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// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/memory.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 */
7
8/*
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
11 */
12
13/*
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
16 *
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
19 * far as I could see.
20 *
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
22 */
23
24/*
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
30 */
31
32/*
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 *
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
38 *
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
40 */
41
42#include <linux/kernel_stat.h>
43#include <linux/mm.h>
44#include <linux/sched/mm.h>
45#include <linux/sched/coredump.h>
46#include <linux/sched/numa_balancing.h>
47#include <linux/sched/task.h>
48#include <linux/hugetlb.h>
49#include <linux/mman.h>
50#include <linux/swap.h>
51#include <linux/highmem.h>
52#include <linux/pagemap.h>
53#include <linux/memremap.h>
54#include <linux/ksm.h>
55#include <linux/rmap.h>
56#include <linux/export.h>
57#include <linux/delayacct.h>
58#include <linux/init.h>
59#include <linux/pfn_t.h>
60#include <linux/writeback.h>
61#include <linux/memcontrol.h>
62#include <linux/mmu_notifier.h>
63#include <linux/swapops.h>
64#include <linux/elf.h>
65#include <linux/gfp.h>
66#include <linux/migrate.h>
67#include <linux/string.h>
68#include <linux/debugfs.h>
69#include <linux/userfaultfd_k.h>
70#include <linux/dax.h>
71#include <linux/oom.h>
72#include <linux/numa.h>
73#include <linux/perf_event.h>
74#include <linux/ptrace.h>
75#include <linux/vmalloc.h>
76
77#include <trace/events/kmem.h>
78
79#include <asm/io.h>
80#include <asm/mmu_context.h>
81#include <asm/pgalloc.h>
82#include <linux/uaccess.h>
83#include <asm/tlb.h>
84#include <asm/tlbflush.h>
85
86#include "pgalloc-track.h"
87#include "internal.h"
88
89#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
90#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
91#endif
92
93#ifndef CONFIG_NUMA
94unsigned long max_mapnr;
95EXPORT_SYMBOL(max_mapnr);
96
97struct page *mem_map;
98EXPORT_SYMBOL(mem_map);
99#endif
100
101/*
102 * A number of key systems in x86 including ioremap() rely on the assumption
103 * that high_memory defines the upper bound on direct map memory, then end
104 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
105 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
106 * and ZONE_HIGHMEM.
107 */
108void *high_memory;
109EXPORT_SYMBOL(high_memory);
110
111/*
112 * Randomize the address space (stacks, mmaps, brk, etc.).
113 *
114 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
115 * as ancient (libc5 based) binaries can segfault. )
116 */
117int randomize_va_space __read_mostly =
118#ifdef CONFIG_COMPAT_BRK
119 1;
120#else
121 2;
122#endif
123
124#ifndef arch_faults_on_old_pte
125static inline bool arch_faults_on_old_pte(void)
126{
127 /*
128 * Those arches which don't have hw access flag feature need to
129 * implement their own helper. By default, "true" means pagefault
130 * will be hit on old pte.
131 */
132 return true;
133}
134#endif
135
136#ifndef arch_wants_old_prefaulted_pte
137static inline bool arch_wants_old_prefaulted_pte(void)
138{
139 /*
140 * Transitioning a PTE from 'old' to 'young' can be expensive on
141 * some architectures, even if it's performed in hardware. By
142 * default, "false" means prefaulted entries will be 'young'.
143 */
144 return false;
145}
146#endif
147
148static int __init disable_randmaps(char *s)
149{
150 randomize_va_space = 0;
151 return 1;
152}
153__setup("norandmaps", disable_randmaps);
154
155unsigned long zero_pfn __read_mostly;
156EXPORT_SYMBOL(zero_pfn);
157
158unsigned long highest_memmap_pfn __read_mostly;
159
160/*
161 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
162 */
163static int __init init_zero_pfn(void)
164{
165 zero_pfn = page_to_pfn(ZERO_PAGE(0));
166 return 0;
167}
168early_initcall(init_zero_pfn);
169
170void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
171{
172 trace_rss_stat(mm, member, count);
173}
174
175#if defined(SPLIT_RSS_COUNTING)
176
177void sync_mm_rss(struct mm_struct *mm)
178{
179 int i;
180
181 for (i = 0; i < NR_MM_COUNTERS; i++) {
182 if (current->rss_stat.count[i]) {
183 add_mm_counter(mm, i, current->rss_stat.count[i]);
184 current->rss_stat.count[i] = 0;
185 }
186 }
187 current->rss_stat.events = 0;
188}
189
190static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
191{
192 struct task_struct *task = current;
193
194 if (likely(task->mm == mm))
195 task->rss_stat.count[member] += val;
196 else
197 add_mm_counter(mm, member, val);
198}
199#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
200#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
201
202/* sync counter once per 64 page faults */
203#define TASK_RSS_EVENTS_THRESH (64)
204static void check_sync_rss_stat(struct task_struct *task)
205{
206 if (unlikely(task != current))
207 return;
208 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
209 sync_mm_rss(task->mm);
210}
211#else /* SPLIT_RSS_COUNTING */
212
213#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
214#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
215
216static void check_sync_rss_stat(struct task_struct *task)
217{
218}
219
220#endif /* SPLIT_RSS_COUNTING */
221
222/*
223 * Note: this doesn't free the actual pages themselves. That
224 * has been handled earlier when unmapping all the memory regions.
225 */
226static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
227 unsigned long addr)
228{
229 pgtable_t token = pmd_pgtable(*pmd);
230 pmd_clear(pmd);
231 pte_free_tlb(tlb, token, addr);
232 mm_dec_nr_ptes(tlb->mm);
233}
234
235static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
236 unsigned long addr, unsigned long end,
237 unsigned long floor, unsigned long ceiling)
238{
239 pmd_t *pmd;
240 unsigned long next;
241 unsigned long start;
242
243 start = addr;
244 pmd = pmd_offset(pud, addr);
245 do {
246 next = pmd_addr_end(addr, end);
247 if (pmd_none_or_clear_bad(pmd))
248 continue;
249 free_pte_range(tlb, pmd, addr);
250 } while (pmd++, addr = next, addr != end);
251
252 start &= PUD_MASK;
253 if (start < floor)
254 return;
255 if (ceiling) {
256 ceiling &= PUD_MASK;
257 if (!ceiling)
258 return;
259 }
260 if (end - 1 > ceiling - 1)
261 return;
262
263 pmd = pmd_offset(pud, start);
264 pud_clear(pud);
265 pmd_free_tlb(tlb, pmd, start);
266 mm_dec_nr_pmds(tlb->mm);
267}
268
269static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
270 unsigned long addr, unsigned long end,
271 unsigned long floor, unsigned long ceiling)
272{
273 pud_t *pud;
274 unsigned long next;
275 unsigned long start;
276
277 start = addr;
278 pud = pud_offset(p4d, addr);
279 do {
280 next = pud_addr_end(addr, end);
281 if (pud_none_or_clear_bad(pud))
282 continue;
283 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
284 } while (pud++, addr = next, addr != end);
285
286 start &= P4D_MASK;
287 if (start < floor)
288 return;
289 if (ceiling) {
290 ceiling &= P4D_MASK;
291 if (!ceiling)
292 return;
293 }
294 if (end - 1 > ceiling - 1)
295 return;
296
297 pud = pud_offset(p4d, start);
298 p4d_clear(p4d);
299 pud_free_tlb(tlb, pud, start);
300 mm_dec_nr_puds(tlb->mm);
301}
302
303static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
304 unsigned long addr, unsigned long end,
305 unsigned long floor, unsigned long ceiling)
306{
307 p4d_t *p4d;
308 unsigned long next;
309 unsigned long start;
310
311 start = addr;
312 p4d = p4d_offset(pgd, addr);
313 do {
314 next = p4d_addr_end(addr, end);
315 if (p4d_none_or_clear_bad(p4d))
316 continue;
317 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
318 } while (p4d++, addr = next, addr != end);
319
320 start &= PGDIR_MASK;
321 if (start < floor)
322 return;
323 if (ceiling) {
324 ceiling &= PGDIR_MASK;
325 if (!ceiling)
326 return;
327 }
328 if (end - 1 > ceiling - 1)
329 return;
330
331 p4d = p4d_offset(pgd, start);
332 pgd_clear(pgd);
333 p4d_free_tlb(tlb, p4d, start);
334}
335
336/*
337 * This function frees user-level page tables of a process.
338 */
339void free_pgd_range(struct mmu_gather *tlb,
340 unsigned long addr, unsigned long end,
341 unsigned long floor, unsigned long ceiling)
342{
343 pgd_t *pgd;
344 unsigned long next;
345
346 /*
347 * The next few lines have given us lots of grief...
348 *
349 * Why are we testing PMD* at this top level? Because often
350 * there will be no work to do at all, and we'd prefer not to
351 * go all the way down to the bottom just to discover that.
352 *
353 * Why all these "- 1"s? Because 0 represents both the bottom
354 * of the address space and the top of it (using -1 for the
355 * top wouldn't help much: the masks would do the wrong thing).
356 * The rule is that addr 0 and floor 0 refer to the bottom of
357 * the address space, but end 0 and ceiling 0 refer to the top
358 * Comparisons need to use "end - 1" and "ceiling - 1" (though
359 * that end 0 case should be mythical).
360 *
361 * Wherever addr is brought up or ceiling brought down, we must
362 * be careful to reject "the opposite 0" before it confuses the
363 * subsequent tests. But what about where end is brought down
364 * by PMD_SIZE below? no, end can't go down to 0 there.
365 *
366 * Whereas we round start (addr) and ceiling down, by different
367 * masks at different levels, in order to test whether a table
368 * now has no other vmas using it, so can be freed, we don't
369 * bother to round floor or end up - the tests don't need that.
370 */
371
372 addr &= PMD_MASK;
373 if (addr < floor) {
374 addr += PMD_SIZE;
375 if (!addr)
376 return;
377 }
378 if (ceiling) {
379 ceiling &= PMD_MASK;
380 if (!ceiling)
381 return;
382 }
383 if (end - 1 > ceiling - 1)
384 end -= PMD_SIZE;
385 if (addr > end - 1)
386 return;
387 /*
388 * We add page table cache pages with PAGE_SIZE,
389 * (see pte_free_tlb()), flush the tlb if we need
390 */
391 tlb_change_page_size(tlb, PAGE_SIZE);
392 pgd = pgd_offset(tlb->mm, addr);
393 do {
394 next = pgd_addr_end(addr, end);
395 if (pgd_none_or_clear_bad(pgd))
396 continue;
397 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
398 } while (pgd++, addr = next, addr != end);
399}
400
401void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
402 unsigned long floor, unsigned long ceiling)
403{
404 while (vma) {
405 struct vm_area_struct *next = vma->vm_next;
406 unsigned long addr = vma->vm_start;
407
408 /*
409 * Hide vma from rmap and truncate_pagecache before freeing
410 * pgtables
411 */
412 unlink_anon_vmas(vma);
413 unlink_file_vma(vma);
414
415 if (is_vm_hugetlb_page(vma)) {
416 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
417 floor, next ? next->vm_start : ceiling);
418 } else {
419 /*
420 * Optimization: gather nearby vmas into one call down
421 */
422 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
423 && !is_vm_hugetlb_page(next)) {
424 vma = next;
425 next = vma->vm_next;
426 unlink_anon_vmas(vma);
427 unlink_file_vma(vma);
428 }
429 free_pgd_range(tlb, addr, vma->vm_end,
430 floor, next ? next->vm_start : ceiling);
431 }
432 vma = next;
433 }
434}
435
436int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
437{
438 spinlock_t *ptl;
439 pgtable_t new = pte_alloc_one(mm);
440 if (!new)
441 return -ENOMEM;
442
443 /*
444 * Ensure all pte setup (eg. pte page lock and page clearing) are
445 * visible before the pte is made visible to other CPUs by being
446 * put into page tables.
447 *
448 * The other side of the story is the pointer chasing in the page
449 * table walking code (when walking the page table without locking;
450 * ie. most of the time). Fortunately, these data accesses consist
451 * of a chain of data-dependent loads, meaning most CPUs (alpha
452 * being the notable exception) will already guarantee loads are
453 * seen in-order. See the alpha page table accessors for the
454 * smp_rmb() barriers in page table walking code.
455 */
456 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
457
458 ptl = pmd_lock(mm, pmd);
459 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
460 mm_inc_nr_ptes(mm);
461 pmd_populate(mm, pmd, new);
462 new = NULL;
463 }
464 spin_unlock(ptl);
465 if (new)
466 pte_free(mm, new);
467 return 0;
468}
469
470int __pte_alloc_kernel(pmd_t *pmd)
471{
472 pte_t *new = pte_alloc_one_kernel(&init_mm);
473 if (!new)
474 return -ENOMEM;
475
476 smp_wmb(); /* See comment in __pte_alloc */
477
478 spin_lock(&init_mm.page_table_lock);
479 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
480 pmd_populate_kernel(&init_mm, pmd, new);
481 new = NULL;
482 }
483 spin_unlock(&init_mm.page_table_lock);
484 if (new)
485 pte_free_kernel(&init_mm, new);
486 return 0;
487}
488
489static inline void init_rss_vec(int *rss)
490{
491 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
492}
493
494static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
495{
496 int i;
497
498 if (current->mm == mm)
499 sync_mm_rss(mm);
500 for (i = 0; i < NR_MM_COUNTERS; i++)
501 if (rss[i])
502 add_mm_counter(mm, i, rss[i]);
503}
504
505/*
506 * This function is called to print an error when a bad pte
507 * is found. For example, we might have a PFN-mapped pte in
508 * a region that doesn't allow it.
509 *
510 * The calling function must still handle the error.
511 */
512static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
513 pte_t pte, struct page *page)
514{
515 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
516 p4d_t *p4d = p4d_offset(pgd, addr);
517 pud_t *pud = pud_offset(p4d, addr);
518 pmd_t *pmd = pmd_offset(pud, addr);
519 struct address_space *mapping;
520 pgoff_t index;
521 static unsigned long resume;
522 static unsigned long nr_shown;
523 static unsigned long nr_unshown;
524
525 /*
526 * Allow a burst of 60 reports, then keep quiet for that minute;
527 * or allow a steady drip of one report per second.
528 */
529 if (nr_shown == 60) {
530 if (time_before(jiffies, resume)) {
531 nr_unshown++;
532 return;
533 }
534 if (nr_unshown) {
535 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
536 nr_unshown);
537 nr_unshown = 0;
538 }
539 nr_shown = 0;
540 }
541 if (nr_shown++ == 0)
542 resume = jiffies + 60 * HZ;
543
544 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
545 index = linear_page_index(vma, addr);
546
547 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
548 current->comm,
549 (long long)pte_val(pte), (long long)pmd_val(*pmd));
550 if (page)
551 dump_page(page, "bad pte");
552 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
553 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
554 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
555 vma->vm_file,
556 vma->vm_ops ? vma->vm_ops->fault : NULL,
557 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
558 mapping ? mapping->a_ops->readpage : NULL);
559 dump_stack();
560 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
561}
562
563/*
564 * vm_normal_page -- This function gets the "struct page" associated with a pte.
565 *
566 * "Special" mappings do not wish to be associated with a "struct page" (either
567 * it doesn't exist, or it exists but they don't want to touch it). In this
568 * case, NULL is returned here. "Normal" mappings do have a struct page.
569 *
570 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
571 * pte bit, in which case this function is trivial. Secondly, an architecture
572 * may not have a spare pte bit, which requires a more complicated scheme,
573 * described below.
574 *
575 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
576 * special mapping (even if there are underlying and valid "struct pages").
577 * COWed pages of a VM_PFNMAP are always normal.
578 *
579 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
580 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
581 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
582 * mapping will always honor the rule
583 *
584 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
585 *
586 * And for normal mappings this is false.
587 *
588 * This restricts such mappings to be a linear translation from virtual address
589 * to pfn. To get around this restriction, we allow arbitrary mappings so long
590 * as the vma is not a COW mapping; in that case, we know that all ptes are
591 * special (because none can have been COWed).
592 *
593 *
594 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
595 *
596 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
597 * page" backing, however the difference is that _all_ pages with a struct
598 * page (that is, those where pfn_valid is true) are refcounted and considered
599 * normal pages by the VM. The disadvantage is that pages are refcounted
600 * (which can be slower and simply not an option for some PFNMAP users). The
601 * advantage is that we don't have to follow the strict linearity rule of
602 * PFNMAP mappings in order to support COWable mappings.
603 *
604 */
605struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
606 pte_t pte)
607{
608 unsigned long pfn = pte_pfn(pte);
609
610 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
611 if (likely(!pte_special(pte)))
612 goto check_pfn;
613 if (vma->vm_ops && vma->vm_ops->find_special_page)
614 return vma->vm_ops->find_special_page(vma, addr);
615 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
616 return NULL;
617 if (is_zero_pfn(pfn))
618 return NULL;
619 if (pte_devmap(pte))
620 return NULL;
621
622 print_bad_pte(vma, addr, pte, NULL);
623 return NULL;
624 }
625
626 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
627
628 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
629 if (vma->vm_flags & VM_MIXEDMAP) {
630 if (!pfn_valid(pfn))
631 return NULL;
632 goto out;
633 } else {
634 unsigned long off;
635 off = (addr - vma->vm_start) >> PAGE_SHIFT;
636 if (pfn == vma->vm_pgoff + off)
637 return NULL;
638 if (!is_cow_mapping(vma->vm_flags))
639 return NULL;
640 }
641 }
642
643 if (is_zero_pfn(pfn))
644 return NULL;
645
646check_pfn:
647 if (unlikely(pfn > highest_memmap_pfn)) {
648 print_bad_pte(vma, addr, pte, NULL);
649 return NULL;
650 }
651
652 /*
653 * NOTE! We still have PageReserved() pages in the page tables.
654 * eg. VDSO mappings can cause them to exist.
655 */
656out:
657 return pfn_to_page(pfn);
658}
659
660#ifdef CONFIG_TRANSPARENT_HUGEPAGE
661struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
662 pmd_t pmd)
663{
664 unsigned long pfn = pmd_pfn(pmd);
665
666 /*
667 * There is no pmd_special() but there may be special pmds, e.g.
668 * in a direct-access (dax) mapping, so let's just replicate the
669 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
670 */
671 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
672 if (vma->vm_flags & VM_MIXEDMAP) {
673 if (!pfn_valid(pfn))
674 return NULL;
675 goto out;
676 } else {
677 unsigned long off;
678 off = (addr - vma->vm_start) >> PAGE_SHIFT;
679 if (pfn == vma->vm_pgoff + off)
680 return NULL;
681 if (!is_cow_mapping(vma->vm_flags))
682 return NULL;
683 }
684 }
685
686 if (pmd_devmap(pmd))
687 return NULL;
688 if (is_huge_zero_pmd(pmd))
689 return NULL;
690 if (unlikely(pfn > highest_memmap_pfn))
691 return NULL;
692
693 /*
694 * NOTE! We still have PageReserved() pages in the page tables.
695 * eg. VDSO mappings can cause them to exist.
696 */
697out:
698 return pfn_to_page(pfn);
699}
700#endif
701
702static void restore_exclusive_pte(struct vm_area_struct *vma,
703 struct page *page, unsigned long address,
704 pte_t *ptep)
705{
706 pte_t pte;
707 swp_entry_t entry;
708
709 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
710 if (pte_swp_soft_dirty(*ptep))
711 pte = pte_mksoft_dirty(pte);
712
713 entry = pte_to_swp_entry(*ptep);
714 if (pte_swp_uffd_wp(*ptep))
715 pte = pte_mkuffd_wp(pte);
716 else if (is_writable_device_exclusive_entry(entry))
717 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
718
719 set_pte_at(vma->vm_mm, address, ptep, pte);
720
721 /*
722 * No need to take a page reference as one was already
723 * created when the swap entry was made.
724 */
725 if (PageAnon(page))
726 page_add_anon_rmap(page, vma, address, false);
727 else
728 /*
729 * Currently device exclusive access only supports anonymous
730 * memory so the entry shouldn't point to a filebacked page.
731 */
732 WARN_ON_ONCE(!PageAnon(page));
733
734 if (vma->vm_flags & VM_LOCKED)
735 mlock_vma_page(page);
736
737 /*
738 * No need to invalidate - it was non-present before. However
739 * secondary CPUs may have mappings that need invalidating.
740 */
741 update_mmu_cache(vma, address, ptep);
742}
743
744/*
745 * Tries to restore an exclusive pte if the page lock can be acquired without
746 * sleeping.
