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/export.h>
51#include <linux/delayacct.h>
52#include <linux/init.h>
53#include <linux/writeback.h>
54#include <linux/memcontrol.h>
55#include <linux/mmu_notifier.h>
56#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
59#include <linux/gfp.h>
60#include <linux/migrate.h>
61#include <linux/string.h>
62#include <linux/dma-debug.h>
63#include <linux/debugfs.h>
64
65#include <asm/io.h>
66#include <asm/pgalloc.h>
67#include <asm/uaccess.h>
68#include <asm/tlb.h>
69#include <asm/tlbflush.h>
70#include <asm/pgtable.h>
71
72#include "internal.h"
73
74#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
76#endif
77
78#ifndef CONFIG_NEED_MULTIPLE_NODES
79/* use the per-pgdat data instead for discontigmem - mbligh */
80unsigned long max_mapnr;
81struct page *mem_map;
82
83EXPORT_SYMBOL(max_mapnr);
84EXPORT_SYMBOL(mem_map);
85#endif
86
87/*
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92 * and ZONE_HIGHMEM.
93 */
94void * high_memory;
95
96EXPORT_SYMBOL(high_memory);
97
98/*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104int randomize_va_space __read_mostly =
105#ifdef CONFIG_COMPAT_BRK
106 1;
107#else
108 2;
109#endif
110
111static int __init disable_randmaps(char *s)
112{
113 randomize_va_space = 0;
114 return 1;
115}
116__setup("norandmaps", disable_randmaps);
117
118unsigned long zero_pfn __read_mostly;
119unsigned long highest_memmap_pfn __read_mostly;
120
121/*
122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 */
124static int __init init_zero_pfn(void)
125{
126 zero_pfn = page_to_pfn(ZERO_PAGE(0));
127 return 0;
128}
129core_initcall(init_zero_pfn);
130
131
132#if defined(SPLIT_RSS_COUNTING)
133
134void sync_mm_rss(struct mm_struct *mm)
135{
136 int i;
137
138 for (i = 0; i < NR_MM_COUNTERS; i++) {
139 if (current->rss_stat.count[i]) {
140 add_mm_counter(mm, i, current->rss_stat.count[i]);
141 current->rss_stat.count[i] = 0;
142 }
143 }
144 current->rss_stat.events = 0;
145}
146
147static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148{
149 struct task_struct *task = current;
150
151 if (likely(task->mm == mm))
152 task->rss_stat.count[member] += val;
153 else
154 add_mm_counter(mm, member, val);
155}
156#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159/* sync counter once per 64 page faults */
160#define TASK_RSS_EVENTS_THRESH (64)
161static void check_sync_rss_stat(struct task_struct *task)
162{
163 if (unlikely(task != current))
164 return;
165 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166 sync_mm_rss(task->mm);
167}
168#else /* SPLIT_RSS_COUNTING */
169
170#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173static void check_sync_rss_stat(struct task_struct *task)
174{
175}
176
177#endif /* SPLIT_RSS_COUNTING */
178
179#ifdef HAVE_GENERIC_MMU_GATHER
180
181static int tlb_next_batch(struct mmu_gather *tlb)
182{
183 struct mmu_gather_batch *batch;
184
185 batch = tlb->active;
186 if (batch->next) {
187 tlb->active = batch->next;
188 return 1;
189 }
190
191 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192 return 0;
193
194 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195 if (!batch)
196 return 0;
197
198 tlb->batch_count++;
199 batch->next = NULL;
200 batch->nr = 0;
201 batch->max = MAX_GATHER_BATCH;
202
203 tlb->active->next = batch;
204 tlb->active = batch;
205
206 return 1;
207}
208
209/* tlb_gather_mmu
210 * Called to initialize an (on-stack) mmu_gather structure for page-table
211 * tear-down from @mm. The @fullmm argument is used when @mm is without
212 * users and we're going to destroy the full address space (exit/execve).
213 */
214void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
215{
216 tlb->mm = mm;
217
218 /* Is it from 0 to ~0? */
219 tlb->fullmm = !(start | (end+1));
220 tlb->need_flush_all = 0;
221 tlb->start = start;
222 tlb->end = end;
223 tlb->need_flush = 0;
224 tlb->local.next = NULL;
225 tlb->local.nr = 0;
226 tlb->local.max = ARRAY_SIZE(tlb->__pages);
227 tlb->active = &tlb->local;
228 tlb->batch_count = 0;
229
230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 tlb->batch = NULL;
232#endif
233}
234
235static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
236{
237 tlb->need_flush = 0;
238 tlb_flush(tlb);
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb_table_flush(tlb);
241#endif
242}
243
244static void tlb_flush_mmu_free(struct mmu_gather *tlb)
245{
246 struct mmu_gather_batch *batch;
247
248 for (batch = &tlb->local; batch; batch = batch->next) {
249 free_pages_and_swap_cache(batch->pages, batch->nr);
250 batch->nr = 0;
251 }
252 tlb->active = &tlb->local;
253}
254
255void tlb_flush_mmu(struct mmu_gather *tlb)
256{
257 if (!tlb->need_flush)
258 return;
259 tlb_flush_mmu_tlbonly(tlb);
260 tlb_flush_mmu_free(tlb);
261}
262
263/* tlb_finish_mmu
264 * Called at the end of the shootdown operation to free up any resources
265 * that were required.
266 */
267void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
268{
269 struct mmu_gather_batch *batch, *next;
270
271 tlb_flush_mmu(tlb);
272
273 /* keep the page table cache within bounds */
274 check_pgt_cache();
275
276 for (batch = tlb->local.next; batch; batch = next) {
277 next = batch->next;
278 free_pages((unsigned long)batch, 0);
279 }
280 tlb->local.next = NULL;
281}
282
283/* __tlb_remove_page
284 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
285 * handling the additional races in SMP caused by other CPUs caching valid
286 * mappings in their TLBs. Returns the number of free page slots left.
287 * When out of page slots we must call tlb_flush_mmu().
288 */
289int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
290{
291 struct mmu_gather_batch *batch;
292
293 VM_BUG_ON(!tlb->need_flush);
294
295 batch = tlb->active;
296 batch->pages[batch->nr++] = page;
297 if (batch->nr == batch->max) {
298 if (!tlb_next_batch(tlb))
299 return 0;
300 batch = tlb->active;
301 }
302 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
303
304 return batch->max - batch->nr;
305}
306
307#endif /* HAVE_GENERIC_MMU_GATHER */
308
309#ifdef CONFIG_HAVE_RCU_TABLE_FREE
310
311/*
312 * See the comment near struct mmu_table_batch.
313 */
314
315static void tlb_remove_table_smp_sync(void *arg)
316{
317 /* Simply deliver the interrupt */
318}
319
320static void tlb_remove_table_one(void *table)
321{
322 /*
323 * This isn't an RCU grace period and hence the page-tables cannot be
324 * assumed to be actually RCU-freed.
325 *
326 * It is however sufficient for software page-table walkers that rely on
327 * IRQ disabling. See the comment near struct mmu_table_batch.
328 */
329 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
330 __tlb_remove_table(table);
331}
332
333static void tlb_remove_table_rcu(struct rcu_head *head)
334{
335 struct mmu_table_batch *batch;
336 int i;
337
338 batch = container_of(head, struct mmu_table_batch, rcu);
339
340 for (i = 0; i < batch->nr; i++)
341 __tlb_remove_table(batch->tables[i]);
342
343 free_page((unsigned long)batch);
344}
345
346void tlb_table_flush(struct mmu_gather *tlb)
347{
348 struct mmu_table_batch **batch = &tlb->batch;
349
350 if (*batch) {
351 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
352 *batch = NULL;
353 }
354}
355
356void tlb_remove_table(struct mmu_gather *tlb, void *table)
357{
358 struct mmu_table_batch **batch = &tlb->batch;
359
360 tlb->need_flush = 1;
361
362 /*
363 * When there's less then two users of this mm there cannot be a
364 * concurrent page-table walk.
365 */
366 if (atomic_read(&tlb->mm->mm_users) < 2) {
367 __tlb_remove_table(table);
368 return;
369 }
370
371 if (*batch == NULL) {
372 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
373 if (*batch == NULL) {
374 tlb_remove_table_one(table);
375 return;
376 }
377 (*batch)->nr = 0;
378 }
379 (*batch)->tables[(*batch)->nr++] = table;
380 if ((*batch)->nr == MAX_TABLE_BATCH)
381 tlb_table_flush(tlb);
382}
383
384#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
385
386/*
387 * Note: this doesn't free the actual pages themselves. That
388 * has been handled earlier when unmapping all the memory regions.
389 */
390static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
391 unsigned long addr)
392{
393 pgtable_t token = pmd_pgtable(*pmd);
394 pmd_clear(pmd);
395 pte_free_tlb(tlb, token, addr);
396 atomic_long_dec(&tlb->mm->nr_ptes);
397}
398
399static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
400 unsigned long addr, unsigned long end,
401 unsigned long floor, unsigned long ceiling)
402{
403 pmd_t *pmd;
404 unsigned long next;
405 unsigned long start;
406
407 start = addr;
408 pmd = pmd_offset(pud, addr);
409 do {
410 next = pmd_addr_end(addr, end);
411 if (pmd_none_or_clear_bad(pmd))
412 continue;
413 free_pte_range(tlb, pmd, addr);
414 } while (pmd++, addr = next, addr != end);
415
416 start &= PUD_MASK;
417 if (start < floor)
418 return;
419 if (ceiling) {
420 ceiling &= PUD_MASK;
421 if (!ceiling)
422 return;
423 }
424 if (end - 1 > ceiling - 1)
425 return;
426
427 pmd = pmd_offset(pud, start);
428 pud_clear(pud);
429 pmd_free_tlb(tlb, pmd, start);
430}
431
432static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
433 unsigned long addr, unsigned long end,
434 unsigned long floor, unsigned long ceiling)
435{
436 pud_t *pud;
437 unsigned long next;
438 unsigned long start;
439
440 start = addr;
441 pud = pud_offset(pgd, addr);
442 do {
443 next = pud_addr_end(addr, end);
444 if (pud_none_or_clear_bad(pud))
445 continue;
446 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
447 } while (pud++, addr = next, addr != end);
448
449 start &= PGDIR_MASK;
450 if (start < floor)
451 return;
452 if (ceiling) {
453 ceiling &= PGDIR_MASK;
454 if (!ceiling)
455 return;
456 }
457 if (end - 1 > ceiling - 1)
458 return;
459
460 pud = pud_offset(pgd, start);
461 pgd_clear(pgd);
462 pud_free_tlb(tlb, pud, start);
463}
464
465/*
466 * This function frees user-level page tables of a process.
467 */
468void free_pgd_range(struct mmu_gather *tlb,
469 unsigned long addr, unsigned long end,
470 unsigned long floor, unsigned long ceiling)
471{
472 pgd_t *pgd;
473 unsigned long next;
474
475 /*
476 * The next few lines have given us lots of grief...
477 *
478 * Why are we testing PMD* at this top level? Because often
479 * there will be no work to do at all, and we'd prefer not to
480 * go all the way down to the bottom just to discover that.
481 *
482 * Why all these "- 1"s? Because 0 represents both the bottom
483 * of the address space and the top of it (using -1 for the
484 * top wouldn't help much: the masks would do the wrong thing).
485 * The rule is that addr 0 and floor 0 refer to the bottom of
486 * the address space, but end 0 and ceiling 0 refer to the top
487 * Comparisons need to use "end - 1" and "ceiling - 1" (though
488 * that end 0 case should be mythical).
489 *
490 * Wherever addr is brought up or ceiling brought down, we must
491 * be careful to reject "the opposite 0" before it confuses the
492 * subsequent tests. But what about where end is brought down
493 * by PMD_SIZE below? no, end can't go down to 0 there.
494 *
495 * Whereas we round start (addr) and ceiling down, by different
496 * masks at different levels, in order to test whether a table
497 * now has no other vmas using it, so can be freed, we don't
498 * bother to round floor or end up - the tests don't need that.
499 */
500
501 addr &= PMD_MASK;
502 if (addr < floor) {
503 addr += PMD_SIZE;
504 if (!addr)
505 return;
506 }
507 if (ceiling) {
508 ceiling &= PMD_MASK;
509 if (!ceiling)
510 return;
511 }
512 if (end - 1 > ceiling - 1)
513 end -= PMD_SIZE;
514 if (addr > end - 1)
515 return;
516
517 pgd = pgd_offset(tlb->mm, addr);
518 do {
519 next = pgd_addr_end(addr, end);
520 if (pgd_none_or_clear_bad(pgd))
521 continue;
522 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
523 } while (pgd++, addr = next, addr != end);
524}
525
526void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
527 unsigned long floor, unsigned long ceiling)
528{
529 while (vma) {
530 struct vm_area_struct *next = vma->vm_next;
531 unsigned long addr = vma->vm_start;
532
533 /*
534 * Hide vma from rmap and truncate_pagecache before freeing
535 * pgtables
536 */
537 unlink_anon_vmas(vma);
538 unlink_file_vma(vma);
539
540 if (is_vm_hugetlb_page(vma)) {
541 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
542 floor, next? next->vm_start: ceiling);
543 } else {
544 /*
545 * Optimization: gather nearby vmas into one call down
546 */
547 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
548 && !is_vm_hugetlb_page(next)) {
549 vma = next;
550 next = vma->vm_next;
551 unlink_anon_vmas(vma);
552 unlink_file_vma(vma);
553 }
554 free_pgd_range(tlb, addr, vma->vm_end,
555 floor, next? next->vm_start: ceiling);
556 }
557 vma = next;
558 }
559}
560
561int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
562 pmd_t *pmd, unsigned long address)
563{
564 spinlock_t *ptl;
565 pgtable_t new = pte_alloc_one(mm, address);
566 int wait_split_huge_page;
567 if (!new)
568 return -ENOMEM;
569
570 /*
571 * Ensure all pte setup (eg. pte page lock and page clearing) are
572 * visible before the pte is made visible to other CPUs by being
573 * put into page tables.
574 *
575 * The other side of the story is the pointer chasing in the page
576 * table walking code (when walking the page table without locking;
577 * ie. most of the time). Fortunately, these data accesses consist
578 * of a chain of data-dependent loads, meaning most CPUs (alpha
579 * being the notable exception) will already guarantee loads are
580 * seen in-order. See the alpha page table accessors for the
581 * smp_read_barrier_depends() barriers in page table walking code.
582 */
583 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
584
585 ptl = pmd_lock(mm, pmd);
586 wait_split_huge_page = 0;
587 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
588 atomic_long_inc(&mm->nr_ptes);
589 pmd_populate(mm, pmd, new);
590 new = NULL;
591 } else if (unlikely(pmd_trans_splitting(*pmd)))
592 wait_split_huge_page = 1;
593 spin_unlock(ptl);
594 if (new)
595 pte_free(mm, new);
596 if (wait_split_huge_page)
597 wait_split_huge_page(vma->anon_vma, pmd);
598 return 0;
599}
600
601int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
602{
603 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
604 if (!new)
605 return -ENOMEM;
606
607 smp_wmb(); /* See comment in __pte_alloc */
608
609 spin_lock(&init_mm.page_table_lock);
610 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
611 pmd_populate_kernel(&init_mm, pmd, new);
612 new = NULL;
613 } else
614 VM_BUG_ON(pmd_trans_splitting(*pmd));
615 spin_unlock(&init_mm.page_table_lock);
616 if (new)
617 pte_free_kernel(&init_mm, new);
618 return 0;
619}
620
621static inline void init_rss_vec(int *rss)
622{
623 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
624}
625
626static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
627{
628 int i;
629
630 if (current->mm == mm)
631 sync_mm_rss(mm);
632 for (i = 0; i < NR_MM_COUNTERS; i++)
633 if (rss[i])
634 add_mm_counter(mm, i, rss[i]);
635}
636
637/*
638 * This function is called to print an error when a bad pte
639 * is found. For example, we might have a PFN-mapped pte in
640 * a region that doesn't allow it.
641 *
642 * The calling function must still handle the error.
643 */
644static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
645 pte_t pte, struct page *page)
646{
647 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
648 pud_t *pud = pud_offset(pgd, addr);
649 pmd_t *pmd = pmd_offset(pud, addr);
650 struct address_space *mapping;
651 pgoff_t index;
652 static unsigned long resume;
653 static unsigned long nr_shown;
654 static unsigned long nr_unshown;
655
656 /*
657 * Allow a burst of 60 reports, then keep quiet for that minute;
658 * or allow a steady drip of one report per second.
659 */
660 if (nr_shown == 60) {
661 if (time_before(jiffies, resume)) {
662 nr_unshown++;
663 return;
664 }
665 if (nr_unshown) {
666 printk(KERN_ALERT
667 "BUG: Bad page map: %lu messages suppressed\n",
668 nr_unshown);
669 nr_unshown = 0;
670 }
671 nr_shown = 0;
672 }
673 if (nr_shown++ == 0)
674 resume = jiffies + 60 * HZ;
675
676 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
677 index = linear_page_index(vma, addr);
678
679 printk(KERN_ALERT
680 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
681 current->comm,
682 (long long)pte_val(pte), (long long)pmd_val(*pmd));
683 if (page)
684 dump_page(page, "bad pte");
685 printk(KERN_ALERT
686 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
687 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
688 /*
689 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
690 */
691 if (vma->vm_ops)
692 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
693 vma->vm_ops->fault);
694 if (vma->vm_file)
695 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
696 vma->vm_file->f_op->mmap);
697 dump_stack();
698 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
699}
700
701static inline bool is_cow_mapping(vm_flags_t flags)
702{
703 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
704}
705
706/*
707 * vm_normal_page -- This function gets the "struct page" associated with a pte.
708 *
709 * "Special" mappings do not wish to be associated with a "struct page" (either
710 * it doesn't exist, or it exists but they don't want to touch it). In this
711 * case, NULL is returned here. "Normal" mappings do have a struct page.
712 *
713 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
714 * pte bit, in which case this function is trivial. Secondly, an architecture
715 * may not have a spare pte bit, which requires a more complicated scheme,
716 * described below.
717 *
718 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
719 * special mapping (even if there are underlying and valid "struct pages").
720 * COWed pages of a VM_PFNMAP are always normal.
721 *
722 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
723 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
724 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
725 * mapping will always honor the rule
726 *
727 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
728 *
729 * And for normal mappings this is false.
730 *
731 * This restricts such mappings to be a linear translation from virtual address
732 * to pfn. To get around this restriction, we allow arbitrary mappings so long
733 * as the vma is not a COW mapping; in that case, we know that all ptes are
734 * special (because none can have been COWed).
735 *
736 *
737 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
738 *
739 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
740 * page" backing, however the difference is that _all_ pages with a struct
741 * page (that is, those where pfn_valid is true) are refcounted and considered
742 * normal pages by the VM. The disadvantage is that pages are refcounted
743 * (which can be slower and simply not an option for some PFNMAP users). The
744 * advantage is that we don't have to follow the strict linearity rule of
745 * PFNMAP mappings in order to support COWable mappings.
746 *
747 */
748#ifdef __HAVE_ARCH_PTE_SPECIAL
749# define HAVE_PTE_SPECIAL 1
750#else
751# define HAVE_PTE_SPECIAL 0
752#endif
753struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
754 pte_t pte)
755{
756 unsigned long pfn = pte_pfn(pte);
757
758 if (HAVE_PTE_SPECIAL) {
759 if (likely(!pte_special(pte)))
760 goto check_pfn;
761 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
762 return NULL;
763 if (!is_zero_pfn(pfn))
764 print_bad_pte(vma, addr, pte, NULL);
765 return NULL;
766 }
767
768 /* !HAVE_PTE_SPECIAL case follows: */
769
770 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
771 if (vma->vm_flags & VM_MIXEDMAP) {
772 if (!pfn_valid(pfn))
773 return NULL;
774 goto out;
775 } else {
776 unsigned long off;
777 off = (addr - vma->vm_start) >> PAGE_SHIFT;
778 if (pfn == vma->vm_pgoff + off)
779 return NULL;
780 if (!is_cow_mapping(vma->vm_flags))
781 return NULL;
782 }
783 }
784
785 if (is_zero_pfn(pfn))
786 return NULL;
787check_pfn:
788 if (unlikely(pfn > highest_memmap_pfn)) {
789 print_bad_pte(vma, addr, pte, NULL);
790 return NULL;
791 }
792
793 /*
794 * NOTE! We still have PageReserved() pages in the page tables.
795 * eg. VDSO mappings can cause them to exist.
796 */
797out:
798 return pfn_to_page(pfn);
799}
800
801/*
802 * copy one vm_area from one task to the other. Assumes the page tables
803 * already present in the new task to be cleared in the whole range
804 * covered by this vma.
805 */
806
807static inline unsigned long
808copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
809 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
810 unsigned long addr, int *rss)
811{
812 unsigned long vm_flags = vma->vm_flags;
813 pte_t pte = *src_pte;
814 struct page *page;
815
816 /* pte contains position in swap or file, so copy. */
817 if (unlikely(!pte_present(pte))) {
818 if (!pte_file(pte)) {
819 swp_entry_t entry = pte_to_swp_entry(pte);
820
821 if (swap_duplicate(entry) < 0)
822 return entry.val;
823
824 /* make sure dst_mm is on swapoff's mmlist. */
825 if (unlikely(list_empty(&dst_mm->mmlist))) {
826 spin_lock(&mmlist_lock);
827 if (list_empty(&dst_mm->mmlist))
828 list_add(&dst_mm->mmlist,
829 &src_mm->mmlist);
830 spin_unlock(&mmlist_lock);
831 }
832 if (likely(!non_swap_entry(entry)))
833 rss[MM_SWAPENTS]++;
834 else if (is_migration_entry(entry)) {
835 page = migration_entry_to_page(entry);
836
837 if (PageAnon(page))
838 rss[MM_ANONPAGES]++;
839 else
840 rss[MM_FILEPAGES]++;
841
842 if (is_write_migration_entry(entry) &&
843 is_cow_mapping(vm_flags)) {
844 /*
845 * COW mappings require pages in both
846 * parent and child to be set to read.
847 */
848 make_migration_entry_read(&entry);
849 pte = swp_entry_to_pte(entry);
850 if (pte_swp_soft_dirty(*src_pte))
851 pte = pte_swp_mksoft_dirty(pte);
852 set_pte_at(src_mm, addr, src_pte, pte);
853 }
854 }
855 }
856 goto out_set_pte;
857 }
858
859 /*
860 * If it's a COW mapping, write protect it both
861 * in the parent and the child
862 */
863 if (is_cow_mapping(vm_flags)) {
864 ptep_set_wrprotect(src_mm, addr, src_pte);
865 pte = pte_wrprotect(pte);
866 }
867
868 /*
869 * If it's a shared mapping, mark it clean in
870 * the child
871 */
872 if (vm_flags & VM_SHARED)
873 pte = pte_mkclean(pte);
874 pte = pte_mkold(pte);
875
876 page = vm_normal_page(vma, addr, pte);
877 if (page) {
878 get_page(page);
879 page_dup_rmap(page);
880 if (PageAnon(page))
881 rss[MM_ANONPAGES]++;
882 else
883 rss[MM_FILEPAGES]++;
884 }
885
886out_set_pte:
887 set_pte_at(dst_mm, addr, dst_pte, pte);
888 return 0;
889}
890
891int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
892 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
893 unsigned long addr, unsigned long end)
894{
895 pte_t *orig_src_pte, *orig_dst_pte;
896 pte_t *src_pte, *dst_pte;
897 spinlock_t *src_ptl, *dst_ptl;
898 int progress = 0;
899 int rss[NR_MM_COUNTERS];
900 swp_entry_t entry = (swp_entry_t){0};
901
902again:
903 init_rss_vec(rss);
904
905 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
906 if (!dst_pte)
907 return -ENOMEM;
908 src_pte = pte_offset_map(src_pmd, addr);
909 src_ptl = pte_lockptr(src_mm, src_pmd);
910 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
911 orig_src_pte = src_pte;
912 orig_dst_pte = dst_pte;
913 arch_enter_lazy_mmu_mode();
914
915 do {
916 /*
917 * We are holding two locks at this point - either of them
918 * could generate latencies in another task on another CPU.
919 */
920 if (progress >= 32) {
921 progress = 0;
922 if (need_resched() ||
923 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
924 break;
925 }
926 if (pte_none(*src_pte)) {
927 progress++;
928 continue;
929 }
930 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
931 vma, addr, rss);
932 if (entry.val)
933 break;
934 progress += 8;
935 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
936
937 arch_leave_lazy_mmu_mode();
938 spin_unlock(src_ptl);
939 pte_unmap(orig_src_pte);
940 add_mm_rss_vec(dst_mm, rss);
941 pte_unmap_unlock(orig_dst_pte, dst_ptl);
942 cond_resched();
943
944 if (entry.val) {
945 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
946 return -ENOMEM;
947 progress = 0;
948 }
949 if (addr != end)
950 goto again;
951 return 0;
952}
953
954static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
955 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
956 unsigned long addr, unsigned long end)
957{
958 pmd_t *src_pmd, *dst_pmd;
959 unsigned long next;
960
961 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
962 if (!dst_pmd)
963 return -ENOMEM;
964 src_pmd = pmd_offset(src_pud, addr);
965 do {
966 next = pmd_addr_end(addr, end);
967 if (pmd_trans_huge(*src_pmd)) {
968 int err;
969 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
970 err = copy_huge_pmd(dst_mm, src_mm,
971 dst_pmd, src_pmd, addr, vma);
972 if (err == -ENOMEM)
973 return -ENOMEM;
974 if (!err)
975 continue;
976 /* fall through */
977 }
978 if (pmd_none_or_clear_bad(src_pmd))
979 continue;
980 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
981 vma, addr, next))
982 return -ENOMEM;
983 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
984 return 0;
985}
986
987static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
988 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
989 unsigned long addr, unsigned long end)
990{
991 pud_t *src_pud, *dst_pud;
992 unsigned long next;
993
994 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
995 if (!dst_pud)
996 return -ENOMEM;
997 src_pud = pud_offset(src_pgd, addr);
998 do {
999 next = pud_addr_end(addr, end);
1000 if (pud_none_or_clear_bad(src_pud))
1001 continue;
1002 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1003 vma, addr, next))
1004 return -ENOMEM;
1005 } while (dst_pud++, src_pud++, addr = next, addr != end);
1006 return 0;
1007}
1008
1009int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1010 struct vm_area_struct *vma)
1011{
1012 pgd_t *src_pgd, *dst_pgd;
1013 unsigned long next;
1014 unsigned long addr = vma->vm_start;
1015 unsigned long end = vma->vm_end;
1016 unsigned long mmun_start; /* For mmu_notifiers */
1017 unsigned long mmun_end; /* For mmu_notifiers */
1018 bool is_cow;
1019 int ret;
1020
1021 /*
1022 * Don't copy ptes where a page fault will fill them correctly.
1023 * Fork becomes much lighter when there are big shared or private
1024 * readonly mappings. The tradeoff is that copy_page_range is more
1025 * efficient than faulting.
1026 */
1027 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1028 VM_PFNMAP | VM_MIXEDMAP))) {
1029 if (!vma->anon_vma)
1030 return 0;
1031 }
1032
1033 if (is_vm_hugetlb_page(vma))
1034 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1035
1036 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1037 /*
1038 * We do not free on error cases below as remove_vma
1039 * gets called on error from higher level routine
1040 */
1041 ret = track_pfn_copy(vma);
1042 if (ret)
1043 return ret;
1044 }
1045
1046 /*
1047 * We need to invalidate the secondary MMU mappings only when
1048 * there could be a permission downgrade on the ptes of the
1049 * parent mm. And a permission downgrade will only happen if
1050 * is_cow_mapping() returns true.
1051 */
1052 is_cow = is_cow_mapping(vma->vm_flags);
1053 mmun_start = addr;
1054 mmun_end = end;
1055 if (is_cow)
1056 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1057 mmun_end);
1058
1059 ret = 0;
1060 dst_pgd = pgd_offset(dst_mm, addr);
1061 src_pgd = pgd_offset(src_mm, addr);
1062 do {
1063 next = pgd_addr_end(addr, end);
1064 if (pgd_none_or_clear_bad(src_pgd))
1065 continue;
1066 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1067 vma, addr, next))) {
1068 ret = -ENOMEM;
1069 break;
1070 }
1071 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1072
1073 if (is_cow)
1074 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1075 return ret;
1076}
1077
1078static unsigned long zap_pte_range(struct mmu_gather *tlb,
1079 struct vm_area_struct *vma, pmd_t *pmd,
1080 unsigned long addr, unsigned long end,
1081 struct zap_details *details)
1082{
1083 struct mm_struct *mm = tlb->mm;
1084 int force_flush = 0;
1085 int rss[NR_MM_COUNTERS];
1086 spinlock_t *ptl;
1087 pte_t *start_pte;
1088 pte_t *pte;
1089
1090again:
1091 init_rss_vec(rss);
1092 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1093 pte = start_pte;
1094 arch_enter_lazy_mmu_mode();
1095 do {
1096 pte_t ptent = *pte;
1097 if (pte_none(ptent)) {
1098 continue;
1099 }
1100
1101 if (pte_present(ptent)) {
1102 struct page *page;
1103
1104 page = vm_normal_page(vma, addr, ptent);
1105 if (unlikely(details) && page) {
1106 /*
1107 * unmap_shared_mapping_pages() wants to
1108 * invalidate cache without truncating:
1109 * unmap shared but keep private pages.
1110 */
1111 if (details->check_mapping &&
1112 details->check_mapping != page->mapping)
1113 continue;
1114 /*
1115 * Each page->index must be checked when
1116 * invalidating or truncating nonlinear.
1117 */
1118 if (details->nonlinear_vma &&
1119 (page->index < details->first_index ||
1120 page->index > details->last_index))
1121 continue;
1122 }
1123 ptent = ptep_get_and_clear_full(mm, addr, pte,
1124 tlb->fullmm);
1125 tlb_remove_tlb_entry(tlb, pte, addr);
1126 if (unlikely(!page))
1127 continue;
1128 if (unlikely(details) && details->nonlinear_vma
1129 && linear_page_index(details->nonlinear_vma,
1130 addr) != page->index) {
1131 pte_t ptfile = pgoff_to_pte(page->index);
1132 if (pte_soft_dirty(ptent))
1133 pte_file_mksoft_dirty(ptfile);
1134 set_pte_at(mm, addr, pte, ptfile);
1135 }
1136 if (PageAnon(page))
1137 rss[MM_ANONPAGES]--;
1138 else {
1139 if (pte_dirty(ptent)) {
1140 force_flush = 1;
1141 set_page_dirty(page);
1142 }
1143 if (pte_young(ptent) &&
1144 likely(!(vma->vm_flags & VM_SEQ_READ)))
1145 mark_page_accessed(page);
1146 rss[MM_FILEPAGES]--;
1147 }
1148 page_remove_rmap(page);
1149 if (unlikely(page_mapcount(page) < 0))
1150 print_bad_pte(vma, addr, ptent, page);
1151 if (unlikely(!__tlb_remove_page(tlb, page))) {
1152 force_flush = 1;
1153 break;
1154 }
1155 continue;
1156 }
1157 /*
1158 * If details->check_mapping, we leave swap entries;
1159 * if details->nonlinear_vma, we leave file entries.
1160 */
1161 if (unlikely(details))
1162 continue;
1163 if (pte_file(ptent)) {
1164 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1165 print_bad_pte(vma, addr, ptent, NULL);
1166 } else {
1167 swp_entry_t entry = pte_to_swp_entry(ptent);
1168
1169 if (!non_swap_entry(entry))
1170 rss[MM_SWAPENTS]--;
1171 else if (is_migration_entry(entry)) {
1172 struct page *page;
1173
1174 page = migration_entry_to_page(entry);
1175
1176 if (PageAnon(page))
1177 rss[MM_ANONPAGES]--;
1178 else
1179 rss[MM_FILEPAGES]--;
1180 }
1181 if (unlikely(!free_swap_and_cache(entry)))
1182 print_bad_pte(vma, addr, ptent, NULL);
1183 }
1184 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1185 } while (pte++, addr += PAGE_SIZE, addr != end);
1186
1187 add_mm_rss_vec(mm, rss);
1188 arch_leave_lazy_mmu_mode();
1189
1190 /* Do the actual TLB flush before dropping ptl */
1191 if (force_flush) {
1192 unsigned long old_end;
1193
1194 /*
1195 * Flush the TLB just for the previous segment,
1196 * then update the range to be the remaining
1197 * TLB range.
1198 */
1199 old_end = tlb->end;
1200 tlb->end = addr;
1201 tlb_flush_mmu_tlbonly(tlb);
1202 tlb->start = addr;
1203 tlb->end = old_end;
1204 }
1205 pte_unmap_unlock(start_pte, ptl);
1206
1207 /*
1208 * If we forced a TLB flush (either due to running out of
1209 * batch buffers or because we needed to flush dirty TLB
1210 * entries before releasing the ptl), free the batched
1211 * memory too. Restart if we didn't do everything.
1212 */
1213 if (force_flush) {
1214 force_flush = 0;
1215 tlb_flush_mmu_free(tlb);
1216
1217 if (addr != end)
1218 goto again;
1219 }
1220
1221 return addr;
1222}
1223
1224static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pud_t *pud,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1228{
1229 pmd_t *pmd;
1230 unsigned long next;
1231
1232 pmd = pmd_offset(pud, addr);
1233 do {
1234 next = pmd_addr_end(addr, end);
1235 if (pmd_trans_huge(*pmd)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237#ifdef CONFIG_DEBUG_VM
1238 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240 __func__, addr, end,
1241 vma->vm_start,
1242 vma->vm_end);
1243 BUG();
1244 }
1245#endif
1246 split_huge_page_pmd(vma, addr, pmd);
1247 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1248 goto next;
1249 /* fall through */
1250 }
1251 /*
1252 * Here there can be other concurrent MADV_DONTNEED or
1253 * trans huge page faults running, and if the pmd is
1254 * none or trans huge it can change under us. This is
1255 * because MADV_DONTNEED holds the mmap_sem in read
1256 * mode.
1257 */
1258 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1259 goto next;
1260 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1261next:
1262 cond_resched();
1263 } while (pmd++, addr = next, addr != end);
1264
1265 return addr;
1266}
1267
1268static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1269 struct vm_area_struct *vma, pgd_t *pgd,
1270 unsigned long addr, unsigned long end,
1271 struct zap_details *details)
1272{
1273 pud_t *pud;
1274 unsigned long next;
1275
1276 pud = pud_offset(pgd, addr);
1277 do {
1278 next = pud_addr_end(addr, end);
1279 if (pud_none_or_clear_bad(pud))
1280 continue;
1281 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282 } while (pud++, addr = next, addr != end);
1283
1284 return addr;
1285}
1286
1287static void unmap_page_range(struct mmu_gather *tlb,
1288 struct vm_area_struct *vma,
1289 unsigned long addr, unsigned long end,
1290 struct zap_details *details)
1291{
1292 pgd_t *pgd;
1293 unsigned long next;
1294
1295 if (details && !details->check_mapping && !details->nonlinear_vma)
1296 details = NULL;
1297
1298 BUG_ON(addr >= end);
1299 mem_cgroup_uncharge_start();
1300 tlb_start_vma(tlb, vma);
1301 pgd = pgd_offset(vma->vm_mm, addr);
1302 do {
1303 next = pgd_addr_end(addr, end);
1304 if (pgd_none_or_clear_bad(pgd))
1305 continue;
1306 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307 } while (pgd++, addr = next, addr != end);
1308 tlb_end_vma(tlb, vma);
1309 mem_cgroup_uncharge_end();
1310}
1311
1312
1313static void unmap_single_vma(struct mmu_gather *tlb,
1314 struct vm_area_struct *vma, unsigned long start_addr,
1315 unsigned long end_addr,
1316 struct zap_details *details)
1317{
1318 unsigned long start = max(vma->vm_start, start_addr);
1319 unsigned long end;
1320
1321 if (start >= vma->vm_end)
1322 return;
1323 end = min(vma->vm_end, end_addr);
1324 if (end <= vma->vm_start)
1325 return;
1326
1327 if (vma->vm_file)
1328 uprobe_munmap(vma, start, end);
1329
1330 if (unlikely(vma->vm_flags & VM_PFNMAP))
1331 untrack_pfn(vma, 0, 0);
1332
1333 if (start != end) {
1334 if (unlikely(is_vm_hugetlb_page(vma))) {
1335 /*
1336 * It is undesirable to test vma->vm_file as it
1337 * should be non-null for valid hugetlb area.
1338 * However, vm_file will be NULL in the error
1339 * cleanup path of mmap_region. When
1340 * hugetlbfs ->mmap method fails,
1341 * mmap_region() nullifies vma->vm_file
1342 * before calling this function to clean up.
1343 * Since no pte has actually been setup, it is
1344 * safe to do nothing in this case.
1345 */
1346 if (vma->vm_file) {
1347 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1348 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1349 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1350 }
1351 } else
1352 unmap_page_range(tlb, vma, start, end, details);
1353 }
1354}
1355
1356/**
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
1358 * @tlb: address of the caller's struct mmu_gather
1359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
1362 *
1363 * Unmap all pages in the vma list.
1364 *
1365 * Only addresses between `start' and `end' will be unmapped.
1366 *
1367 * The VMA list must be sorted in ascending virtual address order.
1368 *
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns. So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1373 */
1374void unmap_vmas(struct mmu_gather *tlb,
1375 struct vm_area_struct *vma, unsigned long start_addr,
1376 unsigned long end_addr)
1377{
1378 struct mm_struct *mm = vma->vm_mm;
1379
1380 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1381 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1382 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1383 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1384}
1385
1386/**
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
1389 * @start: starting address of pages to zap
1390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
1392 *
1393 * Caller must protect the VMA list
1394 */
1395void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1396 unsigned long size, struct zap_details *details)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct mmu_gather tlb;
1400 unsigned long end = start + size;
1401
1402 lru_add_drain();
1403 tlb_gather_mmu(&tlb, mm, start, end);
1404 update_hiwater_rss(mm);
1405 mmu_notifier_invalidate_range_start(mm, start, end);
1406 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1407 unmap_single_vma(&tlb, vma, start, end, details);
1408 mmu_notifier_invalidate_range_end(mm, start, end);
1409 tlb_finish_mmu(&tlb, start, end);
1410}
1411
1412/**
1413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1418 *
1419 * The range must fit into one VMA.
1420 */
1421static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422 unsigned long size, struct zap_details *details)
1423{
1424 struct mm_struct *mm = vma->vm_mm;
1425 struct mmu_gather tlb;
1426 unsigned long end = address + size;
1427
1428 lru_add_drain();
1429 tlb_gather_mmu(&tlb, mm, address, end);
1430 update_hiwater_rss(mm);
1431 mmu_notifier_invalidate_range_start(mm, address, end);
1432 unmap_single_vma(&tlb, vma, address, end, details);
1433 mmu_notifier_invalidate_range_end(mm, address, end);
1434 tlb_finish_mmu(&tlb, address, end);
1435}
1436
1437/**
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1442 *
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 *
1445 * The entire address range must be fully contained within the vma.
1446 *
1447 * Returns 0 if successful.
