mm: avoid allocating CMA memory for stack
[linux-3.10.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/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
63 #include <asm/io.h>
64 #include <asm/pgalloc.h>
65 #include <asm/uaccess.h>
66 #include <asm/tlb.h>
67 #include <asm/tlbflush.h>
68 #include <asm/pgtable.h>
69
70 #include "internal.h"
71
72 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
73 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
74 #endif
75
76 #ifndef CONFIG_NEED_MULTIPLE_NODES
77 /* use the per-pgdat data instead for discontigmem - mbligh */
78 unsigned long max_mapnr;
79 struct page *mem_map;
80
81 EXPORT_SYMBOL(max_mapnr);
82 EXPORT_SYMBOL(mem_map);
83 #endif
84
85 unsigned long num_physpages;
86 /*
87  * A number of key systems in x86 including ioremap() rely on the assumption
88  * that high_memory defines the upper bound on direct map memory, then end
89  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
90  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
91  * and ZONE_HIGHMEM.
92  */
93 void * high_memory;
94
95 EXPORT_SYMBOL(num_physpages);
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 /*
122  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123  */
124 static int __init init_zero_pfn(void)
125 {
126         zero_pfn = page_to_pfn(ZERO_PAGE(0));
127         return 0;
128 }
129 core_initcall(init_zero_pfn);
130
131
132 #if defined(SPLIT_RSS_COUNTING)
133
134 void sync_mm_rss(struct mm_struct *mm)
135 {
136         int i;
137
138         for (i = 0; i < NR_MM_COUNTERS; i++) {
139                 if (current->rss_stat.count[i]) {
140                         add_mm_counter(mm, i, current->rss_stat.count[i]);
141                         current->rss_stat.count[i] = 0;
142                 }
143         }
144         current->rss_stat.events = 0;
145 }
146
147 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148 {
149         struct task_struct *task = current;
150
151         if (likely(task->mm == mm))
152                 task->rss_stat.count[member] += val;
153         else
154                 add_mm_counter(mm, member, val);
155 }
156 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159 /* sync counter once per 64 page faults */
160 #define TASK_RSS_EVENTS_THRESH  (64)
161 static void check_sync_rss_stat(struct task_struct *task)
162 {
163         if (unlikely(task != current))
164                 return;
165         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166                 sync_mm_rss(task->mm);
167 }
168 #else /* SPLIT_RSS_COUNTING */
169
170 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173 static void check_sync_rss_stat(struct task_struct *task)
174 {
175 }
176
177 #endif /* SPLIT_RSS_COUNTING */
178
179 #ifdef HAVE_GENERIC_MMU_GATHER
180
181 static int tlb_next_batch(struct mmu_gather *tlb)
182 {
183         struct mmu_gather_batch *batch;
184
185         batch = tlb->active;
186         if (batch->next) {
187                 tlb->active = batch->next;
188                 return 1;
189         }
190
191         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192                 return 0;
193
194         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195         if (!batch)
196                 return 0;
197
198         tlb->batch_count++;
199         batch->next = NULL;
200         batch->nr   = 0;
201         batch->max  = MAX_GATHER_BATCH;
202
203         tlb->active->next = batch;
204         tlb->active = batch;
205
206         return 1;
207 }
208
209 /* tlb_gather_mmu
210  *      Called to initialize an (on-stack) mmu_gather structure for page-table
211  *      tear-down from @mm. The @fullmm argument is used when @mm is without
212  *      users and we're going to destroy the full address space (exit/execve).
213  */
214 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
215 {
216         tlb->mm = mm;
217
218         /* Is it from 0 to ~0? */
219         tlb->fullmm     = !(start | (end+1));
220         tlb->need_flush_all = 0;
221         tlb->start      = start;
222         tlb->end        = end;
223         tlb->need_flush = 0;
224         tlb->local.next = NULL;
225         tlb->local.nr   = 0;
226         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
227         tlb->active     = &tlb->local;
228         tlb->batch_count = 0;
229
230 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
231         tlb->batch = NULL;
232 #endif
233 }
234
235 void tlb_flush_mmu(struct mmu_gather *tlb)
236 {
237         struct mmu_gather_batch *batch;
238
239         if (!tlb->need_flush)
240                 return;
241         tlb->need_flush = 0;
242         tlb_flush(tlb);
243 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
244         tlb_table_flush(tlb);
245 #endif
246
247         for (batch = &tlb->local; batch; batch = batch->next) {
248                 free_pages_and_swap_cache(batch->pages, batch->nr);
249                 batch->nr = 0;
250         }
251         tlb->active = &tlb->local;
252 }
253
254 /* tlb_finish_mmu
255  *      Called at the end of the shootdown operation to free up any resources
256  *      that were required.
257  */
258 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
259 {
260         struct mmu_gather_batch *batch, *next;
261
262         tlb_flush_mmu(tlb);
263
264         /* keep the page table cache within bounds */
265         check_pgt_cache();
266
267         for (batch = tlb->local.next; batch; batch = next) {
268                 next = batch->next;
269                 free_pages((unsigned long)batch, 0);
270         }
271         tlb->local.next = NULL;
272 }
273
274 /* __tlb_remove_page
275  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
276  *      handling the additional races in SMP caused by other CPUs caching valid
277  *      mappings in their TLBs. Returns the number of free page slots left.
278  *      When out of page slots we must call tlb_flush_mmu().
279  */
280 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
281 {
282         struct mmu_gather_batch *batch;
283
284         VM_BUG_ON(!tlb->need_flush);
285
286         batch = tlb->active;
287         batch->pages[batch->nr++] = page;
288         if (batch->nr == batch->max) {
289                 if (!tlb_next_batch(tlb))
290                         return 0;
291                 batch = tlb->active;
292         }
293         VM_BUG_ON(batch->nr > batch->max);
294
295         return batch->max - batch->nr;
296 }
297
298 #endif /* HAVE_GENERIC_MMU_GATHER */
299
300 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
301
302 /*
303  * See the comment near struct mmu_table_batch.
304  */
305
306 static void tlb_remove_table_smp_sync(void *arg)
307 {
308         /* Simply deliver the interrupt */
309 }
310
311 static void tlb_remove_table_one(void *table)
312 {
313         /*
314          * This isn't an RCU grace period and hence the page-tables cannot be
315          * assumed to be actually RCU-freed.
316          *
317          * It is however sufficient for software page-table walkers that rely on
318          * IRQ disabling. See the comment near struct mmu_table_batch.
319          */
320         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
321         __tlb_remove_table(table);
322 }
323
324 static void tlb_remove_table_rcu(struct rcu_head *head)
325 {
326         struct mmu_table_batch *batch;
327         int i;
328
329         batch = container_of(head, struct mmu_table_batch, rcu);
330
331         for (i = 0; i < batch->nr; i++)
332                 __tlb_remove_table(batch->tables[i]);
333
334         free_page((unsigned long)batch);
335 }
336
337 void tlb_table_flush(struct mmu_gather *tlb)
338 {
339         struct mmu_table_batch **batch = &tlb->batch;
340
341         if (*batch) {
342                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
343                 *batch = NULL;
344         }
345 }
346
347 void tlb_remove_table(struct mmu_gather *tlb, void *table)
348 {
349         struct mmu_table_batch **batch = &tlb->batch;
350
351         tlb->need_flush = 1;
352
353         /*
354          * When there's less then two users of this mm there cannot be a
355          * concurrent page-table walk.
356          */
357         if (atomic_read(&tlb->mm->mm_users) < 2) {
358                 __tlb_remove_table(table);
359                 return;
360         }
361
362         if (*batch == NULL) {
363                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
364                 if (*batch == NULL) {
365                         tlb_remove_table_one(table);
366                         return;
367                 }
368                 (*batch)->nr = 0;
369         }
370         (*batch)->tables[(*batch)->nr++] = table;
371         if ((*batch)->nr == MAX_TABLE_BATCH)
372                 tlb_table_flush(tlb);
373 }
374
375 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
376
377 /*
378  * If a p?d_bad entry is found while walking page tables, report
379  * the error, before resetting entry to p?d_none.  Usually (but
380  * very seldom) called out from the p?d_none_or_clear_bad macros.
381  */
382
383 void pgd_clear_bad(pgd_t *pgd)
384 {
385         pgd_ERROR(*pgd);
386         pgd_clear(pgd);
387 }
388
389 void pud_clear_bad(pud_t *pud)
390 {
391         pud_ERROR(*pud);
392         pud_clear(pud);
393 }
394
395 void pmd_clear_bad(pmd_t *pmd)
396 {
397         pmd_ERROR(*pmd);
398         pmd_clear(pmd);
399 }
400
401 /*
402  * Note: this doesn't free the actual pages themselves. That
403  * has been handled earlier when unmapping all the memory regions.
404  */
405 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
406                            unsigned long addr)
407 {
408         pgtable_t token = pmd_pgtable(*pmd);
409         pmd_clear(pmd);
410         pte_free_tlb(tlb, token, addr);
411         tlb->mm->nr_ptes--;
412 }
413
414 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
415                                 unsigned long addr, unsigned long end,
416                                 unsigned long floor, unsigned long ceiling)
417 {
418         pmd_t *pmd;
419         unsigned long next;
420         unsigned long start;
421
422         start = addr;
423         pmd = pmd_offset(pud, addr);
424         do {
425                 next = pmd_addr_end(addr, end);
426                 if (pmd_none_or_clear_bad(pmd))
427                         continue;
428                 free_pte_range(tlb, pmd, addr);
429         } while (pmd++, addr = next, addr != end);
430
431         start &= PUD_MASK;
432         if (start < floor)
433                 return;
434         if (ceiling) {
435                 ceiling &= PUD_MASK;
436                 if (!ceiling)
437                         return;
438         }
439         if (end - 1 > ceiling - 1)
440                 return;
441
442         pmd = pmd_offset(pud, start);
443         pud_clear(pud);
444         pmd_free_tlb(tlb, pmd, start);
445 }
446
447 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
448                                 unsigned long addr, unsigned long end,
449                                 unsigned long floor, unsigned long ceiling)
450 {
451         pud_t *pud;
452         unsigned long next;
453         unsigned long start;
454
455         start = addr;
456         pud = pud_offset(pgd, addr);
457         do {
458                 next = pud_addr_end(addr, end);
459                 if (pud_none_or_clear_bad(pud))
460                         continue;
461                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
462         } while (pud++, addr = next, addr != end);
463
464         start &= PGDIR_MASK;
465         if (start < floor)
466                 return;
467         if (ceiling) {
468                 ceiling &= PGDIR_MASK;
469                 if (!ceiling)
470                         return;
471         }
472         if (end - 1 > ceiling - 1)
473                 return;
474
475         pud = pud_offset(pgd, start);
476         pgd_clear(pgd);
477         pud_free_tlb(tlb, pud, start);
478 }
479
480 /*
481  * This function frees user-level page tables of a process.
482  *
483  * Must be called with pagetable lock held.
484  */
485 void free_pgd_range(struct mmu_gather *tlb,
486                         unsigned long addr, unsigned long end,
487                         unsigned long floor, unsigned long ceiling)
488 {
489         pgd_t *pgd;
490         unsigned long next;
491
492         /*
493          * The next few lines have given us lots of grief...
494          *
495          * Why are we testing PMD* at this top level?  Because often
496          * there will be no work to do at all, and we'd prefer not to
497          * go all the way down to the bottom just to discover that.
498          *
499          * Why all these "- 1"s?  Because 0 represents both the bottom
500          * of the address space and the top of it (using -1 for the
501          * top wouldn't help much: the masks would do the wrong thing).
502          * The rule is that addr 0 and floor 0 refer to the bottom of
503          * the address space, but end 0 and ceiling 0 refer to the top
504          * Comparisons need to use "end - 1" and "ceiling - 1" (though
505          * that end 0 case should be mythical).
506          *
507          * Wherever addr is brought up or ceiling brought down, we must
508          * be careful to reject "the opposite 0" before it confuses the
509          * subsequent tests.  But what about where end is brought down
510          * by PMD_SIZE below? no, end can't go down to 0 there.
511          *
512          * Whereas we round start (addr) and ceiling down, by different
513          * masks at different levels, in order to test whether a table
514          * now has no other vmas using it, so can be freed, we don't
515          * bother to round floor or end up - the tests don't need that.
516          */
517
518         addr &= PMD_MASK;
519         if (addr < floor) {
520                 addr += PMD_SIZE;
521                 if (!addr)
522                         return;
523         }
524         if (ceiling) {
525                 ceiling &= PMD_MASK;
526                 if (!ceiling)
527                         return;
528         }
529         if (end - 1 > ceiling - 1)
530                 end -= PMD_SIZE;
531         if (addr > end - 1)
532                 return;
533
534         pgd = pgd_offset(tlb->mm, addr);
535         do {
536                 next = pgd_addr_end(addr, end);
537                 if (pgd_none_or_clear_bad(pgd))
538                         continue;
539                 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
540         } while (pgd++, addr = next, addr != end);
541 }
542
543 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
544                 unsigned long floor, unsigned long ceiling)
545 {
546         while (vma) {
547                 struct vm_area_struct *next = vma->vm_next;
548                 unsigned long addr = vma->vm_start;
549
550                 /*
551                  * Hide vma from rmap and truncate_pagecache before freeing
552                  * pgtables
553                  */
554                 unlink_anon_vmas(vma);
555                 unlink_file_vma(vma);
556
557                 if (is_vm_hugetlb_page(vma)) {
558                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
559                                 floor, next? next->vm_start: ceiling);
560                 } else {
561                         /*
562                          * Optimization: gather nearby vmas into one call down
563                          */
564                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
565                                && !is_vm_hugetlb_page(next)) {
566                                 vma = next;
567                                 next = vma->vm_next;
568                                 unlink_anon_vmas(vma);
569                                 unlink_file_vma(vma);
570                         }
571                         free_pgd_range(tlb, addr, vma->vm_end,
572                                 floor, next? next->vm_start: ceiling);
573                 }
574                 vma = next;
575         }
576 }
577
578 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
579                 pmd_t *pmd, unsigned long address)
580 {
581         pgtable_t new = pte_alloc_one(mm, address);
582         int wait_split_huge_page;
583         if (!new)
584                 return -ENOMEM;
585
586         /*
587          * Ensure all pte setup (eg. pte page lock and page clearing) are
588          * visible before the pte is made visible to other CPUs by being
589          * put into page tables.
590          *
591          * The other side of the story is the pointer chasing in the page
592          * table walking code (when walking the page table without locking;
593          * ie. most of the time). Fortunately, these data accesses consist
594          * of a chain of data-dependent loads, meaning most CPUs (alpha
595          * being the notable exception) will already guarantee loads are
596          * seen in-order. See the alpha page table accessors for the
597          * smp_read_barrier_depends() barriers in page table walking code.
598          */
599         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
600
601         spin_lock(&mm->page_table_lock);
602         wait_split_huge_page = 0;
603         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
604                 mm->nr_ptes++;
605                 pmd_populate(mm, pmd, new);
606                 new = NULL;
607         } else if (unlikely(pmd_trans_splitting(*pmd)))
608                 wait_split_huge_page = 1;
609         spin_unlock(&mm->page_table_lock);
610         if (new)
611                 pte_free(mm, new);
612         if (wait_split_huge_page)
613                 wait_split_huge_page(vma->anon_vma, pmd);
614         return 0;
615 }
616
617 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
618 {
619         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
620         if (!new)
621                 return -ENOMEM;
622
623         smp_wmb(); /* See comment in __pte_alloc */
624
625         spin_lock(&init_mm.page_table_lock);
626         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
627                 pmd_populate_kernel(&init_mm, pmd, new);
628                 new = NULL;
629         } else
630                 VM_BUG_ON(pmd_trans_splitting(*pmd));
631         spin_unlock(&init_mm.page_table_lock);
632         if (new)
633                 pte_free_kernel(&init_mm, new);
634         return 0;
635 }
636
637 static inline void init_rss_vec(int *rss)
638 {
639         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
640 }
641
642 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
643 {
644         int i;
645
646         if (current->mm == mm)
647                 sync_mm_rss(mm);
648         for (i = 0; i < NR_MM_COUNTERS; i++)
649                 if (rss[i])
650                         add_mm_counter(mm, i, rss[i]);
651 }
652
653 /*
654  * This function is called to print an error when a bad pte
655  * is found. For example, we might have a PFN-mapped pte in
656  * a region that doesn't allow it.
657  *
658  * The calling function must still handle the error.
659  */
660 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
661                           pte_t pte, struct page *page)
662 {
663         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
664         pud_t *pud = pud_offset(pgd, addr);
665         pmd_t *pmd = pmd_offset(pud, addr);
666         struct address_space *mapping;
667         pgoff_t index;
668         static unsigned long resume;
669         static unsigned long nr_shown;
670         static unsigned long nr_unshown;
671
672         /*
673          * Allow a burst of 60 reports, then keep quiet for that minute;
674          * or allow a steady drip of one report per second.
