Loading...
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/module.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
61#include <asm/io.h>
62#include <asm/pgalloc.h>
63#include <asm/uaccess.h>
64#include <asm/tlb.h>
65#include <asm/tlbflush.h>
66#include <asm/pgtable.h>
67
68#include "internal.h"
69
70#ifndef CONFIG_NEED_MULTIPLE_NODES
71/* use the per-pgdat data instead for discontigmem - mbligh */
72unsigned long max_mapnr;
73struct page *mem_map;
74
75EXPORT_SYMBOL(max_mapnr);
76EXPORT_SYMBOL(mem_map);
77#endif
78
79unsigned long num_physpages;
80/*
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
85 * and ZONE_HIGHMEM.
86 */
87void * high_memory;
88
89EXPORT_SYMBOL(num_physpages);
90EXPORT_SYMBOL(high_memory);
91
92/*
93 * Randomize the address space (stacks, mmaps, brk, etc.).
94 *
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
97 */
98int randomize_va_space __read_mostly =
99#ifdef CONFIG_COMPAT_BRK
100 1;
101#else
102 2;
103#endif
104
105static int __init disable_randmaps(char *s)
106{
107 randomize_va_space = 0;
108 return 1;
109}
110__setup("norandmaps", disable_randmaps);
111
112unsigned long zero_pfn __read_mostly;
113unsigned long highest_memmap_pfn __read_mostly;
114
115/*
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
117 */
118static int __init init_zero_pfn(void)
119{
120 zero_pfn = page_to_pfn(ZERO_PAGE(0));
121 return 0;
122}
123core_initcall(init_zero_pfn);
124
125
126#if defined(SPLIT_RSS_COUNTING)
127
128static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
129{
130 int i;
131
132 for (i = 0; i < NR_MM_COUNTERS; i++) {
133 if (task->rss_stat.count[i]) {
134 add_mm_counter(mm, i, task->rss_stat.count[i]);
135 task->rss_stat.count[i] = 0;
136 }
137 }
138 task->rss_stat.events = 0;
139}
140
141static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
142{
143 struct task_struct *task = current;
144
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
147 else
148 add_mm_counter(mm, member, val);
149}
150#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
152
153/* sync counter once per 64 page faults */
154#define TASK_RSS_EVENTS_THRESH (64)
155static void check_sync_rss_stat(struct task_struct *task)
156{
157 if (unlikely(task != current))
158 return;
159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160 __sync_task_rss_stat(task, task->mm);
161}
162
163unsigned long get_mm_counter(struct mm_struct *mm, int member)
164{
165 long val = 0;
166
167 /*
168 * Don't use task->mm here...for avoiding to use task_get_mm()..
169 * The caller must guarantee task->mm is not invalid.
170 */
171 val = atomic_long_read(&mm->rss_stat.count[member]);
172 /*
173 * counter is updated in asynchronous manner and may go to minus.
174 * But it's never be expected number for users.
175 */
176 if (val < 0)
177 return 0;
178 return (unsigned long)val;
179}
180
181void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
182{
183 __sync_task_rss_stat(task, mm);
184}
185#else /* SPLIT_RSS_COUNTING */
186
187#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
189
190static void check_sync_rss_stat(struct task_struct *task)
191{
192}
193
194#endif /* SPLIT_RSS_COUNTING */
195
196#ifdef HAVE_GENERIC_MMU_GATHER
197
198static int tlb_next_batch(struct mmu_gather *tlb)
199{
200 struct mmu_gather_batch *batch;
201
202 batch = tlb->active;
203 if (batch->next) {
204 tlb->active = batch->next;
205 return 1;
206 }
207
208 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
209 if (!batch)
210 return 0;
211
212 batch->next = NULL;
213 batch->nr = 0;
214 batch->max = MAX_GATHER_BATCH;
215
216 tlb->active->next = batch;
217 tlb->active = batch;
218
219 return 1;
220}
221
222/* tlb_gather_mmu
223 * Called to initialize an (on-stack) mmu_gather structure for page-table
224 * tear-down from @mm. The @fullmm argument is used when @mm is without
225 * users and we're going to destroy the full address space (exit/execve).
226 */
227void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
228{
229 tlb->mm = mm;
230
231 tlb->fullmm = fullmm;
232 tlb->need_flush = 0;
233 tlb->fast_mode = (num_possible_cpus() == 1);
234 tlb->local.next = NULL;
235 tlb->local.nr = 0;
236 tlb->local.max = ARRAY_SIZE(tlb->__pages);
237 tlb->active = &tlb->local;
238
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb->batch = NULL;
241#endif
242}
243
244void tlb_flush_mmu(struct mmu_gather *tlb)
245{
246 struct mmu_gather_batch *batch;
247
248 if (!tlb->need_flush)
249 return;
250 tlb->need_flush = 0;
251 tlb_flush(tlb);
252#ifdef CONFIG_HAVE_RCU_TABLE_FREE
253 tlb_table_flush(tlb);
254#endif
255
256 if (tlb_fast_mode(tlb))
257 return;
258
259 for (batch = &tlb->local; batch; batch = batch->next) {
260 free_pages_and_swap_cache(batch->pages, batch->nr);
261 batch->nr = 0;
262 }
263 tlb->active = &tlb->local;
264}
265
266/* tlb_finish_mmu
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
269 */
270void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
271{
272 struct mmu_gather_batch *batch, *next;
273
274 tlb_flush_mmu(tlb);
275
276 /* keep the page table cache within bounds */
277 check_pgt_cache();
278
279 for (batch = tlb->local.next; batch; batch = next) {
280 next = batch->next;
281 free_pages((unsigned long)batch, 0);
282 }
283 tlb->local.next = NULL;
284}
285
286/* __tlb_remove_page
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
291 */
292int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
293{
294 struct mmu_gather_batch *batch;
295
296 tlb->need_flush = 1;
297
298 if (tlb_fast_mode(tlb)) {
299 free_page_and_swap_cache(page);
300 return 1; /* avoid calling tlb_flush_mmu() */
301 }
302
303 batch = tlb->active;
304 batch->pages[batch->nr++] = page;
305 if (batch->nr == batch->max) {
306 if (!tlb_next_batch(tlb))
307 return 0;
308 batch = tlb->active;
309 }
310 VM_BUG_ON(batch->nr > batch->max);
311
312 return batch->max - batch->nr;
313}
314
315#endif /* HAVE_GENERIC_MMU_GATHER */
316
317#ifdef CONFIG_HAVE_RCU_TABLE_FREE
318
319/*
320 * See the comment near struct mmu_table_batch.
321 */
322
323static void tlb_remove_table_smp_sync(void *arg)
324{
325 /* Simply deliver the interrupt */
326}
327
328static void tlb_remove_table_one(void *table)
329{
330 /*
331 * This isn't an RCU grace period and hence the page-tables cannot be
332 * assumed to be actually RCU-freed.
333 *
334 * It is however sufficient for software page-table walkers that rely on
335 * IRQ disabling. See the comment near struct mmu_table_batch.
336 */
337 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
338 __tlb_remove_table(table);
339}
340
341static void tlb_remove_table_rcu(struct rcu_head *head)
342{
343 struct mmu_table_batch *batch;
344 int i;
345
346 batch = container_of(head, struct mmu_table_batch, rcu);
347
348 for (i = 0; i < batch->nr; i++)
349 __tlb_remove_table(batch->tables[i]);
350
351 free_page((unsigned long)batch);
352}
353
354void tlb_table_flush(struct mmu_gather *tlb)
355{
356 struct mmu_table_batch **batch = &tlb->batch;
357
358 if (*batch) {
359 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
360 *batch = NULL;
361 }
362}
363
364void tlb_remove_table(struct mmu_gather *tlb, void *table)
365{
366 struct mmu_table_batch **batch = &tlb->batch;
367
368 tlb->need_flush = 1;
369
370 /*
371 * When there's less then two users of this mm there cannot be a
372 * concurrent page-table walk.
373 */
374 if (atomic_read(&tlb->mm->mm_users) < 2) {
375 __tlb_remove_table(table);
376 return;
377 }
378
379 if (*batch == NULL) {
380 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
381 if (*batch == NULL) {
382 tlb_remove_table_one(table);
383 return;
384 }
385 (*batch)->nr = 0;
386 }
387 (*batch)->tables[(*batch)->nr++] = table;
388 if ((*batch)->nr == MAX_TABLE_BATCH)
389 tlb_table_flush(tlb);
390}
391
392#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
393
394/*
395 * If a p?d_bad entry is found while walking page tables, report
396 * the error, before resetting entry to p?d_none. Usually (but
397 * very seldom) called out from the p?d_none_or_clear_bad macros.
398 */
399
400void pgd_clear_bad(pgd_t *pgd)
401{
402 pgd_ERROR(*pgd);
403 pgd_clear(pgd);
404}
405
406void pud_clear_bad(pud_t *pud)
407{
408 pud_ERROR(*pud);
409 pud_clear(pud);
410}
411
412void pmd_clear_bad(pmd_t *pmd)
413{
414 pmd_ERROR(*pmd);
415 pmd_clear(pmd);
416}
417
418/*
419 * Note: this doesn't free the actual pages themselves. That
420 * has been handled earlier when unmapping all the memory regions.
421 */
422static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
423 unsigned long addr)
424{
425 pgtable_t token = pmd_pgtable(*pmd);
426 pmd_clear(pmd);
427 pte_free_tlb(tlb, token, addr);
428 tlb->mm->nr_ptes--;
429}
430
431static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
432 unsigned long addr, unsigned long end,
433 unsigned long floor, unsigned long ceiling)
434{
435 pmd_t *pmd;
436 unsigned long next;
437 unsigned long start;
438
439 start = addr;
440 pmd = pmd_offset(pud, addr);
441 do {
442 next = pmd_addr_end(addr, end);
443 if (pmd_none_or_clear_bad(pmd))
444 continue;
445 free_pte_range(tlb, pmd, addr);
446 } while (pmd++, addr = next, addr != end);
447
448 start &= PUD_MASK;
449 if (start < floor)
450 return;
451 if (ceiling) {
452 ceiling &= PUD_MASK;
453 if (!ceiling)
454 return;
455 }
456 if (end - 1 > ceiling - 1)
457 return;
458
459 pmd = pmd_offset(pud, start);
460 pud_clear(pud);
461 pmd_free_tlb(tlb, pmd, start);
462}
463
464static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
465 unsigned long addr, unsigned long end,
466 unsigned long floor, unsigned long ceiling)
467{
468 pud_t *pud;
469 unsigned long next;
470 unsigned long start;
471
472 start = addr;
473 pud = pud_offset(pgd, addr);
474 do {
475 next = pud_addr_end(addr, end);
476 if (pud_none_or_clear_bad(pud))
477 continue;
478 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
479 } while (pud++, addr = next, addr != end);
480
481 start &= PGDIR_MASK;
482 if (start < floor)
483 return;
484 if (ceiling) {
485 ceiling &= PGDIR_MASK;
486 if (!ceiling)
487 return;
488 }
489 if (end - 1 > ceiling - 1)
490 return;
491
492 pud = pud_offset(pgd, start);
493 pgd_clear(pgd);
494 pud_free_tlb(tlb, pud, start);
495}
496
497/*
498 * This function frees user-level page tables of a process.
499 *
500 * Must be called with pagetable lock held.
501 */
502void free_pgd_range(struct mmu_gather *tlb,
503 unsigned long addr, unsigned long end,
504 unsigned long floor, unsigned long ceiling)
505{
506 pgd_t *pgd;
507 unsigned long next;
508
509 /*
510 * The next few lines have given us lots of grief...
511 *
512 * Why are we testing PMD* at this top level? Because often
513 * there will be no work to do at all, and we'd prefer not to
514 * go all the way down to the bottom just to discover that.
515 *
516 * Why all these "- 1"s? Because 0 represents both the bottom
517 * of the address space and the top of it (using -1 for the
518 * top wouldn't help much: the masks would do the wrong thing).
519 * The rule is that addr 0 and floor 0 refer to the bottom of
520 * the address space, but end 0 and ceiling 0 refer to the top
521 * Comparisons need to use "end - 1" and "ceiling - 1" (though
522 * that end 0 case should be mythical).
523 *
524 * Wherever addr is brought up or ceiling brought down, we must
525 * be careful to reject "the opposite 0" before it confuses the
526 * subsequent tests. But what about where end is brought down
527 * by PMD_SIZE below? no, end can't go down to 0 there.
528 *
529 * Whereas we round start (addr) and ceiling down, by different
530 * masks at different levels, in order to test whether a table
531 * now has no other vmas using it, so can be freed, we don't
532 * bother to round floor or end up - the tests don't need that.
533 */
534
535 addr &= PMD_MASK;
536 if (addr < floor) {
537 addr += PMD_SIZE;
538 if (!addr)
539 return;
540 }
541 if (ceiling) {
542 ceiling &= PMD_MASK;
543 if (!ceiling)
544 return;
545 }
546 if (end - 1 > ceiling - 1)
547 end -= PMD_SIZE;
548 if (addr > end - 1)
549 return;
550
551 pgd = pgd_offset(tlb->mm, addr);
552 do {
553 next = pgd_addr_end(addr, end);
554 if (pgd_none_or_clear_bad(pgd))
555 continue;
556 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
557 } while (pgd++, addr = next, addr != end);
558}
559
560void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
561 unsigned long floor, unsigned long ceiling)
562{
563 while (vma) {
564 struct vm_area_struct *next = vma->vm_next;
565 unsigned long addr = vma->vm_start;
566
567 /*
568 * Hide vma from rmap and truncate_pagecache before freeing
569 * pgtables
570 */
571 unlink_anon_vmas(vma);
572 unlink_file_vma(vma);
573
574 if (is_vm_hugetlb_page(vma)) {
575 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
576 floor, next? next->vm_start: ceiling);
577 } else {
578 /*
579 * Optimization: gather nearby vmas into one call down
580 */
581 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
582 && !is_vm_hugetlb_page(next)) {
583 vma = next;
584 next = vma->vm_next;
585 unlink_anon_vmas(vma);
586 unlink_file_vma(vma);
587 }
588 free_pgd_range(tlb, addr, vma->vm_end,
589 floor, next? next->vm_start: ceiling);
590 }
591 vma = next;
592 }
593}
594
595int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
596 pmd_t *pmd, unsigned long address)
597{
598 pgtable_t new = pte_alloc_one(mm, address);
599 int wait_split_huge_page;
600 if (!new)
601 return -ENOMEM;
602
603 /*
604 * Ensure all pte setup (eg. pte page lock and page clearing) are
605 * visible before the pte is made visible to other CPUs by being
606 * put into page tables.
607 *
608 * The other side of the story is the pointer chasing in the page
609 * table walking code (when walking the page table without locking;
610 * ie. most of the time). Fortunately, these data accesses consist
611 * of a chain of data-dependent loads, meaning most CPUs (alpha
612 * being the notable exception) will already guarantee loads are
613 * seen in-order. See the alpha page table accessors for the
614 * smp_read_barrier_depends() barriers in page table walking code.
615 */
616 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
617
618 spin_lock(&mm->page_table_lock);
619 wait_split_huge_page = 0;
620 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
621 mm->nr_ptes++;
622 pmd_populate(mm, pmd, new);
623 new = NULL;
624 } else if (unlikely(pmd_trans_splitting(*pmd)))
625 wait_split_huge_page = 1;
626 spin_unlock(&mm->page_table_lock);
627 if (new)
628 pte_free(mm, new);
629 if (wait_split_huge_page)
630 wait_split_huge_page(vma->anon_vma, pmd);
631 return 0;
632}
633
634int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
635{
636 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
637 if (!new)
638 return -ENOMEM;
639
640 smp_wmb(); /* See comment in __pte_alloc */
641
642 spin_lock(&init_mm.page_table_lock);
643 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
644 pmd_populate_kernel(&init_mm, pmd, new);
645 new = NULL;
646 } else
647 VM_BUG_ON(pmd_trans_splitting(*pmd));
648 spin_unlock(&init_mm.page_table_lock);
649 if (new)
650 pte_free_kernel(&init_mm, new);
651 return 0;
652}
653
654static inline void init_rss_vec(int *rss)
655{
656 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
657}
658
659static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
660{
661 int i;
662
663 if (current->mm == mm)
664 sync_mm_rss(current, mm);
665 for (i = 0; i < NR_MM_COUNTERS; i++)
666 if (rss[i])
667 add_mm_counter(mm, i, rss[i]);
668}
669
670/*
671 * This function is called to print an error when a bad pte
672 * is found. For example, we might have a PFN-mapped pte in
673 * a region that doesn't allow it.
674 *
675 * The calling function must still handle the error.
676 */
677static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
678 pte_t pte, struct page *page)
679{
680 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
681 pud_t *pud = pud_offset(pgd, addr);
682 pmd_t *pmd = pmd_offset(pud, addr);
683 struct address_space *mapping;
684 pgoff_t index;
685 static unsigned long resume;
686 static unsigned long nr_shown;
687 static unsigned long nr_unshown;
688
689 /*
690 * Allow a burst of 60 reports, then keep quiet for that minute;
691 * or allow a steady drip of one report per second.
692 */
693 if (nr_shown == 60) {
694 if (time_before(jiffies, resume)) {
695 nr_unshown++;
696 return;
697 }
698 if (nr_unshown) {
699 printk(KERN_ALERT
700 "BUG: Bad page map: %lu messages suppressed\n",
701 nr_unshown);
702 nr_unshown = 0;
703 }
704 nr_shown = 0;
705 }
706 if (nr_shown++ == 0)
707 resume = jiffies + 60 * HZ;
708
709 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
710 index = linear_page_index(vma, addr);
711
712 printk(KERN_ALERT
713 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
714 current->comm,
715 (long long)pte_val(pte), (long long)pmd_val(*pmd));
716 if (page)
717 dump_page(page);
718 printk(KERN_ALERT
719 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
720 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
721 /*
722 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
723 */
724 if (vma->vm_ops)
725 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
726 (unsigned long)vma->vm_ops->fault);
727 if (vma->vm_file && vma->vm_file->f_op)
728 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
729 (unsigned long)vma->vm_file->f_op->mmap);
730 dump_stack();
731 add_taint(TAINT_BAD_PAGE);
732}
733
734static inline int is_cow_mapping(vm_flags_t flags)
735{
736 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
737}
738
739#ifndef is_zero_pfn
740static inline int is_zero_pfn(unsigned long pfn)
741{
742 return pfn == zero_pfn;
743}
744#endif
745
746#ifndef my_zero_pfn
747static inline unsigned long my_zero_pfn(unsigned long addr)
748{
749 return zero_pfn;
750}
751#endif
752
753/*
754 * vm_normal_page -- This function gets the "struct page" associated with a pte.
755 *
756 * "Special" mappings do not wish to be associated with a "struct page" (either
757 * it doesn't exist, or it exists but they don't want to touch it). In this
758 * case, NULL is returned here. "Normal" mappings do have a struct page.
759 *
760 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
761 * pte bit, in which case this function is trivial. Secondly, an architecture
762 * may not have a spare pte bit, which requires a more complicated scheme,
763 * described below.
764 *
765 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
766 * special mapping (even if there are underlying and valid "struct pages").
767 * COWed pages of a VM_PFNMAP are always normal.
768 *
769 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
770 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
771 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
772 * mapping will always honor the rule
773 *
774 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
775 *
776 * And for normal mappings this is false.
777 *
778 * This restricts such mappings to be a linear translation from virtual address
779 * to pfn. To get around this restriction, we allow arbitrary mappings so long
780 * as the vma is not a COW mapping; in that case, we know that all ptes are
781 * special (because none can have been COWed).
782 *
783 *
784 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
785 *
786 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
787 * page" backing, however the difference is that _all_ pages with a struct
788 * page (that is, those where pfn_valid is true) are refcounted and considered
789 * normal pages by the VM. The disadvantage is that pages are refcounted
790 * (which can be slower and simply not an option for some PFNMAP users). The
791 * advantage is that we don't have to follow the strict linearity rule of
792 * PFNMAP mappings in order to support COWable mappings.
793 *
794 */
795#ifdef __HAVE_ARCH_PTE_SPECIAL
796# define HAVE_PTE_SPECIAL 1
797#else
798# define HAVE_PTE_SPECIAL 0
799#endif
800struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
801 pte_t pte)
802{
803 unsigned long pfn = pte_pfn(pte);
804
805 if (HAVE_PTE_SPECIAL) {
806 if (likely(!pte_special(pte)))
807 goto check_pfn;
808 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
809 return NULL;
810 if (!is_zero_pfn(pfn))
811 print_bad_pte(vma, addr, pte, NULL);
812 return NULL;
813 }
814
815 /* !HAVE_PTE_SPECIAL case follows: */
816
817 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
818 if (vma->vm_flags & VM_MIXEDMAP) {
819 if (!pfn_valid(pfn))
820 return NULL;
821 goto out;
822 } else {
823 unsigned long off;
824 off = (addr - vma->vm_start) >> PAGE_SHIFT;
825 if (pfn == vma->vm_pgoff + off)
826 return NULL;
827 if (!is_cow_mapping(vma->vm_flags))
828 return NULL;
829 }
830 }
831
832 if (is_zero_pfn(pfn))
833 return NULL;
834check_pfn:
835 if (unlikely(pfn > highest_memmap_pfn)) {
836 print_bad_pte(vma, addr, pte, NULL);
837 return NULL;
838 }
839
840 /*
841 * NOTE! We still have PageReserved() pages in the page tables.
842 * eg. VDSO mappings can cause them to exist.
843 */
844out:
845 return pfn_to_page(pfn);
846}
847
848/*
849 * copy one vm_area from one task to the other. Assumes the page tables
850 * already present in the new task to be cleared in the whole range
851 * covered by this vma.
852 */
853
854static inline unsigned long
855copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
856 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
857 unsigned long addr, int *rss)
858{
859 unsigned long vm_flags = vma->vm_flags;
860 pte_t pte = *src_pte;
861 struct page *page;
862
863 /* pte contains position in swap or file, so copy. */
864 if (unlikely(!pte_present(pte))) {
865 if (!pte_file(pte)) {
866 swp_entry_t entry = pte_to_swp_entry(pte);
867
868 if (swap_duplicate(entry) < 0)
869 return entry.val;
870
871 /* make sure dst_mm is on swapoff's mmlist. */
872 if (unlikely(list_empty(&dst_mm->mmlist))) {
873 spin_lock(&mmlist_lock);
874 if (list_empty(&dst_mm->mmlist))
875 list_add(&dst_mm->mmlist,
876 &src_mm->mmlist);
877 spin_unlock(&mmlist_lock);
878 }
879 if (likely(!non_swap_entry(entry)))
880 rss[MM_SWAPENTS]++;
881 else if (is_write_migration_entry(entry) &&
882 is_cow_mapping(vm_flags)) {
883 /*
884 * COW mappings require pages in both parent
885 * and child to be set to read.
886 */
887 make_migration_entry_read(&entry);
888 pte = swp_entry_to_pte(entry);
889 set_pte_at(src_mm, addr, src_pte, pte);
890 }
891 }
892 goto out_set_pte;
893 }
894
895 /*
896 * If it's a COW mapping, write protect it both
897 * in the parent and the child
898 */
899 if (is_cow_mapping(vm_flags)) {
900 ptep_set_wrprotect(src_mm, addr, src_pte);
901 pte = pte_wrprotect(pte);
902 }
903
904 /*
905 * If it's a shared mapping, mark it clean in
906 * the child
907 */
908 if (vm_flags & VM_SHARED)
909 pte = pte_mkclean(pte);
910 pte = pte_mkold(pte);
911
912 page = vm_normal_page(vma, addr, pte);
913 if (page) {
914 get_page(page);
915 page_dup_rmap(page);
916 if (PageAnon(page))
917 rss[MM_ANONPAGES]++;
918 else
919 rss[MM_FILEPAGES]++;
920 }
921
922out_set_pte:
923 set_pte_at(dst_mm, addr, dst_pte, pte);
924 return 0;
925}
926
927int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
928 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
929 unsigned long addr, unsigned long end)
930{
931 pte_t *orig_src_pte, *orig_dst_pte;
932 pte_t *src_pte, *dst_pte;
933 spinlock_t *src_ptl, *dst_ptl;
934 int progress = 0;
935 int rss[NR_MM_COUNTERS];
936 swp_entry_t entry = (swp_entry_t){0};
937
938again:
939 init_rss_vec(rss);
940
941 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
942 if (!dst_pte)
943 return -ENOMEM;
944 src_pte = pte_offset_map(src_pmd, addr);
945 src_ptl = pte_lockptr(src_mm, src_pmd);
946 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
947 orig_src_pte = src_pte;
948 orig_dst_pte = dst_pte;
949 arch_enter_lazy_mmu_mode();
950
951 do {
952 /*
953 * We are holding two locks at this point - either of them
954 * could generate latencies in another task on another CPU.
955 */
956 if (progress >= 32) {
957 progress = 0;
958 if (need_resched() ||
959 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
960 break;
961 }
962 if (pte_none(*src_pte)) {
963 progress++;
964 continue;
965 }
966 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
967 vma, addr, rss);
968 if (entry.val)
969 break;
970 progress += 8;
971 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
972
973 arch_leave_lazy_mmu_mode();
974 spin_unlock(src_ptl);
975 pte_unmap(orig_src_pte);
976 add_mm_rss_vec(dst_mm, rss);
977 pte_unmap_unlock(orig_dst_pte, dst_ptl);
978 cond_resched();
979
980 if (entry.val) {
981 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
982 return -ENOMEM;
983 progress = 0;
984 }
985 if (addr != end)
986 goto again;
987 return 0;
988}
989
990static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
991 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
992 unsigned long addr, unsigned long end)
993{
994 pmd_t *src_pmd, *dst_pmd;
995 unsigned long next;
996
997 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
998 if (!dst_pmd)
999 return -ENOMEM;
1000 src_pmd = pmd_offset(src_pud, addr);
1001 do {
1002 next = pmd_addr_end(addr, end);
1003 if (pmd_trans_huge(*src_pmd)) {
1004 int err;
1005 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1006 err = copy_huge_pmd(dst_mm, src_mm,
1007 dst_pmd, src_pmd, addr, vma);
1008 if (err == -ENOMEM)
1009 return -ENOMEM;
1010 if (!err)
1011 continue;
1012 /* fall through */
1013 }
1014 if (pmd_none_or_clear_bad(src_pmd))
1015 continue;
1016 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1017 vma, addr, next))
1018 return -ENOMEM;
1019 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1020 return 0;
1021}
1022
1023static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1024 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1025 unsigned long addr, unsigned long end)
1026{
1027 pud_t *src_pud, *dst_pud;
1028 unsigned long next;
1029
1030 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1031 if (!dst_pud)
1032 return -ENOMEM;
1033 src_pud = pud_offset(src_pgd, addr);
1034 do {
1035 next = pud_addr_end(addr, end);
1036 if (pud_none_or_clear_bad(src_pud))
1037 continue;
1038 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1039 vma, addr, next))
1040 return -ENOMEM;
1041 } while (dst_pud++, src_pud++, addr = next, addr != end);
1042 return 0;
1043}
1044
1045int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1046 struct vm_area_struct *vma)
1047{
1048 pgd_t *src_pgd, *dst_pgd;
1049 unsigned long next;
1050 unsigned long addr = vma->vm_start;
1051 unsigned long end = vma->vm_end;
1052 int ret;
1053
1054 /*
1055 * Don't copy ptes where a page fault will fill them correctly.
1056 * Fork becomes much lighter when there are big shared or private
1057 * readonly mappings. The tradeoff is that copy_page_range is more
1058 * efficient than faulting.
1059 */
1060 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1061 if (!vma->anon_vma)
1062 return 0;
1063 }
1064
1065 if (is_vm_hugetlb_page(vma))
1066 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1067
1068 if (unlikely(is_pfn_mapping(vma))) {
1069 /*
1070 * We do not free on error cases below as remove_vma
1071 * gets called on error from higher level routine
1072 */
1073 ret = track_pfn_vma_copy(vma);
1074 if (ret)
1075 return ret;
1076 }
1077
1078 /*
1079 * We need to invalidate the secondary MMU mappings only when
1080 * there could be a permission downgrade on the ptes of the
1081 * parent mm. And a permission downgrade will only happen if
1082 * is_cow_mapping() returns true.
1083 */
1084 if (is_cow_mapping(vma->vm_flags))
1085 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1086
1087 ret = 0;
1088 dst_pgd = pgd_offset(dst_mm, addr);
1089 src_pgd = pgd_offset(src_mm, addr);
1090 do {
1091 next = pgd_addr_end(addr, end);
1092 if (pgd_none_or_clear_bad(src_pgd))
1093 continue;
1094 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1095 vma, addr, next))) {
1096 ret = -ENOMEM;
1097 break;
1098 }
1099 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1100
1101 if (is_cow_mapping(vma->vm_flags))
1102 mmu_notifier_invalidate_range_end(src_mm,
1103 vma->vm_start, end);
1104 return ret;
1105}
1106
1107static unsigned long zap_pte_range(struct mmu_gather *tlb,
1108 struct vm_area_struct *vma, pmd_t *pmd,
1109 unsigned long addr, unsigned long end,
1110 struct zap_details *details)
1111{
1112 struct mm_struct *mm = tlb->mm;
1113 int force_flush = 0;
1114 int rss[NR_MM_COUNTERS];
1115 spinlock_t *ptl;
1116 pte_t *start_pte;
1117 pte_t *pte;
1118
1119again:
1120 init_rss_vec(rss);
1121 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1122 pte = start_pte;
1123 arch_enter_lazy_mmu_mode();
1124 do {
1125 pte_t ptent = *pte;
1126 if (pte_none(ptent)) {
1127 continue;
1128 }
1129
1130 if (pte_present(ptent)) {
1131 struct page *page;
1132
1133 page = vm_normal_page(vma, addr, ptent);
1134 if (unlikely(details) && page) {
1135 /*
1136 * unmap_shared_mapping_pages() wants to
1137 * invalidate cache without truncating:
1138 * unmap shared but keep private pages.
1139 */
1140 if (details->check_mapping &&
1141 details->check_mapping != page->mapping)
1142 continue;
1143 /*
1144 * Each page->index must be checked when
1145 * invalidating or truncating nonlinear.
