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/sched/mm.h>
44#include <linux/sched/coredump.h>
45#include <linux/sched/numa_balancing.h>
46#include <linux/sched/task.h>
47#include <linux/hugetlb.h>
48#include <linux/mman.h>
49#include <linux/swap.h>
50#include <linux/highmem.h>
51#include <linux/pagemap.h>
52#include <linux/memremap.h>
53#include <linux/ksm.h>
54#include <linux/rmap.h>
55#include <linux/export.h>
56#include <linux/delayacct.h>
57#include <linux/init.h>
58#include <linux/pfn_t.h>
59#include <linux/writeback.h>
60#include <linux/memcontrol.h>
61#include <linux/mmu_notifier.h>
62#include <linux/swapops.h>
63#include <linux/elf.h>
64#include <linux/gfp.h>
65#include <linux/migrate.h>
66#include <linux/string.h>
67#include <linux/dma-debug.h>
68#include <linux/debugfs.h>
69#include <linux/userfaultfd_k.h>
70#include <linux/dax.h>
71#include <linux/oom.h>
72
73#include <asm/io.h>
74#include <asm/mmu_context.h>
75#include <asm/pgalloc.h>
76#include <linux/uaccess.h>
77#include <asm/tlb.h>
78#include <asm/tlbflush.h>
79#include <asm/pgtable.h>
80
81#include "internal.h"
82
83#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
84#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
85#endif
86
87#ifndef CONFIG_NEED_MULTIPLE_NODES
88/* use the per-pgdat data instead for discontigmem - mbligh */
89unsigned long max_mapnr;
90EXPORT_SYMBOL(max_mapnr);
91
92struct page *mem_map;
93EXPORT_SYMBOL(mem_map);
94#endif
95
96/*
97 * A number of key systems in x86 including ioremap() rely on the assumption
98 * that high_memory defines the upper bound on direct map memory, then end
99 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
100 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
101 * and ZONE_HIGHMEM.
102 */
103void *high_memory;
104EXPORT_SYMBOL(high_memory);
105
106/*
107 * Randomize the address space (stacks, mmaps, brk, etc.).
108 *
109 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110 * as ancient (libc5 based) binaries can segfault. )
111 */
112int randomize_va_space __read_mostly =
113#ifdef CONFIG_COMPAT_BRK
114 1;
115#else
116 2;
117#endif
118
119static int __init disable_randmaps(char *s)
120{
121 randomize_va_space = 0;
122 return 1;
123}
124__setup("norandmaps", disable_randmaps);
125
126unsigned long zero_pfn __read_mostly;
127EXPORT_SYMBOL(zero_pfn);
128
129unsigned long highest_memmap_pfn __read_mostly;
130
131/*
132 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
133 */
134static int __init init_zero_pfn(void)
135{
136 zero_pfn = page_to_pfn(ZERO_PAGE(0));
137 return 0;
138}
139core_initcall(init_zero_pfn);
140
141
142#if defined(SPLIT_RSS_COUNTING)
143
144void sync_mm_rss(struct mm_struct *mm)
145{
146 int i;
147
148 for (i = 0; i < NR_MM_COUNTERS; i++) {
149 if (current->rss_stat.count[i]) {
150 add_mm_counter(mm, i, current->rss_stat.count[i]);
151 current->rss_stat.count[i] = 0;
152 }
153 }
154 current->rss_stat.events = 0;
155}
156
157static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
158{
159 struct task_struct *task = current;
160
161 if (likely(task->mm == mm))
162 task->rss_stat.count[member] += val;
163 else
164 add_mm_counter(mm, member, val);
165}
166#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
167#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
168
169/* sync counter once per 64 page faults */
170#define TASK_RSS_EVENTS_THRESH (64)
171static void check_sync_rss_stat(struct task_struct *task)
172{
173 if (unlikely(task != current))
174 return;
175 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
176 sync_mm_rss(task->mm);
177}
178#else /* SPLIT_RSS_COUNTING */
179
180#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
181#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
182
183static void check_sync_rss_stat(struct task_struct *task)
184{
185}
186
187#endif /* SPLIT_RSS_COUNTING */
188
189#ifdef HAVE_GENERIC_MMU_GATHER
190
191static bool tlb_next_batch(struct mmu_gather *tlb)
192{
193 struct mmu_gather_batch *batch;
194
195 batch = tlb->active;
196 if (batch->next) {
197 tlb->active = batch->next;
198 return true;
199 }
200
201 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
202 return false;
203
204 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
205 if (!batch)
206 return false;
207
208 tlb->batch_count++;
209 batch->next = NULL;
210 batch->nr = 0;
211 batch->max = MAX_GATHER_BATCH;
212
213 tlb->active->next = batch;
214 tlb->active = batch;
215
216 return true;
217}
218
219void arch_tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
220 unsigned long start, unsigned long end)
221{
222 tlb->mm = mm;
223
224 /* Is it from 0 to ~0? */
225 tlb->fullmm = !(start | (end+1));
226 tlb->need_flush_all = 0;
227 tlb->local.next = NULL;
228 tlb->local.nr = 0;
229 tlb->local.max = ARRAY_SIZE(tlb->__pages);
230 tlb->active = &tlb->local;
231 tlb->batch_count = 0;
232
233#ifdef CONFIG_HAVE_RCU_TABLE_FREE
234 tlb->batch = NULL;
235#endif
236 tlb->page_size = 0;
237
238 __tlb_reset_range(tlb);
239}
240
241static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
242{
243 if (!tlb->end)
244 return;
245
246 tlb_flush(tlb);
247 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
248#ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb);
250#endif
251 __tlb_reset_range(tlb);
252}
253
254static void tlb_flush_mmu_free(struct mmu_gather *tlb)
255{
256 struct mmu_gather_batch *batch;
257
258 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
259 free_pages_and_swap_cache(batch->pages, batch->nr);
260 batch->nr = 0;
261 }
262 tlb->active = &tlb->local;
263}
264
265void tlb_flush_mmu(struct mmu_gather *tlb)
266{
267 tlb_flush_mmu_tlbonly(tlb);
268 tlb_flush_mmu_free(tlb);
269}
270
271/* tlb_finish_mmu
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
274 */
275void arch_tlb_finish_mmu(struct mmu_gather *tlb,
276 unsigned long start, unsigned long end, bool force)
277{
278 struct mmu_gather_batch *batch, *next;
279
280 if (force)
281 __tlb_adjust_range(tlb, start, end - start);
282
283 tlb_flush_mmu(tlb);
284
285 /* keep the page table cache within bounds */
286 check_pgt_cache();
287
288 for (batch = tlb->local.next; batch; batch = next) {
289 next = batch->next;
290 free_pages((unsigned long)batch, 0);
291 }
292 tlb->local.next = NULL;
293}
294
295/* __tlb_remove_page
296 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
297 * handling the additional races in SMP caused by other CPUs caching valid
298 * mappings in their TLBs. Returns the number of free page slots left.
299 * When out of page slots we must call tlb_flush_mmu().
300 *returns true if the caller should flush.
301 */
302bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
303{
304 struct mmu_gather_batch *batch;
305
306 VM_BUG_ON(!tlb->end);
307 VM_WARN_ON(tlb->page_size != page_size);
308
309 batch = tlb->active;
310 /*
311 * Add the page and check if we are full. If so
312 * force a flush.
313 */
314 batch->pages[batch->nr++] = page;
315 if (batch->nr == batch->max) {
316 if (!tlb_next_batch(tlb))
317 return true;
318 batch = tlb->active;
319 }
320 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
321
322 return false;
323}
324
325#endif /* HAVE_GENERIC_MMU_GATHER */
326
327#ifdef CONFIG_HAVE_RCU_TABLE_FREE
328
329/*
330 * See the comment near struct mmu_table_batch.
331 */
332
333static void tlb_remove_table_smp_sync(void *arg)
334{
335 /* Simply deliver the interrupt */
336}
337
338static void tlb_remove_table_one(void *table)
339{
340 /*
341 * This isn't an RCU grace period and hence the page-tables cannot be
342 * assumed to be actually RCU-freed.
343 *
344 * It is however sufficient for software page-table walkers that rely on
345 * IRQ disabling. See the comment near struct mmu_table_batch.
346 */
347 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
348 __tlb_remove_table(table);
349}
350
351static void tlb_remove_table_rcu(struct rcu_head *head)
352{
353 struct mmu_table_batch *batch;
354 int i;
355
356 batch = container_of(head, struct mmu_table_batch, rcu);
357
358 for (i = 0; i < batch->nr; i++)
359 __tlb_remove_table(batch->tables[i]);
360
361 free_page((unsigned long)batch);
362}
363
364void tlb_table_flush(struct mmu_gather *tlb)
365{
366 struct mmu_table_batch **batch = &tlb->batch;
367
368 if (*batch) {
369 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
370 *batch = NULL;
371 }
372}
373
374void tlb_remove_table(struct mmu_gather *tlb, void *table)
375{
376 struct mmu_table_batch **batch = &tlb->batch;
377
378 /*
379 * When there's less then two users of this mm there cannot be a
380 * concurrent page-table walk.
381 */
382 if (atomic_read(&tlb->mm->mm_users) < 2) {
383 __tlb_remove_table(table);
384 return;
385 }
386
387 if (*batch == NULL) {
388 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
389 if (*batch == NULL) {
390 tlb_remove_table_one(table);
391 return;
392 }
393 (*batch)->nr = 0;
394 }
395 (*batch)->tables[(*batch)->nr++] = table;
396 if ((*batch)->nr == MAX_TABLE_BATCH)
397 tlb_table_flush(tlb);
398}
399
400#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
401
402/**
403 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
404 * @tlb: the mmu_gather structure to initialize
405 * @mm: the mm_struct of the target address space
406 * @start: start of the region that will be removed from the page-table
407 * @end: end of the region that will be removed from the page-table
408 *
409 * Called to initialize an (on-stack) mmu_gather structure for page-table
410 * tear-down from @mm. The @start and @end are set to 0 and -1
411 * respectively when @mm is without users and we're going to destroy
412 * the full address space (exit/execve).
413 */
414void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
415 unsigned long start, unsigned long end)
416{
417 arch_tlb_gather_mmu(tlb, mm, start, end);
418 inc_tlb_flush_pending(tlb->mm);
419}
420
421void tlb_finish_mmu(struct mmu_gather *tlb,
422 unsigned long start, unsigned long end)
423{
424 /*
425 * If there are parallel threads are doing PTE changes on same range
426 * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
427 * flush by batching, a thread has stable TLB entry can fail to flush
428 * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
429 * forcefully if we detect parallel PTE batching threads.
430 */
431 bool force = mm_tlb_flush_nested(tlb->mm);
432
433 arch_tlb_finish_mmu(tlb, start, end, force);
434 dec_tlb_flush_pending(tlb->mm);
435}
436
437/*
438 * Note: this doesn't free the actual pages themselves. That
439 * has been handled earlier when unmapping all the memory regions.
440 */
441static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
442 unsigned long addr)
443{
444 pgtable_t token = pmd_pgtable(*pmd);
445 pmd_clear(pmd);
446 pte_free_tlb(tlb, token, addr);
447 mm_dec_nr_ptes(tlb->mm);
448}
449
450static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
451 unsigned long addr, unsigned long end,
452 unsigned long floor, unsigned long ceiling)
453{
454 pmd_t *pmd;
455 unsigned long next;
456 unsigned long start;
457
458 start = addr;
459 pmd = pmd_offset(pud, addr);
460 do {
461 next = pmd_addr_end(addr, end);
462 if (pmd_none_or_clear_bad(pmd))
463 continue;
464 free_pte_range(tlb, pmd, addr);
465 } while (pmd++, addr = next, addr != end);
466
467 start &= PUD_MASK;
468 if (start < floor)
469 return;
470 if (ceiling) {
471 ceiling &= PUD_MASK;
472 if (!ceiling)
473 return;
474 }
475 if (end - 1 > ceiling - 1)
476 return;
477
478 pmd = pmd_offset(pud, start);
479 pud_clear(pud);
480 pmd_free_tlb(tlb, pmd, start);
481 mm_dec_nr_pmds(tlb->mm);
482}
483
484static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
485 unsigned long addr, unsigned long end,
486 unsigned long floor, unsigned long ceiling)
487{
488 pud_t *pud;
489 unsigned long next;
490 unsigned long start;
491
492 start = addr;
493 pud = pud_offset(p4d, addr);
494 do {
495 next = pud_addr_end(addr, end);
496 if (pud_none_or_clear_bad(pud))
497 continue;
498 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
499 } while (pud++, addr = next, addr != end);
500
501 start &= P4D_MASK;
502 if (start < floor)
503 return;
504 if (ceiling) {
505 ceiling &= P4D_MASK;
506 if (!ceiling)
507 return;
508 }
509 if (end - 1 > ceiling - 1)
510 return;
511
512 pud = pud_offset(p4d, start);
513 p4d_clear(p4d);
514 pud_free_tlb(tlb, pud, start);
515 mm_dec_nr_puds(tlb->mm);
516}
517
518static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
519 unsigned long addr, unsigned long end,
520 unsigned long floor, unsigned long ceiling)
521{
522 p4d_t *p4d;
523 unsigned long next;
524 unsigned long start;
525
526 start = addr;
527 p4d = p4d_offset(pgd, addr);
528 do {
529 next = p4d_addr_end(addr, end);
530 if (p4d_none_or_clear_bad(p4d))
531 continue;
532 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
533 } while (p4d++, addr = next, addr != end);
534
535 start &= PGDIR_MASK;
536 if (start < floor)
537 return;
538 if (ceiling) {
539 ceiling &= PGDIR_MASK;
540 if (!ceiling)
541 return;
542 }
543 if (end - 1 > ceiling - 1)
544 return;
545
546 p4d = p4d_offset(pgd, start);
547 pgd_clear(pgd);
548 p4d_free_tlb(tlb, p4d, start);
549}
550
551/*
552 * This function frees user-level page tables of a process.
553 */
554void free_pgd_range(struct mmu_gather *tlb,
555 unsigned long addr, unsigned long end,
556 unsigned long floor, unsigned long ceiling)
557{
558 pgd_t *pgd;
559 unsigned long next;
560
561 /*
562 * The next few lines have given us lots of grief...
563 *
564 * Why are we testing PMD* at this top level? Because often
565 * there will be no work to do at all, and we'd prefer not to
566 * go all the way down to the bottom just to discover that.
567 *
568 * Why all these "- 1"s? Because 0 represents both the bottom
569 * of the address space and the top of it (using -1 for the
570 * top wouldn't help much: the masks would do the wrong thing).
571 * The rule is that addr 0 and floor 0 refer to the bottom of
572 * the address space, but end 0 and ceiling 0 refer to the top
573 * Comparisons need to use "end - 1" and "ceiling - 1" (though
574 * that end 0 case should be mythical).
575 *
576 * Wherever addr is brought up or ceiling brought down, we must
577 * be careful to reject "the opposite 0" before it confuses the
578 * subsequent tests. But what about where end is brought down
579 * by PMD_SIZE below? no, end can't go down to 0 there.
580 *
581 * Whereas we round start (addr) and ceiling down, by different
582 * masks at different levels, in order to test whether a table
583 * now has no other vmas using it, so can be freed, we don't
584 * bother to round floor or end up - the tests don't need that.
585 */
586
587 addr &= PMD_MASK;
588 if (addr < floor) {
589 addr += PMD_SIZE;
590 if (!addr)
591 return;
592 }
593 if (ceiling) {
594 ceiling &= PMD_MASK;
595 if (!ceiling)
596 return;
597 }
598 if (end - 1 > ceiling - 1)
599 end -= PMD_SIZE;
600 if (addr > end - 1)
601 return;
602 /*
603 * We add page table cache pages with PAGE_SIZE,
604 * (see pte_free_tlb()), flush the tlb if we need
605 */
606 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
607 pgd = pgd_offset(tlb->mm, addr);
608 do {
609 next = pgd_addr_end(addr, end);
610 if (pgd_none_or_clear_bad(pgd))
611 continue;
612 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
613 } while (pgd++, addr = next, addr != end);
614}
615
616void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
617 unsigned long floor, unsigned long ceiling)
618{
619 while (vma) {
620 struct vm_area_struct *next = vma->vm_next;
621 unsigned long addr = vma->vm_start;
622
623 /*
624 * Hide vma from rmap and truncate_pagecache before freeing
625 * pgtables
626 */
627 unlink_anon_vmas(vma);
628 unlink_file_vma(vma);
629
630 if (is_vm_hugetlb_page(vma)) {
631 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
632 floor, next ? next->vm_start : ceiling);
633 } else {
634 /*
635 * Optimization: gather nearby vmas into one call down
636 */
637 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
638 && !is_vm_hugetlb_page(next)) {
639 vma = next;
640 next = vma->vm_next;
641 unlink_anon_vmas(vma);
642 unlink_file_vma(vma);
643 }
644 free_pgd_range(tlb, addr, vma->vm_end,
645 floor, next ? next->vm_start : ceiling);
646 }
647 vma = next;
648 }
649}
650
651int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
652{
653 spinlock_t *ptl;
654 pgtable_t new = pte_alloc_one(mm, address);
655 if (!new)
656 return -ENOMEM;
657
658 /*
659 * Ensure all pte setup (eg. pte page lock and page clearing) are
660 * visible before the pte is made visible to other CPUs by being
661 * put into page tables.
662 *
663 * The other side of the story is the pointer chasing in the page
664 * table walking code (when walking the page table without locking;
665 * ie. most of the time). Fortunately, these data accesses consist
666 * of a chain of data-dependent loads, meaning most CPUs (alpha
667 * being the notable exception) will already guarantee loads are
668 * seen in-order. See the alpha page table accessors for the
669 * smp_read_barrier_depends() barriers in page table walking code.
670 */
671 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
672
673 ptl = pmd_lock(mm, pmd);
674 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
675 mm_inc_nr_ptes(mm);
676 pmd_populate(mm, pmd, new);
677 new = NULL;
678 }
679 spin_unlock(ptl);
680 if (new)
681 pte_free(mm, new);
682 return 0;
683}
684
685int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
686{
687 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
688 if (!new)
689 return -ENOMEM;
690
691 smp_wmb(); /* See comment in __pte_alloc */
692
693 spin_lock(&init_mm.page_table_lock);
694 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
695 pmd_populate_kernel(&init_mm, pmd, new);
696 new = NULL;
697 }
698 spin_unlock(&init_mm.page_table_lock);
699 if (new)
700 pte_free_kernel(&init_mm, new);
701 return 0;
702}
703
704static inline void init_rss_vec(int *rss)
705{
706 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
707}
708
709static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
710{
711 int i;
712
713 if (current->mm == mm)
714 sync_mm_rss(mm);
715 for (i = 0; i < NR_MM_COUNTERS; i++)
716 if (rss[i])
717 add_mm_counter(mm, i, rss[i]);
718}
719
720/*
721 * This function is called to print an error when a bad pte
722 * is found. For example, we might have a PFN-mapped pte in
723 * a region that doesn't allow it.
724 *
725 * The calling function must still handle the error.
726 */
727static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
728 pte_t pte, struct page *page)
729{
730 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
731 p4d_t *p4d = p4d_offset(pgd, addr);
732 pud_t *pud = pud_offset(p4d, addr);
733 pmd_t *pmd = pmd_offset(pud, addr);
734 struct address_space *mapping;
735 pgoff_t index;
736 static unsigned long resume;
737 static unsigned long nr_shown;
738 static unsigned long nr_unshown;
739
740 /*
741 * Allow a burst of 60 reports, then keep quiet for that minute;
742 * or allow a steady drip of one report per second.
743 */
744 if (nr_shown == 60) {
745 if (time_before(jiffies, resume)) {
746 nr_unshown++;
747 return;
748 }
749 if (nr_unshown) {
750 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
751 nr_unshown);
752 nr_unshown = 0;
753 }
754 nr_shown = 0;
755 }
756 if (nr_shown++ == 0)
757 resume = jiffies + 60 * HZ;
758
759 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
760 index = linear_page_index(vma, addr);
761
762 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
763 current->comm,
764 (long long)pte_val(pte), (long long)pmd_val(*pmd));
765 if (page)
766 dump_page(page, "bad pte");
767 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
768 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
769 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
770 vma->vm_file,
771 vma->vm_ops ? vma->vm_ops->fault : NULL,
772 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
773 mapping ? mapping->a_ops->readpage : NULL);
774 dump_stack();
775 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
776}
777
778/*
779 * vm_normal_page -- This function gets the "struct page" associated with a pte.
780 *
781 * "Special" mappings do not wish to be associated with a "struct page" (either
782 * it doesn't exist, or it exists but they don't want to touch it). In this
783 * case, NULL is returned here. "Normal" mappings do have a struct page.
784 *
785 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
786 * pte bit, in which case this function is trivial. Secondly, an architecture
787 * may not have a spare pte bit, which requires a more complicated scheme,
788 * described below.
789 *
790 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
791 * special mapping (even if there are underlying and valid "struct pages").
792 * COWed pages of a VM_PFNMAP are always normal.
793 *
794 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
795 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
796 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
797 * mapping will always honor the rule
798 *
799 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
800 *
801 * And for normal mappings this is false.
802 *
803 * This restricts such mappings to be a linear translation from virtual address
804 * to pfn. To get around this restriction, we allow arbitrary mappings so long
805 * as the vma is not a COW mapping; in that case, we know that all ptes are
806 * special (because none can have been COWed).
807 *
808 *
809 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
810 *
811 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
812 * page" backing, however the difference is that _all_ pages with a struct
813 * page (that is, those where pfn_valid is true) are refcounted and considered
814 * normal pages by the VM. The disadvantage is that pages are refcounted
815 * (which can be slower and simply not an option for some PFNMAP users). The
816 * advantage is that we don't have to follow the strict linearity rule of
817 * PFNMAP mappings in order to support COWable mappings.
818 *
819 */
820#ifdef __HAVE_ARCH_PTE_SPECIAL
821# define HAVE_PTE_SPECIAL 1
822#else
823# define HAVE_PTE_SPECIAL 0
824#endif
825struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
826 pte_t pte, bool with_public_device)
827{
828 unsigned long pfn = pte_pfn(pte);
829
830 if (HAVE_PTE_SPECIAL) {
831 if (likely(!pte_special(pte)))
832 goto check_pfn;
833 if (vma->vm_ops && vma->vm_ops->find_special_page)
834 return vma->vm_ops->find_special_page(vma, addr);
835 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
836 return NULL;
837 if (is_zero_pfn(pfn))
838 return NULL;
839
840 /*
841 * Device public pages are special pages (they are ZONE_DEVICE
842 * pages but different from persistent memory). They behave
843 * allmost like normal pages. The difference is that they are
844 * not on the lru and thus should never be involve with any-
845 * thing that involve lru manipulation (mlock, numa balancing,
846 * ...).
847 *
848 * This is why we still want to return NULL for such page from
849 * vm_normal_page() so that we do not have to special case all
850 * call site of vm_normal_page().
851 */
852 if (likely(pfn <= highest_memmap_pfn)) {
853 struct page *page = pfn_to_page(pfn);
854
855 if (is_device_public_page(page)) {
856 if (with_public_device)
857 return page;
858 return NULL;
859 }
860 }
861 print_bad_pte(vma, addr, pte, NULL);
862 return NULL;
863 }
864
865 /* !HAVE_PTE_SPECIAL case follows: */
866
867 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
868 if (vma->vm_flags & VM_MIXEDMAP) {
869 if (!pfn_valid(pfn))
870 return NULL;
871 goto out;
872 } else {
873 unsigned long off;
874 off = (addr - vma->vm_start) >> PAGE_SHIFT;
875 if (pfn == vma->vm_pgoff + off)
876 return NULL;
877 if (!is_cow_mapping(vma->vm_flags))
878 return NULL;
879 }
880 }
881
882 if (is_zero_pfn(pfn))
883 return NULL;
884check_pfn:
885 if (unlikely(pfn > highest_memmap_pfn)) {
886 print_bad_pte(vma, addr, pte, NULL);
887 return NULL;
888 }
889
890 /*
891 * NOTE! We still have PageReserved() pages in the page tables.
892 * eg. VDSO mappings can cause them to exist.
893 */
894out:
895 return pfn_to_page(pfn);
896}
897
898#ifdef CONFIG_TRANSPARENT_HUGEPAGE
899struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
900 pmd_t pmd)
901{
902 unsigned long pfn = pmd_pfn(pmd);
903
904 /*
905 * There is no pmd_special() but there may be special pmds, e.g.
906 * in a direct-access (dax) mapping, so let's just replicate the
907 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
908 */
909 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
910 if (vma->vm_flags & VM_MIXEDMAP) {
911 if (!pfn_valid(pfn))
912 return NULL;
913 goto out;
914 } else {
915 unsigned long off;
916 off = (addr - vma->vm_start) >> PAGE_SHIFT;
917 if (pfn == vma->vm_pgoff + off)
918 return NULL;
919 if (!is_cow_mapping(vma->vm_flags))
920 return NULL;
921 }
922 }
923
924 if (is_zero_pfn(pfn))
925 return NULL;
926 if (unlikely(pfn > highest_memmap_pfn))
927 return NULL;
928
929 /*
930 * NOTE! We still have PageReserved() pages in the page tables.
931 * eg. VDSO mappings can cause them to exist.
932 */
933out:
934 return pfn_to_page(pfn);
935}
936#endif
937
938/*
939 * copy one vm_area from one task to the other. Assumes the page tables
940 * already present in the new task to be cleared in the whole range
941 * covered by this vma.
942 */
943
944static inline unsigned long
945copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
946 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
947 unsigned long addr, int *rss)
948{
949 unsigned long vm_flags = vma->vm_flags;
950 pte_t pte = *src_pte;
951 struct page *page;
952
953 /* pte contains position in swap or file, so copy. */
954 if (unlikely(!pte_present(pte))) {
955 swp_entry_t entry = pte_to_swp_entry(pte);
956
957 if (likely(!non_swap_entry(entry))) {
958 if (swap_duplicate(entry) < 0)
959 return entry.val;
960
961 /* make sure dst_mm is on swapoff's mmlist. */
962 if (unlikely(list_empty(&dst_mm->mmlist))) {
963 spin_lock(&mmlist_lock);
964 if (list_empty(&dst_mm->mmlist))
965 list_add(&dst_mm->mmlist,
966 &src_mm->mmlist);
967 spin_unlock(&mmlist_lock);
968 }
969 rss[MM_SWAPENTS]++;
970 } else if (is_migration_entry(entry)) {
971 page = migration_entry_to_page(entry);
972
973 rss[mm_counter(page)]++;
974
975 if (is_write_migration_entry(entry) &&
976 is_cow_mapping(vm_flags)) {
977 /*
978 * COW mappings require pages in both
979 * parent and child to be set to read.
980 */
981 make_migration_entry_read(&entry);
982 pte = swp_entry_to_pte(entry);
983 if (pte_swp_soft_dirty(*src_pte))
984 pte = pte_swp_mksoft_dirty(pte);
985 set_pte_at(src_mm, addr, src_pte, pte);
986 }
987 } else if (is_device_private_entry(entry)) {
988 page = device_private_entry_to_page(entry);
989
990 /*
991 * Update rss count even for unaddressable pages, as
992 * they should treated just like normal pages in this
993 * respect.
994 *
995 * We will likely want to have some new rss counters
996 * for unaddressable pages, at some point. But for now
997 * keep things as they are.
998 */
999 get_page(page);
1000 rss[mm_counter(page)]++;
1001 page_dup_rmap(page, false);
1002
1003 /*
1004 * We do not preserve soft-dirty information, because so
1005 * far, checkpoint/restore is the only feature that
1006 * requires that. And checkpoint/restore does not work
1007 * when a device driver is involved (you cannot easily
1008 * save and restore device driver state).
1009 */
1010 if (is_write_device_private_entry(entry) &&
1011 is_cow_mapping(vm_flags)) {
1012 make_device_private_entry_read(&entry);
1013 pte = swp_entry_to_pte(entry);
1014 set_pte_at(src_mm, addr, src_pte, pte);
1015 }
1016 }
1017 goto out_set_pte;
1018 }
1019
1020 /*
1021 * If it's a COW mapping, write protect it both
1022 * in the parent and the child
1023 */
1024 if (is_cow_mapping(vm_flags)) {
1025 ptep_set_wrprotect(src_mm, addr, src_pte);
1026 pte = pte_wrprotect(pte);
1027 }
1028
1029 /*
1030 * If it's a shared mapping, mark it clean in
1031 * the child
1032 */
1033 if (vm_flags & VM_SHARED)
1034 pte = pte_mkclean(pte);
1035 pte = pte_mkold(pte);
1036
1037 page = vm_normal_page(vma, addr, pte);
1038 if (page) {
1039 get_page(page);
1040 page_dup_rmap(page, false);
1041 rss[mm_counter(page)]++;
1042 } else if (pte_devmap(pte)) {
1043 page = pte_page(pte);
1044
1045 /*
1046 * Cache coherent device memory behave like regular page and
1047 * not like persistent memory page. For more informations see
1048 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
1049 */
1050 if (is_device_public_page(page)) {
1051 get_page(page);
1052 page_dup_rmap(page, false);
1053 rss[mm_counter(page)]++;
1054 }
1055 }
1056
1057out_set_pte:
1058 set_pte_at(dst_mm, addr, dst_pte, pte);
1059 return 0;
1060}
1061
1062static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1063 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
1064 unsigned long addr, unsigned long end)
1065{
1066 pte_t *orig_src_pte, *orig_dst_pte;
1067 pte_t *src_pte, *dst_pte;
1068 spinlock_t *src_ptl, *dst_ptl;
1069 int progress = 0;
1070 int rss[NR_MM_COUNTERS];
1071 swp_entry_t entry = (swp_entry_t){0};
1072
1073again:
1074 init_rss_vec(rss);
1075
1076 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1077 if (!dst_pte)
1078 return -ENOMEM;
1079 src_pte = pte_offset_map(src_pmd, addr);
1080 src_ptl = pte_lockptr(src_mm, src_pmd);
1081 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1082 orig_src_pte = src_pte;
1083 orig_dst_pte = dst_pte;
1084 arch_enter_lazy_mmu_mode();
1085
1086 do {
1087 /*
1088 * We are holding two locks at this point - either of them
1089 * could generate latencies in another task on another CPU.
1090 */
1091 if (progress >= 32) {
1092 progress = 0;
1093 if (need_resched() ||
1094 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1095 break;
1096 }
1097 if (pte_none(*src_pte)) {
1098 progress++;
1099 continue;
1100 }
1101 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
1102 vma, addr, rss);
1103 if (entry.val)
1104 break;
1105 progress += 8;
1106 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1107
1108 arch_leave_lazy_mmu_mode();
1109 spin_unlock(src_ptl);
1110 pte_unmap(orig_src_pte);
1111 add_mm_rss_vec(dst_mm, rss);
1112 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1113 cond_resched();
1114
1115 if (entry.val) {
1116 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
1117 return -ENOMEM;
1118 progress = 0;
1119 }
1120 if (addr != end)
1121 goto again;
1122 return 0;
1123}
1124
1125static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1126 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1127 unsigned long addr, unsigned long end)
1128{
1129 pmd_t *src_pmd, *dst_pmd;
1130 unsigned long next;
1131
1132 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1133 if (!dst_pmd)
1134 return -ENOMEM;
1135 src_pmd = pmd_offset(src_pud, addr);
1136 do {
1137 next = pmd_addr_end(addr, end);
1138 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1139 || pmd_devmap(*src_pmd)) {
1140 int err;
1141 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1142 err = copy_huge_pmd(dst_mm, src_mm,
1143 dst_pmd, src_pmd, addr, vma);
1144 if (err == -ENOMEM)
1145 return -ENOMEM;
1146 if (!err)
1147 continue;
1148 /* fall through */
1149 }
1150 if (pmd_none_or_clear_bad(src_pmd))
1151 continue;
1152 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1153 vma, addr, next))
1154 return -ENOMEM;
1155 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1156 return 0;
1157}
1158
1159static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1160 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1161 unsigned long addr, unsigned long end)
1162{
1163 pud_t *src_pud, *dst_pud;
1164 unsigned long next;
1165
1166 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1167 if (!dst_pud)
1168 return -ENOMEM;
1169 src_pud = pud_offset(src_p4d, addr);
1170 do {
1171 next = pud_addr_end(addr, end);
1172 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1173 int err;
1174
1175 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1176 err = copy_huge_pud(dst_mm, src_mm,
1177 dst_pud, src_pud, addr, vma);
1178 if (err == -ENOMEM)
1179 return -ENOMEM;
1180 if (!err)
1181 continue;
1182 /* fall through */
1183 }
1184 if (pud_none_or_clear_bad(src_pud))
1185 continue;
1186 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1187 vma, addr, next))
1188 return -ENOMEM;
1189 } while (dst_pud++, src_pud++, addr = next, addr != end);
1190 return 0;
1191}
1192
1193static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1194 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1195 unsigned long addr, unsigned long end)
1196{
1197 p4d_t *src_p4d, *dst_p4d;
1198 unsigned long next;
1199
1200 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1201 if (!dst_p4d)
1202 return -ENOMEM;
1203 src_p4d = p4d_offset(src_pgd, addr);
1204 do {
1205 next = p4d_addr_end(addr, end);
1206 if (p4d_none_or_clear_bad(src_p4d))
1207 continue;
1208 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1209 vma, addr, next))
1210 return -ENOMEM;
1211 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1212 return 0;
1213}
1214
1215int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1216 struct vm_area_struct *vma)
1217{
1218 pgd_t *src_pgd, *dst_pgd;
1219 unsigned long next;
1220 unsigned long addr = vma->vm_start;
1221 unsigned long end = vma->vm_end;
1222 unsigned long mmun_start; /* For mmu_notifiers */
1223 unsigned long mmun_end; /* For mmu_notifiers */
1224 bool is_cow;
1225 int ret;
1226
1227 /*
1228 * Don't copy ptes where a page fault will fill them correctly.
1229 * Fork becomes much lighter when there are big shared or private
1230 * readonly mappings. The tradeoff is that copy_page_range is more
1231 * efficient than faulting.
1232 */
1233 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1234 !vma->anon_vma)
1235 return 0;
1236
1237 if (is_vm_hugetlb_page(vma))
1238 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1239
1240 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1241 /*
1242 * We do not free on error cases below as remove_vma
1243 * gets called on error from higher level routine
1244 */
1245 ret = track_pfn_copy(vma);
1246 if (ret)
1247 return ret;
1248 }
1249
1250 /*
1251 * We need to invalidate the secondary MMU mappings only when
1252 * there could be a permission downgrade on the ptes of the
1253 * parent mm. And a permission downgrade will only happen if
1254 * is_cow_mapping() returns true.
1255 */
1256 is_cow = is_cow_mapping(vma->vm_flags);
1257 mmun_start = addr;
1258 mmun_end = end;
1259 if (is_cow)
1260 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1261 mmun_end);
1262
1263 ret = 0;
1264 dst_pgd = pgd_offset(dst_mm, addr);
1265 src_pgd = pgd_offset(src_mm, addr);
1266 do {
1267 next = pgd_addr_end(addr, end);
1268 if (pgd_none_or_clear_bad(src_pgd))
1269 continue;
1270 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1271 vma, addr, next))) {
1272 ret = -ENOMEM;
1273 break;
1274 }
1275 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1276
1277 if (is_cow)
1278 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1279 return ret;
1280}
1281
1282static unsigned long zap_pte_range(struct mmu_gather *tlb,
1283 struct vm_area_struct *vma, pmd_t *pmd,
1284 unsigned long addr, unsigned long end,
1285 struct zap_details *details)
1286{
1287 struct mm_struct *mm = tlb->mm;
1288 int force_flush = 0;
1289 int rss[NR_MM_COUNTERS];
1290 spinlock_t *ptl;
1291 pte_t *start_pte;
1292 pte_t *pte;
1293 swp_entry_t entry;
1294
1295 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1296again:
1297 init_rss_vec(rss);
1298 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1299 pte = start_pte;
1300 flush_tlb_batched_pending(mm);
1301 arch_enter_lazy_mmu_mode();
1302 do {
1303 pte_t ptent = *pte;
1304 if (pte_none(ptent))
1305 continue;
1306
1307 if (pte_present(ptent)) {
1308 struct page *page;
1309
1310 page = _vm_normal_page(vma, addr, ptent, true);
1311 if (unlikely(details) && page) {
1312 /*
1313 * unmap_shared_mapping_pages() wants to
1314 * invalidate cache without truncating:
1315 * unmap shared but keep private pages.
1316 */
1317 if (details->check_mapping &&
1318 details->check_mapping != page_rmapping(page))
1319 continue;
1320 }
1321 ptent = ptep_get_and_clear_full(mm, addr, pte,
1322 tlb->fullmm);
1323 tlb_remove_tlb_entry(tlb, pte, addr);
1324 if (unlikely(!page))
1325 continue;
1326
1327 if (!PageAnon(page)) {
1328 if (pte_dirty(ptent)) {
1329 force_flush = 1;
1330 set_page_dirty(page);
1331 }
1332 if (pte_young(ptent) &&
1333 likely(!(vma->vm_flags & VM_SEQ_READ)))
1334 mark_page_accessed(page);
1335 }
1336 rss[mm_counter(page)]--;
1337 page_remove_rmap(page, false);
1338 if (unlikely(page_mapcount(page) < 0))
1339 print_bad_pte(vma, addr, ptent, page);
1340 if (unlikely(__tlb_remove_page(tlb, page))) {
1341 force_flush = 1;
1342 addr += PAGE_SIZE;
1343 break;
1344 }
1345 continue;
1346 }
1347
1348 entry = pte_to_swp_entry(ptent);
1349 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1350 struct page *page = device_private_entry_to_page(entry);
1351
1352 if (unlikely(details && details->check_mapping)) {
1353 /*
1354 * unmap_shared_mapping_pages() wants to
1355 * invalidate cache without truncating:
1356 * unmap shared but keep private pages.
1357 */
1358 if (details->check_mapping !=
1359 page_rmapping(page))
1360 continue;
1361 }
1362
1363 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1364 rss[mm_counter(page)]--;
1365 page_remove_rmap(page, false);
1366 put_page(page);
1367 continue;
1368 }
1369
1370 /* If details->check_mapping, we leave swap entries. */
1371 if (unlikely(details))
1372 continue;
1373
1374 entry = pte_to_swp_entry(ptent);
1375 if (!non_swap_entry(entry))
1376 rss[MM_SWAPENTS]--;
1377 else if (is_migration_entry(entry)) {
1378 struct page *page;
1379
1380 page = migration_entry_to_page(entry);
1381 rss[mm_counter(page)]--;
1382 }
1383 if (unlikely(!free_swap_and_cache(entry)))
1384 print_bad_pte(vma, addr, ptent, NULL);
1385 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1386 } while (pte++, addr += PAGE_SIZE, addr != end);
1387
1388 add_mm_rss_vec(mm, rss);
1389 arch_leave_lazy_mmu_mode();
1390
1391 /* Do the actual TLB flush before dropping ptl */
1392 if (force_flush)
1393 tlb_flush_mmu_tlbonly(tlb);
1394 pte_unmap_unlock(start_pte, ptl);
1395
1396 /*
1397 * If we forced a TLB flush (either due to running out of
1398 * batch buffers or because we needed to flush dirty TLB
1399 * entries before releasing the ptl), free the batched
1400 * memory too. Restart if we didn't do everything.
1401 */
1402 if (force_flush) {
1403 force_flush = 0;
1404 tlb_flush_mmu_free(tlb);
1405 if (addr != end)
1406 goto again;
1407 }
1408
1409 return addr;
1410}
1411
1412static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1413 struct vm_area_struct *vma, pud_t *pud,
1414 unsigned long addr, unsigned long end,
1415 struct zap_details *details)
1416{
1417 pmd_t *pmd;
1418 unsigned long next;
1419
1420 pmd = pmd_offset(pud, addr);
1421 do {
1422 next = pmd_addr_end(addr, end);
1423 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1424 if (next - addr != HPAGE_PMD_SIZE) {
1425 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1426 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1427 __split_huge_pmd(vma, pmd, addr, false, NULL);
1428 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1429 goto next;
1430 /* fall through */
1431 }
1432 /*
1433 * Here there can be other concurrent MADV_DONTNEED or
1434 * trans huge page faults running, and if the pmd is
1435 * none or trans huge it can change under us. This is
1436 * because MADV_DONTNEED holds the mmap_sem in read
1437 * mode.
1438 */
1439 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1440 goto next;
1441 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1442next:
1443 cond_resched();
1444 } while (pmd++, addr = next, addr != end);
1445
1446 return addr;
1447}
1448
1449static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1450 struct vm_area_struct *vma, p4d_t *p4d,
1451 unsigned long addr, unsigned long end,
1452 struct zap_details *details)
1453{
1454 pud_t *pud;
1455 unsigned long next;
1456
1457 pud = pud_offset(p4d, addr);
1458 do {
1459 next = pud_addr_end(addr, end);
1460 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1461 if (next - addr != HPAGE_PUD_SIZE) {
1462 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1463 split_huge_pud(vma, pud, addr);
1464 } else if (zap_huge_pud(tlb, vma, pud, addr))
1465 goto next;
1466 /* fall through */
1467 }
1468 if (pud_none_or_clear_bad(pud))
1469 continue;
1470 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1471next:
1472 cond_resched();
1473 } while (pud++, addr = next, addr != end);
1474
1475 return addr;
1476}
1477
1478static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1479 struct vm_area_struct *vma, pgd_t *pgd,
1480 unsigned long addr, unsigned long end,
1481 struct zap_details *details)
1482{
1483 p4d_t *p4d;
1484 unsigned long next;
1485
1486 p4d = p4d_offset(pgd, addr);
1487 do {
1488 next = p4d_addr_end(addr, end);
1489 if (p4d_none_or_clear_bad(p4d))
1490 continue;
1491 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1492 } while (p4d++, addr = next, addr != end);
1493
1494 return addr;
1495}
1496
1497void unmap_page_range(struct mmu_gather *tlb,
1498 struct vm_area_struct *vma,
1499 unsigned long addr, unsigned long end,
1500 struct zap_details *details)
1501{
1502 pgd_t *pgd;
1503 unsigned long next;
1504
1505 BUG_ON(addr >= end);
1506 tlb_start_vma(tlb, vma);
1507 pgd = pgd_offset(vma->vm_mm, addr);
1508 do {
1509 next = pgd_addr_end(addr, end);
1510 if (pgd_none_or_clear_bad(pgd))
1511 continue;
1512 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1513 } while (pgd++, addr = next, addr != end);
1514 tlb_end_vma(tlb, vma);
1515}
1516
1517
1518static void unmap_single_vma(struct mmu_gather *tlb,
1519 struct vm_area_struct *vma, unsigned long start_addr,
1520 unsigned long end_addr,
1521 struct zap_details *details)
1522{
1523 unsigned long start = max(vma->vm_start, start_addr);
1524 unsigned long end;
1525
1526 if (start >= vma->vm_end)
1527 return;
1528 end = min(vma->vm_end, end_addr);
1529 if (end <= vma->vm_start)
1530 return;
1531
1532 if (vma->vm_file)
1533 uprobe_munmap(vma, start, end);
1534
1535 if (unlikely(vma->vm_flags & VM_PFNMAP))
1536 untrack_pfn(vma, 0, 0);
1537
1538 if (start != end) {
1539 if (unlikely(is_vm_hugetlb_page(vma))) {
1540 /*
1541 * It is undesirable to test vma->vm_file as it
1542 * should be non-null for valid hugetlb area.