675          */
676         if (nr_shown == 60) {
677                 if (time_before(jiffies, resume)) {
678                         nr_unshown++;
679                         return;
680                 }
681                 if (nr_unshown) {
682                         printk(KERN_ALERT
683                                 "BUG: Bad page map: %lu messages suppressed\n",
684                                 nr_unshown);
685                         nr_unshown = 0;
686                 }
687                 nr_shown = 0;
688         }
689         if (nr_shown++ == 0)
690                 resume = jiffies + 60 * HZ;
691
692         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
693         index = linear_page_index(vma, addr);
694
695         printk(KERN_ALERT
696                 "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
697                 current->comm,
698                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
699         if (page)
700                 dump_page(page);
701         printk(KERN_ALERT
702                 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
703                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
704         /*
705          * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
706          */
707         if (vma->vm_ops)
708                 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
709                        vma->vm_ops->fault);
710         if (vma->vm_file && vma->vm_file->f_op)
711                 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
712                        vma->vm_file->f_op->mmap);
713         dump_stack();
714         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
715 }
716
717 static inline bool is_cow_mapping(vm_flags_t flags)
718 {
719         return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
720 }
721
722 /*
723  * vm_normal_page -- This function gets the "struct page" associated with a pte.
724  *
725  * "Special" mappings do not wish to be associated with a "struct page" (either
726  * it doesn't exist, or it exists but they don't want to touch it). In this
727  * case, NULL is returned here. "Normal" mappings do have a struct page.
728  *
729  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
730  * pte bit, in which case this function is trivial. Secondly, an architecture
731  * may not have a spare pte bit, which requires a more complicated scheme,
732  * described below.
733  *
734  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
735  * special mapping (even if there are underlying and valid "struct pages").
736  * COWed pages of a VM_PFNMAP are always normal.
737  *
738  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
739  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
740  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
741  * mapping will always honor the rule
742  *
743  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
744  *
745  * And for normal mappings this is false.
746  *
747  * This restricts such mappings to be a linear translation from virtual address
748  * to pfn. To get around this restriction, we allow arbitrary mappings so long
749  * as the vma is not a COW mapping; in that case, we know that all ptes are
750  * special (because none can have been COWed).
751  *
752  *
753  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
754  *
755  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
756  * page" backing, however the difference is that _all_ pages with a struct
757  * page (that is, those where pfn_valid is true) are refcounted and considered
758  * normal pages by the VM. The disadvantage is that pages are refcounted
759  * (which can be slower and simply not an option for some PFNMAP users). The
760  * advantage is that we don't have to follow the strict linearity rule of
761  * PFNMAP mappings in order to support COWable mappings.
762  *
763  */
764 #ifdef __HAVE_ARCH_PTE_SPECIAL
765 # define HAVE_PTE_SPECIAL 1
766 #else
767 # define HAVE_PTE_SPECIAL 0
768 #endif
769 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
770                                 pte_t pte)
771 {
772         unsigned long pfn = pte_pfn(pte);
773
774         if (HAVE_PTE_SPECIAL) {
775                 if (likely(!pte_special(pte)))
776                         goto check_pfn;
777                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
778                         return NULL;
779                 if (!is_zero_pfn(pfn))
780                         print_bad_pte(vma, addr, pte, NULL);
781                 return NULL;
782         }
783
784         /* !HAVE_PTE_SPECIAL case follows: */
785
786         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
787                 if (vma->vm_flags & VM_MIXEDMAP) {
788                         if (!pfn_valid(pfn))
789                                 return NULL;
790                         goto out;
791                 } else {
792                         unsigned long off;
793                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
794                         if (pfn == vma->vm_pgoff + off)
795                                 return NULL;
796                         if (!is_cow_mapping(vma->vm_flags))
797                                 return NULL;
798                 }
799         }
800
801         if (is_zero_pfn(pfn))
802                 return NULL;
803 check_pfn:
804         if (unlikely(pfn > highest_memmap_pfn)) {
805                 print_bad_pte(vma, addr, pte, NULL);
806                 return NULL;
807         }
808
809         /*
810          * NOTE! We still have PageReserved() pages in the page tables.
811          * eg. VDSO mappings can cause them to exist.
812          */
813 out:
814         return pfn_to_page(pfn);
815 }
816
817 /*
818  * copy one vm_area from one task to the other. Assumes the page tables
819  * already present in the new task to be cleared in the whole range
820  * covered by this vma.
821  */
822
823 static inline unsigned long
824 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
825                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
826                 unsigned long addr, int *rss)
827 {
828         unsigned long vm_flags = vma->vm_flags;
829         pte_t pte = *src_pte;
830         struct page *page;
831
832         /* pte contains position in swap or file, so copy. */
833         if (unlikely(!pte_present(pte))) {
834                 if (!pte_file(pte)) {
835                         swp_entry_t entry = pte_to_swp_entry(pte);
836
837                         if (swap_duplicate(entry) < 0)
838                                 return entry.val;
839
840                         /* make sure dst_mm is on swapoff's mmlist. */
841                         if (unlikely(list_empty(&dst_mm->mmlist))) {
842                                 spin_lock(&mmlist_lock);
843                                 if (list_empty(&dst_mm->mmlist))
844                                         list_add(&dst_mm->mmlist,
845                                                  &src_mm->mmlist);
846                                 spin_unlock(&mmlist_lock);
847                         }
848                         if (likely(!non_swap_entry(entry)))
849                                 rss[MM_SWAPENTS]++;
850                         else if (is_migration_entry(entry)) {
851                                 page = migration_entry_to_page(entry);
852
853                                 if (PageAnon(page))
854                                         rss[MM_ANONPAGES]++;
855                                 else
856                                         rss[MM_FILEPAGES]++;
857
858                                 if (is_write_migration_entry(entry) &&
859                                     is_cow_mapping(vm_flags)) {
860                                         /*
861                                          * COW mappings require pages in both
862                                          * parent and child to be set to read.
863                                          */
864                                         make_migration_entry_read(&entry);
865                                         pte = swp_entry_to_pte(entry);
866                                         set_pte_at(src_mm, addr, src_pte, pte);
867                                 }
868                         }
869                 }
870                 goto out_set_pte;
871         }
872
873         /*
874          * If it's a COW mapping, write protect it both
875          * in the parent and the child
876          */
877         if (is_cow_mapping(vm_flags)) {
878                 ptep_set_wrprotect(src_mm, addr, src_pte);
879                 pte = pte_wrprotect(pte);
880         }
881
882         /*
883          * If it's a shared mapping, mark it clean in
884          * the child
885          */
886         if (vm_flags & VM_SHARED)
887                 pte = pte_mkclean(pte);
888         pte = pte_mkold(pte);
889
890         page = vm_normal_page(vma, addr, pte);
891         if (page) {
892                 get_page(page);
893                 page_dup_rmap(page);
894                 if (PageAnon(page))
895                         rss[MM_ANONPAGES]++;
896                 else
897                         rss[MM_FILEPAGES]++;
898         }
899
900 out_set_pte:
901         set_pte_at(dst_mm, addr, dst_pte, pte);
902         return 0;
903 }
904
905 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
906                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
907                    unsigned long addr, unsigned long end)
908 {
909         pte_t *orig_src_pte, *orig_dst_pte;
910         pte_t *src_pte, *dst_pte;
911         spinlock_t *src_ptl, *dst_ptl;
912         int progress = 0;
913         int rss[NR_MM_COUNTERS];
914         swp_entry_t entry = (swp_entry_t){0};
915
916 again:
917         init_rss_vec(rss);
918
919         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
920         if (!dst_pte)
921                 return -ENOMEM;
922         src_pte = pte_offset_map(src_pmd, addr);
923         src_ptl = pte_lockptr(src_mm, src_pmd);
924         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
925         orig_src_pte = src_pte;
926         orig_dst_pte = dst_pte;
927         arch_enter_lazy_mmu_mode();
928
929         do {
930                 /*
931                  * We are holding two locks at this point - either of them
932                  * could generate latencies in another task on another CPU.
933                  */
934                 if (progress >= 32) {
935                         progress = 0;
936                         if (need_resched() ||
937                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
938                                 break;
939                 }
940                 if (pte_none(*src_pte)) {
941                         progress++;
942                         continue;
943                 }
944                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
945                                                         vma, addr, rss);
946                 if (entry.val)
947                         break;
948                 progress += 8;
949         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
950
951         arch_leave_lazy_mmu_mode();
952         spin_unlock(src_ptl);
953         pte_unmap(orig_src_pte);
954         add_mm_rss_vec(dst_mm, rss);
955         pte_unmap_unlock(orig_dst_pte, dst_ptl);
956         cond_resched();
957
958         if (entry.val) {
959                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
960                         return -ENOMEM;
961                 progress = 0;
962         }
963         if (addr != end)
964                 goto again;
965         return 0;
966 }
967
968 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
969                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
970                 unsigned long addr, unsigned long end)
971 {
972         pmd_t *src_pmd, *dst_pmd;
973         unsigned long next;
974
975         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
976         if (!dst_pmd)
977                 return -ENOMEM;
978         src_pmd = pmd_offset(src_pud, addr);
979         do {
980                 next = pmd_addr_end(addr, end);
981                 if (pmd_trans_huge(*src_pmd)) {
982                         int err;
983                         VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
984                         err = copy_huge_pmd(dst_mm, src_mm,
985                                             dst_pmd, src_pmd, addr, vma);
986                         if (err == -ENOMEM)
987                                 return -ENOMEM;
988                         if (!err)
989                                 continue;
990                         /* fall through */
991                 }
992                 if (pmd_none_or_clear_bad(src_pmd))
993                         continue;
994                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
995                                                 vma, addr, next))
996                         return -ENOMEM;
997         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
998         return 0;
999 }
1000
1001 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1002                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1003                 unsigned long addr, unsigned long end)
1004 {
1005         pud_t *src_pud, *dst_pud;
1006         unsigned long next;
1007
1008         dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1009         if (!dst_pud)
1010                 return -ENOMEM;
1011         src_pud = pud_offset(src_pgd, addr);
1012         do {
1013                 next = pud_addr_end(addr, end);
1014                 if (pud_none_or_clear_bad(src_pud))
1015                         continue;
1016                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1017                                                 vma, addr, next))
1018                         return -ENOMEM;
1019         } while (dst_pud++, src_pud++, addr = next, addr != end);
1020         return 0;
1021 }
1022
1023 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1024                 struct vm_area_struct *vma)
1025 {
1026         pgd_t *src_pgd, *dst_pgd;
1027         unsigned long next;
1028         unsigned long addr = vma->vm_start;
1029         unsigned long end = vma->vm_end;
1030         unsigned long mmun_start;       /* For mmu_notifiers */
1031         unsigned long mmun_end;         /* For mmu_notifiers */
1032         bool is_cow;
1033         int ret;
1034
1035         /*
1036          * Don't copy ptes where a page fault will fill them correctly.
1037          * Fork becomes much lighter when there are big shared or private
1038          * readonly mappings. The tradeoff is that copy_page_range is more
1039          * efficient than faulting.
1040          */
1041         if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1042                                VM_PFNMAP | VM_MIXEDMAP))) {
1043                 if (!vma->anon_vma)
1044                         return 0;
1045         }
1046
1047         if (is_vm_hugetlb_page(vma))
1048                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1049
1050         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1051                 /*
1052                  * We do not free on error cases below as remove_vma
1053                  * gets called on error from higher level routine
1054                  */
1055                 ret = track_pfn_copy(vma);
1056                 if (ret)
1057                         return ret;
1058         }
1059
1060         /*
1061          * We need to invalidate the secondary MMU mappings only when
1062          * there could be a permission downgrade on the ptes of the
1063          * parent mm. And a permission downgrade will only happen if
1064          * is_cow_mapping() returns true.
1065          */
1066         is_cow = is_cow_mapping(vma->vm_flags);
1067         mmun_start = addr;
1068         mmun_end   = end;
1069         if (is_cow)
1070                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1071                                                     mmun_end);
1072
1073         ret = 0;
1074         dst_pgd = pgd_offset(dst_mm, addr);
1075         src_pgd = pgd_offset(src_mm, addr);
1076         do {
1077                 next = pgd_addr_end(addr, end);
1078                 if (pgd_none_or_clear_bad(src_pgd))
1079                         continue;
1080                 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1081                                             vma, addr, next))) {
1082                         ret = -ENOMEM;
1083                         break;
1084                 }
1085         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1086
1087         if (is_cow)
1088                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1089         return ret;
1090 }
1091
1092 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1093                                 struct vm_area_struct *vma, pmd_t *pmd,
1094                                 unsigned long addr, unsigned long end,
1095                                 struct zap_details *details)
1096 {
1097         struct mm_struct *mm = tlb->mm;
1098         int force_flush = 0;
1099         int rss[NR_MM_COUNTERS];
1100         spinlock_t *ptl;
1101         pte_t *start_pte;
1102         pte_t *pte;
1103
1104 again:
1105         init_rss_vec(rss);
1106         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1107         pte = start_pte;
1108         arch_enter_lazy_mmu_mode();
1109         do {
1110                 pte_t ptent = *pte;
1111                 if (pte_none(ptent)) {
1112                         continue;
1113                 }
1114
1115                 if (pte_present(ptent)) {
1116                         struct page *page;
1117
1118                         page = vm_normal_page(vma, addr, ptent);
1119                         if (unlikely(details) && page) {
1120                                 /*
1121                                  * unmap_shared_mapping_pages() wants to
1122                                  * invalidate cache without truncating:
1123                                  * unmap shared but keep private pages.
1124                                  */
1125                                 if (details->check_mapping &&
1126                                     details->check_mapping != page->mapping)
1127                                         continue;
1128                                 /*
1129                                  * Each page->index must be checked when
1130                                  * invalidating or truncating nonlinear.
1131                                  */
1132                                 if (details->nonlinear_vma &&
1133                                     (page->index < details->first_index ||
1134                                      page->index > details->last_index))
1135                                         continue;
1136                         }
1137                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1138                                                         tlb->fullmm);
1139                         tlb_remove_tlb_entry(tlb, pte, addr);
1140                         if (unlikely(!page))
1141                                 continue;
1142                         if (unlikely(details) && details->nonlinear_vma
1143                             && linear_page_index(details->nonlinear_vma,
1144                                                 addr) != page->index)
1145                                 set_pte_at(mm, addr, pte,
1146                                            pgoff_to_pte(page->index));
1147                         if (PageAnon(page))
1148                                 rss[MM_ANONPAGES]--;
1149                         else {
1150                                 if (pte_dirty(ptent))
1151                                         set_page_dirty(page);
1152                                 if (pte_young(ptent) &&
1153                                     likely(!VM_SequentialReadHint(vma)))
1154                                         mark_page_accessed(page);
1155                                 rss[MM_FILEPAGES]--;
1156                         }
1157                         page_remove_rmap(page);
1158                         if (unlikely(page_mapcount(page) < 0))
1159                                 print_bad_pte(vma, addr, ptent, page);
1160                         force_flush = !__tlb_remove_page(tlb, page);
1161                         if (force_flush)
1162                                 break;
1163                         continue;
1164                 }
1165                 /*
1166                  * If details->check_mapping, we leave swap entries;
1167                  * if details->nonlinear_vma, we leave file entries.
1168                  */
1169                 if (unlikely(details))
1170                         continue;
1171                 if (pte_file(ptent)) {
1172                         if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1173                                 print_bad_pte(vma, addr, ptent, NULL);
1174                 } else {
1175                         swp_entry_t entry = pte_to_swp_entry(ptent);
1176
1177                         if (!non_swap_entry(entry))
1178                                 rss[MM_SWAPENTS]--;
1179                         else if (is_migration_entry(entry)) {
1180                                 struct page *page;
1181
1182                                 page = migration_entry_to_page(entry);
1183
1184                                 if (PageAnon(page))
1185                                         rss[MM_ANONPAGES]--;
1186                                 else
1187                                         rss[MM_FILEPAGES]--;
1188                         }
1189                         if (unlikely(!free_swap_and_cache(entry)))
1190                                 print_bad_pte(vma, addr, ptent, NULL);
1191                 }
1192                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1193         } while (pte++, addr += PAGE_SIZE, addr != end);
1194
1195         add_mm_rss_vec(mm, rss);
1196         arch_leave_lazy_mmu_mode();
1197         pte_unmap_unlock(start_pte, ptl);
1198
1199         /*
1200          * mmu_gather ran out of room to batch pages, we break out of
1201          * the PTE lock to avoid doing the potential expensive TLB invalidate
1202          * and page-free while holding it.
1203          */
1204         if (force_flush) {
1205                 unsigned long old_end;
1206
1207                 force_flush = 0;
1208
1209                 /*
1210                  * Flush the TLB just for the previous segment,
1211                  * then update the range to be the remaining
1212                  * TLB range.
1213                  */
1214                 old_end = tlb->end;
1215                 tlb->end = addr;
1216
1217                 tlb_flush_mmu(tlb);
1218
1219                 tlb->start = addr;
1220                 tlb->end = old_end;
1221
1222                 if (addr != end)
1223                         goto again;
1224         }
1225
1226         return addr;
1227 }
1228
1229 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1230                                 struct vm_area_struct *vma, pud_t *pud,
1231                                 unsigned long addr, unsigned long end,
1232                                 struct zap_details *details)
1233 {
1234         pmd_t *pmd;
1235         unsigned long next;
1236
1237         pmd = pmd_offset(pud, addr);
1238         do {
1239                 next = pmd_addr_end(addr, end);
1240                 if (pmd_trans_huge(*pmd)) {
1241                         if (next - addr != HPAGE_PMD_SIZE) {
1242 #ifdef CONFIG_DEBUG_VM
1243                                 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1244                                         pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1245                                                 __func__, addr, end,
1246                                                 vma->vm_start,
1247                                                 vma->vm_end);
1248                                         BUG();
1249                                 }
1250 #endif
1251                                 split_huge_page_pmd(vma, addr, pmd);
1252                         } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1253                                 goto next;
1254                         /* fall through */
1255                 }
1256                 /*
1257                  * Here there can be other concurrent MADV_DONTNEED or
1258                  * trans huge page faults running, and if the pmd is
1259                  * none or trans huge it can change under us. This is
1260                  * because MADV_DONTNEED holds the mmap_sem in read
1261                  * mode.