1146 */
1147 if (details->nonlinear_vma &&
1148 (page->index < details->first_index ||
1149 page->index > details->last_index))
1150 continue;
1151 }
1152 ptent = ptep_get_and_clear_full(mm, addr, pte,
1153 tlb->fullmm);
1154 tlb_remove_tlb_entry(tlb, pte, addr);
1155 if (unlikely(!page))
1156 continue;
1157 if (unlikely(details) && details->nonlinear_vma
1158 && linear_page_index(details->nonlinear_vma,
1159 addr) != page->index)
1160 set_pte_at(mm, addr, pte,
1161 pgoff_to_pte(page->index));
1162 if (PageAnon(page))
1163 rss[MM_ANONPAGES]--;
1164 else {
1165 if (pte_dirty(ptent))
1166 set_page_dirty(page);
1167 if (pte_young(ptent) &&
1168 likely(!VM_SequentialReadHint(vma)))
1169 mark_page_accessed(page);
1170 rss[MM_FILEPAGES]--;
1171 }
1172 page_remove_rmap(page);
1173 if (unlikely(page_mapcount(page) < 0))
1174 print_bad_pte(vma, addr, ptent, page);
1175 force_flush = !__tlb_remove_page(tlb, page);
1176 if (force_flush)
1177 break;
1178 continue;
1179 }
1180 /*
1181 * If details->check_mapping, we leave swap entries;
1182 * if details->nonlinear_vma, we leave file entries.
1183 */
1184 if (unlikely(details))
1185 continue;
1186 if (pte_file(ptent)) {
1187 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1188 print_bad_pte(vma, addr, ptent, NULL);
1189 } else {
1190 swp_entry_t entry = pte_to_swp_entry(ptent);
1191
1192 if (!non_swap_entry(entry))
1193 rss[MM_SWAPENTS]--;
1194 if (unlikely(!free_swap_and_cache(entry)))
1195 print_bad_pte(vma, addr, ptent, NULL);
1196 }
1197 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198 } while (pte++, addr += PAGE_SIZE, addr != end);
1199
1200 add_mm_rss_vec(mm, rss);
1201 arch_leave_lazy_mmu_mode();
1202 pte_unmap_unlock(start_pte, ptl);
1203
1204 /*
1205 * mmu_gather ran out of room to batch pages, we break out of
1206 * the PTE lock to avoid doing the potential expensive TLB invalidate
1207 * and page-free while holding it.
1208 */
1209 if (force_flush) {
1210 force_flush = 0;
1211 tlb_flush_mmu(tlb);
1212 if (addr != end)
1213 goto again;
1214 }
1215
1216 return addr;
1217}
1218
1219static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1220 struct vm_area_struct *vma, pud_t *pud,
1221 unsigned long addr, unsigned long end,
1222 struct zap_details *details)
1223{
1224 pmd_t *pmd;
1225 unsigned long next;
1226
1227 pmd = pmd_offset(pud, addr);
1228 do {
1229 next = pmd_addr_end(addr, end);
1230 if (pmd_trans_huge(*pmd)) {
1231 if (next-addr != HPAGE_PMD_SIZE) {
1232 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1233 split_huge_page_pmd(vma->vm_mm, pmd);
1234 } else if (zap_huge_pmd(tlb, vma, pmd))
1235 continue;
1236 /* fall through */
1237 }
1238 if (pmd_none_or_clear_bad(pmd))
1239 continue;
1240 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1241 cond_resched();
1242 } while (pmd++, addr = next, addr != end);
1243
1244 return addr;
1245}
1246
1247static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1248 struct vm_area_struct *vma, pgd_t *pgd,
1249 unsigned long addr, unsigned long end,
1250 struct zap_details *details)
1251{
1252 pud_t *pud;
1253 unsigned long next;
1254
1255 pud = pud_offset(pgd, addr);
1256 do {
1257 next = pud_addr_end(addr, end);
1258 if (pud_none_or_clear_bad(pud))
1259 continue;
1260 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1261 } while (pud++, addr = next, addr != end);
1262
1263 return addr;
1264}
1265
1266static unsigned long unmap_page_range(struct mmu_gather *tlb,
1267 struct vm_area_struct *vma,
1268 unsigned long addr, unsigned long end,
1269 struct zap_details *details)
1270{
1271 pgd_t *pgd;
1272 unsigned long next;
1273
1274 if (details && !details->check_mapping && !details->nonlinear_vma)
1275 details = NULL;
1276
1277 BUG_ON(addr >= end);
1278 mem_cgroup_uncharge_start();
1279 tlb_start_vma(tlb, vma);
1280 pgd = pgd_offset(vma->vm_mm, addr);
1281 do {
1282 next = pgd_addr_end(addr, end);
1283 if (pgd_none_or_clear_bad(pgd))
1284 continue;
1285 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1286 } while (pgd++, addr = next, addr != end);
1287 tlb_end_vma(tlb, vma);
1288 mem_cgroup_uncharge_end();
1289
1290 return addr;
1291}
1292
1293/**
1294 * unmap_vmas - unmap a range of memory covered by a list of vma's
1295 * @tlb: address of the caller's struct mmu_gather
1296 * @vma: the starting vma
1297 * @start_addr: virtual address at which to start unmapping
1298 * @end_addr: virtual address at which to end unmapping
1299 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1300 * @details: details of nonlinear truncation or shared cache invalidation
1301 *
1302 * Returns the end address of the unmapping (restart addr if interrupted).
1303 *
1304 * Unmap all pages in the vma list.
1305 *
1306 * Only addresses between `start' and `end' will be unmapped.
1307 *
1308 * The VMA list must be sorted in ascending virtual address order.
1309 *
1310 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1311 * range after unmap_vmas() returns. So the only responsibility here is to
1312 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1313 * drops the lock and schedules.
1314 */
1315unsigned long unmap_vmas(struct mmu_gather *tlb,
1316 struct vm_area_struct *vma, unsigned long start_addr,
1317 unsigned long end_addr, unsigned long *nr_accounted,
1318 struct zap_details *details)
1319{
1320 unsigned long start = start_addr;
1321 struct mm_struct *mm = vma->vm_mm;
1322
1323 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1324 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1325 unsigned long end;
1326
1327 start = max(vma->vm_start, start_addr);
1328 if (start >= vma->vm_end)
1329 continue;
1330 end = min(vma->vm_end, end_addr);
1331 if (end <= vma->vm_start)
1332 continue;
1333
1334 if (vma->vm_flags & VM_ACCOUNT)
1335 *nr_accounted += (end - start) >> PAGE_SHIFT;
1336
1337 if (unlikely(is_pfn_mapping(vma)))
1338 untrack_pfn_vma(vma, 0, 0);
1339
1340 while (start != end) {
1341 if (unlikely(is_vm_hugetlb_page(vma))) {
1342 /*
1343 * It is undesirable to test vma->vm_file as it
1344 * should be non-null for valid hugetlb area.
1345 * However, vm_file will be NULL in the error
1346 * cleanup path of do_mmap_pgoff. When
1347 * hugetlbfs ->mmap method fails,
1348 * do_mmap_pgoff() nullifies vma->vm_file
1349 * before calling this function to clean up.
1350 * Since no pte has actually been setup, it is
1351 * safe to do nothing in this case.
1352 */
1353 if (vma->vm_file)
1354 unmap_hugepage_range(vma, start, end, NULL);
1355
1356 start = end;
1357 } else
1358 start = unmap_page_range(tlb, vma, start, end, details);
1359 }
1360 }
1361
1362 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1363 return start; /* which is now the end (or restart) address */
1364}
1365
1366/**
1367 * zap_page_range - remove user pages in a given range
1368 * @vma: vm_area_struct holding the applicable pages
1369 * @address: starting address of pages to zap
1370 * @size: number of bytes to zap
1371 * @details: details of nonlinear truncation or shared cache invalidation
1372 */
1373unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1374 unsigned long size, struct zap_details *details)
1375{
1376 struct mm_struct *mm = vma->vm_mm;
1377 struct mmu_gather tlb;
1378 unsigned long end = address + size;
1379 unsigned long nr_accounted = 0;
1380
1381 lru_add_drain();
1382 tlb_gather_mmu(&tlb, mm, 0);
1383 update_hiwater_rss(mm);
1384 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1385 tlb_finish_mmu(&tlb, address, end);
1386 return end;
1387}
1388
1389/**
1390 * zap_vma_ptes - remove ptes mapping the vma
1391 * @vma: vm_area_struct holding ptes to be zapped
1392 * @address: starting address of pages to zap
1393 * @size: number of bytes to zap
1394 *
1395 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1396 *
1397 * The entire address range must be fully contained within the vma.
1398 *
1399 * Returns 0 if successful.
1400 */
1401int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1402 unsigned long size)
1403{
1404 if (address < vma->vm_start || address + size > vma->vm_end ||
1405 !(vma->vm_flags & VM_PFNMAP))
1406 return -1;
1407 zap_page_range(vma, address, size, NULL);
1408 return 0;
1409}
1410EXPORT_SYMBOL_GPL(zap_vma_ptes);
1411
1412/**
1413 * follow_page - look up a page descriptor from a user-virtual address
1414 * @vma: vm_area_struct mapping @address
1415 * @address: virtual address to look up
1416 * @flags: flags modifying lookup behaviour
1417 *
1418 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1419 *
1420 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1421 * an error pointer if there is a mapping to something not represented
1422 * by a page descriptor (see also vm_normal_page()).
1423 */
1424struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1425 unsigned int flags)
1426{
1427 pgd_t *pgd;
1428 pud_t *pud;
1429 pmd_t *pmd;
1430 pte_t *ptep, pte;
1431 spinlock_t *ptl;
1432 struct page *page;
1433 struct mm_struct *mm = vma->vm_mm;
1434
1435 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1436 if (!IS_ERR(page)) {
1437 BUG_ON(flags & FOLL_GET);
1438 goto out;
1439 }
1440
1441 page = NULL;
1442 pgd = pgd_offset(mm, address);
1443 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1444 goto no_page_table;
1445
1446 pud = pud_offset(pgd, address);
1447 if (pud_none(*pud))
1448 goto no_page_table;
1449 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1450 BUG_ON(flags & FOLL_GET);
1451 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1452 goto out;
1453 }
1454 if (unlikely(pud_bad(*pud)))
1455 goto no_page_table;
1456
1457 pmd = pmd_offset(pud, address);
1458 if (pmd_none(*pmd))
1459 goto no_page_table;
1460 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1461 BUG_ON(flags & FOLL_GET);
1462 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1463 goto out;
1464 }
1465 if (pmd_trans_huge(*pmd)) {
1466 if (flags & FOLL_SPLIT) {
1467 split_huge_page_pmd(mm, pmd);
1468 goto split_fallthrough;
1469 }
1470 spin_lock(&mm->page_table_lock);
1471 if (likely(pmd_trans_huge(*pmd))) {
1472 if (unlikely(pmd_trans_splitting(*pmd))) {
1473 spin_unlock(&mm->page_table_lock);
1474 wait_split_huge_page(vma->anon_vma, pmd);
1475 } else {
1476 page = follow_trans_huge_pmd(mm, address,
1477 pmd, flags);
1478 spin_unlock(&mm->page_table_lock);
1479 goto out;
1480 }
1481 } else
1482 spin_unlock(&mm->page_table_lock);
1483 /* fall through */
1484 }
1485split_fallthrough:
1486 if (unlikely(pmd_bad(*pmd)))
1487 goto no_page_table;
1488
1489 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1490
1491 pte = *ptep;
1492 if (!pte_present(pte))
1493 goto no_page;
1494 if ((flags & FOLL_WRITE) && !pte_write(pte))
1495 goto unlock;
1496
1497 page = vm_normal_page(vma, address, pte);
1498 if (unlikely(!page)) {
1499 if ((flags & FOLL_DUMP) ||
1500 !is_zero_pfn(pte_pfn(pte)))
1501 goto bad_page;
1502 page = pte_page(pte);
1503 }
1504
1505 if (flags & FOLL_GET)
1506 get_page(page);
1507 if (flags & FOLL_TOUCH) {
1508 if ((flags & FOLL_WRITE) &&
1509 !pte_dirty(pte) && !PageDirty(page))
1510 set_page_dirty(page);
1511 /*
1512 * pte_mkyoung() would be more correct here, but atomic care
1513 * is needed to avoid losing the dirty bit: it is easier to use
1514 * mark_page_accessed().
1515 */
1516 mark_page_accessed(page);
1517 }
1518 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1519 /*
1520 * The preliminary mapping check is mainly to avoid the
1521 * pointless overhead of lock_page on the ZERO_PAGE
1522 * which might bounce very badly if there is contention.
1523 *
1524 * If the page is already locked, we don't need to
1525 * handle it now - vmscan will handle it later if and
1526 * when it attempts to reclaim the page.
1527 */
1528 if (page->mapping && trylock_page(page)) {
1529 lru_add_drain(); /* push cached pages to LRU */
1530 /*
1531 * Because we lock page here and migration is
1532 * blocked by the pte's page reference, we need
1533 * only check for file-cache page truncation.
1534 */
1535 if (page->mapping)
1536 mlock_vma_page(page);
1537 unlock_page(page);
1538 }
1539 }
1540unlock:
1541 pte_unmap_unlock(ptep, ptl);
1542out:
1543 return page;
1544
1545bad_page:
1546 pte_unmap_unlock(ptep, ptl);
1547 return ERR_PTR(-EFAULT);
1548
1549no_page:
1550 pte_unmap_unlock(ptep, ptl);
1551 if (!pte_none(pte))
1552 return page;
1553
1554no_page_table:
1555 /*
1556 * When core dumping an enormous anonymous area that nobody
1557 * has touched so far, we don't want to allocate unnecessary pages or
1558 * page tables. Return error instead of NULL to skip handle_mm_fault,
1559 * then get_dump_page() will return NULL to leave a hole in the dump.
1560 * But we can only make this optimization where a hole would surely
1561 * be zero-filled if handle_mm_fault() actually did handle it.
1562 */
1563 if ((flags & FOLL_DUMP) &&
1564 (!vma->vm_ops || !vma->vm_ops->fault))
1565 return ERR_PTR(-EFAULT);
1566 return page;
1567}
1568
1569static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1570{
1571 return stack_guard_page_start(vma, addr) ||
1572 stack_guard_page_end(vma, addr+PAGE_SIZE);
1573}
1574
1575/**
1576 * __get_user_pages() - pin user pages in memory
1577 * @tsk: task_struct of target task
1578 * @mm: mm_struct of target mm
1579 * @start: starting user address
1580 * @nr_pages: number of pages from start to pin
1581 * @gup_flags: flags modifying pin behaviour
1582 * @pages: array that receives pointers to the pages pinned.
1583 * Should be at least nr_pages long. Or NULL, if caller
1584 * only intends to ensure the pages are faulted in.
1585 * @vmas: array of pointers to vmas corresponding to each page.
1586 * Or NULL if the caller does not require them.
1587 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1588 *
1589 * Returns number of pages pinned. This may be fewer than the number
1590 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1591 * were pinned, returns -errno. Each page returned must be released
1592 * with a put_page() call when it is finished with. vmas will only
1593 * remain valid while mmap_sem is held.
1594 *
1595 * Must be called with mmap_sem held for read or write.
1596 *
1597 * __get_user_pages walks a process's page tables and takes a reference to
1598 * each struct page that each user address corresponds to at a given
1599 * instant. That is, it takes the page that would be accessed if a user
1600 * thread accesses the given user virtual address at that instant.
1601 *
1602 * This does not guarantee that the page exists in the user mappings when
1603 * __get_user_pages returns, and there may even be a completely different
1604 * page there in some cases (eg. if mmapped pagecache has been invalidated
1605 * and subsequently re faulted). However it does guarantee that the page
1606 * won't be freed completely. And mostly callers simply care that the page
1607 * contains data that was valid *at some point in time*. Typically, an IO
1608 * or similar operation cannot guarantee anything stronger anyway because
1609 * locks can't be held over the syscall boundary.
1610 *
1611 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1612 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1613 * appropriate) must be called after the page is finished with, and
1614 * before put_page is called.
1615 *
1616 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1617 * or mmap_sem contention, and if waiting is needed to pin all pages,
1618 * *@nonblocking will be set to 0.
1619 *
1620 * In most cases, get_user_pages or get_user_pages_fast should be used
1621 * instead of __get_user_pages. __get_user_pages should be used only if
1622 * you need some special @gup_flags.
1623 */
1624int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1625 unsigned long start, int nr_pages, unsigned int gup_flags,
1626 struct page **pages, struct vm_area_struct **vmas,
1627 int *nonblocking)
1628{
1629 int i;
1630 unsigned long vm_flags;
1631
1632 if (nr_pages <= 0)
1633 return 0;
1634
1635 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1636
1637 /*
1638 * Require read or write permissions.
1639 * If FOLL_FORCE is set, we only require the "MAY" flags.
1640 */
1641 vm_flags = (gup_flags & FOLL_WRITE) ?
1642 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1643 vm_flags &= (gup_flags & FOLL_FORCE) ?
1644 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1645 i = 0;
1646
1647 do {
1648 struct vm_area_struct *vma;
1649
1650 vma = find_extend_vma(mm, start);
1651 if (!vma && in_gate_area(mm, start)) {
1652 unsigned long pg = start & PAGE_MASK;
1653 pgd_t *pgd;
1654 pud_t *pud;
1655 pmd_t *pmd;
1656 pte_t *pte;
1657
1658 /* user gate pages are read-only */
1659 if (gup_flags & FOLL_WRITE)
1660 return i ? : -EFAULT;
1661 if (pg > TASK_SIZE)
1662 pgd = pgd_offset_k(pg);
1663 else
1664 pgd = pgd_offset_gate(mm, pg);
1665 BUG_ON(pgd_none(*pgd));
1666 pud = pud_offset(pgd, pg);
1667 BUG_ON(pud_none(*pud));
1668 pmd = pmd_offset(pud, pg);
1669 if (pmd_none(*pmd))
1670 return i ? : -EFAULT;
1671 VM_BUG_ON(pmd_trans_huge(*pmd));
1672 pte = pte_offset_map(pmd, pg);
1673 if (pte_none(*pte)) {
1674 pte_unmap(pte);
1675 return i ? : -EFAULT;
1676 }
1677 vma = get_gate_vma(mm);
1678 if (pages) {
1679 struct page *page;
1680
1681 page = vm_normal_page(vma, start, *pte);
1682 if (!page) {
1683 if (!(gup_flags & FOLL_DUMP) &&
1684 is_zero_pfn(pte_pfn(*pte)))
1685 page = pte_page(*pte);
1686 else {
1687 pte_unmap(pte);
1688 return i ? : -EFAULT;
1689 }
1690 }
1691 pages[i] = page;
1692 get_page(page);
1693 }
1694 pte_unmap(pte);
1695 goto next_page;
1696 }
1697
1698 if (!vma ||
1699 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1700 !(vm_flags & vma->vm_flags))
1701 return i ? : -EFAULT;
1702
1703 if (is_vm_hugetlb_page(vma)) {
1704 i = follow_hugetlb_page(mm, vma, pages, vmas,
1705 &start, &nr_pages, i, gup_flags);
1706 continue;
1707 }
1708
1709 do {
1710 struct page *page;
1711 unsigned int foll_flags = gup_flags;
1712
1713 /*
1714 * If we have a pending SIGKILL, don't keep faulting
1715 * pages and potentially allocating memory.
1716 */
1717 if (unlikely(fatal_signal_pending(current)))
1718 return i ? i : -ERESTARTSYS;
1719
1720 cond_resched();
1721 while (!(page = follow_page(vma, start, foll_flags))) {
1722 int ret;
1723 unsigned int fault_flags = 0;
1724
1725 /* For mlock, just skip the stack guard page. */
1726 if (foll_flags & FOLL_MLOCK) {
1727 if (stack_guard_page(vma, start))
1728 goto next_page;
1729 }
1730 if (foll_flags & FOLL_WRITE)
1731 fault_flags |= FAULT_FLAG_WRITE;
1732 if (nonblocking)
1733 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1734 if (foll_flags & FOLL_NOWAIT)
1735 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1736
1737 ret = handle_mm_fault(mm, vma, start,
1738 fault_flags);
1739
1740 if (ret & VM_FAULT_ERROR) {
1741 if (ret & VM_FAULT_OOM)
1742 return i ? i : -ENOMEM;
1743 if (ret & (VM_FAULT_HWPOISON |
1744 VM_FAULT_HWPOISON_LARGE)) {
1745 if (i)
1746 return i;
1747 else if (gup_flags & FOLL_HWPOISON)
1748 return -EHWPOISON;
1749 else
1750 return -EFAULT;
1751 }
1752 if (ret & VM_FAULT_SIGBUS)
1753 return i ? i : -EFAULT;
1754 BUG();
1755 }
1756
1757 if (tsk) {
1758 if (ret & VM_FAULT_MAJOR)
1759 tsk->maj_flt++;
1760 else
1761 tsk->min_flt++;
1762 }
1763
1764 if (ret & VM_FAULT_RETRY) {
1765 if (nonblocking)
1766 *nonblocking = 0;
1767 return i;
1768 }
1769
1770 /*
1771 * The VM_FAULT_WRITE bit tells us that
1772 * do_wp_page has broken COW when necessary,
1773 * even if maybe_mkwrite decided not to set
1774 * pte_write. We can thus safely do subsequent
1775 * page lookups as if they were reads. But only
1776 * do so when looping for pte_write is futile:
1777 * in some cases userspace may also be wanting
1778 * to write to the gotten user page, which a
1779 * read fault here might prevent (a readonly
1780 * page might get reCOWed by userspace write).
1781 */
1782 if ((ret & VM_FAULT_WRITE) &&
1783 !(vma->vm_flags & VM_WRITE))
1784 foll_flags &= ~FOLL_WRITE;
1785
1786 cond_resched();
1787 }
1788 if (IS_ERR(page))
1789 return i ? i : PTR_ERR(page);
1790 if (pages) {
1791 pages[i] = page;
1792
1793 flush_anon_page(vma, page, start);
1794 flush_dcache_page(page);
1795 }
1796next_page:
1797 if (vmas)
1798 vmas[i] = vma;
1799 i++;
1800 start += PAGE_SIZE;
1801 nr_pages--;
1802 } while (nr_pages && start < vma->vm_end);
1803 } while (nr_pages);
1804 return i;
1805}
1806EXPORT_SYMBOL(__get_user_pages);
1807
1808/*
1809 * fixup_user_fault() - manually resolve a user page fault
1810 * @tsk: the task_struct to use for page fault accounting, or
1811 * NULL if faults are not to be recorded.
1812 * @mm: mm_struct of target mm
1813 * @address: user address
1814 * @fault_flags:flags to pass down to handle_mm_fault()
1815 *
1816 * This is meant to be called in the specific scenario where for locking reasons
1817 * we try to access user memory in atomic context (within a pagefault_disable()
1818 * section), this returns -EFAULT, and we want to resolve the user fault before
1819 * trying again.
1820 *
1821 * Typically this is meant to be used by the futex code.
1822 *
1823 * The main difference with get_user_pages() is that this function will
1824 * unconditionally call handle_mm_fault() which will in turn perform all the
1825 * necessary SW fixup of the dirty and young bits in the PTE, while
1826 * handle_mm_fault() only guarantees to update these in the struct page.
1827 *
1828 * This is important for some architectures where those bits also gate the
1829 * access permission to the page because they are maintained in software. On
1830 * such architectures, gup() will not be enough to make a subsequent access
1831 * succeed.
1832 *
1833 * This should be called with the mm_sem held for read.
1834 */
1835int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1836 unsigned long address, unsigned int fault_flags)
1837{
1838 struct vm_area_struct *vma;
1839 int ret;
1840
1841 vma = find_extend_vma(mm, address);
1842 if (!vma || address < vma->vm_start)
1843 return -EFAULT;
1844
1845 ret = handle_mm_fault(mm, vma, address, fault_flags);
1846 if (ret & VM_FAULT_ERROR) {
1847 if (ret & VM_FAULT_OOM)
1848 return -ENOMEM;
1849 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1850 return -EHWPOISON;
1851 if (ret & VM_FAULT_SIGBUS)
1852 return -EFAULT;
1853 BUG();
1854 }
1855 if (tsk) {
1856 if (ret & VM_FAULT_MAJOR)
1857 tsk->maj_flt++;
1858 else
1859 tsk->min_flt++;
1860 }
1861 return 0;
1862}
1863
1864/*
1865 * get_user_pages() - pin user pages in memory
1866 * @tsk: the task_struct to use for page fault accounting, or
1867 * NULL if faults are not to be recorded.
1868 * @mm: mm_struct of target mm
1869 * @start: starting user address
1870 * @nr_pages: number of pages from start to pin
1871 * @write: whether pages will be written to by the caller
1872 * @force: whether to force write access even if user mapping is
1873 * readonly. This will result in the page being COWed even
1874 * in MAP_SHARED mappings. You do not want this.
1875 * @pages: array that receives pointers to the pages pinned.
1876 * Should be at least nr_pages long. Or NULL, if caller
1877 * only intends to ensure the pages are faulted in.
1878 * @vmas: array of pointers to vmas corresponding to each page.
1879 * Or NULL if the caller does not require them.
1880 *
1881 * Returns number of pages pinned. This may be fewer than the number
1882 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1883 * were pinned, returns -errno. Each page returned must be released
1884 * with a put_page() call when it is finished with. vmas will only
1885 * remain valid while mmap_sem is held.
1886 *
1887 * Must be called with mmap_sem held for read or write.
1888 *
1889 * get_user_pages walks a process's page tables and takes a reference to
1890 * each struct page that each user address corresponds to at a given
1891 * instant. That is, it takes the page that would be accessed if a user
1892 * thread accesses the given user virtual address at that instant.
1893 *
1894 * This does not guarantee that the page exists in the user mappings when
1895 * get_user_pages returns, and there may even be a completely different
1896 * page there in some cases (eg. if mmapped pagecache has been invalidated
1897 * and subsequently re faulted). However it does guarantee that the page
1898 * won't be freed completely. And mostly callers simply care that the page
1899 * contains data that was valid *at some point in time*. Typically, an IO
1900 * or similar operation cannot guarantee anything stronger anyway because
1901 * locks can't be held over the syscall boundary.
1902 *
1903 * If write=0, the page must not be written to. If the page is written to,
1904 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1905 * after the page is finished with, and before put_page is called.
1906 *
1907 * get_user_pages is typically used for fewer-copy IO operations, to get a
1908 * handle on the memory by some means other than accesses via the user virtual
1909 * addresses. The pages may be submitted for DMA to devices or accessed via
1910 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1911 * use the correct cache flushing APIs.
1912 *
1913 * See also get_user_pages_fast, for performance critical applications.
1914 */
1915int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1916 unsigned long start, int nr_pages, int write, int force,
1917 struct page **pages, struct vm_area_struct **vmas)
1918{
1919 int flags = FOLL_TOUCH;
1920
1921 if (pages)
1922 flags |= FOLL_GET;
1923 if (write)
1924 flags |= FOLL_WRITE;
1925 if (force)
1926 flags |= FOLL_FORCE;
1927
1928 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1929 NULL);
1930}
1931EXPORT_SYMBOL(get_user_pages);
1932
1933/**
1934 * get_dump_page() - pin user page in memory while writing it to core dump
1935 * @addr: user address
1936 *
1937 * Returns struct page pointer of user page pinned for dump,
1938 * to be freed afterwards by page_cache_release() or put_page().
1939 *
1940 * Returns NULL on any kind of failure - a hole must then be inserted into
1941 * the corefile, to preserve alignment with its headers; and also returns
1942 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1943 * allowing a hole to be left in the corefile to save diskspace.
1944 *
1945 * Called without mmap_sem, but after all other threads have been killed.
1946 */
1947#ifdef CONFIG_ELF_CORE
1948struct page *get_dump_page(unsigned long addr)
1949{
1950 struct vm_area_struct *vma;
1951 struct page *page;
1952
1953 if (__get_user_pages(current, current->mm, addr, 1,
1954 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1955 NULL) < 1)
1956 return NULL;
1957 flush_cache_page(vma, addr, page_to_pfn(page));
1958 return page;
1959}
1960#endif /* CONFIG_ELF_CORE */
1961
1962pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1963 spinlock_t **ptl)
1964{
1965 pgd_t * pgd = pgd_offset(mm, addr);
1966 pud_t * pud = pud_alloc(mm, pgd, addr);
1967 if (pud) {
1968 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1969 if (pmd) {
1970 VM_BUG_ON(pmd_trans_huge(*pmd));
1971 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1972 }
1973 }
1974 return NULL;
1975}
1976
1977/*
1978 * This is the old fallback for page remapping.
1979 *
1980 * For historical reasons, it only allows reserved pages. Only
1981 * old drivers should use this, and they needed to mark their
1982 * pages reserved for the old functions anyway.
1983 */
1984static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1985 struct page *page, pgprot_t prot)
1986{
1987 struct mm_struct *mm = vma->vm_mm;
1988 int retval;
1989 pte_t *pte;
1990 spinlock_t *ptl;
1991
1992 retval = -EINVAL;
1993 if (PageAnon(page))
1994 goto out;
1995 retval = -ENOMEM;
1996 flush_dcache_page(page);
1997 pte = get_locked_pte(mm, addr, &ptl);
1998 if (!pte)
1999 goto out;
2000 retval = -EBUSY;
2001 if (!pte_none(*pte))
2002 goto out_unlock;
2003
2004 /* Ok, finally just insert the thing.. */
2005 get_page(page);
2006 inc_mm_counter_fast(mm, MM_FILEPAGES);
2007 page_add_file_rmap(page);
2008 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2009
2010 retval = 0;
2011 pte_unmap_unlock(pte, ptl);
2012 return retval;
2013out_unlock:
2014 pte_unmap_unlock(pte, ptl);
2015out:
2016 return retval;
2017}
2018
2019/**
2020 * vm_insert_page - insert single page into user vma
2021 * @vma: user vma to map to
2022 * @addr: target user address of this page
2023 * @page: source kernel page
2024 *
2025 * This allows drivers to insert individual pages they've allocated
2026 * into a user vma.
2027 *
2028 * The page has to be a nice clean _individual_ kernel allocation.
2029 * If you allocate a compound page, you need to have marked it as
2030 * such (__GFP_COMP), or manually just split the page up yourself
2031 * (see split_page()).
2032 *
2033 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2034 * took an arbitrary page protection parameter. This doesn't allow
2035 * that. Your vma protection will have to be set up correctly, which
2036 * means that if you want a shared writable mapping, you'd better
2037 * ask for a shared writable mapping!
2038 *
2039 * The page does not need to be reserved.