1543 * However, vm_file will be NULL in the error
1544 * cleanup path of mmap_region. When
1545 * hugetlbfs ->mmap method fails,
1546 * mmap_region() nullifies vma->vm_file
1547 * before calling this function to clean up.
1548 * Since no pte has actually been setup, it is
1549 * safe to do nothing in this case.
1550 */
1551 if (vma->vm_file) {
1552 i_mmap_lock_write(vma->vm_file->f_mapping);
1553 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1554 i_mmap_unlock_write(vma->vm_file->f_mapping);
1555 }
1556 } else
1557 unmap_page_range(tlb, vma, start, end, details);
1558 }
1559}
1560
1561/**
1562 * unmap_vmas - unmap a range of memory covered by a list of vma's
1563 * @tlb: address of the caller's struct mmu_gather
1564 * @vma: the starting vma
1565 * @start_addr: virtual address at which to start unmapping
1566 * @end_addr: virtual address at which to end unmapping
1567 *
1568 * Unmap all pages in the vma list.
1569 *
1570 * Only addresses between `start' and `end' will be unmapped.
1571 *
1572 * The VMA list must be sorted in ascending virtual address order.
1573 *
1574 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1575 * range after unmap_vmas() returns. So the only responsibility here is to
1576 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1577 * drops the lock and schedules.
1578 */
1579void unmap_vmas(struct mmu_gather *tlb,
1580 struct vm_area_struct *vma, unsigned long start_addr,
1581 unsigned long end_addr)
1582{
1583 struct mm_struct *mm = vma->vm_mm;
1584
1585 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1586 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1587 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1588 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1589}
1590
1591/**
1592 * zap_page_range - remove user pages in a given range
1593 * @vma: vm_area_struct holding the applicable pages
1594 * @start: starting address of pages to zap
1595 * @size: number of bytes to zap
1596 *
1597 * Caller must protect the VMA list
1598 */
1599void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1600 unsigned long size)
1601{
1602 struct mm_struct *mm = vma->vm_mm;
1603 struct mmu_gather tlb;
1604 unsigned long end = start + size;
1605
1606 lru_add_drain();
1607 tlb_gather_mmu(&tlb, mm, start, end);
1608 update_hiwater_rss(mm);
1609 mmu_notifier_invalidate_range_start(mm, start, end);
1610 for ( ; vma && vma->vm_start < end; vma = vma->vm_next) {
1611 unmap_single_vma(&tlb, vma, start, end, NULL);
1612
1613 /*
1614 * zap_page_range does not specify whether mmap_sem should be
1615 * held for read or write. That allows parallel zap_page_range
1616 * operations to unmap a PTE and defer a flush meaning that
1617 * this call observes pte_none and fails to flush the TLB.
1618 * Rather than adding a complex API, ensure that no stale
1619 * TLB entries exist when this call returns.
1620 */
1621 flush_tlb_range(vma, start, end);
1622 }
1623
1624 mmu_notifier_invalidate_range_end(mm, start, end);
1625 tlb_finish_mmu(&tlb, start, end);
1626}
1627
1628/**
1629 * zap_page_range_single - remove user pages in a given range
1630 * @vma: vm_area_struct holding the applicable pages
1631 * @address: starting address of pages to zap
1632 * @size: number of bytes to zap
1633 * @details: details of shared cache invalidation
1634 *
1635 * The range must fit into one VMA.
1636 */
1637static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1638 unsigned long size, struct zap_details *details)
1639{
1640 struct mm_struct *mm = vma->vm_mm;
1641 struct mmu_gather tlb;
1642 unsigned long end = address + size;
1643
1644 lru_add_drain();
1645 tlb_gather_mmu(&tlb, mm, address, end);
1646 update_hiwater_rss(mm);
1647 mmu_notifier_invalidate_range_start(mm, address, end);
1648 unmap_single_vma(&tlb, vma, address, end, details);
1649 mmu_notifier_invalidate_range_end(mm, address, end);
1650 tlb_finish_mmu(&tlb, address, end);
1651}
1652
1653/**
1654 * zap_vma_ptes - remove ptes mapping the vma
1655 * @vma: vm_area_struct holding ptes to be zapped
1656 * @address: starting address of pages to zap
1657 * @size: number of bytes to zap
1658 *
1659 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1660 *
1661 * The entire address range must be fully contained within the vma.
1662 *
1663 * Returns 0 if successful.
1664 */
1665int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1666 unsigned long size)
1667{
1668 if (address < vma->vm_start || address + size > vma->vm_end ||
1669 !(vma->vm_flags & VM_PFNMAP))
1670 return -1;
1671 zap_page_range_single(vma, address, size, NULL);
1672 return 0;
1673}
1674EXPORT_SYMBOL_GPL(zap_vma_ptes);
1675
1676pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1677 spinlock_t **ptl)
1678{
1679 pgd_t *pgd;
1680 p4d_t *p4d;
1681 pud_t *pud;
1682 pmd_t *pmd;
1683
1684 pgd = pgd_offset(mm, addr);
1685 p4d = p4d_alloc(mm, pgd, addr);
1686 if (!p4d)
1687 return NULL;
1688 pud = pud_alloc(mm, p4d, addr);
1689 if (!pud)
1690 return NULL;
1691 pmd = pmd_alloc(mm, pud, addr);
1692 if (!pmd)
1693 return NULL;
1694
1695 VM_BUG_ON(pmd_trans_huge(*pmd));
1696 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1697}
1698
1699/*
1700 * This is the old fallback for page remapping.
1701 *
1702 * For historical reasons, it only allows reserved pages. Only
1703 * old drivers should use this, and they needed to mark their
1704 * pages reserved for the old functions anyway.
1705 */
1706static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1707 struct page *page, pgprot_t prot)
1708{
1709 struct mm_struct *mm = vma->vm_mm;
1710 int retval;
1711 pte_t *pte;
1712 spinlock_t *ptl;
1713
1714 retval = -EINVAL;
1715 if (PageAnon(page))
1716 goto out;
1717 retval = -ENOMEM;
1718 flush_dcache_page(page);
1719 pte = get_locked_pte(mm, addr, &ptl);
1720 if (!pte)
1721 goto out;
1722 retval = -EBUSY;
1723 if (!pte_none(*pte))
1724 goto out_unlock;
1725
1726 /* Ok, finally just insert the thing.. */
1727 get_page(page);
1728 inc_mm_counter_fast(mm, mm_counter_file(page));
1729 page_add_file_rmap(page, false);
1730 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1731
1732 retval = 0;
1733 pte_unmap_unlock(pte, ptl);
1734 return retval;
1735out_unlock:
1736 pte_unmap_unlock(pte, ptl);
1737out:
1738 return retval;
1739}
1740
1741/**
1742 * vm_insert_page - insert single page into user vma
1743 * @vma: user vma to map to
1744 * @addr: target user address of this page
1745 * @page: source kernel page
1746 *
1747 * This allows drivers to insert individual pages they've allocated
1748 * into a user vma.
1749 *
1750 * The page has to be a nice clean _individual_ kernel allocation.
1751 * If you allocate a compound page, you need to have marked it as
1752 * such (__GFP_COMP), or manually just split the page up yourself
1753 * (see split_page()).
1754 *
1755 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1756 * took an arbitrary page protection parameter. This doesn't allow
1757 * that. Your vma protection will have to be set up correctly, which
1758 * means that if you want a shared writable mapping, you'd better
1759 * ask for a shared writable mapping!
1760 *
1761 * The page does not need to be reserved.
1762 *
1763 * Usually this function is called from f_op->mmap() handler
1764 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1765 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1766 * function from other places, for example from page-fault handler.
1767 */
1768int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1769 struct page *page)
1770{
1771 if (addr < vma->vm_start || addr >= vma->vm_end)
1772 return -EFAULT;
1773 if (!page_count(page))
1774 return -EINVAL;
1775 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1776 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1777 BUG_ON(vma->vm_flags & VM_PFNMAP);
1778 vma->vm_flags |= VM_MIXEDMAP;
1779 }
1780 return insert_page(vma, addr, page, vma->vm_page_prot);
1781}
1782EXPORT_SYMBOL(vm_insert_page);
1783
1784static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1785 pfn_t pfn, pgprot_t prot, bool mkwrite)
1786{
1787 struct mm_struct *mm = vma->vm_mm;
1788 int retval;
1789 pte_t *pte, entry;
1790 spinlock_t *ptl;
1791
1792 retval = -ENOMEM;
1793 pte = get_locked_pte(mm, addr, &ptl);
1794 if (!pte)
1795 goto out;
1796 retval = -EBUSY;
1797 if (!pte_none(*pte)) {
1798 if (mkwrite) {
1799 /*
1800 * For read faults on private mappings the PFN passed
1801 * in may not match the PFN we have mapped if the
1802 * mapped PFN is a writeable COW page. In the mkwrite
1803 * case we are creating a writable PTE for a shared
1804 * mapping and we expect the PFNs to match.
1805 */
1806 if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1807 goto out_unlock;
1808 entry = *pte;
1809 goto out_mkwrite;
1810 } else
1811 goto out_unlock;
1812 }
1813
1814 /* Ok, finally just insert the thing.. */
1815 if (pfn_t_devmap(pfn))
1816 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1817 else
1818 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1819
1820out_mkwrite:
1821 if (mkwrite) {
1822 entry = pte_mkyoung(entry);
1823 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1824 }
1825
1826 set_pte_at(mm, addr, pte, entry);
1827 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1828
1829 retval = 0;
1830out_unlock:
1831 pte_unmap_unlock(pte, ptl);
1832out:
1833 return retval;
1834}
1835
1836/**
1837 * vm_insert_pfn - insert single pfn into user vma
1838 * @vma: user vma to map to
1839 * @addr: target user address of this page
1840 * @pfn: source kernel pfn
1841 *
1842 * Similar to vm_insert_page, this allows drivers to insert individual pages
1843 * they've allocated into a user vma. Same comments apply.
1844 *
1845 * This function should only be called from a vm_ops->fault handler, and
1846 * in that case the handler should return NULL.
1847 *
1848 * vma cannot be a COW mapping.
1849 *
1850 * As this is called only for pages that do not currently exist, we
1851 * do not need to flush old virtual caches or the TLB.
1852 */
1853int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1854 unsigned long pfn)
1855{
1856 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1857}
1858EXPORT_SYMBOL(vm_insert_pfn);
1859
1860/**
1861 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1862 * @vma: user vma to map to
1863 * @addr: target user address of this page
1864 * @pfn: source kernel pfn
1865 * @pgprot: pgprot flags for the inserted page
1866 *
1867 * This is exactly like vm_insert_pfn, except that it allows drivers to
1868 * to override pgprot on a per-page basis.
1869 *
1870 * This only makes sense for IO mappings, and it makes no sense for
1871 * cow mappings. In general, using multiple vmas is preferable;
1872 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1873 * impractical.
1874 */
1875int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1876 unsigned long pfn, pgprot_t pgprot)
1877{
1878 int ret;
1879 /*
1880 * Technically, architectures with pte_special can avoid all these
1881 * restrictions (same for remap_pfn_range). However we would like
1882 * consistency in testing and feature parity among all, so we should
1883 * try to keep these invariants in place for everybody.
1884 */
1885 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1886 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1887 (VM_PFNMAP|VM_MIXEDMAP));
1888 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1889 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1890
1891 if (addr < vma->vm_start || addr >= vma->vm_end)
1892 return -EFAULT;
1893
1894 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1895
1896 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1897 false);
1898
1899 return ret;
1900}
1901EXPORT_SYMBOL(vm_insert_pfn_prot);
1902
1903static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1904{
1905 /* these checks mirror the abort conditions in vm_normal_page */
1906 if (vma->vm_flags & VM_MIXEDMAP)
1907 return true;
1908 if (pfn_t_devmap(pfn))
1909 return true;
1910 if (pfn_t_special(pfn))
1911 return true;
1912 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1913 return true;
1914 return false;
1915}
1916
1917static int __vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1918 pfn_t pfn, bool mkwrite)
1919{
1920 pgprot_t pgprot = vma->vm_page_prot;
1921
1922 BUG_ON(!vm_mixed_ok(vma, pfn));
1923
1924 if (addr < vma->vm_start || addr >= vma->vm_end)
1925 return -EFAULT;
1926
1927 track_pfn_insert(vma, &pgprot, pfn);
1928
1929 /*
1930 * If we don't have pte special, then we have to use the pfn_valid()
1931 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1932 * refcount the page if pfn_valid is true (hence insert_page rather
1933 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1934 * without pte special, it would there be refcounted as a normal page.
1935 */
1936 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1937 struct page *page;
1938
1939 /*
1940 * At this point we are committed to insert_page()
1941 * regardless of whether the caller specified flags that
1942 * result in pfn_t_has_page() == false.
1943 */
1944 page = pfn_to_page(pfn_t_to_pfn(pfn));
1945 return insert_page(vma, addr, page, pgprot);
1946 }
1947 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1948}
1949
1950int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1951 pfn_t pfn)
1952{
1953 return __vm_insert_mixed(vma, addr, pfn, false);
1954
1955}
1956EXPORT_SYMBOL(vm_insert_mixed);
1957
1958int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
1959 pfn_t pfn)
1960{
1961 return __vm_insert_mixed(vma, addr, pfn, true);
1962}
1963EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
1964
1965/*
1966 * maps a range of physical memory into the requested pages. the old
1967 * mappings are removed. any references to nonexistent pages results
1968 * in null mappings (currently treated as "copy-on-access")
1969 */
1970static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1971 unsigned long addr, unsigned long end,
1972 unsigned long pfn, pgprot_t prot)
1973{
1974 pte_t *pte;
1975 spinlock_t *ptl;
1976
1977 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1978 if (!pte)
1979 return -ENOMEM;
1980 arch_enter_lazy_mmu_mode();
1981 do {
1982 BUG_ON(!pte_none(*pte));
1983 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1984 pfn++;
1985 } while (pte++, addr += PAGE_SIZE, addr != end);
1986 arch_leave_lazy_mmu_mode();
1987 pte_unmap_unlock(pte - 1, ptl);
1988 return 0;
1989}
1990
1991static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1992 unsigned long addr, unsigned long end,
1993 unsigned long pfn, pgprot_t prot)
1994{
1995 pmd_t *pmd;
1996 unsigned long next;
1997
1998 pfn -= addr >> PAGE_SHIFT;
1999 pmd = pmd_alloc(mm, pud, addr);
2000 if (!pmd)
2001 return -ENOMEM;
2002 VM_BUG_ON(pmd_trans_huge(*pmd));
2003 do {
2004 next = pmd_addr_end(addr, end);
2005 if (remap_pte_range(mm, pmd, addr, next,
2006 pfn + (addr >> PAGE_SHIFT), prot))
2007 return -ENOMEM;
2008 } while (pmd++, addr = next, addr != end);
2009 return 0;
2010}
2011
2012static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2013 unsigned long addr, unsigned long end,
2014 unsigned long pfn, pgprot_t prot)
2015{
2016 pud_t *pud;
2017 unsigned long next;
2018
2019 pfn -= addr >> PAGE_SHIFT;
2020 pud = pud_alloc(mm, p4d, addr);
2021 if (!pud)
2022 return -ENOMEM;
2023 do {
2024 next = pud_addr_end(addr, end);
2025 if (remap_pmd_range(mm, pud, addr, next,
2026 pfn + (addr >> PAGE_SHIFT), prot))
2027 return -ENOMEM;
2028 } while (pud++, addr = next, addr != end);
2029 return 0;
2030}
2031
2032static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2033 unsigned long addr, unsigned long end,
2034 unsigned long pfn, pgprot_t prot)
2035{
2036 p4d_t *p4d;
2037 unsigned long next;
2038
2039 pfn -= addr >> PAGE_SHIFT;
2040 p4d = p4d_alloc(mm, pgd, addr);
2041 if (!p4d)
2042 return -ENOMEM;
2043 do {
2044 next = p4d_addr_end(addr, end);
2045 if (remap_pud_range(mm, p4d, addr, next,
2046 pfn + (addr >> PAGE_SHIFT), prot))
2047 return -ENOMEM;
2048 } while (p4d++, addr = next, addr != end);
2049 return 0;
2050}
2051
2052/**
2053 * remap_pfn_range - remap kernel memory to userspace
2054 * @vma: user vma to map to
2055 * @addr: target user address to start at
2056 * @pfn: physical address of kernel memory
2057 * @size: size of map area
2058 * @prot: page protection flags for this mapping
2059 *
2060 * Note: this is only safe if the mm semaphore is held when called.
2061 */
2062int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2063 unsigned long pfn, unsigned long size, pgprot_t prot)
2064{
2065 pgd_t *pgd;
2066 unsigned long next;
2067 unsigned long end = addr + PAGE_ALIGN(size);
2068 struct mm_struct *mm = vma->vm_mm;
2069 unsigned long remap_pfn = pfn;
2070 int err;
2071
2072 /*
2073 * Physically remapped pages are special. Tell the
2074 * rest of the world about it:
2075 * VM_IO tells people not to look at these pages
2076 * (accesses can have side effects).
2077 * VM_PFNMAP tells the core MM that the base pages are just
2078 * raw PFN mappings, and do not have a "struct page" associated
2079 * with them.
2080 * VM_DONTEXPAND
2081 * Disable vma merging and expanding with mremap().
2082 * VM_DONTDUMP
2083 * Omit vma from core dump, even when VM_IO turned off.
2084 *
2085 * There's a horrible special case to handle copy-on-write
2086 * behaviour that some programs depend on. We mark the "original"
2087 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2088 * See vm_normal_page() for details.
2089 */
2090 if (is_cow_mapping(vma->vm_flags)) {
2091 if (addr != vma->vm_start || end != vma->vm_end)
2092 return -EINVAL;
2093 vma->vm_pgoff = pfn;
2094 }
2095
2096 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2097 if (err)
2098 return -EINVAL;
2099
2100 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2101
2102 BUG_ON(addr >= end);
2103 pfn -= addr >> PAGE_SHIFT;
2104 pgd = pgd_offset(mm, addr);
2105 flush_cache_range(vma, addr, end);
2106 do {
2107 next = pgd_addr_end(addr, end);
2108 err = remap_p4d_range(mm, pgd, addr, next,
2109 pfn + (addr >> PAGE_SHIFT), prot);
2110 if (err)
2111 break;
2112 } while (pgd++, addr = next, addr != end);
2113
2114 if (err)
2115 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2116
2117 return err;
2118}
2119EXPORT_SYMBOL(remap_pfn_range);
2120
2121/**
2122 * vm_iomap_memory - remap memory to userspace
2123 * @vma: user vma to map to
2124 * @start: start of area
2125 * @len: size of area
2126 *
2127 * This is a simplified io_remap_pfn_range() for common driver use. The
2128 * driver just needs to give us the physical memory range to be mapped,
2129 * we'll figure out the rest from the vma information.
2130 *
2131 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2132 * whatever write-combining details or similar.
2133 */
2134int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2135{
2136 unsigned long vm_len, pfn, pages;
2137
2138 /* Check that the physical memory area passed in looks valid */
2139 if (start + len < start)
2140 return -EINVAL;
2141 /*
2142 * You *really* shouldn't map things that aren't page-aligned,
2143 * but we've historically allowed it because IO memory might
2144 * just have smaller alignment.
2145 */
2146 len += start & ~PAGE_MASK;
2147 pfn = start >> PAGE_SHIFT;
2148 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2149 if (pfn + pages < pfn)
2150 return -EINVAL;
2151
2152 /* We start the mapping 'vm_pgoff' pages into the area */
2153 if (vma->vm_pgoff > pages)
2154 return -EINVAL;
2155 pfn += vma->vm_pgoff;
2156 pages -= vma->vm_pgoff;
2157
2158 /* Can we fit all of the mapping? */
2159 vm_len = vma->vm_end - vma->vm_start;
2160 if (vm_len >> PAGE_SHIFT > pages)
2161 return -EINVAL;
2162
2163 /* Ok, let it rip */
2164 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2165}
2166EXPORT_SYMBOL(vm_iomap_memory);
2167
2168static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2169 unsigned long addr, unsigned long end,
2170 pte_fn_t fn, void *data)
2171{
2172 pte_t *pte;
2173 int err;
2174 pgtable_t token;
2175 spinlock_t *uninitialized_var(ptl);
2176
2177 pte = (mm == &init_mm) ?
2178 pte_alloc_kernel(pmd, addr) :
2179 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2180 if (!pte)
2181 return -ENOMEM;
2182
2183 BUG_ON(pmd_huge(*pmd));
2184
2185 arch_enter_lazy_mmu_mode();
2186
2187 token = pmd_pgtable(*pmd);
2188
2189 do {
2190 err = fn(pte++, token, addr, data);
2191 if (err)
2192 break;
2193 } while (addr += PAGE_SIZE, addr != end);
2194
2195 arch_leave_lazy_mmu_mode();
2196
2197 if (mm != &init_mm)
2198 pte_unmap_unlock(pte-1, ptl);
2199 return err;
2200}
2201
2202static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2203 unsigned long addr, unsigned long end,
2204 pte_fn_t fn, void *data)
2205{
2206 pmd_t *pmd;
2207 unsigned long next;
2208 int err;
2209
2210 BUG_ON(pud_huge(*pud));
2211
2212 pmd = pmd_alloc(mm, pud, addr);
2213 if (!pmd)
2214 return -ENOMEM;
2215 do {
2216 next = pmd_addr_end(addr, end);
2217 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2218 if (err)
2219 break;
2220 } while (pmd++, addr = next, addr != end);
2221 return err;
2222}
2223
2224static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2225 unsigned long addr, unsigned long end,
2226 pte_fn_t fn, void *data)
2227{
2228 pud_t *pud;
2229 unsigned long next;
2230 int err;
2231
2232 pud = pud_alloc(mm, p4d, addr);
2233 if (!pud)
2234 return -ENOMEM;
2235 do {
2236 next = pud_addr_end(addr, end);
2237 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2238 if (err)
2239 break;
2240 } while (pud++, addr = next, addr != end);
2241 return err;
2242}
2243
2244static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2245 unsigned long addr, unsigned long end,
2246 pte_fn_t fn, void *data)
2247{
2248 p4d_t *p4d;
2249 unsigned long next;
2250 int err;
2251
2252 p4d = p4d_alloc(mm, pgd, addr);
2253 if (!p4d)
2254 return -ENOMEM;
2255 do {
2256 next = p4d_addr_end(addr, end);
2257 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2258 if (err)
2259 break;
2260 } while (p4d++, addr = next, addr != end);
2261 return err;
2262}
2263
2264/*
2265 * Scan a region of virtual memory, filling in page tables as necessary
2266 * and calling a provided function on each leaf page table.
2267 */
2268int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2269 unsigned long size, pte_fn_t fn, void *data)
2270{
2271 pgd_t *pgd;
2272 unsigned long next;
2273 unsigned long end = addr + size;
2274 int err;
2275
2276 if (WARN_ON(addr >= end))
2277 return -EINVAL;
2278
2279 pgd = pgd_offset(mm, addr);
2280 do {
2281 next = pgd_addr_end(addr, end);
2282 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2283 if (err)
2284 break;
2285 } while (pgd++, addr = next, addr != end);
2286
2287 return err;
2288}
2289EXPORT_SYMBOL_GPL(apply_to_page_range);
2290
2291/*
2292 * handle_pte_fault chooses page fault handler according to an entry which was
2293 * read non-atomically. Before making any commitment, on those architectures
2294 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2295 * parts, do_swap_page must check under lock before unmapping the pte and
2296 * proceeding (but do_wp_page is only called after already making such a check;
2297 * and do_anonymous_page can safely check later on).
2298 */
2299static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2300 pte_t *page_table, pte_t orig_pte)
2301{
2302 int same = 1;
2303#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2304 if (sizeof(pte_t) > sizeof(unsigned long)) {
2305 spinlock_t *ptl = pte_lockptr(mm, pmd);
2306 spin_lock(ptl);
2307 same = pte_same(*page_table, orig_pte);
2308 spin_unlock(ptl);
2309 }
2310#endif
2311 pte_unmap(page_table);
2312 return same;
2313}
2314
2315static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2316{
2317 debug_dma_assert_idle(src);
2318
2319 /*
2320 * If the source page was a PFN mapping, we don't have
2321 * a "struct page" for it. We do a best-effort copy by
2322 * just copying from the original user address. If that
2323 * fails, we just zero-fill it. Live with it.
2324 */
2325 if (unlikely(!src)) {
2326 void *kaddr = kmap_atomic(dst);
2327 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2328
2329 /*
2330 * This really shouldn't fail, because the page is there
2331 * in the page tables. But it might just be unreadable,
2332 * in which case we just give up and fill the result with
2333 * zeroes.
2334 */
2335 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2336 clear_page(kaddr);
2337 kunmap_atomic(kaddr);
2338 flush_dcache_page(dst);
2339 } else
2340 copy_user_highpage(dst, src, va, vma);
2341}
2342
2343static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2344{
2345 struct file *vm_file = vma->vm_file;
2346
2347 if (vm_file)
2348 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2349
2350 /*
2351 * Special mappings (e.g. VDSO) do not have any file so fake
2352 * a default GFP_KERNEL for them.
2353 */
2354 return GFP_KERNEL;
2355}
2356
2357/*
2358 * Notify the address space that the page is about to become writable so that
2359 * it can prohibit this or wait for the page to get into an appropriate state.
2360 *
2361 * We do this without the lock held, so that it can sleep if it needs to.
2362 */
2363static int do_page_mkwrite(struct vm_fault *vmf)
2364{
2365 int ret;
2366 struct page *page = vmf->page;
2367 unsigned int old_flags = vmf->flags;
2368
2369 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2370
2371 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2372 /* Restore original flags so that caller is not surprised */
2373 vmf->flags = old_flags;
2374 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2375 return ret;
2376 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2377 lock_page(page);
2378 if (!page->mapping) {
2379 unlock_page(page);
2380 return 0; /* retry */
2381 }
2382 ret |= VM_FAULT_LOCKED;
2383 } else
2384 VM_BUG_ON_PAGE(!PageLocked(page), page);
2385 return ret;
2386}
2387
2388/*
2389 * Handle dirtying of a page in shared file mapping on a write fault.
2390 *
2391 * The function expects the page to be locked and unlocks it.
2392 */
2393static void fault_dirty_shared_page(struct vm_area_struct *vma,
2394 struct page *page)
2395{
2396 struct address_space *mapping;
2397 bool dirtied;
2398 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2399
2400 dirtied = set_page_dirty(page);
2401 VM_BUG_ON_PAGE(PageAnon(page), page);
2402 /*
2403 * Take a local copy of the address_space - page.mapping may be zeroed
2404 * by truncate after unlock_page(). The address_space itself remains
2405 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2406 * release semantics to prevent the compiler from undoing this copying.
2407 */
2408 mapping = page_rmapping(page);
2409 unlock_page(page);
2410
2411 if ((dirtied || page_mkwrite) && mapping) {
2412 /*
2413 * Some device drivers do not set page.mapping
2414 * but still dirty their pages
2415 */
2416 balance_dirty_pages_ratelimited(mapping);
2417 }
2418
2419 if (!page_mkwrite)
2420 file_update_time(vma->vm_file);
2421}
2422
2423/*
2424 * Handle write page faults for pages that can be reused in the current vma
2425 *
2426 * This can happen either due to the mapping being with the VM_SHARED flag,
2427 * or due to us being the last reference standing to the page. In either
2428 * case, all we need to do here is to mark the page as writable and update
2429 * any related book-keeping.
2430 */
2431static inline void wp_page_reuse(struct vm_fault *vmf)
2432 __releases(vmf->ptl)
2433{
2434 struct vm_area_struct *vma = vmf->vma;
2435 struct page *page = vmf->page;
2436 pte_t entry;
2437 /*
2438 * Clear the pages cpupid information as the existing
2439 * information potentially belongs to a now completely
2440 * unrelated process.
2441 */
2442 if (page)
2443 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2444
2445 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2446 entry = pte_mkyoung(vmf->orig_pte);
2447 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2448 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2449 update_mmu_cache(vma, vmf->address, vmf->pte);
2450 pte_unmap_unlock(vmf->pte, vmf->ptl);
2451}
2452
2453/*
2454 * Handle the case of a page which we actually need to copy to a new page.
2455 *
2456 * Called with mmap_sem locked and the old page referenced, but
2457 * without the ptl held.
2458 *
2459 * High level logic flow:
2460 *
2461 * - Allocate a page, copy the content of the old page to the new one.
2462 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2463 * - Take the PTL. If the pte changed, bail out and release the allocated page
2464 * - If the pte is still the way we remember it, update the page table and all
2465 * relevant references. This includes dropping the reference the page-table
2466 * held to the old page, as well as updating the rmap.
2467 * - In any case, unlock the PTL and drop the reference we took to the old page.
2468 */
2469static int wp_page_copy(struct vm_fault *vmf)
2470{
2471 struct vm_area_struct *vma = vmf->vma;
2472 struct mm_struct *mm = vma->vm_mm;
2473 struct page *old_page = vmf->page;
2474 struct page *new_page = NULL;
2475 pte_t entry;
2476 int page_copied = 0;
2477 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2478 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2479 struct mem_cgroup *memcg;
2480
2481 if (unlikely(anon_vma_prepare(vma)))
2482 goto oom;
2483
2484 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2485 new_page = alloc_zeroed_user_highpage_movable(vma,
2486 vmf->address);
2487 if (!new_page)
2488 goto oom;
2489 } else {
2490 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2491 vmf->address);
2492 if (!new_page)
2493 goto oom;
2494 cow_user_page(new_page, old_page, vmf->address, vma);
2495 }
2496
2497 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2498 goto oom_free_new;
2499
2500 __SetPageUptodate(new_page);
2501
2502 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2503
2504 /*
2505 * Re-check the pte - we dropped the lock
2506 */
2507 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2508 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2509 if (old_page) {
2510 if (!PageAnon(old_page)) {
2511 dec_mm_counter_fast(mm,
2512 mm_counter_file(old_page));
2513 inc_mm_counter_fast(mm, MM_ANONPAGES);
2514 }
2515 } else {
2516 inc_mm_counter_fast(mm, MM_ANONPAGES);
2517 }
2518 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2519 entry = mk_pte(new_page, vma->vm_page_prot);
2520 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2521 /*
2522 * Clear the pte entry and flush it first, before updating the
2523 * pte with the new entry. This will avoid a race condition
2524 * seen in the presence of one thread doing SMC and another
2525 * thread doing COW.
2526 */
2527 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2528 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2529 mem_cgroup_commit_charge(new_page, memcg, false, false);
2530 lru_cache_add_active_or_unevictable(new_page, vma);
2531 /*
2532 * We call the notify macro here because, when using secondary
2533 * mmu page tables (such as kvm shadow page tables), we want the
2534 * new page to be mapped directly into the secondary page table.
2535 */
2536 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2537 update_mmu_cache(vma, vmf->address, vmf->pte);
2538 if (old_page) {
2539 /*
2540 * Only after switching the pte to the new page may
2541 * we remove the mapcount here. Otherwise another
2542 * process may come and find the rmap count decremented
2543 * before the pte is switched to the new page, and
2544 * "reuse" the old page writing into it while our pte
2545 * here still points into it and can be read by other
2546 * threads.
2547 *
2548 * The critical issue is to order this
2549 * page_remove_rmap with the ptp_clear_flush above.
2550 * Those stores are ordered by (if nothing else,)
2551 * the barrier present in the atomic_add_negative
2552 * in page_remove_rmap.
2553 *
2554 * Then the TLB flush in ptep_clear_flush ensures that
2555 * no process can access the old page before the
2556 * decremented mapcount is visible. And the old page
2557 * cannot be reused until after the decremented
2558 * mapcount is visible. So transitively, TLBs to
2559 * old page will be flushed before it can be reused.
2560 */
2561 page_remove_rmap(old_page, false);
2562 }
2563
2564 /* Free the old page.. */
2565 new_page = old_page;
2566 page_copied = 1;
2567 } else {
2568 mem_cgroup_cancel_charge(new_page, memcg, false);
2569 }
2570
2571 if (new_page)
2572 put_page(new_page);
2573
2574 pte_unmap_unlock(vmf->pte, vmf->ptl);
2575 /*
2576 * No need to double call mmu_notifier->invalidate_range() callback as
2577 * the above ptep_clear_flush_notify() did already call it.
2578 */
2579 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2580 if (old_page) {
2581 /*
2582 * Don't let another task, with possibly unlocked vma,
2583 * keep the mlocked page.
2584 */
2585 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2586 lock_page(old_page); /* LRU manipulation */
2587 if (PageMlocked(old_page))
2588 munlock_vma_page(old_page);
2589 unlock_page(old_page);
2590 }
2591 put_page(old_page);
2592 }
2593 return page_copied ? VM_FAULT_WRITE : 0;
2594oom_free_new:
2595 put_page(new_page);
2596oom:
2597 if (old_page)
2598 put_page(old_page);
2599 return VM_FAULT_OOM;
2600}
2601
2602/**
2603 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2604 * writeable once the page is prepared
2605 *
2606 * @vmf: structure describing the fault
2607 *
2608 * This function handles all that is needed to finish a write page fault in a
2609 * shared mapping due to PTE being read-only once the mapped page is prepared.
2610 * It handles locking of PTE and modifying it. The function returns
2611 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2612 * lock.
2613 *
2614 * The function expects the page to be locked or other protection against
2615 * concurrent faults / writeback (such as DAX radix tree locks).
2616 */
2617int finish_mkwrite_fault(struct vm_fault *vmf)
2618{
2619 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2620 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2621 &vmf->ptl);
2622 /*
2623 * We might have raced with another page fault while we released the
2624 * pte_offset_map_lock.
2625 */
2626 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2627 pte_unmap_unlock(vmf->pte, vmf->ptl);
2628 return VM_FAULT_NOPAGE;
2629 }
2630 wp_page_reuse(vmf);
2631 return 0;
2632}
2633
2634/*
2635 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2636 * mapping
2637 */
2638static int wp_pfn_shared(struct vm_fault *vmf)
2639{
2640 struct vm_area_struct *vma = vmf->vma;
2641
2642 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2643 int ret;
2644
2645 pte_unmap_unlock(vmf->pte, vmf->ptl);
2646 vmf->flags |= FAULT_FLAG_MKWRITE;
2647 ret = vma->vm_ops->pfn_mkwrite(vmf);
2648 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2649 return ret;
2650 return finish_mkwrite_fault(vmf);
2651 }
2652 wp_page_reuse(vmf);
2653 return VM_FAULT_WRITE;
2654}
2655
2656static int wp_page_shared(struct vm_fault *vmf)
2657 __releases(vmf->ptl)
2658{
2659 struct vm_area_struct *vma = vmf->vma;
2660
2661 get_page(vmf->page);
2662
2663 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2664 int tmp;
2665
2666 pte_unmap_unlock(vmf->pte, vmf->ptl);
2667 tmp = do_page_mkwrite(vmf);
2668 if (unlikely(!tmp || (tmp &
2669 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2670 put_page(vmf->page);
2671 return tmp;
2672 }
2673 tmp = finish_mkwrite_fault(vmf);
2674 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2675 unlock_page(vmf->page);
2676 put_page(vmf->page);
2677 return tmp;
2678 }
2679 } else {
2680 wp_page_reuse(vmf);
2681 lock_page(vmf->page);
2682 }
2683 fault_dirty_shared_page(vma, vmf->page);
2684 put_page(vmf->page);
2685
2686 return VM_FAULT_WRITE;
2687}
2688
2689/*
2690 * This routine handles present pages, when users try to write
2691 * to a shared page. It is done by copying the page to a new address
2692 * and decrementing the shared-page counter for the old page.
2693 *
2694 * Note that this routine assumes that the protection checks have been
2695 * done by the caller (the low-level page fault routine in most cases).
2696 * Thus we can safely just mark it writable once we've done any necessary
2697 * COW.
2698 *
2699 * We also mark the page dirty at this point even though the page will
2700 * change only once the write actually happens. This avoids a few races,
2701 * and potentially makes it more efficient.
2702 *
2703 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2704 * but allow concurrent faults), with pte both mapped and locked.
2705 * We return with mmap_sem still held, but pte unmapped and unlocked.
2706 */
2707static int do_wp_page(struct vm_fault *vmf)
2708 __releases(vmf->ptl)
2709{
2710 struct vm_area_struct *vma = vmf->vma;
2711
2712 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2713 if (!vmf->page) {
2714 /*
2715 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2716 * VM_PFNMAP VMA.
2717 *
2718 * We should not cow pages in a shared writeable mapping.
2719 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2720 */
2721 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2722 (VM_WRITE|VM_SHARED))
2723 return wp_pfn_shared(vmf);
2724
2725 pte_unmap_unlock(vmf->pte, vmf->ptl);
2726 return wp_page_copy(vmf);
2727 }
2728
2729 /*
2730 * Take out anonymous pages first, anonymous shared vmas are
2731 * not dirty accountable.
2732 */
2733 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2734 int total_map_swapcount;
2735 if (!trylock_page(vmf->page)) {
2736 get_page(vmf->page);
2737 pte_unmap_unlock(vmf->pte, vmf->ptl);
2738 lock_page(vmf->page);
2739 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2740 vmf->address, &vmf->ptl);
2741 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2742 unlock_page(vmf->page);
2743 pte_unmap_unlock(vmf->pte, vmf->ptl);
2744 put_page(vmf->page);
2745 return 0;
2746 }
2747 put_page(vmf->page);
2748 }
2749 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2750 if (total_map_swapcount == 1) {
2751 /*
2752 * The page is all ours. Move it to
2753 * our anon_vma so the rmap code will
2754 * not search our parent or siblings.
2755 * Protected against the rmap code by
2756 * the page lock.
2757 */
2758 page_move_anon_rmap(vmf->page, vma);
2759 }
2760 unlock_page(vmf->page);
2761 wp_page_reuse(vmf);
2762 return VM_FAULT_WRITE;
2763 }
2764 unlock_page(vmf->page);
2765 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2766 (VM_WRITE|VM_SHARED))) {
2767 return wp_page_shared(vmf);
2768 }
2769
2770 /*
2771 * Ok, we need to copy. Oh, well..
2772 */
2773 get_page(vmf->page);
2774
2775 pte_unmap_unlock(vmf->pte, vmf->ptl);
2776 return wp_page_copy(vmf);
2777}
2778
2779static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2780 unsigned long start_addr, unsigned long end_addr,
2781 struct zap_details *details)
2782{
2783 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2784}
2785
2786static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2787 struct zap_details *details)
2788{
2789 struct vm_area_struct *vma;
2790 pgoff_t vba, vea, zba, zea;
2791
2792 vma_interval_tree_foreach(vma, root,
2793 details->first_index, details->last_index) {
2794
2795 vba = vma->vm_pgoff;
2796 vea = vba + vma_pages(vma) - 1;
2797 zba = details->first_index;
2798 if (zba < vba)
2799 zba = vba;
2800 zea = details->last_index;
2801 if (zea > vea)
2802 zea = vea;
2803
2804 unmap_mapping_range_vma(vma,
2805 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2806 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2807 details);
2808 }
2809}
2810
2811/**
2812 * unmap_mapping_pages() - Unmap pages from processes.
2813 * @mapping: The address space containing pages to be unmapped.
2814 * @start: Index of first page to be unmapped.
2815 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2816 * @even_cows: Whether to unmap even private COWed pages.
2817 *
2818 * Unmap the pages in this address space from any userspace process which
2819 * has them mmaped. Generally, you want to remove COWed pages as well when
2820 * a file is being truncated, but not when invalidating pages from the page
2821 * cache.
2822 */
2823void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2824 pgoff_t nr, bool even_cows)
2825{
2826 struct zap_details details = { };
2827
2828 details.check_mapping = even_cows ? NULL : mapping;
2829 details.first_index = start;
2830 details.last_index = start + nr - 1;
2831 if (details.last_index < details.first_index)
2832 details.last_index = ULONG_MAX;
2833
2834 i_mmap_lock_write(mapping);
2835 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2836 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2837 i_mmap_unlock_write(mapping);
2838}
2839
2840/**
2841 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2842 * address_space corresponding to the specified byte range in the underlying
2843 * file.
2844 *
2845 * @mapping: the address space containing mmaps to be unmapped.
2846 * @holebegin: byte in first page to unmap, relative to the start of
2847 * the underlying file. This will be rounded down to a PAGE_SIZE
2848 * boundary. Note that this is different from truncate_pagecache(), which
2849 * must keep the partial page. In contrast, we must get rid of
2850 * partial pages.
2851 * @holelen: size of prospective hole in bytes. This will be rounded
2852 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2853 * end of the file.
2854 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2855 * but 0 when invalidating pagecache, don't throw away private data.
2856 */
2857void unmap_mapping_range(struct address_space *mapping,
2858 loff_t const holebegin, loff_t const holelen, int even_cows)
2859{
2860 pgoff_t hba = holebegin >> PAGE_SHIFT;
2861 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2862
2863 /* Check for overflow. */
2864 if (sizeof(holelen) > sizeof(hlen)) {
2865 long long holeend =
2866 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2867 if (holeend & ~(long long)ULONG_MAX)
2868 hlen = ULONG_MAX - hba + 1;
2869 }
2870
2871 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2872}
2873EXPORT_SYMBOL(unmap_mapping_range);
2874
2875/*
2876 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2877 * but allow concurrent faults), and pte mapped but not yet locked.
2878 * We return with pte unmapped and unlocked.
2879 *
2880 * We return with the mmap_sem locked or unlocked in the same cases
2881 * as does filemap_fault().
2882 */
2883int do_swap_page(struct vm_fault *vmf)
2884{
2885 struct vm_area_struct *vma = vmf->vma;
2886 struct page *page = NULL, *swapcache;
2887 struct mem_cgroup *memcg;
2888 swp_entry_t entry;
2889 pte_t pte;
2890 int locked;
2891 int exclusive = 0;
2892 int ret = 0;
2893
2894 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2895 goto out;
2896
2897 entry = pte_to_swp_entry(vmf->orig_pte);
2898 if (unlikely(non_swap_entry(entry))) {
2899 if (is_migration_entry(entry)) {
2900 migration_entry_wait(vma->vm_mm, vmf->pmd,
2901 vmf->address);
2902 } else if (is_device_private_entry(entry)) {
2903 /*
2904 * For un-addressable device memory we call the pgmap
2905 * fault handler callback. The callback must migrate
2906 * the page back to some CPU accessible page.