1262                  */
1263                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1264                         goto next;
1265                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1266 next:
1267                 cond_resched();
1268         } while (pmd++, addr = next, addr != end);
1269
1270         return addr;
1271 }
1272
1273 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1274                                 struct vm_area_struct *vma, pgd_t *pgd,
1275                                 unsigned long addr, unsigned long end,
1276                                 struct zap_details *details)
1277 {
1278         pud_t *pud;
1279         unsigned long next;
1280
1281         pud = pud_offset(pgd, addr);
1282         do {
1283                 next = pud_addr_end(addr, end);
1284                 if (pud_none_or_clear_bad(pud))
1285                         continue;
1286                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1287         } while (pud++, addr = next, addr != end);
1288
1289         return addr;
1290 }
1291
1292 static void unmap_page_range(struct mmu_gather *tlb,
1293                              struct vm_area_struct *vma,
1294                              unsigned long addr, unsigned long end,
1295                              struct zap_details *details)
1296 {
1297         pgd_t *pgd;
1298         unsigned long next;
1299
1300         if (details && !details->check_mapping && !details->nonlinear_vma)
1301                 details = NULL;
1302
1303         BUG_ON(addr >= end);
1304         mem_cgroup_uncharge_start();
1305         tlb_start_vma(tlb, vma);
1306         pgd = pgd_offset(vma->vm_mm, addr);
1307         do {
1308                 next = pgd_addr_end(addr, end);
1309                 if (pgd_none_or_clear_bad(pgd))
1310                         continue;
1311                 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1312         } while (pgd++, addr = next, addr != end);
1313         tlb_end_vma(tlb, vma);
1314         mem_cgroup_uncharge_end();
1315 }
1316
1317
1318 static void unmap_single_vma(struct mmu_gather *tlb,
1319                 struct vm_area_struct *vma, unsigned long start_addr,
1320                 unsigned long end_addr,
1321                 struct zap_details *details)
1322 {
1323         unsigned long start = max(vma->vm_start, start_addr);
1324         unsigned long end;
1325
1326         if (start >= vma->vm_end)
1327                 return;
1328         end = min(vma->vm_end, end_addr);
1329         if (end <= vma->vm_start)
1330                 return;
1331
1332         if (vma->vm_file)
1333                 uprobe_munmap(vma, start, end);
1334
1335         if (unlikely(vma->vm_flags & VM_PFNMAP))
1336                 untrack_pfn(vma, 0, 0);
1337
1338         if (start != end) {
1339                 if (unlikely(is_vm_hugetlb_page(vma))) {
1340                         /*
1341                          * It is undesirable to test vma->vm_file as it
1342                          * should be non-null for valid hugetlb area.
1343                          * However, vm_file will be NULL in the error
1344                          * cleanup path of do_mmap_pgoff. When
1345                          * hugetlbfs ->mmap method fails,
1346                          * do_mmap_pgoff() nullifies vma->vm_file
1347                          * before calling this function to clean up.
1348                          * Since no pte has actually been setup, it is
1349                          * safe to do nothing in this case.
1350                          */
1351                         if (vma->vm_file) {
1352                                 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1353                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1354                                 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1355                         }
1356                 } else
1357                         unmap_page_range(tlb, vma, start, end, details);
1358         }
1359 }
1360
1361 /**
1362  * unmap_vmas - unmap a range of memory covered by a list of vma's
1363  * @tlb: address of the caller's struct mmu_gather
1364  * @vma: the starting vma
1365  * @start_addr: virtual address at which to start unmapping
1366  * @end_addr: virtual address at which to end unmapping
1367  *
1368  * Unmap all pages in the vma list.
1369  *
1370  * Only addresses between `start' and `end' will be unmapped.
1371  *
1372  * The VMA list must be sorted in ascending virtual address order.
1373  *
1374  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1375  * range after unmap_vmas() returns.  So the only responsibility here is to
1376  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1377  * drops the lock and schedules.
1378  */
1379 void unmap_vmas(struct mmu_gather *tlb,
1380                 struct vm_area_struct *vma, unsigned long start_addr,
1381                 unsigned long end_addr)
1382 {
1383         struct mm_struct *mm = vma->vm_mm;
1384
1385         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1386         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1387                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1388         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1389 }
1390
1391 /**
1392  * zap_page_range - remove user pages in a given range
1393  * @vma: vm_area_struct holding the applicable pages
1394  * @start: starting address of pages to zap
1395  * @size: number of bytes to zap
1396  * @details: details of nonlinear truncation or shared cache invalidation
1397  *
1398  * Caller must protect the VMA list
1399  */
1400 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1401                 unsigned long size, struct zap_details *details)
1402 {
1403         struct mm_struct *mm = vma->vm_mm;
1404         struct mmu_gather tlb;
1405         unsigned long end = start + size;
1406
1407         lru_add_drain();
1408         tlb_gather_mmu(&tlb, mm, start, end);
1409         update_hiwater_rss(mm);
1410         mmu_notifier_invalidate_range_start(mm, start, end);
1411         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1412                 unmap_single_vma(&tlb, vma, start, end, details);
1413         mmu_notifier_invalidate_range_end(mm, start, end);
1414         tlb_finish_mmu(&tlb, start, end);
1415 }
1416
1417 /**
1418  * zap_page_range_single - remove user pages in a given range
1419  * @vma: vm_area_struct holding the applicable pages
1420  * @address: starting address of pages to zap
1421  * @size: number of bytes to zap
1422  * @details: details of nonlinear truncation or shared cache invalidation
1423  *
1424  * The range must fit into one VMA.
1425  */
1426 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1427                 unsigned long size, struct zap_details *details)
1428 {
1429         struct mm_struct *mm = vma->vm_mm;
1430         struct mmu_gather tlb;
1431         unsigned long end = address + size;
1432
1433         lru_add_drain();
1434         tlb_gather_mmu(&tlb, mm, address, end);
1435         update_hiwater_rss(mm);
1436         mmu_notifier_invalidate_range_start(mm, address, end);
1437         unmap_single_vma(&tlb, vma, address, end, details);
1438         mmu_notifier_invalidate_range_end(mm, address, end);
1439         tlb_finish_mmu(&tlb, address, end);
1440 }
1441
1442 /**
1443  * zap_vma_ptes - remove ptes mapping the vma
1444  * @vma: vm_area_struct holding ptes to be zapped
1445  * @address: starting address of pages to zap
1446  * @size: number of bytes to zap
1447  *
1448  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1449  *
1450  * The entire address range must be fully contained within the vma.
1451  *
1452  * Returns 0 if successful.
1453  */
1454 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1455                 unsigned long size)
1456 {
1457         if (address < vma->vm_start || address + size > vma->vm_end ||
1458                         !(vma->vm_flags & VM_PFNMAP))
1459                 return -1;
1460         zap_page_range_single(vma, address, size, NULL);
1461         return 0;
1462 }
1463 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1464
1465 /**
1466  * follow_page_mask - look up a page descriptor from a user-virtual address
1467  * @vma: vm_area_struct mapping @address
1468  * @address: virtual address to look up
1469  * @flags: flags modifying lookup behaviour
1470  * @page_mask: on output, *page_mask is set according to the size of the page
1471  *
1472  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1473  *
1474  * Returns the mapped (struct page *), %NULL if no mapping exists, or
1475  * an error pointer if there is a mapping to something not represented
1476  * by a page descriptor (see also vm_normal_page()).
1477  */
1478 struct page *follow_page_mask(struct vm_area_struct *vma,
1479                               unsigned long address, unsigned int flags,
1480                               unsigned int *page_mask)
1481 {
1482         pgd_t *pgd;
1483         pud_t *pud;
1484         pmd_t *pmd;
1485         pte_t *ptep, pte;
1486         spinlock_t *ptl;
1487         struct page *page;
1488         struct mm_struct *mm;
1489
1490         if (WARN_ON(!vma))
1491                 return NULL;
1492
1493         mm = vma->vm_mm;
1494
1495         *page_mask = 0;
1496
1497         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1498         if (!IS_ERR(page)) {
1499                 BUG_ON(flags & FOLL_GET);
1500                 goto out;
1501         }
1502
1503         page = NULL;
1504         pgd = pgd_offset(mm, address);
1505         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1506                 goto no_page_table;
1507
1508         pud = pud_offset(pgd, address);
1509         if (pud_none(*pud))
1510                 goto no_page_table;
1511         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1512                 BUG_ON(flags & FOLL_GET);
1513                 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1514                 goto out;
1515         }
1516         if (unlikely(pud_bad(*pud)))
1517                 goto no_page_table;
1518
1519         pmd = pmd_offset(pud, address);
1520         if (pmd_none(*pmd))
1521                 goto no_page_table;
1522         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1523                 BUG_ON(flags & FOLL_GET);
1524                 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1525                 goto out;
1526         }
1527         if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1528                 goto no_page_table;
1529         if (pmd_trans_huge(*pmd)) {
1530                 if (flags & FOLL_SPLIT) {
1531                         split_huge_page_pmd(vma, address, pmd);
1532                         goto split_fallthrough;
1533                 }
1534                 spin_lock(&mm->page_table_lock);
1535                 if (likely(pmd_trans_huge(*pmd))) {
1536                         if (unlikely(pmd_trans_splitting(*pmd))) {
1537                                 spin_unlock(&mm->page_table_lock);
1538                                 wait_split_huge_page(vma->anon_vma, pmd);
1539                         } else {
1540                                 page = follow_trans_huge_pmd(vma, address,
1541                                                              pmd, flags);
1542                                 spin_unlock(&mm->page_table_lock);
1543                                 *page_mask = HPAGE_PMD_NR - 1;
1544                                 goto out;
1545                         }
1546                 } else
1547                         spin_unlock(&mm->page_table_lock);
1548                 /* fall through */
1549         }
1550 split_fallthrough:
1551         if (unlikely(pmd_bad(*pmd)))
1552                 goto no_page_table;
1553
1554         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1555
1556         pte = *ptep;
1557         if (!pte_present(pte)) {
1558                 swp_entry_t entry;
1559                 /*
1560                  * KSM's break_ksm() relies upon recognizing a ksm page
1561                  * even while it is being migrated, so for that case we
1562                  * need migration_entry_wait().
1563                  */
1564                 if (likely(!(flags & FOLL_MIGRATION)))
1565                         goto no_page;
1566                 if (pte_none(pte) || pte_file(pte))
1567                         goto no_page;
1568                 entry = pte_to_swp_entry(pte);
1569                 if (!is_migration_entry(entry))
1570                         goto no_page;
1571                 pte_unmap_unlock(ptep, ptl);
1572                 migration_entry_wait(mm, pmd, address);
1573                 goto split_fallthrough;
1574         }
1575         if ((flags & FOLL_NUMA) && pte_numa(pte))
1576                 goto no_page;
1577         if ((flags & FOLL_WRITE) && !pte_write(pte))
1578                 goto unlock;
1579
1580         page = vm_normal_page(vma, address, pte);
1581         if (unlikely(!page)) {
1582                 if ((flags & FOLL_DUMP) ||
1583                     !is_zero_pfn(pte_pfn(pte)))
1584                         goto bad_page;
1585                 page = pte_page(pte);
1586         }
1587
1588         if (flags & FOLL_GET)
1589                 get_page_foll(page);
1590         if (flags & FOLL_TOUCH) {
1591                 if ((flags & FOLL_WRITE) &&
1592                     !pte_dirty(pte) && !PageDirty(page))
1593                         set_page_dirty(page);
1594                 /*
1595                  * pte_mkyoung() would be more correct here, but atomic care
1596                  * is needed to avoid losing the dirty bit: it is easier to use
1597                  * mark_page_accessed().
1598                  */
1599                 mark_page_accessed(page);
1600         }
1601         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1602                 /*
1603                  * The preliminary mapping check is mainly to avoid the
1604                  * pointless overhead of lock_page on the ZERO_PAGE
1605                  * which might bounce very badly if there is contention.
1606                  *
1607                  * If the page is already locked, we don't need to
1608                  * handle it now - vmscan will handle it later if and
1609                  * when it attempts to reclaim the page.
1610                  */
1611                 if (page->mapping && trylock_page(page)) {
1612                         lru_add_drain();  /* push cached pages to LRU */
1613                         /*
1614                          * Because we lock page here, and migration is
1615                          * blocked by the pte's page reference, and we
1616                          * know the page is still mapped, we don't even
1617                          * need to check for file-cache page truncation.
1618                          */
1619                         mlock_vma_page(page);
1620                         unlock_page(page);
1621                 }
1622         }
1623 unlock:
1624         pte_unmap_unlock(ptep, ptl);
1625 out:
1626         return page;
1627
1628 bad_page:
1629         pte_unmap_unlock(ptep, ptl);
1630         return ERR_PTR(-EFAULT);
1631
1632 no_page:
1633         pte_unmap_unlock(ptep, ptl);
1634         if (!pte_none(pte))
1635                 return page;
1636
1637 no_page_table:
1638         /*
1639          * When core dumping an enormous anonymous area that nobody
1640          * has touched so far, we don't want to allocate unnecessary pages or
1641          * page tables.  Return error instead of NULL to skip handle_mm_fault,
1642          * then get_dump_page() will return NULL to leave a hole in the dump.
1643          * But we can only make this optimization where a hole would surely
1644          * be zero-filled if handle_mm_fault() actually did handle it.
1645          */
1646         if ((flags & FOLL_DUMP) &&
1647             (!vma->vm_ops || !vma->vm_ops->fault))
1648                 return ERR_PTR(-EFAULT);
1649         return page;
1650 }
1651
1652 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1653 {
1654         return stack_guard_page_start(vma, addr) ||
1655                stack_guard_page_end(vma, addr+PAGE_SIZE);
1656 }
1657
1658 /**
1659  * replace_cma_page() - migrate page out of CMA page blocks
1660  * @page:       source page to be migrated
1661  *
1662  * Returns either the old page (if migration was not possible) or the pointer
1663  * to the newly allocated page (with additional reference taken).
1664  *
1665  * get_user_pages() might take a reference to a page for a long period of time,
1666  * what prevent such page from migration. This is fatal to the preffered usage
1667  * pattern of CMA pageblocks. This function replaces the given user page with
1668  * a new one allocated from NON-MOVABLE pageblock, so locking CMA page can be
1669  * avoided.
1670  */
1671 static inline struct page *migrate_replace_cma_page(struct page *page)
1672 {
1673         struct page *newpage = alloc_page(GFP_HIGHUSER);
1674
1675         if (!newpage)
1676                 goto out;
1677
1678         /*
1679          * Take additional reference to the new page to ensure it won't get
1680          * freed after migration procedure end.
1681          */
1682         get_page_foll(newpage);
1683
1684         if (migrate_replace_page(page, newpage) == 0)
1685                 return newpage;
1686
1687         put_page(newpage);
1688         __free_page(newpage);
1689 out:
1690         /*
1691          * Migration errors in case of get_user_pages() might not
1692          * be fatal to CMA itself, so better don't fail here.
1693          */
1694         return page;
1695 }
1696
1697 /**
1698  * __get_user_pages() - pin user pages in memory
1699  * @tsk:        task_struct of target task
1700  * @mm:         mm_struct of target mm
1701  * @start:      starting user address
1702  * @nr_pages:   number of pages from start to pin
1703  * @gup_flags:  flags modifying pin behaviour
1704  * @pages:      array that receives pointers to the pages pinned.
1705  *              Should be at least nr_pages long. Or NULL, if caller
1706  *              only intends to ensure the pages are faulted in.
1707  * @vmas:       array of pointers to vmas corresponding to each page.
1708  *              Or NULL if the caller does not require them.
1709  * @nonblocking: whether waiting for disk IO or mmap_sem contention
1710  *
1711  * Returns number of pages pinned. This may be fewer than the number
1712  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1713  * were pinned, returns -errno. Each page returned must be released
1714  * with a put_page() call when it is finished with. vmas will only
1715  * remain valid while mmap_sem is held.
1716  *
1717  * Must be called with mmap_sem held for read or write.
1718  *
1719  * __get_user_pages walks a process's page tables and takes a reference to
1720  * each struct page that each user address corresponds to at a given
1721  * instant. That is, it takes the page that would be accessed if a user
1722  * thread accesses the given user virtual address at that instant.
1723  *
1724  * This does not guarantee that the page exists in the user mappings when
1725  * __get_user_pages returns, and there may even be a completely different
1726  * page there in some cases (eg. if mmapped pagecache has been invalidated
1727  * and subsequently re faulted). However it does guarantee that the page
1728  * won't be freed completely. And mostly callers simply care that the page
1729  * contains data that was valid *at some point in time*. Typically, an IO
1730  * or similar operation cannot guarantee anything stronger anyway because
1731  * locks can't be held over the syscall boundary.
