Version:  2.0.40 2.2.26 2.4.37 3.10 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

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

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