2907 */
2908 ret = device_private_entry_fault(vma, vmf->address, entry,
2909 vmf->flags, vmf->pmd);
2910 } else if (is_hwpoison_entry(entry)) {
2911 ret = VM_FAULT_HWPOISON;
2912 } else {
2913 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2914 ret = VM_FAULT_SIGBUS;
2915 }
2916 goto out;
2917 }
2918
2919
2920 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2921 page = lookup_swap_cache(entry, vma, vmf->address);
2922 swapcache = page;
2923
2924 if (!page) {
2925 struct swap_info_struct *si = swp_swap_info(entry);
2926
2927 if (si->flags & SWP_SYNCHRONOUS_IO &&
2928 __swap_count(si, entry) == 1) {
2929 /* skip swapcache */
2930 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2931 vmf->address);
2932 if (page) {
2933 __SetPageLocked(page);
2934 __SetPageSwapBacked(page);
2935 set_page_private(page, entry.val);
2936 lru_cache_add_anon(page);
2937 swap_readpage(page, true);
2938 }
2939 } else {
2940 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2941 vmf);
2942 swapcache = page;
2943 }
2944
2945 if (!page) {
2946 /*
2947 * Back out if somebody else faulted in this pte
2948 * while we released the pte lock.
2949 */
2950 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2951 vmf->address, &vmf->ptl);
2952 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2953 ret = VM_FAULT_OOM;
2954 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2955 goto unlock;
2956 }
2957
2958 /* Had to read the page from swap area: Major fault */
2959 ret = VM_FAULT_MAJOR;
2960 count_vm_event(PGMAJFAULT);
2961 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2962 } else if (PageHWPoison(page)) {
2963 /*
2964 * hwpoisoned dirty swapcache pages are kept for killing
2965 * owner processes (which may be unknown at hwpoison time)
2966 */
2967 ret = VM_FAULT_HWPOISON;
2968 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2969 goto out_release;
2970 }
2971
2972 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2973
2974 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2975 if (!locked) {
2976 ret |= VM_FAULT_RETRY;
2977 goto out_release;
2978 }
2979
2980 /*
2981 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2982 * release the swapcache from under us. The page pin, and pte_same
2983 * test below, are not enough to exclude that. Even if it is still
2984 * swapcache, we need to check that the page's swap has not changed.
2985 */
2986 if (unlikely((!PageSwapCache(page) ||
2987 page_private(page) != entry.val)) && swapcache)
2988 goto out_page;
2989
2990 page = ksm_might_need_to_copy(page, vma, vmf->address);
2991 if (unlikely(!page)) {
2992 ret = VM_FAULT_OOM;
2993 page = swapcache;
2994 goto out_page;
2995 }
2996
2997 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2998 &memcg, false)) {
2999 ret = VM_FAULT_OOM;
3000 goto out_page;
3001 }
3002
3003 /*
3004 * Back out if somebody else already faulted in this pte.
3005 */
3006 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3007 &vmf->ptl);
3008 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3009 goto out_nomap;
3010
3011 if (unlikely(!PageUptodate(page))) {
3012 ret = VM_FAULT_SIGBUS;
3013 goto out_nomap;
3014 }
3015
3016 /*
3017 * The page isn't present yet, go ahead with the fault.
3018 *
3019 * Be careful about the sequence of operations here.
3020 * To get its accounting right, reuse_swap_page() must be called
3021 * while the page is counted on swap but not yet in mapcount i.e.
3022 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3023 * must be called after the swap_free(), or it will never succeed.
3024 */
3025
3026 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3027 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3028 pte = mk_pte(page, vma->vm_page_prot);
3029 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3030 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3031 vmf->flags &= ~FAULT_FLAG_WRITE;
3032 ret |= VM_FAULT_WRITE;
3033 exclusive = RMAP_EXCLUSIVE;
3034 }
3035 flush_icache_page(vma, page);
3036 if (pte_swp_soft_dirty(vmf->orig_pte))
3037 pte = pte_mksoft_dirty(pte);
3038 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3039 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3040 vmf->orig_pte = pte;
3041
3042 /* ksm created a completely new copy */
3043 if (unlikely(page != swapcache && swapcache)) {
3044 page_add_new_anon_rmap(page, vma, vmf->address, false);
3045 mem_cgroup_commit_charge(page, memcg, false, false);
3046 lru_cache_add_active_or_unevictable(page, vma);
3047 } else {
3048 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3049 mem_cgroup_commit_charge(page, memcg, true, false);
3050 activate_page(page);
3051 }
3052
3053 swap_free(entry);
3054 if (mem_cgroup_swap_full(page) ||
3055 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3056 try_to_free_swap(page);
3057 unlock_page(page);
3058 if (page != swapcache && swapcache) {
3059 /*
3060 * Hold the lock to avoid the swap entry to be reused
3061 * until we take the PT lock for the pte_same() check
3062 * (to avoid false positives from pte_same). For
3063 * further safety release the lock after the swap_free
3064 * so that the swap count won't change under a
3065 * parallel locked swapcache.
3066 */
3067 unlock_page(swapcache);
3068 put_page(swapcache);
3069 }
3070
3071 if (vmf->flags & FAULT_FLAG_WRITE) {
3072 ret |= do_wp_page(vmf);
3073 if (ret & VM_FAULT_ERROR)
3074 ret &= VM_FAULT_ERROR;
3075 goto out;
3076 }
3077
3078 /* No need to invalidate - it was non-present before */
3079 update_mmu_cache(vma, vmf->address, vmf->pte);
3080unlock:
3081 pte_unmap_unlock(vmf->pte, vmf->ptl);
3082out:
3083 return ret;
3084out_nomap:
3085 mem_cgroup_cancel_charge(page, memcg, false);
3086 pte_unmap_unlock(vmf->pte, vmf->ptl);
3087out_page:
3088 unlock_page(page);
3089out_release:
3090 put_page(page);
3091 if (page != swapcache && swapcache) {
3092 unlock_page(swapcache);
3093 put_page(swapcache);
3094 }
3095 return ret;
3096}
3097
3098/*
3099 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3100 * but allow concurrent faults), and pte mapped but not yet locked.
3101 * We return with mmap_sem still held, but pte unmapped and unlocked.
3102 */
3103static int do_anonymous_page(struct vm_fault *vmf)
3104{
3105 struct vm_area_struct *vma = vmf->vma;
3106 struct mem_cgroup *memcg;
3107 struct page *page;
3108 int ret = 0;
3109 pte_t entry;
3110
3111 /* File mapping without ->vm_ops ? */
3112 if (vma->vm_flags & VM_SHARED)
3113 return VM_FAULT_SIGBUS;
3114
3115 /*
3116 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3117 * pte_offset_map() on pmds where a huge pmd might be created
3118 * from a different thread.
3119 *
3120 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3121 * parallel threads are excluded by other means.
3122 *
3123 * Here we only have down_read(mmap_sem).
3124 */
3125 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
3126 return VM_FAULT_OOM;
3127
3128 /* See the comment in pte_alloc_one_map() */
3129 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3130 return 0;
3131
3132 /* Use the zero-page for reads */
3133 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3134 !mm_forbids_zeropage(vma->vm_mm)) {
3135 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3136 vma->vm_page_prot));
3137 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3138 vmf->address, &vmf->ptl);
3139 if (!pte_none(*vmf->pte))
3140 goto unlock;
3141 ret = check_stable_address_space(vma->vm_mm);
3142 if (ret)
3143 goto unlock;
3144 /* Deliver the page fault to userland, check inside PT lock */
3145 if (userfaultfd_missing(vma)) {
3146 pte_unmap_unlock(vmf->pte, vmf->ptl);
3147 return handle_userfault(vmf, VM_UFFD_MISSING);
3148 }
3149 goto setpte;
3150 }
3151
3152 /* Allocate our own private page. */
3153 if (unlikely(anon_vma_prepare(vma)))
3154 goto oom;
3155 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3156 if (!page)
3157 goto oom;
3158
3159 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
3160 goto oom_free_page;
3161
3162 /*
3163 * The memory barrier inside __SetPageUptodate makes sure that
3164 * preceeding stores to the page contents become visible before
3165 * the set_pte_at() write.
3166 */
3167 __SetPageUptodate(page);
3168
3169 entry = mk_pte(page, vma->vm_page_prot);
3170 if (vma->vm_flags & VM_WRITE)
3171 entry = pte_mkwrite(pte_mkdirty(entry));
3172
3173 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3174 &vmf->ptl);
3175 if (!pte_none(*vmf->pte))
3176 goto release;
3177
3178 ret = check_stable_address_space(vma->vm_mm);
3179 if (ret)
3180 goto release;
3181
3182 /* Deliver the page fault to userland, check inside PT lock */
3183 if (userfaultfd_missing(vma)) {
3184 pte_unmap_unlock(vmf->pte, vmf->ptl);
3185 mem_cgroup_cancel_charge(page, memcg, false);
3186 put_page(page);
3187 return handle_userfault(vmf, VM_UFFD_MISSING);
3188 }
3189
3190 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3191 page_add_new_anon_rmap(page, vma, vmf->address, false);
3192 mem_cgroup_commit_charge(page, memcg, false, false);
3193 lru_cache_add_active_or_unevictable(page, vma);
3194setpte:
3195 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3196
3197 /* No need to invalidate - it was non-present before */
3198 update_mmu_cache(vma, vmf->address, vmf->pte);
3199unlock:
3200 pte_unmap_unlock(vmf->pte, vmf->ptl);
3201 return ret;
3202release:
3203 mem_cgroup_cancel_charge(page, memcg, false);
3204 put_page(page);
3205 goto unlock;
3206oom_free_page:
3207 put_page(page);
3208oom:
3209 return VM_FAULT_OOM;
3210}
3211
3212/*
3213 * The mmap_sem must have been held on entry, and may have been
3214 * released depending on flags and vma->vm_ops->fault() return value.
3215 * See filemap_fault() and __lock_page_retry().
3216 */
3217static int __do_fault(struct vm_fault *vmf)
3218{
3219 struct vm_area_struct *vma = vmf->vma;
3220 int ret;
3221
3222 ret = vma->vm_ops->fault(vmf);
3223 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3224 VM_FAULT_DONE_COW)))
3225 return ret;
3226
3227 if (unlikely(PageHWPoison(vmf->page))) {
3228 if (ret & VM_FAULT_LOCKED)
3229 unlock_page(vmf->page);
3230 put_page(vmf->page);
3231 vmf->page = NULL;
3232 return VM_FAULT_HWPOISON;
3233 }
3234
3235 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3236 lock_page(vmf->page);
3237 else
3238 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3239
3240 return ret;
3241}
3242
3243/*
3244 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3245 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3246 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3247 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3248 */
3249static int pmd_devmap_trans_unstable(pmd_t *pmd)
3250{
3251 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3252}
3253
3254static int pte_alloc_one_map(struct vm_fault *vmf)
3255{
3256 struct vm_area_struct *vma = vmf->vma;
3257
3258 if (!pmd_none(*vmf->pmd))
3259 goto map_pte;
3260 if (vmf->prealloc_pte) {
3261 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3262 if (unlikely(!pmd_none(*vmf->pmd))) {
3263 spin_unlock(vmf->ptl);
3264 goto map_pte;
3265 }
3266
3267 mm_inc_nr_ptes(vma->vm_mm);
3268 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3269 spin_unlock(vmf->ptl);
3270 vmf->prealloc_pte = NULL;
3271 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3272 return VM_FAULT_OOM;
3273 }
3274map_pte:
3275 /*
3276 * If a huge pmd materialized under us just retry later. Use
3277 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3278 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3279 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3280 * running immediately after a huge pmd fault in a different thread of
3281 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3282 * All we have to ensure is that it is a regular pmd that we can walk
3283 * with pte_offset_map() and we can do that through an atomic read in
3284 * C, which is what pmd_trans_unstable() provides.
3285 */
3286 if (pmd_devmap_trans_unstable(vmf->pmd))
3287 return VM_FAULT_NOPAGE;
3288
3289 /*
3290 * At this point we know that our vmf->pmd points to a page of ptes
3291 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3292 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3293 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3294 * be valid and we will re-check to make sure the vmf->pte isn't
3295 * pte_none() under vmf->ptl protection when we return to
3296 * alloc_set_pte().
3297 */
3298 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3299 &vmf->ptl);
3300 return 0;
3301}
3302
3303#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3304
3305#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3306static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3307 unsigned long haddr)
3308{
3309 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3310 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3311 return false;
3312 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3313 return false;
3314 return true;
3315}
3316
3317static void deposit_prealloc_pte(struct vm_fault *vmf)
3318{
3319 struct vm_area_struct *vma = vmf->vma;
3320
3321 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3322 /*
3323 * We are going to consume the prealloc table,
3324 * count that as nr_ptes.
3325 */
3326 mm_inc_nr_ptes(vma->vm_mm);
3327 vmf->prealloc_pte = NULL;
3328}
3329
3330static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3331{
3332 struct vm_area_struct *vma = vmf->vma;
3333 bool write = vmf->flags & FAULT_FLAG_WRITE;
3334 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3335 pmd_t entry;
3336 int i, ret;
3337
3338 if (!transhuge_vma_suitable(vma, haddr))
3339 return VM_FAULT_FALLBACK;
3340
3341 ret = VM_FAULT_FALLBACK;
3342 page = compound_head(page);
3343
3344 /*
3345 * Archs like ppc64 need additonal space to store information
3346 * related to pte entry. Use the preallocated table for that.
3347 */
3348 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3349 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3350 if (!vmf->prealloc_pte)
3351 return VM_FAULT_OOM;
3352 smp_wmb(); /* See comment in __pte_alloc() */
3353 }
3354
3355 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3356 if (unlikely(!pmd_none(*vmf->pmd)))
3357 goto out;
3358
3359 for (i = 0; i < HPAGE_PMD_NR; i++)
3360 flush_icache_page(vma, page + i);
3361
3362 entry = mk_huge_pmd(page, vma->vm_page_prot);
3363 if (write)
3364 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3365
3366 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3367 page_add_file_rmap(page, true);
3368 /*
3369 * deposit and withdraw with pmd lock held
3370 */
3371 if (arch_needs_pgtable_deposit())
3372 deposit_prealloc_pte(vmf);
3373
3374 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3375
3376 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3377
3378 /* fault is handled */
3379 ret = 0;
3380 count_vm_event(THP_FILE_MAPPED);
3381out:
3382 spin_unlock(vmf->ptl);
3383 return ret;
3384}
3385#else
3386static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3387{
3388 BUILD_BUG();
3389 return 0;
3390}
3391#endif
3392
3393/**
3394 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3395 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3396 *
3397 * @vmf: fault environment
3398 * @memcg: memcg to charge page (only for private mappings)
3399 * @page: page to map
3400 *
3401 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3402 * return.
3403 *
3404 * Target users are page handler itself and implementations of
3405 * vm_ops->map_pages.
3406 */
3407int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3408 struct page *page)
3409{
3410 struct vm_area_struct *vma = vmf->vma;
3411 bool write = vmf->flags & FAULT_FLAG_WRITE;
3412 pte_t entry;
3413 int ret;
3414
3415 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3416 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3417 /* THP on COW? */
3418 VM_BUG_ON_PAGE(memcg, page);
3419
3420 ret = do_set_pmd(vmf, page);
3421 if (ret != VM_FAULT_FALLBACK)
3422 return ret;
3423 }
3424
3425 if (!vmf->pte) {
3426 ret = pte_alloc_one_map(vmf);
3427 if (ret)
3428 return ret;
3429 }
3430
3431 /* Re-check under ptl */
3432 if (unlikely(!pte_none(*vmf->pte)))
3433 return VM_FAULT_NOPAGE;
3434
3435 flush_icache_page(vma, page);
3436 entry = mk_pte(page, vma->vm_page_prot);
3437 if (write)
3438 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3439 /* copy-on-write page */
3440 if (write && !(vma->vm_flags & VM_SHARED)) {
3441 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3442 page_add_new_anon_rmap(page, vma, vmf->address, false);
3443 mem_cgroup_commit_charge(page, memcg, false, false);
3444 lru_cache_add_active_or_unevictable(page, vma);
3445 } else {
3446 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3447 page_add_file_rmap(page, false);
3448 }
3449 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3450
3451 /* no need to invalidate: a not-present page won't be cached */
3452 update_mmu_cache(vma, vmf->address, vmf->pte);
3453
3454 return 0;
3455}
3456
3457
3458/**
3459 * finish_fault - finish page fault once we have prepared the page to fault
3460 *
3461 * @vmf: structure describing the fault
3462 *
3463 * This function handles all that is needed to finish a page fault once the
3464 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3465 * given page, adds reverse page mapping, handles memcg charges and LRU
3466 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3467 * error.
3468 *
3469 * The function expects the page to be locked and on success it consumes a
3470 * reference of a page being mapped (for the PTE which maps it).
3471 */
3472int finish_fault(struct vm_fault *vmf)
3473{
3474 struct page *page;
3475 int ret = 0;
3476
3477 /* Did we COW the page? */
3478 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3479 !(vmf->vma->vm_flags & VM_SHARED))
3480 page = vmf->cow_page;
3481 else
3482 page = vmf->page;
3483
3484 /*
3485 * check even for read faults because we might have lost our CoWed
3486 * page
3487 */
3488 if (!(vmf->vma->vm_flags & VM_SHARED))
3489 ret = check_stable_address_space(vmf->vma->vm_mm);
3490 if (!ret)
3491 ret = alloc_set_pte(vmf, vmf->memcg, page);
3492 if (vmf->pte)
3493 pte_unmap_unlock(vmf->pte, vmf->ptl);
3494 return ret;
3495}
3496
3497static unsigned long fault_around_bytes __read_mostly =
3498 rounddown_pow_of_two(65536);
3499
3500#ifdef CONFIG_DEBUG_FS
3501static int fault_around_bytes_get(void *data, u64 *val)
3502{
3503 *val = fault_around_bytes;
3504 return 0;
3505}
3506
3507/*
3508 * fault_around_bytes must be rounded down to the nearest page order as it's
3509 * what do_fault_around() expects to see.
3510 */
3511static int fault_around_bytes_set(void *data, u64 val)
3512{
3513 if (val / PAGE_SIZE > PTRS_PER_PTE)
3514 return -EINVAL;
3515 if (val > PAGE_SIZE)
3516 fault_around_bytes = rounddown_pow_of_two(val);
3517 else
3518 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3519 return 0;
3520}
3521DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3522 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3523
3524static int __init fault_around_debugfs(void)
3525{
3526 void *ret;
3527
3528 ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3529 &fault_around_bytes_fops);
3530 if (!ret)
3531 pr_warn("Failed to create fault_around_bytes in debugfs");
3532 return 0;
3533}
3534late_initcall(fault_around_debugfs);
3535#endif
3536
3537/*
3538 * do_fault_around() tries to map few pages around the fault address. The hope
3539 * is that the pages will be needed soon and this will lower the number of
3540 * faults to handle.
3541 *
3542 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3543 * not ready to be mapped: not up-to-date, locked, etc.
3544 *
3545 * This function is called with the page table lock taken. In the split ptlock
3546 * case the page table lock only protects only those entries which belong to
3547 * the page table corresponding to the fault address.
3548 *
3549 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3550 * only once.
3551 *
3552 * fault_around_bytes defines how many bytes we'll try to map.
3553 * do_fault_around() expects it to be set to a power of two less than or equal
3554 * to PTRS_PER_PTE.
3555 *
3556 * The virtual address of the area that we map is naturally aligned to
3557 * fault_around_bytes rounded down to the machine page size
3558 * (and therefore to page order). This way it's easier to guarantee
3559 * that we don't cross page table boundaries.
3560 */
3561static int do_fault_around(struct vm_fault *vmf)
3562{
3563 unsigned long address = vmf->address, nr_pages, mask;
3564 pgoff_t start_pgoff = vmf->pgoff;
3565 pgoff_t end_pgoff;
3566 int off, ret = 0;
3567
3568 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3569 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3570
3571 vmf->address = max(address & mask, vmf->vma->vm_start);
3572 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3573 start_pgoff -= off;
3574
3575 /*
3576 * end_pgoff is either the end of the page table, the end of
3577 * the vma or nr_pages from start_pgoff, depending what is nearest.
3578 */
3579 end_pgoff = start_pgoff -
3580 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3581 PTRS_PER_PTE - 1;
3582 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3583 start_pgoff + nr_pages - 1);
3584
3585 if (pmd_none(*vmf->pmd)) {
3586 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3587 vmf->address);
3588 if (!vmf->prealloc_pte)
3589 goto out;
3590 smp_wmb(); /* See comment in __pte_alloc() */
3591 }
3592
3593 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3594
3595 /* Huge page is mapped? Page fault is solved */
3596 if (pmd_trans_huge(*vmf->pmd)) {
3597 ret = VM_FAULT_NOPAGE;
3598 goto out;
3599 }
3600
3601 /* ->map_pages() haven't done anything useful. Cold page cache? */
3602 if (!vmf->pte)
3603 goto out;
3604
3605 /* check if the page fault is solved */
3606 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3607 if (!pte_none(*vmf->pte))
3608 ret = VM_FAULT_NOPAGE;
3609 pte_unmap_unlock(vmf->pte, vmf->ptl);
3610out:
3611 vmf->address = address;
3612 vmf->pte = NULL;
3613 return ret;
3614}
3615
3616static int do_read_fault(struct vm_fault *vmf)
3617{
3618 struct vm_area_struct *vma = vmf->vma;
3619 int ret = 0;
3620
3621 /*
3622 * Let's call ->map_pages() first and use ->fault() as fallback
3623 * if page by the offset is not ready to be mapped (cold cache or
3624 * something).
3625 */
3626 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3627 ret = do_fault_around(vmf);
3628 if (ret)
3629 return ret;
3630 }
3631
3632 ret = __do_fault(vmf);
3633 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3634 return ret;
3635
3636 ret |= finish_fault(vmf);
3637 unlock_page(vmf->page);
3638 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3639 put_page(vmf->page);
3640 return ret;
3641}
3642
3643static int do_cow_fault(struct vm_fault *vmf)
3644{
3645 struct vm_area_struct *vma = vmf->vma;
3646 int ret;
3647
3648 if (unlikely(anon_vma_prepare(vma)))
3649 return VM_FAULT_OOM;
3650
3651 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3652 if (!vmf->cow_page)
3653 return VM_FAULT_OOM;
3654
3655 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3656 &vmf->memcg, false)) {
3657 put_page(vmf->cow_page);
3658 return VM_FAULT_OOM;
3659 }
3660
3661 ret = __do_fault(vmf);
3662 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3663 goto uncharge_out;
3664 if (ret & VM_FAULT_DONE_COW)
3665 return ret;
3666
3667 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3668 __SetPageUptodate(vmf->cow_page);
3669
3670 ret |= finish_fault(vmf);
3671 unlock_page(vmf->page);
3672 put_page(vmf->page);
3673 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3674 goto uncharge_out;
3675 return ret;
3676uncharge_out:
3677 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3678 put_page(vmf->cow_page);
3679 return ret;
3680}
3681
3682static int do_shared_fault(struct vm_fault *vmf)
3683{
3684 struct vm_area_struct *vma = vmf->vma;
3685 int ret, tmp;
3686
3687 ret = __do_fault(vmf);
3688 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3689 return ret;
3690
3691 /*
3692 * Check if the backing address space wants to know that the page is
3693 * about to become writable
3694 */
3695 if (vma->vm_ops->page_mkwrite) {
3696 unlock_page(vmf->page);
3697 tmp = do_page_mkwrite(vmf);
3698 if (unlikely(!tmp ||
3699 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3700 put_page(vmf->page);
3701 return tmp;
3702 }
3703 }
3704
3705 ret |= finish_fault(vmf);
3706 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3707 VM_FAULT_RETRY))) {
3708 unlock_page(vmf->page);
3709 put_page(vmf->page);
3710 return ret;
3711 }
3712
3713 fault_dirty_shared_page(vma, vmf->page);
3714 return ret;
3715}
3716
3717/*
3718 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3719 * but allow concurrent faults).
3720 * The mmap_sem may have been released depending on flags and our
3721 * return value. See filemap_fault() and __lock_page_or_retry().
3722 */
3723static int do_fault(struct vm_fault *vmf)
3724{
3725 struct vm_area_struct *vma = vmf->vma;
3726 int ret;
3727
3728 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3729 if (!vma->vm_ops->fault)
3730 ret = VM_FAULT_SIGBUS;
3731 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3732 ret = do_read_fault(vmf);
3733 else if (!(vma->vm_flags & VM_SHARED))
3734 ret = do_cow_fault(vmf);
3735 else
3736 ret = do_shared_fault(vmf);
3737
3738 /* preallocated pagetable is unused: free it */
3739 if (vmf->prealloc_pte) {
3740 pte_free(vma->vm_mm, vmf->prealloc_pte);
3741 vmf->prealloc_pte = NULL;
3742 }
3743 return ret;
3744}
3745
3746static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3747 unsigned long addr, int page_nid,
3748 int *flags)
3749{
3750 get_page(page);
3751
3752 count_vm_numa_event(NUMA_HINT_FAULTS);
3753 if (page_nid == numa_node_id()) {
3754 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3755 *flags |= TNF_FAULT_LOCAL;
3756 }
3757
3758 return mpol_misplaced(page, vma, addr);
3759}
3760
3761static int do_numa_page(struct vm_fault *vmf)
3762{
3763 struct vm_area_struct *vma = vmf->vma;
3764 struct page *page = NULL;
3765 int page_nid = -1;
3766 int last_cpupid;
3767 int target_nid;
3768 bool migrated = false;
3769 pte_t pte;
3770 bool was_writable = pte_savedwrite(vmf->orig_pte);
3771 int flags = 0;
3772
3773 /*
3774 * The "pte" at this point cannot be used safely without
3775 * validation through pte_unmap_same(). It's of NUMA type but
3776 * the pfn may be screwed if the read is non atomic.
3777 */
3778 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3779 spin_lock(vmf->ptl);
3780 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3781 pte_unmap_unlock(vmf->pte, vmf->ptl);
3782 goto out;
3783 }
3784
3785 /*
3786 * Make it present again, Depending on how arch implementes non
3787 * accessible ptes, some can allow access by kernel mode.
3788 */
3789 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3790 pte = pte_modify(pte, vma->vm_page_prot);
3791 pte = pte_mkyoung(pte);
3792 if (was_writable)
3793 pte = pte_mkwrite(pte);
3794 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3795 update_mmu_cache(vma, vmf->address, vmf->pte);
3796
3797 page = vm_normal_page(vma, vmf->address, pte);
3798 if (!page) {
3799 pte_unmap_unlock(vmf->pte, vmf->ptl);
3800 return 0;
3801 }
3802
3803 /* TODO: handle PTE-mapped THP */
3804 if (PageCompound(page)) {
3805 pte_unmap_unlock(vmf->pte, vmf->ptl);
3806 return 0;
3807 }
3808
3809 /*
3810 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3811 * much anyway since they can be in shared cache state. This misses
3812 * the case where a mapping is writable but the process never writes
3813 * to it but pte_write gets cleared during protection updates and
3814 * pte_dirty has unpredictable behaviour between PTE scan updates,
3815 * background writeback, dirty balancing and application behaviour.
3816 */
3817 if (!pte_write(pte))
3818 flags |= TNF_NO_GROUP;
3819
3820 /*
3821 * Flag if the page is shared between multiple address spaces. This
3822 * is later used when determining whether to group tasks together
3823 */
3824 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3825 flags |= TNF_SHARED;
3826
3827 last_cpupid = page_cpupid_last(page);
3828 page_nid = page_to_nid(page);
3829 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3830 &flags);
3831 pte_unmap_unlock(vmf->pte, vmf->ptl);
3832 if (target_nid == -1) {
3833 put_page(page);
3834 goto out;
3835 }
3836
3837 /* Migrate to the requested node */
3838 migrated = migrate_misplaced_page(page, vma, target_nid);
3839 if (migrated) {
3840 page_nid = target_nid;
3841 flags |= TNF_MIGRATED;
3842 } else
3843 flags |= TNF_MIGRATE_FAIL;
3844
3845out:
3846 if (page_nid != -1)
3847 task_numa_fault(last_cpupid, page_nid, 1, flags);
3848 return 0;
3849}
3850
3851static inline int create_huge_pmd(struct vm_fault *vmf)
3852{
3853 if (vma_is_anonymous(vmf->vma))
3854 return do_huge_pmd_anonymous_page(vmf);
3855 if (vmf->vma->vm_ops->huge_fault)
3856 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3857 return VM_FAULT_FALLBACK;
3858}
3859
3860/* `inline' is required to avoid gcc 4.1.2 build error */
3861static inline int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3862{
3863 if (vma_is_anonymous(vmf->vma))
3864 return do_huge_pmd_wp_page(vmf, orig_pmd);
3865 if (vmf->vma->vm_ops->huge_fault)
3866 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3867
3868 /* COW handled on pte level: split pmd */
3869 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3870 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3871
3872 return VM_FAULT_FALLBACK;
3873}
3874
3875static inline bool vma_is_accessible(struct vm_area_struct *vma)
3876{
3877 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3878}
3879
3880static int create_huge_pud(struct vm_fault *vmf)
3881{
3882#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3883 /* No support for anonymous transparent PUD pages yet */
3884 if (vma_is_anonymous(vmf->vma))
3885 return VM_FAULT_FALLBACK;
3886 if (vmf->vma->vm_ops->huge_fault)
3887 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3888#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3889 return VM_FAULT_FALLBACK;
3890}
3891
3892static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3893{
3894#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3895 /* No support for anonymous transparent PUD pages yet */
3896 if (vma_is_anonymous(vmf->vma))
3897 return VM_FAULT_FALLBACK;
3898 if (vmf->vma->vm_ops->huge_fault)
3899 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3900#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3901 return VM_FAULT_FALLBACK;
3902}
3903
3904/*
3905 * These routines also need to handle stuff like marking pages dirty
3906 * and/or accessed for architectures that don't do it in hardware (most
3907 * RISC architectures). The early dirtying is also good on the i386.
3908 *
3909 * There is also a hook called "update_mmu_cache()" that architectures
3910 * with external mmu caches can use to update those (ie the Sparc or
3911 * PowerPC hashed page tables that act as extended TLBs).
3912 *
3913 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3914 * concurrent faults).
3915 *
3916 * The mmap_sem may have been released depending on flags and our return value.
3917 * See filemap_fault() and __lock_page_or_retry().
3918 */
3919static int handle_pte_fault(struct vm_fault *vmf)
3920{
3921 pte_t entry;
3922
3923 if (unlikely(pmd_none(*vmf->pmd))) {
3924 /*
3925 * Leave __pte_alloc() until later: because vm_ops->fault may
3926 * want to allocate huge page, and if we expose page table
3927 * for an instant, it will be difficult to retract from
3928 * concurrent faults and from rmap lookups.
3929 */
3930 vmf->pte = NULL;
3931 } else {
3932 /* See comment in pte_alloc_one_map() */
3933 if (pmd_devmap_trans_unstable(vmf->pmd))
3934 return 0;
3935 /*
3936 * A regular pmd is established and it can't morph into a huge
3937 * pmd from under us anymore at this point because we hold the
3938 * mmap_sem read mode and khugepaged takes it in write mode.
3939 * So now it's safe to run pte_offset_map().
3940 */
3941 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3942 vmf->orig_pte = *vmf->pte;
3943
3944 /*
3945 * some architectures can have larger ptes than wordsize,
3946 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3947 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3948 * accesses. The code below just needs a consistent view
3949 * for the ifs and we later double check anyway with the
3950 * ptl lock held. So here a barrier will do.
3951 */
3952 barrier();
3953 if (pte_none(vmf->orig_pte)) {
3954 pte_unmap(vmf->pte);
3955 vmf->pte = NULL;
3956 }
3957 }
3958
3959 if (!vmf->pte) {
3960 if (vma_is_anonymous(vmf->vma))
3961 return do_anonymous_page(vmf);
3962 else
3963 return do_fault(vmf);
3964 }
3965
3966 if (!pte_present(vmf->orig_pte))
3967 return do_swap_page(vmf);
3968
3969 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3970 return do_numa_page(vmf);
3971
3972 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3973 spin_lock(vmf->ptl);
3974 entry = vmf->orig_pte;
3975 if (unlikely(!pte_same(*vmf->pte, entry)))
3976 goto unlock;
3977 if (vmf->flags & FAULT_FLAG_WRITE) {
3978 if (!pte_write(entry))
3979 return do_wp_page(vmf);
3980 entry = pte_mkdirty(entry);
3981 }
3982 entry = pte_mkyoung(entry);
3983 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3984 vmf->flags & FAULT_FLAG_WRITE)) {
3985 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3986 } else {
3987 /*
3988 * This is needed only for protection faults but the arch code
3989 * is not yet telling us if this is a protection fault or not.
3990 * This still avoids useless tlb flushes for .text page faults
3991 * with threads.
3992 */
3993 if (vmf->flags & FAULT_FLAG_WRITE)
3994 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3995 }
3996unlock:
3997 pte_unmap_unlock(vmf->pte, vmf->ptl);
3998 return 0;
3999}
4000
4001/*
4002 * By the time we get here, we already hold the mm semaphore
4003 *
4004 * The mmap_sem may have been released depending on flags and our
4005 * return value. See filemap_fault() and __lock_page_or_retry().
4006 */
4007static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4008 unsigned int flags)
4009{
4010 struct vm_fault vmf = {
4011 .vma = vma,
4012 .address = address & PAGE_MASK,
4013 .flags = flags,
4014 .pgoff = linear_page_index(vma, address),
4015 .gfp_mask = __get_fault_gfp_mask(vma),
4016 };
4017 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4018 struct mm_struct *mm = vma->vm_mm;
4019 pgd_t *pgd;
4020 p4d_t *p4d;
4021 int ret;
4022
4023 pgd = pgd_offset(mm, address);
4024 p4d = p4d_alloc(mm, pgd, address);
4025 if (!p4d)
4026 return VM_FAULT_OOM;
4027
4028 vmf.pud = pud_alloc(mm, p4d, address);
4029 if (!vmf.pud)
4030 return VM_FAULT_OOM;
4031 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
4032 ret = create_huge_pud(&vmf);
4033 if (!(ret & VM_FAULT_FALLBACK))
4034 return ret;
4035 } else {
4036 pud_t orig_pud = *vmf.pud;
4037
4038 barrier();
4039 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4040
4041 /* NUMA case for anonymous PUDs would go here */
4042
4043 if (dirty && !pud_write(orig_pud)) {
4044 ret = wp_huge_pud(&vmf, orig_pud);
4045 if (!(ret & VM_FAULT_FALLBACK))
4046 return ret;
4047 } else {
4048 huge_pud_set_accessed(&vmf, orig_pud);
4049 return 0;
4050 }
4051 }
4052 }
4053
4054 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4055 if (!vmf.pmd)
4056 return VM_FAULT_OOM;
4057 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
4058 ret = create_huge_pmd(&vmf);
4059 if (!(ret & VM_FAULT_FALLBACK))
4060 return ret;
4061 } else {
4062 pmd_t orig_pmd = *vmf.pmd;
4063
4064 barrier();
4065 if (unlikely(is_swap_pmd(orig_pmd))) {
4066 VM_BUG_ON(thp_migration_supported() &&
4067 !is_pmd_migration_entry(orig_pmd));
4068 if (is_pmd_migration_entry(orig_pmd))
4069 pmd_migration_entry_wait(mm, vmf.pmd);
4070 return 0;
4071 }
4072 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4073 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4074 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4075
4076 if (dirty && !pmd_write(orig_pmd)) {
4077 ret = wp_huge_pmd(&vmf, orig_pmd);
4078 if (!(ret & VM_FAULT_FALLBACK))
4079 return ret;
4080 } else {
4081 huge_pmd_set_accessed(&vmf, orig_pmd);
4082 return 0;
4083 }
4084 }
4085 }
4086
4087 return handle_pte_fault(&vmf);
4088}
4089
4090/*
4091 * By the time we get here, we already hold the mm semaphore
4092 *
4093 * The mmap_sem may have been released depending on flags and our
4094 * return value. See filemap_fault() and __lock_page_or_retry().
4095 */
4096int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4097 unsigned int flags)
4098{
4099 int ret;
4100
4101 __set_current_state(TASK_RUNNING);
4102
4103 count_vm_event(PGFAULT);
4104 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4105
4106 /* do counter updates before entering really critical section. */
4107 check_sync_rss_stat(current);
4108
4109 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4110 flags & FAULT_FLAG_INSTRUCTION,
4111 flags & FAULT_FLAG_REMOTE))
4112 return VM_FAULT_SIGSEGV;
4113
4114 /*
4115 * Enable the memcg OOM handling for faults triggered in user
4116 * space. Kernel faults are handled more gracefully.
4117 */
4118 if (flags & FAULT_FLAG_USER)
4119 mem_cgroup_oom_enable();
4120
4121 if (unlikely(is_vm_hugetlb_page(vma)))
4122 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4123 else
4124 ret = __handle_mm_fault(vma, address, flags);
4125
4126 if (flags & FAULT_FLAG_USER) {
4127 mem_cgroup_oom_disable();
4128 /*
4129 * The task may have entered a memcg OOM situation but
4130 * if the allocation error was handled gracefully (no
4131 * VM_FAULT_OOM), there is no need to kill anything.
4132 * Just clean up the OOM state peacefully.
4133 */
4134 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4135 mem_cgroup_oom_synchronize(false);
4136 }
4137
4138 return ret;
4139}
4140EXPORT_SYMBOL_GPL(handle_mm_fault);
4141
4142#ifndef __PAGETABLE_P4D_FOLDED
4143/*
4144 * Allocate p4d page table.
4145 * We've already handled the fast-path in-line.
4146 */
4147int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4148{
4149 p4d_t *new = p4d_alloc_one(mm, address);
4150 if (!new)
4151 return -ENOMEM;
4152
4153 smp_wmb(); /* See comment in __pte_alloc */
4154
4155 spin_lock(&mm->page_table_lock);
4156 if (pgd_present(*pgd)) /* Another has populated it */
4157 p4d_free(mm, new);
4158 else
4159 pgd_populate(mm, pgd, new);
4160 spin_unlock(&mm->page_table_lock);
4161 return 0;
4162}
4163#endif /* __PAGETABLE_P4D_FOLDED */
4164
4165#ifndef __PAGETABLE_PUD_FOLDED
4166/*
4167 * Allocate page upper directory.
4168 * We've already handled the fast-path in-line.
4169 */
4170int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4171{
4172 pud_t *new = pud_alloc_one(mm, address);
4173 if (!new)
4174 return -ENOMEM;
4175
4176 smp_wmb(); /* See comment in __pte_alloc */
4177
4178 spin_lock(&mm->page_table_lock);
4179#ifndef __ARCH_HAS_5LEVEL_HACK
4180 if (!p4d_present(*p4d)) {
4181 mm_inc_nr_puds(mm);
4182 p4d_populate(mm, p4d, new);
4183 } else /* Another has populated it */
4184 pud_free(mm, new);
4185#else
4186 if (!pgd_present(*p4d)) {
4187 mm_inc_nr_puds(mm);
4188 pgd_populate(mm, p4d, new);
4189 } else /* Another has populated it */
4190 pud_free(mm, new);
4191#endif /* __ARCH_HAS_5LEVEL_HACK */
4192 spin_unlock(&mm->page_table_lock);
4193 return 0;
4194}
4195#endif /* __PAGETABLE_PUD_FOLDED */
4196
4197#ifndef __PAGETABLE_PMD_FOLDED
4198/*
4199 * Allocate page middle directory.
4200 * We've already handled the fast-path in-line.
4201 */
4202int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4203{
4204 spinlock_t *ptl;
4205 pmd_t *new = pmd_alloc_one(mm, address);
4206 if (!new)
4207 return -ENOMEM;
4208
4209 smp_wmb(); /* See comment in __pte_alloc */
4210
4211 ptl = pud_lock(mm, pud);
4212#ifndef __ARCH_HAS_4LEVEL_HACK
4213 if (!pud_present(*pud)) {
4214 mm_inc_nr_pmds(mm);
4215 pud_populate(mm, pud, new);
4216 } else /* Another has populated it */
4217 pmd_free(mm, new);
4218#else
4219 if (!pgd_present(*pud)) {
4220 mm_inc_nr_pmds(mm);
4221 pgd_populate(mm, pud, new);
4222 } else /* Another has populated it */
4223 pmd_free(mm, new);
4224#endif /* __ARCH_HAS_4LEVEL_HACK */
4225 spin_unlock(ptl);
4226 return 0;
4227}
4228#endif /* __PAGETABLE_PMD_FOLDED */
4229
4230static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4231 unsigned long *start, unsigned long *end,
4232 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4233{
4234 pgd_t *pgd;
4235 p4d_t *p4d;
4236 pud_t *pud;
4237 pmd_t *pmd;
4238 pte_t *ptep;
4239
4240 pgd = pgd_offset(mm, address);
4241 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4242 goto out;
4243
4244 p4d = p4d_offset(pgd, address);
4245 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4246 goto out;
4247
4248 pud = pud_offset(p4d, address);
4249 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4250 goto out;
4251
4252 pmd = pmd_offset(pud, address);
4253 VM_BUG_ON(pmd_trans_huge(*pmd));
4254
4255 if (pmd_huge(*pmd)) {
4256 if (!pmdpp)
4257 goto out;
4258
4259 if (start && end) {
4260 *start = address & PMD_MASK;
4261 *end = *start + PMD_SIZE;
4262 mmu_notifier_invalidate_range_start(mm, *start, *end);
4263 }
4264 *ptlp = pmd_lock(mm, pmd);
4265 if (pmd_huge(*pmd)) {
4266 *pmdpp = pmd;
4267 return 0;
4268 }
4269 spin_unlock(*ptlp);
4270 if (start && end)
4271 mmu_notifier_invalidate_range_end(mm, *start, *end);
4272 }
4273
4274 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4275 goto out;
4276
4277 if (start && end) {
4278 *start = address & PAGE_MASK;
4279 *end = *start + PAGE_SIZE;
4280 mmu_notifier_invalidate_range_start(mm, *start, *end);
4281 }
4282 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4283 if (!pte_present(*ptep))
4284 goto unlock;
4285 *ptepp = ptep;
4286 return 0;
4287unlock:
4288 pte_unmap_unlock(ptep, *ptlp);
4289 if (start && end)
4290 mmu_notifier_invalidate_range_end(mm, *start, *end);
4291out:
4292 return -EINVAL;
4293}
4294
4295static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4296 pte_t **ptepp, spinlock_t **ptlp)
4297{
4298 int res;
4299
4300 /* (void) is needed to make gcc happy */
4301 (void) __cond_lock(*ptlp,
4302 !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4303 ptepp, NULL, ptlp)));
4304 return res;
4305}
4306
4307int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4308 unsigned long *start, unsigned long *end,
4309 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4310{
4311 int res;
4312
4313 /* (void) is needed to make gcc happy */
4314 (void) __cond_lock(*ptlp,
4315 !(res = __follow_pte_pmd(mm, address, start, end,
4316 ptepp, pmdpp, ptlp)));
4317 return res;
4318}
4319EXPORT_SYMBOL(follow_pte_pmd);
4320
4321/**
4322 * follow_pfn - look up PFN at a user virtual address
4323 * @vma: memory mapping
4324 * @address: user virtual address
4325 * @pfn: location to store found PFN
4326 *
4327 * Only IO mappings and raw PFN mappings are allowed.