1732  *
1733  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1734  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1735  * appropriate) must be called after the page is finished with, and
1736  * before put_page is called.
1737  *
1738  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1739  * or mmap_sem contention, and if waiting is needed to pin all pages,
1740  * *@nonblocking will be set to 0.
1741  *
1742  * In most cases, get_user_pages or get_user_pages_fast should be used
1743  * instead of __get_user_pages. __get_user_pages should be used only if
1744  * you need some special @gup_flags.
1745  */
1746 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1747                 unsigned long start, unsigned long nr_pages,
1748                 unsigned int gup_flags, struct page **pages,
1749                 struct vm_area_struct **vmas, int *nonblocking)
1750 {
1751         long i;
1752         unsigned long vm_flags;
1753         unsigned int page_mask;
1754
1755         if (!nr_pages)
1756                 return 0;
1757
1758         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1759
1760         /* 
1761          * Require read or write permissions.
1762          * If FOLL_FORCE is set, we only require the "MAY" flags.
1763          */
1764         vm_flags  = (gup_flags & FOLL_WRITE) ?
1765                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1766         vm_flags &= (gup_flags & FOLL_FORCE) ?
1767                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1768
1769         /*
1770          * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1771          * would be called on PROT_NONE ranges. We must never invoke
1772          * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1773          * page faults would unprotect the PROT_NONE ranges if
1774          * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1775          * bitflag. So to avoid that, don't set FOLL_NUMA if
1776          * FOLL_FORCE is set.
1777          */
1778         if (!(gup_flags & FOLL_FORCE))
1779                 gup_flags |= FOLL_NUMA;
1780
1781         i = 0;
1782
1783         do {
1784                 struct vm_area_struct *vma;
1785
1786                 vma = find_extend_vma(mm, start);
1787                 if (!vma && in_gate_area(mm, start)) {
1788                         unsigned long pg = start & PAGE_MASK;
1789                         pgd_t *pgd;
1790                         pud_t *pud;
1791                         pmd_t *pmd;
1792                         pte_t *pte;
1793
1794                         /* user gate pages are read-only */
1795                         if (gup_flags & FOLL_WRITE)
1796                                 return i ? : -EFAULT;
1797                         if (pg > TASK_SIZE)
1798                                 pgd = pgd_offset_k(pg);
1799                         else
1800                                 pgd = pgd_offset_gate(mm, pg);
1801                         BUG_ON(pgd_none(*pgd));
1802                         pud = pud_offset(pgd, pg);
1803                         BUG_ON(pud_none(*pud));
1804                         pmd = pmd_offset(pud, pg);
1805                         if (pmd_none(*pmd))
1806                                 return i ? : -EFAULT;
1807                         VM_BUG_ON(pmd_trans_huge(*pmd));
1808                         pte = pte_offset_map(pmd, pg);
1809                         if (pte_none(*pte)) {
1810                                 pte_unmap(pte);
1811                                 return i ? : -EFAULT;
1812                         }
1813                         vma = get_gate_vma(mm);
1814                         if (pages) {
1815                                 struct page *page;
1816
1817                                 page = vm_normal_page(vma, start, *pte);
1818                                 if (!page) {
1819                                         if (!(gup_flags & FOLL_DUMP) &&
1820                                              is_zero_pfn(pte_pfn(*pte)))
1821                                                 page = pte_page(*pte);
1822                                         else {
1823                                                 pte_unmap(pte);
1824                                                 return i ? : -EFAULT;
1825                                         }
1826                                 }
1827                                 pages[i] = page;
1828                                 get_page(page);
1829                         }
1830                         pte_unmap(pte);
1831                         page_mask = 0;
1832                         goto next_page;
1833                 }
1834
1835                 if (!vma ||
1836                     (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1837                     !(vm_flags & vma->vm_flags))
1838                         return i ? : -EFAULT;
1839
1840                 if (is_vm_hugetlb_page(vma)) {
1841                         i = follow_hugetlb_page(mm, vma, pages, vmas,
1842                                         &start, &nr_pages, i, gup_flags);
1843                         continue;
1844                 }
1845
1846                 do {
1847                         struct page *page;
1848                         unsigned int foll_flags = gup_flags;
1849                         unsigned int page_increm;
1850
1851                         /*
1852                          * If we have a pending SIGKILL, don't keep faulting
1853                          * pages and potentially allocating memory.
1854                          */
1855                         if (unlikely(fatal_signal_pending(current)))
1856                                 return i ? i : -ERESTARTSYS;
1857
1858                         cond_resched();
1859                         while (!(page = follow_page_mask(vma, start,
1860                                                 foll_flags, &page_mask))) {
1861                                 int ret;
1862                                 unsigned int fault_flags = 0;
1863
1864                                 if (gup_flags & FOLL_DURABLE)
1865                                         fault_flags = FAULT_FLAG_NO_CMA;
1866
1867                                 /* For mlock, just skip the stack guard page. */
1868                                 if (foll_flags & FOLL_MLOCK) {
1869                                         if (stack_guard_page(vma, start))
1870                                                 goto next_page;
1871                                 }
1872                                 if (foll_flags & FOLL_WRITE)
1873                                         fault_flags |= FAULT_FLAG_WRITE;
1874                                 if (nonblocking)
1875                                         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1876                                 if (foll_flags & FOLL_NOWAIT)
1877                                         fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1878
1879                                 ret = handle_mm_fault(mm, vma, start,
1880                                                         fault_flags);
1881
1882                                 if (ret & VM_FAULT_ERROR) {
1883                                         if (ret & VM_FAULT_OOM)
1884                                                 return i ? i : -ENOMEM;
1885                                         if (ret & (VM_FAULT_HWPOISON |
1886                                                    VM_FAULT_HWPOISON_LARGE)) {
1887                                                 if (i)
1888                                                         return i;
1889                                                 else if (gup_flags & FOLL_HWPOISON)
1890                                                         return -EHWPOISON;
1891                                                 else
1892                                                         return -EFAULT;
1893                                         }
1894                                         if (ret & VM_FAULT_SIGBUS)
1895                                                 return i ? i : -EFAULT;
1896                                         BUG();
1897                                 }
1898
1899                                 if (tsk) {
1900                                         if (ret & VM_FAULT_MAJOR)
1901                                                 tsk->maj_flt++;
1902                                         else
1903                                                 tsk->min_flt++;
1904                                 }
1905
1906                                 if (ret & VM_FAULT_RETRY) {
1907                                         if (nonblocking)
1908                                                 *nonblocking = 0;
1909                                         return i;
1910                                 }
1911
1912                                 /*
1913                                  * The VM_FAULT_WRITE bit tells us that
1914                                  * do_wp_page has broken COW when necessary,
1915                                  * even if maybe_mkwrite decided not to set
1916                                  * pte_write. We can thus safely do subsequent
1917                                  * page lookups as if they were reads. But only
1918                                  * do so when looping for pte_write is futile:
1919                                  * in some cases userspace may also be wanting
1920                                  * to write to the gotten user page, which a
1921                                  * read fault here might prevent (a readonly
1922                                  * page might get reCOWed by userspace write).
1923                                  */
1924                                 if ((ret & VM_FAULT_WRITE) &&
1925                                     !(vma->vm_flags & VM_WRITE))
1926                                         foll_flags &= ~FOLL_WRITE;
1927
1928                                 cond_resched();
1929                         }
1930                         if (IS_ERR(page))
1931                                 return i ? i : PTR_ERR(page);
1932
1933                         if ((gup_flags & FOLL_DURABLE) && is_cma_page(page))
1934                                 page = migrate_replace_cma_page(page);
1935
1936                         if (pages) {
1937                                 pages[i] = page;
1938
1939                                 flush_anon_page(vma, page, start);
1940                                 flush_dcache_page(page);
1941                                 page_mask = 0;
1942                         }
1943 next_page:
1944                         if (vmas) {
1945                                 vmas[i] = vma;
1946                                 page_mask = 0;
1947                         }
1948                         page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1949                         if (page_increm > nr_pages)
1950                                 page_increm = nr_pages;
1951                         i += page_increm;
1952                         start += page_increm * PAGE_SIZE;
1953                         nr_pages -= page_increm;
1954                 } while (nr_pages && start < vma->vm_end);
1955         } while (nr_pages);
1956         return i;
1957 }
1958 EXPORT_SYMBOL(__get_user_pages);
1959
1960 /*
1961  * fixup_user_fault() - manually resolve a user page fault
1962  * @tsk:        the task_struct to use for page fault accounting, or
1963  *              NULL if faults are not to be recorded.
1964  * @mm:         mm_struct of target mm
1965  * @address:    user address
1966  * @fault_flags:flags to pass down to handle_mm_fault()
1967  *
1968  * This is meant to be called in the specific scenario where for locking reasons
1969  * we try to access user memory in atomic context (within a pagefault_disable()
1970  * section), this returns -EFAULT, and we want to resolve the user fault before
1971  * trying again.
1972  *
1973  * Typically this is meant to be used by the futex code.
1974  *
1975  * The main difference with get_user_pages() is that this function will
1976  * unconditionally call handle_mm_fault() which will in turn perform all the
1977  * necessary SW fixup of the dirty and young bits in the PTE, while
1978  * handle_mm_fault() only guarantees to update these in the struct page.
1979  *
1980  * This is important for some architectures where those bits also gate the
1981  * access permission to the page because they are maintained in software.  On
1982  * such architectures, gup() will not be enough to make a subsequent access
1983  * succeed.
1984  *
1985  * This should be called with the mm_sem held for read.
1986  */
1987 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1988                      unsigned long address, unsigned int fault_flags)
1989 {
1990         struct vm_area_struct *vma;
1991         int ret;
1992
1993         vma = find_extend_vma(mm, address);
1994         if (!vma || address < vma->vm_start)
1995                 return -EFAULT;
1996
1997         ret = handle_mm_fault(mm, vma, address, fault_flags);
1998         if (ret & VM_FAULT_ERROR) {
1999                 if (ret & VM_FAULT_OOM)
2000                         return -ENOMEM;
2001                 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
2002                         return -EHWPOISON;
2003                 if (ret & VM_FAULT_SIGBUS)
2004                         return -EFAULT;
2005                 BUG();
2006         }
2007         if (tsk) {
2008                 if (ret & VM_FAULT_MAJOR)
2009                         tsk->maj_flt++;
2010                 else
2011                         tsk->min_flt++;
2012         }
2013         return 0;
2014 }
2015
2016 /**
2017  * get_dump_page() - pin user page in memory while writing it to core dump
2018  * @addr: user address
2019  *
2020  * Returns struct page pointer of user page pinned for dump,
2021  * to be freed afterwards by page_cache_release() or put_page().
2022  *
2023  * Returns NULL on any kind of failure - a hole must then be inserted into
2024  * the corefile, to preserve alignment with its headers; and also returns
2025  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2026  * allowing a hole to be left in the corefile to save diskspace.
2027  *
2028  * Called without mmap_sem, but after all other threads have been killed.
2029  */
2030 #ifdef CONFIG_ELF_CORE
2031 struct page *get_dump_page(unsigned long addr)
2032 {
2033         struct vm_area_struct *vma;
2034         struct page *page;
2035
2036         if (__get_user_pages(current, current->mm, addr, 1,
2037                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2038                              NULL) < 1)
2039                 return NULL;
2040         flush_cache_page(vma, addr, page_to_pfn(page));
2041         return page;
2042 }
2043 #endif /* CONFIG_ELF_CORE */
2044
2045 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2046                         spinlock_t **ptl)
2047 {
2048         pgd_t * pgd = pgd_offset(mm, addr);
2049         pud_t * pud = pud_alloc(mm, pgd, addr);
2050         if (pud) {
2051                 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2052                 if (pmd) {
2053                         VM_BUG_ON(pmd_trans_huge(*pmd));
2054                         return pte_alloc_map_lock(mm, pmd, addr, ptl);
2055                 }
2056         }
2057         return NULL;
2058 }
2059
2060 /*
2061  * This is the old fallback for page remapping.
2062  *
2063  * For historical reasons, it only allows reserved pages. Only
2064  * old drivers should use this, and they needed to mark their
2065  * pages reserved for the old functions anyway.
2066  */
2067 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2068                         struct page *page, pgprot_t prot)
2069 {
2070         struct mm_struct *mm = vma->vm_mm;
2071         int retval;
2072         pte_t *pte;
2073         spinlock_t *ptl;
2074
2075         retval = -EINVAL;
2076         if (PageAnon(page))
2077                 goto out;
2078         retval = -ENOMEM;
2079         flush_dcache_page(page);
2080         pte = get_locked_pte(mm, addr, &ptl);
2081         if (!pte)
2082                 goto out;
2083         retval = -EBUSY;
2084         if (!pte_none(*pte))
2085                 goto out_unlock;
2086
2087         /* Ok, finally just insert the thing.. */
2088         get_page(page);
2089         inc_mm_counter_fast(mm, MM_FILEPAGES);
2090         page_add_file_rmap(page);
2091         set_pte_at(mm, addr, pte, mk_pte(page, prot));
2092
2093         retval = 0;
2094         pte_unmap_unlock(pte, ptl);
2095         return retval;
2096 out_unlock:
2097         pte_unmap_unlock(pte, ptl);
2098 out:
2099         return retval;
2100 }
2101
2102 /**
2103  * vm_insert_page - insert single page into user vma
2104  * @vma: user vma to map to
2105  * @addr: target user address of this page
2106  * @page: source kernel page
2107  *
2108  * This allows drivers to insert individual pages they've allocated
2109  * into a user vma.
2110  *
2111  * The page has to be a nice clean _individual_ kernel allocation.
2112  * If you allocate a compound page, you need to have marked it as
2113  * such (__GFP_COMP), or manually just split the page up yourself
2114  * (see split_page()).
2115  *
2116  * NOTE! Traditionally this was done with "remap_pfn_range()" which
2117  * took an arbitrary page protection parameter. This doesn't allow
2118  * that. Your vma protection will have to be set up correctly, which
2119  * means that if you want a shared writable mapping, you'd better
2120  * ask for a shared writable mapping!
2121  *
2122  * The page does not need to be reserved.
2123  *
2124  * Usually this function is called from f_op->mmap() handler
2125  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2126  * Caller must set VM_MIXEDMAP on vma if it wants to call this
2127  * function from other places, for example from page-fault handler.
2128  */
2129 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2130                         struct page *page)
2131 {
2132         if (addr < vma->vm_start || addr >= vma->vm_end)
2133                 return -EFAULT;
2134         if (!page_count(page))
2135                 return -EINVAL;
2136         if (!(vma->vm_flags & VM_MIXEDMAP)) {
2137                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2138                 BUG_ON(vma->vm_flags & VM_PFNMAP);
2139                 vma->vm_flags |= VM_MIXEDMAP;
2140         }
2141         return insert_page(vma, addr, page, vma->vm_page_prot);
2142 }
2143 EXPORT_SYMBOL(vm_insert_page);
2144
2145 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2146                         unsigned long pfn, pgprot_t prot)
2147 {
2148         struct mm_struct *mm = vma->vm_mm;
2149         int retval;
2150         pte_t *pte, entry;
2151         spinlock_t *ptl;
2152
2153         retval = -ENOMEM;
2154         pte = get_locked_pte(mm, addr, &ptl);
2155         if (!pte)
2156                 goto out;
2157         retval = -EBUSY;
2158         if (!pte_none(*pte))
2159                 goto out_unlock;
2160
2161         /* Ok, finally just insert the thing.. */
2162         entry = pte_mkspecial(pfn_pte(pfn, prot));
2163         set_pte_at(mm, addr, pte, entry);
2164         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2165
2166         retval = 0;
2167 out_unlock:
2168         pte_unmap_unlock(pte, ptl);
2169 out:
2170         return retval;
2171 }
2172
2173 /**
2174  * vm_insert_pfn - insert single pfn into user vma
2175  * @vma: user vma to map to
2176  * @addr: target user address of this page
2177  * @pfn: source kernel pfn
2178  *
2179  * Similar to vm_insert_page, this allows drivers to insert individual pages
2180  * they've allocated into a user vma. Same comments apply.
2181  *
2182  * This function should only be called from a vm_ops->fault handler, and
2183  * in that case the handler should return NULL.
2184  *
2185  * vma cannot be a COW mapping.
2186  *
2187  * As this is called only for pages that do not currently exist, we
2188  * do not need to flush old virtual caches or the TLB.
2189  */
2190 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2191                         unsigned long pfn)
2192 {
2193         int ret;
2194         pgprot_t pgprot = vma->vm_page_prot;
2195         /*
2196          * Technically, architectures with pte_special can avoid all these
2197          * restrictions (same for remap_pfn_range).  However we would like
2198          * consistency in testing and feature parity among all, so we should
2199          * try to keep these invariants in place for everybody.