1448 */
1449int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1450 unsigned long size)
1451{
1452 if (address < vma->vm_start || address + size > vma->vm_end ||
1453 !(vma->vm_flags & VM_PFNMAP))
1454 return -1;
1455 zap_page_range_single(vma, address, size, NULL);
1456 return 0;
1457}
1458EXPORT_SYMBOL_GPL(zap_vma_ptes);
1459
1460/**
1461 * follow_page_mask - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1465 * @page_mask: on output, *page_mask is set according to the size of the page
1466 *
1467 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1468 *
1469 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1470 * an error pointer if there is a mapping to something not represented
1471 * by a page descriptor (see also vm_normal_page()).
1472 */
1473struct page *follow_page_mask(struct vm_area_struct *vma,
1474 unsigned long address, unsigned int flags,
1475 unsigned int *page_mask)
1476{
1477 pgd_t *pgd;
1478 pud_t *pud;
1479 pmd_t *pmd;
1480 pte_t *ptep, pte;
1481 spinlock_t *ptl;
1482 struct page *page;
1483 struct mm_struct *mm = vma->vm_mm;
1484
1485 *page_mask = 0;
1486
1487 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1488 if (!IS_ERR(page)) {
1489 BUG_ON(flags & FOLL_GET);
1490 goto out;
1491 }
1492
1493 page = NULL;
1494 pgd = pgd_offset(mm, address);
1495 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1496 goto no_page_table;
1497
1498 pud = pud_offset(pgd, address);
1499 if (pud_none(*pud))
1500 goto no_page_table;
1501 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1502 if (flags & FOLL_GET)
1503 goto out;
1504 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1505 goto out;
1506 }
1507 if (unlikely(pud_bad(*pud)))
1508 goto no_page_table;
1509
1510 pmd = pmd_offset(pud, address);
1511 if (pmd_none(*pmd))
1512 goto no_page_table;
1513 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1514 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1515 if (flags & FOLL_GET) {
1516 /*
1517 * Refcount on tail pages are not well-defined and
1518 * shouldn't be taken. The caller should handle a NULL
1519 * return when trying to follow tail pages.
1520 */
1521 if (PageHead(page))
1522 get_page(page);
1523 else {
1524 page = NULL;
1525 goto out;
1526 }
1527 }
1528 goto out;
1529 }
1530 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1531 goto no_page_table;
1532 if (pmd_trans_huge(*pmd)) {
1533 if (flags & FOLL_SPLIT) {
1534 split_huge_page_pmd(vma, address, pmd);
1535 goto split_fallthrough;
1536 }
1537 ptl = pmd_lock(mm, pmd);
1538 if (likely(pmd_trans_huge(*pmd))) {
1539 if (unlikely(pmd_trans_splitting(*pmd))) {
1540 spin_unlock(ptl);
1541 wait_split_huge_page(vma->anon_vma, pmd);
1542 } else {
1543 page = follow_trans_huge_pmd(vma, address,
1544 pmd, flags);
1545 spin_unlock(ptl);
1546 *page_mask = HPAGE_PMD_NR - 1;
1547 goto out;
1548 }
1549 } else
1550 spin_unlock(ptl);
1551 /* fall through */
1552 }
1553split_fallthrough:
1554 if (unlikely(pmd_bad(*pmd)))
1555 goto no_page_table;
1556
1557 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1558
1559 pte = *ptep;
1560 if (!pte_present(pte)) {
1561 swp_entry_t entry;
1562 /*
1563 * KSM's break_ksm() relies upon recognizing a ksm page
1564 * even while it is being migrated, so for that case we
1565 * need migration_entry_wait().
1566 */
1567 if (likely(!(flags & FOLL_MIGRATION)))
1568 goto no_page;
1569 if (pte_none(pte) || pte_file(pte))
1570 goto no_page;
1571 entry = pte_to_swp_entry(pte);
1572 if (!is_migration_entry(entry))
1573 goto no_page;
1574 pte_unmap_unlock(ptep, ptl);
1575 migration_entry_wait(mm, pmd, address);
1576 goto split_fallthrough;
1577 }
1578 if ((flags & FOLL_NUMA) && pte_numa(pte))
1579 goto no_page;
1580 if ((flags & FOLL_WRITE) && !pte_write(pte))
1581 goto unlock;
1582
1583 page = vm_normal_page(vma, address, pte);
1584 if (unlikely(!page)) {
1585 if ((flags & FOLL_DUMP) ||
1586 !is_zero_pfn(pte_pfn(pte)))
1587 goto bad_page;
1588 page = pte_page(pte);
1589 }
1590
1591 if (flags & FOLL_GET)
1592 get_page_foll(page);
1593 if (flags & FOLL_TOUCH) {
1594 if ((flags & FOLL_WRITE) &&
1595 !pte_dirty(pte) && !PageDirty(page))
1596 set_page_dirty(page);
1597 /*
1598 * pte_mkyoung() would be more correct here, but atomic care
1599 * is needed to avoid losing the dirty bit: it is easier to use
1600 * mark_page_accessed().
1601 */
1602 mark_page_accessed(page);
1603 }
1604 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1605 /*
1606 * The preliminary mapping check is mainly to avoid the
1607 * pointless overhead of lock_page on the ZERO_PAGE
1608 * which might bounce very badly if there is contention.
1609 *
1610 * If the page is already locked, we don't need to
1611 * handle it now - vmscan will handle it later if and
1612 * when it attempts to reclaim the page.
1613 */
1614 if (page->mapping && trylock_page(page)) {
1615 lru_add_drain(); /* push cached pages to LRU */
1616 /*
1617 * Because we lock page here, and migration is
1618 * blocked by the pte's page reference, and we
1619 * know the page is still mapped, we don't even
1620 * need to check for file-cache page truncation.
1621 */
1622 mlock_vma_page(page);
1623 unlock_page(page);
1624 }
1625 }
1626unlock:
1627 pte_unmap_unlock(ptep, ptl);
1628out:
1629 return page;
1630
1631bad_page:
1632 pte_unmap_unlock(ptep, ptl);
1633 return ERR_PTR(-EFAULT);
1634
1635no_page:
1636 pte_unmap_unlock(ptep, ptl);
1637 if (!pte_none(pte))
1638 return page;
1639
1640no_page_table:
1641 /*
1642 * When core dumping an enormous anonymous area that nobody
1643 * has touched so far, we don't want to allocate unnecessary pages or
1644 * page tables. Return error instead of NULL to skip handle_mm_fault,
1645 * then get_dump_page() will return NULL to leave a hole in the dump.
1646 * But we can only make this optimization where a hole would surely
1647 * be zero-filled if handle_mm_fault() actually did handle it.
1648 */
1649 if ((flags & FOLL_DUMP) &&
1650 (!vma->vm_ops || !vma->vm_ops->fault))
1651 return ERR_PTR(-EFAULT);
1652 return page;
1653}
1654
1655static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1656{
1657 return stack_guard_page_start(vma, addr) ||
1658 stack_guard_page_end(vma, addr+PAGE_SIZE);
1659}
1660
1661/**
1662 * __get_user_pages() - pin user pages in memory
1663 * @tsk: task_struct of target task
1664 * @mm: mm_struct of target mm
1665 * @start: starting user address
1666 * @nr_pages: number of pages from start to pin
1667 * @gup_flags: flags modifying pin behaviour
1668 * @pages: array that receives pointers to the pages pinned.
1669 * Should be at least nr_pages long. Or NULL, if caller
1670 * only intends to ensure the pages are faulted in.
1671 * @vmas: array of pointers to vmas corresponding to each page.
1672 * Or NULL if the caller does not require them.
1673 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1674 *
1675 * Returns number of pages pinned. This may be fewer than the number
1676 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1677 * were pinned, returns -errno. Each page returned must be released
1678 * with a put_page() call when it is finished with. vmas will only
1679 * remain valid while mmap_sem is held.
1680 *
1681 * Must be called with mmap_sem held for read or write.
1682 *
1683 * __get_user_pages walks a process's page tables and takes a reference to
1684 * each struct page that each user address corresponds to at a given
1685 * instant. That is, it takes the page that would be accessed if a user
1686 * thread accesses the given user virtual address at that instant.
1687 *
1688 * This does not guarantee that the page exists in the user mappings when
1689 * __get_user_pages returns, and there may even be a completely different
1690 * page there in some cases (eg. if mmapped pagecache has been invalidated
1691 * and subsequently re faulted). However it does guarantee that the page
1692 * won't be freed completely. And mostly callers simply care that the page
1693 * contains data that was valid *at some point in time*. Typically, an IO
1694 * or similar operation cannot guarantee anything stronger anyway because
1695 * locks can't be held over the syscall boundary.
1696 *
1697 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1698 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1699 * appropriate) must be called after the page is finished with, and
1700 * before put_page is called.
1701 *
1702 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1703 * or mmap_sem contention, and if waiting is needed to pin all pages,
1704 * *@nonblocking will be set to 0.
1705 *
1706 * In most cases, get_user_pages or get_user_pages_fast should be used
1707 * instead of __get_user_pages. __get_user_pages should be used only if
1708 * you need some special @gup_flags.
1709 */
1710long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1711 unsigned long start, unsigned long nr_pages,
1712 unsigned int gup_flags, struct page **pages,
1713 struct vm_area_struct **vmas, int *nonblocking)
1714{
1715 long i;
1716 unsigned long vm_flags;
1717 unsigned int page_mask;
1718
1719 if (!nr_pages)
1720 return 0;
1721
1722 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1723
1724 /*
1725 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1726 * would be called on PROT_NONE ranges. We must never invoke
1727 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1728 * page faults would unprotect the PROT_NONE ranges if
1729 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1730 * bitflag. So to avoid that, don't set FOLL_NUMA if
1731 * FOLL_FORCE is set.
1732 */
1733 if (!(gup_flags & FOLL_FORCE))
1734 gup_flags |= FOLL_NUMA;
1735
1736 i = 0;
1737
1738 do {
1739 struct vm_area_struct *vma;
1740
1741 vma = find_extend_vma(mm, start);
1742 if (!vma && in_gate_area(mm, start)) {
1743 unsigned long pg = start & PAGE_MASK;
1744 pgd_t *pgd;
1745 pud_t *pud;
1746 pmd_t *pmd;
1747 pte_t *pte;
1748
1749 /* user gate pages are read-only */
1750 if (gup_flags & FOLL_WRITE)
1751 goto efault;
1752 if (pg > TASK_SIZE)
1753 pgd = pgd_offset_k(pg);
1754 else
1755 pgd = pgd_offset_gate(mm, pg);
1756 BUG_ON(pgd_none(*pgd));
1757 pud = pud_offset(pgd, pg);
1758 BUG_ON(pud_none(*pud));
1759 pmd = pmd_offset(pud, pg);
1760 if (pmd_none(*pmd))
1761 goto efault;
1762 VM_BUG_ON(pmd_trans_huge(*pmd));
1763 pte = pte_offset_map(pmd, pg);
1764 if (pte_none(*pte)) {
1765 pte_unmap(pte);
1766 goto efault;
1767 }
1768 vma = get_gate_vma(mm);
1769 if (pages) {
1770 struct page *page;
1771
1772 page = vm_normal_page(vma, start, *pte);
1773 if (!page) {
1774 if (!(gup_flags & FOLL_DUMP) &&
1775 is_zero_pfn(pte_pfn(*pte)))
1776 page = pte_page(*pte);
1777 else {
1778 pte_unmap(pte);
1779 goto efault;
1780 }
1781 }
1782 pages[i] = page;
1783 get_page(page);
1784 }
1785 pte_unmap(pte);
1786 page_mask = 0;
1787 goto next_page;
1788 }
1789
1790 if (!vma)
1791 goto efault;
1792 vm_flags = vma->vm_flags;
1793 if (vm_flags & (VM_IO | VM_PFNMAP))
1794 goto efault;
1795
1796 if (gup_flags & FOLL_WRITE) {
1797 if (!(vm_flags & VM_WRITE)) {
1798 if (!(gup_flags & FOLL_FORCE))
1799 goto efault;
1800 /*
1801 * We used to let the write,force case do COW
1802 * in a VM_MAYWRITE VM_SHARED !VM_WRITE vma, so
1803 * ptrace could set a breakpoint in a read-only
1804 * mapping of an executable, without corrupting
1805 * the file (yet only when that file had been
1806 * opened for writing!). Anon pages in shared
1807 * mappings are surprising: now just reject it.
1808 */
1809 if (!is_cow_mapping(vm_flags)) {
1810 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
1811 goto efault;
1812 }
1813 }
1814 } else {
1815 if (!(vm_flags & VM_READ)) {
1816 if (!(gup_flags & FOLL_FORCE))
1817 goto efault;
1818 /*
1819 * Is there actually any vma we can reach here
1820 * which does not have VM_MAYREAD set?
1821 */
1822 if (!(vm_flags & VM_MAYREAD))
1823 goto efault;
1824 }
1825 }
1826
1827 if (is_vm_hugetlb_page(vma)) {
1828 i = follow_hugetlb_page(mm, vma, pages, vmas,
1829 &start, &nr_pages, i, gup_flags);
1830 continue;
1831 }
1832
1833 do {
1834 struct page *page;
1835 unsigned int foll_flags = gup_flags;
1836 unsigned int page_increm;
1837
1838 /*
1839 * If we have a pending SIGKILL, don't keep faulting
1840 * pages and potentially allocating memory.
1841 */
1842 if (unlikely(fatal_signal_pending(current)))
1843 return i ? i : -ERESTARTSYS;
1844
1845 cond_resched();
1846 while (!(page = follow_page_mask(vma, start,
1847 foll_flags, &page_mask))) {
1848 int ret;
1849 unsigned int fault_flags = 0;
1850
1851 /* For mlock, just skip the stack guard page. */
1852 if (foll_flags & FOLL_MLOCK) {
1853 if (stack_guard_page(vma, start))
1854 goto next_page;
1855 }
1856 if (foll_flags & FOLL_WRITE)
1857 fault_flags |= FAULT_FLAG_WRITE;
1858 if (nonblocking)
1859 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1860 if (foll_flags & FOLL_NOWAIT)
1861 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1862
1863 ret = handle_mm_fault(mm, vma, start,
1864 fault_flags);
1865
1866 if (ret & VM_FAULT_ERROR) {
1867 if (ret & VM_FAULT_OOM)
1868 return i ? i : -ENOMEM;
1869 if (ret & (VM_FAULT_HWPOISON |
1870 VM_FAULT_HWPOISON_LARGE)) {
1871 if (i)
1872 return i;
1873 else if (gup_flags & FOLL_HWPOISON)
1874 return -EHWPOISON;
1875 else
1876 return -EFAULT;
1877 }
1878 if (ret & VM_FAULT_SIGBUS)
1879 goto efault;
1880 BUG();
1881 }
1882
1883 if (tsk) {
1884 if (ret & VM_FAULT_MAJOR)
1885 tsk->maj_flt++;
1886 else
1887 tsk->min_flt++;
1888 }
1889
1890 if (ret & VM_FAULT_RETRY) {
1891 if (nonblocking)
1892 *nonblocking = 0;
1893 return i;
1894 }
1895
1896 /*
1897 * The VM_FAULT_WRITE bit tells us that
1898 * do_wp_page has broken COW when necessary,
1899 * even if maybe_mkwrite decided not to set
1900 * pte_write. We can thus safely do subsequent
1901 * page lookups as if they were reads. But only
1902 * do so when looping for pte_write is futile:
1903 * in some cases userspace may also be wanting
1904 * to write to the gotten user page, which a
1905 * read fault here might prevent (a readonly
1906 * page might get reCOWed by userspace write).
1907 */
1908 if ((ret & VM_FAULT_WRITE) &&
1909 !(vma->vm_flags & VM_WRITE))
1910 foll_flags &= ~FOLL_WRITE;
1911
1912 cond_resched();
1913 }
1914 if (IS_ERR(page))
1915 return i ? i : PTR_ERR(page);
1916 if (pages) {
1917 pages[i] = page;
1918
1919 flush_anon_page(vma, page, start);
1920 flush_dcache_page(page);
1921 page_mask = 0;
1922 }
1923next_page:
1924 if (vmas) {
1925 vmas[i] = vma;
1926 page_mask = 0;
1927 }
1928 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1929 if (page_increm > nr_pages)
1930 page_increm = nr_pages;
1931 i += page_increm;
1932 start += page_increm * PAGE_SIZE;
1933 nr_pages -= page_increm;
1934 } while (nr_pages && start < vma->vm_end);
1935 } while (nr_pages);
1936 return i;
1937efault:
1938 return i ? : -EFAULT;
1939}
1940EXPORT_SYMBOL(__get_user_pages);
1941
1942/*
1943 * fixup_user_fault() - manually resolve a user page fault
1944 * @tsk: the task_struct to use for page fault accounting, or
1945 * NULL if faults are not to be recorded.
1946 * @mm: mm_struct of target mm
1947 * @address: user address
1948 * @fault_flags:flags to pass down to handle_mm_fault()
1949 *
1950 * This is meant to be called in the specific scenario where for locking reasons
1951 * we try to access user memory in atomic context (within a pagefault_disable()
1952 * section), this returns -EFAULT, and we want to resolve the user fault before
1953 * trying again.
1954 *
1955 * Typically this is meant to be used by the futex code.
1956 *
1957 * The main difference with get_user_pages() is that this function will
1958 * unconditionally call handle_mm_fault() which will in turn perform all the
1959 * necessary SW fixup of the dirty and young bits in the PTE, while
1960 * handle_mm_fault() only guarantees to update these in the struct page.
1961 *
1962 * This is important for some architectures where those bits also gate the
1963 * access permission to the page because they are maintained in software. On
1964 * such architectures, gup() will not be enough to make a subsequent access
1965 * succeed.
1966 *
1967 * This should be called with the mm_sem held for read.
1968 */
1969int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1970 unsigned long address, unsigned int fault_flags)
1971{
1972 struct vm_area_struct *vma;
1973 vm_flags_t vm_flags;
1974 int ret;
1975
1976 vma = find_extend_vma(mm, address);
1977 if (!vma || address < vma->vm_start)
1978 return -EFAULT;
1979
1980 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1981 if (!(vm_flags & vma->vm_flags))
1982 return -EFAULT;
1983
1984 ret = handle_mm_fault(mm, vma, address, fault_flags);
1985 if (ret & VM_FAULT_ERROR) {
1986 if (ret & VM_FAULT_OOM)
1987 return -ENOMEM;
1988 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1989 return -EHWPOISON;
1990 if (ret & VM_FAULT_SIGBUS)
1991 return -EFAULT;
1992 BUG();
1993 }
1994 if (tsk) {
1995 if (ret & VM_FAULT_MAJOR)
1996 tsk->maj_flt++;
1997 else
1998 tsk->min_flt++;
1999 }
2000 return 0;
2001}
2002
2003/*
2004 * get_user_pages() - pin user pages in memory
2005 * @tsk: the task_struct to use for page fault accounting, or
2006 * NULL if faults are not to be recorded.
2007 * @mm: mm_struct of target mm
2008 * @start: starting user address
2009 * @nr_pages: number of pages from start to pin
2010 * @write: whether pages will be written to by the caller
2011 * @force: whether to force access even when user mapping is currently
2012 * protected (but never forces write access to shared mapping).
2013 * @pages: array that receives pointers to the pages pinned.
2014 * Should be at least nr_pages long. Or NULL, if caller
2015 * only intends to ensure the pages are faulted in.
2016 * @vmas: array of pointers to vmas corresponding to each page.
2017 * Or NULL if the caller does not require them.
2018 *
2019 * Returns number of pages pinned. This may be fewer than the number
2020 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2021 * were pinned, returns -errno. Each page returned must be released
2022 * with a put_page() call when it is finished with. vmas will only
2023 * remain valid while mmap_sem is held.
2024 *
2025 * Must be called with mmap_sem held for read or write.
2026 *
2027 * get_user_pages walks a process's page tables and takes a reference to
2028 * each struct page that each user address corresponds to at a given
2029 * instant. That is, it takes the page that would be accessed if a user
2030 * thread accesses the given user virtual address at that instant.
2031 *
2032 * This does not guarantee that the page exists in the user mappings when
2033 * get_user_pages returns, and there may even be a completely different
2034 * page there in some cases (eg. if mmapped pagecache has been invalidated
2035 * and subsequently re faulted). However it does guarantee that the page
2036 * won't be freed completely. And mostly callers simply care that the page
2037 * contains data that was valid *at some point in time*. Typically, an IO
2038 * or similar operation cannot guarantee anything stronger anyway because
2039 * locks can't be held over the syscall boundary.
2040 *
2041 * If write=0, the page must not be written to. If the page is written to,
2042 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2043 * after the page is finished with, and before put_page is called.
2044 *
2045 * get_user_pages is typically used for fewer-copy IO operations, to get a
2046 * handle on the memory by some means other than accesses via the user virtual
2047 * addresses. The pages may be submitted for DMA to devices or accessed via
2048 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2049 * use the correct cache flushing APIs.
2050 *
2051 * See also get_user_pages_fast, for performance critical applications.
2052 */
2053long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2054 unsigned long start, unsigned long nr_pages, int write,
2055 int force, struct page **pages, struct vm_area_struct **vmas)
2056{
2057 int flags = FOLL_TOUCH;
2058
2059 if (pages)
2060 flags |= FOLL_GET;
2061 if (write)
2062 flags |= FOLL_WRITE;
2063 if (force)
2064 flags |= FOLL_FORCE;
2065
2066 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2067 NULL);
2068}
2069EXPORT_SYMBOL(get_user_pages);
2070
2071/**
2072 * get_dump_page() - pin user page in memory while writing it to core dump
2073 * @addr: user address
2074 *
2075 * Returns struct page pointer of user page pinned for dump,
2076 * to be freed afterwards by page_cache_release() or put_page().
2077 *
2078 * Returns NULL on any kind of failure - a hole must then be inserted into
2079 * the corefile, to preserve alignment with its headers; and also returns
2080 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2081 * allowing a hole to be left in the corefile to save diskspace.
2082 *
2083 * Called without mmap_sem, but after all other threads have been killed.
2084 */
2085#ifdef CONFIG_ELF_CORE
2086struct page *get_dump_page(unsigned long addr)
2087{
2088 struct vm_area_struct *vma;
2089 struct page *page;
2090
2091 if (__get_user_pages(current, current->mm, addr, 1,
2092 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2093 NULL) < 1)
2094 return NULL;
2095 flush_cache_page(vma, addr, page_to_pfn(page));
2096 return page;
2097}
2098#endif /* CONFIG_ELF_CORE */
2099
2100pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2101 spinlock_t **ptl)
2102{
2103 pgd_t * pgd = pgd_offset(mm, addr);
2104 pud_t * pud = pud_alloc(mm, pgd, addr);
2105 if (pud) {
2106 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2107 if (pmd) {
2108 VM_BUG_ON(pmd_trans_huge(*pmd));
2109 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2110 }
2111 }
2112 return NULL;
2113}
2114
2115/*
2116 * This is the old fallback for page remapping.
2117 *
2118 * For historical reasons, it only allows reserved pages. Only
2119 * old drivers should use this, and they needed to mark their
2120 * pages reserved for the old functions anyway.
2121 */
2122static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2123 struct page *page, pgprot_t prot)
2124{
2125 struct mm_struct *mm = vma->vm_mm;
2126 int retval;
2127 pte_t *pte;
2128 spinlock_t *ptl;
2129
2130 retval = -EINVAL;
2131 if (PageAnon(page))
2132 goto out;
2133 retval = -ENOMEM;
2134 flush_dcache_page(page);
2135 pte = get_locked_pte(mm, addr, &ptl);
2136 if (!pte)
2137 goto out;
2138 retval = -EBUSY;
2139 if (!pte_none(*pte))
2140 goto out_unlock;
2141
2142 /* Ok, finally just insert the thing.. */
2143 get_page(page);
2144 inc_mm_counter_fast(mm, MM_FILEPAGES);
2145 page_add_file_rmap(page);
2146 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2147
2148 retval = 0;
2149 pte_unmap_unlock(pte, ptl);
2150 return retval;
2151out_unlock:
2152 pte_unmap_unlock(pte, ptl);
2153out:
2154 return retval;
2155}
2156
2157/**
2158 * vm_insert_page - insert single page into user vma
2159 * @vma: user vma to map to
2160 * @addr: target user address of this page
2161 * @page: source kernel page
2162 *
2163 * This allows drivers to insert individual pages they've allocated
2164 * into a user vma.
2165 *
2166 * The page has to be a nice clean _individual_ kernel allocation.
2167 * If you allocate a compound page, you need to have marked it as
2168 * such (__GFP_COMP), or manually just split the page up yourself
2169 * (see split_page()).
2170 *
2171 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2172 * took an arbitrary page protection parameter. This doesn't allow
2173 * that. Your vma protection will have to be set up correctly, which
2174 * means that if you want a shared writable mapping, you'd better
2175 * ask for a shared writable mapping!
2176 *
2177 * The page does not need to be reserved.
2178 *
2179 * Usually this function is called from f_op->mmap() handler
2180 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2181 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2182 * function from other places, for example from page-fault handler.
2183 */
2184int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2185 struct page *page)
2186{
2187 if (addr < vma->vm_start || addr >= vma->vm_end)
2188 return -EFAULT;
2189 if (!page_count(page))
2190 return -EINVAL;
2191 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2192 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2193 BUG_ON(vma->vm_flags & VM_PFNMAP);
2194 vma->vm_flags |= VM_MIXEDMAP;
2195 }
2196 return insert_page(vma, addr, page, vma->vm_page_prot);
2197}
2198EXPORT_SYMBOL(vm_insert_page);
2199
2200static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2201 unsigned long pfn, pgprot_t prot)
2202{
2203 struct mm_struct *mm = vma->vm_mm;
2204 int retval;
2205 pte_t *pte, entry;
2206 spinlock_t *ptl;
2207
2208 retval = -ENOMEM;
2209 pte = get_locked_pte(mm, addr, &ptl);
2210 if (!pte)
2211 goto out;
2212 retval = -EBUSY;
2213 if (!pte_none(*pte))
2214 goto out_unlock;
2215
2216 /* Ok, finally just insert the thing.. */
2217 entry = pte_mkspecial(pfn_pte(pfn, prot));
2218 set_pte_at(mm, addr, pte, entry);
2219 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2220
2221 retval = 0;
2222out_unlock:
2223 pte_unmap_unlock(pte, ptl);
2224out:
2225 return retval;
2226}
2227
2228/**
2229 * vm_insert_pfn - insert single pfn into user vma
2230 * @vma: user vma to map to
2231 * @addr: target user address of this page
2232 * @pfn: source kernel pfn
2233 *
2234 * Similar to vm_insert_page, this allows drivers to insert individual pages
2235 * they've allocated into a user vma. Same comments apply.
2236 *
2237 * This function should only be called from a vm_ops->fault handler, and
2238 * in that case the handler should return NULL.
2239 *
2240 * vma cannot be a COW mapping.
2241 *
2242 * As this is called only for pages that do not currently exist, we
2243 * do not need to flush old virtual caches or the TLB.
2244 */
2245int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2246 unsigned long pfn)
2247{
2248 int ret;
2249 pgprot_t pgprot = vma->vm_page_prot;
2250 /*
2251 * Technically, architectures with pte_special can avoid all these
2252 * restrictions (same for remap_pfn_range). However we would like
2253 * consistency in testing and feature parity among all, so we should
2254 * try to keep these invariants in place for everybody.
2255 */
2256 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2257 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2258 (VM_PFNMAP|VM_MIXEDMAP));
2259 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2260 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2261
2262 if (addr < vma->vm_start || addr >= vma->vm_end)
2263 return -EFAULT;
2264 if (track_pfn_insert(vma, &pgprot, pfn))
2265 return -EINVAL;
2266
2267 ret = insert_pfn(vma, addr, pfn, pgprot);
2268
2269 return ret;
2270}
2271EXPORT_SYMBOL(vm_insert_pfn);
2272
2273int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2274 unsigned long pfn)
2275{
2276 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2277
2278 if (addr < vma->vm_start || addr >= vma->vm_end)
2279 return -EFAULT;
2280
2281 /*
2282 * If we don't have pte special, then we have to use the pfn_valid()
2283 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2284 * refcount the page if pfn_valid is true (hence insert_page rather
2285 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2286 * without pte special, it would there be refcounted as a normal page.
2287 */
2288 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2289 struct page *page;
2290
2291 page = pfn_to_page(pfn);
2292 return insert_page(vma, addr, page, vma->vm_page_prot);
2293 }
2294 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2295}
2296EXPORT_SYMBOL(vm_insert_mixed);
2297
2298/*
2299 * maps a range of physical memory into the requested pages. the old
2300 * mappings are removed. any references to nonexistent pages results
2301 * in null mappings (currently treated as "copy-on-access")
2302 */
2303static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2304 unsigned long addr, unsigned long end,
2305 unsigned long pfn, pgprot_t prot)
2306{
2307 pte_t *pte;
2308 spinlock_t *ptl;
2309
2310 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2311 if (!pte)
2312 return -ENOMEM;
2313 arch_enter_lazy_mmu_mode();
2314 do {
2315 BUG_ON(!pte_none(*pte));
2316 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2317 pfn++;
2318 } while (pte++, addr += PAGE_SIZE, addr != end);
2319 arch_leave_lazy_mmu_mode();
2320 pte_unmap_unlock(pte - 1, ptl);
2321 return 0;
2322}
2323
2324static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2325 unsigned long addr, unsigned long end,
2326 unsigned long pfn, pgprot_t prot)
2327{
2328 pmd_t *pmd;
2329 unsigned long next;
2330
2331 pfn -= addr >> PAGE_SHIFT;
2332 pmd = pmd_alloc(mm, pud, addr);
2333 if (!pmd)
2334 return -ENOMEM;
2335 VM_BUG_ON(pmd_trans_huge(*pmd));
2336 do {
2337 next = pmd_addr_end(addr, end);
2338 if (remap_pte_range(mm, pmd, addr, next,
2339 pfn + (addr >> PAGE_SHIFT), prot))
2340 return -ENOMEM;
2341 } while (pmd++, addr = next, addr != end);
2342 return 0;
2343}
2344
2345static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2346 unsigned long addr, unsigned long end,
2347 unsigned long pfn, pgprot_t prot)
2348{
2349 pud_t *pud;
2350 unsigned long next;
2351
2352 pfn -= addr >> PAGE_SHIFT;
2353 pud = pud_alloc(mm, pgd, addr);
2354 if (!pud)
2355 return -ENOMEM;
2356 do {
2357 next = pud_addr_end(addr, end);
2358 if (remap_pmd_range(mm, pud, addr, next,
2359 pfn + (addr >> PAGE_SHIFT), prot))
2360 return -ENOMEM;
2361 } while (pud++, addr = next, addr != end);
2362 return 0;
2363}
2364
2365/**
2366 * remap_pfn_range - remap kernel memory to userspace
2367 * @vma: user vma to map to
2368 * @addr: target user address to start at
2369 * @pfn: physical address of kernel memory
2370 * @size: size of map area
2371 * @prot: page protection flags for this mapping
2372 *
2373 * Note: this is only safe if the mm semaphore is held when called.
2374 */
2375int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2376 unsigned long pfn, unsigned long size, pgprot_t prot)
2377{
2378 pgd_t *pgd;
2379 unsigned long next;
2380 unsigned long end = addr + PAGE_ALIGN(size);
2381 struct mm_struct *mm = vma->vm_mm;
2382 int err;
2383
2384 /*
2385 * Physically remapped pages are special. Tell the
2386 * rest of the world about it:
2387 * VM_IO tells people not to look at these pages
2388 * (accesses can have side effects).
2389 * VM_PFNMAP tells the core MM that the base pages are just
2390 * raw PFN mappings, and do not have a "struct page" associated
2391 * with them.
2392 * VM_DONTEXPAND
2393 * Disable vma merging and expanding with mremap().
2394 * VM_DONTDUMP
2395 * Omit vma from core dump, even when VM_IO turned off.
2396 *
2397 * There's a horrible special case to handle copy-on-write
2398 * behaviour that some programs depend on. We mark the "original"
2399 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2400 * See vm_normal_page() for details.
2401 */
2402 if (is_cow_mapping(vma->vm_flags)) {
2403 if (addr != vma->vm_start || end != vma->vm_end)
2404 return -EINVAL;
2405 vma->vm_pgoff = pfn;
2406 }
2407
2408 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2409 if (err)
2410 return -EINVAL;
2411
2412 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2413
2414 BUG_ON(addr >= end);
2415 pfn -= addr >> PAGE_SHIFT;
2416 pgd = pgd_offset(mm, addr);
2417 flush_cache_range(vma, addr, end);
2418 do {
2419 next = pgd_addr_end(addr, end);
2420 err = remap_pud_range(mm, pgd, addr, next,
2421 pfn + (addr >> PAGE_SHIFT), prot);
2422 if (err)
2423 break;
2424 } while (pgd++, addr = next, addr != end);
2425
2426 if (err)
2427 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2428
2429 return err;
2430}
2431EXPORT_SYMBOL(remap_pfn_range);
2432
2433/**
2434 * vm_iomap_memory - remap memory to userspace
2435 * @vma: user vma to map to
2436 * @start: start of area
2437 * @len: size of area
2438 *
2439 * This is a simplified io_remap_pfn_range() for common driver use. The
2440 * driver just needs to give us the physical memory range to be mapped,
2441 * we'll figure out the rest from the vma information.
2442 *
2443 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2444 * whatever write-combining details or similar.
2445 */
2446int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2447{
2448 unsigned long vm_len, pfn, pages;
2449
2450 /* Check that the physical memory area passed in looks valid */
2451 if (start + len < start)
2452 return -EINVAL;
2453 /*
2454 * You *really* shouldn't map things that aren't page-aligned,
2455 * but we've historically allowed it because IO memory might
2456 * just have smaller alignment.
2457 */
2458 len += start & ~PAGE_MASK;
2459 pfn = start >> PAGE_SHIFT;
2460 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2461 if (pfn + pages < pfn)
2462 return -EINVAL;
2463
2464 /* We start the mapping 'vm_pgoff' pages into the area */
2465 if (vma->vm_pgoff > pages)
2466 return -EINVAL;
2467 pfn += vma->vm_pgoff;
2468 pages -= vma->vm_pgoff;
2469
2470 /* Can we fit all of the mapping? */
2471 vm_len = vma->vm_end - vma->vm_start;
2472 if (vm_len >> PAGE_SHIFT > pages)
2473 return -EINVAL;
2474
2475 /* Ok, let it rip */
2476 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2477}
2478EXPORT_SYMBOL(vm_iomap_memory);
2479
2480static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2481 unsigned long addr, unsigned long end,
2482 pte_fn_t fn, void *data)
2483{
2484 pte_t *pte;
2485 int err;
2486 pgtable_t token;
2487 spinlock_t *uninitialized_var(ptl);
2488
2489 pte = (mm == &init_mm) ?
2490 pte_alloc_kernel(pmd, addr) :
2491 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2492 if (!pte)
2493 return -ENOMEM;
2494
2495 BUG_ON(pmd_huge(*pmd));
2496
2497 arch_enter_lazy_mmu_mode();
2498
2499 token = pmd_pgtable(*pmd);
2500
2501 do {
2502 err = fn(pte++, token, addr, data);
2503 if (err)
2504 break;
2505 } while (addr += PAGE_SIZE, addr != end);
2506
2507 arch_leave_lazy_mmu_mode();
2508
2509 if (mm != &init_mm)
2510 pte_unmap_unlock(pte-1, ptl);
2511 return err;
2512}
2513
2514static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2515 unsigned long addr, unsigned long end,
2516 pte_fn_t fn, void *data)
2517{
2518 pmd_t *pmd;
2519 unsigned long next;
2520 int err;
2521
2522 BUG_ON(pud_huge(*pud));
2523
2524 pmd = pmd_alloc(mm, pud, addr);
2525 if (!pmd)
2526 return -ENOMEM;
2527 do {
2528 next = pmd_addr_end(addr, end);
2529 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2530 if (err)
2531 break;
2532 } while (pmd++, addr = next, addr != end);
2533 return err;
2534}
2535
2536static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2537 unsigned long addr, unsigned long end,
2538 pte_fn_t fn, void *data)
2539{
2540 pud_t *pud;
2541 unsigned long next;
2542 int err;
2543
2544 pud = pud_alloc(mm, pgd, addr);
2545 if (!pud)
2546 return -ENOMEM;
2547 do {
2548 next = pud_addr_end(addr, end);
2549 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2550 if (err)
2551 break;
2552 } while (pud++, addr = next, addr != end);
2553 return err;
2554}
2555
2556/*
2557 * Scan a region of virtual memory, filling in page tables as necessary
2558 * and calling a provided function on each leaf page table.
2559 */
2560int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2561 unsigned long size, pte_fn_t fn, void *data)
2562{
2563 pgd_t *pgd;
2564 unsigned long next;
2565 unsigned long end = addr + size;
2566 int err;
2567
2568 BUG_ON(addr >= end);
2569 pgd = pgd_offset(mm, addr);
2570 do {
2571 next = pgd_addr_end(addr, end);
2572 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2573 if (err)
2574 break;
2575 } while (pgd++, addr = next, addr != end);
2576
2577 return err;
2578}
2579EXPORT_SYMBOL_GPL(apply_to_page_range);
2580
2581/*
2582 * handle_pte_fault chooses page fault handler according to an entry
2583 * which was read non-atomically. Before making any commitment, on
2584 * those architectures or configurations (e.g. i386 with PAE) which
2585 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2586 * must check under lock before unmapping the pte and proceeding
2587 * (but do_wp_page is only called after already making such a check;
2588 * and do_anonymous_page can safely check later on).
2589 */
2590static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2591 pte_t *page_table, pte_t orig_pte)
2592{
2593 int same = 1;
2594#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2595 if (sizeof(pte_t) > sizeof(unsigned long)) {
2596 spinlock_t *ptl = pte_lockptr(mm, pmd);
2597 spin_lock(ptl);
2598 same = pte_same(*page_table, orig_pte);
2599 spin_unlock(ptl);
2600 }
2601#endif
2602 pte_unmap(page_table);
2603 return same;
2604}
2605
2606static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2607{
2608 debug_dma_assert_idle(src);
2609
2610 /*
2611 * If the source page was a PFN mapping, we don't have
2612 * a "struct page" for it. We do a best-effort copy by
2613 * just copying from the original user address. If that
2614 * fails, we just zero-fill it. Live with it.
2615 */
2616 if (unlikely(!src)) {
2617 void *kaddr = kmap_atomic(dst);
2618 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2619
2620 /*
2621 * This really shouldn't fail, because the page is there
2622 * in the page tables. But it might just be unreadable,
2623 * in which case we just give up and fill the result with
2624 * zeroes.
2625 */
2626 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2627 clear_page(kaddr);
2628 kunmap_atomic(kaddr);
2629 flush_dcache_page(dst);
2630 } else
2631 copy_user_highpage(dst, src, va, vma);
2632}
2633
2634/*
2635 * Notify the address space that the page is about to become writable so that
2636 * it can prohibit this or wait for the page to get into an appropriate state.
2637 *
2638 * We do this without the lock held, so that it can sleep if it needs to.