747 */
748static int
749try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
750 unsigned long addr)
751{
752 swp_entry_t entry = pte_to_swp_entry(*src_pte);
753 struct page *page = pfn_swap_entry_to_page(entry);
754
755 if (trylock_page(page)) {
756 restore_exclusive_pte(vma, page, addr, src_pte);
757 unlock_page(page);
758 return 0;
759 }
760
761 return -EBUSY;
762}
763
764/*
765 * copy one vm_area from one task to the other. Assumes the page tables
766 * already present in the new task to be cleared in the whole range
767 * covered by this vma.
768 */
769
770static unsigned long
771copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
772 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
773 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
774{
775 unsigned long vm_flags = dst_vma->vm_flags;
776 pte_t pte = *src_pte;
777 struct page *page;
778 swp_entry_t entry = pte_to_swp_entry(pte);
779
780 if (likely(!non_swap_entry(entry))) {
781 if (swap_duplicate(entry) < 0)
782 return -EIO;
783
784 /* make sure dst_mm is on swapoff's mmlist. */
785 if (unlikely(list_empty(&dst_mm->mmlist))) {
786 spin_lock(&mmlist_lock);
787 if (list_empty(&dst_mm->mmlist))
788 list_add(&dst_mm->mmlist,
789 &src_mm->mmlist);
790 spin_unlock(&mmlist_lock);
791 }
792 rss[MM_SWAPENTS]++;
793 } else if (is_migration_entry(entry)) {
794 page = pfn_swap_entry_to_page(entry);
795
796 rss[mm_counter(page)]++;
797
798 if (is_writable_migration_entry(entry) &&
799 is_cow_mapping(vm_flags)) {
800 /*
801 * COW mappings require pages in both
802 * parent and child to be set to read.
803 */
804 entry = make_readable_migration_entry(
805 swp_offset(entry));
806 pte = swp_entry_to_pte(entry);
807 if (pte_swp_soft_dirty(*src_pte))
808 pte = pte_swp_mksoft_dirty(pte);
809 if (pte_swp_uffd_wp(*src_pte))
810 pte = pte_swp_mkuffd_wp(pte);
811 set_pte_at(src_mm, addr, src_pte, pte);
812 }
813 } else if (is_device_private_entry(entry)) {
814 page = pfn_swap_entry_to_page(entry);
815
816 /*
817 * Update rss count even for unaddressable pages, as
818 * they should treated just like normal pages in this
819 * respect.
820 *
821 * We will likely want to have some new rss counters
822 * for unaddressable pages, at some point. But for now
823 * keep things as they are.
824 */
825 get_page(page);
826 rss[mm_counter(page)]++;
827 page_dup_rmap(page, false);
828
829 /*
830 * We do not preserve soft-dirty information, because so
831 * far, checkpoint/restore is the only feature that
832 * requires that. And checkpoint/restore does not work
833 * when a device driver is involved (you cannot easily
834 * save and restore device driver state).
835 */
836 if (is_writable_device_private_entry(entry) &&
837 is_cow_mapping(vm_flags)) {
838 entry = make_readable_device_private_entry(
839 swp_offset(entry));
840 pte = swp_entry_to_pte(entry);
841 if (pte_swp_uffd_wp(*src_pte))
842 pte = pte_swp_mkuffd_wp(pte);
843 set_pte_at(src_mm, addr, src_pte, pte);
844 }
845 } else if (is_device_exclusive_entry(entry)) {
846 /*
847 * Make device exclusive entries present by restoring the
848 * original entry then copying as for a present pte. Device
849 * exclusive entries currently only support private writable
850 * (ie. COW) mappings.
851 */
852 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
853 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
854 return -EBUSY;
855 return -ENOENT;
856 }
857 if (!userfaultfd_wp(dst_vma))
858 pte = pte_swp_clear_uffd_wp(pte);
859 set_pte_at(dst_mm, addr, dst_pte, pte);
860 return 0;
861}
862
863/*
864 * Copy a present and normal page if necessary.
865 *
866 * NOTE! The usual case is that this doesn't need to do
867 * anything, and can just return a positive value. That
868 * will let the caller know that it can just increase
869 * the page refcount and re-use the pte the traditional
870 * way.
871 *
872 * But _if_ we need to copy it because it needs to be
873 * pinned in the parent (and the child should get its own
874 * copy rather than just a reference to the same page),
875 * we'll do that here and return zero to let the caller
876 * know we're done.
877 *
878 * And if we need a pre-allocated page but don't yet have
879 * one, return a negative error to let the preallocation
880 * code know so that it can do so outside the page table
881 * lock.
882 */
883static inline int
884copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
885 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
886 struct page **prealloc, pte_t pte, struct page *page)
887{
888 struct page *new_page;
889
890 /*
891 * What we want to do is to check whether this page may
892 * have been pinned by the parent process. If so,
893 * instead of wrprotect the pte on both sides, we copy
894 * the page immediately so that we'll always guarantee
895 * the pinned page won't be randomly replaced in the
896 * future.
897 *
898 * The page pinning checks are just "has this mm ever
899 * seen pinning", along with the (inexact) check of
900 * the page count. That might give false positives for
901 * for pinning, but it will work correctly.
902 */
903 if (likely(!page_needs_cow_for_dma(src_vma, page)))
904 return 1;
905
906 new_page = *prealloc;
907 if (!new_page)
908 return -EAGAIN;
909
910 /*
911 * We have a prealloc page, all good! Take it
912 * over and copy the page & arm it.
913 */
914 *prealloc = NULL;
915 copy_user_highpage(new_page, page, addr, src_vma);
916 __SetPageUptodate(new_page);
917 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
918 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
919 rss[mm_counter(new_page)]++;
920
921 /* All done, just insert the new page copy in the child */
922 pte = mk_pte(new_page, dst_vma->vm_page_prot);
923 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
924 if (userfaultfd_pte_wp(dst_vma, *src_pte))
925 /* Uffd-wp needs to be delivered to dest pte as well */
926 pte = pte_wrprotect(pte_mkuffd_wp(pte));
927 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
928 return 0;
929}
930
931/*
932 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
933 * is required to copy this pte.
934 */
935static inline int
936copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
937 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
938 struct page **prealloc)
939{
940 struct mm_struct *src_mm = src_vma->vm_mm;
941 unsigned long vm_flags = src_vma->vm_flags;
942 pte_t pte = *src_pte;
943 struct page *page;
944
945 page = vm_normal_page(src_vma, addr, pte);
946 if (page) {
947 int retval;
948
949 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
950 addr, rss, prealloc, pte, page);
951 if (retval <= 0)
952 return retval;
953
954 get_page(page);
955 page_dup_rmap(page, false);
956 rss[mm_counter(page)]++;
957 }
958
959 /*
960 * If it's a COW mapping, write protect it both
961 * in the parent and the child
962 */
963 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
964 ptep_set_wrprotect(src_mm, addr, src_pte);
965 pte = pte_wrprotect(pte);
966 }
967
968 /*
969 * If it's a shared mapping, mark it clean in
970 * the child
971 */
972 if (vm_flags & VM_SHARED)
973 pte = pte_mkclean(pte);
974 pte = pte_mkold(pte);
975
976 if (!userfaultfd_wp(dst_vma))
977 pte = pte_clear_uffd_wp(pte);
978
979 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
980 return 0;
981}
982
983static inline struct page *
984page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
985 unsigned long addr)
986{
987 struct page *new_page;
988
989 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
990 if (!new_page)
991 return NULL;
992
993 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
994 put_page(new_page);
995 return NULL;
996 }
997 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
998
999 return new_page;
1000}
1001
1002static int
1003copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1004 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1005 unsigned long end)
1006{
1007 struct mm_struct *dst_mm = dst_vma->vm_mm;
1008 struct mm_struct *src_mm = src_vma->vm_mm;
1009 pte_t *orig_src_pte, *orig_dst_pte;
1010 pte_t *src_pte, *dst_pte;
1011 spinlock_t *src_ptl, *dst_ptl;
1012 int progress, ret = 0;
1013 int rss[NR_MM_COUNTERS];
1014 swp_entry_t entry = (swp_entry_t){0};
1015 struct page *prealloc = NULL;
1016
1017again:
1018 progress = 0;
1019 init_rss_vec(rss);
1020
1021 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1022 if (!dst_pte) {
1023 ret = -ENOMEM;
1024 goto out;
1025 }
1026 src_pte = pte_offset_map(src_pmd, addr);
1027 src_ptl = pte_lockptr(src_mm, src_pmd);
1028 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1029 orig_src_pte = src_pte;
1030 orig_dst_pte = dst_pte;
1031 arch_enter_lazy_mmu_mode();
1032
1033 do {
1034 /*
1035 * We are holding two locks at this point - either of them
1036 * could generate latencies in another task on another CPU.
1037 */
1038 if (progress >= 32) {
1039 progress = 0;
1040 if (need_resched() ||
1041 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1042 break;
1043 }
1044 if (pte_none(*src_pte)) {
1045 progress++;
1046 continue;
1047 }
1048 if (unlikely(!pte_present(*src_pte))) {
1049 ret = copy_nonpresent_pte(dst_mm, src_mm,
1050 dst_pte, src_pte,
1051 dst_vma, src_vma,
1052 addr, rss);
1053 if (ret == -EIO) {
1054 entry = pte_to_swp_entry(*src_pte);
1055 break;
1056 } else if (ret == -EBUSY) {
1057 break;
1058 } else if (!ret) {
1059 progress += 8;
1060 continue;
1061 }
1062
1063 /*
1064 * Device exclusive entry restored, continue by copying
1065 * the now present pte.
1066 */
1067 WARN_ON_ONCE(ret != -ENOENT);
1068 }
1069 /* copy_present_pte() will clear `*prealloc' if consumed */
1070 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1071 addr, rss, &prealloc);
1072 /*
1073 * If we need a pre-allocated page for this pte, drop the
1074 * locks, allocate, and try again.
1075 */
1076 if (unlikely(ret == -EAGAIN))
1077 break;
1078 if (unlikely(prealloc)) {
1079 /*
1080 * pre-alloc page cannot be reused by next time so as
1081 * to strictly follow mempolicy (e.g., alloc_page_vma()
1082 * will allocate page according to address). This
1083 * could only happen if one pinned pte changed.
1084 */
1085 put_page(prealloc);
1086 prealloc = NULL;
1087 }
1088 progress += 8;
1089 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1090
1091 arch_leave_lazy_mmu_mode();
1092 spin_unlock(src_ptl);
1093 pte_unmap(orig_src_pte);
1094 add_mm_rss_vec(dst_mm, rss);
1095 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1096 cond_resched();
1097
1098 if (ret == -EIO) {
1099 VM_WARN_ON_ONCE(!entry.val);
1100 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1101 ret = -ENOMEM;
1102 goto out;
1103 }
1104 entry.val = 0;
1105 } else if (ret == -EBUSY) {
1106 goto out;
1107 } else if (ret == -EAGAIN) {
1108 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1109 if (!prealloc)
1110 return -ENOMEM;
1111 } else if (ret) {
1112 VM_WARN_ON_ONCE(1);
1113 }
1114
1115 /* We've captured and resolved the error. Reset, try again. */
1116 ret = 0;
1117
1118 if (addr != end)
1119 goto again;
1120out:
1121 if (unlikely(prealloc))
1122 put_page(prealloc);
1123 return ret;
1124}
1125
1126static inline int
1127copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1128 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1129 unsigned long end)
1130{
1131 struct mm_struct *dst_mm = dst_vma->vm_mm;
1132 struct mm_struct *src_mm = src_vma->vm_mm;
1133 pmd_t *src_pmd, *dst_pmd;
1134 unsigned long next;
1135
1136 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1137 if (!dst_pmd)
1138 return -ENOMEM;
1139 src_pmd = pmd_offset(src_pud, addr);
1140 do {
1141 next = pmd_addr_end(addr, end);
1142 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1143 || pmd_devmap(*src_pmd)) {
1144 int err;
1145 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1146 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1147 addr, dst_vma, src_vma);
1148 if (err == -ENOMEM)
1149 return -ENOMEM;
1150 if (!err)
1151 continue;
1152 /* fall through */
1153 }
1154 if (pmd_none_or_clear_bad(src_pmd))
1155 continue;
1156 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1157 addr, next))
1158 return -ENOMEM;
1159 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1160 return 0;
1161}
1162
1163static inline int
1164copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1165 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1166 unsigned long end)
1167{
1168 struct mm_struct *dst_mm = dst_vma->vm_mm;
1169 struct mm_struct *src_mm = src_vma->vm_mm;
1170 pud_t *src_pud, *dst_pud;
1171 unsigned long next;
1172
1173 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1174 if (!dst_pud)
1175 return -ENOMEM;
1176 src_pud = pud_offset(src_p4d, addr);
1177 do {
1178 next = pud_addr_end(addr, end);
1179 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1180 int err;
1181
1182 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1183 err = copy_huge_pud(dst_mm, src_mm,
1184 dst_pud, src_pud, addr, src_vma);
1185 if (err == -ENOMEM)
1186 return -ENOMEM;
1187 if (!err)
1188 continue;
1189 /* fall through */
1190 }
1191 if (pud_none_or_clear_bad(src_pud))
1192 continue;
1193 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1194 addr, next))
1195 return -ENOMEM;
1196 } while (dst_pud++, src_pud++, addr = next, addr != end);
1197 return 0;
1198}
1199
1200static inline int
1201copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1202 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1203 unsigned long end)
1204{
1205 struct mm_struct *dst_mm = dst_vma->vm_mm;
1206 p4d_t *src_p4d, *dst_p4d;
1207 unsigned long next;
1208
1209 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1210 if (!dst_p4d)
1211 return -ENOMEM;
1212 src_p4d = p4d_offset(src_pgd, addr);
1213 do {
1214 next = p4d_addr_end(addr, end);
1215 if (p4d_none_or_clear_bad(src_p4d))
1216 continue;
1217 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1218 addr, next))
1219 return -ENOMEM;
1220 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1221 return 0;
1222}
1223
1224int
1225copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1226{
1227 pgd_t *src_pgd, *dst_pgd;
1228 unsigned long next;
1229 unsigned long addr = src_vma->vm_start;
1230 unsigned long end = src_vma->vm_end;
1231 struct mm_struct *dst_mm = dst_vma->vm_mm;
1232 struct mm_struct *src_mm = src_vma->vm_mm;
1233 struct mmu_notifier_range range;
1234 bool is_cow;
1235 int ret;
1236
1237 /*
1238 * Don't copy ptes where a page fault will fill them correctly.
1239 * Fork becomes much lighter when there are big shared or private
1240 * readonly mappings. The tradeoff is that copy_page_range is more
1241 * efficient than faulting.
1242 */
1243 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1244 !src_vma->anon_vma)
1245 return 0;
1246
1247 if (is_vm_hugetlb_page(src_vma))
1248 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1249
1250 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1251 /*
1252 * We do not free on error cases below as remove_vma
1253 * gets called on error from higher level routine
1254 */
1255 ret = track_pfn_copy(src_vma);
1256 if (ret)
1257 return ret;
1258 }
1259
1260 /*
1261 * We need to invalidate the secondary MMU mappings only when
1262 * there could be a permission downgrade on the ptes of the
1263 * parent mm. And a permission downgrade will only happen if
1264 * is_cow_mapping() returns true.
1265 */
1266 is_cow = is_cow_mapping(src_vma->vm_flags);
1267
1268 if (is_cow) {
1269 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1270 0, src_vma, src_mm, addr, end);
1271 mmu_notifier_invalidate_range_start(&range);
1272 /*
1273 * Disabling preemption is not needed for the write side, as
1274 * the read side doesn't spin, but goes to the mmap_lock.
1275 *
1276 * Use the raw variant of the seqcount_t write API to avoid
1277 * lockdep complaining about preemptibility.
1278 */
1279 mmap_assert_write_locked(src_mm);
1280 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1281 }
1282
1283 ret = 0;
1284 dst_pgd = pgd_offset(dst_mm, addr);
1285 src_pgd = pgd_offset(src_mm, addr);
1286 do {
1287 next = pgd_addr_end(addr, end);
1288 if (pgd_none_or_clear_bad(src_pgd))
1289 continue;
1290 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1291 addr, next))) {
1292 ret = -ENOMEM;
1293 break;
1294 }
1295 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1296
1297 if (is_cow) {
1298 raw_write_seqcount_end(&src_mm->write_protect_seq);
1299 mmu_notifier_invalidate_range_end(&range);
1300 }
1301 return ret;
1302}
1303
1304static unsigned long zap_pte_range(struct mmu_gather *tlb,
1305 struct vm_area_struct *vma, pmd_t *pmd,
1306 unsigned long addr, unsigned long end,
1307 struct zap_details *details)
1308{
1309 struct mm_struct *mm = tlb->mm;
1310 int force_flush = 0;
1311 int rss[NR_MM_COUNTERS];
1312 spinlock_t *ptl;
1313 pte_t *start_pte;
1314 pte_t *pte;
1315 swp_entry_t entry;
1316
1317 tlb_change_page_size(tlb, PAGE_SIZE);
1318again:
1319 init_rss_vec(rss);
1320 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1321 pte = start_pte;
1322 flush_tlb_batched_pending(mm);
1323 arch_enter_lazy_mmu_mode();
1324 do {
1325 pte_t ptent = *pte;
1326 if (pte_none(ptent))
1327 continue;
1328
1329 if (need_resched())
1330 break;
1331
1332 if (pte_present(ptent)) {
1333 struct page *page;
1334
1335 page = vm_normal_page(vma, addr, ptent);
1336 if (unlikely(details) && page) {
1337 /*
1338 * unmap_shared_mapping_pages() wants to
1339 * invalidate cache without truncating:
1340 * unmap shared but keep private pages.
1341 */
1342 if (details->check_mapping &&
1343 details->check_mapping != page_rmapping(page))
1344 continue;
1345 }
1346 ptent = ptep_get_and_clear_full(mm, addr, pte,
1347 tlb->fullmm);
1348 tlb_remove_tlb_entry(tlb, pte, addr);
1349 if (unlikely(!page))
1350 continue;
1351
1352 if (!PageAnon(page)) {
1353 if (pte_dirty(ptent)) {
1354 force_flush = 1;
1355 set_page_dirty(page);
1356 }
1357 if (pte_young(ptent) &&
1358 likely(!(vma->vm_flags & VM_SEQ_READ)))
1359 mark_page_accessed(page);
1360 }
1361 rss[mm_counter(page)]--;
1362 page_remove_rmap(page, false);
1363 if (unlikely(page_mapcount(page) < 0))
1364 print_bad_pte(vma, addr, ptent, page);
1365 if (unlikely(__tlb_remove_page(tlb, page))) {
1366 force_flush = 1;
1367 addr += PAGE_SIZE;
1368 break;
1369 }
1370 continue;
1371 }
1372
1373 entry = pte_to_swp_entry(ptent);
1374 if (is_device_private_entry(entry) ||
1375 is_device_exclusive_entry(entry)) {
1376 struct page *page = pfn_swap_entry_to_page(entry);
1377
1378 if (unlikely(details && details->check_mapping)) {
1379 /*
1380 * unmap_shared_mapping_pages() wants to
1381 * invalidate cache without truncating:
1382 * unmap shared but keep private pages.
1383 */
1384 if (details->check_mapping !=
1385 page_rmapping(page))
1386 continue;
1387 }
1388
1389 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1390 rss[mm_counter(page)]--;
1391
1392 if (is_device_private_entry(entry))
1393 page_remove_rmap(page, false);
1394
1395 put_page(page);
1396 continue;
1397 }
1398
1399 /* If details->check_mapping, we leave swap entries. */
1400 if (unlikely(details))
1401 continue;
1402
1403 if (!non_swap_entry(entry))
1404 rss[MM_SWAPENTS]--;
1405 else if (is_migration_entry(entry)) {
1406 struct page *page;
1407
1408 page = pfn_swap_entry_to_page(entry);
1409 rss[mm_counter(page)]--;
1410 }
1411 if (unlikely(!free_swap_and_cache(entry)))
1412 print_bad_pte(vma, addr, ptent, NULL);
1413 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1414 } while (pte++, addr += PAGE_SIZE, addr != end);
1415
1416 add_mm_rss_vec(mm, rss);
1417 arch_leave_lazy_mmu_mode();
1418
1419 /* Do the actual TLB flush before dropping ptl */
1420 if (force_flush)
1421 tlb_flush_mmu_tlbonly(tlb);
1422 pte_unmap_unlock(start_pte, ptl);
1423
1424 /*
1425 * If we forced a TLB flush (either due to running out of
1426 * batch buffers or because we needed to flush dirty TLB
1427 * entries before releasing the ptl), free the batched
1428 * memory too. Restart if we didn't do everything.