2040 */
2041int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2042 struct page *page)
2043{
2044 if (addr < vma->vm_start || addr >= vma->vm_end)
2045 return -EFAULT;
2046 if (!page_count(page))
2047 return -EINVAL;
2048 vma->vm_flags |= VM_INSERTPAGE;
2049 return insert_page(vma, addr, page, vma->vm_page_prot);
2050}
2051EXPORT_SYMBOL(vm_insert_page);
2052
2053static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2054 unsigned long pfn, pgprot_t prot)
2055{
2056 struct mm_struct *mm = vma->vm_mm;
2057 int retval;
2058 pte_t *pte, entry;
2059 spinlock_t *ptl;
2060
2061 retval = -ENOMEM;
2062 pte = get_locked_pte(mm, addr, &ptl);
2063 if (!pte)
2064 goto out;
2065 retval = -EBUSY;
2066 if (!pte_none(*pte))
2067 goto out_unlock;
2068
2069 /* Ok, finally just insert the thing.. */
2070 entry = pte_mkspecial(pfn_pte(pfn, prot));
2071 set_pte_at(mm, addr, pte, entry);
2072 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2073
2074 retval = 0;
2075out_unlock:
2076 pte_unmap_unlock(pte, ptl);
2077out:
2078 return retval;
2079}
2080
2081/**
2082 * vm_insert_pfn - insert single pfn into user vma
2083 * @vma: user vma to map to
2084 * @addr: target user address of this page
2085 * @pfn: source kernel pfn
2086 *
2087 * Similar to vm_inert_page, this allows drivers to insert individual pages
2088 * they've allocated into a user vma. Same comments apply.
2089 *
2090 * This function should only be called from a vm_ops->fault handler, and
2091 * in that case the handler should return NULL.
2092 *
2093 * vma cannot be a COW mapping.
2094 *
2095 * As this is called only for pages that do not currently exist, we
2096 * do not need to flush old virtual caches or the TLB.
2097 */
2098int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2099 unsigned long pfn)
2100{
2101 int ret;
2102 pgprot_t pgprot = vma->vm_page_prot;
2103 /*
2104 * Technically, architectures with pte_special can avoid all these
2105 * restrictions (same for remap_pfn_range). However we would like
2106 * consistency in testing and feature parity among all, so we should
2107 * try to keep these invariants in place for everybody.
2108 */
2109 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2110 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2111 (VM_PFNMAP|VM_MIXEDMAP));
2112 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2113 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2114
2115 if (addr < vma->vm_start || addr >= vma->vm_end)
2116 return -EFAULT;
2117 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2118 return -EINVAL;
2119
2120 ret = insert_pfn(vma, addr, pfn, pgprot);
2121
2122 if (ret)
2123 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2124
2125 return ret;
2126}
2127EXPORT_SYMBOL(vm_insert_pfn);
2128
2129int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2130 unsigned long pfn)
2131{
2132 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2133
2134 if (addr < vma->vm_start || addr >= vma->vm_end)
2135 return -EFAULT;
2136
2137 /*
2138 * If we don't have pte special, then we have to use the pfn_valid()
2139 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2140 * refcount the page if pfn_valid is true (hence insert_page rather
2141 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2142 * without pte special, it would there be refcounted as a normal page.
2143 */
2144 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2145 struct page *page;
2146
2147 page = pfn_to_page(pfn);
2148 return insert_page(vma, addr, page, vma->vm_page_prot);
2149 }
2150 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2151}
2152EXPORT_SYMBOL(vm_insert_mixed);
2153
2154/*
2155 * maps a range of physical memory into the requested pages. the old
2156 * mappings are removed. any references to nonexistent pages results
2157 * in null mappings (currently treated as "copy-on-access")
2158 */
2159static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2160 unsigned long addr, unsigned long end,
2161 unsigned long pfn, pgprot_t prot)
2162{
2163 pte_t *pte;
2164 spinlock_t *ptl;
2165
2166 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2167 if (!pte)
2168 return -ENOMEM;
2169 arch_enter_lazy_mmu_mode();
2170 do {
2171 BUG_ON(!pte_none(*pte));
2172 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2173 pfn++;
2174 } while (pte++, addr += PAGE_SIZE, addr != end);
2175 arch_leave_lazy_mmu_mode();
2176 pte_unmap_unlock(pte - 1, ptl);
2177 return 0;
2178}
2179
2180static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2181 unsigned long addr, unsigned long end,
2182 unsigned long pfn, pgprot_t prot)
2183{
2184 pmd_t *pmd;
2185 unsigned long next;
2186
2187 pfn -= addr >> PAGE_SHIFT;
2188 pmd = pmd_alloc(mm, pud, addr);
2189 if (!pmd)
2190 return -ENOMEM;
2191 VM_BUG_ON(pmd_trans_huge(*pmd));
2192 do {
2193 next = pmd_addr_end(addr, end);
2194 if (remap_pte_range(mm, pmd, addr, next,
2195 pfn + (addr >> PAGE_SHIFT), prot))
2196 return -ENOMEM;
2197 } while (pmd++, addr = next, addr != end);
2198 return 0;
2199}
2200
2201static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2202 unsigned long addr, unsigned long end,
2203 unsigned long pfn, pgprot_t prot)
2204{
2205 pud_t *pud;
2206 unsigned long next;
2207
2208 pfn -= addr >> PAGE_SHIFT;
2209 pud = pud_alloc(mm, pgd, addr);
2210 if (!pud)
2211 return -ENOMEM;
2212 do {
2213 next = pud_addr_end(addr, end);
2214 if (remap_pmd_range(mm, pud, addr, next,
2215 pfn + (addr >> PAGE_SHIFT), prot))
2216 return -ENOMEM;
2217 } while (pud++, addr = next, addr != end);
2218 return 0;
2219}
2220
2221/**
2222 * remap_pfn_range - remap kernel memory to userspace
2223 * @vma: user vma to map to
2224 * @addr: target user address to start at
2225 * @pfn: physical address of kernel memory
2226 * @size: size of map area
2227 * @prot: page protection flags for this mapping
2228 *
2229 * Note: this is only safe if the mm semaphore is held when called.
2230 */
2231int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2232 unsigned long pfn, unsigned long size, pgprot_t prot)
2233{
2234 pgd_t *pgd;
2235 unsigned long next;
2236 unsigned long end = addr + PAGE_ALIGN(size);
2237 struct mm_struct *mm = vma->vm_mm;
2238 int err;
2239
2240 /*
2241 * Physically remapped pages are special. Tell the
2242 * rest of the world about it:
2243 * VM_IO tells people not to look at these pages
2244 * (accesses can have side effects).
2245 * VM_RESERVED is specified all over the place, because
2246 * in 2.4 it kept swapout's vma scan off this vma; but
2247 * in 2.6 the LRU scan won't even find its pages, so this
2248 * flag means no more than count its pages in reserved_vm,
2249 * and omit it from core dump, even when VM_IO turned off.
2250 * VM_PFNMAP tells the core MM that the base pages are just
2251 * raw PFN mappings, and do not have a "struct page" associated
2252 * with them.
2253 *
2254 * There's a horrible special case to handle copy-on-write
2255 * behaviour that some programs depend on. We mark the "original"
2256 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2257 */
2258 if (addr == vma->vm_start && end == vma->vm_end) {
2259 vma->vm_pgoff = pfn;
2260 vma->vm_flags |= VM_PFN_AT_MMAP;
2261 } else if (is_cow_mapping(vma->vm_flags))
2262 return -EINVAL;
2263
2264 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2265
2266 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2267 if (err) {
2268 /*
2269 * To indicate that track_pfn related cleanup is not
2270 * needed from higher level routine calling unmap_vmas
2271 */
2272 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2273 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2274 return -EINVAL;
2275 }
2276
2277 BUG_ON(addr >= end);
2278 pfn -= addr >> PAGE_SHIFT;
2279 pgd = pgd_offset(mm, addr);
2280 flush_cache_range(vma, addr, end);
2281 do {
2282 next = pgd_addr_end(addr, end);
2283 err = remap_pud_range(mm, pgd, addr, next,
2284 pfn + (addr >> PAGE_SHIFT), prot);
2285 if (err)
2286 break;
2287 } while (pgd++, addr = next, addr != end);
2288
2289 if (err)
2290 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2291
2292 return err;
2293}
2294EXPORT_SYMBOL(remap_pfn_range);
2295
2296static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2297 unsigned long addr, unsigned long end,
2298 pte_fn_t fn, void *data)
2299{
2300 pte_t *pte;
2301 int err;
2302 pgtable_t token;
2303 spinlock_t *uninitialized_var(ptl);
2304
2305 pte = (mm == &init_mm) ?
2306 pte_alloc_kernel(pmd, addr) :
2307 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2308 if (!pte)
2309 return -ENOMEM;
2310
2311 BUG_ON(pmd_huge(*pmd));
2312
2313 arch_enter_lazy_mmu_mode();
2314
2315 token = pmd_pgtable(*pmd);
2316
2317 do {
2318 err = fn(pte++, token, addr, data);
2319 if (err)
2320 break;
2321 } while (addr += PAGE_SIZE, addr != end);
2322
2323 arch_leave_lazy_mmu_mode();
2324
2325 if (mm != &init_mm)
2326 pte_unmap_unlock(pte-1, ptl);
2327 return err;
2328}
2329
2330static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2331 unsigned long addr, unsigned long end,
2332 pte_fn_t fn, void *data)
2333{
2334 pmd_t *pmd;
2335 unsigned long next;
2336 int err;
2337
2338 BUG_ON(pud_huge(*pud));
2339
2340 pmd = pmd_alloc(mm, pud, addr);
2341 if (!pmd)
2342 return -ENOMEM;
2343 do {
2344 next = pmd_addr_end(addr, end);
2345 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2346 if (err)
2347 break;
2348 } while (pmd++, addr = next, addr != end);
2349 return err;
2350}
2351
2352static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2353 unsigned long addr, unsigned long end,
2354 pte_fn_t fn, void *data)
2355{
2356 pud_t *pud;
2357 unsigned long next;
2358 int err;
2359
2360 pud = pud_alloc(mm, pgd, addr);
2361 if (!pud)
2362 return -ENOMEM;
2363 do {
2364 next = pud_addr_end(addr, end);
2365 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2366 if (err)
2367 break;
2368 } while (pud++, addr = next, addr != end);
2369 return err;
2370}
2371
2372/*
2373 * Scan a region of virtual memory, filling in page tables as necessary
2374 * and calling a provided function on each leaf page table.
2375 */
2376int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2377 unsigned long size, pte_fn_t fn, void *data)
2378{
2379 pgd_t *pgd;
2380 unsigned long next;
2381 unsigned long end = addr + size;
2382 int err;
2383
2384 BUG_ON(addr >= end);
2385 pgd = pgd_offset(mm, addr);
2386 do {
2387 next = pgd_addr_end(addr, end);
2388 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2389 if (err)
2390 break;
2391 } while (pgd++, addr = next, addr != end);
2392
2393 return err;
2394}
2395EXPORT_SYMBOL_GPL(apply_to_page_range);
2396
2397/*
2398 * handle_pte_fault chooses page fault handler according to an entry
2399 * which was read non-atomically. Before making any commitment, on
2400 * those architectures or configurations (e.g. i386 with PAE) which
2401 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2402 * must check under lock before unmapping the pte and proceeding
2403 * (but do_wp_page is only called after already making such a check;
2404 * and do_anonymous_page can safely check later on).
2405 */
2406static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2407 pte_t *page_table, pte_t orig_pte)
2408{
2409 int same = 1;
2410#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2411 if (sizeof(pte_t) > sizeof(unsigned long)) {
2412 spinlock_t *ptl = pte_lockptr(mm, pmd);
2413 spin_lock(ptl);
2414 same = pte_same(*page_table, orig_pte);
2415 spin_unlock(ptl);
2416 }
2417#endif
2418 pte_unmap(page_table);
2419 return same;
2420}
2421
2422static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2423{
2424 /*
2425 * If the source page was a PFN mapping, we don't have
2426 * a "struct page" for it. We do a best-effort copy by
2427 * just copying from the original user address. If that
2428 * fails, we just zero-fill it. Live with it.
2429 */
2430 if (unlikely(!src)) {
2431 void *kaddr = kmap_atomic(dst, KM_USER0);
2432 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2433
2434 /*
2435 * This really shouldn't fail, because the page is there
2436 * in the page tables. But it might just be unreadable,
2437 * in which case we just give up and fill the result with
2438 * zeroes.
2439 */
2440 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2441 clear_page(kaddr);
2442 kunmap_atomic(kaddr, KM_USER0);
2443 flush_dcache_page(dst);
2444 } else
2445 copy_user_highpage(dst, src, va, vma);
2446}
2447
2448/*
2449 * This routine handles present pages, when users try to write
2450 * to a shared page. It is done by copying the page to a new address
2451 * and decrementing the shared-page counter for the old page.
2452 *
2453 * Note that this routine assumes that the protection checks have been
2454 * done by the caller (the low-level page fault routine in most cases).
2455 * Thus we can safely just mark it writable once we've done any necessary
2456 * COW.
2457 *
2458 * We also mark the page dirty at this point even though the page will
2459 * change only once the write actually happens. This avoids a few races,
2460 * and potentially makes it more efficient.
2461 *
2462 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2463 * but allow concurrent faults), with pte both mapped and locked.
2464 * We return with mmap_sem still held, but pte unmapped and unlocked.
2465 */
2466static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2467 unsigned long address, pte_t *page_table, pmd_t *pmd,
2468 spinlock_t *ptl, pte_t orig_pte)
2469 __releases(ptl)
2470{
2471 struct page *old_page, *new_page;
2472 pte_t entry;
2473 int ret = 0;
2474 int page_mkwrite = 0;
2475 struct page *dirty_page = NULL;
2476
2477 old_page = vm_normal_page(vma, address, orig_pte);
2478 if (!old_page) {
2479 /*
2480 * VM_MIXEDMAP !pfn_valid() case
2481 *
2482 * We should not cow pages in a shared writeable mapping.
2483 * Just mark the pages writable as we can't do any dirty
2484 * accounting on raw pfn maps.
2485 */
2486 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2487 (VM_WRITE|VM_SHARED))
2488 goto reuse;
2489 goto gotten;
2490 }
2491
2492 /*
2493 * Take out anonymous pages first, anonymous shared vmas are
2494 * not dirty accountable.
2495 */
2496 if (PageAnon(old_page) && !PageKsm(old_page)) {
2497 if (!trylock_page(old_page)) {
2498 page_cache_get(old_page);
2499 pte_unmap_unlock(page_table, ptl);
2500 lock_page(old_page);
2501 page_table = pte_offset_map_lock(mm, pmd, address,
2502 &ptl);
2503 if (!pte_same(*page_table, orig_pte)) {
2504 unlock_page(old_page);
2505 goto unlock;
2506 }
2507 page_cache_release(old_page);
2508 }
2509 if (reuse_swap_page(old_page)) {
2510 /*
2511 * The page is all ours. Move it to our anon_vma so
2512 * the rmap code will not search our parent or siblings.
2513 * Protected against the rmap code by the page lock.
2514 */
2515 page_move_anon_rmap(old_page, vma, address);
2516 unlock_page(old_page);
2517 goto reuse;
2518 }
2519 unlock_page(old_page);
2520 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2521 (VM_WRITE|VM_SHARED))) {
2522 /*
2523 * Only catch write-faults on shared writable pages,
2524 * read-only shared pages can get COWed by
2525 * get_user_pages(.write=1, .force=1).
2526 */
2527 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2528 struct vm_fault vmf;
2529 int tmp;
2530
2531 vmf.virtual_address = (void __user *)(address &
2532 PAGE_MASK);
2533 vmf.pgoff = old_page->index;
2534 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2535 vmf.page = old_page;
2536
2537 /*
2538 * Notify the address space that the page is about to
2539 * become writable so that it can prohibit this or wait
2540 * for the page to get into an appropriate state.
2541 *
2542 * We do this without the lock held, so that it can
2543 * sleep if it needs to.
2544 */
2545 page_cache_get(old_page);
2546 pte_unmap_unlock(page_table, ptl);
2547
2548 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2549 if (unlikely(tmp &
2550 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2551 ret = tmp;
2552 goto unwritable_page;
2553 }
2554 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2555 lock_page(old_page);
2556 if (!old_page->mapping) {
2557 ret = 0; /* retry the fault */
2558 unlock_page(old_page);
2559 goto unwritable_page;
2560 }
2561 } else
2562 VM_BUG_ON(!PageLocked(old_page));
2563
2564 /*
2565 * Since we dropped the lock we need to revalidate
2566 * the PTE as someone else may have changed it. If
2567 * they did, we just return, as we can count on the
2568 * MMU to tell us if they didn't also make it writable.
2569 */
2570 page_table = pte_offset_map_lock(mm, pmd, address,
2571 &ptl);
2572 if (!pte_same(*page_table, orig_pte)) {
2573 unlock_page(old_page);
2574 goto unlock;
2575 }
2576
2577 page_mkwrite = 1;
2578 }
2579 dirty_page = old_page;
2580 get_page(dirty_page);
2581
2582reuse:
2583 flush_cache_page(vma, address, pte_pfn(orig_pte));
2584 entry = pte_mkyoung(orig_pte);
2585 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2586 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2587 update_mmu_cache(vma, address, page_table);
2588 pte_unmap_unlock(page_table, ptl);
2589 ret |= VM_FAULT_WRITE;
2590
2591 if (!dirty_page)
2592 return ret;
2593
2594 /*
2595 * Yes, Virginia, this is actually required to prevent a race
2596 * with clear_page_dirty_for_io() from clearing the page dirty
2597 * bit after it clear all dirty ptes, but before a racing
2598 * do_wp_page installs a dirty pte.
2599 *
2600 * __do_fault is protected similarly.
2601 */
2602 if (!page_mkwrite) {
2603 wait_on_page_locked(dirty_page);
2604 set_page_dirty_balance(dirty_page, page_mkwrite);
2605 }
2606 put_page(dirty_page);
2607 if (page_mkwrite) {
2608 struct address_space *mapping = dirty_page->mapping;
2609
2610 set_page_dirty(dirty_page);
2611 unlock_page(dirty_page);
2612 page_cache_release(dirty_page);
2613 if (mapping) {
2614 /*
2615 * Some device drivers do not set page.mapping
2616 * but still dirty their pages
2617 */
2618 balance_dirty_pages_ratelimited(mapping);
2619 }
2620 }
2621
2622 /* file_update_time outside page_lock */
2623 if (vma->vm_file)
2624 file_update_time(vma->vm_file);
2625
2626 return ret;
2627 }
2628
2629 /*
2630 * Ok, we need to copy. Oh, well..
2631 */
2632 page_cache_get(old_page);
2633gotten:
2634 pte_unmap_unlock(page_table, ptl);
2635
2636 if (unlikely(anon_vma_prepare(vma)))
2637 goto oom;
2638
2639 if (is_zero_pfn(pte_pfn(orig_pte))) {
2640 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2641 if (!new_page)
2642 goto oom;
2643 } else {
2644 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2645 if (!new_page)
2646 goto oom;
2647 cow_user_page(new_page, old_page, address, vma);
2648 }
2649 __SetPageUptodate(new_page);
2650
2651 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2652 goto oom_free_new;
2653
2654 /*
2655 * Re-check the pte - we dropped the lock
2656 */
2657 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2658 if (likely(pte_same(*page_table, orig_pte))) {
2659 if (old_page) {
2660 if (!PageAnon(old_page)) {
2661 dec_mm_counter_fast(mm, MM_FILEPAGES);
2662 inc_mm_counter_fast(mm, MM_ANONPAGES);
2663 }
2664 } else
2665 inc_mm_counter_fast(mm, MM_ANONPAGES);
2666 flush_cache_page(vma, address, pte_pfn(orig_pte));
2667 entry = mk_pte(new_page, vma->vm_page_prot);
2668 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2669 /*
2670 * Clear the pte entry and flush it first, before updating the
2671 * pte with the new entry. This will avoid a race condition
2672 * seen in the presence of one thread doing SMC and another
2673 * thread doing COW.
2674 */
2675 ptep_clear_flush(vma, address, page_table);
2676 page_add_new_anon_rmap(new_page, vma, address);
2677 /*
2678 * We call the notify macro here because, when using secondary
2679 * mmu page tables (such as kvm shadow page tables), we want the
2680 * new page to be mapped directly into the secondary page table.
2681 */
2682 set_pte_at_notify(mm, address, page_table, entry);
2683 update_mmu_cache(vma, address, page_table);
2684 if (old_page) {
2685 /*
2686 * Only after switching the pte to the new page may
2687 * we remove the mapcount here. Otherwise another
2688 * process may come and find the rmap count decremented
2689 * before the pte is switched to the new page, and
2690 * "reuse" the old page writing into it while our pte
2691 * here still points into it and can be read by other
2692 * threads.
2693 *
2694 * The critical issue is to order this
2695 * page_remove_rmap with the ptp_clear_flush above.
2696 * Those stores are ordered by (if nothing else,)
2697 * the barrier present in the atomic_add_negative
2698 * in page_remove_rmap.
2699 *
2700 * Then the TLB flush in ptep_clear_flush ensures that
2701 * no process can access the old page before the
2702 * decremented mapcount is visible. And the old page
2703 * cannot be reused until after the decremented
2704 * mapcount is visible. So transitively, TLBs to
2705 * old page will be flushed before it can be reused.
2706 */
2707 page_remove_rmap(old_page);
2708 }
2709
2710 /* Free the old page.. */
2711 new_page = old_page;
2712 ret |= VM_FAULT_WRITE;
2713 } else
2714 mem_cgroup_uncharge_page(new_page);
2715
2716 if (new_page)
2717 page_cache_release(new_page);
2718unlock:
2719 pte_unmap_unlock(page_table, ptl);
2720 if (old_page) {
2721 /*
2722 * Don't let another task, with possibly unlocked vma,
2723 * keep the mlocked page.
2724 */
2725 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2726 lock_page(old_page); /* LRU manipulation */
2727 munlock_vma_page(old_page);
2728 unlock_page(old_page);
2729 }
2730 page_cache_release(old_page);
2731 }
2732 return ret;
2733oom_free_new:
2734 page_cache_release(new_page);
2735oom:
2736 if (old_page) {
2737 if (page_mkwrite) {
2738 unlock_page(old_page);
2739 page_cache_release(old_page);
2740 }
2741 page_cache_release(old_page);
2742 }
2743 return VM_FAULT_OOM;
2744
2745unwritable_page:
2746 page_cache_release(old_page);
2747 return ret;
2748}
2749
2750static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2751 unsigned long start_addr, unsigned long end_addr,
2752 struct zap_details *details)
2753{
2754 zap_page_range(vma, start_addr, end_addr - start_addr, details);
2755}
2756
2757static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2758 struct zap_details *details)
2759{
2760 struct vm_area_struct *vma;
2761 struct prio_tree_iter iter;
2762 pgoff_t vba, vea, zba, zea;
2763
2764 vma_prio_tree_foreach(vma, &iter, root,
2765 details->first_index, details->last_index) {
2766
2767 vba = vma->vm_pgoff;
2768 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2769 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2770 zba = details->first_index;
2771 if (zba < vba)
2772 zba = vba;
2773 zea = details->last_index;
2774 if (zea > vea)
2775 zea = vea;
2776
2777 unmap_mapping_range_vma(vma,
2778 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2779 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2780 details);
2781 }
2782}
2783
2784static inline void unmap_mapping_range_list(struct list_head *head,
2785 struct zap_details *details)
2786{
2787 struct vm_area_struct *vma;
2788
2789 /*
2790 * In nonlinear VMAs there is no correspondence between virtual address
2791 * offset and file offset. So we must perform an exhaustive search
2792 * across *all* the pages in each nonlinear VMA, not just the pages
2793 * whose virtual address lies outside the file truncation point.
2794 */
2795 list_for_each_entry(vma, head, shared.vm_set.list) {
2796 details->nonlinear_vma = vma;
2797 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2798 }
2799}
2800
2801/**
2802 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2803 * @mapping: the address space containing mmaps to be unmapped.
2804 * @holebegin: byte in first page to unmap, relative to the start of
2805 * the underlying file. This will be rounded down to a PAGE_SIZE
2806 * boundary. Note that this is different from truncate_pagecache(), which
2807 * must keep the partial page. In contrast, we must get rid of
2808 * partial pages.
2809 * @holelen: size of prospective hole in bytes. This will be rounded
2810 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2811 * end of the file.
2812 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2813 * but 0 when invalidating pagecache, don't throw away private data.
2814 */
2815void unmap_mapping_range(struct address_space *mapping,
2816 loff_t const holebegin, loff_t const holelen, int even_cows)
2817{
2818 struct zap_details details;
2819 pgoff_t hba = holebegin >> PAGE_SHIFT;
2820 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2821
2822 /* Check for overflow. */
2823 if (sizeof(holelen) > sizeof(hlen)) {
2824 long long holeend =
2825 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2826 if (holeend & ~(long long)ULONG_MAX)
2827 hlen = ULONG_MAX - hba + 1;
2828 }
2829
2830 details.check_mapping = even_cows? NULL: mapping;
2831 details.nonlinear_vma = NULL;
2832 details.first_index = hba;
2833 details.last_index = hba + hlen - 1;
2834 if (details.last_index < details.first_index)
2835 details.last_index = ULONG_MAX;
2836
2837
2838 mutex_lock(&mapping->i_mmap_mutex);
2839 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2840 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2841 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2842 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2843 mutex_unlock(&mapping->i_mmap_mutex);
2844}
2845EXPORT_SYMBOL(unmap_mapping_range);
2846
2847/*
2848 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2849 * but allow concurrent faults), and pte mapped but not yet locked.
2850 * We return with mmap_sem still held, but pte unmapped and unlocked.
2851 */
2852static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2853 unsigned long address, pte_t *page_table, pmd_t *pmd,
2854 unsigned int flags, pte_t orig_pte)
2855{
2856 spinlock_t *ptl;
2857 struct page *page, *swapcache = NULL;
2858 swp_entry_t entry;
2859 pte_t pte;
2860 int locked;
2861 struct mem_cgroup *ptr;
2862 int exclusive = 0;
2863 int ret = 0;
2864
2865 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2866 goto out;
2867
2868 entry = pte_to_swp_entry(orig_pte);
2869 if (unlikely(non_swap_entry(entry))) {
2870 if (is_migration_entry(entry)) {
2871 migration_entry_wait(mm, pmd, address);
2872 } else if (is_hwpoison_entry(entry)) {
2873 ret = VM_FAULT_HWPOISON;
2874 } else {
2875 print_bad_pte(vma, address, orig_pte, NULL);
2876 ret = VM_FAULT_SIGBUS;
2877 }
2878 goto out;
2879 }
2880 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2881 page = lookup_swap_cache(entry);
2882 if (!page) {
2883 grab_swap_token(mm); /* Contend for token _before_ read-in */
2884 page = swapin_readahead(entry,
2885 GFP_HIGHUSER_MOVABLE, vma, address);
2886 if (!page) {
2887 /*
2888 * Back out if somebody else faulted in this pte
2889 * while we released the pte lock.
2890 */
2891 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2892 if (likely(pte_same(*page_table, orig_pte)))
2893 ret = VM_FAULT_OOM;
2894 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2895 goto unlock;
2896 }
2897
2898 /* Had to read the page from swap area: Major fault */
2899 ret = VM_FAULT_MAJOR;
2900 count_vm_event(PGMAJFAULT);
2901 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2902 } else if (PageHWPoison(page)) {
2903 /*
2904 * hwpoisoned dirty swapcache pages are kept for killing
2905 * owner processes (which may be unknown at hwpoison time)
2906 */
2907 ret = VM_FAULT_HWPOISON;
2908 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2909 goto out_release;
2910 }
2911
2912 locked = lock_page_or_retry(page, mm, flags);
2913 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2914 if (!locked) {
2915 ret |= VM_FAULT_RETRY;
2916 goto out_release;
2917 }
2918
2919 /*
2920 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2921 * release the swapcache from under us. The page pin, and pte_same
2922 * test below, are not enough to exclude that. Even if it is still
2923 * swapcache, we need to check that the page's swap has not changed.
2924 */
2925 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2926 goto out_page;
2927
2928 if (ksm_might_need_to_copy(page, vma, address)) {
2929 swapcache = page;
2930 page = ksm_does_need_to_copy(page, vma, address);
2931
2932 if (unlikely(!page)) {
2933 ret = VM_FAULT_OOM;
2934 page = swapcache;
2935 swapcache = NULL;
2936 goto out_page;
2937 }
2938 }
2939
2940 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2941 ret = VM_FAULT_OOM;
2942 goto out_page;
2943 }
2944
2945 /*
2946 * Back out if somebody else already faulted in this pte.
2947 */
2948 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2949 if (unlikely(!pte_same(*page_table, orig_pte)))
2950 goto out_nomap;
2951
2952 if (unlikely(!PageUptodate(page))) {
2953 ret = VM_FAULT_SIGBUS;
2954 goto out_nomap;
2955 }
2956
2957 /*
2958 * The page isn't present yet, go ahead with the fault.
2959 *
2960 * Be careful about the sequence of operations here.
2961 * To get its accounting right, reuse_swap_page() must be called
2962 * while the page is counted on swap but not yet in mapcount i.e.
2963 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2964 * must be called after the swap_free(), or it will never succeed.
2965 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2966 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2967 * in page->private. In this case, a record in swap_cgroup is silently
2968 * discarded at swap_free().
2969 */
2970
2971 inc_mm_counter_fast(mm, MM_ANONPAGES);
2972 dec_mm_counter_fast(mm, MM_SWAPENTS);
2973 pte = mk_pte(page, vma->vm_page_prot);
2974 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2975 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2976 flags &= ~FAULT_FLAG_WRITE;
2977 ret |= VM_FAULT_WRITE;
2978 exclusive = 1;
2979 }
2980 flush_icache_page(vma, page);
2981 set_pte_at(mm, address, page_table, pte);
2982 do_page_add_anon_rmap(page, vma, address, exclusive);
2983 /* It's better to call commit-charge after rmap is established */
2984 mem_cgroup_commit_charge_swapin(page, ptr);
2985
2986 swap_free(entry);
2987 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2988 try_to_free_swap(page);
2989 unlock_page(page);
2990 if (swapcache) {
2991 /*
2992 * Hold the lock to avoid the swap entry to be reused
2993 * until we take the PT lock for the pte_same() check
2994 * (to avoid false positives from pte_same). For
2995 * further safety release the lock after the swap_free
2996 * so that the swap count won't change under a
2997 * parallel locked swapcache.
