]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - mm/memory.c
Merge branch 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs
[karo-tx-linux.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/ksm.h>
53 #include <linux/rmap.h>
54 #include <linux/export.h>
55 #include <linux/delayacct.h>
56 #include <linux/init.h>
57 #include <linux/pfn_t.h>
58 #include <linux/writeback.h>
59 #include <linux/memcontrol.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/kallsyms.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71
72 #include <asm/io.h>
73 #include <asm/mmu_context.h>
74 #include <asm/pgalloc.h>
75 #include <linux/uaccess.h>
76 #include <asm/tlb.h>
77 #include <asm/tlbflush.h>
78 #include <asm/pgtable.h>
79
80 #include "internal.h"
81
82 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
83 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
84 #endif
85
86 #ifndef CONFIG_NEED_MULTIPLE_NODES
87 /* use the per-pgdat data instead for discontigmem - mbligh */
88 unsigned long max_mapnr;
89 EXPORT_SYMBOL(max_mapnr);
90
91 struct page *mem_map;
92 EXPORT_SYMBOL(mem_map);
93 #endif
94
95 /*
96  * A number of key systems in x86 including ioremap() rely on the assumption
97  * that high_memory defines the upper bound on direct map memory, then end
98  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
99  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
100  * and ZONE_HIGHMEM.
101  */
102 void *high_memory;
103 EXPORT_SYMBOL(high_memory);
104
105 /*
106  * Randomize the address space (stacks, mmaps, brk, etc.).
107  *
108  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
109  *   as ancient (libc5 based) binaries can segfault. )
110  */
111 int randomize_va_space __read_mostly =
112 #ifdef CONFIG_COMPAT_BRK
113                                         1;
114 #else
115                                         2;
116 #endif
117
118 static int __init disable_randmaps(char *s)
119 {
120         randomize_va_space = 0;
121         return 1;
122 }
123 __setup("norandmaps", disable_randmaps);
124
125 unsigned long zero_pfn __read_mostly;
126 EXPORT_SYMBOL(zero_pfn);
127
128 unsigned long highest_memmap_pfn __read_mostly;
129
130 /*
131  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
132  */
133 static int __init init_zero_pfn(void)
134 {
135         zero_pfn = page_to_pfn(ZERO_PAGE(0));
136         return 0;
137 }
138 core_initcall(init_zero_pfn);
139
140
141 #if defined(SPLIT_RSS_COUNTING)
142
143 void sync_mm_rss(struct mm_struct *mm)
144 {
145         int i;
146
147         for (i = 0; i < NR_MM_COUNTERS; i++) {
148                 if (current->rss_stat.count[i]) {
149                         add_mm_counter(mm, i, current->rss_stat.count[i]);
150                         current->rss_stat.count[i] = 0;
151                 }
152         }
153         current->rss_stat.events = 0;
154 }
155
156 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
157 {
158         struct task_struct *task = current;
159
160         if (likely(task->mm == mm))
161                 task->rss_stat.count[member] += val;
162         else
163                 add_mm_counter(mm, member, val);
164 }
165 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
166 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
167
168 /* sync counter once per 64 page faults */
169 #define TASK_RSS_EVENTS_THRESH  (64)
170 static void check_sync_rss_stat(struct task_struct *task)
171 {
172         if (unlikely(task != current))
173                 return;
174         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
175                 sync_mm_rss(task->mm);
176 }
177 #else /* SPLIT_RSS_COUNTING */
178
179 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
180 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
181
182 static void check_sync_rss_stat(struct task_struct *task)
183 {
184 }
185
186 #endif /* SPLIT_RSS_COUNTING */
187
188 #ifdef HAVE_GENERIC_MMU_GATHER
189
190 static bool tlb_next_batch(struct mmu_gather *tlb)
191 {
192         struct mmu_gather_batch *batch;
193
194         batch = tlb->active;
195         if (batch->next) {
196                 tlb->active = batch->next;
197                 return true;
198         }
199
200         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
201                 return false;
202
203         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
204         if (!batch)
205                 return false;
206
207         tlb->batch_count++;
208         batch->next = NULL;
209         batch->nr   = 0;
210         batch->max  = MAX_GATHER_BATCH;
211
212         tlb->active->next = batch;
213         tlb->active = batch;
214
215         return true;
216 }
217
218 void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
219                                 unsigned long start, unsigned long end)
220 {
221         tlb->mm = mm;
222
223         /* Is it from 0 to ~0? */
224         tlb->fullmm     = !(start | (end+1));
225         tlb->need_flush_all = 0;
226         tlb->local.next = NULL;
227         tlb->local.nr   = 0;
228         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
229         tlb->active     = &tlb->local;
230         tlb->batch_count = 0;
231
232 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
233         tlb->batch = NULL;
234 #endif
235         tlb->page_size = 0;
236
237         __tlb_reset_range(tlb);
238 }
239
240 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
241 {
242         if (!tlb->end)
243                 return;
244
245         tlb_flush(tlb);
246         mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
247 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
248         tlb_table_flush(tlb);
249 #endif
250         __tlb_reset_range(tlb);
251 }
252
253 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
254 {
255         struct mmu_gather_batch *batch;
256
257         for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
258                 free_pages_and_swap_cache(batch->pages, batch->nr);
259                 batch->nr = 0;
260         }
261         tlb->active = &tlb->local;
262 }
263
264 void tlb_flush_mmu(struct mmu_gather *tlb)
265 {
266         tlb_flush_mmu_tlbonly(tlb);
267         tlb_flush_mmu_free(tlb);
268 }
269
270 /* tlb_finish_mmu
271  *      Called at the end of the shootdown operation to free up any resources
272  *      that were required.
273  */
274 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
275                 unsigned long start, unsigned long end, bool force)
276 {
277         struct mmu_gather_batch *batch, *next;
278
279         if (force)
280                 __tlb_adjust_range(tlb, start, end - start);
281
282         tlb_flush_mmu(tlb);
283
284         /* keep the page table cache within bounds */
285         check_pgt_cache();
286
287         for (batch = tlb->local.next; batch; batch = next) {
288                 next = batch->next;
289                 free_pages((unsigned long)batch, 0);
290         }
291         tlb->local.next = NULL;
292 }
293
294 /* __tlb_remove_page
295  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
296  *      handling the additional races in SMP caused by other CPUs caching valid
297  *      mappings in their TLBs. Returns the number of free page slots left.
298  *      When out of page slots we must call tlb_flush_mmu().
299  *returns true if the caller should flush.
300  */
301 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
302 {
303         struct mmu_gather_batch *batch;
304
305         VM_BUG_ON(!tlb->end);
306         VM_WARN_ON(tlb->page_size != page_size);
307
308         batch = tlb->active;
309         /*
310          * Add the page and check if we are full. If so
311          * force a flush.
312          */
313         batch->pages[batch->nr++] = page;
314         if (batch->nr == batch->max) {
315                 if (!tlb_next_batch(tlb))
316                         return true;
317                 batch = tlb->active;
318         }
319         VM_BUG_ON_PAGE(batch->nr > batch->max, page);
320
321         return false;
322 }
323
324 #endif /* HAVE_GENERIC_MMU_GATHER */
325
326 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
327
328 /*
329  * See the comment near struct mmu_table_batch.
330  */
331
332 static void tlb_remove_table_smp_sync(void *arg)
333 {
334         /* Simply deliver the interrupt */
335 }
336
337 static void tlb_remove_table_one(void *table)
338 {
339         /*
340          * This isn't an RCU grace period and hence the page-tables cannot be
341          * assumed to be actually RCU-freed.
342          *
343          * It is however sufficient for software page-table walkers that rely on
344          * IRQ disabling. See the comment near struct mmu_table_batch.
345          */
346         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
347         __tlb_remove_table(table);
348 }
349
350 static void tlb_remove_table_rcu(struct rcu_head *head)
351 {
352         struct mmu_table_batch *batch;
353         int i;
354
355         batch = container_of(head, struct mmu_table_batch, rcu);
356
357         for (i = 0; i < batch->nr; i++)
358                 __tlb_remove_table(batch->tables[i]);
359
360         free_page((unsigned long)batch);
361 }
362
363 void tlb_table_flush(struct mmu_gather *tlb)
364 {
365         struct mmu_table_batch **batch = &tlb->batch;
366
367         if (*batch) {
368                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
369                 *batch = NULL;
370         }
371 }
372
373 void tlb_remove_table(struct mmu_gather *tlb, void *table)
374 {
375         struct mmu_table_batch **batch = &tlb->batch;
376
377         /*
378          * When there's less then two users of this mm there cannot be a
379          * concurrent page-table walk.
380          */
381         if (atomic_read(&tlb->mm->mm_users) < 2) {
382                 __tlb_remove_table(table);
383                 return;
384         }
385
386         if (*batch == NULL) {
387                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
388                 if (*batch == NULL) {
389                         tlb_remove_table_one(table);
390                         return;
391                 }
392                 (*batch)->nr = 0;
393         }
394         (*batch)->tables[(*batch)->nr++] = table;
395         if ((*batch)->nr == MAX_TABLE_BATCH)
396                 tlb_table_flush(tlb);
397 }
398
399 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
400
401 /* tlb_gather_mmu
402  *      Called to initialize an (on-stack) mmu_gather structure for page-table
403  *      tear-down from @mm. The @fullmm argument is used when @mm is without
404  *      users and we're going to destroy the full address space (exit/execve).
405  */
406 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
407                         unsigned long start, unsigned long end)
408 {
409         arch_tlb_gather_mmu(tlb, mm, start, end);
410         inc_tlb_flush_pending(tlb->mm);
411 }
412
413 void tlb_finish_mmu(struct mmu_gather *tlb,
414                 unsigned long start, unsigned long end)
415 {
416         /*
417          * If there are parallel threads are doing PTE changes on same range
418          * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
419          * flush by batching, a thread has stable TLB entry can fail to flush
420          * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
421          * forcefully if we detect parallel PTE batching threads.
422          */
423         bool force = mm_tlb_flush_nested(tlb->mm);
424
425         arch_tlb_finish_mmu(tlb, start, end, force);
426         dec_tlb_flush_pending(tlb->mm);
427 }
428
429 /*
430  * Note: this doesn't free the actual pages themselves. That
431  * has been handled earlier when unmapping all the memory regions.
432  */
433 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
434                            unsigned long addr)
435 {
436         pgtable_t token = pmd_pgtable(*pmd);
437         pmd_clear(pmd);
438         pte_free_tlb(tlb, token, addr);
439         atomic_long_dec(&tlb->mm->nr_ptes);
440 }
441
442 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
443                                 unsigned long addr, unsigned long end,
444                                 unsigned long floor, unsigned long ceiling)
445 {
446         pmd_t *pmd;
447         unsigned long next;
448         unsigned long start;
449
450         start = addr;
451         pmd = pmd_offset(pud, addr);
452         do {
453                 next = pmd_addr_end(addr, end);
454                 if (pmd_none_or_clear_bad(pmd))
455                         continue;
456                 free_pte_range(tlb, pmd, addr);
457         } while (pmd++, addr = next, addr != end);
458
459         start &= PUD_MASK;
460         if (start < floor)
461                 return;
462         if (ceiling) {
463                 ceiling &= PUD_MASK;
464                 if (!ceiling)
465                         return;
466         }
467         if (end - 1 > ceiling - 1)
468                 return;
469
470         pmd = pmd_offset(pud, start);
471         pud_clear(pud);
472         pmd_free_tlb(tlb, pmd, start);
473         mm_dec_nr_pmds(tlb->mm);
474 }
475
476 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
477                                 unsigned long addr, unsigned long end,
478                                 unsigned long floor, unsigned long ceiling)
479 {
480         pud_t *pud;
481         unsigned long next;
482         unsigned long start;
483
484         start = addr;
485         pud = pud_offset(p4d, addr);
486         do {
487                 next = pud_addr_end(addr, end);
488                 if (pud_none_or_clear_bad(pud))
489                         continue;
490                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
491         } while (pud++, addr = next, addr != end);
492
493         start &= P4D_MASK;
494         if (start < floor)
495                 return;
496         if (ceiling) {
497                 ceiling &= P4D_MASK;
498                 if (!ceiling)
499                         return;
500         }
501         if (end - 1 > ceiling - 1)
502                 return;
503
504         pud = pud_offset(p4d, start);
505         p4d_clear(p4d);
506         pud_free_tlb(tlb, pud, start);
507 }
508
509 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
510                                 unsigned long addr, unsigned long end,
511                                 unsigned long floor, unsigned long ceiling)
512 {
513         p4d_t *p4d;
514         unsigned long next;
515         unsigned long start;
516
517         start = addr;
518         p4d = p4d_offset(pgd, addr);
519         do {
520                 next = p4d_addr_end(addr, end);
521                 if (p4d_none_or_clear_bad(p4d))
522                         continue;
523                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
524         } while (p4d++, addr = next, addr != end);
525
526         start &= PGDIR_MASK;
527         if (start < floor)
528                 return;
529         if (ceiling) {
530                 ceiling &= PGDIR_MASK;
531                 if (!ceiling)
532                         return;
533         }
534         if (end - 1 > ceiling - 1)
535                 return;
536
537         p4d = p4d_offset(pgd, start);
538         pgd_clear(pgd);
539         p4d_free_tlb(tlb, p4d, start);
540 }
541
542 /*
543  * This function frees user-level page tables of a process.
544  */
545 void free_pgd_range(struct mmu_gather *tlb,
546                         unsigned long addr, unsigned long end,
547                         unsigned long floor, unsigned long ceiling)
548 {
549         pgd_t *pgd;
550         unsigned long next;
551
552         /*
553          * The next few lines have given us lots of grief...
554          *
555          * Why are we testing PMD* at this top level?  Because often
556          * there will be no work to do at all, and we'd prefer not to
557          * go all the way down to the bottom just to discover that.
558          *
559          * Why all these "- 1"s?  Because 0 represents both the bottom
560          * of the address space and the top of it (using -1 for the
561          * top wouldn't help much: the masks would do the wrong thing).
562          * The rule is that addr 0 and floor 0 refer to the bottom of
563          * the address space, but end 0 and ceiling 0 refer to the top
564          * Comparisons need to use "end - 1" and "ceiling - 1" (though
565          * that end 0 case should be mythical).
566          *
567          * Wherever addr is brought up or ceiling brought down, we must
568          * be careful to reject "the opposite 0" before it confuses the
569          * subsequent tests.  But what about where end is brought down
570          * by PMD_SIZE below? no, end can't go down to 0 there.
571          *
572          * Whereas we round start (addr) and ceiling down, by different
573          * masks at different levels, in order to test whether a table
574          * now has no other vmas using it, so can be freed, we don't
575          * bother to round floor or end up - the tests don't need that.
576          */
577
578         addr &= PMD_MASK;
579         if (addr < floor) {
580                 addr += PMD_SIZE;
581                 if (!addr)
582                         return;
583         }
584         if (ceiling) {
585                 ceiling &= PMD_MASK;
586                 if (!ceiling)
587                         return;
588         }
589         if (end - 1 > ceiling - 1)
590                 end -= PMD_SIZE;
591         if (addr > end - 1)
592                 return;
593         /*
594          * We add page table cache pages with PAGE_SIZE,
595          * (see pte_free_tlb()), flush the tlb if we need
596          */
597         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
598         pgd = pgd_offset(tlb->mm, addr);
599         do {
600                 next = pgd_addr_end(addr, end);
601                 if (pgd_none_or_clear_bad(pgd))
602                         continue;
603                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
604         } while (pgd++, addr = next, addr != end);
605 }
606
607 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
608                 unsigned long floor, unsigned long ceiling)
609 {
610         while (vma) {
611                 struct vm_area_struct *next = vma->vm_next;
612                 unsigned long addr = vma->vm_start;
613
614                 /*
615                  * Hide vma from rmap and truncate_pagecache before freeing
616                  * pgtables
617                  */
618                 unlink_anon_vmas(vma);
619                 unlink_file_vma(vma);
620
621                 if (is_vm_hugetlb_page(vma)) {
622                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
623                                 floor, next ? next->vm_start : ceiling);
624                 } else {
625                         /*
626                          * Optimization: gather nearby vmas into one call down
627                          */
628                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
629                                && !is_vm_hugetlb_page(next)) {
630                                 vma = next;
631                                 next = vma->vm_next;
632                                 unlink_anon_vmas(vma);
633                                 unlink_file_vma(vma);
634                         }
635                         free_pgd_range(tlb, addr, vma->vm_end,
636                                 floor, next ? next->vm_start : ceiling);
637                 }
638                 vma = next;
639         }
640 }
641
642 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
643 {
644         spinlock_t *ptl;
645         pgtable_t new = pte_alloc_one(mm, address);
646         if (!new)
647                 return -ENOMEM;
648
649         /*
650          * Ensure all pte setup (eg. pte page lock and page clearing) are
651          * visible before the pte is made visible to other CPUs by being
652          * put into page tables.