4328 *
4329 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4330 */
4331int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4332 unsigned long *pfn)
4333{
4334 int ret = -EINVAL;
4335 spinlock_t *ptl;
4336 pte_t *ptep;
4337
4338 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4339 return ret;
4340
4341 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4342 if (ret)
4343 return ret;
4344 *pfn = pte_pfn(*ptep);
4345 pte_unmap_unlock(ptep, ptl);
4346 return 0;
4347}
4348EXPORT_SYMBOL(follow_pfn);
4349
4350#ifdef CONFIG_HAVE_IOREMAP_PROT
4351int follow_phys(struct vm_area_struct *vma,
4352 unsigned long address, unsigned int flags,
4353 unsigned long *prot, resource_size_t *phys)
4354{
4355 int ret = -EINVAL;
4356 pte_t *ptep, pte;
4357 spinlock_t *ptl;
4358
4359 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4360 goto out;
4361
4362 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4363 goto out;
4364 pte = *ptep;
4365
4366 if ((flags & FOLL_WRITE) && !pte_write(pte))
4367 goto unlock;
4368
4369 *prot = pgprot_val(pte_pgprot(pte));
4370 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4371
4372 ret = 0;
4373unlock:
4374 pte_unmap_unlock(ptep, ptl);
4375out:
4376 return ret;
4377}
4378
4379int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4380 void *buf, int len, int write)
4381{
4382 resource_size_t phys_addr;
4383 unsigned long prot = 0;
4384 void __iomem *maddr;
4385 int offset = addr & (PAGE_SIZE-1);
4386
4387 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4388 return -EINVAL;
4389
4390 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4391 if (write)
4392 memcpy_toio(maddr + offset, buf, len);
4393 else
4394 memcpy_fromio(buf, maddr + offset, len);
4395 iounmap(maddr);
4396
4397 return len;
4398}
4399EXPORT_SYMBOL_GPL(generic_access_phys);
4400#endif
4401
4402/*
4403 * Access another process' address space as given in mm. If non-NULL, use the
4404 * given task for page fault accounting.
4405 */
4406int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4407 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4408{
4409 struct vm_area_struct *vma;
4410 void *old_buf = buf;
4411 int write = gup_flags & FOLL_WRITE;
4412
4413 down_read(&mm->mmap_sem);
4414 /* ignore errors, just check how much was successfully transferred */
4415 while (len) {
4416 int bytes, ret, offset;
4417 void *maddr;
4418 struct page *page = NULL;
4419
4420 ret = get_user_pages_remote(tsk, mm, addr, 1,
4421 gup_flags, &page, &vma, NULL);
4422 if (ret <= 0) {
4423#ifndef CONFIG_HAVE_IOREMAP_PROT
4424 break;
4425#else
4426 /*
4427 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4428 * we can access using slightly different code.
4429 */
4430 vma = find_vma(mm, addr);
4431 if (!vma || vma->vm_start > addr)
4432 break;
4433 if (vma->vm_ops && vma->vm_ops->access)
4434 ret = vma->vm_ops->access(vma, addr, buf,
4435 len, write);
4436 if (ret <= 0)
4437 break;
4438 bytes = ret;
4439#endif
4440 } else {
4441 bytes = len;
4442 offset = addr & (PAGE_SIZE-1);
4443 if (bytes > PAGE_SIZE-offset)
4444 bytes = PAGE_SIZE-offset;
4445
4446 maddr = kmap(page);
4447 if (write) {
4448 copy_to_user_page(vma, page, addr,
4449 maddr + offset, buf, bytes);
4450 set_page_dirty_lock(page);
4451 } else {
4452 copy_from_user_page(vma, page, addr,
4453 buf, maddr + offset, bytes);
4454 }
4455 kunmap(page);
4456 put_page(page);
4457 }
4458 len -= bytes;
4459 buf += bytes;
4460 addr += bytes;
4461 }
4462 up_read(&mm->mmap_sem);
4463
4464 return buf - old_buf;
4465}
4466
4467/**
4468 * access_remote_vm - access another process' address space
4469 * @mm: the mm_struct of the target address space
4470 * @addr: start address to access
4471 * @buf: source or destination buffer
4472 * @len: number of bytes to transfer
4473 * @gup_flags: flags modifying lookup behaviour
4474 *
4475 * The caller must hold a reference on @mm.
4476 */
4477int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4478 void *buf, int len, unsigned int gup_flags)
4479{
4480 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4481}
4482
4483/*
4484 * Access another process' address space.
4485 * Source/target buffer must be kernel space,
4486 * Do not walk the page table directly, use get_user_pages
4487 */
4488int access_process_vm(struct task_struct *tsk, unsigned long addr,
4489 void *buf, int len, unsigned int gup_flags)
4490{
4491 struct mm_struct *mm;
4492 int ret;
4493
4494 mm = get_task_mm(tsk);
4495 if (!mm)
4496 return 0;
4497
4498 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4499
4500 mmput(mm);
4501
4502 return ret;
4503}
4504EXPORT_SYMBOL_GPL(access_process_vm);
4505
4506/*
4507 * Print the name of a VMA.
4508 */
4509void print_vma_addr(char *prefix, unsigned long ip)
4510{
4511 struct mm_struct *mm = current->mm;
4512 struct vm_area_struct *vma;
4513
4514 /*
4515 * we might be running from an atomic context so we cannot sleep
4516 */
4517 if (!down_read_trylock(&mm->mmap_sem))
4518 return;
4519
4520 vma = find_vma(mm, ip);
4521 if (vma && vma->vm_file) {
4522 struct file *f = vma->vm_file;
4523 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4524 if (buf) {
4525 char *p;
4526
4527 p = file_path(f, buf, PAGE_SIZE);
4528 if (IS_ERR(p))
4529 p = "?";
4530 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4531 vma->vm_start,
4532 vma->vm_end - vma->vm_start);
4533 free_page((unsigned long)buf);
4534 }
4535 }
4536 up_read(&mm->mmap_sem);
4537}
4538
4539#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4540void __might_fault(const char *file, int line)
4541{
4542 /*
4543 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4544 * holding the mmap_sem, this is safe because kernel memory doesn't
4545 * get paged out, therefore we'll never actually fault, and the
4546 * below annotations will generate false positives.
4547 */
4548 if (uaccess_kernel())
4549 return;
4550 if (pagefault_disabled())
4551 return;
4552 __might_sleep(file, line, 0);
4553#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4554 if (current->mm)
4555 might_lock_read(¤t->mm->mmap_sem);
4556#endif
4557}
4558EXPORT_SYMBOL(__might_fault);
4559#endif
4560
4561#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4562static void clear_gigantic_page(struct page *page,
4563 unsigned long addr,
4564 unsigned int pages_per_huge_page)
4565{
4566 int i;
4567 struct page *p = page;
4568
4569 might_sleep();
4570 for (i = 0; i < pages_per_huge_page;
4571 i++, p = mem_map_next(p, page, i)) {
4572 cond_resched();
4573 clear_user_highpage(p, addr + i * PAGE_SIZE);
4574 }
4575}
4576void clear_huge_page(struct page *page,
4577 unsigned long addr_hint, unsigned int pages_per_huge_page)
4578{
4579 int i, n, base, l;
4580 unsigned long addr = addr_hint &
4581 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4582
4583 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4584 clear_gigantic_page(page, addr, pages_per_huge_page);
4585 return;
4586 }
4587
4588 /* Clear sub-page to access last to keep its cache lines hot */
4589 might_sleep();
4590 n = (addr_hint - addr) / PAGE_SIZE;
4591 if (2 * n <= pages_per_huge_page) {
4592 /* If sub-page to access in first half of huge page */
4593 base = 0;
4594 l = n;
4595 /* Clear sub-pages at the end of huge page */
4596 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4597 cond_resched();
4598 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4599 }
4600 } else {
4601 /* If sub-page to access in second half of huge page */
4602 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4603 l = pages_per_huge_page - n;
4604 /* Clear sub-pages at the begin of huge page */
4605 for (i = 0; i < base; i++) {
4606 cond_resched();
4607 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4608 }
4609 }
4610 /*
4611 * Clear remaining sub-pages in left-right-left-right pattern
4612 * towards the sub-page to access
4613 */
4614 for (i = 0; i < l; i++) {
4615 int left_idx = base + i;
4616 int right_idx = base + 2 * l - 1 - i;
4617
4618 cond_resched();
4619 clear_user_highpage(page + left_idx,
4620 addr + left_idx * PAGE_SIZE);
4621 cond_resched();
4622 clear_user_highpage(page + right_idx,
4623 addr + right_idx * PAGE_SIZE);
4624 }
4625}
4626
4627static void copy_user_gigantic_page(struct page *dst, struct page *src,
4628 unsigned long addr,
4629 struct vm_area_struct *vma,
4630 unsigned int pages_per_huge_page)
4631{
4632 int i;
4633 struct page *dst_base = dst;
4634 struct page *src_base = src;
4635
4636 for (i = 0; i < pages_per_huge_page; ) {
4637 cond_resched();
4638 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4639
4640 i++;
4641 dst = mem_map_next(dst, dst_base, i);
4642 src = mem_map_next(src, src_base, i);
4643 }
4644}
4645
4646void copy_user_huge_page(struct page *dst, struct page *src,
4647 unsigned long addr, struct vm_area_struct *vma,
4648 unsigned int pages_per_huge_page)
4649{
4650 int i;
4651
4652 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4653 copy_user_gigantic_page(dst, src, addr, vma,
4654 pages_per_huge_page);
4655 return;
4656 }
4657
4658 might_sleep();
4659 for (i = 0; i < pages_per_huge_page; i++) {
4660 cond_resched();
4661 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4662 }
4663}
4664
4665long copy_huge_page_from_user(struct page *dst_page,
4666 const void __user *usr_src,
4667 unsigned int pages_per_huge_page,
4668 bool allow_pagefault)
4669{
4670 void *src = (void *)usr_src;
4671 void *page_kaddr;
4672 unsigned long i, rc = 0;
4673 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4674
4675 for (i = 0; i < pages_per_huge_page; i++) {
4676 if (allow_pagefault)
4677 page_kaddr = kmap(dst_page + i);
4678 else
4679 page_kaddr = kmap_atomic(dst_page + i);
4680 rc = copy_from_user(page_kaddr,
4681 (const void __user *)(src + i * PAGE_SIZE),
4682 PAGE_SIZE);
4683 if (allow_pagefault)
4684 kunmap(dst_page + i);
4685 else
4686 kunmap_atomic(page_kaddr);
4687
4688 ret_val -= (PAGE_SIZE - rc);
4689 if (rc)
4690 break;
4691
4692 cond_resched();
4693 }
4694 return ret_val;
4695}
4696#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4697
4698#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4699
4700static struct kmem_cache *page_ptl_cachep;
4701
4702void __init ptlock_cache_init(void)
4703{
4704 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4705 SLAB_PANIC, NULL);
4706}
4707
4708bool ptlock_alloc(struct page *page)
4709{
4710 spinlock_t *ptl;
4711
4712 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4713 if (!ptl)
4714 return false;
4715 page->ptl = ptl;
4716 return true;
4717}
4718
4719void ptlock_free(struct page *page)
4720{
4721 kmem_cache_free(page_ptl_cachep, page->ptl);
4722}
4723#endif
1
2// SPDX-License-Identifier: GPL-2.0-only
3/*
4 * linux/mm/memory.c
5 *
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 */
8
9/*
10 * demand-loading started 01.12.91 - seems it is high on the list of
11 * things wanted, and it should be easy to implement. - Linus
12 */
13
14/*
15 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
16 * pages started 02.12.91, seems to work. - Linus.
17 *
18 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
19 * would have taken more than the 6M I have free, but it worked well as
20 * far as I could see.
21 *
22 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
23 */
24
25/*
26 * Real VM (paging to/from disk) started 18.12.91. Much more work and
27 * thought has to go into this. Oh, well..
28 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
29 * Found it. Everything seems to work now.
30 * 20.12.91 - Ok, making the swap-device changeable like the root.
31 */
32
33/*
34 * 05.04.94 - Multi-page memory management added for v1.1.
35 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 *
37 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
38 * (Gerhard.Wichert@pdb.siemens.de)
39 *
40 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 */
42
43#include <linux/kernel_stat.h>
44#include <linux/mm.h>
45#include <linux/mm_inline.h>
46#include <linux/sched/mm.h>
47#include <linux/sched/coredump.h>
48#include <linux/sched/numa_balancing.h>
49#include <linux/sched/task.h>
50#include <linux/hugetlb.h>
51#include <linux/mman.h>
52#include <linux/swap.h>
53#include <linux/highmem.h>
54#include <linux/pagemap.h>
55#include <linux/memremap.h>
56#include <linux/kmsan.h>
57#include <linux/ksm.h>
58#include <linux/rmap.h>
59#include <linux/export.h>
60#include <linux/delayacct.h>
61#include <linux/init.h>
62#include <linux/pfn_t.h>
63#include <linux/writeback.h>
64#include <linux/memcontrol.h>
65#include <linux/mmu_notifier.h>
66#include <linux/swapops.h>
67#include <linux/elf.h>
68#include <linux/gfp.h>
69#include <linux/migrate.h>
70#include <linux/string.h>
71#include <linux/memory-tiers.h>
72#include <linux/debugfs.h>
73#include <linux/userfaultfd_k.h>
74#include <linux/dax.h>
75#include <linux/oom.h>
76#include <linux/numa.h>
77#include <linux/perf_event.h>
78#include <linux/ptrace.h>
79#include <linux/vmalloc.h>
80#include <linux/sched/sysctl.h>
81
82#include <trace/events/kmem.h>
83
84#include <asm/io.h>
85#include <asm/mmu_context.h>
86#include <asm/pgalloc.h>
87#include <linux/uaccess.h>
88#include <asm/tlb.h>
89#include <asm/tlbflush.h>
90
91#include "pgalloc-track.h"
92#include "internal.h"
93#include "swap.h"
94
95#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
96#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
97#endif
98
99#ifndef CONFIG_NUMA
100unsigned long max_mapnr;
101EXPORT_SYMBOL(max_mapnr);
102
103struct page *mem_map;
104EXPORT_SYMBOL(mem_map);
105#endif
106
107static vm_fault_t do_fault(struct vm_fault *vmf);
108static vm_fault_t do_anonymous_page(struct vm_fault *vmf);
109static bool vmf_pte_changed(struct vm_fault *vmf);
110
111/*
112 * Return true if the original pte was a uffd-wp pte marker (so the pte was
113 * wr-protected).
114 */
115static bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
116{
117 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
118 return false;
119
120 return pte_marker_uffd_wp(vmf->orig_pte);
121}
122
123/*
124 * A number of key systems in x86 including ioremap() rely on the assumption
125 * that high_memory defines the upper bound on direct map memory, then end
126 * of ZONE_NORMAL.
127 */
128void *high_memory;
129EXPORT_SYMBOL(high_memory);
130
131/*
132 * Randomize the address space (stacks, mmaps, brk, etc.).
133 *
134 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
135 * as ancient (libc5 based) binaries can segfault. )
136 */
137int randomize_va_space __read_mostly =
138#ifdef CONFIG_COMPAT_BRK
139 1;
140#else
141 2;
142#endif
143
144#ifndef arch_wants_old_prefaulted_pte
145static inline bool arch_wants_old_prefaulted_pte(void)
146{
147 /*
148 * Transitioning a PTE from 'old' to 'young' can be expensive on
149 * some architectures, even if it's performed in hardware. By
150 * default, "false" means prefaulted entries will be 'young'.
151 */
152 return false;
153}
154#endif
155
156static int __init disable_randmaps(char *s)
157{
158 randomize_va_space = 0;
159 return 1;
160}
161__setup("norandmaps", disable_randmaps);
162
163unsigned long zero_pfn __read_mostly;
164EXPORT_SYMBOL(zero_pfn);
165
166unsigned long highest_memmap_pfn __read_mostly;
167
168/*
169 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
170 */
171static int __init init_zero_pfn(void)
172{
173 zero_pfn = page_to_pfn(ZERO_PAGE(0));
174 return 0;
175}
176early_initcall(init_zero_pfn);
177
178void mm_trace_rss_stat(struct mm_struct *mm, int member)
179{
180 trace_rss_stat(mm, member);
181}
182
183/*
184 * Note: this doesn't free the actual pages themselves. That
185 * has been handled earlier when unmapping all the memory regions.
186 */
187static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
188 unsigned long addr)
189{
190 pgtable_t token = pmd_pgtable(*pmd);
191 pmd_clear(pmd);
192 pte_free_tlb(tlb, token, addr);
193 mm_dec_nr_ptes(tlb->mm);
194}
195
196static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
197 unsigned long addr, unsigned long end,
198 unsigned long floor, unsigned long ceiling)
199{
200 pmd_t *pmd;
201 unsigned long next;
202 unsigned long start;
203
204 start = addr;
205 pmd = pmd_offset(pud, addr);
206 do {
207 next = pmd_addr_end(addr, end);
208 if (pmd_none_or_clear_bad(pmd))
209 continue;
210 free_pte_range(tlb, pmd, addr);
211 } while (pmd++, addr = next, addr != end);
212
213 start &= PUD_MASK;
214 if (start < floor)
215 return;
216 if (ceiling) {
217 ceiling &= PUD_MASK;
218 if (!ceiling)
219 return;
220 }
221 if (end - 1 > ceiling - 1)
222 return;
223
224 pmd = pmd_offset(pud, start);
225 pud_clear(pud);
226 pmd_free_tlb(tlb, pmd, start);
227 mm_dec_nr_pmds(tlb->mm);
228}
229
230static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
231 unsigned long addr, unsigned long end,
232 unsigned long floor, unsigned long ceiling)
233{
234 pud_t *pud;
235 unsigned long next;
236 unsigned long start;
237
238 start = addr;
239 pud = pud_offset(p4d, addr);
240 do {
241 next = pud_addr_end(addr, end);
242 if (pud_none_or_clear_bad(pud))
243 continue;
244 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
245 } while (pud++, addr = next, addr != end);
246
247 start &= P4D_MASK;
248 if (start < floor)
249 return;
250 if (ceiling) {
251 ceiling &= P4D_MASK;
252 if (!ceiling)
253 return;
254 }
255 if (end - 1 > ceiling - 1)
256 return;
257
258 pud = pud_offset(p4d, start);
259 p4d_clear(p4d);
260 pud_free_tlb(tlb, pud, start);
261 mm_dec_nr_puds(tlb->mm);
262}
263
264static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
265 unsigned long addr, unsigned long end,
266 unsigned long floor, unsigned long ceiling)
267{
268 p4d_t *p4d;
269 unsigned long next;
270 unsigned long start;
271
272 start = addr;
273 p4d = p4d_offset(pgd, addr);
274 do {
275 next = p4d_addr_end(addr, end);
276 if (p4d_none_or_clear_bad(p4d))
277 continue;
278 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
279 } while (p4d++, addr = next, addr != end);
280
281 start &= PGDIR_MASK;
282 if (start < floor)
283 return;
284 if (ceiling) {
285 ceiling &= PGDIR_MASK;
286 if (!ceiling)
287 return;
288 }
289 if (end - 1 > ceiling - 1)
290 return;
291
292 p4d = p4d_offset(pgd, start);
293 pgd_clear(pgd);
294 p4d_free_tlb(tlb, p4d, start);
295}
296
297/*
298 * This function frees user-level page tables of a process.
299 */
300void free_pgd_range(struct mmu_gather *tlb,
301 unsigned long addr, unsigned long end,
302 unsigned long floor, unsigned long ceiling)
303{
304 pgd_t *pgd;
305 unsigned long next;
306
307 /*
308 * The next few lines have given us lots of grief...
309 *
310 * Why are we testing PMD* at this top level? Because often
311 * there will be no work to do at all, and we'd prefer not to
312 * go all the way down to the bottom just to discover that.
313 *
314 * Why all these "- 1"s? Because 0 represents both the bottom
315 * of the address space and the top of it (using -1 for the
316 * top wouldn't help much: the masks would do the wrong thing).
317 * The rule is that addr 0 and floor 0 refer to the bottom of
318 * the address space, but end 0 and ceiling 0 refer to the top
319 * Comparisons need to use "end - 1" and "ceiling - 1" (though
320 * that end 0 case should be mythical).
321 *
322 * Wherever addr is brought up or ceiling brought down, we must
323 * be careful to reject "the opposite 0" before it confuses the
324 * subsequent tests. But what about where end is brought down
325 * by PMD_SIZE below? no, end can't go down to 0 there.
326 *
327 * Whereas we round start (addr) and ceiling down, by different
328 * masks at different levels, in order to test whether a table
329 * now has no other vmas using it, so can be freed, we don't
330 * bother to round floor or end up - the tests don't need that.
331 */
332
333 addr &= PMD_MASK;
334 if (addr < floor) {
335 addr += PMD_SIZE;
336 if (!addr)
337 return;
338 }
339 if (ceiling) {
340 ceiling &= PMD_MASK;
341 if (!ceiling)
342 return;
343 }
344 if (end - 1 > ceiling - 1)
345 end -= PMD_SIZE;
346 if (addr > end - 1)
347 return;
348 /*
349 * We add page table cache pages with PAGE_SIZE,
350 * (see pte_free_tlb()), flush the tlb if we need
351 */
352 tlb_change_page_size(tlb, PAGE_SIZE);
353 pgd = pgd_offset(tlb->mm, addr);
354 do {
355 next = pgd_addr_end(addr, end);
356 if (pgd_none_or_clear_bad(pgd))
357 continue;
358 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
359 } while (pgd++, addr = next, addr != end);
360}
361
362void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
363 struct vm_area_struct *vma, unsigned long floor,
364 unsigned long ceiling, bool mm_wr_locked)
365{
366 do {
367 unsigned long addr = vma->vm_start;
368 struct vm_area_struct *next;
369
370 /*
371 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
372 * be 0. This will underflow and is okay.
373 */
374 next = mas_find(mas, ceiling - 1);
375 if (unlikely(xa_is_zero(next)))
376 next = NULL;
377
378 /*
379 * Hide vma from rmap and truncate_pagecache before freeing
380 * pgtables
381 */
382 if (mm_wr_locked)
383 vma_start_write(vma);
384 unlink_anon_vmas(vma);
385 unlink_file_vma(vma);
386
387 if (is_vm_hugetlb_page(vma)) {
388 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
389 floor, next ? next->vm_start : ceiling);
390 } else {
391 /*
392 * Optimization: gather nearby vmas into one call down
393 */
394 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
395 && !is_vm_hugetlb_page(next)) {
396 vma = next;
397 next = mas_find(mas, ceiling - 1);
398 if (unlikely(xa_is_zero(next)))
399 next = NULL;
400 if (mm_wr_locked)
401 vma_start_write(vma);
402 unlink_anon_vmas(vma);
403 unlink_file_vma(vma);
404 }
405 free_pgd_range(tlb, addr, vma->vm_end,
406 floor, next ? next->vm_start : ceiling);
407 }
408 vma = next;
409 } while (vma);
410}
411
412void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
413{
414 spinlock_t *ptl = pmd_lock(mm, pmd);
415
416 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
417 mm_inc_nr_ptes(mm);
418 /*
419 * Ensure all pte setup (eg. pte page lock and page clearing) are
420 * visible before the pte is made visible to other CPUs by being
421 * put into page tables.
422 *
423 * The other side of the story is the pointer chasing in the page
424 * table walking code (when walking the page table without locking;
425 * ie. most of the time). Fortunately, these data accesses consist
426 * of a chain of data-dependent loads, meaning most CPUs (alpha
427 * being the notable exception) will already guarantee loads are
428 * seen in-order. See the alpha page table accessors for the
429 * smp_rmb() barriers in page table walking code.
430 */
431 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
432 pmd_populate(mm, pmd, *pte);
433 *pte = NULL;
434 }
435 spin_unlock(ptl);
436}
437
438int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
439{
440 pgtable_t new = pte_alloc_one(mm);
441 if (!new)
442 return -ENOMEM;
443
444 pmd_install(mm, pmd, &new);
445 if (new)
446 pte_free(mm, new);
447 return 0;
448}
449
450int __pte_alloc_kernel(pmd_t *pmd)
451{
452 pte_t *new = pte_alloc_one_kernel(&init_mm);
453 if (!new)
454 return -ENOMEM;
455
456 spin_lock(&init_mm.page_table_lock);
457 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
458 smp_wmb(); /* See comment in pmd_install() */
459 pmd_populate_kernel(&init_mm, pmd, new);
460 new = NULL;
461 }
462 spin_unlock(&init_mm.page_table_lock);
463 if (new)
464 pte_free_kernel(&init_mm, new);
465 return 0;
466}
467
468static inline void init_rss_vec(int *rss)
469{
470 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
471}
472
473static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
474{
475 int i;
476
477 for (i = 0; i < NR_MM_COUNTERS; i++)
478 if (rss[i])
479 add_mm_counter(mm, i, rss[i]);
480}
481
482/*
483 * This function is called to print an error when a bad pte
484 * is found. For example, we might have a PFN-mapped pte in
485 * a region that doesn't allow it.
486 *
487 * The calling function must still handle the error.
488 */
489static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
490 pte_t pte, struct page *page)
491{
492 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
493 p4d_t *p4d = p4d_offset(pgd, addr);
494 pud_t *pud = pud_offset(p4d, addr);
495 pmd_t *pmd = pmd_offset(pud, addr);
496 struct address_space *mapping;
497 pgoff_t index;
498 static unsigned long resume;
499 static unsigned long nr_shown;
500 static unsigned long nr_unshown;
501
502 /*
503 * Allow a burst of 60 reports, then keep quiet for that minute;
504 * or allow a steady drip of one report per second.
505 */
506 if (nr_shown == 60) {
507 if (time_before(jiffies, resume)) {
508 nr_unshown++;
509 return;
510 }
511 if (nr_unshown) {
512 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
513 nr_unshown);
514 nr_unshown = 0;
515 }
516 nr_shown = 0;
517 }
518 if (nr_shown++ == 0)
519 resume = jiffies + 60 * HZ;
520
521 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
522 index = linear_page_index(vma, addr);
523
524 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
525 current->comm,
526 (long long)pte_val(pte), (long long)pmd_val(*pmd));
527 if (page)
528 dump_page(page, "bad pte");
529 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
530 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
531 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
532 vma->vm_file,
533 vma->vm_ops ? vma->vm_ops->fault : NULL,
534 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
535 mapping ? mapping->a_ops->read_folio : NULL);
536 dump_stack();
537 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
538}
539
540/*
541 * vm_normal_page -- This function gets the "struct page" associated with a pte.
542 *
543 * "Special" mappings do not wish to be associated with a "struct page" (either
544 * it doesn't exist, or it exists but they don't want to touch it). In this
545 * case, NULL is returned here. "Normal" mappings do have a struct page.
546 *
547 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
548 * pte bit, in which case this function is trivial. Secondly, an architecture
549 * may not have a spare pte bit, which requires a more complicated scheme,
550 * described below.
551 *
552 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
553 * special mapping (even if there are underlying and valid "struct pages").
554 * COWed pages of a VM_PFNMAP are always normal.
555 *
556 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
557 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
558 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
559 * mapping will always honor the rule
560 *
561 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
562 *
563 * And for normal mappings this is false.
564 *
565 * This restricts such mappings to be a linear translation from virtual address
566 * to pfn. To get around this restriction, we allow arbitrary mappings so long
567 * as the vma is not a COW mapping; in that case, we know that all ptes are
568 * special (because none can have been COWed).
569 *
570 *
571 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
572 *
573 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
574 * page" backing, however the difference is that _all_ pages with a struct
575 * page (that is, those where pfn_valid is true) are refcounted and considered
576 * normal pages by the VM. The disadvantage is that pages are refcounted
577 * (which can be slower and simply not an option for some PFNMAP users). The
578 * advantage is that we don't have to follow the strict linearity rule of
579 * PFNMAP mappings in order to support COWable mappings.
580 *
581 */
582struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
583 pte_t pte)
584{
585 unsigned long pfn = pte_pfn(pte);
586
587 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
588 if (likely(!pte_special(pte)))
589 goto check_pfn;
590 if (vma->vm_ops && vma->vm_ops->find_special_page)
591 return vma->vm_ops->find_special_page(vma, addr);
592 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
593 return NULL;
594 if (is_zero_pfn(pfn))
595 return NULL;
596 if (pte_devmap(pte))
597 /*
598 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
599 * and will have refcounts incremented on their struct pages
600 * when they are inserted into PTEs, thus they are safe to
601 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
602 * do not have refcounts. Example of legacy ZONE_DEVICE is
603 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
604 */
605 return NULL;
606
607 print_bad_pte(vma, addr, pte, NULL);
608 return NULL;
609 }
610
611 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
612
613 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
614 if (vma->vm_flags & VM_MIXEDMAP) {
615 if (!pfn_valid(pfn))
616 return NULL;
617 goto out;
618 } else {
619 unsigned long off;
620 off = (addr - vma->vm_start) >> PAGE_SHIFT;
621 if (pfn == vma->vm_pgoff + off)
622 return NULL;
623 if (!is_cow_mapping(vma->vm_flags))
624 return NULL;
625 }
626 }
627
628 if (is_zero_pfn(pfn))
629 return NULL;
630
631check_pfn:
632 if (unlikely(pfn > highest_memmap_pfn)) {
633 print_bad_pte(vma, addr, pte, NULL);
634 return NULL;
635 }
636
637 /*
638 * NOTE! We still have PageReserved() pages in the page tables.
639 * eg. VDSO mappings can cause them to exist.
640 */
641out:
642 return pfn_to_page(pfn);
643}
644
645struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr,
646 pte_t pte)
647{
648 struct page *page = vm_normal_page(vma, addr, pte);
649
650 if (page)
651 return page_folio(page);
652 return NULL;
653}
654
655#ifdef CONFIG_TRANSPARENT_HUGEPAGE
656struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
657 pmd_t pmd)
658{
659 unsigned long pfn = pmd_pfn(pmd);
660
661 /*
662 * There is no pmd_special() but there may be special pmds, e.g.
663 * in a direct-access (dax) mapping, so let's just replicate the
664 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
665 */
666 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
667 if (vma->vm_flags & VM_MIXEDMAP) {
668 if (!pfn_valid(pfn))
669 return NULL;
670 goto out;
671 } else {
672 unsigned long off;
673 off = (addr - vma->vm_start) >> PAGE_SHIFT;
674 if (pfn == vma->vm_pgoff + off)
675 return NULL;
676 if (!is_cow_mapping(vma->vm_flags))
677 return NULL;
678 }
679 }
680
681 if (pmd_devmap(pmd))
682 return NULL;
683 if (is_huge_zero_pmd(pmd))
684 return NULL;
685 if (unlikely(pfn > highest_memmap_pfn))
686 return NULL;
687
688 /*
689 * NOTE! We still have PageReserved() pages in the page tables.
690 * eg. VDSO mappings can cause them to exist.
691 */
692out:
693 return pfn_to_page(pfn);
694}
695
696struct folio *vm_normal_folio_pmd(struct vm_area_struct *vma,
697 unsigned long addr, pmd_t pmd)
698{
699 struct page *page = vm_normal_page_pmd(vma, addr, pmd);
700
701 if (page)
702 return page_folio(page);
703 return NULL;
704}
705#endif
706
707static void restore_exclusive_pte(struct vm_area_struct *vma,
708 struct page *page, unsigned long address,
709 pte_t *ptep)
710{
711 struct folio *folio = page_folio(page);
712 pte_t orig_pte;
713 pte_t pte;
714 swp_entry_t entry;
715
716 orig_pte = ptep_get(ptep);
717 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
718 if (pte_swp_soft_dirty(orig_pte))
719 pte = pte_mksoft_dirty(pte);
720
721 entry = pte_to_swp_entry(orig_pte);
722 if (pte_swp_uffd_wp(orig_pte))
723 pte = pte_mkuffd_wp(pte);
724 else if (is_writable_device_exclusive_entry(entry))
725 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
726
727 VM_BUG_ON_FOLIO(pte_write(pte) && (!folio_test_anon(folio) &&
728 PageAnonExclusive(page)), folio);
729
730 /*
731 * No need to take a page reference as one was already
732 * created when the swap entry was made.
733 */
734 if (folio_test_anon(folio))
735 folio_add_anon_rmap_pte(folio, page, vma, address, RMAP_NONE);
736 else
737 /*
738 * Currently device exclusive access only supports anonymous
739 * memory so the entry shouldn't point to a filebacked page.
740 */
741 WARN_ON_ONCE(1);
742
743 set_pte_at(vma->vm_mm, address, ptep, pte);
744
745 /*
746 * No need to invalidate - it was non-present before. However
747 * secondary CPUs may have mappings that need invalidating.
748 */
749 update_mmu_cache(vma, address, ptep);
750}
751
752/*
753 * Tries to restore an exclusive pte if the page lock can be acquired without
754 * sleeping.
755 */
756static int
757try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
758 unsigned long addr)
759{
760 swp_entry_t entry = pte_to_swp_entry(ptep_get(src_pte));
761 struct page *page = pfn_swap_entry_to_page(entry);
762
763 if (trylock_page(page)) {
764 restore_exclusive_pte(vma, page, addr, src_pte);
765 unlock_page(page);
766 return 0;
767 }
768
769 return -EBUSY;
770}
771
772/*
773 * copy one vm_area from one task to the other. Assumes the page tables
774 * already present in the new task to be cleared in the whole range
775 * covered by this vma.
776 */
777
778static unsigned long
779copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
780 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
781 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
782{
783 unsigned long vm_flags = dst_vma->vm_flags;
784 pte_t orig_pte = ptep_get(src_pte);
785 pte_t pte = orig_pte;
786 struct folio *folio;
787 struct page *page;
788 swp_entry_t entry = pte_to_swp_entry(orig_pte);
789
790 if (likely(!non_swap_entry(entry))) {
791 if (swap_duplicate(entry) < 0)
792 return -EIO;
793
794 /* make sure dst_mm is on swapoff's mmlist. */
795 if (unlikely(list_empty(&dst_mm->mmlist))) {
796 spin_lock(&mmlist_lock);
797 if (list_empty(&dst_mm->mmlist))
798 list_add(&dst_mm->mmlist,
799 &src_mm->mmlist);
800 spin_unlock(&mmlist_lock);
801 }
802 /* Mark the swap entry as shared. */
803 if (pte_swp_exclusive(orig_pte)) {
804 pte = pte_swp_clear_exclusive(orig_pte);
805 set_pte_at(src_mm, addr, src_pte, pte);
806 }
807 rss[MM_SWAPENTS]++;
808 } else if (is_migration_entry(entry)) {
809 folio = pfn_swap_entry_folio(entry);
810
811 rss[mm_counter(folio)]++;
812
813 if (!is_readable_migration_entry(entry) &&
814 is_cow_mapping(vm_flags)) {
815 /*
816 * COW mappings require pages in both parent and child
817 * to be set to read. A previously exclusive entry is
818 * now shared.
819 */
820 entry = make_readable_migration_entry(
821 swp_offset(entry));
822 pte = swp_entry_to_pte(entry);
823 if (pte_swp_soft_dirty(orig_pte))
824 pte = pte_swp_mksoft_dirty(pte);
825 if (pte_swp_uffd_wp(orig_pte))
826 pte = pte_swp_mkuffd_wp(pte);
827 set_pte_at(src_mm, addr, src_pte, pte);
828 }
829 } else if (is_device_private_entry(entry)) {
830 page = pfn_swap_entry_to_page(entry);
831 folio = page_folio(page);
832
833 /*
834 * Update rss count even for unaddressable pages, as
835 * they should treated just like normal pages in this
836 * respect.
837 *
838 * We will likely want to have some new rss counters
839 * for unaddressable pages, at some point. But for now
840 * keep things as they are.
841 */
842 folio_get(folio);
843 rss[mm_counter(folio)]++;
844 /* Cannot fail as these pages cannot get pinned. */
845 folio_try_dup_anon_rmap_pte(folio, page, src_vma);
846
847 /*
848 * We do not preserve soft-dirty information, because so
849 * far, checkpoint/restore is the only feature that
850 * requires that. And checkpoint/restore does not work
851 * when a device driver is involved (you cannot easily
852 * save and restore device driver state).
853 */
854 if (is_writable_device_private_entry(entry) &&
855 is_cow_mapping(vm_flags)) {
856 entry = make_readable_device_private_entry(
857 swp_offset(entry));
858 pte = swp_entry_to_pte(entry);
859 if (pte_swp_uffd_wp(orig_pte))
860 pte = pte_swp_mkuffd_wp(pte);
861 set_pte_at(src_mm, addr, src_pte, pte);
862 }
863 } else if (is_device_exclusive_entry(entry)) {
864 /*
865 * Make device exclusive entries present by restoring the
866 * original entry then copying as for a present pte. Device
867 * exclusive entries currently only support private writable
868 * (ie. COW) mappings.
869 */
870 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
871 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
872 return -EBUSY;
873 return -ENOENT;
874 } else if (is_pte_marker_entry(entry)) {
875 pte_marker marker = copy_pte_marker(entry, dst_vma);
876
877 if (marker)
878 set_pte_at(dst_mm, addr, dst_pte,
879 make_pte_marker(marker));
880 return 0;
881 }
882 if (!userfaultfd_wp(dst_vma))
883 pte = pte_swp_clear_uffd_wp(pte);
884 set_pte_at(dst_mm, addr, dst_pte, pte);
885 return 0;
886}
887
888/*
889 * Copy a present and normal page.
890 *
891 * NOTE! The usual case is that this isn't required;
892 * instead, the caller can just increase the page refcount
893 * and re-use the pte the traditional way.
894 *
895 * And if we need a pre-allocated page but don't yet have
896 * one, return a negative error to let the preallocation
897 * code know so that it can do so outside the page table
898 * lock.
899 */
900static inline int
901copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
902 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
903 struct folio **prealloc, struct page *page)
904{
905 struct folio *new_folio;
906 pte_t pte;
907
908 new_folio = *prealloc;
909 if (!new_folio)
910 return -EAGAIN;
911
912 /*
913 * We have a prealloc page, all good! Take it
914 * over and copy the page & arm it.
915 */
916 *prealloc = NULL;
917 copy_user_highpage(&new_folio->page, page, addr, src_vma);
918 __folio_mark_uptodate(new_folio);
919 folio_add_new_anon_rmap(new_folio, dst_vma, addr);
920 folio_add_lru_vma(new_folio, dst_vma);
921 rss[MM_ANONPAGES]++;
922
923 /* All done, just insert the new page copy in the child */
924 pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot);
925 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
926 if (userfaultfd_pte_wp(dst_vma, ptep_get(src_pte)))
927 /* Uffd-wp needs to be delivered to dest pte as well */
928 pte = pte_mkuffd_wp(pte);
929 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
930 return 0;
931}
932
933static __always_inline void __copy_present_ptes(struct vm_area_struct *dst_vma,
934 struct vm_area_struct *src_vma, pte_t *dst_pte, pte_t *src_pte,
935 pte_t pte, unsigned long addr, int nr)
936{
937 struct mm_struct *src_mm = src_vma->vm_mm;
938
939 /* If it's a COW mapping, write protect it both processes. */
940 if (is_cow_mapping(src_vma->vm_flags) && pte_write(pte)) {
941 wrprotect_ptes(src_mm, addr, src_pte, nr);
942 pte = pte_wrprotect(pte);
943 }
944
945 /* If it's a shared mapping, mark it clean in the child. */
946 if (src_vma->vm_flags & VM_SHARED)
947 pte = pte_mkclean(pte);
948 pte = pte_mkold(pte);
949
950 if (!userfaultfd_wp(dst_vma))
951 pte = pte_clear_uffd_wp(pte);
952
953 set_ptes(dst_vma->vm_mm, addr, dst_pte, pte, nr);
954}
955
956/*
957 * Copy one present PTE, trying to batch-process subsequent PTEs that map
958 * consecutive pages of the same folio by copying them as well.
959 *
960 * Returns -EAGAIN if one preallocated page is required to copy the next PTE.
961 * Otherwise, returns the number of copied PTEs (at least 1).
962 */
963static inline int
964copy_present_ptes(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
965 pte_t *dst_pte, pte_t *src_pte, pte_t pte, unsigned long addr,
966 int max_nr, int *rss, struct folio **prealloc)
967{
968 struct page *page;
969 struct folio *folio;
970 bool any_writable;
971 fpb_t flags = 0;
972 int err, nr;
973
974 page = vm_normal_page(src_vma, addr, pte);
975 if (unlikely(!page))
976 goto copy_pte;
977
978 folio = page_folio(page);
979
980 /*
981 * If we likely have to copy, just don't bother with batching. Make
982 * sure that the common "small folio" case is as fast as possible
983 * by keeping the batching logic separate.
984 */
985 if (unlikely(!*prealloc && folio_test_large(folio) && max_nr != 1)) {
986 if (src_vma->vm_flags & VM_SHARED)
987 flags |= FPB_IGNORE_DIRTY;
988 if (!vma_soft_dirty_enabled(src_vma))
989 flags |= FPB_IGNORE_SOFT_DIRTY;
990
991 nr = folio_pte_batch(folio, addr, src_pte, pte, max_nr, flags,
992 &any_writable);
993 folio_ref_add(folio, nr);
994 if (folio_test_anon(folio)) {
995 if (unlikely(folio_try_dup_anon_rmap_ptes(folio, page,
996 nr, src_vma))) {
997 folio_ref_sub(folio, nr);
998 return -EAGAIN;
999 }
1000 rss[MM_ANONPAGES] += nr;
1001 VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio);
1002 } else {
1003 folio_dup_file_rmap_ptes(folio, page, nr);
1004 rss[mm_counter_file(folio)] += nr;
1005 }
1006 if (any_writable)
1007 pte = pte_mkwrite(pte, src_vma);
1008 __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte,
1009 addr, nr);
1010 return nr;
1011 }
1012
1013 folio_get(folio);
1014 if (folio_test_anon(folio)) {
1015 /*
1016 * If this page may have been pinned by the parent process,
1017 * copy the page immediately for the child so that we'll always
1018 * guarantee the pinned page won't be randomly replaced in the
1019 * future.
