vmf->page);
2665         }
2666         fault_dirty_shared_page(vma, vmf->page);
2667         put_page(vmf->page);
2668
2669         return VM_FAULT_WRITE;
2670 }
2671
2672 /*
2673  * This routine handles present pages, when users try to write
2674  * to a shared page. It is done by copying the page to a new address
2675  * and decrementing the shared-page counter for the old page.
2676  *
2677  * Note that this routine assumes that the protection checks have been
2678  * done by the caller (the low-level page fault routine in most cases).
2679  * Thus we can safely just mark it writable once we've done any necessary
2680  * COW.
2681  *
2682  * We also mark the page dirty at this point even though the page will
2683  * change only once the write actually happens. This avoids a few races,
2684  * and potentially makes it more efficient.
2685  *
2686  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2687  * but allow concurrent faults), with pte both mapped and locked.
2688  * We return with mmap_sem still held, but pte unmapped and unlocked.
2689  */
2690 static int do_wp_page(struct vm_fault *vmf)
2691         __releases(vmf->ptl)
2692 {
2693         struct vm_area_struct *vma = vmf->vma;
2694
2695         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2696         if (!vmf->page) {
2697                 /*
2698                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2699                  * VM_PFNMAP VMA.
2700                  *
2701                  * We should not cow pages in a shared writeable mapping.
2702                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2703                  */
2704                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2705                                      (VM_WRITE|VM_SHARED))
2706                         return wp_pfn_shared(vmf);
2707
2708                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2709                 return wp_page_copy(vmf);
2710         }
2711
2712         /*
2713          * Take out anonymous pages first, anonymous shared vmas are
2714          * not dirty accountable.
2715          */
2716         if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2717                 int total_map_swapcount;
2718                 if (!trylock_page(vmf->page)) {
2719                         get_page(vmf->page);
2720                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2721                         lock_page(vmf->page);
2722                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2723                                         vmf->address, &vmf->ptl);
2724                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2725                                 unlock_page(vmf->page);
2726                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2727                                 put_page(vmf->page);
2728                                 return 0;
2729                         }
2730                         put_page(vmf->page);
2731                 }
2732                 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2733                         if (total_map_swapcount == 1) {
2734                                 /*
2735                                  * The page is all ours. Move it to
2736                                  * our anon_vma so the rmap code will
2737                                  * not search our parent or siblings.
2738                                  * Protected against the rmap code by
2739                                  * the page lock.
2740                                  */
2741                                 page_move_anon_rmap(vmf->page, vma);
2742                         }
2743                         unlock_page(vmf->page);
2744                         wp_page_reuse(vmf);
2745                         return VM_FAULT_WRITE;
2746                 }
2747                 unlock_page(vmf->page);
2748         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2749                                         (VM_WRITE|VM_SHARED))) {
2750                 return wp_page_shared(vmf);
2751         }
2752
2753         /*
2754          * Ok, we need to copy. Oh, well..
2755          */
2756         get_page(vmf->page);
2757
2758         pte_unmap_unlock(vmf->pte, vmf->ptl);
2759         return wp_page_copy(vmf);
2760 }
2761
2762 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2763                 unsigned long start_addr, unsigned long end_addr,
2764                 struct zap_details *details)
2765 {
2766         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2767 }
2768
2769 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2770                                             struct zap_details *details)
2771 {
2772         struct vm_area_struct *vma;
2773         pgoff_t vba, vea, zba, zea;
2774
2775         vma_interval_tree_foreach(vma, root,
2776                         details->first_index, details->last_index) {
2777
2778                 vba = vma->vm_pgoff;
2779                 vea = vba + vma_pages(vma) - 1;
2780                 zba = details->first_index;
2781                 if (zba < vba)
2782                         zba = vba;
2783                 zea = details->last_index;
2784                 if (zea > vea)
2785                         zea = vea;
2786
2787                 unmap_mapping_range_vma(vma,
2788                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2789                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2790                                 details);
2791         }
2792 }
2793
2794 /**
2795  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2796  * address_space corresponding to the specified page range in the underlying
2797  * file.
2798  *
2799  * @mapping: the address space containing mmaps to be unmapped.