2200          */
2201         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2202         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2203                                                 (VM_PFNMAP|VM_MIXEDMAP));
2204         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2205         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2206
2207         if (addr < vma->vm_start || addr >= vma->vm_end)
2208                 return -EFAULT;
2209         if (track_pfn_insert(vma, &pgprot, pfn))
2210                 return -EINVAL;
2211
2212         ret = insert_pfn(vma, addr, pfn, pgprot);
2213
2214         return ret;
2215 }
2216 EXPORT_SYMBOL(vm_insert_pfn);
2217
2218 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2219                         unsigned long pfn)
2220 {
2221         BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2222
2223         if (addr < vma->vm_start || addr >= vma->vm_end)
2224                 return -EFAULT;
2225
2226         /*
2227          * If we don't have pte special, then we have to use the pfn_valid()
2228          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2229          * refcount the page if pfn_valid is true (hence insert_page rather
2230          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2231          * without pte special, it would there be refcounted as a normal page.
2232          */
2233         if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2234                 struct page *page;
2235
2236                 page = pfn_to_page(pfn);
2237                 return insert_page(vma, addr, page, vma->vm_page_prot);
2238         }
2239         return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2240 }
2241 EXPORT_SYMBOL(vm_insert_mixed);
2242
2243 /*
2244  * maps a range of physical memory into the requested pages. the old
2245  * mappings are removed. any references to nonexistent pages results
2246  * in null mappings (currently treated as "copy-on-access")
2247  */
2248 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2249                         unsigned long addr, unsigned long end,
2250                         unsigned long pfn, pgprot_t prot)
2251 {
2252         pte_t *pte;
2253         spinlock_t *ptl;
2254
2255         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2256         if (!pte)
2257                 return -ENOMEM;
2258         arch_enter_lazy_mmu_mode();
2259         do {
2260                 BUG_ON(!pte_none(*pte));
2261                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2262                 pfn++;
2263         } while (pte++, addr += PAGE_SIZE, addr != end);
2264         arch_leave_lazy_mmu_mode();
2265         pte_unmap_unlock(pte - 1, ptl);
2266         return 0;
2267 }
2268
2269 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2270                         unsigned long addr, unsigned long end,
2271                         unsigned long pfn, pgprot_t prot)
2272 {
2273         pmd_t *pmd;
2274         unsigned long next;
2275
2276         pfn -= addr >> PAGE_SHIFT;
2277         pmd = pmd_alloc(mm, pud, addr);
2278         if (!pmd)
2279                 return -ENOMEM;
2280         VM_BUG_ON(pmd_trans_huge(*pmd));
2281         do {
2282                 next = pmd_addr_end(addr, end);
2283                 if (remap_pte_range(mm, pmd, addr, next,
2284                                 pfn + (addr >> PAGE_SHIFT), prot))
2285                         return -ENOMEM;
2286         } while (pmd++, addr = next, addr != end);
2287         return 0;
2288 }
2289
2290 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2291                         unsigned long addr, unsigned long end,
2292                         unsigned long pfn, pgprot_t prot)
2293 {
2294         pud_t *pud;
2295         unsigned long next;
2296
2297         pfn -= addr >> PAGE_SHIFT;
2298         pud = pud_alloc(mm, pgd, addr);
2299         if (!pud)
2300                 return -ENOMEM;
2301         do {
2302                 next = pud_addr_end(addr, end);
2303                 if (remap_pmd_range(mm, pud, addr, next,
2304                                 pfn + (addr >> PAGE_SHIFT), prot))
2305                         return -ENOMEM;
2306         } while (pud++, addr = next, addr != end);
2307         return 0;
2308 }
2309
2310 /**
2311  * remap_pfn_range - remap kernel memory to userspace
2312  * @vma: user vma to map to
2313  * @addr: target user address to start at
2314  * @pfn: physical address of kernel memory
2315  * @size: size of map area
2316  * @prot: page protection flags for this mapping
2317  *
2318  *  Note: this is only safe if the mm semaphore is held when called.
2319  */
2320 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2321                     unsigned long pfn, unsigned long size, pgprot_t prot)
2322 {
2323         pgd_t *pgd;
2324         unsigned long next;
2325         unsigned long end = addr + PAGE_ALIGN(size);
2326         struct mm_struct *mm = vma->vm_mm;
2327         int err;
2328
2329         /*
2330          * Physically remapped pages are special. Tell the
2331          * rest of the world about it:
2332          *   VM_IO tells people not to look at these pages
2333          *      (accesses can have side effects).
2334          *   VM_PFNMAP tells the core MM that the base pages are just
2335          *      raw PFN mappings, and do not have a "struct page" associated
2336          *      with them.
2337          *   VM_DONTEXPAND
2338          *      Disable vma merging and expanding with mremap().
2339          *   VM_DONTDUMP
2340          *      Omit vma from core dump, even when VM_IO turned off.
2341          *
2342          * There's a horrible special case to handle copy-on-write
2343          * behaviour that some programs depend on. We mark the "original"
2344          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2345          * See vm_normal_page() for details.
2346          */
2347         if (is_cow_mapping(vma->vm_flags)) {
2348                 if (addr != vma->vm_start || end != vma->vm_end)
2349                         return -EINVAL;
2350                 vma->vm_pgoff = pfn;
2351         }
2352
2353         err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2354         if (err)
2355                 return -EINVAL;
2356
2357         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2358
2359         BUG_ON(addr >= end);
2360         pfn -= addr >> PAGE_SHIFT;
2361         pgd = pgd_offset(mm, addr);
2362         flush_cache_range(vma, addr, end);
2363         do {
2364                 next = pgd_addr_end(addr, end);
2365                 err = remap_pud_range(mm, pgd, addr, next,
2366                                 pfn + (addr >> PAGE_SHIFT), prot);
2367                 if (err)
2368                         break;
2369         } while (pgd++, addr = next, addr != end);
2370
2371         if (err)
2372                 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2373
2374         return err;
2375 }
2376 EXPORT_SYMBOL(remap_pfn_range);
2377
2378 /**
2379  * vm_iomap_memory - remap memory to userspace
2380  * @vma: user vma to map to
2381  * @start: start of area
2382  * @len: size of area
2383  *
2384  * This is a simplified io_remap_pfn_range() for common driver use. The
2385  * driver just needs to give us the physical memory range to be mapped,
2386  * we'll figure out the rest from the vma information.
2387  *
2388  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2389  * whatever write-combining details or similar.
2390  */
2391 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2392 {
2393         unsigned long vm_len, pfn, pages;
2394
2395         /* Check that the physical memory area passed in looks valid */
2396         if (start + len < start)
2397                 return -EINVAL;
2398         /*
2399          * You *really* shouldn't map things that aren't page-aligned,
2400          * but we've historically allowed it because IO memory might
2401          * just have smaller alignment.
2402          */
2403         len += start & ~PAGE_MASK;
2404         pfn = start >> PAGE_SHIFT;
2405         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2406         if (pfn + pages < pfn)
2407                 return -EINVAL;
2408
2409         /* We start the mapping 'vm_pgoff' pages into the area */
2410         if (vma->vm_pgoff > pages)
2411                 return -EINVAL;
2412         pfn += vma->vm_pgoff;
2413         pages -= vma->vm_pgoff;
2414
2415         /* Can we fit all of the mapping? */
2416         vm_len = vma->vm_end - vma->vm_start;
2417         if (vm_len >> PAGE_SHIFT > pages)
2418                 return -EINVAL;
2419
2420         /* Ok, let it rip */
2421         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2422 }
2423 EXPORT_SYMBOL(vm_iomap_memory);
2424
2425 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2426                                      unsigned long addr, unsigned long end,
2427                                      pte_fn_t fn, void *data)
2428 {
2429         pte_t *pte;
2430         int err;
2431         pgtable_t token;
2432         spinlock_t *uninitialized_var(ptl);
2433
2434         pte = (mm == &init_mm) ?
2435                 pte_alloc_kernel(pmd, addr) :
2436                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2437         if (!pte)
2438                 return -ENOMEM;
2439
2440         BUG_ON(pmd_huge(*pmd));
2441
2442         arch_enter_lazy_mmu_mode();
2443
2444         token = pmd_pgtable(*pmd);
2445
2446         do {
2447                 err = fn(pte++, token, addr, data);
2448                 if (err)
2449                         break;
2450         } while (addr += PAGE_SIZE, addr != end);
2451
2452         arch_leave_lazy_mmu_mode();
2453
2454         if (mm != &init_mm)
2455                 pte_unmap_unlock(pte-1, ptl);
2456         return err;
2457 }
2458
2459 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2460                                      unsigned long addr, unsigned long end,
2461                                      pte_fn_t fn, void *data)
2462 {
2463         pmd_t *pmd;
2464         unsigned long next;
2465         int err;
2466
2467         BUG_ON(pud_huge(*pud));
2468
2469         pmd = pmd_alloc(mm, pud, addr);
2470         if (!pmd)
2471                 return -ENOMEM;
2472         do {
2473                 next = pmd_addr_end(addr, end);
2474                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2475                 if (err)
2476                         break;
2477         } while (pmd++, addr = next, addr != end);
2478         return err;
2479 }
2480
2481 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2482                                      unsigned long addr, unsigned long end,
2483                                      pte_fn_t fn, void *data)
2484 {
2485         pud_t *pud;
2486         unsigned long next;
2487         int err;
2488
2489         pud = pud_alloc(mm, pgd, addr);
2490         if (!pud)
2491                 return -ENOMEM;
2492         do {
2493                 next = pud_addr_end(addr, end);
2494                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2495                 if (err)
2496                         break;
2497         } while (pud++, addr = next, addr != end);
2498         return err;
2499 }
2500
2501 /*
2502  * Scan a region of virtual memory, filling in page tables as necessary
2503  * and calling a provided function on each leaf page table.
2504  */
2505 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2506                         unsigned long size, pte_fn_t fn, void *data)
2507 {
2508         pgd_t *pgd;
2509         unsigned long next;
2510         unsigned long end = addr + size;
2511         int err;
2512
2513         BUG_ON(addr >= end);
2514         pgd = pgd_offset(mm, addr);
2515         do {
2516                 next = pgd_addr_end(addr, end);
2517                 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2518                 if (err)
2519                         break;
2520         } while (pgd++, addr = next, addr != end);
2521
2522         return err;
2523 }
2524 EXPORT_SYMBOL_GPL(apply_to_page_range);
2525
2526 /*
2527  * handle_pte_fault chooses page fault handler according to an entry
2528  * which was read non-atomically.  Before making any commitment, on
2529  * those architectures or configurations (e.g. i386 with PAE) which
2530  * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2531  * must check under lock before unmapping the pte and proceeding
2532  * (but do_wp_page is only called after already making such a check;
2533  * and do_anonymous_page can safely check later on).
2534  */
2535 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2536                                 pte_t *page_table, pte_t orig_pte)
2537 {
2538         int same = 1;
2539 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2540         if (sizeof(pte_t) > sizeof(unsigned long)) {
2541                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2542                 spin_lock(ptl);
2543                 same = pte_same(*page_table, orig_pte);
2544                 spin_unlock(ptl);
2545         }
2546 #endif
2547         pte_unmap(page_table);
2548         return same;
2549 }
2550
2551 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2552 {
2553         /*
2554          * If the source page was a PFN mapping, we don't have
2555          * a "struct page" for it. We do a best-effort copy by
2556          * just copying from the original user address. If that
2557          * fails, we just zero-fill it. Live with it.
2558          */
2559         if (unlikely(!src)) {
2560                 void *kaddr = kmap_atomic(dst);
2561                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2562
2563                 /*
2564                  * This really shouldn't fail, because the page is there
2565                  * in the page tables. But it might just be unreadable,
2566                  * in which case we just give up and fill the result with
2567                  * zeroes.
2568                  */
2569                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2570                         clear_page(kaddr);
2571                 kunmap_atomic(kaddr);
2572                 flush_dcache_page(dst);
2573         } else
2574                 copy_user_highpage(dst, src, va, vma);
2575 }
2576
2577 /*
2578  * This routine handles present pages, when users try to write
2579  * to a shared page. It is done by copying the page to a new address
2580  * and decrementing the shared-page counter for the old page.
2581  *
2582  * Note that this routine assumes that the protection checks have been
2583  * done by the caller (the low-level page fault routine in most cases).
2584  * Thus we can safely just mark it writable once we've done any necessary
2585  * COW.
2586  *
2587  * We also mark the page dirty at this point even though the page will
2588  * change only once the write actually happens. This avoids a few races,
2589  * and potentially makes it more efficient.
2590  *
2591  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2592  * but allow concurrent faults), with pte both mapped and locked.
2593  * We return with mmap_sem still held, but pte unmapped and unlocked.
2594  */
2595 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2596                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2597                 spinlock_t *ptl, pte_t orig_pte, unsigned int flags)
2598         __releases(ptl)
2599 {
2600         struct page *old_page, *new_page = NULL;
2601         pte_t entry;
2602         int ret = 0;
2603         int page_mkwrite = 0;
2604         struct page *dirty_page = NULL;
2605         unsigned long mmun_start = 0;   /* For mmu_notifiers */
2606         unsigned long mmun_end = 0;     /* For mmu_notifiers */
2607         gfp_t gfp = GFP_HIGHUSER_MOVABLE;
2608
2609         if (IS_ENABLED(CONFIG_CMA) && (flags & FAULT_FLAG_NO_CMA))
2610                 gfp &= ~__GFP_MOVABLE;
2611
2612         old_page = vm_normal_page(vma, address, orig_pte);
2613         if (!old_page) {
2614                 /*
2615                  * VM_MIXEDMAP !pfn_valid() case
2616                  *
2617                  * We should not cow pages in a shared writeable mapping.
2618                  * Just mark the pages writable as we can't do any dirty
2619                  * accounting on raw pfn maps.
2620                  */
2621                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2622                                      (VM_WRITE|VM_SHARED))
2623                         goto reuse;
2624                 goto gotten;
2625         }
2626
2627         /*
2628          * Take out anonymous pages first, anonymous shared vmas are
2629          * not dirty accountable.
2630          */
2631         if (PageAnon(old_page) && !PageKsm(old_page)) {
2632                 if (!trylock_page(old_page)) {
2633                         page_cache_get(old_page);
2634                         pte_unmap_unlock(page_table, ptl);
2635                         lock_page(old_page);
2636                         page_table = pte_offset_map_lock(mm, pmd, address,
2637                                                          &ptl);
2638                         if (!pte_same(*page_table, orig_pte)) {
2639                                 unlock_page(old_page);
2640                                 goto unlock;
2641                         }
2642                         page_cache_release(old_page);
2643                 }
2644                 if (reuse_swap_page(old_page)) {
2645                         /*
2646                          * The page is all ours.  Move it to our anon_vma so
2647                          * the rmap code will not search our parent or siblings.
2648                          * Protected against the rmap code by the page lock.
2649                          */
2650                         page_move_anon_rmap(old_page, vma, address);
2651                         unlock_page(old_page);
2652                         goto reuse;
2653                 }
2654                 unlock_page(old_page);
2655         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2656                                         (VM_WRITE|VM_SHARED))) {
2657                 /*
2658                  * Only catch write-faults on shared writable pages,
2659                  * read-only shared pages can get COWed by
2660                  * get_user_pages(.write=1, .force=1).
2661                  */
2662                 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2663                         struct vm_fault vmf;
2664                         int tmp;
2665
2666                         vmf.virtual_address = (void __user *)(address &
2667                                                                 PAGE_MASK);
2668                         vmf.pgoff = old_page->index;
2669                         vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2670                         vmf.page = old_page;
2671
2672                         /*
2673                          * Notify the address space that the page is about to
2674                          * become writable so that it can prohibit this or wait
2675                          * for the page to get into an appropriate state.
2676                          *
2677                          * We do this without the lock held, so that it can
2678                          * sleep if it needs to.
2679                          */
2680                         page_cache_get(old_page);
2681                         pte_unmap_unlock(page_table, ptl);
2682
2683                         tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2684                         if (unlikely(tmp &
2685                                         (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2686                                 ret = tmp;
2687                                 goto unwritable_page;
2688                         }
2689                         if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2690                                 lock_page(old_page);
2691                                 if (!old_page->mapping) {
2692                                         ret = 0; /* retry the fault */
2693                                         unlock_page(old_page);
2694                                         goto unwritable_page;
2695                                 }
2696                         } else
2697                                 VM_BUG_ON(!PageLocked(old_page));
2698
2699                         /*
2700                          * Since we dropped the lock we need to revalidate
2701                          * the PTE as someone else may have changed it.  If
2702                          * they did, we just return, as we can count on the
2703                          * MMU to tell us if they didn't also make it writable.
2704                          */
2705                         page_table = pte_offset_map_lock(mm, pmd, address,
2706                                                          &ptl);
2707                         if (!pte_same(*page_table, orig_pte)) {
2708                                 unlock_page(old_page);
2709                                 goto unlock;
2710                         }
2711
2712                         page_mkwrite = 1;
2713                 }
2714                 dirty_page = old_page;
2715                 get_page(dirty_page);
2716
2717 reuse:
2718                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2719                 entry = pte_mkyoung(orig_pte);
2720                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2721                 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2722                         update_mmu_cache(vma, address, page_table);
2723                 pte_unmap_unlock(page_table, ptl);
2724                 ret |= VM_FAULT_WRITE;
2725
2726                 if (!dirty_page)
2727                         return ret;
2728
2729                 /*
2730                  * Yes, Virginia, this is actually required to prevent a race
2731                  * with clear_page_dirty_for_io() from clearing the page dirty
2732                  * bit after it clear all dirty ptes, but before a racing
2733                  * do_wp_page installs a dirty pte.