2639 */
2640static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2641 unsigned long address)
2642{
2643 struct vm_fault vmf;
2644 int ret;
2645
2646 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2647 vmf.pgoff = page->index;
2648 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2649 vmf.page = page;
2650
2651 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2652 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2653 return ret;
2654 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2655 lock_page(page);
2656 if (!page->mapping) {
2657 unlock_page(page);
2658 return 0; /* retry */
2659 }
2660 ret |= VM_FAULT_LOCKED;
2661 } else
2662 VM_BUG_ON_PAGE(!PageLocked(page), page);
2663 return ret;
2664}
2665
2666/*
2667 * This routine handles present pages, when users try to write
2668 * to a shared page. It is done by copying the page to a new address
2669 * and decrementing the shared-page counter for the old page.
2670 *
2671 * Note that this routine assumes that the protection checks have been
2672 * done by the caller (the low-level page fault routine in most cases).
2673 * Thus we can safely just mark it writable once we've done any necessary
2674 * COW.
2675 *
2676 * We also mark the page dirty at this point even though the page will
2677 * change only once the write actually happens. This avoids a few races,
2678 * and potentially makes it more efficient.
2679 *
2680 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2681 * but allow concurrent faults), with pte both mapped and locked.
2682 * We return with mmap_sem still held, but pte unmapped and unlocked.
2683 */
2684static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2685 unsigned long address, pte_t *page_table, pmd_t *pmd,
2686 spinlock_t *ptl, pte_t orig_pte)
2687 __releases(ptl)
2688{
2689 struct page *old_page, *new_page = NULL;
2690 pte_t entry;
2691 int ret = 0;
2692 int page_mkwrite = 0;
2693 struct page *dirty_page = NULL;
2694 unsigned long mmun_start = 0; /* For mmu_notifiers */
2695 unsigned long mmun_end = 0; /* For mmu_notifiers */
2696
2697 old_page = vm_normal_page(vma, address, orig_pte);
2698 if (!old_page) {
2699 /*
2700 * VM_MIXEDMAP !pfn_valid() case
2701 *
2702 * We should not cow pages in a shared writeable mapping.
2703 * Just mark the pages writable as we can't do any dirty
2704 * accounting on raw pfn maps.
2705 */
2706 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2707 (VM_WRITE|VM_SHARED))
2708 goto reuse;
2709 goto gotten;
2710 }
2711
2712 /*
2713 * Take out anonymous pages first, anonymous shared vmas are
2714 * not dirty accountable.
2715 */
2716 if (PageAnon(old_page) && !PageKsm(old_page)) {
2717 if (!trylock_page(old_page)) {
2718 page_cache_get(old_page);
2719 pte_unmap_unlock(page_table, ptl);
2720 lock_page(old_page);
2721 page_table = pte_offset_map_lock(mm, pmd, address,
2722 &ptl);
2723 if (!pte_same(*page_table, orig_pte)) {
2724 unlock_page(old_page);
2725 goto unlock;
2726 }
2727 page_cache_release(old_page);
2728 }
2729 if (reuse_swap_page(old_page)) {
2730 /*
2731 * The page is all ours. Move it to our anon_vma so
2732 * the rmap code will not search our parent or siblings.
2733 * Protected against the rmap code by the page lock.
2734 */
2735 page_move_anon_rmap(old_page, vma, address);
2736 unlock_page(old_page);
2737 goto reuse;
2738 }
2739 unlock_page(old_page);
2740 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2741 (VM_WRITE|VM_SHARED))) {
2742 /*
2743 * Only catch write-faults on shared writable pages,
2744 * read-only shared pages can get COWed by
2745 * get_user_pages(.write=1, .force=1).
2746 */
2747 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2748 int tmp;
2749 page_cache_get(old_page);
2750 pte_unmap_unlock(page_table, ptl);
2751 tmp = do_page_mkwrite(vma, old_page, address);
2752 if (unlikely(!tmp || (tmp &
2753 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2754 page_cache_release(old_page);
2755 return tmp;
2756 }
2757 /*
2758 * Since we dropped the lock we need to revalidate
2759 * the PTE as someone else may have changed it. If
2760 * they did, we just return, as we can count on the
2761 * MMU to tell us if they didn't also make it writable.
2762 */
2763 page_table = pte_offset_map_lock(mm, pmd, address,
2764 &ptl);
2765 if (!pte_same(*page_table, orig_pte)) {
2766 unlock_page(old_page);
2767 goto unlock;
2768 }
2769
2770 page_mkwrite = 1;
2771 }
2772 dirty_page = old_page;
2773 get_page(dirty_page);
2774
2775reuse:
2776 /*
2777 * Clear the pages cpupid information as the existing
2778 * information potentially belongs to a now completely
2779 * unrelated process.
2780 */
2781 if (old_page)
2782 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2783
2784 flush_cache_page(vma, address, pte_pfn(orig_pte));
2785 entry = pte_mkyoung(orig_pte);
2786 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2787 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2788 update_mmu_cache(vma, address, page_table);
2789 pte_unmap_unlock(page_table, ptl);
2790 ret |= VM_FAULT_WRITE;
2791
2792 if (!dirty_page)
2793 return ret;
2794
2795 /*
2796 * Yes, Virginia, this is actually required to prevent a race
2797 * with clear_page_dirty_for_io() from clearing the page dirty
2798 * bit after it clear all dirty ptes, but before a racing
2799 * do_wp_page installs a dirty pte.
2800 *
2801 * do_shared_fault is protected similarly.
2802 */
2803 if (!page_mkwrite) {
2804 wait_on_page_locked(dirty_page);
2805 set_page_dirty_balance(dirty_page);
2806 /* file_update_time outside page_lock */
2807 if (vma->vm_file)
2808 file_update_time(vma->vm_file);
2809 }
2810 put_page(dirty_page);
2811 if (page_mkwrite) {
2812 struct address_space *mapping = dirty_page->mapping;
2813
2814 set_page_dirty(dirty_page);
2815 unlock_page(dirty_page);
2816 page_cache_release(dirty_page);
2817 if (mapping) {
2818 /*
2819 * Some device drivers do not set page.mapping
2820 * but still dirty their pages
2821 */
2822 balance_dirty_pages_ratelimited(mapping);
2823 }
2824 }
2825
2826 return ret;
2827 }
2828
2829 /*
2830 * Ok, we need to copy. Oh, well..
2831 */
2832 page_cache_get(old_page);
2833gotten:
2834 pte_unmap_unlock(page_table, ptl);
2835
2836 if (unlikely(anon_vma_prepare(vma)))
2837 goto oom;
2838
2839 if (is_zero_pfn(pte_pfn(orig_pte))) {
2840 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2841 if (!new_page)
2842 goto oom;
2843 } else {
2844 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2845 if (!new_page)
2846 goto oom;
2847 cow_user_page(new_page, old_page, address, vma);
2848 }
2849 __SetPageUptodate(new_page);
2850
2851 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))
2852 goto oom_free_new;
2853
2854 mmun_start = address & PAGE_MASK;
2855 mmun_end = mmun_start + PAGE_SIZE;
2856 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2857
2858 /*
2859 * Re-check the pte - we dropped the lock
2860 */
2861 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2862 if (likely(pte_same(*page_table, orig_pte))) {
2863 if (old_page) {
2864 if (!PageAnon(old_page)) {
2865 dec_mm_counter_fast(mm, MM_FILEPAGES);
2866 inc_mm_counter_fast(mm, MM_ANONPAGES);
2867 }
2868 } else
2869 inc_mm_counter_fast(mm, MM_ANONPAGES);
2870 flush_cache_page(vma, address, pte_pfn(orig_pte));
2871 entry = mk_pte(new_page, vma->vm_page_prot);
2872 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2873 /*
2874 * Clear the pte entry and flush it first, before updating the
2875 * pte with the new entry. This will avoid a race condition
2876 * seen in the presence of one thread doing SMC and another
2877 * thread doing COW.
2878 */
2879 ptep_clear_flush(vma, address, page_table);
2880 page_add_new_anon_rmap(new_page, vma, address);
2881 /*
2882 * We call the notify macro here because, when using secondary
2883 * mmu page tables (such as kvm shadow page tables), we want the
2884 * new page to be mapped directly into the secondary page table.
2885 */
2886 set_pte_at_notify(mm, address, page_table, entry);
2887 update_mmu_cache(vma, address, page_table);
2888 if (old_page) {
2889 /*
2890 * Only after switching the pte to the new page may
2891 * we remove the mapcount here. Otherwise another
2892 * process may come and find the rmap count decremented
2893 * before the pte is switched to the new page, and
2894 * "reuse" the old page writing into it while our pte
2895 * here still points into it and can be read by other
2896 * threads.
2897 *
2898 * The critical issue is to order this
2899 * page_remove_rmap with the ptp_clear_flush above.
2900 * Those stores are ordered by (if nothing else,)
2901 * the barrier present in the atomic_add_negative
2902 * in page_remove_rmap.
2903 *
2904 * Then the TLB flush in ptep_clear_flush ensures that
2905 * no process can access the old page before the
2906 * decremented mapcount is visible. And the old page
2907 * cannot be reused until after the decremented
2908 * mapcount is visible. So transitively, TLBs to
2909 * old page will be flushed before it can be reused.
2910 */
2911 page_remove_rmap(old_page);
2912 }
2913
2914 /* Free the old page.. */
2915 new_page = old_page;
2916 ret |= VM_FAULT_WRITE;
2917 } else
2918 mem_cgroup_uncharge_page(new_page);
2919
2920 if (new_page)
2921 page_cache_release(new_page);
2922unlock:
2923 pte_unmap_unlock(page_table, ptl);
2924 if (mmun_end > mmun_start)
2925 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2926 if (old_page) {
2927 /*
2928 * Don't let another task, with possibly unlocked vma,
2929 * keep the mlocked page.
2930 */
2931 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2932 lock_page(old_page); /* LRU manipulation */
2933 munlock_vma_page(old_page);
2934 unlock_page(old_page);
2935 }
2936 page_cache_release(old_page);
2937 }
2938 return ret;
2939oom_free_new:
2940 page_cache_release(new_page);
2941oom:
2942 if (old_page)
2943 page_cache_release(old_page);
2944 return VM_FAULT_OOM;
2945}
2946
2947static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2948 unsigned long start_addr, unsigned long end_addr,
2949 struct zap_details *details)
2950{
2951 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2952}
2953
2954static inline void unmap_mapping_range_tree(struct rb_root *root,
2955 struct zap_details *details)
2956{
2957 struct vm_area_struct *vma;
2958 pgoff_t vba, vea, zba, zea;
2959
2960 vma_interval_tree_foreach(vma, root,
2961 details->first_index, details->last_index) {
2962
2963 vba = vma->vm_pgoff;
2964 vea = vba + vma_pages(vma) - 1;
2965 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2966 zba = details->first_index;
2967 if (zba < vba)
2968 zba = vba;
2969 zea = details->last_index;
2970 if (zea > vea)
2971 zea = vea;
2972
2973 unmap_mapping_range_vma(vma,
2974 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2975 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2976 details);
2977 }
2978}
2979
2980static inline void unmap_mapping_range_list(struct list_head *head,
2981 struct zap_details *details)
2982{
2983 struct vm_area_struct *vma;
2984
2985 /*
2986 * In nonlinear VMAs there is no correspondence between virtual address
2987 * offset and file offset. So we must perform an exhaustive search
2988 * across *all* the pages in each nonlinear VMA, not just the pages
2989 * whose virtual address lies outside the file truncation point.
2990 */
2991 list_for_each_entry(vma, head, shared.nonlinear) {
2992 details->nonlinear_vma = vma;
2993 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2994 }
2995}
2996
2997/**
2998 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2999 * @mapping: the address space containing mmaps to be unmapped.
3000 * @holebegin: byte in first page to unmap, relative to the start of
3001 * the underlying file. This will be rounded down to a PAGE_SIZE
3002 * boundary. Note that this is different from truncate_pagecache(), which
3003 * must keep the partial page. In contrast, we must get rid of
3004 * partial pages.
3005 * @holelen: size of prospective hole in bytes. This will be rounded
3006 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3007 * end of the file.
3008 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3009 * but 0 when invalidating pagecache, don't throw away private data.
3010 */
3011void unmap_mapping_range(struct address_space *mapping,
3012 loff_t const holebegin, loff_t const holelen, int even_cows)
3013{
3014 struct zap_details details;
3015 pgoff_t hba = holebegin >> PAGE_SHIFT;
3016 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3017
3018 /* Check for overflow. */
3019 if (sizeof(holelen) > sizeof(hlen)) {
3020 long long holeend =
3021 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3022 if (holeend & ~(long long)ULONG_MAX)
3023 hlen = ULONG_MAX - hba + 1;
3024 }
3025
3026 details.check_mapping = even_cows? NULL: mapping;
3027 details.nonlinear_vma = NULL;
3028 details.first_index = hba;
3029 details.last_index = hba + hlen - 1;
3030 if (details.last_index < details.first_index)
3031 details.last_index = ULONG_MAX;
3032
3033
3034 mutex_lock(&mapping->i_mmap_mutex);
3035 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
3036 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3037 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
3038 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
3039 mutex_unlock(&mapping->i_mmap_mutex);
3040}
3041EXPORT_SYMBOL(unmap_mapping_range);
3042
3043/*
3044 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3045 * but allow concurrent faults), and pte mapped but not yet locked.
3046 * We return with mmap_sem still held, but pte unmapped and unlocked.
3047 */
3048static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3049 unsigned long address, pte_t *page_table, pmd_t *pmd,
3050 unsigned int flags, pte_t orig_pte)
3051{
3052 spinlock_t *ptl;
3053 struct page *page, *swapcache;
3054 swp_entry_t entry;
3055 pte_t pte;
3056 int locked;
3057 struct mem_cgroup *ptr;
3058 int exclusive = 0;
3059 int ret = 0;
3060
3061 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3062 goto out;
3063
3064 entry = pte_to_swp_entry(orig_pte);
3065 if (unlikely(non_swap_entry(entry))) {
3066 if (is_migration_entry(entry)) {
3067 migration_entry_wait(mm, pmd, address);
3068 } else if (is_hwpoison_entry(entry)) {
3069 ret = VM_FAULT_HWPOISON;
3070 } else {
3071 print_bad_pte(vma, address, orig_pte, NULL);
3072 ret = VM_FAULT_SIGBUS;
3073 }
3074 goto out;
3075 }
3076 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3077 page = lookup_swap_cache(entry);
3078 if (!page) {
3079 page = swapin_readahead(entry,
3080 GFP_HIGHUSER_MOVABLE, vma, address);
3081 if (!page) {
3082 /*
3083 * Back out if somebody else faulted in this pte
3084 * while we released the pte lock.
3085 */
3086 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3087 if (likely(pte_same(*page_table, orig_pte)))
3088 ret = VM_FAULT_OOM;
3089 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3090 goto unlock;
3091 }
3092
3093 /* Had to read the page from swap area: Major fault */
3094 ret = VM_FAULT_MAJOR;
3095 count_vm_event(PGMAJFAULT);
3096 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3097 } else if (PageHWPoison(page)) {
3098 /*
3099 * hwpoisoned dirty swapcache pages are kept for killing
3100 * owner processes (which may be unknown at hwpoison time)
3101 */
3102 ret = VM_FAULT_HWPOISON;
3103 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3104 swapcache = page;
3105 goto out_release;
3106 }
3107
3108 swapcache = page;
3109 locked = lock_page_or_retry(page, mm, flags);
3110
3111 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3112 if (!locked) {
3113 ret |= VM_FAULT_RETRY;
3114 goto out_release;
3115 }
3116
3117 /*
3118 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3119 * release the swapcache from under us. The page pin, and pte_same
3120 * test below, are not enough to exclude that. Even if it is still
3121 * swapcache, we need to check that the page's swap has not changed.
3122 */
3123 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3124 goto out_page;
3125
3126 page = ksm_might_need_to_copy(page, vma, address);
3127 if (unlikely(!page)) {
3128 ret = VM_FAULT_OOM;
3129 page = swapcache;
3130 goto out_page;
3131 }
3132
3133 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3134 ret = VM_FAULT_OOM;
3135 goto out_page;
3136 }
3137
3138 /*
3139 * Back out if somebody else already faulted in this pte.
3140 */
3141 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3142 if (unlikely(!pte_same(*page_table, orig_pte)))
3143 goto out_nomap;
3144
3145 if (unlikely(!PageUptodate(page))) {
3146 ret = VM_FAULT_SIGBUS;
3147 goto out_nomap;
3148 }
3149
3150 /*
3151 * The page isn't present yet, go ahead with the fault.
3152 *
3153 * Be careful about the sequence of operations here.
3154 * To get its accounting right, reuse_swap_page() must be called
3155 * while the page is counted on swap but not yet in mapcount i.e.
3156 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3157 * must be called after the swap_free(), or it will never succeed.
3158 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3159 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3160 * in page->private. In this case, a record in swap_cgroup is silently
3161 * discarded at swap_free().
3162 */
3163
3164 inc_mm_counter_fast(mm, MM_ANONPAGES);
3165 dec_mm_counter_fast(mm, MM_SWAPENTS);
3166 pte = mk_pte(page, vma->vm_page_prot);
3167 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3168 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3169 flags &= ~FAULT_FLAG_WRITE;
3170 ret |= VM_FAULT_WRITE;
3171 exclusive = 1;
3172 }
3173 flush_icache_page(vma, page);
3174 if (pte_swp_soft_dirty(orig_pte))
3175 pte = pte_mksoft_dirty(pte);
3176 set_pte_at(mm, address, page_table, pte);
3177 if (page == swapcache)
3178 do_page_add_anon_rmap(page, vma, address, exclusive);
3179 else /* ksm created a completely new copy */
3180 page_add_new_anon_rmap(page, vma, address);
3181 /* It's better to call commit-charge after rmap is established */
3182 mem_cgroup_commit_charge_swapin(page, ptr);
3183
3184 swap_free(entry);
3185 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3186 try_to_free_swap(page);
3187 unlock_page(page);
3188 if (page != swapcache) {
3189 /*
3190 * Hold the lock to avoid the swap entry to be reused
3191 * until we take the PT lock for the pte_same() check
3192 * (to avoid false positives from pte_same). For
3193 * further safety release the lock after the swap_free
3194 * so that the swap count won't change under a
3195 * parallel locked swapcache.
3196 */
3197 unlock_page(swapcache);
3198 page_cache_release(swapcache);
3199 }
3200
3201 if (flags & FAULT_FLAG_WRITE) {
3202 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3203 if (ret & VM_FAULT_ERROR)
3204 ret &= VM_FAULT_ERROR;
3205 goto out;
3206 }
3207
3208 /* No need to invalidate - it was non-present before */
3209 update_mmu_cache(vma, address, page_table);
3210unlock:
3211 pte_unmap_unlock(page_table, ptl);
3212out:
3213 return ret;
3214out_nomap:
3215 mem_cgroup_cancel_charge_swapin(ptr);
3216 pte_unmap_unlock(page_table, ptl);
3217out_page:
3218 unlock_page(page);
3219out_release:
3220 page_cache_release(page);
3221 if (page != swapcache) {
3222 unlock_page(swapcache);
3223 page_cache_release(swapcache);
3224 }
3225 return ret;
3226}
3227
3228/*
3229 * This is like a special single-page "expand_{down|up}wards()",
3230 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3231 * doesn't hit another vma.
3232 */
3233static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3234{
3235 address &= PAGE_MASK;
3236 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3237 struct vm_area_struct *prev = vma->vm_prev;
3238
3239 /*
3240 * Is there a mapping abutting this one below?
3241 *
3242 * That's only ok if it's the same stack mapping
3243 * that has gotten split..
3244 */
3245 if (prev && prev->vm_end == address)
3246 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3247
3248 expand_downwards(vma, address - PAGE_SIZE);
3249 }
3250 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3251 struct vm_area_struct *next = vma->vm_next;
3252
3253 /* As VM_GROWSDOWN but s/below/above/ */
3254 if (next && next->vm_start == address + PAGE_SIZE)
3255 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3256
3257 expand_upwards(vma, address + PAGE_SIZE);
3258 }
3259 return 0;
3260}
3261
3262/*
3263 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3264 * but allow concurrent faults), and pte mapped but not yet locked.
3265 * We return with mmap_sem still held, but pte unmapped and unlocked.
3266 */
3267static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3268 unsigned long address, pte_t *page_table, pmd_t *pmd,
3269 unsigned int flags)
3270{
3271 struct page *page;
3272 spinlock_t *ptl;
3273 pte_t entry;
3274
3275 pte_unmap(page_table);
3276
3277 /* Check if we need to add a guard page to the stack */
3278 if (check_stack_guard_page(vma, address) < 0)
3279 return VM_FAULT_SIGBUS;
3280
3281 /* Use the zero-page for reads */
3282 if (!(flags & FAULT_FLAG_WRITE)) {
3283 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3284 vma->vm_page_prot));
3285 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3286 if (!pte_none(*page_table))
3287 goto unlock;
3288 goto setpte;
3289 }
3290
3291 /* Allocate our own private page. */
3292 if (unlikely(anon_vma_prepare(vma)))
3293 goto oom;
3294 page = alloc_zeroed_user_highpage_movable(vma, address);
3295 if (!page)
3296 goto oom;
3297 /*
3298 * The memory barrier inside __SetPageUptodate makes sure that
3299 * preceeding stores to the page contents become visible before
3300 * the set_pte_at() write.
3301 */
3302 __SetPageUptodate(page);
3303
3304 if (mem_cgroup_charge_anon(page, mm, GFP_KERNEL))
3305 goto oom_free_page;
3306
3307 entry = mk_pte(page, vma->vm_page_prot);
3308 if (vma->vm_flags & VM_WRITE)
3309 entry = pte_mkwrite(pte_mkdirty(entry));
3310
3311 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3312 if (!pte_none(*page_table))
3313 goto release;
3314
3315 inc_mm_counter_fast(mm, MM_ANONPAGES);
3316 page_add_new_anon_rmap(page, vma, address);
3317setpte:
3318 set_pte_at(mm, address, page_table, entry);
3319
3320 /* No need to invalidate - it was non-present before */
3321 update_mmu_cache(vma, address, page_table);
3322unlock:
3323 pte_unmap_unlock(page_table, ptl);
3324 return 0;
3325release:
3326 mem_cgroup_uncharge_page(page);
3327 page_cache_release(page);
3328 goto unlock;
3329oom_free_page:
3330 page_cache_release(page);
3331oom:
3332 return VM_FAULT_OOM;
3333}
3334
3335static int __do_fault(struct vm_area_struct *vma, unsigned long address,
3336 pgoff_t pgoff, unsigned int flags, struct page **page)
3337{
3338 struct vm_fault vmf;
3339 int ret;
3340
3341 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3342 vmf.pgoff = pgoff;
3343 vmf.flags = flags;
3344 vmf.page = NULL;
3345
3346 ret = vma->vm_ops->fault(vma, &vmf);
3347 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3348 return ret;
3349
3350 if (unlikely(PageHWPoison(vmf.page))) {
3351 if (ret & VM_FAULT_LOCKED)
3352 unlock_page(vmf.page);
3353 page_cache_release(vmf.page);
3354 return VM_FAULT_HWPOISON;
3355 }
3356
3357 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3358 lock_page(vmf.page);
3359 else
3360 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
3361
3362 *page = vmf.page;
3363 return ret;
3364}
3365
3366/**
3367 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
3368 *
3369 * @vma: virtual memory area
3370 * @address: user virtual address
3371 * @page: page to map
3372 * @pte: pointer to target page table entry
3373 * @write: true, if new entry is writable
3374 * @anon: true, if it's anonymous page
3375 *
3376 * Caller must hold page table lock relevant for @pte.
3377 *
3378 * Target users are page handler itself and implementations of
3379 * vm_ops->map_pages.
3380 */
3381void do_set_pte(struct vm_area_struct *vma, unsigned long address,
3382 struct page *page, pte_t *pte, bool write, bool anon)
3383{
3384 pte_t entry;
3385
3386 flush_icache_page(vma, page);
3387 entry = mk_pte(page, vma->vm_page_prot);
3388 if (write)
3389 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3390 else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
3391 pte_mksoft_dirty(entry);
3392 if (anon) {
3393 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3394 page_add_new_anon_rmap(page, vma, address);
3395 } else {
3396 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
3397 page_add_file_rmap(page);
3398 }
3399 set_pte_at(vma->vm_mm, address, pte, entry);
3400
3401 /* no need to invalidate: a not-present page won't be cached */
3402 update_mmu_cache(vma, address, pte);
3403}
3404
3405#define FAULT_AROUND_ORDER 4
3406
3407#ifdef CONFIG_DEBUG_FS
3408static unsigned int fault_around_order = FAULT_AROUND_ORDER;
3409
3410static int fault_around_order_get(void *data, u64 *val)
3411{
3412 *val = fault_around_order;
3413 return 0;
3414}
3415
3416static int fault_around_order_set(void *data, u64 val)
3417{
3418 BUILD_BUG_ON((1UL << FAULT_AROUND_ORDER) > PTRS_PER_PTE);
3419 if (1UL << val > PTRS_PER_PTE)
3420 return -EINVAL;
3421 fault_around_order = val;
3422 return 0;
3423}
3424DEFINE_SIMPLE_ATTRIBUTE(fault_around_order_fops,
3425 fault_around_order_get, fault_around_order_set, "%llu\n");
3426
3427static int __init fault_around_debugfs(void)
3428{
3429 void *ret;
3430
3431 ret = debugfs_create_file("fault_around_order", 0644, NULL, NULL,
3432 &fault_around_order_fops);
3433 if (!ret)
3434 pr_warn("Failed to create fault_around_order in debugfs");
3435 return 0;
3436}
3437late_initcall(fault_around_debugfs);
3438
3439static inline unsigned long fault_around_pages(void)
3440{
3441 return 1UL << fault_around_order;
3442}
3443
3444static inline unsigned long fault_around_mask(void)
3445{
3446 return ~((1UL << (PAGE_SHIFT + fault_around_order)) - 1);
3447}
3448#else
3449static inline unsigned long fault_around_pages(void)
3450{
3451 unsigned long nr_pages;
3452
3453 nr_pages = 1UL << FAULT_AROUND_ORDER;
3454 BUILD_BUG_ON(nr_pages > PTRS_PER_PTE);
3455 return nr_pages;
3456}
3457
3458static inline unsigned long fault_around_mask(void)
3459{
3460 return ~((1UL << (PAGE_SHIFT + FAULT_AROUND_ORDER)) - 1);
3461}
3462#endif
3463
3464static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
3465 pte_t *pte, pgoff_t pgoff, unsigned int flags)
3466{
3467 unsigned long start_addr;
3468 pgoff_t max_pgoff;
3469 struct vm_fault vmf;
3470 int off;
3471
3472 start_addr = max(address & fault_around_mask(), vma->vm_start);
3473 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3474 pte -= off;
3475 pgoff -= off;
3476
3477 /*
3478 * max_pgoff is either end of page table or end of vma
3479 * or fault_around_pages() from pgoff, depending what is neast.
3480 */
3481 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3482 PTRS_PER_PTE - 1;
3483 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
3484 pgoff + fault_around_pages() - 1);
3485
3486 /* Check if it makes any sense to call ->map_pages */
3487 while (!pte_none(*pte)) {
3488 if (++pgoff > max_pgoff)
3489 return;
3490 start_addr += PAGE_SIZE;
3491 if (start_addr >= vma->vm_end)
3492 return;
3493 pte++;
3494 }
3495
3496 vmf.virtual_address = (void __user *) start_addr;
3497 vmf.pte = pte;
3498 vmf.pgoff = pgoff;
3499 vmf.max_pgoff = max_pgoff;
3500 vmf.flags = flags;
3501 vma->vm_ops->map_pages(vma, &vmf);
3502}
3503
3504static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3505 unsigned long address, pmd_t *pmd,
3506 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3507{
3508 struct page *fault_page;
3509 spinlock_t *ptl;
3510 pte_t *pte;
3511 int ret = 0;
3512
3513 /*
3514 * Let's call ->map_pages() first and use ->fault() as fallback
3515 * if page by the offset is not ready to be mapped (cold cache or
3516 * something).
3517 */
3518 if (vma->vm_ops->map_pages) {
3519 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3520 do_fault_around(vma, address, pte, pgoff, flags);
3521 if (!pte_same(*pte, orig_pte))
3522 goto unlock_out;
3523 pte_unmap_unlock(pte, ptl);
3524 }
3525
3526 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3527 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3528 return ret;
3529
3530 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3531 if (unlikely(!pte_same(*pte, orig_pte))) {
3532 pte_unmap_unlock(pte, ptl);
3533 unlock_page(fault_page);
3534 page_cache_release(fault_page);
3535 return ret;
3536 }
3537 do_set_pte(vma, address, fault_page, pte, false, false);
3538 unlock_page(fault_page);
3539unlock_out:
3540 pte_unmap_unlock(pte, ptl);
3541 return ret;
3542}
3543
3544static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3545 unsigned long address, pmd_t *pmd,
3546 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3547{
3548 struct page *fault_page, *new_page;
3549 spinlock_t *ptl;
3550 pte_t *pte;
3551 int ret;
3552
3553 if (unlikely(anon_vma_prepare(vma)))
3554 return VM_FAULT_OOM;
3555
3556 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3557 if (!new_page)
3558 return VM_FAULT_OOM;
3559
3560 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)) {
3561 page_cache_release(new_page);
3562 return VM_FAULT_OOM;
3563 }
3564
3565 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3566 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3567 goto uncharge_out;
3568
3569 copy_user_highpage(new_page, fault_page, address, vma);
3570 __SetPageUptodate(new_page);
3571
3572 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3573 if (unlikely(!pte_same(*pte, orig_pte))) {
3574 pte_unmap_unlock(pte, ptl);
3575 unlock_page(fault_page);
3576 page_cache_release(fault_page);
3577 goto uncharge_out;
3578 }
3579 do_set_pte(vma, address, new_page, pte, true, true);
3580 pte_unmap_unlock(pte, ptl);
3581 unlock_page(fault_page);
3582 page_cache_release(fault_page);
3583 return ret;
3584uncharge_out:
3585 mem_cgroup_uncharge_page(new_page);
3586 page_cache_release(new_page);
3587 return ret;
3588}
3589
3590static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3591 unsigned long address, pmd_t *pmd,
3592 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3593{
3594 struct page *fault_page;
3595 struct address_space *mapping;
3596 spinlock_t *ptl;
3597 pte_t *pte;
3598 int dirtied = 0;
3599 int ret, tmp;
3600
3601 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3602 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3603 return ret;
3604
3605 /*
3606 * Check if the backing address space wants to know that the page is
3607 * about to become writable
3608 */
3609 if (vma->vm_ops->page_mkwrite) {
3610 unlock_page(fault_page);
3611 tmp = do_page_mkwrite(vma, fault_page, address);
3612 if (unlikely(!tmp ||
3613 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3614 page_cache_release(fault_page);
3615 return tmp;
3616 }
3617 }
3618
3619 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3620 if (unlikely(!pte_same(*pte, orig_pte))) {
3621 pte_unmap_unlock(pte, ptl);
3622 unlock_page(fault_page);
3623 page_cache_release(fault_page);
3624 return ret;
3625 }
3626 do_set_pte(vma, address, fault_page, pte, true, false);
3627 pte_unmap_unlock(pte, ptl);
3628
3629 if (set_page_dirty(fault_page))
3630 dirtied = 1;
3631 mapping = fault_page->mapping;
3632 unlock_page(fault_page);
3633 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3634 /*
3635 * Some device drivers do not set page.mapping but still
3636 * dirty their pages
3637 */
3638 balance_dirty_pages_ratelimited(mapping);
3639 }
3640
3641 /* file_update_time outside page_lock */
3642 if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3643 file_update_time(vma->vm_file);
3644
3645 return ret;
3646}
3647
3648static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3649 unsigned long address, pte_t *page_table, pmd_t *pmd,
3650 unsigned int flags, pte_t orig_pte)
3651{
3652 pgoff_t pgoff = (((address & PAGE_MASK)
3653 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3654
3655 pte_unmap(page_table);
3656 if (!(flags & FAULT_FLAG_WRITE))
3657 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3658 orig_pte);
3659 if (!(vma->vm_flags & VM_SHARED))
3660 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3661 orig_pte);
3662 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3663}
3664
3665/*
3666 * Fault of a previously existing named mapping. Repopulate the pte
3667 * from the encoded file_pte if possible. This enables swappable
3668 * nonlinear vmas.
3669 *
3670 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3671 * but allow concurrent faults), and pte mapped but not yet locked.
3672 * We return with mmap_sem still held, but pte unmapped and unlocked.
3673 */
3674static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3675 unsigned long address, pte_t *page_table, pmd_t *pmd,
3676 unsigned int flags, pte_t orig_pte)
3677{
3678 pgoff_t pgoff;
3679
3680 flags |= FAULT_FLAG_NONLINEAR;
3681
3682 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3683 return 0;
3684
3685 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3686 /*
3687 * Page table corrupted: show pte and kill process.
3688 */
3689 print_bad_pte(vma, address, orig_pte, NULL);
3690 return VM_FAULT_SIGBUS;
3691 }
3692
3693 pgoff = pte_to_pgoff(orig_pte);
3694 if (!(flags & FAULT_FLAG_WRITE))
3695 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3696 orig_pte);
3697 if (!(vma->vm_flags & VM_SHARED))
3698 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3699 orig_pte);
3700 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3701}
3702
3703static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3704 unsigned long addr, int page_nid,
3705 int *flags)
3706{
3707 get_page(page);
3708
3709 count_vm_numa_event(NUMA_HINT_FAULTS);
3710 if (page_nid == numa_node_id()) {
3711 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3712 *flags |= TNF_FAULT_LOCAL;
3713 }
3714
3715 return mpol_misplaced(page, vma, addr);
3716}
3717
3718static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3719 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3720{
3721 struct page *page = NULL;
3722 spinlock_t *ptl;
3723 int page_nid = -1;
3724 int last_cpupid;
3725 int target_nid;
3726 bool migrated = false;
3727 int flags = 0;
3728
3729 /*
3730 * The "pte" at this point cannot be used safely without
3731 * validation through pte_unmap_same(). It's of NUMA type but
3732 * the pfn may be screwed if the read is non atomic.
3733 *
3734 * ptep_modify_prot_start is not called as this is clearing
3735 * the _PAGE_NUMA bit and it is not really expected that there
3736 * would be concurrent hardware modifications to the PTE.
3737 */
3738 ptl = pte_lockptr(mm, pmd);
3739 spin_lock(ptl);
3740 if (unlikely(!pte_same(*ptep, pte))) {
3741 pte_unmap_unlock(ptep, ptl);
3742 goto out;
3743 }
3744
3745 pte = pte_mknonnuma(pte);
3746 set_pte_at(mm, addr, ptep, pte);
3747 update_mmu_cache(vma, addr, ptep);
3748
3749 page = vm_normal_page(vma, addr, pte);
3750 if (!page) {
3751 pte_unmap_unlock(ptep, ptl);
3752 return 0;
3753 }
3754 BUG_ON(is_zero_pfn(page_to_pfn(page)));
3755
3756 /*
3757 * Avoid grouping on DSO/COW pages in specific and RO pages
3758 * in general, RO pages shouldn't hurt as much anyway since
3759 * they can be in shared cache state.
3760 */
3761 if (!pte_write(pte))
3762 flags |= TNF_NO_GROUP;
3763
3764 /*
3765 * Flag if the page is shared between multiple address spaces. This
3766 * is later used when determining whether to group tasks together
3767 */
3768 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3769 flags |= TNF_SHARED;
3770
3771 last_cpupid = page_cpupid_last(page);
3772 page_nid = page_to_nid(page);
3773 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3774 pte_unmap_unlock(ptep, ptl);
3775 if (target_nid == -1) {
3776 put_page(page);
3777 goto out;
3778 }
3779
3780 /* Migrate to the requested node */
3781 migrated = migrate_misplaced_page(page, vma, target_nid);
3782 if (migrated) {
3783 page_nid = target_nid;
3784 flags |= TNF_MIGRATED;
3785 }
3786
3787out:
3788 if (page_nid != -1)
3789 task_numa_fault(last_cpupid, page_nid, 1, flags);
3790 return 0;
3791}
3792
3793/*
3794 * These routines also need to handle stuff like marking pages dirty
3795 * and/or accessed for architectures that don't do it in hardware (most
3796 * RISC architectures). The early dirtying is also good on the i386.
3797 *
3798 * There is also a hook called "update_mmu_cache()" that architectures
3799 * with external mmu caches can use to update those (ie the Sparc or
3800 * PowerPC hashed page tables that act as extended TLBs).
3801 *
3802 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3803 * but allow concurrent faults), and pte mapped but not yet locked.
3804 * We return with mmap_sem still held, but pte unmapped and unlocked.
3805 */
3806static int handle_pte_fault(struct mm_struct *mm,
3807 struct vm_area_struct *vma, unsigned long address,
3808 pte_t *pte, pmd_t *pmd, unsigned int flags)
3809{
3810 pte_t entry;
3811 spinlock_t *ptl;
3812
3813 entry = *pte;
3814 if (!pte_present(entry)) {
3815 if (pte_none(entry)) {
3816 if (vma->vm_ops) {
3817 if (likely(vma->vm_ops->fault))
3818 return do_linear_fault(mm, vma, address,
3819 pte, pmd, flags, entry);
3820 }
3821 return do_anonymous_page(mm, vma, address,
3822 pte, pmd, flags);
3823 }
3824 if (pte_file(entry))
3825 return do_nonlinear_fault(mm, vma, address,
3826 pte, pmd, flags, entry);
3827 return do_swap_page(mm, vma, address,
3828 pte, pmd, flags, entry);
3829 }
3830
3831 if (pte_numa(entry))
3832 return do_numa_page(mm, vma, address, entry, pte, pmd);
3833
3834 ptl = pte_lockptr(mm, pmd);
3835 spin_lock(ptl);
3836 if (unlikely(!pte_same(*pte, entry)))
3837 goto unlock;
3838 if (flags & FAULT_FLAG_WRITE) {
3839 if (!pte_write(entry))
3840 return do_wp_page(mm, vma, address,
3841 pte, pmd, ptl, entry);
3842 entry = pte_mkdirty(entry);
3843 }
3844 entry = pte_mkyoung(entry);
3845 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3846 update_mmu_cache(vma, address, pte);
3847 } else {
3848 /*
3849 * This is needed only for protection faults but the arch code
3850 * is not yet telling us if this is a protection fault or not.
3851 * This still avoids useless tlb flushes for .text page faults
3852 * with threads.
3853 */
3854 if (flags & FAULT_FLAG_WRITE)
3855 flush_tlb_fix_spurious_fault(vma, address);
3856 }
3857unlock:
3858 pte_unmap_unlock(pte, ptl);
3859 return 0;
3860}
3861
3862/*
3863 * By the time we get here, we already hold the mm semaphore
3864 */
3865static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3866 unsigned long address, unsigned int flags)
3867{
3868 pgd_t *pgd;
3869 pud_t *pud;
3870 pmd_t *pmd;
3871 pte_t *pte;
3872
3873 if (unlikely(is_vm_hugetlb_page(vma)))
3874 return hugetlb_fault(mm, vma, address, flags);
3875
3876 pgd = pgd_offset(mm, address);
3877 pud = pud_alloc(mm, pgd, address);
3878 if (!pud)
3879 return VM_FAULT_OOM;
3880 pmd = pmd_alloc(mm, pud, address);
3881 if (!pmd)
3882 return VM_FAULT_OOM;
3883 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3884 int ret = VM_FAULT_FALLBACK;
3885 if (!vma->vm_ops)
3886 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3887 pmd, flags);
3888 if (!(ret & VM_FAULT_FALLBACK))
3889 return ret;
3890 } else {
3891 pmd_t orig_pmd = *pmd;
3892 int ret;
3893
3894 barrier();
3895 if (pmd_trans_huge(orig_pmd)) {
3896 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3897
3898 /*
3899 * If the pmd is splitting, return and retry the
3900 * the fault. Alternative: wait until the split
3901 * is done, and goto retry.