2800  * @holebegin: byte in first page to unmap, relative to the start of
2801  * the underlying file.  This will be rounded down to a PAGE_SIZE
2802  * boundary.  Note that this is different from truncate_pagecache(), which
2803  * must keep the partial page.  In contrast, we must get rid of
2804  * partial pages.
2805  * @holelen: size of prospective hole in bytes.  This will be rounded
2806  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2807  * end of the file.
2808  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2809  * but 0 when invalidating pagecache, don't throw away private data.
2810  */
2811 void unmap_mapping_range(struct address_space *mapping,
2812                 loff_t const holebegin, loff_t const holelen, int even_cows)
2813 {
2814         struct zap_details details = { };
2815         pgoff_t hba = holebegin >> PAGE_SHIFT;
2816         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2817
2818         /* Check for overflow. */
2819         if (sizeof(holelen) > sizeof(hlen)) {
2820                 long long holeend =
2821                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2822                 if (holeend & ~(long long)ULONG_MAX)
2823                         hlen = ULONG_MAX - hba + 1;
2824         }
2825
2826         details.check_mapping = even_cows ? NULL : mapping;
2827         details.first_index = hba;
2828         details.last_index = hba + hlen - 1;
2829         if (details.last_index < details.first_index)
2830                 details.last_index = ULONG_MAX;
2831
2832         i_mmap_lock_write(mapping);
2833         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2834                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2835         i_mmap_unlock_write(mapping);
2836 }
2837 EXPORT_SYMBOL(unmap_mapping_range);
2838
2839 /*
2840  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2841  * but allow concurrent faults), and pte mapped but not yet locked.
2842  * We return with pte unmapped and unlocked.
2843  *
2844  * We return with the mmap_sem locked or unlocked in the same cases
2845  * as does filemap_fault().
2846  */
2847 int do_swap_page(struct vm_fault *vmf)
2848 {
2849         struct vm_area_struct *vma = vmf->vma;
2850         struct page *page = NULL, *swapcache = NULL;
2851         struct mem_cgroup *memcg;
2852         struct vma_swap_readahead swap_ra;
2853         swp_entry_t entry;
2854         pte_t pte;
2855         int locked;
2856         int exclusive = 0;
2857         int ret = 0;
2858         bool vma_readahead = swap_use_vma_readahead();
2859
2860         if (vma_readahead)
2861                 page = swap_readahead_detect(vmf, &swap_ra);
2862         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
2863                 if (page)
2864                         put_page(page);
2865                 goto out;
2866         }
2867
2868         entry = pte_to_swp_entry(vmf->orig_pte);
2869         if (unlikely(non_swap_entry(entry))) {
2870                 if (is_migration_entry(entry)) {
2871                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2872                                              vmf->address);
2873                 } else if (is_device_private_entry(entry)) {
2874                         /*
2875                          * For un-addressable device memory we call the pgmap
2876                          * fault handler callback. The callback must migrate
2877                          * the page back to some CPU accessible page.
2878                          */
2879                         ret = device_private_entry_fault(vma, vmf->address, entry,
2880                                                  vmf->flags, vmf->pmd);
2881                 } else if (is_hwpoison_entry(entry)) {
2882                         ret = VM_FAULT_HWPOISON;
2883                 } else {
2884                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2885                         ret = VM_FAULT_SIGBUS;
2886                 }
2887                 goto out;
2888         }
2889
2890
2891         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2892         if (!page)
2893                 page = lookup_swap_cache(entry, vma_readahead ? vma : NULL,
2894                                          vmf->address);
2895         if (!page) {
2896                 struct swap_info_struct *si = swp_swap_info(entry);
2897
2898                 if (si->flags & SWP_SYNCHRONOUS_IO &&
2899                                 __swap_count(si, entry) == 1) {
2900                         /* skip swapcache */
2901                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2902                         if (page) {
2903                                 __SetPageLocked(page);
2904                                 __SetPageSwapBacked(page);
2905                                 set_page_private(page, entry.val);
2906                                 lru_cache_add_anon(page);
2907                                 swap_readpage(page, true);
2908                         }
2909                 } else {
2910                         if (vma_readahead)
2911                                 page = do_swap_page_readahead(entry,
2912                                         GFP_HIGHUSER_MOVABLE, vmf, &swap_ra);
2913                         else
2914                                 page = swapin_readahead(entry,
2915                                        GFP_HIGHUSER_MOVABLE, vma, vmf->address);
2916                         swapcache = page;
2917                 }
2918
2919                 if (!page) {
2920                         /*
2921                          * Back out if somebody else faulted in this pte
2922                          * while we released the pte lock.