1020 */
1021 if (unlikely(folio_try_dup_anon_rmap_pte(folio, page, src_vma))) {
1022 /* Page may be pinned, we have to copy. */
1023 folio_put(folio);
1024 err = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
1025 addr, rss, prealloc, page);
1026 return err ? err : 1;
1027 }
1028 rss[MM_ANONPAGES]++;
1029 VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio);
1030 } else {
1031 folio_dup_file_rmap_pte(folio, page);
1032 rss[mm_counter_file(folio)]++;
1033 }
1034
1035copy_pte:
1036 __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte, addr, 1);
1037 return 1;
1038}
1039
1040static inline struct folio *folio_prealloc(struct mm_struct *src_mm,
1041 struct vm_area_struct *vma, unsigned long addr, bool need_zero)
1042{
1043 struct folio *new_folio;
1044
1045 if (need_zero)
1046 new_folio = vma_alloc_zeroed_movable_folio(vma, addr);
1047 else
1048 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma,
1049 addr, false);
1050
1051 if (!new_folio)
1052 return NULL;
1053
1054 if (mem_cgroup_charge(new_folio, src_mm, GFP_KERNEL)) {
1055 folio_put(new_folio);
1056 return NULL;
1057 }
1058 folio_throttle_swaprate(new_folio, GFP_KERNEL);
1059
1060 return new_folio;
1061}
1062
1063static int
1064copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1065 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1066 unsigned long end)
1067{
1068 struct mm_struct *dst_mm = dst_vma->vm_mm;
1069 struct mm_struct *src_mm = src_vma->vm_mm;
1070 pte_t *orig_src_pte, *orig_dst_pte;
1071 pte_t *src_pte, *dst_pte;
1072 pte_t ptent;
1073 spinlock_t *src_ptl, *dst_ptl;
1074 int progress, max_nr, ret = 0;
1075 int rss[NR_MM_COUNTERS];
1076 swp_entry_t entry = (swp_entry_t){0};
1077 struct folio *prealloc = NULL;
1078 int nr;
1079
1080again:
1081 progress = 0;
1082 init_rss_vec(rss);
1083
1084 /*
1085 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1086 * error handling here, assume that exclusive mmap_lock on dst and src
1087 * protects anon from unexpected THP transitions; with shmem and file
1088 * protected by mmap_lock-less collapse skipping areas with anon_vma
1089 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1090 * can remove such assumptions later, but this is good enough for now.
1091 */
1092 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1093 if (!dst_pte) {
1094 ret = -ENOMEM;
1095 goto out;
1096 }
1097 src_pte = pte_offset_map_nolock(src_mm, src_pmd, addr, &src_ptl);
1098 if (!src_pte) {
1099 pte_unmap_unlock(dst_pte, dst_ptl);
1100 /* ret == 0 */
1101 goto out;
1102 }
1103 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1104 orig_src_pte = src_pte;
1105 orig_dst_pte = dst_pte;
1106 arch_enter_lazy_mmu_mode();
1107
1108 do {
1109 nr = 1;
1110
1111 /*
1112 * We are holding two locks at this point - either of them
1113 * could generate latencies in another task on another CPU.
1114 */
1115 if (progress >= 32) {
1116 progress = 0;
1117 if (need_resched() ||
1118 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1119 break;
1120 }
1121 ptent = ptep_get(src_pte);
1122 if (pte_none(ptent)) {
1123 progress++;
1124 continue;
1125 }
1126 if (unlikely(!pte_present(ptent))) {
1127 ret = copy_nonpresent_pte(dst_mm, src_mm,
1128 dst_pte, src_pte,
1129 dst_vma, src_vma,
1130 addr, rss);
1131 if (ret == -EIO) {
1132 entry = pte_to_swp_entry(ptep_get(src_pte));
1133 break;
1134 } else if (ret == -EBUSY) {
1135 break;
1136 } else if (!ret) {
1137 progress += 8;
1138 continue;
1139 }
1140 ptent = ptep_get(src_pte);
1141 VM_WARN_ON_ONCE(!pte_present(ptent));
1142
1143 /*
1144 * Device exclusive entry restored, continue by copying
1145 * the now present pte.
1146 */
1147 WARN_ON_ONCE(ret != -ENOENT);
1148 }
1149 /* copy_present_ptes() will clear `*prealloc' if consumed */
1150 max_nr = (end - addr) / PAGE_SIZE;
1151 ret = copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte,
1152 ptent, addr, max_nr, rss, &prealloc);
1153 /*
1154 * If we need a pre-allocated page for this pte, drop the
1155 * locks, allocate, and try again.
1156 */
1157 if (unlikely(ret == -EAGAIN))
1158 break;
1159 if (unlikely(prealloc)) {
1160 /*
1161 * pre-alloc page cannot be reused by next time so as
1162 * to strictly follow mempolicy (e.g., alloc_page_vma()
1163 * will allocate page according to address). This
1164 * could only happen if one pinned pte changed.
1165 */
1166 folio_put(prealloc);
1167 prealloc = NULL;
1168 }
1169 nr = ret;
1170 progress += 8 * nr;
1171 } while (dst_pte += nr, src_pte += nr, addr += PAGE_SIZE * nr,
1172 addr != end);
1173
1174 arch_leave_lazy_mmu_mode();
1175 pte_unmap_unlock(orig_src_pte, src_ptl);
1176 add_mm_rss_vec(dst_mm, rss);
1177 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1178 cond_resched();
1179
1180 if (ret == -EIO) {
1181 VM_WARN_ON_ONCE(!entry.val);
1182 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1183 ret = -ENOMEM;
1184 goto out;
1185 }
1186 entry.val = 0;
1187 } else if (ret == -EBUSY) {
1188 goto out;
1189 } else if (ret == -EAGAIN) {
1190 prealloc = folio_prealloc(src_mm, src_vma, addr, false);
1191 if (!prealloc)
1192 return -ENOMEM;
1193 } else if (ret < 0) {
1194 VM_WARN_ON_ONCE(1);
1195 }
1196
1197 /* We've captured and resolved the error. Reset, try again. */
1198 ret = 0;
1199
1200 if (addr != end)
1201 goto again;
1202out:
1203 if (unlikely(prealloc))
1204 folio_put(prealloc);
1205 return ret;
1206}
1207
1208static inline int
1209copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1210 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1211 unsigned long end)
1212{
1213 struct mm_struct *dst_mm = dst_vma->vm_mm;
1214 struct mm_struct *src_mm = src_vma->vm_mm;
1215 pmd_t *src_pmd, *dst_pmd;
1216 unsigned long next;
1217
1218 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1219 if (!dst_pmd)
1220 return -ENOMEM;
1221 src_pmd = pmd_offset(src_pud, addr);
1222 do {
1223 next = pmd_addr_end(addr, end);
1224 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1225 || pmd_devmap(*src_pmd)) {
1226 int err;
1227 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1228 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1229 addr, dst_vma, src_vma);
1230 if (err == -ENOMEM)
1231 return -ENOMEM;
1232 if (!err)
1233 continue;
1234 /* fall through */
1235 }
1236 if (pmd_none_or_clear_bad(src_pmd))
1237 continue;
1238 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1239 addr, next))
1240 return -ENOMEM;
1241 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1242 return 0;
1243}
1244
1245static inline int
1246copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1247 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1248 unsigned long end)
1249{
1250 struct mm_struct *dst_mm = dst_vma->vm_mm;
1251 struct mm_struct *src_mm = src_vma->vm_mm;
1252 pud_t *src_pud, *dst_pud;
1253 unsigned long next;
1254
1255 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1256 if (!dst_pud)
1257 return -ENOMEM;
1258 src_pud = pud_offset(src_p4d, addr);
1259 do {
1260 next = pud_addr_end(addr, end);
1261 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1262 int err;
1263
1264 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1265 err = copy_huge_pud(dst_mm, src_mm,
1266 dst_pud, src_pud, addr, src_vma);
1267 if (err == -ENOMEM)
1268 return -ENOMEM;
1269 if (!err)
1270 continue;
1271 /* fall through */
1272 }
1273 if (pud_none_or_clear_bad(src_pud))
1274 continue;
1275 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1276 addr, next))
1277 return -ENOMEM;
1278 } while (dst_pud++, src_pud++, addr = next, addr != end);
1279 return 0;
1280}
1281
1282static inline int
1283copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1284 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1285 unsigned long end)
1286{
1287 struct mm_struct *dst_mm = dst_vma->vm_mm;
1288 p4d_t *src_p4d, *dst_p4d;
1289 unsigned long next;
1290
1291 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1292 if (!dst_p4d)
1293 return -ENOMEM;
1294 src_p4d = p4d_offset(src_pgd, addr);
1295 do {
1296 next = p4d_addr_end(addr, end);
1297 if (p4d_none_or_clear_bad(src_p4d))
1298 continue;
1299 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1300 addr, next))
1301 return -ENOMEM;
1302 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1303 return 0;
1304}
1305
1306/*
1307 * Return true if the vma needs to copy the pgtable during this fork(). Return
1308 * false when we can speed up fork() by allowing lazy page faults later until
1309 * when the child accesses the memory range.
1310 */
1311static bool
1312vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1313{
1314 /*
1315 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1316 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1317 * contains uffd-wp protection information, that's something we can't
1318 * retrieve from page cache, and skip copying will lose those info.
1319 */
1320 if (userfaultfd_wp(dst_vma))
1321 return true;
1322
1323 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1324 return true;
1325
1326 if (src_vma->anon_vma)
1327 return true;
1328
1329 /*
1330 * Don't copy ptes where a page fault will fill them correctly. Fork
1331 * becomes much lighter when there are big shared or private readonly
1332 * mappings. The tradeoff is that copy_page_range is more efficient
1333 * than faulting.
1334 */
1335 return false;
1336}
1337
1338int
1339copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1340{
1341 pgd_t *src_pgd, *dst_pgd;
1342 unsigned long next;
1343 unsigned long addr = src_vma->vm_start;
1344 unsigned long end = src_vma->vm_end;
1345 struct mm_struct *dst_mm = dst_vma->vm_mm;
1346 struct mm_struct *src_mm = src_vma->vm_mm;
1347 struct mmu_notifier_range range;
1348 bool is_cow;
1349 int ret;
1350
1351 if (!vma_needs_copy(dst_vma, src_vma))
1352 return 0;
1353
1354 if (is_vm_hugetlb_page(src_vma))
1355 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1356
1357 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1358 /*
1359 * We do not free on error cases below as remove_vma
1360 * gets called on error from higher level routine
1361 */
1362 ret = track_pfn_copy(src_vma);
1363 if (ret)
1364 return ret;
1365 }
1366
1367 /*
1368 * We need to invalidate the secondary MMU mappings only when
1369 * there could be a permission downgrade on the ptes of the
1370 * parent mm. And a permission downgrade will only happen if
1371 * is_cow_mapping() returns true.
1372 */
1373 is_cow = is_cow_mapping(src_vma->vm_flags);
1374
1375 if (is_cow) {
1376 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1377 0, src_mm, addr, end);
1378 mmu_notifier_invalidate_range_start(&range);
1379 /*
1380 * Disabling preemption is not needed for the write side, as
1381 * the read side doesn't spin, but goes to the mmap_lock.
1382 *
1383 * Use the raw variant of the seqcount_t write API to avoid
1384 * lockdep complaining about preemptibility.
1385 */
1386 vma_assert_write_locked(src_vma);
1387 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1388 }
1389
1390 ret = 0;
1391 dst_pgd = pgd_offset(dst_mm, addr);
1392 src_pgd = pgd_offset(src_mm, addr);
1393 do {
1394 next = pgd_addr_end(addr, end);
1395 if (pgd_none_or_clear_bad(src_pgd))
1396 continue;
1397 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1398 addr, next))) {
1399 untrack_pfn_clear(dst_vma);
1400 ret = -ENOMEM;
1401 break;
1402 }
1403 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1404
1405 if (is_cow) {
1406 raw_write_seqcount_end(&src_mm->write_protect_seq);
1407 mmu_notifier_invalidate_range_end(&range);
1408 }
1409 return ret;
1410}
1411
1412/* Whether we should zap all COWed (private) pages too */
1413static inline bool should_zap_cows(struct zap_details *details)
1414{
1415 /* By default, zap all pages */
1416 if (!details)
1417 return true;
1418
1419 /* Or, we zap COWed pages only if the caller wants to */
1420 return details->even_cows;
1421}
1422
1423/* Decides whether we should zap this folio with the folio pointer specified */
1424static inline bool should_zap_folio(struct zap_details *details,
1425 struct folio *folio)
1426{
1427 /* If we can make a decision without *folio.. */
1428 if (should_zap_cows(details))
1429 return true;
1430
1431 /* Otherwise we should only zap non-anon folios */
1432 return !folio_test_anon(folio);
1433}
1434
1435static inline bool zap_drop_file_uffd_wp(struct zap_details *details)
1436{
1437 if (!details)
1438 return false;
1439
1440 return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1441}
1442
1443/*
1444 * This function makes sure that we'll replace the none pte with an uffd-wp
1445 * swap special pte marker when necessary. Must be with the pgtable lock held.
1446 */
1447static inline void
1448zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1449 unsigned long addr, pte_t *pte, int nr,
1450 struct zap_details *details, pte_t pteval)
1451{
1452 /* Zap on anonymous always means dropping everything */
1453 if (vma_is_anonymous(vma))
1454 return;
1455
1456 if (zap_drop_file_uffd_wp(details))
1457 return;
1458
1459 for (;;) {
1460 /* the PFN in the PTE is irrelevant. */
1461 pte_install_uffd_wp_if_needed(vma, addr, pte, pteval);
1462 if (--nr == 0)
1463 break;
1464 pte++;
1465 addr += PAGE_SIZE;
1466 }
1467}
1468
1469static __always_inline void zap_present_folio_ptes(struct mmu_gather *tlb,
1470 struct vm_area_struct *vma, struct folio *folio,
1471 struct page *page, pte_t *pte, pte_t ptent, unsigned int nr,
1472 unsigned long addr, struct zap_details *details, int *rss,
1473 bool *force_flush, bool *force_break)
1474{
1475 struct mm_struct *mm = tlb->mm;
1476 bool delay_rmap = false;
1477
1478 if (!folio_test_anon(folio)) {
1479 ptent = get_and_clear_full_ptes(mm, addr, pte, nr, tlb->fullmm);
1480 if (pte_dirty(ptent)) {
1481 folio_mark_dirty(folio);
1482 if (tlb_delay_rmap(tlb)) {
1483 delay_rmap = true;
1484 *force_flush = true;
1485 }
1486 }
1487 if (pte_young(ptent) && likely(vma_has_recency(vma)))
1488 folio_mark_accessed(folio);
1489 rss[mm_counter(folio)] -= nr;
1490 } else {
1491 /* We don't need up-to-date accessed/dirty bits. */
1492 clear_full_ptes(mm, addr, pte, nr, tlb->fullmm);
1493 rss[MM_ANONPAGES] -= nr;
1494 }
1495 /* Checking a single PTE in a batch is sufficient. */
1496 arch_check_zapped_pte(vma, ptent);
1497 tlb_remove_tlb_entries(tlb, pte, nr, addr);
1498 if (unlikely(userfaultfd_pte_wp(vma, ptent)))
1499 zap_install_uffd_wp_if_needed(vma, addr, pte, nr, details,
1500 ptent);
1501
1502 if (!delay_rmap) {
1503 folio_remove_rmap_ptes(folio, page, nr, vma);
1504
1505 /* Only sanity-check the first page in a batch. */
1506 if (unlikely(page_mapcount(page) < 0))
1507 print_bad_pte(vma, addr, ptent, page);
1508 }
1509 if (unlikely(__tlb_remove_folio_pages(tlb, page, nr, delay_rmap))) {
1510 *force_flush = true;
1511 *force_break = true;
1512 }
1513}
1514
1515/*
1516 * Zap or skip at least one present PTE, trying to batch-process subsequent
1517 * PTEs that map consecutive pages of the same folio.
1518 *
1519 * Returns the number of processed (skipped or zapped) PTEs (at least 1).
1520 */
1521static inline int zap_present_ptes(struct mmu_gather *tlb,
1522 struct vm_area_struct *vma, pte_t *pte, pte_t ptent,
1523 unsigned int max_nr, unsigned long addr,
1524 struct zap_details *details, int *rss, bool *force_flush,
1525 bool *force_break)
1526{
1527 const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
1528 struct mm_struct *mm = tlb->mm;
1529 struct folio *folio;
1530 struct page *page;
1531 int nr;
1532
1533 page = vm_normal_page(vma, addr, ptent);
1534 if (!page) {
1535 /* We don't need up-to-date accessed/dirty bits. */
1536 ptep_get_and_clear_full(mm, addr, pte, tlb->fullmm);
1537 arch_check_zapped_pte(vma, ptent);
1538 tlb_remove_tlb_entry(tlb, pte, addr);
1539 if (userfaultfd_pte_wp(vma, ptent))
1540 zap_install_uffd_wp_if_needed(vma, addr, pte, 1,
1541 details, ptent);
1542 ksm_might_unmap_zero_page(mm, ptent);
1543 return 1;
1544 }
1545
1546 folio = page_folio(page);
1547 if (unlikely(!should_zap_folio(details, folio)))
1548 return 1;
1549
1550 /*
1551 * Make sure that the common "small folio" case is as fast as possible
1552 * by keeping the batching logic separate.
1553 */
1554 if (unlikely(folio_test_large(folio) && max_nr != 1)) {
1555 nr = folio_pte_batch(folio, addr, pte, ptent, max_nr, fpb_flags,
1556 NULL);
1557
1558 zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, nr,
1559 addr, details, rss, force_flush,
1560 force_break);
1561 return nr;
1562 }
1563 zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, 1, addr,
1564 details, rss, force_flush, force_break);
1565 return 1;
1566}
1567
1568static unsigned long zap_pte_range(struct mmu_gather *tlb,
1569 struct vm_area_struct *vma, pmd_t *pmd,
1570 unsigned long addr, unsigned long end,
1571 struct zap_details *details)
1572{
1573 bool force_flush = false, force_break = false;
1574 struct mm_struct *mm = tlb->mm;
1575 int rss[NR_MM_COUNTERS];
1576 spinlock_t *ptl;
1577 pte_t *start_pte;
1578 pte_t *pte;
1579 swp_entry_t entry;
1580 int nr;
1581
1582 tlb_change_page_size(tlb, PAGE_SIZE);
1583 init_rss_vec(rss);
1584 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1585 if (!pte)
1586 return addr;
1587
1588 flush_tlb_batched_pending(mm);
1589 arch_enter_lazy_mmu_mode();
1590 do {
1591 pte_t ptent = ptep_get(pte);
1592 struct folio *folio;
1593 struct page *page;
1594 int max_nr;
1595
1596 nr = 1;
1597 if (pte_none(ptent))
1598 continue;
1599
1600 if (need_resched())
1601 break;
1602
1603 if (pte_present(ptent)) {
1604 max_nr = (end - addr) / PAGE_SIZE;
1605 nr = zap_present_ptes(tlb, vma, pte, ptent, max_nr,
1606 addr, details, rss, &force_flush,
1607 &force_break);
1608 if (unlikely(force_break)) {
1609 addr += nr * PAGE_SIZE;
1610 break;
1611 }
1612 continue;
1613 }
1614
1615 entry = pte_to_swp_entry(ptent);
1616 if (is_device_private_entry(entry) ||
1617 is_device_exclusive_entry(entry)) {
1618 page = pfn_swap_entry_to_page(entry);
1619 folio = page_folio(page);
1620 if (unlikely(!should_zap_folio(details, folio)))
1621 continue;
1622 /*
1623 * Both device private/exclusive mappings should only
1624 * work with anonymous page so far, so we don't need to
1625 * consider uffd-wp bit when zap. For more information,
1626 * see zap_install_uffd_wp_if_needed().
1627 */
1628 WARN_ON_ONCE(!vma_is_anonymous(vma));
1629 rss[mm_counter(folio)]--;
1630 if (is_device_private_entry(entry))
1631 folio_remove_rmap_pte(folio, page, vma);
1632 folio_put(folio);
1633 } else if (!non_swap_entry(entry)) {
1634 /* Genuine swap entry, hence a private anon page */
1635 if (!should_zap_cows(details))
1636 continue;
1637 rss[MM_SWAPENTS]--;
1638 if (unlikely(!free_swap_and_cache(entry)))
1639 print_bad_pte(vma, addr, ptent, NULL);
1640 } else if (is_migration_entry(entry)) {
1641 folio = pfn_swap_entry_folio(entry);
1642 if (!should_zap_folio(details, folio))
1643 continue;
1644 rss[mm_counter(folio)]--;
1645 } else if (pte_marker_entry_uffd_wp(entry)) {
1646 /*
1647 * For anon: always drop the marker; for file: only
1648 * drop the marker if explicitly requested.
1649 */
1650 if (!vma_is_anonymous(vma) &&
1651 !zap_drop_file_uffd_wp(details))
1652 continue;
1653 } else if (is_hwpoison_entry(entry) ||
1654 is_poisoned_swp_entry(entry)) {
1655 if (!should_zap_cows(details))
1656 continue;
1657 } else {
1658 /* We should have covered all the swap entry types */
1659 pr_alert("unrecognized swap entry 0x%lx\n", entry.val);
1660 WARN_ON_ONCE(1);
1661 }
1662 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1663 zap_install_uffd_wp_if_needed(vma, addr, pte, 1, details, ptent);
1664 } while (pte += nr, addr += PAGE_SIZE * nr, addr != end);
1665
1666 add_mm_rss_vec(mm, rss);
1667 arch_leave_lazy_mmu_mode();
1668
1669 /* Do the actual TLB flush before dropping ptl */
1670 if (force_flush) {
1671 tlb_flush_mmu_tlbonly(tlb);
1672 tlb_flush_rmaps(tlb, vma);
1673 }
1674 pte_unmap_unlock(start_pte, ptl);
1675
1676 /*
1677 * If we forced a TLB flush (either due to running out of
1678 * batch buffers or because we needed to flush dirty TLB
1679 * entries before releasing the ptl), free the batched
1680 * memory too. Come back again if we didn't do everything.
1681 */
1682 if (force_flush)
1683 tlb_flush_mmu(tlb);
1684
1685 return addr;
1686}
1687
1688static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1689 struct vm_area_struct *vma, pud_t *pud,
1690 unsigned long addr, unsigned long end,
1691 struct zap_details *details)
1692{
1693 pmd_t *pmd;
1694 unsigned long next;
1695
1696 pmd = pmd_offset(pud, addr);
1697 do {
1698 next = pmd_addr_end(addr, end);
1699 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1700 if (next - addr != HPAGE_PMD_SIZE)
1701 __split_huge_pmd(vma, pmd, addr, false, NULL);
1702 else if (zap_huge_pmd(tlb, vma, pmd, addr)) {
1703 addr = next;
1704 continue;
1705 }
1706 /* fall through */
1707 } else if (details && details->single_folio &&
1708 folio_test_pmd_mappable(details->single_folio) &&
1709 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1710 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1711 /*
1712 * Take and drop THP pmd lock so that we cannot return
1713 * prematurely, while zap_huge_pmd() has cleared *pmd,
1714 * but not yet decremented compound_mapcount().
1715 */
1716 spin_unlock(ptl);
1717 }
1718 if (pmd_none(*pmd)) {
1719 addr = next;
1720 continue;
1721 }
1722 addr = zap_pte_range(tlb, vma, pmd, addr, next, details);
1723 if (addr != next)
1724 pmd--;
1725 } while (pmd++, cond_resched(), addr != end);
1726
1727 return addr;
1728}
1729
1730static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1731 struct vm_area_struct *vma, p4d_t *p4d,
1732 unsigned long addr, unsigned long end,
1733 struct zap_details *details)
1734{
1735 pud_t *pud;
1736 unsigned long next;
1737
1738 pud = pud_offset(p4d, addr);
1739 do {
1740 next = pud_addr_end(addr, end);
1741 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1742 if (next - addr != HPAGE_PUD_SIZE) {
1743 mmap_assert_locked(tlb->mm);
1744 split_huge_pud(vma, pud, addr);
1745 } else if (zap_huge_pud(tlb, vma, pud, addr))
1746 goto next;
1747 /* fall through */
1748 }
1749 if (pud_none_or_clear_bad(pud))
1750 continue;
1751 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1752next:
1753 cond_resched();
1754 } while (pud++, addr = next, addr != end);
1755
1756 return addr;
1757}
1758
1759static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1760 struct vm_area_struct *vma, pgd_t *pgd,
1761 unsigned long addr, unsigned long end,
1762 struct zap_details *details)
1763{
1764 p4d_t *p4d;
1765 unsigned long next;
1766
1767 p4d = p4d_offset(pgd, addr);
1768 do {
1769 next = p4d_addr_end(addr, end);
1770 if (p4d_none_or_clear_bad(p4d))
1771 continue;
1772 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1773 } while (p4d++, addr = next, addr != end);
1774
1775 return addr;
1776}
1777
1778void unmap_page_range(struct mmu_gather *tlb,
1779 struct vm_area_struct *vma,
1780 unsigned long addr, unsigned long end,
1781 struct zap_details *details)
1782{
1783 pgd_t *pgd;
1784 unsigned long next;
1785
1786 BUG_ON(addr >= end);
1787 tlb_start_vma(tlb, vma);
1788 pgd = pgd_offset(vma->vm_mm, addr);
1789 do {
1790 next = pgd_addr_end(addr, end);
1791 if (pgd_none_or_clear_bad(pgd))
1792 continue;
1793 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1794 } while (pgd++, addr = next, addr != end);
1795 tlb_end_vma(tlb, vma);
1796}
1797
1798
1799static void unmap_single_vma(struct mmu_gather *tlb,
1800 struct vm_area_struct *vma, unsigned long start_addr,
1801 unsigned long end_addr,
1802 struct zap_details *details, bool mm_wr_locked)
1803{
1804 unsigned long start = max(vma->vm_start, start_addr);
1805 unsigned long end;
1806
1807 if (start >= vma->vm_end)
1808 return;
1809 end = min(vma->vm_end, end_addr);
1810 if (end <= vma->vm_start)
1811 return;
1812
1813 if (vma->vm_file)
1814 uprobe_munmap(vma, start, end);
1815
1816 if (unlikely(vma->vm_flags & VM_PFNMAP))
1817 untrack_pfn(vma, 0, 0, mm_wr_locked);
1818
1819 if (start != end) {
1820 if (unlikely(is_vm_hugetlb_page(vma))) {
1821 /*
1822 * It is undesirable to test vma->vm_file as it
1823 * should be non-null for valid hugetlb area.
1824 * However, vm_file will be NULL in the error
1825 * cleanup path of mmap_region. When
1826 * hugetlbfs ->mmap method fails,
1827 * mmap_region() nullifies vma->vm_file
1828 * before calling this function to clean up.
1829 * Since no pte has actually been setup, it is
1830 * safe to do nothing in this case.
1831 */
1832 if (vma->vm_file) {
1833 zap_flags_t zap_flags = details ?
1834 details->zap_flags : 0;
1835 __unmap_hugepage_range(tlb, vma, start, end,
1836 NULL, zap_flags);
1837 }
1838 } else
1839 unmap_page_range(tlb, vma, start, end, details);
1840 }
1841}
1842
1843/**
1844 * unmap_vmas - unmap a range of memory covered by a list of vma's
1845 * @tlb: address of the caller's struct mmu_gather
1846 * @mas: the maple state
1847 * @vma: the starting vma
1848 * @start_addr: virtual address at which to start unmapping
1849 * @end_addr: virtual address at which to end unmapping
1850 * @tree_end: The maximum index to check
1851 * @mm_wr_locked: lock flag
1852 *
1853 * Unmap all pages in the vma list.
1854 *
1855 * Only addresses between `start' and `end' will be unmapped.
1856 *
1857 * The VMA list must be sorted in ascending virtual address order.
1858 *
1859 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1860 * range after unmap_vmas() returns. So the only responsibility here is to
1861 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1862 * drops the lock and schedules.
1863 */
1864void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas,
1865 struct vm_area_struct *vma, unsigned long start_addr,
1866 unsigned long end_addr, unsigned long tree_end,
1867 bool mm_wr_locked)
1868{
1869 struct mmu_notifier_range range;
1870 struct zap_details details = {
1871 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
1872 /* Careful - we need to zap private pages too! */
1873 .even_cows = true,
1874 };
1875
1876 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm,
1877 start_addr, end_addr);
1878 mmu_notifier_invalidate_range_start(&range);
1879 do {
1880 unsigned long start = start_addr;
1881 unsigned long end = end_addr;
1882 hugetlb_zap_begin(vma, &start, &end);
1883 unmap_single_vma(tlb, vma, start, end, &details,
1884 mm_wr_locked);
1885 hugetlb_zap_end(vma, &details);
1886 vma = mas_find(mas, tree_end - 1);
1887 } while (vma && likely(!xa_is_zero(vma)));
1888 mmu_notifier_invalidate_range_end(&range);
1889}
1890
1891/**
1892 * zap_page_range_single - remove user pages in a given range
1893 * @vma: vm_area_struct holding the applicable pages
1894 * @address: starting address of pages to zap
1895 * @size: number of bytes to zap
1896 * @details: details of shared cache invalidation
1897 *
1898 * The range must fit into one VMA.
1899 */
1900void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1901 unsigned long size, struct zap_details *details)
1902{
1903 const unsigned long end = address + size;
1904 struct mmu_notifier_range range;
1905 struct mmu_gather tlb;
1906
1907 lru_add_drain();
1908 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1909 address, end);
1910 hugetlb_zap_begin(vma, &range.start, &range.end);
1911 tlb_gather_mmu(&tlb, vma->vm_mm);
1912 update_hiwater_rss(vma->vm_mm);
1913 mmu_notifier_invalidate_range_start(&range);
1914 /*
1915 * unmap 'address-end' not 'range.start-range.end' as range
1916 * could have been expanded for hugetlb pmd sharing.
1917 */
1918 unmap_single_vma(&tlb, vma, address, end, details, false);
1919 mmu_notifier_invalidate_range_end(&range);
1920 tlb_finish_mmu(&tlb);
1921 hugetlb_zap_end(vma, details);
1922}
1923
1924/**
1925 * zap_vma_ptes - remove ptes mapping the vma
1926 * @vma: vm_area_struct holding ptes to be zapped
1927 * @address: starting address of pages to zap
1928 * @size: number of bytes to zap
1929 *
1930 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1931 *
1932 * The entire address range must be fully contained within the vma.
1933 *
1934 */
1935void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1936 unsigned long size)
1937{
1938 if (!range_in_vma(vma, address, address + size) ||
1939 !(vma->vm_flags & VM_PFNMAP))
1940 return;
1941
1942 zap_page_range_single(vma, address, size, NULL);
1943}
1944EXPORT_SYMBOL_GPL(zap_vma_ptes);
1945
1946static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1947{
1948 pgd_t *pgd;
1949 p4d_t *p4d;
1950 pud_t *pud;
1951 pmd_t *pmd;
1952
1953 pgd = pgd_offset(mm, addr);
1954 p4d = p4d_alloc(mm, pgd, addr);
1955 if (!p4d)
1956 return NULL;
1957 pud = pud_alloc(mm, p4d, addr);
1958 if (!pud)
1959 return NULL;
1960 pmd = pmd_alloc(mm, pud, addr);
1961 if (!pmd)
1962 return NULL;
1963
1964 VM_BUG_ON(pmd_trans_huge(*pmd));
1965 return pmd;
1966}
1967
1968pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1969 spinlock_t **ptl)
1970{
1971 pmd_t *pmd = walk_to_pmd(mm, addr);
1972
1973 if (!pmd)
1974 return NULL;
1975 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1976}
1977
1978static int validate_page_before_insert(struct page *page)
1979{
1980 struct folio *folio = page_folio(page);
1981
1982 if (folio_test_anon(folio) || folio_test_slab(folio) ||
1983 page_has_type(page))
1984 return -EINVAL;
1985 flush_dcache_folio(folio);
1986 return 0;
1987}
1988
1989static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1990 unsigned long addr, struct page *page, pgprot_t prot)
1991{
1992 struct folio *folio = page_folio(page);
1993
1994 if (!pte_none(ptep_get(pte)))
1995 return -EBUSY;
1996 /* Ok, finally just insert the thing.. */
1997 folio_get(folio);
1998 inc_mm_counter(vma->vm_mm, mm_counter_file(folio));
1999 folio_add_file_rmap_pte(folio, page, vma);
2000 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
2001 return 0;
2002}
2003
2004/*
2005 * This is the old fallback for page remapping.
2006 *
2007 * For historical reasons, it only allows reserved pages. Only
2008 * old drivers should use this, and they needed to mark their
2009 * pages reserved for the old functions anyway.
2010 */
2011static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2012 struct page *page, pgprot_t prot)
2013{
2014 int retval;
2015 pte_t *pte;
2016 spinlock_t *ptl;
2017
2018 retval = validate_page_before_insert(page);
2019 if (retval)
2020 goto out;
2021 retval = -ENOMEM;
2022 pte = get_locked_pte(vma->vm_mm, addr, &ptl);
2023 if (!pte)
2024 goto out;
2025 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
2026 pte_unmap_unlock(pte, ptl);
2027out:
2028 return retval;
2029}
2030
2031static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
2032 unsigned long addr, struct page *page, pgprot_t prot)
2033{
2034 int err;
2035
2036 if (!page_count(page))
2037 return -EINVAL;
2038 err = validate_page_before_insert(page);
2039 if (err)
2040 return err;
2041 return insert_page_into_pte_locked(vma, pte, addr, page, prot);
2042}
2043
2044/* insert_pages() amortizes the cost of spinlock operations
2045 * when inserting pages in a loop.
2046 */
2047static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
2048 struct page **pages, unsigned long *num, pgprot_t prot)
2049{
2050 pmd_t *pmd = NULL;
2051 pte_t *start_pte, *pte;
2052 spinlock_t *pte_lock;
2053 struct mm_struct *const mm = vma->vm_mm;
2054 unsigned long curr_page_idx = 0;
2055 unsigned long remaining_pages_total = *num;
2056 unsigned long pages_to_write_in_pmd;
2057 int ret;
2058more:
2059 ret = -EFAULT;
2060 pmd = walk_to_pmd(mm, addr);
2061 if (!pmd)
2062 goto out;
2063
2064 pages_to_write_in_pmd = min_t(unsigned long,
2065 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
2066
2067 /* Allocate the PTE if necessary; takes PMD lock once only. */
2068 ret = -ENOMEM;
2069 if (pte_alloc(mm, pmd))
2070 goto out;
2071
2072 while (pages_to_write_in_pmd) {
2073 int pte_idx = 0;
2074 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
2075
2076 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
2077 if (!start_pte) {
2078 ret = -EFAULT;
2079 goto out;
2080 }
2081 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
2082 int err = insert_page_in_batch_locked(vma, pte,
2083 addr, pages[curr_page_idx], prot);
2084 if (unlikely(err)) {
2085 pte_unmap_unlock(start_pte, pte_lock);
2086 ret = err;
2087 remaining_pages_total -= pte_idx;
2088 goto out;
2089 }
2090 addr += PAGE_SIZE;
2091 ++curr_page_idx;
2092 }
2093 pte_unmap_unlock(start_pte, pte_lock);
2094 pages_to_write_in_pmd -= batch_size;
2095 remaining_pages_total -= batch_size;
2096 }
2097 if (remaining_pages_total)
2098 goto more;
2099 ret = 0;
2100out:
2101 *num = remaining_pages_total;
2102 return ret;
2103}
2104
2105/**
2106 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
2107 * @vma: user vma to map to
2108 * @addr: target start user address of these pages
2109 * @pages: source kernel pages
2110 * @num: in: number of pages to map. out: number of pages that were *not*
2111 * mapped. (0 means all pages were successfully mapped).
2112 *
2113 * Preferred over vm_insert_page() when inserting multiple pages.
2114 *
2115 * In case of error, we may have mapped a subset of the provided
2116 * pages. It is the caller's responsibility to account for this case.
2117 *
2118 * The same restrictions apply as in vm_insert_page().
2119 */
2120int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
2121 struct page **pages, unsigned long *num)
2122{
2123 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
2124
2125 if (addr < vma->vm_start || end_addr >= vma->vm_end)
2126 return -EFAULT;
2127 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2128 BUG_ON(mmap_read_trylock(vma->vm_mm));
2129 BUG_ON(vma->vm_flags & VM_PFNMAP);
2130 vm_flags_set(vma, VM_MIXEDMAP);
2131 }
2132 /* Defer page refcount checking till we're about to map that page. */
2133 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
2134}
2135EXPORT_SYMBOL(vm_insert_pages);
2136
2137/**
2138 * vm_insert_page - insert single page into user vma
2139 * @vma: user vma to map to
2140 * @addr: target user address of this page
2141 * @page: source kernel page
2142 *
2143 * This allows drivers to insert individual pages they've allocated
2144 * into a user vma.
2145 *
2146 * The page has to be a nice clean _individual_ kernel allocation.
2147 * If you allocate a compound page, you need to have marked it as
2148 * such (__GFP_COMP), or manually just split the page up yourself
2149 * (see split_page()).
2150 *
2151 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2152 * took an arbitrary page protection parameter. This doesn't allow
2153 * that. Your vma protection will have to be set up correctly, which
2154 * means that if you want a shared writable mapping, you'd better
2155 * ask for a shared writable mapping!
2156 *
2157 * The page does not need to be reserved.
2158 *
2159 * Usually this function is called from f_op->mmap() handler
2160 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2161 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2162 * function from other places, for example from page-fault handler.
2163 *
2164 * Return: %0 on success, negative error code otherwise.
2165 */
2166int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2167 struct page *page)
2168{
2169 if (addr < vma->vm_start || addr >= vma->vm_end)
2170 return -EFAULT;
2171 if (!page_count(page))
2172 return -EINVAL;
2173 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2174 BUG_ON(mmap_read_trylock(vma->vm_mm));
2175 BUG_ON(vma->vm_flags & VM_PFNMAP);
2176 vm_flags_set(vma, VM_MIXEDMAP);
2177 }
2178 return insert_page(vma, addr, page, vma->vm_page_prot);
2179}
2180EXPORT_SYMBOL(vm_insert_page);
2181
2182/*
2183 * __vm_map_pages - maps range of kernel pages into user vma
2184 * @vma: user vma to map to
2185 * @pages: pointer to array of source kernel pages
2186 * @num: number of pages in page array
2187 * @offset: user's requested vm_pgoff
2188 *
2189 * This allows drivers to map range of kernel pages into a user vma.
2190 *
2191 * Return: 0 on success and error code otherwise.
2192 */
2193static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2194 unsigned long num, unsigned long offset)
2195{
2196 unsigned long count = vma_pages(vma);
2197 unsigned long uaddr = vma->vm_start;
2198 int ret, i;
2199
2200 /* Fail if the user requested offset is beyond the end of the object */
2201 if (offset >= num)
2202 return -ENXIO;
2203
2204 /* Fail if the user requested size exceeds available object size */
2205 if (count > num - offset)
2206 return -ENXIO;
2207
2208 for (i = 0; i < count; i++) {
2209 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2210 if (ret < 0)
2211 return ret;
2212 uaddr += PAGE_SIZE;
2213 }
2214
2215 return 0;
2216}
2217
2218/**
2219 * vm_map_pages - maps range of kernel pages starts with non zero offset
2220 * @vma: user vma to map to
2221 * @pages: pointer to array of source kernel pages
2222 * @num: number of pages in page array
2223 *
2224 * Maps an object consisting of @num pages, catering for the user's
2225 * requested vm_pgoff
2226 *
2227 * If we fail to insert any page into the vma, the function will return
2228 * immediately leaving any previously inserted pages present. Callers
2229 * from the mmap handler may immediately return the error as their caller
2230 * will destroy the vma, removing any successfully inserted pages. Other
2231 * callers should make their own arrangements for calling unmap_region().
2232 *
2233 * Context: Process context. Called by mmap handlers.
2234 * Return: 0 on success and error code otherwise.
2235 */
2236int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2237 unsigned long num)
2238{
2239 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2240}
2241EXPORT_SYMBOL(vm_map_pages);
2242
2243/**
2244 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2245 * @vma: user vma to map to
2246 * @pages: pointer to array of source kernel pages
2247 * @num: number of pages in page array
2248 *
2249 * Similar to vm_map_pages(), except that it explicitly sets the offset
2250 * to 0. This function is intended for the drivers that did not consider
2251 * vm_pgoff.
2252 *
2253 * Context: Process context. Called by mmap handlers.
2254 * Return: 0 on success and error code otherwise.
2255 */
2256int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2257 unsigned long num)
2258{
2259 return __vm_map_pages(vma, pages, num, 0);
2260}
2261EXPORT_SYMBOL(vm_map_pages_zero);
2262
2263static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2264 pfn_t pfn, pgprot_t prot, bool mkwrite)
2265{
2266 struct mm_struct *mm = vma->vm_mm;
2267 pte_t *pte, entry;
2268 spinlock_t *ptl;
2269
2270 pte = get_locked_pte(mm, addr, &ptl);
2271 if (!pte)
2272 return VM_FAULT_OOM;
2273 entry = ptep_get(pte);
2274 if (!pte_none(entry)) {
2275 if (mkwrite) {
2276 /*
2277 * For read faults on private mappings the PFN passed
2278 * in may not match the PFN we have mapped if the
2279 * mapped PFN is a writeable COW page. In the mkwrite
2280 * case we are creating a writable PTE for a shared
2281 * mapping and we expect the PFNs to match. If they
2282 * don't match, we are likely racing with block
2283 * allocation and mapping invalidation so just skip the
2284 * update.
2285 */
2286 if (pte_pfn(entry) != pfn_t_to_pfn(pfn)) {
2287 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
2288 goto out_unlock;
2289 }
2290 entry = pte_mkyoung(entry);
2291 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2292 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2293 update_mmu_cache(vma, addr, pte);
2294 }
2295 goto out_unlock;
2296 }
2297
2298 /* Ok, finally just insert the thing.. */
2299 if (pfn_t_devmap(pfn))
2300 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2301 else
2302 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2303
2304 if (mkwrite) {
2305 entry = pte_mkyoung(entry);
2306 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2307 }
2308
2309 set_pte_at(mm, addr, pte, entry);
2310 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2311
2312out_unlock:
2313 pte_unmap_unlock(pte, ptl);
2314 return VM_FAULT_NOPAGE;
2315}
2316
2317/**
2318 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2319 * @vma: user vma to map to
2320 * @addr: target user address of this page
2321 * @pfn: source kernel pfn
2322 * @pgprot: pgprot flags for the inserted page
2323 *
2324 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2325 * to override pgprot on a per-page basis.
2326 *
2327 * This only makes sense for IO mappings, and it makes no sense for
2328 * COW mappings. In general, using multiple vmas is preferable;
2329 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2330 * impractical.
2331 *
2332 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2333 * caching- and encryption bits different than those of @vma->vm_page_prot,
2334 * because the caching- or encryption mode may not be known at mmap() time.
2335 *
2336 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2337 * to set caching and encryption bits for those vmas (except for COW pages).
2338 * This is ensured by core vm only modifying these page table entries using
2339 * functions that don't touch caching- or encryption bits, using pte_modify()
2340 * if needed. (See for example mprotect()).
2341 *
2342 * Also when new page-table entries are created, this is only done using the
2343 * fault() callback, and never using the value of vma->vm_page_prot,
2344 * except for page-table entries that point to anonymous pages as the result
2345 * of COW.
2346 *
2347 * Context: Process context. May allocate using %GFP_KERNEL.
2348 * Return: vm_fault_t value.
2349 */
2350vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2351 unsigned long pfn, pgprot_t pgprot)
2352{
2353 /*
2354 * Technically, architectures with pte_special can avoid all these
2355 * restrictions (same for remap_pfn_range). However we would like
2356 * consistency in testing and feature parity among all, so we should
2357 * try to keep these invariants in place for everybody.