3902 */
3903 if (pmd_trans_splitting(orig_pmd))
3904 return 0;
3905
3906 if (pmd_numa(orig_pmd))
3907 return do_huge_pmd_numa_page(mm, vma, address,
3908 orig_pmd, pmd);
3909
3910 if (dirty && !pmd_write(orig_pmd)) {
3911 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3912 orig_pmd);
3913 if (!(ret & VM_FAULT_FALLBACK))
3914 return ret;
3915 } else {
3916 huge_pmd_set_accessed(mm, vma, address, pmd,
3917 orig_pmd, dirty);
3918 return 0;
3919 }
3920 }
3921 }
3922
3923 /* THP should already have been handled */
3924 BUG_ON(pmd_numa(*pmd));
3925
3926 /*
3927 * Use __pte_alloc instead of pte_alloc_map, because we can't
3928 * run pte_offset_map on the pmd, if an huge pmd could
3929 * materialize from under us from a different thread.
3930 */
3931 if (unlikely(pmd_none(*pmd)) &&
3932 unlikely(__pte_alloc(mm, vma, pmd, address)))
3933 return VM_FAULT_OOM;
3934 /* if an huge pmd materialized from under us just retry later */
3935 if (unlikely(pmd_trans_huge(*pmd)))
3936 return 0;
3937 /*
3938 * A regular pmd is established and it can't morph into a huge pmd
3939 * from under us anymore at this point because we hold the mmap_sem
3940 * read mode and khugepaged takes it in write mode. So now it's
3941 * safe to run pte_offset_map().
3942 */
3943 pte = pte_offset_map(pmd, address);
3944
3945 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3946}
3947
3948int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3949 unsigned long address, unsigned int flags)
3950{
3951 int ret;
3952
3953 __set_current_state(TASK_RUNNING);
3954
3955 count_vm_event(PGFAULT);
3956 mem_cgroup_count_vm_event(mm, PGFAULT);
3957
3958 /* do counter updates before entering really critical section. */
3959 check_sync_rss_stat(current);
3960
3961 /*
3962 * Enable the memcg OOM handling for faults triggered in user
3963 * space. Kernel faults are handled more gracefully.
3964 */
3965 if (flags & FAULT_FLAG_USER)
3966 mem_cgroup_oom_enable();
3967
3968 ret = __handle_mm_fault(mm, vma, address, flags);
3969
3970 if (flags & FAULT_FLAG_USER) {
3971 mem_cgroup_oom_disable();
3972 /*
3973 * The task may have entered a memcg OOM situation but
3974 * if the allocation error was handled gracefully (no
3975 * VM_FAULT_OOM), there is no need to kill anything.
3976 * Just clean up the OOM state peacefully.
3977 */
3978 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3979 mem_cgroup_oom_synchronize(false);
3980 }
3981
3982 return ret;
3983}
3984
3985#ifndef __PAGETABLE_PUD_FOLDED
3986/*
3987 * Allocate page upper directory.
3988 * We've already handled the fast-path in-line.
3989 */
3990int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3991{
3992 pud_t *new = pud_alloc_one(mm, address);
3993 if (!new)
3994 return -ENOMEM;
3995
3996 smp_wmb(); /* See comment in __pte_alloc */
3997
3998 spin_lock(&mm->page_table_lock);
3999 if (pgd_present(*pgd)) /* Another has populated it */
4000 pud_free(mm, new);
4001 else
4002 pgd_populate(mm, pgd, new);
4003 spin_unlock(&mm->page_table_lock);
4004 return 0;
4005}
4006#endif /* __PAGETABLE_PUD_FOLDED */
4007
4008#ifndef __PAGETABLE_PMD_FOLDED
4009/*
4010 * Allocate page middle directory.
4011 * We've already handled the fast-path in-line.
4012 */
4013int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4014{
4015 pmd_t *new = pmd_alloc_one(mm, address);
4016 if (!new)
4017 return -ENOMEM;
4018
4019 smp_wmb(); /* See comment in __pte_alloc */
4020
4021 spin_lock(&mm->page_table_lock);
4022#ifndef __ARCH_HAS_4LEVEL_HACK
4023 if (pud_present(*pud)) /* Another has populated it */
4024 pmd_free(mm, new);
4025 else
4026 pud_populate(mm, pud, new);
4027#else
4028 if (pgd_present(*pud)) /* Another has populated it */
4029 pmd_free(mm, new);
4030 else
4031 pgd_populate(mm, pud, new);
4032#endif /* __ARCH_HAS_4LEVEL_HACK */
4033 spin_unlock(&mm->page_table_lock);
4034 return 0;
4035}
4036#endif /* __PAGETABLE_PMD_FOLDED */
4037
4038#if !defined(__HAVE_ARCH_GATE_AREA)
4039
4040#if defined(AT_SYSINFO_EHDR)
4041static struct vm_area_struct gate_vma;
4042
4043static int __init gate_vma_init(void)
4044{
4045 gate_vma.vm_mm = NULL;
4046 gate_vma.vm_start = FIXADDR_USER_START;
4047 gate_vma.vm_end = FIXADDR_USER_END;
4048 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
4049 gate_vma.vm_page_prot = __P101;
4050
4051 return 0;
4052}
4053__initcall(gate_vma_init);
4054#endif
4055
4056struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
4057{
4058#ifdef AT_SYSINFO_EHDR
4059 return &gate_vma;
4060#else
4061 return NULL;
4062#endif
4063}
4064
4065int in_gate_area_no_mm(unsigned long addr)
4066{
4067#ifdef AT_SYSINFO_EHDR
4068 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
4069 return 1;
4070#endif
4071 return 0;
4072}
4073
4074#endif /* __HAVE_ARCH_GATE_AREA */
4075
4076static int __follow_pte(struct mm_struct *mm, unsigned long address,
4077 pte_t **ptepp, spinlock_t **ptlp)
4078{
4079 pgd_t *pgd;
4080 pud_t *pud;
4081 pmd_t *pmd;
4082 pte_t *ptep;
4083
4084 pgd = pgd_offset(mm, address);
4085 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4086 goto out;
4087
4088 pud = pud_offset(pgd, address);
4089 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4090 goto out;
4091
4092 pmd = pmd_offset(pud, address);
4093 VM_BUG_ON(pmd_trans_huge(*pmd));
4094 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4095 goto out;
4096
4097 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
4098 if (pmd_huge(*pmd))
4099 goto out;
4100
4101 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4102 if (!ptep)
4103 goto out;
4104 if (!pte_present(*ptep))
4105 goto unlock;
4106 *ptepp = ptep;
4107 return 0;
4108unlock:
4109 pte_unmap_unlock(ptep, *ptlp);
4110out:
4111 return -EINVAL;
4112}
4113
4114static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4115 pte_t **ptepp, spinlock_t **ptlp)
4116{
4117 int res;
4118
4119 /* (void) is needed to make gcc happy */
4120 (void) __cond_lock(*ptlp,
4121 !(res = __follow_pte(mm, address, ptepp, ptlp)));
4122 return res;
4123}
4124
4125/**
4126 * follow_pfn - look up PFN at a user virtual address
4127 * @vma: memory mapping
4128 * @address: user virtual address
4129 * @pfn: location to store found PFN
4130 *
4131 * Only IO mappings and raw PFN mappings are allowed.
4132 *
4133 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4134 */
4135int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4136 unsigned long *pfn)
4137{
4138 int ret = -EINVAL;
4139 spinlock_t *ptl;
4140 pte_t *ptep;
4141
4142 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4143 return ret;
4144
4145 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4146 if (ret)
4147 return ret;
4148 *pfn = pte_pfn(*ptep);
4149 pte_unmap_unlock(ptep, ptl);
4150 return 0;
4151}
4152EXPORT_SYMBOL(follow_pfn);
4153
4154#ifdef CONFIG_HAVE_IOREMAP_PROT
4155int follow_phys(struct vm_area_struct *vma,
4156 unsigned long address, unsigned int flags,
4157 unsigned long *prot, resource_size_t *phys)
4158{
4159 int ret = -EINVAL;
4160 pte_t *ptep, pte;
4161 spinlock_t *ptl;
4162
4163 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4164 goto out;
4165
4166 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4167 goto out;
4168 pte = *ptep;
4169
4170 if ((flags & FOLL_WRITE) && !pte_write(pte))
4171 goto unlock;
4172
4173 *prot = pgprot_val(pte_pgprot(pte));
4174 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4175
4176 ret = 0;
4177unlock:
4178 pte_unmap_unlock(ptep, ptl);
4179out:
4180 return ret;
4181}
4182
4183int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4184 void *buf, int len, int write)
4185{
4186 resource_size_t phys_addr;
4187 unsigned long prot = 0;
4188 void __iomem *maddr;
4189 int offset = addr & (PAGE_SIZE-1);
4190
4191 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4192 return -EINVAL;
4193
4194 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4195 if (write)
4196 memcpy_toio(maddr + offset, buf, len);
4197 else
4198 memcpy_fromio(buf, maddr + offset, len);
4199 iounmap(maddr);
4200
4201 return len;
4202}
4203EXPORT_SYMBOL_GPL(generic_access_phys);
4204#endif
4205
4206/*
4207 * Access another process' address space as given in mm. If non-NULL, use the
4208 * given task for page fault accounting.
4209 */
4210static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4211 unsigned long addr, void *buf, int len, int write)
4212{
4213 struct vm_area_struct *vma;
4214 void *old_buf = buf;
4215
4216 down_read(&mm->mmap_sem);
4217 /* ignore errors, just check how much was successfully transferred */
4218 while (len) {
4219 int bytes, ret, offset;
4220 void *maddr;
4221 struct page *page = NULL;
4222
4223 ret = get_user_pages(tsk, mm, addr, 1,
4224 write, 1, &page, &vma);
4225 if (ret <= 0) {
4226 /*
4227 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4228 * we can access using slightly different code.
4229 */
4230#ifdef CONFIG_HAVE_IOREMAP_PROT
4231 vma = find_vma(mm, addr);
4232 if (!vma || vma->vm_start > addr)
4233 break;
4234 if (vma->vm_ops && vma->vm_ops->access)
4235 ret = vma->vm_ops->access(vma, addr, buf,
4236 len, write);
4237 if (ret <= 0)
4238#endif
4239 break;
4240 bytes = ret;
4241 } else {
4242 bytes = len;
4243 offset = addr & (PAGE_SIZE-1);
4244 if (bytes > PAGE_SIZE-offset)
4245 bytes = PAGE_SIZE-offset;
4246
4247 maddr = kmap(page);
4248 if (write) {
4249 copy_to_user_page(vma, page, addr,
4250 maddr + offset, buf, bytes);
4251 set_page_dirty_lock(page);
4252 } else {
4253 copy_from_user_page(vma, page, addr,
4254 buf, maddr + offset, bytes);
4255 }
4256 kunmap(page);
4257 page_cache_release(page);
4258 }
4259 len -= bytes;
4260 buf += bytes;
4261 addr += bytes;
4262 }
4263 up_read(&mm->mmap_sem);
4264
4265 return buf - old_buf;
4266}
4267
4268/**
4269 * access_remote_vm - access another process' address space
4270 * @mm: the mm_struct of the target address space
4271 * @addr: start address to access
4272 * @buf: source or destination buffer
4273 * @len: number of bytes to transfer
4274 * @write: whether the access is a write
4275 *
4276 * The caller must hold a reference on @mm.
4277 */
4278int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4279 void *buf, int len, int write)
4280{
4281 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4282}
4283
4284/*
4285 * Access another process' address space.
4286 * Source/target buffer must be kernel space,
4287 * Do not walk the page table directly, use get_user_pages
4288 */
4289int access_process_vm(struct task_struct *tsk, unsigned long addr,
4290 void *buf, int len, int write)
4291{
4292 struct mm_struct *mm;
4293 int ret;
4294
4295 mm = get_task_mm(tsk);
4296 if (!mm)
4297 return 0;
4298
4299 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4300 mmput(mm);
4301
4302 return ret;
4303}
4304
4305/*
4306 * Print the name of a VMA.
4307 */
4308void print_vma_addr(char *prefix, unsigned long ip)
4309{
4310 struct mm_struct *mm = current->mm;
4311 struct vm_area_struct *vma;
4312
4313 /*
4314 * Do not print if we are in atomic
4315 * contexts (in exception stacks, etc.):
4316 */
4317 if (preempt_count())
4318 return;
4319
4320 down_read(&mm->mmap_sem);
4321 vma = find_vma(mm, ip);
4322 if (vma && vma->vm_file) {
4323 struct file *f = vma->vm_file;
4324 char *buf = (char *)__get_free_page(GFP_KERNEL);
4325 if (buf) {
4326 char *p;
4327
4328 p = d_path(&f->f_path, buf, PAGE_SIZE);
4329 if (IS_ERR(p))
4330 p = "?";
4331 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4332 vma->vm_start,
4333 vma->vm_end - vma->vm_start);
4334 free_page((unsigned long)buf);
4335 }
4336 }
4337 up_read(&mm->mmap_sem);
4338}
4339
4340#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4341void might_fault(void)
4342{
4343 /*
4344 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4345 * holding the mmap_sem, this is safe because kernel memory doesn't
4346 * get paged out, therefore we'll never actually fault, and the
4347 * below annotations will generate false positives.
4348 */
4349 if (segment_eq(get_fs(), KERNEL_DS))
4350 return;
4351
4352 /*
4353 * it would be nicer only to annotate paths which are not under
4354 * pagefault_disable, however that requires a larger audit and
4355 * providing helpers like get_user_atomic.
4356 */
4357 if (in_atomic())
4358 return;
4359
4360 __might_sleep(__FILE__, __LINE__, 0);
4361
4362 if (current->mm)
4363 might_lock_read(¤t->mm->mmap_sem);
4364}
4365EXPORT_SYMBOL(might_fault);
4366#endif
4367
4368#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4369static void clear_gigantic_page(struct page *page,
4370 unsigned long addr,
4371 unsigned int pages_per_huge_page)
4372{
4373 int i;
4374 struct page *p = page;
4375
4376 might_sleep();
4377 for (i = 0; i < pages_per_huge_page;
4378 i++, p = mem_map_next(p, page, i)) {
4379 cond_resched();
4380 clear_user_highpage(p, addr + i * PAGE_SIZE);
4381 }
4382}
4383void clear_huge_page(struct page *page,
4384 unsigned long addr, unsigned int pages_per_huge_page)
4385{
4386 int i;
4387
4388 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4389 clear_gigantic_page(page, addr, pages_per_huge_page);
4390 return;
4391 }
4392
4393 might_sleep();
4394 for (i = 0; i < pages_per_huge_page; i++) {
4395 cond_resched();
4396 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4397 }
4398}
4399
4400static void copy_user_gigantic_page(struct page *dst, struct page *src,
4401 unsigned long addr,
4402 struct vm_area_struct *vma,
4403 unsigned int pages_per_huge_page)
4404{
4405 int i;
4406 struct page *dst_base = dst;
4407 struct page *src_base = src;
4408
4409 for (i = 0; i < pages_per_huge_page; ) {
4410 cond_resched();
4411 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4412
4413 i++;
4414 dst = mem_map_next(dst, dst_base, i);
4415 src = mem_map_next(src, src_base, i);
4416 }
4417}
4418
4419void copy_user_huge_page(struct page *dst, struct page *src,
4420 unsigned long addr, struct vm_area_struct *vma,
4421 unsigned int pages_per_huge_page)
4422{
4423 int i;
4424
4425 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4426 copy_user_gigantic_page(dst, src, addr, vma,
4427 pages_per_huge_page);
4428 return;
4429 }
4430
4431 might_sleep();
4432 for (i = 0; i < pages_per_huge_page; i++) {
4433 cond_resched();
4434 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4435 }
4436}
4437#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4438
4439#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4440
4441static struct kmem_cache *page_ptl_cachep;
4442
4443void __init ptlock_cache_init(void)
4444{
4445 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4446 SLAB_PANIC, NULL);
4447}
4448
4449bool ptlock_alloc(struct page *page)
4450{
4451 spinlock_t *ptl;
4452
4453 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4454 if (!ptl)
4455 return false;
4456 page->ptl = ptl;
4457 return true;
4458}
4459
4460void ptlock_free(struct page *page)
4461{
4462 kmem_cache_free(page_ptl_cachep, page->ptl);
4463}
4464#endif
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/mm_inline.h>
45#include <linux/sched/mm.h>
46#include <linux/sched/coredump.h>
47#include <linux/sched/numa_balancing.h>
48#include <linux/sched/task.h>
49#include <linux/hugetlb.h>
50#include <linux/mman.h>
51#include <linux/swap.h>
52#include <linux/highmem.h>
53#include <linux/pagemap.h>
54#include <linux/memremap.h>
55#include <linux/kmsan.h>
56#include <linux/ksm.h>
57#include <linux/rmap.h>
58#include <linux/export.h>
59#include <linux/delayacct.h>
60#include <linux/init.h>
61#include <linux/pfn_t.h>
62#include <linux/writeback.h>
63#include <linux/memcontrol.h>
64#include <linux/mmu_notifier.h>
65#include <linux/swapops.h>
66#include <linux/elf.h>
67#include <linux/gfp.h>
68#include <linux/migrate.h>
69#include <linux/string.h>
70#include <linux/memory-tiers.h>
71#include <linux/debugfs.h>
72#include <linux/userfaultfd_k.h>
73#include <linux/dax.h>
74#include <linux/oom.h>
75#include <linux/numa.h>
76#include <linux/perf_event.h>
77#include <linux/ptrace.h>
78#include <linux/vmalloc.h>
79#include <linux/sched/sysctl.h>
80
81#include <trace/events/kmem.h>
82
83#include <asm/io.h>
84#include <asm/mmu_context.h>
85#include <asm/pgalloc.h>
86#include <linux/uaccess.h>
87#include <asm/tlb.h>
88#include <asm/tlbflush.h>
89
90#include "pgalloc-track.h"
91#include "internal.h"
92#include "swap.h"
93
94#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
95#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
96#endif
97
98#ifndef CONFIG_NUMA
99unsigned long max_mapnr;
100EXPORT_SYMBOL(max_mapnr);
101
102struct page *mem_map;
103EXPORT_SYMBOL(mem_map);
104#endif
105
106static vm_fault_t do_fault(struct vm_fault *vmf);
107
108/*
109 * A number of key systems in x86 including ioremap() rely on the assumption
110 * that high_memory defines the upper bound on direct map memory, then end
111 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
112 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
113 * and ZONE_HIGHMEM.
114 */
115void *high_memory;
116EXPORT_SYMBOL(high_memory);
117
118/*
119 * Randomize the address space (stacks, mmaps, brk, etc.).
120 *
121 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
122 * as ancient (libc5 based) binaries can segfault. )
123 */
124int randomize_va_space __read_mostly =
125#ifdef CONFIG_COMPAT_BRK
126 1;
127#else
128 2;
129#endif
130
131#ifndef arch_wants_old_prefaulted_pte
132static inline bool arch_wants_old_prefaulted_pte(void)
133{
134 /*
135 * Transitioning a PTE from 'old' to 'young' can be expensive on
136 * some architectures, even if it's performed in hardware. By
137 * default, "false" means prefaulted entries will be 'young'.
138 */
139 return false;
140}
141#endif
142
143static int __init disable_randmaps(char *s)
144{
145 randomize_va_space = 0;
146 return 1;
147}
148__setup("norandmaps", disable_randmaps);
149
150unsigned long zero_pfn __read_mostly;
151EXPORT_SYMBOL(zero_pfn);
152
153unsigned long highest_memmap_pfn __read_mostly;
154
155/*
156 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
157 */
158static int __init init_zero_pfn(void)
159{
160 zero_pfn = page_to_pfn(ZERO_PAGE(0));
161 return 0;
162}
163early_initcall(init_zero_pfn);
164
165void mm_trace_rss_stat(struct mm_struct *mm, int member)
166{
167 trace_rss_stat(mm, member);
168}
169
170/*
171 * Note: this doesn't free the actual pages themselves. That
172 * has been handled earlier when unmapping all the memory regions.
173 */
174static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
175 unsigned long addr)
176{
177 pgtable_t token = pmd_pgtable(*pmd);
178 pmd_clear(pmd);
179 pte_free_tlb(tlb, token, addr);
180 mm_dec_nr_ptes(tlb->mm);
181}
182
183static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
184 unsigned long addr, unsigned long end,
185 unsigned long floor, unsigned long ceiling)
186{
187 pmd_t *pmd;
188 unsigned long next;
189 unsigned long start;
190
191 start = addr;
192 pmd = pmd_offset(pud, addr);
193 do {
194 next = pmd_addr_end(addr, end);
195 if (pmd_none_or_clear_bad(pmd))
196 continue;
197 free_pte_range(tlb, pmd, addr);
198 } while (pmd++, addr = next, addr != end);
199
200 start &= PUD_MASK;
201 if (start < floor)
202 return;
203 if (ceiling) {
204 ceiling &= PUD_MASK;
205 if (!ceiling)
206 return;
207 }
208 if (end - 1 > ceiling - 1)
209 return;
210
211 pmd = pmd_offset(pud, start);
212 pud_clear(pud);
213 pmd_free_tlb(tlb, pmd, start);
214 mm_dec_nr_pmds(tlb->mm);
215}
216
217static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
218 unsigned long addr, unsigned long end,
219 unsigned long floor, unsigned long ceiling)
220{
221 pud_t *pud;
222 unsigned long next;
223 unsigned long start;
224
225 start = addr;
226 pud = pud_offset(p4d, addr);
227 do {
228 next = pud_addr_end(addr, end);
229 if (pud_none_or_clear_bad(pud))
230 continue;
231 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
232 } while (pud++, addr = next, addr != end);
233
234 start &= P4D_MASK;
235 if (start < floor)
236 return;
237 if (ceiling) {
238 ceiling &= P4D_MASK;
239 if (!ceiling)
240 return;
241 }
242 if (end - 1 > ceiling - 1)
243 return;
244
245 pud = pud_offset(p4d, start);
246 p4d_clear(p4d);
247 pud_free_tlb(tlb, pud, start);
248 mm_dec_nr_puds(tlb->mm);
249}
250
251static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
252 unsigned long addr, unsigned long end,
253 unsigned long floor, unsigned long ceiling)
254{
255 p4d_t *p4d;
256 unsigned long next;
257 unsigned long start;
258
259 start = addr;
260 p4d = p4d_offset(pgd, addr);
261 do {
262 next = p4d_addr_end(addr, end);
263 if (p4d_none_or_clear_bad(p4d))
264 continue;
265 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
266 } while (p4d++, addr = next, addr != end);
267
268 start &= PGDIR_MASK;
269 if (start < floor)
270 return;
271 if (ceiling) {
272 ceiling &= PGDIR_MASK;
273 if (!ceiling)
274 return;
275 }
276 if (end - 1 > ceiling - 1)
277 return;
278
279 p4d = p4d_offset(pgd, start);
280 pgd_clear(pgd);
281 p4d_free_tlb(tlb, p4d, start);
282}
283
284/*
285 * This function frees user-level page tables of a process.
286 */
287void free_pgd_range(struct mmu_gather *tlb,
288 unsigned long addr, unsigned long end,
289 unsigned long floor, unsigned long ceiling)
290{
291 pgd_t *pgd;
292 unsigned long next;
293
294 /*
295 * The next few lines have given us lots of grief...
296 *
297 * Why are we testing PMD* at this top level? Because often
298 * there will be no work to do at all, and we'd prefer not to
299 * go all the way down to the bottom just to discover that.
300 *
301 * Why all these "- 1"s? Because 0 represents both the bottom
302 * of the address space and the top of it (using -1 for the
303 * top wouldn't help much: the masks would do the wrong thing).
304 * The rule is that addr 0 and floor 0 refer to the bottom of
305 * the address space, but end 0 and ceiling 0 refer to the top
306 * Comparisons need to use "end - 1" and "ceiling - 1" (though
307 * that end 0 case should be mythical).
308 *
309 * Wherever addr is brought up or ceiling brought down, we must
310 * be careful to reject "the opposite 0" before it confuses the
311 * subsequent tests. But what about where end is brought down
312 * by PMD_SIZE below? no, end can't go down to 0 there.
313 *
314 * Whereas we round start (addr) and ceiling down, by different
315 * masks at different levels, in order to test whether a table
316 * now has no other vmas using it, so can be freed, we don't
317 * bother to round floor or end up - the tests don't need that.
318 */
319
320 addr &= PMD_MASK;
321 if (addr < floor) {
322 addr += PMD_SIZE;
323 if (!addr)
324 return;
325 }
326 if (ceiling) {
327 ceiling &= PMD_MASK;
328 if (!ceiling)
329 return;
330 }
331 if (end - 1 > ceiling - 1)
332 end -= PMD_SIZE;
333 if (addr > end - 1)
334 return;
335 /*
336 * We add page table cache pages with PAGE_SIZE,
337 * (see pte_free_tlb()), flush the tlb if we need
338 */
339 tlb_change_page_size(tlb, PAGE_SIZE);
340 pgd = pgd_offset(tlb->mm, addr);
341 do {
342 next = pgd_addr_end(addr, end);
343 if (pgd_none_or_clear_bad(pgd))
344 continue;
345 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
346 } while (pgd++, addr = next, addr != end);
347}
348
349void free_pgtables(struct mmu_gather *tlb, struct maple_tree *mt,
350 struct vm_area_struct *vma, unsigned long floor,
351 unsigned long ceiling)
352{
353 MA_STATE(mas, mt, vma->vm_end, vma->vm_end);
354
355 do {
356 unsigned long addr = vma->vm_start;
357 struct vm_area_struct *next;
358
359 /*
360 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
361 * be 0. This will underflow and is okay.
362 */
363 next = mas_find(&mas, ceiling - 1);
364
365 /*
366 * Hide vma from rmap and truncate_pagecache before freeing
367 * pgtables
368 */
369 unlink_anon_vmas(vma);
370 unlink_file_vma(vma);
371
372 if (is_vm_hugetlb_page(vma)) {
373 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
374 floor, next ? next->vm_start : ceiling);
375 } else {
376 /*
377 * Optimization: gather nearby vmas into one call down
378 */
379 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
380 && !is_vm_hugetlb_page(next)) {
381 vma = next;
382 next = mas_find(&mas, ceiling - 1);
383 unlink_anon_vmas(vma);
384 unlink_file_vma(vma);
385 }
386 free_pgd_range(tlb, addr, vma->vm_end,
387 floor, next ? next->vm_start : ceiling);
388 }
389 vma = next;
390 } while (vma);
391}
392
393void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
394{
395 spinlock_t *ptl = pmd_lock(mm, pmd);
396
397 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
398 mm_inc_nr_ptes(mm);
399 /*
400 * Ensure all pte setup (eg. pte page lock and page clearing) are
401 * visible before the pte is made visible to other CPUs by being
402 * put into page tables.
403 *
404 * The other side of the story is the pointer chasing in the page
405 * table walking code (when walking the page table without locking;
406 * ie. most of the time). Fortunately, these data accesses consist
407 * of a chain of data-dependent loads, meaning most CPUs (alpha
408 * being the notable exception) will already guarantee loads are
409 * seen in-order. See the alpha page table accessors for the
410 * smp_rmb() barriers in page table walking code.
411 */
412 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
413 pmd_populate(mm, pmd, *pte);
414 *pte = NULL;
415 }
416 spin_unlock(ptl);
417}
418
419int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
420{
421 pgtable_t new = pte_alloc_one(mm);
422 if (!new)
423 return -ENOMEM;
424
425 pmd_install(mm, pmd, &new);
426 if (new)
427 pte_free(mm, new);
428 return 0;
429}
430
431int __pte_alloc_kernel(pmd_t *pmd)
432{
433 pte_t *new = pte_alloc_one_kernel(&init_mm);
434 if (!new)
435 return -ENOMEM;
436
437 spin_lock(&init_mm.page_table_lock);
438 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
439 smp_wmb(); /* See comment in pmd_install() */
440 pmd_populate_kernel(&init_mm, pmd, new);
441 new = NULL;
442 }
443 spin_unlock(&init_mm.page_table_lock);
444 if (new)
445 pte_free_kernel(&init_mm, new);
446 return 0;
447}
448
449static inline void init_rss_vec(int *rss)
450{
451 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
452}
453
454static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
455{
456 int i;
457
458 if (current->mm == mm)
459 sync_mm_rss(mm);
460 for (i = 0; i < NR_MM_COUNTERS; i++)
461 if (rss[i])
462 add_mm_counter(mm, i, rss[i]);
463}
464
465/*
466 * This function is called to print an error when a bad pte
467 * is found. For example, we might have a PFN-mapped pte in
468 * a region that doesn't allow it.
469 *
470 * The calling function must still handle the error.
471 */
472static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
473 pte_t pte, struct page *page)
474{
475 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
476 p4d_t *p4d = p4d_offset(pgd, addr);
477 pud_t *pud = pud_offset(p4d, addr);
478 pmd_t *pmd = pmd_offset(pud, addr);
479 struct address_space *mapping;
480 pgoff_t index;
481 static unsigned long resume;
482 static unsigned long nr_shown;
483 static unsigned long nr_unshown;
484
485 /*
486 * Allow a burst of 60 reports, then keep quiet for that minute;
487 * or allow a steady drip of one report per second.
488 */
489 if (nr_shown == 60) {
490 if (time_before(jiffies, resume)) {
491 nr_unshown++;
492 return;
493 }
494 if (nr_unshown) {
495 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
496 nr_unshown);
497 nr_unshown = 0;
498 }
499 nr_shown = 0;
500 }
501 if (nr_shown++ == 0)
502 resume = jiffies + 60 * HZ;
503
504 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
505 index = linear_page_index(vma, addr);
506
507 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
508 current->comm,
509 (long long)pte_val(pte), (long long)pmd_val(*pmd));
510 if (page)
511 dump_page(page, "bad pte");
512 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
513 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
514 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
515 vma->vm_file,
516 vma->vm_ops ? vma->vm_ops->fault : NULL,
517 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
518 mapping ? mapping->a_ops->read_folio : NULL);
519 dump_stack();
520 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
521}
522
523/*
524 * vm_normal_page -- This function gets the "struct page" associated with a pte.
525 *
526 * "Special" mappings do not wish to be associated with a "struct page" (either
527 * it doesn't exist, or it exists but they don't want to touch it). In this
528 * case, NULL is returned here. "Normal" mappings do have a struct page.
529 *
530 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
531 * pte bit, in which case this function is trivial. Secondly, an architecture
532 * may not have a spare pte bit, which requires a more complicated scheme,
533 * described below.
534 *
535 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
536 * special mapping (even if there are underlying and valid "struct pages").
537 * COWed pages of a VM_PFNMAP are always normal.
538 *
539 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
540 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
541 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
542 * mapping will always honor the rule
543 *
544 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
545 *
546 * And for normal mappings this is false.
547 *
548 * This restricts such mappings to be a linear translation from virtual address
549 * to pfn. To get around this restriction, we allow arbitrary mappings so long
550 * as the vma is not a COW mapping; in that case, we know that all ptes are
551 * special (because none can have been COWed).
552 *
553 *
554 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
555 *
556 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
557 * page" backing, however the difference is that _all_ pages with a struct
558 * page (that is, those where pfn_valid is true) are refcounted and considered
559 * normal pages by the VM. The disadvantage is that pages are refcounted
560 * (which can be slower and simply not an option for some PFNMAP users). The
561 * advantage is that we don't have to follow the strict linearity rule of
562 * PFNMAP mappings in order to support COWable mappings.
563 *
564 */
565struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
566 pte_t pte)
567{
568 unsigned long pfn = pte_pfn(pte);
569
570 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
571 if (likely(!pte_special(pte)))
572 goto check_pfn;
573 if (vma->vm_ops && vma->vm_ops->find_special_page)
574 return vma->vm_ops->find_special_page(vma, addr);
575 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
576 return NULL;
577 if (is_zero_pfn(pfn))
578 return NULL;
579 if (pte_devmap(pte))
580 /*
581 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
582 * and will have refcounts incremented on their struct pages
583 * when they are inserted into PTEs, thus they are safe to
584 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
585 * do not have refcounts. Example of legacy ZONE_DEVICE is
586 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
587 */
588 return NULL;
589
590 print_bad_pte(vma, addr, pte, NULL);
591 return NULL;
592 }
593
594 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
595
596 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
597 if (vma->vm_flags & VM_MIXEDMAP) {
598 if (!pfn_valid(pfn))
599 return NULL;
600 goto out;
601 } else {
602 unsigned long off;
603 off = (addr - vma->vm_start) >> PAGE_SHIFT;
604 if (pfn == vma->vm_pgoff + off)
605 return NULL;
606 if (!is_cow_mapping(vma->vm_flags))
607 return NULL;
608 }
609 }
610
611 if (is_zero_pfn(pfn))
612 return NULL;
613
614check_pfn:
615 if (unlikely(pfn > highest_memmap_pfn)) {
616 print_bad_pte(vma, addr, pte, NULL);
617 return NULL;
618 }
619
620 /*
621 * NOTE! We still have PageReserved() pages in the page tables.
622 * eg. VDSO mappings can cause them to exist.
623 */
624out:
625 return pfn_to_page(pfn);
626}
627
628#ifdef CONFIG_TRANSPARENT_HUGEPAGE
629struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
630 pmd_t pmd)
631{
632 unsigned long pfn = pmd_pfn(pmd);
633
634 /*
635 * There is no pmd_special() but there may be special pmds, e.g.
636 * in a direct-access (dax) mapping, so let's just replicate the
637 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
638 */
639 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
640 if (vma->vm_flags & VM_MIXEDMAP) {
641 if (!pfn_valid(pfn))
642 return NULL;
643 goto out;
644 } else {
645 unsigned long off;
646 off = (addr - vma->vm_start) >> PAGE_SHIFT;
647 if (pfn == vma->vm_pgoff + off)
648 return NULL;
649 if (!is_cow_mapping(vma->vm_flags))
650 return NULL;
651 }
652 }
653
654 if (pmd_devmap(pmd))
655 return NULL;
656 if (is_huge_zero_pmd(pmd))
657 return NULL;
658 if (unlikely(pfn > highest_memmap_pfn))
659 return NULL;
660
661 /*
662 * NOTE! We still have PageReserved() pages in the page tables.
663 * eg. VDSO mappings can cause them to exist.
664 */
665out:
666 return pfn_to_page(pfn);
667}
668#endif
669
670static void restore_exclusive_pte(struct vm_area_struct *vma,
671 struct page *page, unsigned long address,
672 pte_t *ptep)
673{
674 pte_t pte;
675 swp_entry_t entry;
676
677 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
678 if (pte_swp_soft_dirty(*ptep))
679 pte = pte_mksoft_dirty(pte);
680
681 entry = pte_to_swp_entry(*ptep);
682 if (pte_swp_uffd_wp(*ptep))
683 pte = pte_mkuffd_wp(pte);
684 else if (is_writable_device_exclusive_entry(entry))
685 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
686
687 VM_BUG_ON(pte_write(pte) && !(PageAnon(page) && PageAnonExclusive(page)));
688
689 /*
690 * No need to take a page reference as one was already
691 * created when the swap entry was made.
692 */
693 if (PageAnon(page))
694 page_add_anon_rmap(page, vma, address, RMAP_NONE);
695 else
696 /*
697 * Currently device exclusive access only supports anonymous
698 * memory so the entry shouldn't point to a filebacked page.
699 */
700 WARN_ON_ONCE(1);
701
702 set_pte_at(vma->vm_mm, address, ptep, pte);
703
704 /*
705 * No need to invalidate - it was non-present before. However
706 * secondary CPUs may have mappings that need invalidating.
707 */
708 update_mmu_cache(vma, address, ptep);
709}
710
711/*
712 * Tries to restore an exclusive pte if the page lock can be acquired without
713 * sleeping.
714 */
715static int
716try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
717 unsigned long addr)
718{
719 swp_entry_t entry = pte_to_swp_entry(*src_pte);
720 struct page *page = pfn_swap_entry_to_page(entry);
721
722 if (trylock_page(page)) {
723 restore_exclusive_pte(vma, page, addr, src_pte);
724 unlock_page(page);
725 return 0;
726 }
727
728 return -EBUSY;
729}
730
731/*
732 * copy one vm_area from one task to the other. Assumes the page tables
733 * already present in the new task to be cleared in the whole range
734 * covered by this vma.
735 */
736
737static unsigned long
738copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
739 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
740 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
741{
742 unsigned long vm_flags = dst_vma->vm_flags;
743 pte_t pte = *src_pte;
744 struct page *page;
745 swp_entry_t entry = pte_to_swp_entry(pte);
746
747 if (likely(!non_swap_entry(entry))) {
748 if (swap_duplicate(entry) < 0)
749 return -EIO;
750
751 /* make sure dst_mm is on swapoff's mmlist. */
752 if (unlikely(list_empty(&dst_mm->mmlist))) {
753 spin_lock(&mmlist_lock);
754 if (list_empty(&dst_mm->mmlist))
755 list_add(&dst_mm->mmlist,
756 &src_mm->mmlist);
757 spin_unlock(&mmlist_lock);
758 }
759 /* Mark the swap entry as shared. */
760 if (pte_swp_exclusive(*src_pte)) {
761 pte = pte_swp_clear_exclusive(*src_pte);
762 set_pte_at(src_mm, addr, src_pte, pte);
763 }
764 rss[MM_SWAPENTS]++;
765 } else if (is_migration_entry(entry)) {
766 page = pfn_swap_entry_to_page(entry);
767
768 rss[mm_counter(page)]++;
769
770 if (!is_readable_migration_entry(entry) &&
771 is_cow_mapping(vm_flags)) {
772 /*
773 * COW mappings require pages in both parent and child
774 * to be set to read. A previously exclusive entry is
775 * now shared.
776 */
777 entry = make_readable_migration_entry(
778 swp_offset(entry));
779 pte = swp_entry_to_pte(entry);
780 if (pte_swp_soft_dirty(*src_pte))
781 pte = pte_swp_mksoft_dirty(pte);
782 if (pte_swp_uffd_wp(*src_pte))
783 pte = pte_swp_mkuffd_wp(pte);
784 set_pte_at(src_mm, addr, src_pte, pte);
785 }
786 } else if (is_device_private_entry(entry)) {
787 page = pfn_swap_entry_to_page(entry);
788
789 /*
790 * Update rss count even for unaddressable pages, as
791 * they should treated just like normal pages in this
792 * respect.
793 *
794 * We will likely want to have some new rss counters
795 * for unaddressable pages, at some point. But for now
796 * keep things as they are.
