Version:  2.0.40 2.2.26 2.4.37 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8

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

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us