653          *
654          * The other side of the story is the pointer chasing in the page
655          * table walking code (when walking the page table without locking;
656          * ie. most of the time). Fortunately, these data accesses consist
657          * of a chain of data-dependent loads, meaning most CPUs (alpha
658          * being the notable exception) will already guarantee loads are
659          * seen in-order. See the alpha page table accessors for the
660          * smp_read_barrier_depends() barriers in page table walking code.
661          */
662         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
663
664         ptl = pmd_lock(mm, pmd);
665         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
666                 atomic_long_inc(&mm->nr_ptes);
667                 pmd_populate(mm, pmd, new);
668                 new = NULL;
669         }
670         spin_unlock(ptl);
671         if (new)
672                 pte_free(mm, new);
673         return 0;
674 }
675
676 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
677 {
678         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
679         if (!new)
680                 return -ENOMEM;
681
682         smp_wmb(); /* See comment in __pte_alloc */
683
684         spin_lock(&init_mm.page_table_lock);
685         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
686                 pmd_populate_kernel(&init_mm, pmd, new);
687                 new = NULL;
688         }
689         spin_unlock(&init_mm.page_table_lock);
690         if (new)
691                 pte_free_kernel(&init_mm, new);
692         return 0;
693 }
694
695 static inline void init_rss_vec(int *rss)
696 {
697         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
698 }
699
700 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
701 {
702         int i;
703
704         if (current->mm == mm)
705                 sync_mm_rss(mm);
706         for (i = 0; i < NR_MM_COUNTERS; i++)
707                 if (rss[i])
708                         add_mm_counter(mm, i, rss[i]);
709 }
710
711 /*
712  * This function is called to print an error when a bad pte
713  * is found. For example, we might have a PFN-mapped pte in
714  * a region that doesn't allow it.
715  *
716  * The calling function must still handle the error.
717  */
718 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
719                           pte_t pte, struct page *page)
720 {
721         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
722         p4d_t *p4d = p4d_offset(pgd, addr);
723         pud_t *pud = pud_offset(p4d, addr);
724         pmd_t *pmd = pmd_offset(pud, addr);
725         struct address_space *mapping;
726         pgoff_t index;
727         static unsigned long resume;
728         static unsigned long nr_shown;
729         static unsigned long nr_unshown;
730
731         /*
732          * Allow a burst of 60 reports, then keep quiet for that minute;
733          * or allow a steady drip of one report per second.
734          */
735         if (nr_shown == 60) {
736                 if (time_before(jiffies, resume)) {
737                         nr_unshown++;
738                         return;
739                 }
740                 if (nr_unshown) {
741                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
742                                  nr_unshown);
743                         nr_unshown = 0;
744                 }
745                 nr_shown = 0;
746         }
747         if (nr_shown++ == 0)
748                 resume = jiffies + 60 * HZ;
749
750         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
751         index = linear_page_index(vma, addr);
752
753         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
754                  current->comm,
755                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
756         if (page)
757                 dump_page(page, "bad pte");
758         pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
759                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
760         /*
761          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
762          */
763         pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
764                  vma->vm_file,
765                  vma->vm_ops ? vma->vm_ops->fault : NULL,
766                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
767                  mapping ? mapping->a_ops->readpage : NULL);
768         dump_stack();
769         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
770 }
771
772 /*
773  * vm_normal_page -- This function gets the "struct page" associated with a pte.
774  *
775  * "Special" mappings do not wish to be associated with a "struct page" (either
776  * it doesn't exist, or it exists but they don't want to touch it). In this
777  * case, NULL is returned here. "Normal" mappings do have a struct page.
778  *
779  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
780  * pte bit, in which case this function is trivial. Secondly, an architecture
781  * may not have a spare pte bit, which requires a more complicated scheme,
782  * described below.
783  *
784  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
785  * special mapping (even if there are underlying and valid "struct pages").
786  * COWed pages of a VM_PFNMAP are always normal.
787  *
788  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
789  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
790  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
791  * mapping will always honor the rule
792  *
793  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
794  *
795  * And for normal mappings this is false.
796  *
797  * This restricts such mappings to be a linear translation from virtual address
798  * to pfn. To get around this restriction, we allow arbitrary mappings so long
799  * as the vma is not a COW mapping; in that case, we know that all ptes are
800  * special (because none can have been COWed).
801  *
802  *
803  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
804  *
805  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
806  * page" backing, however the difference is that _all_ pages with a struct
807  * page (that is, those where pfn_valid is true) are refcounted and considered
808  * normal pages by the VM. The disadvantage is that pages are refcounted
809  * (which can be slower and simply not an option for some PFNMAP users). The
810  * advantage is that we don't have to follow the strict linearity rule of
811  * PFNMAP mappings in order to support COWable mappings.
812  *
813  */
814 #ifdef __HAVE_ARCH_PTE_SPECIAL
815 # define HAVE_PTE_SPECIAL 1
816 #else
817 # define HAVE_PTE_SPECIAL 0
818 #endif
819 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
820                                 pte_t pte)
821 {
822         unsigned long pfn = pte_pfn(pte);
823
824         if (HAVE_PTE_SPECIAL) {
825                 if (likely(!pte_special(pte)))
826                         goto check_pfn;
827                 if (vma->vm_ops && vma->vm_ops->find_special_page)
828                         return vma->vm_ops->find_special_page(vma, addr);
829                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
830                         return NULL;
831                 if (!is_zero_pfn(pfn))
832                         print_bad_pte(vma, addr, pte, NULL);
833                 return NULL;
834         }
835
836         /* !HAVE_PTE_SPECIAL case follows: */
837
838         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
839                 if (vma->vm_flags & VM_MIXEDMAP) {
840                         if (!pfn_valid(pfn))
841                                 return NULL;
842                         goto out;
843                 } else {
844                         unsigned long off;
845                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
846                         if (pfn == vma->vm_pgoff + off)
847                                 return NULL;
848                         if (!is_cow_mapping(vma->vm_flags))
849                                 return NULL;
850                 }
851         }
852
853         if (is_zero_pfn(pfn))
854                 return NULL;
855 check_pfn:
856         if (unlikely(pfn > highest_memmap_pfn)) {
857                 print_bad_pte(vma, addr, pte, NULL);
858                 return NULL;
859         }
860
861         /*
862          * NOTE! We still have PageReserved() pages in the page tables.
863          * eg. VDSO mappings can cause them to exist.
864          */
865 out:
866         return pfn_to_page(pfn);
867 }
868
869 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
870 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
871                                 pmd_t pmd)
872 {
873         unsigned long pfn = pmd_pfn(pmd);
874
875         /*
876          * There is no pmd_special() but there may be special pmds, e.g.
877          * in a direct-access (dax) mapping, so let's just replicate the
878          * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
879          */
880         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
881                 if (vma->vm_flags & VM_MIXEDMAP) {
882                         if (!pfn_valid(pfn))
883                                 return NULL;
884                         goto out;
885                 } else {
886                         unsigned long off;
887                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
888                         if (pfn == vma->vm_pgoff + off)
889                                 return NULL;
890                         if (!is_cow_mapping(vma->vm_flags))
891                                 return NULL;
892                 }
893         }
894
895         if (is_zero_pfn(pfn))
896                 return NULL;
897         if (unlikely(pfn > highest_memmap_pfn))
898                 return NULL;
899
900         /*
901          * NOTE! We still have PageReserved() pages in the page tables.
902          * eg. VDSO mappings can cause them to exist.
903          */
904 out:
905         return pfn_to_page(pfn);
906 }
907 #endif
908
909 /*
910  * copy one vm_area from one task to the other. Assumes the page tables
911  * already present in the new task to be cleared in the whole range
912  * covered by this vma.
913  */
914
915 static inline unsigned long
916 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
917                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
918                 unsigned long addr, int *rss)
919 {
920         unsigned long vm_flags = vma->vm_flags;
921         pte_t pte = *src_pte;
922         struct page *page;
923
924         /* pte contains position in swap or file, so copy. */
925         if (unlikely(!pte_present(pte))) {
926                 swp_entry_t entry = pte_to_swp_entry(pte);
927
928                 if (likely(!non_swap_entry(entry))) {
929                         if (swap_duplicate(entry) < 0)
930                                 return entry.val;
931
932                         /* make sure dst_mm is on swapoff's mmlist. */
933                         if (unlikely(list_empty(&dst_mm->mmlist))) {
934                                 spin_lock(&mmlist_lock);
935                                 if (list_empty(&dst_mm->mmlist))
936                                         list_add(&dst_mm->mmlist,
937                                                         &src_mm->mmlist);
938                                 spin_unlock(&mmlist_lock);
939                         }
940                         rss[MM_SWAPENTS]++;
941                 } else if (is_migration_entry(entry)) {
942                         page = migration_entry_to_page(entry);
943
944                         rss[mm_counter(page)]++;
945
946                         if (is_write_migration_entry(entry) &&
947                                         is_cow_mapping(vm_flags)) {
948                                 /*
949                                  * COW mappings require pages in both
950                                  * parent and child to be set to read.
951                                  */
952                                 make_migration_entry_read(&entry);
953                                 pte = swp_entry_to_pte(entry);
954                                 if (pte_swp_soft_dirty(*src_pte))
955                                         pte = pte_swp_mksoft_dirty(pte);
956                                 set_pte_at(src_mm, addr, src_pte, pte);
957                         }
958                 }
959                 goto out_set_pte;
960         }
961
962         /*
963          * If it's a COW mapping, write protect it both
964          * in the parent and the child
965          */
966         if (is_cow_mapping(vm_flags)) {
967                 ptep_set_wrprotect(src_mm, addr, src_pte);
968                 pte = pte_wrprotect(pte);
969         }
970
971         /*
972          * If it's a shared mapping, mark it clean in
973          * the child
974          */
975         if (vm_flags & VM_SHARED)
976                 pte = pte_mkclean(pte);
977         pte = pte_mkold(pte);
978
979         page = vm_normal_page(vma, addr, pte);
980         if (page) {
981                 get_page(page);
982                 page_dup_rmap(page, false);
983                 rss[mm_counter(page)]++;
984         }
985
986 out_set_pte:
987         set_pte_at(dst_mm, addr, dst_pte, pte);
988         return 0;
989 }
990
991 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
992                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
993                    unsigned long addr, unsigned long end)
994 {
995         pte_t *orig_src_pte, *orig_dst_pte;
996         pte_t *src_pte, *dst_pte;
997         spinlock_t *src_ptl, *dst_ptl;
998         int progress = 0;
999         int rss[NR_MM_COUNTERS];
1000         swp_entry_t entry = (swp_entry_t){0};
1001
1002 again:
1003         init_rss_vec(rss);
1004
1005         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1006         if (!dst_pte)
1007                 return -ENOMEM;
1008         src_pte = pte_offset_map(src_pmd, addr);
1009         src_ptl = pte_lockptr(src_mm, src_pmd);
1010         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1011         orig_src_pte = src_pte;
1012         orig_dst_pte = dst_pte;
1013         arch_enter_lazy_mmu_mode();
1014
1015         do {
1016                 /*
1017                  * We are holding two locks at this point - either of them
1018                  * could generate latencies in another task on another CPU.
1019                  */
1020                 if (progress >= 32) {
1021                         progress = 0;
1022                         if (need_resched() ||
1023                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1024                                 break;
1025                 }
1026                 if (pte_none(*src_pte)) {
1027                         progress++;
1028                         continue;
1029                 }
1030                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1031                                                         vma, addr, rss);
1032                 if (entry.val)
1033                         break;
1034                 progress += 8;
1035         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1036
1037         arch_leave_lazy_mmu_mode();
1038         spin_unlock(src_ptl);
1039         pte_unmap(orig_src_pte);
1040         add_mm_rss_vec(dst_mm, rss);
1041         pte_unmap_unlock(orig_dst_pte, dst_ptl);
1042         cond_resched();
1043
1044         if (entry.val) {
1045                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1046                         return -ENOMEM;
1047                 progress = 0;
1048         }
1049         if (addr != end)
1050                 goto again;
1051         return 0;
1052 }
1053
1054 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1055                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1056                 unsigned long addr, unsigned long end)
1057 {
1058         pmd_t *src_pmd, *dst_pmd;
1059         unsigned long next;
1060
1061         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1062         if (!dst_pmd)
1063                 return -ENOMEM;
1064         src_pmd = pmd_offset(src_pud, addr);
1065         do {
1066                 next = pmd_addr_end(addr, end);
1067                 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
1068                         int err;
1069                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1070                         err = copy_huge_pmd(dst_mm, src_mm,
1071                                             dst_pmd, src_pmd, addr, vma);
1072                         if (err == -ENOMEM)
1073                                 return -ENOMEM;
1074                         if (!err)
1075                                 continue;
1076                         /* fall through */
1077                 }
1078                 if (pmd_none_or_clear_bad(src_pmd))
1079                         continue;
1080                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1081                                                 vma, addr, next))
1082                         return -ENOMEM;
1083         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1084         return 0;
1085 }
1086
1087 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1088                 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1089                 unsigned long addr, unsigned long end)
1090 {
1091         pud_t *src_pud, *dst_pud;
1092         unsigned long next;
1093
1094         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1095         if (!dst_pud)
1096                 return -ENOMEM;
1097         src_pud = pud_offset(src_p4d, addr);
1098         do {
1099                 next = pud_addr_end(addr, end);
1100                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1101                         int err;
1102
1103                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1104                         err = copy_huge_pud(dst_mm, src_mm,
1105                                             dst_pud, src_pud, addr, vma);
1106                         if (err == -ENOMEM)
1107                                 return -ENOMEM;
1108                         if (!err)
1109                                 continue;
1110                         /* fall through */
1111                 }
1112                 if (pud_none_or_clear_bad(src_pud))
1113                         continue;
1114                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1115                                                 vma, addr, next))
1116                         return -ENOMEM;
1117         } while (dst_pud++, src_pud++, addr = next, addr != end);
1118         return 0;
1119 }
1120
1121 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1122                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1123                 unsigned long addr, unsigned long end)
1124 {
1125         p4d_t *src_p4d, *dst_p4d;
1126         unsigned long next;
1127
1128         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1129         if (!dst_p4d)
1130                 return -ENOMEM;
1131         src_p4d = p4d_offset(src_pgd, addr);
1132         do {
1133                 next = p4d_addr_end(addr, end);
1134                 if (p4d_none_or_clear_bad(src_p4d))
1135                         continue;
1136                 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1137                                                 vma, addr, next))
1138                         return -ENOMEM;
1139         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1140         return 0;
1141 }
1142
1143 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1144                 struct vm_area_struct *vma)
1145 {
1146         pgd_t *src_pgd, *dst_pgd;
1147         unsigned long next;
1148         unsigned long addr = vma->vm_start;
1149         unsigned long end = vma->vm_end;
1150         unsigned long mmun_start;       /* For mmu_notifiers */
1151         unsigned long mmun_end;         /* For mmu_notifiers */
1152         bool is_cow;
1153         int ret;
1154
1155         /*
1156          * Don't copy ptes where a page fault will fill them correctly.
1157          * Fork becomes much lighter when there are big shared or private
1158          * readonly mappings. The tradeoff is that copy_page_range is more
1159          * efficient than faulting.
1160          */
1161         if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1162                         !vma->anon_vma)
1163                 return 0;
1164
1165         if (is_vm_hugetlb_page(vma))
1166                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1167
1168         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1169                 /*
1170                  * We do not free on error cases below as remove_vma
1171                  * gets called on error from higher level routine
1172                  */
1173                 ret = track_pfn_copy(vma);
1174                 if (ret)
1175                         return ret;
1176         }
1177
1178         /*
1179          * We need to invalidate the secondary MMU mappings only when
1180          * there could be a permission downgrade on the ptes of the
1181          * parent mm. And a permission downgrade will only happen if
1182          * is_cow_mapping() returns true.
1183          */
1184         is_cow = is_cow_mapping(vma->vm_flags);
1185         mmun_start = addr;
1186         mmun_end   = end;
1187         if (is_cow)
1188                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1189                                                     mmun_end);
1190
1191         ret = 0;
1192         dst_pgd = pgd_offset(dst_mm, addr);
1193         src_pgd = pgd_offset(src_mm, addr);
1194         do {
1195                 next = pgd_addr_end(addr, end);
1196                 if (pgd_none_or_clear_bad(src_pgd))
1197                         continue;
1198                 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1199                                             vma, addr, next))) {
1200                         ret = -ENOMEM;
1201                         break;
1202                 }
1203         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1204
1205         if (is_cow)
1206                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1207         return ret;
1208 }
1209
1210 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1211                                 struct vm_area_struct *vma, pmd_t *pmd,
1212                                 unsigned long addr, unsigned long end,
1213                                 struct zap_details *details)
1214 {
1215         struct mm_struct *mm = tlb->mm;
1216         int force_flush = 0;
1217         int rss[NR_MM_COUNTERS];
1218         spinlock_t *ptl;
1219         pte_t *start_pte;
1220         pte_t *pte;
1221         swp_entry_t entry;
1222
1223         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1224 again:
1225         init_rss_vec(rss);
1226         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1227         pte = start_pte;
1228         flush_tlb_batched_pending(mm);
1229         arch_enter_lazy_mmu_mode();
1230         do {
1231                 pte_t ptent = *pte;
1232                 if (pte_none(ptent))
1233                         continue;
1234
1235                 if (pte_present(ptent)) {
1236                         struct page *page;
1237
1238                         page = vm_normal_page(vma, addr, ptent);
1239                         if (unlikely(details) && page) {
1240                                 /*
1241                                  * unmap_shared_mapping_pages() wants to
1242                                  * invalidate cache without truncating:
1243                                  * unmap shared but keep private pages.