797 */
798 get_page(page);
799 rss[mm_counter(page)]++;
800 /* Cannot fail as these pages cannot get pinned. */
801 BUG_ON(page_try_dup_anon_rmap(page, false, src_vma));
802
803 /*
804 * We do not preserve soft-dirty information, because so
805 * far, checkpoint/restore is the only feature that
806 * requires that. And checkpoint/restore does not work
807 * when a device driver is involved (you cannot easily
808 * save and restore device driver state).
809 */
810 if (is_writable_device_private_entry(entry) &&
811 is_cow_mapping(vm_flags)) {
812 entry = make_readable_device_private_entry(
813 swp_offset(entry));
814 pte = swp_entry_to_pte(entry);
815 if (pte_swp_uffd_wp(*src_pte))
816 pte = pte_swp_mkuffd_wp(pte);
817 set_pte_at(src_mm, addr, src_pte, pte);
818 }
819 } else if (is_device_exclusive_entry(entry)) {
820 /*
821 * Make device exclusive entries present by restoring the
822 * original entry then copying as for a present pte. Device
823 * exclusive entries currently only support private writable
824 * (ie. COW) mappings.
825 */
826 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
827 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
828 return -EBUSY;
829 return -ENOENT;
830 } else if (is_pte_marker_entry(entry)) {
831 if (is_swapin_error_entry(entry) || userfaultfd_wp(dst_vma))
832 set_pte_at(dst_mm, addr, dst_pte, pte);
833 return 0;
834 }
835 if (!userfaultfd_wp(dst_vma))
836 pte = pte_swp_clear_uffd_wp(pte);
837 set_pte_at(dst_mm, addr, dst_pte, pte);
838 return 0;
839}
840
841/*
842 * Copy a present and normal page.
843 *
844 * NOTE! The usual case is that this isn't required;
845 * instead, the caller can just increase the page refcount
846 * and re-use the pte the traditional way.
847 *
848 * And if we need a pre-allocated page but don't yet have
849 * one, return a negative error to let the preallocation
850 * code know so that it can do so outside the page table
851 * lock.
852 */
853static inline int
854copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
855 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
856 struct page **prealloc, struct page *page)
857{
858 struct page *new_page;
859 pte_t pte;
860
861 new_page = *prealloc;
862 if (!new_page)
863 return -EAGAIN;
864
865 /*
866 * We have a prealloc page, all good! Take it
867 * over and copy the page & arm it.
868 */
869 *prealloc = NULL;
870 copy_user_highpage(new_page, page, addr, src_vma);
871 __SetPageUptodate(new_page);
872 page_add_new_anon_rmap(new_page, dst_vma, addr);
873 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
874 rss[mm_counter(new_page)]++;
875
876 /* All done, just insert the new page copy in the child */
877 pte = mk_pte(new_page, dst_vma->vm_page_prot);
878 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
879 if (userfaultfd_pte_wp(dst_vma, *src_pte))
880 /* Uffd-wp needs to be delivered to dest pte as well */
881 pte = pte_wrprotect(pte_mkuffd_wp(pte));
882 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
883 return 0;
884}
885
886/*
887 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
888 * is required to copy this pte.
889 */
890static inline int
891copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
892 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
893 struct page **prealloc)
894{
895 struct mm_struct *src_mm = src_vma->vm_mm;
896 unsigned long vm_flags = src_vma->vm_flags;
897 pte_t pte = *src_pte;
898 struct page *page;
899
900 page = vm_normal_page(src_vma, addr, pte);
901 if (page && PageAnon(page)) {
902 /*
903 * If this page may have been pinned by the parent process,
904 * copy the page immediately for the child so that we'll always
905 * guarantee the pinned page won't be randomly replaced in the
906 * future.
907 */
908 get_page(page);
909 if (unlikely(page_try_dup_anon_rmap(page, false, src_vma))) {
910 /* Page maybe pinned, we have to copy. */
911 put_page(page);
912 return copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
913 addr, rss, prealloc, page);
914 }
915 rss[mm_counter(page)]++;
916 } else if (page) {
917 get_page(page);
918 page_dup_file_rmap(page, false);
919 rss[mm_counter(page)]++;
920 }
921
922 /*
923 * If it's a COW mapping, write protect it both
924 * in the parent and the child
925 */
926 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
927 ptep_set_wrprotect(src_mm, addr, src_pte);
928 pte = pte_wrprotect(pte);
929 }
930 VM_BUG_ON(page && PageAnon(page) && PageAnonExclusive(page));
931
932 /*
933 * If it's a shared mapping, mark it clean in
934 * the child
935 */
936 if (vm_flags & VM_SHARED)
937 pte = pte_mkclean(pte);
938 pte = pte_mkold(pte);
939
940 if (!userfaultfd_wp(dst_vma))
941 pte = pte_clear_uffd_wp(pte);
942
943 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
944 return 0;
945}
946
947static inline struct page *
948page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
949 unsigned long addr)
950{
951 struct page *new_page;
952
953 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
954 if (!new_page)
955 return NULL;
956
957 if (mem_cgroup_charge(page_folio(new_page), src_mm, GFP_KERNEL)) {
958 put_page(new_page);
959 return NULL;
960 }
961 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
962
963 return new_page;
964}
965
966static int
967copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
968 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
969 unsigned long end)
970{
971 struct mm_struct *dst_mm = dst_vma->vm_mm;
972 struct mm_struct *src_mm = src_vma->vm_mm;
973 pte_t *orig_src_pte, *orig_dst_pte;
974 pte_t *src_pte, *dst_pte;
975 spinlock_t *src_ptl, *dst_ptl;
976 int progress, ret = 0;
977 int rss[NR_MM_COUNTERS];
978 swp_entry_t entry = (swp_entry_t){0};
979 struct page *prealloc = NULL;
980
981again:
982 progress = 0;
983 init_rss_vec(rss);
984
985 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
986 if (!dst_pte) {
987 ret = -ENOMEM;
988 goto out;
989 }
990 src_pte = pte_offset_map(src_pmd, addr);
991 src_ptl = pte_lockptr(src_mm, src_pmd);
992 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
993 orig_src_pte = src_pte;
994 orig_dst_pte = dst_pte;
995 arch_enter_lazy_mmu_mode();
996
997 do {
998 /*
999 * We are holding two locks at this point - either of them
1000 * could generate latencies in another task on another CPU.
1001 */
1002 if (progress >= 32) {
1003 progress = 0;
1004 if (need_resched() ||
1005 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1006 break;
1007 }
1008 if (pte_none(*src_pte)) {
1009 progress++;
1010 continue;
1011 }
1012 if (unlikely(!pte_present(*src_pte))) {
1013 ret = copy_nonpresent_pte(dst_mm, src_mm,
1014 dst_pte, src_pte,
1015 dst_vma, src_vma,
1016 addr, rss);
1017 if (ret == -EIO) {
1018 entry = pte_to_swp_entry(*src_pte);
1019 break;
1020 } else if (ret == -EBUSY) {
1021 break;
1022 } else if (!ret) {
1023 progress += 8;
1024 continue;
1025 }
1026
1027 /*
1028 * Device exclusive entry restored, continue by copying
1029 * the now present pte.
1030 */
1031 WARN_ON_ONCE(ret != -ENOENT);
1032 }
1033 /* copy_present_pte() will clear `*prealloc' if consumed */
1034 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1035 addr, rss, &prealloc);
1036 /*
1037 * If we need a pre-allocated page for this pte, drop the
1038 * locks, allocate, and try again.
1039 */
1040 if (unlikely(ret == -EAGAIN))
1041 break;
1042 if (unlikely(prealloc)) {
1043 /*
1044 * pre-alloc page cannot be reused by next time so as
1045 * to strictly follow mempolicy (e.g., alloc_page_vma()
1046 * will allocate page according to address). This
1047 * could only happen if one pinned pte changed.
1048 */
1049 put_page(prealloc);
1050 prealloc = NULL;
1051 }
1052 progress += 8;
1053 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1054
1055 arch_leave_lazy_mmu_mode();
1056 spin_unlock(src_ptl);
1057 pte_unmap(orig_src_pte);
1058 add_mm_rss_vec(dst_mm, rss);
1059 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1060 cond_resched();
1061
1062 if (ret == -EIO) {
1063 VM_WARN_ON_ONCE(!entry.val);
1064 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1065 ret = -ENOMEM;
1066 goto out;
1067 }
1068 entry.val = 0;
1069 } else if (ret == -EBUSY) {
1070 goto out;
1071 } else if (ret == -EAGAIN) {
1072 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1073 if (!prealloc)
1074 return -ENOMEM;
1075 } else if (ret) {
1076 VM_WARN_ON_ONCE(1);
1077 }
1078
1079 /* We've captured and resolved the error. Reset, try again. */
1080 ret = 0;
1081
1082 if (addr != end)
1083 goto again;
1084out:
1085 if (unlikely(prealloc))
1086 put_page(prealloc);
1087 return ret;
1088}
1089
1090static inline int
1091copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1092 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1093 unsigned long end)
1094{
1095 struct mm_struct *dst_mm = dst_vma->vm_mm;
1096 struct mm_struct *src_mm = src_vma->vm_mm;
1097 pmd_t *src_pmd, *dst_pmd;
1098 unsigned long next;
1099
1100 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1101 if (!dst_pmd)
1102 return -ENOMEM;
1103 src_pmd = pmd_offset(src_pud, addr);
1104 do {
1105 next = pmd_addr_end(addr, end);
1106 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1107 || pmd_devmap(*src_pmd)) {
1108 int err;
1109 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1110 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1111 addr, dst_vma, src_vma);
1112 if (err == -ENOMEM)
1113 return -ENOMEM;
1114 if (!err)
1115 continue;
1116 /* fall through */
1117 }
1118 if (pmd_none_or_clear_bad(src_pmd))
1119 continue;
1120 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1121 addr, next))
1122 return -ENOMEM;
1123 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1124 return 0;
1125}
1126
1127static inline int
1128copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1129 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1130 unsigned long end)
1131{
1132 struct mm_struct *dst_mm = dst_vma->vm_mm;
1133 struct mm_struct *src_mm = src_vma->vm_mm;
1134 pud_t *src_pud, *dst_pud;
1135 unsigned long next;
1136
1137 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1138 if (!dst_pud)
1139 return -ENOMEM;
1140 src_pud = pud_offset(src_p4d, addr);
1141 do {
1142 next = pud_addr_end(addr, end);
1143 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1144 int err;
1145
1146 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1147 err = copy_huge_pud(dst_mm, src_mm,
1148 dst_pud, src_pud, addr, src_vma);
1149 if (err == -ENOMEM)
1150 return -ENOMEM;
1151 if (!err)
1152 continue;
1153 /* fall through */
1154 }
1155 if (pud_none_or_clear_bad(src_pud))
1156 continue;
1157 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1158 addr, next))
1159 return -ENOMEM;
1160 } while (dst_pud++, src_pud++, addr = next, addr != end);
1161 return 0;
1162}
1163
1164static inline int
1165copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1166 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1167 unsigned long end)
1168{
1169 struct mm_struct *dst_mm = dst_vma->vm_mm;
1170 p4d_t *src_p4d, *dst_p4d;
1171 unsigned long next;
1172
1173 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1174 if (!dst_p4d)
1175 return -ENOMEM;
1176 src_p4d = p4d_offset(src_pgd, addr);
1177 do {
1178 next = p4d_addr_end(addr, end);
1179 if (p4d_none_or_clear_bad(src_p4d))
1180 continue;
1181 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1182 addr, next))
1183 return -ENOMEM;
1184 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1185 return 0;
1186}
1187
1188/*
1189 * Return true if the vma needs to copy the pgtable during this fork(). Return
1190 * false when we can speed up fork() by allowing lazy page faults later until
1191 * when the child accesses the memory range.
1192 */
1193static bool
1194vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1195{
1196 /*
1197 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1198 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1199 * contains uffd-wp protection information, that's something we can't
1200 * retrieve from page cache, and skip copying will lose those info.
1201 */
1202 if (userfaultfd_wp(dst_vma))
1203 return true;
1204
1205 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1206 return true;
1207
1208 if (src_vma->anon_vma)
1209 return true;
1210
1211 /*
1212 * Don't copy ptes where a page fault will fill them correctly. Fork
1213 * becomes much lighter when there are big shared or private readonly
1214 * mappings. The tradeoff is that copy_page_range is more efficient
1215 * than faulting.
1216 */
1217 return false;
1218}
1219
1220int
1221copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1222{
1223 pgd_t *src_pgd, *dst_pgd;
1224 unsigned long next;
1225 unsigned long addr = src_vma->vm_start;
1226 unsigned long end = src_vma->vm_end;
1227 struct mm_struct *dst_mm = dst_vma->vm_mm;
1228 struct mm_struct *src_mm = src_vma->vm_mm;
1229 struct mmu_notifier_range range;
1230 bool is_cow;
1231 int ret;
1232
1233 if (!vma_needs_copy(dst_vma, src_vma))
1234 return 0;
1235
1236 if (is_vm_hugetlb_page(src_vma))
1237 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1238
1239 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1240 /*
1241 * We do not free on error cases below as remove_vma
1242 * gets called on error from higher level routine
1243 */
1244 ret = track_pfn_copy(src_vma);
1245 if (ret)
1246 return ret;
1247 }
1248
1249 /*
1250 * We need to invalidate the secondary MMU mappings only when
1251 * there could be a permission downgrade on the ptes of the
1252 * parent mm. And a permission downgrade will only happen if
1253 * is_cow_mapping() returns true.
1254 */
1255 is_cow = is_cow_mapping(src_vma->vm_flags);
1256
1257 if (is_cow) {
1258 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1259 0, src_vma, src_mm, addr, end);
1260 mmu_notifier_invalidate_range_start(&range);
1261 /*
1262 * Disabling preemption is not needed for the write side, as
1263 * the read side doesn't spin, but goes to the mmap_lock.
1264 *
1265 * Use the raw variant of the seqcount_t write API to avoid
1266 * lockdep complaining about preemptibility.
1267 */
1268 mmap_assert_write_locked(src_mm);
1269 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1270 }
1271
1272 ret = 0;
1273 dst_pgd = pgd_offset(dst_mm, addr);
1274 src_pgd = pgd_offset(src_mm, addr);
1275 do {
1276 next = pgd_addr_end(addr, end);
1277 if (pgd_none_or_clear_bad(src_pgd))
1278 continue;
1279 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1280 addr, next))) {
1281 ret = -ENOMEM;
1282 break;
1283 }
1284 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1285
1286 if (is_cow) {
1287 raw_write_seqcount_end(&src_mm->write_protect_seq);
1288 mmu_notifier_invalidate_range_end(&range);
1289 }
1290 return ret;
1291}
1292
1293/* Whether we should zap all COWed (private) pages too */
1294static inline bool should_zap_cows(struct zap_details *details)
1295{
1296 /* By default, zap all pages */
1297 if (!details)
1298 return true;
1299
1300 /* Or, we zap COWed pages only if the caller wants to */
1301 return details->even_cows;
1302}
1303
1304/* Decides whether we should zap this page with the page pointer specified */
1305static inline bool should_zap_page(struct zap_details *details, struct page *page)
1306{
1307 /* If we can make a decision without *page.. */
1308 if (should_zap_cows(details))
1309 return true;
1310
1311 /* E.g. the caller passes NULL for the case of a zero page */
1312 if (!page)
1313 return true;
1314
1315 /* Otherwise we should only zap non-anon pages */
1316 return !PageAnon(page);
1317}
1318
1319static inline bool zap_drop_file_uffd_wp(struct zap_details *details)
1320{
1321 if (!details)
1322 return false;
1323
1324 return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1325}
1326
1327/*
1328 * This function makes sure that we'll replace the none pte with an uffd-wp
1329 * swap special pte marker when necessary. Must be with the pgtable lock held.
1330 */
1331static inline void
1332zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1333 unsigned long addr, pte_t *pte,
1334 struct zap_details *details, pte_t pteval)
1335{
1336 if (zap_drop_file_uffd_wp(details))
1337 return;
1338
1339 pte_install_uffd_wp_if_needed(vma, addr, pte, pteval);
1340}
1341
1342static unsigned long zap_pte_range(struct mmu_gather *tlb,
1343 struct vm_area_struct *vma, pmd_t *pmd,
1344 unsigned long addr, unsigned long end,
1345 struct zap_details *details)
1346{
1347 struct mm_struct *mm = tlb->mm;
1348 int force_flush = 0;
1349 int rss[NR_MM_COUNTERS];
1350 spinlock_t *ptl;
1351 pte_t *start_pte;
1352 pte_t *pte;
1353 swp_entry_t entry;
1354
1355 tlb_change_page_size(tlb, PAGE_SIZE);
1356again:
1357 init_rss_vec(rss);
1358 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1359 pte = start_pte;
1360 flush_tlb_batched_pending(mm);
1361 arch_enter_lazy_mmu_mode();
1362 do {
1363 pte_t ptent = *pte;
1364 struct page *page;
1365
1366 if (pte_none(ptent))
1367 continue;
1368
1369 if (need_resched())
1370 break;
1371
1372 if (pte_present(ptent)) {
1373 unsigned int delay_rmap;
1374
1375 page = vm_normal_page(vma, addr, ptent);
1376 if (unlikely(!should_zap_page(details, page)))
1377 continue;
1378 ptent = ptep_get_and_clear_full(mm, addr, pte,
1379 tlb->fullmm);
1380 tlb_remove_tlb_entry(tlb, pte, addr);
1381 zap_install_uffd_wp_if_needed(vma, addr, pte, details,
1382 ptent);
1383 if (unlikely(!page))
1384 continue;
1385
1386 delay_rmap = 0;
1387 if (!PageAnon(page)) {
1388 if (pte_dirty(ptent)) {
1389 set_page_dirty(page);
1390 if (tlb_delay_rmap(tlb)) {
1391 delay_rmap = 1;
1392 force_flush = 1;
1393 }
1394 }
1395 if (pte_young(ptent) &&
1396 likely(!(vma->vm_flags & VM_SEQ_READ)))
1397 mark_page_accessed(page);
1398 }
1399 rss[mm_counter(page)]--;
1400 if (!delay_rmap) {
1401 page_remove_rmap(page, vma, false);
1402 if (unlikely(page_mapcount(page) < 0))
1403 print_bad_pte(vma, addr, ptent, page);
1404 }
1405 if (unlikely(__tlb_remove_page(tlb, page, delay_rmap))) {
1406 force_flush = 1;
1407 addr += PAGE_SIZE;
1408 break;
1409 }
1410 continue;
1411 }
1412
1413 entry = pte_to_swp_entry(ptent);
1414 if (is_device_private_entry(entry) ||
1415 is_device_exclusive_entry(entry)) {
1416 page = pfn_swap_entry_to_page(entry);
1417 if (unlikely(!should_zap_page(details, page)))
1418 continue;
1419 /*
1420 * Both device private/exclusive mappings should only
1421 * work with anonymous page so far, so we don't need to
1422 * consider uffd-wp bit when zap. For more information,
1423 * see zap_install_uffd_wp_if_needed().
1424 */
1425 WARN_ON_ONCE(!vma_is_anonymous(vma));
1426 rss[mm_counter(page)]--;
1427 if (is_device_private_entry(entry))
1428 page_remove_rmap(page, vma, false);
1429 put_page(page);
1430 } else if (!non_swap_entry(entry)) {
1431 /* Genuine swap entry, hence a private anon page */
1432 if (!should_zap_cows(details))
1433 continue;
1434 rss[MM_SWAPENTS]--;
1435 if (unlikely(!free_swap_and_cache(entry)))
1436 print_bad_pte(vma, addr, ptent, NULL);
1437 } else if (is_migration_entry(entry)) {
1438 page = pfn_swap_entry_to_page(entry);
1439 if (!should_zap_page(details, page))
1440 continue;
1441 rss[mm_counter(page)]--;
1442 } else if (pte_marker_entry_uffd_wp(entry)) {
1443 /* Only drop the uffd-wp marker if explicitly requested */
1444 if (!zap_drop_file_uffd_wp(details))
1445 continue;
1446 } else if (is_hwpoison_entry(entry) ||
1447 is_swapin_error_entry(entry)) {
1448 if (!should_zap_cows(details))
1449 continue;
1450 } else {
1451 /* We should have covered all the swap entry types */
1452 WARN_ON_ONCE(1);
1453 }
1454 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1455 zap_install_uffd_wp_if_needed(vma, addr, pte, details, ptent);
1456 } while (pte++, addr += PAGE_SIZE, addr != end);
1457
1458 add_mm_rss_vec(mm, rss);
1459 arch_leave_lazy_mmu_mode();
1460
1461 /* Do the actual TLB flush before dropping ptl */
1462 if (force_flush) {
1463 tlb_flush_mmu_tlbonly(tlb);
1464 tlb_flush_rmaps(tlb, vma);
1465 }
1466 pte_unmap_unlock(start_pte, ptl);
1467
1468 /*
1469 * If we forced a TLB flush (either due to running out of
1470 * batch buffers or because we needed to flush dirty TLB
1471 * entries before releasing the ptl), free the batched
1472 * memory too. Restart if we didn't do everything.
1473 */
1474 if (force_flush) {
1475 force_flush = 0;
1476 tlb_flush_mmu(tlb);
1477 }
1478
1479 if (addr != end) {
1480 cond_resched();
1481 goto again;
1482 }
1483
1484 return addr;
1485}
1486
1487static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1488 struct vm_area_struct *vma, pud_t *pud,
1489 unsigned long addr, unsigned long end,
1490 struct zap_details *details)
1491{
1492 pmd_t *pmd;
1493 unsigned long next;
1494
1495 pmd = pmd_offset(pud, addr);
1496 do {
1497 next = pmd_addr_end(addr, end);
1498 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1499 if (next - addr != HPAGE_PMD_SIZE)
1500 __split_huge_pmd(vma, pmd, addr, false, NULL);
1501 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1502 goto next;
1503 /* fall through */
1504 } else if (details && details->single_folio &&
1505 folio_test_pmd_mappable(details->single_folio) &&
1506 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1507 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1508 /*
1509 * Take and drop THP pmd lock so that we cannot return
1510 * prematurely, while zap_huge_pmd() has cleared *pmd,
1511 * but not yet decremented compound_mapcount().
1512 */
1513 spin_unlock(ptl);
1514 }
1515
1516 /*
1517 * Here there can be other concurrent MADV_DONTNEED or
1518 * trans huge page faults running, and if the pmd is
1519 * none or trans huge it can change under us. This is
1520 * because MADV_DONTNEED holds the mmap_lock in read
1521 * mode.
1522 */
1523 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1524 goto next;
1525 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1526next:
1527 cond_resched();
1528 } while (pmd++, addr = next, addr != end);
1529
1530 return addr;
1531}
1532
1533static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1534 struct vm_area_struct *vma, p4d_t *p4d,
1535 unsigned long addr, unsigned long end,
1536 struct zap_details *details)
1537{
1538 pud_t *pud;
1539 unsigned long next;
1540
1541 pud = pud_offset(p4d, addr);
1542 do {
1543 next = pud_addr_end(addr, end);
1544 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1545 if (next - addr != HPAGE_PUD_SIZE) {
1546 mmap_assert_locked(tlb->mm);
1547 split_huge_pud(vma, pud, addr);
1548 } else if (zap_huge_pud(tlb, vma, pud, addr))
1549 goto next;
1550 /* fall through */
1551 }
1552 if (pud_none_or_clear_bad(pud))
1553 continue;
1554 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1555next:
1556 cond_resched();
1557 } while (pud++, addr = next, addr != end);
1558
1559 return addr;
1560}
1561
1562static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1563 struct vm_area_struct *vma, pgd_t *pgd,
1564 unsigned long addr, unsigned long end,
1565 struct zap_details *details)
1566{
1567 p4d_t *p4d;
1568 unsigned long next;
1569
1570 p4d = p4d_offset(pgd, addr);
1571 do {
1572 next = p4d_addr_end(addr, end);
1573 if (p4d_none_or_clear_bad(p4d))
1574 continue;
1575 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1576 } while (p4d++, addr = next, addr != end);
1577
1578 return addr;
1579}
1580
1581void unmap_page_range(struct mmu_gather *tlb,
1582 struct vm_area_struct *vma,
1583 unsigned long addr, unsigned long end,
1584 struct zap_details *details)
1585{
1586 pgd_t *pgd;
1587 unsigned long next;
1588
1589 BUG_ON(addr >= end);
1590 tlb_start_vma(tlb, vma);
1591 pgd = pgd_offset(vma->vm_mm, addr);
1592 do {
1593 next = pgd_addr_end(addr, end);
1594 if (pgd_none_or_clear_bad(pgd))
1595 continue;
1596 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1597 } while (pgd++, addr = next, addr != end);
1598 tlb_end_vma(tlb, vma);
1599}
1600
1601
1602static void unmap_single_vma(struct mmu_gather *tlb,
1603 struct vm_area_struct *vma, unsigned long start_addr,
1604 unsigned long end_addr,
1605 struct zap_details *details)
1606{
1607 unsigned long start = max(vma->vm_start, start_addr);
1608 unsigned long end;
1609
1610 if (start >= vma->vm_end)
1611 return;
1612 end = min(vma->vm_end, end_addr);
1613 if (end <= vma->vm_start)
1614 return;
1615
1616 if (vma->vm_file)
1617 uprobe_munmap(vma, start, end);
1618
1619 if (unlikely(vma->vm_flags & VM_PFNMAP))
1620 untrack_pfn(vma, 0, 0);
1621
1622 if (start != end) {
1623 if (unlikely(is_vm_hugetlb_page(vma))) {
1624 /*
1625 * It is undesirable to test vma->vm_file as it
1626 * should be non-null for valid hugetlb area.
1627 * However, vm_file will be NULL in the error
1628 * cleanup path of mmap_region. When
1629 * hugetlbfs ->mmap method fails,
1630 * mmap_region() nullifies vma->vm_file
1631 * before calling this function to clean up.
1632 * Since no pte has actually been setup, it is
1633 * safe to do nothing in this case.
1634 */
1635 if (vma->vm_file) {
1636 zap_flags_t zap_flags = details ?
1637 details->zap_flags : 0;
1638 __unmap_hugepage_range_final(tlb, vma, start, end,
1639 NULL, zap_flags);
1640 }
1641 } else
1642 unmap_page_range(tlb, vma, start, end, details);
1643 }
1644}
1645
1646/**
1647 * unmap_vmas - unmap a range of memory covered by a list of vma's
1648 * @tlb: address of the caller's struct mmu_gather
1649 * @mt: the maple tree
1650 * @vma: the starting vma
1651 * @start_addr: virtual address at which to start unmapping
1652 * @end_addr: virtual address at which to end unmapping
1653 *
1654 * Unmap all pages in the vma list.
1655 *
1656 * Only addresses between `start' and `end' will be unmapped.
1657 *
1658 * The VMA list must be sorted in ascending virtual address order.
1659 *
1660 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1661 * range after unmap_vmas() returns. So the only responsibility here is to
1662 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1663 * drops the lock and schedules.
1664 */
1665void unmap_vmas(struct mmu_gather *tlb, struct maple_tree *mt,
1666 struct vm_area_struct *vma, unsigned long start_addr,
1667 unsigned long end_addr)
1668{
1669 struct mmu_notifier_range range;
1670 struct zap_details details = {
1671 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
1672 /* Careful - we need to zap private pages too! */
1673 .even_cows = true,
1674 };
1675 MA_STATE(mas, mt, vma->vm_end, vma->vm_end);
1676
1677 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1678 start_addr, end_addr);
1679 mmu_notifier_invalidate_range_start(&range);
1680 do {
1681 unmap_single_vma(tlb, vma, start_addr, end_addr, &details);
1682 } while ((vma = mas_find(&mas, end_addr - 1)) != NULL);
1683 mmu_notifier_invalidate_range_end(&range);
1684}
1685
1686/**
1687 * zap_page_range - remove user pages in a given range
1688 * @vma: vm_area_struct holding the applicable pages
1689 * @start: starting address of pages to zap
1690 * @size: number of bytes to zap
1691 *
1692 * Caller must protect the VMA list
1693 */
1694void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1695 unsigned long size)
1696{
1697 struct maple_tree *mt = &vma->vm_mm->mm_mt;
1698 unsigned long end = start + size;
1699 struct mmu_notifier_range range;
1700 struct mmu_gather tlb;
1701 MA_STATE(mas, mt, vma->vm_end, vma->vm_end);
1702
1703 lru_add_drain();
1704 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1705 start, start + size);
1706 tlb_gather_mmu(&tlb, vma->vm_mm);
1707 update_hiwater_rss(vma->vm_mm);
1708 mmu_notifier_invalidate_range_start(&range);
1709 do {
1710 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1711 } while ((vma = mas_find(&mas, end - 1)) != NULL);
1712 mmu_notifier_invalidate_range_end(&range);
1713 tlb_finish_mmu(&tlb);
1714}
1715
1716/**
1717 * zap_page_range_single - remove user pages in a given range
1718 * @vma: vm_area_struct holding the applicable pages
1719 * @address: starting address of pages to zap
1720 * @size: number of bytes to zap
1721 * @details: details of shared cache invalidation
1722 *
1723 * The range must fit into one VMA.
1724 */
1725void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1726 unsigned long size, struct zap_details *details)
1727{
1728 const unsigned long end = address + size;
1729 struct mmu_notifier_range range;
1730 struct mmu_gather tlb;
1731
1732 lru_add_drain();
1733 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1734 address, end);
1735 if (is_vm_hugetlb_page(vma))
1736 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1737 &range.end);
1738 tlb_gather_mmu(&tlb, vma->vm_mm);
1739 update_hiwater_rss(vma->vm_mm);
1740 mmu_notifier_invalidate_range_start(&range);
1741 /*
1742 * unmap 'address-end' not 'range.start-range.end' as range
1743 * could have been expanded for hugetlb pmd sharing.
1744 */
1745 unmap_single_vma(&tlb, vma, address, end, details);
1746 mmu_notifier_invalidate_range_end(&range);
1747 tlb_finish_mmu(&tlb);
1748}
1749
1750/**
1751 * zap_vma_ptes - remove ptes mapping the vma
1752 * @vma: vm_area_struct holding ptes to be zapped
1753 * @address: starting address of pages to zap
1754 * @size: number of bytes to zap
1755 *
1756 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1757 *
1758 * The entire address range must be fully contained within the vma.
1759 *
1760 */
1761void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1762 unsigned long size)
1763{
1764 if (!range_in_vma(vma, address, address + size) ||
1765 !(vma->vm_flags & VM_PFNMAP))
1766 return;
1767
1768 zap_page_range_single(vma, address, size, NULL);
1769}
1770EXPORT_SYMBOL_GPL(zap_vma_ptes);
1771
1772static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1773{
1774 pgd_t *pgd;
1775 p4d_t *p4d;
1776 pud_t *pud;
1777 pmd_t *pmd;
1778
1779 pgd = pgd_offset(mm, addr);
1780 p4d = p4d_alloc(mm, pgd, addr);
1781 if (!p4d)
1782 return NULL;
1783 pud = pud_alloc(mm, p4d, addr);
1784 if (!pud)
1785 return NULL;
1786 pmd = pmd_alloc(mm, pud, addr);
1787 if (!pmd)
1788 return NULL;
1789
1790 VM_BUG_ON(pmd_trans_huge(*pmd));
1791 return pmd;
1792}
1793
1794pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1795 spinlock_t **ptl)
1796{
1797 pmd_t *pmd = walk_to_pmd(mm, addr);
1798
1799 if (!pmd)
1800 return NULL;
1801 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1802}
1803
1804static int validate_page_before_insert(struct page *page)
1805{
1806 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1807 return -EINVAL;
1808 flush_dcache_page(page);
1809 return 0;
1810}
1811
1812static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1813 unsigned long addr, struct page *page, pgprot_t prot)
1814{
1815 if (!pte_none(*pte))
1816 return -EBUSY;
1817 /* Ok, finally just insert the thing.. */
1818 get_page(page);
1819 inc_mm_counter(vma->vm_mm, mm_counter_file(page));
1820 page_add_file_rmap(page, vma, false);
1821 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
1822 return 0;
1823}
1824
1825/*
1826 * This is the old fallback for page remapping.
1827 *
1828 * For historical reasons, it only allows reserved pages. Only
1829 * old drivers should use this, and they needed to mark their
1830 * pages reserved for the old functions anyway.
1831 */
1832static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1833 struct page *page, pgprot_t prot)
1834{
1835 int retval;
1836 pte_t *pte;
1837 spinlock_t *ptl;
1838
1839 retval = validate_page_before_insert(page);
1840 if (retval)
1841 goto out;
1842 retval = -ENOMEM;
1843 pte = get_locked_pte(vma->vm_mm, addr, &ptl);
1844 if (!pte)
1845 goto out;
1846 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
1847 pte_unmap_unlock(pte, ptl);
1848out:
1849 return retval;
1850}
1851
1852#ifdef pte_index
1853static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
1854 unsigned long addr, struct page *page, pgprot_t prot)
1855{
1856 int err;
1857
1858 if (!page_count(page))
1859 return -EINVAL;
1860 err = validate_page_before_insert(page);
1861 if (err)
1862 return err;
1863 return insert_page_into_pte_locked(vma, pte, addr, page, prot);
1864}
1865
1866/* insert_pages() amortizes the cost of spinlock operations
1867 * when inserting pages in a loop. Arch *must* define pte_index.
1868 */
1869static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1870 struct page **pages, unsigned long *num, pgprot_t prot)
1871{
1872 pmd_t *pmd = NULL;
1873 pte_t *start_pte, *pte;
1874 spinlock_t *pte_lock;
1875 struct mm_struct *const mm = vma->vm_mm;
1876 unsigned long curr_page_idx = 0;
1877 unsigned long remaining_pages_total = *num;
1878 unsigned long pages_to_write_in_pmd;
1879 int ret;
1880more:
1881 ret = -EFAULT;
1882 pmd = walk_to_pmd(mm, addr);
1883 if (!pmd)
1884 goto out;
1885
1886 pages_to_write_in_pmd = min_t(unsigned long,
1887 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1888
1889 /* Allocate the PTE if necessary; takes PMD lock once only. */
1890 ret = -ENOMEM;
1891 if (pte_alloc(mm, pmd))
1892 goto out;
1893
1894 while (pages_to_write_in_pmd) {
1895 int pte_idx = 0;
1896 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1897
1898 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1899 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1900 int err = insert_page_in_batch_locked(vma, pte,
1901 addr, pages[curr_page_idx], prot);
1902 if (unlikely(err)) {
1903 pte_unmap_unlock(start_pte, pte_lock);
1904 ret = err;
1905 remaining_pages_total -= pte_idx;
1906 goto out;
1907 }
1908 addr += PAGE_SIZE;
1909 ++curr_page_idx;
1910 }
1911 pte_unmap_unlock(start_pte, pte_lock);
1912 pages_to_write_in_pmd -= batch_size;
1913 remaining_pages_total -= batch_size;
1914 }
1915 if (remaining_pages_total)
1916 goto more;
1917 ret = 0;
1918out:
1919 *num = remaining_pages_total;
1920 return ret;
1921}
1922#endif /* ifdef pte_index */
1923
1924/**
1925 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1926 * @vma: user vma to map to
1927 * @addr: target start user address of these pages
1928 * @pages: source kernel pages
1929 * @num: in: number of pages to map. out: number of pages that were *not*
1930 * mapped. (0 means all pages were successfully mapped).
1931 *
1932 * Preferred over vm_insert_page() when inserting multiple pages.
1933 *
1934 * In case of error, we may have mapped a subset of the provided
1935 * pages. It is the caller's responsibility to account for this case.
1936 *
1937 * The same restrictions apply as in vm_insert_page().
1938 */
1939int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1940 struct page **pages, unsigned long *num)
1941{
1942#ifdef pte_index
1943 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1944
1945 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1946 return -EFAULT;
1947 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1948 BUG_ON(mmap_read_trylock(vma->vm_mm));
1949 BUG_ON(vma->vm_flags & VM_PFNMAP);
1950 vma->vm_flags |= VM_MIXEDMAP;
1951 }
1952 /* Defer page refcount checking till we're about to map that page. */
1953 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1954#else
1955 unsigned long idx = 0, pgcount = *num;
1956 int err = -EINVAL;
1957
1958 for (; idx < pgcount; ++idx) {
1959 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1960 if (err)
1961 break;
1962 }
1963 *num = pgcount - idx;
1964 return err;
1965#endif /* ifdef pte_index */
1966}
1967EXPORT_SYMBOL(vm_insert_pages);
1968
1969/**
1970 * vm_insert_page - insert single page into user vma
1971 * @vma: user vma to map to
1972 * @addr: target user address of this page
1973 * @page: source kernel page
1974 *
1975 * This allows drivers to insert individual pages they've allocated
1976 * into a user vma.
1977 *
1978 * The page has to be a nice clean _individual_ kernel allocation.
1979 * If you allocate a compound page, you need to have marked it as
1980 * such (__GFP_COMP), or manually just split the page up yourself
1981 * (see split_page()).
1982 *
1983 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1984 * took an arbitrary page protection parameter. This doesn't allow
1985 * that. Your vma protection will have to be set up correctly, which
1986 * means that if you want a shared writable mapping, you'd better
1987 * ask for a shared writable mapping!
1988 *
1989 * The page does not need to be reserved.
1990 *
1991 * Usually this function is called from f_op->mmap() handler
1992 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1993 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1994 * function from other places, for example from page-fault handler.
1995 *
1996 * Return: %0 on success, negative error code otherwise.
1997 */
1998int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1999 struct page *page)
2000{
2001 if (addr < vma->vm_start || addr >= vma->vm_end)
2002 return -EFAULT;
2003 if (!page_count(page))
2004 return -EINVAL;
2005 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2006 BUG_ON(mmap_read_trylock(vma->vm_mm));
2007 BUG_ON(vma->vm_flags & VM_PFNMAP);
2008 vma->vm_flags |= VM_MIXEDMAP;
2009 }
2010 return insert_page(vma, addr, page, vma->vm_page_prot);
2011}
2012EXPORT_SYMBOL(vm_insert_page);
2013
2014/*
2015 * __vm_map_pages - maps range of kernel pages into user vma
2016 * @vma: user vma to map to
2017 * @pages: pointer to array of source kernel pages
2018 * @num: number of pages in page array
2019 * @offset: user's requested vm_pgoff
2020 *
2021 * This allows drivers to map range of kernel pages into a user vma.
2022 *
2023 * Return: 0 on success and error code otherwise.
2024 */
2025static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2026 unsigned long num, unsigned long offset)
2027{
2028 unsigned long count = vma_pages(vma);
2029 unsigned long uaddr = vma->vm_start;
2030 int ret, i;
2031
2032 /* Fail if the user requested offset is beyond the end of the object */
2033 if (offset >= num)
2034 return -ENXIO;
2035
2036 /* Fail if the user requested size exceeds available object size */
2037 if (count > num - offset)
2038 return -ENXIO;
2039
2040 for (i = 0; i < count; i++) {
2041 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2042 if (ret < 0)
2043 return ret;
2044 uaddr += PAGE_SIZE;
2045 }
2046
2047 return 0;
2048}
2049
2050/**
2051 * vm_map_pages - maps range of kernel pages starts with non zero offset
2052 * @vma: user vma to map to
2053 * @pages: pointer to array of source kernel pages
2054 * @num: number of pages in page array
2055 *
2056 * Maps an object consisting of @num pages, catering for the user's
2057 * requested vm_pgoff
2058 *
2059 * If we fail to insert any page into the vma, the function will return
2060 * immediately leaving any previously inserted pages present. Callers
2061 * from the mmap handler may immediately return the error as their caller
2062 * will destroy the vma, removing any successfully inserted pages. Other
2063 * callers should make their own arrangements for calling unmap_region().