2358 */
2359 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2360 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2361 (VM_PFNMAP|VM_MIXEDMAP));
2362 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2363 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2364
2365 if (addr < vma->vm_start || addr >= vma->vm_end)
2366 return VM_FAULT_SIGBUS;
2367
2368 if (!pfn_modify_allowed(pfn, pgprot))
2369 return VM_FAULT_SIGBUS;
2370
2371 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2372
2373 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2374 false);
2375}
2376EXPORT_SYMBOL(vmf_insert_pfn_prot);
2377
2378/**
2379 * vmf_insert_pfn - insert single pfn into user vma
2380 * @vma: user vma to map to
2381 * @addr: target user address of this page
2382 * @pfn: source kernel pfn
2383 *
2384 * Similar to vm_insert_page, this allows drivers to insert individual pages
2385 * they've allocated into a user vma. Same comments apply.
2386 *
2387 * This function should only be called from a vm_ops->fault handler, and
2388 * in that case the handler should return the result of this function.
2389 *
2390 * vma cannot be a COW mapping.
2391 *
2392 * As this is called only for pages that do not currently exist, we
2393 * do not need to flush old virtual caches or the TLB.
2394 *
2395 * Context: Process context. May allocate using %GFP_KERNEL.
2396 * Return: vm_fault_t value.
2397 */
2398vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2399 unsigned long pfn)
2400{
2401 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2402}
2403EXPORT_SYMBOL(vmf_insert_pfn);
2404
2405static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2406{
2407 /* these checks mirror the abort conditions in vm_normal_page */
2408 if (vma->vm_flags & VM_MIXEDMAP)
2409 return true;
2410 if (pfn_t_devmap(pfn))
2411 return true;
2412 if (pfn_t_special(pfn))
2413 return true;
2414 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2415 return true;
2416 return false;
2417}
2418
2419static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2420 unsigned long addr, pfn_t pfn, bool mkwrite)
2421{
2422 pgprot_t pgprot = vma->vm_page_prot;
2423 int err;
2424
2425 BUG_ON(!vm_mixed_ok(vma, pfn));
2426
2427 if (addr < vma->vm_start || addr >= vma->vm_end)
2428 return VM_FAULT_SIGBUS;
2429
2430 track_pfn_insert(vma, &pgprot, pfn);
2431
2432 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2433 return VM_FAULT_SIGBUS;
2434
2435 /*
2436 * If we don't have pte special, then we have to use the pfn_valid()
2437 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2438 * refcount the page if pfn_valid is true (hence insert_page rather
2439 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2440 * without pte special, it would there be refcounted as a normal page.
2441 */
2442 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2443 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2444 struct page *page;
2445
2446 /*
2447 * At this point we are committed to insert_page()
2448 * regardless of whether the caller specified flags that
2449 * result in pfn_t_has_page() == false.
2450 */
2451 page = pfn_to_page(pfn_t_to_pfn(pfn));
2452 err = insert_page(vma, addr, page, pgprot);
2453 } else {
2454 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2455 }
2456
2457 if (err == -ENOMEM)
2458 return VM_FAULT_OOM;
2459 if (err < 0 && err != -EBUSY)
2460 return VM_FAULT_SIGBUS;
2461
2462 return VM_FAULT_NOPAGE;
2463}
2464
2465vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2466 pfn_t pfn)
2467{
2468 return __vm_insert_mixed(vma, addr, pfn, false);
2469}
2470EXPORT_SYMBOL(vmf_insert_mixed);
2471
2472/*
2473 * If the insertion of PTE failed because someone else already added a
2474 * different entry in the mean time, we treat that as success as we assume
2475 * the same entry was actually inserted.
2476 */
2477vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2478 unsigned long addr, pfn_t pfn)
2479{
2480 return __vm_insert_mixed(vma, addr, pfn, true);
2481}
2482EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2483
2484/*
2485 * maps a range of physical memory into the requested pages. the old
2486 * mappings are removed. any references to nonexistent pages results
2487 * in null mappings (currently treated as "copy-on-access")
2488 */
2489static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2490 unsigned long addr, unsigned long end,
2491 unsigned long pfn, pgprot_t prot)
2492{
2493 pte_t *pte, *mapped_pte;
2494 spinlock_t *ptl;
2495 int err = 0;
2496
2497 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2498 if (!pte)
2499 return -ENOMEM;
2500 arch_enter_lazy_mmu_mode();
2501 do {
2502 BUG_ON(!pte_none(ptep_get(pte)));
2503 if (!pfn_modify_allowed(pfn, prot)) {
2504 err = -EACCES;
2505 break;
2506 }
2507 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2508 pfn++;
2509 } while (pte++, addr += PAGE_SIZE, addr != end);
2510 arch_leave_lazy_mmu_mode();
2511 pte_unmap_unlock(mapped_pte, ptl);
2512 return err;
2513}
2514
2515static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2516 unsigned long addr, unsigned long end,
2517 unsigned long pfn, pgprot_t prot)
2518{
2519 pmd_t *pmd;
2520 unsigned long next;
2521 int err;
2522
2523 pfn -= addr >> PAGE_SHIFT;
2524 pmd = pmd_alloc(mm, pud, addr);
2525 if (!pmd)
2526 return -ENOMEM;
2527 VM_BUG_ON(pmd_trans_huge(*pmd));
2528 do {
2529 next = pmd_addr_end(addr, end);
2530 err = remap_pte_range(mm, pmd, addr, next,
2531 pfn + (addr >> PAGE_SHIFT), prot);
2532 if (err)
2533 return err;
2534 } while (pmd++, addr = next, addr != end);
2535 return 0;
2536}
2537
2538static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2539 unsigned long addr, unsigned long end,
2540 unsigned long pfn, pgprot_t prot)
2541{
2542 pud_t *pud;
2543 unsigned long next;
2544 int err;
2545
2546 pfn -= addr >> PAGE_SHIFT;
2547 pud = pud_alloc(mm, p4d, addr);
2548 if (!pud)
2549 return -ENOMEM;
2550 do {
2551 next = pud_addr_end(addr, end);
2552 err = remap_pmd_range(mm, pud, addr, next,
2553 pfn + (addr >> PAGE_SHIFT), prot);
2554 if (err)
2555 return err;
2556 } while (pud++, addr = next, addr != end);
2557 return 0;
2558}
2559
2560static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2561 unsigned long addr, unsigned long end,
2562 unsigned long pfn, pgprot_t prot)
2563{
2564 p4d_t *p4d;
2565 unsigned long next;
2566 int err;
2567
2568 pfn -= addr >> PAGE_SHIFT;
2569 p4d = p4d_alloc(mm, pgd, addr);
2570 if (!p4d)
2571 return -ENOMEM;
2572 do {
2573 next = p4d_addr_end(addr, end);
2574 err = remap_pud_range(mm, p4d, addr, next,
2575 pfn + (addr >> PAGE_SHIFT), prot);
2576 if (err)
2577 return err;
2578 } while (p4d++, addr = next, addr != end);
2579 return 0;
2580}
2581
2582/*
2583 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2584 * must have pre-validated the caching bits of the pgprot_t.
2585 */
2586int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2587 unsigned long pfn, unsigned long size, pgprot_t prot)
2588{
2589 pgd_t *pgd;
2590 unsigned long next;
2591 unsigned long end = addr + PAGE_ALIGN(size);
2592 struct mm_struct *mm = vma->vm_mm;
2593 int err;
2594
2595 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2596 return -EINVAL;
2597
2598 /*
2599 * Physically remapped pages are special. Tell the
2600 * rest of the world about it:
2601 * VM_IO tells people not to look at these pages
2602 * (accesses can have side effects).
2603 * VM_PFNMAP tells the core MM that the base pages are just
2604 * raw PFN mappings, and do not have a "struct page" associated
2605 * with them.
2606 * VM_DONTEXPAND
2607 * Disable vma merging and expanding with mremap().
2608 * VM_DONTDUMP
2609 * Omit vma from core dump, even when VM_IO turned off.
2610 *
2611 * There's a horrible special case to handle copy-on-write
2612 * behaviour that some programs depend on. We mark the "original"
2613 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2614 * See vm_normal_page() for details.
2615 */
2616 if (is_cow_mapping(vma->vm_flags)) {
2617 if (addr != vma->vm_start || end != vma->vm_end)
2618 return -EINVAL;
2619 vma->vm_pgoff = pfn;
2620 }
2621
2622 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
2623
2624 BUG_ON(addr >= end);
2625 pfn -= addr >> PAGE_SHIFT;
2626 pgd = pgd_offset(mm, addr);
2627 flush_cache_range(vma, addr, end);
2628 do {
2629 next = pgd_addr_end(addr, end);
2630 err = remap_p4d_range(mm, pgd, addr, next,
2631 pfn + (addr >> PAGE_SHIFT), prot);
2632 if (err)
2633 return err;
2634 } while (pgd++, addr = next, addr != end);
2635
2636 return 0;
2637}
2638
2639/**
2640 * remap_pfn_range - remap kernel memory to userspace
2641 * @vma: user vma to map to
2642 * @addr: target page aligned user address to start at
2643 * @pfn: page frame number of kernel physical memory address
2644 * @size: size of mapping area
2645 * @prot: page protection flags for this mapping
2646 *
2647 * Note: this is only safe if the mm semaphore is held when called.
2648 *
2649 * Return: %0 on success, negative error code otherwise.
2650 */
2651int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2652 unsigned long pfn, unsigned long size, pgprot_t prot)
2653{
2654 int err;
2655
2656 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2657 if (err)
2658 return -EINVAL;
2659
2660 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2661 if (err)
2662 untrack_pfn(vma, pfn, PAGE_ALIGN(size), true);
2663 return err;
2664}
2665EXPORT_SYMBOL(remap_pfn_range);
2666
2667/**
2668 * vm_iomap_memory - remap memory to userspace
2669 * @vma: user vma to map to
2670 * @start: start of the physical memory to be mapped
2671 * @len: size of area
2672 *
2673 * This is a simplified io_remap_pfn_range() for common driver use. The
2674 * driver just needs to give us the physical memory range to be mapped,
2675 * we'll figure out the rest from the vma information.
2676 *
2677 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2678 * whatever write-combining details or similar.
2679 *
2680 * Return: %0 on success, negative error code otherwise.
2681 */
2682int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2683{
2684 unsigned long vm_len, pfn, pages;
2685
2686 /* Check that the physical memory area passed in looks valid */
2687 if (start + len < start)
2688 return -EINVAL;
2689 /*
2690 * You *really* shouldn't map things that aren't page-aligned,
2691 * but we've historically allowed it because IO memory might
2692 * just have smaller alignment.
2693 */
2694 len += start & ~PAGE_MASK;
2695 pfn = start >> PAGE_SHIFT;
2696 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2697 if (pfn + pages < pfn)
2698 return -EINVAL;
2699
2700 /* We start the mapping 'vm_pgoff' pages into the area */
2701 if (vma->vm_pgoff > pages)
2702 return -EINVAL;
2703 pfn += vma->vm_pgoff;
2704 pages -= vma->vm_pgoff;
2705
2706 /* Can we fit all of the mapping? */
2707 vm_len = vma->vm_end - vma->vm_start;
2708 if (vm_len >> PAGE_SHIFT > pages)
2709 return -EINVAL;
2710
2711 /* Ok, let it rip */
2712 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2713}
2714EXPORT_SYMBOL(vm_iomap_memory);
2715
2716static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2717 unsigned long addr, unsigned long end,
2718 pte_fn_t fn, void *data, bool create,
2719 pgtbl_mod_mask *mask)
2720{
2721 pte_t *pte, *mapped_pte;
2722 int err = 0;
2723 spinlock_t *ptl;
2724
2725 if (create) {
2726 mapped_pte = pte = (mm == &init_mm) ?
2727 pte_alloc_kernel_track(pmd, addr, mask) :
2728 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2729 if (!pte)
2730 return -ENOMEM;
2731 } else {
2732 mapped_pte = pte = (mm == &init_mm) ?
2733 pte_offset_kernel(pmd, addr) :
2734 pte_offset_map_lock(mm, pmd, addr, &ptl);
2735 if (!pte)
2736 return -EINVAL;
2737 }
2738
2739 arch_enter_lazy_mmu_mode();
2740
2741 if (fn) {
2742 do {
2743 if (create || !pte_none(ptep_get(pte))) {
2744 err = fn(pte++, addr, data);
2745 if (err)
2746 break;
2747 }
2748 } while (addr += PAGE_SIZE, addr != end);
2749 }
2750 *mask |= PGTBL_PTE_MODIFIED;
2751
2752 arch_leave_lazy_mmu_mode();
2753
2754 if (mm != &init_mm)
2755 pte_unmap_unlock(mapped_pte, ptl);
2756 return err;
2757}
2758
2759static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2760 unsigned long addr, unsigned long end,
2761 pte_fn_t fn, void *data, bool create,
2762 pgtbl_mod_mask *mask)
2763{
2764 pmd_t *pmd;
2765 unsigned long next;
2766 int err = 0;
2767
2768 BUG_ON(pud_huge(*pud));
2769
2770 if (create) {
2771 pmd = pmd_alloc_track(mm, pud, addr, mask);
2772 if (!pmd)
2773 return -ENOMEM;
2774 } else {
2775 pmd = pmd_offset(pud, addr);
2776 }
2777 do {
2778 next = pmd_addr_end(addr, end);
2779 if (pmd_none(*pmd) && !create)
2780 continue;
2781 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2782 return -EINVAL;
2783 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2784 if (!create)
2785 continue;
2786 pmd_clear_bad(pmd);
2787 }
2788 err = apply_to_pte_range(mm, pmd, addr, next,
2789 fn, data, create, mask);
2790 if (err)
2791 break;
2792 } while (pmd++, addr = next, addr != end);
2793
2794 return err;
2795}
2796
2797static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2798 unsigned long addr, unsigned long end,
2799 pte_fn_t fn, void *data, bool create,
2800 pgtbl_mod_mask *mask)
2801{
2802 pud_t *pud;
2803 unsigned long next;
2804 int err = 0;
2805
2806 if (create) {
2807 pud = pud_alloc_track(mm, p4d, addr, mask);
2808 if (!pud)
2809 return -ENOMEM;
2810 } else {
2811 pud = pud_offset(p4d, addr);
2812 }
2813 do {
2814 next = pud_addr_end(addr, end);
2815 if (pud_none(*pud) && !create)
2816 continue;
2817 if (WARN_ON_ONCE(pud_leaf(*pud)))
2818 return -EINVAL;
2819 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2820 if (!create)
2821 continue;
2822 pud_clear_bad(pud);
2823 }
2824 err = apply_to_pmd_range(mm, pud, addr, next,
2825 fn, data, create, mask);
2826 if (err)
2827 break;
2828 } while (pud++, addr = next, addr != end);
2829
2830 return err;
2831}
2832
2833static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2834 unsigned long addr, unsigned long end,
2835 pte_fn_t fn, void *data, bool create,
2836 pgtbl_mod_mask *mask)
2837{
2838 p4d_t *p4d;
2839 unsigned long next;
2840 int err = 0;
2841
2842 if (create) {
2843 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2844 if (!p4d)
2845 return -ENOMEM;
2846 } else {
2847 p4d = p4d_offset(pgd, addr);
2848 }
2849 do {
2850 next = p4d_addr_end(addr, end);
2851 if (p4d_none(*p4d) && !create)
2852 continue;
2853 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2854 return -EINVAL;
2855 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2856 if (!create)
2857 continue;
2858 p4d_clear_bad(p4d);
2859 }
2860 err = apply_to_pud_range(mm, p4d, addr, next,
2861 fn, data, create, mask);
2862 if (err)
2863 break;
2864 } while (p4d++, addr = next, addr != end);
2865
2866 return err;
2867}
2868
2869static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2870 unsigned long size, pte_fn_t fn,
2871 void *data, bool create)
2872{
2873 pgd_t *pgd;
2874 unsigned long start = addr, next;
2875 unsigned long end = addr + size;
2876 pgtbl_mod_mask mask = 0;
2877 int err = 0;
2878
2879 if (WARN_ON(addr >= end))
2880 return -EINVAL;
2881
2882 pgd = pgd_offset(mm, addr);
2883 do {
2884 next = pgd_addr_end(addr, end);
2885 if (pgd_none(*pgd) && !create)
2886 continue;
2887 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2888 return -EINVAL;
2889 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2890 if (!create)
2891 continue;
2892 pgd_clear_bad(pgd);
2893 }
2894 err = apply_to_p4d_range(mm, pgd, addr, next,
2895 fn, data, create, &mask);
2896 if (err)
2897 break;
2898 } while (pgd++, addr = next, addr != end);
2899
2900 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2901 arch_sync_kernel_mappings(start, start + size);
2902
2903 return err;
2904}
2905
2906/*
2907 * Scan a region of virtual memory, filling in page tables as necessary
2908 * and calling a provided function on each leaf page table.
2909 */
2910int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2911 unsigned long size, pte_fn_t fn, void *data)
2912{
2913 return __apply_to_page_range(mm, addr, size, fn, data, true);
2914}
2915EXPORT_SYMBOL_GPL(apply_to_page_range);
2916
2917/*
2918 * Scan a region of virtual memory, calling a provided function on
2919 * each leaf page table where it exists.
2920 *
2921 * Unlike apply_to_page_range, this does _not_ fill in page tables
2922 * where they are absent.
2923 */
2924int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2925 unsigned long size, pte_fn_t fn, void *data)
2926{
2927 return __apply_to_page_range(mm, addr, size, fn, data, false);
2928}
2929EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2930
2931/*
2932 * handle_pte_fault chooses page fault handler according to an entry which was
2933 * read non-atomically. Before making any commitment, on those architectures
2934 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2935 * parts, do_swap_page must check under lock before unmapping the pte and
2936 * proceeding (but do_wp_page is only called after already making such a check;
2937 * and do_anonymous_page can safely check later on).
2938 */
2939static inline int pte_unmap_same(struct vm_fault *vmf)
2940{
2941 int same = 1;
2942#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2943 if (sizeof(pte_t) > sizeof(unsigned long)) {
2944 spin_lock(vmf->ptl);
2945 same = pte_same(ptep_get(vmf->pte), vmf->orig_pte);
2946 spin_unlock(vmf->ptl);
2947 }
2948#endif
2949 pte_unmap(vmf->pte);
2950 vmf->pte = NULL;
2951 return same;
2952}
2953
2954/*
2955 * Return:
2956 * 0: copied succeeded
2957 * -EHWPOISON: copy failed due to hwpoison in source page
2958 * -EAGAIN: copied failed (some other reason)
2959 */
2960static inline int __wp_page_copy_user(struct page *dst, struct page *src,
2961 struct vm_fault *vmf)
2962{
2963 int ret;
2964 void *kaddr;
2965 void __user *uaddr;
2966 struct vm_area_struct *vma = vmf->vma;
2967 struct mm_struct *mm = vma->vm_mm;
2968 unsigned long addr = vmf->address;
2969
2970 if (likely(src)) {
2971 if (copy_mc_user_highpage(dst, src, addr, vma)) {
2972 memory_failure_queue(page_to_pfn(src), 0);
2973 return -EHWPOISON;
2974 }
2975 return 0;
2976 }
2977
2978 /*
2979 * If the source page was a PFN mapping, we don't have
2980 * a "struct page" for it. We do a best-effort copy by
2981 * just copying from the original user address. If that
2982 * fails, we just zero-fill it. Live with it.
2983 */
2984 kaddr = kmap_local_page(dst);
2985 pagefault_disable();
2986 uaddr = (void __user *)(addr & PAGE_MASK);
2987
2988 /*
2989 * On architectures with software "accessed" bits, we would
2990 * take a double page fault, so mark it accessed here.
2991 */
2992 vmf->pte = NULL;
2993 if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) {
2994 pte_t entry;
2995
2996 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2997 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2998 /*
2999 * Other thread has already handled the fault
3000 * and update local tlb only
3001 */
3002 if (vmf->pte)
3003 update_mmu_tlb(vma, addr, vmf->pte);
3004 ret = -EAGAIN;
3005 goto pte_unlock;
3006 }
3007
3008 entry = pte_mkyoung(vmf->orig_pte);
3009 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
3010 update_mmu_cache_range(vmf, vma, addr, vmf->pte, 1);
3011 }
3012
3013 /*
3014 * This really shouldn't fail, because the page is there
3015 * in the page tables. But it might just be unreadable,
3016 * in which case we just give up and fill the result with
3017 * zeroes.
3018 */
3019 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
3020 if (vmf->pte)
3021 goto warn;
3022
3023 /* Re-validate under PTL if the page is still mapped */
3024 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
3025 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3026 /* The PTE changed under us, update local tlb */
3027 if (vmf->pte)
3028 update_mmu_tlb(vma, addr, vmf->pte);
3029 ret = -EAGAIN;
3030 goto pte_unlock;
3031 }
3032
3033 /*
3034 * The same page can be mapped back since last copy attempt.
3035 * Try to copy again under PTL.
3036 */
3037 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
3038 /*
3039 * Give a warn in case there can be some obscure
3040 * use-case
3041 */
3042warn:
3043 WARN_ON_ONCE(1);
3044 clear_page(kaddr);
3045 }
3046 }
3047
3048 ret = 0;
3049
3050pte_unlock:
3051 if (vmf->pte)
3052 pte_unmap_unlock(vmf->pte, vmf->ptl);
3053 pagefault_enable();
3054 kunmap_local(kaddr);
3055 flush_dcache_page(dst);
3056
3057 return ret;
3058}
3059
3060static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
3061{
3062 struct file *vm_file = vma->vm_file;
3063
3064 if (vm_file)
3065 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
3066
3067 /*
3068 * Special mappings (e.g. VDSO) do not have any file so fake
3069 * a default GFP_KERNEL for them.
3070 */
3071 return GFP_KERNEL;
3072}
3073
3074/*
3075 * Notify the address space that the page is about to become writable so that
3076 * it can prohibit this or wait for the page to get into an appropriate state.
3077 *
3078 * We do this without the lock held, so that it can sleep if it needs to.
3079 */
3080static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio)
3081{
3082 vm_fault_t ret;
3083 unsigned int old_flags = vmf->flags;
3084
3085 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3086
3087 if (vmf->vma->vm_file &&
3088 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
3089 return VM_FAULT_SIGBUS;
3090
3091 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
3092 /* Restore original flags so that caller is not surprised */
3093 vmf->flags = old_flags;
3094 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
3095 return ret;
3096 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
3097 folio_lock(folio);
3098 if (!folio->mapping) {
3099 folio_unlock(folio);
3100 return 0; /* retry */
3101 }
3102 ret |= VM_FAULT_LOCKED;
3103 } else
3104 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3105 return ret;
3106}
3107
3108/*
3109 * Handle dirtying of a page in shared file mapping on a write fault.
3110 *
3111 * The function expects the page to be locked and unlocks it.
3112 */
3113static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
3114{
3115 struct vm_area_struct *vma = vmf->vma;
3116 struct address_space *mapping;
3117 struct folio *folio = page_folio(vmf->page);
3118 bool dirtied;
3119 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
3120
3121 dirtied = folio_mark_dirty(folio);
3122 VM_BUG_ON_FOLIO(folio_test_anon(folio), folio);
3123 /*
3124 * Take a local copy of the address_space - folio.mapping may be zeroed
3125 * by truncate after folio_unlock(). The address_space itself remains
3126 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
3127 * release semantics to prevent the compiler from undoing this copying.
3128 */
3129 mapping = folio_raw_mapping(folio);
3130 folio_unlock(folio);
3131
3132 if (!page_mkwrite)
3133 file_update_time(vma->vm_file);
3134
3135 /*
3136 * Throttle page dirtying rate down to writeback speed.
3137 *
3138 * mapping may be NULL here because some device drivers do not
3139 * set page.mapping but still dirty their pages
3140 *
3141 * Drop the mmap_lock before waiting on IO, if we can. The file
3142 * is pinning the mapping, as per above.
3143 */
3144 if ((dirtied || page_mkwrite) && mapping) {
3145 struct file *fpin;
3146
3147 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
3148 balance_dirty_pages_ratelimited(mapping);
3149 if (fpin) {
3150 fput(fpin);
3151 return VM_FAULT_COMPLETED;
3152 }
3153 }
3154
3155 return 0;
3156}
3157
3158/*
3159 * Handle write page faults for pages that can be reused in the current vma
3160 *
3161 * This can happen either due to the mapping being with the VM_SHARED flag,
3162 * or due to us being the last reference standing to the page. In either
3163 * case, all we need to do here is to mark the page as writable and update
3164 * any related book-keeping.
3165 */
3166static inline void wp_page_reuse(struct vm_fault *vmf, struct folio *folio)
3167 __releases(vmf->ptl)
3168{
3169 struct vm_area_struct *vma = vmf->vma;
3170 pte_t entry;
3171
3172 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3173
3174 if (folio) {
3175 VM_BUG_ON(folio_test_anon(folio) &&
3176 !PageAnonExclusive(vmf->page));
3177 /*
3178 * Clear the folio's cpupid information as the existing
3179 * information potentially belongs to a now completely
3180 * unrelated process.
3181 */
3182 folio_xchg_last_cpupid(folio, (1 << LAST_CPUPID_SHIFT) - 1);
3183 }
3184
3185 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3186 entry = pte_mkyoung(vmf->orig_pte);
3187 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3188 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
3189 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3190 pte_unmap_unlock(vmf->pte, vmf->ptl);
3191 count_vm_event(PGREUSE);
3192}
3193
3194/*
3195 * We could add a bitflag somewhere, but for now, we know that all
3196 * vm_ops that have a ->map_pages have been audited and don't need
3197 * the mmap_lock to be held.
3198 */
3199static inline vm_fault_t vmf_can_call_fault(const struct vm_fault *vmf)
3200{
3201 struct vm_area_struct *vma = vmf->vma;
3202
3203 if (vma->vm_ops->map_pages || !(vmf->flags & FAULT_FLAG_VMA_LOCK))
3204 return 0;
3205 vma_end_read(vma);
3206 return VM_FAULT_RETRY;
3207}
3208
3209vm_fault_t vmf_anon_prepare(struct vm_fault *vmf)
3210{
3211 struct vm_area_struct *vma = vmf->vma;
3212
3213 if (likely(vma->anon_vma))
3214 return 0;
3215 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3216 vma_end_read(vma);
3217 return VM_FAULT_RETRY;
3218 }
3219 if (__anon_vma_prepare(vma))
3220 return VM_FAULT_OOM;
3221 return 0;
3222}
3223
3224/*
3225 * Handle the case of a page which we actually need to copy to a new page,
3226 * either due to COW or unsharing.
3227 *
3228 * Called with mmap_lock locked and the old page referenced, but
3229 * without the ptl held.
3230 *
3231 * High level logic flow:
3232 *
3233 * - Allocate a page, copy the content of the old page to the new one.
3234 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3235 * - Take the PTL. If the pte changed, bail out and release the allocated page
3236 * - If the pte is still the way we remember it, update the page table and all
3237 * relevant references. This includes dropping the reference the page-table
3238 * held to the old page, as well as updating the rmap.
3239 * - In any case, unlock the PTL and drop the reference we took to the old page.
3240 */
3241static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3242{
3243 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3244 struct vm_area_struct *vma = vmf->vma;
3245 struct mm_struct *mm = vma->vm_mm;
3246 struct folio *old_folio = NULL;
3247 struct folio *new_folio = NULL;
3248 pte_t entry;
3249 int page_copied = 0;
3250 struct mmu_notifier_range range;
3251 vm_fault_t ret;
3252 bool pfn_is_zero;
3253
3254 delayacct_wpcopy_start();
3255
3256 if (vmf->page)
3257 old_folio = page_folio(vmf->page);
3258 ret = vmf_anon_prepare(vmf);
3259 if (unlikely(ret))
3260 goto out;
3261
3262 pfn_is_zero = is_zero_pfn(pte_pfn(vmf->orig_pte));
3263 new_folio = folio_prealloc(mm, vma, vmf->address, pfn_is_zero);
3264 if (!new_folio)
3265 goto oom;
3266
3267 if (!pfn_is_zero) {
3268 int err;
3269
3270 err = __wp_page_copy_user(&new_folio->page, vmf->page, vmf);
3271 if (err) {
3272 /*
3273 * COW failed, if the fault was solved by other,
3274 * it's fine. If not, userspace would re-fault on
3275 * the same address and we will handle the fault
3276 * from the second attempt.
3277 * The -EHWPOISON case will not be retried.
3278 */
3279 folio_put(new_folio);
3280 if (old_folio)
3281 folio_put(old_folio);
3282
3283 delayacct_wpcopy_end();
3284 return err == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3285 }
3286 kmsan_copy_page_meta(&new_folio->page, vmf->page);
3287 }
3288
3289 __folio_mark_uptodate(new_folio);
3290
3291 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
3292 vmf->address & PAGE_MASK,
3293 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3294 mmu_notifier_invalidate_range_start(&range);
3295
3296 /*
3297 * Re-check the pte - we dropped the lock
3298 */
3299 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3300 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3301 if (old_folio) {
3302 if (!folio_test_anon(old_folio)) {
3303 dec_mm_counter(mm, mm_counter_file(old_folio));
3304 inc_mm_counter(mm, MM_ANONPAGES);
3305 }
3306 } else {
3307 ksm_might_unmap_zero_page(mm, vmf->orig_pte);
3308 inc_mm_counter(mm, MM_ANONPAGES);
3309 }
3310 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3311 entry = mk_pte(&new_folio->page, vma->vm_page_prot);
3312 entry = pte_sw_mkyoung(entry);
3313 if (unlikely(unshare)) {
3314 if (pte_soft_dirty(vmf->orig_pte))
3315 entry = pte_mksoft_dirty(entry);
3316 if (pte_uffd_wp(vmf->orig_pte))
3317 entry = pte_mkuffd_wp(entry);
3318 } else {
3319 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3320 }
3321
3322 /*
3323 * Clear the pte entry and flush it first, before updating the
3324 * pte with the new entry, to keep TLBs on different CPUs in
3325 * sync. This code used to set the new PTE then flush TLBs, but
3326 * that left a window where the new PTE could be loaded into
3327 * some TLBs while the old PTE remains in others.
3328 */
3329 ptep_clear_flush(vma, vmf->address, vmf->pte);
3330 folio_add_new_anon_rmap(new_folio, vma, vmf->address);
3331 folio_add_lru_vma(new_folio, vma);
3332 /*
3333 * We call the notify macro here because, when using secondary
3334 * mmu page tables (such as kvm shadow page tables), we want the
3335 * new page to be mapped directly into the secondary page table.
3336 */
3337 BUG_ON(unshare && pte_write(entry));
3338 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3339 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3340 if (old_folio) {
3341 /*
3342 * Only after switching the pte to the new page may
3343 * we remove the mapcount here. Otherwise another
3344 * process may come and find the rmap count decremented
3345 * before the pte is switched to the new page, and
3346 * "reuse" the old page writing into it while our pte
3347 * here still points into it and can be read by other
3348 * threads.
3349 *
3350 * The critical issue is to order this
3351 * folio_remove_rmap_pte() with the ptp_clear_flush
3352 * above. Those stores are ordered by (if nothing else,)
3353 * the barrier present in the atomic_add_negative
3354 * in folio_remove_rmap_pte();
3355 *
3356 * Then the TLB flush in ptep_clear_flush ensures that
3357 * no process can access the old page before the
3358 * decremented mapcount is visible. And the old page
3359 * cannot be reused until after the decremented
3360 * mapcount is visible. So transitively, TLBs to
3361 * old page will be flushed before it can be reused.
3362 */
3363 folio_remove_rmap_pte(old_folio, vmf->page, vma);
3364 }
3365
3366 /* Free the old page.. */
3367 new_folio = old_folio;
3368 page_copied = 1;
3369 pte_unmap_unlock(vmf->pte, vmf->ptl);
3370 } else if (vmf->pte) {
3371 update_mmu_tlb(vma, vmf->address, vmf->pte);
3372 pte_unmap_unlock(vmf->pte, vmf->ptl);
3373 }
3374
3375 mmu_notifier_invalidate_range_end(&range);
3376
3377 if (new_folio)
3378 folio_put(new_folio);
3379 if (old_folio) {
3380 if (page_copied)
3381 free_swap_cache(old_folio);
3382 folio_put(old_folio);
3383 }
3384
3385 delayacct_wpcopy_end();
3386 return 0;
3387oom:
3388 ret = VM_FAULT_OOM;
3389out:
3390 if (old_folio)
3391 folio_put(old_folio);
3392
3393 delayacct_wpcopy_end();
3394 return ret;
3395}
3396
3397/**
3398 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3399 * writeable once the page is prepared
3400 *
3401 * @vmf: structure describing the fault
3402 * @folio: the folio of vmf->page
3403 *
3404 * This function handles all that is needed to finish a write page fault in a
3405 * shared mapping due to PTE being read-only once the mapped page is prepared.
3406 * It handles locking of PTE and modifying it.
3407 *
3408 * The function expects the page to be locked or other protection against
3409 * concurrent faults / writeback (such as DAX radix tree locks).
3410 *
3411 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3412 * we acquired PTE lock.
3413 */
3414static vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf, struct folio *folio)
3415{
3416 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3417 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3418 &vmf->ptl);
3419 if (!vmf->pte)
3420 return VM_FAULT_NOPAGE;
3421 /*
3422 * We might have raced with another page fault while we released the
3423 * pte_offset_map_lock.
3424 */
3425 if (!pte_same(ptep_get(vmf->pte), vmf->orig_pte)) {
3426 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3427 pte_unmap_unlock(vmf->pte, vmf->ptl);
3428 return VM_FAULT_NOPAGE;
3429 }
3430 wp_page_reuse(vmf, folio);
3431 return 0;
3432}
3433
3434/*
3435 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3436 * mapping
3437 */
3438static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3439{
3440 struct vm_area_struct *vma = vmf->vma;
3441
3442 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3443 vm_fault_t ret;
3444
3445 pte_unmap_unlock(vmf->pte, vmf->ptl);
3446 ret = vmf_can_call_fault(vmf);
3447 if (ret)
3448 return ret;
3449
3450 vmf->flags |= FAULT_FLAG_MKWRITE;
3451 ret = vma->vm_ops->pfn_mkwrite(vmf);
3452 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3453 return ret;
3454 return finish_mkwrite_fault(vmf, NULL);
3455 }
3456 wp_page_reuse(vmf, NULL);
3457 return 0;
3458}
3459
3460static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio)
3461 __releases(vmf->ptl)
3462{
3463 struct vm_area_struct *vma = vmf->vma;
3464 vm_fault_t ret = 0;
3465
3466 folio_get(folio);
3467
3468 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3469 vm_fault_t tmp;
3470
3471 pte_unmap_unlock(vmf->pte, vmf->ptl);
3472 tmp = vmf_can_call_fault(vmf);
3473 if (tmp) {
3474 folio_put(folio);
3475 return tmp;
3476 }
3477
3478 tmp = do_page_mkwrite(vmf, folio);
3479 if (unlikely(!tmp || (tmp &
3480 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3481 folio_put(folio);
3482 return tmp;
3483 }
3484 tmp = finish_mkwrite_fault(vmf, folio);
3485 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3486 folio_unlock(folio);
3487 folio_put(folio);
3488 return tmp;
3489 }
3490 } else {
3491 wp_page_reuse(vmf, folio);
3492 folio_lock(folio);
3493 }
3494 ret |= fault_dirty_shared_page(vmf);
3495 folio_put(folio);
3496
3497 return ret;
3498}
3499
3500static bool wp_can_reuse_anon_folio(struct folio *folio,
3501 struct vm_area_struct *vma)
3502{
3503 /*
3504 * We could currently only reuse a subpage of a large folio if no
3505 * other subpages of the large folios are still mapped. However,
3506 * let's just consistently not reuse subpages even if we could
3507 * reuse in that scenario, and give back a large folio a bit
3508 * sooner.
3509 */
3510 if (folio_test_large(folio))
3511 return false;
3512
3513 /*
3514 * We have to verify under folio lock: these early checks are
3515 * just an optimization to avoid locking the folio and freeing
3516 * the swapcache if there is little hope that we can reuse.
3517 *
3518 * KSM doesn't necessarily raise the folio refcount.
3519 */
3520 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3521 return false;
3522 if (!folio_test_lru(folio))
3523 /*
3524 * We cannot easily detect+handle references from
3525 * remote LRU caches or references to LRU folios.
3526 */
3527 lru_add_drain();
3528 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
3529 return false;
3530 if (!folio_trylock(folio))
3531 return false;
3532 if (folio_test_swapcache(folio))
3533 folio_free_swap(folio);
3534 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
3535 folio_unlock(folio);
3536 return false;
3537 }
3538 /*
3539 * Ok, we've got the only folio reference from our mapping
3540 * and the folio is locked, it's dark out, and we're wearing
3541 * sunglasses. Hit it.
3542 */
3543 folio_move_anon_rmap(folio, vma);
3544 folio_unlock(folio);
3545 return true;
3546}
3547
3548/*
3549 * This routine handles present pages, when
3550 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3551 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3552 * (FAULT_FLAG_UNSHARE)
3553 *
3554 * It is done by copying the page to a new address and decrementing the
3555 * shared-page counter for the old page.
3556 *
3557 * Note that this routine assumes that the protection checks have been
3558 * done by the caller (the low-level page fault routine in most cases).
3559 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3560 * done any necessary COW.
3561 *
3562 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3563 * though the page will change only once the write actually happens. This
3564 * avoids a few races, and potentially makes it more efficient.
3565 *
3566 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3567 * but allow concurrent faults), with pte both mapped and locked.
3568 * We return with mmap_lock still held, but pte unmapped and unlocked.
3569 */
3570static vm_fault_t do_wp_page(struct vm_fault *vmf)
3571 __releases(vmf->ptl)
3572{
3573 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3574 struct vm_area_struct *vma = vmf->vma;
3575 struct folio *folio = NULL;
3576 pte_t pte;
3577
3578 if (likely(!unshare)) {
3579 if (userfaultfd_pte_wp(vma, ptep_get(vmf->pte))) {
3580 if (!userfaultfd_wp_async(vma)) {
3581 pte_unmap_unlock(vmf->pte, vmf->ptl);
3582 return handle_userfault(vmf, VM_UFFD_WP);
3583 }
3584
3585 /*
3586 * Nothing needed (cache flush, TLB invalidations,
3587 * etc.) because we're only removing the uffd-wp bit,
3588 * which is completely invisible to the user.
3589 */
3590 pte = pte_clear_uffd_wp(ptep_get(vmf->pte));
3591
3592 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3593 /*
3594 * Update this to be prepared for following up CoW
3595 * handling
3596 */
3597 vmf->orig_pte = pte;
3598 }
3599
3600 /*
3601 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3602 * is flushed in this case before copying.
3603 */
3604 if (unlikely(userfaultfd_wp(vmf->vma) &&
3605 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3606 flush_tlb_page(vmf->vma, vmf->address);
3607 }
3608
3609 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3610
3611 if (vmf->page)
3612 folio = page_folio(vmf->page);
3613
3614 /*
3615 * Shared mapping: we are guaranteed to have VM_WRITE and
3616 * FAULT_FLAG_WRITE set at this point.
3617 */
3618 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
3619 /*
3620 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3621 * VM_PFNMAP VMA.
3622 *
3623 * We should not cow pages in a shared writeable mapping.
3624 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3625 */
3626 if (!vmf->page)
3627 return wp_pfn_shared(vmf);
3628 return wp_page_shared(vmf, folio);
3629 }
3630
3631 /*
3632 * Private mapping: create an exclusive anonymous page copy if reuse
3633 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3634 *
3635 * If we encounter a page that is marked exclusive, we must reuse
3636 * the page without further checks.
3637 */
3638 if (folio && folio_test_anon(folio) &&
3639 (PageAnonExclusive(vmf->page) || wp_can_reuse_anon_folio(folio, vma))) {
3640 if (!PageAnonExclusive(vmf->page))
3641 SetPageAnonExclusive(vmf->page);
3642 if (unlikely(unshare)) {
3643 pte_unmap_unlock(vmf->pte, vmf->ptl);
3644 return 0;
3645 }
3646 wp_page_reuse(vmf, folio);
3647 return 0;
3648 }
3649 /*
3650 * Ok, we need to copy. Oh, well..
3651 */
3652 if (folio)
3653 folio_get(folio);
3654
3655 pte_unmap_unlock(vmf->pte, vmf->ptl);
3656#ifdef CONFIG_KSM
3657 if (folio && folio_test_ksm(folio))
3658 count_vm_event(COW_KSM);
3659#endif
3660 return wp_page_copy(vmf);
3661}
3662
3663static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3664 unsigned long start_addr, unsigned long end_addr,
3665 struct zap_details *details)
3666{
3667 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3668}
3669
3670static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3671 pgoff_t first_index,
3672 pgoff_t last_index,
3673 struct zap_details *details)
3674{
3675 struct vm_area_struct *vma;
3676 pgoff_t vba, vea, zba, zea;
3677
3678 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3679 vba = vma->vm_pgoff;
3680 vea = vba + vma_pages(vma) - 1;
3681 zba = max(first_index, vba);
3682 zea = min(last_index, vea);
3683
3684 unmap_mapping_range_vma(vma,
3685 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3686 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3687 details);
3688 }
3689}
3690
3691/**
3692 * unmap_mapping_folio() - Unmap single folio from processes.
3693 * @folio: The locked folio to be unmapped.
3694 *
3695 * Unmap this folio from any userspace process which still has it mmaped.
3696 * Typically, for efficiency, the range of nearby pages has already been
3697 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3698 * truncation or invalidation holds the lock on a folio, it may find that
3699 * the page has been remapped again: and then uses unmap_mapping_folio()
3700 * to unmap it finally.
3701 */
3702void unmap_mapping_folio(struct folio *folio)
3703{
3704 struct address_space *mapping = folio->mapping;
3705 struct zap_details details = { };
3706 pgoff_t first_index;
3707 pgoff_t last_index;
3708
3709 VM_BUG_ON(!folio_test_locked(folio));
3710
3711 first_index = folio->index;
3712 last_index = folio_next_index(folio) - 1;
3713
3714 details.even_cows = false;
3715 details.single_folio = folio;
3716 details.zap_flags = ZAP_FLAG_DROP_MARKER;
3717
3718 i_mmap_lock_read(mapping);
3719 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3720 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3721 last_index, &details);
3722 i_mmap_unlock_read(mapping);
3723}
3724
3725/**
3726 * unmap_mapping_pages() - Unmap pages from processes.
3727 * @mapping: The address space containing pages to be unmapped.
3728 * @start: Index of first page to be unmapped.
3729 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3730 * @even_cows: Whether to unmap even private COWed pages.
3731 *
3732 * Unmap the pages in this address space from any userspace process which
3733 * has them mmaped. Generally, you want to remove COWed pages as well when
3734 * a file is being truncated, but not when invalidating pages from the page
3735 * cache.