1244                                  */
1245                                 if (details->check_mapping &&
1246                                     details->check_mapping != page_rmapping(page))
1247                                         continue;
1248                         }
1249                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1250                                                         tlb->fullmm);
1251                         tlb_remove_tlb_entry(tlb, pte, addr);
1252                         if (unlikely(!page))
1253                                 continue;
1254
1255                         if (!PageAnon(page)) {
1256                                 if (pte_dirty(ptent)) {
1257                                         force_flush = 1;
1258                                         set_page_dirty(page);
1259                                 }
1260                                 if (pte_young(ptent) &&
1261                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1262                                         mark_page_accessed(page);
1263                         }
1264                         rss[mm_counter(page)]--;
1265                         page_remove_rmap(page, false);
1266                         if (unlikely(page_mapcount(page) < 0))
1267                                 print_bad_pte(vma, addr, ptent, page);
1268                         if (unlikely(__tlb_remove_page(tlb, page))) {
1269                                 force_flush = 1;
1270                                 addr += PAGE_SIZE;
1271                                 break;
1272                         }
1273                         continue;
1274                 }
1275                 /* If details->check_mapping, we leave swap entries. */
1276                 if (unlikely(details))
1277                         continue;
1278
1279                 entry = pte_to_swp_entry(ptent);
1280                 if (!non_swap_entry(entry))
1281                         rss[MM_SWAPENTS]--;
1282                 else if (is_migration_entry(entry)) {
1283                         struct page *page;
1284
1285                         page = migration_entry_to_page(entry);
1286                         rss[mm_counter(page)]--;
1287                 }
1288                 if (unlikely(!free_swap_and_cache(entry)))
1289                         print_bad_pte(vma, addr, ptent, NULL);
1290                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1291         } while (pte++, addr += PAGE_SIZE, addr != end);
1292
1293         add_mm_rss_vec(mm, rss);
1294         arch_leave_lazy_mmu_mode();
1295
1296         /* Do the actual TLB flush before dropping ptl */
1297         if (force_flush)
1298                 tlb_flush_mmu_tlbonly(tlb);
1299         pte_unmap_unlock(start_pte, ptl);
1300
1301         /*
1302          * If we forced a TLB flush (either due to running out of
1303          * batch buffers or because we needed to flush dirty TLB
1304          * entries before releasing the ptl), free the batched
1305          * memory too. Restart if we didn't do everything.
1306          */
1307         if (force_flush) {
1308                 force_flush = 0;
1309                 tlb_flush_mmu_free(tlb);
1310                 if (addr != end)
1311                         goto again;
1312         }
1313
1314         return addr;
1315 }
1316
1317 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1318                                 struct vm_area_struct *vma, pud_t *pud,
1319                                 unsigned long addr, unsigned long end,
1320                                 struct zap_details *details)
1321 {
1322         pmd_t *pmd;
1323         unsigned long next;
1324
1325         pmd = pmd_offset(pud, addr);
1326         do {
1327                 next = pmd_addr_end(addr, end);
1328                 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1329                         if (next - addr != HPAGE_PMD_SIZE) {
1330                                 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1331                                     !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1332                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1333                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1334                                 goto next;
1335                         /* fall through */
1336                 }
1337                 /*
1338                  * Here there can be other concurrent MADV_DONTNEED or
1339                  * trans huge page faults running, and if the pmd is
1340                  * none or trans huge it can change under us. This is
1341                  * because MADV_DONTNEED holds the mmap_sem in read
1342                  * mode.
1343                  */
1344                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1345                         goto next;
1346                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1347 next:
1348                 cond_resched();
1349         } while (pmd++, addr = next, addr != end);
1350
1351         return addr;
1352 }
1353
1354 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1355                                 struct vm_area_struct *vma, p4d_t *p4d,
1356                                 unsigned long addr, unsigned long end,
1357                                 struct zap_details *details)
1358 {
1359         pud_t *pud;
1360         unsigned long next;
1361
1362         pud = pud_offset(p4d, addr);
1363         do {
1364                 next = pud_addr_end(addr, end);
1365                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1366                         if (next - addr != HPAGE_PUD_SIZE) {
1367                                 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1368                                 split_huge_pud(vma, pud, addr);
1369                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1370                                 goto next;
1371                         /* fall through */
1372                 }
1373                 if (pud_none_or_clear_bad(pud))
1374                         continue;
1375                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1376 next:
1377                 cond_resched();
1378         } while (pud++, addr = next, addr != end);
1379
1380         return addr;
1381 }
1382
1383 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1384                                 struct vm_area_struct *vma, pgd_t *pgd,
1385                                 unsigned long addr, unsigned long end,
1386                                 struct zap_details *details)
1387 {
1388         p4d_t *p4d;
1389         unsigned long next;
1390
1391         p4d = p4d_offset(pgd, addr);
1392         do {
1393                 next = p4d_addr_end(addr, end);
1394                 if (p4d_none_or_clear_bad(p4d))
1395                         continue;
1396                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1397         } while (p4d++, addr = next, addr != end);
1398
1399         return addr;
1400 }
1401
1402 void unmap_page_range(struct mmu_gather *tlb,
1403                              struct vm_area_struct *vma,
1404                              unsigned long addr, unsigned long end,
1405                              struct zap_details *details)
1406 {
1407         pgd_t *pgd;
1408         unsigned long next;
1409
1410         BUG_ON(addr >= end);
1411         tlb_start_vma(tlb, vma);
1412         pgd = pgd_offset(vma->vm_mm, addr);
1413         do {
1414                 next = pgd_addr_end(addr, end);
1415                 if (pgd_none_or_clear_bad(pgd))
1416                         continue;
1417                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1418         } while (pgd++, addr = next, addr != end);
1419         tlb_end_vma(tlb, vma);
1420 }
1421
1422
1423 static void unmap_single_vma(struct mmu_gather *tlb,
1424                 struct vm_area_struct *vma, unsigned long start_addr,
1425                 unsigned long end_addr,
1426                 struct zap_details *details)
1427 {
1428         unsigned long start = max(vma->vm_start, start_addr);
1429         unsigned long end;
1430
1431         if (start >= vma->vm_end)
1432                 return;
1433         end = min(vma->vm_end, end_addr);
1434         if (end <= vma->vm_start)
1435                 return;
1436
1437         if (vma->vm_file)
1438                 uprobe_munmap(vma, start, end);
1439
1440         if (unlikely(vma->vm_flags & VM_PFNMAP))
1441                 untrack_pfn(vma, 0, 0);
1442
1443         if (start != end) {
1444                 if (unlikely(is_vm_hugetlb_page(vma))) {
1445                         /*
1446                          * It is undesirable to test vma->vm_file as it
1447                          * should be non-null for valid hugetlb area.
1448                          * However, vm_file will be NULL in the error
1449                          * cleanup path of mmap_region. When
1450                          * hugetlbfs ->mmap method fails,
1451                          * mmap_region() nullifies vma->vm_file
1452                          * before calling this function to clean up.
1453                          * Since no pte has actually been setup, it is
1454                          * safe to do nothing in this case.
1455                          */
1456                         if (vma->vm_file) {
1457                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1458                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1459                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1460                         }
1461                 } else
1462                         unmap_page_range(tlb, vma, start, end, details);
1463         }
1464 }
1465
1466 /**
1467  * unmap_vmas - unmap a range of memory covered by a list of vma's
1468  * @tlb: address of the caller's struct mmu_gather
1469  * @vma: the starting vma
1470  * @start_addr: virtual address at which to start unmapping
1471  * @end_addr: virtual address at which to end unmapping
1472  *
1473  * Unmap all pages in the vma list.
1474  *
1475  * Only addresses between `start' and `end' will be unmapped.
1476  *
1477  * The VMA list must be sorted in ascending virtual address order.
1478  *
1479  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1480  * range after unmap_vmas() returns.  So the only responsibility here is to
1481  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1482  * drops the lock and schedules.
1483  */
1484 void unmap_vmas(struct mmu_gather *tlb,
1485                 struct vm_area_struct *vma, unsigned long start_addr,
1486                 unsigned long end_addr)
1487 {
1488         struct mm_struct *mm = vma->vm_mm;
1489
1490         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1491         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1492                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1493         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1494 }
1495
1496 /**
1497  * zap_page_range - remove user pages in a given range
1498  * @vma: vm_area_struct holding the applicable pages
1499  * @start: starting address of pages to zap
1500  * @size: number of bytes to zap
1501  *
1502  * Caller must protect the VMA list
1503  */
1504 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1505                 unsigned long size)
1506 {
1507         struct mm_struct *mm = vma->vm_mm;
1508         struct mmu_gather tlb;
1509         unsigned long end = start + size;
1510
1511         lru_add_drain();
1512         tlb_gather_mmu(&tlb, mm, start, end);
1513         update_hiwater_rss(mm);
1514         mmu_notifier_invalidate_range_start(mm, start, end);
1515         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1516                 unmap_single_vma(&tlb, vma, start, end, NULL);
1517         mmu_notifier_invalidate_range_end(mm, start, end);
1518         tlb_finish_mmu(&tlb, start, end);
1519 }
1520
1521 /**
1522  * zap_page_range_single - remove user pages in a given range
1523  * @vma: vm_area_struct holding the applicable pages
1524  * @address: starting address of pages to zap
1525  * @size: number of bytes to zap
1526  * @details: details of shared cache invalidation
1527  *
1528  * The range must fit into one VMA.
1529  */
1530 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1531                 unsigned long size, struct zap_details *details)
1532 {
1533         struct mm_struct *mm = vma->vm_mm;
1534         struct mmu_gather tlb;
1535         unsigned long end = address + size;
1536
1537         lru_add_drain();
1538         tlb_gather_mmu(&tlb, mm, address, end);
1539         update_hiwater_rss(mm);
1540         mmu_notifier_invalidate_range_start(mm, address, end);
1541         unmap_single_vma(&tlb, vma, address, end, details);
1542         mmu_notifier_invalidate_range_end(mm, address, end);
1543         tlb_finish_mmu(&tlb, address, end);
1544 }
1545
1546 /**
1547  * zap_vma_ptes - remove ptes mapping the vma
1548  * @vma: vm_area_struct holding ptes to be zapped
1549  * @address: starting address of pages to zap
1550  * @size: number of bytes to zap
1551  *
1552  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1553  *
1554  * The entire address range must be fully contained within the vma.
1555  *
1556  * Returns 0 if successful.
1557  */
1558 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1559                 unsigned long size)
1560 {
1561         if (address < vma->vm_start || address + size > vma->vm_end ||
1562                         !(vma->vm_flags & VM_PFNMAP))
1563                 return -1;
1564         zap_page_range_single(vma, address, size, NULL);
1565         return 0;
1566 }
1567 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1568
1569 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1570                         spinlock_t **ptl)
1571 {
1572         pgd_t *pgd;
1573         p4d_t *p4d;
1574         pud_t *pud;
1575         pmd_t *pmd;
1576
1577         pgd = pgd_offset(mm, addr);
1578         p4d = p4d_alloc(mm, pgd, addr);
1579         if (!p4d)
1580                 return NULL;
1581         pud = pud_alloc(mm, p4d, addr);
1582         if (!pud)
1583                 return NULL;
1584         pmd = pmd_alloc(mm, pud, addr);
1585         if (!pmd)
1586                 return NULL;
1587
1588         VM_BUG_ON(pmd_trans_huge(*pmd));
1589         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1590 }
1591
1592 /*
1593  * This is the old fallback for page remapping.
1594  *
1595  * For historical reasons, it only allows reserved pages. Only
1596  * old drivers should use this, and they needed to mark their
1597  * pages reserved for the old functions anyway.
1598  */
1599 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1600                         struct page *page, pgprot_t prot)
1601 {
1602         struct mm_struct *mm = vma->vm_mm;
1603         int retval;
1604         pte_t *pte;
1605         spinlock_t *ptl;
1606
1607         retval = -EINVAL;
1608         if (PageAnon(page))
1609                 goto out;
1610         retval = -ENOMEM;
1611         flush_dcache_page(page);
1612         pte = get_locked_pte(mm, addr, &ptl);
1613         if (!pte)
1614                 goto out;
1615         retval = -EBUSY;
1616         if (!pte_none(*pte))
1617                 goto out_unlock;
1618
1619         /* Ok, finally just insert the thing.. */
1620         get_page(page);
1621         inc_mm_counter_fast(mm, mm_counter_file(page));
1622         page_add_file_rmap(page, false);
1623         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1624
1625         retval = 0;
1626         pte_unmap_unlock(pte, ptl);
1627         return retval;
1628 out_unlock:
1629         pte_unmap_unlock(pte, ptl);
1630 out:
1631         return retval;
1632 }
1633
1634 /**
1635  * vm_insert_page - insert single page into user vma
1636  * @vma: user vma to map to
1637  * @addr: target user address of this page
1638  * @page: source kernel page
1639  *
1640  * This allows drivers to insert individual pages they've allocated
1641  * into a user vma.
1642  *
1643  * The page has to be a nice clean _individual_ kernel allocation.
1644  * If you allocate a compound page, you need to have marked it as
1645  * such (__GFP_COMP), or manually just split the page up yourself
1646  * (see split_page()).
1647  *
1648  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1649  * took an arbitrary page protection parameter. This doesn't allow
1650  * that. Your vma protection will have to be set up correctly, which
1651  * means that if you want a shared writable mapping, you'd better
1652  * ask for a shared writable mapping!
1653  *
1654  * The page does not need to be reserved.
1655  *
1656  * Usually this function is called from f_op->mmap() handler
1657  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1658  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1659  * function from other places, for example from page-fault handler.
1660  */
1661 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1662                         struct page *page)
1663 {
1664         if (addr < vma->vm_start || addr >= vma->vm_end)
1665                 return -EFAULT;
1666         if (!page_count(page))
1667                 return -EINVAL;
1668         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1669                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1670                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1671                 vma->vm_flags |= VM_MIXEDMAP;
1672         }
1673         return insert_page(vma, addr, page, vma->vm_page_prot);
1674 }
1675 EXPORT_SYMBOL(vm_insert_page);
1676
1677 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1678                         pfn_t pfn, pgprot_t prot)
1679 {
1680         struct mm_struct *mm = vma->vm_mm;
1681         int retval;
1682         pte_t *pte, entry;
1683         spinlock_t *ptl;
1684
1685         retval = -ENOMEM;
1686         pte = get_locked_pte(mm, addr, &ptl);
1687         if (!pte)
1688                 goto out;
1689         retval = -EBUSY;
1690         if (!pte_none(*pte))
1691                 goto out_unlock;
1692
1693         /* Ok, finally just insert the thing.. */
1694         if (pfn_t_devmap(pfn))
1695                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1696         else
1697                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1698         set_pte_at(mm, addr, pte, entry);
1699         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1700
1701         retval = 0;
1702 out_unlock:
1703         pte_unmap_unlock(pte, ptl);
1704 out:
1705         return retval;
1706 }
1707
1708 /**
1709  * vm_insert_pfn - insert single pfn into user vma
1710  * @vma: user vma to map to
1711  * @addr: target user address of this page
1712  * @pfn: source kernel pfn
1713  *
1714  * Similar to vm_insert_page, this allows drivers to insert individual pages
1715  * they've allocated into a user vma. Same comments apply.
1716  *
1717  * This function should only be called from a vm_ops->fault handler, and
1718  * in that case the handler should return NULL.
1719  *
1720  * vma cannot be a COW mapping.
1721  *
1722  * As this is called only for pages that do not currently exist, we
1723  * do not need to flush old virtual caches or the TLB.
1724  */
1725 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1726                         unsigned long pfn)
1727 {
1728         return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1729 }
1730 EXPORT_SYMBOL(vm_insert_pfn);
1731
1732 /**
1733  * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1734  * @vma: user vma to map to
1735  * @addr: target user address of this page
1736  * @pfn: source kernel pfn
1737  * @pgprot: pgprot flags for the inserted page
1738  *
1739  * This is exactly like vm_insert_pfn, except that it allows drivers to
1740  * to override pgprot on a per-page basis.