2064 *
2065 * Context: Process context. Called by mmap handlers.
2066 * Return: 0 on success and error code otherwise.
2067 */
2068int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2069 unsigned long num)
2070{
2071 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2072}
2073EXPORT_SYMBOL(vm_map_pages);
2074
2075/**
2076 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2077 * @vma: user vma to map to
2078 * @pages: pointer to array of source kernel pages
2079 * @num: number of pages in page array
2080 *
2081 * Similar to vm_map_pages(), except that it explicitly sets the offset
2082 * to 0. This function is intended for the drivers that did not consider
2083 * vm_pgoff.
2084 *
2085 * Context: Process context. Called by mmap handlers.
2086 * Return: 0 on success and error code otherwise.
2087 */
2088int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2089 unsigned long num)
2090{
2091 return __vm_map_pages(vma, pages, num, 0);
2092}
2093EXPORT_SYMBOL(vm_map_pages_zero);
2094
2095static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2096 pfn_t pfn, pgprot_t prot, bool mkwrite)
2097{
2098 struct mm_struct *mm = vma->vm_mm;
2099 pte_t *pte, entry;
2100 spinlock_t *ptl;
2101
2102 pte = get_locked_pte(mm, addr, &ptl);
2103 if (!pte)
2104 return VM_FAULT_OOM;
2105 if (!pte_none(*pte)) {
2106 if (mkwrite) {
2107 /*
2108 * For read faults on private mappings the PFN passed
2109 * in may not match the PFN we have mapped if the
2110 * mapped PFN is a writeable COW page. In the mkwrite
2111 * case we are creating a writable PTE for a shared
2112 * mapping and we expect the PFNs to match. If they
2113 * don't match, we are likely racing with block
2114 * allocation and mapping invalidation so just skip the
2115 * update.
2116 */
2117 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
2118 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
2119 goto out_unlock;
2120 }
2121 entry = pte_mkyoung(*pte);
2122 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2123 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2124 update_mmu_cache(vma, addr, pte);
2125 }
2126 goto out_unlock;
2127 }
2128
2129 /* Ok, finally just insert the thing.. */
2130 if (pfn_t_devmap(pfn))
2131 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2132 else
2133 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2134
2135 if (mkwrite) {
2136 entry = pte_mkyoung(entry);
2137 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2138 }
2139
2140 set_pte_at(mm, addr, pte, entry);
2141 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2142
2143out_unlock:
2144 pte_unmap_unlock(pte, ptl);
2145 return VM_FAULT_NOPAGE;
2146}
2147
2148/**
2149 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2150 * @vma: user vma to map to
2151 * @addr: target user address of this page
2152 * @pfn: source kernel pfn
2153 * @pgprot: pgprot flags for the inserted page
2154 *
2155 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2156 * to override pgprot on a per-page basis.
2157 *
2158 * This only makes sense for IO mappings, and it makes no sense for
2159 * COW mappings. In general, using multiple vmas is preferable;
2160 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2161 * impractical.
2162 *
2163 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2164 * a value of @pgprot different from that of @vma->vm_page_prot.
2165 *
2166 * Context: Process context. May allocate using %GFP_KERNEL.
2167 * Return: vm_fault_t value.
2168 */
2169vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2170 unsigned long pfn, pgprot_t pgprot)
2171{
2172 /*
2173 * Technically, architectures with pte_special can avoid all these
2174 * restrictions (same for remap_pfn_range). However we would like
2175 * consistency in testing and feature parity among all, so we should
2176 * try to keep these invariants in place for everybody.
2177 */
2178 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2179 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2180 (VM_PFNMAP|VM_MIXEDMAP));
2181 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2182 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2183
2184 if (addr < vma->vm_start || addr >= vma->vm_end)
2185 return VM_FAULT_SIGBUS;
2186
2187 if (!pfn_modify_allowed(pfn, pgprot))
2188 return VM_FAULT_SIGBUS;
2189
2190 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2191
2192 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2193 false);
2194}
2195EXPORT_SYMBOL(vmf_insert_pfn_prot);
2196
2197/**
2198 * vmf_insert_pfn - insert single pfn into user vma
2199 * @vma: user vma to map to
2200 * @addr: target user address of this page
2201 * @pfn: source kernel pfn
2202 *
2203 * Similar to vm_insert_page, this allows drivers to insert individual pages
2204 * they've allocated into a user vma. Same comments apply.
2205 *
2206 * This function should only be called from a vm_ops->fault handler, and
2207 * in that case the handler should return the result of this function.
2208 *
2209 * vma cannot be a COW mapping.
2210 *
2211 * As this is called only for pages that do not currently exist, we
2212 * do not need to flush old virtual caches or the TLB.
2213 *
2214 * Context: Process context. May allocate using %GFP_KERNEL.
2215 * Return: vm_fault_t value.
2216 */
2217vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2218 unsigned long pfn)
2219{
2220 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2221}
2222EXPORT_SYMBOL(vmf_insert_pfn);
2223
2224static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2225{
2226 /* these checks mirror the abort conditions in vm_normal_page */
2227 if (vma->vm_flags & VM_MIXEDMAP)
2228 return true;
2229 if (pfn_t_devmap(pfn))
2230 return true;
2231 if (pfn_t_special(pfn))
2232 return true;
2233 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2234 return true;
2235 return false;
2236}
2237
2238static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2239 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2240 bool mkwrite)
2241{
2242 int err;
2243
2244 BUG_ON(!vm_mixed_ok(vma, pfn));
2245
2246 if (addr < vma->vm_start || addr >= vma->vm_end)
2247 return VM_FAULT_SIGBUS;
2248
2249 track_pfn_insert(vma, &pgprot, pfn);
2250
2251 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2252 return VM_FAULT_SIGBUS;
2253
2254 /*
2255 * If we don't have pte special, then we have to use the pfn_valid()
2256 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2257 * refcount the page if pfn_valid is true (hence insert_page rather
2258 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2259 * without pte special, it would there be refcounted as a normal page.
2260 */
2261 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2262 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2263 struct page *page;
2264
2265 /*
2266 * At this point we are committed to insert_page()
2267 * regardless of whether the caller specified flags that
2268 * result in pfn_t_has_page() == false.
2269 */
2270 page = pfn_to_page(pfn_t_to_pfn(pfn));
2271 err = insert_page(vma, addr, page, pgprot);
2272 } else {
2273 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2274 }
2275
2276 if (err == -ENOMEM)
2277 return VM_FAULT_OOM;
2278 if (err < 0 && err != -EBUSY)
2279 return VM_FAULT_SIGBUS;
2280
2281 return VM_FAULT_NOPAGE;
2282}
2283
2284/**
2285 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2286 * @vma: user vma to map to
2287 * @addr: target user address of this page
2288 * @pfn: source kernel pfn
2289 * @pgprot: pgprot flags for the inserted page
2290 *
2291 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2292 * to override pgprot on a per-page basis.
2293 *
2294 * Typically this function should be used by drivers to set caching- and
2295 * encryption bits different than those of @vma->vm_page_prot, because
2296 * the caching- or encryption mode may not be known at mmap() time.
2297 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2298 * to set caching and encryption bits for those vmas (except for COW pages).
2299 * This is ensured by core vm only modifying these page table entries using
2300 * functions that don't touch caching- or encryption bits, using pte_modify()
2301 * if needed. (See for example mprotect()).
2302 * Also when new page-table entries are created, this is only done using the
2303 * fault() callback, and never using the value of vma->vm_page_prot,
2304 * except for page-table entries that point to anonymous pages as the result
2305 * of COW.
2306 *
2307 * Context: Process context. May allocate using %GFP_KERNEL.
2308 * Return: vm_fault_t value.
2309 */
2310vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2311 pfn_t pfn, pgprot_t pgprot)
2312{
2313 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2314}
2315EXPORT_SYMBOL(vmf_insert_mixed_prot);
2316
2317vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2318 pfn_t pfn)
2319{
2320 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2321}
2322EXPORT_SYMBOL(vmf_insert_mixed);
2323
2324/*
2325 * If the insertion of PTE failed because someone else already added a
2326 * different entry in the mean time, we treat that as success as we assume
2327 * the same entry was actually inserted.
2328 */
2329vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2330 unsigned long addr, pfn_t pfn)
2331{
2332 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2333}
2334EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2335
2336/*
2337 * maps a range of physical memory into the requested pages. the old
2338 * mappings are removed. any references to nonexistent pages results
2339 * in null mappings (currently treated as "copy-on-access")
2340 */
2341static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2342 unsigned long addr, unsigned long end,
2343 unsigned long pfn, pgprot_t prot)
2344{
2345 pte_t *pte, *mapped_pte;
2346 spinlock_t *ptl;
2347 int err = 0;
2348
2349 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2350 if (!pte)
2351 return -ENOMEM;
2352 arch_enter_lazy_mmu_mode();
2353 do {
2354 BUG_ON(!pte_none(*pte));
2355 if (!pfn_modify_allowed(pfn, prot)) {
2356 err = -EACCES;
2357 break;
2358 }
2359 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2360 pfn++;
2361 } while (pte++, addr += PAGE_SIZE, addr != end);
2362 arch_leave_lazy_mmu_mode();
2363 pte_unmap_unlock(mapped_pte, ptl);
2364 return err;
2365}
2366
2367static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2368 unsigned long addr, unsigned long end,
2369 unsigned long pfn, pgprot_t prot)
2370{
2371 pmd_t *pmd;
2372 unsigned long next;
2373 int err;
2374
2375 pfn -= addr >> PAGE_SHIFT;
2376 pmd = pmd_alloc(mm, pud, addr);
2377 if (!pmd)
2378 return -ENOMEM;
2379 VM_BUG_ON(pmd_trans_huge(*pmd));
2380 do {
2381 next = pmd_addr_end(addr, end);
2382 err = remap_pte_range(mm, pmd, addr, next,
2383 pfn + (addr >> PAGE_SHIFT), prot);
2384 if (err)
2385 return err;
2386 } while (pmd++, addr = next, addr != end);
2387 return 0;
2388}
2389
2390static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2391 unsigned long addr, unsigned long end,
2392 unsigned long pfn, pgprot_t prot)
2393{
2394 pud_t *pud;
2395 unsigned long next;
2396 int err;
2397
2398 pfn -= addr >> PAGE_SHIFT;
2399 pud = pud_alloc(mm, p4d, addr);
2400 if (!pud)
2401 return -ENOMEM;
2402 do {
2403 next = pud_addr_end(addr, end);
2404 err = remap_pmd_range(mm, pud, addr, next,
2405 pfn + (addr >> PAGE_SHIFT), prot);
2406 if (err)
2407 return err;
2408 } while (pud++, addr = next, addr != end);
2409 return 0;
2410}
2411
2412static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2413 unsigned long addr, unsigned long end,
2414 unsigned long pfn, pgprot_t prot)
2415{
2416 p4d_t *p4d;
2417 unsigned long next;
2418 int err;
2419
2420 pfn -= addr >> PAGE_SHIFT;
2421 p4d = p4d_alloc(mm, pgd, addr);
2422 if (!p4d)
2423 return -ENOMEM;
2424 do {
2425 next = p4d_addr_end(addr, end);
2426 err = remap_pud_range(mm, p4d, addr, next,
2427 pfn + (addr >> PAGE_SHIFT), prot);
2428 if (err)
2429 return err;
2430 } while (p4d++, addr = next, addr != end);
2431 return 0;
2432}
2433
2434/*
2435 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2436 * must have pre-validated the caching bits of the pgprot_t.
2437 */
2438int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2439 unsigned long pfn, unsigned long size, pgprot_t prot)
2440{
2441 pgd_t *pgd;
2442 unsigned long next;
2443 unsigned long end = addr + PAGE_ALIGN(size);
2444 struct mm_struct *mm = vma->vm_mm;
2445 int err;
2446
2447 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2448 return -EINVAL;
2449
2450 /*
2451 * Physically remapped pages are special. Tell the
2452 * rest of the world about it:
2453 * VM_IO tells people not to look at these pages
2454 * (accesses can have side effects).
2455 * VM_PFNMAP tells the core MM that the base pages are just
2456 * raw PFN mappings, and do not have a "struct page" associated
2457 * with them.
2458 * VM_DONTEXPAND
2459 * Disable vma merging and expanding with mremap().
2460 * VM_DONTDUMP
2461 * Omit vma from core dump, even when VM_IO turned off.
2462 *
2463 * There's a horrible special case to handle copy-on-write
2464 * behaviour that some programs depend on. We mark the "original"
2465 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2466 * See vm_normal_page() for details.
2467 */
2468 if (is_cow_mapping(vma->vm_flags)) {
2469 if (addr != vma->vm_start || end != vma->vm_end)
2470 return -EINVAL;
2471 vma->vm_pgoff = pfn;
2472 }
2473
2474 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2475
2476 BUG_ON(addr >= end);
2477 pfn -= addr >> PAGE_SHIFT;
2478 pgd = pgd_offset(mm, addr);
2479 flush_cache_range(vma, addr, end);
2480 do {
2481 next = pgd_addr_end(addr, end);
2482 err = remap_p4d_range(mm, pgd, addr, next,
2483 pfn + (addr >> PAGE_SHIFT), prot);
2484 if (err)
2485 return err;
2486 } while (pgd++, addr = next, addr != end);
2487
2488 return 0;
2489}
2490
2491/**
2492 * remap_pfn_range - remap kernel memory to userspace
2493 * @vma: user vma to map to
2494 * @addr: target page aligned user address to start at
2495 * @pfn: page frame number of kernel physical memory address
2496 * @size: size of mapping area
2497 * @prot: page protection flags for this mapping
2498 *
2499 * Note: this is only safe if the mm semaphore is held when called.
2500 *
2501 * Return: %0 on success, negative error code otherwise.
2502 */
2503int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2504 unsigned long pfn, unsigned long size, pgprot_t prot)
2505{
2506 int err;
2507
2508 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2509 if (err)
2510 return -EINVAL;
2511
2512 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2513 if (err)
2514 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2515 return err;
2516}
2517EXPORT_SYMBOL(remap_pfn_range);
2518
2519/**
2520 * vm_iomap_memory - remap memory to userspace
2521 * @vma: user vma to map to
2522 * @start: start of the physical memory to be mapped
2523 * @len: size of area
2524 *
2525 * This is a simplified io_remap_pfn_range() for common driver use. The
2526 * driver just needs to give us the physical memory range to be mapped,
2527 * we'll figure out the rest from the vma information.
2528 *
2529 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2530 * whatever write-combining details or similar.
2531 *
2532 * Return: %0 on success, negative error code otherwise.
2533 */
2534int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2535{
2536 unsigned long vm_len, pfn, pages;
2537
2538 /* Check that the physical memory area passed in looks valid */
2539 if (start + len < start)
2540 return -EINVAL;
2541 /*
2542 * You *really* shouldn't map things that aren't page-aligned,
2543 * but we've historically allowed it because IO memory might
2544 * just have smaller alignment.
2545 */
2546 len += start & ~PAGE_MASK;
2547 pfn = start >> PAGE_SHIFT;
2548 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2549 if (pfn + pages < pfn)
2550 return -EINVAL;
2551
2552 /* We start the mapping 'vm_pgoff' pages into the area */
2553 if (vma->vm_pgoff > pages)
2554 return -EINVAL;
2555 pfn += vma->vm_pgoff;
2556 pages -= vma->vm_pgoff;
2557
2558 /* Can we fit all of the mapping? */
2559 vm_len = vma->vm_end - vma->vm_start;
2560 if (vm_len >> PAGE_SHIFT > pages)
2561 return -EINVAL;
2562
2563 /* Ok, let it rip */
2564 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2565}
2566EXPORT_SYMBOL(vm_iomap_memory);
2567
2568static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2569 unsigned long addr, unsigned long end,
2570 pte_fn_t fn, void *data, bool create,
2571 pgtbl_mod_mask *mask)
2572{
2573 pte_t *pte, *mapped_pte;
2574 int err = 0;
2575 spinlock_t *ptl;
2576
2577 if (create) {
2578 mapped_pte = pte = (mm == &init_mm) ?
2579 pte_alloc_kernel_track(pmd, addr, mask) :
2580 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2581 if (!pte)
2582 return -ENOMEM;
2583 } else {
2584 mapped_pte = pte = (mm == &init_mm) ?
2585 pte_offset_kernel(pmd, addr) :
2586 pte_offset_map_lock(mm, pmd, addr, &ptl);
2587 }
2588
2589 BUG_ON(pmd_huge(*pmd));
2590
2591 arch_enter_lazy_mmu_mode();
2592
2593 if (fn) {
2594 do {
2595 if (create || !pte_none(*pte)) {
2596 err = fn(pte++, addr, data);
2597 if (err)
2598 break;
2599 }
2600 } while (addr += PAGE_SIZE, addr != end);
2601 }
2602 *mask |= PGTBL_PTE_MODIFIED;
2603
2604 arch_leave_lazy_mmu_mode();
2605
2606 if (mm != &init_mm)
2607 pte_unmap_unlock(mapped_pte, ptl);
2608 return err;
2609}
2610
2611static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2612 unsigned long addr, unsigned long end,
2613 pte_fn_t fn, void *data, bool create,
2614 pgtbl_mod_mask *mask)
2615{
2616 pmd_t *pmd;
2617 unsigned long next;
2618 int err = 0;
2619
2620 BUG_ON(pud_huge(*pud));
2621
2622 if (create) {
2623 pmd = pmd_alloc_track(mm, pud, addr, mask);
2624 if (!pmd)
2625 return -ENOMEM;
2626 } else {
2627 pmd = pmd_offset(pud, addr);
2628 }
2629 do {
2630 next = pmd_addr_end(addr, end);
2631 if (pmd_none(*pmd) && !create)
2632 continue;
2633 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2634 return -EINVAL;
2635 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2636 if (!create)
2637 continue;
2638 pmd_clear_bad(pmd);
2639 }
2640 err = apply_to_pte_range(mm, pmd, addr, next,
2641 fn, data, create, mask);
2642 if (err)
2643 break;
2644 } while (pmd++, addr = next, addr != end);
2645
2646 return err;
2647}
2648
2649static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2650 unsigned long addr, unsigned long end,
2651 pte_fn_t fn, void *data, bool create,
2652 pgtbl_mod_mask *mask)
2653{
2654 pud_t *pud;
2655 unsigned long next;
2656 int err = 0;
2657
2658 if (create) {
2659 pud = pud_alloc_track(mm, p4d, addr, mask);
2660 if (!pud)
2661 return -ENOMEM;
2662 } else {
2663 pud = pud_offset(p4d, addr);
2664 }
2665 do {
2666 next = pud_addr_end(addr, end);
2667 if (pud_none(*pud) && !create)
2668 continue;
2669 if (WARN_ON_ONCE(pud_leaf(*pud)))
2670 return -EINVAL;
2671 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2672 if (!create)
2673 continue;
2674 pud_clear_bad(pud);
2675 }
2676 err = apply_to_pmd_range(mm, pud, addr, next,
2677 fn, data, create, mask);
2678 if (err)
2679 break;
2680 } while (pud++, addr = next, addr != end);
2681
2682 return err;
2683}
2684
2685static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2686 unsigned long addr, unsigned long end,
2687 pte_fn_t fn, void *data, bool create,
2688 pgtbl_mod_mask *mask)
2689{
2690 p4d_t *p4d;
2691 unsigned long next;
2692 int err = 0;
2693
2694 if (create) {
2695 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2696 if (!p4d)
2697 return -ENOMEM;
2698 } else {
2699 p4d = p4d_offset(pgd, addr);
2700 }
2701 do {
2702 next = p4d_addr_end(addr, end);
2703 if (p4d_none(*p4d) && !create)
2704 continue;
2705 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2706 return -EINVAL;
2707 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2708 if (!create)
2709 continue;
2710 p4d_clear_bad(p4d);
2711 }
2712 err = apply_to_pud_range(mm, p4d, addr, next,
2713 fn, data, create, mask);
2714 if (err)
2715 break;
2716 } while (p4d++, addr = next, addr != end);
2717
2718 return err;
2719}
2720
2721static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2722 unsigned long size, pte_fn_t fn,
2723 void *data, bool create)
2724{
2725 pgd_t *pgd;
2726 unsigned long start = addr, next;
2727 unsigned long end = addr + size;
2728 pgtbl_mod_mask mask = 0;
2729 int err = 0;
2730
2731 if (WARN_ON(addr >= end))
2732 return -EINVAL;
2733
2734 pgd = pgd_offset(mm, addr);
2735 do {
2736 next = pgd_addr_end(addr, end);
2737 if (pgd_none(*pgd) && !create)
2738 continue;
2739 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2740 return -EINVAL;
2741 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2742 if (!create)
2743 continue;
2744 pgd_clear_bad(pgd);
2745 }
2746 err = apply_to_p4d_range(mm, pgd, addr, next,
2747 fn, data, create, &mask);
2748 if (err)
2749 break;
2750 } while (pgd++, addr = next, addr != end);
2751
2752 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2753 arch_sync_kernel_mappings(start, start + size);
2754
2755 return err;
2756}
2757
2758/*
2759 * Scan a region of virtual memory, filling in page tables as necessary
2760 * and calling a provided function on each leaf page table.
2761 */
2762int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2763 unsigned long size, pte_fn_t fn, void *data)
2764{
2765 return __apply_to_page_range(mm, addr, size, fn, data, true);
2766}
2767EXPORT_SYMBOL_GPL(apply_to_page_range);
2768
2769/*
2770 * Scan a region of virtual memory, calling a provided function on
2771 * each leaf page table where it exists.
2772 *
2773 * Unlike apply_to_page_range, this does _not_ fill in page tables
2774 * where they are absent.
2775 */
2776int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2777 unsigned long size, pte_fn_t fn, void *data)
2778{
2779 return __apply_to_page_range(mm, addr, size, fn, data, false);
2780}
2781EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2782
2783/*
2784 * handle_pte_fault chooses page fault handler according to an entry which was
2785 * read non-atomically. Before making any commitment, on those architectures
2786 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2787 * parts, do_swap_page must check under lock before unmapping the pte and
2788 * proceeding (but do_wp_page is only called after already making such a check;
2789 * and do_anonymous_page can safely check later on).
2790 */
2791static inline int pte_unmap_same(struct vm_fault *vmf)
2792{
2793 int same = 1;
2794#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2795 if (sizeof(pte_t) > sizeof(unsigned long)) {
2796 spinlock_t *ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
2797 spin_lock(ptl);
2798 same = pte_same(*vmf->pte, vmf->orig_pte);
2799 spin_unlock(ptl);
2800 }
2801#endif
2802 pte_unmap(vmf->pte);
2803 vmf->pte = NULL;
2804 return same;
2805}
2806
2807/*
2808 * Return:
2809 * 0: copied succeeded
2810 * -EHWPOISON: copy failed due to hwpoison in source page
2811 * -EAGAIN: copied failed (some other reason)
2812 */
2813static inline int __wp_page_copy_user(struct page *dst, struct page *src,
2814 struct vm_fault *vmf)
2815{
2816 int ret;
2817 void *kaddr;
2818 void __user *uaddr;
2819 bool locked = false;
2820 struct vm_area_struct *vma = vmf->vma;
2821 struct mm_struct *mm = vma->vm_mm;
2822 unsigned long addr = vmf->address;
2823
2824 if (likely(src)) {
2825 if (copy_mc_user_highpage(dst, src, addr, vma)) {
2826 memory_failure_queue(page_to_pfn(src), 0);
2827 return -EHWPOISON;
2828 }
2829 return 0;
2830 }
2831
2832 /*
2833 * If the source page was a PFN mapping, we don't have
2834 * a "struct page" for it. We do a best-effort copy by
2835 * just copying from the original user address. If that
2836 * fails, we just zero-fill it. Live with it.
2837 */
2838 kaddr = kmap_atomic(dst);
2839 uaddr = (void __user *)(addr & PAGE_MASK);
2840
2841 /*
2842 * On architectures with software "accessed" bits, we would
2843 * take a double page fault, so mark it accessed here.
2844 */
2845 if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) {
2846 pte_t entry;
2847
2848 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2849 locked = true;
2850 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2851 /*
2852 * Other thread has already handled the fault
2853 * and update local tlb only
2854 */
2855 update_mmu_tlb(vma, addr, vmf->pte);
2856 ret = -EAGAIN;
2857 goto pte_unlock;
2858 }
2859
2860 entry = pte_mkyoung(vmf->orig_pte);
2861 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2862 update_mmu_cache(vma, addr, vmf->pte);
2863 }
2864
2865 /*
2866 * This really shouldn't fail, because the page is there
2867 * in the page tables. But it might just be unreadable,
2868 * in which case we just give up and fill the result with
2869 * zeroes.
2870 */
2871 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2872 if (locked)
2873 goto warn;
2874
2875 /* Re-validate under PTL if the page is still mapped */
2876 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2877 locked = true;
2878 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2879 /* The PTE changed under us, update local tlb */
2880 update_mmu_tlb(vma, addr, vmf->pte);
2881 ret = -EAGAIN;
2882 goto pte_unlock;
2883 }
2884
2885 /*
2886 * The same page can be mapped back since last copy attempt.
2887 * Try to copy again under PTL.
2888 */
2889 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2890 /*
2891 * Give a warn in case there can be some obscure
2892 * use-case
2893 */
2894warn:
2895 WARN_ON_ONCE(1);
2896 clear_page(kaddr);
2897 }
2898 }
2899
2900 ret = 0;
2901
2902pte_unlock:
2903 if (locked)
2904 pte_unmap_unlock(vmf->pte, vmf->ptl);
2905 kunmap_atomic(kaddr);
2906 flush_dcache_page(dst);
2907
2908 return ret;
2909}
2910
2911static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2912{
2913 struct file *vm_file = vma->vm_file;
2914
2915 if (vm_file)
2916 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2917
2918 /*
2919 * Special mappings (e.g. VDSO) do not have any file so fake
2920 * a default GFP_KERNEL for them.
2921 */
2922 return GFP_KERNEL;
2923}
2924
2925/*
2926 * Notify the address space that the page is about to become writable so that
2927 * it can prohibit this or wait for the page to get into an appropriate state.
2928 *
2929 * We do this without the lock held, so that it can sleep if it needs to.
2930 */
2931static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2932{
2933 vm_fault_t ret;
2934 struct page *page = vmf->page;
2935 unsigned int old_flags = vmf->flags;
2936
2937 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2938
2939 if (vmf->vma->vm_file &&
2940 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2941 return VM_FAULT_SIGBUS;
2942
2943 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2944 /* Restore original flags so that caller is not surprised */
2945 vmf->flags = old_flags;
2946 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2947 return ret;
2948 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2949 lock_page(page);
2950 if (!page->mapping) {
2951 unlock_page(page);
2952 return 0; /* retry */
2953 }
2954 ret |= VM_FAULT_LOCKED;
2955 } else
2956 VM_BUG_ON_PAGE(!PageLocked(page), page);
2957 return ret;
2958}
2959
2960/*
2961 * Handle dirtying of a page in shared file mapping on a write fault.
2962 *
2963 * The function expects the page to be locked and unlocks it.
2964 */
2965static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2966{
2967 struct vm_area_struct *vma = vmf->vma;
2968 struct address_space *mapping;
2969 struct page *page = vmf->page;
2970 bool dirtied;
2971 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2972
2973 dirtied = set_page_dirty(page);
2974 VM_BUG_ON_PAGE(PageAnon(page), page);
2975 /*
2976 * Take a local copy of the address_space - page.mapping may be zeroed
2977 * by truncate after unlock_page(). The address_space itself remains
2978 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2979 * release semantics to prevent the compiler from undoing this copying.
2980 */
2981 mapping = page_rmapping(page);
2982 unlock_page(page);
2983
2984 if (!page_mkwrite)
2985 file_update_time(vma->vm_file);
2986
2987 /*
2988 * Throttle page dirtying rate down to writeback speed.
2989 *
2990 * mapping may be NULL here because some device drivers do not
2991 * set page.mapping but still dirty their pages
2992 *
2993 * Drop the mmap_lock before waiting on IO, if we can. The file
2994 * is pinning the mapping, as per above.
2995 */
2996 if ((dirtied || page_mkwrite) && mapping) {
2997 struct file *fpin;
2998
2999 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
3000 balance_dirty_pages_ratelimited(mapping);
3001 if (fpin) {
3002 fput(fpin);
3003 return VM_FAULT_COMPLETED;
3004 }
3005 }
3006
3007 return 0;
3008}
3009
3010/*
3011 * Handle write page faults for pages that can be reused in the current vma
3012 *
3013 * This can happen either due to the mapping being with the VM_SHARED flag,
3014 * or due to us being the last reference standing to the page. In either
3015 * case, all we need to do here is to mark the page as writable and update
3016 * any related book-keeping.
3017 */
3018static inline void wp_page_reuse(struct vm_fault *vmf)
3019 __releases(vmf->ptl)
3020{
3021 struct vm_area_struct *vma = vmf->vma;
3022 struct page *page = vmf->page;
3023 pte_t entry;
3024
3025 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3026 VM_BUG_ON(page && PageAnon(page) && !PageAnonExclusive(page));
3027
3028 /*
3029 * Clear the pages cpupid information as the existing
3030 * information potentially belongs to a now completely
3031 * unrelated process.
3032 */
3033 if (page)
3034 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
3035
3036 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3037 entry = pte_mkyoung(vmf->orig_pte);
3038 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3039 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
3040 update_mmu_cache(vma, vmf->address, vmf->pte);
3041 pte_unmap_unlock(vmf->pte, vmf->ptl);
3042 count_vm_event(PGREUSE);
3043}
3044
3045/*
3046 * Handle the case of a page which we actually need to copy to a new page,
3047 * either due to COW or unsharing.
3048 *
3049 * Called with mmap_lock locked and the old page referenced, but
3050 * without the ptl held.
3051 *
3052 * High level logic flow:
3053 *
3054 * - Allocate a page, copy the content of the old page to the new one.
3055 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3056 * - Take the PTL. If the pte changed, bail out and release the allocated page
3057 * - If the pte is still the way we remember it, update the page table and all
3058 * relevant references. This includes dropping the reference the page-table
3059 * held to the old page, as well as updating the rmap.
3060 * - In any case, unlock the PTL and drop the reference we took to the old page.
3061 */
3062static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3063{
3064 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3065 struct vm_area_struct *vma = vmf->vma;
3066 struct mm_struct *mm = vma->vm_mm;
3067 struct page *old_page = vmf->page;
3068 struct page *new_page = NULL;
3069 pte_t entry;
3070 int page_copied = 0;
3071 struct mmu_notifier_range range;
3072 int ret;
3073
3074 delayacct_wpcopy_start();
3075
3076 if (unlikely(anon_vma_prepare(vma)))
3077 goto oom;
3078
3079 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
3080 new_page = alloc_zeroed_user_highpage_movable(vma,
3081 vmf->address);
3082 if (!new_page)
3083 goto oom;
3084 } else {
3085 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3086 vmf->address);
3087 if (!new_page)
3088 goto oom;
3089
3090 ret = __wp_page_copy_user(new_page, old_page, vmf);
3091 if (ret) {
3092 /*
3093 * COW failed, if the fault was solved by other,
3094 * it's fine. If not, userspace would re-fault on
3095 * the same address and we will handle the fault
3096 * from the second attempt.
3097 * The -EHWPOISON case will not be retried.
3098 */
3099 put_page(new_page);
3100 if (old_page)
3101 put_page(old_page);
3102
3103 delayacct_wpcopy_end();
3104 return ret == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3105 }
3106 kmsan_copy_page_meta(new_page, old_page);
3107 }
3108
3109 if (mem_cgroup_charge(page_folio(new_page), mm, GFP_KERNEL))
3110 goto oom_free_new;
3111 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
3112
3113 __SetPageUptodate(new_page);
3114
3115 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
3116 vmf->address & PAGE_MASK,
3117 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3118 mmu_notifier_invalidate_range_start(&range);
3119
3120 /*
3121 * Re-check the pte - we dropped the lock
3122 */
3123 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3124 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
3125 if (old_page) {
3126 if (!PageAnon(old_page)) {
3127 dec_mm_counter(mm, mm_counter_file(old_page));
3128 inc_mm_counter(mm, MM_ANONPAGES);
3129 }
3130 } else {
3131 inc_mm_counter(mm, MM_ANONPAGES);
3132 }
3133 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3134 entry = mk_pte(new_page, vma->vm_page_prot);
3135 entry = pte_sw_mkyoung(entry);
3136 if (unlikely(unshare)) {
3137 if (pte_soft_dirty(vmf->orig_pte))
3138 entry = pte_mksoft_dirty(entry);
3139 if (pte_uffd_wp(vmf->orig_pte))
3140 entry = pte_mkuffd_wp(entry);
3141 } else {
3142 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3143 }
3144
3145 /*
3146 * Clear the pte entry and flush it first, before updating the
3147 * pte with the new entry, to keep TLBs on different CPUs in
3148 * sync. This code used to set the new PTE then flush TLBs, but
3149 * that left a window where the new PTE could be loaded into
3150 * some TLBs while the old PTE remains in others.
3151 */
3152 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
3153 page_add_new_anon_rmap(new_page, vma, vmf->address);
3154 lru_cache_add_inactive_or_unevictable(new_page, vma);
3155 /*
3156 * We call the notify macro here because, when using secondary
3157 * mmu page tables (such as kvm shadow page tables), we want the
3158 * new page to be mapped directly into the secondary page table.
3159 */
3160 BUG_ON(unshare && pte_write(entry));
3161 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3162 update_mmu_cache(vma, vmf->address, vmf->pte);
3163 if (old_page) {
3164 /*
3165 * Only after switching the pte to the new page may
3166 * we remove the mapcount here. Otherwise another
3167 * process may come and find the rmap count decremented
3168 * before the pte is switched to the new page, and
3169 * "reuse" the old page writing into it while our pte
3170 * here still points into it and can be read by other
3171 * threads.
3172 *
3173 * The critical issue is to order this
3174 * page_remove_rmap with the ptp_clear_flush above.
3175 * Those stores are ordered by (if nothing else,)
3176 * the barrier present in the atomic_add_negative
3177 * in page_remove_rmap.
3178 *
3179 * Then the TLB flush in ptep_clear_flush ensures that
3180 * no process can access the old page before the
3181 * decremented mapcount is visible. And the old page
3182 * cannot be reused until after the decremented
3183 * mapcount is visible. So transitively, TLBs to
3184 * old page will be flushed before it can be reused.
3185 */
3186 page_remove_rmap(old_page, vma, false);
3187 }
3188
3189 /* Free the old page.. */
3190 new_page = old_page;
3191 page_copied = 1;
3192 } else {
3193 update_mmu_tlb(vma, vmf->address, vmf->pte);
3194 }
3195
3196 if (new_page)
3197 put_page(new_page);
3198
3199 pte_unmap_unlock(vmf->pte, vmf->ptl);
3200 /*
3201 * No need to double call mmu_notifier->invalidate_range() callback as
3202 * the above ptep_clear_flush_notify() did already call it.
3203 */
3204 mmu_notifier_invalidate_range_only_end(&range);
3205 if (old_page) {
3206 if (page_copied)
3207 free_swap_cache(old_page);
3208 put_page(old_page);
3209 }
3210
3211 delayacct_wpcopy_end();
3212 return 0;
3213oom_free_new:
3214 put_page(new_page);
3215oom:
3216 if (old_page)
3217 put_page(old_page);
3218
3219 delayacct_wpcopy_end();
3220 return VM_FAULT_OOM;
3221}
3222
3223/**
3224 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3225 * writeable once the page is prepared
3226 *
3227 * @vmf: structure describing the fault
3228 *
3229 * This function handles all that is needed to finish a write page fault in a
3230 * shared mapping due to PTE being read-only once the mapped page is prepared.
3231 * It handles locking of PTE and modifying it.
3232 *
3233 * The function expects the page to be locked or other protection against
3234 * concurrent faults / writeback (such as DAX radix tree locks).
3235 *
3236 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3237 * we acquired PTE lock.
3238 */
3239vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3240{
3241 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3242 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3243 &vmf->ptl);
3244 /*
3245 * We might have raced with another page fault while we released the
3246 * pte_offset_map_lock.
3247 */
3248 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3249 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3250 pte_unmap_unlock(vmf->pte, vmf->ptl);
3251 return VM_FAULT_NOPAGE;
3252 }
3253 wp_page_reuse(vmf);
3254 return 0;
3255}
3256
3257/*
3258 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3259 * mapping
3260 */
3261static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3262{
3263 struct vm_area_struct *vma = vmf->vma;
3264
3265 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3266 vm_fault_t ret;
3267
3268 pte_unmap_unlock(vmf->pte, vmf->ptl);
3269 vmf->flags |= FAULT_FLAG_MKWRITE;
3270 ret = vma->vm_ops->pfn_mkwrite(vmf);
3271 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3272 return ret;
3273 return finish_mkwrite_fault(vmf);
3274 }
3275 wp_page_reuse(vmf);
3276 return 0;
3277}
3278
3279static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3280 __releases(vmf->ptl)
3281{
3282 struct vm_area_struct *vma = vmf->vma;
3283 vm_fault_t ret = 0;
3284
3285 get_page(vmf->page);
3286
3287 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3288 vm_fault_t tmp;
3289
3290 pte_unmap_unlock(vmf->pte, vmf->ptl);
3291 tmp = do_page_mkwrite(vmf);
3292 if (unlikely(!tmp || (tmp &
3293 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3294 put_page(vmf->page);
3295 return tmp;
3296 }
3297 tmp = finish_mkwrite_fault(vmf);
3298 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3299 unlock_page(vmf->page);
3300 put_page(vmf->page);
3301 return tmp;
3302 }
3303 } else {
3304 wp_page_reuse(vmf);
3305 lock_page(vmf->page);
3306 }
3307 ret |= fault_dirty_shared_page(vmf);
3308 put_page(vmf->page);
3309
3310 return ret;
3311}
3312
3313/*
3314 * This routine handles present pages, when
3315 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3316 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3317 * (FAULT_FLAG_UNSHARE)
3318 *
3319 * It is done by copying the page to a new address and decrementing the
3320 * shared-page counter for the old page.
3321 *
3322 * Note that this routine assumes that the protection checks have been
3323 * done by the caller (the low-level page fault routine in most cases).
3324 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3325 * done any necessary COW.
3326 *
3327 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3328 * though the page will change only once the write actually happens. This
3329 * avoids a few races, and potentially makes it more efficient.
3330 *
3331 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3332 * but allow concurrent faults), with pte both mapped and locked.
3333 * We return with mmap_lock still held, but pte unmapped and unlocked.