2923                          */
2924                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2925                                         vmf->address, &vmf->ptl);
2926                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2927                                 ret = VM_FAULT_OOM;
2928                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2929                         goto unlock;
2930                 }
2931
2932                 /* Had to read the page from swap area: Major fault */
2933                 ret = VM_FAULT_MAJOR;
2934                 count_vm_event(PGMAJFAULT);
2935                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2936         } else if (PageHWPoison(page)) {
2937                 /*
2938                  * hwpoisoned dirty swapcache pages are kept for killing
2939                  * owner processes (which may be unknown at hwpoison time)
2940                  */
2941                 ret = VM_FAULT_HWPOISON;
2942                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2943                 swapcache = page;
2944                 goto out_release;
2945         }
2946
2947         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2948
2949         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2950         if (!locked) {
2951                 ret |= VM_FAULT_RETRY;
2952                 goto out_release;
2953         }
2954
2955         /*
2956          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2957          * release the swapcache from under us.  The page pin, and pte_same
2958          * test below, are not enough to exclude that.  Even if it is still
2959          * swapcache, we need to check that the page's swap has not changed.
2960          */
2961         if (unlikely((!PageSwapCache(page) ||
2962                         page_private(page) != entry.val)) && swapcache)
2963                 goto out_page;
2964
2965         page = ksm_might_need_to_copy(page, vma, vmf->address);
2966         if (unlikely(!page)) {
2967                 ret = VM_FAULT_OOM;
2968                 page = swapcache;
2969                 goto out_page;
2970         }
2971
2972         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2973                                 &memcg, false)) {
2974                 ret = VM_FAULT_OOM;
2975                 goto out_page;
2976         }
2977
2978         /*
2979          * Back out if somebody else already faulted in this pte.
2980          */
2981         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2982                         &vmf->ptl);
2983         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2984                 goto out_nomap;
2985
2986         if (unlikely(!PageUptodate(page))) {
2987                 ret = VM_FAULT_SIGBUS;
2988                 goto out_nomap;
2989         }
2990
2991         /*
2992          * The page isn't present yet, go ahead with the fault.
2993          *
2994          * Be careful about the sequence of operations here.
2995          * To get its accounting right, reuse_swap_page() must be called
2996          * while the page is counted on swap but not yet in mapcount i.e.
2997          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2998          * must be called after the swap_free(), or it will never succeed.
2999          */
3000
3001         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3002         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3003         pte = mk_pte(page, vma->vm_page_prot);
3004         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3005                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3006                 vmf->flags &= ~FAULT_FLAG_WRITE;
3007                 ret |= VM_FAULT_WRITE;
3008                 exclusive = RMAP_EXCLUSIVE;
3009         }
3010         flush_icache_page(vma, page);
3011         if (pte_swp_soft_dirty(vmf->orig_pte))
3012                 pte = pte_mksoft_dirty(pte);
3013         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3014         vmf->orig_pte = pte;
3015
3016         /* ksm created a completely new copy */
3017         if (unlikely(page != swapcache && swapcache)) {
3018                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3019                 mem_cgroup_commit_charge(page, memcg, false, false);
3020                 lru_cache_add_active_or_unevictable(page, vma);
3021         } else {
3022                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3023                 mem_cgroup_commit_charge(page, memcg, true, false);
3024                 activate_page(page);
3025         }
3026
3027         swap_free(entry);
3028         if (mem_cgroup_swap_full(page) ||
3029             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3030                 try_to_free_swap(page);
3031         unlock_page(page);
3032         if (page != swapcache && swapcache) {
3033                 /*
3034                  * Hold the lock to avoid the swap entry to be reused
3035                  * until we take the PT lock for the pte_same() check
3036                  * (to avoid false positives from pte_same). For
3037                  * further safety release the lock after the swap_free
3038                  * so that the swap count won't change under a
3039                  * parallel locked swapcache.