2734                  *
2735                  * __do_fault is protected similarly.
2736                  */
2737                 if (!page_mkwrite) {
2738                         wait_on_page_locked(dirty_page);
2739                         set_page_dirty_balance(dirty_page, page_mkwrite);
2740                         /* file_update_time outside page_lock */
2741                         if (vma->vm_file)
2742                                 file_update_time(vma->vm_file);
2743                 }
2744                 put_page(dirty_page);
2745                 if (page_mkwrite) {
2746                         struct address_space *mapping = dirty_page->mapping;
2747
2748                         set_page_dirty(dirty_page);
2749                         unlock_page(dirty_page);
2750                         page_cache_release(dirty_page);
2751                         if (mapping)    {
2752                                 /*
2753                                  * Some device drivers do not set page.mapping
2754                                  * but still dirty their pages
2755                                  */
2756                                 balance_dirty_pages_ratelimited(mapping);
2757                         }
2758                 }
2759
2760                 return ret;
2761         }
2762
2763         /*
2764          * Ok, we need to copy. Oh, well..
2765          */
2766         page_cache_get(old_page);
2767 gotten:
2768         pte_unmap_unlock(page_table, ptl);
2769
2770         if (unlikely(anon_vma_prepare(vma)))
2771                 goto oom;
2772
2773         if (is_zero_pfn(pte_pfn(orig_pte))) {
2774                 new_page = alloc_zeroed_user_highpage(gfp, vma, address);
2775                 if (!new_page)
2776                         goto oom;
2777         } else {
2778                 new_page = alloc_page_vma(gfp, vma, address);
2779                 if (!new_page)
2780                         goto oom;
2781                 cow_user_page(new_page, old_page, address, vma);
2782         }
2783         __SetPageUptodate(new_page);
2784
2785         if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2786                 goto oom_free_new;
2787
2788         mmun_start  = address & PAGE_MASK;
2789         mmun_end    = mmun_start + PAGE_SIZE;
2790         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2791
2792         /*
2793          * Re-check the pte - we dropped the lock
2794          */
2795         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2796         if (likely(pte_same(*page_table, orig_pte))) {
2797                 if (old_page) {
2798                         if (!PageAnon(old_page)) {
2799                                 dec_mm_counter_fast(mm, MM_FILEPAGES);
2800                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2801                         }
2802                 } else
2803                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2804                 flush_cache_page(vma, address, pte_pfn(orig_pte));
2805                 entry = mk_pte(new_page, vma->vm_page_prot);
2806                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2807                 /*
2808                  * Clear the pte entry and flush it first, before updating the
2809                  * pte with the new entry. This will avoid a race condition
2810                  * seen in the presence of one thread doing SMC and another
2811                  * thread doing COW.
2812                  */
2813                 ptep_clear_flush(vma, address, page_table);
2814                 page_add_new_anon_rmap(new_page, vma, address);
2815                 /*
2816                  * We call the notify macro here because, when using secondary
2817                  * mmu page tables (such as kvm shadow page tables), we want the
2818                  * new page to be mapped directly into the secondary page table.
2819                  */
2820                 set_pte_at_notify(mm, address, page_table, entry);
2821                 update_mmu_cache(vma, address, page_table);
2822                 if (old_page) {
2823                         /*
2824                          * Only after switching the pte to the new page may
2825                          * we remove the mapcount here. Otherwise another
2826                          * process may come and find the rmap count decremented
2827                          * before the pte is switched to the new page, and
2828                          * "reuse" the old page writing into it while our pte
2829                          * here still points into it and can be read by other
2830                          * threads.
2831                          *
2832                          * The critical issue is to order this
2833                          * page_remove_rmap with the ptp_clear_flush above.
2834                          * Those stores are ordered by (if nothing else,)
2835                          * the barrier present in the atomic_add_negative
2836                          * in page_remove_rmap.
2837                          *
2838                          * Then the TLB flush in ptep_clear_flush ensures that
2839                          * no process can access the old page before the
2840                          * decremented mapcount is visible. And the old page
2841                          * cannot be reused until after the decremented
2842                          * mapcount is visible. So transitively, TLBs to
2843                          * old page will be flushed before it can be reused.
2844                          */
2845                         page_remove_rmap(old_page);
2846                 }
2847
2848                 /* Free the old page.. */
2849                 new_page = old_page;
2850                 ret |= VM_FAULT_WRITE;
2851         } else
2852                 mem_cgroup_uncharge_page(new_page);
2853
2854         if (new_page)
2855                 page_cache_release(new_page);
2856 unlock:
2857         pte_unmap_unlock(page_table, ptl);
2858         if (mmun_end > mmun_start)
2859                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2860         if (old_page) {
2861                 /*
2862                  * Don't let another task, with possibly unlocked vma,
2863                  * keep the mlocked page.
2864                  */
2865                 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2866                         lock_page(old_page);    /* LRU manipulation */
2867                         munlock_vma_page(old_page);
2868                         unlock_page(old_page);
2869                 }
2870                 page_cache_release(old_page);
2871         }
2872         return ret;
2873 oom_free_new:
2874         page_cache_release(new_page);
2875 oom:
2876         if (old_page)
2877                 page_cache_release(old_page);
2878         return VM_FAULT_OOM;
2879
2880 unwritable_page:
2881         page_cache_release(old_page);
2882         return ret;
2883 }
2884
2885 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2886                 unsigned long start_addr, unsigned long end_addr,
2887                 struct zap_details *details)
2888 {
2889         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2890 }
2891
2892 static inline void unmap_mapping_range_tree(struct rb_root *root,
2893                                             struct zap_details *details)
2894 {
2895         struct vm_area_struct *vma;
2896         pgoff_t vba, vea, zba, zea;
2897
2898         vma_interval_tree_foreach(vma, root,
2899                         details->first_index, details->last_index) {
2900
2901                 vba = vma->vm_pgoff;
2902                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2903                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2904                 zba = details->first_index;
2905                 if (zba < vba)
2906                         zba = vba;
2907                 zea = details->last_index;
2908                 if (zea > vea)
2909                         zea = vea;
2910
2911                 unmap_mapping_range_vma(vma,
2912                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2913                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2914                                 details);
2915         }
2916 }
2917
2918 static inline void unmap_mapping_range_list(struct list_head *head,
2919                                             struct zap_details *details)
2920 {
2921         struct vm_area_struct *vma;
2922
2923         /*
2924          * In nonlinear VMAs there is no correspondence between virtual address
2925          * offset and file offset.  So we must perform an exhaustive search
2926          * across *all* the pages in each nonlinear VMA, not just the pages
2927          * whose virtual address lies outside the file truncation point.
2928          */
2929         list_for_each_entry(vma, head, shared.nonlinear) {
2930                 details->nonlinear_vma = vma;
2931                 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2932         }
2933 }
2934
2935 /**
2936  * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2937  * @mapping: the address space containing mmaps to be unmapped.
2938  * @holebegin: byte in first page to unmap, relative to the start of
2939  * the underlying file.  This will be rounded down to a PAGE_SIZE
2940  * boundary.  Note that this is different from truncate_pagecache(), which
2941  * must keep the partial page.  In contrast, we must get rid of
2942  * partial pages.
2943  * @holelen: size of prospective hole in bytes.  This will be rounded
2944  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2945  * end of the file.
2946  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2947  * but 0 when invalidating pagecache, don't throw away private data.
2948  */
2949 void unmap_mapping_range(struct address_space *mapping,
2950                 loff_t const holebegin, loff_t const holelen, int even_cows)
2951 {
2952         struct zap_details details;
2953         pgoff_t hba = holebegin >> PAGE_SHIFT;
2954         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2955
2956         /* Check for overflow. */
2957         if (sizeof(holelen) > sizeof(hlen)) {
2958                 long long holeend =
2959                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2960                 if (holeend & ~(long long)ULONG_MAX)
2961                         hlen = ULONG_MAX - hba + 1;
2962         }
2963
2964         details.check_mapping = even_cows? NULL: mapping;
2965         details.nonlinear_vma = NULL;
2966         details.first_index = hba;
2967         details.last_index = hba + hlen - 1;
2968         if (details.last_index < details.first_index)
2969                 details.last_index = ULONG_MAX;
2970
2971
2972         mutex_lock(&mapping->i_mmap_mutex);
2973         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2974                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2975         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2976                 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2977         mutex_unlock(&mapping->i_mmap_mutex);
2978 }
2979 EXPORT_SYMBOL(unmap_mapping_range);
2980
2981 /*
2982  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2983  * but allow concurrent faults), and pte mapped but not yet locked.
2984  * We return with mmap_sem still held, but pte unmapped and unlocked.
2985  */
2986 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2987                 unsigned long address, pte_t *page_table, pmd_t *pmd,
2988                 unsigned int flags, pte_t orig_pte)
2989 {
2990         spinlock_t *ptl;
2991         struct page *page, *swapcache;
2992         swp_entry_t entry;
2993         pte_t pte;
2994         int locked;
2995         struct mem_cgroup *ptr;
2996         int exclusive = 0;
2997         int ret = 0;
2998
2999         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3000                 goto out;
3001
3002         entry = pte_to_swp_entry(orig_pte);
3003         if (unlikely(non_swap_entry(entry))) {
3004                 if (is_migration_entry(entry)) {
3005                         migration_entry_wait(mm, pmd, address);
3006                 } else if (is_hwpoison_entry(entry)) {
3007                         ret = VM_FAULT_HWPOISON;
3008                 } else {
3009                         print_bad_pte(vma, address, orig_pte, NULL);
3010                         ret = VM_FAULT_SIGBUS;
3011                 }
3012                 goto out;
3013         }
3014         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3015         page = lookup_swap_cache(entry);
3016         if (!page) {
3017                 page = swapin_readahead(entry,
3018                                         GFP_HIGHUSER_MOVABLE, vma, address);
3019                 if (!page) {
3020                         /*
3021                          * Back out if somebody else faulted in this pte
3022                          * while we released the pte lock.
3023                          */
3024                         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3025                         if (likely(pte_same(*page_table, orig_pte)))
3026                                 ret = VM_FAULT_OOM;
3027                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3028                         goto unlock;
3029                 }
3030
3031                 /* Had to read the page from swap area: Major fault */
3032                 ret = VM_FAULT_MAJOR;
3033                 count_vm_event(PGMAJFAULT);
3034                 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3035         } else if (PageHWPoison(page)) {
3036                 /*
3037                  * hwpoisoned dirty swapcache pages are kept for killing
3038                  * owner processes (which may be unknown at hwpoison time)
3039                  */
3040                 ret = VM_FAULT_HWPOISON;
3041                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3042                 swapcache = page;
3043                 goto out_release;
3044         }
3045
3046         swapcache = page;
3047         locked = lock_page_or_retry(page, mm, flags);
3048
3049         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3050         if (!locked) {
3051                 ret |= VM_FAULT_RETRY;
3052                 goto out_release;
3053         }
3054
3055         /*
3056          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3057          * release the swapcache from under us.  The page pin, and pte_same
3058          * test below, are not enough to exclude that.  Even if it is still
3059          * swapcache, we need to check that the page's swap has not changed.
3060          */
3061         if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3062                 goto out_page;
3063
3064         page = ksm_might_need_to_copy(page, vma, address);
3065         if (unlikely(!page)) {
3066                 ret = VM_FAULT_OOM;
3067                 page = swapcache;
3068                 goto out_page;
3069         }
3070
3071         if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3072                 ret = VM_FAULT_OOM;
3073                 goto out_page;
3074         }
3075
3076         /*
3077          * Back out if somebody else already faulted in this pte.
3078          */
3079         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3080         if (unlikely(!pte_same(*page_table, orig_pte)))
3081                 goto out_nomap;
3082
3083         if (unlikely(!PageUptodate(page))) {
3084                 ret = VM_FAULT_SIGBUS;
3085                 goto out_nomap;
3086         }
3087
3088         /*
3089          * The page isn't present yet, go ahead with the fault.
3090          *
3091          * Be careful about the sequence of operations here.
3092          * To get its accounting right, reuse_swap_page() must be called
3093          * while the page is counted on swap but not yet in mapcount i.e.
3094          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3095          * must be called after the swap_free(), or it will never succeed.
3096          * Because delete_from_swap_page() may be called by reuse_swap_page(),
3097          * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3098          * in page->private. In this case, a record in swap_cgroup  is silently
3099          * discarded at swap_free().
3100          */
3101
3102         inc_mm_counter_fast(mm, MM_ANONPAGES);
3103         dec_mm_counter_fast(mm, MM_SWAPENTS);
3104         pte = mk_pte(page, vma->vm_page_prot);
3105         if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3106                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3107                 flags &= ~FAULT_FLAG_WRITE;
3108                 ret |= VM_FAULT_WRITE;
3109                 exclusive = 1;
3110         }
3111         flush_icache_page(vma, page);
3112         set_pte_at(mm, address, page_table, pte);
3113         if (page == swapcache)
3114                 do_page_add_anon_rmap(page, vma, address, exclusive);
3115         else /* ksm created a completely new copy */
3116                 page_add_new_anon_rmap(page, vma, address);
3117         /* It's better to call commit-charge after rmap is established */
3118         mem_cgroup_commit_charge_swapin(page, ptr);
3119
3120         swap_free(entry);
3121         if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3122                 try_to_free_swap(page);
3123         unlock_page(page);
3124         if (page != swapcache) {
3125                 /*
3126                  * Hold the lock to avoid the swap entry to be reused
3127                  * until we take the PT lock for the pte_same() check
3128                  * (to avoid false positives from pte_same). For
3129                  * further safety release the lock after the swap_free
3130                  * so that the swap count won't change under a
3131                  * parallel locked swapcache.
3132                  */
3133                 unlock_page(swapcache);
3134                 page_cache_release(swapcache);
3135         }
3136
3137         if (flags & FAULT_FLAG_WRITE) {
3138                 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte, flags);
3139                 if (ret & VM_FAULT_ERROR)
3140                         ret &= VM_FAULT_ERROR;
3141                 goto out;
3142         }
3143
3144         /* No need to invalidate - it was non-present before */
3145         update_mmu_cache(vma, address, page_table);
3146 unlock:
3147         pte_unmap_unlock(page_table, ptl);
3148 out:
3149         return ret;
3150 out_nomap:
3151         mem_cgroup_cancel_charge_swapin(ptr);
3152         pte_unmap_unlock(page_table, ptl);
3153 out_page:
3154         unlock_page(page);
3155 out_release:
3156         page_cache_release(page);
3157         if (page != swapcache) {
3158                 unlock_page(swapcache);
3159                 page_cache_release(swapcache);
3160         }
3161         return ret;
3162 }
3163
3164 /*
3165  * This is like a special single-page "expand_{down|up}wards()",
3166  * except we must first make sure that 'address{-|+}PAGE_SIZE'
3167  * doesn't hit another vma.
3168  */
3169 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3170 {
3171         address &= PAGE_MASK;
3172         if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3173                 struct vm_area_struct *prev = vma->vm_prev;
3174
3175                 /*
3176                  * Is there a mapping abutting this one below?
3177                  *
3178                  * That's only ok if it's the same stack mapping
3179                  * that has gotten split..
3180                  */
3181                 if (prev && prev->vm_end == address)
3182                         return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3183
3184                 expand_downwards(vma, address - PAGE_SIZE);
3185         }
3186         if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3187                 struct vm_area_struct *next = vma->vm_next;
3188
3189                 /* As VM_GROWSDOWN but s/below/above/ */
3190                 if (next && next->vm_start == address + PAGE_SIZE)
3191                         return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3192
3193                 expand_upwards(vma, address + PAGE_SIZE);
3194         }
3195         return 0;
3196 }
3197
3198 bool is_vma_temporary_stack(struct vm_area_struct *vma);
3199 /*
3200  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3201  * but allow concurrent faults), and pte mapped but not yet locked.
3202  * We return with mmap_sem still held, but pte unmapped and unlocked.
3203  */
3204 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3205                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3206                 unsigned int flags)
3207 {
3208         struct page *page;
3209         spinlock_t *ptl;
3210         pte_t entry;
3211
3212         pte_unmap(page_table);
3213
3214         /* Check if we need to add a guard page to the stack */
3215         if (check_stack_guard_page(vma, address) < 0)
3216                 return VM_FAULT_SIGBUS;
3217
3218         /* Use the zero-page for reads */
3219         if (!(flags & FAULT_FLAG_WRITE)) {
3220                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3221                                                 vma->vm_page_prot));
3222                 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3223                 if (!pte_none(*page_table))
3224                         goto unlock;
3225                 goto setpte;
3226         }
3227
3228         /* Allocate our own private page. */
3229         if (unlikely(anon_vma_prepare(vma)))
3230                 goto oom;
3231         if (vma->vm_flags & VM_LOCKED || flags & FAULT_FLAG_NO_CMA ||
3232             is_vma_temporary_stack(vma)) {
3233                 page = alloc_zeroed_user_highpage(GFP_HIGHUSER, vma, address);
3234         } else {
3235                 page = alloc_zeroed_user_highpage_movable(vma, address);
3236         }
3237         if (!page)
3238                 goto oom;
3239         /*
3240          * The memory barrier inside __SetPageUptodate makes sure that
3241          * preceeding stores to the page contents become visible before
3242          * the set_pte_at() write.