1741  *
1742  * This only makes sense for IO mappings, and it makes no sense for
1743  * cow mappings.  In general, using multiple vmas is preferable;
1744  * vm_insert_pfn_prot should only be used if using multiple VMAs is
1745  * impractical.
1746  */
1747 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1748                         unsigned long pfn, pgprot_t pgprot)
1749 {
1750         int ret;
1751         /*
1752          * Technically, architectures with pte_special can avoid all these
1753          * restrictions (same for remap_pfn_range).  However we would like
1754          * consistency in testing and feature parity among all, so we should
1755          * try to keep these invariants in place for everybody.
1756          */
1757         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1758         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1759                                                 (VM_PFNMAP|VM_MIXEDMAP));
1760         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1761         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1762
1763         if (addr < vma->vm_start || addr >= vma->vm_end)
1764                 return -EFAULT;
1765
1766         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1767
1768         ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1769
1770         return ret;
1771 }
1772 EXPORT_SYMBOL(vm_insert_pfn_prot);
1773
1774 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1775                         pfn_t pfn)
1776 {
1777         pgprot_t pgprot = vma->vm_page_prot;
1778
1779         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1780
1781         if (addr < vma->vm_start || addr >= vma->vm_end)
1782                 return -EFAULT;
1783
1784         track_pfn_insert(vma, &pgprot, pfn);
1785
1786         /*
1787          * If we don't have pte special, then we have to use the pfn_valid()
1788          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1789          * refcount the page if pfn_valid is true (hence insert_page rather
1790          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1791          * without pte special, it would there be refcounted as a normal page.
1792          */
1793         if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1794                 struct page *page;
1795
1796                 /*
1797                  * At this point we are committed to insert_page()
1798                  * regardless of whether the caller specified flags that
1799                  * result in pfn_t_has_page() == false.
1800                  */
1801                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1802                 return insert_page(vma, addr, page, pgprot);
1803         }
1804         return insert_pfn(vma, addr, pfn, pgprot);
1805 }
1806 EXPORT_SYMBOL(vm_insert_mixed);
1807
1808 /*
1809  * maps a range of physical memory into the requested pages. the old
1810  * mappings are removed. any references to nonexistent pages results
1811  * in null mappings (currently treated as "copy-on-access")
1812  */
1813 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1814                         unsigned long addr, unsigned long end,
1815                         unsigned long pfn, pgprot_t prot)
1816 {
1817         pte_t *pte;
1818         spinlock_t *ptl;
1819
1820         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1821         if (!pte)
1822                 return -ENOMEM;
1823         arch_enter_lazy_mmu_mode();
1824         do {
1825                 BUG_ON(!pte_none(*pte));
1826                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1827                 pfn++;
1828         } while (pte++, addr += PAGE_SIZE, addr != end);
1829         arch_leave_lazy_mmu_mode();
1830         pte_unmap_unlock(pte - 1, ptl);
1831         return 0;
1832 }
1833
1834 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1835                         unsigned long addr, unsigned long end,
1836                         unsigned long pfn, pgprot_t prot)
1837 {
1838         pmd_t *pmd;
1839         unsigned long next;
1840
1841         pfn -= addr >> PAGE_SHIFT;
1842         pmd = pmd_alloc(mm, pud, addr);
1843         if (!pmd)
1844                 return -ENOMEM;
1845         VM_BUG_ON(pmd_trans_huge(*pmd));
1846         do {
1847                 next = pmd_addr_end(addr, end);
1848                 if (remap_pte_range(mm, pmd, addr, next,
1849                                 pfn + (addr >> PAGE_SHIFT), prot))
1850                         return -ENOMEM;
1851         } while (pmd++, addr = next, addr != end);
1852         return 0;
1853 }
1854
1855 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1856                         unsigned long addr, unsigned long end,
1857                         unsigned long pfn, pgprot_t prot)
1858 {
1859         pud_t *pud;
1860         unsigned long next;
1861
1862         pfn -= addr >> PAGE_SHIFT;
1863         pud = pud_alloc(mm, p4d, addr);
1864         if (!pud)
1865                 return -ENOMEM;
1866         do {
1867                 next = pud_addr_end(addr, end);
1868                 if (remap_pmd_range(mm, pud, addr, next,
1869                                 pfn + (addr >> PAGE_SHIFT), prot))
1870                         return -ENOMEM;
1871         } while (pud++, addr = next, addr != end);
1872         return 0;
1873 }
1874
1875 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1876                         unsigned long addr, unsigned long end,
1877                         unsigned long pfn, pgprot_t prot)
1878 {
1879         p4d_t *p4d;
1880         unsigned long next;
1881
1882         pfn -= addr >> PAGE_SHIFT;
1883         p4d = p4d_alloc(mm, pgd, addr);
1884         if (!p4d)
1885                 return -ENOMEM;
1886         do {
1887                 next = p4d_addr_end(addr, end);
1888                 if (remap_pud_range(mm, p4d, addr, next,
1889                                 pfn + (addr >> PAGE_SHIFT), prot))
1890                         return -ENOMEM;
1891         } while (p4d++, addr = next, addr != end);
1892         return 0;
1893 }
1894
1895 /**
1896  * remap_pfn_range - remap kernel memory to userspace
1897  * @vma: user vma to map to
1898  * @addr: target user address to start at
1899  * @pfn: physical address of kernel memory
1900  * @size: size of map area
1901  * @prot: page protection flags for this mapping
1902  *
1903  *  Note: this is only safe if the mm semaphore is held when called.
1904  */
1905 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1906                     unsigned long pfn, unsigned long size, pgprot_t prot)
1907 {
1908         pgd_t *pgd;
1909         unsigned long next;
1910         unsigned long end = addr + PAGE_ALIGN(size);
1911         struct mm_struct *mm = vma->vm_mm;
1912         unsigned long remap_pfn = pfn;
1913         int err;
1914
1915         /*
1916          * Physically remapped pages are special. Tell the
1917          * rest of the world about it:
1918          *   VM_IO tells people not to look at these pages
1919          *      (accesses can have side effects).
1920          *   VM_PFNMAP tells the core MM that the base pages are just
1921          *      raw PFN mappings, and do not have a "struct page" associated
1922          *      with them.
1923          *   VM_DONTEXPAND
1924          *      Disable vma merging and expanding with mremap().
1925          *   VM_DONTDUMP
1926          *      Omit vma from core dump, even when VM_IO turned off.
1927          *
1928          * There's a horrible special case to handle copy-on-write
1929          * behaviour that some programs depend on. We mark the "original"
1930          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1931          * See vm_normal_page() for details.
1932          */
1933         if (is_cow_mapping(vma->vm_flags)) {
1934                 if (addr != vma->vm_start || end != vma->vm_end)
1935                         return -EINVAL;
1936                 vma->vm_pgoff = pfn;
1937         }
1938
1939         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1940         if (err)
1941                 return -EINVAL;
1942
1943         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1944
1945         BUG_ON(addr >= end);
1946         pfn -= addr >> PAGE_SHIFT;
1947         pgd = pgd_offset(mm, addr);
1948         flush_cache_range(vma, addr, end);
1949         do {
1950                 next = pgd_addr_end(addr, end);
1951                 err = remap_p4d_range(mm, pgd, addr, next,
1952                                 pfn + (addr >> PAGE_SHIFT), prot);
1953                 if (err)
1954                         break;
1955         } while (pgd++, addr = next, addr != end);
1956
1957         if (err)
1958                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1959
1960         return err;
1961 }
1962 EXPORT_SYMBOL(remap_pfn_range);
1963
1964 /**
1965  * vm_iomap_memory - remap memory to userspace
1966  * @vma: user vma to map to
1967  * @start: start of area
1968  * @len: size of area
1969  *
1970  * This is a simplified io_remap_pfn_range() for common driver use. The
1971  * driver just needs to give us the physical memory range to be mapped,
1972  * we'll figure out the rest from the vma information.
1973  *
1974  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1975  * whatever write-combining details or similar.
1976  */
1977 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1978 {
1979         unsigned long vm_len, pfn, pages;
1980
1981         /* Check that the physical memory area passed in looks valid */
1982         if (start + len < start)
1983                 return -EINVAL;
1984         /*
1985          * You *really* shouldn't map things that aren't page-aligned,
1986          * but we've historically allowed it because IO memory might
1987          * just have smaller alignment.
1988          */
1989         len += start & ~PAGE_MASK;
1990         pfn = start >> PAGE_SHIFT;
1991         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1992         if (pfn + pages < pfn)
1993                 return -EINVAL;
1994
1995         /* We start the mapping 'vm_pgoff' pages into the area */
1996         if (vma->vm_pgoff > pages)
1997                 return -EINVAL;
1998         pfn += vma->vm_pgoff;
1999         pages -= vma->vm_pgoff;
2000
2001         /* Can we fit all of the mapping? */
2002         vm_len = vma->vm_end - vma->vm_start;
2003         if (vm_len >> PAGE_SHIFT > pages)
2004                 return -EINVAL;
2005
2006         /* Ok, let it rip */
2007         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2008 }
2009 EXPORT_SYMBOL(vm_iomap_memory);
2010
2011 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2012                                      unsigned long addr, unsigned long end,
2013                                      pte_fn_t fn, void *data)
2014 {
2015         pte_t *pte;
2016         int err;
2017         pgtable_t token;
2018         spinlock_t *uninitialized_var(ptl);
2019
2020         pte = (mm == &init_mm) ?
2021                 pte_alloc_kernel(pmd, addr) :
2022                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2023         if (!pte)
2024                 return -ENOMEM;
2025
2026         BUG_ON(pmd_huge(*pmd));
2027
2028         arch_enter_lazy_mmu_mode();
2029
2030         token = pmd_pgtable(*pmd);
2031
2032         do {
2033                 err = fn(pte++, token, addr, data);
2034                 if (err)
2035                         break;
2036         } while (addr += PAGE_SIZE, addr != end);
2037
2038         arch_leave_lazy_mmu_mode();
2039
2040         if (mm != &init_mm)
2041                 pte_unmap_unlock(pte-1, ptl);
2042         return err;
2043 }
2044
2045 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2046                                      unsigned long addr, unsigned long end,
2047                                      pte_fn_t fn, void *data)
2048 {
2049         pmd_t *pmd;
2050         unsigned long next;
2051         int err;
2052
2053         BUG_ON(pud_huge(*pud));
2054
2055         pmd = pmd_alloc(mm, pud, addr);
2056         if (!pmd)
2057                 return -ENOMEM;
2058         do {
2059                 next = pmd_addr_end(addr, end);
2060                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2061                 if (err)
2062                         break;
2063         } while (pmd++, addr = next, addr != end);
2064         return err;
2065 }
2066
2067 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2068                                      unsigned long addr, unsigned long end,
2069                                      pte_fn_t fn, void *data)
2070 {
2071         pud_t *pud;
2072         unsigned long next;
2073         int err;
2074
2075         pud = pud_alloc(mm, p4d, addr);
2076         if (!pud)
2077                 return -ENOMEM;
2078         do {
2079                 next = pud_addr_end(addr, end);
2080                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2081                 if (err)
2082                         break;
2083         } while (pud++, addr = next, addr != end);
2084         return err;
2085 }
2086
2087 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2088                                      unsigned long addr, unsigned long end,
2089                                      pte_fn_t fn, void *data)
2090 {
2091         p4d_t *p4d;
2092         unsigned long next;
2093         int err;
2094
2095         p4d = p4d_alloc(mm, pgd, addr);
2096         if (!p4d)
2097                 return -ENOMEM;
2098         do {
2099                 next = p4d_addr_end(addr, end);
2100                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2101                 if (err)
2102                         break;
2103         } while (p4d++, addr = next, addr != end);
2104         return err;
2105 }
2106
2107 /*
2108  * Scan a region of virtual memory, filling in page tables as necessary
2109  * and calling a provided function on each leaf page table.
2110  */
2111 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2112                         unsigned long size, pte_fn_t fn, void *data)
2113 {
2114         pgd_t *pgd;
2115         unsigned long next;
2116         unsigned long end = addr + size;
2117         int err;
2118
2119         if (WARN_ON(addr >= end))
2120                 return -EINVAL;
2121
2122         pgd = pgd_offset(mm, addr);
2123         do {
2124                 next = pgd_addr_end(addr, end);
2125                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2126                 if (err)
2127                         break;
2128         } while (pgd++, addr = next, addr != end);
2129
2130         return err;
2131 }
2132 EXPORT_SYMBOL_GPL(apply_to_page_range);
2133
2134 /*
2135  * handle_pte_fault chooses page fault handler according to an entry which was
2136  * read non-atomically.  Before making any commitment, on those architectures
2137  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2138  * parts, do_swap_page must check under lock before unmapping the pte and
2139  * proceeding (but do_wp_page is only called after already making such a check;
2140  * and do_anonymous_page can safely check later on).
2141  */
2142 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2143                                 pte_t *page_table, pte_t orig_pte)
2144 {
2145         int same = 1;
2146 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2147         if (sizeof(pte_t) > sizeof(unsigned long)) {
2148                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2149                 spin_lock(ptl);
2150                 same = pte_same(*page_table, orig_pte);
2151                 spin_unlock(ptl);
2152         }
2153 #endif
2154         pte_unmap(page_table);
2155         return same;
2156 }
2157
2158 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2159 {
2160         debug_dma_assert_idle(src);
2161
2162         /*
2163          * If the source page was a PFN mapping, we don't have
2164          * a "struct page" for it. We do a best-effort copy by
2165          * just copying from the original user address. If that
2166          * fails, we just zero-fill it. Live with it.
2167          */
2168         if (unlikely(!src)) {
2169                 void *kaddr = kmap_atomic(dst);
2170                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2171
2172                 /*
2173                  * This really shouldn't fail, because the page is there
2174                  * in the page tables. But it might just be unreadable,
2175                  * in which case we just give up and fill the result with
2176                  * zeroes.
2177                  */
2178                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2179                         clear_page(kaddr);
2180                 kunmap_atomic(kaddr);
2181                 flush_dcache_page(dst);
2182         } else
2183                 copy_user_highpage(dst, src, va, vma);
2184 }
2185
2186 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2187 {
2188         struct file *vm_file = vma->vm_file;
2189
2190         if (vm_file)
2191                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2192
2193         /*
2194          * Special mappings (e.g. VDSO) do not have any file so fake
2195          * a default GFP_KERNEL for them.
2196          */
2197         return GFP_KERNEL;
2198 }
2199
2200 /*
2201  * Notify the address space that the page is about to become writable so that
2202  * it can prohibit this or wait for the page to get into an appropriate state.
2203  *
2204  * We do this without the lock held, so that it can sleep if it needs to.
2205  */
2206 static int do_page_mkwrite(struct vm_fault *vmf)
2207 {
2208         int ret;
2209         struct page *page = vmf->page;
2210         unsigned int old_flags = vmf->flags;
2211
2212         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2213
2214         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2215         /* Restore original flags so that caller is not surprised */
2216         vmf->flags = old_flags;
2217         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2218                 return ret;
2219         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2220                 lock_page(page);
2221                 if (!page->mapping) {
2222                         unlock_page(page);
2223                         return 0; /* retry */
2224                 }
2225                 ret |= VM_FAULT_LOCKED;
2226         } else
2227                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2228         return ret;
2229 }
2230
2231 /*
2232  * Handle dirtying of a page in shared file mapping on a write fault.
2233  *
2234  * The function expects the page to be locked and unlocks it.
2235  */
2236 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2237                                     struct page *page)
2238 {
2239         struct address_space *mapping;
2240         bool dirtied;
2241         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2242
2243         dirtied = set_page_dirty(page);
2244         VM_BUG_ON_PAGE(PageAnon(page), page);
2245         /*
2246          * Take a local copy of the address_space - page.mapping may be zeroed
2247          * by truncate after unlock_page().   The address_space itself remains
2248          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2249          * release semantics to prevent the compiler from undoing this copying.
2250          */
2251         mapping = page_rmapping(page);
2252         unlock_page(page);
2253
2254         if ((dirtied || page_mkwrite) && mapping) {
2255                 /*
2256                  * Some device drivers do not set page.mapping
2257                  * but still dirty their pages
2258                  */
2259                 balance_dirty_pages_ratelimited(mapping);
2260         }
2261
2262         if (!page_mkwrite)
2263                 file_update_time(vma->vm_file);
2264 }
2265
2266 /*
2267  * Handle write page faults for pages that can be reused in the current vma
2268  *
2269  * This can happen either due to the mapping being with the VM_SHARED flag,
2270  * or due to us being the last reference standing to the page. In either
2271  * case, all we need to do here is to mark the page as writable and update
2272  * any related book-keeping.
2273  */
2274 static inline void wp_page_reuse(struct vm_fault *vmf)
2275         __releases(vmf->ptl)
2276 {
2277         struct vm_area_struct *vma = vmf->vma;
2278         struct page *page = vmf->page;
2279         pte_t entry;
2280         /*
2281          * Clear the pages cpupid information as the existing
2282          * information potentially belongs to a now completely
2283          * unrelated process.