2998 */
2999 unlock_page(swapcache);
3000 page_cache_release(swapcache);
3001 }
3002
3003 if (flags & FAULT_FLAG_WRITE) {
3004 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3005 if (ret & VM_FAULT_ERROR)
3006 ret &= VM_FAULT_ERROR;
3007 goto out;
3008 }
3009
3010 /* No need to invalidate - it was non-present before */
3011 update_mmu_cache(vma, address, page_table);
3012unlock:
3013 pte_unmap_unlock(page_table, ptl);
3014out:
3015 return ret;
3016out_nomap:
3017 mem_cgroup_cancel_charge_swapin(ptr);
3018 pte_unmap_unlock(page_table, ptl);
3019out_page:
3020 unlock_page(page);
3021out_release:
3022 page_cache_release(page);
3023 if (swapcache) {
3024 unlock_page(swapcache);
3025 page_cache_release(swapcache);
3026 }
3027 return ret;
3028}
3029
3030/*
3031 * This is like a special single-page "expand_{down|up}wards()",
3032 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3033 * doesn't hit another vma.
3034 */
3035static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3036{
3037 address &= PAGE_MASK;
3038 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3039 struct vm_area_struct *prev = vma->vm_prev;
3040
3041 /*
3042 * Is there a mapping abutting this one below?
3043 *
3044 * That's only ok if it's the same stack mapping
3045 * that has gotten split..
3046 */
3047 if (prev && prev->vm_end == address)
3048 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3049
3050 expand_downwards(vma, address - PAGE_SIZE);
3051 }
3052 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3053 struct vm_area_struct *next = vma->vm_next;
3054
3055 /* As VM_GROWSDOWN but s/below/above/ */
3056 if (next && next->vm_start == address + PAGE_SIZE)
3057 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3058
3059 expand_upwards(vma, address + PAGE_SIZE);
3060 }
3061 return 0;
3062}
3063
3064/*
3065 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3066 * but allow concurrent faults), and pte mapped but not yet locked.
3067 * We return with mmap_sem still held, but pte unmapped and unlocked.
3068 */
3069static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3070 unsigned long address, pte_t *page_table, pmd_t *pmd,
3071 unsigned int flags)
3072{
3073 struct page *page;
3074 spinlock_t *ptl;
3075 pte_t entry;
3076
3077 pte_unmap(page_table);
3078
3079 /* Check if we need to add a guard page to the stack */
3080 if (check_stack_guard_page(vma, address) < 0)
3081 return VM_FAULT_SIGBUS;
3082
3083 /* Use the zero-page for reads */
3084 if (!(flags & FAULT_FLAG_WRITE)) {
3085 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3086 vma->vm_page_prot));
3087 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3088 if (!pte_none(*page_table))
3089 goto unlock;
3090 goto setpte;
3091 }
3092
3093 /* Allocate our own private page. */
3094 if (unlikely(anon_vma_prepare(vma)))
3095 goto oom;
3096 page = alloc_zeroed_user_highpage_movable(vma, address);
3097 if (!page)
3098 goto oom;
3099 __SetPageUptodate(page);
3100
3101 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3102 goto oom_free_page;
3103
3104 entry = mk_pte(page, vma->vm_page_prot);
3105 if (vma->vm_flags & VM_WRITE)
3106 entry = pte_mkwrite(pte_mkdirty(entry));
3107
3108 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3109 if (!pte_none(*page_table))
3110 goto release;
3111
3112 inc_mm_counter_fast(mm, MM_ANONPAGES);
3113 page_add_new_anon_rmap(page, vma, address);
3114setpte:
3115 set_pte_at(mm, address, page_table, entry);
3116
3117 /* No need to invalidate - it was non-present before */
3118 update_mmu_cache(vma, address, page_table);
3119unlock:
3120 pte_unmap_unlock(page_table, ptl);
3121 return 0;
3122release:
3123 mem_cgroup_uncharge_page(page);
3124 page_cache_release(page);
3125 goto unlock;
3126oom_free_page:
3127 page_cache_release(page);
3128oom:
3129 return VM_FAULT_OOM;
3130}
3131
3132/*
3133 * __do_fault() tries to create a new page mapping. It aggressively
3134 * tries to share with existing pages, but makes a separate copy if
3135 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3136 * the next page fault.
3137 *
3138 * As this is called only for pages that do not currently exist, we
3139 * do not need to flush old virtual caches or the TLB.
3140 *
3141 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3142 * but allow concurrent faults), and pte neither mapped nor locked.
3143 * We return with mmap_sem still held, but pte unmapped and unlocked.
3144 */
3145static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3146 unsigned long address, pmd_t *pmd,
3147 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3148{
3149 pte_t *page_table;
3150 spinlock_t *ptl;
3151 struct page *page;
3152 struct page *cow_page;
3153 pte_t entry;
3154 int anon = 0;
3155 struct page *dirty_page = NULL;
3156 struct vm_fault vmf;
3157 int ret;
3158 int page_mkwrite = 0;
3159
3160 /*
3161 * If we do COW later, allocate page befor taking lock_page()
3162 * on the file cache page. This will reduce lock holding time.
3163 */
3164 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3165
3166 if (unlikely(anon_vma_prepare(vma)))
3167 return VM_FAULT_OOM;
3168
3169 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3170 if (!cow_page)
3171 return VM_FAULT_OOM;
3172
3173 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3174 page_cache_release(cow_page);
3175 return VM_FAULT_OOM;
3176 }
3177 } else
3178 cow_page = NULL;
3179
3180 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3181 vmf.pgoff = pgoff;
3182 vmf.flags = flags;
3183 vmf.page = NULL;
3184
3185 ret = vma->vm_ops->fault(vma, &vmf);
3186 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3187 VM_FAULT_RETRY)))
3188 goto uncharge_out;
3189
3190 if (unlikely(PageHWPoison(vmf.page))) {
3191 if (ret & VM_FAULT_LOCKED)
3192 unlock_page(vmf.page);
3193 ret = VM_FAULT_HWPOISON;
3194 goto uncharge_out;
3195 }
3196
3197 /*
3198 * For consistency in subsequent calls, make the faulted page always
3199 * locked.
3200 */
3201 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3202 lock_page(vmf.page);
3203 else
3204 VM_BUG_ON(!PageLocked(vmf.page));
3205
3206 /*
3207 * Should we do an early C-O-W break?
3208 */
3209 page = vmf.page;
3210 if (flags & FAULT_FLAG_WRITE) {
3211 if (!(vma->vm_flags & VM_SHARED)) {
3212 page = cow_page;
3213 anon = 1;
3214 copy_user_highpage(page, vmf.page, address, vma);
3215 __SetPageUptodate(page);
3216 } else {
3217 /*
3218 * If the page will be shareable, see if the backing
3219 * address space wants to know that the page is about
3220 * to become writable
3221 */
3222 if (vma->vm_ops->page_mkwrite) {
3223 int tmp;
3224
3225 unlock_page(page);
3226 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3227 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3228 if (unlikely(tmp &
3229 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3230 ret = tmp;
3231 goto unwritable_page;
3232 }
3233 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3234 lock_page(page);
3235 if (!page->mapping) {
3236 ret = 0; /* retry the fault */
3237 unlock_page(page);
3238 goto unwritable_page;
3239 }
3240 } else
3241 VM_BUG_ON(!PageLocked(page));
3242 page_mkwrite = 1;
3243 }
3244 }
3245
3246 }
3247
3248 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3249
3250 /*
3251 * This silly early PAGE_DIRTY setting removes a race
3252 * due to the bad i386 page protection. But it's valid
3253 * for other architectures too.
3254 *
3255 * Note that if FAULT_FLAG_WRITE is set, we either now have
3256 * an exclusive copy of the page, or this is a shared mapping,
3257 * so we can make it writable and dirty to avoid having to
3258 * handle that later.
3259 */
3260 /* Only go through if we didn't race with anybody else... */
3261 if (likely(pte_same(*page_table, orig_pte))) {
3262 flush_icache_page(vma, page);
3263 entry = mk_pte(page, vma->vm_page_prot);
3264 if (flags & FAULT_FLAG_WRITE)
3265 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3266 if (anon) {
3267 inc_mm_counter_fast(mm, MM_ANONPAGES);
3268 page_add_new_anon_rmap(page, vma, address);
3269 } else {
3270 inc_mm_counter_fast(mm, MM_FILEPAGES);
3271 page_add_file_rmap(page);
3272 if (flags & FAULT_FLAG_WRITE) {
3273 dirty_page = page;
3274 get_page(dirty_page);
3275 }
3276 }
3277 set_pte_at(mm, address, page_table, entry);
3278
3279 /* no need to invalidate: a not-present page won't be cached */
3280 update_mmu_cache(vma, address, page_table);
3281 } else {
3282 if (cow_page)
3283 mem_cgroup_uncharge_page(cow_page);
3284 if (anon)
3285 page_cache_release(page);
3286 else
3287 anon = 1; /* no anon but release faulted_page */
3288 }
3289
3290 pte_unmap_unlock(page_table, ptl);
3291
3292 if (dirty_page) {
3293 struct address_space *mapping = page->mapping;
3294
3295 if (set_page_dirty(dirty_page))
3296 page_mkwrite = 1;
3297 unlock_page(dirty_page);
3298 put_page(dirty_page);
3299 if (page_mkwrite && mapping) {
3300 /*
3301 * Some device drivers do not set page.mapping but still
3302 * dirty their pages
3303 */
3304 balance_dirty_pages_ratelimited(mapping);
3305 }
3306
3307 /* file_update_time outside page_lock */
3308 if (vma->vm_file)
3309 file_update_time(vma->vm_file);
3310 } else {
3311 unlock_page(vmf.page);
3312 if (anon)
3313 page_cache_release(vmf.page);
3314 }
3315
3316 return ret;
3317
3318unwritable_page:
3319 page_cache_release(page);
3320 return ret;
3321uncharge_out:
3322 /* fs's fault handler get error */
3323 if (cow_page) {
3324 mem_cgroup_uncharge_page(cow_page);
3325 page_cache_release(cow_page);
3326 }
3327 return ret;
3328}
3329
3330static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3331 unsigned long address, pte_t *page_table, pmd_t *pmd,
3332 unsigned int flags, pte_t orig_pte)
3333{
3334 pgoff_t pgoff = (((address & PAGE_MASK)
3335 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3336
3337 pte_unmap(page_table);
3338 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3339}
3340
3341/*
3342 * Fault of a previously existing named mapping. Repopulate the pte
3343 * from the encoded file_pte if possible. This enables swappable
3344 * nonlinear vmas.
3345 *
3346 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3347 * but allow concurrent faults), and pte mapped but not yet locked.
3348 * We return with mmap_sem still held, but pte unmapped and unlocked.
3349 */
3350static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3351 unsigned long address, pte_t *page_table, pmd_t *pmd,
3352 unsigned int flags, pte_t orig_pte)
3353{
3354 pgoff_t pgoff;
3355
3356 flags |= FAULT_FLAG_NONLINEAR;
3357
3358 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3359 return 0;
3360
3361 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3362 /*
3363 * Page table corrupted: show pte and kill process.
3364 */
3365 print_bad_pte(vma, address, orig_pte, NULL);
3366 return VM_FAULT_SIGBUS;
3367 }
3368
3369 pgoff = pte_to_pgoff(orig_pte);
3370 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3371}
3372
3373/*
3374 * These routines also need to handle stuff like marking pages dirty
3375 * and/or accessed for architectures that don't do it in hardware (most
3376 * RISC architectures). The early dirtying is also good on the i386.
3377 *
3378 * There is also a hook called "update_mmu_cache()" that architectures
3379 * with external mmu caches can use to update those (ie the Sparc or
3380 * PowerPC hashed page tables that act as extended TLBs).
3381 *
3382 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3383 * but allow concurrent faults), and pte mapped but not yet locked.
3384 * We return with mmap_sem still held, but pte unmapped and unlocked.
3385 */
3386int handle_pte_fault(struct mm_struct *mm,
3387 struct vm_area_struct *vma, unsigned long address,
3388 pte_t *pte, pmd_t *pmd, unsigned int flags)
3389{
3390 pte_t entry;
3391 spinlock_t *ptl;
3392
3393 entry = *pte;
3394 if (!pte_present(entry)) {
3395 if (pte_none(entry)) {
3396 if (vma->vm_ops) {
3397 if (likely(vma->vm_ops->fault))
3398 return do_linear_fault(mm, vma, address,
3399 pte, pmd, flags, entry);
3400 }
3401 return do_anonymous_page(mm, vma, address,
3402 pte, pmd, flags);
3403 }
3404 if (pte_file(entry))
3405 return do_nonlinear_fault(mm, vma, address,
3406 pte, pmd, flags, entry);
3407 return do_swap_page(mm, vma, address,
3408 pte, pmd, flags, entry);
3409 }
3410
3411 ptl = pte_lockptr(mm, pmd);
3412 spin_lock(ptl);
3413 if (unlikely(!pte_same(*pte, entry)))
3414 goto unlock;
3415 if (flags & FAULT_FLAG_WRITE) {
3416 if (!pte_write(entry))
3417 return do_wp_page(mm, vma, address,
3418 pte, pmd, ptl, entry);
3419 entry = pte_mkdirty(entry);
3420 }
3421 entry = pte_mkyoung(entry);
3422 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3423 update_mmu_cache(vma, address, pte);
3424 } else {
3425 /*
3426 * This is needed only for protection faults but the arch code
3427 * is not yet telling us if this is a protection fault or not.
3428 * This still avoids useless tlb flushes for .text page faults
3429 * with threads.
3430 */
3431 if (flags & FAULT_FLAG_WRITE)
3432 flush_tlb_fix_spurious_fault(vma, address);
3433 }
3434unlock:
3435 pte_unmap_unlock(pte, ptl);
3436 return 0;
3437}
3438
3439/*
3440 * By the time we get here, we already hold the mm semaphore
3441 */
3442int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3443 unsigned long address, unsigned int flags)
3444{
3445 pgd_t *pgd;
3446 pud_t *pud;
3447 pmd_t *pmd;
3448 pte_t *pte;
3449
3450 __set_current_state(TASK_RUNNING);
3451
3452 count_vm_event(PGFAULT);
3453 mem_cgroup_count_vm_event(mm, PGFAULT);
3454
3455 /* do counter updates before entering really critical section. */
3456 check_sync_rss_stat(current);
3457
3458 if (unlikely(is_vm_hugetlb_page(vma)))
3459 return hugetlb_fault(mm, vma, address, flags);
3460
3461 pgd = pgd_offset(mm, address);
3462 pud = pud_alloc(mm, pgd, address);
3463 if (!pud)
3464 return VM_FAULT_OOM;
3465 pmd = pmd_alloc(mm, pud, address);
3466 if (!pmd)
3467 return VM_FAULT_OOM;
3468 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3469 if (!vma->vm_ops)
3470 return do_huge_pmd_anonymous_page(mm, vma, address,
3471 pmd, flags);
3472 } else {
3473 pmd_t orig_pmd = *pmd;
3474 barrier();
3475 if (pmd_trans_huge(orig_pmd)) {
3476 if (flags & FAULT_FLAG_WRITE &&
3477 !pmd_write(orig_pmd) &&
3478 !pmd_trans_splitting(orig_pmd))
3479 return do_huge_pmd_wp_page(mm, vma, address,
3480 pmd, orig_pmd);
3481 return 0;
3482 }
3483 }
3484
3485 /*
3486 * Use __pte_alloc instead of pte_alloc_map, because we can't
3487 * run pte_offset_map on the pmd, if an huge pmd could
3488 * materialize from under us from a different thread.
3489 */
3490 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3491 return VM_FAULT_OOM;
3492 /* if an huge pmd materialized from under us just retry later */
3493 if (unlikely(pmd_trans_huge(*pmd)))
3494 return 0;
3495 /*
3496 * A regular pmd is established and it can't morph into a huge pmd
3497 * from under us anymore at this point because we hold the mmap_sem
3498 * read mode and khugepaged takes it in write mode. So now it's
3499 * safe to run pte_offset_map().
3500 */
3501 pte = pte_offset_map(pmd, address);
3502
3503 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3504}
3505
3506#ifndef __PAGETABLE_PUD_FOLDED
3507/*
3508 * Allocate page upper directory.
3509 * We've already handled the fast-path in-line.
3510 */
3511int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3512{
3513 pud_t *new = pud_alloc_one(mm, address);
3514 if (!new)
3515 return -ENOMEM;
3516
3517 smp_wmb(); /* See comment in __pte_alloc */
3518
3519 spin_lock(&mm->page_table_lock);
3520 if (pgd_present(*pgd)) /* Another has populated it */
3521 pud_free(mm, new);
3522 else
3523 pgd_populate(mm, pgd, new);
3524 spin_unlock(&mm->page_table_lock);
3525 return 0;
3526}
3527#endif /* __PAGETABLE_PUD_FOLDED */
3528
3529#ifndef __PAGETABLE_PMD_FOLDED
3530/*
3531 * Allocate page middle directory.
3532 * We've already handled the fast-path in-line.
3533 */
3534int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3535{
3536 pmd_t *new = pmd_alloc_one(mm, address);
3537 if (!new)
3538 return -ENOMEM;
3539
3540 smp_wmb(); /* See comment in __pte_alloc */
3541
3542 spin_lock(&mm->page_table_lock);
3543#ifndef __ARCH_HAS_4LEVEL_HACK
3544 if (pud_present(*pud)) /* Another has populated it */
3545 pmd_free(mm, new);
3546 else
3547 pud_populate(mm, pud, new);
3548#else
3549 if (pgd_present(*pud)) /* Another has populated it */
3550 pmd_free(mm, new);
3551 else
3552 pgd_populate(mm, pud, new);
3553#endif /* __ARCH_HAS_4LEVEL_HACK */
3554 spin_unlock(&mm->page_table_lock);
3555 return 0;
3556}
3557#endif /* __PAGETABLE_PMD_FOLDED */
3558
3559int make_pages_present(unsigned long addr, unsigned long end)
3560{
3561 int ret, len, write;
3562 struct vm_area_struct * vma;
3563
3564 vma = find_vma(current->mm, addr);
3565 if (!vma)
3566 return -ENOMEM;
3567 /*
3568 * We want to touch writable mappings with a write fault in order
3569 * to break COW, except for shared mappings because these don't COW
3570 * and we would not want to dirty them for nothing.
3571 */
3572 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3573 BUG_ON(addr >= end);
3574 BUG_ON(end > vma->vm_end);
3575 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3576 ret = get_user_pages(current, current->mm, addr,
3577 len, write, 0, NULL, NULL);
3578 if (ret < 0)
3579 return ret;
3580 return ret == len ? 0 : -EFAULT;
3581}
3582
3583#if !defined(__HAVE_ARCH_GATE_AREA)
3584
3585#if defined(AT_SYSINFO_EHDR)
3586static struct vm_area_struct gate_vma;
3587
3588static int __init gate_vma_init(void)
3589{
3590 gate_vma.vm_mm = NULL;
3591 gate_vma.vm_start = FIXADDR_USER_START;
3592 gate_vma.vm_end = FIXADDR_USER_END;
3593 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3594 gate_vma.vm_page_prot = __P101;
3595 /*
3596 * Make sure the vDSO gets into every core dump.
3597 * Dumping its contents makes post-mortem fully interpretable later
3598 * without matching up the same kernel and hardware config to see
3599 * what PC values meant.
3600 */
3601 gate_vma.vm_flags |= VM_ALWAYSDUMP;
3602 return 0;
3603}
3604__initcall(gate_vma_init);
3605#endif
3606
3607struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3608{
3609#ifdef AT_SYSINFO_EHDR
3610 return &gate_vma;
3611#else
3612 return NULL;
3613#endif
3614}
3615
3616int in_gate_area_no_mm(unsigned long addr)
3617{
3618#ifdef AT_SYSINFO_EHDR
3619 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3620 return 1;
3621#endif
3622 return 0;
3623}
3624
3625#endif /* __HAVE_ARCH_GATE_AREA */
3626
3627static int __follow_pte(struct mm_struct *mm, unsigned long address,
3628 pte_t **ptepp, spinlock_t **ptlp)
3629{
3630 pgd_t *pgd;
3631 pud_t *pud;
3632 pmd_t *pmd;
3633 pte_t *ptep;
3634
3635 pgd = pgd_offset(mm, address);
3636 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3637 goto out;
3638
3639 pud = pud_offset(pgd, address);
3640 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3641 goto out;
3642
3643 pmd = pmd_offset(pud, address);
3644 VM_BUG_ON(pmd_trans_huge(*pmd));
3645 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3646 goto out;
3647
3648 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3649 if (pmd_huge(*pmd))
3650 goto out;
3651
3652 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3653 if (!ptep)
3654 goto out;
3655 if (!pte_present(*ptep))
3656 goto unlock;
3657 *ptepp = ptep;
3658 return 0;
3659unlock:
3660 pte_unmap_unlock(ptep, *ptlp);
3661out:
3662 return -EINVAL;
3663}
3664
3665static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3666 pte_t **ptepp, spinlock_t **ptlp)
3667{
3668 int res;
3669
3670 /* (void) is needed to make gcc happy */
3671 (void) __cond_lock(*ptlp,
3672 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3673 return res;
3674}
3675
3676/**
3677 * follow_pfn - look up PFN at a user virtual address
3678 * @vma: memory mapping
3679 * @address: user virtual address
3680 * @pfn: location to store found PFN
3681 *
3682 * Only IO mappings and raw PFN mappings are allowed.
3683 *
3684 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3685 */
3686int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3687 unsigned long *pfn)
3688{
3689 int ret = -EINVAL;
3690 spinlock_t *ptl;
3691 pte_t *ptep;
3692
3693 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3694 return ret;
3695
3696 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3697 if (ret)
3698 return ret;
3699 *pfn = pte_pfn(*ptep);
3700 pte_unmap_unlock(ptep, ptl);
3701 return 0;
3702}
3703EXPORT_SYMBOL(follow_pfn);
3704
3705#ifdef CONFIG_HAVE_IOREMAP_PROT
3706int follow_phys(struct vm_area_struct *vma,
3707 unsigned long address, unsigned int flags,
3708 unsigned long *prot, resource_size_t *phys)
3709{
3710 int ret = -EINVAL;
3711 pte_t *ptep, pte;
3712 spinlock_t *ptl;
3713
3714 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3715 goto out;
3716
3717 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3718 goto out;
3719 pte = *ptep;
3720
3721 if ((flags & FOLL_WRITE) && !pte_write(pte))
3722 goto unlock;
3723
3724 *prot = pgprot_val(pte_pgprot(pte));
3725 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3726
3727 ret = 0;
3728unlock:
3729 pte_unmap_unlock(ptep, ptl);
3730out:
3731 return ret;
3732}
3733
3734int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3735 void *buf, int len, int write)
3736{
3737 resource_size_t phys_addr;
3738 unsigned long prot = 0;
3739 void __iomem *maddr;
3740 int offset = addr & (PAGE_SIZE-1);
3741
3742 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3743 return -EINVAL;
3744
3745 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3746 if (write)
3747 memcpy_toio(maddr + offset, buf, len);
3748 else
3749 memcpy_fromio(buf, maddr + offset, len);
3750 iounmap(maddr);
3751
3752 return len;
3753}
3754#endif
3755
3756/*
3757 * Access another process' address space as given in mm. If non-NULL, use the
3758 * given task for page fault accounting.
3759 */
3760static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3761 unsigned long addr, void *buf, int len, int write)
3762{
3763 struct vm_area_struct *vma;
3764 void *old_buf = buf;
3765
3766 down_read(&mm->mmap_sem);
3767 /* ignore errors, just check how much was successfully transferred */
3768 while (len) {
3769 int bytes, ret, offset;
3770 void *maddr;
3771 struct page *page = NULL;
3772
3773 ret = get_user_pages(tsk, mm, addr, 1,
3774 write, 1, &page, &vma);
3775 if (ret <= 0) {
3776 /*
3777 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3778 * we can access using slightly different code.
3779 */
3780#ifdef CONFIG_HAVE_IOREMAP_PROT
3781 vma = find_vma(mm, addr);
3782 if (!vma || vma->vm_start > addr)
3783 break;
3784 if (vma->vm_ops && vma->vm_ops->access)
3785 ret = vma->vm_ops->access(vma, addr, buf,
3786 len, write);
3787 if (ret <= 0)
3788#endif
3789 break;
3790 bytes = ret;
3791 } else {
3792 bytes = len;
3793 offset = addr & (PAGE_SIZE-1);
3794 if (bytes > PAGE_SIZE-offset)
3795 bytes = PAGE_SIZE-offset;
3796
3797 maddr = kmap(page);
3798 if (write) {
3799 copy_to_user_page(vma, page, addr,
3800 maddr + offset, buf, bytes);
3801 set_page_dirty_lock(page);
3802 } else {
3803 copy_from_user_page(vma, page, addr,
3804 buf, maddr + offset, bytes);
3805 }
3806 kunmap(page);
3807 page_cache_release(page);
3808 }
3809 len -= bytes;
3810 buf += bytes;
3811 addr += bytes;
3812 }
3813 up_read(&mm->mmap_sem);
3814
3815 return buf - old_buf;
3816}
3817
3818/**
3819 * access_remote_vm - access another process' address space
3820 * @mm: the mm_struct of the target address space
3821 * @addr: start address to access
3822 * @buf: source or destination buffer
3823 * @len: number of bytes to transfer
3824 * @write: whether the access is a write
3825 *
3826 * The caller must hold a reference on @mm.
3827 */
3828int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3829 void *buf, int len, int write)
3830{
3831 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3832}
3833
3834/*
3835 * Access another process' address space.
3836 * Source/target buffer must be kernel space,
3837 * Do not walk the page table directly, use get_user_pages
3838 */
3839int access_process_vm(struct task_struct *tsk, unsigned long addr,
3840 void *buf, int len, int write)
3841{
3842 struct mm_struct *mm;
3843 int ret;
3844
3845 mm = get_task_mm(tsk);
3846 if (!mm)
3847 return 0;
3848
3849 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3850 mmput(mm);
3851
3852 return ret;
3853}
3854
3855/*
3856 * Print the name of a VMA.
3857 */
3858void print_vma_addr(char *prefix, unsigned long ip)
3859{
3860 struct mm_struct *mm = current->mm;
3861 struct vm_area_struct *vma;
3862
3863 /*
3864 * Do not print if we are in atomic
3865 * contexts (in exception stacks, etc.):
3866 */
3867 if (preempt_count())
3868 return;
3869
3870 down_read(&mm->mmap_sem);
3871 vma = find_vma(mm, ip);
3872 if (vma && vma->vm_file) {
3873 struct file *f = vma->vm_file;
3874 char *buf = (char *)__get_free_page(GFP_KERNEL);
3875 if (buf) {
3876 char *p, *s;
3877
3878 p = d_path(&f->f_path, buf, PAGE_SIZE);
3879 if (IS_ERR(p))
3880 p = "?";
3881 s = strrchr(p, '/');
3882 if (s)
3883 p = s+1;
3884 printk("%s%s[%lx+%lx]", prefix, p,
3885 vma->vm_start,
3886 vma->vm_end - vma->vm_start);
3887 free_page((unsigned long)buf);
3888 }
3889 }
3890 up_read(¤t->mm->mmap_sem);
3891}
3892
3893#ifdef CONFIG_PROVE_LOCKING
3894void might_fault(void)
3895{
3896 /*
3897 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3898 * holding the mmap_sem, this is safe because kernel memory doesn't
3899 * get paged out, therefore we'll never actually fault, and the
3900 * below annotations will generate false positives.
3901 */
3902 if (segment_eq(get_fs(), KERNEL_DS))
3903 return;
3904
3905 might_sleep();
3906 /*
3907 * it would be nicer only to annotate paths which are not under
3908 * pagefault_disable, however that requires a larger audit and
3909 * providing helpers like get_user_atomic.
3910 */
3911 if (!in_atomic() && current->mm)
3912 might_lock_read(¤t->mm->mmap_sem);
3913}
3914EXPORT_SYMBOL(might_fault);
3915#endif
3916
3917#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3918static void clear_gigantic_page(struct page *page,
3919 unsigned long addr,
3920 unsigned int pages_per_huge_page)
3921{
3922 int i;
3923 struct page *p = page;
3924
3925 might_sleep();
3926 for (i = 0; i < pages_per_huge_page;
3927 i++, p = mem_map_next(p, page, i)) {
3928 cond_resched();
3929 clear_user_highpage(p, addr + i * PAGE_SIZE);
3930 }
3931}
3932void clear_huge_page(struct page *page,
3933 unsigned long addr, unsigned int pages_per_huge_page)
3934{
3935 int i;
3936
3937 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3938 clear_gigantic_page(page, addr, pages_per_huge_page);
3939 return;
3940 }
3941
3942 might_sleep();
3943 for (i = 0; i < pages_per_huge_page; i++) {
3944 cond_resched();
3945 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3946 }
3947}
3948
3949static void copy_user_gigantic_page(struct page *dst, struct page *src,
3950 unsigned long addr,
3951 struct vm_area_struct *vma,
3952 unsigned int pages_per_huge_page)
3953{
3954 int i;
3955 struct page *dst_base = dst;
3956 struct page *src_base = src;
3957
3958 for (i = 0; i < pages_per_huge_page; ) {
3959 cond_resched();
3960 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3961
3962 i++;
3963 dst = mem_map_next(dst, dst_base, i);
3964 src = mem_map_next(src, src_base, i);
3965 }
3966}
3967
3968void copy_user_huge_page(struct page *dst, struct page *src,
3969 unsigned long addr, struct vm_area_struct *vma,
3970 unsigned int pages_per_huge_page)
3971{
3972 int i;
3973
3974 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3975 copy_user_gigantic_page(dst, src, addr, vma,
3976 pages_per_huge_page);
3977 return;
3978 }
3979
3980 might_sleep();
3981 for (i = 0; i < pages_per_huge_page; i++) {
3982 cond_resched();
3983 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3984 }
3985}
3986#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
1/*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
29 */
30
31/*
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
41#include <linux/kernel_stat.h>
42#include <linux/mm.h>
43#include <linux/hugetlb.h>
44#include <linux/mman.h>
45#include <linux/swap.h>
46#include <linux/highmem.h>
47#include <linux/pagemap.h>
48#include <linux/ksm.h>
49#include <linux/rmap.h>
50#include <linux/export.h>
51#include <linux/delayacct.h>
52#include <linux/init.h>
53#include <linux/pfn_t.h>
54#include <linux/writeback.h>
55#include <linux/memcontrol.h>
56#include <linux/mmu_notifier.h>
57#include <linux/kallsyms.h>
58#include <linux/swapops.h>
59#include <linux/elf.h>
60#include <linux/gfp.h>
61#include <linux/migrate.h>
62#include <linux/string.h>
63#include <linux/dma-debug.h>
64#include <linux/debugfs.h>
65#include <linux/userfaultfd_k.h>
66#include <linux/dax.h>
67
68#include <asm/io.h>
69#include <asm/mmu_context.h>
70#include <asm/pgalloc.h>
71#include <linux/uaccess.h>
72#include <asm/tlb.h>
73#include <asm/tlbflush.h>
74#include <asm/pgtable.h>
75
76#include "internal.h"
77
78#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
79#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
80#endif
81
82#ifndef CONFIG_NEED_MULTIPLE_NODES
83/* use the per-pgdat data instead for discontigmem - mbligh */
84unsigned long max_mapnr;
85struct page *mem_map;
86
87EXPORT_SYMBOL(max_mapnr);
88EXPORT_SYMBOL(mem_map);
89#endif
90
91/*
92 * A number of key systems in x86 including ioremap() rely on the assumption
93 * that high_memory defines the upper bound on direct map memory, then end
94 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
95 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
96 * and ZONE_HIGHMEM.