3243          */
3244         __SetPageUptodate(page);
3245
3246         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3247                 goto oom_free_page;
3248
3249         entry = mk_pte(page, vma->vm_page_prot);
3250         if (vma->vm_flags & VM_WRITE)
3251                 entry = pte_mkwrite(pte_mkdirty(entry));
3252
3253         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3254         if (!pte_none(*page_table))
3255                 goto release;
3256
3257         inc_mm_counter_fast(mm, MM_ANONPAGES);
3258         page_add_new_anon_rmap(page, vma, address);
3259 setpte:
3260         set_pte_at(mm, address, page_table, entry);
3261
3262         /* No need to invalidate - it was non-present before */
3263         update_mmu_cache(vma, address, page_table);
3264 unlock:
3265         pte_unmap_unlock(page_table, ptl);
3266         return 0;
3267 release:
3268         mem_cgroup_uncharge_page(page);
3269         page_cache_release(page);
3270         goto unlock;
3271 oom_free_page:
3272         page_cache_release(page);
3273 oom:
3274         return VM_FAULT_OOM;
3275 }
3276
3277 /*
3278  * __do_fault() tries to create a new page mapping. It aggressively
3279  * tries to share with existing pages, but makes a separate copy if
3280  * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3281  * the next page fault.
3282  *
3283  * As this is called only for pages that do not currently exist, we
3284  * do not need to flush old virtual caches or the TLB.
3285  *
3286  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3287  * but allow concurrent faults), and pte neither mapped nor locked.
3288  * We return with mmap_sem still held, but pte unmapped and unlocked.
3289  */
3290 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3291                 unsigned long address, pmd_t *pmd,
3292                 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3293 {
3294         pte_t *page_table;
3295         spinlock_t *ptl;
3296         struct page *page;
3297         struct page *cow_page;
3298         pte_t entry;
3299         int anon = 0;
3300         struct page *dirty_page = NULL;
3301         struct vm_fault vmf;
3302         int ret;
3303         int page_mkwrite = 0;
3304         gfp_t gfp = GFP_HIGHUSER_MOVABLE;
3305
3306         if (IS_ENABLED(CONFIG_CMA) && (flags & FAULT_FLAG_NO_CMA))
3307                 gfp &= ~__GFP_MOVABLE;
3308
3309
3310         /*
3311          * If we do COW later, allocate page befor taking lock_page()
3312          * on the file cache page. This will reduce lock holding time.
3313          */
3314         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3315
3316                 if (unlikely(anon_vma_prepare(vma)))
3317                         return VM_FAULT_OOM;
3318
3319                 cow_page = alloc_page_vma(gfp, vma, address);
3320                 if (!cow_page)
3321                         return VM_FAULT_OOM;
3322
3323                 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3324                         page_cache_release(cow_page);
3325                         return VM_FAULT_OOM;
3326                 }
3327         } else
3328                 cow_page = NULL;
3329
3330         vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3331         vmf.pgoff = pgoff;
3332         vmf.flags = flags;
3333         vmf.page = NULL;
3334
3335         ret = vma->vm_ops->fault(vma, &vmf);
3336         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3337                             VM_FAULT_RETRY)))
3338                 goto uncharge_out;
3339
3340         if (unlikely(PageHWPoison(vmf.page))) {
3341                 if (ret & VM_FAULT_LOCKED)
3342                         unlock_page(vmf.page);
3343                 ret = VM_FAULT_HWPOISON;
3344                 goto uncharge_out;
3345         }
3346
3347         /*
3348          * For consistency in subsequent calls, make the faulted page always
3349          * locked.
3350          */
3351         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3352                 lock_page(vmf.page);
3353         else
3354                 VM_BUG_ON(!PageLocked(vmf.page));
3355
3356         /*
3357          * Should we do an early C-O-W break?
3358          */
3359         page = vmf.page;
3360         if (flags & FAULT_FLAG_WRITE) {
3361                 if (!(vma->vm_flags & VM_SHARED)) {
3362                         page = cow_page;
3363                         anon = 1;
3364                         copy_user_highpage(page, vmf.page, address, vma);
3365                         __SetPageUptodate(page);
3366                 } else {
3367                         /*
3368                          * If the page will be shareable, see if the backing
3369                          * address space wants to know that the page is about
3370                          * to become writable
3371                          */
3372                         if (vma->vm_ops->page_mkwrite) {
3373                                 int tmp;
3374
3375                                 unlock_page(page);
3376                                 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3377                                 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3378                                 if (unlikely(tmp &
3379                                           (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3380                                         ret = tmp;
3381                                         goto unwritable_page;
3382                                 }
3383                                 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3384                                         lock_page(page);
3385                                         if (!page->mapping) {
3386                                                 ret = 0; /* retry the fault */
3387                                                 unlock_page(page);
3388                                                 goto unwritable_page;
3389                                         }
3390                                 } else
3391                                         VM_BUG_ON(!PageLocked(page));
3392                                 page_mkwrite = 1;
3393                         }
3394                 }
3395
3396         }
3397
3398         page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3399
3400         /*
3401          * This silly early PAGE_DIRTY setting removes a race
3402          * due to the bad i386 page protection. But it's valid
3403          * for other architectures too.
3404          *
3405          * Note that if FAULT_FLAG_WRITE is set, we either now have
3406          * an exclusive copy of the page, or this is a shared mapping,
3407          * so we can make it writable and dirty to avoid having to
3408          * handle that later.
3409          */
3410         /* Only go through if we didn't race with anybody else... */
3411         if (likely(pte_same(*page_table, orig_pte))) {
3412                 flush_icache_page(vma, page);
3413                 entry = mk_pte(page, vma->vm_page_prot);
3414                 if (flags & FAULT_FLAG_WRITE)
3415                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3416                 if (anon) {
3417                         inc_mm_counter_fast(mm, MM_ANONPAGES);
3418                         page_add_new_anon_rmap(page, vma, address);
3419                 } else {
3420                         inc_mm_counter_fast(mm, MM_FILEPAGES);
3421                         page_add_file_rmap(page);
3422                         if (flags & FAULT_FLAG_WRITE) {
3423                                 dirty_page = page;
3424                                 get_page(dirty_page);
3425                         }
3426                 }
3427                 set_pte_at(mm, address, page_table, entry);
3428
3429                 /* no need to invalidate: a not-present page won't be cached */
3430                 update_mmu_cache(vma, address, page_table);
3431         } else {
3432                 if (cow_page)
3433                         mem_cgroup_uncharge_page(cow_page);
3434                 if (anon)
3435                         page_cache_release(page);
3436                 else
3437                         anon = 1; /* no anon but release faulted_page */
3438         }
3439
3440         pte_unmap_unlock(page_table, ptl);
3441
3442         if (dirty_page) {
3443                 struct address_space *mapping = page->mapping;
3444                 int dirtied = 0;
3445
3446                 if (set_page_dirty(dirty_page))
3447                         dirtied = 1;
3448                 unlock_page(dirty_page);
3449                 put_page(dirty_page);
3450                 if ((dirtied || page_mkwrite) && mapping) {
3451                         /*
3452                          * Some device drivers do not set page.mapping but still
3453                          * dirty their pages
3454                          */
3455                         balance_dirty_pages_ratelimited(mapping);
3456                 }
3457
3458                 /* file_update_time outside page_lock */
3459                 if (vma->vm_file && !page_mkwrite)
3460                         file_update_time(vma->vm_file);
3461         } else {
3462                 unlock_page(vmf.page);
3463                 if (anon)
3464                         page_cache_release(vmf.page);
3465         }
3466
3467         return ret;
3468
3469 unwritable_page:
3470         page_cache_release(page);
3471         return ret;
3472 uncharge_out:
3473         /* fs's fault handler get error */
3474         if (cow_page) {
3475                 mem_cgroup_uncharge_page(cow_page);
3476                 page_cache_release(cow_page);
3477         }
3478         return ret;
3479 }
3480
3481 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3482                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3483                 unsigned int flags, pte_t orig_pte)
3484 {
3485         pgoff_t pgoff = (((address & PAGE_MASK)
3486                         - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3487
3488         pte_unmap(page_table);
3489         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3490 }
3491
3492 /*
3493  * Fault of a previously existing named mapping. Repopulate the pte
3494  * from the encoded file_pte if possible. This enables swappable
3495  * nonlinear vmas.
3496  *
3497  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3498  * but allow concurrent faults), and pte mapped but not yet locked.
3499  * We return with mmap_sem still held, but pte unmapped and unlocked.
3500  */
3501 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3502                 unsigned long address, pte_t *page_table, pmd_t *pmd,
3503                 unsigned int flags, pte_t orig_pte)
3504 {
3505         pgoff_t pgoff;
3506
3507         flags |= FAULT_FLAG_NONLINEAR;
3508
3509         if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3510                 return 0;
3511
3512         if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3513                 /*
3514                  * Page table corrupted: show pte and kill process.
3515                  */
3516                 print_bad_pte(vma, address, orig_pte, NULL);
3517                 return VM_FAULT_SIGBUS;
3518         }
3519
3520         pgoff = pte_to_pgoff(orig_pte);
3521         return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3522 }
3523
3524 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3525                                 unsigned long addr, int page_nid)
3526 {
3527         get_page(page);
3528
3529         count_vm_numa_event(NUMA_HINT_FAULTS);
3530         if (page_nid == numa_node_id())
3531                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3532
3533         return mpol_misplaced(page, vma, addr);
3534 }
3535
3536 int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3537                    unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3538 {
3539         struct page *page = NULL;
3540         spinlock_t *ptl;
3541         int page_nid = -1;
3542         int target_nid;
3543         bool migrated = false;
3544
3545         /*
3546         * The "pte" at this point cannot be used safely without
3547         * validation through pte_unmap_same(). It's of NUMA type but
3548         * the pfn may be screwed if the read is non atomic.
3549         *
3550         * ptep_modify_prot_start is not called as this is clearing
3551         * the _PAGE_NUMA bit and it is not really expected that there
3552         * would be concurrent hardware modifications to the PTE.
3553         */
3554         ptl = pte_lockptr(mm, pmd);
3555         spin_lock(ptl);
3556         if (unlikely(!pte_same(*ptep, pte))) {
3557                 pte_unmap_unlock(ptep, ptl);
3558                 goto out;
3559         }
3560
3561         pte = pte_mknonnuma(pte);
3562         set_pte_at(mm, addr, ptep, pte);
3563         update_mmu_cache(vma, addr, ptep);
3564
3565         page = vm_normal_page(vma, addr, pte);
3566         if (!page) {
3567                 pte_unmap_unlock(ptep, ptl);
3568                 return 0;
3569         }
3570
3571         page_nid = page_to_nid(page);
3572         target_nid = numa_migrate_prep(page, vma, addr, page_nid);
3573         pte_unmap_unlock(ptep, ptl);
3574         if (target_nid == -1) {
3575                 put_page(page);
3576                 goto out;
3577         }
3578
3579         /* Migrate to the requested node */
3580         migrated = migrate_misplaced_page(page, target_nid);
3581         if (migrated)
3582                 page_nid = target_nid;
3583
3584 out:
3585         if (page_nid != -1)
3586                 task_numa_fault(page_nid, 1, migrated);
3587         return 0;
3588 }
3589
3590 /* NUMA hinting page fault entry point for regular pmds */
3591 #ifdef CONFIG_NUMA_BALANCING
3592 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3593                      unsigned long addr, pmd_t *pmdp)
3594 {
3595         pmd_t pmd;
3596         pte_t *pte, *orig_pte;
3597         unsigned long _addr = addr & PMD_MASK;
3598         unsigned long offset;
3599         spinlock_t *ptl;
3600         bool numa = false;
3601
3602         spin_lock(&mm->page_table_lock);
3603         pmd = *pmdp;
3604         if (pmd_numa(pmd)) {
3605                 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3606                 numa = true;
3607         }
3608         spin_unlock(&mm->page_table_lock);
3609
3610         if (!numa)
3611                 return 0;
3612
3613         /* we're in a page fault so some vma must be in the range */
3614         BUG_ON(!vma);
3615         BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3616         offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3617         VM_BUG_ON(offset >= PMD_SIZE);
3618         orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3619         pte += offset >> PAGE_SHIFT;
3620         for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3621                 pte_t pteval = *pte;
3622                 struct page *page;
3623                 int page_nid = -1;
3624                 int target_nid;
3625                 bool migrated = false;
3626
3627                 if (!pte_present(pteval))
3628                         continue;
3629                 if (!pte_numa(pteval))
3630                         continue;
3631                 if (addr >= vma->vm_end) {
3632                         vma = find_vma(mm, addr);
3633                         /* there's a pte present so there must be a vma */
3634                         BUG_ON(!vma);
3635                         BUG_ON(addr < vma->vm_start);
3636                 }
3637                 if (pte_numa(pteval)) {
3638                         pteval = pte_mknonnuma(pteval);
3639                         set_pte_at(mm, addr, pte, pteval);
3640                 }
3641                 page = vm_normal_page(vma, addr, pteval);
3642                 if (unlikely(!page))
3643                         continue;
3644                 /* only check non-shared pages */
3645                 if (unlikely(page_mapcount(page) != 1))
3646                         continue;
3647
3648                 page_nid = page_to_nid(page);
3649                 target_nid = numa_migrate_prep(page, vma, addr, page_nid);
3650                 pte_unmap_unlock(pte, ptl);
3651                 if (target_nid != -1) {
3652                         migrated = migrate_misplaced_page(page, target_nid);
3653                         if (migrated)
3654                                 page_nid = target_nid;
3655                 } else {
3656                         put_page(page);
3657                 }
3658
3659                 if (page_nid != -1)
3660                         task_numa_fault(page_nid, 1, migrated);
3661
3662                 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
3663         }
3664         pte_unmap_unlock(orig_pte, ptl);
3665
3666         return 0;
3667 }
3668 #else
3669 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3670                      unsigned long addr, pmd_t *pmdp)
3671 {
3672         BUG();
3673         return 0;
3674 }
3675 #endif /* CONFIG_NUMA_BALANCING */
3676
3677 /*
3678  * These routines also need to handle stuff like marking pages dirty
3679  * and/or accessed for architectures that don't do it in hardware (most
3680  * RISC architectures).  The early dirtying is also good on the i386.
3681  *
3682  * There is also a hook called "update_mmu_cache()" that architectures
3683  * with external mmu caches can use to update those (ie the Sparc or
3684  * PowerPC hashed page tables that act as extended TLBs).
3685  *
3686  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3687  * but allow concurrent faults), and pte mapped but not yet locked.
3688  * We return with mmap_sem still held, but pte unmapped and unlocked.
3689  */
3690 int handle_pte_fault(struct mm_struct *mm,
3691                      struct vm_area_struct *vma, unsigned long address,
3692                      pte_t *pte, pmd_t *pmd, unsigned int flags)
3693 {
3694         pte_t entry;
3695         spinlock_t *ptl;
3696
3697         entry = *pte;
3698         if (!pte_present(entry)) {
3699                 if (pte_none(entry)) {
3700                         if (vma->vm_ops) {
3701                                 if (likely(vma->vm_ops->fault))
3702                                         return do_linear_fault(mm, vma, address,
3703                                                 pte, pmd, flags, entry);
3704                         }
3705                         return do_anonymous_page(mm, vma, address,
3706                                                  pte, pmd, flags);
3707                 }
3708                 if (pte_file(entry))
3709                         return do_nonlinear_fault(mm, vma, address,
3710                                         pte, pmd, flags, entry);
3711                 return do_swap_page(mm, vma, address,
3712                                         pte, pmd, flags, entry);
3713         }
3714
3715         if (pte_numa(entry))
3716                 return do_numa_page(mm, vma, address, entry, pte, pmd);
3717
3718         ptl = pte_lockptr(mm, pmd);
3719         spin_lock(ptl);
3720         if (unlikely(!pte_same(*pte, entry)))
3721                 goto unlock;
3722         if (flags & FAULT_FLAG_WRITE) {
3723                 if (!pte_write(entry))
3724                         return do_wp_page(mm, vma, address,
3725                                         pte, pmd, ptl, entry, flags);
3726                 entry = pte_mkdirty(entry);
3727         }
3728         entry = pte_mkyoung(entry);
3729         if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3730                 update_mmu_cache(vma, address, pte);
3731         } else {
3732                 /*
3733                  * This is needed only for protection faults but the arch code
3734                  * is not yet telling us if this is a protection fault or not.
3735                  * This still avoids useless tlb flushes for .text page faults
3736                  * with threads.