2284          */
2285         if (page)
2286                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2287
2288         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2289         entry = pte_mkyoung(vmf->orig_pte);
2290         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2291         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2292                 update_mmu_cache(vma, vmf->address, vmf->pte);
2293         pte_unmap_unlock(vmf->pte, vmf->ptl);
2294 }
2295
2296 /*
2297  * Handle the case of a page which we actually need to copy to a new page.
2298  *
2299  * Called with mmap_sem locked and the old page referenced, but
2300  * without the ptl held.
2301  *
2302  * High level logic flow:
2303  *
2304  * - Allocate a page, copy the content of the old page to the new one.
2305  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2306  * - Take the PTL. If the pte changed, bail out and release the allocated page
2307  * - If the pte is still the way we remember it, update the page table and all
2308  *   relevant references. This includes dropping the reference the page-table
2309  *   held to the old page, as well as updating the rmap.
2310  * - In any case, unlock the PTL and drop the reference we took to the old page.
2311  */
2312 static int wp_page_copy(struct vm_fault *vmf)
2313 {
2314         struct vm_area_struct *vma = vmf->vma;
2315         struct mm_struct *mm = vma->vm_mm;
2316         struct page *old_page = vmf->page;
2317         struct page *new_page = NULL;
2318         pte_t entry;
2319         int page_copied = 0;
2320         const unsigned long mmun_start = vmf->address & PAGE_MASK;
2321         const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2322         struct mem_cgroup *memcg;
2323
2324         if (unlikely(anon_vma_prepare(vma)))
2325                 goto oom;
2326
2327         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2328                 new_page = alloc_zeroed_user_highpage_movable(vma,
2329                                                               vmf->address);
2330                 if (!new_page)
2331                         goto oom;
2332         } else {
2333                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2334                                 vmf->address);
2335                 if (!new_page)
2336                         goto oom;
2337                 cow_user_page(new_page, old_page, vmf->address, vma);
2338         }
2339
2340         if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2341                 goto oom_free_new;
2342
2343         __SetPageUptodate(new_page);
2344
2345         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2346
2347         /*
2348          * Re-check the pte - we dropped the lock
2349          */
2350         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2351         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2352                 if (old_page) {
2353                         if (!PageAnon(old_page)) {
2354                                 dec_mm_counter_fast(mm,
2355                                                 mm_counter_file(old_page));
2356                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2357                         }
2358                 } else {
2359                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2360                 }
2361                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2362                 entry = mk_pte(new_page, vma->vm_page_prot);
2363                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2364                 /*
2365                  * Clear the pte entry and flush it first, before updating the
2366                  * pte with the new entry. This will avoid a race condition
2367                  * seen in the presence of one thread doing SMC and another
2368                  * thread doing COW.
2369                  */
2370                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2371                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2372                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2373                 lru_cache_add_active_or_unevictable(new_page, vma);
2374                 /*
2375                  * We call the notify macro here because, when using secondary
2376                  * mmu page tables (such as kvm shadow page tables), we want the
2377                  * new page to be mapped directly into the secondary page table.
2378                  */
2379                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2380                 update_mmu_cache(vma, vmf->address, vmf->pte);
2381                 if (old_page) {
2382                         /*
2383                          * Only after switching the pte to the new page may
2384                          * we remove the mapcount here. Otherwise another
2385                          * process may come and find the rmap count decremented
2386                          * before the pte is switched to the new page, and
2387                          * "reuse" the old page writing into it while our pte
2388                          * here still points into it and can be read by other
2389                          * threads.
2390                          *
2391                          * The critical issue is to order this
2392                          * page_remove_rmap with the ptp_clear_flush above.
2393                          * Those stores are ordered by (if nothing else,)
2394                          * the barrier present in the atomic_add_negative
2395                          * in page_remove_rmap.
2396                          *
2397                          * Then the TLB flush in ptep_clear_flush ensures that
2398                          * no process can access the old page before the
2399                          * decremented mapcount is visible. And the old page
2400                          * cannot be reused until after the decremented
2401                          * mapcount is visible. So transitively, TLBs to
2402                          * old page will be flushed before it can be reused.
2403                          */
2404                         page_remove_rmap(old_page, false);
2405                 }
2406
2407                 /* Free the old page.. */
2408                 new_page = old_page;
2409                 page_copied = 1;
2410         } else {
2411                 mem_cgroup_cancel_charge(new_page, memcg, false);
2412         }
2413
2414         if (new_page)
2415                 put_page(new_page);
2416
2417         pte_unmap_unlock(vmf->pte, vmf->ptl);
2418         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2419         if (old_page) {
2420                 /*
2421                  * Don't let another task, with possibly unlocked vma,
2422                  * keep the mlocked page.
2423                  */
2424                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2425                         lock_page(old_page);    /* LRU manipulation */
2426                         if (PageMlocked(old_page))
2427                                 munlock_vma_page(old_page);
2428                         unlock_page(old_page);
2429                 }
2430                 put_page(old_page);
2431         }
2432         return page_copied ? VM_FAULT_WRITE : 0;
2433 oom_free_new:
2434         put_page(new_page);
2435 oom:
2436         if (old_page)
2437                 put_page(old_page);
2438         return VM_FAULT_OOM;
2439 }
2440
2441 /**
2442  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2443  *                        writeable once the page is prepared
2444  *
2445  * @vmf: structure describing the fault
2446  *
2447  * This function handles all that is needed to finish a write page fault in a
2448  * shared mapping due to PTE being read-only once the mapped page is prepared.
2449  * It handles locking of PTE and modifying it. The function returns
2450  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2451  * lock.
2452  *
2453  * The function expects the page to be locked or other protection against
2454  * concurrent faults / writeback (such as DAX radix tree locks).
2455  */
2456 int finish_mkwrite_fault(struct vm_fault *vmf)
2457 {
2458         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2459         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2460                                        &vmf->ptl);
2461         /*
2462          * We might have raced with another page fault while we released the
2463          * pte_offset_map_lock.
2464          */
2465         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2466                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2467                 return VM_FAULT_NOPAGE;
2468         }
2469         wp_page_reuse(vmf);
2470         return 0;
2471 }
2472
2473 /*
2474  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2475  * mapping
2476  */
2477 static int wp_pfn_shared(struct vm_fault *vmf)
2478 {
2479         struct vm_area_struct *vma = vmf->vma;
2480
2481         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2482                 int ret;
2483
2484                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2485                 vmf->flags |= FAULT_FLAG_MKWRITE;
2486                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2487                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2488                         return ret;
2489                 return finish_mkwrite_fault(vmf);
2490         }
2491         wp_page_reuse(vmf);
2492         return VM_FAULT_WRITE;
2493 }
2494
2495 static int wp_page_shared(struct vm_fault *vmf)
2496         __releases(vmf->ptl)
2497 {
2498         struct vm_area_struct *vma = vmf->vma;
2499
2500         get_page(vmf->page);
2501
2502         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2503                 int tmp;
2504
2505                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2506                 tmp = do_page_mkwrite(vmf);
2507                 if (unlikely(!tmp || (tmp &
2508                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2509                         put_page(vmf->page);
2510                         return tmp;
2511                 }
2512                 tmp = finish_mkwrite_fault(vmf);
2513                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2514                         unlock_page(vmf->page);
2515                         put_page(vmf->page);
2516                         return tmp;
2517                 }
2518         } else {
2519                 wp_page_reuse(vmf);
2520                 lock_page(vmf->page);
2521         }
2522         fault_dirty_shared_page(vma, vmf->page);
2523         put_page(vmf->page);
2524
2525         return VM_FAULT_WRITE;
2526 }
2527
2528 /*
2529  * This routine handles present pages, when users try to write
2530  * to a shared page. It is done by copying the page to a new address
2531  * and decrementing the shared-page counter for the old page.
2532  *
2533  * Note that this routine assumes that the protection checks have been
2534  * done by the caller (the low-level page fault routine in most cases).
2535  * Thus we can safely just mark it writable once we've done any necessary
2536  * COW.
2537  *
2538  * We also mark the page dirty at this point even though the page will
2539  * change only once the write actually happens. This avoids a few races,
2540  * and potentially makes it more efficient.
2541  *
2542  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2543  * but allow concurrent faults), with pte both mapped and locked.
2544  * We return with mmap_sem still held, but pte unmapped and unlocked.
2545  */
2546 static int do_wp_page(struct vm_fault *vmf)
2547         __releases(vmf->ptl)
2548 {
2549         struct vm_area_struct *vma = vmf->vma;
2550
2551         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2552         if (!vmf->page) {
2553                 /*
2554                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2555                  * VM_PFNMAP VMA.
2556                  *
2557                  * We should not cow pages in a shared writeable mapping.
2558                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2559                  */
2560                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2561                                      (VM_WRITE|VM_SHARED))
2562                         return wp_pfn_shared(vmf);
2563
2564                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2565                 return wp_page_copy(vmf);
2566         }
2567
2568         /*
2569          * Take out anonymous pages first, anonymous shared vmas are
2570          * not dirty accountable.
2571          */
2572         if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2573                 int total_mapcount;
2574                 if (!trylock_page(vmf->page)) {
2575                         get_page(vmf->page);
2576                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2577                         lock_page(vmf->page);
2578                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2579                                         vmf->address, &vmf->ptl);
2580                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2581                                 unlock_page(vmf->page);
2582                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2583                                 put_page(vmf->page);
2584                                 return 0;
2585                         }
2586                         put_page(vmf->page);
2587                 }
2588                 if (reuse_swap_page(vmf->page, &total_mapcount)) {
2589                         if (total_mapcount == 1) {
2590                                 /*
2591                                  * The page is all ours. Move it to
2592                                  * our anon_vma so the rmap code will
2593                                  * not search our parent or siblings.
2594                                  * Protected against the rmap code by
2595                                  * the page lock.
2596                                  */
2597                                 page_move_anon_rmap(vmf->page, vma);
2598                         }
2599                         unlock_page(vmf->page);
2600                         wp_page_reuse(vmf);
2601                         return VM_FAULT_WRITE;
2602                 }
2603                 unlock_page(vmf->page);
2604         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2605                                         (VM_WRITE|VM_SHARED))) {
2606                 return wp_page_shared(vmf);
2607         }
2608
2609         /*
2610          * Ok, we need to copy. Oh, well..
2611          */
2612         get_page(vmf->page);
2613
2614         pte_unmap_unlock(vmf->pte, vmf->ptl);
2615         return wp_page_copy(vmf);
2616 }
2617
2618 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2619                 unsigned long start_addr, unsigned long end_addr,
2620                 struct zap_details *details)
2621 {
2622         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2623 }
2624
2625 static inline void unmap_mapping_range_tree(struct rb_root *root,
2626                                             struct zap_details *details)
2627 {
2628         struct vm_area_struct *vma;
2629         pgoff_t vba, vea, zba, zea;
2630
2631         vma_interval_tree_foreach(vma, root,
2632                         details->first_index, details->last_index) {
2633
2634                 vba = vma->vm_pgoff;
2635                 vea = vba + vma_pages(vma) - 1;
2636                 zba = details->first_index;
2637                 if (zba < vba)
2638                         zba = vba;
2639                 zea = details->last_index;
2640                 if (zea > vea)
2641                         zea = vea;
2642
2643                 unmap_mapping_range_vma(vma,
2644                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2645                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2646                                 details);
2647         }
2648 }
2649
2650 /**
2651  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2652  * address_space corresponding to the specified page range in the underlying
2653  * file.
2654  *
2655  * @mapping: the address space containing mmaps to be unmapped.
2656  * @holebegin: byte in first page to unmap, relative to the start of
2657  * the underlying file.  This will be rounded down to a PAGE_SIZE
2658  * boundary.  Note that this is different from truncate_pagecache(), which
2659  * must keep the partial page.  In contrast, we must get rid of
2660  * partial pages.
2661  * @holelen: size of prospective hole in bytes.  This will be rounded
2662  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2663  * end of the file.
2664  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2665  * but 0 when invalidating pagecache, don't throw away private data.
2666  */
2667 void unmap_mapping_range(struct address_space *mapping,
2668                 loff_t const holebegin, loff_t const holelen, int even_cows)
2669 {
2670         struct zap_details details = { };
2671         pgoff_t hba = holebegin >> PAGE_SHIFT;
2672         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2673
2674         /* Check for overflow. */
2675         if (sizeof(holelen) > sizeof(hlen)) {
2676                 long long holeend =
2677                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2678                 if (holeend & ~(long long)ULONG_MAX)
2679                         hlen = ULONG_MAX - hba + 1;
2680         }
2681
2682         details.check_mapping = even_cows ? NULL : mapping;
2683         details.first_index = hba;
2684         details.last_index = hba + hlen - 1;
2685         if (details.last_index < details.first_index)
2686                 details.last_index = ULONG_MAX;
2687
2688         i_mmap_lock_write(mapping);
2689         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2690                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2691         i_mmap_unlock_write(mapping);
2692 }
2693 EXPORT_SYMBOL(unmap_mapping_range);
2694
2695 /*
2696  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2697  * but allow concurrent faults), and pte mapped but not yet locked.
2698  * We return with pte unmapped and unlocked.
2699  *
2700  * We return with the mmap_sem locked or unlocked in the same cases
2701  * as does filemap_fault().
2702  */
2703 int do_swap_page(struct vm_fault *vmf)
2704 {
2705         struct vm_area_struct *vma = vmf->vma;
2706         struct page *page, *swapcache;
2707         struct mem_cgroup *memcg;
2708         swp_entry_t entry;
2709         pte_t pte;
2710         int locked;
2711         int exclusive = 0;
2712         int ret = 0;
2713
2714         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2715                 goto out;
2716
2717         entry = pte_to_swp_entry(vmf->orig_pte);
2718         if (unlikely(non_swap_entry(entry))) {
2719                 if (is_migration_entry(entry)) {
2720                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2721                                              vmf->address);
2722                 } else if (is_hwpoison_entry(entry)) {
2723                         ret = VM_FAULT_HWPOISON;
2724                 } else {
2725                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2726                         ret = VM_FAULT_SIGBUS;
2727                 }
2728                 goto out;
2729         }
2730         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2731         page = lookup_swap_cache(entry);
2732         if (!page) {
2733                 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, vma,
2734                                         vmf->address);
2735                 if (!page) {
2736                         /*
2737                          * Back out if somebody else faulted in this pte
2738                          * while we released the pte lock.
2739                          */
2740                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2741                                         vmf->address, &vmf->ptl);
2742                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2743                                 ret = VM_FAULT_OOM;
2744                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2745                         goto unlock;
2746                 }
2747
2748                 /* Had to read the page from swap area: Major fault */
2749                 ret = VM_FAULT_MAJOR;
2750                 count_vm_event(PGMAJFAULT);
2751                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2752         } else if (PageHWPoison(page)) {
2753                 /*
2754                  * hwpoisoned dirty swapcache pages are kept for killing
2755                  * owner processes (which may be unknown at hwpoison time)
2756                  */
2757                 ret = VM_FAULT_HWPOISON;
2758                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2759                 swapcache = page;
2760                 goto out_release;
2761         }
2762
2763         swapcache = page;
2764         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2765
2766         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2767         if (!locked) {
2768                 ret |= VM_FAULT_RETRY;
2769                 goto out_release;
2770         }
2771
2772         /*
2773          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2774          * release the swapcache from under us.  The page pin, and pte_same
2775          * test below, are not enough to exclude that.  Even if it is still
2776          * swapcache, we need to check that the page's swap has not changed.
2777          */
2778         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2779                 goto out_page;
2780
2781         page = ksm_might_need_to_copy(page, vma, vmf->address);
2782         if (unlikely(!page)) {
2783                 ret = VM_FAULT_OOM;
2784                 page = swapcache;
2785                 goto out_page;
2786         }
2787
2788         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2789                                 &memcg, false)) {
2790                 ret = VM_FAULT_OOM;
2791                 goto out_page;
2792         }
2793
2794         /*
2795          * Back out if somebody else already faulted in this pte.
2796          */
2797         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2798                         &vmf->ptl);
2799         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2800                 goto out_nomap;
2801
2802         if (unlikely(!PageUptodate(page))) {
2803                 ret = VM_FAULT_SIGBUS;
2804                 goto out_nomap;
2805         }
2806
2807         /*
2808          * The page isn't present yet, go ahead with the fault.
2809          *
2810          * Be careful about the sequence of operations here.
2811          * To get its accounting right, reuse_swap_page() must be called
2812          * while the page is counted on swap but not yet in mapcount i.e.
2813          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2814          * must be called after the swap_free(), or it will never succeed.