3334 */
3335static vm_fault_t do_wp_page(struct vm_fault *vmf)
3336 __releases(vmf->ptl)
3337{
3338 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3339 struct vm_area_struct *vma = vmf->vma;
3340 struct folio *folio = NULL;
3341
3342 if (likely(!unshare)) {
3343 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3344 pte_unmap_unlock(vmf->pte, vmf->ptl);
3345 return handle_userfault(vmf, VM_UFFD_WP);
3346 }
3347
3348 /*
3349 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3350 * is flushed in this case before copying.
3351 */
3352 if (unlikely(userfaultfd_wp(vmf->vma) &&
3353 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3354 flush_tlb_page(vmf->vma, vmf->address);
3355 }
3356
3357 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3358
3359 /*
3360 * Shared mapping: we are guaranteed to have VM_WRITE and
3361 * FAULT_FLAG_WRITE set at this point.
3362 */
3363 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
3364 /*
3365 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3366 * VM_PFNMAP VMA.
3367 *
3368 * We should not cow pages in a shared writeable mapping.
3369 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3370 */
3371 if (!vmf->page)
3372 return wp_pfn_shared(vmf);
3373 return wp_page_shared(vmf);
3374 }
3375
3376 if (vmf->page)
3377 folio = page_folio(vmf->page);
3378
3379 /*
3380 * Private mapping: create an exclusive anonymous page copy if reuse
3381 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3382 */
3383 if (folio && folio_test_anon(folio)) {
3384 /*
3385 * If the page is exclusive to this process we must reuse the
3386 * page without further checks.
3387 */
3388 if (PageAnonExclusive(vmf->page))
3389 goto reuse;
3390
3391 /*
3392 * We have to verify under folio lock: these early checks are
3393 * just an optimization to avoid locking the folio and freeing
3394 * the swapcache if there is little hope that we can reuse.
3395 *
3396 * KSM doesn't necessarily raise the folio refcount.
3397 */
3398 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3399 goto copy;
3400 if (!folio_test_lru(folio))
3401 /*
3402 * Note: We cannot easily detect+handle references from
3403 * remote LRU pagevecs or references to LRU folios.
3404 */
3405 lru_add_drain();
3406 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
3407 goto copy;
3408 if (!folio_trylock(folio))
3409 goto copy;
3410 if (folio_test_swapcache(folio))
3411 folio_free_swap(folio);
3412 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
3413 folio_unlock(folio);
3414 goto copy;
3415 }
3416 /*
3417 * Ok, we've got the only folio reference from our mapping
3418 * and the folio is locked, it's dark out, and we're wearing
3419 * sunglasses. Hit it.
3420 */
3421 page_move_anon_rmap(vmf->page, vma);
3422 folio_unlock(folio);
3423reuse:
3424 if (unlikely(unshare)) {
3425 pte_unmap_unlock(vmf->pte, vmf->ptl);
3426 return 0;
3427 }
3428 wp_page_reuse(vmf);
3429 return 0;
3430 }
3431copy:
3432 /*
3433 * Ok, we need to copy. Oh, well..
3434 */
3435 if (folio)
3436 folio_get(folio);
3437
3438 pte_unmap_unlock(vmf->pte, vmf->ptl);
3439#ifdef CONFIG_KSM
3440 if (folio && folio_test_ksm(folio))
3441 count_vm_event(COW_KSM);
3442#endif
3443 return wp_page_copy(vmf);
3444}
3445
3446static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3447 unsigned long start_addr, unsigned long end_addr,
3448 struct zap_details *details)
3449{
3450 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3451}
3452
3453static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3454 pgoff_t first_index,
3455 pgoff_t last_index,
3456 struct zap_details *details)
3457{
3458 struct vm_area_struct *vma;
3459 pgoff_t vba, vea, zba, zea;
3460
3461 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3462 vba = vma->vm_pgoff;
3463 vea = vba + vma_pages(vma) - 1;
3464 zba = max(first_index, vba);
3465 zea = min(last_index, vea);
3466
3467 unmap_mapping_range_vma(vma,
3468 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3469 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3470 details);
3471 }
3472}
3473
3474/**
3475 * unmap_mapping_folio() - Unmap single folio from processes.
3476 * @folio: The locked folio to be unmapped.
3477 *
3478 * Unmap this folio from any userspace process which still has it mmaped.
3479 * Typically, for efficiency, the range of nearby pages has already been
3480 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3481 * truncation or invalidation holds the lock on a folio, it may find that
3482 * the page has been remapped again: and then uses unmap_mapping_folio()
3483 * to unmap it finally.
3484 */
3485void unmap_mapping_folio(struct folio *folio)
3486{
3487 struct address_space *mapping = folio->mapping;
3488 struct zap_details details = { };
3489 pgoff_t first_index;
3490 pgoff_t last_index;
3491
3492 VM_BUG_ON(!folio_test_locked(folio));
3493
3494 first_index = folio->index;
3495 last_index = folio->index + folio_nr_pages(folio) - 1;
3496
3497 details.even_cows = false;
3498 details.single_folio = folio;
3499 details.zap_flags = ZAP_FLAG_DROP_MARKER;
3500
3501 i_mmap_lock_read(mapping);
3502 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3503 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3504 last_index, &details);
3505 i_mmap_unlock_read(mapping);
3506}
3507
3508/**
3509 * unmap_mapping_pages() - Unmap pages from processes.
3510 * @mapping: The address space containing pages to be unmapped.
3511 * @start: Index of first page to be unmapped.
3512 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3513 * @even_cows: Whether to unmap even private COWed pages.
3514 *
3515 * Unmap the pages in this address space from any userspace process which
3516 * has them mmaped. Generally, you want to remove COWed pages as well when
3517 * a file is being truncated, but not when invalidating pages from the page
3518 * cache.
3519 */
3520void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3521 pgoff_t nr, bool even_cows)
3522{
3523 struct zap_details details = { };
3524 pgoff_t first_index = start;
3525 pgoff_t last_index = start + nr - 1;
3526
3527 details.even_cows = even_cows;
3528 if (last_index < first_index)
3529 last_index = ULONG_MAX;
3530
3531 i_mmap_lock_read(mapping);
3532 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3533 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3534 last_index, &details);
3535 i_mmap_unlock_read(mapping);
3536}
3537EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3538
3539/**
3540 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3541 * address_space corresponding to the specified byte range in the underlying
3542 * file.
3543 *
3544 * @mapping: the address space containing mmaps to be unmapped.
3545 * @holebegin: byte in first page to unmap, relative to the start of
3546 * the underlying file. This will be rounded down to a PAGE_SIZE
3547 * boundary. Note that this is different from truncate_pagecache(), which
3548 * must keep the partial page. In contrast, we must get rid of
3549 * partial pages.
3550 * @holelen: size of prospective hole in bytes. This will be rounded
3551 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3552 * end of the file.
3553 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3554 * but 0 when invalidating pagecache, don't throw away private data.
3555 */
3556void unmap_mapping_range(struct address_space *mapping,
3557 loff_t const holebegin, loff_t const holelen, int even_cows)
3558{
3559 pgoff_t hba = holebegin >> PAGE_SHIFT;
3560 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3561
3562 /* Check for overflow. */
3563 if (sizeof(holelen) > sizeof(hlen)) {
3564 long long holeend =
3565 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3566 if (holeend & ~(long long)ULONG_MAX)
3567 hlen = ULONG_MAX - hba + 1;
3568 }
3569
3570 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3571}
3572EXPORT_SYMBOL(unmap_mapping_range);
3573
3574/*
3575 * Restore a potential device exclusive pte to a working pte entry
3576 */
3577static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3578{
3579 struct folio *folio = page_folio(vmf->page);
3580 struct vm_area_struct *vma = vmf->vma;
3581 struct mmu_notifier_range range;
3582
3583 if (!folio_lock_or_retry(folio, vma->vm_mm, vmf->flags))
3584 return VM_FAULT_RETRY;
3585 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
3586 vma->vm_mm, vmf->address & PAGE_MASK,
3587 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3588 mmu_notifier_invalidate_range_start(&range);
3589
3590 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3591 &vmf->ptl);
3592 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3593 restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte);
3594
3595 pte_unmap_unlock(vmf->pte, vmf->ptl);
3596 folio_unlock(folio);
3597
3598 mmu_notifier_invalidate_range_end(&range);
3599 return 0;
3600}
3601
3602static inline bool should_try_to_free_swap(struct folio *folio,
3603 struct vm_area_struct *vma,
3604 unsigned int fault_flags)
3605{
3606 if (!folio_test_swapcache(folio))
3607 return false;
3608 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
3609 folio_test_mlocked(folio))
3610 return true;
3611 /*
3612 * If we want to map a page that's in the swapcache writable, we
3613 * have to detect via the refcount if we're really the exclusive
3614 * user. Try freeing the swapcache to get rid of the swapcache
3615 * reference only in case it's likely that we'll be the exlusive user.
3616 */
3617 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
3618 folio_ref_count(folio) == 2;
3619}
3620
3621static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
3622{
3623 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
3624 vmf->address, &vmf->ptl);
3625 /*
3626 * Be careful so that we will only recover a special uffd-wp pte into a
3627 * none pte. Otherwise it means the pte could have changed, so retry.
3628 *
3629 * This should also cover the case where e.g. the pte changed
3630 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_SWAPIN_ERROR.
3631 * So is_pte_marker() check is not enough to safely drop the pte.
3632 */
3633 if (pte_same(vmf->orig_pte, *vmf->pte))
3634 pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
3635 pte_unmap_unlock(vmf->pte, vmf->ptl);
3636 return 0;
3637}
3638
3639/*
3640 * This is actually a page-missing access, but with uffd-wp special pte
3641 * installed. It means this pte was wr-protected before being unmapped.
3642 */
3643static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
3644{
3645 /*
3646 * Just in case there're leftover special ptes even after the region
3647 * got unregistered - we can simply clear them. We can also do that
3648 * proactively when e.g. when we do UFFDIO_UNREGISTER upon some uffd-wp
3649 * ranges, but it should be more efficient to be done lazily here.
3650 */
3651 if (unlikely(!userfaultfd_wp(vmf->vma) || vma_is_anonymous(vmf->vma)))
3652 return pte_marker_clear(vmf);
3653
3654 /* do_fault() can handle pte markers too like none pte */
3655 return do_fault(vmf);
3656}
3657
3658static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
3659{
3660 swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte);
3661 unsigned long marker = pte_marker_get(entry);
3662
3663 /*
3664 * PTE markers should never be empty. If anything weird happened,
3665 * the best thing to do is to kill the process along with its mm.
3666 */
3667 if (WARN_ON_ONCE(!marker))
3668 return VM_FAULT_SIGBUS;
3669
3670 /* Higher priority than uffd-wp when data corrupted */
3671 if (marker & PTE_MARKER_SWAPIN_ERROR)
3672 return VM_FAULT_SIGBUS;
3673
3674 if (pte_marker_entry_uffd_wp(entry))
3675 return pte_marker_handle_uffd_wp(vmf);
3676
3677 /* This is an unknown pte marker */
3678 return VM_FAULT_SIGBUS;
3679}
3680
3681/*
3682 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3683 * but allow concurrent faults), and pte mapped but not yet locked.
3684 * We return with pte unmapped and unlocked.
3685 *
3686 * We return with the mmap_lock locked or unlocked in the same cases
3687 * as does filemap_fault().
3688 */
3689vm_fault_t do_swap_page(struct vm_fault *vmf)
3690{
3691 struct vm_area_struct *vma = vmf->vma;
3692 struct folio *swapcache, *folio = NULL;
3693 struct page *page;
3694 struct swap_info_struct *si = NULL;
3695 rmap_t rmap_flags = RMAP_NONE;
3696 bool exclusive = false;
3697 swp_entry_t entry;
3698 pte_t pte;
3699 int locked;
3700 vm_fault_t ret = 0;
3701 void *shadow = NULL;
3702
3703 if (!pte_unmap_same(vmf))
3704 goto out;
3705
3706 entry = pte_to_swp_entry(vmf->orig_pte);
3707 if (unlikely(non_swap_entry(entry))) {
3708 if (is_migration_entry(entry)) {
3709 migration_entry_wait(vma->vm_mm, vmf->pmd,
3710 vmf->address);
3711 } else if (is_device_exclusive_entry(entry)) {
3712 vmf->page = pfn_swap_entry_to_page(entry);
3713 ret = remove_device_exclusive_entry(vmf);
3714 } else if (is_device_private_entry(entry)) {
3715 vmf->page = pfn_swap_entry_to_page(entry);
3716 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3717 vmf->address, &vmf->ptl);
3718 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3719 spin_unlock(vmf->ptl);
3720 goto out;
3721 }
3722
3723 /*
3724 * Get a page reference while we know the page can't be
3725 * freed.
3726 */
3727 get_page(vmf->page);
3728 pte_unmap_unlock(vmf->pte, vmf->ptl);
3729 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3730 put_page(vmf->page);
3731 } else if (is_hwpoison_entry(entry)) {
3732 ret = VM_FAULT_HWPOISON;
3733 } else if (is_pte_marker_entry(entry)) {
3734 ret = handle_pte_marker(vmf);
3735 } else {
3736 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3737 ret = VM_FAULT_SIGBUS;
3738 }
3739 goto out;
3740 }
3741
3742 /* Prevent swapoff from happening to us. */
3743 si = get_swap_device(entry);
3744 if (unlikely(!si))
3745 goto out;
3746
3747 folio = swap_cache_get_folio(entry, vma, vmf->address);
3748 if (folio)
3749 page = folio_file_page(folio, swp_offset(entry));
3750 swapcache = folio;
3751
3752 if (!folio) {
3753 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3754 __swap_count(entry) == 1) {
3755 /* skip swapcache */
3756 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0,
3757 vma, vmf->address, false);
3758 page = &folio->page;
3759 if (folio) {
3760 __folio_set_locked(folio);
3761 __folio_set_swapbacked(folio);
3762
3763 if (mem_cgroup_swapin_charge_folio(folio,
3764 vma->vm_mm, GFP_KERNEL,
3765 entry)) {
3766 ret = VM_FAULT_OOM;
3767 goto out_page;
3768 }
3769 mem_cgroup_swapin_uncharge_swap(entry);
3770
3771 shadow = get_shadow_from_swap_cache(entry);
3772 if (shadow)
3773 workingset_refault(folio, shadow);
3774
3775 folio_add_lru(folio);
3776
3777 /* To provide entry to swap_readpage() */
3778 folio_set_swap_entry(folio, entry);
3779 swap_readpage(page, true, NULL);
3780 folio->private = NULL;
3781 }
3782 } else {
3783 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3784 vmf);
3785 if (page)
3786 folio = page_folio(page);
3787 swapcache = folio;
3788 }
3789
3790 if (!folio) {
3791 /*
3792 * Back out if somebody else faulted in this pte
3793 * while we released the pte lock.
3794 */
3795 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3796 vmf->address, &vmf->ptl);
3797 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3798 ret = VM_FAULT_OOM;
3799 goto unlock;
3800 }
3801
3802 /* Had to read the page from swap area: Major fault */
3803 ret = VM_FAULT_MAJOR;
3804 count_vm_event(PGMAJFAULT);
3805 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3806 } else if (PageHWPoison(page)) {
3807 /*
3808 * hwpoisoned dirty swapcache pages are kept for killing
3809 * owner processes (which may be unknown at hwpoison time)
3810 */
3811 ret = VM_FAULT_HWPOISON;
3812 goto out_release;
3813 }
3814
3815 locked = folio_lock_or_retry(folio, vma->vm_mm, vmf->flags);
3816
3817 if (!locked) {
3818 ret |= VM_FAULT_RETRY;
3819 goto out_release;
3820 }
3821
3822 if (swapcache) {
3823 /*
3824 * Make sure folio_free_swap() or swapoff did not release the
3825 * swapcache from under us. The page pin, and pte_same test
3826 * below, are not enough to exclude that. Even if it is still
3827 * swapcache, we need to check that the page's swap has not
3828 * changed.
3829 */
3830 if (unlikely(!folio_test_swapcache(folio) ||
3831 page_private(page) != entry.val))
3832 goto out_page;
3833
3834 /*
3835 * KSM sometimes has to copy on read faults, for example, if
3836 * page->index of !PageKSM() pages would be nonlinear inside the
3837 * anon VMA -- PageKSM() is lost on actual swapout.
3838 */
3839 page = ksm_might_need_to_copy(page, vma, vmf->address);
3840 if (unlikely(!page)) {
3841 ret = VM_FAULT_OOM;
3842 goto out_page;
3843 } else if (unlikely(PTR_ERR(page) == -EHWPOISON)) {
3844 ret = VM_FAULT_HWPOISON;
3845 goto out_page;
3846 }
3847 folio = page_folio(page);
3848
3849 /*
3850 * If we want to map a page that's in the swapcache writable, we
3851 * have to detect via the refcount if we're really the exclusive
3852 * owner. Try removing the extra reference from the local LRU
3853 * pagevecs if required.
3854 */
3855 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
3856 !folio_test_ksm(folio) && !folio_test_lru(folio))
3857 lru_add_drain();
3858 }
3859
3860 cgroup_throttle_swaprate(page, GFP_KERNEL);
3861
3862 /*
3863 * Back out if somebody else already faulted in this pte.
3864 */
3865 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3866 &vmf->ptl);
3867 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3868 goto out_nomap;
3869
3870 if (unlikely(!folio_test_uptodate(folio))) {
3871 ret = VM_FAULT_SIGBUS;
3872 goto out_nomap;
3873 }
3874
3875 /*
3876 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
3877 * must never point at an anonymous page in the swapcache that is
3878 * PG_anon_exclusive. Sanity check that this holds and especially, that
3879 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
3880 * check after taking the PT lock and making sure that nobody
3881 * concurrently faulted in this page and set PG_anon_exclusive.
3882 */
3883 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
3884 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
3885
3886 /*
3887 * Check under PT lock (to protect against concurrent fork() sharing
3888 * the swap entry concurrently) for certainly exclusive pages.
3889 */
3890 if (!folio_test_ksm(folio)) {
3891 /*
3892 * Note that pte_swp_exclusive() == false for architectures
3893 * without __HAVE_ARCH_PTE_SWP_EXCLUSIVE.
3894 */
3895 exclusive = pte_swp_exclusive(vmf->orig_pte);
3896 if (folio != swapcache) {
3897 /*
3898 * We have a fresh page that is not exposed to the
3899 * swapcache -> certainly exclusive.
3900 */
3901 exclusive = true;
3902 } else if (exclusive && folio_test_writeback(folio) &&
3903 data_race(si->flags & SWP_STABLE_WRITES)) {
3904 /*
3905 * This is tricky: not all swap backends support
3906 * concurrent page modifications while under writeback.
3907 *
3908 * So if we stumble over such a page in the swapcache
3909 * we must not set the page exclusive, otherwise we can
3910 * map it writable without further checks and modify it
3911 * while still under writeback.
3912 *
3913 * For these problematic swap backends, simply drop the
3914 * exclusive marker: this is perfectly fine as we start
3915 * writeback only if we fully unmapped the page and
3916 * there are no unexpected references on the page after
3917 * unmapping succeeded. After fully unmapped, no
3918 * further GUP references (FOLL_GET and FOLL_PIN) can
3919 * appear, so dropping the exclusive marker and mapping
3920 * it only R/O is fine.
3921 */
3922 exclusive = false;
3923 }
3924 }
3925
3926 /*
3927 * Remove the swap entry and conditionally try to free up the swapcache.
3928 * We're already holding a reference on the page but haven't mapped it
3929 * yet.
3930 */
3931 swap_free(entry);
3932 if (should_try_to_free_swap(folio, vma, vmf->flags))
3933 folio_free_swap(folio);
3934
3935 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
3936 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
3937 pte = mk_pte(page, vma->vm_page_prot);
3938
3939 /*
3940 * Same logic as in do_wp_page(); however, optimize for pages that are
3941 * certainly not shared either because we just allocated them without
3942 * exposing them to the swapcache or because the swap entry indicates
3943 * exclusivity.
3944 */
3945 if (!folio_test_ksm(folio) &&
3946 (exclusive || folio_ref_count(folio) == 1)) {
3947 if (vmf->flags & FAULT_FLAG_WRITE) {
3948 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3949 vmf->flags &= ~FAULT_FLAG_WRITE;
3950 }
3951 rmap_flags |= RMAP_EXCLUSIVE;
3952 }
3953 flush_icache_page(vma, page);
3954 if (pte_swp_soft_dirty(vmf->orig_pte))
3955 pte = pte_mksoft_dirty(pte);
3956 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3957 pte = pte_mkuffd_wp(pte);
3958 pte = pte_wrprotect(pte);
3959 }
3960 vmf->orig_pte = pte;
3961
3962 /* ksm created a completely new copy */
3963 if (unlikely(folio != swapcache && swapcache)) {
3964 page_add_new_anon_rmap(page, vma, vmf->address);
3965 folio_add_lru_vma(folio, vma);
3966 } else {
3967 page_add_anon_rmap(page, vma, vmf->address, rmap_flags);
3968 }
3969
3970 VM_BUG_ON(!folio_test_anon(folio) ||
3971 (pte_write(pte) && !PageAnonExclusive(page)));
3972 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3973 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3974
3975 folio_unlock(folio);
3976 if (folio != swapcache && swapcache) {
3977 /*
3978 * Hold the lock to avoid the swap entry to be reused
3979 * until we take the PT lock for the pte_same() check
3980 * (to avoid false positives from pte_same). For
3981 * further safety release the lock after the swap_free
3982 * so that the swap count won't change under a
3983 * parallel locked swapcache.
3984 */
3985 folio_unlock(swapcache);
3986 folio_put(swapcache);
3987 }
3988
3989 if (vmf->flags & FAULT_FLAG_WRITE) {
3990 ret |= do_wp_page(vmf);
3991 if (ret & VM_FAULT_ERROR)
3992 ret &= VM_FAULT_ERROR;
3993 goto out;
3994 }
3995
3996 /* No need to invalidate - it was non-present before */
3997 update_mmu_cache(vma, vmf->address, vmf->pte);
3998unlock:
3999 pte_unmap_unlock(vmf->pte, vmf->ptl);
4000out:
4001 if (si)
4002 put_swap_device(si);
4003 return ret;
4004out_nomap:
4005 pte_unmap_unlock(vmf->pte, vmf->ptl);
4006out_page:
4007 folio_unlock(folio);
4008out_release:
4009 folio_put(folio);
4010 if (folio != swapcache && swapcache) {
4011 folio_unlock(swapcache);
4012 folio_put(swapcache);
4013 }
4014 if (si)
4015 put_swap_device(si);
4016 return ret;
4017}
4018
4019/*
4020 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4021 * but allow concurrent faults), and pte mapped but not yet locked.
4022 * We return with mmap_lock still held, but pte unmapped and unlocked.
4023 */
4024static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
4025{
4026 struct vm_area_struct *vma = vmf->vma;
4027 struct page *page;
4028 vm_fault_t ret = 0;
4029 pte_t entry;
4030
4031 /* File mapping without ->vm_ops ? */
4032 if (vma->vm_flags & VM_SHARED)
4033 return VM_FAULT_SIGBUS;
4034
4035 /*
4036 * Use pte_alloc() instead of pte_alloc_map(). We can't run
4037 * pte_offset_map() on pmds where a huge pmd might be created
4038 * from a different thread.
4039 *
4040 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
4041 * parallel threads are excluded by other means.
4042 *
4043 * Here we only have mmap_read_lock(mm).
4044 */
4045 if (pte_alloc(vma->vm_mm, vmf->pmd))
4046 return VM_FAULT_OOM;
4047
4048 /* See comment in handle_pte_fault() */
4049 if (unlikely(pmd_trans_unstable(vmf->pmd)))
4050 return 0;
4051
4052 /* Use the zero-page for reads */
4053 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
4054 !mm_forbids_zeropage(vma->vm_mm)) {
4055 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
4056 vma->vm_page_prot));
4057 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4058 vmf->address, &vmf->ptl);
4059 if (!pte_none(*vmf->pte)) {
4060 update_mmu_tlb(vma, vmf->address, vmf->pte);
4061 goto unlock;
4062 }
4063 ret = check_stable_address_space(vma->vm_mm);
4064 if (ret)
4065 goto unlock;
4066 /* Deliver the page fault to userland, check inside PT lock */
4067 if (userfaultfd_missing(vma)) {
4068 pte_unmap_unlock(vmf->pte, vmf->ptl);
4069 return handle_userfault(vmf, VM_UFFD_MISSING);
4070 }
4071 goto setpte;
4072 }
4073
4074 /* Allocate our own private page. */
4075 if (unlikely(anon_vma_prepare(vma)))
4076 goto oom;
4077 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
4078 if (!page)
4079 goto oom;
4080
4081 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
4082 goto oom_free_page;
4083 cgroup_throttle_swaprate(page, GFP_KERNEL);
4084
4085 /*
4086 * The memory barrier inside __SetPageUptodate makes sure that
4087 * preceding stores to the page contents become visible before
4088 * the set_pte_at() write.
4089 */
4090 __SetPageUptodate(page);
4091
4092 entry = mk_pte(page, vma->vm_page_prot);
4093 entry = pte_sw_mkyoung(entry);
4094 if (vma->vm_flags & VM_WRITE)
4095 entry = pte_mkwrite(pte_mkdirty(entry));
4096
4097 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
4098 &vmf->ptl);
4099 if (!pte_none(*vmf->pte)) {
4100 update_mmu_tlb(vma, vmf->address, vmf->pte);
4101 goto release;
4102 }
4103
4104 ret = check_stable_address_space(vma->vm_mm);
4105 if (ret)
4106 goto release;
4107
4108 /* Deliver the page fault to userland, check inside PT lock */
4109 if (userfaultfd_missing(vma)) {
4110 pte_unmap_unlock(vmf->pte, vmf->ptl);
4111 put_page(page);
4112 return handle_userfault(vmf, VM_UFFD_MISSING);
4113 }
4114
4115 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4116 page_add_new_anon_rmap(page, vma, vmf->address);
4117 lru_cache_add_inactive_or_unevictable(page, vma);
4118setpte:
4119 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
4120
4121 /* No need to invalidate - it was non-present before */
4122 update_mmu_cache(vma, vmf->address, vmf->pte);
4123unlock:
4124 pte_unmap_unlock(vmf->pte, vmf->ptl);
4125 return ret;
4126release:
4127 put_page(page);
4128 goto unlock;
4129oom_free_page:
4130 put_page(page);
4131oom:
4132 return VM_FAULT_OOM;
4133}
4134
4135/*
4136 * The mmap_lock must have been held on entry, and may have been
4137 * released depending on flags and vma->vm_ops->fault() return value.
4138 * See filemap_fault() and __lock_page_retry().
4139 */
4140static vm_fault_t __do_fault(struct vm_fault *vmf)
4141{
4142 struct vm_area_struct *vma = vmf->vma;
4143 vm_fault_t ret;
4144
4145 /*
4146 * Preallocate pte before we take page_lock because this might lead to
4147 * deadlocks for memcg reclaim which waits for pages under writeback:
4148 * lock_page(A)
4149 * SetPageWriteback(A)
4150 * unlock_page(A)
4151 * lock_page(B)
4152 * lock_page(B)
4153 * pte_alloc_one
4154 * shrink_page_list
4155 * wait_on_page_writeback(A)
4156 * SetPageWriteback(B)
4157 * unlock_page(B)
4158 * # flush A, B to clear the writeback
4159 */
4160 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
4161 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4162 if (!vmf->prealloc_pte)
4163 return VM_FAULT_OOM;
4164 }
4165
4166 ret = vma->vm_ops->fault(vmf);
4167 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
4168 VM_FAULT_DONE_COW)))
4169 return ret;
4170
4171 if (unlikely(PageHWPoison(vmf->page))) {
4172 struct page *page = vmf->page;
4173 vm_fault_t poisonret = VM_FAULT_HWPOISON;
4174 if (ret & VM_FAULT_LOCKED) {
4175 if (page_mapped(page))
4176 unmap_mapping_pages(page_mapping(page),
4177 page->index, 1, false);
4178 /* Retry if a clean page was removed from the cache. */
4179 if (invalidate_inode_page(page))
4180 poisonret = VM_FAULT_NOPAGE;
4181 unlock_page(page);
4182 }
4183 put_page(page);
4184 vmf->page = NULL;
4185 return poisonret;
4186 }
4187
4188 if (unlikely(!(ret & VM_FAULT_LOCKED)))
4189 lock_page(vmf->page);
4190 else
4191 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
4192
4193 return ret;
4194}
4195
4196#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4197static void deposit_prealloc_pte(struct vm_fault *vmf)
4198{
4199 struct vm_area_struct *vma = vmf->vma;
4200
4201 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4202 /*
4203 * We are going to consume the prealloc table,
4204 * count that as nr_ptes.
4205 */
4206 mm_inc_nr_ptes(vma->vm_mm);
4207 vmf->prealloc_pte = NULL;
4208}
4209
4210vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4211{
4212 struct vm_area_struct *vma = vmf->vma;
4213 bool write = vmf->flags & FAULT_FLAG_WRITE;
4214 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
4215 pmd_t entry;
4216 int i;
4217 vm_fault_t ret = VM_FAULT_FALLBACK;
4218
4219 if (!transhuge_vma_suitable(vma, haddr))
4220 return ret;
4221
4222 page = compound_head(page);
4223 if (compound_order(page) != HPAGE_PMD_ORDER)
4224 return ret;
4225
4226 /*
4227 * Just backoff if any subpage of a THP is corrupted otherwise
4228 * the corrupted page may mapped by PMD silently to escape the
4229 * check. This kind of THP just can be PTE mapped. Access to
4230 * the corrupted subpage should trigger SIGBUS as expected.
4231 */
4232 if (unlikely(PageHasHWPoisoned(page)))
4233 return ret;
4234
4235 /*
4236 * Archs like ppc64 need additional space to store information
4237 * related to pte entry. Use the preallocated table for that.
4238 */
4239 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
4240 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4241 if (!vmf->prealloc_pte)
4242 return VM_FAULT_OOM;
4243 }
4244
4245 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4246 if (unlikely(!pmd_none(*vmf->pmd)))
4247 goto out;
4248
4249 for (i = 0; i < HPAGE_PMD_NR; i++)
4250 flush_icache_page(vma, page + i);
4251
4252 entry = mk_huge_pmd(page, vma->vm_page_prot);
4253 if (write)
4254 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
4255
4256 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
4257 page_add_file_rmap(page, vma, true);
4258
4259 /*
4260 * deposit and withdraw with pmd lock held
4261 */
4262 if (arch_needs_pgtable_deposit())
4263 deposit_prealloc_pte(vmf);
4264
4265 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
4266
4267 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
4268
4269 /* fault is handled */
4270 ret = 0;
4271 count_vm_event(THP_FILE_MAPPED);
4272out:
4273 spin_unlock(vmf->ptl);
4274 return ret;
4275}
4276#else
4277vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4278{
4279 return VM_FAULT_FALLBACK;
4280}
4281#endif
4282
4283void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
4284{
4285 struct vm_area_struct *vma = vmf->vma;
4286 bool uffd_wp = pte_marker_uffd_wp(vmf->orig_pte);
4287 bool write = vmf->flags & FAULT_FLAG_WRITE;
4288 bool prefault = vmf->address != addr;
4289 pte_t entry;
4290
4291 flush_icache_page(vma, page);
4292 entry = mk_pte(page, vma->vm_page_prot);
4293
4294 if (prefault && arch_wants_old_prefaulted_pte())
4295 entry = pte_mkold(entry);
4296 else
4297 entry = pte_sw_mkyoung(entry);
4298
4299 if (write)
4300 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4301 if (unlikely(uffd_wp))
4302 entry = pte_mkuffd_wp(pte_wrprotect(entry));
4303 /* copy-on-write page */
4304 if (write && !(vma->vm_flags & VM_SHARED)) {
4305 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4306 page_add_new_anon_rmap(page, vma, addr);
4307 lru_cache_add_inactive_or_unevictable(page, vma);
4308 } else {
4309 inc_mm_counter(vma->vm_mm, mm_counter_file(page));
4310 page_add_file_rmap(page, vma, false);
4311 }
4312 set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
4313}
4314
4315static bool vmf_pte_changed(struct vm_fault *vmf)
4316{
4317 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
4318 return !pte_same(*vmf->pte, vmf->orig_pte);
4319
4320 return !pte_none(*vmf->pte);
4321}
4322
4323/**
4324 * finish_fault - finish page fault once we have prepared the page to fault
4325 *
4326 * @vmf: structure describing the fault
4327 *
4328 * This function handles all that is needed to finish a page fault once the
4329 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4330 * given page, adds reverse page mapping, handles memcg charges and LRU
4331 * addition.
4332 *
4333 * The function expects the page to be locked and on success it consumes a
4334 * reference of a page being mapped (for the PTE which maps it).
4335 *
4336 * Return: %0 on success, %VM_FAULT_ code in case of error.
4337 */
4338vm_fault_t finish_fault(struct vm_fault *vmf)
4339{
4340 struct vm_area_struct *vma = vmf->vma;
4341 struct page *page;
4342 vm_fault_t ret;
4343
4344 /* Did we COW the page? */
4345 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4346 page = vmf->cow_page;
4347 else
4348 page = vmf->page;
4349
4350 /*
4351 * check even for read faults because we might have lost our CoWed
4352 * page
4353 */
4354 if (!(vma->vm_flags & VM_SHARED)) {
4355 ret = check_stable_address_space(vma->vm_mm);
4356 if (ret)
4357 return ret;
4358 }
4359
4360 if (pmd_none(*vmf->pmd)) {
4361 if (PageTransCompound(page)) {
4362 ret = do_set_pmd(vmf, page);
4363 if (ret != VM_FAULT_FALLBACK)
4364 return ret;
4365 }
4366
4367 if (vmf->prealloc_pte)
4368 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4369 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4370 return VM_FAULT_OOM;
4371 }
4372
4373 /*
4374 * See comment in handle_pte_fault() for how this scenario happens, we
4375 * need to return NOPAGE so that we drop this page.
4376 */
4377 if (pmd_devmap_trans_unstable(vmf->pmd))
4378 return VM_FAULT_NOPAGE;
4379
4380 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4381 vmf->address, &vmf->ptl);
4382
4383 /* Re-check under ptl */
4384 if (likely(!vmf_pte_changed(vmf))) {
4385 do_set_pte(vmf, page, vmf->address);
4386
4387 /* no need to invalidate: a not-present page won't be cached */
4388 update_mmu_cache(vma, vmf->address, vmf->pte);
4389
4390 ret = 0;
4391 } else {
4392 update_mmu_tlb(vma, vmf->address, vmf->pte);
4393 ret = VM_FAULT_NOPAGE;
4394 }
4395
4396 pte_unmap_unlock(vmf->pte, vmf->ptl);
4397 return ret;
4398}
4399
4400static unsigned long fault_around_bytes __read_mostly =
4401 rounddown_pow_of_two(65536);
4402
4403#ifdef CONFIG_DEBUG_FS
4404static int fault_around_bytes_get(void *data, u64 *val)
4405{
4406 *val = fault_around_bytes;
4407 return 0;
4408}
4409
4410/*
4411 * fault_around_bytes must be rounded down to the nearest page order as it's
4412 * what do_fault_around() expects to see.
4413 */
4414static int fault_around_bytes_set(void *data, u64 val)
4415{
4416 if (val / PAGE_SIZE > PTRS_PER_PTE)
4417 return -EINVAL;
4418 if (val > PAGE_SIZE)
4419 fault_around_bytes = rounddown_pow_of_two(val);
4420 else
4421 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
4422 return 0;
4423}
4424DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4425 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4426
4427static int __init fault_around_debugfs(void)
4428{
4429 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4430 &fault_around_bytes_fops);
4431 return 0;
4432}
4433late_initcall(fault_around_debugfs);
4434#endif
4435
4436/*
4437 * do_fault_around() tries to map few pages around the fault address. The hope
4438 * is that the pages will be needed soon and this will lower the number of
4439 * faults to handle.
4440 *
4441 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4442 * not ready to be mapped: not up-to-date, locked, etc.
4443 *
4444 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4445 * only once.
4446 *
4447 * fault_around_bytes defines how many bytes we'll try to map.
4448 * do_fault_around() expects it to be set to a power of two less than or equal
4449 * to PTRS_PER_PTE.
4450 *
4451 * The virtual address of the area that we map is naturally aligned to
4452 * fault_around_bytes rounded down to the machine page size
4453 * (and therefore to page order). This way it's easier to guarantee
4454 * that we don't cross page table boundaries.
4455 */
4456static vm_fault_t do_fault_around(struct vm_fault *vmf)
4457{
4458 unsigned long address = vmf->address, nr_pages, mask;
4459 pgoff_t start_pgoff = vmf->pgoff;
4460 pgoff_t end_pgoff;
4461 int off;
4462
4463 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4464 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4465
4466 address = max(address & mask, vmf->vma->vm_start);
4467 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4468 start_pgoff -= off;
4469
4470 /*
4471 * end_pgoff is either the end of the page table, the end of
4472 * the vma or nr_pages from start_pgoff, depending what is nearest.
4473 */
4474 end_pgoff = start_pgoff -
4475 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4476 PTRS_PER_PTE - 1;
4477 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4478 start_pgoff + nr_pages - 1);
4479
4480 if (pmd_none(*vmf->pmd)) {
4481 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4482 if (!vmf->prealloc_pte)
4483 return VM_FAULT_OOM;
4484 }
4485
4486 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4487}
4488
4489/* Return true if we should do read fault-around, false otherwise */
4490static inline bool should_fault_around(struct vm_fault *vmf)
4491{
4492 /* No ->map_pages? No way to fault around... */
4493 if (!vmf->vma->vm_ops->map_pages)
4494 return false;
4495
4496 if (uffd_disable_fault_around(vmf->vma))
4497 return false;
4498
4499 return fault_around_bytes >> PAGE_SHIFT > 1;
4500}
4501
4502static vm_fault_t do_read_fault(struct vm_fault *vmf)
4503{
4504 vm_fault_t ret = 0;
4505
4506 /*
4507 * Let's call ->map_pages() first and use ->fault() as fallback
4508 * if page by the offset is not ready to be mapped (cold cache or
4509 * something).
4510 */
4511 if (should_fault_around(vmf)) {
4512 ret = do_fault_around(vmf);
4513 if (ret)
4514 return ret;
4515 }
4516
4517 ret = __do_fault(vmf);
4518 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4519 return ret;
4520
4521 ret |= finish_fault(vmf);
4522 unlock_page(vmf->page);
4523 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4524 put_page(vmf->page);
4525 return ret;
4526}
4527
4528static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4529{
4530 struct vm_area_struct *vma = vmf->vma;
4531 vm_fault_t ret;
4532
4533 if (unlikely(anon_vma_prepare(vma)))
4534 return VM_FAULT_OOM;
4535
4536 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4537 if (!vmf->cow_page)
4538 return VM_FAULT_OOM;
4539
4540 if (mem_cgroup_charge(page_folio(vmf->cow_page), vma->vm_mm,
4541 GFP_KERNEL)) {
4542 put_page(vmf->cow_page);
4543 return VM_FAULT_OOM;
4544 }
4545 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4546
4547 ret = __do_fault(vmf);
4548 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4549 goto uncharge_out;
4550 if (ret & VM_FAULT_DONE_COW)
4551 return ret;
4552
4553 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4554 __SetPageUptodate(vmf->cow_page);
4555
4556 ret |= finish_fault(vmf);
4557 unlock_page(vmf->page);
4558 put_page(vmf->page);
4559 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4560 goto uncharge_out;
4561 return ret;
4562uncharge_out:
4563 put_page(vmf->cow_page);
4564 return ret;
4565}
4566
4567static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4568{
4569 struct vm_area_struct *vma = vmf->vma;
4570 vm_fault_t ret, tmp;
4571
4572 ret = __do_fault(vmf);
4573 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4574 return ret;
4575
4576 /*
4577 * Check if the backing address space wants to know that the page is
4578 * about to become writable
4579 */
4580 if (vma->vm_ops->page_mkwrite) {
4581 unlock_page(vmf->page);
4582 tmp = do_page_mkwrite(vmf);
4583 if (unlikely(!tmp ||
4584 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4585 put_page(vmf->page);
4586 return tmp;
4587 }
4588 }
4589
4590 ret |= finish_fault(vmf);
4591 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4592 VM_FAULT_RETRY))) {
4593 unlock_page(vmf->page);
4594 put_page(vmf->page);
4595 return ret;
4596 }
4597
4598 ret |= fault_dirty_shared_page(vmf);
4599 return ret;
4600}
4601
4602/*
4603 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4604 * but allow concurrent faults).