3737                  */
3738                 if (flags & FAULT_FLAG_WRITE)
3739                         flush_tlb_fix_spurious_fault(vma, address);
3740         }
3741 unlock:
3742         pte_unmap_unlock(pte, ptl);
3743         return 0;
3744 }
3745
3746 /*
3747  * By the time we get here, we already hold the mm semaphore
3748  */
3749 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3750                              unsigned long address, unsigned int flags)
3751 {
3752         pgd_t *pgd;
3753         pud_t *pud;
3754         pmd_t *pmd;
3755         pte_t *pte;
3756
3757         if (unlikely(is_vm_hugetlb_page(vma)))
3758                 return hugetlb_fault(mm, vma, address, flags);
3759
3760 retry:
3761         pgd = pgd_offset(mm, address);
3762         pud = pud_alloc(mm, pgd, address);
3763         if (!pud)
3764                 return VM_FAULT_OOM;
3765         pmd = pmd_alloc(mm, pud, address);
3766         if (!pmd)
3767                 return VM_FAULT_OOM;
3768         if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3769                 if (!vma->vm_ops)
3770                         return do_huge_pmd_anonymous_page(mm, vma, address,
3771                                                           pmd, flags);
3772         } else {
3773                 pmd_t orig_pmd = *pmd;
3774                 int ret;
3775
3776                 barrier();
3777                 if (pmd_trans_huge(orig_pmd)) {
3778                         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3779
3780                         /*
3781                          * If the pmd is splitting, return and retry the
3782                          * the fault.  Alternative: wait until the split
3783                          * is done, and goto retry.
3784                          */
3785                         if (pmd_trans_splitting(orig_pmd))
3786                                 return 0;
3787
3788                         if (pmd_numa(orig_pmd))
3789                                 return do_huge_pmd_numa_page(mm, vma, address,
3790                                                              orig_pmd, pmd);
3791
3792                         if (dirty && !pmd_write(orig_pmd)) {
3793                                 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3794                                                           orig_pmd);
3795                                 /*
3796                                  * If COW results in an oom, the huge pmd will
3797                                  * have been split, so retry the fault on the
3798                                  * pte for a smaller charge.
3799                                  */
3800                                 if (unlikely(ret & VM_FAULT_OOM))
3801                                         goto retry;
3802                                 return ret;
3803                         } else {
3804                                 huge_pmd_set_accessed(mm, vma, address, pmd,
3805                                                       orig_pmd, dirty);
3806                         }
3807
3808                         return 0;
3809                 }
3810         }
3811
3812         if (pmd_numa(*pmd))
3813                 return do_pmd_numa_page(mm, vma, address, pmd);
3814
3815         /*
3816          * Use __pte_alloc instead of pte_alloc_map, because we can't
3817          * run pte_offset_map on the pmd, if an huge pmd could
3818          * materialize from under us from a different thread.
3819          */
3820         if (unlikely(pmd_none(*pmd)) &&
3821             unlikely(__pte_alloc(mm, vma, pmd, address)))
3822                 return VM_FAULT_OOM;
3823         /* if an huge pmd materialized from under us just retry later */
3824         if (unlikely(pmd_trans_huge(*pmd)))
3825                 return 0;
3826         /*
3827          * A regular pmd is established and it can't morph into a huge pmd
3828          * from under us anymore at this point because we hold the mmap_sem
3829          * read mode and khugepaged takes it in write mode. So now it's
3830          * safe to run pte_offset_map().
3831          */
3832         pte = pte_offset_map(pmd, address);
3833
3834         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3835 }
3836
3837 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3838                     unsigned long address, unsigned int flags)
3839 {
3840         int ret;
3841
3842         __set_current_state(TASK_RUNNING);
3843
3844         count_vm_event(PGFAULT);
3845         mem_cgroup_count_vm_event(mm, PGFAULT);
3846
3847         /* do counter updates before entering really critical section. */
3848         check_sync_rss_stat(current);
3849
3850         /*
3851          * Enable the memcg OOM handling for faults triggered in user
3852          * space.  Kernel faults are handled more gracefully.
3853          */
3854         if (flags & FAULT_FLAG_USER)
3855                 mem_cgroup_enable_oom();
3856
3857         ret = __handle_mm_fault(mm, vma, address, flags);
3858
3859         if (flags & FAULT_FLAG_USER)
3860                 mem_cgroup_disable_oom();
3861
3862         if (WARN_ON(task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)))
3863                 mem_cgroup_oom_synchronize();
3864
3865         return ret;
3866 }
3867
3868 #ifndef __PAGETABLE_PUD_FOLDED
3869 /*
3870  * Allocate page upper directory.
3871  * We've already handled the fast-path in-line.
3872  */
3873 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3874 {
3875         pud_t *new = pud_alloc_one(mm, address);
3876         if (!new)
3877                 return -ENOMEM;
3878
3879         smp_wmb(); /* See comment in __pte_alloc */
3880
3881         spin_lock(&mm->page_table_lock);
3882         if (pgd_present(*pgd))          /* Another has populated it */
3883                 pud_free(mm, new);
3884         else
3885                 pgd_populate(mm, pgd, new);
3886         spin_unlock(&mm->page_table_lock);
3887         return 0;
3888 }
3889 #endif /* __PAGETABLE_PUD_FOLDED */
3890
3891 #ifndef __PAGETABLE_PMD_FOLDED
3892 /*
3893  * Allocate page middle directory.
3894  * We've already handled the fast-path in-line.
3895  */
3896 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3897 {
3898         pmd_t *new = pmd_alloc_one(mm, address);
3899         if (!new)
3900                 return -ENOMEM;
3901
3902         smp_wmb(); /* See comment in __pte_alloc */
3903
3904         spin_lock(&mm->page_table_lock);
3905 #ifndef __ARCH_HAS_4LEVEL_HACK
3906         if (pud_present(*pud))          /* Another has populated it */
3907                 pmd_free(mm, new);
3908         else
3909                 pud_populate(mm, pud, new);
3910 #else
3911         if (pgd_present(*pud))          /* Another has populated it */
3912                 pmd_free(mm, new);
3913         else
3914                 pgd_populate(mm, pud, new);
3915 #endif /* __ARCH_HAS_4LEVEL_HACK */
3916         spin_unlock(&mm->page_table_lock);
3917         return 0;
3918 }
3919 #endif /* __PAGETABLE_PMD_FOLDED */
3920
3921 #if !defined(__HAVE_ARCH_GATE_AREA)
3922
3923 #if defined(AT_SYSINFO_EHDR)
3924 static struct vm_area_struct gate_vma;
3925
3926 static int __init gate_vma_init(void)
3927 {
3928         gate_vma.vm_mm = NULL;
3929         gate_vma.vm_start = FIXADDR_USER_START;
3930         gate_vma.vm_end = FIXADDR_USER_END;
3931         gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3932         gate_vma.vm_page_prot = __P101;
3933
3934         return 0;
3935 }
3936 __initcall(gate_vma_init);
3937 #endif
3938
3939 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3940 {
3941 #ifdef AT_SYSINFO_EHDR
3942         return &gate_vma;
3943 #else
3944         return NULL;
3945 #endif
3946 }
3947
3948 int in_gate_area_no_mm(unsigned long addr)
3949 {
3950 #ifdef AT_SYSINFO_EHDR
3951         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3952                 return 1;
3953 #endif
3954         return 0;
3955 }
3956
3957 #endif  /* __HAVE_ARCH_GATE_AREA */
3958
3959 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3960                 pte_t **ptepp, spinlock_t **ptlp)
3961 {
3962         pgd_t *pgd;
3963         pud_t *pud;
3964         pmd_t *pmd;
3965         pte_t *ptep;
3966
3967         pgd = pgd_offset(mm, address);
3968         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3969                 goto out;
3970
3971         pud = pud_offset(pgd, address);
3972         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3973                 goto out;
3974
3975         pmd = pmd_offset(pud, address);
3976         VM_BUG_ON(pmd_trans_huge(*pmd));
3977         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3978                 goto out;
3979
3980         /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3981         if (pmd_huge(*pmd))
3982                 goto out;
3983
3984         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3985         if (!ptep)
3986                 goto out;
3987         if (!pte_present(*ptep))
3988                 goto unlock;
3989         *ptepp = ptep;
3990         return 0;
3991 unlock:
3992         pte_unmap_unlock(ptep, *ptlp);
3993 out:
3994         return -EINVAL;
3995 }
3996
3997 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3998                              pte_t **ptepp, spinlock_t **ptlp)
3999 {
4000         int res;
4001
4002         /* (void) is needed to make gcc happy */
4003         (void) __cond_lock(*ptlp,
4004                            !(res = __follow_pte(mm, address, ptepp, ptlp)));
4005         return res;
4006 }
4007
4008 /**
4009  * follow_pfn - look up PFN at a user virtual address
4010  * @vma: memory mapping
4011  * @address: user virtual address
4012  * @pfn: location to store found PFN
4013  *
4014  * Only IO mappings and raw PFN mappings are allowed.
4015  *
4016  * Returns zero and the pfn at @pfn on success, -ve otherwise.
4017  */
4018 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4019         unsigned long *pfn)
4020 {
4021         int ret = -EINVAL;
4022         spinlock_t *ptl;
4023         pte_t *ptep;
4024
4025         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4026                 return ret;
4027
4028         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4029         if (ret)
4030                 return ret;
4031         *pfn = pte_pfn(*ptep);
4032         pte_unmap_unlock(ptep, ptl);
4033         return 0;
4034 }
4035 EXPORT_SYMBOL(follow_pfn);
4036
4037 #ifdef CONFIG_HAVE_IOREMAP_PROT
4038 int follow_phys(struct vm_area_struct *vma,
4039                 unsigned long address, unsigned int flags,
4040                 unsigned long *prot, resource_size_t *phys)
4041 {
4042         int ret = -EINVAL;
4043         pte_t *ptep, pte;
4044         spinlock_t *ptl;
4045
4046         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4047                 goto out;
4048
4049         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4050                 goto out;
4051         pte = *ptep;
4052
4053         if ((flags & FOLL_WRITE) && !pte_write(pte))
4054                 goto unlock;
4055
4056         *prot = pgprot_val(pte_pgprot(pte));
4057         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4058
4059         ret = 0;
4060 unlock:
4061         pte_unmap_unlock(ptep, ptl);
4062 out:
4063         return ret;
4064 }
4065
4066 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4067                         void *buf, int len, int write)
4068 {
4069         resource_size_t phys_addr;
4070         unsigned long prot = 0;
4071         void __iomem *maddr;
4072         int offset = addr & (PAGE_SIZE-1);
4073
4074         if (follow_phys(vma, addr, write, &prot, &phys_addr))
4075                 return -EINVAL;
4076
4077         maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4078         if (write)
4079                 memcpy_toio(maddr + offset, buf, len);
4080         else
4081                 memcpy_fromio(buf, maddr + offset, len);
4082         iounmap(maddr);
4083
4084         return len;
4085 }
4086 EXPORT_SYMBOL_GPL(generic_access_phys);
4087 #endif
4088
4089 /*
4090  * Access another process' address space as given in mm.  If non-NULL, use the
4091  * given task for page fault accounting.
4092  */
4093 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4094                 unsigned long addr, void *buf, int len, int write)
4095 {
4096         struct vm_area_struct *vma;
4097         void *old_buf = buf;
4098
4099         down_read(&mm->mmap_sem);
4100         /* ignore errors, just check how much was successfully transferred */
4101         while (len) {
4102                 int bytes, ret, offset;
4103                 void *maddr;
4104                 struct page *page = NULL;
4105
4106                 ret = get_user_pages(tsk, mm, addr, 1,
4107                                 write, 1, &page, &vma);
4108                 if (ret <= 0) {
4109                         /*
4110                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
4111                          * we can access using slightly different code.
4112                          */
4113 #ifdef CONFIG_HAVE_IOREMAP_PROT
4114                         vma = find_vma(mm, addr);
4115                         if (!vma || vma->vm_start > addr)
4116                                 break;
4117                         if (vma->vm_ops && vma->vm_ops->access)
4118                                 ret = vma->vm_ops->access(vma, addr, buf,
4119                                                           len, write);
4120                         if (ret <= 0)
4121 #endif
4122                                 break;
4123                         bytes = ret;
4124                 } else {
4125                         bytes = len;
4126                         offset = addr & (PAGE_SIZE-1);
4127                         if (bytes > PAGE_SIZE-offset)
4128                                 bytes = PAGE_SIZE-offset;
4129
4130                         maddr = kmap(page);
4131                         if (write) {
4132                                 copy_to_user_page(vma, page, addr,
4133                                                   maddr + offset, buf, bytes);
4134                                 set_page_dirty_lock(page);
4135                         } else {
4136                                 copy_from_user_page(vma, page, addr,
4137                                                     buf, maddr + offset, bytes);
4138                         }
4139                         kunmap(page);
4140                         page_cache_release(page);
4141                 }
4142                 len -= bytes;
4143                 buf += bytes;
4144                 addr += bytes;
4145         }
4146         up_read(&mm->mmap_sem);
4147
4148         return buf - old_buf;
4149 }
4150
4151 /**
4152  * access_remote_vm - access another process' address space
4153  * @mm:         the mm_struct of the target address space
4154  * @addr:       start address to access
4155  * @buf:        source or destination buffer
4156  * @len:        number of bytes to transfer
4157  * @write:      whether the access is a write
4158  *
4159  * The caller must hold a reference on @mm.
4160  */
4161 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4162                 void *buf, int len, int write)
4163 {
4164         return __access_remote_vm(NULL, mm, addr, buf, len, write);
4165 }
4166
4167 /*
4168  * Access another process' address space.
4169  * Source/target buffer must be kernel space,
4170  * Do not walk the page table directly, use get_user_pages
4171  */
4172 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4173                 void *buf, int len, int write)
4174 {
4175         struct mm_struct *mm;
4176         int ret;
4177
4178         mm = get_task_mm(tsk);
4179         if (!mm)
4180                 return 0;
4181
4182         ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4183         mmput(mm);
4184
4185         return ret;
4186 }
4187
4188 /*
4189  * Print the name of a VMA.
4190  */
4191 void print_vma_addr(char *prefix, unsigned long ip)
4192 {
4193         struct mm_struct *mm = current->mm;
4194         struct vm_area_struct *vma;
4195
4196         /*
4197          * Do not print if we are in atomic
4198          * contexts (in exception stacks, etc.):
4199          */
4200         if (preempt_count())
4201                 return;
4202
4203         down_read(&mm->mmap_sem);
4204         vma = find_vma(mm, ip);
4205         if (vma && vma->vm_file) {
4206                 struct file *f = vma->vm_file;
4207                 char *buf = (char *)__get_free_page(GFP_KERNEL);
4208                 if (buf) {
4209                         char *p;
4210
4211                         p = d_path(&f->f_path, buf, PAGE_SIZE);
4212                         if (IS_ERR(p))
4213                                 p = "?";
4214                         printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4215                                         vma->vm_start,
4216                                         vma->vm_end - vma->vm_start);
4217                         free_page((unsigned long)buf);
4218                 }
4219         }
4220         up_read(&mm->mmap_sem);
4221 }
4222
4223 #ifdef CONFIG_PROVE_LOCKING
4224 void might_fault(void)
4225 {
4226         /*
4227          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4228          * holding the mmap_sem, this is safe because kernel memory doesn't
4229          * get paged out, therefore we'll never actually fault, and the
4230          * below annotations will generate false positives.
4231          */
4232         if (segment_eq(get_fs(), KERNEL_DS))
4233                 return;
4234
4235         might_sleep();
4236         /*
4237          * it would be nicer only to annotate paths which are not under
4238          * pagefault_disable, however that requires a larger audit and
4239          * providing helpers like get_user_atomic.
4240          */
4241         if (!in_atomic() && current->mm)
4242                 might_lock_read(&current->mm->mmap_sem);
4243 }
4244 EXPORT_SYMBOL(might_fault);
4245 #endif
4246
4247 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4248 static void clear_gigantic_page(struct page *page,
4249                                 unsigned long addr,
4250                                 unsigned int pages_per_huge_page)
4251 {
4252         int i;
4253         struct page *p = page;
4254
4255         might_sleep();
4256         for (i = 0; i < pages_per_huge_page;
4257              i++, p = mem_map_next(p, page, i)) {
4258                 cond_resched();
4259                 clear_user_highpage(p, addr + i * PAGE_SIZE);
4260         }
4261 }
4262 void clear_huge_page(struct page *page,
4263                      unsigned long addr, unsigned int pages_per_huge_page)
4264 {
4265         int i;
4266
4267         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4268                 clear_gigantic_page(page, addr, pages_per_huge_page);
4269                 return;
4270         }
4271
4272         might_sleep();
4273         for (i = 0; i < pages_per_huge_page; i++) {
4274                 cond_resched();
4275                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4276         }
4277 }
4278
4279 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4280                                     unsigned long addr,
4281                                     struct vm_area_struct *vma,
4282                                     unsigned int pages_per_huge_page)
4283 {
4284         int i;
4285         struct page *dst_base = dst;
4286         struct page *src_base = src;
4287
4288         for (i = 0; i < pages_per_huge_page; ) {
4289                 cond_resched();
4290                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4291
4292                 i++;
4293                 dst = mem_map_next(dst, dst_base, i);
4294                 src = mem_map_next(src, src_base, i);
4295         }
4296 }
4297
4298 void copy_user_huge_page(struct page *dst, struct page *src,
4299                          unsigned long addr, struct vm_area_struct *vma,
4300                          unsigned int pages_per_huge_page)
4301 {
4302         int i;
4303
4304         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4305                 copy_user_gigantic_page(dst, src, addr, vma,
4306                                         pages_per_huge_page);
4307                 return;
4308         }
4309
4310         might_sleep();
4311         for (i = 0; i < pages_per_huge_page; i++) {
4312                 cond_resched();
4313                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4314         }
4315 }
4316 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */