Loading...
1/*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
29 */
30
31/*
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
41#include <linux/kernel_stat.h>
42#include <linux/mm.h>
43#include <linux/hugetlb.h>
44#include <linux/mman.h>
45#include <linux/swap.h>
46#include <linux/highmem.h>
47#include <linux/pagemap.h>
48#include <linux/ksm.h>
49#include <linux/rmap.h>
50#include <linux/export.h>
51#include <linux/delayacct.h>
52#include <linux/init.h>
53#include <linux/writeback.h>
54#include <linux/memcontrol.h>
55#include <linux/mmu_notifier.h>
56#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
59#include <linux/gfp.h>
60#include <linux/migrate.h>
61#include <linux/string.h>
62#include <linux/dma-debug.h>
63#include <linux/debugfs.h>
64
65#include <asm/io.h>
66#include <asm/pgalloc.h>
67#include <asm/uaccess.h>
68#include <asm/tlb.h>
69#include <asm/tlbflush.h>
70#include <asm/pgtable.h>
71
72#include "internal.h"
73
74#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
76#endif
77
78#ifndef CONFIG_NEED_MULTIPLE_NODES
79/* use the per-pgdat data instead for discontigmem - mbligh */
80unsigned long max_mapnr;
81struct page *mem_map;
82
83EXPORT_SYMBOL(max_mapnr);
84EXPORT_SYMBOL(mem_map);
85#endif
86
87/*
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92 * and ZONE_HIGHMEM.
93 */
94void * high_memory;
95
96EXPORT_SYMBOL(high_memory);
97
98/*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104int randomize_va_space __read_mostly =
105#ifdef CONFIG_COMPAT_BRK
106 1;
107#else
108 2;
109#endif
110
111static int __init disable_randmaps(char *s)
112{
113 randomize_va_space = 0;
114 return 1;
115}
116__setup("norandmaps", disable_randmaps);
117
118unsigned long zero_pfn __read_mostly;
119unsigned long highest_memmap_pfn __read_mostly;
120
121/*
122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 */
124static int __init init_zero_pfn(void)
125{
126 zero_pfn = page_to_pfn(ZERO_PAGE(0));
127 return 0;
128}
129core_initcall(init_zero_pfn);
130
131
132#if defined(SPLIT_RSS_COUNTING)
133
134void sync_mm_rss(struct mm_struct *mm)
135{
136 int i;
137
138 for (i = 0; i < NR_MM_COUNTERS; i++) {
139 if (current->rss_stat.count[i]) {
140 add_mm_counter(mm, i, current->rss_stat.count[i]);
141 current->rss_stat.count[i] = 0;
142 }
143 }
144 current->rss_stat.events = 0;
145}
146
147static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148{
149 struct task_struct *task = current;
150
151 if (likely(task->mm == mm))
152 task->rss_stat.count[member] += val;
153 else
154 add_mm_counter(mm, member, val);
155}
156#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159/* sync counter once per 64 page faults */
160#define TASK_RSS_EVENTS_THRESH (64)
161static void check_sync_rss_stat(struct task_struct *task)
162{
163 if (unlikely(task != current))
164 return;
165 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166 sync_mm_rss(task->mm);
167}
168#else /* SPLIT_RSS_COUNTING */
169
170#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173static void check_sync_rss_stat(struct task_struct *task)
174{
175}
176
177#endif /* SPLIT_RSS_COUNTING */
178
179#ifdef HAVE_GENERIC_MMU_GATHER
180
181static int tlb_next_batch(struct mmu_gather *tlb)
182{
183 struct mmu_gather_batch *batch;
184
185 batch = tlb->active;
186 if (batch->next) {
187 tlb->active = batch->next;
188 return 1;
189 }
190
191 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192 return 0;
193
194 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195 if (!batch)
196 return 0;
197
198 tlb->batch_count++;
199 batch->next = NULL;
200 batch->nr = 0;
201 batch->max = MAX_GATHER_BATCH;
202
203 tlb->active->next = batch;
204 tlb->active = batch;
205
206 return 1;
207}
208
209/* tlb_gather_mmu
210 * Called to initialize an (on-stack) mmu_gather structure for page-table
211 * tear-down from @mm. The @fullmm argument is used when @mm is without
212 * users and we're going to destroy the full address space (exit/execve).
213 */
214void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
215{
216 tlb->mm = mm;
217
218 /* Is it from 0 to ~0? */
219 tlb->fullmm = !(start | (end+1));
220 tlb->need_flush_all = 0;
221 tlb->start = start;
222 tlb->end = end;
223 tlb->need_flush = 0;
224 tlb->local.next = NULL;
225 tlb->local.nr = 0;
226 tlb->local.max = ARRAY_SIZE(tlb->__pages);
227 tlb->active = &tlb->local;
228 tlb->batch_count = 0;
229
230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 tlb->batch = NULL;
232#endif
233}
234
235static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
236{
237 tlb->need_flush = 0;
238 tlb_flush(tlb);
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb_table_flush(tlb);
241#endif
242}
243
244static void tlb_flush_mmu_free(struct mmu_gather *tlb)
245{
246 struct mmu_gather_batch *batch;
247
248 for (batch = &tlb->local; batch; batch = batch->next) {
249 free_pages_and_swap_cache(batch->pages, batch->nr);
250 batch->nr = 0;
251 }
252 tlb->active = &tlb->local;
253}
254
255void tlb_flush_mmu(struct mmu_gather *tlb)
256{
257 if (!tlb->need_flush)
258 return;
259 tlb_flush_mmu_tlbonly(tlb);
260 tlb_flush_mmu_free(tlb);
261}
262
263/* tlb_finish_mmu
264 * Called at the end of the shootdown operation to free up any resources
265 * that were required.
266 */
267void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
268{
269 struct mmu_gather_batch *batch, *next;
270
271 tlb_flush_mmu(tlb);
272
273 /* keep the page table cache within bounds */
274 check_pgt_cache();
275
276 for (batch = tlb->local.next; batch; batch = next) {
277 next = batch->next;
278 free_pages((unsigned long)batch, 0);
279 }
280 tlb->local.next = NULL;
281}
282
283/* __tlb_remove_page
284 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
285 * handling the additional races in SMP caused by other CPUs caching valid
286 * mappings in their TLBs. Returns the number of free page slots left.
287 * When out of page slots we must call tlb_flush_mmu().
288 */
289int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
290{
291 struct mmu_gather_batch *batch;
292
293 VM_BUG_ON(!tlb->need_flush);
294
295 batch = tlb->active;
296 batch->pages[batch->nr++] = page;
297 if (batch->nr == batch->max) {
298 if (!tlb_next_batch(tlb))
299 return 0;
300 batch = tlb->active;
301 }
302 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
303
304 return batch->max - batch->nr;
305}
306
307#endif /* HAVE_GENERIC_MMU_GATHER */
308
309#ifdef CONFIG_HAVE_RCU_TABLE_FREE
310
311/*
312 * See the comment near struct mmu_table_batch.
313 */
314
315static void tlb_remove_table_smp_sync(void *arg)
316{
317 /* Simply deliver the interrupt */
318}
319
320static void tlb_remove_table_one(void *table)
321{
322 /*
323 * This isn't an RCU grace period and hence the page-tables cannot be
324 * assumed to be actually RCU-freed.
325 *
326 * It is however sufficient for software page-table walkers that rely on
327 * IRQ disabling. See the comment near struct mmu_table_batch.
328 */
329 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
330 __tlb_remove_table(table);
331}
332
333static void tlb_remove_table_rcu(struct rcu_head *head)
334{
335 struct mmu_table_batch *batch;
336 int i;
337
338 batch = container_of(head, struct mmu_table_batch, rcu);
339
340 for (i = 0; i < batch->nr; i++)
341 __tlb_remove_table(batch->tables[i]);
342
343 free_page((unsigned long)batch);
344}
345
346void tlb_table_flush(struct mmu_gather *tlb)
347{
348 struct mmu_table_batch **batch = &tlb->batch;
349
350 if (*batch) {
351 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
352 *batch = NULL;
353 }
354}
355
356void tlb_remove_table(struct mmu_gather *tlb, void *table)
357{
358 struct mmu_table_batch **batch = &tlb->batch;
359
360 tlb->need_flush = 1;
361
362 /*
363 * When there's less then two users of this mm there cannot be a
364 * concurrent page-table walk.
365 */
366 if (atomic_read(&tlb->mm->mm_users) < 2) {
367 __tlb_remove_table(table);
368 return;
369 }
370
371 if (*batch == NULL) {
372 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
373 if (*batch == NULL) {
374 tlb_remove_table_one(table);
375 return;
376 }
377 (*batch)->nr = 0;
378 }
379 (*batch)->tables[(*batch)->nr++] = table;
380 if ((*batch)->nr == MAX_TABLE_BATCH)
381 tlb_table_flush(tlb);
382}
383
384#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
385
386/*
387 * Note: this doesn't free the actual pages themselves. That
388 * has been handled earlier when unmapping all the memory regions.
389 */
390static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
391 unsigned long addr)
392{
393 pgtable_t token = pmd_pgtable(*pmd);
394 pmd_clear(pmd);
395 pte_free_tlb(tlb, token, addr);
396 atomic_long_dec(&tlb->mm->nr_ptes);
397}
398
399static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
400 unsigned long addr, unsigned long end,
401 unsigned long floor, unsigned long ceiling)
402{
403 pmd_t *pmd;
404 unsigned long next;
405 unsigned long start;
406
407 start = addr;
408 pmd = pmd_offset(pud, addr);
409 do {
410 next = pmd_addr_end(addr, end);
411 if (pmd_none_or_clear_bad(pmd))
412 continue;
413 free_pte_range(tlb, pmd, addr);
414 } while (pmd++, addr = next, addr != end);
415
416 start &= PUD_MASK;
417 if (start < floor)
418 return;
419 if (ceiling) {
420 ceiling &= PUD_MASK;
421 if (!ceiling)
422 return;
423 }
424 if (end - 1 > ceiling - 1)
425 return;
426
427 pmd = pmd_offset(pud, start);
428 pud_clear(pud);
429 pmd_free_tlb(tlb, pmd, start);
430}
431
432static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
433 unsigned long addr, unsigned long end,
434 unsigned long floor, unsigned long ceiling)
435{
436 pud_t *pud;
437 unsigned long next;
438 unsigned long start;
439
440 start = addr;
441 pud = pud_offset(pgd, addr);
442 do {
443 next = pud_addr_end(addr, end);
444 if (pud_none_or_clear_bad(pud))
445 continue;
446 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
447 } while (pud++, addr = next, addr != end);
448
449 start &= PGDIR_MASK;
450 if (start < floor)
451 return;
452 if (ceiling) {
453 ceiling &= PGDIR_MASK;
454 if (!ceiling)
455 return;
456 }
457 if (end - 1 > ceiling - 1)
458 return;
459
460 pud = pud_offset(pgd, start);
461 pgd_clear(pgd);
462 pud_free_tlb(tlb, pud, start);
463}
464
465/*
466 * This function frees user-level page tables of a process.
467 */
468void free_pgd_range(struct mmu_gather *tlb,
469 unsigned long addr, unsigned long end,
470 unsigned long floor, unsigned long ceiling)
471{
472 pgd_t *pgd;
473 unsigned long next;
474
475 /*
476 * The next few lines have given us lots of grief...
477 *
478 * Why are we testing PMD* at this top level? Because often
479 * there will be no work to do at all, and we'd prefer not to
480 * go all the way down to the bottom just to discover that.
481 *
482 * Why all these "- 1"s? Because 0 represents both the bottom
483 * of the address space and the top of it (using -1 for the
484 * top wouldn't help much: the masks would do the wrong thing).
485 * The rule is that addr 0 and floor 0 refer to the bottom of
486 * the address space, but end 0 and ceiling 0 refer to the top
487 * Comparisons need to use "end - 1" and "ceiling - 1" (though
488 * that end 0 case should be mythical).
489 *
490 * Wherever addr is brought up or ceiling brought down, we must
491 * be careful to reject "the opposite 0" before it confuses the
492 * subsequent tests. But what about where end is brought down
493 * by PMD_SIZE below? no, end can't go down to 0 there.
494 *
495 * Whereas we round start (addr) and ceiling down, by different
496 * masks at different levels, in order to test whether a table
497 * now has no other vmas using it, so can be freed, we don't
498 * bother to round floor or end up - the tests don't need that.
499 */
500
501 addr &= PMD_MASK;
502 if (addr < floor) {
503 addr += PMD_SIZE;
504 if (!addr)
505 return;
506 }
507 if (ceiling) {
508 ceiling &= PMD_MASK;
509 if (!ceiling)
510 return;
511 }
512 if (end - 1 > ceiling - 1)
513 end -= PMD_SIZE;
514 if (addr > end - 1)
515 return;
516
517 pgd = pgd_offset(tlb->mm, addr);
518 do {
519 next = pgd_addr_end(addr, end);
520 if (pgd_none_or_clear_bad(pgd))
521 continue;
522 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
523 } while (pgd++, addr = next, addr != end);
524}
525
526void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
527 unsigned long floor, unsigned long ceiling)
528{
529 while (vma) {
530 struct vm_area_struct *next = vma->vm_next;
531 unsigned long addr = vma->vm_start;
532
533 /*
534 * Hide vma from rmap and truncate_pagecache before freeing
535 * pgtables
536 */
537 unlink_anon_vmas(vma);
538 unlink_file_vma(vma);
539
540 if (is_vm_hugetlb_page(vma)) {
541 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
542 floor, next? next->vm_start: ceiling);
543 } else {
544 /*
545 * Optimization: gather nearby vmas into one call down
546 */
547 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
548 && !is_vm_hugetlb_page(next)) {
549 vma = next;
550 next = vma->vm_next;
551 unlink_anon_vmas(vma);
552 unlink_file_vma(vma);
553 }
554 free_pgd_range(tlb, addr, vma->vm_end,
555 floor, next? next->vm_start: ceiling);
556 }
557 vma = next;
558 }
559}
560
561int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
562 pmd_t *pmd, unsigned long address)
563{
564 spinlock_t *ptl;
565 pgtable_t new = pte_alloc_one(mm, address);
566 int wait_split_huge_page;
567 if (!new)
568 return -ENOMEM;
569
570 /*
571 * Ensure all pte setup (eg. pte page lock and page clearing) are
572 * visible before the pte is made visible to other CPUs by being
573 * put into page tables.
574 *
575 * The other side of the story is the pointer chasing in the page
576 * table walking code (when walking the page table without locking;
577 * ie. most of the time). Fortunately, these data accesses consist
578 * of a chain of data-dependent loads, meaning most CPUs (alpha
579 * being the notable exception) will already guarantee loads are
580 * seen in-order. See the alpha page table accessors for the
581 * smp_read_barrier_depends() barriers in page table walking code.
582 */
583 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
584
585 ptl = pmd_lock(mm, pmd);
586 wait_split_huge_page = 0;
587 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
588 atomic_long_inc(&mm->nr_ptes);
589 pmd_populate(mm, pmd, new);
590 new = NULL;
591 } else if (unlikely(pmd_trans_splitting(*pmd)))
592 wait_split_huge_page = 1;
593 spin_unlock(ptl);
594 if (new)
595 pte_free(mm, new);
596 if (wait_split_huge_page)
597 wait_split_huge_page(vma->anon_vma, pmd);
598 return 0;
599}
600
601int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
602{
603 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
604 if (!new)
605 return -ENOMEM;
606
607 smp_wmb(); /* See comment in __pte_alloc */
608
609 spin_lock(&init_mm.page_table_lock);
610 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
611 pmd_populate_kernel(&init_mm, pmd, new);
612 new = NULL;
613 } else
614 VM_BUG_ON(pmd_trans_splitting(*pmd));
615 spin_unlock(&init_mm.page_table_lock);
616 if (new)
617 pte_free_kernel(&init_mm, new);
618 return 0;
619}
620
621static inline void init_rss_vec(int *rss)
622{
623 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
624}
625
626static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
627{
628 int i;
629
630 if (current->mm == mm)
631 sync_mm_rss(mm);
632 for (i = 0; i < NR_MM_COUNTERS; i++)
633 if (rss[i])
634 add_mm_counter(mm, i, rss[i]);
635}
636
637/*
638 * This function is called to print an error when a bad pte
639 * is found. For example, we might have a PFN-mapped pte in
640 * a region that doesn't allow it.
641 *
642 * The calling function must still handle the error.
643 */
644static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
645 pte_t pte, struct page *page)
646{
647 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
648 pud_t *pud = pud_offset(pgd, addr);
649 pmd_t *pmd = pmd_offset(pud, addr);
650 struct address_space *mapping;
651 pgoff_t index;
652 static unsigned long resume;
653 static unsigned long nr_shown;
654 static unsigned long nr_unshown;
655
656 /*
657 * Allow a burst of 60 reports, then keep quiet for that minute;
658 * or allow a steady drip of one report per second.
659 */
660 if (nr_shown == 60) {
661 if (time_before(jiffies, resume)) {
662 nr_unshown++;
663 return;
664 }
665 if (nr_unshown) {
666 printk(KERN_ALERT
667 "BUG: Bad page map: %lu messages suppressed\n",
668 nr_unshown);
669 nr_unshown = 0;
670 }
671 nr_shown = 0;
672 }
673 if (nr_shown++ == 0)
674 resume = jiffies + 60 * HZ;
675
676 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
677 index = linear_page_index(vma, addr);
678
679 printk(KERN_ALERT
680 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
681 current->comm,
682 (long long)pte_val(pte), (long long)pmd_val(*pmd));
683 if (page)
684 dump_page(page, "bad pte");
685 printk(KERN_ALERT
686 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
687 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
688 /*
689 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
690 */
691 if (vma->vm_ops)
692 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
693 vma->vm_ops->fault);
694 if (vma->vm_file)
695 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
696 vma->vm_file->f_op->mmap);
697 dump_stack();
698 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
699}
700
701static inline bool is_cow_mapping(vm_flags_t flags)
702{
703 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
704}
705
706/*
707 * vm_normal_page -- This function gets the "struct page" associated with a pte.
708 *
709 * "Special" mappings do not wish to be associated with a "struct page" (either
710 * it doesn't exist, or it exists but they don't want to touch it). In this
711 * case, NULL is returned here. "Normal" mappings do have a struct page.
712 *
713 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
714 * pte bit, in which case this function is trivial. Secondly, an architecture
715 * may not have a spare pte bit, which requires a more complicated scheme,
716 * described below.
717 *
718 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
719 * special mapping (even if there are underlying and valid "struct pages").
720 * COWed pages of a VM_PFNMAP are always normal.
721 *
722 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
723 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
724 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
725 * mapping will always honor the rule
726 *
727 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
728 *
729 * And for normal mappings this is false.
730 *
731 * This restricts such mappings to be a linear translation from virtual address
732 * to pfn. To get around this restriction, we allow arbitrary mappings so long
733 * as the vma is not a COW mapping; in that case, we know that all ptes are
734 * special (because none can have been COWed).
735 *
736 *
737 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
738 *
739 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
740 * page" backing, however the difference is that _all_ pages with a struct
741 * page (that is, those where pfn_valid is true) are refcounted and considered
742 * normal pages by the VM. The disadvantage is that pages are refcounted
743 * (which can be slower and simply not an option for some PFNMAP users). The
744 * advantage is that we don't have to follow the strict linearity rule of
745 * PFNMAP mappings in order to support COWable mappings.
746 *
747 */
748#ifdef __HAVE_ARCH_PTE_SPECIAL
749# define HAVE_PTE_SPECIAL 1
750#else
751# define HAVE_PTE_SPECIAL 0
752#endif
753struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
754 pte_t pte)
755{
756 unsigned long pfn = pte_pfn(pte);
757
758 if (HAVE_PTE_SPECIAL) {
759 if (likely(!pte_special(pte)))
760 goto check_pfn;
761 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
762 return NULL;
763 if (!is_zero_pfn(pfn))
764 print_bad_pte(vma, addr, pte, NULL);
765 return NULL;
766 }
767
768 /* !HAVE_PTE_SPECIAL case follows: */
769
770 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
771 if (vma->vm_flags & VM_MIXEDMAP) {
772 if (!pfn_valid(pfn))
773 return NULL;
774 goto out;
775 } else {
776 unsigned long off;
777 off = (addr - vma->vm_start) >> PAGE_SHIFT;
778 if (pfn == vma->vm_pgoff + off)
779 return NULL;
780 if (!is_cow_mapping(vma->vm_flags))
781 return NULL;
782 }
783 }
784
785 if (is_zero_pfn(pfn))
786 return NULL;
787check_pfn:
788 if (unlikely(pfn > highest_memmap_pfn)) {
789 print_bad_pte(vma, addr, pte, NULL);
790 return NULL;
791 }
792
793 /*
794 * NOTE! We still have PageReserved() pages in the page tables.
795 * eg. VDSO mappings can cause them to exist.
796 */
797out:
798 return pfn_to_page(pfn);
799}
800
801/*
802 * copy one vm_area from one task to the other. Assumes the page tables
803 * already present in the new task to be cleared in the whole range
804 * covered by this vma.
805 */
806
807static inline unsigned long
808copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
809 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
810 unsigned long addr, int *rss)
811{
812 unsigned long vm_flags = vma->vm_flags;
813 pte_t pte = *src_pte;
814 struct page *page;
815
816 /* pte contains position in swap or file, so copy. */
817 if (unlikely(!pte_present(pte))) {
818 if (!pte_file(pte)) {
819 swp_entry_t entry = pte_to_swp_entry(pte);
820
821 if (swap_duplicate(entry) < 0)
822 return entry.val;
823
824 /* make sure dst_mm is on swapoff's mmlist. */
825 if (unlikely(list_empty(&dst_mm->mmlist))) {
826 spin_lock(&mmlist_lock);
827 if (list_empty(&dst_mm->mmlist))
828 list_add(&dst_mm->mmlist,
829 &src_mm->mmlist);
830 spin_unlock(&mmlist_lock);
831 }
832 if (likely(!non_swap_entry(entry)))
833 rss[MM_SWAPENTS]++;
834 else if (is_migration_entry(entry)) {
835 page = migration_entry_to_page(entry);
836
837 if (PageAnon(page))
838 rss[MM_ANONPAGES]++;
839 else
840 rss[MM_FILEPAGES]++;
841
842 if (is_write_migration_entry(entry) &&
843 is_cow_mapping(vm_flags)) {
844 /*
845 * COW mappings require pages in both
846 * parent and child to be set to read.
847 */
848 make_migration_entry_read(&entry);
849 pte = swp_entry_to_pte(entry);
850 if (pte_swp_soft_dirty(*src_pte))
851 pte = pte_swp_mksoft_dirty(pte);
852 set_pte_at(src_mm, addr, src_pte, pte);
853 }
854 }
855 }
856 goto out_set_pte;
857 }
858
859 /*
860 * If it's a COW mapping, write protect it both
861 * in the parent and the child
862 */
863 if (is_cow_mapping(vm_flags)) {
864 ptep_set_wrprotect(src_mm, addr, src_pte);
865 pte = pte_wrprotect(pte);
866 }
867
868 /*
869 * If it's a shared mapping, mark it clean in
870 * the child
871 */
872 if (vm_flags & VM_SHARED)
873 pte = pte_mkclean(pte);
874 pte = pte_mkold(pte);
875
876 page = vm_normal_page(vma, addr, pte);
877 if (page) {
878 get_page(page);
879 page_dup_rmap(page);
880 if (PageAnon(page))
881 rss[MM_ANONPAGES]++;
882 else
883 rss[MM_FILEPAGES]++;
884 }
885
886out_set_pte:
887 set_pte_at(dst_mm, addr, dst_pte, pte);
888 return 0;
889}
890
891int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
892 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
893 unsigned long addr, unsigned long end)
894{
895 pte_t *orig_src_pte, *orig_dst_pte;
896 pte_t *src_pte, *dst_pte;
897 spinlock_t *src_ptl, *dst_ptl;
898 int progress = 0;
899 int rss[NR_MM_COUNTERS];
900 swp_entry_t entry = (swp_entry_t){0};
901
902again:
903 init_rss_vec(rss);
904
905 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
906 if (!dst_pte)
907 return -ENOMEM;
908 src_pte = pte_offset_map(src_pmd, addr);
909 src_ptl = pte_lockptr(src_mm, src_pmd);
910 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
911 orig_src_pte = src_pte;
912 orig_dst_pte = dst_pte;
913 arch_enter_lazy_mmu_mode();
914
915 do {
916 /*
917 * We are holding two locks at this point - either of them
918 * could generate latencies in another task on another CPU.
919 */
920 if (progress >= 32) {
921 progress = 0;
922 if (need_resched() ||
923 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
924 break;
925 }
926 if (pte_none(*src_pte)) {
927 progress++;
928 continue;
929 }
930 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
931 vma, addr, rss);
932 if (entry.val)
933 break;
934 progress += 8;
935 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
936
937 arch_leave_lazy_mmu_mode();
938 spin_unlock(src_ptl);
939 pte_unmap(orig_src_pte);
940 add_mm_rss_vec(dst_mm, rss);
941 pte_unmap_unlock(orig_dst_pte, dst_ptl);
942 cond_resched();
943
944 if (entry.val) {
945 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
946 return -ENOMEM;
947 progress = 0;
948 }
949 if (addr != end)
950 goto again;
951 return 0;
952}
953
954static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
955 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
956 unsigned long addr, unsigned long end)
957{
958 pmd_t *src_pmd, *dst_pmd;
959 unsigned long next;
960
961 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
962 if (!dst_pmd)
963 return -ENOMEM;
964 src_pmd = pmd_offset(src_pud, addr);
965 do {
966 next = pmd_addr_end(addr, end);
967 if (pmd_trans_huge(*src_pmd)) {
968 int err;
969 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
970 err = copy_huge_pmd(dst_mm, src_mm,
971 dst_pmd, src_pmd, addr, vma);
972 if (err == -ENOMEM)
973 return -ENOMEM;
974 if (!err)
975 continue;
976 /* fall through */
977 }
978 if (pmd_none_or_clear_bad(src_pmd))
979 continue;
980 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
981 vma, addr, next))
982 return -ENOMEM;
983 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
984 return 0;
985}
986
987static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
988 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
989 unsigned long addr, unsigned long end)
990{
991 pud_t *src_pud, *dst_pud;
992 unsigned long next;
993
994 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
995 if (!dst_pud)
996 return -ENOMEM;
997 src_pud = pud_offset(src_pgd, addr);
998 do {
999 next = pud_addr_end(addr, end);
1000 if (pud_none_or_clear_bad(src_pud))
1001 continue;
1002 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1003 vma, addr, next))
1004 return -ENOMEM;
1005 } while (dst_pud++, src_pud++, addr = next, addr != end);
1006 return 0;
1007}
1008
1009int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1010 struct vm_area_struct *vma)
1011{
1012 pgd_t *src_pgd, *dst_pgd;
1013 unsigned long next;
1014 unsigned long addr = vma->vm_start;
1015 unsigned long end = vma->vm_end;
1016 unsigned long mmun_start; /* For mmu_notifiers */
1017 unsigned long mmun_end; /* For mmu_notifiers */
1018 bool is_cow;
1019 int ret;
1020
1021 /*
1022 * Don't copy ptes where a page fault will fill them correctly.
1023 * Fork becomes much lighter when there are big shared or private
1024 * readonly mappings. The tradeoff is that copy_page_range is more
1025 * efficient than faulting.
1026 */
1027 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1028 VM_PFNMAP | VM_MIXEDMAP))) {
1029 if (!vma->anon_vma)
1030 return 0;
1031 }
1032
1033 if (is_vm_hugetlb_page(vma))
1034 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1035
1036 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1037 /*
1038 * We do not free on error cases below as remove_vma
1039 * gets called on error from higher level routine
1040 */
1041 ret = track_pfn_copy(vma);
1042 if (ret)
1043 return ret;
1044 }
1045
1046 /*
1047 * We need to invalidate the secondary MMU mappings only when
1048 * there could be a permission downgrade on the ptes of the
1049 * parent mm. And a permission downgrade will only happen if
1050 * is_cow_mapping() returns true.
1051 */
1052 is_cow = is_cow_mapping(vma->vm_flags);
1053 mmun_start = addr;
1054 mmun_end = end;
1055 if (is_cow)
1056 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1057 mmun_end);
1058
1059 ret = 0;
1060 dst_pgd = pgd_offset(dst_mm, addr);
1061 src_pgd = pgd_offset(src_mm, addr);
1062 do {
1063 next = pgd_addr_end(addr, end);
1064 if (pgd_none_or_clear_bad(src_pgd))
1065 continue;
1066 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1067 vma, addr, next))) {
1068 ret = -ENOMEM;
1069 break;
1070 }
1071 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1072
1073 if (is_cow)
1074 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1075 return ret;
1076}
1077
1078static unsigned long zap_pte_range(struct mmu_gather *tlb,
1079 struct vm_area_struct *vma, pmd_t *pmd,
1080 unsigned long addr, unsigned long end,
1081 struct zap_details *details)
1082{
1083 struct mm_struct *mm = tlb->mm;
1084 int force_flush = 0;
1085 int rss[NR_MM_COUNTERS];
1086 spinlock_t *ptl;
1087 pte_t *start_pte;
1088 pte_t *pte;
1089
1090again:
1091 init_rss_vec(rss);
1092 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1093 pte = start_pte;
1094 arch_enter_lazy_mmu_mode();
1095 do {
1096 pte_t ptent = *pte;
1097 if (pte_none(ptent)) {
1098 continue;
1099 }
1100
1101 if (pte_present(ptent)) {
1102 struct page *page;
1103
1104 page = vm_normal_page(vma, addr, ptent);
1105 if (unlikely(details) && page) {
1106 /*
1107 * unmap_shared_mapping_pages() wants to
1108 * invalidate cache without truncating:
1109 * unmap shared but keep private pages.
1110 */
1111 if (details->check_mapping &&
1112 details->check_mapping != page->mapping)
1113 continue;
1114 /*
1115 * Each page->index must be checked when
1116 * invalidating or truncating nonlinear.
1117 */
1118 if (details->nonlinear_vma &&
1119 (page->index < details->first_index ||
1120 page->index > details->last_index))
1121 continue;
1122 }
1123 ptent = ptep_get_and_clear_full(mm, addr, pte,
1124 tlb->fullmm);
1125 tlb_remove_tlb_entry(tlb, pte, addr);
1126 if (unlikely(!page))
1127 continue;
1128 if (unlikely(details) && details->nonlinear_vma
1129 && linear_page_index(details->nonlinear_vma,
1130 addr) != page->index) {
1131 pte_t ptfile = pgoff_to_pte(page->index);
1132 if (pte_soft_dirty(ptent))
1133 pte_file_mksoft_dirty(ptfile);
1134 set_pte_at(mm, addr, pte, ptfile);
1135 }
1136 if (PageAnon(page))
1137 rss[MM_ANONPAGES]--;
1138 else {
1139 if (pte_dirty(ptent)) {
1140 force_flush = 1;
1141 set_page_dirty(page);
1142 }
1143 if (pte_young(ptent) &&
1144 likely(!(vma->vm_flags & VM_SEQ_READ)))
1145 mark_page_accessed(page);
1146 rss[MM_FILEPAGES]--;
1147 }
1148 page_remove_rmap(page);
1149 if (unlikely(page_mapcount(page) < 0))
1150 print_bad_pte(vma, addr, ptent, page);
1151 if (unlikely(!__tlb_remove_page(tlb, page))) {
1152 force_flush = 1;
1153 break;
1154 }
1155 continue;
1156 }
1157 /*
1158 * If details->check_mapping, we leave swap entries;
1159 * if details->nonlinear_vma, we leave file entries.
1160 */
1161 if (unlikely(details))
1162 continue;
1163 if (pte_file(ptent)) {
1164 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1165 print_bad_pte(vma, addr, ptent, NULL);
1166 } else {
1167 swp_entry_t entry = pte_to_swp_entry(ptent);
1168
1169 if (!non_swap_entry(entry))
1170 rss[MM_SWAPENTS]--;
1171 else if (is_migration_entry(entry)) {
1172 struct page *page;
1173
1174 page = migration_entry_to_page(entry);
1175
1176 if (PageAnon(page))
1177 rss[MM_ANONPAGES]--;
1178 else
1179 rss[MM_FILEPAGES]--;
1180 }
1181 if (unlikely(!free_swap_and_cache(entry)))
1182 print_bad_pte(vma, addr, ptent, NULL);
1183 }
1184 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1185 } while (pte++, addr += PAGE_SIZE, addr != end);
1186
1187 add_mm_rss_vec(mm, rss);
1188 arch_leave_lazy_mmu_mode();
1189
1190 /* Do the actual TLB flush before dropping ptl */
1191 if (force_flush) {
1192 unsigned long old_end;
1193
1194 /*
1195 * Flush the TLB just for the previous segment,
1196 * then update the range to be the remaining
1197 * TLB range.
1198 */
1199 old_end = tlb->end;
1200 tlb->end = addr;
1201 tlb_flush_mmu_tlbonly(tlb);
1202 tlb->start = addr;
1203 tlb->end = old_end;
1204 }
1205 pte_unmap_unlock(start_pte, ptl);
1206
1207 /*
1208 * If we forced a TLB flush (either due to running out of
1209 * batch buffers or because we needed to flush dirty TLB
1210 * entries before releasing the ptl), free the batched
1211 * memory too. Restart if we didn't do everything.
1212 */
1213 if (force_flush) {
1214 force_flush = 0;
1215 tlb_flush_mmu_free(tlb);
1216
1217 if (addr != end)
1218 goto again;
1219 }
1220
1221 return addr;
1222}
1223
1224static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pud_t *pud,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1228{
1229 pmd_t *pmd;
1230 unsigned long next;
1231
1232 pmd = pmd_offset(pud, addr);
1233 do {
1234 next = pmd_addr_end(addr, end);
1235 if (pmd_trans_huge(*pmd)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237#ifdef CONFIG_DEBUG_VM
1238 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240 __func__, addr, end,
1241 vma->vm_start,
1242 vma->vm_end);
1243 BUG();
1244 }
1245#endif
1246 split_huge_page_pmd(vma, addr, pmd);
1247 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1248 goto next;
1249 /* fall through */
1250 }
1251 /*
1252 * Here there can be other concurrent MADV_DONTNEED or
1253 * trans huge page faults running, and if the pmd is
1254 * none or trans huge it can change under us. This is
1255 * because MADV_DONTNEED holds the mmap_sem in read
1256 * mode.
1257 */
1258 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1259 goto next;
1260 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1261next:
1262 cond_resched();
1263 } while (pmd++, addr = next, addr != end);
1264
1265 return addr;
1266}
1267
1268static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1269 struct vm_area_struct *vma, pgd_t *pgd,
1270 unsigned long addr, unsigned long end,
1271 struct zap_details *details)
1272{
1273 pud_t *pud;
1274 unsigned long next;
1275
1276 pud = pud_offset(pgd, addr);
1277 do {
1278 next = pud_addr_end(addr, end);
1279 if (pud_none_or_clear_bad(pud))
1280 continue;
1281 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282 } while (pud++, addr = next, addr != end);
1283
1284 return addr;
1285}
1286
1287static void unmap_page_range(struct mmu_gather *tlb,
1288 struct vm_area_struct *vma,
1289 unsigned long addr, unsigned long end,
1290 struct zap_details *details)
1291{
1292 pgd_t *pgd;
1293 unsigned long next;
1294
1295 if (details && !details->check_mapping && !details->nonlinear_vma)
1296 details = NULL;
1297
1298 BUG_ON(addr >= end);
1299 mem_cgroup_uncharge_start();
1300 tlb_start_vma(tlb, vma);
1301 pgd = pgd_offset(vma->vm_mm, addr);
1302 do {
1303 next = pgd_addr_end(addr, end);
1304 if (pgd_none_or_clear_bad(pgd))
1305 continue;
1306 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307 } while (pgd++, addr = next, addr != end);
1308 tlb_end_vma(tlb, vma);
1309 mem_cgroup_uncharge_end();
1310}
1311
1312
1313static void unmap_single_vma(struct mmu_gather *tlb,
1314 struct vm_area_struct *vma, unsigned long start_addr,
1315 unsigned long end_addr,
1316 struct zap_details *details)
1317{
1318 unsigned long start = max(vma->vm_start, start_addr);
1319 unsigned long end;
1320
1321 if (start >= vma->vm_end)
1322 return;
1323 end = min(vma->vm_end, end_addr);
1324 if (end <= vma->vm_start)
1325 return;
1326
1327 if (vma->vm_file)
1328 uprobe_munmap(vma, start, end);
1329
1330 if (unlikely(vma->vm_flags & VM_PFNMAP))
1331 untrack_pfn(vma, 0, 0);
1332
1333 if (start != end) {
1334 if (unlikely(is_vm_hugetlb_page(vma))) {
1335 /*
1336 * It is undesirable to test vma->vm_file as it
1337 * should be non-null for valid hugetlb area.
1338 * However, vm_file will be NULL in the error
1339 * cleanup path of mmap_region. When
1340 * hugetlbfs ->mmap method fails,
1341 * mmap_region() nullifies vma->vm_file
1342 * before calling this function to clean up.
1343 * Since no pte has actually been setup, it is
1344 * safe to do nothing in this case.
1345 */
1346 if (vma->vm_file) {
1347 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1348 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1349 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1350 }
1351 } else
1352 unmap_page_range(tlb, vma, start, end, details);
1353 }
1354}
1355
1356/**
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
1358 * @tlb: address of the caller's struct mmu_gather
1359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
1362 *
1363 * Unmap all pages in the vma list.
1364 *
1365 * Only addresses between `start' and `end' will be unmapped.
1366 *
1367 * The VMA list must be sorted in ascending virtual address order.
1368 *
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns. So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1373 */
1374void unmap_vmas(struct mmu_gather *tlb,
1375 struct vm_area_struct *vma, unsigned long start_addr,
1376 unsigned long end_addr)
1377{
1378 struct mm_struct *mm = vma->vm_mm;
1379
1380 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1381 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1382 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1383 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1384}
1385
1386/**
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
1389 * @start: starting address of pages to zap
1390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
1392 *
1393 * Caller must protect the VMA list
1394 */
1395void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1396 unsigned long size, struct zap_details *details)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct mmu_gather tlb;
1400 unsigned long end = start + size;
1401
1402 lru_add_drain();
1403 tlb_gather_mmu(&tlb, mm, start, end);
1404 update_hiwater_rss(mm);
1405 mmu_notifier_invalidate_range_start(mm, start, end);
1406 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1407 unmap_single_vma(&tlb, vma, start, end, details);
1408 mmu_notifier_invalidate_range_end(mm, start, end);
1409 tlb_finish_mmu(&tlb, start, end);
1410}
1411
1412/**
1413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1418 *
1419 * The range must fit into one VMA.
1420 */
1421static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422 unsigned long size, struct zap_details *details)
1423{
1424 struct mm_struct *mm = vma->vm_mm;
1425 struct mmu_gather tlb;
1426 unsigned long end = address + size;
1427
1428 lru_add_drain();
1429 tlb_gather_mmu(&tlb, mm, address, end);
1430 update_hiwater_rss(mm);
1431 mmu_notifier_invalidate_range_start(mm, address, end);
1432 unmap_single_vma(&tlb, vma, address, end, details);
1433 mmu_notifier_invalidate_range_end(mm, address, end);
1434 tlb_finish_mmu(&tlb, address, end);
1435}
1436
1437/**
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1442 *
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 *
1445 * The entire address range must be fully contained within the vma.
1446 *
1447 * Returns 0 if successful.
1448 */
1449int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1450 unsigned long size)
1451{
1452 if (address < vma->vm_start || address + size > vma->vm_end ||
1453 !(vma->vm_flags & VM_PFNMAP))
1454 return -1;
1455 zap_page_range_single(vma, address, size, NULL);
1456 return 0;
1457}
1458EXPORT_SYMBOL_GPL(zap_vma_ptes);
1459
1460/**
1461 * follow_page_mask - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1465 * @page_mask: on output, *page_mask is set according to the size of the page
1466 *
1467 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1468 *
1469 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1470 * an error pointer if there is a mapping to something not represented
1471 * by a page descriptor (see also vm_normal_page()).
1472 */
1473struct page *follow_page_mask(struct vm_area_struct *vma,
1474 unsigned long address, unsigned int flags,
1475 unsigned int *page_mask)
1476{
1477 pgd_t *pgd;
1478 pud_t *pud;
1479 pmd_t *pmd;
1480 pte_t *ptep, pte;
1481 spinlock_t *ptl;
1482 struct page *page;
1483 struct mm_struct *mm = vma->vm_mm;
1484
1485 *page_mask = 0;
1486
1487 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1488 if (!IS_ERR(page)) {
1489 BUG_ON(flags & FOLL_GET);
1490 goto out;
1491 }
1492
1493 page = NULL;
1494 pgd = pgd_offset(mm, address);
1495 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1496 goto no_page_table;
1497
1498 pud = pud_offset(pgd, address);
1499 if (pud_none(*pud))
1500 goto no_page_table;
1501 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1502 if (flags & FOLL_GET)
1503 goto out;
1504 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1505 goto out;
1506 }
1507 if (unlikely(pud_bad(*pud)))
1508 goto no_page_table;
1509
1510 pmd = pmd_offset(pud, address);
1511 if (pmd_none(*pmd))
1512 goto no_page_table;
1513 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1514 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1515 if (flags & FOLL_GET) {
1516 /*
1517 * Refcount on tail pages are not well-defined and
1518 * shouldn't be taken. The caller should handle a NULL
1519 * return when trying to follow tail pages.
1520 */
1521 if (PageHead(page))
1522 get_page(page);
1523 else {
1524 page = NULL;
1525 goto out;
1526 }
1527 }
1528 goto out;
1529 }
1530 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1531 goto no_page_table;
1532 if (pmd_trans_huge(*pmd)) {
1533 if (flags & FOLL_SPLIT) {
1534 split_huge_page_pmd(vma, address, pmd);
1535 goto split_fallthrough;
1536 }
1537 ptl = pmd_lock(mm, pmd);
1538 if (likely(pmd_trans_huge(*pmd))) {
1539 if (unlikely(pmd_trans_splitting(*pmd))) {
1540 spin_unlock(ptl);
1541 wait_split_huge_page(vma->anon_vma, pmd);
1542 } else {
1543 page = follow_trans_huge_pmd(vma, address,
1544 pmd, flags);
1545 spin_unlock(ptl);
1546 *page_mask = HPAGE_PMD_NR - 1;
1547 goto out;
1548 }
1549 } else
1550 spin_unlock(ptl);
1551 /* fall through */
1552 }
1553split_fallthrough:
1554 if (unlikely(pmd_bad(*pmd)))
1555 goto no_page_table;
1556
1557 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1558
1559 pte = *ptep;
1560 if (!pte_present(pte)) {
1561 swp_entry_t entry;
1562 /*
1563 * KSM's break_ksm() relies upon recognizing a ksm page
1564 * even while it is being migrated, so for that case we
1565 * need migration_entry_wait().
1566 */
1567 if (likely(!(flags & FOLL_MIGRATION)))
1568 goto no_page;
1569 if (pte_none(pte) || pte_file(pte))
1570 goto no_page;
1571 entry = pte_to_swp_entry(pte);
1572 if (!is_migration_entry(entry))
1573 goto no_page;
1574 pte_unmap_unlock(ptep, ptl);
1575 migration_entry_wait(mm, pmd, address);
1576 goto split_fallthrough;
1577 }
1578 if ((flags & FOLL_NUMA) && pte_numa(pte))
1579 goto no_page;
1580 if ((flags & FOLL_WRITE) && !pte_write(pte))
1581 goto unlock;
1582
1583 page = vm_normal_page(vma, address, pte);
1584 if (unlikely(!page)) {
1585 if ((flags & FOLL_DUMP) ||
1586 !is_zero_pfn(pte_pfn(pte)))
1587 goto bad_page;
1588 page = pte_page(pte);
1589 }
1590
1591 if (flags & FOLL_GET)
1592 get_page_foll(page);
1593 if (flags & FOLL_TOUCH) {
1594 if ((flags & FOLL_WRITE) &&
1595 !pte_dirty(pte) && !PageDirty(page))
1596 set_page_dirty(page);
1597 /*
1598 * pte_mkyoung() would be more correct here, but atomic care
1599 * is needed to avoid losing the dirty bit: it is easier to use
1600 * mark_page_accessed().
1601 */
1602 mark_page_accessed(page);
1603 }
1604 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1605 /*
1606 * The preliminary mapping check is mainly to avoid the
1607 * pointless overhead of lock_page on the ZERO_PAGE
1608 * which might bounce very badly if there is contention.
1609 *
1610 * If the page is already locked, we don't need to
1611 * handle it now - vmscan will handle it later if and
1612 * when it attempts to reclaim the page.
1613 */
1614 if (page->mapping && trylock_page(page)) {
1615 lru_add_drain(); /* push cached pages to LRU */
1616 /*
1617 * Because we lock page here, and migration is
1618 * blocked by the pte's page reference, and we
1619 * know the page is still mapped, we don't even
1620 * need to check for file-cache page truncation.
1621 */
1622 mlock_vma_page(page);
1623 unlock_page(page);
1624 }
1625 }
1626unlock:
1627 pte_unmap_unlock(ptep, ptl);
1628out:
1629 return page;
1630
1631bad_page:
1632 pte_unmap_unlock(ptep, ptl);
1633 return ERR_PTR(-EFAULT);
1634
1635no_page:
1636 pte_unmap_unlock(ptep, ptl);
1637 if (!pte_none(pte))
1638 return page;
1639
1640no_page_table:
1641 /*
1642 * When core dumping an enormous anonymous area that nobody
1643 * has touched so far, we don't want to allocate unnecessary pages or
1644 * page tables. Return error instead of NULL to skip handle_mm_fault,
1645 * then get_dump_page() will return NULL to leave a hole in the dump.
1646 * But we can only make this optimization where a hole would surely
1647 * be zero-filled if handle_mm_fault() actually did handle it.
1648 */
1649 if ((flags & FOLL_DUMP) &&
1650 (!vma->vm_ops || !vma->vm_ops->fault))
1651 return ERR_PTR(-EFAULT);
1652 return page;
1653}
1654
1655static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1656{
1657 return stack_guard_page_start(vma, addr) ||
1658 stack_guard_page_end(vma, addr+PAGE_SIZE);
1659}
1660
1661/**
1662 * __get_user_pages() - pin user pages in memory
1663 * @tsk: task_struct of target task
1664 * @mm: mm_struct of target mm
1665 * @start: starting user address
1666 * @nr_pages: number of pages from start to pin
1667 * @gup_flags: flags modifying pin behaviour
1668 * @pages: array that receives pointers to the pages pinned.
1669 * Should be at least nr_pages long. Or NULL, if caller
1670 * only intends to ensure the pages are faulted in.
1671 * @vmas: array of pointers to vmas corresponding to each page.
1672 * Or NULL if the caller does not require them.
1673 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1674 *
1675 * Returns number of pages pinned. This may be fewer than the number
1676 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1677 * were pinned, returns -errno. Each page returned must be released
1678 * with a put_page() call when it is finished with. vmas will only
1679 * remain valid while mmap_sem is held.
1680 *
1681 * Must be called with mmap_sem held for read or write.
1682 *
1683 * __get_user_pages walks a process's page tables and takes a reference to
1684 * each struct page that each user address corresponds to at a given
1685 * instant. That is, it takes the page that would be accessed if a user
1686 * thread accesses the given user virtual address at that instant.
1687 *
1688 * This does not guarantee that the page exists in the user mappings when
1689 * __get_user_pages returns, and there may even be a completely different
1690 * page there in some cases (eg. if mmapped pagecache has been invalidated
1691 * and subsequently re faulted). However it does guarantee that the page
1692 * won't be freed completely. And mostly callers simply care that the page
1693 * contains data that was valid *at some point in time*. Typically, an IO
1694 * or similar operation cannot guarantee anything stronger anyway because
1695 * locks can't be held over the syscall boundary.
1696 *
1697 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1698 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1699 * appropriate) must be called after the page is finished with, and
1700 * before put_page is called.
1701 *
1702 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1703 * or mmap_sem contention, and if waiting is needed to pin all pages,
1704 * *@nonblocking will be set to 0.
1705 *
1706 * In most cases, get_user_pages or get_user_pages_fast should be used
1707 * instead of __get_user_pages. __get_user_pages should be used only if
1708 * you need some special @gup_flags.
1709 */
1710long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1711 unsigned long start, unsigned long nr_pages,
1712 unsigned int gup_flags, struct page **pages,
1713 struct vm_area_struct **vmas, int *nonblocking)
1714{
1715 long i;
1716 unsigned long vm_flags;
1717 unsigned int page_mask;
1718
1719 if (!nr_pages)
1720 return 0;
1721
1722 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1723
1724 /*
1725 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1726 * would be called on PROT_NONE ranges. We must never invoke
1727 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1728 * page faults would unprotect the PROT_NONE ranges if
1729 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1730 * bitflag. So to avoid that, don't set FOLL_NUMA if
1731 * FOLL_FORCE is set.
1732 */
1733 if (!(gup_flags & FOLL_FORCE))
1734 gup_flags |= FOLL_NUMA;
1735
1736 i = 0;
1737
1738 do {
1739 struct vm_area_struct *vma;
1740
1741 vma = find_extend_vma(mm, start);
1742 if (!vma && in_gate_area(mm, start)) {
1743 unsigned long pg = start & PAGE_MASK;
1744 pgd_t *pgd;
1745 pud_t *pud;
1746 pmd_t *pmd;
1747 pte_t *pte;
1748
1749 /* user gate pages are read-only */
1750 if (gup_flags & FOLL_WRITE)
1751 goto efault;
1752 if (pg > TASK_SIZE)
1753 pgd = pgd_offset_k(pg);
1754 else
1755 pgd = pgd_offset_gate(mm, pg);
1756 BUG_ON(pgd_none(*pgd));
1757 pud = pud_offset(pgd, pg);
1758 BUG_ON(pud_none(*pud));
1759 pmd = pmd_offset(pud, pg);
1760 if (pmd_none(*pmd))
1761 goto efault;
1762 VM_BUG_ON(pmd_trans_huge(*pmd));
1763 pte = pte_offset_map(pmd, pg);
1764 if (pte_none(*pte)) {
1765 pte_unmap(pte);
1766 goto efault;
1767 }
1768 vma = get_gate_vma(mm);
1769 if (pages) {
1770 struct page *page;
1771
1772 page = vm_normal_page(vma, start, *pte);
1773 if (!page) {
1774 if (!(gup_flags & FOLL_DUMP) &&
1775 is_zero_pfn(pte_pfn(*pte)))
1776 page = pte_page(*pte);
1777 else {
1778 pte_unmap(pte);
1779 goto efault;
1780 }
1781 }
1782 pages[i] = page;
1783 get_page(page);
1784 }
1785 pte_unmap(pte);
1786 page_mask = 0;
1787 goto next_page;
1788 }
1789
1790 if (!vma)
1791 goto efault;
1792 vm_flags = vma->vm_flags;
1793 if (vm_flags & (VM_IO | VM_PFNMAP))
1794 goto efault;
1795
1796 if (gup_flags & FOLL_WRITE) {
1797 if (!(vm_flags & VM_WRITE)) {
1798 if (!(gup_flags & FOLL_FORCE))
1799 goto efault;
1800 /*
1801 * We used to let the write,force case do COW
1802 * in a VM_MAYWRITE VM_SHARED !VM_WRITE vma, so
1803 * ptrace could set a breakpoint in a read-only
1804 * mapping of an executable, without corrupting
1805 * the file (yet only when that file had been
1806 * opened for writing!). Anon pages in shared
1807 * mappings are surprising: now just reject it.
1808 */
1809 if (!is_cow_mapping(vm_flags)) {
1810 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
1811 goto efault;
1812 }
1813 }
1814 } else {
1815 if (!(vm_flags & VM_READ)) {
1816 if (!(gup_flags & FOLL_FORCE))
1817 goto efault;
1818 /*
1819 * Is there actually any vma we can reach here
1820 * which does not have VM_MAYREAD set?
1821 */
1822 if (!(vm_flags & VM_MAYREAD))
1823 goto efault;
1824 }
1825 }
1826
1827 if (is_vm_hugetlb_page(vma)) {
1828 i = follow_hugetlb_page(mm, vma, pages, vmas,
1829 &start, &nr_pages, i, gup_flags);
1830 continue;
1831 }
1832
1833 do {
1834 struct page *page;
1835 unsigned int foll_flags = gup_flags;
1836 unsigned int page_increm;
1837
1838 /*
1839 * If we have a pending SIGKILL, don't keep faulting
1840 * pages and potentially allocating memory.
1841 */
1842 if (unlikely(fatal_signal_pending(current)))
1843 return i ? i : -ERESTARTSYS;
1844
1845 cond_resched();
1846 while (!(page = follow_page_mask(vma, start,
1847 foll_flags, &page_mask))) {
1848 int ret;
1849 unsigned int fault_flags = 0;
1850
1851 /* For mlock, just skip the stack guard page. */
1852 if (foll_flags & FOLL_MLOCK) {
1853 if (stack_guard_page(vma, start))
1854 goto next_page;
1855 }
1856 if (foll_flags & FOLL_WRITE)
1857 fault_flags |= FAULT_FLAG_WRITE;
1858 if (nonblocking)
1859 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1860 if (foll_flags & FOLL_NOWAIT)
1861 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1862
1863 ret = handle_mm_fault(mm, vma, start,
1864 fault_flags);
1865
1866 if (ret & VM_FAULT_ERROR) {
1867 if (ret & VM_FAULT_OOM)
1868 return i ? i : -ENOMEM;
1869 if (ret & (VM_FAULT_HWPOISON |
1870 VM_FAULT_HWPOISON_LARGE)) {
1871 if (i)
1872 return i;
1873 else if (gup_flags & FOLL_HWPOISON)
1874 return -EHWPOISON;
1875 else
1876 return -EFAULT;
1877 }
1878 if (ret & VM_FAULT_SIGBUS)
1879 goto efault;
1880 BUG();
1881 }
1882
1883 if (tsk) {
1884 if (ret & VM_FAULT_MAJOR)
1885 tsk->maj_flt++;
1886 else
1887 tsk->min_flt++;
1888 }
1889
1890 if (ret & VM_FAULT_RETRY) {
1891 if (nonblocking)
1892 *nonblocking = 0;
1893 return i;
1894 }
1895
1896 /*
1897 * The VM_FAULT_WRITE bit tells us that
1898 * do_wp_page has broken COW when necessary,
1899 * even if maybe_mkwrite decided not to set
1900 * pte_write. We can thus safely do subsequent
1901 * page lookups as if they were reads. But only
1902 * do so when looping for pte_write is futile:
1903 * in some cases userspace may also be wanting
1904 * to write to the gotten user page, which a
1905 * read fault here might prevent (a readonly
1906 * page might get reCOWed by userspace write).
1907 */
1908 if ((ret & VM_FAULT_WRITE) &&
1909 !(vma->vm_flags & VM_WRITE))
1910 foll_flags &= ~FOLL_WRITE;
1911
1912 cond_resched();
1913 }
1914 if (IS_ERR(page))
1915 return i ? i : PTR_ERR(page);
1916 if (pages) {
1917 pages[i] = page;
1918
1919 flush_anon_page(vma, page, start);
1920 flush_dcache_page(page);
1921 page_mask = 0;
1922 }
1923next_page:
1924 if (vmas) {
1925 vmas[i] = vma;
1926 page_mask = 0;
1927 }
1928 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1929 if (page_increm > nr_pages)
1930 page_increm = nr_pages;
1931 i += page_increm;
1932 start += page_increm * PAGE_SIZE;
1933 nr_pages -= page_increm;
1934 } while (nr_pages && start < vma->vm_end);
1935 } while (nr_pages);
1936 return i;
1937efault:
1938 return i ? : -EFAULT;
1939}
1940EXPORT_SYMBOL(__get_user_pages);
1941
1942/*
1943 * fixup_user_fault() - manually resolve a user page fault
1944 * @tsk: the task_struct to use for page fault accounting, or
1945 * NULL if faults are not to be recorded.
1946 * @mm: mm_struct of target mm
1947 * @address: user address
1948 * @fault_flags:flags to pass down to handle_mm_fault()
1949 *
1950 * This is meant to be called in the specific scenario where for locking reasons
1951 * we try to access user memory in atomic context (within a pagefault_disable()
1952 * section), this returns -EFAULT, and we want to resolve the user fault before
1953 * trying again.
1954 *
1955 * Typically this is meant to be used by the futex code.
1956 *
1957 * The main difference with get_user_pages() is that this function will
1958 * unconditionally call handle_mm_fault() which will in turn perform all the
1959 * necessary SW fixup of the dirty and young bits in the PTE, while
1960 * handle_mm_fault() only guarantees to update these in the struct page.
1961 *
1962 * This is important for some architectures where those bits also gate the
1963 * access permission to the page because they are maintained in software. On
1964 * such architectures, gup() will not be enough to make a subsequent access
1965 * succeed.
1966 *
1967 * This should be called with the mm_sem held for read.
1968 */
1969int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1970 unsigned long address, unsigned int fault_flags)
1971{
1972 struct vm_area_struct *vma;
1973 vm_flags_t vm_flags;
1974 int ret;
1975
1976 vma = find_extend_vma(mm, address);
1977 if (!vma || address < vma->vm_start)
1978 return -EFAULT;
1979
1980 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1981 if (!(vm_flags & vma->vm_flags))
1982 return -EFAULT;
1983
1984 ret = handle_mm_fault(mm, vma, address, fault_flags);
1985 if (ret & VM_FAULT_ERROR) {
1986 if (ret & VM_FAULT_OOM)
1987 return -ENOMEM;
1988 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1989 return -EHWPOISON;
1990 if (ret & VM_FAULT_SIGBUS)
1991 return -EFAULT;
1992 BUG();
1993 }
1994 if (tsk) {
1995 if (ret & VM_FAULT_MAJOR)
1996 tsk->maj_flt++;
1997 else
1998 tsk->min_flt++;
1999 }
2000 return 0;
2001}
2002
2003/*
2004 * get_user_pages() - pin user pages in memory
2005 * @tsk: the task_struct to use for page fault accounting, or
2006 * NULL if faults are not to be recorded.
2007 * @mm: mm_struct of target mm
2008 * @start: starting user address
2009 * @nr_pages: number of pages from start to pin
2010 * @write: whether pages will be written to by the caller
2011 * @force: whether to force access even when user mapping is currently
2012 * protected (but never forces write access to shared mapping).
2013 * @pages: array that receives pointers to the pages pinned.
2014 * Should be at least nr_pages long. Or NULL, if caller
2015 * only intends to ensure the pages are faulted in.
2016 * @vmas: array of pointers to vmas corresponding to each page.
2017 * Or NULL if the caller does not require them.
2018 *
2019 * Returns number of pages pinned. This may be fewer than the number
2020 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2021 * were pinned, returns -errno. Each page returned must be released
2022 * with a put_page() call when it is finished with. vmas will only
2023 * remain valid while mmap_sem is held.
2024 *
2025 * Must be called with mmap_sem held for read or write.
2026 *
2027 * get_user_pages walks a process's page tables and takes a reference to
2028 * each struct page that each user address corresponds to at a given
2029 * instant. That is, it takes the page that would be accessed if a user
2030 * thread accesses the given user virtual address at that instant.
2031 *
2032 * This does not guarantee that the page exists in the user mappings when
2033 * get_user_pages returns, and there may even be a completely different
2034 * page there in some cases (eg. if mmapped pagecache has been invalidated
2035 * and subsequently re faulted). However it does guarantee that the page
2036 * won't be freed completely. And mostly callers simply care that the page
2037 * contains data that was valid *at some point in time*. Typically, an IO
2038 * or similar operation cannot guarantee anything stronger anyway because
2039 * locks can't be held over the syscall boundary.
2040 *
2041 * If write=0, the page must not be written to. If the page is written to,
2042 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2043 * after the page is finished with, and before put_page is called.
2044 *
2045 * get_user_pages is typically used for fewer-copy IO operations, to get a
2046 * handle on the memory by some means other than accesses via the user virtual
2047 * addresses. The pages may be submitted for DMA to devices or accessed via
2048 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2049 * use the correct cache flushing APIs.
2050 *
2051 * See also get_user_pages_fast, for performance critical applications.
2052 */
2053long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2054 unsigned long start, unsigned long nr_pages, int write,
2055 int force, struct page **pages, struct vm_area_struct **vmas)
2056{
2057 int flags = FOLL_TOUCH;
2058
2059 if (pages)
2060 flags |= FOLL_GET;
2061 if (write)
2062 flags |= FOLL_WRITE;
2063 if (force)
2064 flags |= FOLL_FORCE;
2065
2066 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2067 NULL);
2068}
2069EXPORT_SYMBOL(get_user_pages);
2070
2071/**
2072 * get_dump_page() - pin user page in memory while writing it to core dump
2073 * @addr: user address
2074 *
2075 * Returns struct page pointer of user page pinned for dump,
2076 * to be freed afterwards by page_cache_release() or put_page().
2077 *
2078 * Returns NULL on any kind of failure - a hole must then be inserted into
2079 * the corefile, to preserve alignment with its headers; and also returns
2080 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2081 * allowing a hole to be left in the corefile to save diskspace.
2082 *
2083 * Called without mmap_sem, but after all other threads have been killed.
2084 */
2085#ifdef CONFIG_ELF_CORE
2086struct page *get_dump_page(unsigned long addr)
2087{
2088 struct vm_area_struct *vma;
2089 struct page *page;
2090
2091 if (__get_user_pages(current, current->mm, addr, 1,
2092 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2093 NULL) < 1)
2094 return NULL;
2095 flush_cache_page(vma, addr, page_to_pfn(page));
2096 return page;
2097}
2098#endif /* CONFIG_ELF_CORE */
2099
2100pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2101 spinlock_t **ptl)
2102{
2103 pgd_t * pgd = pgd_offset(mm, addr);
2104 pud_t * pud = pud_alloc(mm, pgd, addr);
2105 if (pud) {
2106 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2107 if (pmd) {
2108 VM_BUG_ON(pmd_trans_huge(*pmd));
2109 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2110 }
2111 }
2112 return NULL;
2113}
2114
2115/*
2116 * This is the old fallback for page remapping.
2117 *
2118 * For historical reasons, it only allows reserved pages. Only
2119 * old drivers should use this, and they needed to mark their
2120 * pages reserved for the old functions anyway.
2121 */
2122static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2123 struct page *page, pgprot_t prot)
2124{
2125 struct mm_struct *mm = vma->vm_mm;
2126 int retval;
2127 pte_t *pte;
2128 spinlock_t *ptl;
2129
2130 retval = -EINVAL;
2131 if (PageAnon(page))
2132 goto out;
2133 retval = -ENOMEM;
2134 flush_dcache_page(page);
2135 pte = get_locked_pte(mm, addr, &ptl);
2136 if (!pte)
2137 goto out;
2138 retval = -EBUSY;
2139 if (!pte_none(*pte))
2140 goto out_unlock;
2141
2142 /* Ok, finally just insert the thing.. */
2143 get_page(page);
2144 inc_mm_counter_fast(mm, MM_FILEPAGES);
2145 page_add_file_rmap(page);
2146 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2147
2148 retval = 0;
2149 pte_unmap_unlock(pte, ptl);
2150 return retval;
2151out_unlock:
2152 pte_unmap_unlock(pte, ptl);
2153out:
2154 return retval;
2155}
2156
2157/**
2158 * vm_insert_page - insert single page into user vma
2159 * @vma: user vma to map to
2160 * @addr: target user address of this page
2161 * @page: source kernel page
2162 *
2163 * This allows drivers to insert individual pages they've allocated
2164 * into a user vma.
2165 *
2166 * The page has to be a nice clean _individual_ kernel allocation.
2167 * If you allocate a compound page, you need to have marked it as
2168 * such (__GFP_COMP), or manually just split the page up yourself
2169 * (see split_page()).
2170 *
2171 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2172 * took an arbitrary page protection parameter. This doesn't allow
2173 * that. Your vma protection will have to be set up correctly, which
2174 * means that if you want a shared writable mapping, you'd better
2175 * ask for a shared writable mapping!
2176 *
2177 * The page does not need to be reserved.
2178 *
2179 * Usually this function is called from f_op->mmap() handler
2180 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2181 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2182 * function from other places, for example from page-fault handler.
2183 */
2184int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2185 struct page *page)
2186{
2187 if (addr < vma->vm_start || addr >= vma->vm_end)
2188 return -EFAULT;
2189 if (!page_count(page))
2190 return -EINVAL;
2191 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2192 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2193 BUG_ON(vma->vm_flags & VM_PFNMAP);
2194 vma->vm_flags |= VM_MIXEDMAP;
2195 }
2196 return insert_page(vma, addr, page, vma->vm_page_prot);
2197}
2198EXPORT_SYMBOL(vm_insert_page);
2199
2200static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2201 unsigned long pfn, pgprot_t prot)
2202{
2203 struct mm_struct *mm = vma->vm_mm;
2204 int retval;
2205 pte_t *pte, entry;
2206 spinlock_t *ptl;
2207
2208 retval = -ENOMEM;
2209 pte = get_locked_pte(mm, addr, &ptl);
2210 if (!pte)
2211 goto out;
2212 retval = -EBUSY;
2213 if (!pte_none(*pte))
2214 goto out_unlock;
2215
2216 /* Ok, finally just insert the thing.. */
2217 entry = pte_mkspecial(pfn_pte(pfn, prot));
2218 set_pte_at(mm, addr, pte, entry);
2219 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2220
2221 retval = 0;
2222out_unlock:
2223 pte_unmap_unlock(pte, ptl);
2224out:
2225 return retval;
2226}
2227
2228/**
2229 * vm_insert_pfn - insert single pfn into user vma
2230 * @vma: user vma to map to
2231 * @addr: target user address of this page
2232 * @pfn: source kernel pfn
2233 *
2234 * Similar to vm_insert_page, this allows drivers to insert individual pages
2235 * they've allocated into a user vma. Same comments apply.
2236 *
2237 * This function should only be called from a vm_ops->fault handler, and
2238 * in that case the handler should return NULL.
2239 *
2240 * vma cannot be a COW mapping.
2241 *
2242 * As this is called only for pages that do not currently exist, we
2243 * do not need to flush old virtual caches or the TLB.
2244 */
2245int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2246 unsigned long pfn)
2247{
2248 int ret;
2249 pgprot_t pgprot = vma->vm_page_prot;
2250 /*
2251 * Technically, architectures with pte_special can avoid all these
2252 * restrictions (same for remap_pfn_range). However we would like
2253 * consistency in testing and feature parity among all, so we should
2254 * try to keep these invariants in place for everybody.
2255 */
2256 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2257 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2258 (VM_PFNMAP|VM_MIXEDMAP));
2259 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2260 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2261
2262 if (addr < vma->vm_start || addr >= vma->vm_end)
2263 return -EFAULT;
2264 if (track_pfn_insert(vma, &pgprot, pfn))
2265 return -EINVAL;
2266
2267 ret = insert_pfn(vma, addr, pfn, pgprot);
2268
2269 return ret;
2270}
2271EXPORT_SYMBOL(vm_insert_pfn);
2272
2273int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2274 unsigned long pfn)
2275{
2276 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2277
2278 if (addr < vma->vm_start || addr >= vma->vm_end)
2279 return -EFAULT;
2280
2281 /*
2282 * If we don't have pte special, then we have to use the pfn_valid()
2283 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2284 * refcount the page if pfn_valid is true (hence insert_page rather
2285 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2286 * without pte special, it would there be refcounted as a normal page.
2287 */
2288 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2289 struct page *page;
2290
2291 page = pfn_to_page(pfn);
2292 return insert_page(vma, addr, page, vma->vm_page_prot);
2293 }
2294 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2295}
2296EXPORT_SYMBOL(vm_insert_mixed);
2297
2298/*
2299 * maps a range of physical memory into the requested pages. the old
2300 * mappings are removed. any references to nonexistent pages results
2301 * in null mappings (currently treated as "copy-on-access")
2302 */
2303static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2304 unsigned long addr, unsigned long end,
2305 unsigned long pfn, pgprot_t prot)
2306{
2307 pte_t *pte;
2308 spinlock_t *ptl;
2309
2310 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2311 if (!pte)
2312 return -ENOMEM;
2313 arch_enter_lazy_mmu_mode();
2314 do {
2315 BUG_ON(!pte_none(*pte));
2316 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2317 pfn++;
2318 } while (pte++, addr += PAGE_SIZE, addr != end);
2319 arch_leave_lazy_mmu_mode();
2320 pte_unmap_unlock(pte - 1, ptl);
2321 return 0;
2322}
2323
2324static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2325 unsigned long addr, unsigned long end,
2326 unsigned long pfn, pgprot_t prot)
2327{
2328 pmd_t *pmd;
2329 unsigned long next;
2330
2331 pfn -= addr >> PAGE_SHIFT;
2332 pmd = pmd_alloc(mm, pud, addr);
2333 if (!pmd)
2334 return -ENOMEM;
2335 VM_BUG_ON(pmd_trans_huge(*pmd));
2336 do {
2337 next = pmd_addr_end(addr, end);
2338 if (remap_pte_range(mm, pmd, addr, next,
2339 pfn + (addr >> PAGE_SHIFT), prot))
2340 return -ENOMEM;
2341 } while (pmd++, addr = next, addr != end);
2342 return 0;
2343}
2344
2345static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2346 unsigned long addr, unsigned long end,
2347 unsigned long pfn, pgprot_t prot)
2348{
2349 pud_t *pud;
2350 unsigned long next;
2351
2352 pfn -= addr >> PAGE_SHIFT;
2353 pud = pud_alloc(mm, pgd, addr);
2354 if (!pud)
2355 return -ENOMEM;
2356 do {
2357 next = pud_addr_end(addr, end);
2358 if (remap_pmd_range(mm, pud, addr, next,
2359 pfn + (addr >> PAGE_SHIFT), prot))
2360 return -ENOMEM;
2361 } while (pud++, addr = next, addr != end);
2362 return 0;
2363}
2364
2365/**
2366 * remap_pfn_range - remap kernel memory to userspace
2367 * @vma: user vma to map to
2368 * @addr: target user address to start at
2369 * @pfn: physical address of kernel memory
2370 * @size: size of map area
2371 * @prot: page protection flags for this mapping
2372 *
2373 * Note: this is only safe if the mm semaphore is held when called.
2374 */
2375int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2376 unsigned long pfn, unsigned long size, pgprot_t prot)
2377{
2378 pgd_t *pgd;
2379 unsigned long next;
2380 unsigned long end = addr + PAGE_ALIGN(size);
2381 struct mm_struct *mm = vma->vm_mm;
2382 int err;
2383
2384 /*
2385 * Physically remapped pages are special. Tell the
2386 * rest of the world about it:
2387 * VM_IO tells people not to look at these pages
2388 * (accesses can have side effects).
2389 * VM_PFNMAP tells the core MM that the base pages are just
2390 * raw PFN mappings, and do not have a "struct page" associated
2391 * with them.
2392 * VM_DONTEXPAND
2393 * Disable vma merging and expanding with mremap().
2394 * VM_DONTDUMP
2395 * Omit vma from core dump, even when VM_IO turned off.
2396 *
2397 * There's a horrible special case to handle copy-on-write
2398 * behaviour that some programs depend on. We mark the "original"
2399 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2400 * See vm_normal_page() for details.
2401 */
2402 if (is_cow_mapping(vma->vm_flags)) {
2403 if (addr != vma->vm_start || end != vma->vm_end)
2404 return -EINVAL;
2405 vma->vm_pgoff = pfn;
2406 }
2407
2408 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2409 if (err)
2410 return -EINVAL;
2411
2412 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2413
2414 BUG_ON(addr >= end);
2415 pfn -= addr >> PAGE_SHIFT;
2416 pgd = pgd_offset(mm, addr);
2417 flush_cache_range(vma, addr, end);
2418 do {
2419 next = pgd_addr_end(addr, end);
2420 err = remap_pud_range(mm, pgd, addr, next,
2421 pfn + (addr >> PAGE_SHIFT), prot);
2422 if (err)
2423 break;
2424 } while (pgd++, addr = next, addr != end);
2425
2426 if (err)
2427 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2428
2429 return err;
2430}
2431EXPORT_SYMBOL(remap_pfn_range);
2432
2433/**
2434 * vm_iomap_memory - remap memory to userspace
2435 * @vma: user vma to map to
2436 * @start: start of area
2437 * @len: size of area
2438 *
2439 * This is a simplified io_remap_pfn_range() for common driver use. The
2440 * driver just needs to give us the physical memory range to be mapped,
2441 * we'll figure out the rest from the vma information.
2442 *
2443 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2444 * whatever write-combining details or similar.
2445 */
2446int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2447{
2448 unsigned long vm_len, pfn, pages;
2449
2450 /* Check that the physical memory area passed in looks valid */
2451 if (start + len < start)
2452 return -EINVAL;
2453 /*
2454 * You *really* shouldn't map things that aren't page-aligned,
2455 * but we've historically allowed it because IO memory might
2456 * just have smaller alignment.
2457 */
2458 len += start & ~PAGE_MASK;
2459 pfn = start >> PAGE_SHIFT;
2460 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2461 if (pfn + pages < pfn)
2462 return -EINVAL;
2463
2464 /* We start the mapping 'vm_pgoff' pages into the area */
2465 if (vma->vm_pgoff > pages)
2466 return -EINVAL;
2467 pfn += vma->vm_pgoff;
2468 pages -= vma->vm_pgoff;
2469
2470 /* Can we fit all of the mapping? */
2471 vm_len = vma->vm_end - vma->vm_start;
2472 if (vm_len >> PAGE_SHIFT > pages)
2473 return -EINVAL;
2474
2475 /* Ok, let it rip */
2476 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2477}
2478EXPORT_SYMBOL(vm_iomap_memory);
2479
2480static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2481 unsigned long addr, unsigned long end,
2482 pte_fn_t fn, void *data)
2483{
2484 pte_t *pte;
2485 int err;
2486 pgtable_t token;
2487 spinlock_t *uninitialized_var(ptl);
2488
2489 pte = (mm == &init_mm) ?
2490 pte_alloc_kernel(pmd, addr) :
2491 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2492 if (!pte)
2493 return -ENOMEM;
2494
2495 BUG_ON(pmd_huge(*pmd));
2496
2497 arch_enter_lazy_mmu_mode();
2498
2499 token = pmd_pgtable(*pmd);
2500
2501 do {
2502 err = fn(pte++, token, addr, data);
2503 if (err)
2504 break;
2505 } while (addr += PAGE_SIZE, addr != end);
2506
2507 arch_leave_lazy_mmu_mode();
2508
2509 if (mm != &init_mm)
2510 pte_unmap_unlock(pte-1, ptl);
2511 return err;
2512}
2513
2514static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2515 unsigned long addr, unsigned long end,
2516 pte_fn_t fn, void *data)
2517{
2518 pmd_t *pmd;
2519 unsigned long next;
2520 int err;
2521
2522 BUG_ON(pud_huge(*pud));
2523
2524 pmd = pmd_alloc(mm, pud, addr);
2525 if (!pmd)
2526 return -ENOMEM;
2527 do {
2528 next = pmd_addr_end(addr, end);
2529 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2530 if (err)
2531 break;
2532 } while (pmd++, addr = next, addr != end);
2533 return err;
2534}
2535
2536static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2537 unsigned long addr, unsigned long end,
2538 pte_fn_t fn, void *data)
2539{
2540 pud_t *pud;
2541 unsigned long next;
2542 int err;
2543
2544 pud = pud_alloc(mm, pgd, addr);
2545 if (!pud)
2546 return -ENOMEM;
2547 do {
2548 next = pud_addr_end(addr, end);
2549 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2550 if (err)
2551 break;
2552 } while (pud++, addr = next, addr != end);
2553 return err;
2554}
2555
2556/*
2557 * Scan a region of virtual memory, filling in page tables as necessary
2558 * and calling a provided function on each leaf page table.
2559 */
2560int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2561 unsigned long size, pte_fn_t fn, void *data)
2562{
2563 pgd_t *pgd;
2564 unsigned long next;
2565 unsigned long end = addr + size;
2566 int err;
2567
2568 BUG_ON(addr >= end);
2569 pgd = pgd_offset(mm, addr);
2570 do {
2571 next = pgd_addr_end(addr, end);
2572 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2573 if (err)
2574 break;
2575 } while (pgd++, addr = next, addr != end);
2576
2577 return err;
2578}
2579EXPORT_SYMBOL_GPL(apply_to_page_range);
2580
2581/*
2582 * handle_pte_fault chooses page fault handler according to an entry
2583 * which was read non-atomically. Before making any commitment, on
2584 * those architectures or configurations (e.g. i386 with PAE) which
2585 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2586 * must check under lock before unmapping the pte and proceeding
2587 * (but do_wp_page is only called after already making such a check;
2588 * and do_anonymous_page can safely check later on).
2589 */
2590static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2591 pte_t *page_table, pte_t orig_pte)
2592{
2593 int same = 1;
2594#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2595 if (sizeof(pte_t) > sizeof(unsigned long)) {
2596 spinlock_t *ptl = pte_lockptr(mm, pmd);
2597 spin_lock(ptl);
2598 same = pte_same(*page_table, orig_pte);
2599 spin_unlock(ptl);
2600 }
2601#endif
2602 pte_unmap(page_table);
2603 return same;
2604}
2605
2606static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2607{
2608 debug_dma_assert_idle(src);
2609
2610 /*
2611 * If the source page was a PFN mapping, we don't have
2612 * a "struct page" for it. We do a best-effort copy by
2613 * just copying from the original user address. If that
2614 * fails, we just zero-fill it. Live with it.
2615 */
2616 if (unlikely(!src)) {
2617 void *kaddr = kmap_atomic(dst);
2618 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2619
2620 /*
2621 * This really shouldn't fail, because the page is there
2622 * in the page tables. But it might just be unreadable,
2623 * in which case we just give up and fill the result with
2624 * zeroes.
2625 */
2626 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2627 clear_page(kaddr);
2628 kunmap_atomic(kaddr);
2629 flush_dcache_page(dst);
2630 } else
2631 copy_user_highpage(dst, src, va, vma);
2632}
2633
2634/*
2635 * Notify the address space that the page is about to become writable so that
2636 * it can prohibit this or wait for the page to get into an appropriate state.
2637 *
2638 * We do this without the lock held, so that it can sleep if it needs to.
2639 */
2640static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2641 unsigned long address)
2642{
2643 struct vm_fault vmf;
2644 int ret;
2645
2646 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2647 vmf.pgoff = page->index;
2648 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2649 vmf.page = page;
2650
2651 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2652 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2653 return ret;
2654 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2655 lock_page(page);
2656 if (!page->mapping) {
2657 unlock_page(page);
2658 return 0; /* retry */
2659 }
2660 ret |= VM_FAULT_LOCKED;
2661 } else
2662 VM_BUG_ON_PAGE(!PageLocked(page), page);
2663 return ret;
2664}
2665
2666/*
2667 * This routine handles present pages, when users try to write
2668 * to a shared page. It is done by copying the page to a new address
2669 * and decrementing the shared-page counter for the old page.
2670 *
2671 * Note that this routine assumes that the protection checks have been
2672 * done by the caller (the low-level page fault routine in most cases).
2673 * Thus we can safely just mark it writable once we've done any necessary
2674 * COW.
2675 *
2676 * We also mark the page dirty at this point even though the page will
2677 * change only once the write actually happens. This avoids a few races,
2678 * and potentially makes it more efficient.
2679 *
2680 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2681 * but allow concurrent faults), with pte both mapped and locked.
2682 * We return with mmap_sem still held, but pte unmapped and unlocked.
2683 */
2684static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2685 unsigned long address, pte_t *page_table, pmd_t *pmd,
2686 spinlock_t *ptl, pte_t orig_pte)
2687 __releases(ptl)
2688{
2689 struct page *old_page, *new_page = NULL;
2690 pte_t entry;
2691 int ret = 0;
2692 int page_mkwrite = 0;
2693 struct page *dirty_page = NULL;
2694 unsigned long mmun_start = 0; /* For mmu_notifiers */
2695 unsigned long mmun_end = 0; /* For mmu_notifiers */
2696
2697 old_page = vm_normal_page(vma, address, orig_pte);
2698 if (!old_page) {
2699 /*
2700 * VM_MIXEDMAP !pfn_valid() case
2701 *
2702 * We should not cow pages in a shared writeable mapping.
2703 * Just mark the pages writable as we can't do any dirty
2704 * accounting on raw pfn maps.
2705 */
2706 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2707 (VM_WRITE|VM_SHARED))
2708 goto reuse;
2709 goto gotten;
2710 }
2711
2712 /*
2713 * Take out anonymous pages first, anonymous shared vmas are
2714 * not dirty accountable.
2715 */
2716 if (PageAnon(old_page) && !PageKsm(old_page)) {
2717 if (!trylock_page(old_page)) {
2718 page_cache_get(old_page);
2719 pte_unmap_unlock(page_table, ptl);
2720 lock_page(old_page);
2721 page_table = pte_offset_map_lock(mm, pmd, address,
2722 &ptl);
2723 if (!pte_same(*page_table, orig_pte)) {
2724 unlock_page(old_page);
2725 goto unlock;
2726 }
2727 page_cache_release(old_page);
2728 }
2729 if (reuse_swap_page(old_page)) {
2730 /*
2731 * The page is all ours. Move it to our anon_vma so
2732 * the rmap code will not search our parent or siblings.
2733 * Protected against the rmap code by the page lock.
2734 */
2735 page_move_anon_rmap(old_page, vma, address);
2736 unlock_page(old_page);
2737 goto reuse;
2738 }
2739 unlock_page(old_page);
2740 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2741 (VM_WRITE|VM_SHARED))) {
2742 /*
2743 * Only catch write-faults on shared writable pages,
2744 * read-only shared pages can get COWed by
2745 * get_user_pages(.write=1, .force=1).
2746 */
2747 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2748 int tmp;
2749 page_cache_get(old_page);
2750 pte_unmap_unlock(page_table, ptl);
2751 tmp = do_page_mkwrite(vma, old_page, address);
2752 if (unlikely(!tmp || (tmp &
2753 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2754 page_cache_release(old_page);
2755 return tmp;
2756 }
2757 /*
2758 * Since we dropped the lock we need to revalidate
2759 * the PTE as someone else may have changed it. If
2760 * they did, we just return, as we can count on the
2761 * MMU to tell us if they didn't also make it writable.
2762 */
2763 page_table = pte_offset_map_lock(mm, pmd, address,
2764 &ptl);
2765 if (!pte_same(*page_table, orig_pte)) {
2766 unlock_page(old_page);
2767 goto unlock;
2768 }
2769
2770 page_mkwrite = 1;
2771 }
2772 dirty_page = old_page;
2773 get_page(dirty_page);
2774
2775reuse:
2776 /*
2777 * Clear the pages cpupid information as the existing
2778 * information potentially belongs to a now completely
2779 * unrelated process.
2780 */
2781 if (old_page)
2782 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2783
2784 flush_cache_page(vma, address, pte_pfn(orig_pte));
2785 entry = pte_mkyoung(orig_pte);
2786 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2787 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2788 update_mmu_cache(vma, address, page_table);
2789 pte_unmap_unlock(page_table, ptl);
2790 ret |= VM_FAULT_WRITE;
2791
2792 if (!dirty_page)
2793 return ret;
2794
2795 /*
2796 * Yes, Virginia, this is actually required to prevent a race
2797 * with clear_page_dirty_for_io() from clearing the page dirty
2798 * bit after it clear all dirty ptes, but before a racing
2799 * do_wp_page installs a dirty pte.
2800 *
2801 * do_shared_fault is protected similarly.
2802 */
2803 if (!page_mkwrite) {
2804 wait_on_page_locked(dirty_page);
2805 set_page_dirty_balance(dirty_page);
2806 /* file_update_time outside page_lock */
2807 if (vma->vm_file)
2808 file_update_time(vma->vm_file);
2809 }
2810 put_page(dirty_page);
2811 if (page_mkwrite) {
2812 struct address_space *mapping = dirty_page->mapping;
2813
2814 set_page_dirty(dirty_page);
2815 unlock_page(dirty_page);
2816 page_cache_release(dirty_page);
2817 if (mapping) {
2818 /*
2819 * Some device drivers do not set page.mapping
2820 * but still dirty their pages
2821 */
2822 balance_dirty_pages_ratelimited(mapping);
2823 }
2824 }
2825
2826 return ret;
2827 }
2828
2829 /*
2830 * Ok, we need to copy. Oh, well..
2831 */
2832 page_cache_get(old_page);
2833gotten:
2834 pte_unmap_unlock(page_table, ptl);
2835
2836 if (unlikely(anon_vma_prepare(vma)))
2837 goto oom;
2838
2839 if (is_zero_pfn(pte_pfn(orig_pte))) {
2840 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2841 if (!new_page)
2842 goto oom;
2843 } else {
2844 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2845 if (!new_page)
2846 goto oom;
2847 cow_user_page(new_page, old_page, address, vma);
2848 }
2849 __SetPageUptodate(new_page);
2850
2851 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))
2852 goto oom_free_new;
2853
2854 mmun_start = address & PAGE_MASK;
2855 mmun_end = mmun_start + PAGE_SIZE;
2856 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2857
2858 /*
2859 * Re-check the pte - we dropped the lock
2860 */
2861 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2862 if (likely(pte_same(*page_table, orig_pte))) {
2863 if (old_page) {
2864 if (!PageAnon(old_page)) {
2865 dec_mm_counter_fast(mm, MM_FILEPAGES);
2866 inc_mm_counter_fast(mm, MM_ANONPAGES);
2867 }
2868 } else
2869 inc_mm_counter_fast(mm, MM_ANONPAGES);
2870 flush_cache_page(vma, address, pte_pfn(orig_pte));
2871 entry = mk_pte(new_page, vma->vm_page_prot);
2872 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2873 /*
2874 * Clear the pte entry and flush it first, before updating the
2875 * pte with the new entry. This will avoid a race condition
2876 * seen in the presence of one thread doing SMC and another
2877 * thread doing COW.
2878 */
2879 ptep_clear_flush(vma, address, page_table);
2880 page_add_new_anon_rmap(new_page, vma, address);
2881 /*
2882 * We call the notify macro here because, when using secondary
2883 * mmu page tables (such as kvm shadow page tables), we want the
2884 * new page to be mapped directly into the secondary page table.
2885 */
2886 set_pte_at_notify(mm, address, page_table, entry);
2887 update_mmu_cache(vma, address, page_table);
2888 if (old_page) {
2889 /*
2890 * Only after switching the pte to the new page may
2891 * we remove the mapcount here. Otherwise another
2892 * process may come and find the rmap count decremented
2893 * before the pte is switched to the new page, and
2894 * "reuse" the old page writing into it while our pte
2895 * here still points into it and can be read by other
2896 * threads.
2897 *
2898 * The critical issue is to order this
2899 * page_remove_rmap with the ptp_clear_flush above.
2900 * Those stores are ordered by (if nothing else,)
2901 * the barrier present in the atomic_add_negative
2902 * in page_remove_rmap.
2903 *
2904 * Then the TLB flush in ptep_clear_flush ensures that
2905 * no process can access the old page before the
2906 * decremented mapcount is visible. And the old page
2907 * cannot be reused until after the decremented
2908 * mapcount is visible. So transitively, TLBs to
2909 * old page will be flushed before it can be reused.
2910 */
2911 page_remove_rmap(old_page);
2912 }
2913
2914 /* Free the old page.. */
2915 new_page = old_page;
2916 ret |= VM_FAULT_WRITE;
2917 } else
2918 mem_cgroup_uncharge_page(new_page);
2919
2920 if (new_page)
2921 page_cache_release(new_page);
2922unlock:
2923 pte_unmap_unlock(page_table, ptl);
2924 if (mmun_end > mmun_start)
2925 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2926 if (old_page) {
2927 /*
2928 * Don't let another task, with possibly unlocked vma,
2929 * keep the mlocked page.
2930 */
2931 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2932 lock_page(old_page); /* LRU manipulation */
2933 munlock_vma_page(old_page);
2934 unlock_page(old_page);
2935 }
2936 page_cache_release(old_page);
2937 }
2938 return ret;
2939oom_free_new:
2940 page_cache_release(new_page);
2941oom:
2942 if (old_page)
2943 page_cache_release(old_page);
2944 return VM_FAULT_OOM;
2945}
2946
2947static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2948 unsigned long start_addr, unsigned long end_addr,
2949 struct zap_details *details)
2950{
2951 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2952}
2953
2954static inline void unmap_mapping_range_tree(struct rb_root *root,
2955 struct zap_details *details)
2956{
2957 struct vm_area_struct *vma;
2958 pgoff_t vba, vea, zba, zea;
2959
2960 vma_interval_tree_foreach(vma, root,
2961 details->first_index, details->last_index) {
2962
2963 vba = vma->vm_pgoff;
2964 vea = vba + vma_pages(vma) - 1;
2965 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2966 zba = details->first_index;
2967 if (zba < vba)
2968 zba = vba;
2969 zea = details->last_index;
2970 if (zea > vea)
2971 zea = vea;
2972
2973 unmap_mapping_range_vma(vma,
2974 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2975 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2976 details);
2977 }
2978}
2979
2980static inline void unmap_mapping_range_list(struct list_head *head,
2981 struct zap_details *details)
2982{
2983 struct vm_area_struct *vma;
2984
2985 /*
2986 * In nonlinear VMAs there is no correspondence between virtual address
2987 * offset and file offset. So we must perform an exhaustive search
2988 * across *all* the pages in each nonlinear VMA, not just the pages
2989 * whose virtual address lies outside the file truncation point.
2990 */
2991 list_for_each_entry(vma, head, shared.nonlinear) {
2992 details->nonlinear_vma = vma;
2993 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2994 }
2995}
2996
2997/**
2998 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2999 * @mapping: the address space containing mmaps to be unmapped.
3000 * @holebegin: byte in first page to unmap, relative to the start of
3001 * the underlying file. This will be rounded down to a PAGE_SIZE
3002 * boundary. Note that this is different from truncate_pagecache(), which
3003 * must keep the partial page. In contrast, we must get rid of
3004 * partial pages.
3005 * @holelen: size of prospective hole in bytes. This will be rounded
3006 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3007 * end of the file.
3008 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3009 * but 0 when invalidating pagecache, don't throw away private data.
3010 */
3011void unmap_mapping_range(struct address_space *mapping,
3012 loff_t const holebegin, loff_t const holelen, int even_cows)
3013{
3014 struct zap_details details;
3015 pgoff_t hba = holebegin >> PAGE_SHIFT;
3016 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3017
3018 /* Check for overflow. */
3019 if (sizeof(holelen) > sizeof(hlen)) {
3020 long long holeend =
3021 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3022 if (holeend & ~(long long)ULONG_MAX)
3023 hlen = ULONG_MAX - hba + 1;
3024 }
3025
3026 details.check_mapping = even_cows? NULL: mapping;
3027 details.nonlinear_vma = NULL;
3028 details.first_index = hba;
3029 details.last_index = hba + hlen - 1;
3030 if (details.last_index < details.first_index)
3031 details.last_index = ULONG_MAX;
3032
3033
3034 mutex_lock(&mapping->i_mmap_mutex);
3035 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
3036 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3037 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
3038 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
3039 mutex_unlock(&mapping->i_mmap_mutex);
3040}
3041EXPORT_SYMBOL(unmap_mapping_range);
3042
3043/*
3044 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3045 * but allow concurrent faults), and pte mapped but not yet locked.
3046 * We return with mmap_sem still held, but pte unmapped and unlocked.
3047 */
3048static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3049 unsigned long address, pte_t *page_table, pmd_t *pmd,
3050 unsigned int flags, pte_t orig_pte)
3051{
3052 spinlock_t *ptl;
3053 struct page *page, *swapcache;
3054 swp_entry_t entry;
3055 pte_t pte;
3056 int locked;
3057 struct mem_cgroup *ptr;
3058 int exclusive = 0;
3059 int ret = 0;
3060
3061 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3062 goto out;
3063
3064 entry = pte_to_swp_entry(orig_pte);
3065 if (unlikely(non_swap_entry(entry))) {
3066 if (is_migration_entry(entry)) {
3067 migration_entry_wait(mm, pmd, address);
3068 } else if (is_hwpoison_entry(entry)) {
3069 ret = VM_FAULT_HWPOISON;
3070 } else {
3071 print_bad_pte(vma, address, orig_pte, NULL);
3072 ret = VM_FAULT_SIGBUS;
3073 }
3074 goto out;
3075 }
3076 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3077 page = lookup_swap_cache(entry);
3078 if (!page) {
3079 page = swapin_readahead(entry,
3080 GFP_HIGHUSER_MOVABLE, vma, address);
3081 if (!page) {
3082 /*
3083 * Back out if somebody else faulted in this pte
3084 * while we released the pte lock.
3085 */
3086 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3087 if (likely(pte_same(*page_table, orig_pte)))
3088 ret = VM_FAULT_OOM;
3089 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3090 goto unlock;
3091 }
3092
3093 /* Had to read the page from swap area: Major fault */
3094 ret = VM_FAULT_MAJOR;
3095 count_vm_event(PGMAJFAULT);
3096 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3097 } else if (PageHWPoison(page)) {
3098 /*
3099 * hwpoisoned dirty swapcache pages are kept for killing
3100 * owner processes (which may be unknown at hwpoison time)
3101 */
3102 ret = VM_FAULT_HWPOISON;
3103 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3104 swapcache = page;
3105 goto out_release;
3106 }
3107
3108 swapcache = page;
3109 locked = lock_page_or_retry(page, mm, flags);
3110
3111 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3112 if (!locked) {
3113 ret |= VM_FAULT_RETRY;
3114 goto out_release;
3115 }
3116
3117 /*
3118 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3119 * release the swapcache from under us. The page pin, and pte_same
3120 * test below, are not enough to exclude that. Even if it is still
3121 * swapcache, we need to check that the page's swap has not changed.
3122 */
3123 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3124 goto out_page;
3125
3126 page = ksm_might_need_to_copy(page, vma, address);
3127 if (unlikely(!page)) {
3128 ret = VM_FAULT_OOM;
3129 page = swapcache;
3130 goto out_page;
3131 }
3132
3133 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3134 ret = VM_FAULT_OOM;
3135 goto out_page;
3136 }
3137
3138 /*
3139 * Back out if somebody else already faulted in this pte.
3140 */
3141 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3142 if (unlikely(!pte_same(*page_table, orig_pte)))
3143 goto out_nomap;
3144
3145 if (unlikely(!PageUptodate(page))) {
3146 ret = VM_FAULT_SIGBUS;
3147 goto out_nomap;
3148 }
3149
3150 /*
3151 * The page isn't present yet, go ahead with the fault.
3152 *
3153 * Be careful about the sequence of operations here.
3154 * To get its accounting right, reuse_swap_page() must be called
3155 * while the page is counted on swap but not yet in mapcount i.e.
3156 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3157 * must be called after the swap_free(), or it will never succeed.
3158 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3159 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3160 * in page->private. In this case, a record in swap_cgroup is silently
3161 * discarded at swap_free().
3162 */
3163
3164 inc_mm_counter_fast(mm, MM_ANONPAGES);
3165 dec_mm_counter_fast(mm, MM_SWAPENTS);
3166 pte = mk_pte(page, vma->vm_page_prot);
3167 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3168 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3169 flags &= ~FAULT_FLAG_WRITE;
3170 ret |= VM_FAULT_WRITE;
3171 exclusive = 1;
3172 }
3173 flush_icache_page(vma, page);
3174 if (pte_swp_soft_dirty(orig_pte))
3175 pte = pte_mksoft_dirty(pte);
3176 set_pte_at(mm, address, page_table, pte);
3177 if (page == swapcache)
3178 do_page_add_anon_rmap(page, vma, address, exclusive);
3179 else /* ksm created a completely new copy */
3180 page_add_new_anon_rmap(page, vma, address);
3181 /* It's better to call commit-charge after rmap is established */
3182 mem_cgroup_commit_charge_swapin(page, ptr);
3183
3184 swap_free(entry);
3185 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3186 try_to_free_swap(page);
3187 unlock_page(page);
3188 if (page != swapcache) {
3189 /*
3190 * Hold the lock to avoid the swap entry to be reused
3191 * until we take the PT lock for the pte_same() check
3192 * (to avoid false positives from pte_same). For
3193 * further safety release the lock after the swap_free
3194 * so that the swap count won't change under a
3195 * parallel locked swapcache.
3196 */
3197 unlock_page(swapcache);
3198 page_cache_release(swapcache);
3199 }
3200
3201 if (flags & FAULT_FLAG_WRITE) {
3202 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3203 if (ret & VM_FAULT_ERROR)
3204 ret &= VM_FAULT_ERROR;
3205 goto out;
3206 }
3207
3208 /* No need to invalidate - it was non-present before */
3209 update_mmu_cache(vma, address, page_table);
3210unlock:
3211 pte_unmap_unlock(page_table, ptl);
3212out:
3213 return ret;
3214out_nomap:
3215 mem_cgroup_cancel_charge_swapin(ptr);
3216 pte_unmap_unlock(page_table, ptl);
3217out_page:
3218 unlock_page(page);
3219out_release:
3220 page_cache_release(page);
3221 if (page != swapcache) {
3222 unlock_page(swapcache);
3223 page_cache_release(swapcache);
3224 }
3225 return ret;
3226}
3227
3228/*
3229 * This is like a special single-page "expand_{down|up}wards()",
3230 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3231 * doesn't hit another vma.