1429 */
1430 if (force_flush) {
1431 force_flush = 0;
1432 tlb_flush_mmu(tlb);
1433 }
1434
1435 if (addr != end) {
1436 cond_resched();
1437 goto again;
1438 }
1439
1440 return addr;
1441}
1442
1443static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1444 struct vm_area_struct *vma, pud_t *pud,
1445 unsigned long addr, unsigned long end,
1446 struct zap_details *details)
1447{
1448 pmd_t *pmd;
1449 unsigned long next;
1450
1451 pmd = pmd_offset(pud, addr);
1452 do {
1453 next = pmd_addr_end(addr, end);
1454 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1455 if (next - addr != HPAGE_PMD_SIZE)
1456 __split_huge_pmd(vma, pmd, addr, false, NULL);
1457 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1458 goto next;
1459 /* fall through */
1460 } else if (details && details->single_page &&
1461 PageTransCompound(details->single_page) &&
1462 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1463 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1464 /*
1465 * Take and drop THP pmd lock so that we cannot return
1466 * prematurely, while zap_huge_pmd() has cleared *pmd,
1467 * but not yet decremented compound_mapcount().
1468 */
1469 spin_unlock(ptl);
1470 }
1471
1472 /*
1473 * Here there can be other concurrent MADV_DONTNEED or
1474 * trans huge page faults running, and if the pmd is
1475 * none or trans huge it can change under us. This is
1476 * because MADV_DONTNEED holds the mmap_lock in read
1477 * mode.
1478 */
1479 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1480 goto next;
1481 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1482next:
1483 cond_resched();
1484 } while (pmd++, addr = next, addr != end);
1485
1486 return addr;
1487}
1488
1489static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1490 struct vm_area_struct *vma, p4d_t *p4d,
1491 unsigned long addr, unsigned long end,
1492 struct zap_details *details)
1493{
1494 pud_t *pud;
1495 unsigned long next;
1496
1497 pud = pud_offset(p4d, addr);
1498 do {
1499 next = pud_addr_end(addr, end);
1500 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1501 if (next - addr != HPAGE_PUD_SIZE) {
1502 mmap_assert_locked(tlb->mm);
1503 split_huge_pud(vma, pud, addr);
1504 } else if (zap_huge_pud(tlb, vma, pud, addr))
1505 goto next;
1506 /* fall through */
1507 }
1508 if (pud_none_or_clear_bad(pud))
1509 continue;
1510 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1511next:
1512 cond_resched();
1513 } while (pud++, addr = next, addr != end);
1514
1515 return addr;
1516}
1517
1518static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1519 struct vm_area_struct *vma, pgd_t *pgd,
1520 unsigned long addr, unsigned long end,
1521 struct zap_details *details)
1522{
1523 p4d_t *p4d;
1524 unsigned long next;
1525
1526 p4d = p4d_offset(pgd, addr);
1527 do {
1528 next = p4d_addr_end(addr, end);
1529 if (p4d_none_or_clear_bad(p4d))
1530 continue;
1531 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1532 } while (p4d++, addr = next, addr != end);
1533
1534 return addr;
1535}
1536
1537void unmap_page_range(struct mmu_gather *tlb,
1538 struct vm_area_struct *vma,
1539 unsigned long addr, unsigned long end,
1540 struct zap_details *details)
1541{
1542 pgd_t *pgd;
1543 unsigned long next;
1544
1545 BUG_ON(addr >= end);
1546 tlb_start_vma(tlb, vma);
1547 pgd = pgd_offset(vma->vm_mm, addr);
1548 do {
1549 next = pgd_addr_end(addr, end);
1550 if (pgd_none_or_clear_bad(pgd))
1551 continue;
1552 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1553 } while (pgd++, addr = next, addr != end);
1554 tlb_end_vma(tlb, vma);
1555}
1556
1557
1558static void unmap_single_vma(struct mmu_gather *tlb,
1559 struct vm_area_struct *vma, unsigned long start_addr,
1560 unsigned long end_addr,
1561 struct zap_details *details)
1562{
1563 unsigned long start = max(vma->vm_start, start_addr);
1564 unsigned long end;
1565
1566 if (start >= vma->vm_end)
1567 return;
1568 end = min(vma->vm_end, end_addr);
1569 if (end <= vma->vm_start)
1570 return;
1571
1572 if (vma->vm_file)
1573 uprobe_munmap(vma, start, end);
1574
1575 if (unlikely(vma->vm_flags & VM_PFNMAP))
1576 untrack_pfn(vma, 0, 0);
1577
1578 if (start != end) {
1579 if (unlikely(is_vm_hugetlb_page(vma))) {
1580 /*
1581 * It is undesirable to test vma->vm_file as it
1582 * should be non-null for valid hugetlb area.
1583 * However, vm_file will be NULL in the error
1584 * cleanup path of mmap_region. When
1585 * hugetlbfs ->mmap method fails,
1586 * mmap_region() nullifies vma->vm_file
1587 * before calling this function to clean up.
1588 * Since no pte has actually been setup, it is
1589 * safe to do nothing in this case.
1590 */
1591 if (vma->vm_file) {
1592 i_mmap_lock_write(vma->vm_file->f_mapping);
1593 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1594 i_mmap_unlock_write(vma->vm_file->f_mapping);
1595 }
1596 } else
1597 unmap_page_range(tlb, vma, start, end, details);
1598 }
1599}
1600
1601/**
1602 * unmap_vmas - unmap a range of memory covered by a list of vma's
1603 * @tlb: address of the caller's struct mmu_gather
1604 * @vma: the starting vma
1605 * @start_addr: virtual address at which to start unmapping
1606 * @end_addr: virtual address at which to end unmapping
1607 *
1608 * Unmap all pages in the vma list.
1609 *
1610 * Only addresses between `start' and `end' will be unmapped.
1611 *
1612 * The VMA list must be sorted in ascending virtual address order.
1613 *
1614 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1615 * range after unmap_vmas() returns. So the only responsibility here is to
1616 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1617 * drops the lock and schedules.
1618 */
1619void unmap_vmas(struct mmu_gather *tlb,
1620 struct vm_area_struct *vma, unsigned long start_addr,
1621 unsigned long end_addr)
1622{
1623 struct mmu_notifier_range range;
1624
1625 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1626 start_addr, end_addr);
1627 mmu_notifier_invalidate_range_start(&range);
1628 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1629 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1630 mmu_notifier_invalidate_range_end(&range);
1631}
1632
1633/**
1634 * zap_page_range - remove user pages in a given range
1635 * @vma: vm_area_struct holding the applicable pages
1636 * @start: starting address of pages to zap
1637 * @size: number of bytes to zap
1638 *
1639 * Caller must protect the VMA list
1640 */
1641void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1642 unsigned long size)
1643{
1644 struct mmu_notifier_range range;
1645 struct mmu_gather tlb;
1646
1647 lru_add_drain();
1648 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1649 start, start + size);
1650 tlb_gather_mmu(&tlb, vma->vm_mm);
1651 update_hiwater_rss(vma->vm_mm);
1652 mmu_notifier_invalidate_range_start(&range);
1653 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1654 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1655 mmu_notifier_invalidate_range_end(&range);
1656 tlb_finish_mmu(&tlb);
1657}
1658
1659/**
1660 * zap_page_range_single - remove user pages in a given range
1661 * @vma: vm_area_struct holding the applicable pages
1662 * @address: starting address of pages to zap
1663 * @size: number of bytes to zap
1664 * @details: details of shared cache invalidation
1665 *
1666 * The range must fit into one VMA.
1667 */
1668static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1669 unsigned long size, struct zap_details *details)
1670{
1671 struct mmu_notifier_range range;
1672 struct mmu_gather tlb;
1673
1674 lru_add_drain();
1675 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1676 address, address + size);
1677 tlb_gather_mmu(&tlb, vma->vm_mm);
1678 update_hiwater_rss(vma->vm_mm);
1679 mmu_notifier_invalidate_range_start(&range);
1680 unmap_single_vma(&tlb, vma, address, range.end, details);
1681 mmu_notifier_invalidate_range_end(&range);
1682 tlb_finish_mmu(&tlb);
1683}
1684
1685/**
1686 * zap_vma_ptes - remove ptes mapping the vma
1687 * @vma: vm_area_struct holding ptes to be zapped
1688 * @address: starting address of pages to zap
1689 * @size: number of bytes to zap
1690 *
1691 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1692 *
1693 * The entire address range must be fully contained within the vma.
1694 *
1695 */
1696void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1697 unsigned long size)
1698{
1699 if (address < vma->vm_start || address + size > vma->vm_end ||
1700 !(vma->vm_flags & VM_PFNMAP))
1701 return;
1702
1703 zap_page_range_single(vma, address, size, NULL);
1704}
1705EXPORT_SYMBOL_GPL(zap_vma_ptes);
1706
1707static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1708{
1709 pgd_t *pgd;
1710 p4d_t *p4d;
1711 pud_t *pud;
1712 pmd_t *pmd;
1713
1714 pgd = pgd_offset(mm, addr);
1715 p4d = p4d_alloc(mm, pgd, addr);
1716 if (!p4d)
1717 return NULL;
1718 pud = pud_alloc(mm, p4d, addr);
1719 if (!pud)
1720 return NULL;
1721 pmd = pmd_alloc(mm, pud, addr);
1722 if (!pmd)
1723 return NULL;
1724
1725 VM_BUG_ON(pmd_trans_huge(*pmd));
1726 return pmd;
1727}
1728
1729pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1730 spinlock_t **ptl)
1731{
1732 pmd_t *pmd = walk_to_pmd(mm, addr);
1733
1734 if (!pmd)
1735 return NULL;
1736 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1737}
1738
1739static int validate_page_before_insert(struct page *page)
1740{
1741 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1742 return -EINVAL;
1743 flush_dcache_page(page);
1744 return 0;
1745}
1746
1747static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1748 unsigned long addr, struct page *page, pgprot_t prot)
1749{
1750 if (!pte_none(*pte))
1751 return -EBUSY;
1752 /* Ok, finally just insert the thing.. */
1753 get_page(page);
1754 inc_mm_counter_fast(mm, mm_counter_file(page));
1755 page_add_file_rmap(page, false);
1756 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1757 return 0;
1758}
1759
1760/*
1761 * This is the old fallback for page remapping.
1762 *
1763 * For historical reasons, it only allows reserved pages. Only
1764 * old drivers should use this, and they needed to mark their
1765 * pages reserved for the old functions anyway.
1766 */
1767static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1768 struct page *page, pgprot_t prot)
1769{
1770 struct mm_struct *mm = vma->vm_mm;
1771 int retval;
1772 pte_t *pte;
1773 spinlock_t *ptl;
1774
1775 retval = validate_page_before_insert(page);
1776 if (retval)
1777 goto out;
1778 retval = -ENOMEM;
1779 pte = get_locked_pte(mm, addr, &ptl);
1780 if (!pte)
1781 goto out;
1782 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1783 pte_unmap_unlock(pte, ptl);
1784out:
1785 return retval;
1786}
1787
1788#ifdef pte_index
1789static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1790 unsigned long addr, struct page *page, pgprot_t prot)
1791{
1792 int err;
1793
1794 if (!page_count(page))
1795 return -EINVAL;
1796 err = validate_page_before_insert(page);
1797 if (err)
1798 return err;
1799 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1800}
1801
1802/* insert_pages() amortizes the cost of spinlock operations
1803 * when inserting pages in a loop. Arch *must* define pte_index.
1804 */
1805static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1806 struct page **pages, unsigned long *num, pgprot_t prot)
1807{
1808 pmd_t *pmd = NULL;
1809 pte_t *start_pte, *pte;
1810 spinlock_t *pte_lock;
1811 struct mm_struct *const mm = vma->vm_mm;
1812 unsigned long curr_page_idx = 0;
1813 unsigned long remaining_pages_total = *num;
1814 unsigned long pages_to_write_in_pmd;
1815 int ret;
1816more:
1817 ret = -EFAULT;
1818 pmd = walk_to_pmd(mm, addr);
1819 if (!pmd)
1820 goto out;
1821
1822 pages_to_write_in_pmd = min_t(unsigned long,
1823 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1824
1825 /* Allocate the PTE if necessary; takes PMD lock once only. */
1826 ret = -ENOMEM;
1827 if (pte_alloc(mm, pmd))
1828 goto out;
1829
1830 while (pages_to_write_in_pmd) {
1831 int pte_idx = 0;
1832 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1833
1834 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1835 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1836 int err = insert_page_in_batch_locked(mm, pte,
1837 addr, pages[curr_page_idx], prot);
1838 if (unlikely(err)) {
1839 pte_unmap_unlock(start_pte, pte_lock);
1840 ret = err;
1841 remaining_pages_total -= pte_idx;
1842 goto out;
1843 }
1844 addr += PAGE_SIZE;
1845 ++curr_page_idx;
1846 }
1847 pte_unmap_unlock(start_pte, pte_lock);
1848 pages_to_write_in_pmd -= batch_size;
1849 remaining_pages_total -= batch_size;
1850 }
1851 if (remaining_pages_total)
1852 goto more;
1853 ret = 0;
1854out:
1855 *num = remaining_pages_total;
1856 return ret;
1857}
1858#endif /* ifdef pte_index */
1859
1860/**
1861 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1862 * @vma: user vma to map to
1863 * @addr: target start user address of these pages
1864 * @pages: source kernel pages
1865 * @num: in: number of pages to map. out: number of pages that were *not*
1866 * mapped. (0 means all pages were successfully mapped).
1867 *
1868 * Preferred over vm_insert_page() when inserting multiple pages.
1869 *
1870 * In case of error, we may have mapped a subset of the provided
1871 * pages. It is the caller's responsibility to account for this case.
1872 *
1873 * The same restrictions apply as in vm_insert_page().
1874 */
1875int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1876 struct page **pages, unsigned long *num)
1877{
1878#ifdef pte_index
1879 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1880
1881 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1882 return -EFAULT;
1883 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1884 BUG_ON(mmap_read_trylock(vma->vm_mm));
1885 BUG_ON(vma->vm_flags & VM_PFNMAP);
1886 vma->vm_flags |= VM_MIXEDMAP;
1887 }
1888 /* Defer page refcount checking till we're about to map that page. */
1889 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1890#else
1891 unsigned long idx = 0, pgcount = *num;
1892 int err = -EINVAL;
1893
1894 for (; idx < pgcount; ++idx) {
1895 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1896 if (err)
1897 break;
1898 }
1899 *num = pgcount - idx;
1900 return err;
1901#endif /* ifdef pte_index */
1902}
1903EXPORT_SYMBOL(vm_insert_pages);
1904
1905/**
1906 * vm_insert_page - insert single page into user vma
1907 * @vma: user vma to map to
1908 * @addr: target user address of this page
1909 * @page: source kernel page
1910 *
1911 * This allows drivers to insert individual pages they've allocated
1912 * into a user vma.
1913 *
1914 * The page has to be a nice clean _individual_ kernel allocation.
1915 * If you allocate a compound page, you need to have marked it as
1916 * such (__GFP_COMP), or manually just split the page up yourself
1917 * (see split_page()).
1918 *
1919 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1920 * took an arbitrary page protection parameter. This doesn't allow
1921 * that. Your vma protection will have to be set up correctly, which
1922 * means that if you want a shared writable mapping, you'd better
1923 * ask for a shared writable mapping!
1924 *
1925 * The page does not need to be reserved.
1926 *
1927 * Usually this function is called from f_op->mmap() handler
1928 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1929 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1930 * function from other places, for example from page-fault handler.
1931 *
1932 * Return: %0 on success, negative error code otherwise.
1933 */
1934int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1935 struct page *page)
1936{
1937 if (addr < vma->vm_start || addr >= vma->vm_end)
1938 return -EFAULT;
1939 if (!page_count(page))
1940 return -EINVAL;
1941 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1942 BUG_ON(mmap_read_trylock(vma->vm_mm));
1943 BUG_ON(vma->vm_flags & VM_PFNMAP);
1944 vma->vm_flags |= VM_MIXEDMAP;
1945 }
1946 return insert_page(vma, addr, page, vma->vm_page_prot);
1947}
1948EXPORT_SYMBOL(vm_insert_page);
1949
1950/*
1951 * __vm_map_pages - maps range of kernel pages into user vma
1952 * @vma: user vma to map to
1953 * @pages: pointer to array of source kernel pages
1954 * @num: number of pages in page array
1955 * @offset: user's requested vm_pgoff
1956 *
1957 * This allows drivers to map range of kernel pages into a user vma.
1958 *
1959 * Return: 0 on success and error code otherwise.
1960 */
1961static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1962 unsigned long num, unsigned long offset)
1963{
1964 unsigned long count = vma_pages(vma);
1965 unsigned long uaddr = vma->vm_start;
1966 int ret, i;
1967
1968 /* Fail if the user requested offset is beyond the end of the object */
1969 if (offset >= num)
1970 return -ENXIO;
1971
1972 /* Fail if the user requested size exceeds available object size */
1973 if (count > num - offset)
1974 return -ENXIO;
1975
1976 for (i = 0; i < count; i++) {
1977 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1978 if (ret < 0)
1979 return ret;
1980 uaddr += PAGE_SIZE;
1981 }
1982
1983 return 0;
1984}
1985
1986/**
1987 * vm_map_pages - maps range of kernel pages starts with non zero offset
1988 * @vma: user vma to map to
1989 * @pages: pointer to array of source kernel pages
1990 * @num: number of pages in page array
1991 *
1992 * Maps an object consisting of @num pages, catering for the user's
1993 * requested vm_pgoff
1994 *
1995 * If we fail to insert any page into the vma, the function will return
1996 * immediately leaving any previously inserted pages present. Callers
1997 * from the mmap handler may immediately return the error as their caller
1998 * will destroy the vma, removing any successfully inserted pages. Other
1999 * callers should make their own arrangements for calling unmap_region().
2000 *
2001 * Context: Process context. Called by mmap handlers.
2002 * Return: 0 on success and error code otherwise.
2003 */
2004int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2005 unsigned long num)
2006{
2007 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2008}
2009EXPORT_SYMBOL(vm_map_pages);
2010
2011/**
2012 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2013 * @vma: user vma to map to
2014 * @pages: pointer to array of source kernel pages
2015 * @num: number of pages in page array
2016 *
2017 * Similar to vm_map_pages(), except that it explicitly sets the offset
2018 * to 0. This function is intended for the drivers that did not consider
2019 * vm_pgoff.
2020 *
2021 * Context: Process context. Called by mmap handlers.
2022 * Return: 0 on success and error code otherwise.
2023 */
2024int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2025 unsigned long num)
2026{
2027 return __vm_map_pages(vma, pages, num, 0);
2028}
2029EXPORT_SYMBOL(vm_map_pages_zero);
2030
2031static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2032 pfn_t pfn, pgprot_t prot, bool mkwrite)
2033{
2034 struct mm_struct *mm = vma->vm_mm;
2035 pte_t *pte, entry;
2036 spinlock_t *ptl;
2037
2038 pte = get_locked_pte(mm, addr, &ptl);
2039 if (!pte)
2040 return VM_FAULT_OOM;
2041 if (!pte_none(*pte)) {
2042 if (mkwrite) {
2043 /*
2044 * For read faults on private mappings the PFN passed
2045 * in may not match the PFN we have mapped if the
2046 * mapped PFN is a writeable COW page. In the mkwrite
2047 * case we are creating a writable PTE for a shared
2048 * mapping and we expect the PFNs to match. If they
2049 * don't match, we are likely racing with block
2050 * allocation and mapping invalidation so just skip the
2051 * update.
2052 */
2053 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
2054 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
2055 goto out_unlock;
2056 }
2057 entry = pte_mkyoung(*pte);
2058 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2059 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2060 update_mmu_cache(vma, addr, pte);
2061 }
2062 goto out_unlock;
2063 }
2064
2065 /* Ok, finally just insert the thing.. */
2066 if (pfn_t_devmap(pfn))
2067 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2068 else
2069 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2070
2071 if (mkwrite) {
2072 entry = pte_mkyoung(entry);
2073 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2074 }
2075
2076 set_pte_at(mm, addr, pte, entry);
2077 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2078
2079out_unlock:
2080 pte_unmap_unlock(pte, ptl);
2081 return VM_FAULT_NOPAGE;
2082}
2083
2084/**
2085 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2086 * @vma: user vma to map to
2087 * @addr: target user address of this page
2088 * @pfn: source kernel pfn
2089 * @pgprot: pgprot flags for the inserted page
2090 *
2091 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2092 * to override pgprot on a per-page basis.
2093 *
2094 * This only makes sense for IO mappings, and it makes no sense for
2095 * COW mappings. In general, using multiple vmas is preferable;
2096 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2097 * impractical.
2098 *
2099 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2100 * a value of @pgprot different from that of @vma->vm_page_prot.
2101 *
2102 * Context: Process context. May allocate using %GFP_KERNEL.
2103 * Return: vm_fault_t value.
2104 */
2105vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2106 unsigned long pfn, pgprot_t pgprot)
2107{
2108 /*
2109 * Technically, architectures with pte_special can avoid all these
2110 * restrictions (same for remap_pfn_range). However we would like
2111 * consistency in testing and feature parity among all, so we should
2112 * try to keep these invariants in place for everybody.
2113 */
2114 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2115 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2116 (VM_PFNMAP|VM_MIXEDMAP));
2117 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2118 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2119
2120 if (addr < vma->vm_start || addr >= vma->vm_end)
2121 return VM_FAULT_SIGBUS;
2122
2123 if (!pfn_modify_allowed(pfn, pgprot))
2124 return VM_FAULT_SIGBUS;
2125
2126 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2127
2128 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2129 false);
2130}
2131EXPORT_SYMBOL(vmf_insert_pfn_prot);
2132
2133/**
2134 * vmf_insert_pfn - insert single pfn into user vma
2135 * @vma: user vma to map to
2136 * @addr: target user address of this page
2137 * @pfn: source kernel pfn
2138 *
2139 * Similar to vm_insert_page, this allows drivers to insert individual pages
2140 * they've allocated into a user vma. Same comments apply.
2141 *
2142 * This function should only be called from a vm_ops->fault handler, and
2143 * in that case the handler should return the result of this function.
2144 *
2145 * vma cannot be a COW mapping.
2146 *
2147 * As this is called only for pages that do not currently exist, we
2148 * do not need to flush old virtual caches or the TLB.
2149 *
2150 * Context: Process context. May allocate using %GFP_KERNEL.
2151 * Return: vm_fault_t value.
2152 */
2153vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2154 unsigned long pfn)
2155{
2156 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2157}
2158EXPORT_SYMBOL(vmf_insert_pfn);
2159
2160static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2161{
2162 /* these checks mirror the abort conditions in vm_normal_page */
2163 if (vma->vm_flags & VM_MIXEDMAP)
2164 return true;
2165 if (pfn_t_devmap(pfn))
2166 return true;
2167 if (pfn_t_special(pfn))
2168 return true;
2169 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2170 return true;
2171 return false;
2172}
2173
2174static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2175 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2176 bool mkwrite)
2177{
2178 int err;
2179
2180 BUG_ON(!vm_mixed_ok(vma, pfn));
2181
2182 if (addr < vma->vm_start || addr >= vma->vm_end)
2183 return VM_FAULT_SIGBUS;
2184
2185 track_pfn_insert(vma, &pgprot, pfn);
2186
2187 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2188 return VM_FAULT_SIGBUS;
2189
2190 /*
2191 * If we don't have pte special, then we have to use the pfn_valid()
2192 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2193 * refcount the page if pfn_valid is true (hence insert_page rather
2194 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2195 * without pte special, it would there be refcounted as a normal page.
2196 */
2197 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2198 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2199 struct page *page;
2200
2201 /*
2202 * At this point we are committed to insert_page()
2203 * regardless of whether the caller specified flags that
2204 * result in pfn_t_has_page() == false.
2205 */
2206 page = pfn_to_page(pfn_t_to_pfn(pfn));
2207 err = insert_page(vma, addr, page, pgprot);
2208 } else {
2209 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2210 }
2211
2212 if (err == -ENOMEM)
2213 return VM_FAULT_OOM;
2214 if (err < 0 && err != -EBUSY)
2215 return VM_FAULT_SIGBUS;
2216
2217 return VM_FAULT_NOPAGE;
2218}
2219
2220/**
2221 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2222 * @vma: user vma to map to
2223 * @addr: target user address of this page
2224 * @pfn: source kernel pfn
2225 * @pgprot: pgprot flags for the inserted page
2226 *
2227 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2228 * to override pgprot on a per-page basis.
2229 *
2230 * Typically this function should be used by drivers to set caching- and
2231 * encryption bits different than those of @vma->vm_page_prot, because
2232 * the caching- or encryption mode may not be known at mmap() time.
2233 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2234 * to set caching and encryption bits for those vmas (except for COW pages).
2235 * This is ensured by core vm only modifying these page table entries using
2236 * functions that don't touch caching- or encryption bits, using pte_modify()
2237 * if needed. (See for example mprotect()).
2238 * Also when new page-table entries are created, this is only done using the
2239 * fault() callback, and never using the value of vma->vm_page_prot,
2240 * except for page-table entries that point to anonymous pages as the result
2241 * of COW.
2242 *
2243 * Context: Process context. May allocate using %GFP_KERNEL.
2244 * Return: vm_fault_t value.
2245 */
2246vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2247 pfn_t pfn, pgprot_t pgprot)
2248{
2249 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2250}
2251EXPORT_SYMBOL(vmf_insert_mixed_prot);
2252
2253vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2254 pfn_t pfn)
2255{
2256 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2257}
2258EXPORT_SYMBOL(vmf_insert_mixed);
2259
2260/*
2261 * If the insertion of PTE failed because someone else already added a
2262 * different entry in the mean time, we treat that as success as we assume
2263 * the same entry was actually inserted.
2264 */
2265vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2266 unsigned long addr, pfn_t pfn)
2267{
2268 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2269}
2270EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2271
2272/*
2273 * maps a range of physical memory into the requested pages. the old
2274 * mappings are removed. any references to nonexistent pages results
2275 * in null mappings (currently treated as "copy-on-access")
2276 */
2277static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2278 unsigned long addr, unsigned long end,
2279 unsigned long pfn, pgprot_t prot)
2280{
2281 pte_t *pte, *mapped_pte;
2282 spinlock_t *ptl;
2283 int err = 0;
2284
2285 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2286 if (!pte)
2287 return -ENOMEM;
2288 arch_enter_lazy_mmu_mode();
2289 do {
2290 BUG_ON(!pte_none(*pte));
2291 if (!pfn_modify_allowed(pfn, prot)) {
2292 err = -EACCES;
2293 break;
2294 }
2295 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2296 pfn++;
2297 } while (pte++, addr += PAGE_SIZE, addr != end);
2298 arch_leave_lazy_mmu_mode();
2299 pte_unmap_unlock(mapped_pte, ptl);
2300 return err;
2301}
2302
2303static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2304 unsigned long addr, unsigned long end,
2305 unsigned long pfn, pgprot_t prot)
2306{
2307 pmd_t *pmd;
2308 unsigned long next;
2309 int err;
2310
2311 pfn -= addr >> PAGE_SHIFT;
2312 pmd = pmd_alloc(mm, pud, addr);
2313 if (!pmd)
2314 return -ENOMEM;
2315 VM_BUG_ON(pmd_trans_huge(*pmd));
2316 do {
2317 next = pmd_addr_end(addr, end);
2318 err = remap_pte_range(mm, pmd, addr, next,
2319 pfn + (addr >> PAGE_SHIFT), prot);
2320 if (err)
2321 return err;
2322 } while (pmd++, addr = next, addr != end);
2323 return 0;
2324}
2325
2326static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2327 unsigned long addr, unsigned long end,
2328 unsigned long pfn, pgprot_t prot)
2329{
2330 pud_t *pud;
2331 unsigned long next;
2332 int err;
2333
2334 pfn -= addr >> PAGE_SHIFT;
2335 pud = pud_alloc(mm, p4d, addr);
2336 if (!pud)
2337 return -ENOMEM;
2338 do {
2339 next = pud_addr_end(addr, end);
2340 err = remap_pmd_range(mm, pud, addr, next,
2341 pfn + (addr >> PAGE_SHIFT), prot);
2342 if (err)
2343 return err;
2344 } while (pud++, addr = next, addr != end);
2345 return 0;
2346}
2347
2348static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2349 unsigned long addr, unsigned long end,
2350 unsigned long pfn, pgprot_t prot)
2351{
2352 p4d_t *p4d;
2353 unsigned long next;
2354 int err;
2355
2356 pfn -= addr >> PAGE_SHIFT;
2357 p4d = p4d_alloc(mm, pgd, addr);
2358 if (!p4d)
2359 return -ENOMEM;
2360 do {
2361 next = p4d_addr_end(addr, end);
2362 err = remap_pud_range(mm, p4d, addr, next,
2363 pfn + (addr >> PAGE_SHIFT), prot);
2364 if (err)
2365 return err;
2366 } while (p4d++, addr = next, addr != end);
2367 return 0;
2368}
2369
2370/*
2371 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2372 * must have pre-validated the caching bits of the pgprot_t.
2373 */
2374int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2375 unsigned long pfn, unsigned long size, pgprot_t prot)
2376{
2377 pgd_t *pgd;
2378 unsigned long next;
2379 unsigned long end = addr + PAGE_ALIGN(size);
2380 struct mm_struct *mm = vma->vm_mm;
2381 int err;
2382
2383 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2384 return -EINVAL;
2385
2386 /*
2387 * Physically remapped pages are special. Tell the
2388 * rest of the world about it:
2389 * VM_IO tells people not to look at these pages
2390 * (accesses can have side effects).
2391 * VM_PFNMAP tells the core MM that the base pages are just
2392 * raw PFN mappings, and do not have a "struct page" associated
2393 * with them.
2394 * VM_DONTEXPAND
2395 * Disable vma merging and expanding with mremap().
2396 * VM_DONTDUMP
2397 * Omit vma from core dump, even when VM_IO turned off.
2398 *
2399 * There's a horrible special case to handle copy-on-write
2400 * behaviour that some programs depend on. We mark the "original"
2401 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2402 * See vm_normal_page() for details.
2403 */
2404 if (is_cow_mapping(vma->vm_flags)) {
2405 if (addr != vma->vm_start || end != vma->vm_end)
2406 return -EINVAL;
2407 vma->vm_pgoff = pfn;
2408 }
2409
2410 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2411
2412 BUG_ON(addr >= end);
2413 pfn -= addr >> PAGE_SHIFT;
2414 pgd = pgd_offset(mm, addr);
2415 flush_cache_range(vma, addr, end);
2416 do {
2417 next = pgd_addr_end(addr, end);
2418 err = remap_p4d_range(mm, pgd, addr, next,
2419 pfn + (addr >> PAGE_SHIFT), prot);
2420 if (err)
2421 return err;
2422 } while (pgd++, addr = next, addr != end);
2423
2424 return 0;
2425}
2426
2427/**
2428 * remap_pfn_range - remap kernel memory to userspace
2429 * @vma: user vma to map to
2430 * @addr: target page aligned user address to start at
2431 * @pfn: page frame number of kernel physical memory address
2432 * @size: size of mapping area
2433 * @prot: page protection flags for this mapping
2434 *
2435 * Note: this is only safe if the mm semaphore is held when called.
2436 *
2437 * Return: %0 on success, negative error code otherwise.
2438 */
2439int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2440 unsigned long pfn, unsigned long size, pgprot_t prot)
2441{
2442 int err;
2443
2444 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2445 if (err)
2446 return -EINVAL;
2447
2448 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2449 if (err)
2450 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2451 return err;
2452}
2453EXPORT_SYMBOL(remap_pfn_range);
2454
2455/**
2456 * vm_iomap_memory - remap memory to userspace
2457 * @vma: user vma to map to
2458 * @start: start of the physical memory to be mapped
2459 * @len: size of area
2460 *
2461 * This is a simplified io_remap_pfn_range() for common driver use. The
2462 * driver just needs to give us the physical memory range to be mapped,
2463 * we'll figure out the rest from the vma information.
2464 *
2465 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2466 * whatever write-combining details or similar.
2467 *
2468 * Return: %0 on success, negative error code otherwise.
2469 */
2470int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2471{
2472 unsigned long vm_len, pfn, pages;
2473
2474 /* Check that the physical memory area passed in looks valid */
2475 if (start + len < start)
2476 return -EINVAL;
2477 /*
2478 * You *really* shouldn't map things that aren't page-aligned,
2479 * but we've historically allowed it because IO memory might
2480 * just have smaller alignment.
2481 */
2482 len += start & ~PAGE_MASK;
2483 pfn = start >> PAGE_SHIFT;
2484 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2485 if (pfn + pages < pfn)
2486 return -EINVAL;
2487
2488 /* We start the mapping 'vm_pgoff' pages into the area */
2489 if (vma->vm_pgoff > pages)
2490 return -EINVAL;
2491 pfn += vma->vm_pgoff;
2492 pages -= vma->vm_pgoff;
2493
2494 /* Can we fit all of the mapping? */
2495 vm_len = vma->vm_end - vma->vm_start;
2496 if (vm_len >> PAGE_SHIFT > pages)
2497 return -EINVAL;
2498
2499 /* Ok, let it rip */
2500 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2501}
2502EXPORT_SYMBOL(vm_iomap_memory);
2503
2504static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2505 unsigned long addr, unsigned long end,
2506 pte_fn_t fn, void *data, bool create,
2507 pgtbl_mod_mask *mask)
2508{
2509 pte_t *pte, *mapped_pte;
2510 int err = 0;
2511 spinlock_t *ptl;
2512
2513 if (create) {
2514 mapped_pte = pte = (mm == &init_mm) ?
2515 pte_alloc_kernel_track(pmd, addr, mask) :
2516 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2517 if (!pte)
2518 return -ENOMEM;
2519 } else {
2520 mapped_pte = pte = (mm == &init_mm) ?
2521 pte_offset_kernel(pmd, addr) :
2522 pte_offset_map_lock(mm, pmd, addr, &ptl);
2523 }
2524
2525 BUG_ON(pmd_huge(*pmd));
2526
2527 arch_enter_lazy_mmu_mode();
2528
2529 if (fn) {
2530 do {
2531 if (create || !pte_none(*pte)) {
2532 err = fn(pte++, addr, data);
2533 if (err)
2534 break;
2535 }
2536 } while (addr += PAGE_SIZE, addr != end);
2537 }
2538 *mask |= PGTBL_PTE_MODIFIED;
2539
2540 arch_leave_lazy_mmu_mode();
2541
2542 if (mm != &init_mm)
2543 pte_unmap_unlock(mapped_pte, ptl);
2544 return err;
2545}
2546
2547static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2548 unsigned long addr, unsigned long end,
2549 pte_fn_t fn, void *data, bool create,
2550 pgtbl_mod_mask *mask)
2551{
2552 pmd_t *pmd;
2553 unsigned long next;
2554 int err = 0;
2555
2556 BUG_ON(pud_huge(*pud));
2557
2558 if (create) {
2559 pmd = pmd_alloc_track(mm, pud, addr, mask);
2560 if (!pmd)
2561 return -ENOMEM;
2562 } else {
2563 pmd = pmd_offset(pud, addr);
2564 }
2565 do {
2566 next = pmd_addr_end(addr, end);
2567 if (pmd_none(*pmd) && !create)
2568 continue;
2569 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2570 return -EINVAL;
2571 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2572 if (!create)
2573 continue;
2574 pmd_clear_bad(pmd);
2575 }
2576 err = apply_to_pte_range(mm, pmd, addr, next,
2577 fn, data, create, mask);
2578 if (err)
2579 break;
2580 } while (pmd++, addr = next, addr != end);
2581
2582 return err;
2583}
2584
2585static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2586 unsigned long addr, unsigned long end,
2587 pte_fn_t fn, void *data, bool create,
2588 pgtbl_mod_mask *mask)
2589{
2590 pud_t *pud;
2591 unsigned long next;
2592 int err = 0;
2593
2594 if (create) {
2595 pud = pud_alloc_track(mm, p4d, addr, mask);
2596 if (!pud)
2597 return -ENOMEM;
2598 } else {
2599 pud = pud_offset(p4d, addr);
2600 }
2601 do {
2602 next = pud_addr_end(addr, end);
2603 if (pud_none(*pud) && !create)
2604 continue;
2605 if (WARN_ON_ONCE(pud_leaf(*pud)))
2606 return -EINVAL;
2607 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2608 if (!create)
2609 continue;
2610 pud_clear_bad(pud);
2611 }
2612 err = apply_to_pmd_range(mm, pud, addr, next,
2613 fn, data, create, mask);
2614 if (err)
2615 break;
2616 } while (pud++, addr = next, addr != end);
2617
2618 return err;
2619}
2620
2621static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2622 unsigned long addr, unsigned long end,
2623 pte_fn_t fn, void *data, bool create,
2624 pgtbl_mod_mask *mask)
2625{
2626 p4d_t *p4d;
2627 unsigned long next;
2628 int err = 0;
2629
2630 if (create) {
2631 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2632 if (!p4d)
2633 return -ENOMEM;
2634 } else {
2635 p4d = p4d_offset(pgd, addr);
2636 }
2637 do {
2638 next = p4d_addr_end(addr, end);
2639 if (p4d_none(*p4d) && !create)
2640 continue;
2641 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2642 return -EINVAL;
2643 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2644 if (!create)
2645 continue;
2646 p4d_clear_bad(p4d);
2647 }
2648 err = apply_to_pud_range(mm, p4d, addr, next,
2649 fn, data, create, mask);
2650 if (err)
2651 break;
2652 } while (p4d++, addr = next, addr != end);
2653
2654 return err;
2655}
2656
2657static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2658 unsigned long size, pte_fn_t fn,
2659 void *data, bool create)
2660{
2661 pgd_t *pgd;
2662 unsigned long start = addr, next;
2663 unsigned long end = addr + size;
2664 pgtbl_mod_mask mask = 0;
2665 int err = 0;
2666
2667 if (WARN_ON(addr >= end))
2668 return -EINVAL;
2669
2670 pgd = pgd_offset(mm, addr);
2671 do {
2672 next = pgd_addr_end(addr, end);
2673 if (pgd_none(*pgd) && !create)
2674 continue;
2675 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2676 return -EINVAL;
2677 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2678 if (!create)
2679 continue;
2680 pgd_clear_bad(pgd);
2681 }
2682 err = apply_to_p4d_range(mm, pgd, addr, next,
2683 fn, data, create, &mask);
2684 if (err)
2685 break;
2686 } while (pgd++, addr = next, addr != end);
2687
2688 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2689 arch_sync_kernel_mappings(start, start + size);
2690
2691 return err;
2692}
2693
2694/*
2695 * Scan a region of virtual memory, filling in page tables as necessary
2696 * and calling a provided function on each leaf page table.
2697 */
2698int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2699 unsigned long size, pte_fn_t fn, void *data)
2700{
2701 return __apply_to_page_range(mm, addr, size, fn, data, true);
2702}
2703EXPORT_SYMBOL_GPL(apply_to_page_range);
2704
2705/*
2706 * Scan a region of virtual memory, calling a provided function on
2707 * each leaf page table where it exists.
2708 *
2709 * Unlike apply_to_page_range, this does _not_ fill in page tables
2710 * where they are absent.
2711 */
2712int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2713 unsigned long size, pte_fn_t fn, void *data)
2714{
2715 return __apply_to_page_range(mm, addr, size, fn, data, false);
2716}
2717EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2718
2719/*
2720 * handle_pte_fault chooses page fault handler according to an entry which was
2721 * read non-atomically. Before making any commitment, on those architectures
2722 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2723 * parts, do_swap_page must check under lock before unmapping the pte and
2724 * proceeding (but do_wp_page is only called after already making such a check;
2725 * and do_anonymous_page can safely check later on).
2726 */
2727static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2728 pte_t *page_table, pte_t orig_pte)
2729{
2730 int same = 1;
2731#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2732 if (sizeof(pte_t) > sizeof(unsigned long)) {
2733 spinlock_t *ptl = pte_lockptr(mm, pmd);
2734 spin_lock(ptl);
2735 same = pte_same(*page_table, orig_pte);
2736 spin_unlock(ptl);
2737 }
2738#endif
2739 pte_unmap(page_table);
2740 return same;
2741}
2742
2743static inline bool cow_user_page(struct page *dst, struct page *src,
2744 struct vm_fault *vmf)
2745{
2746 bool ret;
2747 void *kaddr;
2748 void __user *uaddr;
2749 bool locked = false;
2750 struct vm_area_struct *vma = vmf->vma;
2751 struct mm_struct *mm = vma->vm_mm;
2752 unsigned long addr = vmf->address;
2753
2754 if (likely(src)) {
2755 copy_user_highpage(dst, src, addr, vma);
2756 return true;
2757 }
2758
2759 /*
2760 * If the source page was a PFN mapping, we don't have
2761 * a "struct page" for it. We do a best-effort copy by
2762 * just copying from the original user address. If that
2763 * fails, we just zero-fill it. Live with it.
2764 */
2765 kaddr = kmap_atomic(dst);
2766 uaddr = (void __user *)(addr & PAGE_MASK);
2767
2768 /*
2769 * On architectures with software "accessed" bits, we would
2770 * take a double page fault, so mark it accessed here.
2771 */
2772 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2773 pte_t entry;
2774
2775 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2776 locked = true;
2777 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2778 /*
2779 * Other thread has already handled the fault
2780 * and update local tlb only
2781 */
2782 update_mmu_tlb(vma, addr, vmf->pte);
2783 ret = false;
2784 goto pte_unlock;
2785 }
2786
2787 entry = pte_mkyoung(vmf->orig_pte);
2788 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2789 update_mmu_cache(vma, addr, vmf->pte);
2790 }
2791
2792 /*
2793 * This really shouldn't fail, because the page is there
2794 * in the page tables. But it might just be unreadable,
2795 * in which case we just give up and fill the result with
2796 * zeroes.
2797 */
2798 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2799 if (locked)
2800 goto warn;
2801
2802 /* Re-validate under PTL if the page is still mapped */
2803 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2804 locked = true;
2805 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2806 /* The PTE changed under us, update local tlb */
2807 update_mmu_tlb(vma, addr, vmf->pte);
2808 ret = false;
2809 goto pte_unlock;
2810 }
2811
2812 /*
2813 * The same page can be mapped back since last copy attempt.
2814 * Try to copy again under PTL.
2815 */
2816 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2817 /*
2818 * Give a warn in case there can be some obscure
2819 * use-case
2820 */
2821warn:
2822 WARN_ON_ONCE(1);
2823 clear_page(kaddr);
2824 }
2825 }
2826
2827 ret = true;
2828
2829pte_unlock:
2830 if (locked)
2831 pte_unmap_unlock(vmf->pte, vmf->ptl);
2832 kunmap_atomic(kaddr);
2833 flush_dcache_page(dst);
2834
2835 return ret;
2836}
2837
2838static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2839{
2840 struct file *vm_file = vma->vm_file;
2841
2842 if (vm_file)
2843 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2844
2845 /*
2846 * Special mappings (e.g. VDSO) do not have any file so fake
2847 * a default GFP_KERNEL for them.
2848 */
2849 return GFP_KERNEL;
2850}
2851
2852/*
2853 * Notify the address space that the page is about to become writable so that
2854 * it can prohibit this or wait for the page to get into an appropriate state.
2855 *
2856 * We do this without the lock held, so that it can sleep if it needs to.
2857 */
2858static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2859{
2860 vm_fault_t ret;
2861 struct page *page = vmf->page;
2862 unsigned int old_flags = vmf->flags;
2863
2864 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2865
2866 if (vmf->vma->vm_file &&
2867 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2868 return VM_FAULT_SIGBUS;
2869
2870 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2871 /* Restore original flags so that caller is not surprised */
2872 vmf->flags = old_flags;
2873 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2874 return ret;
2875 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2876 lock_page(page);
2877 if (!page->mapping) {
2878 unlock_page(page);
2879 return 0; /* retry */
2880 }
2881 ret |= VM_FAULT_LOCKED;
2882 } else
2883 VM_BUG_ON_PAGE(!PageLocked(page), page);
2884 return ret;
2885}
2886
2887/*
2888 * Handle dirtying of a page in shared file mapping on a write fault.
2889 *
2890 * The function expects the page to be locked and unlocks it.
2891 */
2892static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2893{
2894 struct vm_area_struct *vma = vmf->vma;
2895 struct address_space *mapping;
2896 struct page *page = vmf->page;
2897 bool dirtied;
2898 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2899
2900 dirtied = set_page_dirty(page);
2901 VM_BUG_ON_PAGE(PageAnon(page), page);
2902 /*
2903 * Take a local copy of the address_space - page.mapping may be zeroed
2904 * by truncate after unlock_page(). The address_space itself remains
2905 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2906 * release semantics to prevent the compiler from undoing this copying.
2907 */
2908 mapping = page_rmapping(page);
2909 unlock_page(page);
2910
2911 if (!page_mkwrite)
2912 file_update_time(vma->vm_file);
2913
2914 /*
2915 * Throttle page dirtying rate down to writeback speed.
2916 *
2917 * mapping may be NULL here because some device drivers do not
2918 * set page.mapping but still dirty their pages
2919 *
2920 * Drop the mmap_lock before waiting on IO, if we can. The file
2921 * is pinning the mapping, as per above.
2922 */
2923 if ((dirtied || page_mkwrite) && mapping) {
2924 struct file *fpin;
2925
2926 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2927 balance_dirty_pages_ratelimited(mapping);
2928 if (fpin) {
2929 fput(fpin);
2930 return VM_FAULT_RETRY;
2931 }
2932 }
2933
2934 return 0;
2935}
2936
2937/*
2938 * Handle write page faults for pages that can be reused in the current vma
2939 *
2940 * This can happen either due to the mapping being with the VM_SHARED flag,
2941 * or due to us being the last reference standing to the page. In either
2942 * case, all we need to do here is to mark the page as writable and update
2943 * any related book-keeping.
2944 */
2945static inline void wp_page_reuse(struct vm_fault *vmf)
2946 __releases(vmf->ptl)
2947{
2948 struct vm_area_struct *vma = vmf->vma;
2949 struct page *page = vmf->page;
2950 pte_t entry;
2951 /*
2952 * Clear the pages cpupid information as the existing
2953 * information potentially belongs to a now completely
2954 * unrelated process.
2955 */
2956 if (page)
2957 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2958
2959 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2960 entry = pte_mkyoung(vmf->orig_pte);
2961 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2962 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2963 update_mmu_cache(vma, vmf->address, vmf->pte);
2964 pte_unmap_unlock(vmf->pte, vmf->ptl);
2965 count_vm_event(PGREUSE);
2966}
2967
2968/*
2969 * Handle the case of a page which we actually need to copy to a new page.
2970 *
2971 * Called with mmap_lock locked and the old page referenced, but
2972 * without the ptl held.
2973 *
2974 * High level logic flow:
2975 *
2976 * - Allocate a page, copy the content of the old page to the new one.
2977 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2978 * - Take the PTL. If the pte changed, bail out and release the allocated page
2979 * - If the pte is still the way we remember it, update the page table and all
2980 * relevant references. This includes dropping the reference the page-table
2981 * held to the old page, as well as updating the rmap.
2982 * - In any case, unlock the PTL and drop the reference we took to the old page.
2983 */
2984static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2985{
2986 struct vm_area_struct *vma = vmf->vma;
2987 struct mm_struct *mm = vma->vm_mm;
2988 struct page *old_page = vmf->page;
2989 struct page *new_page = NULL;
2990 pte_t entry;
2991 int page_copied = 0;
2992 struct mmu_notifier_range range;
2993
2994 if (unlikely(anon_vma_prepare(vma)))
2995 goto oom;
2996
2997 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2998 new_page = alloc_zeroed_user_highpage_movable(vma,
2999 vmf->address);
3000 if (!new_page)
3001 goto oom;
3002 } else {
3003 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3004 vmf->address);
3005 if (!new_page)
3006 goto oom;
3007
3008 if (!cow_user_page(new_page, old_page, vmf)) {
3009 /*
3010 * COW failed, if the fault was solved by other,
3011 * it's fine. If not, userspace would re-fault on
3012 * the same address and we will handle the fault
3013 * from the second attempt.
3014 */
3015 put_page(new_page);
3016 if (old_page)
3017 put_page(old_page);
3018 return 0;
3019 }
3020 }
3021
3022 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
3023 goto oom_free_new;
3024 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
3025
3026 __SetPageUptodate(new_page);
3027
3028 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
3029 vmf->address & PAGE_MASK,
3030 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3031 mmu_notifier_invalidate_range_start(&range);
3032
3033 /*
3034 * Re-check the pte - we dropped the lock
3035 */
3036 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3037 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
3038 if (old_page) {
3039 if (!PageAnon(old_page)) {
3040 dec_mm_counter_fast(mm,
3041 mm_counter_file(old_page));
3042 inc_mm_counter_fast(mm, MM_ANONPAGES);
3043 }
3044 } else {
3045 inc_mm_counter_fast(mm, MM_ANONPAGES);
3046 }
3047 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3048 entry = mk_pte(new_page, vma->vm_page_prot);
3049 entry = pte_sw_mkyoung(entry);
3050 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3051
3052 /*
3053 * Clear the pte entry and flush it first, before updating the
3054 * pte with the new entry, to keep TLBs on different CPUs in
3055 * sync. This code used to set the new PTE then flush TLBs, but
3056 * that left a window where the new PTE could be loaded into
3057 * some TLBs while the old PTE remains in others.
3058 */
3059 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
3060 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
3061 lru_cache_add_inactive_or_unevictable(new_page, vma);
3062 /*
3063 * We call the notify macro here because, when using secondary
3064 * mmu page tables (such as kvm shadow page tables), we want the
3065 * new page to be mapped directly into the secondary page table.
3066 */
3067 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3068 update_mmu_cache(vma, vmf->address, vmf->pte);
3069 if (old_page) {
3070 /*
3071 * Only after switching the pte to the new page may
3072 * we remove the mapcount here. Otherwise another
3073 * process may come and find the rmap count decremented
3074 * before the pte is switched to the new page, and
3075 * "reuse" the old page writing into it while our pte
3076 * here still points into it and can be read by other
3077 * threads.
3078 *
3079 * The critical issue is to order this
3080 * page_remove_rmap with the ptp_clear_flush above.
3081 * Those stores are ordered by (if nothing else,)
3082 * the barrier present in the atomic_add_negative
3083 * in page_remove_rmap.
3084 *
3085 * Then the TLB flush in ptep_clear_flush ensures that
3086 * no process can access the old page before the
3087 * decremented mapcount is visible. And the old page
3088 * cannot be reused until after the decremented
3089 * mapcount is visible. So transitively, TLBs to
3090 * old page will be flushed before it can be reused.
3091 */
3092 page_remove_rmap(old_page, false);
3093 }
3094
3095 /* Free the old page.. */
3096 new_page = old_page;
3097 page_copied = 1;
3098 } else {
3099 update_mmu_tlb(vma, vmf->address, vmf->pte);
3100 }
3101
3102 if (new_page)
3103 put_page(new_page);
3104
3105 pte_unmap_unlock(vmf->pte, vmf->ptl);
3106 /*
3107 * No need to double call mmu_notifier->invalidate_range() callback as
3108 * the above ptep_clear_flush_notify() did already call it.
3109 */
3110 mmu_notifier_invalidate_range_only_end(&range);
3111 if (old_page) {
3112 /*
3113 * Don't let another task, with possibly unlocked vma,
3114 * keep the mlocked page.
3115 */
3116 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
3117 lock_page(old_page); /* LRU manipulation */
3118 if (PageMlocked(old_page))
3119 munlock_vma_page(old_page);
3120 unlock_page(old_page);
3121 }
3122 if (page_copied)
3123 free_swap_cache(old_page);
3124 put_page(old_page);
3125 }
3126 return page_copied ? VM_FAULT_WRITE : 0;
3127oom_free_new:
3128 put_page(new_page);
3129oom:
3130 if (old_page)
3131 put_page(old_page);
3132 return VM_FAULT_OOM;
3133}
3134
3135/**
3136 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3137 * writeable once the page is prepared
3138 *
3139 * @vmf: structure describing the fault
3140 *
3141 * This function handles all that is needed to finish a write page fault in a
3142 * shared mapping due to PTE being read-only once the mapped page is prepared.
3143 * It handles locking of PTE and modifying it.
3144 *
3145 * The function expects the page to be locked or other protection against
3146 * concurrent faults / writeback (such as DAX radix tree locks).
3147 *
3148 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3149 * we acquired PTE lock.
3150 */
3151vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3152{
3153 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3154 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3155 &vmf->ptl);
3156 /*
3157 * We might have raced with another page fault while we released the
3158 * pte_offset_map_lock.
3159 */
3160 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3161 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3162 pte_unmap_unlock(vmf->pte, vmf->ptl);
3163 return VM_FAULT_NOPAGE;
3164 }
3165 wp_page_reuse(vmf);
3166 return 0;
3167}
3168
3169/*
3170 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3171 * mapping
3172 */
3173static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3174{
3175 struct vm_area_struct *vma = vmf->vma;
3176
3177 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3178 vm_fault_t ret;
3179
3180 pte_unmap_unlock(vmf->pte, vmf->ptl);
3181 vmf->flags |= FAULT_FLAG_MKWRITE;
3182 ret = vma->vm_ops->pfn_mkwrite(vmf);
3183 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3184 return ret;
3185 return finish_mkwrite_fault(vmf);
3186 }
3187 wp_page_reuse(vmf);
3188 return VM_FAULT_WRITE;
3189}
3190
3191static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3192 __releases(vmf->ptl)
3193{
3194 struct vm_area_struct *vma = vmf->vma;
3195 vm_fault_t ret = VM_FAULT_WRITE;
3196
3197 get_page(vmf->page);
3198
3199 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3200 vm_fault_t tmp;
3201
3202 pte_unmap_unlock(vmf->pte, vmf->ptl);
3203 tmp = do_page_mkwrite(vmf);
3204 if (unlikely(!tmp || (tmp &
3205 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3206 put_page(vmf->page);
3207 return tmp;
3208 }
3209 tmp = finish_mkwrite_fault(vmf);
3210 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3211 unlock_page(vmf->page);
3212 put_page(vmf->page);
3213 return tmp;
3214 }
3215 } else {
3216 wp_page_reuse(vmf);
3217 lock_page(vmf->page);
3218 }
3219 ret |= fault_dirty_shared_page(vmf);
3220 put_page(vmf->page);
3221
3222 return ret;
3223}
3224
3225/*
3226 * This routine handles present pages, when users try to write
3227 * to a shared page. It is done by copying the page to a new address
3228 * and decrementing the shared-page counter for the old page.
3229 *
3230 * Note that this routine assumes that the protection checks have been
3231 * done by the caller (the low-level page fault routine in most cases).
3232 * Thus we can safely just mark it writable once we've done any necessary
3233 * COW.
3234 *
3235 * We also mark the page dirty at this point even though the page will
3236 * change only once the write actually happens. This avoids a few races,
3237 * and potentially makes it more efficient.
3238 *
3239 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3240 * but allow concurrent faults), with pte both mapped and locked.
3241 * We return with mmap_lock still held, but pte unmapped and unlocked.
3242 */
3243static vm_fault_t do_wp_page(struct vm_fault *vmf)
3244 __releases(vmf->ptl)
3245{
3246 struct vm_area_struct *vma = vmf->vma;
3247
3248 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3249 pte_unmap_unlock(vmf->pte, vmf->ptl);
3250 return handle_userfault(vmf, VM_UFFD_WP);
3251 }
3252
3253 /*
3254 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3255 * is flushed in this case before copying.
3256 */
3257 if (unlikely(userfaultfd_wp(vmf->vma) &&
3258 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3259 flush_tlb_page(vmf->vma, vmf->address);
3260
3261 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3262 if (!vmf->page) {
3263 /*
3264 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3265 * VM_PFNMAP VMA.
3266 *
3267 * We should not cow pages in a shared writeable mapping.
3268 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3269 */
3270 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3271 (VM_WRITE|VM_SHARED))
3272 return wp_pfn_shared(vmf);
3273
3274 pte_unmap_unlock(vmf->pte, vmf->ptl);
3275 return wp_page_copy(vmf);
3276 }
3277
3278 /*
3279 * Take out anonymous pages first, anonymous shared vmas are
3280 * not dirty accountable.
3281 */
3282 if (PageAnon(vmf->page)) {
3283 struct page *page = vmf->page;
3284
3285 /* PageKsm() doesn't necessarily raise the page refcount */
3286 if (PageKsm(page) || page_count(page) != 1)
3287 goto copy;
3288 if (!trylock_page(page))
3289 goto copy;
3290 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3291 unlock_page(page);
3292 goto copy;
3293 }
3294 /*
3295 * Ok, we've got the only map reference, and the only
3296 * page count reference, and the page is locked,
3297 * it's dark out, and we're wearing sunglasses. Hit it.
3298 */
3299 unlock_page(page);
3300 wp_page_reuse(vmf);
3301 return VM_FAULT_WRITE;
3302 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3303 (VM_WRITE|VM_SHARED))) {
3304 return wp_page_shared(vmf);
3305 }
3306copy:
3307 /*
3308 * Ok, we need to copy. Oh, well..
3309 */
3310 get_page(vmf->page);
3311
3312 pte_unmap_unlock(vmf->pte, vmf->ptl);
3313 return wp_page_copy(vmf);
3314}
3315
3316static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3317 unsigned long start_addr, unsigned long end_addr,
3318 struct zap_details *details)
3319{
3320 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3321}
3322
3323static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3324 struct zap_details *details)
3325{
3326 struct vm_area_struct *vma;
3327 pgoff_t vba, vea, zba, zea;
3328
3329 vma_interval_tree_foreach(vma, root,
3330 details->first_index, details->last_index) {
3331
3332 vba = vma->vm_pgoff;
3333 vea = vba + vma_pages(vma) - 1;
3334 zba = details->first_index;
3335 if (zba < vba)
3336 zba = vba;
3337 zea = details->last_index;
3338 if (zea > vea)
3339 zea = vea;
3340
3341 unmap_mapping_range_vma(vma,
3342 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3343 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3344 details);
3345 }
3346}
3347
3348/**
3349 * unmap_mapping_page() - Unmap single page from processes.
3350 * @page: The locked page to be unmapped.
3351 *
3352 * Unmap this page from any userspace process which still has it mmaped.
3353 * Typically, for efficiency, the range of nearby pages has already been
3354 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3355 * truncation or invalidation holds the lock on a page, it may find that
3356 * the page has been remapped again: and then uses unmap_mapping_page()
3357 * to unmap it finally.
3358 */
3359void unmap_mapping_page(struct page *page)
3360{
3361 struct address_space *mapping = page->mapping;
3362 struct zap_details details = { };
3363
3364 VM_BUG_ON(!PageLocked(page));
3365 VM_BUG_ON(PageTail(page));
3366
3367 details.check_mapping = mapping;
3368 details.first_index = page->index;
3369 details.last_index = page->index + thp_nr_pages(page) - 1;
3370 details.single_page = page;
3371
3372 i_mmap_lock_write(mapping);
3373 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3374 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3375 i_mmap_unlock_write(mapping);
3376}
3377
3378/**
3379 * unmap_mapping_pages() - Unmap pages from processes.
3380 * @mapping: The address space containing pages to be unmapped.
3381 * @start: Index of first page to be unmapped.
3382 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3383 * @even_cows: Whether to unmap even private COWed pages.
3384 *
3385 * Unmap the pages in this address space from any userspace process which
3386 * has them mmaped. Generally, you want to remove COWed pages as well when
3387 * a file is being truncated, but not when invalidating pages from the page
3388 * cache.
3389 */
3390void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3391 pgoff_t nr, bool even_cows)
3392{
3393 struct zap_details details = { };
3394
3395 details.check_mapping = even_cows ? NULL : mapping;
3396 details.first_index = start;
3397 details.last_index = start + nr - 1;
3398 if (details.last_index < details.first_index)
3399 details.last_index = ULONG_MAX;
3400
3401 i_mmap_lock_write(mapping);
3402 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3403 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3404 i_mmap_unlock_write(mapping);
3405}
3406
3407/**
3408 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3409 * address_space corresponding to the specified byte range in the underlying
3410 * file.
3411 *
3412 * @mapping: the address space containing mmaps to be unmapped.
3413 * @holebegin: byte in first page to unmap, relative to the start of
3414 * the underlying file. This will be rounded down to a PAGE_SIZE
3415 * boundary. Note that this is different from truncate_pagecache(), which
3416 * must keep the partial page. In contrast, we must get rid of
3417 * partial pages.
3418 * @holelen: size of prospective hole in bytes. This will be rounded
3419 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3420 * end of the file.
3421 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3422 * but 0 when invalidating pagecache, don't throw away private data.
3423 */
3424void unmap_mapping_range(struct address_space *mapping,
3425 loff_t const holebegin, loff_t const holelen, int even_cows)
3426{
3427 pgoff_t hba = holebegin >> PAGE_SHIFT;
3428 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3429
3430 /* Check for overflow. */
3431 if (sizeof(holelen) > sizeof(hlen)) {
3432 long long holeend =
3433 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3434 if (holeend & ~(long long)ULONG_MAX)
3435 hlen = ULONG_MAX - hba + 1;
3436 }
3437
3438 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3439}
3440EXPORT_SYMBOL(unmap_mapping_range);
3441
3442/*
3443 * Restore a potential device exclusive pte to a working pte entry
3444 */
3445static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3446{
3447 struct page *page = vmf->page;
3448 struct vm_area_struct *vma = vmf->vma;
3449 struct mmu_notifier_range range;
3450
3451 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags))
3452 return VM_FAULT_RETRY;
3453 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
3454 vma->vm_mm, vmf->address & PAGE_MASK,
3455 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3456 mmu_notifier_invalidate_range_start(&range);
3457
3458 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3459 &vmf->ptl);
3460 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3461 restore_exclusive_pte(vma, page, vmf->address, vmf->pte);
3462
3463 pte_unmap_unlock(vmf->pte, vmf->ptl);
3464 unlock_page(page);
3465
3466 mmu_notifier_invalidate_range_end(&range);
3467 return 0;
3468}
3469
3470/*
3471 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3472 * but allow concurrent faults), and pte mapped but not yet locked.
3473 * We return with pte unmapped and unlocked.
3474 *
3475 * We return with the mmap_lock locked or unlocked in the same cases
3476 * as does filemap_fault().
3477 */
3478vm_fault_t do_swap_page(struct vm_fault *vmf)
3479{
3480 struct vm_area_struct *vma = vmf->vma;
3481 struct page *page = NULL, *swapcache;
3482 struct swap_info_struct *si = NULL;
3483 swp_entry_t entry;
3484 pte_t pte;
3485 int locked;
3486 int exclusive = 0;
3487 vm_fault_t ret = 0;
3488 void *shadow = NULL;
3489
3490 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3491 goto out;
3492
3493 entry = pte_to_swp_entry(vmf->orig_pte);
3494 if (unlikely(non_swap_entry(entry))) {
3495 if (is_migration_entry(entry)) {
3496 migration_entry_wait(vma->vm_mm, vmf->pmd,
3497 vmf->address);
3498 } else if (is_device_exclusive_entry(entry)) {
3499 vmf->page = pfn_swap_entry_to_page(entry);
3500 ret = remove_device_exclusive_entry(vmf);
3501 } else if (is_device_private_entry(entry)) {
3502 vmf->page = pfn_swap_entry_to_page(entry);
3503 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3504 } else if (is_hwpoison_entry(entry)) {
3505 ret = VM_FAULT_HWPOISON;
3506 } else {
3507 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3508 ret = VM_FAULT_SIGBUS;
3509 }
3510 goto out;
3511 }
3512
3513 /* Prevent swapoff from happening to us. */
3514 si = get_swap_device(entry);
3515 if (unlikely(!si))
3516 goto out;
3517
3518 delayacct_set_flag(current, DELAYACCT_PF_SWAPIN);
3519 page = lookup_swap_cache(entry, vma, vmf->address);
3520 swapcache = page;
3521
3522 if (!page) {
3523 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3524 __swap_count(entry) == 1) {
3525 /* skip swapcache */
3526 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3527 vmf->address);
3528 if (page) {
3529 __SetPageLocked(page);
3530 __SetPageSwapBacked(page);
3531
3532 if (mem_cgroup_swapin_charge_page(page,
3533 vma->vm_mm, GFP_KERNEL, entry)) {
3534 ret = VM_FAULT_OOM;
3535 goto out_page;
3536 }
3537 mem_cgroup_swapin_uncharge_swap(entry);
3538
3539 shadow = get_shadow_from_swap_cache(entry);
3540 if (shadow)
3541 workingset_refault(page, shadow);
3542
3543 lru_cache_add(page);
3544
3545 /* To provide entry to swap_readpage() */
3546 set_page_private(page, entry.val);
3547 swap_readpage(page, true);
3548 set_page_private(page, 0);
3549 }
3550 } else {
3551 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3552 vmf);
3553 swapcache = page;
3554 }
3555
3556 if (!page) {
3557 /*
3558 * Back out if somebody else faulted in this pte
3559 * while we released the pte lock.
3560 */
3561 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3562 vmf->address, &vmf->ptl);
3563 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3564 ret = VM_FAULT_OOM;
3565 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3566 goto unlock;
3567 }
3568
3569 /* Had to read the page from swap area: Major fault */
3570 ret = VM_FAULT_MAJOR;
3571 count_vm_event(PGMAJFAULT);
3572 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3573 } else if (PageHWPoison(page)) {
3574 /*
3575 * hwpoisoned dirty swapcache pages are kept for killing
3576 * owner processes (which may be unknown at hwpoison time)
3577 */
3578 ret = VM_FAULT_HWPOISON;
3579 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3580 goto out_release;
3581 }
3582
3583 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3584
3585 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3586 if (!locked) {
3587 ret |= VM_FAULT_RETRY;
3588 goto out_release;
3589 }
3590
3591 /*
3592 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3593 * release the swapcache from under us. The page pin, and pte_same
3594 * test below, are not enough to exclude that. Even if it is still
3595 * swapcache, we need to check that the page's swap has not changed.
3596 */
3597 if (unlikely((!PageSwapCache(page) ||
3598 page_private(page) != entry.val)) && swapcache)
3599 goto out_page;
3600
3601 page = ksm_might_need_to_copy(page, vma, vmf->address);
3602 if (unlikely(!page)) {
3603 ret = VM_FAULT_OOM;
3604 page = swapcache;
3605 goto out_page;
3606 }
3607
3608 cgroup_throttle_swaprate(page, GFP_KERNEL);
3609
3610 /*
3611 * Back out if somebody else already faulted in this pte.
3612 */
3613 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3614 &vmf->ptl);
3615 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3616 goto out_nomap;
3617
3618 if (unlikely(!PageUptodate(page))) {
3619 ret = VM_FAULT_SIGBUS;
3620 goto out_nomap;
3621 }
3622
3623 /*
3624 * The page isn't present yet, go ahead with the fault.
3625 *
3626 * Be careful about the sequence of operations here.
3627 * To get its accounting right, reuse_swap_page() must be called
3628 * while the page is counted on swap but not yet in mapcount i.e.
3629 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3630 * must be called after the swap_free(), or it will never succeed.
3631 */
3632
3633 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3634 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3635 pte = mk_pte(page, vma->vm_page_prot);
3636 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3637 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3638 vmf->flags &= ~FAULT_FLAG_WRITE;
3639 ret |= VM_FAULT_WRITE;
3640 exclusive = RMAP_EXCLUSIVE;
3641 }
3642 flush_icache_page(vma, page);
3643 if (pte_swp_soft_dirty(vmf->orig_pte))
3644 pte = pte_mksoft_dirty(pte);
3645 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3646 pte = pte_mkuffd_wp(pte);
3647 pte = pte_wrprotect(pte);
3648 }
3649 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3650 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3651 vmf->orig_pte = pte;
3652
3653 /* ksm created a completely new copy */
3654 if (unlikely(page != swapcache && swapcache)) {
3655 page_add_new_anon_rmap(page, vma, vmf->address, false);
3656 lru_cache_add_inactive_or_unevictable(page, vma);
3657 } else {
3658 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3659 }
3660
3661 swap_free(entry);
3662 if (mem_cgroup_swap_full(page) ||
3663 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3664 try_to_free_swap(page);
3665 unlock_page(page);
3666 if (page != swapcache && swapcache) {
3667 /*
3668 * Hold the lock to avoid the swap entry to be reused
3669 * until we take the PT lock for the pte_same() check
3670 * (to avoid false positives from pte_same). For
3671 * further safety release the lock after the swap_free
3672 * so that the swap count won't change under a
3673 * parallel locked swapcache.
3674 */
3675 unlock_page(swapcache);
3676 put_page(swapcache);
3677 }
3678
3679 if (vmf->flags & FAULT_FLAG_WRITE) {
3680 ret |= do_wp_page(vmf);
3681 if (ret & VM_FAULT_ERROR)
3682 ret &= VM_FAULT_ERROR;
3683 goto out;
3684 }
3685
3686 /* No need to invalidate - it was non-present before */
3687 update_mmu_cache(vma, vmf->address, vmf->pte);
3688unlock:
3689 pte_unmap_unlock(vmf->pte, vmf->ptl);
3690out:
3691 if (si)
3692 put_swap_device(si);
3693 return ret;
3694out_nomap:
3695 pte_unmap_unlock(vmf->pte, vmf->ptl);
3696out_page:
3697 unlock_page(page);
3698out_release:
3699 put_page(page);
3700 if (page != swapcache && swapcache) {
3701 unlock_page(swapcache);
3702 put_page(swapcache);
3703 }
3704 if (si)
3705 put_swap_device(si);
3706 return ret;
3707}
3708
3709/*
3710 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3711 * but allow concurrent faults), and pte mapped but not yet locked.
3712 * We return with mmap_lock still held, but pte unmapped and unlocked.
3713 */
3714static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3715{
3716 struct vm_area_struct *vma = vmf->vma;
3717 struct page *page;
3718 vm_fault_t ret = 0;
3719 pte_t entry;
3720
3721 /* File mapping without ->vm_ops ? */
3722 if (vma->vm_flags & VM_SHARED)
3723 return VM_FAULT_SIGBUS;
3724
3725 /*
3726 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3727 * pte_offset_map() on pmds where a huge pmd might be created
3728 * from a different thread.
3729 *
3730 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3731 * parallel threads are excluded by other means.
3732 *
3733 * Here we only have mmap_read_lock(mm).
3734 */
3735 if (pte_alloc(vma->vm_mm, vmf->pmd))
3736 return VM_FAULT_OOM;
3737
3738 /* See comment in handle_pte_fault() */
3739 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3740 return 0;
3741
3742 /* Use the zero-page for reads */
3743 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3744 !mm_forbids_zeropage(vma->vm_mm)) {
3745 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3746 vma->vm_page_prot));
3747 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3748 vmf->address, &vmf->ptl);
3749 if (!pte_none(*vmf->pte)) {
3750 update_mmu_tlb(vma, vmf->address, vmf->pte);
3751 goto unlock;
3752 }
3753 ret = check_stable_address_space(vma->vm_mm);
3754 if (ret)
3755 goto unlock;
3756 /* Deliver the page fault to userland, check inside PT lock */
3757 if (userfaultfd_missing(vma)) {
3758 pte_unmap_unlock(vmf->pte, vmf->ptl);
3759 return handle_userfault(vmf, VM_UFFD_MISSING);
3760 }
3761 goto setpte;
3762 }
3763
3764 /* Allocate our own private page. */
3765 if (unlikely(anon_vma_prepare(vma)))
3766 goto oom;
3767 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3768 if (!page)
3769 goto oom;
3770
3771 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3772 goto oom_free_page;
3773 cgroup_throttle_swaprate(page, GFP_KERNEL);
3774
3775 /*
3776 * The memory barrier inside __SetPageUptodate makes sure that
3777 * preceding stores to the page contents become visible before
3778 * the set_pte_at() write.
3779 */
3780 __SetPageUptodate(page);
3781
3782 entry = mk_pte(page, vma->vm_page_prot);
3783 entry = pte_sw_mkyoung(entry);
3784 if (vma->vm_flags & VM_WRITE)
3785 entry = pte_mkwrite(pte_mkdirty(entry));
3786
3787 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3788 &vmf->ptl);
3789 if (!pte_none(*vmf->pte)) {
3790 update_mmu_cache(vma, vmf->address, vmf->pte);
3791 goto release;
3792 }
3793
3794 ret = check_stable_address_space(vma->vm_mm);
3795 if (ret)
3796 goto release;
3797
3798 /* Deliver the page fault to userland, check inside PT lock */
3799 if (userfaultfd_missing(vma)) {
3800 pte_unmap_unlock(vmf->pte, vmf->ptl);
3801 put_page(page);
3802 return handle_userfault(vmf, VM_UFFD_MISSING);
3803 }
3804
3805 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3806 page_add_new_anon_rmap(page, vma, vmf->address, false);
3807 lru_cache_add_inactive_or_unevictable(page, vma);
3808setpte:
3809 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3810
3811 /* No need to invalidate - it was non-present before */
3812 update_mmu_cache(vma, vmf->address, vmf->pte);
3813unlock:
3814 pte_unmap_unlock(vmf->pte, vmf->ptl);
3815 return ret;
3816release:
3817 put_page(page);
3818 goto unlock;
3819oom_free_page:
3820 put_page(page);
3821oom:
3822 return VM_FAULT_OOM;
3823}
3824
3825/*
3826 * The mmap_lock must have been held on entry, and may have been
3827 * released depending on flags and vma->vm_ops->fault() return value.
3828 * See filemap_fault() and __lock_page_retry().
3829 */
3830static vm_fault_t __do_fault(struct vm_fault *vmf)
3831{
3832 struct vm_area_struct *vma = vmf->vma;
3833 vm_fault_t ret;
3834
3835 /*
3836 * Preallocate pte before we take page_lock because this might lead to
3837 * deadlocks for memcg reclaim which waits for pages under writeback:
3838 * lock_page(A)
3839 * SetPageWriteback(A)
3840 * unlock_page(A)
3841 * lock_page(B)
3842 * lock_page(B)
3843 * pte_alloc_one
3844 * shrink_page_list
3845 * wait_on_page_writeback(A)
3846 * SetPageWriteback(B)
3847 * unlock_page(B)
3848 * # flush A, B to clear the writeback
3849 */
3850 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3851 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3852 if (!vmf->prealloc_pte)
3853 return VM_FAULT_OOM;
3854 smp_wmb(); /* See comment in __pte_alloc() */
3855 }
3856
3857 ret = vma->vm_ops->fault(vmf);
3858 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3859 VM_FAULT_DONE_COW)))
3860 return ret;
3861
3862 if (unlikely(PageHWPoison(vmf->page))) {
3863 if (ret & VM_FAULT_LOCKED)
3864 unlock_page(vmf->page);
3865 put_page(vmf->page);
3866 vmf->page = NULL;
3867 return VM_FAULT_HWPOISON;
3868 }
3869
3870 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3871 lock_page(vmf->page);
3872 else
3873 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3874
3875 return ret;
3876}
3877
3878#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3879static void deposit_prealloc_pte(struct vm_fault *vmf)
3880{
3881 struct vm_area_struct *vma = vmf->vma;
3882
3883 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3884 /*
3885 * We are going to consume the prealloc table,
3886 * count that as nr_ptes.
3887 */
3888 mm_inc_nr_ptes(vma->vm_mm);
3889 vmf->prealloc_pte = NULL;
3890}
3891
3892vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3893{
3894 struct vm_area_struct *vma = vmf->vma;
3895 bool write = vmf->flags & FAULT_FLAG_WRITE;
3896 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3897 pmd_t entry;
3898 int i;
3899 vm_fault_t ret = VM_FAULT_FALLBACK;
3900
3901 if (!transhuge_vma_suitable(vma, haddr))
3902 return ret;
3903
3904 page = compound_head(page);
3905 if (compound_order(page) != HPAGE_PMD_ORDER)
3906 return ret;
3907
3908 /*
3909 * Archs like ppc64 need additional space to store information
3910 * related to pte entry. Use the preallocated table for that.
3911 */
3912 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3913 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3914 if (!vmf->prealloc_pte)
3915 return VM_FAULT_OOM;
3916 smp_wmb(); /* See comment in __pte_alloc() */
3917 }
3918
3919 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3920 if (unlikely(!pmd_none(*vmf->pmd)))
3921 goto out;
3922
3923 for (i = 0; i < HPAGE_PMD_NR; i++)
3924 flush_icache_page(vma, page + i);
3925
3926 entry = mk_huge_pmd(page, vma->vm_page_prot);
3927 if (write)
3928 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3929
3930 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3931 page_add_file_rmap(page, true);
3932 /*
3933 * deposit and withdraw with pmd lock held
3934 */
3935 if (arch_needs_pgtable_deposit())
3936 deposit_prealloc_pte(vmf);
3937
3938 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3939
3940 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3941
3942 /* fault is handled */
3943 ret = 0;
3944 count_vm_event(THP_FILE_MAPPED);
3945out:
3946 spin_unlock(vmf->ptl);
3947 return ret;
3948}
3949#else
3950vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3951{
3952 return VM_FAULT_FALLBACK;
3953}
3954#endif
3955
3956void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
3957{
3958 struct vm_area_struct *vma = vmf->vma;
3959 bool write = vmf->flags & FAULT_FLAG_WRITE;
3960 bool prefault = vmf->address != addr;
3961 pte_t entry;
3962
3963 flush_icache_page(vma, page);
3964 entry = mk_pte(page, vma->vm_page_prot);
3965
3966 if (prefault && arch_wants_old_prefaulted_pte())
3967 entry = pte_mkold(entry);
3968 else
3969 entry = pte_sw_mkyoung(entry);
3970
3971 if (write)
3972 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3973 /* copy-on-write page */
3974 if (write && !(vma->vm_flags & VM_SHARED)) {
3975 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3976 page_add_new_anon_rmap(page, vma, addr, false);
3977 lru_cache_add_inactive_or_unevictable(page, vma);
3978 } else {
3979 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3980 page_add_file_rmap(page, false);
3981 }
3982 set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
3983}
3984
3985/**
3986 * finish_fault - finish page fault once we have prepared the page to fault
3987 *
3988 * @vmf: structure describing the fault
3989 *
3990 * This function handles all that is needed to finish a page fault once the
3991 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3992 * given page, adds reverse page mapping, handles memcg charges and LRU
3993 * addition.
3994 *
3995 * The function expects the page to be locked and on success it consumes a
3996 * reference of a page being mapped (for the PTE which maps it).
3997 *
3998 * Return: %0 on success, %VM_FAULT_ code in case of error.
3999 */
4000vm_fault_t finish_fault(struct vm_fault *vmf)
4001{
4002 struct vm_area_struct *vma = vmf->vma;
4003 struct page *page;
4004 vm_fault_t ret;
4005
4006 /* Did we COW the page? */
4007 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4008 page = vmf->cow_page;
4009 else
4010 page = vmf->page;
4011
4012 /*
4013 * check even for read faults because we might have lost our CoWed
4014 * page
4015 */
4016 if (!(vma->vm_flags & VM_SHARED)) {
4017 ret = check_stable_address_space(vma->vm_mm);
4018 if (ret)
4019 return ret;
4020 }
4021
4022 if (pmd_none(*vmf->pmd)) {
4023 if (PageTransCompound(page)) {
4024 ret = do_set_pmd(vmf, page);
4025 if (ret != VM_FAULT_FALLBACK)
4026 return ret;
4027 }
4028
4029 if (vmf->prealloc_pte) {
4030 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4031 if (likely(pmd_none(*vmf->pmd))) {
4032 mm_inc_nr_ptes(vma->vm_mm);
4033 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4034 vmf->prealloc_pte = NULL;
4035 }
4036 spin_unlock(vmf->ptl);
4037 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
4038 return VM_FAULT_OOM;
4039 }
4040 }
4041
4042 /* See comment in handle_pte_fault() */
4043 if (pmd_devmap_trans_unstable(vmf->pmd))
4044 return 0;
4045
4046 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4047 vmf->address, &vmf->ptl);
4048 ret = 0;
4049 /* Re-check under ptl */
4050 if (likely(pte_none(*vmf->pte)))
4051 do_set_pte(vmf, page, vmf->address);
4052 else
4053 ret = VM_FAULT_NOPAGE;
4054
4055 update_mmu_tlb(vma, vmf->address, vmf->pte);
4056 pte_unmap_unlock(vmf->pte, vmf->ptl);
4057 return ret;
4058}
4059
4060static unsigned long fault_around_bytes __read_mostly =
4061 rounddown_pow_of_two(65536);
4062
4063#ifdef CONFIG_DEBUG_FS
4064static int fault_around_bytes_get(void *data, u64 *val)
4065{
4066 *val = fault_around_bytes;
4067 return 0;
4068}
4069
4070/*
4071 * fault_around_bytes must be rounded down to the nearest page order as it's
4072 * what do_fault_around() expects to see.
4073 */
4074static int fault_around_bytes_set(void *data, u64 val)
4075{
4076 if (val / PAGE_SIZE > PTRS_PER_PTE)
4077 return -EINVAL;
4078 if (val > PAGE_SIZE)
4079 fault_around_bytes = rounddown_pow_of_two(val);
4080 else
4081 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
4082 return 0;
4083}
4084DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4085 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4086
4087static int __init fault_around_debugfs(void)
4088{
4089 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4090 &fault_around_bytes_fops);
4091 return 0;
4092}
4093late_initcall(fault_around_debugfs);
4094#endif
4095
4096/*
4097 * do_fault_around() tries to map few pages around the fault address. The hope
4098 * is that the pages will be needed soon and this will lower the number of
4099 * faults to handle.
4100 *
4101 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4102 * not ready to be mapped: not up-to-date, locked, etc.
4103 *
4104 * This function is called with the page table lock taken. In the split ptlock
4105 * case the page table lock only protects only those entries which belong to
4106 * the page table corresponding to the fault address.
4107 *
4108 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4109 * only once.
4110 *
4111 * fault_around_bytes defines how many bytes we'll try to map.
4112 * do_fault_around() expects it to be set to a power of two less than or equal
4113 * to PTRS_PER_PTE.
4114 *
4115 * The virtual address of the area that we map is naturally aligned to
4116 * fault_around_bytes rounded down to the machine page size
4117 * (and therefore to page order). This way it's easier to guarantee
4118 * that we don't cross page table boundaries.
4119 */
4120static vm_fault_t do_fault_around(struct vm_fault *vmf)
4121{
4122 unsigned long address = vmf->address, nr_pages, mask;
4123 pgoff_t start_pgoff = vmf->pgoff;
4124 pgoff_t end_pgoff;
4125 int off;
4126
4127 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4128 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4129
4130 address = max(address & mask, vmf->vma->vm_start);
4131 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4132 start_pgoff -= off;
4133
4134 /*
4135 * end_pgoff is either the end of the page table, the end of
4136 * the vma or nr_pages from start_pgoff, depending what is nearest.
4137 */
4138 end_pgoff = start_pgoff -
4139 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4140 PTRS_PER_PTE - 1;
4141 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4142 start_pgoff + nr_pages - 1);
4143
4144 if (pmd_none(*vmf->pmd)) {
4145 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4146 if (!vmf->prealloc_pte)
4147 return VM_FAULT_OOM;
4148 smp_wmb(); /* See comment in __pte_alloc() */
4149 }
4150
4151 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4152}
4153
4154static vm_fault_t do_read_fault(struct vm_fault *vmf)
4155{
4156 struct vm_area_struct *vma = vmf->vma;
4157 vm_fault_t ret = 0;
4158
4159 /*
4160 * Let's call ->map_pages() first and use ->fault() as fallback
4161 * if page by the offset is not ready to be mapped (cold cache or
4162 * something).
4163 */
4164 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4165 if (likely(!userfaultfd_minor(vmf->vma))) {
4166 ret = do_fault_around(vmf);
4167 if (ret)
4168 return ret;
4169 }
4170 }
4171
4172 ret = __do_fault(vmf);
4173 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4174 return ret;
4175
4176 ret |= finish_fault(vmf);
4177 unlock_page(vmf->page);
4178 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4179 put_page(vmf->page);
4180 return ret;
4181}
4182
4183static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4184{
4185 struct vm_area_struct *vma = vmf->vma;
4186 vm_fault_t ret;
4187
4188 if (unlikely(anon_vma_prepare(vma)))
4189 return VM_FAULT_OOM;
4190
4191 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4192 if (!vmf->cow_page)
4193 return VM_FAULT_OOM;
4194
4195 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4196 put_page(vmf->cow_page);
4197 return VM_FAULT_OOM;
4198 }
4199 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4200
4201 ret = __do_fault(vmf);
4202 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4203 goto uncharge_out;
4204 if (ret & VM_FAULT_DONE_COW)
4205 return ret;
4206
4207 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4208 __SetPageUptodate(vmf->cow_page);
4209
4210 ret |= finish_fault(vmf);
4211 unlock_page(vmf->page);
4212 put_page(vmf->page);
4213 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4214 goto uncharge_out;
4215 return ret;
4216uncharge_out:
4217 put_page(vmf->cow_page);
4218 return ret;
4219}
4220
4221static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4222{
4223 struct vm_area_struct *vma = vmf->vma;
4224 vm_fault_t ret, tmp;
4225
4226 ret = __do_fault(vmf);
4227 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4228 return ret;
4229
4230 /*
4231 * Check if the backing address space wants to know that the page is
4232 * about to become writable
4233 */
4234 if (vma->vm_ops->page_mkwrite) {
4235 unlock_page(vmf->page);
4236 tmp = do_page_mkwrite(vmf);
4237 if (unlikely(!tmp ||
4238 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4239 put_page(vmf->page);
4240 return tmp;
4241 }
4242 }
4243
4244 ret |= finish_fault(vmf);
4245 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4246 VM_FAULT_RETRY))) {
4247 unlock_page(vmf->page);
4248 put_page(vmf->page);
4249 return ret;
4250 }
4251
4252 ret |= fault_dirty_shared_page(vmf);
4253 return ret;
4254}
4255
4256/*
4257 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4258 * but allow concurrent faults).
4259 * The mmap_lock may have been released depending on flags and our
4260 * return value. See filemap_fault() and __lock_page_or_retry().
4261 * If mmap_lock is released, vma may become invalid (for example
4262 * by other thread calling munmap()).
4263 */
4264static vm_fault_t do_fault(struct vm_fault *vmf)
4265{
4266 struct vm_area_struct *vma = vmf->vma;
4267 struct mm_struct *vm_mm = vma->vm_mm;
4268 vm_fault_t ret;
4269
4270 /*
4271 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4272 */
4273 if (!vma->vm_ops->fault) {
4274 /*
4275 * If we find a migration pmd entry or a none pmd entry, which
4276 * should never happen, return SIGBUS
4277 */
4278 if (unlikely(!pmd_present(*vmf->pmd)))
4279 ret = VM_FAULT_SIGBUS;
4280 else {
4281 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4282 vmf->pmd,
4283 vmf->address,
4284 &vmf->ptl);
4285 /*
4286 * Make sure this is not a temporary clearing of pte
4287 * by holding ptl and checking again. A R/M/W update
4288 * of pte involves: take ptl, clearing the pte so that
4289 * we don't have concurrent modification by hardware
4290 * followed by an update.
4291 */
4292 if (unlikely(pte_none(*vmf->pte)))
4293 ret = VM_FAULT_SIGBUS;
4294 else
4295 ret = VM_FAULT_NOPAGE;
4296
4297 pte_unmap_unlock(vmf->pte, vmf->ptl);
4298 }
4299 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4300 ret = do_read_fault(vmf);
4301 else if (!(vma->vm_flags & VM_SHARED))
4302 ret = do_cow_fault(vmf);
4303 else
4304 ret = do_shared_fault(vmf);
4305
4306 /* preallocated pagetable is unused: free it */
4307 if (vmf->prealloc_pte) {
4308 pte_free(vm_mm, vmf->prealloc_pte);
4309 vmf->prealloc_pte = NULL;
4310 }
4311 return ret;
4312}
4313
4314int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4315 unsigned long addr, int page_nid, int *flags)
4316{
4317 get_page(page);
4318
4319 count_vm_numa_event(NUMA_HINT_FAULTS);
4320 if (page_nid == numa_node_id()) {
4321 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4322 *flags |= TNF_FAULT_LOCAL;
4323 }
4324
4325 return mpol_misplaced(page, vma, addr);
4326}
4327
4328static vm_fault_t do_numa_page(struct vm_fault *vmf)
4329{
4330 struct vm_area_struct *vma = vmf->vma;
4331 struct page *page = NULL;
4332 int page_nid = NUMA_NO_NODE;
4333 int last_cpupid;
4334 int target_nid;
4335 pte_t pte, old_pte;
4336 bool was_writable = pte_savedwrite(vmf->orig_pte);
4337 int flags = 0;
4338
4339 /*
4340 * The "pte" at this point cannot be used safely without
4341 * validation through pte_unmap_same(). It's of NUMA type but
4342 * the pfn may be screwed if the read is non atomic.
4343 */
4344 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4345 spin_lock(vmf->ptl);
4346 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4347 pte_unmap_unlock(vmf->pte, vmf->ptl);
4348 goto out;
4349 }
4350
4351 /* Get the normal PTE */
4352 old_pte = ptep_get(vmf->pte);
4353 pte = pte_modify(old_pte, vma->vm_page_prot);
4354
4355 page = vm_normal_page(vma, vmf->address, pte);
4356 if (!page)
4357 goto out_map;
4358
4359 /* TODO: handle PTE-mapped THP */
4360 if (PageCompound(page))
4361 goto out_map;
4362
4363 /*
4364 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4365 * much anyway since they can be in shared cache state. This misses
4366 * the case where a mapping is writable but the process never writes
4367 * to it but pte_write gets cleared during protection updates and
4368 * pte_dirty has unpredictable behaviour between PTE scan updates,
4369 * background writeback, dirty balancing and application behaviour.
4370 */
4371 if (!was_writable)
4372 flags |= TNF_NO_GROUP;
4373
4374 /*
4375 * Flag if the page is shared between multiple address spaces. This
4376 * is later used when determining whether to group tasks together
4377 */
4378 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4379 flags |= TNF_SHARED;
4380
4381 last_cpupid = page_cpupid_last(page);
4382 page_nid = page_to_nid(page);
4383 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4384 &flags);
4385 if (target_nid == NUMA_NO_NODE) {
4386 put_page(page);
4387 goto out_map;
4388 }
4389 pte_unmap_unlock(vmf->pte, vmf->ptl);
4390
4391 /* Migrate to the requested node */
4392 if (migrate_misplaced_page(page, vma, target_nid)) {
4393 page_nid = target_nid;
4394 flags |= TNF_MIGRATED;
4395 } else {
4396 flags |= TNF_MIGRATE_FAIL;
4397 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4398 spin_lock(vmf->ptl);
4399 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4400 pte_unmap_unlock(vmf->pte, vmf->ptl);
4401 goto out;
4402 }
4403 goto out_map;
4404 }
4405
4406out:
4407 if (page_nid != NUMA_NO_NODE)
4408 task_numa_fault(last_cpupid, page_nid, 1, flags);
4409 return 0;
4410out_map:
4411 /*
4412 * Make it present again, depending on how arch implements
4413 * non-accessible ptes, some can allow access by kernel mode.
4414 */
4415 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4416 pte = pte_modify(old_pte, vma->vm_page_prot);
4417 pte = pte_mkyoung(pte);
4418 if (was_writable)
4419 pte = pte_mkwrite(pte);
4420 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4421 update_mmu_cache(vma, vmf->address, vmf->pte);
4422 pte_unmap_unlock(vmf->pte, vmf->ptl);
4423 goto out;
4424}
4425
4426static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4427{
4428 if (vma_is_anonymous(vmf->vma))
4429 return do_huge_pmd_anonymous_page(vmf);
4430 if (vmf->vma->vm_ops->huge_fault)
4431 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4432 return VM_FAULT_FALLBACK;
4433}
4434
4435/* `inline' is required to avoid gcc 4.1.2 build error */
4436static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4437{
4438 if (vma_is_anonymous(vmf->vma)) {
4439 if (userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd))
4440 return handle_userfault(vmf, VM_UFFD_WP);
4441 return do_huge_pmd_wp_page(vmf);
4442 }
4443 if (vmf->vma->vm_ops->huge_fault) {
4444 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4445
4446 if (!(ret & VM_FAULT_FALLBACK))
4447 return ret;
4448 }
4449
4450 /* COW or write-notify handled on pte level: split pmd. */
4451 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4452
4453 return VM_FAULT_FALLBACK;
4454}
4455
4456static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4457{
4458#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4459 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4460 /* No support for anonymous transparent PUD pages yet */
4461 if (vma_is_anonymous(vmf->vma))
4462 goto split;
4463 if (vmf->vma->vm_ops->huge_fault) {
4464 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4465
4466 if (!(ret & VM_FAULT_FALLBACK))
4467 return ret;
4468 }
4469split:
4470 /* COW or write-notify not handled on PUD level: split pud.*/
4471 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4472#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4473 return VM_FAULT_FALLBACK;
4474}
4475
4476static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4477{
4478#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4479 /* No support for anonymous transparent PUD pages yet */
4480 if (vma_is_anonymous(vmf->vma))
4481 return VM_FAULT_FALLBACK;
4482 if (vmf->vma->vm_ops->huge_fault)
4483 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4484#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4485 return VM_FAULT_FALLBACK;
4486}
4487
4488/*
4489 * These routines also need to handle stuff like marking pages dirty
4490 * and/or accessed for architectures that don't do it in hardware (most
4491 * RISC architectures). The early dirtying is also good on the i386.
4492 *
4493 * There is also a hook called "update_mmu_cache()" that architectures
4494 * with external mmu caches can use to update those (ie the Sparc or
4495 * PowerPC hashed page tables that act as extended TLBs).
4496 *
4497 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4498 * concurrent faults).
4499 *
4500 * The mmap_lock may have been released depending on flags and our return value.
4501 * See filemap_fault() and __lock_page_or_retry().
4502 */
4503static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4504{
4505 pte_t entry;
4506
4507 if (unlikely(pmd_none(*vmf->pmd))) {
4508 /*
4509 * Leave __pte_alloc() until later: because vm_ops->fault may
4510 * want to allocate huge page, and if we expose page table
4511 * for an instant, it will be difficult to retract from
4512 * concurrent faults and from rmap lookups.
4513 */
4514 vmf->pte = NULL;
4515 } else {
4516 /*
4517 * If a huge pmd materialized under us just retry later. Use
4518 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4519 * of pmd_trans_huge() to ensure the pmd didn't become
4520 * pmd_trans_huge under us and then back to pmd_none, as a
4521 * result of MADV_DONTNEED running immediately after a huge pmd
4522 * fault in a different thread of this mm, in turn leading to a
4523 * misleading pmd_trans_huge() retval. All we have to ensure is
4524 * that it is a regular pmd that we can walk with
4525 * pte_offset_map() and we can do that through an atomic read
4526 * in C, which is what pmd_trans_unstable() provides.
4527 */
4528 if (pmd_devmap_trans_unstable(vmf->pmd))
4529 return 0;
4530 /*
4531 * A regular pmd is established and it can't morph into a huge
4532 * pmd from under us anymore at this point because we hold the
4533 * mmap_lock read mode and khugepaged takes it in write mode.
4534 * So now it's safe to run pte_offset_map().
4535 */
4536 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4537 vmf->orig_pte = *vmf->pte;
4538
4539 /*
4540 * some architectures can have larger ptes than wordsize,
4541 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4542 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4543 * accesses. The code below just needs a consistent view
4544 * for the ifs and we later double check anyway with the
4545 * ptl lock held. So here a barrier will do.
4546 */
4547 barrier();
4548 if (pte_none(vmf->orig_pte)) {
4549 pte_unmap(vmf->pte);
4550 vmf->pte = NULL;
4551 }
4552 }
4553
4554 if (!vmf->pte) {
4555 if (vma_is_anonymous(vmf->vma))
4556 return do_anonymous_page(vmf);
4557 else
4558 return do_fault(vmf);
4559 }
4560
4561 if (!pte_present(vmf->orig_pte))
4562 return do_swap_page(vmf);
4563
4564 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4565 return do_numa_page(vmf);
4566
4567 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4568 spin_lock(vmf->ptl);
4569 entry = vmf->orig_pte;
4570 if (unlikely(!pte_same(*vmf->pte, entry))) {
4571 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4572 goto unlock;
4573 }
4574 if (vmf->flags & FAULT_FLAG_WRITE) {
4575 if (!pte_write(entry))
4576 return do_wp_page(vmf);
4577 entry = pte_mkdirty(entry);
4578 }
4579 entry = pte_mkyoung(entry);
4580 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4581 vmf->flags & FAULT_FLAG_WRITE)) {
4582 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4583 } else {
4584 /* Skip spurious TLB flush for retried page fault */
4585 if (vmf->flags & FAULT_FLAG_TRIED)
4586 goto unlock;
4587 /*
4588 * This is needed only for protection faults but the arch code
4589 * is not yet telling us if this is a protection fault or not.
4590 * This still avoids useless tlb flushes for .text page faults
4591 * with threads.
4592 */
4593 if (vmf->flags & FAULT_FLAG_WRITE)
4594 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4595 }
4596unlock:
4597 pte_unmap_unlock(vmf->pte, vmf->ptl);
4598 return 0;
4599}
4600
4601/*
4602 * By the time we get here, we already hold the mm semaphore
4603 *
4604 * The mmap_lock may have been released depending on flags and our
4605 * return value. See filemap_fault() and __lock_page_or_retry().
4606 */
4607static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4608 unsigned long address, unsigned int flags)
4609{
4610 struct vm_fault vmf = {
4611 .vma = vma,
4612 .address = address & PAGE_MASK,
4613 .flags = flags,
4614 .pgoff = linear_page_index(vma, address),
4615 .gfp_mask = __get_fault_gfp_mask(vma),
4616 };
4617 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4618 struct mm_struct *mm = vma->vm_mm;
4619 pgd_t *pgd;
4620 p4d_t *p4d;
4621 vm_fault_t ret;
4622
4623 pgd = pgd_offset(mm, address);
4624 p4d = p4d_alloc(mm, pgd, address);
4625 if (!p4d)
4626 return VM_FAULT_OOM;
4627
4628 vmf.pud = pud_alloc(mm, p4d, address);
4629 if (!vmf.pud)
4630 return VM_FAULT_OOM;
4631retry_pud:
4632 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4633 ret = create_huge_pud(&vmf);
4634 if (!(ret & VM_FAULT_FALLBACK))
4635 return ret;
4636 } else {
4637 pud_t orig_pud = *vmf.pud;
4638
4639 barrier();
4640 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4641
4642 /* NUMA case for anonymous PUDs would go here */
4643
4644 if (dirty && !pud_write(orig_pud)) {
4645 ret = wp_huge_pud(&vmf, orig_pud);
4646 if (!(ret & VM_FAULT_FALLBACK))
4647 return ret;
4648 } else {
4649 huge_pud_set_accessed(&vmf, orig_pud);
4650 return 0;
4651 }
4652 }
4653 }
4654
4655 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4656 if (!vmf.pmd)
4657 return VM_FAULT_OOM;
4658
4659 /* Huge pud page fault raced with pmd_alloc? */
4660 if (pud_trans_unstable(vmf.pud))
4661 goto retry_pud;
4662
4663 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4664 ret = create_huge_pmd(&vmf);
4665 if (!(ret & VM_FAULT_FALLBACK))
4666 return ret;
4667 } else {
4668 vmf.orig_pmd = *vmf.pmd;
4669
4670 barrier();
4671 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
4672 VM_BUG_ON(thp_migration_supported() &&
4673 !is_pmd_migration_entry(vmf.orig_pmd));
4674 if (is_pmd_migration_entry(vmf.orig_pmd))
4675 pmd_migration_entry_wait(mm, vmf.pmd);
4676 return 0;
4677 }
4678 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
4679 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
4680 return do_huge_pmd_numa_page(&vmf);
4681
4682 if (dirty && !pmd_write(vmf.orig_pmd)) {
4683 ret = wp_huge_pmd(&vmf);
4684 if (!(ret & VM_FAULT_FALLBACK))
4685 return ret;
4686 } else {
4687 huge_pmd_set_accessed(&vmf);
4688 return 0;
4689 }
4690 }
4691 }
4692
4693 return handle_pte_fault(&vmf);
4694}
4695
4696/**
4697 * mm_account_fault - Do page fault accounting
4698 *
4699 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4700 * of perf event counters, but we'll still do the per-task accounting to
4701 * the task who triggered this page fault.
4702 * @address: the faulted address.
4703 * @flags: the fault flags.
4704 * @ret: the fault retcode.
4705 *
4706 * This will take care of most of the page fault accounting. Meanwhile, it
4707 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4708 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4709 * still be in per-arch page fault handlers at the entry of page fault.
4710 */
4711static inline void mm_account_fault(struct pt_regs *regs,
4712 unsigned long address, unsigned int flags,
4713 vm_fault_t ret)
4714{
4715 bool major;
4716
4717 /*
4718 * We don't do accounting for some specific faults:
4719 *
4720 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4721 * includes arch_vma_access_permitted() failing before reaching here.
4722 * So this is not a "this many hardware page faults" counter. We
4723 * should use the hw profiling for that.
4724 *
4725 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4726 * once they're completed.
4727 */
4728 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4729 return;
4730
4731 /*
4732 * We define the fault as a major fault when the final successful fault
4733 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4734 * handle it immediately previously).
4735 */
4736 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4737
4738 if (major)
4739 current->maj_flt++;
4740 else
4741 current->min_flt++;
4742
4743 /*
4744 * If the fault is done for GUP, regs will be NULL. We only do the
4745 * accounting for the per thread fault counters who triggered the
4746 * fault, and we skip the perf event updates.
4747 */
4748 if (!regs)
4749 return;
4750
4751 if (major)
4752 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4753 else
4754 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4755}
4756
4757/*
4758 * By the time we get here, we already hold the mm semaphore
4759 *
4760 * The mmap_lock may have been released depending on flags and our
4761 * return value. See filemap_fault() and __lock_page_or_retry().