2815          */
2816
2817         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2818         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2819         pte = mk_pte(page, vma->vm_page_prot);
2820         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2821                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2822                 vmf->flags &= ~FAULT_FLAG_WRITE;
2823                 ret |= VM_FAULT_WRITE;
2824                 exclusive = RMAP_EXCLUSIVE;
2825         }
2826         flush_icache_page(vma, page);
2827         if (pte_swp_soft_dirty(vmf->orig_pte))
2828                 pte = pte_mksoft_dirty(pte);
2829         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2830         vmf->orig_pte = pte;
2831         if (page == swapcache) {
2832                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2833                 mem_cgroup_commit_charge(page, memcg, true, false);
2834                 activate_page(page);
2835         } else { /* ksm created a completely new copy */
2836                 page_add_new_anon_rmap(page, vma, vmf->address, false);
2837                 mem_cgroup_commit_charge(page, memcg, false, false);
2838                 lru_cache_add_active_or_unevictable(page, vma);
2839         }
2840
2841         swap_free(entry);
2842         if (mem_cgroup_swap_full(page) ||
2843             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2844                 try_to_free_swap(page);
2845         unlock_page(page);
2846         if (page != swapcache) {
2847                 /*
2848                  * Hold the lock to avoid the swap entry to be reused
2849                  * until we take the PT lock for the pte_same() check
2850                  * (to avoid false positives from pte_same). For
2851                  * further safety release the lock after the swap_free
2852                  * so that the swap count won't change under a
2853                  * parallel locked swapcache.
2854                  */
2855                 unlock_page(swapcache);
2856                 put_page(swapcache);
2857         }
2858
2859         if (vmf->flags & FAULT_FLAG_WRITE) {
2860                 ret |= do_wp_page(vmf);
2861                 if (ret & VM_FAULT_ERROR)
2862                         ret &= VM_FAULT_ERROR;
2863                 goto out;
2864         }
2865
2866         /* No need to invalidate - it was non-present before */
2867         update_mmu_cache(vma, vmf->address, vmf->pte);
2868 unlock:
2869         pte_unmap_unlock(vmf->pte, vmf->ptl);
2870 out:
2871         return ret;
2872 out_nomap:
2873         mem_cgroup_cancel_charge(page, memcg, false);
2874         pte_unmap_unlock(vmf->pte, vmf->ptl);
2875 out_page:
2876         unlock_page(page);
2877 out_release:
2878         put_page(page);
2879         if (page != swapcache) {
2880                 unlock_page(swapcache);
2881                 put_page(swapcache);
2882         }
2883         return ret;
2884 }
2885
2886 /*
2887  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2888  * but allow concurrent faults), and pte mapped but not yet locked.
2889  * We return with mmap_sem still held, but pte unmapped and unlocked.
2890  */
2891 static int do_anonymous_page(struct vm_fault *vmf)
2892 {
2893         struct vm_area_struct *vma = vmf->vma;
2894         struct mem_cgroup *memcg;
2895         struct page *page;
2896         pte_t entry;
2897
2898         /* File mapping without ->vm_ops ? */
2899         if (vma->vm_flags & VM_SHARED)
2900                 return VM_FAULT_SIGBUS;
2901
2902         /*
2903          * Use pte_alloc() instead of pte_alloc_map().  We can't run
2904          * pte_offset_map() on pmds where a huge pmd might be created
2905          * from a different thread.
2906          *
2907          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2908          * parallel threads are excluded by other means.
2909          *
2910          * Here we only have down_read(mmap_sem).
2911          */
2912         if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2913                 return VM_FAULT_OOM;
2914
2915         /* See the comment in pte_alloc_one_map() */
2916         if (unlikely(pmd_trans_unstable(vmf->pmd)))
2917                 return 0;
2918
2919         /* Use the zero-page for reads */
2920         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2921                         !mm_forbids_zeropage(vma->vm_mm)) {
2922                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2923                                                 vma->vm_page_prot));
2924                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2925                                 vmf->address, &vmf->ptl);
2926                 if (!pte_none(*vmf->pte))
2927                         goto unlock;
2928                 /* Deliver the page fault to userland, check inside PT lock */
2929                 if (userfaultfd_missing(vma)) {
2930                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2931                         return handle_userfault(vmf, VM_UFFD_MISSING);
2932                 }
2933                 goto setpte;
2934         }
2935
2936         /* Allocate our own private page. */
2937         if (unlikely(anon_vma_prepare(vma)))
2938                 goto oom;
2939         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2940         if (!page)
2941                 goto oom;
2942
2943         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2944                 goto oom_free_page;
2945
2946         /*
2947          * The memory barrier inside __SetPageUptodate makes sure that
2948          * preceeding stores to the page contents become visible before
2949          * the set_pte_at() write.
2950          */
2951         __SetPageUptodate(page);
2952
2953         entry = mk_pte(page, vma->vm_page_prot);
2954         if (vma->vm_flags & VM_WRITE)
2955                 entry = pte_mkwrite(pte_mkdirty(entry));
2956
2957         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2958                         &vmf->ptl);
2959         if (!pte_none(*vmf->pte))
2960                 goto release;
2961
2962         /* Deliver the page fault to userland, check inside PT lock */
2963         if (userfaultfd_missing(vma)) {
2964                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2965                 mem_cgroup_cancel_charge(page, memcg, false);
2966                 put_page(page);
2967                 return handle_userfault(vmf, VM_UFFD_MISSING);
2968         }
2969
2970         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2971         page_add_new_anon_rmap(page, vma, vmf->address, false);
2972         mem_cgroup_commit_charge(page, memcg, false, false);
2973         lru_cache_add_active_or_unevictable(page, vma);
2974 setpte:
2975         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2976
2977         /* No need to invalidate - it was non-present before */
2978         update_mmu_cache(vma, vmf->address, vmf->pte);
2979 unlock:
2980         pte_unmap_unlock(vmf->pte, vmf->ptl);
2981         return 0;
2982 release:
2983         mem_cgroup_cancel_charge(page, memcg, false);
2984         put_page(page);
2985         goto unlock;
2986 oom_free_page:
2987         put_page(page);
2988 oom:
2989         return VM_FAULT_OOM;
2990 }
2991
2992 /*
2993  * The mmap_sem must have been held on entry, and may have been
2994  * released depending on flags and vma->vm_ops->fault() return value.
2995  * See filemap_fault() and __lock_page_retry().
2996  */
2997 static int __do_fault(struct vm_fault *vmf)
2998 {
2999         struct vm_area_struct *vma = vmf->vma;
3000         int ret;
3001
3002         ret = vma->vm_ops->fault(vmf);
3003         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3004                             VM_FAULT_DONE_COW)))
3005                 return ret;
3006
3007         if (unlikely(PageHWPoison(vmf->page))) {
3008                 if (ret & VM_FAULT_LOCKED)
3009                         unlock_page(vmf->page);
3010                 put_page(vmf->page);
3011                 vmf->page = NULL;
3012                 return VM_FAULT_HWPOISON;
3013         }
3014
3015         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3016                 lock_page(vmf->page);
3017         else
3018                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3019
3020         return ret;
3021 }
3022
3023 /*
3024  * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3025  * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3026  * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3027  * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3028  */
3029 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3030 {
3031         return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3032 }
3033
3034 static int pte_alloc_one_map(struct vm_fault *vmf)
3035 {
3036         struct vm_area_struct *vma = vmf->vma;
3037
3038         if (!pmd_none(*vmf->pmd))
3039                 goto map_pte;
3040         if (vmf->prealloc_pte) {
3041                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3042                 if (unlikely(!pmd_none(*vmf->pmd))) {
3043                         spin_unlock(vmf->ptl);
3044                         goto map_pte;
3045                 }
3046
3047                 atomic_long_inc(&vma->vm_mm->nr_ptes);
3048                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3049                 spin_unlock(vmf->ptl);
3050                 vmf->prealloc_pte = NULL;
3051         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3052                 return VM_FAULT_OOM;
3053         }
3054 map_pte:
3055         /*
3056          * If a huge pmd materialized under us just retry later.  Use
3057          * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3058          * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3059          * under us and then back to pmd_none, as a result of MADV_DONTNEED
3060          * running immediately after a huge pmd fault in a different thread of
3061          * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3062          * All we have to ensure is that it is a regular pmd that we can walk
3063          * with pte_offset_map() and we can do that through an atomic read in
3064          * C, which is what pmd_trans_unstable() provides.
3065          */
3066         if (pmd_devmap_trans_unstable(vmf->pmd))
3067                 return VM_FAULT_NOPAGE;
3068
3069         /*
3070          * At this point we know that our vmf->pmd points to a page of ptes
3071          * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3072          * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3073          * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3074          * be valid and we will re-check to make sure the vmf->pte isn't
3075          * pte_none() under vmf->ptl protection when we return to
3076          * alloc_set_pte().
3077          */
3078         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3079                         &vmf->ptl);
3080         return 0;
3081 }
3082
3083 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3084
3085 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3086 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3087                 unsigned long haddr)
3088 {
3089         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3090                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3091                 return false;
3092         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3093                 return false;
3094         return true;
3095 }
3096
3097 static void deposit_prealloc_pte(struct vm_fault *vmf)
3098 {
3099         struct vm_area_struct *vma = vmf->vma;
3100
3101         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3102         /*
3103          * We are going to consume the prealloc table,
3104          * count that as nr_ptes.
3105          */
3106         atomic_long_inc(&vma->vm_mm->nr_ptes);
3107         vmf->prealloc_pte = NULL;
3108 }
3109
3110 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3111 {
3112         struct vm_area_struct *vma = vmf->vma;
3113         bool write = vmf->flags & FAULT_FLAG_WRITE;
3114         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3115         pmd_t entry;
3116         int i, ret;
3117
3118         if (!transhuge_vma_suitable(vma, haddr))
3119                 return VM_FAULT_FALLBACK;
3120
3121         ret = VM_FAULT_FALLBACK;
3122         page = compound_head(page);
3123
3124         /*
3125          * Archs like ppc64 need additonal space to store information
3126          * related to pte entry. Use the preallocated table for that.
3127          */
3128         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3129                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3130                 if (!vmf->prealloc_pte)
3131                         return VM_FAULT_OOM;
3132                 smp_wmb(); /* See comment in __pte_alloc() */
3133         }
3134
3135         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3136         if (unlikely(!pmd_none(*vmf->pmd)))
3137                 goto out;
3138
3139         for (i = 0; i < HPAGE_PMD_NR; i++)
3140                 flush_icache_page(vma, page + i);
3141
3142         entry = mk_huge_pmd(page, vma->vm_page_prot);
3143         if (write)
3144                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3145
3146         add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3147         page_add_file_rmap(page, true);
3148         /*
3149          * deposit and withdraw with pmd lock held
3150          */
3151         if (arch_needs_pgtable_deposit())
3152                 deposit_prealloc_pte(vmf);
3153
3154         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3155
3156         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3157
3158         /* fault is handled */
3159         ret = 0;
3160         count_vm_event(THP_FILE_MAPPED);
3161 out:
3162         spin_unlock(vmf->ptl);
3163         return ret;
3164 }
3165 #else
3166 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3167 {
3168         BUILD_BUG();
3169         return 0;
3170 }
3171 #endif
3172
3173 /**
3174  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3175  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3176  *
3177  * @vmf: fault environment
3178  * @memcg: memcg to charge page (only for private mappings)
3179  * @page: page to map
3180  *
3181  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3182  * return.
3183  *
3184  * Target users are page handler itself and implementations of
3185  * vm_ops->map_pages.
3186  */
3187 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3188                 struct page *page)
3189 {
3190         struct vm_area_struct *vma = vmf->vma;
3191         bool write = vmf->flags & FAULT_FLAG_WRITE;
3192         pte_t entry;
3193         int ret;
3194
3195         if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3196                         IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3197                 /* THP on COW? */
3198                 VM_BUG_ON_PAGE(memcg, page);
3199
3200                 ret = do_set_pmd(vmf, page);
3201                 if (ret != VM_FAULT_FALLBACK)
3202                         return ret;
3203         }
3204
3205         if (!vmf->pte) {
3206                 ret = pte_alloc_one_map(vmf);
3207                 if (ret)
3208                         return ret;
3209         }
3210
3211         /* Re-check under ptl */
3212         if (unlikely(!pte_none(*vmf->pte)))
3213                 return VM_FAULT_NOPAGE;
3214
3215         flush_icache_page(vma, page);
3216         entry = mk_pte(page, vma->vm_page_prot);
3217         if (write)
3218                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3219         /* copy-on-write page */
3220         if (write && !(vma->vm_flags & VM_SHARED)) {
3221                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3222                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3223                 mem_cgroup_commit_charge(page, memcg, false, false);
3224                 lru_cache_add_active_or_unevictable(page, vma);
3225         } else {
3226                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3227                 page_add_file_rmap(page, false);
3228         }
3229         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3230
3231         /* no need to invalidate: a not-present page won't be cached */
3232         update_mmu_cache(vma, vmf->address, vmf->pte);
3233
3234         return 0;
3235 }
3236
3237
3238 /**
3239  * finish_fault - finish page fault once we have prepared the page to fault
3240  *
3241  * @vmf: structure describing the fault
3242  *
3243  * This function handles all that is needed to finish a page fault once the
3244  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3245  * given page, adds reverse page mapping, handles memcg charges and LRU
3246  * addition. The function returns 0 on success, VM_FAULT_ code in case of
3247  * error.
3248  *
3249  * The function expects the page to be locked and on success it consumes a
3250  * reference of a page being mapped (for the PTE which maps it).
3251  */
3252 int finish_fault(struct vm_fault *vmf)
3253 {
3254         struct page *page;
3255         int ret;
3256
3257         /* Did we COW the page? */
3258         if ((vmf->flags & FAULT_FLAG_WRITE) &&
3259             !(vmf->vma->vm_flags & VM_SHARED))
3260                 page = vmf->cow_page;
3261         else
3262                 page = vmf->page;
3263         ret = alloc_set_pte(vmf, vmf->memcg, page);
3264         if (vmf->pte)
3265                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3266         return ret;
3267 }
3268
3269 static unsigned long fault_around_bytes __read_mostly =
3270         rounddown_pow_of_two(65536);
3271
3272 #ifdef CONFIG_DEBUG_FS
3273 static int fault_around_bytes_get(void *data, u64 *val)
3274 {
3275         *val = fault_around_bytes;
3276         return 0;
3277 }
3278
3279 /*
3280  * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3281  * rounded down to nearest page order. It's what do_fault_around() expects to
3282  * see.
3283  */
3284 static int fault_around_bytes_set(void *data, u64 val)
3285 {
3286         if (val / PAGE_SIZE > PTRS_PER_PTE)
3287                 return -EINVAL;
3288         if (val > PAGE_SIZE)
3289                 fault_around_bytes = rounddown_pow_of_two(val);
3290         else
3291                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3292         return 0;
3293 }
3294 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3295                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3296
3297 static int __init fault_around_debugfs(void)
3298 {
3299         void *ret;
3300
3301         ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3302                         &fault_around_bytes_fops);
3303         if (!ret)
3304                 pr_warn("Failed to create fault_around_bytes in debugfs");
3305         return 0;
3306 }
3307 late_initcall(fault_around_debugfs);
3308 #endif
3309
3310 /*
3311  * do_fault_around() tries to map few pages around the fault address. The hope
3312  * is that the pages will be needed soon and this will lower the number of
3313  * faults to handle.
3314  *
3315  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3316  * not ready to be mapped: not up-to-date, locked, etc.
3317  *
3318  * This function is called with the page table lock taken. In the split ptlock
3319  * case the page table lock only protects only those entries which belong to
3320  * the page table corresponding to the fault address.
3321  *
3322  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3323  * only once.
3324  *
3325  * fault_around_pages() defines how many pages we'll try to map.
3326  * do_fault_around() expects it to return a power of two less than or equal to
3327  * PTRS_PER_PTE.
3328  *
3329  * The virtual address of the area that we map is naturally aligned to the
3330  * fault_around_pages() value (and therefore to page order).  This way it's
3331  * easier to guarantee that we don't cross page table boundaries.
3332  */
3333 static int do_fault_around(struct vm_fault *vmf)
3334 {
3335         unsigned long address = vmf->address, nr_pages, mask;
3336         pgoff_t start_pgoff = vmf->pgoff;
3337         pgoff_t end_pgoff;
3338         int off, ret = 0;
3339
3340         nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3341         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3342
3343         vmf->address = max(address & mask, vmf->vma->vm_start);
3344         off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3345         start_pgoff -= off;
3346
3347         /*
3348          *  end_pgoff is either end of page table or end of vma
3349          *  or fault_around_pages() from start_pgoff, depending what is nearest.