3736 */
3737void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3738 pgoff_t nr, bool even_cows)
3739{
3740 struct zap_details details = { };
3741 pgoff_t first_index = start;
3742 pgoff_t last_index = start + nr - 1;
3743
3744 details.even_cows = even_cows;
3745 if (last_index < first_index)
3746 last_index = ULONG_MAX;
3747
3748 i_mmap_lock_read(mapping);
3749 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3750 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3751 last_index, &details);
3752 i_mmap_unlock_read(mapping);
3753}
3754EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3755
3756/**
3757 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3758 * address_space corresponding to the specified byte range in the underlying
3759 * file.
3760 *
3761 * @mapping: the address space containing mmaps to be unmapped.
3762 * @holebegin: byte in first page to unmap, relative to the start of
3763 * the underlying file. This will be rounded down to a PAGE_SIZE
3764 * boundary. Note that this is different from truncate_pagecache(), which
3765 * must keep the partial page. In contrast, we must get rid of
3766 * partial pages.
3767 * @holelen: size of prospective hole in bytes. This will be rounded
3768 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3769 * end of the file.
3770 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3771 * but 0 when invalidating pagecache, don't throw away private data.
3772 */
3773void unmap_mapping_range(struct address_space *mapping,
3774 loff_t const holebegin, loff_t const holelen, int even_cows)
3775{
3776 pgoff_t hba = (pgoff_t)(holebegin) >> PAGE_SHIFT;
3777 pgoff_t hlen = ((pgoff_t)(holelen) + PAGE_SIZE - 1) >> PAGE_SHIFT;
3778
3779 /* Check for overflow. */
3780 if (sizeof(holelen) > sizeof(hlen)) {
3781 long long holeend =
3782 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3783 if (holeend & ~(long long)ULONG_MAX)
3784 hlen = ULONG_MAX - hba + 1;
3785 }
3786
3787 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3788}
3789EXPORT_SYMBOL(unmap_mapping_range);
3790
3791/*
3792 * Restore a potential device exclusive pte to a working pte entry
3793 */
3794static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3795{
3796 struct folio *folio = page_folio(vmf->page);
3797 struct vm_area_struct *vma = vmf->vma;
3798 struct mmu_notifier_range range;
3799 vm_fault_t ret;
3800
3801 /*
3802 * We need a reference to lock the folio because we don't hold
3803 * the PTL so a racing thread can remove the device-exclusive
3804 * entry and unmap it. If the folio is free the entry must
3805 * have been removed already. If it happens to have already
3806 * been re-allocated after being freed all we do is lock and
3807 * unlock it.
3808 */
3809 if (!folio_try_get(folio))
3810 return 0;
3811
3812 ret = folio_lock_or_retry(folio, vmf);
3813 if (ret) {
3814 folio_put(folio);
3815 return ret;
3816 }
3817 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
3818 vma->vm_mm, vmf->address & PAGE_MASK,
3819 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3820 mmu_notifier_invalidate_range_start(&range);
3821
3822 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3823 &vmf->ptl);
3824 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3825 restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte);
3826
3827 if (vmf->pte)
3828 pte_unmap_unlock(vmf->pte, vmf->ptl);
3829 folio_unlock(folio);
3830 folio_put(folio);
3831
3832 mmu_notifier_invalidate_range_end(&range);
3833 return 0;
3834}
3835
3836static inline bool should_try_to_free_swap(struct folio *folio,
3837 struct vm_area_struct *vma,
3838 unsigned int fault_flags)
3839{
3840 if (!folio_test_swapcache(folio))
3841 return false;
3842 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
3843 folio_test_mlocked(folio))
3844 return true;
3845 /*
3846 * If we want to map a page that's in the swapcache writable, we
3847 * have to detect via the refcount if we're really the exclusive
3848 * user. Try freeing the swapcache to get rid of the swapcache
3849 * reference only in case it's likely that we'll be the exlusive user.
3850 */
3851 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
3852 folio_ref_count(folio) == 2;
3853}
3854
3855static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
3856{
3857 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
3858 vmf->address, &vmf->ptl);
3859 if (!vmf->pte)
3860 return 0;
3861 /*
3862 * Be careful so that we will only recover a special uffd-wp pte into a
3863 * none pte. Otherwise it means the pte could have changed, so retry.
3864 *
3865 * This should also cover the case where e.g. the pte changed
3866 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3867 * So is_pte_marker() check is not enough to safely drop the pte.
3868 */
3869 if (pte_same(vmf->orig_pte, ptep_get(vmf->pte)))
3870 pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
3871 pte_unmap_unlock(vmf->pte, vmf->ptl);
3872 return 0;
3873}
3874
3875static vm_fault_t do_pte_missing(struct vm_fault *vmf)
3876{
3877 if (vma_is_anonymous(vmf->vma))
3878 return do_anonymous_page(vmf);
3879 else
3880 return do_fault(vmf);
3881}
3882
3883/*
3884 * This is actually a page-missing access, but with uffd-wp special pte
3885 * installed. It means this pte was wr-protected before being unmapped.
3886 */
3887static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
3888{
3889 /*
3890 * Just in case there're leftover special ptes even after the region
3891 * got unregistered - we can simply clear them.
3892 */
3893 if (unlikely(!userfaultfd_wp(vmf->vma)))
3894 return pte_marker_clear(vmf);
3895
3896 return do_pte_missing(vmf);
3897}
3898
3899static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
3900{
3901 swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte);
3902 unsigned long marker = pte_marker_get(entry);
3903
3904 /*
3905 * PTE markers should never be empty. If anything weird happened,
3906 * the best thing to do is to kill the process along with its mm.
3907 */
3908 if (WARN_ON_ONCE(!marker))
3909 return VM_FAULT_SIGBUS;
3910
3911 /* Higher priority than uffd-wp when data corrupted */
3912 if (marker & PTE_MARKER_POISONED)
3913 return VM_FAULT_HWPOISON;
3914
3915 if (pte_marker_entry_uffd_wp(entry))
3916 return pte_marker_handle_uffd_wp(vmf);
3917
3918 /* This is an unknown pte marker */
3919 return VM_FAULT_SIGBUS;
3920}
3921
3922/*
3923 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3924 * but allow concurrent faults), and pte mapped but not yet locked.
3925 * We return with pte unmapped and unlocked.
3926 *
3927 * We return with the mmap_lock locked or unlocked in the same cases
3928 * as does filemap_fault().
3929 */
3930vm_fault_t do_swap_page(struct vm_fault *vmf)
3931{
3932 struct vm_area_struct *vma = vmf->vma;
3933 struct folio *swapcache, *folio = NULL;
3934 struct page *page;
3935 struct swap_info_struct *si = NULL;
3936 rmap_t rmap_flags = RMAP_NONE;
3937 bool need_clear_cache = false;
3938 bool exclusive = false;
3939 swp_entry_t entry;
3940 pte_t pte;
3941 vm_fault_t ret = 0;
3942 void *shadow = NULL;
3943
3944 if (!pte_unmap_same(vmf))
3945 goto out;
3946
3947 entry = pte_to_swp_entry(vmf->orig_pte);
3948 if (unlikely(non_swap_entry(entry))) {
3949 if (is_migration_entry(entry)) {
3950 migration_entry_wait(vma->vm_mm, vmf->pmd,
3951 vmf->address);
3952 } else if (is_device_exclusive_entry(entry)) {
3953 vmf->page = pfn_swap_entry_to_page(entry);
3954 ret = remove_device_exclusive_entry(vmf);
3955 } else if (is_device_private_entry(entry)) {
3956 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3957 /*
3958 * migrate_to_ram is not yet ready to operate
3959 * under VMA lock.
3960 */
3961 vma_end_read(vma);
3962 ret = VM_FAULT_RETRY;
3963 goto out;
3964 }
3965
3966 vmf->page = pfn_swap_entry_to_page(entry);
3967 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3968 vmf->address, &vmf->ptl);
3969 if (unlikely(!vmf->pte ||
3970 !pte_same(ptep_get(vmf->pte),
3971 vmf->orig_pte)))
3972 goto unlock;
3973
3974 /*
3975 * Get a page reference while we know the page can't be
3976 * freed.
3977 */
3978 get_page(vmf->page);
3979 pte_unmap_unlock(vmf->pte, vmf->ptl);
3980 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3981 put_page(vmf->page);
3982 } else if (is_hwpoison_entry(entry)) {
3983 ret = VM_FAULT_HWPOISON;
3984 } else if (is_pte_marker_entry(entry)) {
3985 ret = handle_pte_marker(vmf);
3986 } else {
3987 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3988 ret = VM_FAULT_SIGBUS;
3989 }
3990 goto out;
3991 }
3992
3993 /* Prevent swapoff from happening to us. */
3994 si = get_swap_device(entry);
3995 if (unlikely(!si))
3996 goto out;
3997
3998 folio = swap_cache_get_folio(entry, vma, vmf->address);
3999 if (folio)
4000 page = folio_file_page(folio, swp_offset(entry));
4001 swapcache = folio;
4002
4003 if (!folio) {
4004 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
4005 __swap_count(entry) == 1) {
4006 /*
4007 * Prevent parallel swapin from proceeding with
4008 * the cache flag. Otherwise, another thread may
4009 * finish swapin first, free the entry, and swapout
4010 * reusing the same entry. It's undetectable as
4011 * pte_same() returns true due to entry reuse.
4012 */
4013 if (swapcache_prepare(entry)) {
4014 /* Relax a bit to prevent rapid repeated page faults */
4015 schedule_timeout_uninterruptible(1);
4016 goto out;
4017 }
4018 need_clear_cache = true;
4019
4020 /* skip swapcache */
4021 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0,
4022 vma, vmf->address, false);
4023 page = &folio->page;
4024 if (folio) {
4025 __folio_set_locked(folio);
4026 __folio_set_swapbacked(folio);
4027
4028 if (mem_cgroup_swapin_charge_folio(folio,
4029 vma->vm_mm, GFP_KERNEL,
4030 entry)) {
4031 ret = VM_FAULT_OOM;
4032 goto out_page;
4033 }
4034 mem_cgroup_swapin_uncharge_swap(entry);
4035
4036 shadow = get_shadow_from_swap_cache(entry);
4037 if (shadow)
4038 workingset_refault(folio, shadow);
4039
4040 folio_add_lru(folio);
4041
4042 /* To provide entry to swap_read_folio() */
4043 folio->swap = entry;
4044 swap_read_folio(folio, true, NULL);
4045 folio->private = NULL;
4046 }
4047 } else {
4048 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
4049 vmf);
4050 if (page)
4051 folio = page_folio(page);
4052 swapcache = folio;
4053 }
4054
4055 if (!folio) {
4056 /*
4057 * Back out if somebody else faulted in this pte
4058 * while we released the pte lock.
4059 */
4060 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4061 vmf->address, &vmf->ptl);
4062 if (likely(vmf->pte &&
4063 pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
4064 ret = VM_FAULT_OOM;
4065 goto unlock;
4066 }
4067
4068 /* Had to read the page from swap area: Major fault */
4069 ret = VM_FAULT_MAJOR;
4070 count_vm_event(PGMAJFAULT);
4071 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
4072 } else if (PageHWPoison(page)) {
4073 /*
4074 * hwpoisoned dirty swapcache pages are kept for killing
4075 * owner processes (which may be unknown at hwpoison time)
4076 */
4077 ret = VM_FAULT_HWPOISON;
4078 goto out_release;
4079 }
4080
4081 ret |= folio_lock_or_retry(folio, vmf);
4082 if (ret & VM_FAULT_RETRY)
4083 goto out_release;
4084
4085 if (swapcache) {
4086 /*
4087 * Make sure folio_free_swap() or swapoff did not release the
4088 * swapcache from under us. The page pin, and pte_same test
4089 * below, are not enough to exclude that. Even if it is still
4090 * swapcache, we need to check that the page's swap has not
4091 * changed.
4092 */
4093 if (unlikely(!folio_test_swapcache(folio) ||
4094 page_swap_entry(page).val != entry.val))
4095 goto out_page;
4096
4097 /*
4098 * KSM sometimes has to copy on read faults, for example, if
4099 * page->index of !PageKSM() pages would be nonlinear inside the
4100 * anon VMA -- PageKSM() is lost on actual swapout.
4101 */
4102 folio = ksm_might_need_to_copy(folio, vma, vmf->address);
4103 if (unlikely(!folio)) {
4104 ret = VM_FAULT_OOM;
4105 folio = swapcache;
4106 goto out_page;
4107 } else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
4108 ret = VM_FAULT_HWPOISON;
4109 folio = swapcache;
4110 goto out_page;
4111 }
4112 if (folio != swapcache)
4113 page = folio_page(folio, 0);
4114
4115 /*
4116 * If we want to map a page that's in the swapcache writable, we
4117 * have to detect via the refcount if we're really the exclusive
4118 * owner. Try removing the extra reference from the local LRU
4119 * caches if required.
4120 */
4121 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
4122 !folio_test_ksm(folio) && !folio_test_lru(folio))
4123 lru_add_drain();
4124 }
4125
4126 folio_throttle_swaprate(folio, GFP_KERNEL);
4127
4128 /*
4129 * Back out if somebody else already faulted in this pte.
4130 */
4131 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
4132 &vmf->ptl);
4133 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
4134 goto out_nomap;
4135
4136 if (unlikely(!folio_test_uptodate(folio))) {
4137 ret = VM_FAULT_SIGBUS;
4138 goto out_nomap;
4139 }
4140
4141 /*
4142 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4143 * must never point at an anonymous page in the swapcache that is
4144 * PG_anon_exclusive. Sanity check that this holds and especially, that
4145 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4146 * check after taking the PT lock and making sure that nobody
4147 * concurrently faulted in this page and set PG_anon_exclusive.
4148 */
4149 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
4150 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
4151
4152 /*
4153 * Check under PT lock (to protect against concurrent fork() sharing
4154 * the swap entry concurrently) for certainly exclusive pages.
4155 */
4156 if (!folio_test_ksm(folio)) {
4157 exclusive = pte_swp_exclusive(vmf->orig_pte);
4158 if (folio != swapcache) {
4159 /*
4160 * We have a fresh page that is not exposed to the
4161 * swapcache -> certainly exclusive.
4162 */
4163 exclusive = true;
4164 } else if (exclusive && folio_test_writeback(folio) &&
4165 data_race(si->flags & SWP_STABLE_WRITES)) {
4166 /*
4167 * This is tricky: not all swap backends support
4168 * concurrent page modifications while under writeback.
4169 *
4170 * So if we stumble over such a page in the swapcache
4171 * we must not set the page exclusive, otherwise we can
4172 * map it writable without further checks and modify it
4173 * while still under writeback.
4174 *
4175 * For these problematic swap backends, simply drop the
4176 * exclusive marker: this is perfectly fine as we start
4177 * writeback only if we fully unmapped the page and
4178 * there are no unexpected references on the page after
4179 * unmapping succeeded. After fully unmapped, no
4180 * further GUP references (FOLL_GET and FOLL_PIN) can
4181 * appear, so dropping the exclusive marker and mapping
4182 * it only R/O is fine.
4183 */
4184 exclusive = false;
4185 }
4186 }
4187
4188 /*
4189 * Some architectures may have to restore extra metadata to the page
4190 * when reading from swap. This metadata may be indexed by swap entry
4191 * so this must be called before swap_free().
4192 */
4193 arch_swap_restore(entry, folio);
4194
4195 /*
4196 * Remove the swap entry and conditionally try to free up the swapcache.
4197 * We're already holding a reference on the page but haven't mapped it
4198 * yet.
4199 */
4200 swap_free(entry);
4201 if (should_try_to_free_swap(folio, vma, vmf->flags))
4202 folio_free_swap(folio);
4203
4204 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4205 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
4206 pte = mk_pte(page, vma->vm_page_prot);
4207
4208 /*
4209 * Same logic as in do_wp_page(); however, optimize for pages that are
4210 * certainly not shared either because we just allocated them without
4211 * exposing them to the swapcache or because the swap entry indicates
4212 * exclusivity.
4213 */
4214 if (!folio_test_ksm(folio) &&
4215 (exclusive || folio_ref_count(folio) == 1)) {
4216 if (vmf->flags & FAULT_FLAG_WRITE) {
4217 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
4218 vmf->flags &= ~FAULT_FLAG_WRITE;
4219 }
4220 rmap_flags |= RMAP_EXCLUSIVE;
4221 }
4222 flush_icache_page(vma, page);
4223 if (pte_swp_soft_dirty(vmf->orig_pte))
4224 pte = pte_mksoft_dirty(pte);
4225 if (pte_swp_uffd_wp(vmf->orig_pte))
4226 pte = pte_mkuffd_wp(pte);
4227 vmf->orig_pte = pte;
4228
4229 /* ksm created a completely new copy */
4230 if (unlikely(folio != swapcache && swapcache)) {
4231 folio_add_new_anon_rmap(folio, vma, vmf->address);
4232 folio_add_lru_vma(folio, vma);
4233 } else {
4234 folio_add_anon_rmap_pte(folio, page, vma, vmf->address,
4235 rmap_flags);
4236 }
4237
4238 VM_BUG_ON(!folio_test_anon(folio) ||
4239 (pte_write(pte) && !PageAnonExclusive(page)));
4240 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
4241 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
4242
4243 folio_unlock(folio);
4244 if (folio != swapcache && swapcache) {
4245 /*
4246 * Hold the lock to avoid the swap entry to be reused
4247 * until we take the PT lock for the pte_same() check
4248 * (to avoid false positives from pte_same). For
4249 * further safety release the lock after the swap_free
4250 * so that the swap count won't change under a
4251 * parallel locked swapcache.
4252 */
4253 folio_unlock(swapcache);
4254 folio_put(swapcache);
4255 }
4256
4257 if (vmf->flags & FAULT_FLAG_WRITE) {
4258 ret |= do_wp_page(vmf);
4259 if (ret & VM_FAULT_ERROR)
4260 ret &= VM_FAULT_ERROR;
4261 goto out;
4262 }
4263
4264 /* No need to invalidate - it was non-present before */
4265 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
4266unlock:
4267 if (vmf->pte)
4268 pte_unmap_unlock(vmf->pte, vmf->ptl);
4269out:
4270 /* Clear the swap cache pin for direct swapin after PTL unlock */
4271 if (need_clear_cache)
4272 swapcache_clear(si, entry);
4273 if (si)
4274 put_swap_device(si);
4275 return ret;
4276out_nomap:
4277 if (vmf->pte)
4278 pte_unmap_unlock(vmf->pte, vmf->ptl);
4279out_page:
4280 folio_unlock(folio);
4281out_release:
4282 folio_put(folio);
4283 if (folio != swapcache && swapcache) {
4284 folio_unlock(swapcache);
4285 folio_put(swapcache);
4286 }
4287 if (need_clear_cache)
4288 swapcache_clear(si, entry);
4289 if (si)
4290 put_swap_device(si);
4291 return ret;
4292}
4293
4294static bool pte_range_none(pte_t *pte, int nr_pages)
4295{
4296 int i;
4297
4298 for (i = 0; i < nr_pages; i++) {
4299 if (!pte_none(ptep_get_lockless(pte + i)))
4300 return false;
4301 }
4302
4303 return true;
4304}
4305
4306static struct folio *alloc_anon_folio(struct vm_fault *vmf)
4307{
4308 struct vm_area_struct *vma = vmf->vma;
4309#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4310 unsigned long orders;
4311 struct folio *folio;
4312 unsigned long addr;
4313 pte_t *pte;
4314 gfp_t gfp;
4315 int order;
4316
4317 /*
4318 * If uffd is active for the vma we need per-page fault fidelity to
4319 * maintain the uffd semantics.
4320 */
4321 if (unlikely(userfaultfd_armed(vma)))
4322 goto fallback;
4323
4324 /*
4325 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4326 * for this vma. Then filter out the orders that can't be allocated over
4327 * the faulting address and still be fully contained in the vma.
4328 */
4329 orders = thp_vma_allowable_orders(vma, vma->vm_flags, false, true, true,
4330 BIT(PMD_ORDER) - 1);
4331 orders = thp_vma_suitable_orders(vma, vmf->address, orders);
4332
4333 if (!orders)
4334 goto fallback;
4335
4336 pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK);
4337 if (!pte)
4338 return ERR_PTR(-EAGAIN);
4339
4340 /*
4341 * Find the highest order where the aligned range is completely
4342 * pte_none(). Note that all remaining orders will be completely
4343 * pte_none().
4344 */
4345 order = highest_order(orders);
4346 while (orders) {
4347 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4348 if (pte_range_none(pte + pte_index(addr), 1 << order))
4349 break;
4350 order = next_order(&orders, order);
4351 }
4352
4353 pte_unmap(pte);
4354
4355 /* Try allocating the highest of the remaining orders. */
4356 gfp = vma_thp_gfp_mask(vma);
4357 while (orders) {
4358 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4359 folio = vma_alloc_folio(gfp, order, vma, addr, true);
4360 if (folio) {
4361 if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
4362 folio_put(folio);
4363 goto next;
4364 }
4365 folio_throttle_swaprate(folio, gfp);
4366 clear_huge_page(&folio->page, vmf->address, 1 << order);
4367 return folio;
4368 }
4369next:
4370 order = next_order(&orders, order);
4371 }
4372
4373fallback:
4374#endif
4375 return folio_prealloc(vma->vm_mm, vma, vmf->address, true);
4376}
4377
4378/*
4379 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4380 * but allow concurrent faults), and pte mapped but not yet locked.
4381 * We return with mmap_lock still held, but pte unmapped and unlocked.
4382 */
4383static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
4384{
4385 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4386 struct vm_area_struct *vma = vmf->vma;
4387 unsigned long addr = vmf->address;
4388 struct folio *folio;
4389 vm_fault_t ret = 0;
4390 int nr_pages = 1;
4391 pte_t entry;
4392 int i;
4393
4394 /* File mapping without ->vm_ops ? */
4395 if (vma->vm_flags & VM_SHARED)
4396 return VM_FAULT_SIGBUS;
4397
4398 /*
4399 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4400 * be distinguished from a transient failure of pte_offset_map().
4401 */
4402 if (pte_alloc(vma->vm_mm, vmf->pmd))
4403 return VM_FAULT_OOM;
4404
4405 /* Use the zero-page for reads */
4406 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
4407 !mm_forbids_zeropage(vma->vm_mm)) {
4408 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
4409 vma->vm_page_prot));
4410 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4411 vmf->address, &vmf->ptl);
4412 if (!vmf->pte)
4413 goto unlock;
4414 if (vmf_pte_changed(vmf)) {
4415 update_mmu_tlb(vma, vmf->address, vmf->pte);
4416 goto unlock;
4417 }
4418 ret = check_stable_address_space(vma->vm_mm);
4419 if (ret)
4420 goto unlock;
4421 /* Deliver the page fault to userland, check inside PT lock */
4422 if (userfaultfd_missing(vma)) {
4423 pte_unmap_unlock(vmf->pte, vmf->ptl);
4424 return handle_userfault(vmf, VM_UFFD_MISSING);
4425 }
4426 goto setpte;
4427 }
4428
4429 /* Allocate our own private page. */
4430 if (unlikely(anon_vma_prepare(vma)))
4431 goto oom;
4432 /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
4433 folio = alloc_anon_folio(vmf);
4434 if (IS_ERR(folio))
4435 return 0;
4436 if (!folio)
4437 goto oom;
4438
4439 nr_pages = folio_nr_pages(folio);
4440 addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE);
4441
4442 /*
4443 * The memory barrier inside __folio_mark_uptodate makes sure that
4444 * preceding stores to the page contents become visible before
4445 * the set_pte_at() write.
4446 */
4447 __folio_mark_uptodate(folio);
4448
4449 entry = mk_pte(&folio->page, vma->vm_page_prot);
4450 entry = pte_sw_mkyoung(entry);
4451 if (vma->vm_flags & VM_WRITE)
4452 entry = pte_mkwrite(pte_mkdirty(entry), vma);
4453
4454 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
4455 if (!vmf->pte)
4456 goto release;
4457 if (nr_pages == 1 && vmf_pte_changed(vmf)) {
4458 update_mmu_tlb(vma, addr, vmf->pte);
4459 goto release;
4460 } else if (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages)) {
4461 for (i = 0; i < nr_pages; i++)
4462 update_mmu_tlb(vma, addr + PAGE_SIZE * i, vmf->pte + i);
4463 goto release;
4464 }
4465
4466 ret = check_stable_address_space(vma->vm_mm);
4467 if (ret)
4468 goto release;
4469
4470 /* Deliver the page fault to userland, check inside PT lock */
4471 if (userfaultfd_missing(vma)) {
4472 pte_unmap_unlock(vmf->pte, vmf->ptl);
4473 folio_put(folio);
4474 return handle_userfault(vmf, VM_UFFD_MISSING);
4475 }
4476
4477 folio_ref_add(folio, nr_pages - 1);
4478 add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages);
4479 folio_add_new_anon_rmap(folio, vma, addr);
4480 folio_add_lru_vma(folio, vma);
4481setpte:
4482 if (uffd_wp)
4483 entry = pte_mkuffd_wp(entry);
4484 set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr_pages);
4485
4486 /* No need to invalidate - it was non-present before */
4487 update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr_pages);
4488unlock:
4489 if (vmf->pte)
4490 pte_unmap_unlock(vmf->pte, vmf->ptl);
4491 return ret;
4492release:
4493 folio_put(folio);
4494 goto unlock;
4495oom:
4496 return VM_FAULT_OOM;
4497}
4498
4499/*
4500 * The mmap_lock must have been held on entry, and may have been
4501 * released depending on flags and vma->vm_ops->fault() return value.
4502 * See filemap_fault() and __lock_page_retry().
4503 */
4504static vm_fault_t __do_fault(struct vm_fault *vmf)
4505{
4506 struct vm_area_struct *vma = vmf->vma;
4507 struct folio *folio;
4508 vm_fault_t ret;
4509
4510 /*
4511 * Preallocate pte before we take page_lock because this might lead to
4512 * deadlocks for memcg reclaim which waits for pages under writeback:
4513 * lock_page(A)
4514 * SetPageWriteback(A)
4515 * unlock_page(A)
4516 * lock_page(B)
4517 * lock_page(B)
4518 * pte_alloc_one
4519 * shrink_page_list
4520 * wait_on_page_writeback(A)
4521 * SetPageWriteback(B)
4522 * unlock_page(B)
4523 * # flush A, B to clear the writeback
4524 */
4525 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
4526 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4527 if (!vmf->prealloc_pte)
4528 return VM_FAULT_OOM;
4529 }
4530
4531 ret = vma->vm_ops->fault(vmf);
4532 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
4533 VM_FAULT_DONE_COW)))
4534 return ret;
4535
4536 folio = page_folio(vmf->page);
4537 if (unlikely(PageHWPoison(vmf->page))) {
4538 vm_fault_t poisonret = VM_FAULT_HWPOISON;
4539 if (ret & VM_FAULT_LOCKED) {
4540 if (page_mapped(vmf->page))
4541 unmap_mapping_folio(folio);
4542 /* Retry if a clean folio was removed from the cache. */
4543 if (mapping_evict_folio(folio->mapping, folio))
4544 poisonret = VM_FAULT_NOPAGE;
4545 folio_unlock(folio);
4546 }
4547 folio_put(folio);
4548 vmf->page = NULL;
4549 return poisonret;
4550 }
4551
4552 if (unlikely(!(ret & VM_FAULT_LOCKED)))
4553 folio_lock(folio);
4554 else
4555 VM_BUG_ON_PAGE(!folio_test_locked(folio), vmf->page);
4556
4557 return ret;
4558}
4559
4560#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4561static void deposit_prealloc_pte(struct vm_fault *vmf)
4562{
4563 struct vm_area_struct *vma = vmf->vma;
4564
4565 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4566 /*
4567 * We are going to consume the prealloc table,
4568 * count that as nr_ptes.
4569 */
4570 mm_inc_nr_ptes(vma->vm_mm);
4571 vmf->prealloc_pte = NULL;
4572}
4573
4574vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4575{
4576 struct folio *folio = page_folio(page);
4577 struct vm_area_struct *vma = vmf->vma;
4578 bool write = vmf->flags & FAULT_FLAG_WRITE;
4579 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
4580 pmd_t entry;
4581 vm_fault_t ret = VM_FAULT_FALLBACK;
4582
4583 if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
4584 return ret;
4585
4586 if (page != &folio->page || folio_order(folio) != HPAGE_PMD_ORDER)
4587 return ret;
4588
4589 /*
4590 * Just backoff if any subpage of a THP is corrupted otherwise
4591 * the corrupted page may mapped by PMD silently to escape the
4592 * check. This kind of THP just can be PTE mapped. Access to
4593 * the corrupted subpage should trigger SIGBUS as expected.
4594 */
4595 if (unlikely(folio_test_has_hwpoisoned(folio)))
4596 return ret;
4597
4598 /*
4599 * Archs like ppc64 need additional space to store information
4600 * related to pte entry. Use the preallocated table for that.
4601 */
4602 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
4603 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4604 if (!vmf->prealloc_pte)
4605 return VM_FAULT_OOM;
4606 }
4607
4608 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4609 if (unlikely(!pmd_none(*vmf->pmd)))
4610 goto out;
4611
4612 flush_icache_pages(vma, page, HPAGE_PMD_NR);
4613
4614 entry = mk_huge_pmd(page, vma->vm_page_prot);
4615 if (write)
4616 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
4617
4618 add_mm_counter(vma->vm_mm, mm_counter_file(folio), HPAGE_PMD_NR);
4619 folio_add_file_rmap_pmd(folio, page, vma);
4620
4621 /*
4622 * deposit and withdraw with pmd lock held
4623 */
4624 if (arch_needs_pgtable_deposit())
4625 deposit_prealloc_pte(vmf);
4626
4627 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
4628
4629 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
4630
4631 /* fault is handled */
4632 ret = 0;
4633 count_vm_event(THP_FILE_MAPPED);
4634out:
4635 spin_unlock(vmf->ptl);
4636 return ret;
4637}
4638#else
4639vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4640{
4641 return VM_FAULT_FALLBACK;
4642}
4643#endif
4644
4645/**
4646 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4647 * @vmf: Fault decription.
4648 * @folio: The folio that contains @page.
4649 * @page: The first page to create a PTE for.
4650 * @nr: The number of PTEs to create.
4651 * @addr: The first address to create a PTE for.
4652 */
4653void set_pte_range(struct vm_fault *vmf, struct folio *folio,
4654 struct page *page, unsigned int nr, unsigned long addr)
4655{
4656 struct vm_area_struct *vma = vmf->vma;
4657 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4658 bool write = vmf->flags & FAULT_FLAG_WRITE;
4659 bool prefault = in_range(vmf->address, addr, nr * PAGE_SIZE);
4660 pte_t entry;
4661
4662 flush_icache_pages(vma, page, nr);
4663 entry = mk_pte(page, vma->vm_page_prot);
4664
4665 if (prefault && arch_wants_old_prefaulted_pte())
4666 entry = pte_mkold(entry);
4667 else
4668 entry = pte_sw_mkyoung(entry);
4669
4670 if (write)
4671 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4672 if (unlikely(uffd_wp))
4673 entry = pte_mkuffd_wp(entry);
4674 /* copy-on-write page */
4675 if (write && !(vma->vm_flags & VM_SHARED)) {
4676 add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr);
4677 VM_BUG_ON_FOLIO(nr != 1, folio);
4678 folio_add_new_anon_rmap(folio, vma, addr);
4679 folio_add_lru_vma(folio, vma);
4680 } else {
4681 add_mm_counter(vma->vm_mm, mm_counter_file(folio), nr);
4682 folio_add_file_rmap_ptes(folio, page, nr, vma);
4683 }
4684 set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr);
4685
4686 /* no need to invalidate: a not-present page won't be cached */
4687 update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr);
4688}
4689
4690static bool vmf_pte_changed(struct vm_fault *vmf)
4691{
4692 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
4693 return !pte_same(ptep_get(vmf->pte), vmf->orig_pte);
4694
4695 return !pte_none(ptep_get(vmf->pte));
4696}
4697
4698/**
4699 * finish_fault - finish page fault once we have prepared the page to fault
4700 *
4701 * @vmf: structure describing the fault
4702 *
4703 * This function handles all that is needed to finish a page fault once the
4704 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4705 * given page, adds reverse page mapping, handles memcg charges and LRU
4706 * addition.
4707 *
4708 * The function expects the page to be locked and on success it consumes a
4709 * reference of a page being mapped (for the PTE which maps it).
4710 *
4711 * Return: %0 on success, %VM_FAULT_ code in case of error.
4712 */
4713vm_fault_t finish_fault(struct vm_fault *vmf)
4714{
4715 struct vm_area_struct *vma = vmf->vma;
4716 struct page *page;
4717 vm_fault_t ret;
4718
4719 /* Did we COW the page? */
4720 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4721 page = vmf->cow_page;
4722 else
4723 page = vmf->page;
4724
4725 /*
4726 * check even for read faults because we might have lost our CoWed
4727 * page
4728 */
4729 if (!(vma->vm_flags & VM_SHARED)) {
4730 ret = check_stable_address_space(vma->vm_mm);
4731 if (ret)
4732 return ret;
4733 }
4734
4735 if (pmd_none(*vmf->pmd)) {
4736 if (PageTransCompound(page)) {
4737 ret = do_set_pmd(vmf, page);
4738 if (ret != VM_FAULT_FALLBACK)
4739 return ret;
4740 }
4741
4742 if (vmf->prealloc_pte)
4743 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4744 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4745 return VM_FAULT_OOM;
4746 }
4747
4748 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4749 vmf->address, &vmf->ptl);
4750 if (!vmf->pte)
4751 return VM_FAULT_NOPAGE;
4752
4753 /* Re-check under ptl */
4754 if (likely(!vmf_pte_changed(vmf))) {
4755 struct folio *folio = page_folio(page);
4756
4757 set_pte_range(vmf, folio, page, 1, vmf->address);
4758 ret = 0;
4759 } else {
4760 update_mmu_tlb(vma, vmf->address, vmf->pte);
4761 ret = VM_FAULT_NOPAGE;
4762 }
4763
4764 pte_unmap_unlock(vmf->pte, vmf->ptl);
4765 return ret;
4766}
4767
4768static unsigned long fault_around_pages __read_mostly =
4769 65536 >> PAGE_SHIFT;
4770
4771#ifdef CONFIG_DEBUG_FS
4772static int fault_around_bytes_get(void *data, u64 *val)
4773{
4774 *val = fault_around_pages << PAGE_SHIFT;
4775 return 0;
4776}
4777
4778/*
4779 * fault_around_bytes must be rounded down to the nearest page order as it's
4780 * what do_fault_around() expects to see.
4781 */
4782static int fault_around_bytes_set(void *data, u64 val)
4783{
4784 if (val / PAGE_SIZE > PTRS_PER_PTE)
4785 return -EINVAL;
4786
4787 /*
4788 * The minimum value is 1 page, however this results in no fault-around
4789 * at all. See should_fault_around().
4790 */
4791 val = max(val, PAGE_SIZE);
4792 fault_around_pages = rounddown_pow_of_two(val) >> PAGE_SHIFT;
4793
4794 return 0;
4795}
4796DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4797 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4798
4799static int __init fault_around_debugfs(void)
4800{
4801 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4802 &fault_around_bytes_fops);
4803 return 0;
4804}
4805late_initcall(fault_around_debugfs);
4806#endif
4807
4808/*
4809 * do_fault_around() tries to map few pages around the fault address. The hope
4810 * is that the pages will be needed soon and this will lower the number of
4811 * faults to handle.
4812 *
4813 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4814 * not ready to be mapped: not up-to-date, locked, etc.
4815 *
4816 * This function doesn't cross VMA or page table boundaries, in order to call
4817 * map_pages() and acquire a PTE lock only once.
4818 *
4819 * fault_around_pages defines how many pages we'll try to map.
4820 * do_fault_around() expects it to be set to a power of two less than or equal
4821 * to PTRS_PER_PTE.
4822 *
4823 * The virtual address of the area that we map is naturally aligned to
4824 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4825 * (and therefore to page order). This way it's easier to guarantee
4826 * that we don't cross page table boundaries.
4827 */
4828static vm_fault_t do_fault_around(struct vm_fault *vmf)
4829{
4830 pgoff_t nr_pages = READ_ONCE(fault_around_pages);
4831 pgoff_t pte_off = pte_index(vmf->address);
4832 /* The page offset of vmf->address within the VMA. */
4833 pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
4834 pgoff_t from_pte, to_pte;
4835 vm_fault_t ret;
4836
4837 /* The PTE offset of the start address, clamped to the VMA. */
4838 from_pte = max(ALIGN_DOWN(pte_off, nr_pages),
4839 pte_off - min(pte_off, vma_off));
4840
4841 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4842 to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE,
4843 pte_off + vma_pages(vmf->vma) - vma_off) - 1;
4844
4845 if (pmd_none(*vmf->pmd)) {
4846 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4847 if (!vmf->prealloc_pte)
4848 return VM_FAULT_OOM;
4849 }
4850
4851 rcu_read_lock();
4852 ret = vmf->vma->vm_ops->map_pages(vmf,
4853 vmf->pgoff + from_pte - pte_off,
4854 vmf->pgoff + to_pte - pte_off);
4855 rcu_read_unlock();
4856
4857 return ret;
4858}
4859
4860/* Return true if we should do read fault-around, false otherwise */
4861static inline bool should_fault_around(struct vm_fault *vmf)
4862{
4863 /* No ->map_pages? No way to fault around... */
4864 if (!vmf->vma->vm_ops->map_pages)
4865 return false;
4866
4867 if (uffd_disable_fault_around(vmf->vma))
4868 return false;
4869
4870 /* A single page implies no faulting 'around' at all. */
4871 return fault_around_pages > 1;
4872}
4873
4874static vm_fault_t do_read_fault(struct vm_fault *vmf)
4875{
4876 vm_fault_t ret = 0;
4877 struct folio *folio;
4878
4879 /*
4880 * Let's call ->map_pages() first and use ->fault() as fallback
4881 * if page by the offset is not ready to be mapped (cold cache or
4882 * something).
4883 */
4884 if (should_fault_around(vmf)) {
4885 ret = do_fault_around(vmf);
4886 if (ret)
4887 return ret;
4888 }
4889
4890 ret = vmf_can_call_fault(vmf);
4891 if (ret)
4892 return ret;
4893
4894 ret = __do_fault(vmf);
4895 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4896 return ret;
4897
4898 ret |= finish_fault(vmf);
4899 folio = page_folio(vmf->page);
4900 folio_unlock(folio);
4901 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4902 folio_put(folio);
4903 return ret;
4904}
4905
4906static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4907{
4908 struct vm_area_struct *vma = vmf->vma;
4909 struct folio *folio;
4910 vm_fault_t ret;
4911
4912 ret = vmf_can_call_fault(vmf);
4913 if (!ret)
4914 ret = vmf_anon_prepare(vmf);
4915 if (ret)
4916 return ret;
4917
4918 folio = folio_prealloc(vma->vm_mm, vma, vmf->address, false);
4919 if (!folio)
4920 return VM_FAULT_OOM;
4921
4922 vmf->cow_page = &folio->page;
4923
4924 ret = __do_fault(vmf);
4925 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4926 goto uncharge_out;
4927 if (ret & VM_FAULT_DONE_COW)
4928 return ret;
4929
4930 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4931 __folio_mark_uptodate(folio);
4932
4933 ret |= finish_fault(vmf);
4934 unlock_page(vmf->page);
4935 put_page(vmf->page);
4936 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4937 goto uncharge_out;
4938 return ret;
4939uncharge_out:
4940 folio_put(folio);
4941 return ret;
4942}
4943
4944static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4945{
4946 struct vm_area_struct *vma = vmf->vma;
4947 vm_fault_t ret, tmp;
4948 struct folio *folio;
4949
4950 ret = vmf_can_call_fault(vmf);
4951 if (ret)
4952 return ret;
4953
4954 ret = __do_fault(vmf);
4955 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4956 return ret;
4957
4958 folio = page_folio(vmf->page);
4959
4960 /*
4961 * Check if the backing address space wants to know that the page is
4962 * about to become writable
4963 */
4964 if (vma->vm_ops->page_mkwrite) {
4965 folio_unlock(folio);
4966 tmp = do_page_mkwrite(vmf, folio);
4967 if (unlikely(!tmp ||
4968 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4969 folio_put(folio);
4970 return tmp;
4971 }
4972 }
4973
4974 ret |= finish_fault(vmf);
4975 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4976 VM_FAULT_RETRY))) {
4977 folio_unlock(folio);
4978 folio_put(folio);
4979 return ret;
4980 }
4981
4982 ret |= fault_dirty_shared_page(vmf);
4983 return ret;
4984}
4985
4986/*
4987 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4988 * but allow concurrent faults).
4989 * The mmap_lock may have been released depending on flags and our
4990 * return value. See filemap_fault() and __folio_lock_or_retry().
4991 * If mmap_lock is released, vma may become invalid (for example
4992 * by other thread calling munmap()).
4993 */
4994static vm_fault_t do_fault(struct vm_fault *vmf)
4995{
4996 struct vm_area_struct *vma = vmf->vma;
4997 struct mm_struct *vm_mm = vma->vm_mm;
4998 vm_fault_t ret;
4999
5000 /*
5001 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
5002 */
5003 if (!vma->vm_ops->fault) {
5004 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
5005 vmf->address, &vmf->ptl);
5006 if (unlikely(!vmf->pte))
5007 ret = VM_FAULT_SIGBUS;
5008 else {
5009 /*
5010 * Make sure this is not a temporary clearing of pte
5011 * by holding ptl and checking again. A R/M/W update
5012 * of pte involves: take ptl, clearing the pte so that
5013 * we don't have concurrent modification by hardware
5014 * followed by an update.
5015 */
5016 if (unlikely(pte_none(ptep_get(vmf->pte))))
5017 ret = VM_FAULT_SIGBUS;
5018 else
5019 ret = VM_FAULT_NOPAGE;
5020
5021 pte_unmap_unlock(vmf->pte, vmf->ptl);
5022 }
5023 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
5024 ret = do_read_fault(vmf);
5025 else if (!(vma->vm_flags & VM_SHARED))
5026 ret = do_cow_fault(vmf);
5027 else
5028 ret = do_shared_fault(vmf);
5029
5030 /* preallocated pagetable is unused: free it */
5031 if (vmf->prealloc_pte) {
5032 pte_free(vm_mm, vmf->prealloc_pte);
5033 vmf->prealloc_pte = NULL;
5034 }
5035 return ret;
5036}
5037
5038int numa_migrate_prep(struct folio *folio, struct vm_area_struct *vma,
5039 unsigned long addr, int page_nid, int *flags)
5040{
5041 folio_get(folio);
5042
5043 /* Record the current PID acceesing VMA */
5044 vma_set_access_pid_bit(vma);
5045
5046 count_vm_numa_event(NUMA_HINT_FAULTS);
5047 if (page_nid == numa_node_id()) {
5048 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
5049 *flags |= TNF_FAULT_LOCAL;
5050 }
5051
5052 return mpol_misplaced(folio, vma, addr);
5053}
5054
5055static vm_fault_t do_numa_page(struct vm_fault *vmf)
5056{
5057 struct vm_area_struct *vma = vmf->vma;
5058 struct folio *folio = NULL;
5059 int nid = NUMA_NO_NODE;
5060 bool writable = false;
5061 int last_cpupid;
5062 int target_nid;
5063 pte_t pte, old_pte;
5064 int flags = 0;
5065
5066 /*
5067 * The pte cannot be used safely until we verify, while holding the page
5068 * table lock, that its contents have not changed during fault handling.