4762 */
4763vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4764 unsigned int flags, struct pt_regs *regs)
4765{
4766 vm_fault_t ret;
4767
4768 __set_current_state(TASK_RUNNING);
4769
4770 count_vm_event(PGFAULT);
4771 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4772
4773 /* do counter updates before entering really critical section. */
4774 check_sync_rss_stat(current);
4775
4776 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4777 flags & FAULT_FLAG_INSTRUCTION,
4778 flags & FAULT_FLAG_REMOTE))
4779 return VM_FAULT_SIGSEGV;
4780
4781 /*
4782 * Enable the memcg OOM handling for faults triggered in user
4783 * space. Kernel faults are handled more gracefully.
4784 */
4785 if (flags & FAULT_FLAG_USER)
4786 mem_cgroup_enter_user_fault();
4787
4788 if (unlikely(is_vm_hugetlb_page(vma)))
4789 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4790 else
4791 ret = __handle_mm_fault(vma, address, flags);
4792
4793 if (flags & FAULT_FLAG_USER) {
4794 mem_cgroup_exit_user_fault();
4795 /*
4796 * The task may have entered a memcg OOM situation but
4797 * if the allocation error was handled gracefully (no
4798 * VM_FAULT_OOM), there is no need to kill anything.
4799 * Just clean up the OOM state peacefully.
4800 */
4801 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4802 mem_cgroup_oom_synchronize(false);
4803 }
4804
4805 mm_account_fault(regs, address, flags, ret);
4806
4807 return ret;
4808}
4809EXPORT_SYMBOL_GPL(handle_mm_fault);
4810
4811#ifndef __PAGETABLE_P4D_FOLDED
4812/*
4813 * Allocate p4d page table.
4814 * We've already handled the fast-path in-line.
4815 */
4816int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4817{
4818 p4d_t *new = p4d_alloc_one(mm, address);
4819 if (!new)
4820 return -ENOMEM;
4821
4822 smp_wmb(); /* See comment in __pte_alloc */
4823
4824 spin_lock(&mm->page_table_lock);
4825 if (pgd_present(*pgd)) /* Another has populated it */
4826 p4d_free(mm, new);
4827 else
4828 pgd_populate(mm, pgd, new);
4829 spin_unlock(&mm->page_table_lock);
4830 return 0;
4831}
4832#endif /* __PAGETABLE_P4D_FOLDED */
4833
4834#ifndef __PAGETABLE_PUD_FOLDED
4835/*
4836 * Allocate page upper directory.
4837 * We've already handled the fast-path in-line.
4838 */
4839int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4840{
4841 pud_t *new = pud_alloc_one(mm, address);
4842 if (!new)
4843 return -ENOMEM;
4844
4845 smp_wmb(); /* See comment in __pte_alloc */
4846
4847 spin_lock(&mm->page_table_lock);
4848 if (!p4d_present(*p4d)) {
4849 mm_inc_nr_puds(mm);
4850 p4d_populate(mm, p4d, new);
4851 } else /* Another has populated it */
4852 pud_free(mm, new);
4853 spin_unlock(&mm->page_table_lock);
4854 return 0;
4855}
4856#endif /* __PAGETABLE_PUD_FOLDED */
4857
4858#ifndef __PAGETABLE_PMD_FOLDED
4859/*
4860 * Allocate page middle directory.
4861 * We've already handled the fast-path in-line.
4862 */
4863int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4864{
4865 spinlock_t *ptl;
4866 pmd_t *new = pmd_alloc_one(mm, address);
4867 if (!new)
4868 return -ENOMEM;
4869
4870 smp_wmb(); /* See comment in __pte_alloc */
4871
4872 ptl = pud_lock(mm, pud);
4873 if (!pud_present(*pud)) {
4874 mm_inc_nr_pmds(mm);
4875 pud_populate(mm, pud, new);
4876 } else /* Another has populated it */
4877 pmd_free(mm, new);
4878 spin_unlock(ptl);
4879 return 0;
4880}
4881#endif /* __PAGETABLE_PMD_FOLDED */
4882
4883int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4884 struct mmu_notifier_range *range, pte_t **ptepp,
4885 pmd_t **pmdpp, spinlock_t **ptlp)
4886{
4887 pgd_t *pgd;
4888 p4d_t *p4d;
4889 pud_t *pud;
4890 pmd_t *pmd;
4891 pte_t *ptep;
4892
4893 pgd = pgd_offset(mm, address);
4894 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4895 goto out;
4896
4897 p4d = p4d_offset(pgd, address);
4898 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4899 goto out;
4900
4901 pud = pud_offset(p4d, address);
4902 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4903 goto out;
4904
4905 pmd = pmd_offset(pud, address);
4906 VM_BUG_ON(pmd_trans_huge(*pmd));
4907
4908 if (pmd_huge(*pmd)) {
4909 if (!pmdpp)
4910 goto out;
4911
4912 if (range) {
4913 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4914 NULL, mm, address & PMD_MASK,
4915 (address & PMD_MASK) + PMD_SIZE);
4916 mmu_notifier_invalidate_range_start(range);
4917 }
4918 *ptlp = pmd_lock(mm, pmd);
4919 if (pmd_huge(*pmd)) {
4920 *pmdpp = pmd;
4921 return 0;
4922 }
4923 spin_unlock(*ptlp);
4924 if (range)
4925 mmu_notifier_invalidate_range_end(range);
4926 }
4927
4928 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4929 goto out;
4930
4931 if (range) {
4932 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4933 address & PAGE_MASK,
4934 (address & PAGE_MASK) + PAGE_SIZE);
4935 mmu_notifier_invalidate_range_start(range);
4936 }
4937 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4938 if (!pte_present(*ptep))
4939 goto unlock;
4940 *ptepp = ptep;
4941 return 0;
4942unlock:
4943 pte_unmap_unlock(ptep, *ptlp);
4944 if (range)
4945 mmu_notifier_invalidate_range_end(range);
4946out:
4947 return -EINVAL;
4948}
4949
4950/**
4951 * follow_pte - look up PTE at a user virtual address
4952 * @mm: the mm_struct of the target address space
4953 * @address: user virtual address
4954 * @ptepp: location to store found PTE
4955 * @ptlp: location to store the lock for the PTE
4956 *
4957 * On a successful return, the pointer to the PTE is stored in @ptepp;
4958 * the corresponding lock is taken and its location is stored in @ptlp.
4959 * The contents of the PTE are only stable until @ptlp is released;
4960 * any further use, if any, must be protected against invalidation
4961 * with MMU notifiers.
4962 *
4963 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4964 * should be taken for read.
4965 *
4966 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4967 * it is not a good general-purpose API.
4968 *
4969 * Return: zero on success, -ve otherwise.
4970 */
4971int follow_pte(struct mm_struct *mm, unsigned long address,
4972 pte_t **ptepp, spinlock_t **ptlp)
4973{
4974 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4975}
4976EXPORT_SYMBOL_GPL(follow_pte);
4977
4978/**
4979 * follow_pfn - look up PFN at a user virtual address
4980 * @vma: memory mapping
4981 * @address: user virtual address
4982 * @pfn: location to store found PFN
4983 *
4984 * Only IO mappings and raw PFN mappings are allowed.
4985 *
4986 * This function does not allow the caller to read the permissions
4987 * of the PTE. Do not use it.
4988 *
4989 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4990 */
4991int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4992 unsigned long *pfn)
4993{
4994 int ret = -EINVAL;
4995 spinlock_t *ptl;
4996 pte_t *ptep;
4997
4998 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4999 return ret;
5000
5001 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5002 if (ret)
5003 return ret;
5004 *pfn = pte_pfn(*ptep);
5005 pte_unmap_unlock(ptep, ptl);
5006 return 0;
5007}
5008EXPORT_SYMBOL(follow_pfn);
5009
5010#ifdef CONFIG_HAVE_IOREMAP_PROT
5011int follow_phys(struct vm_area_struct *vma,
5012 unsigned long address, unsigned int flags,
5013 unsigned long *prot, resource_size_t *phys)
5014{
5015 int ret = -EINVAL;
5016 pte_t *ptep, pte;
5017 spinlock_t *ptl;
5018
5019 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5020 goto out;
5021
5022 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5023 goto out;
5024 pte = *ptep;
5025
5026 if ((flags & FOLL_WRITE) && !pte_write(pte))
5027 goto unlock;
5028
5029 *prot = pgprot_val(pte_pgprot(pte));
5030 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5031
5032 ret = 0;
5033unlock:
5034 pte_unmap_unlock(ptep, ptl);
5035out:
5036 return ret;
5037}
5038
5039/**
5040 * generic_access_phys - generic implementation for iomem mmap access
5041 * @vma: the vma to access
5042 * @addr: userspace address, not relative offset within @vma
5043 * @buf: buffer to read/write
5044 * @len: length of transfer
5045 * @write: set to FOLL_WRITE when writing, otherwise reading
5046 *
5047 * This is a generic implementation for &vm_operations_struct.access for an
5048 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5049 * not page based.
5050 */
5051int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5052 void *buf, int len, int write)
5053{
5054 resource_size_t phys_addr;
5055 unsigned long prot = 0;
5056 void __iomem *maddr;
5057 pte_t *ptep, pte;
5058 spinlock_t *ptl;
5059 int offset = offset_in_page(addr);
5060 int ret = -EINVAL;
5061
5062 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5063 return -EINVAL;
5064
5065retry:
5066 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5067 return -EINVAL;
5068 pte = *ptep;
5069 pte_unmap_unlock(ptep, ptl);
5070
5071 prot = pgprot_val(pte_pgprot(pte));
5072 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5073
5074 if ((write & FOLL_WRITE) && !pte_write(pte))
5075 return -EINVAL;
5076
5077 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5078 if (!maddr)
5079 return -ENOMEM;
5080
5081 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5082 goto out_unmap;
5083
5084 if (!pte_same(pte, *ptep)) {
5085 pte_unmap_unlock(ptep, ptl);
5086 iounmap(maddr);
5087
5088 goto retry;
5089 }
5090
5091 if (write)
5092 memcpy_toio(maddr + offset, buf, len);
5093 else
5094 memcpy_fromio(buf, maddr + offset, len);
5095 ret = len;
5096 pte_unmap_unlock(ptep, ptl);
5097out_unmap:
5098 iounmap(maddr);
5099
5100 return ret;
5101}
5102EXPORT_SYMBOL_GPL(generic_access_phys);
5103#endif
5104
5105/*
5106 * Access another process' address space as given in mm.
5107 */
5108int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
5109 int len, unsigned int gup_flags)
5110{
5111 struct vm_area_struct *vma;
5112 void *old_buf = buf;
5113 int write = gup_flags & FOLL_WRITE;
5114
5115 if (mmap_read_lock_killable(mm))
5116 return 0;
5117
5118 /* ignore errors, just check how much was successfully transferred */
5119 while (len) {
5120 int bytes, ret, offset;
5121 void *maddr;
5122 struct page *page = NULL;
5123
5124 ret = get_user_pages_remote(mm, addr, 1,
5125 gup_flags, &page, &vma, NULL);
5126 if (ret <= 0) {
5127#ifndef CONFIG_HAVE_IOREMAP_PROT
5128 break;
5129#else
5130 /*
5131 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5132 * we can access using slightly different code.
5133 */
5134 vma = vma_lookup(mm, addr);
5135 if (!vma)
5136 break;
5137 if (vma->vm_ops && vma->vm_ops->access)
5138 ret = vma->vm_ops->access(vma, addr, buf,
5139 len, write);
5140 if (ret <= 0)
5141 break;
5142 bytes = ret;
5143#endif
5144 } else {
5145 bytes = len;
5146 offset = addr & (PAGE_SIZE-1);
5147 if (bytes > PAGE_SIZE-offset)
5148 bytes = PAGE_SIZE-offset;
5149
5150 maddr = kmap(page);
5151 if (write) {
5152 copy_to_user_page(vma, page, addr,
5153 maddr + offset, buf, bytes);
5154 set_page_dirty_lock(page);
5155 } else {
5156 copy_from_user_page(vma, page, addr,
5157 buf, maddr + offset, bytes);
5158 }
5159 kunmap(page);
5160 put_page(page);
5161 }
5162 len -= bytes;
5163 buf += bytes;
5164 addr += bytes;
5165 }
5166 mmap_read_unlock(mm);
5167
5168 return buf - old_buf;
5169}
5170
5171/**
5172 * access_remote_vm - access another process' address space
5173 * @mm: the mm_struct of the target address space
5174 * @addr: start address to access
5175 * @buf: source or destination buffer
5176 * @len: number of bytes to transfer
5177 * @gup_flags: flags modifying lookup behaviour
5178 *
5179 * The caller must hold a reference on @mm.
5180 *
5181 * Return: number of bytes copied from source to destination.
5182 */
5183int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5184 void *buf, int len, unsigned int gup_flags)
5185{
5186 return __access_remote_vm(mm, addr, buf, len, gup_flags);
5187}
5188
5189/*
5190 * Access another process' address space.
5191 * Source/target buffer must be kernel space,
5192 * Do not walk the page table directly, use get_user_pages
5193 */
5194int access_process_vm(struct task_struct *tsk, unsigned long addr,
5195 void *buf, int len, unsigned int gup_flags)
5196{
5197 struct mm_struct *mm;
5198 int ret;
5199
5200 mm = get_task_mm(tsk);
5201 if (!mm)
5202 return 0;
5203
5204 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5205
5206 mmput(mm);
5207
5208 return ret;
5209}
5210EXPORT_SYMBOL_GPL(access_process_vm);
5211
5212/*
5213 * Print the name of a VMA.
5214 */
5215void print_vma_addr(char *prefix, unsigned long ip)
5216{
5217 struct mm_struct *mm = current->mm;
5218 struct vm_area_struct *vma;
5219
5220 /*
5221 * we might be running from an atomic context so we cannot sleep
5222 */
5223 if (!mmap_read_trylock(mm))
5224 return;
5225
5226 vma = find_vma(mm, ip);
5227 if (vma && vma->vm_file) {
5228 struct file *f = vma->vm_file;
5229 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5230 if (buf) {
5231 char *p;
5232
5233 p = file_path(f, buf, PAGE_SIZE);
5234 if (IS_ERR(p))
5235 p = "?";
5236 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5237 vma->vm_start,
5238 vma->vm_end - vma->vm_start);
5239 free_page((unsigned long)buf);
5240 }
5241 }
5242 mmap_read_unlock(mm);
5243}
5244
5245#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5246void __might_fault(const char *file, int line)
5247{
5248 /*
5249 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5250 * holding the mmap_lock, this is safe because kernel memory doesn't
5251 * get paged out, therefore we'll never actually fault, and the
5252 * below annotations will generate false positives.
5253 */
5254 if (uaccess_kernel())
5255 return;
5256 if (pagefault_disabled())
5257 return;
5258 __might_sleep(file, line, 0);
5259#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5260 if (current->mm)
5261 might_lock_read(¤t->mm->mmap_lock);
5262#endif
5263}
5264EXPORT_SYMBOL(__might_fault);
5265#endif
5266
5267#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5268/*
5269 * Process all subpages of the specified huge page with the specified
5270 * operation. The target subpage will be processed last to keep its
5271 * cache lines hot.
5272 */
5273static inline void process_huge_page(
5274 unsigned long addr_hint, unsigned int pages_per_huge_page,
5275 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5276 void *arg)
5277{
5278 int i, n, base, l;
5279 unsigned long addr = addr_hint &
5280 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5281
5282 /* Process target subpage last to keep its cache lines hot */
5283 might_sleep();
5284 n = (addr_hint - addr) / PAGE_SIZE;
5285 if (2 * n <= pages_per_huge_page) {
5286 /* If target subpage in first half of huge page */
5287 base = 0;
5288 l = n;
5289 /* Process subpages at the end of huge page */
5290 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5291 cond_resched();
5292 process_subpage(addr + i * PAGE_SIZE, i, arg);
5293 }
5294 } else {
5295 /* If target subpage in second half of huge page */
5296 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5297 l = pages_per_huge_page - n;
5298 /* Process subpages at the begin of huge page */
5299 for (i = 0; i < base; i++) {
5300 cond_resched();
5301 process_subpage(addr + i * PAGE_SIZE, i, arg);
5302 }
5303 }
5304 /*
5305 * Process remaining subpages in left-right-left-right pattern
5306 * towards the target subpage
5307 */
5308 for (i = 0; i < l; i++) {
5309 int left_idx = base + i;
5310 int right_idx = base + 2 * l - 1 - i;
5311
5312 cond_resched();
5313 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5314 cond_resched();
5315 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5316 }
5317}
5318
5319static void clear_gigantic_page(struct page *page,
5320 unsigned long addr,
5321 unsigned int pages_per_huge_page)
5322{
5323 int i;
5324 struct page *p = page;
5325
5326 might_sleep();
5327 for (i = 0; i < pages_per_huge_page;
5328 i++, p = mem_map_next(p, page, i)) {
5329 cond_resched();
5330 clear_user_highpage(p, addr + i * PAGE_SIZE);
5331 }
5332}
5333
5334static void clear_subpage(unsigned long addr, int idx, void *arg)
5335{
5336 struct page *page = arg;
5337
5338 clear_user_highpage(page + idx, addr);
5339}
5340
5341void clear_huge_page(struct page *page,
5342 unsigned long addr_hint, unsigned int pages_per_huge_page)
5343{
5344 unsigned long addr = addr_hint &
5345 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5346
5347 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5348 clear_gigantic_page(page, addr, pages_per_huge_page);
5349 return;
5350 }
5351
5352 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5353}
5354
5355static void copy_user_gigantic_page(struct page *dst, struct page *src,
5356 unsigned long addr,
5357 struct vm_area_struct *vma,
5358 unsigned int pages_per_huge_page)
5359{
5360 int i;
5361 struct page *dst_base = dst;
5362 struct page *src_base = src;
5363
5364 for (i = 0; i < pages_per_huge_page; ) {
5365 cond_resched();
5366 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5367
5368 i++;
5369 dst = mem_map_next(dst, dst_base, i);
5370 src = mem_map_next(src, src_base, i);
5371 }
5372}
5373
5374struct copy_subpage_arg {
5375 struct page *dst;
5376 struct page *src;
5377 struct vm_area_struct *vma;
5378};
5379
5380static void copy_subpage(unsigned long addr, int idx, void *arg)
5381{
5382 struct copy_subpage_arg *copy_arg = arg;
5383
5384 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5385 addr, copy_arg->vma);
5386}
5387
5388void copy_user_huge_page(struct page *dst, struct page *src,
5389 unsigned long addr_hint, struct vm_area_struct *vma,
5390 unsigned int pages_per_huge_page)
5391{
5392 unsigned long addr = addr_hint &
5393 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5394 struct copy_subpage_arg arg = {
5395 .dst = dst,
5396 .src = src,
5397 .vma = vma,
5398 };
5399
5400 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5401 copy_user_gigantic_page(dst, src, addr, vma,
5402 pages_per_huge_page);
5403 return;
5404 }
5405
5406 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5407}
5408
5409long copy_huge_page_from_user(struct page *dst_page,
5410 const void __user *usr_src,
5411 unsigned int pages_per_huge_page,
5412 bool allow_pagefault)
5413{
5414 void *src = (void *)usr_src;
5415 void *page_kaddr;
5416 unsigned long i, rc = 0;
5417 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5418 struct page *subpage = dst_page;
5419
5420 for (i = 0; i < pages_per_huge_page;
5421 i++, subpage = mem_map_next(subpage, dst_page, i)) {
5422 if (allow_pagefault)
5423 page_kaddr = kmap(subpage);
5424 else
5425 page_kaddr = kmap_atomic(subpage);
5426 rc = copy_from_user(page_kaddr,
5427 (const void __user *)(src + i * PAGE_SIZE),
5428 PAGE_SIZE);
5429 if (allow_pagefault)
5430 kunmap(subpage);
5431 else
5432 kunmap_atomic(page_kaddr);
5433
5434 ret_val -= (PAGE_SIZE - rc);
5435 if (rc)
5436 break;
5437
5438 cond_resched();
5439 }
5440 return ret_val;
5441}
5442#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5443
5444#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5445
5446static struct kmem_cache *page_ptl_cachep;
5447
5448void __init ptlock_cache_init(void)
5449{
5450 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5451 SLAB_PANIC, NULL);
5452}
5453
5454bool ptlock_alloc(struct page *page)
5455{
5456 spinlock_t *ptl;
5457
5458 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5459 if (!ptl)
5460 return false;
5461 page->ptl = ptl;
5462 return true;
5463}
5464
5465void ptlock_free(struct page *page)
5466{
5467 kmem_cache_free(page_ptl_cachep, page->ptl);
5468}
5469#endif