4605 * The mmap_lock may have been released depending on flags and our
4606 * return value. See filemap_fault() and __folio_lock_or_retry().
4607 * If mmap_lock is released, vma may become invalid (for example
4608 * by other thread calling munmap()).
4609 */
4610static vm_fault_t do_fault(struct vm_fault *vmf)
4611{
4612 struct vm_area_struct *vma = vmf->vma;
4613 struct mm_struct *vm_mm = vma->vm_mm;
4614 vm_fault_t ret;
4615
4616 /*
4617 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4618 */
4619 if (!vma->vm_ops->fault) {
4620 /*
4621 * If we find a migration pmd entry or a none pmd entry, which
4622 * should never happen, return SIGBUS
4623 */
4624 if (unlikely(!pmd_present(*vmf->pmd)))
4625 ret = VM_FAULT_SIGBUS;
4626 else {
4627 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4628 vmf->pmd,
4629 vmf->address,
4630 &vmf->ptl);
4631 /*
4632 * Make sure this is not a temporary clearing of pte
4633 * by holding ptl and checking again. A R/M/W update
4634 * of pte involves: take ptl, clearing the pte so that
4635 * we don't have concurrent modification by hardware
4636 * followed by an update.
4637 */
4638 if (unlikely(pte_none(*vmf->pte)))
4639 ret = VM_FAULT_SIGBUS;
4640 else
4641 ret = VM_FAULT_NOPAGE;
4642
4643 pte_unmap_unlock(vmf->pte, vmf->ptl);
4644 }
4645 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4646 ret = do_read_fault(vmf);
4647 else if (!(vma->vm_flags & VM_SHARED))
4648 ret = do_cow_fault(vmf);
4649 else
4650 ret = do_shared_fault(vmf);
4651
4652 /* preallocated pagetable is unused: free it */
4653 if (vmf->prealloc_pte) {
4654 pte_free(vm_mm, vmf->prealloc_pte);
4655 vmf->prealloc_pte = NULL;
4656 }
4657 return ret;
4658}
4659
4660int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4661 unsigned long addr, int page_nid, int *flags)
4662{
4663 get_page(page);
4664
4665 count_vm_numa_event(NUMA_HINT_FAULTS);
4666 if (page_nid == numa_node_id()) {
4667 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4668 *flags |= TNF_FAULT_LOCAL;
4669 }
4670
4671 return mpol_misplaced(page, vma, addr);
4672}
4673
4674static vm_fault_t do_numa_page(struct vm_fault *vmf)
4675{
4676 struct vm_area_struct *vma = vmf->vma;
4677 struct page *page = NULL;
4678 int page_nid = NUMA_NO_NODE;
4679 bool writable = false;
4680 int last_cpupid;
4681 int target_nid;
4682 pte_t pte, old_pte;
4683 int flags = 0;
4684
4685 /*
4686 * The "pte" at this point cannot be used safely without
4687 * validation through pte_unmap_same(). It's of NUMA type but
4688 * the pfn may be screwed if the read is non atomic.
4689 */
4690 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4691 spin_lock(vmf->ptl);
4692 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4693 pte_unmap_unlock(vmf->pte, vmf->ptl);
4694 goto out;
4695 }
4696
4697 /* Get the normal PTE */
4698 old_pte = ptep_get(vmf->pte);
4699 pte = pte_modify(old_pte, vma->vm_page_prot);
4700
4701 /*
4702 * Detect now whether the PTE could be writable; this information
4703 * is only valid while holding the PT lock.
4704 */
4705 writable = pte_write(pte);
4706 if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
4707 can_change_pte_writable(vma, vmf->address, pte))
4708 writable = true;
4709
4710 page = vm_normal_page(vma, vmf->address, pte);
4711 if (!page || is_zone_device_page(page))
4712 goto out_map;
4713
4714 /* TODO: handle PTE-mapped THP */
4715 if (PageCompound(page))
4716 goto out_map;
4717
4718 /*
4719 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4720 * much anyway since they can be in shared cache state. This misses
4721 * the case where a mapping is writable but the process never writes
4722 * to it but pte_write gets cleared during protection updates and
4723 * pte_dirty has unpredictable behaviour between PTE scan updates,
4724 * background writeback, dirty balancing and application behaviour.
4725 */
4726 if (!writable)
4727 flags |= TNF_NO_GROUP;
4728
4729 /*
4730 * Flag if the page is shared between multiple address spaces. This
4731 * is later used when determining whether to group tasks together
4732 */
4733 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4734 flags |= TNF_SHARED;
4735
4736 page_nid = page_to_nid(page);
4737 /*
4738 * For memory tiering mode, cpupid of slow memory page is used
4739 * to record page access time. So use default value.
4740 */
4741 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4742 !node_is_toptier(page_nid))
4743 last_cpupid = (-1 & LAST_CPUPID_MASK);
4744 else
4745 last_cpupid = page_cpupid_last(page);
4746 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4747 &flags);
4748 if (target_nid == NUMA_NO_NODE) {
4749 put_page(page);
4750 goto out_map;
4751 }
4752 pte_unmap_unlock(vmf->pte, vmf->ptl);
4753 writable = false;
4754
4755 /* Migrate to the requested node */
4756 if (migrate_misplaced_page(page, vma, target_nid)) {
4757 page_nid = target_nid;
4758 flags |= TNF_MIGRATED;
4759 } else {
4760 flags |= TNF_MIGRATE_FAIL;
4761 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4762 spin_lock(vmf->ptl);
4763 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4764 pte_unmap_unlock(vmf->pte, vmf->ptl);
4765 goto out;
4766 }
4767 goto out_map;
4768 }
4769
4770out:
4771 if (page_nid != NUMA_NO_NODE)
4772 task_numa_fault(last_cpupid, page_nid, 1, flags);
4773 return 0;
4774out_map:
4775 /*
4776 * Make it present again, depending on how arch implements
4777 * non-accessible ptes, some can allow access by kernel mode.
4778 */
4779 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4780 pte = pte_modify(old_pte, vma->vm_page_prot);
4781 pte = pte_mkyoung(pte);
4782 if (writable)
4783 pte = pte_mkwrite(pte);
4784 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4785 update_mmu_cache(vma, vmf->address, vmf->pte);
4786 pte_unmap_unlock(vmf->pte, vmf->ptl);
4787 goto out;
4788}
4789
4790static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4791{
4792 if (vma_is_anonymous(vmf->vma))
4793 return do_huge_pmd_anonymous_page(vmf);
4794 if (vmf->vma->vm_ops->huge_fault)
4795 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4796 return VM_FAULT_FALLBACK;
4797}
4798
4799/* `inline' is required to avoid gcc 4.1.2 build error */
4800static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4801{
4802 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
4803 vm_fault_t ret;
4804
4805 if (vma_is_anonymous(vmf->vma)) {
4806 if (likely(!unshare) &&
4807 userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd))
4808 return handle_userfault(vmf, VM_UFFD_WP);
4809 return do_huge_pmd_wp_page(vmf);
4810 }
4811
4812 if (vmf->vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
4813 if (vmf->vma->vm_ops->huge_fault) {
4814 ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4815 if (!(ret & VM_FAULT_FALLBACK))
4816 return ret;
4817 }
4818 }
4819
4820 /* COW or write-notify handled on pte level: split pmd. */
4821 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4822
4823 return VM_FAULT_FALLBACK;
4824}
4825
4826static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4827{
4828#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4829 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4830 /* No support for anonymous transparent PUD pages yet */
4831 if (vma_is_anonymous(vmf->vma))
4832 return VM_FAULT_FALLBACK;
4833 if (vmf->vma->vm_ops->huge_fault)
4834 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4835#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4836 return VM_FAULT_FALLBACK;
4837}
4838
4839static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4840{
4841#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4842 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4843 vm_fault_t ret;
4844
4845 /* No support for anonymous transparent PUD pages yet */
4846 if (vma_is_anonymous(vmf->vma))
4847 goto split;
4848 if (vmf->vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
4849 if (vmf->vma->vm_ops->huge_fault) {
4850 ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4851 if (!(ret & VM_FAULT_FALLBACK))
4852 return ret;
4853 }
4854 }
4855split:
4856 /* COW or write-notify not handled on PUD level: split pud.*/
4857 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4858#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
4859 return VM_FAULT_FALLBACK;
4860}
4861
4862/*
4863 * These routines also need to handle stuff like marking pages dirty
4864 * and/or accessed for architectures that don't do it in hardware (most
4865 * RISC architectures). The early dirtying is also good on the i386.
4866 *
4867 * There is also a hook called "update_mmu_cache()" that architectures
4868 * with external mmu caches can use to update those (ie the Sparc or
4869 * PowerPC hashed page tables that act as extended TLBs).
4870 *
4871 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4872 * concurrent faults).
4873 *
4874 * The mmap_lock may have been released depending on flags and our return value.
4875 * See filemap_fault() and __folio_lock_or_retry().
4876 */
4877static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4878{
4879 pte_t entry;
4880
4881 if (unlikely(pmd_none(*vmf->pmd))) {
4882 /*
4883 * Leave __pte_alloc() until later: because vm_ops->fault may
4884 * want to allocate huge page, and if we expose page table
4885 * for an instant, it will be difficult to retract from
4886 * concurrent faults and from rmap lookups.
4887 */
4888 vmf->pte = NULL;
4889 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
4890 } else {
4891 /*
4892 * If a huge pmd materialized under us just retry later. Use
4893 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4894 * of pmd_trans_huge() to ensure the pmd didn't become
4895 * pmd_trans_huge under us and then back to pmd_none, as a
4896 * result of MADV_DONTNEED running immediately after a huge pmd
4897 * fault in a different thread of this mm, in turn leading to a
4898 * misleading pmd_trans_huge() retval. All we have to ensure is
4899 * that it is a regular pmd that we can walk with
4900 * pte_offset_map() and we can do that through an atomic read
4901 * in C, which is what pmd_trans_unstable() provides.
4902 */
4903 if (pmd_devmap_trans_unstable(vmf->pmd))
4904 return 0;
4905 /*
4906 * A regular pmd is established and it can't morph into a huge
4907 * pmd from under us anymore at this point because we hold the
4908 * mmap_lock read mode and khugepaged takes it in write mode.
4909 * So now it's safe to run pte_offset_map().
4910 */
4911 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4912 vmf->orig_pte = *vmf->pte;
4913 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
4914
4915 /*
4916 * some architectures can have larger ptes than wordsize,
4917 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4918 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4919 * accesses. The code below just needs a consistent view
4920 * for the ifs and we later double check anyway with the
4921 * ptl lock held. So here a barrier will do.
4922 */
4923 barrier();
4924 if (pte_none(vmf->orig_pte)) {
4925 pte_unmap(vmf->pte);
4926 vmf->pte = NULL;
4927 }
4928 }
4929
4930 if (!vmf->pte) {
4931 if (vma_is_anonymous(vmf->vma))
4932 return do_anonymous_page(vmf);
4933 else
4934 return do_fault(vmf);
4935 }
4936
4937 if (!pte_present(vmf->orig_pte))
4938 return do_swap_page(vmf);
4939
4940 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4941 return do_numa_page(vmf);
4942
4943 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4944 spin_lock(vmf->ptl);
4945 entry = vmf->orig_pte;
4946 if (unlikely(!pte_same(*vmf->pte, entry))) {
4947 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4948 goto unlock;
4949 }
4950 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
4951 if (!pte_write(entry))
4952 return do_wp_page(vmf);
4953 else if (likely(vmf->flags & FAULT_FLAG_WRITE))
4954 entry = pte_mkdirty(entry);
4955 }
4956 entry = pte_mkyoung(entry);
4957 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4958 vmf->flags & FAULT_FLAG_WRITE)) {
4959 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4960 } else {
4961 /* Skip spurious TLB flush for retried page fault */
4962 if (vmf->flags & FAULT_FLAG_TRIED)
4963 goto unlock;
4964 /*
4965 * This is needed only for protection faults but the arch code
4966 * is not yet telling us if this is a protection fault or not.
4967 * This still avoids useless tlb flushes for .text page faults
4968 * with threads.
4969 */
4970 if (vmf->flags & FAULT_FLAG_WRITE)
4971 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4972 }
4973unlock:
4974 pte_unmap_unlock(vmf->pte, vmf->ptl);
4975 return 0;
4976}
4977
4978/*
4979 * By the time we get here, we already hold the mm semaphore
4980 *
4981 * The mmap_lock may have been released depending on flags and our
4982 * return value. See filemap_fault() and __folio_lock_or_retry().
4983 */
4984static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4985 unsigned long address, unsigned int flags)
4986{
4987 struct vm_fault vmf = {
4988 .vma = vma,
4989 .address = address & PAGE_MASK,
4990 .real_address = address,
4991 .flags = flags,
4992 .pgoff = linear_page_index(vma, address),
4993 .gfp_mask = __get_fault_gfp_mask(vma),
4994 };
4995 struct mm_struct *mm = vma->vm_mm;
4996 unsigned long vm_flags = vma->vm_flags;
4997 pgd_t *pgd;
4998 p4d_t *p4d;
4999 vm_fault_t ret;
5000
5001 pgd = pgd_offset(mm, address);
5002 p4d = p4d_alloc(mm, pgd, address);
5003 if (!p4d)
5004 return VM_FAULT_OOM;
5005
5006 vmf.pud = pud_alloc(mm, p4d, address);
5007 if (!vmf.pud)
5008 return VM_FAULT_OOM;
5009retry_pud:
5010 if (pud_none(*vmf.pud) &&
5011 hugepage_vma_check(vma, vm_flags, false, true, true)) {
5012 ret = create_huge_pud(&vmf);
5013 if (!(ret & VM_FAULT_FALLBACK))
5014 return ret;
5015 } else {
5016 pud_t orig_pud = *vmf.pud;
5017
5018 barrier();
5019 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
5020
5021 /*
5022 * TODO once we support anonymous PUDs: NUMA case and
5023 * FAULT_FLAG_UNSHARE handling.
5024 */
5025 if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) {
5026 ret = wp_huge_pud(&vmf, orig_pud);
5027 if (!(ret & VM_FAULT_FALLBACK))
5028 return ret;
5029 } else {
5030 huge_pud_set_accessed(&vmf, orig_pud);
5031 return 0;
5032 }
5033 }
5034 }
5035
5036 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
5037 if (!vmf.pmd)
5038 return VM_FAULT_OOM;
5039
5040 /* Huge pud page fault raced with pmd_alloc? */
5041 if (pud_trans_unstable(vmf.pud))
5042 goto retry_pud;
5043
5044 if (pmd_none(*vmf.pmd) &&
5045 hugepage_vma_check(vma, vm_flags, false, true, true)) {
5046 ret = create_huge_pmd(&vmf);
5047 if (!(ret & VM_FAULT_FALLBACK))
5048 return ret;
5049 } else {
5050 vmf.orig_pmd = *vmf.pmd;
5051
5052 barrier();
5053 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
5054 VM_BUG_ON(thp_migration_supported() &&
5055 !is_pmd_migration_entry(vmf.orig_pmd));
5056 if (is_pmd_migration_entry(vmf.orig_pmd))
5057 pmd_migration_entry_wait(mm, vmf.pmd);
5058 return 0;
5059 }
5060 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
5061 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
5062 return do_huge_pmd_numa_page(&vmf);
5063
5064 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5065 !pmd_write(vmf.orig_pmd)) {
5066 ret = wp_huge_pmd(&vmf);
5067 if (!(ret & VM_FAULT_FALLBACK))
5068 return ret;
5069 } else {
5070 huge_pmd_set_accessed(&vmf);
5071 return 0;
5072 }
5073 }
5074 }
5075
5076 return handle_pte_fault(&vmf);
5077}
5078
5079/**
5080 * mm_account_fault - Do page fault accounting
5081 *
5082 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5083 * of perf event counters, but we'll still do the per-task accounting to
5084 * the task who triggered this page fault.
5085 * @address: the faulted address.
5086 * @flags: the fault flags.
5087 * @ret: the fault retcode.
5088 *
5089 * This will take care of most of the page fault accounting. Meanwhile, it
5090 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5091 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5092 * still be in per-arch page fault handlers at the entry of page fault.
5093 */
5094static inline void mm_account_fault(struct pt_regs *regs,
5095 unsigned long address, unsigned int flags,
5096 vm_fault_t ret)
5097{
5098 bool major;
5099
5100 /*
5101 * We don't do accounting for some specific faults:
5102 *
5103 * - Unsuccessful faults (e.g. when the address wasn't valid). That
5104 * includes arch_vma_access_permitted() failing before reaching here.
5105 * So this is not a "this many hardware page faults" counter. We
5106 * should use the hw profiling for that.
5107 *
5108 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
5109 * once they're completed.
5110 */
5111 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
5112 return;
5113
5114 /*
5115 * We define the fault as a major fault when the final successful fault
5116 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5117 * handle it immediately previously).
5118 */
5119 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
5120
5121 if (major)
5122 current->maj_flt++;
5123 else
5124 current->min_flt++;
5125
5126 /*
5127 * If the fault is done for GUP, regs will be NULL. We only do the
5128 * accounting for the per thread fault counters who triggered the
5129 * fault, and we skip the perf event updates.
5130 */
5131 if (!regs)
5132 return;
5133
5134 if (major)
5135 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
5136 else
5137 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
5138}
5139
5140#ifdef CONFIG_LRU_GEN
5141static void lru_gen_enter_fault(struct vm_area_struct *vma)
5142{
5143 /* the LRU algorithm doesn't apply to sequential or random reads */
5144 current->in_lru_fault = !(vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ));
5145}
5146
5147static void lru_gen_exit_fault(void)
5148{
5149 current->in_lru_fault = false;
5150}
5151#else
5152static void lru_gen_enter_fault(struct vm_area_struct *vma)
5153{
5154}
5155
5156static void lru_gen_exit_fault(void)
5157{
5158}
5159#endif /* CONFIG_LRU_GEN */
5160
5161static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
5162 unsigned int *flags)
5163{
5164 if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
5165 if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
5166 return VM_FAULT_SIGSEGV;
5167 /*
5168 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5169 * just treat it like an ordinary read-fault otherwise.
5170 */
5171 if (!is_cow_mapping(vma->vm_flags))
5172 *flags &= ~FAULT_FLAG_UNSHARE;
5173 } else if (*flags & FAULT_FLAG_WRITE) {
5174 /* Write faults on read-only mappings are impossible ... */
5175 if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
5176 return VM_FAULT_SIGSEGV;
5177 /* ... and FOLL_FORCE only applies to COW mappings. */
5178 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
5179 !is_cow_mapping(vma->vm_flags)))
5180 return VM_FAULT_SIGSEGV;
5181 }
5182 return 0;
5183}
5184
5185/*
5186 * By the time we get here, we already hold the mm semaphore
5187 *
5188 * The mmap_lock may have been released depending on flags and our
5189 * return value. See filemap_fault() and __folio_lock_or_retry().
5190 */
5191vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
5192 unsigned int flags, struct pt_regs *regs)
5193{
5194 vm_fault_t ret;
5195
5196 __set_current_state(TASK_RUNNING);
5197
5198 count_vm_event(PGFAULT);
5199 count_memcg_event_mm(vma->vm_mm, PGFAULT);
5200
5201 ret = sanitize_fault_flags(vma, &flags);
5202 if (ret)
5203 return ret;
5204
5205 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
5206 flags & FAULT_FLAG_INSTRUCTION,
5207 flags & FAULT_FLAG_REMOTE))
5208 return VM_FAULT_SIGSEGV;
5209
5210 /*
5211 * Enable the memcg OOM handling for faults triggered in user
5212 * space. Kernel faults are handled more gracefully.
5213 */
5214 if (flags & FAULT_FLAG_USER)
5215 mem_cgroup_enter_user_fault();
5216
5217 lru_gen_enter_fault(vma);
5218
5219 if (unlikely(is_vm_hugetlb_page(vma)))
5220 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
5221 else
5222 ret = __handle_mm_fault(vma, address, flags);
5223
5224 lru_gen_exit_fault();
5225
5226 if (flags & FAULT_FLAG_USER) {
5227 mem_cgroup_exit_user_fault();
5228 /*
5229 * The task may have entered a memcg OOM situation but
5230 * if the allocation error was handled gracefully (no
5231 * VM_FAULT_OOM), there is no need to kill anything.
5232 * Just clean up the OOM state peacefully.
5233 */
5234 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
5235 mem_cgroup_oom_synchronize(false);
5236 }
5237
5238 mm_account_fault(regs, address, flags, ret);
5239
5240 return ret;
5241}
5242EXPORT_SYMBOL_GPL(handle_mm_fault);
5243
5244#ifndef __PAGETABLE_P4D_FOLDED
5245/*
5246 * Allocate p4d page table.
5247 * We've already handled the fast-path in-line.
5248 */
5249int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
5250{
5251 p4d_t *new = p4d_alloc_one(mm, address);
5252 if (!new)
5253 return -ENOMEM;
5254
5255 spin_lock(&mm->page_table_lock);
5256 if (pgd_present(*pgd)) { /* Another has populated it */
5257 p4d_free(mm, new);
5258 } else {
5259 smp_wmb(); /* See comment in pmd_install() */
5260 pgd_populate(mm, pgd, new);
5261 }
5262 spin_unlock(&mm->page_table_lock);
5263 return 0;
5264}
5265#endif /* __PAGETABLE_P4D_FOLDED */
5266
5267#ifndef __PAGETABLE_PUD_FOLDED
5268/*
5269 * Allocate page upper directory.
5270 * We've already handled the fast-path in-line.
5271 */
5272int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
5273{
5274 pud_t *new = pud_alloc_one(mm, address);
5275 if (!new)
5276 return -ENOMEM;
5277
5278 spin_lock(&mm->page_table_lock);
5279 if (!p4d_present(*p4d)) {
5280 mm_inc_nr_puds(mm);
5281 smp_wmb(); /* See comment in pmd_install() */
5282 p4d_populate(mm, p4d, new);
5283 } else /* Another has populated it */
5284 pud_free(mm, new);
5285 spin_unlock(&mm->page_table_lock);
5286 return 0;
5287}
5288#endif /* __PAGETABLE_PUD_FOLDED */
5289
5290#ifndef __PAGETABLE_PMD_FOLDED
5291/*
5292 * Allocate page middle directory.
5293 * We've already handled the fast-path in-line.
5294 */
5295int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
5296{
5297 spinlock_t *ptl;
5298 pmd_t *new = pmd_alloc_one(mm, address);
5299 if (!new)
5300 return -ENOMEM;
5301
5302 ptl = pud_lock(mm, pud);
5303 if (!pud_present(*pud)) {
5304 mm_inc_nr_pmds(mm);
5305 smp_wmb(); /* See comment in pmd_install() */
5306 pud_populate(mm, pud, new);
5307 } else { /* Another has populated it */
5308 pmd_free(mm, new);
5309 }
5310 spin_unlock(ptl);
5311 return 0;
5312}
5313#endif /* __PAGETABLE_PMD_FOLDED */
5314
5315/**
5316 * follow_pte - look up PTE at a user virtual address
5317 * @mm: the mm_struct of the target address space
5318 * @address: user virtual address
5319 * @ptepp: location to store found PTE
5320 * @ptlp: location to store the lock for the PTE
5321 *
5322 * On a successful return, the pointer to the PTE is stored in @ptepp;
5323 * the corresponding lock is taken and its location is stored in @ptlp.
5324 * The contents of the PTE are only stable until @ptlp is released;
5325 * any further use, if any, must be protected against invalidation
5326 * with MMU notifiers.
5327 *
5328 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5329 * should be taken for read.
5330 *
5331 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5332 * it is not a good general-purpose API.
5333 *
5334 * Return: zero on success, -ve otherwise.
5335 */
5336int follow_pte(struct mm_struct *mm, unsigned long address,
5337 pte_t **ptepp, spinlock_t **ptlp)
5338{
5339 pgd_t *pgd;
5340 p4d_t *p4d;
5341 pud_t *pud;
5342 pmd_t *pmd;
5343 pte_t *ptep;
5344
5345 pgd = pgd_offset(mm, address);
5346 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
5347 goto out;
5348
5349 p4d = p4d_offset(pgd, address);
5350 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
5351 goto out;
5352
5353 pud = pud_offset(p4d, address);
5354 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
5355 goto out;
5356
5357 pmd = pmd_offset(pud, address);
5358 VM_BUG_ON(pmd_trans_huge(*pmd));
5359
5360 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
5361 goto out;
5362
5363 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
5364 if (!pte_present(*ptep))
5365 goto unlock;
5366 *ptepp = ptep;
5367 return 0;
5368unlock:
5369 pte_unmap_unlock(ptep, *ptlp);
5370out:
5371 return -EINVAL;
5372}
5373EXPORT_SYMBOL_GPL(follow_pte);
5374
5375/**
5376 * follow_pfn - look up PFN at a user virtual address
5377 * @vma: memory mapping
5378 * @address: user virtual address
5379 * @pfn: location to store found PFN
5380 *
5381 * Only IO mappings and raw PFN mappings are allowed.
5382 *
5383 * This function does not allow the caller to read the permissions
5384 * of the PTE. Do not use it.
5385 *
5386 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5387 */
5388int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5389 unsigned long *pfn)
5390{
5391 int ret = -EINVAL;
5392 spinlock_t *ptl;
5393 pte_t *ptep;
5394
5395 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5396 return ret;
5397
5398 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5399 if (ret)
5400 return ret;
5401 *pfn = pte_pfn(*ptep);
5402 pte_unmap_unlock(ptep, ptl);
5403 return 0;
5404}
5405EXPORT_SYMBOL(follow_pfn);
5406
5407#ifdef CONFIG_HAVE_IOREMAP_PROT
5408int follow_phys(struct vm_area_struct *vma,
5409 unsigned long address, unsigned int flags,
5410 unsigned long *prot, resource_size_t *phys)
5411{
5412 int ret = -EINVAL;
5413 pte_t *ptep, pte;
5414 spinlock_t *ptl;
5415
5416 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5417 goto out;
5418
5419 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5420 goto out;
5421 pte = *ptep;
5422
5423 if ((flags & FOLL_WRITE) && !pte_write(pte))
5424 goto unlock;
5425
5426 *prot = pgprot_val(pte_pgprot(pte));
5427 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5428
5429 ret = 0;
5430unlock:
5431 pte_unmap_unlock(ptep, ptl);
5432out:
5433 return ret;
5434}
5435
5436/**
5437 * generic_access_phys - generic implementation for iomem mmap access
5438 * @vma: the vma to access
5439 * @addr: userspace address, not relative offset within @vma
5440 * @buf: buffer to read/write
5441 * @len: length of transfer
5442 * @write: set to FOLL_WRITE when writing, otherwise reading
5443 *
5444 * This is a generic implementation for &vm_operations_struct.access for an
5445 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5446 * not page based.
5447 */
5448int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5449 void *buf, int len, int write)
5450{
5451 resource_size_t phys_addr;
5452 unsigned long prot = 0;
5453 void __iomem *maddr;
5454 pte_t *ptep, pte;
5455 spinlock_t *ptl;
5456 int offset = offset_in_page(addr);
5457 int ret = -EINVAL;
5458
5459 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5460 return -EINVAL;
5461
5462retry:
5463 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5464 return -EINVAL;
5465 pte = *ptep;
5466 pte_unmap_unlock(ptep, ptl);
5467
5468 prot = pgprot_val(pte_pgprot(pte));
5469 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5470
5471 if ((write & FOLL_WRITE) && !pte_write(pte))
5472 return -EINVAL;
5473
5474 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5475 if (!maddr)
5476 return -ENOMEM;
5477
5478 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5479 goto out_unmap;
5480
5481 if (!pte_same(pte, *ptep)) {
5482 pte_unmap_unlock(ptep, ptl);
5483 iounmap(maddr);
5484
5485 goto retry;
5486 }
5487
5488 if (write)
5489 memcpy_toio(maddr + offset, buf, len);
5490 else
5491 memcpy_fromio(buf, maddr + offset, len);
5492 ret = len;
5493 pte_unmap_unlock(ptep, ptl);
5494out_unmap:
5495 iounmap(maddr);
5496
5497 return ret;
5498}
5499EXPORT_SYMBOL_GPL(generic_access_phys);
5500#endif
5501
5502/*
5503 * Access another process' address space as given in mm.
5504 */
5505int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
5506 int len, unsigned int gup_flags)
5507{
5508 struct vm_area_struct *vma;
5509 void *old_buf = buf;
5510 int write = gup_flags & FOLL_WRITE;
5511
5512 if (mmap_read_lock_killable(mm))
5513 return 0;
5514
5515 /* ignore errors, just check how much was successfully transferred */
5516 while (len) {
5517 int bytes, ret, offset;
5518 void *maddr;
5519 struct page *page = NULL;
5520
5521 ret = get_user_pages_remote(mm, addr, 1,
5522 gup_flags, &page, &vma, NULL);
5523 if (ret <= 0) {
5524#ifndef CONFIG_HAVE_IOREMAP_PROT
5525 break;
5526#else
5527 /*
5528 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5529 * we can access using slightly different code.
5530 */
5531 vma = vma_lookup(mm, addr);
5532 if (!vma)
5533 break;
5534 if (vma->vm_ops && vma->vm_ops->access)
5535 ret = vma->vm_ops->access(vma, addr, buf,
5536 len, write);
5537 if (ret <= 0)
5538 break;
5539 bytes = ret;
5540#endif
5541 } else {
5542 bytes = len;
5543 offset = addr & (PAGE_SIZE-1);
5544 if (bytes > PAGE_SIZE-offset)
5545 bytes = PAGE_SIZE-offset;
5546
5547 maddr = kmap(page);
5548 if (write) {
5549 copy_to_user_page(vma, page, addr,
5550 maddr + offset, buf, bytes);
5551 set_page_dirty_lock(page);
5552 } else {
5553 copy_from_user_page(vma, page, addr,
5554 buf, maddr + offset, bytes);
5555 }
5556 kunmap(page);
5557 put_page(page);
5558 }
5559 len -= bytes;
5560 buf += bytes;
5561 addr += bytes;
5562 }
5563 mmap_read_unlock(mm);
5564
5565 return buf - old_buf;
5566}
5567
5568/**
5569 * access_remote_vm - access another process' address space
5570 * @mm: the mm_struct of the target address space
5571 * @addr: start address to access
5572 * @buf: source or destination buffer
5573 * @len: number of bytes to transfer
5574 * @gup_flags: flags modifying lookup behaviour
5575 *
5576 * The caller must hold a reference on @mm.
5577 *
5578 * Return: number of bytes copied from source to destination.
5579 */
5580int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5581 void *buf, int len, unsigned int gup_flags)
5582{
5583 return __access_remote_vm(mm, addr, buf, len, gup_flags);
5584}
5585
5586/*
5587 * Access another process' address space.
5588 * Source/target buffer must be kernel space,
5589 * Do not walk the page table directly, use get_user_pages
5590 */
5591int access_process_vm(struct task_struct *tsk, unsigned long addr,
5592 void *buf, int len, unsigned int gup_flags)
5593{
5594 struct mm_struct *mm;
5595 int ret;
5596
5597 mm = get_task_mm(tsk);
5598 if (!mm)
5599 return 0;
5600
5601 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5602
5603 mmput(mm);
5604
5605 return ret;
5606}
5607EXPORT_SYMBOL_GPL(access_process_vm);
5608
5609/*
5610 * Print the name of a VMA.
5611 */
5612void print_vma_addr(char *prefix, unsigned long ip)
5613{
5614 struct mm_struct *mm = current->mm;
5615 struct vm_area_struct *vma;
5616
5617 /*
5618 * we might be running from an atomic context so we cannot sleep
5619 */
5620 if (!mmap_read_trylock(mm))
5621 return;
5622
5623 vma = find_vma(mm, ip);
5624 if (vma && vma->vm_file) {
5625 struct file *f = vma->vm_file;
5626 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5627 if (buf) {
5628 char *p;
5629
5630 p = file_path(f, buf, PAGE_SIZE);
5631 if (IS_ERR(p))
5632 p = "?";
5633 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5634 vma->vm_start,
5635 vma->vm_end - vma->vm_start);
5636 free_page((unsigned long)buf);
5637 }
5638 }
5639 mmap_read_unlock(mm);
5640}
5641
5642#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5643void __might_fault(const char *file, int line)
5644{
5645 if (pagefault_disabled())
5646 return;
5647 __might_sleep(file, line);
5648#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5649 if (current->mm)
5650 might_lock_read(¤t->mm->mmap_lock);
5651#endif
5652}
5653EXPORT_SYMBOL(__might_fault);
5654#endif
5655
5656#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5657/*
5658 * Process all subpages of the specified huge page with the specified
5659 * operation. The target subpage will be processed last to keep its
5660 * cache lines hot.
5661 */
5662static inline void process_huge_page(
5663 unsigned long addr_hint, unsigned int pages_per_huge_page,
5664 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5665 void *arg)
5666{
5667 int i, n, base, l;
5668 unsigned long addr = addr_hint &
5669 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5670
5671 /* Process target subpage last to keep its cache lines hot */
5672 might_sleep();
5673 n = (addr_hint - addr) / PAGE_SIZE;
5674 if (2 * n <= pages_per_huge_page) {
5675 /* If target subpage in first half of huge page */
5676 base = 0;
5677 l = n;
5678 /* Process subpages at the end of huge page */
5679 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5680 cond_resched();
5681 process_subpage(addr + i * PAGE_SIZE, i, arg);
5682 }
5683 } else {
5684 /* If target subpage in second half of huge page */
5685 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5686 l = pages_per_huge_page - n;
5687 /* Process subpages at the begin of huge page */
5688 for (i = 0; i < base; i++) {
5689 cond_resched();
5690 process_subpage(addr + i * PAGE_SIZE, i, arg);
5691 }
5692 }
5693 /*
5694 * Process remaining subpages in left-right-left-right pattern
5695 * towards the target subpage
5696 */
5697 for (i = 0; i < l; i++) {
5698 int left_idx = base + i;
5699 int right_idx = base + 2 * l - 1 - i;
5700
5701 cond_resched();
5702 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5703 cond_resched();
5704 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5705 }
5706}
5707
5708static void clear_gigantic_page(struct page *page,
5709 unsigned long addr,
5710 unsigned int pages_per_huge_page)
5711{
5712 int i;
5713 struct page *p;
5714
5715 might_sleep();
5716 for (i = 0; i < pages_per_huge_page; i++) {
5717 p = nth_page(page, i);
5718 cond_resched();
5719 clear_user_highpage(p, addr + i * PAGE_SIZE);
5720 }
5721}
5722
5723static void clear_subpage(unsigned long addr, int idx, void *arg)
5724{
5725 struct page *page = arg;
5726
5727 clear_user_highpage(page + idx, addr);
5728}
5729
5730void clear_huge_page(struct page *page,
5731 unsigned long addr_hint, unsigned int pages_per_huge_page)
5732{
5733 unsigned long addr = addr_hint &
5734 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5735
5736 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5737 clear_gigantic_page(page, addr, pages_per_huge_page);
5738 return;
5739 }
5740
5741 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5742}
5743
5744static void copy_user_gigantic_page(struct page *dst, struct page *src,
5745 unsigned long addr,
5746 struct vm_area_struct *vma,
5747 unsigned int pages_per_huge_page)
5748{
5749 int i;
5750 struct page *dst_base = dst;
5751 struct page *src_base = src;
5752
5753 for (i = 0; i < pages_per_huge_page; i++) {
5754 dst = nth_page(dst_base, i);
5755 src = nth_page(src_base, i);
5756
5757 cond_resched();
5758 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5759 }
5760}
5761
5762struct copy_subpage_arg {
5763 struct page *dst;
5764 struct page *src;
5765 struct vm_area_struct *vma;
5766};
5767
5768static void copy_subpage(unsigned long addr, int idx, void *arg)
5769{
5770 struct copy_subpage_arg *copy_arg = arg;
5771
5772 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5773 addr, copy_arg->vma);
5774}
5775
5776void copy_user_huge_page(struct page *dst, struct page *src,
5777 unsigned long addr_hint, struct vm_area_struct *vma,
5778 unsigned int pages_per_huge_page)
5779{
5780 unsigned long addr = addr_hint &
5781 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5782 struct copy_subpage_arg arg = {
5783 .dst = dst,
5784 .src = src,
5785 .vma = vma,
5786 };
5787
5788 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5789 copy_user_gigantic_page(dst, src, addr, vma,
5790 pages_per_huge_page);
5791 return;
5792 }
5793
5794 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5795}
5796
5797long copy_huge_page_from_user(struct page *dst_page,
5798 const void __user *usr_src,
5799 unsigned int pages_per_huge_page,
5800 bool allow_pagefault)
5801{
5802 void *page_kaddr;
5803 unsigned long i, rc = 0;
5804 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5805 struct page *subpage;
5806
5807 for (i = 0; i < pages_per_huge_page; i++) {
5808 subpage = nth_page(dst_page, i);
5809 if (allow_pagefault)
5810 page_kaddr = kmap(subpage);
5811 else
5812 page_kaddr = kmap_atomic(subpage);
5813 rc = copy_from_user(page_kaddr,
5814 usr_src + i * PAGE_SIZE, PAGE_SIZE);
5815 if (allow_pagefault)
5816 kunmap(subpage);
5817 else
5818 kunmap_atomic(page_kaddr);
5819
5820 ret_val -= (PAGE_SIZE - rc);
5821 if (rc)
5822 break;
5823
5824 flush_dcache_page(subpage);
5825
5826 cond_resched();
5827 }
5828 return ret_val;
5829}
5830#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5831
5832#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5833
5834static struct kmem_cache *page_ptl_cachep;
5835
5836void __init ptlock_cache_init(void)
5837{
5838 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5839 SLAB_PANIC, NULL);
5840}
5841
5842bool ptlock_alloc(struct page *page)
5843{
5844 spinlock_t *ptl;
5845
5846 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5847 if (!ptl)
5848 return false;
5849 page->ptl = ptl;
5850 return true;
5851}
5852
5853void ptlock_free(struct page *page)
5854{
5855 kmem_cache_free(page_ptl_cachep, page->ptl);
5856}
5857#endif