Version:  2.0.40 2.2.26 2.4.37 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0

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

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