97 */
98void * high_memory;
99
100EXPORT_SYMBOL(high_memory);
101
102/*
103 * Randomize the address space (stacks, mmaps, brk, etc.).
104 *
105 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
106 * as ancient (libc5 based) binaries can segfault. )
107 */
108int randomize_va_space __read_mostly =
109#ifdef CONFIG_COMPAT_BRK
110 1;
111#else
112 2;
113#endif
114
115static int __init disable_randmaps(char *s)
116{
117 randomize_va_space = 0;
118 return 1;
119}
120__setup("norandmaps", disable_randmaps);
121
122unsigned long zero_pfn __read_mostly;
123unsigned long highest_memmap_pfn __read_mostly;
124
125EXPORT_SYMBOL(zero_pfn);
126
127/*
128 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
129 */
130static int __init init_zero_pfn(void)
131{
132 zero_pfn = page_to_pfn(ZERO_PAGE(0));
133 return 0;
134}
135core_initcall(init_zero_pfn);
136
137
138#if defined(SPLIT_RSS_COUNTING)
139
140void sync_mm_rss(struct mm_struct *mm)
141{
142 int i;
143
144 for (i = 0; i < NR_MM_COUNTERS; i++) {
145 if (current->rss_stat.count[i]) {
146 add_mm_counter(mm, i, current->rss_stat.count[i]);
147 current->rss_stat.count[i] = 0;
148 }
149 }
150 current->rss_stat.events = 0;
151}
152
153static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
154{
155 struct task_struct *task = current;
156
157 if (likely(task->mm == mm))
158 task->rss_stat.count[member] += val;
159 else
160 add_mm_counter(mm, member, val);
161}
162#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
163#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
164
165/* sync counter once per 64 page faults */
166#define TASK_RSS_EVENTS_THRESH (64)
167static void check_sync_rss_stat(struct task_struct *task)
168{
169 if (unlikely(task != current))
170 return;
171 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
172 sync_mm_rss(task->mm);
173}
174#else /* SPLIT_RSS_COUNTING */
175
176#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
177#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
178
179static void check_sync_rss_stat(struct task_struct *task)
180{
181}
182
183#endif /* SPLIT_RSS_COUNTING */
184
185#ifdef HAVE_GENERIC_MMU_GATHER
186
187static bool tlb_next_batch(struct mmu_gather *tlb)
188{
189 struct mmu_gather_batch *batch;
190
191 batch = tlb->active;
192 if (batch->next) {
193 tlb->active = batch->next;
194 return true;
195 }
196
197 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
198 return false;
199
200 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
201 if (!batch)
202 return false;
203
204 tlb->batch_count++;
205 batch->next = NULL;
206 batch->nr = 0;
207 batch->max = MAX_GATHER_BATCH;
208
209 tlb->active->next = batch;
210 tlb->active = batch;
211
212 return true;
213}
214
215/* tlb_gather_mmu
216 * Called to initialize an (on-stack) mmu_gather structure for page-table
217 * tear-down from @mm. The @fullmm argument is used when @mm is without
218 * users and we're going to destroy the full address space (exit/execve).
219 */
220void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
221{
222 tlb->mm = mm;
223
224 /* Is it from 0 to ~0? */
225 tlb->fullmm = !(start | (end+1));
226 tlb->need_flush_all = 0;
227 tlb->local.next = NULL;
228 tlb->local.nr = 0;
229 tlb->local.max = ARRAY_SIZE(tlb->__pages);
230 tlb->active = &tlb->local;
231 tlb->batch_count = 0;
232
233#ifdef CONFIG_HAVE_RCU_TABLE_FREE
234 tlb->batch = NULL;
235#endif
236 tlb->page_size = 0;
237
238 __tlb_reset_range(tlb);
239}
240
241static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
242{
243 if (!tlb->end)
244 return;
245
246 tlb_flush(tlb);
247 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
248#ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb);
250#endif
251 __tlb_reset_range(tlb);
252}
253
254static void tlb_flush_mmu_free(struct mmu_gather *tlb)
255{
256 struct mmu_gather_batch *batch;
257
258 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
259 free_pages_and_swap_cache(batch->pages, batch->nr);
260 batch->nr = 0;
261 }
262 tlb->active = &tlb->local;
263}
264
265void tlb_flush_mmu(struct mmu_gather *tlb)
266{
267 tlb_flush_mmu_tlbonly(tlb);
268 tlb_flush_mmu_free(tlb);
269}
270
271/* tlb_finish_mmu
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
274 */
275void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
276{
277 struct mmu_gather_batch *batch, *next;
278
279 tlb_flush_mmu(tlb);
280
281 /* keep the page table cache within bounds */
282 check_pgt_cache();
283
284 for (batch = tlb->local.next; batch; batch = next) {
285 next = batch->next;
286 free_pages((unsigned long)batch, 0);
287 }
288 tlb->local.next = NULL;
289}
290
291/* __tlb_remove_page
292 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
293 * handling the additional races in SMP caused by other CPUs caching valid
294 * mappings in their TLBs. Returns the number of free page slots left.
295 * When out of page slots we must call tlb_flush_mmu().
296 *returns true if the caller should flush.
297 */
298bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
299{
300 struct mmu_gather_batch *batch;
301
302 VM_BUG_ON(!tlb->end);
303 VM_WARN_ON(tlb->page_size != page_size);
304
305 batch = tlb->active;
306 /*
307 * Add the page and check if we are full. If so
308 * force a flush.
309 */
310 batch->pages[batch->nr++] = page;
311 if (batch->nr == batch->max) {
312 if (!tlb_next_batch(tlb))
313 return true;
314 batch = tlb->active;
315 }
316 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
317
318 return false;
319}
320
321#endif /* HAVE_GENERIC_MMU_GATHER */
322
323#ifdef CONFIG_HAVE_RCU_TABLE_FREE
324
325/*
326 * See the comment near struct mmu_table_batch.
327 */
328
329static void tlb_remove_table_smp_sync(void *arg)
330{
331 /* Simply deliver the interrupt */
332}
333
334static void tlb_remove_table_one(void *table)
335{
336 /*
337 * This isn't an RCU grace period and hence the page-tables cannot be
338 * assumed to be actually RCU-freed.
339 *
340 * It is however sufficient for software page-table walkers that rely on
341 * IRQ disabling. See the comment near struct mmu_table_batch.
342 */
343 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
344 __tlb_remove_table(table);
345}
346
347static void tlb_remove_table_rcu(struct rcu_head *head)
348{
349 struct mmu_table_batch *batch;
350 int i;
351
352 batch = container_of(head, struct mmu_table_batch, rcu);
353
354 for (i = 0; i < batch->nr; i++)
355 __tlb_remove_table(batch->tables[i]);
356
357 free_page((unsigned long)batch);
358}
359
360void tlb_table_flush(struct mmu_gather *tlb)
361{
362 struct mmu_table_batch **batch = &tlb->batch;
363
364 if (*batch) {
365 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
366 *batch = NULL;
367 }
368}
369
370void tlb_remove_table(struct mmu_gather *tlb, void *table)
371{
372 struct mmu_table_batch **batch = &tlb->batch;
373
374 /*
375 * When there's less then two users of this mm there cannot be a
376 * concurrent page-table walk.
377 */
378 if (atomic_read(&tlb->mm->mm_users) < 2) {
379 __tlb_remove_table(table);
380 return;
381 }
382
383 if (*batch == NULL) {
384 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
385 if (*batch == NULL) {
386 tlb_remove_table_one(table);
387 return;
388 }
389 (*batch)->nr = 0;
390 }
391 (*batch)->tables[(*batch)->nr++] = table;
392 if ((*batch)->nr == MAX_TABLE_BATCH)
393 tlb_table_flush(tlb);
394}
395
396#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
397
398/*
399 * Note: this doesn't free the actual pages themselves. That
400 * has been handled earlier when unmapping all the memory regions.
401 */
402static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
403 unsigned long addr)
404{
405 pgtable_t token = pmd_pgtable(*pmd);
406 pmd_clear(pmd);
407 pte_free_tlb(tlb, token, addr);
408 atomic_long_dec(&tlb->mm->nr_ptes);
409}
410
411static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
412 unsigned long addr, unsigned long end,
413 unsigned long floor, unsigned long ceiling)
414{
415 pmd_t *pmd;
416 unsigned long next;
417 unsigned long start;
418
419 start = addr;
420 pmd = pmd_offset(pud, addr);
421 do {
422 next = pmd_addr_end(addr, end);
423 if (pmd_none_or_clear_bad(pmd))
424 continue;
425 free_pte_range(tlb, pmd, addr);
426 } while (pmd++, addr = next, addr != end);
427
428 start &= PUD_MASK;
429 if (start < floor)
430 return;
431 if (ceiling) {
432 ceiling &= PUD_MASK;
433 if (!ceiling)
434 return;
435 }
436 if (end - 1 > ceiling - 1)
437 return;
438
439 pmd = pmd_offset(pud, start);
440 pud_clear(pud);
441 pmd_free_tlb(tlb, pmd, start);
442 mm_dec_nr_pmds(tlb->mm);
443}
444
445static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
446 unsigned long addr, unsigned long end,
447 unsigned long floor, unsigned long ceiling)
448{
449 pud_t *pud;
450 unsigned long next;
451 unsigned long start;
452
453 start = addr;
454 pud = pud_offset(pgd, addr);
455 do {
456 next = pud_addr_end(addr, end);
457 if (pud_none_or_clear_bad(pud))
458 continue;
459 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
460 } while (pud++, addr = next, addr != end);
461
462 start &= PGDIR_MASK;
463 if (start < floor)
464 return;
465 if (ceiling) {
466 ceiling &= PGDIR_MASK;
467 if (!ceiling)
468 return;
469 }
470 if (end - 1 > ceiling - 1)
471 return;
472
473 pud = pud_offset(pgd, start);
474 pgd_clear(pgd);
475 pud_free_tlb(tlb, pud, start);
476}
477
478/*
479 * This function frees user-level page tables of a process.
480 */
481void free_pgd_range(struct mmu_gather *tlb,
482 unsigned long addr, unsigned long end,
483 unsigned long floor, unsigned long ceiling)
484{
485 pgd_t *pgd;
486 unsigned long next;
487
488 /*
489 * The next few lines have given us lots of grief...
490 *
491 * Why are we testing PMD* at this top level? Because often
492 * there will be no work to do at all, and we'd prefer not to
493 * go all the way down to the bottom just to discover that.
494 *
495 * Why all these "- 1"s? Because 0 represents both the bottom
496 * of the address space and the top of it (using -1 for the
497 * top wouldn't help much: the masks would do the wrong thing).
498 * The rule is that addr 0 and floor 0 refer to the bottom of
499 * the address space, but end 0 and ceiling 0 refer to the top
500 * Comparisons need to use "end - 1" and "ceiling - 1" (though
501 * that end 0 case should be mythical).
502 *
503 * Wherever addr is brought up or ceiling brought down, we must
504 * be careful to reject "the opposite 0" before it confuses the
505 * subsequent tests. But what about where end is brought down
506 * by PMD_SIZE below? no, end can't go down to 0 there.
507 *
508 * Whereas we round start (addr) and ceiling down, by different
509 * masks at different levels, in order to test whether a table
510 * now has no other vmas using it, so can be freed, we don't
511 * bother to round floor or end up - the tests don't need that.
512 */
513
514 addr &= PMD_MASK;
515 if (addr < floor) {
516 addr += PMD_SIZE;
517 if (!addr)
518 return;
519 }
520 if (ceiling) {
521 ceiling &= PMD_MASK;
522 if (!ceiling)
523 return;
524 }
525 if (end - 1 > ceiling - 1)
526 end -= PMD_SIZE;
527 if (addr > end - 1)
528 return;
529 /*
530 * We add page table cache pages with PAGE_SIZE,
531 * (see pte_free_tlb()), flush the tlb if we need
532 */
533 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
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
543void 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
578int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
579{
580 spinlock_t *ptl;
581 pgtable_t new = pte_alloc_one(mm, address);
582 if (!new)
583 return -ENOMEM;
584
585 /*
586 * Ensure all pte setup (eg. pte page lock and page clearing) are
587 * visible before the pte is made visible to other CPUs by being
588 * put into page tables.
589 *
590 * The other side of the story is the pointer chasing in the page
591 * table walking code (when walking the page table without locking;
592 * ie. most of the time). Fortunately, these data accesses consist
593 * of a chain of data-dependent loads, meaning most CPUs (alpha
594 * being the notable exception) will already guarantee loads are
595 * seen in-order. See the alpha page table accessors for the
596 * smp_read_barrier_depends() barriers in page table walking code.
597 */
598 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
599
600 ptl = pmd_lock(mm, pmd);
601 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
602 atomic_long_inc(&mm->nr_ptes);
603 pmd_populate(mm, pmd, new);
604 new = NULL;
605 }
606 spin_unlock(ptl);
607 if (new)
608 pte_free(mm, new);
609 return 0;
610}
611
612int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
613{
614 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
615 if (!new)
616 return -ENOMEM;
617
618 smp_wmb(); /* See comment in __pte_alloc */
619
620 spin_lock(&init_mm.page_table_lock);
621 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
622 pmd_populate_kernel(&init_mm, pmd, new);
623 new = NULL;
624 }
625 spin_unlock(&init_mm.page_table_lock);
626 if (new)
627 pte_free_kernel(&init_mm, new);
628 return 0;
629}
630
631static inline void init_rss_vec(int *rss)
632{
633 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
634}
635
636static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
637{
638 int i;
639
640 if (current->mm == mm)
641 sync_mm_rss(mm);
642 for (i = 0; i < NR_MM_COUNTERS; i++)
643 if (rss[i])
644 add_mm_counter(mm, i, rss[i]);
645}
646
647/*
648 * This function is called to print an error when a bad pte
649 * is found. For example, we might have a PFN-mapped pte in
650 * a region that doesn't allow it.
651 *
652 * The calling function must still handle the error.
653 */
654static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
655 pte_t pte, struct page *page)
656{
657 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
658 pud_t *pud = pud_offset(pgd, addr);
659 pmd_t *pmd = pmd_offset(pud, addr);
660 struct address_space *mapping;
661 pgoff_t index;
662 static unsigned long resume;
663 static unsigned long nr_shown;
664 static unsigned long nr_unshown;
665
666 /*
667 * Allow a burst of 60 reports, then keep quiet for that minute;
668 * or allow a steady drip of one report per second.
669 */
670 if (nr_shown == 60) {
671 if (time_before(jiffies, resume)) {
672 nr_unshown++;
673 return;
674 }
675 if (nr_unshown) {
676 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
677 nr_unshown);
678 nr_unshown = 0;
679 }
680 nr_shown = 0;
681 }
682 if (nr_shown++ == 0)
683 resume = jiffies + 60 * HZ;
684
685 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
686 index = linear_page_index(vma, addr);
687
688 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
689 current->comm,
690 (long long)pte_val(pte), (long long)pmd_val(*pmd));
691 if (page)
692 dump_page(page, "bad pte");
693 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
694 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
695 /*
696 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
697 */
698 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
699 vma->vm_file,
700 vma->vm_ops ? vma->vm_ops->fault : NULL,
701 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
702 mapping ? mapping->a_ops->readpage : NULL);
703 dump_stack();
704 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
705}
706
707/*
708 * vm_normal_page -- This function gets the "struct page" associated with a pte.
709 *
710 * "Special" mappings do not wish to be associated with a "struct page" (either
711 * it doesn't exist, or it exists but they don't want to touch it). In this
712 * case, NULL is returned here. "Normal" mappings do have a struct page.
713 *
714 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
715 * pte bit, in which case this function is trivial. Secondly, an architecture
716 * may not have a spare pte bit, which requires a more complicated scheme,
717 * described below.
718 *
719 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
720 * special mapping (even if there are underlying and valid "struct pages").
721 * COWed pages of a VM_PFNMAP are always normal.
722 *
723 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
724 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
725 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
726 * mapping will always honor the rule
727 *
728 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
729 *
730 * And for normal mappings this is false.
731 *
732 * This restricts such mappings to be a linear translation from virtual address
733 * to pfn. To get around this restriction, we allow arbitrary mappings so long
734 * as the vma is not a COW mapping; in that case, we know that all ptes are
735 * special (because none can have been COWed).
736 *
737 *
738 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
739 *
740 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
741 * page" backing, however the difference is that _all_ pages with a struct
742 * page (that is, those where pfn_valid is true) are refcounted and considered
743 * normal pages by the VM. The disadvantage is that pages are refcounted
744 * (which can be slower and simply not an option for some PFNMAP users). The
745 * advantage is that we don't have to follow the strict linearity rule of
746 * PFNMAP mappings in order to support COWable mappings.
747 *
748 */
749#ifdef __HAVE_ARCH_PTE_SPECIAL
750# define HAVE_PTE_SPECIAL 1
751#else
752# define HAVE_PTE_SPECIAL 0
753#endif
754struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
755 pte_t pte)
756{
757 unsigned long pfn = pte_pfn(pte);
758
759 if (HAVE_PTE_SPECIAL) {
760 if (likely(!pte_special(pte)))
761 goto check_pfn;
762 if (vma->vm_ops && vma->vm_ops->find_special_page)
763 return vma->vm_ops->find_special_page(vma, addr);
764 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
765 return NULL;
766 if (!is_zero_pfn(pfn))
767 print_bad_pte(vma, addr, pte, NULL);
768 return NULL;
769 }
770
771 /* !HAVE_PTE_SPECIAL case follows: */
772
773 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
774 if (vma->vm_flags & VM_MIXEDMAP) {
775 if (!pfn_valid(pfn))
776 return NULL;
777 goto out;
778 } else {
779 unsigned long off;
780 off = (addr - vma->vm_start) >> PAGE_SHIFT;
781 if (pfn == vma->vm_pgoff + off)
782 return NULL;
783 if (!is_cow_mapping(vma->vm_flags))
784 return NULL;
785 }
786 }
787
788 if (is_zero_pfn(pfn))
789 return NULL;
790check_pfn:
791 if (unlikely(pfn > highest_memmap_pfn)) {
792 print_bad_pte(vma, addr, pte, NULL);
793 return NULL;
794 }
795
796 /*
797 * NOTE! We still have PageReserved() pages in the page tables.
798 * eg. VDSO mappings can cause them to exist.
799 */
800out:
801 return pfn_to_page(pfn);
802}
803
804#ifdef CONFIG_TRANSPARENT_HUGEPAGE
805struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
806 pmd_t pmd)
807{
808 unsigned long pfn = pmd_pfn(pmd);
809
810 /*
811 * There is no pmd_special() but there may be special pmds, e.g.
812 * in a direct-access (dax) mapping, so let's just replicate the
813 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
814 */
815 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
816 if (vma->vm_flags & VM_MIXEDMAP) {
817 if (!pfn_valid(pfn))
818 return NULL;
819 goto out;
820 } else {
821 unsigned long off;
822 off = (addr - vma->vm_start) >> PAGE_SHIFT;
823 if (pfn == vma->vm_pgoff + off)
824 return NULL;
825 if (!is_cow_mapping(vma->vm_flags))
826 return NULL;
827 }
828 }
829
830 if (is_zero_pfn(pfn))
831 return NULL;
832 if (unlikely(pfn > highest_memmap_pfn))
833 return NULL;
834
835 /*
836 * NOTE! We still have PageReserved() pages in the page tables.
837 * eg. VDSO mappings can cause them to exist.
838 */
839out:
840 return pfn_to_page(pfn);
841}
842#endif
843
844/*
845 * copy one vm_area from one task to the other. Assumes the page tables
846 * already present in the new task to be cleared in the whole range
847 * covered by this vma.
848 */
849
850static inline unsigned long
851copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
852 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
853 unsigned long addr, int *rss)
854{
855 unsigned long vm_flags = vma->vm_flags;
856 pte_t pte = *src_pte;
857 struct page *page;
858
859 /* pte contains position in swap or file, so copy. */
860 if (unlikely(!pte_present(pte))) {
861 swp_entry_t entry = pte_to_swp_entry(pte);
862
863 if (likely(!non_swap_entry(entry))) {
864 if (swap_duplicate(entry) < 0)
865 return entry.val;
866
867 /* make sure dst_mm is on swapoff's mmlist. */
868 if (unlikely(list_empty(&dst_mm->mmlist))) {
869 spin_lock(&mmlist_lock);
870 if (list_empty(&dst_mm->mmlist))
871 list_add(&dst_mm->mmlist,
872 &src_mm->mmlist);
873 spin_unlock(&mmlist_lock);
874 }
875 rss[MM_SWAPENTS]++;
876 } else if (is_migration_entry(entry)) {
877 page = migration_entry_to_page(entry);
878
879 rss[mm_counter(page)]++;
880
881 if (is_write_migration_entry(entry) &&
882 is_cow_mapping(vm_flags)) {
883 /*
884 * COW mappings require pages in both
885 * parent and child to be set to read.
886 */
887 make_migration_entry_read(&entry);
888 pte = swp_entry_to_pte(entry);
889 if (pte_swp_soft_dirty(*src_pte))
890 pte = pte_swp_mksoft_dirty(pte);
891 set_pte_at(src_mm, addr, src_pte, pte);
892 }
893 }
894 goto out_set_pte;
895 }
896
897 /*
898 * If it's a COW mapping, write protect it both
899 * in the parent and the child
900 */
901 if (is_cow_mapping(vm_flags)) {
902 ptep_set_wrprotect(src_mm, addr, src_pte);
903 pte = pte_wrprotect(pte);
904 }
905
906 /*
907 * If it's a shared mapping, mark it clean in
908 * the child
909 */
910 if (vm_flags & VM_SHARED)
911 pte = pte_mkclean(pte);
912 pte = pte_mkold(pte);
913
914 page = vm_normal_page(vma, addr, pte);
915 if (page) {
916 get_page(page);
917 page_dup_rmap(page, false);
918 rss[mm_counter(page)]++;
919 }
920
921out_set_pte:
922 set_pte_at(dst_mm, addr, dst_pte, pte);
923 return 0;
924}
925
926static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
927 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
928 unsigned long addr, unsigned long end)
929{
930 pte_t *orig_src_pte, *orig_dst_pte;
931 pte_t *src_pte, *dst_pte;
932 spinlock_t *src_ptl, *dst_ptl;
933 int progress = 0;
934 int rss[NR_MM_COUNTERS];
935 swp_entry_t entry = (swp_entry_t){0};
936
937again:
938 init_rss_vec(rss);
939
940 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
941 if (!dst_pte)
942 return -ENOMEM;
943 src_pte = pte_offset_map(src_pmd, addr);
944 src_ptl = pte_lockptr(src_mm, src_pmd);
945 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
946 orig_src_pte = src_pte;
947 orig_dst_pte = dst_pte;
948 arch_enter_lazy_mmu_mode();
949
950 do {
951 /*
952 * We are holding two locks at this point - either of them
953 * could generate latencies in another task on another CPU.
954 */
955 if (progress >= 32) {
956 progress = 0;
957 if (need_resched() ||
958 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
959 break;
960 }
961 if (pte_none(*src_pte)) {
962 progress++;
963 continue;
964 }
965 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
966 vma, addr, rss);
967 if (entry.val)
968 break;
969 progress += 8;
970 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
971
972 arch_leave_lazy_mmu_mode();
973 spin_unlock(src_ptl);
974 pte_unmap(orig_src_pte);
975 add_mm_rss_vec(dst_mm, rss);
976 pte_unmap_unlock(orig_dst_pte, dst_ptl);
977 cond_resched();
978
979 if (entry.val) {
980 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
981 return -ENOMEM;
982 progress = 0;
983 }
984 if (addr != end)
985 goto again;
986 return 0;
987}
988
989static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
990 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
991 unsigned long addr, unsigned long end)
992{
993 pmd_t *src_pmd, *dst_pmd;
994 unsigned long next;
995
996 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
997 if (!dst_pmd)
998 return -ENOMEM;
999 src_pmd = pmd_offset(src_pud, addr);
1000 do {
1001 next = pmd_addr_end(addr, end);
1002 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
1003 int err;
1004 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1005 err = copy_huge_pmd(dst_mm, src_mm,
1006 dst_pmd, src_pmd, addr, vma);
1007 if (err == -ENOMEM)
1008 return -ENOMEM;
1009 if (!err)
1010 continue;
1011 /* fall through */
1012 }
1013 if (pmd_none_or_clear_bad(src_pmd))
1014 continue;
1015 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1016 vma, addr, next))
1017 return -ENOMEM;
1018 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1019 return 0;
1020}
1021
1022static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1024 unsigned long addr, unsigned long end)
1025{
1026 pud_t *src_pud, *dst_pud;
1027 unsigned long next;
1028
1029 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1030 if (!dst_pud)
1031 return -ENOMEM;
1032 src_pud = pud_offset(src_pgd, addr);
1033 do {
1034 next = pud_addr_end(addr, end);
1035 if (pud_none_or_clear_bad(src_pud))
1036 continue;
1037 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1038 vma, addr, next))
1039 return -ENOMEM;
1040 } while (dst_pud++, src_pud++, addr = next, addr != end);
1041 return 0;
1042}
1043
1044int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045 struct vm_area_struct *vma)
1046{
1047 pgd_t *src_pgd, *dst_pgd;
1048 unsigned long next;
1049 unsigned long addr = vma->vm_start;
1050 unsigned long end = vma->vm_end;
1051 unsigned long mmun_start; /* For mmu_notifiers */
1052 unsigned long mmun_end; /* For mmu_notifiers */
1053 bool is_cow;
1054 int ret;
1055
1056 /*
1057 * Don't copy ptes where a page fault will fill them correctly.
1058 * Fork becomes much lighter when there are big shared or private
1059 * readonly mappings. The tradeoff is that copy_page_range is more
1060 * efficient than faulting.
1061 */
1062 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1063 !vma->anon_vma)
1064 return 0;
1065
1066 if (is_vm_hugetlb_page(vma))
1067 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1068
1069 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1070 /*
1071 * We do not free on error cases below as remove_vma
1072 * gets called on error from higher level routine
1073 */
1074 ret = track_pfn_copy(vma);
1075 if (ret)
1076 return ret;
1077 }
1078
1079 /*
1080 * We need to invalidate the secondary MMU mappings only when
1081 * there could be a permission downgrade on the ptes of the
1082 * parent mm. And a permission downgrade will only happen if
1083 * is_cow_mapping() returns true.
1084 */
1085 is_cow = is_cow_mapping(vma->vm_flags);
1086 mmun_start = addr;
1087 mmun_end = end;
1088 if (is_cow)
1089 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1090 mmun_end);
1091
1092 ret = 0;
1093 dst_pgd = pgd_offset(dst_mm, addr);
1094 src_pgd = pgd_offset(src_mm, addr);
1095 do {
1096 next = pgd_addr_end(addr, end);
1097 if (pgd_none_or_clear_bad(src_pgd))
1098 continue;
1099 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1100 vma, addr, next))) {
1101 ret = -ENOMEM;
1102 break;
1103 }
1104 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1105
1106 if (is_cow)
1107 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1108 return ret;
1109}
1110
1111static unsigned long zap_pte_range(struct mmu_gather *tlb,
1112 struct vm_area_struct *vma, pmd_t *pmd,
1113 unsigned long addr, unsigned long end,
1114 struct zap_details *details)
1115{
1116 struct mm_struct *mm = tlb->mm;
1117 int force_flush = 0;
1118 int rss[NR_MM_COUNTERS];
1119 spinlock_t *ptl;
1120 pte_t *start_pte;
1121 pte_t *pte;
1122 swp_entry_t entry;
1123
1124 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1125again:
1126 init_rss_vec(rss);
1127 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1128 pte = start_pte;
1129 arch_enter_lazy_mmu_mode();
1130 do {
1131 pte_t ptent = *pte;
1132 if (pte_none(ptent)) {
1133 continue;
1134 }
1135
1136 if (pte_present(ptent)) {
1137 struct page *page;
1138
1139 page = vm_normal_page(vma, addr, ptent);
1140 if (unlikely(details) && page) {
1141 /*
1142 * unmap_shared_mapping_pages() wants to
1143 * invalidate cache without truncating:
1144 * unmap shared but keep private pages.
1145 */
1146 if (details->check_mapping &&
1147 details->check_mapping != page_rmapping(page))
1148 continue;
1149 }
1150 ptent = ptep_get_and_clear_full(mm, addr, pte,
1151 tlb->fullmm);
1152 tlb_remove_tlb_entry(tlb, pte, addr);
1153 if (unlikely(!page))
1154 continue;
1155
1156 if (!PageAnon(page)) {
1157 if (pte_dirty(ptent)) {
1158 /*
1159 * oom_reaper cannot tear down dirty
1160 * pages
1161 */
1162 if (unlikely(details && details->ignore_dirty))
1163 continue;
1164 force_flush = 1;
1165 set_page_dirty(page);
1166 }
1167 if (pte_young(ptent) &&
1168 likely(!(vma->vm_flags & VM_SEQ_READ)))
1169 mark_page_accessed(page);
1170 }
1171 rss[mm_counter(page)]--;
1172 page_remove_rmap(page, false);
1173 if (unlikely(page_mapcount(page) < 0))
1174 print_bad_pte(vma, addr, ptent, page);
1175 if (unlikely(__tlb_remove_page(tlb, page))) {
1176 force_flush = 1;
1177 addr += PAGE_SIZE;
1178 break;
1179 }
1180 continue;
1181 }
1182 /* only check swap_entries if explicitly asked for in details */
1183 if (unlikely(details && !details->check_swap_entries))
1184 continue;
1185
1186 entry = pte_to_swp_entry(ptent);
1187 if (!non_swap_entry(entry))
1188 rss[MM_SWAPENTS]--;
1189 else if (is_migration_entry(entry)) {
1190 struct page *page;
1191
1192 page = migration_entry_to_page(entry);
1193 rss[mm_counter(page)]--;
1194 }
1195 if (unlikely(!free_swap_and_cache(entry)))
1196 print_bad_pte(vma, addr, ptent, NULL);
1197 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198 } while (pte++, addr += PAGE_SIZE, addr != end);
1199
1200 add_mm_rss_vec(mm, rss);
1201 arch_leave_lazy_mmu_mode();
1202
1203 /* Do the actual TLB flush before dropping ptl */
1204 if (force_flush)
1205 tlb_flush_mmu_tlbonly(tlb);
1206 pte_unmap_unlock(start_pte, ptl);
1207
1208 /*
1209 * If we forced a TLB flush (either due to running out of
1210 * batch buffers or because we needed to flush dirty TLB
1211 * entries before releasing the ptl), free the batched
1212 * memory too. Restart if we didn't do everything.