3232 */
3233static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3234{
3235 address &= PAGE_MASK;
3236 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3237 struct vm_area_struct *prev = vma->vm_prev;
3238
3239 /*
3240 * Is there a mapping abutting this one below?
3241 *
3242 * That's only ok if it's the same stack mapping
3243 * that has gotten split..
3244 */
3245 if (prev && prev->vm_end == address)
3246 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3247
3248 expand_downwards(vma, address - PAGE_SIZE);
3249 }
3250 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3251 struct vm_area_struct *next = vma->vm_next;
3252
3253 /* As VM_GROWSDOWN but s/below/above/ */
3254 if (next && next->vm_start == address + PAGE_SIZE)
3255 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3256
3257 expand_upwards(vma, address + PAGE_SIZE);
3258 }
3259 return 0;
3260}
3261
3262/*
3263 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3264 * but allow concurrent faults), and pte mapped but not yet locked.
3265 * We return with mmap_sem still held, but pte unmapped and unlocked.
3266 */
3267static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3268 unsigned long address, pte_t *page_table, pmd_t *pmd,
3269 unsigned int flags)
3270{
3271 struct page *page;
3272 spinlock_t *ptl;
3273 pte_t entry;
3274
3275 pte_unmap(page_table);
3276
3277 /* Check if we need to add a guard page to the stack */
3278 if (check_stack_guard_page(vma, address) < 0)
3279 return VM_FAULT_SIGBUS;
3280
3281 /* Use the zero-page for reads */
3282 if (!(flags & FAULT_FLAG_WRITE)) {
3283 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3284 vma->vm_page_prot));
3285 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3286 if (!pte_none(*page_table))
3287 goto unlock;
3288 goto setpte;
3289 }
3290
3291 /* Allocate our own private page. */
3292 if (unlikely(anon_vma_prepare(vma)))
3293 goto oom;
3294 page = alloc_zeroed_user_highpage_movable(vma, address);
3295 if (!page)
3296 goto oom;
3297 /*
3298 * The memory barrier inside __SetPageUptodate makes sure that
3299 * preceeding stores to the page contents become visible before
3300 * the set_pte_at() write.
3301 */
3302 __SetPageUptodate(page);
3303
3304 if (mem_cgroup_charge_anon(page, mm, GFP_KERNEL))
3305 goto oom_free_page;
3306
3307 entry = mk_pte(page, vma->vm_page_prot);
3308 if (vma->vm_flags & VM_WRITE)
3309 entry = pte_mkwrite(pte_mkdirty(entry));
3310
3311 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3312 if (!pte_none(*page_table))
3313 goto release;
3314
3315 inc_mm_counter_fast(mm, MM_ANONPAGES);
3316 page_add_new_anon_rmap(page, vma, address);
3317setpte:
3318 set_pte_at(mm, address, page_table, entry);
3319
3320 /* No need to invalidate - it was non-present before */
3321 update_mmu_cache(vma, address, page_table);
3322unlock:
3323 pte_unmap_unlock(page_table, ptl);
3324 return 0;
3325release:
3326 mem_cgroup_uncharge_page(page);
3327 page_cache_release(page);
3328 goto unlock;
3329oom_free_page:
3330 page_cache_release(page);
3331oom:
3332 return VM_FAULT_OOM;
3333}
3334
3335static int __do_fault(struct vm_area_struct *vma, unsigned long address,
3336 pgoff_t pgoff, unsigned int flags, struct page **page)
3337{
3338 struct vm_fault vmf;
3339 int ret;
3340
3341 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3342 vmf.pgoff = pgoff;
3343 vmf.flags = flags;
3344 vmf.page = NULL;
3345
3346 ret = vma->vm_ops->fault(vma, &vmf);
3347 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3348 return ret;
3349
3350 if (unlikely(PageHWPoison(vmf.page))) {
3351 if (ret & VM_FAULT_LOCKED)
3352 unlock_page(vmf.page);
3353 page_cache_release(vmf.page);
3354 return VM_FAULT_HWPOISON;
3355 }
3356
3357 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3358 lock_page(vmf.page);
3359 else
3360 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
3361
3362 *page = vmf.page;
3363 return ret;
3364}
3365
3366/**
3367 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
3368 *
3369 * @vma: virtual memory area
3370 * @address: user virtual address
3371 * @page: page to map
3372 * @pte: pointer to target page table entry
3373 * @write: true, if new entry is writable
3374 * @anon: true, if it's anonymous page
3375 *
3376 * Caller must hold page table lock relevant for @pte.
3377 *
3378 * Target users are page handler itself and implementations of
3379 * vm_ops->map_pages.
3380 */
3381void do_set_pte(struct vm_area_struct *vma, unsigned long address,
3382 struct page *page, pte_t *pte, bool write, bool anon)
3383{
3384 pte_t entry;
3385
3386 flush_icache_page(vma, page);
3387 entry = mk_pte(page, vma->vm_page_prot);
3388 if (write)
3389 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3390 else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
3391 pte_mksoft_dirty(entry);
3392 if (anon) {
3393 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3394 page_add_new_anon_rmap(page, vma, address);
3395 } else {
3396 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
3397 page_add_file_rmap(page);
3398 }
3399 set_pte_at(vma->vm_mm, address, pte, entry);
3400
3401 /* no need to invalidate: a not-present page won't be cached */
3402 update_mmu_cache(vma, address, pte);
3403}
3404
3405#define FAULT_AROUND_ORDER 4
3406
3407#ifdef CONFIG_DEBUG_FS
3408static unsigned int fault_around_order = FAULT_AROUND_ORDER;
3409
3410static int fault_around_order_get(void *data, u64 *val)
3411{
3412 *val = fault_around_order;
3413 return 0;
3414}
3415
3416static int fault_around_order_set(void *data, u64 val)
3417{
3418 BUILD_BUG_ON((1UL << FAULT_AROUND_ORDER) > PTRS_PER_PTE);
3419 if (1UL << val > PTRS_PER_PTE)
3420 return -EINVAL;
3421 fault_around_order = val;
3422 return 0;
3423}
3424DEFINE_SIMPLE_ATTRIBUTE(fault_around_order_fops,
3425 fault_around_order_get, fault_around_order_set, "%llu\n");
3426
3427static int __init fault_around_debugfs(void)
3428{
3429 void *ret;
3430
3431 ret = debugfs_create_file("fault_around_order", 0644, NULL, NULL,
3432 &fault_around_order_fops);
3433 if (!ret)
3434 pr_warn("Failed to create fault_around_order in debugfs");
3435 return 0;
3436}
3437late_initcall(fault_around_debugfs);
3438
3439static inline unsigned long fault_around_pages(void)
3440{
3441 return 1UL << fault_around_order;
3442}
3443
3444static inline unsigned long fault_around_mask(void)
3445{
3446 return ~((1UL << (PAGE_SHIFT + fault_around_order)) - 1);
3447}
3448#else
3449static inline unsigned long fault_around_pages(void)
3450{
3451 unsigned long nr_pages;
3452
3453 nr_pages = 1UL << FAULT_AROUND_ORDER;
3454 BUILD_BUG_ON(nr_pages > PTRS_PER_PTE);
3455 return nr_pages;
3456}
3457
3458static inline unsigned long fault_around_mask(void)
3459{
3460 return ~((1UL << (PAGE_SHIFT + FAULT_AROUND_ORDER)) - 1);
3461}
3462#endif
3463
3464static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
3465 pte_t *pte, pgoff_t pgoff, unsigned int flags)
3466{
3467 unsigned long start_addr;
3468 pgoff_t max_pgoff;
3469 struct vm_fault vmf;
3470 int off;
3471
3472 start_addr = max(address & fault_around_mask(), vma->vm_start);
3473 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3474 pte -= off;
3475 pgoff -= off;
3476
3477 /*
3478 * max_pgoff is either end of page table or end of vma
3479 * or fault_around_pages() from pgoff, depending what is neast.
3480 */
3481 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3482 PTRS_PER_PTE - 1;
3483 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
3484 pgoff + fault_around_pages() - 1);
3485
3486 /* Check if it makes any sense to call ->map_pages */
3487 while (!pte_none(*pte)) {
3488 if (++pgoff > max_pgoff)
3489 return;
3490 start_addr += PAGE_SIZE;
3491 if (start_addr >= vma->vm_end)
3492 return;
3493 pte++;
3494 }
3495
3496 vmf.virtual_address = (void __user *) start_addr;
3497 vmf.pte = pte;
3498 vmf.pgoff = pgoff;
3499 vmf.max_pgoff = max_pgoff;
3500 vmf.flags = flags;
3501 vma->vm_ops->map_pages(vma, &vmf);
3502}
3503
3504static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3505 unsigned long address, pmd_t *pmd,
3506 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3507{
3508 struct page *fault_page;
3509 spinlock_t *ptl;
3510 pte_t *pte;
3511 int ret = 0;
3512
3513 /*
3514 * Let's call ->map_pages() first and use ->fault() as fallback
3515 * if page by the offset is not ready to be mapped (cold cache or
3516 * something).
3517 */
3518 if (vma->vm_ops->map_pages) {
3519 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3520 do_fault_around(vma, address, pte, pgoff, flags);
3521 if (!pte_same(*pte, orig_pte))
3522 goto unlock_out;
3523 pte_unmap_unlock(pte, ptl);
3524 }
3525
3526 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3527 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3528 return ret;
3529
3530 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3531 if (unlikely(!pte_same(*pte, orig_pte))) {
3532 pte_unmap_unlock(pte, ptl);
3533 unlock_page(fault_page);
3534 page_cache_release(fault_page);
3535 return ret;
3536 }
3537 do_set_pte(vma, address, fault_page, pte, false, false);
3538 unlock_page(fault_page);
3539unlock_out:
3540 pte_unmap_unlock(pte, ptl);
3541 return ret;
3542}
3543
3544static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3545 unsigned long address, pmd_t *pmd,
3546 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3547{
3548 struct page *fault_page, *new_page;
3549 spinlock_t *ptl;
3550 pte_t *pte;
3551 int ret;
3552
3553 if (unlikely(anon_vma_prepare(vma)))
3554 return VM_FAULT_OOM;
3555
3556 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3557 if (!new_page)
3558 return VM_FAULT_OOM;
3559
3560 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)) {
3561 page_cache_release(new_page);
3562 return VM_FAULT_OOM;
3563 }
3564
3565 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3566 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3567 goto uncharge_out;
3568
3569 copy_user_highpage(new_page, fault_page, address, vma);
3570 __SetPageUptodate(new_page);
3571
3572 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3573 if (unlikely(!pte_same(*pte, orig_pte))) {
3574 pte_unmap_unlock(pte, ptl);
3575 unlock_page(fault_page);
3576 page_cache_release(fault_page);
3577 goto uncharge_out;
3578 }
3579 do_set_pte(vma, address, new_page, pte, true, true);
3580 pte_unmap_unlock(pte, ptl);
3581 unlock_page(fault_page);
3582 page_cache_release(fault_page);
3583 return ret;
3584uncharge_out:
3585 mem_cgroup_uncharge_page(new_page);
3586 page_cache_release(new_page);
3587 return ret;
3588}
3589
3590static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3591 unsigned long address, pmd_t *pmd,
3592 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3593{
3594 struct page *fault_page;
3595 struct address_space *mapping;
3596 spinlock_t *ptl;
3597 pte_t *pte;
3598 int dirtied = 0;
3599 int ret, tmp;
3600
3601 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3602 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3603 return ret;
3604
3605 /*
3606 * Check if the backing address space wants to know that the page is
3607 * about to become writable
3608 */
3609 if (vma->vm_ops->page_mkwrite) {
3610 unlock_page(fault_page);
3611 tmp = do_page_mkwrite(vma, fault_page, address);
3612 if (unlikely(!tmp ||
3613 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3614 page_cache_release(fault_page);
3615 return tmp;
3616 }
3617 }
3618
3619 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3620 if (unlikely(!pte_same(*pte, orig_pte))) {
3621 pte_unmap_unlock(pte, ptl);
3622 unlock_page(fault_page);
3623 page_cache_release(fault_page);
3624 return ret;
3625 }
3626 do_set_pte(vma, address, fault_page, pte, true, false);
3627 pte_unmap_unlock(pte, ptl);
3628
3629 if (set_page_dirty(fault_page))
3630 dirtied = 1;
3631 mapping = fault_page->mapping;
3632 unlock_page(fault_page);
3633 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3634 /*
3635 * Some device drivers do not set page.mapping but still
3636 * dirty their pages
3637 */
3638 balance_dirty_pages_ratelimited(mapping);
3639 }
3640
3641 /* file_update_time outside page_lock */
3642 if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3643 file_update_time(vma->vm_file);
3644
3645 return ret;
3646}
3647
3648static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3649 unsigned long address, pte_t *page_table, pmd_t *pmd,
3650 unsigned int flags, pte_t orig_pte)
3651{
3652 pgoff_t pgoff = (((address & PAGE_MASK)
3653 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3654
3655 pte_unmap(page_table);
3656 if (!(flags & FAULT_FLAG_WRITE))
3657 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3658 orig_pte);
3659 if (!(vma->vm_flags & VM_SHARED))
3660 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3661 orig_pte);
3662 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3663}
3664
3665/*
3666 * Fault of a previously existing named mapping. Repopulate the pte
3667 * from the encoded file_pte if possible. This enables swappable
3668 * nonlinear vmas.
3669 *
3670 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3671 * but allow concurrent faults), and pte mapped but not yet locked.
3672 * We return with mmap_sem still held, but pte unmapped and unlocked.
3673 */
3674static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3675 unsigned long address, pte_t *page_table, pmd_t *pmd,
3676 unsigned int flags, pte_t orig_pte)
3677{
3678 pgoff_t pgoff;
3679
3680 flags |= FAULT_FLAG_NONLINEAR;
3681
3682 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3683 return 0;
3684
3685 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3686 /*
3687 * Page table corrupted: show pte and kill process.
3688 */
3689 print_bad_pte(vma, address, orig_pte, NULL);
3690 return VM_FAULT_SIGBUS;
3691 }
3692
3693 pgoff = pte_to_pgoff(orig_pte);
3694 if (!(flags & FAULT_FLAG_WRITE))
3695 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3696 orig_pte);
3697 if (!(vma->vm_flags & VM_SHARED))
3698 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3699 orig_pte);
3700 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3701}
3702
3703static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3704 unsigned long addr, int page_nid,
3705 int *flags)
3706{
3707 get_page(page);
3708
3709 count_vm_numa_event(NUMA_HINT_FAULTS);
3710 if (page_nid == numa_node_id()) {
3711 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3712 *flags |= TNF_FAULT_LOCAL;
3713 }
3714
3715 return mpol_misplaced(page, vma, addr);
3716}
3717
3718static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3719 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3720{
3721 struct page *page = NULL;
3722 spinlock_t *ptl;
3723 int page_nid = -1;
3724 int last_cpupid;
3725 int target_nid;
3726 bool migrated = false;
3727 int flags = 0;
3728
3729 /*
3730 * The "pte" at this point cannot be used safely without
3731 * validation through pte_unmap_same(). It's of NUMA type but
3732 * the pfn may be screwed if the read is non atomic.
3733 *
3734 * ptep_modify_prot_start is not called as this is clearing
3735 * the _PAGE_NUMA bit and it is not really expected that there
3736 * would be concurrent hardware modifications to the PTE.
3737 */
3738 ptl = pte_lockptr(mm, pmd);
3739 spin_lock(ptl);
3740 if (unlikely(!pte_same(*ptep, pte))) {
3741 pte_unmap_unlock(ptep, ptl);
3742 goto out;
3743 }
3744
3745 pte = pte_mknonnuma(pte);
3746 set_pte_at(mm, addr, ptep, pte);
3747 update_mmu_cache(vma, addr, ptep);
3748
3749 page = vm_normal_page(vma, addr, pte);
3750 if (!page) {
3751 pte_unmap_unlock(ptep, ptl);
3752 return 0;
3753 }
3754 BUG_ON(is_zero_pfn(page_to_pfn(page)));
3755
3756 /*
3757 * Avoid grouping on DSO/COW pages in specific and RO pages
3758 * in general, RO pages shouldn't hurt as much anyway since
3759 * they can be in shared cache state.