5069 */
5070 spin_lock(vmf->ptl);
5071 /* Read the live PTE from the page tables: */
5072 old_pte = ptep_get(vmf->pte);
5073
5074 if (unlikely(!pte_same(old_pte, vmf->orig_pte))) {
5075 pte_unmap_unlock(vmf->pte, vmf->ptl);
5076 goto out;
5077 }
5078
5079 pte = pte_modify(old_pte, vma->vm_page_prot);
5080
5081 /*
5082 * Detect now whether the PTE could be writable; this information
5083 * is only valid while holding the PT lock.
5084 */
5085 writable = pte_write(pte);
5086 if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
5087 can_change_pte_writable(vma, vmf->address, pte))
5088 writable = true;
5089
5090 folio = vm_normal_folio(vma, vmf->address, pte);
5091 if (!folio || folio_is_zone_device(folio))
5092 goto out_map;
5093
5094 /* TODO: handle PTE-mapped THP */
5095 if (folio_test_large(folio))
5096 goto out_map;
5097
5098 /*
5099 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
5100 * much anyway since they can be in shared cache state. This misses
5101 * the case where a mapping is writable but the process never writes
5102 * to it but pte_write gets cleared during protection updates and
5103 * pte_dirty has unpredictable behaviour between PTE scan updates,
5104 * background writeback, dirty balancing and application behaviour.
5105 */
5106 if (!writable)
5107 flags |= TNF_NO_GROUP;
5108
5109 /*
5110 * Flag if the folio is shared between multiple address spaces. This
5111 * is later used when determining whether to group tasks together
5112 */
5113 if (folio_estimated_sharers(folio) > 1 && (vma->vm_flags & VM_SHARED))
5114 flags |= TNF_SHARED;
5115
5116 nid = folio_nid(folio);
5117 /*
5118 * For memory tiering mode, cpupid of slow memory page is used
5119 * to record page access time. So use default value.
5120 */
5121 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
5122 !node_is_toptier(nid))
5123 last_cpupid = (-1 & LAST_CPUPID_MASK);
5124 else
5125 last_cpupid = folio_last_cpupid(folio);
5126 target_nid = numa_migrate_prep(folio, vma, vmf->address, nid, &flags);
5127 if (target_nid == NUMA_NO_NODE) {
5128 folio_put(folio);
5129 goto out_map;
5130 }
5131 pte_unmap_unlock(vmf->pte, vmf->ptl);
5132 writable = false;
5133
5134 /* Migrate to the requested node */
5135 if (migrate_misplaced_folio(folio, vma, target_nid)) {
5136 nid = target_nid;
5137 flags |= TNF_MIGRATED;
5138 } else {
5139 flags |= TNF_MIGRATE_FAIL;
5140 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
5141 vmf->address, &vmf->ptl);
5142 if (unlikely(!vmf->pte))
5143 goto out;
5144 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
5145 pte_unmap_unlock(vmf->pte, vmf->ptl);
5146 goto out;
5147 }
5148 goto out_map;
5149 }
5150
5151out:
5152 if (nid != NUMA_NO_NODE)
5153 task_numa_fault(last_cpupid, nid, 1, flags);
5154 return 0;
5155out_map:
5156 /*
5157 * Make it present again, depending on how arch implements
5158 * non-accessible ptes, some can allow access by kernel mode.
5159 */
5160 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
5161 pte = pte_modify(old_pte, vma->vm_page_prot);
5162 pte = pte_mkyoung(pte);
5163 if (writable)
5164 pte = pte_mkwrite(pte, vma);
5165 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
5166 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
5167 pte_unmap_unlock(vmf->pte, vmf->ptl);
5168 goto out;
5169}
5170
5171static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
5172{
5173 struct vm_area_struct *vma = vmf->vma;
5174 if (vma_is_anonymous(vma))
5175 return do_huge_pmd_anonymous_page(vmf);
5176 if (vma->vm_ops->huge_fault)
5177 return vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5178 return VM_FAULT_FALLBACK;
5179}
5180
5181/* `inline' is required to avoid gcc 4.1.2 build error */
5182static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
5183{
5184 struct vm_area_struct *vma = vmf->vma;
5185 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
5186 vm_fault_t ret;
5187
5188 if (vma_is_anonymous(vma)) {
5189 if (likely(!unshare) &&
5190 userfaultfd_huge_pmd_wp(vma, vmf->orig_pmd)) {
5191 if (userfaultfd_wp_async(vmf->vma))
5192 goto split;
5193 return handle_userfault(vmf, VM_UFFD_WP);
5194 }
5195 return do_huge_pmd_wp_page(vmf);
5196 }
5197
5198 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5199 if (vma->vm_ops->huge_fault) {
5200 ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5201 if (!(ret & VM_FAULT_FALLBACK))
5202 return ret;
5203 }
5204 }
5205
5206split:
5207 /* COW or write-notify handled on pte level: split pmd. */
5208 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
5209
5210 return VM_FAULT_FALLBACK;
5211}
5212
5213static vm_fault_t create_huge_pud(struct vm_fault *vmf)
5214{
5215#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5216 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5217 struct vm_area_struct *vma = vmf->vma;
5218 /* No support for anonymous transparent PUD pages yet */
5219 if (vma_is_anonymous(vma))
5220 return VM_FAULT_FALLBACK;
5221 if (vma->vm_ops->huge_fault)
5222 return vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5223#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5224 return VM_FAULT_FALLBACK;
5225}
5226
5227static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
5228{
5229#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5230 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5231 struct vm_area_struct *vma = vmf->vma;
5232 vm_fault_t ret;
5233
5234 /* No support for anonymous transparent PUD pages yet */
5235 if (vma_is_anonymous(vma))
5236 goto split;
5237 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5238 if (vma->vm_ops->huge_fault) {
5239 ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5240 if (!(ret & VM_FAULT_FALLBACK))
5241 return ret;
5242 }
5243 }
5244split:
5245 /* COW or write-notify not handled on PUD level: split pud.*/
5246 __split_huge_pud(vma, vmf->pud, vmf->address);
5247#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
5248 return VM_FAULT_FALLBACK;
5249}
5250
5251/*
5252 * These routines also need to handle stuff like marking pages dirty
5253 * and/or accessed for architectures that don't do it in hardware (most
5254 * RISC architectures). The early dirtying is also good on the i386.
5255 *
5256 * There is also a hook called "update_mmu_cache()" that architectures
5257 * with external mmu caches can use to update those (ie the Sparc or
5258 * PowerPC hashed page tables that act as extended TLBs).
5259 *
5260 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
5261 * concurrent faults).
5262 *
5263 * The mmap_lock may have been released depending on flags and our return value.
5264 * See filemap_fault() and __folio_lock_or_retry().
5265 */
5266static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
5267{
5268 pte_t entry;
5269
5270 if (unlikely(pmd_none(*vmf->pmd))) {
5271 /*
5272 * Leave __pte_alloc() until later: because vm_ops->fault may
5273 * want to allocate huge page, and if we expose page table
5274 * for an instant, it will be difficult to retract from
5275 * concurrent faults and from rmap lookups.
5276 */
5277 vmf->pte = NULL;
5278 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
5279 } else {
5280 /*
5281 * A regular pmd is established and it can't morph into a huge
5282 * pmd by anon khugepaged, since that takes mmap_lock in write
5283 * mode; but shmem or file collapse to THP could still morph
5284 * it into a huge pmd: just retry later if so.
5285 */
5286 vmf->pte = pte_offset_map_nolock(vmf->vma->vm_mm, vmf->pmd,
5287 vmf->address, &vmf->ptl);
5288 if (unlikely(!vmf->pte))
5289 return 0;
5290 vmf->orig_pte = ptep_get_lockless(vmf->pte);
5291 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
5292
5293 if (pte_none(vmf->orig_pte)) {
5294 pte_unmap(vmf->pte);
5295 vmf->pte = NULL;
5296 }
5297 }
5298
5299 if (!vmf->pte)
5300 return do_pte_missing(vmf);
5301
5302 if (!pte_present(vmf->orig_pte))
5303 return do_swap_page(vmf);
5304
5305 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
5306 return do_numa_page(vmf);
5307
5308 spin_lock(vmf->ptl);
5309 entry = vmf->orig_pte;
5310 if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) {
5311 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
5312 goto unlock;
5313 }
5314 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
5315 if (!pte_write(entry))
5316 return do_wp_page(vmf);
5317 else if (likely(vmf->flags & FAULT_FLAG_WRITE))
5318 entry = pte_mkdirty(entry);
5319 }
5320 entry = pte_mkyoung(entry);
5321 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
5322 vmf->flags & FAULT_FLAG_WRITE)) {
5323 update_mmu_cache_range(vmf, vmf->vma, vmf->address,
5324 vmf->pte, 1);
5325 } else {
5326 /* Skip spurious TLB flush for retried page fault */
5327 if (vmf->flags & FAULT_FLAG_TRIED)
5328 goto unlock;
5329 /*
5330 * This is needed only for protection faults but the arch code
5331 * is not yet telling us if this is a protection fault or not.
5332 * This still avoids useless tlb flushes for .text page faults
5333 * with threads.
5334 */
5335 if (vmf->flags & FAULT_FLAG_WRITE)
5336 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address,
5337 vmf->pte);
5338 }
5339unlock:
5340 pte_unmap_unlock(vmf->pte, vmf->ptl);
5341 return 0;
5342}
5343
5344/*
5345 * On entry, we hold either the VMA lock or the mmap_lock
5346 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5347 * the result, the mmap_lock is not held on exit. See filemap_fault()
5348 * and __folio_lock_or_retry().
5349 */
5350static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
5351 unsigned long address, unsigned int flags)
5352{
5353 struct vm_fault vmf = {
5354 .vma = vma,
5355 .address = address & PAGE_MASK,
5356 .real_address = address,
5357 .flags = flags,
5358 .pgoff = linear_page_index(vma, address),
5359 .gfp_mask = __get_fault_gfp_mask(vma),
5360 };
5361 struct mm_struct *mm = vma->vm_mm;
5362 unsigned long vm_flags = vma->vm_flags;
5363 pgd_t *pgd;
5364 p4d_t *p4d;
5365 vm_fault_t ret;
5366
5367 pgd = pgd_offset(mm, address);
5368 p4d = p4d_alloc(mm, pgd, address);
5369 if (!p4d)
5370 return VM_FAULT_OOM;
5371
5372 vmf.pud = pud_alloc(mm, p4d, address);
5373 if (!vmf.pud)
5374 return VM_FAULT_OOM;
5375retry_pud:
5376 if (pud_none(*vmf.pud) &&
5377 thp_vma_allowable_order(vma, vm_flags, false, true, true, PUD_ORDER)) {
5378 ret = create_huge_pud(&vmf);
5379 if (!(ret & VM_FAULT_FALLBACK))
5380 return ret;
5381 } else {
5382 pud_t orig_pud = *vmf.pud;
5383
5384 barrier();
5385 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
5386
5387 /*
5388 * TODO once we support anonymous PUDs: NUMA case and
5389 * FAULT_FLAG_UNSHARE handling.
5390 */
5391 if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) {
5392 ret = wp_huge_pud(&vmf, orig_pud);
5393 if (!(ret & VM_FAULT_FALLBACK))
5394 return ret;
5395 } else {
5396 huge_pud_set_accessed(&vmf, orig_pud);
5397 return 0;
5398 }
5399 }
5400 }
5401
5402 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
5403 if (!vmf.pmd)
5404 return VM_FAULT_OOM;
5405
5406 /* Huge pud page fault raced with pmd_alloc? */
5407 if (pud_trans_unstable(vmf.pud))
5408 goto retry_pud;
5409
5410 if (pmd_none(*vmf.pmd) &&
5411 thp_vma_allowable_order(vma, vm_flags, false, true, true, PMD_ORDER)) {
5412 ret = create_huge_pmd(&vmf);
5413 if (!(ret & VM_FAULT_FALLBACK))
5414 return ret;
5415 } else {
5416 vmf.orig_pmd = pmdp_get_lockless(vmf.pmd);
5417
5418 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
5419 VM_BUG_ON(thp_migration_supported() &&
5420 !is_pmd_migration_entry(vmf.orig_pmd));
5421 if (is_pmd_migration_entry(vmf.orig_pmd))
5422 pmd_migration_entry_wait(mm, vmf.pmd);
5423 return 0;
5424 }
5425 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
5426 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
5427 return do_huge_pmd_numa_page(&vmf);
5428
5429 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5430 !pmd_write(vmf.orig_pmd)) {
5431 ret = wp_huge_pmd(&vmf);
5432 if (!(ret & VM_FAULT_FALLBACK))
5433 return ret;
5434 } else {
5435 huge_pmd_set_accessed(&vmf);
5436 return 0;
5437 }
5438 }
5439 }
5440
5441 return handle_pte_fault(&vmf);
5442}
5443
5444/**
5445 * mm_account_fault - Do page fault accounting
5446 * @mm: mm from which memcg should be extracted. It can be NULL.
5447 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5448 * of perf event counters, but we'll still do the per-task accounting to
5449 * the task who triggered this page fault.
5450 * @address: the faulted address.
5451 * @flags: the fault flags.
5452 * @ret: the fault retcode.
5453 *
5454 * This will take care of most of the page fault accounting. Meanwhile, it
5455 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5456 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5457 * still be in per-arch page fault handlers at the entry of page fault.
5458 */
5459static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs,
5460 unsigned long address, unsigned int flags,
5461 vm_fault_t ret)
5462{
5463 bool major;
5464
5465 /* Incomplete faults will be accounted upon completion. */
5466 if (ret & VM_FAULT_RETRY)
5467 return;
5468
5469 /*
5470 * To preserve the behavior of older kernels, PGFAULT counters record
5471 * both successful and failed faults, as opposed to perf counters,
5472 * which ignore failed cases.
5473 */
5474 count_vm_event(PGFAULT);
5475 count_memcg_event_mm(mm, PGFAULT);
5476
5477 /*
5478 * Do not account for unsuccessful faults (e.g. when the address wasn't
5479 * valid). That includes arch_vma_access_permitted() failing before
5480 * reaching here. So this is not a "this many hardware page faults"
5481 * counter. We should use the hw profiling for that.
5482 */
5483 if (ret & VM_FAULT_ERROR)
5484 return;
5485
5486 /*
5487 * We define the fault as a major fault when the final successful fault
5488 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5489 * handle it immediately previously).
5490 */
5491 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
5492
5493 if (major)
5494 current->maj_flt++;
5495 else
5496 current->min_flt++;
5497
5498 /*
5499 * If the fault is done for GUP, regs will be NULL. We only do the
5500 * accounting for the per thread fault counters who triggered the
5501 * fault, and we skip the perf event updates.
5502 */
5503 if (!regs)
5504 return;
5505
5506 if (major)
5507 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
5508 else
5509 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
5510}
5511
5512#ifdef CONFIG_LRU_GEN
5513static void lru_gen_enter_fault(struct vm_area_struct *vma)
5514{
5515 /* the LRU algorithm only applies to accesses with recency */
5516 current->in_lru_fault = vma_has_recency(vma);
5517}
5518
5519static void lru_gen_exit_fault(void)
5520{
5521 current->in_lru_fault = false;
5522}
5523#else
5524static void lru_gen_enter_fault(struct vm_area_struct *vma)
5525{
5526}
5527
5528static void lru_gen_exit_fault(void)
5529{
5530}
5531#endif /* CONFIG_LRU_GEN */
5532
5533static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
5534 unsigned int *flags)
5535{
5536 if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
5537 if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
5538 return VM_FAULT_SIGSEGV;
5539 /*
5540 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5541 * just treat it like an ordinary read-fault otherwise.
5542 */
5543 if (!is_cow_mapping(vma->vm_flags))
5544 *flags &= ~FAULT_FLAG_UNSHARE;
5545 } else if (*flags & FAULT_FLAG_WRITE) {
5546 /* Write faults on read-only mappings are impossible ... */
5547 if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
5548 return VM_FAULT_SIGSEGV;
5549 /* ... and FOLL_FORCE only applies to COW mappings. */
5550 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
5551 !is_cow_mapping(vma->vm_flags)))
5552 return VM_FAULT_SIGSEGV;
5553 }
5554#ifdef CONFIG_PER_VMA_LOCK
5555 /*
5556 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5557 * the assumption that lock is dropped on VM_FAULT_RETRY.
5558 */
5559 if (WARN_ON_ONCE((*flags &
5560 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) ==
5561 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)))
5562 return VM_FAULT_SIGSEGV;
5563#endif
5564
5565 return 0;
5566}
5567
5568/*
5569 * By the time we get here, we already hold the mm semaphore
5570 *
5571 * The mmap_lock may have been released depending on flags and our
5572 * return value. See filemap_fault() and __folio_lock_or_retry().
5573 */
5574vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
5575 unsigned int flags, struct pt_regs *regs)
5576{
5577 /* If the fault handler drops the mmap_lock, vma may be freed */
5578 struct mm_struct *mm = vma->vm_mm;
5579 vm_fault_t ret;
5580
5581 __set_current_state(TASK_RUNNING);
5582
5583 ret = sanitize_fault_flags(vma, &flags);
5584 if (ret)
5585 goto out;
5586
5587 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
5588 flags & FAULT_FLAG_INSTRUCTION,
5589 flags & FAULT_FLAG_REMOTE)) {
5590 ret = VM_FAULT_SIGSEGV;
5591 goto out;
5592 }
5593
5594 /*
5595 * Enable the memcg OOM handling for faults triggered in user
5596 * space. Kernel faults are handled more gracefully.
5597 */
5598 if (flags & FAULT_FLAG_USER)
5599 mem_cgroup_enter_user_fault();
5600
5601 lru_gen_enter_fault(vma);
5602
5603 if (unlikely(is_vm_hugetlb_page(vma)))
5604 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
5605 else
5606 ret = __handle_mm_fault(vma, address, flags);
5607
5608 lru_gen_exit_fault();
5609
5610 if (flags & FAULT_FLAG_USER) {
5611 mem_cgroup_exit_user_fault();
5612 /*
5613 * The task may have entered a memcg OOM situation but
5614 * if the allocation error was handled gracefully (no
5615 * VM_FAULT_OOM), there is no need to kill anything.
5616 * Just clean up the OOM state peacefully.
5617 */
5618 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
5619 mem_cgroup_oom_synchronize(false);
5620 }
5621out:
5622 mm_account_fault(mm, regs, address, flags, ret);
5623
5624 return ret;
5625}
5626EXPORT_SYMBOL_GPL(handle_mm_fault);
5627
5628#ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5629#include <linux/extable.h>
5630
5631static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5632{
5633 if (likely(mmap_read_trylock(mm)))
5634 return true;
5635
5636 if (regs && !user_mode(regs)) {
5637 unsigned long ip = exception_ip(regs);
5638 if (!search_exception_tables(ip))
5639 return false;
5640 }
5641
5642 return !mmap_read_lock_killable(mm);
5643}
5644
5645static inline bool mmap_upgrade_trylock(struct mm_struct *mm)
5646{
5647 /*
5648 * We don't have this operation yet.
5649 *
5650 * It should be easy enough to do: it's basically a
5651 * atomic_long_try_cmpxchg_acquire()
5652 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5653 * it also needs the proper lockdep magic etc.
5654 */
5655 return false;
5656}
5657
5658static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5659{
5660 mmap_read_unlock(mm);
5661 if (regs && !user_mode(regs)) {
5662 unsigned long ip = exception_ip(regs);
5663 if (!search_exception_tables(ip))
5664 return false;
5665 }
5666 return !mmap_write_lock_killable(mm);
5667}
5668
5669/*
5670 * Helper for page fault handling.
5671 *
5672 * This is kind of equivalend to "mmap_read_lock()" followed
5673 * by "find_extend_vma()", except it's a lot more careful about
5674 * the locking (and will drop the lock on failure).
5675 *
5676 * For example, if we have a kernel bug that causes a page
5677 * fault, we don't want to just use mmap_read_lock() to get
5678 * the mm lock, because that would deadlock if the bug were
5679 * to happen while we're holding the mm lock for writing.
5680 *
5681 * So this checks the exception tables on kernel faults in
5682 * order to only do this all for instructions that are actually
5683 * expected to fault.
5684 *
5685 * We can also actually take the mm lock for writing if we
5686 * need to extend the vma, which helps the VM layer a lot.
5687 */
5688struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm,
5689 unsigned long addr, struct pt_regs *regs)
5690{
5691 struct vm_area_struct *vma;
5692
5693 if (!get_mmap_lock_carefully(mm, regs))
5694 return NULL;
5695
5696 vma = find_vma(mm, addr);
5697 if (likely(vma && (vma->vm_start <= addr)))
5698 return vma;
5699
5700 /*
5701 * Well, dang. We might still be successful, but only
5702 * if we can extend a vma to do so.
5703 */
5704 if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) {
5705 mmap_read_unlock(mm);
5706 return NULL;
5707 }
5708
5709 /*
5710 * We can try to upgrade the mmap lock atomically,
5711 * in which case we can continue to use the vma
5712 * we already looked up.
5713 *
5714 * Otherwise we'll have to drop the mmap lock and
5715 * re-take it, and also look up the vma again,
5716 * re-checking it.
5717 */
5718 if (!mmap_upgrade_trylock(mm)) {
5719 if (!upgrade_mmap_lock_carefully(mm, regs))
5720 return NULL;
5721
5722 vma = find_vma(mm, addr);
5723 if (!vma)
5724 goto fail;
5725 if (vma->vm_start <= addr)
5726 goto success;
5727 if (!(vma->vm_flags & VM_GROWSDOWN))
5728 goto fail;
5729 }
5730
5731 if (expand_stack_locked(vma, addr))
5732 goto fail;
5733
5734success:
5735 mmap_write_downgrade(mm);
5736 return vma;
5737
5738fail:
5739 mmap_write_unlock(mm);
5740 return NULL;
5741}
5742#endif
5743
5744#ifdef CONFIG_PER_VMA_LOCK
5745/*
5746 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5747 * stable and not isolated. If the VMA is not found or is being modified the
5748 * function returns NULL.
5749 */
5750struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
5751 unsigned long address)
5752{
5753 MA_STATE(mas, &mm->mm_mt, address, address);
5754 struct vm_area_struct *vma;
5755
5756 rcu_read_lock();
5757retry:
5758 vma = mas_walk(&mas);
5759 if (!vma)
5760 goto inval;
5761
5762 if (!vma_start_read(vma))
5763 goto inval;
5764
5765 /*
5766 * find_mergeable_anon_vma uses adjacent vmas which are not locked.
5767 * This check must happen after vma_start_read(); otherwise, a
5768 * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA
5769 * from its anon_vma.
5770 */
5771 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma))
5772 goto inval_end_read;
5773
5774 /* Check since vm_start/vm_end might change before we lock the VMA */
5775 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5776 goto inval_end_read;
5777
5778 /* Check if the VMA got isolated after we found it */
5779 if (vma->detached) {
5780 vma_end_read(vma);
5781 count_vm_vma_lock_event(VMA_LOCK_MISS);
5782 /* The area was replaced with another one */
5783 goto retry;
5784 }
5785
5786 rcu_read_unlock();
5787 return vma;
5788
5789inval_end_read:
5790 vma_end_read(vma);
5791inval:
5792 rcu_read_unlock();
5793 count_vm_vma_lock_event(VMA_LOCK_ABORT);
5794 return NULL;
5795}
5796#endif /* CONFIG_PER_VMA_LOCK */
5797
5798#ifndef __PAGETABLE_P4D_FOLDED
5799/*
5800 * Allocate p4d page table.
5801 * We've already handled the fast-path in-line.
5802 */
5803int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
5804{
5805 p4d_t *new = p4d_alloc_one(mm, address);
5806 if (!new)
5807 return -ENOMEM;
5808
5809 spin_lock(&mm->page_table_lock);
5810 if (pgd_present(*pgd)) { /* Another has populated it */
5811 p4d_free(mm, new);
5812 } else {
5813 smp_wmb(); /* See comment in pmd_install() */
5814 pgd_populate(mm, pgd, new);
5815 }
5816 spin_unlock(&mm->page_table_lock);
5817 return 0;
5818}
5819#endif /* __PAGETABLE_P4D_FOLDED */
5820
5821#ifndef __PAGETABLE_PUD_FOLDED
5822/*
5823 * Allocate page upper directory.
5824 * We've already handled the fast-path in-line.
5825 */
5826int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
5827{
5828 pud_t *new = pud_alloc_one(mm, address);
5829 if (!new)
5830 return -ENOMEM;
5831
5832 spin_lock(&mm->page_table_lock);
5833 if (!p4d_present(*p4d)) {
5834 mm_inc_nr_puds(mm);
5835 smp_wmb(); /* See comment in pmd_install() */
5836 p4d_populate(mm, p4d, new);
5837 } else /* Another has populated it */
5838 pud_free(mm, new);
5839 spin_unlock(&mm->page_table_lock);
5840 return 0;
5841}
5842#endif /* __PAGETABLE_PUD_FOLDED */
5843
5844#ifndef __PAGETABLE_PMD_FOLDED
5845/*
5846 * Allocate page middle directory.
5847 * We've already handled the fast-path in-line.
5848 */
5849int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
5850{
5851 spinlock_t *ptl;
5852 pmd_t *new = pmd_alloc_one(mm, address);
5853 if (!new)
5854 return -ENOMEM;
5855
5856 ptl = pud_lock(mm, pud);
5857 if (!pud_present(*pud)) {
5858 mm_inc_nr_pmds(mm);
5859 smp_wmb(); /* See comment in pmd_install() */
5860 pud_populate(mm, pud, new);
5861 } else { /* Another has populated it */
5862 pmd_free(mm, new);
5863 }
5864 spin_unlock(ptl);
5865 return 0;
5866}
5867#endif /* __PAGETABLE_PMD_FOLDED */
5868
5869/**
5870 * follow_pte - look up PTE at a user virtual address
5871 * @mm: the mm_struct of the target address space
5872 * @address: user virtual address
5873 * @ptepp: location to store found PTE
5874 * @ptlp: location to store the lock for the PTE
5875 *
5876 * On a successful return, the pointer to the PTE is stored in @ptepp;
5877 * the corresponding lock is taken and its location is stored in @ptlp.
5878 * The contents of the PTE are only stable until @ptlp is released;
5879 * any further use, if any, must be protected against invalidation
5880 * with MMU notifiers.
5881 *
5882 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5883 * should be taken for read.
5884 *
5885 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5886 * it is not a good general-purpose API.
5887 *
5888 * Return: zero on success, -ve otherwise.
5889 */
5890int follow_pte(struct mm_struct *mm, unsigned long address,
5891 pte_t **ptepp, spinlock_t **ptlp)
5892{
5893 pgd_t *pgd;
5894 p4d_t *p4d;
5895 pud_t *pud;
5896 pmd_t *pmd;
5897 pte_t *ptep;
5898
5899 pgd = pgd_offset(mm, address);
5900 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
5901 goto out;
5902
5903 p4d = p4d_offset(pgd, address);
5904 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
5905 goto out;
5906
5907 pud = pud_offset(p4d, address);
5908 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
5909 goto out;
5910
5911 pmd = pmd_offset(pud, address);
5912 VM_BUG_ON(pmd_trans_huge(*pmd));
5913
5914 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
5915 if (!ptep)
5916 goto out;
5917 if (!pte_present(ptep_get(ptep)))
5918 goto unlock;
5919 *ptepp = ptep;
5920 return 0;
5921unlock:
5922 pte_unmap_unlock(ptep, *ptlp);
5923out:
5924 return -EINVAL;
5925}
5926EXPORT_SYMBOL_GPL(follow_pte);
5927
5928/**
5929 * follow_pfn - look up PFN at a user virtual address
5930 * @vma: memory mapping
5931 * @address: user virtual address
5932 * @pfn: location to store found PFN
5933 *
5934 * Only IO mappings and raw PFN mappings are allowed.
5935 *
5936 * This function does not allow the caller to read the permissions
5937 * of the PTE. Do not use it.
5938 *
5939 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5940 */
5941int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5942 unsigned long *pfn)
5943{
5944 int ret = -EINVAL;
5945 spinlock_t *ptl;
5946 pte_t *ptep;
5947
5948 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5949 return ret;
5950
5951 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5952 if (ret)
5953 return ret;
5954 *pfn = pte_pfn(ptep_get(ptep));
5955 pte_unmap_unlock(ptep, ptl);
5956 return 0;
5957}
5958EXPORT_SYMBOL(follow_pfn);
5959
5960#ifdef CONFIG_HAVE_IOREMAP_PROT
5961int follow_phys(struct vm_area_struct *vma,
5962 unsigned long address, unsigned int flags,
5963 unsigned long *prot, resource_size_t *phys)
5964{
5965 int ret = -EINVAL;
5966 pte_t *ptep, pte;
5967 spinlock_t *ptl;
5968
5969 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5970 goto out;
5971
5972 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5973 goto out;
5974 pte = ptep_get(ptep);
5975
5976 /* Never return PFNs of anon folios in COW mappings. */
5977 if (vm_normal_folio(vma, address, pte))
5978 goto unlock;
5979
5980 if ((flags & FOLL_WRITE) && !pte_write(pte))
5981 goto unlock;
5982
5983 *prot = pgprot_val(pte_pgprot(pte));
5984 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5985
5986 ret = 0;
5987unlock:
5988 pte_unmap_unlock(ptep, ptl);
5989out:
5990 return ret;
5991}
5992
5993/**
5994 * generic_access_phys - generic implementation for iomem mmap access
5995 * @vma: the vma to access
5996 * @addr: userspace address, not relative offset within @vma
5997 * @buf: buffer to read/write
5998 * @len: length of transfer
5999 * @write: set to FOLL_WRITE when writing, otherwise reading
6000 *
6001 * This is a generic implementation for &vm_operations_struct.access for an
6002 * iomem mapping. This callback is used by access_process_vm() when the @vma is
6003 * not page based.
6004 */
6005int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
6006 void *buf, int len, int write)
6007{
6008 resource_size_t phys_addr;
6009 unsigned long prot = 0;
6010 void __iomem *maddr;
6011 pte_t *ptep, pte;
6012 spinlock_t *ptl;
6013 int offset = offset_in_page(addr);
6014 int ret = -EINVAL;
6015
6016 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
6017 return -EINVAL;
6018
6019retry:
6020 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
6021 return -EINVAL;
6022 pte = ptep_get(ptep);
6023 pte_unmap_unlock(ptep, ptl);
6024
6025 prot = pgprot_val(pte_pgprot(pte));
6026 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
6027
6028 if ((write & FOLL_WRITE) && !pte_write(pte))
6029 return -EINVAL;
6030
6031 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
6032 if (!maddr)
6033 return -ENOMEM;
6034
6035 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
6036 goto out_unmap;
6037
6038 if (!pte_same(pte, ptep_get(ptep))) {
6039 pte_unmap_unlock(ptep, ptl);
6040 iounmap(maddr);
6041
6042 goto retry;
6043 }
6044
6045 if (write)
6046 memcpy_toio(maddr + offset, buf, len);
6047 else
6048 memcpy_fromio(buf, maddr + offset, len);
6049 ret = len;
6050 pte_unmap_unlock(ptep, ptl);
6051out_unmap:
6052 iounmap(maddr);
6053
6054 return ret;
6055}
6056EXPORT_SYMBOL_GPL(generic_access_phys);
6057#endif
6058
6059/*
6060 * Access another process' address space as given in mm.
6061 */
6062static int __access_remote_vm(struct mm_struct *mm, unsigned long addr,
6063 void *buf, int len, unsigned int gup_flags)
6064{
6065 void *old_buf = buf;
6066 int write = gup_flags & FOLL_WRITE;
6067
6068 if (mmap_read_lock_killable(mm))
6069 return 0;
6070
6071 /* Untag the address before looking up the VMA */
6072 addr = untagged_addr_remote(mm, addr);
6073
6074 /* Avoid triggering the temporary warning in __get_user_pages */
6075 if (!vma_lookup(mm, addr) && !expand_stack(mm, addr))
6076 return 0;
6077
6078 /* ignore errors, just check how much was successfully transferred */
6079 while (len) {
6080 int bytes, offset;
6081 void *maddr;
6082 struct vm_area_struct *vma = NULL;
6083 struct page *page = get_user_page_vma_remote(mm, addr,
6084 gup_flags, &vma);
6085
6086 if (IS_ERR(page)) {
6087 /* We might need to expand the stack to access it */
6088 vma = vma_lookup(mm, addr);
6089 if (!vma) {
6090 vma = expand_stack(mm, addr);
6091
6092 /* mmap_lock was dropped on failure */
6093 if (!vma)
6094 return buf - old_buf;
6095
6096 /* Try again if stack expansion worked */
6097 continue;
6098 }
6099
6100 /*
6101 * Check if this is a VM_IO | VM_PFNMAP VMA, which
6102 * we can access using slightly different code.
6103 */
6104 bytes = 0;
6105#ifdef CONFIG_HAVE_IOREMAP_PROT
6106 if (vma->vm_ops && vma->vm_ops->access)
6107 bytes = vma->vm_ops->access(vma, addr, buf,
6108 len, write);
6109#endif
6110 if (bytes <= 0)
6111 break;
6112 } else {
6113 bytes = len;
6114 offset = addr & (PAGE_SIZE-1);
6115 if (bytes > PAGE_SIZE-offset)
6116 bytes = PAGE_SIZE-offset;
6117
6118 maddr = kmap_local_page(page);
6119 if (write) {
6120 copy_to_user_page(vma, page, addr,
6121 maddr + offset, buf, bytes);
6122 set_page_dirty_lock(page);
6123 } else {
6124 copy_from_user_page(vma, page, addr,
6125 buf, maddr + offset, bytes);
6126 }
6127 unmap_and_put_page(page, maddr);
6128 }
6129 len -= bytes;
6130 buf += bytes;
6131 addr += bytes;
6132 }
6133 mmap_read_unlock(mm);
6134
6135 return buf - old_buf;
6136}
6137
6138/**
6139 * access_remote_vm - access another process' address space
6140 * @mm: the mm_struct of the target address space
6141 * @addr: start address to access
6142 * @buf: source or destination buffer
6143 * @len: number of bytes to transfer
6144 * @gup_flags: flags modifying lookup behaviour
6145 *
6146 * The caller must hold a reference on @mm.
6147 *
6148 * Return: number of bytes copied from source to destination.
6149 */
6150int access_remote_vm(struct mm_struct *mm, unsigned long addr,
6151 void *buf, int len, unsigned int gup_flags)
6152{
6153 return __access_remote_vm(mm, addr, buf, len, gup_flags);
6154}
6155
6156/*
6157 * Access another process' address space.
6158 * Source/target buffer must be kernel space,
6159 * Do not walk the page table directly, use get_user_pages
6160 */
6161int access_process_vm(struct task_struct *tsk, unsigned long addr,
6162 void *buf, int len, unsigned int gup_flags)
6163{
6164 struct mm_struct *mm;
6165 int ret;
6166
6167 mm = get_task_mm(tsk);
6168 if (!mm)
6169 return 0;
6170
6171 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
6172
6173 mmput(mm);
6174
6175 return ret;
6176}
6177EXPORT_SYMBOL_GPL(access_process_vm);
6178
6179/*
6180 * Print the name of a VMA.
6181 */
6182void print_vma_addr(char *prefix, unsigned long ip)
6183{
6184 struct mm_struct *mm = current->mm;
6185 struct vm_area_struct *vma;
6186
6187 /*
6188 * we might be running from an atomic context so we cannot sleep
6189 */
6190 if (!mmap_read_trylock(mm))
6191 return;
6192
6193 vma = find_vma(mm, ip);
6194 if (vma && vma->vm_file) {
6195 struct file *f = vma->vm_file;
6196 char *buf = (char *)__get_free_page(GFP_NOWAIT);
6197 if (buf) {
6198 char *p;
6199
6200 p = file_path(f, buf, PAGE_SIZE);
6201 if (IS_ERR(p))
6202 p = "?";
6203 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
6204 vma->vm_start,
6205 vma->vm_end - vma->vm_start);
6206 free_page((unsigned long)buf);
6207 }
6208 }
6209 mmap_read_unlock(mm);
6210}
6211
6212#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6213void __might_fault(const char *file, int line)
6214{
6215 if (pagefault_disabled())
6216 return;
6217 __might_sleep(file, line);
6218#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6219 if (current->mm)
6220 might_lock_read(¤t->mm->mmap_lock);
6221#endif
6222}
6223EXPORT_SYMBOL(__might_fault);
6224#endif
6225
6226#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
6227/*
6228 * Process all subpages of the specified huge page with the specified
6229 * operation. The target subpage will be processed last to keep its
6230 * cache lines hot.
6231 */
6232static inline int process_huge_page(
6233 unsigned long addr_hint, unsigned int pages_per_huge_page,
6234 int (*process_subpage)(unsigned long addr, int idx, void *arg),
6235 void *arg)
6236{
6237 int i, n, base, l, ret;
6238 unsigned long addr = addr_hint &
6239 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6240
6241 /* Process target subpage last to keep its cache lines hot */
6242 might_sleep();
6243 n = (addr_hint - addr) / PAGE_SIZE;
6244 if (2 * n <= pages_per_huge_page) {
6245 /* If target subpage in first half of huge page */
6246 base = 0;
6247 l = n;
6248 /* Process subpages at the end of huge page */
6249 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
6250 cond_resched();
6251 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6252 if (ret)
6253 return ret;
6254 }
6255 } else {
6256 /* If target subpage in second half of huge page */
6257 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
6258 l = pages_per_huge_page - n;
6259 /* Process subpages at the begin of huge page */
6260 for (i = 0; i < base; i++) {
6261 cond_resched();
6262 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6263 if (ret)
6264 return ret;
6265 }
6266 }
6267 /*
6268 * Process remaining subpages in left-right-left-right pattern
6269 * towards the target subpage
6270 */
6271 for (i = 0; i < l; i++) {
6272 int left_idx = base + i;
6273 int right_idx = base + 2 * l - 1 - i;
6274
6275 cond_resched();
6276 ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
6277 if (ret)
6278 return ret;
6279 cond_resched();
6280 ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
6281 if (ret)
6282 return ret;
6283 }
6284 return 0;
6285}
6286
6287static void clear_gigantic_page(struct page *page,
6288 unsigned long addr,
6289 unsigned int pages_per_huge_page)
6290{
6291 int i;
6292 struct page *p;
6293
6294 might_sleep();
6295 for (i = 0; i < pages_per_huge_page; i++) {
6296 p = nth_page(page, i);
6297 cond_resched();
6298 clear_user_highpage(p, addr + i * PAGE_SIZE);
6299 }
6300}
6301
6302static int clear_subpage(unsigned long addr, int idx, void *arg)
6303{
6304 struct page *page = arg;
6305
6306 clear_user_highpage(nth_page(page, idx), addr);
6307 return 0;
6308}
6309
6310void clear_huge_page(struct page *page,
6311 unsigned long addr_hint, unsigned int pages_per_huge_page)
6312{
6313 unsigned long addr = addr_hint &
6314 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6315
6316 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
6317 clear_gigantic_page(page, addr, pages_per_huge_page);
6318 return;
6319 }
6320
6321 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
6322}
6323
6324static int copy_user_gigantic_page(struct folio *dst, struct folio *src,
6325 unsigned long addr,
6326 struct vm_area_struct *vma,
6327 unsigned int pages_per_huge_page)
6328{
6329 int i;
6330 struct page *dst_page;
6331 struct page *src_page;
6332
6333 for (i = 0; i < pages_per_huge_page; i++) {
6334 dst_page = folio_page(dst, i);
6335 src_page = folio_page(src, i);
6336
6337 cond_resched();
6338 if (copy_mc_user_highpage(dst_page, src_page,
6339 addr + i*PAGE_SIZE, vma)) {
6340 memory_failure_queue(page_to_pfn(src_page), 0);
6341 return -EHWPOISON;
6342 }
6343 }
6344 return 0;
6345}
6346
6347struct copy_subpage_arg {
6348 struct page *dst;
6349 struct page *src;
6350 struct vm_area_struct *vma;
6351};
6352
6353static int copy_subpage(unsigned long addr, int idx, void *arg)
6354{
6355 struct copy_subpage_arg *copy_arg = arg;
6356 struct page *dst = nth_page(copy_arg->dst, idx);
6357 struct page *src = nth_page(copy_arg->src, idx);
6358
6359 if (copy_mc_user_highpage(dst, src, addr, copy_arg->vma)) {
6360 memory_failure_queue(page_to_pfn(src), 0);
6361 return -EHWPOISON;
6362 }
6363 return 0;
6364}
6365
6366int copy_user_large_folio(struct folio *dst, struct folio *src,
6367 unsigned long addr_hint, struct vm_area_struct *vma)
6368{
6369 unsigned int pages_per_huge_page = folio_nr_pages(dst);
6370 unsigned long addr = addr_hint &
6371 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6372 struct copy_subpage_arg arg = {
6373 .dst = &dst->page,
6374 .src = &src->page,
6375 .vma = vma,
6376 };
6377
6378 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES))
6379 return copy_user_gigantic_page(dst, src, addr, vma,
6380 pages_per_huge_page);
6381
6382 return process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
6383}
6384
6385long copy_folio_from_user(struct folio *dst_folio,
6386 const void __user *usr_src,
6387 bool allow_pagefault)
6388{
6389 void *kaddr;
6390 unsigned long i, rc = 0;
6391 unsigned int nr_pages = folio_nr_pages(dst_folio);
6392 unsigned long ret_val = nr_pages * PAGE_SIZE;
6393 struct page *subpage;
6394
6395 for (i = 0; i < nr_pages; i++) {
6396 subpage = folio_page(dst_folio, i);
6397 kaddr = kmap_local_page(subpage);
6398 if (!allow_pagefault)
6399 pagefault_disable();
6400 rc = copy_from_user(kaddr, usr_src + i * PAGE_SIZE, PAGE_SIZE);
6401 if (!allow_pagefault)
6402 pagefault_enable();
6403 kunmap_local(kaddr);
6404
6405 ret_val -= (PAGE_SIZE - rc);
6406 if (rc)
6407 break;
6408
6409 flush_dcache_page(subpage);
6410
6411 cond_resched();
6412 }
6413 return ret_val;
6414}
6415#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6416
6417#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6418
6419static struct kmem_cache *page_ptl_cachep;
6420
6421void __init ptlock_cache_init(void)
6422{
6423 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
6424 SLAB_PANIC, NULL);
6425}
6426
6427bool ptlock_alloc(struct ptdesc *ptdesc)
6428{
6429 spinlock_t *ptl;
6430
6431 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
6432 if (!ptl)
6433 return false;
6434 ptdesc->ptl = ptl;
6435 return true;
6436}
6437
6438void ptlock_free(struct ptdesc *ptdesc)
6439{
6440 kmem_cache_free(page_ptl_cachep, ptdesc->ptl);
6441}
6442#endif