1213 */
1214 if (force_flush) {
1215 force_flush = 0;
1216 tlb_flush_mmu_free(tlb);
1217 if (addr != end)
1218 goto again;
1219 }
1220
1221 return addr;
1222}
1223
1224static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pud_t *pud,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1228{
1229 pmd_t *pmd;
1230 unsigned long next;
1231
1232 pmd = pmd_offset(pud, addr);
1233 do {
1234 next = pmd_addr_end(addr, end);
1235 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1238 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1239 __split_huge_pmd(vma, pmd, addr, false, NULL);
1240 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1241 goto next;
1242 /* fall through */
1243 }
1244 /*
1245 * Here there can be other concurrent MADV_DONTNEED or
1246 * trans huge page faults running, and if the pmd is
1247 * none or trans huge it can change under us. This is
1248 * because MADV_DONTNEED holds the mmap_sem in read
1249 * mode.
1250 */
1251 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1252 goto next;
1253 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1254next:
1255 cond_resched();
1256 } while (pmd++, addr = next, addr != end);
1257
1258 return addr;
1259}
1260
1261static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1262 struct vm_area_struct *vma, pgd_t *pgd,
1263 unsigned long addr, unsigned long end,
1264 struct zap_details *details)
1265{
1266 pud_t *pud;
1267 unsigned long next;
1268
1269 pud = pud_offset(pgd, addr);
1270 do {
1271 next = pud_addr_end(addr, end);
1272 if (pud_none_or_clear_bad(pud))
1273 continue;
1274 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1275 } while (pud++, addr = next, addr != end);
1276
1277 return addr;
1278}
1279
1280void unmap_page_range(struct mmu_gather *tlb,
1281 struct vm_area_struct *vma,
1282 unsigned long addr, unsigned long end,
1283 struct zap_details *details)
1284{
1285 pgd_t *pgd;
1286 unsigned long next;
1287
1288 BUG_ON(addr >= end);
1289 tlb_start_vma(tlb, vma);
1290 pgd = pgd_offset(vma->vm_mm, addr);
1291 do {
1292 next = pgd_addr_end(addr, end);
1293 if (pgd_none_or_clear_bad(pgd))
1294 continue;
1295 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1296 } while (pgd++, addr = next, addr != end);
1297 tlb_end_vma(tlb, vma);
1298}
1299
1300
1301static void unmap_single_vma(struct mmu_gather *tlb,
1302 struct vm_area_struct *vma, unsigned long start_addr,
1303 unsigned long end_addr,
1304 struct zap_details *details)
1305{
1306 unsigned long start = max(vma->vm_start, start_addr);
1307 unsigned long end;
1308
1309 if (start >= vma->vm_end)
1310 return;
1311 end = min(vma->vm_end, end_addr);
1312 if (end <= vma->vm_start)
1313 return;
1314
1315 if (vma->vm_file)
1316 uprobe_munmap(vma, start, end);
1317
1318 if (unlikely(vma->vm_flags & VM_PFNMAP))
1319 untrack_pfn(vma, 0, 0);
1320
1321 if (start != end) {
1322 if (unlikely(is_vm_hugetlb_page(vma))) {
1323 /*
1324 * It is undesirable to test vma->vm_file as it
1325 * should be non-null for valid hugetlb area.
1326 * However, vm_file will be NULL in the error
1327 * cleanup path of mmap_region. When
1328 * hugetlbfs ->mmap method fails,
1329 * mmap_region() nullifies vma->vm_file
1330 * before calling this function to clean up.
1331 * Since no pte has actually been setup, it is
1332 * safe to do nothing in this case.
1333 */
1334 if (vma->vm_file) {
1335 i_mmap_lock_write(vma->vm_file->f_mapping);
1336 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1337 i_mmap_unlock_write(vma->vm_file->f_mapping);
1338 }
1339 } else
1340 unmap_page_range(tlb, vma, start, end, details);
1341 }
1342}
1343
1344/**
1345 * unmap_vmas - unmap a range of memory covered by a list of vma's
1346 * @tlb: address of the caller's struct mmu_gather
1347 * @vma: the starting vma
1348 * @start_addr: virtual address at which to start unmapping
1349 * @end_addr: virtual address at which to end unmapping
1350 *
1351 * Unmap all pages in the vma list.
1352 *
1353 * Only addresses between `start' and `end' will be unmapped.
1354 *
1355 * The VMA list must be sorted in ascending virtual address order.
1356 *
1357 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1358 * range after unmap_vmas() returns. So the only responsibility here is to
1359 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1360 * drops the lock and schedules.
1361 */
1362void unmap_vmas(struct mmu_gather *tlb,
1363 struct vm_area_struct *vma, unsigned long start_addr,
1364 unsigned long end_addr)
1365{
1366 struct mm_struct *mm = vma->vm_mm;
1367
1368 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1369 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1370 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1371 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1372}
1373
1374/**
1375 * zap_page_range - remove user pages in a given range
1376 * @vma: vm_area_struct holding the applicable pages
1377 * @start: starting address of pages to zap
1378 * @size: number of bytes to zap
1379 * @details: details of shared cache invalidation
1380 *
1381 * Caller must protect the VMA list
1382 */
1383void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1384 unsigned long size, struct zap_details *details)
1385{
1386 struct mm_struct *mm = vma->vm_mm;
1387 struct mmu_gather tlb;
1388 unsigned long end = start + size;
1389
1390 lru_add_drain();
1391 tlb_gather_mmu(&tlb, mm, start, end);
1392 update_hiwater_rss(mm);
1393 mmu_notifier_invalidate_range_start(mm, start, end);
1394 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1395 unmap_single_vma(&tlb, vma, start, end, details);
1396 mmu_notifier_invalidate_range_end(mm, start, end);
1397 tlb_finish_mmu(&tlb, start, end);
1398}
1399
1400/**
1401 * zap_page_range_single - remove user pages in a given range
1402 * @vma: vm_area_struct holding the applicable pages
1403 * @address: starting address of pages to zap
1404 * @size: number of bytes to zap
1405 * @details: details of shared cache invalidation
1406 *
1407 * The range must fit into one VMA.
1408 */
1409static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1410 unsigned long size, struct zap_details *details)
1411{
1412 struct mm_struct *mm = vma->vm_mm;
1413 struct mmu_gather tlb;
1414 unsigned long end = address + size;
1415
1416 lru_add_drain();
1417 tlb_gather_mmu(&tlb, mm, address, end);
1418 update_hiwater_rss(mm);
1419 mmu_notifier_invalidate_range_start(mm, address, end);
1420 unmap_single_vma(&tlb, vma, address, end, details);
1421 mmu_notifier_invalidate_range_end(mm, address, end);
1422 tlb_finish_mmu(&tlb, address, end);
1423}
1424
1425/**
1426 * zap_vma_ptes - remove ptes mapping the vma
1427 * @vma: vm_area_struct holding ptes to be zapped
1428 * @address: starting address of pages to zap
1429 * @size: number of bytes to zap
1430 *
1431 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1432 *
1433 * The entire address range must be fully contained within the vma.
1434 *
1435 * Returns 0 if successful.
1436 */
1437int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1438 unsigned long size)
1439{
1440 if (address < vma->vm_start || address + size > vma->vm_end ||
1441 !(vma->vm_flags & VM_PFNMAP))
1442 return -1;
1443 zap_page_range_single(vma, address, size, NULL);
1444 return 0;
1445}
1446EXPORT_SYMBOL_GPL(zap_vma_ptes);
1447
1448pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1449 spinlock_t **ptl)
1450{
1451 pgd_t * pgd = pgd_offset(mm, addr);
1452 pud_t * pud = pud_alloc(mm, pgd, addr);
1453 if (pud) {
1454 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1455 if (pmd) {
1456 VM_BUG_ON(pmd_trans_huge(*pmd));
1457 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1458 }
1459 }
1460 return NULL;
1461}
1462
1463/*
1464 * This is the old fallback for page remapping.
1465 *
1466 * For historical reasons, it only allows reserved pages. Only
1467 * old drivers should use this, and they needed to mark their
1468 * pages reserved for the old functions anyway.
1469 */
1470static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1471 struct page *page, pgprot_t prot)
1472{
1473 struct mm_struct *mm = vma->vm_mm;
1474 int retval;
1475 pte_t *pte;
1476 spinlock_t *ptl;
1477
1478 retval = -EINVAL;
1479 if (PageAnon(page))
1480 goto out;
1481 retval = -ENOMEM;
1482 flush_dcache_page(page);
1483 pte = get_locked_pte(mm, addr, &ptl);
1484 if (!pte)
1485 goto out;
1486 retval = -EBUSY;
1487 if (!pte_none(*pte))
1488 goto out_unlock;
1489
1490 /* Ok, finally just insert the thing.. */
1491 get_page(page);
1492 inc_mm_counter_fast(mm, mm_counter_file(page));
1493 page_add_file_rmap(page, false);
1494 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1495
1496 retval = 0;
1497 pte_unmap_unlock(pte, ptl);
1498 return retval;
1499out_unlock:
1500 pte_unmap_unlock(pte, ptl);
1501out:
1502 return retval;
1503}
1504
1505/**
1506 * vm_insert_page - insert single page into user vma
1507 * @vma: user vma to map to
1508 * @addr: target user address of this page
1509 * @page: source kernel page
1510 *
1511 * This allows drivers to insert individual pages they've allocated
1512 * into a user vma.
1513 *
1514 * The page has to be a nice clean _individual_ kernel allocation.
1515 * If you allocate a compound page, you need to have marked it as
1516 * such (__GFP_COMP), or manually just split the page up yourself
1517 * (see split_page()).
1518 *
1519 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1520 * took an arbitrary page protection parameter. This doesn't allow
1521 * that. Your vma protection will have to be set up correctly, which
1522 * means that if you want a shared writable mapping, you'd better
1523 * ask for a shared writable mapping!
1524 *
1525 * The page does not need to be reserved.
1526 *
1527 * Usually this function is called from f_op->mmap() handler
1528 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1529 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1530 * function from other places, for example from page-fault handler.
1531 */
1532int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1533 struct page *page)
1534{
1535 if (addr < vma->vm_start || addr >= vma->vm_end)
1536 return -EFAULT;
1537 if (!page_count(page))
1538 return -EINVAL;
1539 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1540 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1541 BUG_ON(vma->vm_flags & VM_PFNMAP);
1542 vma->vm_flags |= VM_MIXEDMAP;
1543 }
1544 return insert_page(vma, addr, page, vma->vm_page_prot);
1545}
1546EXPORT_SYMBOL(vm_insert_page);
1547
1548static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1549 pfn_t pfn, pgprot_t prot)
1550{
1551 struct mm_struct *mm = vma->vm_mm;
1552 int retval;
1553 pte_t *pte, entry;
1554 spinlock_t *ptl;
1555
1556 retval = -ENOMEM;
1557 pte = get_locked_pte(mm, addr, &ptl);
1558 if (!pte)
1559 goto out;
1560 retval = -EBUSY;
1561 if (!pte_none(*pte))
1562 goto out_unlock;
1563
1564 /* Ok, finally just insert the thing.. */
1565 if (pfn_t_devmap(pfn))
1566 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1567 else
1568 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1569 set_pte_at(mm, addr, pte, entry);
1570 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1571
1572 retval = 0;
1573out_unlock:
1574 pte_unmap_unlock(pte, ptl);
1575out:
1576 return retval;
1577}
1578
1579/**
1580 * vm_insert_pfn - insert single pfn into user vma
1581 * @vma: user vma to map to
1582 * @addr: target user address of this page
1583 * @pfn: source kernel pfn
1584 *
1585 * Similar to vm_insert_page, this allows drivers to insert individual pages
1586 * they've allocated into a user vma. Same comments apply.
1587 *
1588 * This function should only be called from a vm_ops->fault handler, and
1589 * in that case the handler should return NULL.
1590 *
1591 * vma cannot be a COW mapping.
1592 *
1593 * As this is called only for pages that do not currently exist, we
1594 * do not need to flush old virtual caches or the TLB.
1595 */
1596int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1597 unsigned long pfn)
1598{
1599 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1600}
1601EXPORT_SYMBOL(vm_insert_pfn);
1602
1603/**
1604 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1605 * @vma: user vma to map to
1606 * @addr: target user address of this page
1607 * @pfn: source kernel pfn
1608 * @pgprot: pgprot flags for the inserted page
1609 *
1610 * This is exactly like vm_insert_pfn, except that it allows drivers to
1611 * to override pgprot on a per-page basis.
1612 *
1613 * This only makes sense for IO mappings, and it makes no sense for
1614 * cow mappings. In general, using multiple vmas is preferable;
1615 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1616 * impractical.
1617 */
1618int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1619 unsigned long pfn, pgprot_t pgprot)
1620{
1621 int ret;
1622 /*
1623 * Technically, architectures with pte_special can avoid all these
1624 * restrictions (same for remap_pfn_range). However we would like
1625 * consistency in testing and feature parity among all, so we should
1626 * try to keep these invariants in place for everybody.
1627 */
1628 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1629 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1630 (VM_PFNMAP|VM_MIXEDMAP));
1631 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1632 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1633
1634 if (addr < vma->vm_start || addr >= vma->vm_end)
1635 return -EFAULT;
1636
1637 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1638
1639 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1640
1641 return ret;
1642}
1643EXPORT_SYMBOL(vm_insert_pfn_prot);
1644
1645int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1646 pfn_t pfn)
1647{
1648 pgprot_t pgprot = vma->vm_page_prot;
1649
1650 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1651
1652 if (addr < vma->vm_start || addr >= vma->vm_end)
1653 return -EFAULT;
1654
1655 track_pfn_insert(vma, &pgprot, pfn);
1656
1657 /*
1658 * If we don't have pte special, then we have to use the pfn_valid()
1659 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1660 * refcount the page if pfn_valid is true (hence insert_page rather
1661 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1662 * without pte special, it would there be refcounted as a normal page.
1663 */
1664 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1665 struct page *page;
1666
1667 /*
1668 * At this point we are committed to insert_page()
1669 * regardless of whether the caller specified flags that
1670 * result in pfn_t_has_page() == false.
1671 */
1672 page = pfn_to_page(pfn_t_to_pfn(pfn));
1673 return insert_page(vma, addr, page, pgprot);
1674 }
1675 return insert_pfn(vma, addr, pfn, pgprot);
1676}
1677EXPORT_SYMBOL(vm_insert_mixed);
1678
1679/*
1680 * maps a range of physical memory into the requested pages. the old
1681 * mappings are removed. any references to nonexistent pages results
1682 * in null mappings (currently treated as "copy-on-access")
1683 */
1684static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1685 unsigned long addr, unsigned long end,
1686 unsigned long pfn, pgprot_t prot)
1687{
1688 pte_t *pte;
1689 spinlock_t *ptl;
1690
1691 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1692 if (!pte)
1693 return -ENOMEM;
1694 arch_enter_lazy_mmu_mode();
1695 do {
1696 BUG_ON(!pte_none(*pte));
1697 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1698 pfn++;
1699 } while (pte++, addr += PAGE_SIZE, addr != end);
1700 arch_leave_lazy_mmu_mode();
1701 pte_unmap_unlock(pte - 1, ptl);
1702 return 0;
1703}
1704
1705static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1706 unsigned long addr, unsigned long end,
1707 unsigned long pfn, pgprot_t prot)
1708{
1709 pmd_t *pmd;
1710 unsigned long next;
1711
1712 pfn -= addr >> PAGE_SHIFT;
1713 pmd = pmd_alloc(mm, pud, addr);
1714 if (!pmd)
1715 return -ENOMEM;
1716 VM_BUG_ON(pmd_trans_huge(*pmd));
1717 do {
1718 next = pmd_addr_end(addr, end);
1719 if (remap_pte_range(mm, pmd, addr, next,
1720 pfn + (addr >> PAGE_SHIFT), prot))
1721 return -ENOMEM;
1722 } while (pmd++, addr = next, addr != end);
1723 return 0;
1724}
1725
1726static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1727 unsigned long addr, unsigned long end,
1728 unsigned long pfn, pgprot_t prot)
1729{
1730 pud_t *pud;
1731 unsigned long next;
1732
1733 pfn -= addr >> PAGE_SHIFT;
1734 pud = pud_alloc(mm, pgd, addr);
1735 if (!pud)
1736 return -ENOMEM;
1737 do {
1738 next = pud_addr_end(addr, end);
1739 if (remap_pmd_range(mm, pud, addr, next,
1740 pfn + (addr >> PAGE_SHIFT), prot))
1741 return -ENOMEM;
1742 } while (pud++, addr = next, addr != end);
1743 return 0;
1744}
1745
1746/**
1747 * remap_pfn_range - remap kernel memory to userspace
1748 * @vma: user vma to map to
1749 * @addr: target user address to start at
1750 * @pfn: physical address of kernel memory
1751 * @size: size of map area
1752 * @prot: page protection flags for this mapping
1753 *
1754 * Note: this is only safe if the mm semaphore is held when called.
1755 */
1756int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1757 unsigned long pfn, unsigned long size, pgprot_t prot)
1758{
1759 pgd_t *pgd;
1760 unsigned long next;
1761 unsigned long end = addr + PAGE_ALIGN(size);
1762 struct mm_struct *mm = vma->vm_mm;
1763 unsigned long remap_pfn = pfn;
1764 int err;
1765
1766 /*
1767 * Physically remapped pages are special. Tell the
1768 * rest of the world about it:
1769 * VM_IO tells people not to look at these pages
1770 * (accesses can have side effects).
1771 * VM_PFNMAP tells the core MM that the base pages are just
1772 * raw PFN mappings, and do not have a "struct page" associated
1773 * with them.
1774 * VM_DONTEXPAND
1775 * Disable vma merging and expanding with mremap().
1776 * VM_DONTDUMP
1777 * Omit vma from core dump, even when VM_IO turned off.
1778 *
1779 * There's a horrible special case to handle copy-on-write
1780 * behaviour that some programs depend on. We mark the "original"
1781 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1782 * See vm_normal_page() for details.
1783 */
1784 if (is_cow_mapping(vma->vm_flags)) {
1785 if (addr != vma->vm_start || end != vma->vm_end)
1786 return -EINVAL;
1787 vma->vm_pgoff = pfn;
1788 }
1789
1790 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1791 if (err)
1792 return -EINVAL;
1793
1794 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1795
1796 BUG_ON(addr >= end);
1797 pfn -= addr >> PAGE_SHIFT;
1798 pgd = pgd_offset(mm, addr);
1799 flush_cache_range(vma, addr, end);
1800 do {
1801 next = pgd_addr_end(addr, end);
1802 err = remap_pud_range(mm, pgd, addr, next,
1803 pfn + (addr >> PAGE_SHIFT), prot);
1804 if (err)
1805 break;
1806 } while (pgd++, addr = next, addr != end);
1807
1808 if (err)
1809 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1810
1811 return err;
1812}
1813EXPORT_SYMBOL(remap_pfn_range);
1814
1815/**
1816 * vm_iomap_memory - remap memory to userspace
1817 * @vma: user vma to map to
1818 * @start: start of area
1819 * @len: size of area
1820 *
1821 * This is a simplified io_remap_pfn_range() for common driver use. The
1822 * driver just needs to give us the physical memory range to be mapped,
1823 * we'll figure out the rest from the vma information.
1824 *
1825 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1826 * whatever write-combining details or similar.
1827 */
1828int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1829{
1830 unsigned long vm_len, pfn, pages;
1831
1832 /* Check that the physical memory area passed in looks valid */
1833 if (start + len < start)
1834 return -EINVAL;
1835 /*
1836 * You *really* shouldn't map things that aren't page-aligned,
1837 * but we've historically allowed it because IO memory might
1838 * just have smaller alignment.
1839 */
1840 len += start & ~PAGE_MASK;
1841 pfn = start >> PAGE_SHIFT;
1842 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1843 if (pfn + pages < pfn)
1844 return -EINVAL;
1845
1846 /* We start the mapping 'vm_pgoff' pages into the area */
1847 if (vma->vm_pgoff > pages)
1848 return -EINVAL;
1849 pfn += vma->vm_pgoff;
1850 pages -= vma->vm_pgoff;
1851
1852 /* Can we fit all of the mapping? */
1853 vm_len = vma->vm_end - vma->vm_start;
1854 if (vm_len >> PAGE_SHIFT > pages)
1855 return -EINVAL;
1856
1857 /* Ok, let it rip */
1858 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1859}
1860EXPORT_SYMBOL(vm_iomap_memory);
1861
1862static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1863 unsigned long addr, unsigned long end,
1864 pte_fn_t fn, void *data)
1865{
1866 pte_t *pte;
1867 int err;
1868 pgtable_t token;
1869 spinlock_t *uninitialized_var(ptl);
1870
1871 pte = (mm == &init_mm) ?
1872 pte_alloc_kernel(pmd, addr) :
1873 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1874 if (!pte)
1875 return -ENOMEM;
1876
1877 BUG_ON(pmd_huge(*pmd));
1878
1879 arch_enter_lazy_mmu_mode();
1880
1881 token = pmd_pgtable(*pmd);
1882
1883 do {
1884 err = fn(pte++, token, addr, data);
1885 if (err)
1886 break;
1887 } while (addr += PAGE_SIZE, addr != end);
1888
1889 arch_leave_lazy_mmu_mode();
1890
1891 if (mm != &init_mm)
1892 pte_unmap_unlock(pte-1, ptl);
1893 return err;
1894}
1895
1896static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1897 unsigned long addr, unsigned long end,
1898 pte_fn_t fn, void *data)
1899{
1900 pmd_t *pmd;
1901 unsigned long next;
1902 int err;
1903
1904 BUG_ON(pud_huge(*pud));
1905
1906 pmd = pmd_alloc(mm, pud, addr);
1907 if (!pmd)
1908 return -ENOMEM;
1909 do {
1910 next = pmd_addr_end(addr, end);
1911 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1912 if (err)
1913 break;
1914 } while (pmd++, addr = next, addr != end);
1915 return err;
1916}
1917
1918static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1919 unsigned long addr, unsigned long end,
1920 pte_fn_t fn, void *data)
1921{
1922 pud_t *pud;
1923 unsigned long next;
1924 int err;
1925
1926 pud = pud_alloc(mm, pgd, addr);
1927 if (!pud)
1928 return -ENOMEM;
1929 do {
1930 next = pud_addr_end(addr, end);
1931 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1932 if (err)
1933 break;
1934 } while (pud++, addr = next, addr != end);
1935 return err;
1936}
1937
1938/*
1939 * Scan a region of virtual memory, filling in page tables as necessary
1940 * and calling a provided function on each leaf page table.
1941 */
1942int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1943 unsigned long size, pte_fn_t fn, void *data)
1944{
1945 pgd_t *pgd;
1946 unsigned long next;
1947 unsigned long end = addr + size;
1948 int err;
1949
1950 if (WARN_ON(addr >= end))
1951 return -EINVAL;
1952
1953 pgd = pgd_offset(mm, addr);
1954 do {
1955 next = pgd_addr_end(addr, end);
1956 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1957 if (err)
1958 break;
1959 } while (pgd++, addr = next, addr != end);
1960
1961 return err;
1962}
1963EXPORT_SYMBOL_GPL(apply_to_page_range);
1964
1965/*
1966 * handle_pte_fault chooses page fault handler according to an entry which was
1967 * read non-atomically. Before making any commitment, on those architectures
1968 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1969 * parts, do_swap_page must check under lock before unmapping the pte and
1970 * proceeding (but do_wp_page is only called after already making such a check;
1971 * and do_anonymous_page can safely check later on).
1972 */
1973static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1974 pte_t *page_table, pte_t orig_pte)
1975{
1976 int same = 1;
1977#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1978 if (sizeof(pte_t) > sizeof(unsigned long)) {
1979 spinlock_t *ptl = pte_lockptr(mm, pmd);
1980 spin_lock(ptl);
1981 same = pte_same(*page_table, orig_pte);
1982 spin_unlock(ptl);
1983 }
1984#endif
1985 pte_unmap(page_table);
1986 return same;
1987}
1988
1989static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1990{
1991 debug_dma_assert_idle(src);
1992
1993 /*
1994 * If the source page was a PFN mapping, we don't have
1995 * a "struct page" for it. We do a best-effort copy by
1996 * just copying from the original user address. If that
1997 * fails, we just zero-fill it. Live with it.
1998 */
1999 if (unlikely(!src)) {
2000 void *kaddr = kmap_atomic(dst);
2001 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2002
2003 /*
2004 * This really shouldn't fail, because the page is there
2005 * in the page tables. But it might just be unreadable,
2006 * in which case we just give up and fill the result with
2007 * zeroes.
2008 */
2009 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2010 clear_page(kaddr);
2011 kunmap_atomic(kaddr);
2012 flush_dcache_page(dst);
2013 } else
2014 copy_user_highpage(dst, src, va, vma);
2015}
2016
2017static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2018{
2019 struct file *vm_file = vma->vm_file;
2020
2021 if (vm_file)
2022 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2023
2024 /*
2025 * Special mappings (e.g. VDSO) do not have any file so fake
2026 * a default GFP_KERNEL for them.
2027 */
2028 return GFP_KERNEL;
2029}
2030
2031/*
2032 * Notify the address space that the page is about to become writable so that
2033 * it can prohibit this or wait for the page to get into an appropriate state.
2034 *
2035 * We do this without the lock held, so that it can sleep if it needs to.
2036 */
2037static int do_page_mkwrite(struct vm_fault *vmf)
2038{
2039 int ret;
2040 struct page *page = vmf->page;
2041 unsigned int old_flags = vmf->flags;
2042
2043 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2044
2045 ret = vmf->vma->vm_ops->page_mkwrite(vmf->vma, vmf);
2046 /* Restore original flags so that caller is not surprised */
2047 vmf->flags = old_flags;
2048 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2049 return ret;
2050 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2051 lock_page(page);
2052 if (!page->mapping) {
2053 unlock_page(page);
2054 return 0; /* retry */
2055 }
2056 ret |= VM_FAULT_LOCKED;
2057 } else
2058 VM_BUG_ON_PAGE(!PageLocked(page), page);
2059 return ret;
2060}
2061
2062/*
2063 * Handle dirtying of a page in shared file mapping on a write fault.
2064 *
2065 * The function expects the page to be locked and unlocks it.
2066 */
2067static void fault_dirty_shared_page(struct vm_area_struct *vma,
2068 struct page *page)
2069{
2070 struct address_space *mapping;
2071 bool dirtied;
2072 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2073
2074 dirtied = set_page_dirty(page);
2075 VM_BUG_ON_PAGE(PageAnon(page), page);
2076 /*
2077 * Take a local copy of the address_space - page.mapping may be zeroed
2078 * by truncate after unlock_page(). The address_space itself remains
2079 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2080 * release semantics to prevent the compiler from undoing this copying.
2081 */
2082 mapping = page_rmapping(page);
2083 unlock_page(page);
2084
2085 if ((dirtied || page_mkwrite) && mapping) {
2086 /*
2087 * Some device drivers do not set page.mapping
2088 * but still dirty their pages
2089 */
2090 balance_dirty_pages_ratelimited(mapping);
2091 }
2092
2093 if (!page_mkwrite)
2094 file_update_time(vma->vm_file);
2095}
2096
2097/*
2098 * Handle write page faults for pages that can be reused in the current vma
2099 *
2100 * This can happen either due to the mapping being with the VM_SHARED flag,
2101 * or due to us being the last reference standing to the page. In either
2102 * case, all we need to do here is to mark the page as writable and update
2103 * any related book-keeping.
2104 */
2105static inline void wp_page_reuse(struct vm_fault *vmf)
2106 __releases(vmf->ptl)
2107{
2108 struct vm_area_struct *vma = vmf->vma;
2109 struct page *page = vmf->page;
2110 pte_t entry;
2111 /*
2112 * Clear the pages cpupid information as the existing
2113 * information potentially belongs to a now completely
2114 * unrelated process.
2115 */
2116 if (page)
2117 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2118
2119 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2120 entry = pte_mkyoung(vmf->orig_pte);
2121 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2122 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2123 update_mmu_cache(vma, vmf->address, vmf->pte);
2124 pte_unmap_unlock(vmf->pte, vmf->ptl);
2125}
2126
2127/*
2128 * Handle the case of a page which we actually need to copy to a new page.
2129 *
2130 * Called with mmap_sem locked and the old page referenced, but
2131 * without the ptl held.
2132 *
2133 * High level logic flow:
2134 *
2135 * - Allocate a page, copy the content of the old page to the new one.
2136 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2137 * - Take the PTL. If the pte changed, bail out and release the allocated page
2138 * - If the pte is still the way we remember it, update the page table and all
2139 * relevant references. This includes dropping the reference the page-table
2140 * held to the old page, as well as updating the rmap.
2141 * - In any case, unlock the PTL and drop the reference we took to the old page.
2142 */
2143static int wp_page_copy(struct vm_fault *vmf)
2144{
2145 struct vm_area_struct *vma = vmf->vma;
2146 struct mm_struct *mm = vma->vm_mm;
2147 struct page *old_page = vmf->page;
2148 struct page *new_page = NULL;
2149 pte_t entry;
2150 int page_copied = 0;
2151 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2152 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2153 struct mem_cgroup *memcg;
2154
2155 if (unlikely(anon_vma_prepare(vma)))
2156 goto oom;
2157
2158 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2159 new_page = alloc_zeroed_user_highpage_movable(vma,
2160 vmf->address);
2161 if (!new_page)
2162 goto oom;
2163 } else {
2164 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2165 vmf->address);
2166 if (!new_page)
2167 goto oom;
2168 cow_user_page(new_page, old_page, vmf->address, vma);
2169 }
2170
2171 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2172 goto oom_free_new;
2173
2174 __SetPageUptodate(new_page);
2175
2176 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2177
2178 /*
2179 * Re-check the pte - we dropped the lock
2180 */
2181 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2182 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2183 if (old_page) {
2184 if (!PageAnon(old_page)) {
2185 dec_mm_counter_fast(mm,
2186 mm_counter_file(old_page));
2187 inc_mm_counter_fast(mm, MM_ANONPAGES);
2188 }
2189 } else {
2190 inc_mm_counter_fast(mm, MM_ANONPAGES);
2191 }
2192 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2193 entry = mk_pte(new_page, vma->vm_page_prot);
2194 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2195 /*
2196 * Clear the pte entry and flush it first, before updating the
2197 * pte with the new entry. This will avoid a race condition
2198 * seen in the presence of one thread doing SMC and another
2199 * thread doing COW.