3040                  */
3041                 unlock_page(swapcache);
3042                 put_page(swapcache);
3043         }
3044
3045         if (vmf->flags & FAULT_FLAG_WRITE) {
3046                 ret |= do_wp_page(vmf);
3047                 if (ret & VM_FAULT_ERROR)
3048                         ret &= VM_FAULT_ERROR;
3049                 goto out;
3050         }
3051
3052         /* No need to invalidate - it was non-present before */
3053         update_mmu_cache(vma, vmf->address, vmf->pte);
3054 unlock:
3055         pte_unmap_unlock(vmf->pte, vmf->ptl);
3056 out:
3057         return ret;
3058 out_nomap:
3059         mem_cgroup_cancel_charge(page, memcg, false);
3060         pte_unmap_unlock(vmf->pte, vmf->ptl);
3061 out_page:
3062         unlock_page(page);
3063 out_release:
3064         put_page(page);
3065         if (page != swapcache && swapcache) {
3066                 unlock_page(swapcache);
3067                 put_page(swapcache);
3068         }
3069         return ret;
3070 }
3071
3072 /*
3073  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3074  * but allow concurrent faults), and pte mapped but not yet locked.
3075  * We return with mmap_sem still held, but pte unmapped and unlocked.
3076  */
3077 static int do_anonymous_page(struct vm_fault *vmf)
3078 {
3079         struct vm_area_struct *vma = vmf->vma;
3080         struct mem_cgroup *memcg;
3081         struct page *page;
3082         int ret = 0;
3083         pte_t entry;
3084
3085         /* File mapping without ->vm_ops ? */
3086         if (vma->vm_flags & VM_SHARED)
3087                 return VM_FAULT_SIGBUS;
3088
3089         /*
3090          * Use pte_alloc() instead of pte_alloc_map().  We can't run
3091          * pte_offset_map() on pmds where a huge pmd might be created
3092          * from a different thread.
3093          *
3094          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3095          * parallel threads are excluded by other means.
3096          *
3097          * Here we only have down_read(mmap_sem).
3098          */
3099         if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3100                 return VM_FAULT_OOM;
3101
3102         /* See the comment in pte_alloc_one_map() */
3103         if (unlikely(pmd_trans_unstable(vmf->pmd)))
3104                 return 0;
3105
3106         /* Use the zero-page for reads */
3107         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3108                         !mm_forbids_zeropage(vma->vm_mm)) {
3109                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3110                                                 vma->vm_page_prot));
3111                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3112                                 vmf->address, &vmf->ptl);
3113                 if (!pte_none(*vmf->pte))
3114                         goto unlock;
3115                 ret = check_stable_address_space(vma->vm_mm);
3116                 if (ret)
3117                         goto unlock;
3118                 /* Deliver the page fault to userland, check inside PT lock */
3119                 if (userfaultfd_missing(vma)) {
3120                         pte_unmap_unlock(vmf->pte, vmf->ptl);
3121                         return handle_userfault(vmf, VM_UFFD_MISSING);
3122                 }
3123                 goto setpte;
3124         }
3125
3126         /* Allocate our own private page. */
3127         if (unlikely(anon_vma_prepare(vma)))
3128                 goto oom;
3129         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3130         if (!page)
3131                 goto oom;
3132
3133         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3134                 goto oom_free_page;
3135
3136         /*
3137          * The memory barrier inside __SetPageUptodate makes sure that
3138          * preceeding stores to the page contents become visible before
3139          * the set_pte_at() write.