3350          */
3351         end_pgoff = start_pgoff -
3352                 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3353                 PTRS_PER_PTE - 1;
3354         end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3355                         start_pgoff + nr_pages - 1);
3356
3357         if (pmd_none(*vmf->pmd)) {
3358                 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3359                                                   vmf->address);
3360                 if (!vmf->prealloc_pte)
3361                         goto out;
3362                 smp_wmb(); /* See comment in __pte_alloc() */
3363         }
3364
3365         vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3366
3367         /* Huge page is mapped? Page fault is solved */
3368         if (pmd_trans_huge(*vmf->pmd)) {
3369                 ret = VM_FAULT_NOPAGE;
3370                 goto out;
3371         }
3372
3373         /* ->map_pages() haven't done anything useful. Cold page cache? */
3374         if (!vmf->pte)
3375                 goto out;
3376
3377         /* check if the page fault is solved */
3378         vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3379         if (!pte_none(*vmf->pte))
3380                 ret = VM_FAULT_NOPAGE;
3381         pte_unmap_unlock(vmf->pte, vmf->ptl);
3382 out:
3383         vmf->address = address;
3384         vmf->pte = NULL;
3385         return ret;
3386 }
3387
3388 static int do_read_fault(struct vm_fault *vmf)
3389 {
3390         struct vm_area_struct *vma = vmf->vma;
3391         int ret = 0;
3392
3393         /*
3394          * Let's call ->map_pages() first and use ->fault() as fallback
3395          * if page by the offset is not ready to be mapped (cold cache or
3396          * something).
3397          */
3398         if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3399                 ret = do_fault_around(vmf);
3400                 if (ret)
3401                         return ret;
3402         }
3403
3404         ret = __do_fault(vmf);
3405         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3406                 return ret;
3407
3408         ret |= finish_fault(vmf);
3409         unlock_page(vmf->page);
3410         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3411                 put_page(vmf->page);
3412         return ret;
3413 }
3414
3415 static int do_cow_fault(struct vm_fault *vmf)
3416 {
3417         struct vm_area_struct *vma = vmf->vma;
3418         int ret;
3419
3420         if (unlikely(anon_vma_prepare(vma)))
3421                 return VM_FAULT_OOM;
3422
3423         vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3424         if (!vmf->cow_page)
3425                 return VM_FAULT_OOM;
3426
3427         if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3428                                 &vmf->memcg, false)) {
3429                 put_page(vmf->cow_page);
3430                 return VM_FAULT_OOM;
3431         }
3432
3433         ret = __do_fault(vmf);
3434         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3435                 goto uncharge_out;
3436         if (ret & VM_FAULT_DONE_COW)
3437                 return ret;
3438
3439         copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3440         __SetPageUptodate(vmf->cow_page);
3441
3442         ret |= finish_fault(vmf);
3443         unlock_page(vmf->page);
3444         put_page(vmf->page);
3445         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3446                 goto uncharge_out;
3447         return ret;
3448 uncharge_out:
3449         mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3450         put_page(vmf->cow_page);
3451         return ret;
3452 }
3453
3454 static int do_shared_fault(struct vm_fault *vmf)
3455 {
3456         struct vm_area_struct *vma = vmf->vma;
3457         int ret, tmp;
3458
3459         ret = __do_fault(vmf);
3460         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3461                 return ret;
3462
3463         /*
3464          * Check if the backing address space wants to know that the page is
3465          * about to become writable
3466          */
3467         if (vma->vm_ops->page_mkwrite) {
3468                 unlock_page(vmf->page);
3469                 tmp = do_page_mkwrite(vmf);
3470                 if (unlikely(!tmp ||
3471                                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3472                         put_page(vmf->page);
3473                         return tmp;
3474                 }
3475         }
3476
3477         ret |= finish_fault(vmf);
3478         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3479                                         VM_FAULT_RETRY))) {
3480                 unlock_page(vmf->page);
3481                 put_page(vmf->page);
3482                 return ret;
3483         }
3484
3485         fault_dirty_shared_page(vma, vmf->page);
3486         return ret;
3487 }
3488
3489 /*
3490  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3491  * but allow concurrent faults).
3492  * The mmap_sem may have been released depending on flags and our
3493  * return value.  See filemap_fault() and __lock_page_or_retry().
3494  */
3495 static int do_fault(struct vm_fault *vmf)
3496 {
3497         struct vm_area_struct *vma = vmf->vma;
3498         int ret;
3499
3500         /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3501         if (!vma->vm_ops->fault)
3502                 ret = VM_FAULT_SIGBUS;
3503         else if (!(vmf->flags & FAULT_FLAG_WRITE))
3504                 ret = do_read_fault(vmf);
3505         else if (!(vma->vm_flags & VM_SHARED))
3506                 ret = do_cow_fault(vmf);
3507         else
3508                 ret = do_shared_fault(vmf);
3509
3510         /* preallocated pagetable is unused: free it */
3511         if (vmf->prealloc_pte) {
3512                 pte_free(vma->vm_mm, vmf->prealloc_pte);
3513                 vmf->prealloc_pte = NULL;
3514         }
3515         return ret;
3516 }
3517
3518 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3519                                 unsigned long addr, int page_nid,
3520                                 int *flags)
3521 {
3522         get_page(page);
3523
3524         count_vm_numa_event(NUMA_HINT_FAULTS);
3525         if (page_nid == numa_node_id()) {
3526                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3527                 *flags |= TNF_FAULT_LOCAL;
3528         }
3529
3530         return mpol_misplaced(page, vma, addr);
3531 }
3532
3533 static int do_numa_page(struct vm_fault *vmf)
3534 {
3535         struct vm_area_struct *vma = vmf->vma;
3536         struct page *page = NULL;
3537         int page_nid = -1;
3538         int last_cpupid;
3539         int target_nid;
3540         bool migrated = false;
3541         pte_t pte;
3542         bool was_writable = pte_savedwrite(vmf->orig_pte);
3543         int flags = 0;
3544
3545         /*
3546          * The "pte" at this point cannot be used safely without
3547          * validation through pte_unmap_same(). It's of NUMA type but
3548          * the pfn may be screwed if the read is non atomic.
3549          */
3550         vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3551         spin_lock(vmf->ptl);
3552         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3553                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3554                 goto out;
3555         }
3556
3557         /*
3558          * Make it present again, Depending on how arch implementes non
3559          * accessible ptes, some can allow access by kernel mode.
3560          */
3561         pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3562         pte = pte_modify(pte, vma->vm_page_prot);
3563         pte = pte_mkyoung(pte);
3564         if (was_writable)
3565                 pte = pte_mkwrite(pte);
3566         ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3567         update_mmu_cache(vma, vmf->address, vmf->pte);
3568
3569         page = vm_normal_page(vma, vmf->address, pte);
3570         if (!page) {
3571                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3572                 return 0;
3573         }
3574
3575         /* TODO: handle PTE-mapped THP */
3576         if (PageCompound(page)) {
3577                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3578                 return 0;
3579         }
3580
3581         /*
3582          * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3583          * much anyway since they can be in shared cache state. This misses
3584          * the case where a mapping is writable but the process never writes
3585          * to it but pte_write gets cleared during protection updates and
3586          * pte_dirty has unpredictable behaviour between PTE scan updates,
3587          * background writeback, dirty balancing and application behaviour.
3588          */
3589         if (!pte_write(pte))
3590                 flags |= TNF_NO_GROUP;
3591
3592         /*
3593          * Flag if the page is shared between multiple address spaces. This
3594          * is later used when determining whether to group tasks together
3595          */
3596         if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3597                 flags |= TNF_SHARED;
3598
3599         last_cpupid = page_cpupid_last(page);
3600         page_nid = page_to_nid(page);
3601         target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3602                         &flags);
3603         pte_unmap_unlock(vmf->pte, vmf->ptl);
3604         if (target_nid == -1) {
3605                 put_page(page);
3606                 goto out;
3607         }
3608
3609         /* Migrate to the requested node */
3610         migrated = migrate_misplaced_page(page, vma, target_nid);
3611         if (migrated) {
3612                 page_nid = target_nid;
3613                 flags |= TNF_MIGRATED;
3614         } else
3615                 flags |= TNF_MIGRATE_FAIL;
3616
3617 out:
3618         if (page_nid != -1)
3619                 task_numa_fault(last_cpupid, page_nid, 1, flags);
3620         return 0;
3621 }
3622
3623 static inline int create_huge_pmd(struct vm_fault *vmf)
3624 {
3625         if (vma_is_anonymous(vmf->vma))
3626                 return do_huge_pmd_anonymous_page(vmf);
3627         if (vmf->vma->vm_ops->huge_fault)
3628                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3629         return VM_FAULT_FALLBACK;
3630 }
3631
3632 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3633 {
3634         if (vma_is_anonymous(vmf->vma))
3635                 return do_huge_pmd_wp_page(vmf, orig_pmd);
3636         if (vmf->vma->vm_ops->huge_fault)
3637                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3638
3639         /* COW handled on pte level: split pmd */
3640         VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3641         __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3642
3643         return VM_FAULT_FALLBACK;
3644 }
3645
3646 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3647 {
3648         return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3649 }
3650
3651 static int create_huge_pud(struct vm_fault *vmf)
3652 {
3653 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3654         /* No support for anonymous transparent PUD pages yet */
3655         if (vma_is_anonymous(vmf->vma))
3656                 return VM_FAULT_FALLBACK;
3657         if (vmf->vma->vm_ops->huge_fault)
3658                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3659 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3660         return VM_FAULT_FALLBACK;
3661 }
3662
3663 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3664 {
3665 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3666         /* No support for anonymous transparent PUD pages yet */
3667         if (vma_is_anonymous(vmf->vma))
3668                 return VM_FAULT_FALLBACK;
3669         if (vmf->vma->vm_ops->huge_fault)
3670                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3671 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3672         return VM_FAULT_FALLBACK;
3673 }
3674
3675 /*
3676  * These routines also need to handle stuff like marking pages dirty
3677  * and/or accessed for architectures that don't do it in hardware (most
3678  * RISC architectures).  The early dirtying is also good on the i386.
3679  *
3680  * There is also a hook called "update_mmu_cache()" that architectures
3681  * with external mmu caches can use to update those (ie the Sparc or
3682  * PowerPC hashed page tables that act as extended TLBs).
3683  *
3684  * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3685  * concurrent faults).
3686  *
3687  * The mmap_sem may have been released depending on flags and our return value.
3688  * See filemap_fault() and __lock_page_or_retry().
3689  */
3690 static int handle_pte_fault(struct vm_fault *vmf)
3691 {
3692         pte_t entry;
3693
3694         if (unlikely(pmd_none(*vmf->pmd))) {
3695                 /*
3696                  * Leave __pte_alloc() until later: because vm_ops->fault may
3697                  * want to allocate huge page, and if we expose page table
3698                  * for an instant, it will be difficult to retract from
3699                  * concurrent faults and from rmap lookups.
3700                  */
3701                 vmf->pte = NULL;
3702         } else {
3703                 /* See comment in pte_alloc_one_map() */
3704                 if (pmd_devmap_trans_unstable(vmf->pmd))
3705                         return 0;
3706                 /*
3707                  * A regular pmd is established and it can't morph into a huge
3708                  * pmd from under us anymore at this point because we hold the
3709                  * mmap_sem read mode and khugepaged takes it in write mode.
3710                  * So now it's safe to run pte_offset_map().
3711                  */
3712                 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3713                 vmf->orig_pte = *vmf->pte;
3714
3715                 /*
3716                  * some architectures can have larger ptes than wordsize,
3717                  * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3718                  * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3719                  * atomic accesses.  The code below just needs a consistent
3720                  * view for the ifs and we later double check anyway with the
3721                  * ptl lock held. So here a barrier will do.
3722                  */
3723                 barrier();
3724                 if (pte_none(vmf->orig_pte)) {
3725                         pte_unmap(vmf->pte);
3726                         vmf->pte = NULL;
3727                 }
3728         }
3729
3730         if (!vmf->pte) {
3731                 if (vma_is_anonymous(vmf->vma))
3732                         return do_anonymous_page(vmf);
3733                 else
3734                         return do_fault(vmf);
3735         }
3736
3737         if (!pte_present(vmf->orig_pte))
3738                 return do_swap_page(vmf);
3739
3740         if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3741                 return do_numa_page(vmf);
3742
3743         vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3744         spin_lock(vmf->ptl);
3745         entry = vmf->orig_pte;
3746         if (unlikely(!pte_same(*vmf->pte, entry)))
3747                 goto unlock;
3748         if (vmf->flags & FAULT_FLAG_WRITE) {
3749                 if (!pte_write(entry))
3750                         return do_wp_page(vmf);
3751                 entry = pte_mkdirty(entry);
3752         }
3753         entry = pte_mkyoung(entry);
3754         if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3755                                 vmf->flags & FAULT_FLAG_WRITE)) {
3756                 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3757         } else {
3758                 /*
3759                  * This is needed only for protection faults but the arch code
3760                  * is not yet telling us if this is a protection fault or not.
3761                  * This still avoids useless tlb flushes for .text page faults
3762                  * with threads.
3763                  */
3764                 if (vmf->flags & FAULT_FLAG_WRITE)
3765                         flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3766         }
3767 unlock:
3768         pte_unmap_unlock(vmf->pte, vmf->ptl);
3769         return 0;
3770 }
3771
3772 /*
3773  * By the time we get here, we already hold the mm semaphore
3774  *
3775  * The mmap_sem may have been released depending on flags and our
3776  * return value.  See filemap_fault() and __lock_page_or_retry().
3777  */
3778 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3779                 unsigned int flags)
3780 {
3781         struct vm_fault vmf = {
3782                 .vma = vma,
3783                 .address = address & PAGE_MASK,
3784                 .flags = flags,
3785                 .pgoff = linear_page_index(vma, address),
3786                 .gfp_mask = __get_fault_gfp_mask(vma),
3787         };
3788         struct mm_struct *mm = vma->vm_mm;
3789         pgd_t *pgd;
3790         p4d_t *p4d;
3791         int ret;
3792
3793         pgd = pgd_offset(mm, address);
3794         p4d = p4d_alloc(mm, pgd, address);
3795         if (!p4d)
3796                 return VM_FAULT_OOM;
3797
3798         vmf.pud = pud_alloc(mm, p4d, address);
3799         if (!vmf.pud)
3800                 return VM_FAULT_OOM;
3801         if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
3802                 ret = create_huge_pud(&vmf);
3803                 if (!(ret & VM_FAULT_FALLBACK))
3804                         return ret;
3805         } else {
3806                 pud_t orig_pud = *vmf.pud;
3807
3808                 barrier();
3809                 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3810                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3811
3812                         /* NUMA case for anonymous PUDs would go here */
3813
3814                         if (dirty && !pud_write(orig_pud)) {
3815                                 ret = wp_huge_pud(&vmf, orig_pud);
3816                                 if (!(ret & VM_FAULT_FALLBACK))
3817                                         return ret;
3818                         } else {
3819                                 huge_pud_set_accessed(&vmf, orig_pud);
3820                                 return 0;
3821                         }
3822                 }
3823         }
3824
3825         vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3826         if (!vmf.pmd)
3827                 return VM_FAULT_OOM;
3828         if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3829                 ret = create_huge_pmd(&vmf);
3830                 if (!(ret & VM_FAULT_FALLBACK))
3831                         return ret;
3832         } else {
3833                 pmd_t orig_pmd = *vmf.pmd;
3834
3835                 barrier();
3836                 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3837                         if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3838                                 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3839
3840                         if ((vmf.flags & FAULT_FLAG_WRITE) &&
3841                                         !pmd_write(orig_pmd)) {
3842                                 ret = wp_huge_pmd(&vmf, orig_pmd);
3843                                 if (!(ret & VM_FAULT_FALLBACK))
3844                                         return ret;
3845                         } else {
3846                                 huge_pmd_set_accessed(&vmf, orig_pmd);
3847                                 return 0;
3848                         }
3849                 }
3850         }
3851
3852         return handle_pte_fault(&vmf);
3853 }
3854
3855 /*
3856  * By the time we get here, we already hold the mm semaphore
3857  *
3858  * The mmap_sem may have been released depending on flags and our
3859  * return value.  See filemap_fault() and __lock_page_or_retry().
3860  */
3861 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3862                 unsigned int flags)
3863 {
3864         int ret;
3865
3866         __set_current_state(TASK_RUNNING);
3867
3868         count_vm_event(PGFAULT);
3869         count_memcg_event_mm(vma->vm_mm, PGFAULT);
3870
3871         /* do counter updates before entering really critical section. */
3872         check_sync_rss_stat(current);
3873
3874         /*
3875          * Enable the memcg OOM handling for faults triggered in user
3876          * space.  Kernel faults are handled more gracefully.
3877          */
3878         if (flags & FAULT_FLAG_USER)
3879                 mem_cgroup_oom_enable();
3880
3881         if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3882                                             flags & FAULT_FLAG_INSTRUCTION,
3883                                             flags & FAULT_FLAG_REMOTE))
3884                 return VM_FAULT_SIGSEGV;
3885
3886         if (unlikely(is_vm_hugetlb_page(vma)))
3887                 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3888         else
3889                 ret = __handle_mm_fault(vma, address, flags);
3890
3891         if (flags & FAULT_FLAG_USER) {
3892                 mem_cgroup_oom_disable();
3893                 /*
3894                  * The task may have entered a memcg OOM situation but
3895                  * if the allocation error was handled gracefully (no
3896                  * VM_FAULT_OOM), there is no need to kill anything.