2200 */
2201 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2202 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2203 mem_cgroup_commit_charge(new_page, memcg, false, false);
2204 lru_cache_add_active_or_unevictable(new_page, vma);
2205 /*
2206 * We call the notify macro here because, when using secondary
2207 * mmu page tables (such as kvm shadow page tables), we want the
2208 * new page to be mapped directly into the secondary page table.
2209 */
2210 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2211 update_mmu_cache(vma, vmf->address, vmf->pte);
2212 if (old_page) {
2213 /*
2214 * Only after switching the pte to the new page may
2215 * we remove the mapcount here. Otherwise another
2216 * process may come and find the rmap count decremented
2217 * before the pte is switched to the new page, and
2218 * "reuse" the old page writing into it while our pte
2219 * here still points into it and can be read by other
2220 * threads.
2221 *
2222 * The critical issue is to order this
2223 * page_remove_rmap with the ptp_clear_flush above.
2224 * Those stores are ordered by (if nothing else,)
2225 * the barrier present in the atomic_add_negative
2226 * in page_remove_rmap.
2227 *
2228 * Then the TLB flush in ptep_clear_flush ensures that
2229 * no process can access the old page before the
2230 * decremented mapcount is visible. And the old page
2231 * cannot be reused until after the decremented
2232 * mapcount is visible. So transitively, TLBs to
2233 * old page will be flushed before it can be reused.
2234 */
2235 page_remove_rmap(old_page, false);
2236 }
2237
2238 /* Free the old page.. */
2239 new_page = old_page;
2240 page_copied = 1;
2241 } else {
2242 mem_cgroup_cancel_charge(new_page, memcg, false);
2243 }
2244
2245 if (new_page)
2246 put_page(new_page);
2247
2248 pte_unmap_unlock(vmf->pte, vmf->ptl);
2249 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2250 if (old_page) {
2251 /*
2252 * Don't let another task, with possibly unlocked vma,
2253 * keep the mlocked page.
2254 */
2255 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2256 lock_page(old_page); /* LRU manipulation */
2257 if (PageMlocked(old_page))
2258 munlock_vma_page(old_page);
2259 unlock_page(old_page);
2260 }
2261 put_page(old_page);
2262 }
2263 return page_copied ? VM_FAULT_WRITE : 0;
2264oom_free_new:
2265 put_page(new_page);
2266oom:
2267 if (old_page)
2268 put_page(old_page);
2269 return VM_FAULT_OOM;
2270}
2271
2272/**
2273 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2274 * writeable once the page is prepared
2275 *
2276 * @vmf: structure describing the fault
2277 *
2278 * This function handles all that is needed to finish a write page fault in a
2279 * shared mapping due to PTE being read-only once the mapped page is prepared.
2280 * It handles locking of PTE and modifying it. The function returns
2281 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2282 * lock.
2283 *
2284 * The function expects the page to be locked or other protection against
2285 * concurrent faults / writeback (such as DAX radix tree locks).
2286 */
2287int finish_mkwrite_fault(struct vm_fault *vmf)
2288{
2289 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2290 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2291 &vmf->ptl);
2292 /*
2293 * We might have raced with another page fault while we released the
2294 * pte_offset_map_lock.
2295 */
2296 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2297 pte_unmap_unlock(vmf->pte, vmf->ptl);
2298 return VM_FAULT_NOPAGE;
2299 }
2300 wp_page_reuse(vmf);
2301 return 0;
2302}
2303
2304/*
2305 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2306 * mapping
2307 */
2308static int wp_pfn_shared(struct vm_fault *vmf)
2309{
2310 struct vm_area_struct *vma = vmf->vma;
2311
2312 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2313 int ret;
2314
2315 pte_unmap_unlock(vmf->pte, vmf->ptl);
2316 vmf->flags |= FAULT_FLAG_MKWRITE;
2317 ret = vma->vm_ops->pfn_mkwrite(vma, vmf);
2318 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2319 return ret;
2320 return finish_mkwrite_fault(vmf);
2321 }
2322 wp_page_reuse(vmf);
2323 return VM_FAULT_WRITE;
2324}
2325
2326static int wp_page_shared(struct vm_fault *vmf)
2327 __releases(vmf->ptl)
2328{
2329 struct vm_area_struct *vma = vmf->vma;
2330
2331 get_page(vmf->page);
2332
2333 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2334 int tmp;
2335
2336 pte_unmap_unlock(vmf->pte, vmf->ptl);
2337 tmp = do_page_mkwrite(vmf);
2338 if (unlikely(!tmp || (tmp &
2339 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2340 put_page(vmf->page);
2341 return tmp;
2342 }
2343 tmp = finish_mkwrite_fault(vmf);
2344 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2345 unlock_page(vmf->page);
2346 put_page(vmf->page);
2347 return tmp;
2348 }
2349 } else {
2350 wp_page_reuse(vmf);
2351 lock_page(vmf->page);
2352 }
2353 fault_dirty_shared_page(vma, vmf->page);
2354 put_page(vmf->page);
2355
2356 return VM_FAULT_WRITE;
2357}
2358
2359/*
2360 * This routine handles present pages, when users try to write
2361 * to a shared page. It is done by copying the page to a new address
2362 * and decrementing the shared-page counter for the old page.
2363 *
2364 * Note that this routine assumes that the protection checks have been
2365 * done by the caller (the low-level page fault routine in most cases).
2366 * Thus we can safely just mark it writable once we've done any necessary
2367 * COW.
2368 *
2369 * We also mark the page dirty at this point even though the page will
2370 * change only once the write actually happens. This avoids a few races,
2371 * and potentially makes it more efficient.
2372 *
2373 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2374 * but allow concurrent faults), with pte both mapped and locked.
2375 * We return with mmap_sem still held, but pte unmapped and unlocked.
2376 */
2377static int do_wp_page(struct vm_fault *vmf)
2378 __releases(vmf->ptl)
2379{
2380 struct vm_area_struct *vma = vmf->vma;
2381
2382 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2383 if (!vmf->page) {
2384 /*
2385 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2386 * VM_PFNMAP VMA.
2387 *
2388 * We should not cow pages in a shared writeable mapping.
2389 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2390 */
2391 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2392 (VM_WRITE|VM_SHARED))
2393 return wp_pfn_shared(vmf);
2394
2395 pte_unmap_unlock(vmf->pte, vmf->ptl);
2396 return wp_page_copy(vmf);
2397 }
2398
2399 /*
2400 * Take out anonymous pages first, anonymous shared vmas are
2401 * not dirty accountable.
2402 */
2403 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2404 int total_mapcount;
2405 if (!trylock_page(vmf->page)) {
2406 get_page(vmf->page);
2407 pte_unmap_unlock(vmf->pte, vmf->ptl);
2408 lock_page(vmf->page);
2409 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2410 vmf->address, &vmf->ptl);
2411 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2412 unlock_page(vmf->page);
2413 pte_unmap_unlock(vmf->pte, vmf->ptl);
2414 put_page(vmf->page);
2415 return 0;
2416 }
2417 put_page(vmf->page);
2418 }
2419 if (reuse_swap_page(vmf->page, &total_mapcount)) {
2420 if (total_mapcount == 1) {
2421 /*
2422 * The page is all ours. Move it to
2423 * our anon_vma so the rmap code will
2424 * not search our parent or siblings.
2425 * Protected against the rmap code by
2426 * the page lock.
2427 */
2428 page_move_anon_rmap(vmf->page, vma);
2429 }
2430 unlock_page(vmf->page);
2431 wp_page_reuse(vmf);
2432 return VM_FAULT_WRITE;
2433 }
2434 unlock_page(vmf->page);
2435 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2436 (VM_WRITE|VM_SHARED))) {
2437 return wp_page_shared(vmf);
2438 }
2439
2440 /*
2441 * Ok, we need to copy. Oh, well..
2442 */
2443 get_page(vmf->page);
2444
2445 pte_unmap_unlock(vmf->pte, vmf->ptl);
2446 return wp_page_copy(vmf);
2447}
2448
2449static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2450 unsigned long start_addr, unsigned long end_addr,
2451 struct zap_details *details)
2452{
2453 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2454}
2455
2456static inline void unmap_mapping_range_tree(struct rb_root *root,
2457 struct zap_details *details)
2458{
2459 struct vm_area_struct *vma;
2460 pgoff_t vba, vea, zba, zea;
2461
2462 vma_interval_tree_foreach(vma, root,
2463 details->first_index, details->last_index) {
2464
2465 vba = vma->vm_pgoff;
2466 vea = vba + vma_pages(vma) - 1;
2467 zba = details->first_index;
2468 if (zba < vba)
2469 zba = vba;
2470 zea = details->last_index;
2471 if (zea > vea)
2472 zea = vea;
2473
2474 unmap_mapping_range_vma(vma,
2475 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2476 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2477 details);
2478 }
2479}
2480
2481/**
2482 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2483 * address_space corresponding to the specified page range in the underlying
2484 * file.
2485 *
2486 * @mapping: the address space containing mmaps to be unmapped.
2487 * @holebegin: byte in first page to unmap, relative to the start of
2488 * the underlying file. This will be rounded down to a PAGE_SIZE
2489 * boundary. Note that this is different from truncate_pagecache(), which
2490 * must keep the partial page. In contrast, we must get rid of
2491 * partial pages.
2492 * @holelen: size of prospective hole in bytes. This will be rounded
2493 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2494 * end of the file.
2495 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2496 * but 0 when invalidating pagecache, don't throw away private data.
2497 */
2498void unmap_mapping_range(struct address_space *mapping,
2499 loff_t const holebegin, loff_t const holelen, int even_cows)
2500{
2501 struct zap_details details = { };
2502 pgoff_t hba = holebegin >> PAGE_SHIFT;
2503 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2504
2505 /* Check for overflow. */
2506 if (sizeof(holelen) > sizeof(hlen)) {
2507 long long holeend =
2508 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2509 if (holeend & ~(long long)ULONG_MAX)
2510 hlen = ULONG_MAX - hba + 1;
2511 }
2512
2513 details.check_mapping = even_cows? NULL: mapping;
2514 details.first_index = hba;
2515 details.last_index = hba + hlen - 1;
2516 if (details.last_index < details.first_index)
2517 details.last_index = ULONG_MAX;
2518
2519 i_mmap_lock_write(mapping);
2520 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2521 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2522 i_mmap_unlock_write(mapping);
2523}
2524EXPORT_SYMBOL(unmap_mapping_range);
2525
2526/*
2527 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2528 * but allow concurrent faults), and pte mapped but not yet locked.
2529 * We return with pte unmapped and unlocked.
2530 *
2531 * We return with the mmap_sem locked or unlocked in the same cases
2532 * as does filemap_fault().
2533 */
2534int do_swap_page(struct vm_fault *vmf)
2535{
2536 struct vm_area_struct *vma = vmf->vma;
2537 struct page *page, *swapcache;
2538 struct mem_cgroup *memcg;
2539 swp_entry_t entry;
2540 pte_t pte;
2541 int locked;
2542 int exclusive = 0;
2543 int ret = 0;
2544
2545 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2546 goto out;
2547
2548 entry = pte_to_swp_entry(vmf->orig_pte);
2549 if (unlikely(non_swap_entry(entry))) {
2550 if (is_migration_entry(entry)) {
2551 migration_entry_wait(vma->vm_mm, vmf->pmd,
2552 vmf->address);
2553 } else if (is_hwpoison_entry(entry)) {
2554 ret = VM_FAULT_HWPOISON;
2555 } else {
2556 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2557 ret = VM_FAULT_SIGBUS;
2558 }
2559 goto out;
2560 }
2561 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2562 page = lookup_swap_cache(entry);
2563 if (!page) {
2564 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, vma,
2565 vmf->address);
2566 if (!page) {
2567 /*
2568 * Back out if somebody else faulted in this pte
2569 * while we released the pte lock.
2570 */
2571 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2572 vmf->address, &vmf->ptl);
2573 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2574 ret = VM_FAULT_OOM;
2575 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2576 goto unlock;
2577 }
2578
2579 /* Had to read the page from swap area: Major fault */
2580 ret = VM_FAULT_MAJOR;
2581 count_vm_event(PGMAJFAULT);
2582 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
2583 } else if (PageHWPoison(page)) {
2584 /*
2585 * hwpoisoned dirty swapcache pages are kept for killing
2586 * owner processes (which may be unknown at hwpoison time)
2587 */
2588 ret = VM_FAULT_HWPOISON;
2589 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2590 swapcache = page;
2591 goto out_release;
2592 }
2593
2594 swapcache = page;
2595 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2596
2597 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2598 if (!locked) {
2599 ret |= VM_FAULT_RETRY;
2600 goto out_release;
2601 }
2602
2603 /*
2604 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2605 * release the swapcache from under us. The page pin, and pte_same
2606 * test below, are not enough to exclude that. Even if it is still
2607 * swapcache, we need to check that the page's swap has not changed.
2608 */
2609 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2610 goto out_page;
2611
2612 page = ksm_might_need_to_copy(page, vma, vmf->address);
2613 if (unlikely(!page)) {
2614 ret = VM_FAULT_OOM;
2615 page = swapcache;
2616 goto out_page;
2617 }
2618
2619 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2620 &memcg, false)) {
2621 ret = VM_FAULT_OOM;
2622 goto out_page;
2623 }
2624
2625 /*
2626 * Back out if somebody else already faulted in this pte.
2627 */
2628 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2629 &vmf->ptl);
2630 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2631 goto out_nomap;
2632
2633 if (unlikely(!PageUptodate(page))) {
2634 ret = VM_FAULT_SIGBUS;
2635 goto out_nomap;
2636 }
2637
2638 /*
2639 * The page isn't present yet, go ahead with the fault.
2640 *
2641 * Be careful about the sequence of operations here.
2642 * To get its accounting right, reuse_swap_page() must be called
2643 * while the page is counted on swap but not yet in mapcount i.e.
2644 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2645 * must be called after the swap_free(), or it will never succeed.
2646 */
2647
2648 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2649 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2650 pte = mk_pte(page, vma->vm_page_prot);
2651 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2652 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2653 vmf->flags &= ~FAULT_FLAG_WRITE;
2654 ret |= VM_FAULT_WRITE;
2655 exclusive = RMAP_EXCLUSIVE;
2656 }
2657 flush_icache_page(vma, page);
2658 if (pte_swp_soft_dirty(vmf->orig_pte))
2659 pte = pte_mksoft_dirty(pte);
2660 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2661 vmf->orig_pte = pte;
2662 if (page == swapcache) {
2663 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2664 mem_cgroup_commit_charge(page, memcg, true, false);
2665 activate_page(page);
2666 } else { /* ksm created a completely new copy */
2667 page_add_new_anon_rmap(page, vma, vmf->address, false);
2668 mem_cgroup_commit_charge(page, memcg, false, false);
2669 lru_cache_add_active_or_unevictable(page, vma);
2670 }
2671
2672 swap_free(entry);
2673 if (mem_cgroup_swap_full(page) ||
2674 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2675 try_to_free_swap(page);
2676 unlock_page(page);
2677 if (page != swapcache) {
2678 /*
2679 * Hold the lock to avoid the swap entry to be reused
2680 * until we take the PT lock for the pte_same() check
2681 * (to avoid false positives from pte_same). For
2682 * further safety release the lock after the swap_free
2683 * so that the swap count won't change under a
2684 * parallel locked swapcache.
2685 */
2686 unlock_page(swapcache);
2687 put_page(swapcache);
2688 }
2689
2690 if (vmf->flags & FAULT_FLAG_WRITE) {
2691 ret |= do_wp_page(vmf);
2692 if (ret & VM_FAULT_ERROR)
2693 ret &= VM_FAULT_ERROR;
2694 goto out;
2695 }
2696
2697 /* No need to invalidate - it was non-present before */
2698 update_mmu_cache(vma, vmf->address, vmf->pte);
2699unlock:
2700 pte_unmap_unlock(vmf->pte, vmf->ptl);
2701out:
2702 return ret;
2703out_nomap:
2704 mem_cgroup_cancel_charge(page, memcg, false);
2705 pte_unmap_unlock(vmf->pte, vmf->ptl);
2706out_page:
2707 unlock_page(page);
2708out_release:
2709 put_page(page);
2710 if (page != swapcache) {
2711 unlock_page(swapcache);
2712 put_page(swapcache);
2713 }
2714 return ret;
2715}
2716
2717/*
2718 * This is like a special single-page "expand_{down|up}wards()",
2719 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2720 * doesn't hit another vma.
2721 */
2722static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2723{
2724 address &= PAGE_MASK;
2725 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2726 struct vm_area_struct *prev = vma->vm_prev;
2727
2728 /*
2729 * Is there a mapping abutting this one below?
2730 *
2731 * That's only ok if it's the same stack mapping
2732 * that has gotten split..
2733 */
2734 if (prev && prev->vm_end == address)
2735 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2736
2737 return expand_downwards(vma, address - PAGE_SIZE);
2738 }
2739 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2740 struct vm_area_struct *next = vma->vm_next;
2741
2742 /* As VM_GROWSDOWN but s/below/above/ */
2743 if (next && next->vm_start == address + PAGE_SIZE)
2744 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2745
2746 return expand_upwards(vma, address + PAGE_SIZE);
2747 }
2748 return 0;
2749}
2750
2751/*
2752 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2753 * but allow concurrent faults), and pte mapped but not yet locked.
2754 * We return with mmap_sem still held, but pte unmapped and unlocked.
2755 */
2756static int do_anonymous_page(struct vm_fault *vmf)
2757{
2758 struct vm_area_struct *vma = vmf->vma;
2759 struct mem_cgroup *memcg;
2760 struct page *page;
2761 pte_t entry;
2762
2763 /* File mapping without ->vm_ops ? */
2764 if (vma->vm_flags & VM_SHARED)
2765 return VM_FAULT_SIGBUS;
2766
2767 /* Check if we need to add a guard page to the stack */
2768 if (check_stack_guard_page(vma, vmf->address) < 0)
2769 return VM_FAULT_SIGSEGV;
2770
2771 /*
2772 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2773 * pte_offset_map() on pmds where a huge pmd might be created
2774 * from a different thread.
2775 *
2776 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2777 * parallel threads are excluded by other means.
2778 *
2779 * Here we only have down_read(mmap_sem).
2780 */
2781 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2782 return VM_FAULT_OOM;
2783
2784 /* See the comment in pte_alloc_one_map() */
2785 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2786 return 0;
2787
2788 /* Use the zero-page for reads */
2789 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2790 !mm_forbids_zeropage(vma->vm_mm)) {
2791 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2792 vma->vm_page_prot));
2793 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2794 vmf->address, &vmf->ptl);
2795 if (!pte_none(*vmf->pte))
2796 goto unlock;
2797 /* Deliver the page fault to userland, check inside PT lock */
2798 if (userfaultfd_missing(vma)) {
2799 pte_unmap_unlock(vmf->pte, vmf->ptl);
2800 return handle_userfault(vmf, VM_UFFD_MISSING);
2801 }
2802 goto setpte;
2803 }
2804
2805 /* Allocate our own private page. */
2806 if (unlikely(anon_vma_prepare(vma)))
2807 goto oom;
2808 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2809 if (!page)
2810 goto oom;
2811
2812 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2813 goto oom_free_page;
2814
2815 /*
2816 * The memory barrier inside __SetPageUptodate makes sure that
2817 * preceeding stores to the page contents become visible before
2818 * the set_pte_at() write.
2819 */
2820 __SetPageUptodate(page);
2821
2822 entry = mk_pte(page, vma->vm_page_prot);
2823 if (vma->vm_flags & VM_WRITE)
2824 entry = pte_mkwrite(pte_mkdirty(entry));
2825
2826 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2827 &vmf->ptl);
2828 if (!pte_none(*vmf->pte))
2829 goto release;
2830
2831 /* Deliver the page fault to userland, check inside PT lock */
2832 if (userfaultfd_missing(vma)) {
2833 pte_unmap_unlock(vmf->pte, vmf->ptl);
2834 mem_cgroup_cancel_charge(page, memcg, false);
2835 put_page(page);
2836 return handle_userfault(vmf, VM_UFFD_MISSING);
2837 }
2838
2839 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2840 page_add_new_anon_rmap(page, vma, vmf->address, false);
2841 mem_cgroup_commit_charge(page, memcg, false, false);
2842 lru_cache_add_active_or_unevictable(page, vma);
2843setpte:
2844 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2845
2846 /* No need to invalidate - it was non-present before */
2847 update_mmu_cache(vma, vmf->address, vmf->pte);
2848unlock:
2849 pte_unmap_unlock(vmf->pte, vmf->ptl);
2850 return 0;
2851release:
2852 mem_cgroup_cancel_charge(page, memcg, false);
2853 put_page(page);
2854 goto unlock;
2855oom_free_page:
2856 put_page(page);
2857oom:
2858 return VM_FAULT_OOM;
2859}
2860
2861/*
2862 * The mmap_sem must have been held on entry, and may have been
2863 * released depending on flags and vma->vm_ops->fault() return value.
2864 * See filemap_fault() and __lock_page_retry().
2865 */
2866static int __do_fault(struct vm_fault *vmf)
2867{
2868 struct vm_area_struct *vma = vmf->vma;
2869 int ret;
2870
2871 ret = vma->vm_ops->fault(vma, vmf);
2872 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
2873 VM_FAULT_DONE_COW)))
2874 return ret;
2875
2876 if (unlikely(PageHWPoison(vmf->page))) {
2877 if (ret & VM_FAULT_LOCKED)
2878 unlock_page(vmf->page);
2879 put_page(vmf->page);
2880 vmf->page = NULL;
2881 return VM_FAULT_HWPOISON;
2882 }
2883
2884 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2885 lock_page(vmf->page);
2886 else
2887 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
2888
2889 return ret;
2890}
2891
2892static int pte_alloc_one_map(struct vm_fault *vmf)
2893{
2894 struct vm_area_struct *vma = vmf->vma;
2895
2896 if (!pmd_none(*vmf->pmd))
2897 goto map_pte;
2898 if (vmf->prealloc_pte) {
2899 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2900 if (unlikely(!pmd_none(*vmf->pmd))) {
2901 spin_unlock(vmf->ptl);
2902 goto map_pte;
2903 }
2904
2905 atomic_long_inc(&vma->vm_mm->nr_ptes);
2906 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2907 spin_unlock(vmf->ptl);
2908 vmf->prealloc_pte = 0;
2909 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
2910 return VM_FAULT_OOM;
2911 }
2912map_pte:
2913 /*
2914 * If a huge pmd materialized under us just retry later. Use
2915 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
2916 * didn't become pmd_trans_huge under us and then back to pmd_none, as
2917 * a result of MADV_DONTNEED running immediately after a huge pmd fault
2918 * in a different thread of this mm, in turn leading to a misleading
2919 * pmd_trans_huge() retval. All we have to ensure is that it is a
2920 * regular pmd that we can walk with pte_offset_map() and we can do that
2921 * through an atomic read in C, which is what pmd_trans_unstable()
2922 * provides.
2923 */
2924 if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
2925 return VM_FAULT_NOPAGE;
2926
2927 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2928 &vmf->ptl);
2929 return 0;
2930}
2931
2932#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
2933
2934#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2935static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2936 unsigned long haddr)
2937{
2938 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2939 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2940 return false;
2941 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
2942 return false;
2943 return true;
2944}
2945
2946static void deposit_prealloc_pte(struct vm_fault *vmf)
2947{
2948 struct vm_area_struct *vma = vmf->vma;
2949
2950 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2951 /*
2952 * We are going to consume the prealloc table,
2953 * count that as nr_ptes.
2954 */
2955 atomic_long_inc(&vma->vm_mm->nr_ptes);
2956 vmf->prealloc_pte = 0;
2957}
2958
2959static int do_set_pmd(struct vm_fault *vmf, struct page *page)
2960{
2961 struct vm_area_struct *vma = vmf->vma;
2962 bool write = vmf->flags & FAULT_FLAG_WRITE;
2963 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
2964 pmd_t entry;
2965 int i, ret;
2966
2967 if (!transhuge_vma_suitable(vma, haddr))
2968 return VM_FAULT_FALLBACK;
2969
2970 ret = VM_FAULT_FALLBACK;
2971 page = compound_head(page);
2972
2973 /*
2974 * Archs like ppc64 need additonal space to store information
2975 * related to pte entry. Use the preallocated table for that.
2976 */
2977 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
2978 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
2979 if (!vmf->prealloc_pte)
2980 return VM_FAULT_OOM;
2981 smp_wmb(); /* See comment in __pte_alloc() */
2982 }
2983
2984 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2985 if (unlikely(!pmd_none(*vmf->pmd)))
2986 goto out;
2987
2988 for (i = 0; i < HPAGE_PMD_NR; i++)
2989 flush_icache_page(vma, page + i);
2990
2991 entry = mk_huge_pmd(page, vma->vm_page_prot);
2992 if (write)
2993 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2994
2995 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
2996 page_add_file_rmap(page, true);
2997 /*
2998 * deposit and withdraw with pmd lock held
2999 */
3000 if (arch_needs_pgtable_deposit())
3001 deposit_prealloc_pte(vmf);
3002
3003 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3004
3005 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3006
3007 /* fault is handled */
3008 ret = 0;
3009 count_vm_event(THP_FILE_MAPPED);
3010out:
3011 spin_unlock(vmf->ptl);
3012 return ret;
3013}
3014#else
3015static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3016{
3017 BUILD_BUG();
3018 return 0;
3019}
3020#endif
3021
3022/**
3023 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3024 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3025 *
3026 * @vmf: fault environment
3027 * @memcg: memcg to charge page (only for private mappings)
3028 * @page: page to map
3029 *
3030 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3031 * return.
3032 *
3033 * Target users are page handler itself and implementations of
3034 * vm_ops->map_pages.
3035 */
3036int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3037 struct page *page)
3038{
3039 struct vm_area_struct *vma = vmf->vma;
3040 bool write = vmf->flags & FAULT_FLAG_WRITE;
3041 pte_t entry;
3042 int ret;
3043
3044 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3045 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3046 /* THP on COW? */
3047 VM_BUG_ON_PAGE(memcg, page);
3048
3049 ret = do_set_pmd(vmf, page);
3050 if (ret != VM_FAULT_FALLBACK)
3051 return ret;
3052 }
3053
3054 if (!vmf->pte) {
3055 ret = pte_alloc_one_map(vmf);
3056 if (ret)
3057 return ret;
3058 }
3059
3060 /* Re-check under ptl */
3061 if (unlikely(!pte_none(*vmf->pte)))
3062 return VM_FAULT_NOPAGE;
3063
3064 flush_icache_page(vma, page);
3065 entry = mk_pte(page, vma->vm_page_prot);
3066 if (write)
3067 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3068 /* copy-on-write page */
3069 if (write && !(vma->vm_flags & VM_SHARED)) {
3070 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3071 page_add_new_anon_rmap(page, vma, vmf->address, false);
3072 mem_cgroup_commit_charge(page, memcg, false, false);
3073 lru_cache_add_active_or_unevictable(page, vma);
3074 } else {
3075 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3076 page_add_file_rmap(page, false);
3077 }
3078 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3079
3080 /* no need to invalidate: a not-present page won't be cached */
3081 update_mmu_cache(vma, vmf->address, vmf->pte);
3082
3083 return 0;
3084}
3085
3086
3087/**
3088 * finish_fault - finish page fault once we have prepared the page to fault
3089 *
3090 * @vmf: structure describing the fault
3091 *
3092 * This function handles all that is needed to finish a page fault once the
3093 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3094 * given page, adds reverse page mapping, handles memcg charges and LRU
3095 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3096 * error.
3097 *
3098 * The function expects the page to be locked and on success it consumes a
3099 * reference of a page being mapped (for the PTE which maps it).
3100 */
3101int finish_fault(struct vm_fault *vmf)
3102{
3103 struct page *page;
3104 int ret;
3105
3106 /* Did we COW the page? */
3107 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3108 !(vmf->vma->vm_flags & VM_SHARED))
3109 page = vmf->cow_page;
3110 else
3111 page = vmf->page;
3112 ret = alloc_set_pte(vmf, vmf->memcg, page);
3113 if (vmf->pte)
3114 pte_unmap_unlock(vmf->pte, vmf->ptl);
3115 return ret;
3116}
3117
3118static unsigned long fault_around_bytes __read_mostly =
3119 rounddown_pow_of_two(65536);
3120
3121#ifdef CONFIG_DEBUG_FS
3122static int fault_around_bytes_get(void *data, u64 *val)
3123{
3124 *val = fault_around_bytes;
3125 return 0;
3126}
3127
3128/*
3129 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3130 * rounded down to nearest page order. It's what do_fault_around() expects to
3131 * see.
3132 */
3133static int fault_around_bytes_set(void *data, u64 val)
3134{
3135 if (val / PAGE_SIZE > PTRS_PER_PTE)
3136 return -EINVAL;
3137 if (val > PAGE_SIZE)
3138 fault_around_bytes = rounddown_pow_of_two(val);
3139 else
3140 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3141 return 0;
3142}
3143DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3144 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3145
3146static int __init fault_around_debugfs(void)
3147{
3148 void *ret;
3149
3150 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3151 &fault_around_bytes_fops);
3152 if (!ret)
3153 pr_warn("Failed to create fault_around_bytes in debugfs");
3154 return 0;
3155}
3156late_initcall(fault_around_debugfs);
3157#endif
3158
3159/*
3160 * do_fault_around() tries to map few pages around the fault address. The hope
3161 * is that the pages will be needed soon and this will lower the number of
3162 * faults to handle.
3163 *
3164 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3165 * not ready to be mapped: not up-to-date, locked, etc.
3166 *
3167 * This function is called with the page table lock taken. In the split ptlock
3168 * case the page table lock only protects only those entries which belong to
3169 * the page table corresponding to the fault address.