3140          */
3141         __SetPageUptodate(page);
3142
3143         entry = mk_pte(page, vma->vm_page_prot);
3144         if (vma->vm_flags & VM_WRITE)
3145                 entry = pte_mkwrite(pte_mkdirty(entry));
3146
3147         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3148                         &vmf->ptl);
3149         if (!pte_none(*vmf->pte))
3150                 goto release;
3151
3152         ret = check_stable_address_space(vma->vm_mm);
3153         if (ret)
3154                 goto release;
3155
3156         /* Deliver the page fault to userland, check inside PT lock */
3157         if (userfaultfd_missing(vma)) {
3158                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3159                 mem_cgroup_cancel_charge(page, memcg, false);
3160                 put_page(page);
3161                 return handle_userfault(vmf, VM_UFFD_MISSING);
3162         }
3163
3164         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3165         page_add_new_anon_rmap(page, vma, vmf->address, false);
3166         mem_cgroup_commit_charge(page, memcg, false, false);
3167         lru_cache_add_active_or_unevictable(page, vma);
3168 setpte:
3169         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3170
3171         /* No need to invalidate - it was non-present before */
3172         update_mmu_cache(vma, vmf->address, vmf->pte);
3173 unlock:
3174         pte_unmap_unlock(vmf->pte, vmf->ptl);
3175         return ret;
3176 release:
3177         mem_cgroup_cancel_charge(page, memcg, false);
3178         put_page(page);
3179         goto unlock;
3180 oom_free_page:
3181         put_page(page);
3182 oom:
3183         return VM_FAULT_OOM;
3184 }
3185
3186 /*
3187  * The mmap_sem must have been held on entry, and may have been
3188  * released depending on flags and vma->vm_ops->fault() return value.
3189  * See filemap_fault() and __lock_page_retry().
3190  */
3191 static int __do_fault(struct vm_fault *vmf)
3192 {
3193         struct vm_area_struct *vma = vmf->vma;
3194         int ret;
3195
3196         ret = vma->vm_ops->fault(vmf);
3197         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3198                             VM_FAULT_DONE_COW)))
3199                 return ret;
3200
3201         if (unlikely(PageHWPoison(vmf->page))) {
3202                 if (ret & VM_FAULT_LOCKED)
3203                         unlock_page(vmf->page);
3204                 put_page(vmf->page);
3205                 vmf->page = NULL;
3206                 return VM_FAULT_HWPOISON;
3207         }
3208
3209         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3210                 lock_page(vmf->page);
3211         else
3212                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3213
3214         return ret;
3215 }
3216
3217 /*
3218  * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3219  * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3220  * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3221  * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3222  */
3223 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3224 {
3225         return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3226 }
3227
3228 static int pte_alloc_one_map(struct vm_fault *vmf)
3229 {
3230         struct vm_area_struct *vma = vmf->vma;
3231
3232         if (!pmd_none(*vmf->pmd))
3233                 goto map_pte;
3234         if (vmf->prealloc_pte) {
3235                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3236                 if (unlikely(!pmd_none(*vmf->pmd))) {
3237                         spin_unlock(vmf->ptl);
3238                         goto map_pte;
3239                 }
3240
3241                 mm_inc_nr_ptes(vma->vm_mm);
3242                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3243                 spin_unlock(vmf->ptl);
3244                 vmf->prealloc_pte = NULL;
3245         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3246                 return VM_FAULT_OOM;
3247         }
3248 map_pte:
3249         /*
3250          * If a huge pmd materialized under us just retry later.  Use
3251          * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3252          * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3253          * under us and then back to pmd_none, as a result of MADV_DONTNEED
3254          * running immediately after a huge pmd fault in a different thread of
3255          * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3256          * All we have to ensure is that it is a regular pmd that we can walk
3257          * with pte_offset_map() and we can do that through an atomic read in
3258          * C, which is what pmd_trans_unstable() provides.
3259          */
3260         if (pmd_devmap_trans_unstable(vmf->pmd))
3261                 return VM_FAULT_NOPAGE;
3262
3263         /*
3264          * At this point we know that our vmf->pmd points to a page of ptes
3265          * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3266          * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3267          * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3268          * be valid and we will re-check to make sure the vmf->pte isn't
3269          * pte_none() under vmf->ptl protection when we return to
3270          * alloc_set_pte().
3271          */
3272         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3273                         &vmf->ptl);
3274         return 0;
3275 }
3276
3277 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3278
3279 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3280 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3281                 unsigned long haddr)
3282 {
3283         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3284                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3285                 return false;
3286         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3287                 return false;
3288         return true;
3289 }
3290
3291 static void deposit_prealloc_pte(struct vm_fault *vmf)
3292 {
3293         struct vm_area_struct *vma = vmf->vma;
3294
3295         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3296         /*
3297          * We are going to consume the prealloc table,