3897                  * Just clean up the OOM state peacefully.
3898                  */
3899                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3900                         mem_cgroup_oom_synchronize(false);
3901         }
3902
3903         /*
3904          * This mm has been already reaped by the oom reaper and so the
3905          * refault cannot be trusted in general. Anonymous refaults would
3906          * lose data and give a zero page instead e.g. This is especially
3907          * problem for use_mm() because regular tasks will just die and
3908          * the corrupted data will not be visible anywhere while kthread
3909          * will outlive the oom victim and potentially propagate the date
3910          * further.
3911          */
3912         if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3913                                 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3914                 ret = VM_FAULT_SIGBUS;
3915
3916         return ret;
3917 }
3918 EXPORT_SYMBOL_GPL(handle_mm_fault);
3919
3920 #ifndef __PAGETABLE_P4D_FOLDED
3921 /*
3922  * Allocate p4d page table.
3923  * We've already handled the fast-path in-line.
3924  */
3925 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3926 {
3927         p4d_t *new = p4d_alloc_one(mm, address);
3928         if (!new)
3929                 return -ENOMEM;
3930
3931         smp_wmb(); /* See comment in __pte_alloc */
3932
3933         spin_lock(&mm->page_table_lock);
3934         if (pgd_present(*pgd))          /* Another has populated it */
3935                 p4d_free(mm, new);
3936         else
3937                 pgd_populate(mm, pgd, new);
3938         spin_unlock(&mm->page_table_lock);
3939         return 0;
3940 }
3941 #endif /* __PAGETABLE_P4D_FOLDED */
3942
3943 #ifndef __PAGETABLE_PUD_FOLDED
3944 /*
3945  * Allocate page upper directory.
3946  * We've already handled the fast-path in-line.
3947  */
3948 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
3949 {
3950         pud_t *new = pud_alloc_one(mm, address);
3951         if (!new)
3952                 return -ENOMEM;
3953
3954         smp_wmb(); /* See comment in __pte_alloc */
3955
3956         spin_lock(&mm->page_table_lock);
3957 #ifndef __ARCH_HAS_5LEVEL_HACK
3958         if (p4d_present(*p4d))          /* Another has populated it */
3959                 pud_free(mm, new);
3960         else
3961                 p4d_populate(mm, p4d, new);
3962 #else
3963         if (pgd_present(*p4d))          /* Another has populated it */
3964                 pud_free(mm, new);
3965         else
3966                 pgd_populate(mm, p4d, new);
3967 #endif /* __ARCH_HAS_5LEVEL_HACK */
3968         spin_unlock(&mm->page_table_lock);
3969         return 0;
3970 }
3971 #endif /* __PAGETABLE_PUD_FOLDED */
3972
3973 #ifndef __PAGETABLE_PMD_FOLDED
3974 /*
3975  * Allocate page middle directory.
3976  * We've already handled the fast-path in-line.
3977  */
3978 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3979 {
3980         spinlock_t *ptl;
3981         pmd_t *new = pmd_alloc_one(mm, address);
3982         if (!new)
3983                 return -ENOMEM;
3984
3985         smp_wmb(); /* See comment in __pte_alloc */
3986
3987         ptl = pud_lock(mm, pud);
3988 #ifndef __ARCH_HAS_4LEVEL_HACK
3989         if (!pud_present(*pud)) {
3990                 mm_inc_nr_pmds(mm);
3991                 pud_populate(mm, pud, new);
3992         } else  /* Another has populated it */
3993                 pmd_free(mm, new);
3994 #else
3995         if (!pgd_present(*pud)) {
3996                 mm_inc_nr_pmds(mm);
3997                 pgd_populate(mm, pud, new);
3998         } else /* Another has populated it */
3999                 pmd_free(mm, new);
4000 #endif /* __ARCH_HAS_4LEVEL_HACK */
4001         spin_unlock(ptl);
4002         return 0;
4003 }
4004 #endif /* __PAGETABLE_PMD_FOLDED */
4005
4006 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4007                 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4008 {
4009         pgd_t *pgd;
4010         p4d_t *p4d;
4011         pud_t *pud;
4012         pmd_t *pmd;
4013         pte_t *ptep;
4014
4015         pgd = pgd_offset(mm, address);
4016         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4017                 goto out;
4018
4019         p4d = p4d_offset(pgd, address);
4020         if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4021                 goto out;
4022
4023         pud = pud_offset(p4d, address);
4024         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4025                 goto out;
4026
4027         pmd = pmd_offset(pud, address);
4028         VM_BUG_ON(pmd_trans_huge(*pmd));
4029
4030         if (pmd_huge(*pmd)) {
4031                 if (!pmdpp)
4032                         goto out;
4033
4034                 *ptlp = pmd_lock(mm, pmd);
4035                 if (pmd_huge(*pmd)) {
4036                         *pmdpp = pmd;
4037                         return 0;
4038                 }
4039                 spin_unlock(*ptlp);
4040         }
4041
4042         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4043                 goto out;
4044
4045         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4046         if (!pte_present(*ptep))
4047                 goto unlock;
4048         *ptepp = ptep;
4049         return 0;
4050 unlock:
4051         pte_unmap_unlock(ptep, *ptlp);
4052 out:
4053         return -EINVAL;
4054 }
4055
4056 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4057                              pte_t **ptepp, spinlock_t **ptlp)
4058 {
4059         int res;
4060
4061         /* (void) is needed to make gcc happy */
4062         (void) __cond_lock(*ptlp,
4063                            !(res = __follow_pte_pmd(mm, address, ptepp, NULL,
4064                                            ptlp)));
4065         return res;
4066 }
4067
4068 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4069                              pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4070 {
4071         int res;
4072
4073         /* (void) is needed to make gcc happy */
4074         (void) __cond_lock(*ptlp,
4075                            !(res = __follow_pte_pmd(mm, address, ptepp, pmdpp,
4076                                            ptlp)));
4077         return res;
4078 }
4079 EXPORT_SYMBOL(follow_pte_pmd);
4080
4081 /**
4082  * follow_pfn - look up PFN at a user virtual address
4083  * @vma: memory mapping
4084  * @address: user virtual address
4085  * @pfn: location to store found PFN
4086  *
4087  * Only IO mappings and raw PFN mappings are allowed.
4088  *
4089  * Returns zero and the pfn at @pfn on success, -ve otherwise.
4090  */
4091 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4092         unsigned long *pfn)
4093 {
4094         int ret = -EINVAL;
4095         spinlock_t *ptl;
4096         pte_t *ptep;
4097
4098         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4099                 return ret;
4100
4101         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4102         if (ret)
4103                 return ret;
4104         *pfn = pte_pfn(*ptep);
4105         pte_unmap_unlock(ptep, ptl);
4106         return 0;
4107 }
4108 EXPORT_SYMBOL(follow_pfn);
4109
4110 #ifdef CONFIG_HAVE_IOREMAP_PROT
4111 int follow_phys(struct vm_area_struct *vma,
4112                 unsigned long address, unsigned int flags,
4113                 unsigned long *prot, resource_size_t *phys)
4114 {
4115         int ret = -EINVAL;
4116         pte_t *ptep, pte;
4117         spinlock_t *ptl;
4118
4119         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4120                 goto out;
4121
4122         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4123                 goto out;
4124         pte = *ptep;
4125
4126         if ((flags & FOLL_WRITE) && !pte_write(pte))
4127                 goto unlock;
4128
4129         *prot = pgprot_val(pte_pgprot(pte));
4130         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4131
4132         ret = 0;
4133 unlock:
4134         pte_unmap_unlock(ptep, ptl);
4135 out:
4136         return ret;
4137 }
4138
4139 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4140                         void *buf, int len, int write)
4141 {
4142         resource_size_t phys_addr;
4143         unsigned long prot = 0;
4144         void __iomem *maddr;
4145         int offset = addr & (PAGE_SIZE-1);
4146
4147         if (follow_phys(vma, addr, write, &prot, &phys_addr))
4148                 return -EINVAL;
4149
4150         maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4151         if (write)
4152                 memcpy_toio(maddr + offset, buf, len);
4153         else
4154                 memcpy_fromio(buf, maddr + offset, len);
4155         iounmap(maddr);
4156
4157         return len;
4158 }
4159 EXPORT_SYMBOL_GPL(generic_access_phys);
4160 #endif
4161
4162 /*
4163  * Access another process' address space as given in mm.  If non-NULL, use the
4164  * given task for page fault accounting.
4165  */
4166 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4167                 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4168 {
4169         struct vm_area_struct *vma;
4170         void *old_buf = buf;
4171         int write = gup_flags & FOLL_WRITE;
4172
4173         down_read(&mm->mmap_sem);
4174         /* ignore errors, just check how much was successfully transferred */
4175         while (len) {
4176                 int bytes, ret, offset;
4177                 void *maddr;
4178                 struct page *page = NULL;
4179
4180                 ret = get_user_pages_remote(tsk, mm, addr, 1,
4181                                 gup_flags, &page, &vma, NULL);
4182                 if (ret <= 0) {
4183 #ifndef CONFIG_HAVE_IOREMAP_PROT
4184                         break;
4185 #else
4186                         /*
4187                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
4188                          * we can access using slightly different code.
4189                          */
4190                         vma = find_vma(mm, addr);
4191                         if (!vma || vma->vm_start > addr)
4192                                 break;
4193                         if (vma->vm_ops && vma->vm_ops->access)
4194                                 ret = vma->vm_ops->access(vma, addr, buf,
4195                                                           len, write);
4196                         if (ret <= 0)
4197                                 break;
4198                         bytes = ret;
4199 #endif
4200                 } else {
4201                         bytes = len;
4202                         offset = addr & (PAGE_SIZE-1);
4203                         if (bytes > PAGE_SIZE-offset)
4204                                 bytes = PAGE_SIZE-offset;
4205
4206                         maddr = kmap(page);
4207                         if (write) {
4208                                 copy_to_user_page(vma, page, addr,
4209                                                   maddr + offset, buf, bytes);
4210                                 set_page_dirty_lock(page);
4211                         } else {
4212                                 copy_from_user_page(vma, page, addr,
4213                                                     buf, maddr + offset, bytes);
4214                         }
4215                         kunmap(page);
4216                         put_page(page);
4217                 }
4218                 len -= bytes;
4219                 buf += bytes;
4220                 addr += bytes;
4221         }
4222         up_read(&mm->mmap_sem);
4223
4224         return buf - old_buf;
4225 }
4226
4227 /**
4228  * access_remote_vm - access another process' address space
4229  * @mm:         the mm_struct of the target address space
4230  * @addr:       start address to access
4231  * @buf:        source or destination buffer
4232  * @len:        number of bytes to transfer
4233  * @gup_flags:  flags modifying lookup behaviour
4234  *
4235  * The caller must hold a reference on @mm.
4236  */
4237 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4238                 void *buf, int len, unsigned int gup_flags)
4239 {
4240         return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4241 }
4242
4243 /*
4244  * Access another process' address space.
4245  * Source/target buffer must be kernel space,
4246  * Do not walk the page table directly, use get_user_pages
4247  */
4248 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4249                 void *buf, int len, unsigned int gup_flags)
4250 {
4251         struct mm_struct *mm;
4252         int ret;
4253
4254         mm = get_task_mm(tsk);
4255         if (!mm)
4256                 return 0;
4257
4258         ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4259
4260         mmput(mm);
4261
4262         return ret;
4263 }
4264 EXPORT_SYMBOL_GPL(access_process_vm);
4265
4266 /*
4267  * Print the name of a VMA.
4268  */
4269 void print_vma_addr(char *prefix, unsigned long ip)
4270 {
4271         struct mm_struct *mm = current->mm;
4272         struct vm_area_struct *vma;
4273
4274         /*
4275          * Do not print if we are in atomic
4276          * contexts (in exception stacks, etc.):
4277          */
4278         if (preempt_count())
4279                 return;
4280
4281         down_read(&mm->mmap_sem);
4282         vma = find_vma(mm, ip);
4283         if (vma && vma->vm_file) {
4284                 struct file *f = vma->vm_file;
4285                 char *buf = (char *)__get_free_page(GFP_KERNEL);
4286                 if (buf) {
4287                         char *p;
4288
4289                         p = file_path(f, buf, PAGE_SIZE);
4290                         if (IS_ERR(p))
4291                                 p = "?";
4292                         printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4293                                         vma->vm_start,
4294                                         vma->vm_end - vma->vm_start);
4295                         free_page((unsigned long)buf);
4296                 }
4297         }
4298         up_read(&mm->mmap_sem);
4299 }
4300
4301 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4302 void __might_fault(const char *file, int line)
4303 {
4304         /*
4305          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4306          * holding the mmap_sem, this is safe because kernel memory doesn't
4307          * get paged out, therefore we'll never actually fault, and the
4308          * below annotations will generate false positives.
4309          */
4310         if (uaccess_kernel())
4311                 return;
4312         if (pagefault_disabled())
4313                 return;
4314         __might_sleep(file, line, 0);
4315 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4316         if (current->mm)
4317                 might_lock_read(&current->mm->mmap_sem);
4318 #endif
4319 }
4320 EXPORT_SYMBOL(__might_fault);
4321 #endif
4322
4323 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4324 static void clear_gigantic_page(struct page *page,
4325                                 unsigned long addr,
4326                                 unsigned int pages_per_huge_page)
4327 {
4328         int i;
4329         struct page *p = page;
4330
4331         might_sleep();
4332         for (i = 0; i < pages_per_huge_page;
4333              i++, p = mem_map_next(p, page, i)) {
4334                 cond_resched();
4335                 clear_user_highpage(p, addr + i * PAGE_SIZE);
4336         }
4337 }
4338 void clear_huge_page(struct page *page,
4339                      unsigned long addr, unsigned int pages_per_huge_page)
4340 {
4341         int i;
4342
4343         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4344                 clear_gigantic_page(page, addr, pages_per_huge_page);
4345                 return;
4346         }
4347
4348         might_sleep();
4349         for (i = 0; i < pages_per_huge_page; i++) {
4350                 cond_resched();
4351                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4352         }
4353 }
4354
4355 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4356                                     unsigned long addr,
4357                                     struct vm_area_struct *vma,
4358                                     unsigned int pages_per_huge_page)
4359 {
4360         int i;
4361         struct page *dst_base = dst;
4362         struct page *src_base = src;
4363
4364         for (i = 0; i < pages_per_huge_page; ) {
4365                 cond_resched();
4366                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4367
4368                 i++;
4369                 dst = mem_map_next(dst, dst_base, i);
4370                 src = mem_map_next(src, src_base, i);
4371         }
4372 }
4373
4374 void copy_user_huge_page(struct page *dst, struct page *src,
4375                          unsigned long addr, struct vm_area_struct *vma,
4376                          unsigned int pages_per_huge_page)
4377 {
4378         int i;
4379
4380         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4381                 copy_user_gigantic_page(dst, src, addr, vma,
4382                                         pages_per_huge_page);
4383                 return;
4384         }
4385
4386         might_sleep();
4387         for (i = 0; i < pages_per_huge_page; i++) {
4388                 cond_resched();
4389                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4390         }
4391 }
4392
4393 long copy_huge_page_from_user(struct page *dst_page,
4394                                 const void __user *usr_src,
4395                                 unsigned int pages_per_huge_page,
4396                                 bool allow_pagefault)
4397 {
4398         void *src = (void *)usr_src;
4399         void *page_kaddr;
4400         unsigned long i, rc = 0;
4401         unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4402
4403         for (i = 0; i < pages_per_huge_page; i++) {
4404                 if (allow_pagefault)
4405                         page_kaddr = kmap(dst_page + i);
4406                 else
4407                         page_kaddr = kmap_atomic(dst_page + i);
4408                 rc = copy_from_user(page_kaddr,
4409                                 (const void __user *)(src + i * PAGE_SIZE),
4410                                 PAGE_SIZE);
4411                 if (allow_pagefault)
4412                         kunmap(dst_page + i);
4413                 else
4414                         kunmap_atomic(page_kaddr);
4415
4416                 ret_val -= (PAGE_SIZE - rc);
4417                 if (rc)
4418                         break;
4419
4420                 cond_resched();
4421         }
4422         return ret_val;
4423 }
4424 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4425
4426 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4427
4428 static struct kmem_cache *page_ptl_cachep;
4429
4430 void __init ptlock_cache_init(void)
4431 {
4432         page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4433                         SLAB_PANIC, NULL);
4434 }
4435
4436 bool ptlock_alloc(struct page *page)
4437 {
4438         spinlock_t *ptl;
4439
4440         ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4441         if (!ptl)
4442                 return false;
4443         page->ptl = ptl;
4444         return true;
4445 }
4446
4447 void ptlock_free(struct page *page)
4448 {
4449         kmem_cache_free(page_ptl_cachep, page->ptl);
4450 }
4451 #endif