3170 *
3171 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3172 * only once.
3173 *
3174 * fault_around_pages() defines how many pages we'll try to map.
3175 * do_fault_around() expects it to return a power of two less than or equal to
3176 * PTRS_PER_PTE.
3177 *
3178 * The virtual address of the area that we map is naturally aligned to the
3179 * fault_around_pages() value (and therefore to page order). This way it's
3180 * easier to guarantee that we don't cross page table boundaries.
3181 */
3182static int do_fault_around(struct vm_fault *vmf)
3183{
3184 unsigned long address = vmf->address, nr_pages, mask;
3185 pgoff_t start_pgoff = vmf->pgoff;
3186 pgoff_t end_pgoff;
3187 int off, ret = 0;
3188
3189 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3190 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3191
3192 vmf->address = max(address & mask, vmf->vma->vm_start);
3193 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3194 start_pgoff -= off;
3195
3196 /*
3197 * end_pgoff is either end of page table or end of vma
3198 * or fault_around_pages() from start_pgoff, depending what is nearest.
3199 */
3200 end_pgoff = start_pgoff -
3201 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3202 PTRS_PER_PTE - 1;
3203 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3204 start_pgoff + nr_pages - 1);
3205
3206 if (pmd_none(*vmf->pmd)) {
3207 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3208 vmf->address);
3209 if (!vmf->prealloc_pte)
3210 goto out;
3211 smp_wmb(); /* See comment in __pte_alloc() */
3212 }
3213
3214 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3215
3216 /* Huge page is mapped? Page fault is solved */
3217 if (pmd_trans_huge(*vmf->pmd)) {
3218 ret = VM_FAULT_NOPAGE;
3219 goto out;
3220 }
3221
3222 /* ->map_pages() haven't done anything useful. Cold page cache? */
3223 if (!vmf->pte)
3224 goto out;
3225
3226 /* check if the page fault is solved */
3227 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3228 if (!pte_none(*vmf->pte))
3229 ret = VM_FAULT_NOPAGE;
3230 pte_unmap_unlock(vmf->pte, vmf->ptl);
3231out:
3232 vmf->address = address;
3233 vmf->pte = NULL;
3234 return ret;
3235}
3236
3237static int do_read_fault(struct vm_fault *vmf)
3238{
3239 struct vm_area_struct *vma = vmf->vma;
3240 int ret = 0;
3241
3242 /*
3243 * Let's call ->map_pages() first and use ->fault() as fallback
3244 * if page by the offset is not ready to be mapped (cold cache or
3245 * something).
3246 */
3247 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3248 ret = do_fault_around(vmf);
3249 if (ret)
3250 return ret;
3251 }
3252
3253 ret = __do_fault(vmf);
3254 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3255 return ret;
3256
3257 ret |= finish_fault(vmf);
3258 unlock_page(vmf->page);
3259 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3260 put_page(vmf->page);
3261 return ret;
3262}
3263
3264static int do_cow_fault(struct vm_fault *vmf)
3265{
3266 struct vm_area_struct *vma = vmf->vma;
3267 int ret;
3268
3269 if (unlikely(anon_vma_prepare(vma)))
3270 return VM_FAULT_OOM;
3271
3272 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3273 if (!vmf->cow_page)
3274 return VM_FAULT_OOM;
3275
3276 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3277 &vmf->memcg, false)) {
3278 put_page(vmf->cow_page);
3279 return VM_FAULT_OOM;
3280 }
3281
3282 ret = __do_fault(vmf);
3283 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3284 goto uncharge_out;
3285 if (ret & VM_FAULT_DONE_COW)
3286 return ret;
3287
3288 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3289 __SetPageUptodate(vmf->cow_page);
3290
3291 ret |= finish_fault(vmf);
3292 unlock_page(vmf->page);
3293 put_page(vmf->page);
3294 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3295 goto uncharge_out;
3296 return ret;
3297uncharge_out:
3298 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3299 put_page(vmf->cow_page);
3300 return ret;
3301}
3302
3303static int do_shared_fault(struct vm_fault *vmf)
3304{
3305 struct vm_area_struct *vma = vmf->vma;
3306 int ret, tmp;
3307
3308 ret = __do_fault(vmf);
3309 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3310 return ret;
3311
3312 /*
3313 * Check if the backing address space wants to know that the page is
3314 * about to become writable
3315 */
3316 if (vma->vm_ops->page_mkwrite) {
3317 unlock_page(vmf->page);
3318 tmp = do_page_mkwrite(vmf);
3319 if (unlikely(!tmp ||
3320 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3321 put_page(vmf->page);
3322 return tmp;
3323 }
3324 }
3325
3326 ret |= finish_fault(vmf);
3327 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3328 VM_FAULT_RETRY))) {
3329 unlock_page(vmf->page);
3330 put_page(vmf->page);
3331 return ret;
3332 }
3333
3334 fault_dirty_shared_page(vma, vmf->page);
3335 return ret;
3336}
3337
3338/*
3339 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3340 * but allow concurrent faults).
3341 * The mmap_sem may have been released depending on flags and our
3342 * return value. See filemap_fault() and __lock_page_or_retry().
3343 */
3344static int do_fault(struct vm_fault *vmf)
3345{
3346 struct vm_area_struct *vma = vmf->vma;
3347 int ret;
3348
3349 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3350 if (!vma->vm_ops->fault)
3351 ret = VM_FAULT_SIGBUS;
3352 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3353 ret = do_read_fault(vmf);
3354 else if (!(vma->vm_flags & VM_SHARED))
3355 ret = do_cow_fault(vmf);
3356 else
3357 ret = do_shared_fault(vmf);
3358
3359 /* preallocated pagetable is unused: free it */
3360 if (vmf->prealloc_pte) {
3361 pte_free(vma->vm_mm, vmf->prealloc_pte);
3362 vmf->prealloc_pte = 0;
3363 }
3364 return ret;
3365}
3366
3367static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3368 unsigned long addr, int page_nid,
3369 int *flags)
3370{
3371 get_page(page);
3372
3373 count_vm_numa_event(NUMA_HINT_FAULTS);
3374 if (page_nid == numa_node_id()) {
3375 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3376 *flags |= TNF_FAULT_LOCAL;
3377 }
3378
3379 return mpol_misplaced(page, vma, addr);
3380}
3381
3382static int do_numa_page(struct vm_fault *vmf)
3383{
3384 struct vm_area_struct *vma = vmf->vma;
3385 struct page *page = NULL;
3386 int page_nid = -1;
3387 int last_cpupid;
3388 int target_nid;
3389 bool migrated = false;
3390 pte_t pte = vmf->orig_pte;
3391 bool was_writable = pte_write(pte);
3392 int flags = 0;
3393
3394 /*
3395 * The "pte" at this point cannot be used safely without
3396 * validation through pte_unmap_same(). It's of NUMA type but
3397 * the pfn may be screwed if the read is non atomic.
3398 *
3399 * We can safely just do a "set_pte_at()", because the old
3400 * page table entry is not accessible, so there would be no
3401 * concurrent hardware modifications to the PTE.
3402 */
3403 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3404 spin_lock(vmf->ptl);
3405 if (unlikely(!pte_same(*vmf->pte, pte))) {
3406 pte_unmap_unlock(vmf->pte, vmf->ptl);
3407 goto out;
3408 }
3409
3410 /* Make it present again */
3411 pte = pte_modify(pte, vma->vm_page_prot);
3412 pte = pte_mkyoung(pte);
3413 if (was_writable)
3414 pte = pte_mkwrite(pte);
3415 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3416 update_mmu_cache(vma, vmf->address, vmf->pte);
3417
3418 page = vm_normal_page(vma, vmf->address, pte);
3419 if (!page) {
3420 pte_unmap_unlock(vmf->pte, vmf->ptl);
3421 return 0;
3422 }
3423
3424 /* TODO: handle PTE-mapped THP */
3425 if (PageCompound(page)) {
3426 pte_unmap_unlock(vmf->pte, vmf->ptl);
3427 return 0;
3428 }
3429
3430 /*
3431 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3432 * much anyway since they can be in shared cache state. This misses
3433 * the case where a mapping is writable but the process never writes
3434 * to it but pte_write gets cleared during protection updates and
3435 * pte_dirty has unpredictable behaviour between PTE scan updates,
3436 * background writeback, dirty balancing and application behaviour.
3437 */
3438 if (!pte_write(pte))
3439 flags |= TNF_NO_GROUP;
3440
3441 /*
3442 * Flag if the page is shared between multiple address spaces. This
3443 * is later used when determining whether to group tasks together
3444 */
3445 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3446 flags |= TNF_SHARED;
3447
3448 last_cpupid = page_cpupid_last(page);
3449 page_nid = page_to_nid(page);
3450 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3451 &flags);
3452 pte_unmap_unlock(vmf->pte, vmf->ptl);
3453 if (target_nid == -1) {
3454 put_page(page);
3455 goto out;
3456 }
3457
3458 /* Migrate to the requested node */
3459 migrated = migrate_misplaced_page(page, vma, target_nid);
3460 if (migrated) {
3461 page_nid = target_nid;
3462 flags |= TNF_MIGRATED;
3463 } else
3464 flags |= TNF_MIGRATE_FAIL;
3465
3466out:
3467 if (page_nid != -1)
3468 task_numa_fault(last_cpupid, page_nid, 1, flags);
3469 return 0;
3470}
3471
3472static int create_huge_pmd(struct vm_fault *vmf)
3473{
3474 struct vm_area_struct *vma = vmf->vma;
3475 if (vma_is_anonymous(vma))
3476 return do_huge_pmd_anonymous_page(vmf);
3477 if (vma->vm_ops->pmd_fault)
3478 return vma->vm_ops->pmd_fault(vma, vmf->address, vmf->pmd,
3479 vmf->flags);
3480 return VM_FAULT_FALLBACK;
3481}
3482
3483static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3484{
3485 if (vma_is_anonymous(vmf->vma))
3486 return do_huge_pmd_wp_page(vmf, orig_pmd);
3487 if (vmf->vma->vm_ops->pmd_fault)
3488 return vmf->vma->vm_ops->pmd_fault(vmf->vma, vmf->address,
3489 vmf->pmd, vmf->flags);
3490
3491 /* COW handled on pte level: split pmd */
3492 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3493 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3494
3495 return VM_FAULT_FALLBACK;
3496}
3497
3498static inline bool vma_is_accessible(struct vm_area_struct *vma)
3499{
3500 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3501}
3502
3503/*
3504 * These routines also need to handle stuff like marking pages dirty
3505 * and/or accessed for architectures that don't do it in hardware (most
3506 * RISC architectures). The early dirtying is also good on the i386.
3507 *
3508 * There is also a hook called "update_mmu_cache()" that architectures
3509 * with external mmu caches can use to update those (ie the Sparc or
3510 * PowerPC hashed page tables that act as extended TLBs).
3511 *
3512 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3513 * concurrent faults).
3514 *
3515 * The mmap_sem may have been released depending on flags and our return value.
3516 * See filemap_fault() and __lock_page_or_retry().
3517 */
3518static int handle_pte_fault(struct vm_fault *vmf)
3519{
3520 pte_t entry;
3521
3522 if (unlikely(pmd_none(*vmf->pmd))) {
3523 /*
3524 * Leave __pte_alloc() until later: because vm_ops->fault may
3525 * want to allocate huge page, and if we expose page table
3526 * for an instant, it will be difficult to retract from
3527 * concurrent faults and from rmap lookups.
3528 */
3529 vmf->pte = NULL;
3530 } else {
3531 /* See comment in pte_alloc_one_map() */
3532 if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
3533 return 0;
3534 /*
3535 * A regular pmd is established and it can't morph into a huge
3536 * pmd from under us anymore at this point because we hold the
3537 * mmap_sem read mode and khugepaged takes it in write mode.
3538 * So now it's safe to run pte_offset_map().
3539 */
3540 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3541 vmf->orig_pte = *vmf->pte;
3542
3543 /*
3544 * some architectures can have larger ptes than wordsize,
3545 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3546 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3547 * atomic accesses. The code below just needs a consistent
3548 * view for the ifs and we later double check anyway with the
3549 * ptl lock held. So here a barrier will do.
3550 */
3551 barrier();
3552 if (pte_none(vmf->orig_pte)) {
3553 pte_unmap(vmf->pte);
3554 vmf->pte = NULL;
3555 }
3556 }
3557
3558 if (!vmf->pte) {
3559 if (vma_is_anonymous(vmf->vma))
3560 return do_anonymous_page(vmf);
3561 else
3562 return do_fault(vmf);
3563 }
3564
3565 if (!pte_present(vmf->orig_pte))
3566 return do_swap_page(vmf);
3567
3568 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3569 return do_numa_page(vmf);
3570
3571 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3572 spin_lock(vmf->ptl);
3573 entry = vmf->orig_pte;
3574 if (unlikely(!pte_same(*vmf->pte, entry)))
3575 goto unlock;
3576 if (vmf->flags & FAULT_FLAG_WRITE) {
3577 if (!pte_write(entry))
3578 return do_wp_page(vmf);
3579 entry = pte_mkdirty(entry);
3580 }
3581 entry = pte_mkyoung(entry);
3582 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3583 vmf->flags & FAULT_FLAG_WRITE)) {
3584 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3585 } else {
3586 /*
3587 * This is needed only for protection faults but the arch code
3588 * is not yet telling us if this is a protection fault or not.
3589 * This still avoids useless tlb flushes for .text page faults
3590 * with threads.
3591 */
3592 if (vmf->flags & FAULT_FLAG_WRITE)
3593 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3594 }
3595unlock:
3596 pte_unmap_unlock(vmf->pte, vmf->ptl);
3597 return 0;
3598}
3599
3600/*
3601 * By the time we get here, we already hold the mm semaphore
3602 *
3603 * The mmap_sem may have been released depending on flags and our
3604 * return value. See filemap_fault() and __lock_page_or_retry().
3605 */
3606static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3607 unsigned int flags)
3608{
3609 struct vm_fault vmf = {
3610 .vma = vma,
3611 .address = address & PAGE_MASK,
3612 .flags = flags,
3613 .pgoff = linear_page_index(vma, address),
3614 .gfp_mask = __get_fault_gfp_mask(vma),
3615 };
3616 struct mm_struct *mm = vma->vm_mm;
3617 pgd_t *pgd;
3618 pud_t *pud;
3619
3620 pgd = pgd_offset(mm, address);
3621 pud = pud_alloc(mm, pgd, address);
3622 if (!pud)
3623 return VM_FAULT_OOM;
3624 vmf.pmd = pmd_alloc(mm, pud, address);
3625 if (!vmf.pmd)
3626 return VM_FAULT_OOM;
3627 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3628 int ret = create_huge_pmd(&vmf);
3629 if (!(ret & VM_FAULT_FALLBACK))
3630 return ret;
3631 } else {
3632 pmd_t orig_pmd = *vmf.pmd;
3633 int ret;
3634
3635 barrier();
3636 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3637 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3638 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3639
3640 if ((vmf.flags & FAULT_FLAG_WRITE) &&
3641 !pmd_write(orig_pmd)) {
3642 ret = wp_huge_pmd(&vmf, orig_pmd);
3643 if (!(ret & VM_FAULT_FALLBACK))
3644 return ret;
3645 } else {
3646 huge_pmd_set_accessed(&vmf, orig_pmd);
3647 return 0;
3648 }
3649 }
3650 }
3651
3652 return handle_pte_fault(&vmf);
3653}
3654
3655/*
3656 * By the time we get here, we already hold the mm semaphore
3657 *
3658 * The mmap_sem may have been released depending on flags and our
3659 * return value. See filemap_fault() and __lock_page_or_retry().
3660 */
3661int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3662 unsigned int flags)
3663{
3664 int ret;
3665
3666 __set_current_state(TASK_RUNNING);
3667
3668 count_vm_event(PGFAULT);
3669 mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
3670
3671 /* do counter updates before entering really critical section. */
3672 check_sync_rss_stat(current);
3673
3674 /*
3675 * Enable the memcg OOM handling for faults triggered in user
3676 * space. Kernel faults are handled more gracefully.
3677 */
3678 if (flags & FAULT_FLAG_USER)
3679 mem_cgroup_oom_enable();
3680
3681 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3682 flags & FAULT_FLAG_INSTRUCTION,
3683 flags & FAULT_FLAG_REMOTE))
3684 return VM_FAULT_SIGSEGV;
3685
3686 if (unlikely(is_vm_hugetlb_page(vma)))
3687 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3688 else
3689 ret = __handle_mm_fault(vma, address, flags);
3690
3691 if (flags & FAULT_FLAG_USER) {
3692 mem_cgroup_oom_disable();
3693 /*
3694 * The task may have entered a memcg OOM situation but
3695 * if the allocation error was handled gracefully (no
3696 * VM_FAULT_OOM), there is no need to kill anything.
3697 * Just clean up the OOM state peacefully.
3698 */
3699 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3700 mem_cgroup_oom_synchronize(false);
3701 }
3702
3703 /*
3704 * This mm has been already reaped by the oom reaper and so the
3705 * refault cannot be trusted in general. Anonymous refaults would
3706 * lose data and give a zero page instead e.g. This is especially
3707 * problem for use_mm() because regular tasks will just die and
3708 * the corrupted data will not be visible anywhere while kthread
3709 * will outlive the oom victim and potentially propagate the date
3710 * further.
3711 */
3712 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3713 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3714 ret = VM_FAULT_SIGBUS;
3715
3716 return ret;
3717}
3718EXPORT_SYMBOL_GPL(handle_mm_fault);
3719
3720#ifndef __PAGETABLE_PUD_FOLDED
3721/*
3722 * Allocate page upper directory.
3723 * We've already handled the fast-path in-line.
3724 */
3725int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3726{
3727 pud_t *new = pud_alloc_one(mm, address);
3728 if (!new)
3729 return -ENOMEM;
3730
3731 smp_wmb(); /* See comment in __pte_alloc */
3732
3733 spin_lock(&mm->page_table_lock);
3734 if (pgd_present(*pgd)) /* Another has populated it */
3735 pud_free(mm, new);
3736 else
3737 pgd_populate(mm, pgd, new);
3738 spin_unlock(&mm->page_table_lock);
3739 return 0;
3740}
3741#endif /* __PAGETABLE_PUD_FOLDED */
3742
3743#ifndef __PAGETABLE_PMD_FOLDED
3744/*
3745 * Allocate page middle directory.
3746 * We've already handled the fast-path in-line.
3747 */
3748int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3749{
3750 pmd_t *new = pmd_alloc_one(mm, address);
3751 if (!new)
3752 return -ENOMEM;
3753
3754 smp_wmb(); /* See comment in __pte_alloc */
3755
3756 spin_lock(&mm->page_table_lock);
3757#ifndef __ARCH_HAS_4LEVEL_HACK
3758 if (!pud_present(*pud)) {
3759 mm_inc_nr_pmds(mm);
3760 pud_populate(mm, pud, new);
3761 } else /* Another has populated it */
3762 pmd_free(mm, new);
3763#else
3764 if (!pgd_present(*pud)) {
3765 mm_inc_nr_pmds(mm);
3766 pgd_populate(mm, pud, new);
3767 } else /* Another has populated it */
3768 pmd_free(mm, new);
3769#endif /* __ARCH_HAS_4LEVEL_HACK */
3770 spin_unlock(&mm->page_table_lock);
3771 return 0;
3772}
3773#endif /* __PAGETABLE_PMD_FOLDED */
3774
3775static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3776 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3777{
3778 pgd_t *pgd;
3779 pud_t *pud;
3780 pmd_t *pmd;
3781 pte_t *ptep;
3782
3783 pgd = pgd_offset(mm, address);
3784 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3785 goto out;
3786
3787 pud = pud_offset(pgd, address);
3788 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3789 goto out;
3790
3791 pmd = pmd_offset(pud, address);
3792 VM_BUG_ON(pmd_trans_huge(*pmd));
3793
3794 if (pmd_huge(*pmd)) {
3795 if (!pmdpp)
3796 goto out;
3797
3798 *ptlp = pmd_lock(mm, pmd);
3799 if (pmd_huge(*pmd)) {
3800 *pmdpp = pmd;
3801 return 0;
3802 }
3803 spin_unlock(*ptlp);
3804 }
3805
3806 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3807 goto out;
3808
3809 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3810 if (!ptep)
3811 goto out;
3812 if (!pte_present(*ptep))
3813 goto unlock;
3814 *ptepp = ptep;
3815 return 0;
3816unlock:
3817 pte_unmap_unlock(ptep, *ptlp);
3818out:
3819 return -EINVAL;
3820}
3821
3822static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3823 pte_t **ptepp, spinlock_t **ptlp)
3824{
3825 int res;
3826
3827 /* (void) is needed to make gcc happy */
3828 (void) __cond_lock(*ptlp,
3829 !(res = __follow_pte_pmd(mm, address, ptepp, NULL,
3830 ptlp)));
3831 return res;
3832}
3833
3834int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3835 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3836{
3837 int res;
3838
3839 /* (void) is needed to make gcc happy */
3840 (void) __cond_lock(*ptlp,
3841 !(res = __follow_pte_pmd(mm, address, ptepp, pmdpp,
3842 ptlp)));
3843 return res;
3844}
3845EXPORT_SYMBOL(follow_pte_pmd);
3846
3847/**
3848 * follow_pfn - look up PFN at a user virtual address
3849 * @vma: memory mapping
3850 * @address: user virtual address
3851 * @pfn: location to store found PFN
3852 *
3853 * Only IO mappings and raw PFN mappings are allowed.
3854 *
3855 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3856 */
3857int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3858 unsigned long *pfn)
3859{
3860 int ret = -EINVAL;
3861 spinlock_t *ptl;
3862 pte_t *ptep;
3863
3864 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3865 return ret;
3866
3867 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3868 if (ret)
3869 return ret;
3870 *pfn = pte_pfn(*ptep);
3871 pte_unmap_unlock(ptep, ptl);
3872 return 0;
3873}
3874EXPORT_SYMBOL(follow_pfn);
3875
3876#ifdef CONFIG_HAVE_IOREMAP_PROT
3877int follow_phys(struct vm_area_struct *vma,
3878 unsigned long address, unsigned int flags,
3879 unsigned long *prot, resource_size_t *phys)
3880{
3881 int ret = -EINVAL;
3882 pte_t *ptep, pte;
3883 spinlock_t *ptl;
3884
3885 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3886 goto out;
3887
3888 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3889 goto out;
3890 pte = *ptep;
3891
3892 if ((flags & FOLL_WRITE) && !pte_write(pte))
3893 goto unlock;
3894
3895 *prot = pgprot_val(pte_pgprot(pte));
3896 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3897
3898 ret = 0;
3899unlock:
3900 pte_unmap_unlock(ptep, ptl);
3901out:
3902 return ret;
3903}
3904
3905int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3906 void *buf, int len, int write)
3907{
3908 resource_size_t phys_addr;
3909 unsigned long prot = 0;
3910 void __iomem *maddr;
3911 int offset = addr & (PAGE_SIZE-1);
3912
3913 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3914 return -EINVAL;
3915
3916 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3917 if (write)
3918 memcpy_toio(maddr + offset, buf, len);
3919 else
3920 memcpy_fromio(buf, maddr + offset, len);
3921 iounmap(maddr);
3922
3923 return len;
3924}
3925EXPORT_SYMBOL_GPL(generic_access_phys);
3926#endif
3927
3928/*
3929 * Access another process' address space as given in mm. If non-NULL, use the
3930 * given task for page fault accounting.
3931 */
3932int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3933 unsigned long addr, void *buf, int len, unsigned int gup_flags)
3934{
3935 struct vm_area_struct *vma;
3936 void *old_buf = buf;
3937 int write = gup_flags & FOLL_WRITE;
3938
3939 down_read(&mm->mmap_sem);
3940 /* ignore errors, just check how much was successfully transferred */
3941 while (len) {
3942 int bytes, ret, offset;
3943 void *maddr;
3944 struct page *page = NULL;
3945
3946 ret = get_user_pages_remote(tsk, mm, addr, 1,
3947 gup_flags, &page, &vma, NULL);
3948 if (ret <= 0) {
3949#ifndef CONFIG_HAVE_IOREMAP_PROT
3950 break;
3951#else
3952 /*
3953 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3954 * we can access using slightly different code.
3955 */
3956 vma = find_vma(mm, addr);
3957 if (!vma || vma->vm_start > addr)
3958 break;
3959 if (vma->vm_ops && vma->vm_ops->access)
3960 ret = vma->vm_ops->access(vma, addr, buf,
3961 len, write);
3962 if (ret <= 0)
3963 break;
3964 bytes = ret;
3965#endif
3966 } else {
3967 bytes = len;
3968 offset = addr & (PAGE_SIZE-1);
3969 if (bytes > PAGE_SIZE-offset)
3970 bytes = PAGE_SIZE-offset;
3971
3972 maddr = kmap(page);
3973 if (write) {
3974 copy_to_user_page(vma, page, addr,
3975 maddr + offset, buf, bytes);
3976 set_page_dirty_lock(page);
3977 } else {
3978 copy_from_user_page(vma, page, addr,
3979 buf, maddr + offset, bytes);
3980 }
3981 kunmap(page);
3982 put_page(page);
3983 }
3984 len -= bytes;
3985 buf += bytes;
3986 addr += bytes;
3987 }
3988 up_read(&mm->mmap_sem);
3989
3990 return buf - old_buf;
3991}
3992
3993/**
3994 * access_remote_vm - access another process' address space
3995 * @mm: the mm_struct of the target address space
3996 * @addr: start address to access
3997 * @buf: source or destination buffer
3998 * @len: number of bytes to transfer
3999 * @gup_flags: flags modifying lookup behaviour
4000 *
4001 * The caller must hold a reference on @mm.
4002 */
4003int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4004 void *buf, int len, unsigned int gup_flags)
4005{
4006 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4007}
4008
4009/*
4010 * Access another process' address space.
4011 * Source/target buffer must be kernel space,
4012 * Do not walk the page table directly, use get_user_pages
4013 */
4014int access_process_vm(struct task_struct *tsk, unsigned long addr,
4015 void *buf, int len, unsigned int gup_flags)
4016{
4017 struct mm_struct *mm;
4018 int ret;
4019
4020 mm = get_task_mm(tsk);
4021 if (!mm)
4022 return 0;
4023
4024 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4025
4026 mmput(mm);
4027
4028 return ret;
4029}
4030EXPORT_SYMBOL_GPL(access_process_vm);
4031
4032/*
4033 * Print the name of a VMA.
4034 */
4035void print_vma_addr(char *prefix, unsigned long ip)
4036{
4037 struct mm_struct *mm = current->mm;
4038 struct vm_area_struct *vma;
4039
4040 /*
4041 * Do not print if we are in atomic
4042 * contexts (in exception stacks, etc.):
4043 */
4044 if (preempt_count())
4045 return;
4046
4047 down_read(&mm->mmap_sem);
4048 vma = find_vma(mm, ip);
4049 if (vma && vma->vm_file) {
4050 struct file *f = vma->vm_file;
4051 char *buf = (char *)__get_free_page(GFP_KERNEL);
4052 if (buf) {
4053 char *p;
4054
4055 p = file_path(f, buf, PAGE_SIZE);
4056 if (IS_ERR(p))
4057 p = "?";
4058 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4059 vma->vm_start,
4060 vma->vm_end - vma->vm_start);
4061 free_page((unsigned long)buf);
4062 }
4063 }
4064 up_read(&mm->mmap_sem);
4065}
4066
4067#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4068void __might_fault(const char *file, int line)
4069{
4070 /*
4071 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4072 * holding the mmap_sem, this is safe because kernel memory doesn't
4073 * get paged out, therefore we'll never actually fault, and the
4074 * below annotations will generate false positives.
4075 */
4076 if (segment_eq(get_fs(), KERNEL_DS))
4077 return;
4078 if (pagefault_disabled())
4079 return;
4080 __might_sleep(file, line, 0);
4081#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4082 if (current->mm)
4083 might_lock_read(¤t->mm->mmap_sem);
4084#endif
4085}
4086EXPORT_SYMBOL(__might_fault);
4087#endif
4088
4089#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4090static void clear_gigantic_page(struct page *page,
4091 unsigned long addr,
4092 unsigned int pages_per_huge_page)
4093{
4094 int i;
4095 struct page *p = page;
4096
4097 might_sleep();
4098 for (i = 0; i < pages_per_huge_page;
4099 i++, p = mem_map_next(p, page, i)) {
4100 cond_resched();
4101 clear_user_highpage(p, addr + i * PAGE_SIZE);
4102 }
4103}
4104void clear_huge_page(struct page *page,
4105 unsigned long addr, unsigned int pages_per_huge_page)
4106{
4107 int i;
4108
4109 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4110 clear_gigantic_page(page, addr, pages_per_huge_page);
4111 return;
4112 }
4113
4114 might_sleep();
4115 for (i = 0; i < pages_per_huge_page; i++) {
4116 cond_resched();
4117 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4118 }
4119}
4120
4121static void copy_user_gigantic_page(struct page *dst, struct page *src,
4122 unsigned long addr,
4123 struct vm_area_struct *vma,
4124 unsigned int pages_per_huge_page)
4125{
4126 int i;
4127 struct page *dst_base = dst;
4128 struct page *src_base = src;
4129
4130 for (i = 0; i < pages_per_huge_page; ) {
4131 cond_resched();
4132 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4133
4134 i++;
4135 dst = mem_map_next(dst, dst_base, i);
4136 src = mem_map_next(src, src_base, i);
4137 }
4138}
4139
4140void copy_user_huge_page(struct page *dst, struct page *src,
4141 unsigned long addr, struct vm_area_struct *vma,
4142 unsigned int pages_per_huge_page)
4143{
4144 int i;
4145
4146 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4147 copy_user_gigantic_page(dst, src, addr, vma,
4148 pages_per_huge_page);
4149 return;
4150 }
4151
4152 might_sleep();
4153 for (i = 0; i < pages_per_huge_page; i++) {
4154 cond_resched();
4155 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4156 }
4157}
4158#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4159
4160#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4161
4162static struct kmem_cache *page_ptl_cachep;
4163
4164void __init ptlock_cache_init(void)
4165{
4166 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4167 SLAB_PANIC, NULL);
4168}
4169
4170bool ptlock_alloc(struct page *page)
4171{
4172 spinlock_t *ptl;
4173
4174 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4175 if (!ptl)
4176 return false;
4177 page->ptl = ptl;
4178 return true;
4179}
4180
4181void ptlock_free(struct page *page)
4182{
4183 kmem_cache_free(page_ptl_cachep, page->ptl);
4184}
4185#endif