3760 */
3761 if (!pte_write(pte))
3762 flags |= TNF_NO_GROUP;
3763
3764 /*
3765 * Flag if the page is shared between multiple address spaces. This
3766 * is later used when determining whether to group tasks together
3767 */
3768 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3769 flags |= TNF_SHARED;
3770
3771 last_cpupid = page_cpupid_last(page);
3772 page_nid = page_to_nid(page);
3773 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3774 pte_unmap_unlock(ptep, ptl);
3775 if (target_nid == -1) {
3776 put_page(page);
3777 goto out;
3778 }
3779
3780 /* Migrate to the requested node */
3781 migrated = migrate_misplaced_page(page, vma, target_nid);
3782 if (migrated) {
3783 page_nid = target_nid;
3784 flags |= TNF_MIGRATED;
3785 }
3786
3787out:
3788 if (page_nid != -1)
3789 task_numa_fault(last_cpupid, page_nid, 1, flags);
3790 return 0;
3791}
3792
3793/*
3794 * These routines also need to handle stuff like marking pages dirty
3795 * and/or accessed for architectures that don't do it in hardware (most
3796 * RISC architectures). The early dirtying is also good on the i386.
3797 *
3798 * There is also a hook called "update_mmu_cache()" that architectures
3799 * with external mmu caches can use to update those (ie the Sparc or
3800 * PowerPC hashed page tables that act as extended TLBs).
3801 *
3802 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3803 * but allow concurrent faults), and pte mapped but not yet locked.
3804 * We return with mmap_sem still held, but pte unmapped and unlocked.
3805 */
3806static int handle_pte_fault(struct mm_struct *mm,
3807 struct vm_area_struct *vma, unsigned long address,
3808 pte_t *pte, pmd_t *pmd, unsigned int flags)
3809{
3810 pte_t entry;
3811 spinlock_t *ptl;
3812
3813 entry = *pte;
3814 if (!pte_present(entry)) {
3815 if (pte_none(entry)) {
3816 if (vma->vm_ops) {
3817 if (likely(vma->vm_ops->fault))
3818 return do_linear_fault(mm, vma, address,
3819 pte, pmd, flags, entry);
3820 }
3821 return do_anonymous_page(mm, vma, address,
3822 pte, pmd, flags);
3823 }
3824 if (pte_file(entry))
3825 return do_nonlinear_fault(mm, vma, address,
3826 pte, pmd, flags, entry);
3827 return do_swap_page(mm, vma, address,
3828 pte, pmd, flags, entry);
3829 }
3830
3831 if (pte_numa(entry))
3832 return do_numa_page(mm, vma, address, entry, pte, pmd);
3833
3834 ptl = pte_lockptr(mm, pmd);
3835 spin_lock(ptl);
3836 if (unlikely(!pte_same(*pte, entry)))
3837 goto unlock;
3838 if (flags & FAULT_FLAG_WRITE) {
3839 if (!pte_write(entry))
3840 return do_wp_page(mm, vma, address,
3841 pte, pmd, ptl, entry);
3842 entry = pte_mkdirty(entry);
3843 }
3844 entry = pte_mkyoung(entry);
3845 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3846 update_mmu_cache(vma, address, pte);
3847 } else {
3848 /*
3849 * This is needed only for protection faults but the arch code
3850 * is not yet telling us if this is a protection fault or not.
3851 * This still avoids useless tlb flushes for .text page faults
3852 * with threads.
3853 */
3854 if (flags & FAULT_FLAG_WRITE)
3855 flush_tlb_fix_spurious_fault(vma, address);
3856 }
3857unlock:
3858 pte_unmap_unlock(pte, ptl);
3859 return 0;
3860}
3861
3862/*
3863 * By the time we get here, we already hold the mm semaphore
3864 */
3865static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3866 unsigned long address, unsigned int flags)
3867{
3868 pgd_t *pgd;
3869 pud_t *pud;
3870 pmd_t *pmd;
3871 pte_t *pte;
3872
3873 if (unlikely(is_vm_hugetlb_page(vma)))
3874 return hugetlb_fault(mm, vma, address, flags);
3875
3876 pgd = pgd_offset(mm, address);
3877 pud = pud_alloc(mm, pgd, address);
3878 if (!pud)
3879 return VM_FAULT_OOM;
3880 pmd = pmd_alloc(mm, pud, address);
3881 if (!pmd)
3882 return VM_FAULT_OOM;
3883 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3884 int ret = VM_FAULT_FALLBACK;
3885 if (!vma->vm_ops)
3886 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3887 pmd, flags);
3888 if (!(ret & VM_FAULT_FALLBACK))
3889 return ret;
3890 } else {
3891 pmd_t orig_pmd = *pmd;
3892 int ret;
3893
3894 barrier();
3895 if (pmd_trans_huge(orig_pmd)) {
3896 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3897
3898 /*
3899 * If the pmd is splitting, return and retry the
3900 * the fault. Alternative: wait until the split
3901 * is done, and goto retry.
3902 */
3903 if (pmd_trans_splitting(orig_pmd))
3904 return 0;
3905
3906 if (pmd_numa(orig_pmd))
3907 return do_huge_pmd_numa_page(mm, vma, address,
3908 orig_pmd, pmd);
3909
3910 if (dirty && !pmd_write(orig_pmd)) {
3911 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3912 orig_pmd);
3913 if (!(ret & VM_FAULT_FALLBACK))
3914 return ret;
3915 } else {
3916 huge_pmd_set_accessed(mm, vma, address, pmd,
3917 orig_pmd, dirty);
3918 return 0;
3919 }
3920 }
3921 }
3922
3923 /* THP should already have been handled */
3924 BUG_ON(pmd_numa(*pmd));
3925
3926 /*
3927 * Use __pte_alloc instead of pte_alloc_map, because we can't
3928 * run pte_offset_map on the pmd, if an huge pmd could
3929 * materialize from under us from a different thread.
3930 */
3931 if (unlikely(pmd_none(*pmd)) &&
3932 unlikely(__pte_alloc(mm, vma, pmd, address)))
3933 return VM_FAULT_OOM;
3934 /* if an huge pmd materialized from under us just retry later */
3935 if (unlikely(pmd_trans_huge(*pmd)))
3936 return 0;
3937 /*
3938 * A regular pmd is established and it can't morph into a huge pmd
3939 * from under us anymore at this point because we hold the mmap_sem
3940 * read mode and khugepaged takes it in write mode. So now it's
3941 * safe to run pte_offset_map().
3942 */
3943 pte = pte_offset_map(pmd, address);
3944
3945 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3946}
3947
3948int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3949 unsigned long address, unsigned int flags)
3950{
3951 int ret;
3952
3953 __set_current_state(TASK_RUNNING);
3954
3955 count_vm_event(PGFAULT);
3956 mem_cgroup_count_vm_event(mm, PGFAULT);
3957
3958 /* do counter updates before entering really critical section. */
3959 check_sync_rss_stat(current);
3960
3961 /*
3962 * Enable the memcg OOM handling for faults triggered in user
3963 * space. Kernel faults are handled more gracefully.
3964 */
3965 if (flags & FAULT_FLAG_USER)
3966 mem_cgroup_oom_enable();
3967
3968 ret = __handle_mm_fault(mm, vma, address, flags);
3969
3970 if (flags & FAULT_FLAG_USER) {
3971 mem_cgroup_oom_disable();
3972 /*
3973 * The task may have entered a memcg OOM situation but
3974 * if the allocation error was handled gracefully (no
3975 * VM_FAULT_OOM), there is no need to kill anything.
3976 * Just clean up the OOM state peacefully.
3977 */
3978 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3979 mem_cgroup_oom_synchronize(false);
3980 }
3981
3982 return ret;
3983}
3984
3985#ifndef __PAGETABLE_PUD_FOLDED
3986/*
3987 * Allocate page upper directory.
3988 * We've already handled the fast-path in-line.
3989 */
3990int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3991{
3992 pud_t *new = pud_alloc_one(mm, address);
3993 if (!new)
3994 return -ENOMEM;
3995
3996 smp_wmb(); /* See comment in __pte_alloc */
3997
3998 spin_lock(&mm->page_table_lock);
3999 if (pgd_present(*pgd)) /* Another has populated it */
4000 pud_free(mm, new);
4001 else
4002 pgd_populate(mm, pgd, new);
4003 spin_unlock(&mm->page_table_lock);
4004 return 0;
4005}
4006#endif /* __PAGETABLE_PUD_FOLDED */
4007
4008#ifndef __PAGETABLE_PMD_FOLDED
4009/*
4010 * Allocate page middle directory.
4011 * We've already handled the fast-path in-line.
4012 */
4013int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4014{
4015 pmd_t *new = pmd_alloc_one(mm, address);
4016 if (!new)
4017 return -ENOMEM;
4018
4019 smp_wmb(); /* See comment in __pte_alloc */
4020
4021 spin_lock(&mm->page_table_lock);
4022#ifndef __ARCH_HAS_4LEVEL_HACK
4023 if (pud_present(*pud)) /* Another has populated it */
4024 pmd_free(mm, new);
4025 else
4026 pud_populate(mm, pud, new);
4027#else
4028 if (pgd_present(*pud)) /* Another has populated it */
4029 pmd_free(mm, new);
4030 else
4031 pgd_populate(mm, pud, new);
4032#endif /* __ARCH_HAS_4LEVEL_HACK */
4033 spin_unlock(&mm->page_table_lock);
4034 return 0;
4035}
4036#endif /* __PAGETABLE_PMD_FOLDED */
4037
4038#if !defined(__HAVE_ARCH_GATE_AREA)
4039
4040#if defined(AT_SYSINFO_EHDR)
4041static struct vm_area_struct gate_vma;
4042
4043static int __init gate_vma_init(void)
4044{
4045 gate_vma.vm_mm = NULL;
4046 gate_vma.vm_start = FIXADDR_USER_START;
4047 gate_vma.vm_end = FIXADDR_USER_END;
4048 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
4049 gate_vma.vm_page_prot = __P101;
4050
4051 return 0;
4052}
4053__initcall(gate_vma_init);
4054#endif
4055
4056struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
4057{
4058#ifdef AT_SYSINFO_EHDR
4059 return &gate_vma;
4060#else
4061 return NULL;
4062#endif
4063}
4064
4065int in_gate_area_no_mm(unsigned long addr)
4066{
4067#ifdef AT_SYSINFO_EHDR
4068 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
4069 return 1;
4070#endif
4071 return 0;
4072}
4073
4074#endif /* __HAVE_ARCH_GATE_AREA */
4075
4076static int __follow_pte(struct mm_struct *mm, unsigned long address,
4077 pte_t **ptepp, spinlock_t **ptlp)
4078{
4079 pgd_t *pgd;
4080 pud_t *pud;
4081 pmd_t *pmd;
4082 pte_t *ptep;
4083
4084 pgd = pgd_offset(mm, address);
4085 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4086 goto out;
4087
4088 pud = pud_offset(pgd, address);
4089 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4090 goto out;
4091
4092 pmd = pmd_offset(pud, address);
4093 VM_BUG_ON(pmd_trans_huge(*pmd));
4094 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4095 goto out;
4096
4097 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
4098 if (pmd_huge(*pmd))
4099 goto out;
4100
4101 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4102 if (!ptep)
4103 goto out;
4104 if (!pte_present(*ptep))
4105 goto unlock;
4106 *ptepp = ptep;
4107 return 0;
4108unlock:
4109 pte_unmap_unlock(ptep, *ptlp);
4110out:
4111 return -EINVAL;
4112}
4113
4114static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4115 pte_t **ptepp, spinlock_t **ptlp)
4116{
4117 int res;
4118
4119 /* (void) is needed to make gcc happy */
4120 (void) __cond_lock(*ptlp,
4121 !(res = __follow_pte(mm, address, ptepp, ptlp)));
4122 return res;
4123}
4124
4125/**
4126 * follow_pfn - look up PFN at a user virtual address
4127 * @vma: memory mapping
4128 * @address: user virtual address
4129 * @pfn: location to store found PFN
4130 *
4131 * Only IO mappings and raw PFN mappings are allowed.
4132 *
4133 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4134 */
4135int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4136 unsigned long *pfn)
4137{
4138 int ret = -EINVAL;
4139 spinlock_t *ptl;
4140 pte_t *ptep;
4141
4142 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4143 return ret;
4144
4145 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4146 if (ret)
4147 return ret;
4148 *pfn = pte_pfn(*ptep);
4149 pte_unmap_unlock(ptep, ptl);
4150 return 0;
4151}
4152EXPORT_SYMBOL(follow_pfn);
4153
4154#ifdef CONFIG_HAVE_IOREMAP_PROT
4155int follow_phys(struct vm_area_struct *vma,
4156 unsigned long address, unsigned int flags,
4157 unsigned long *prot, resource_size_t *phys)
4158{
4159 int ret = -EINVAL;
4160 pte_t *ptep, pte;
4161 spinlock_t *ptl;
4162
4163 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4164 goto out;
4165
4166 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4167 goto out;
4168 pte = *ptep;
4169
4170 if ((flags & FOLL_WRITE) && !pte_write(pte))
4171 goto unlock;
4172
4173 *prot = pgprot_val(pte_pgprot(pte));
4174 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4175
4176 ret = 0;
4177unlock:
4178 pte_unmap_unlock(ptep, ptl);
4179out:
4180 return ret;
4181}
4182
4183int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4184 void *buf, int len, int write)
4185{
4186 resource_size_t phys_addr;
4187 unsigned long prot = 0;
4188 void __iomem *maddr;
4189 int offset = addr & (PAGE_SIZE-1);
4190
4191 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4192 return -EINVAL;
4193
4194 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4195 if (write)
4196 memcpy_toio(maddr + offset, buf, len);
4197 else
4198 memcpy_fromio(buf, maddr + offset, len);
4199 iounmap(maddr);
4200
4201 return len;
4202}
4203EXPORT_SYMBOL_GPL(generic_access_phys);
4204#endif
4205
4206/*
4207 * Access another process' address space as given in mm. If non-NULL, use the
4208 * given task for page fault accounting.
4209 */
4210static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4211 unsigned long addr, void *buf, int len, int write)
4212{
4213 struct vm_area_struct *vma;
4214 void *old_buf = buf;
4215
4216 down_read(&mm->mmap_sem);
4217 /* ignore errors, just check how much was successfully transferred */
4218 while (len) {
4219 int bytes, ret, offset;
4220 void *maddr;
4221 struct page *page = NULL;
4222
4223 ret = get_user_pages(tsk, mm, addr, 1,
4224 write, 1, &page, &vma);
4225 if (ret <= 0) {
4226 /*
4227 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4228 * we can access using slightly different code.
4229 */
4230#ifdef CONFIG_HAVE_IOREMAP_PROT
4231 vma = find_vma(mm, addr);
4232 if (!vma || vma->vm_start > addr)
4233 break;
4234 if (vma->vm_ops && vma->vm_ops->access)
4235 ret = vma->vm_ops->access(vma, addr, buf,
4236 len, write);
4237 if (ret <= 0)
4238#endif
4239 break;
4240 bytes = ret;
4241 } else {
4242 bytes = len;
4243 offset = addr & (PAGE_SIZE-1);
4244 if (bytes > PAGE_SIZE-offset)
4245 bytes = PAGE_SIZE-offset;
4246
4247 maddr = kmap(page);
4248 if (write) {
4249 copy_to_user_page(vma, page, addr,
4250 maddr + offset, buf, bytes);
4251 set_page_dirty_lock(page);
4252 } else {
4253 copy_from_user_page(vma, page, addr,
4254 buf, maddr + offset, bytes);
4255 }
4256 kunmap(page);
4257 page_cache_release(page);
4258 }
4259 len -= bytes;
4260 buf += bytes;
4261 addr += bytes;
4262 }
4263 up_read(&mm->mmap_sem);
4264
4265 return buf - old_buf;
4266}
4267
4268/**
4269 * access_remote_vm - access another process' address space
4270 * @mm: the mm_struct of the target address space
4271 * @addr: start address to access
4272 * @buf: source or destination buffer
4273 * @len: number of bytes to transfer
4274 * @write: whether the access is a write
4275 *
4276 * The caller must hold a reference on @mm.
4277 */
4278int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4279 void *buf, int len, int write)
4280{
4281 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4282}
4283
4284/*
4285 * Access another process' address space.
4286 * Source/target buffer must be kernel space,
4287 * Do not walk the page table directly, use get_user_pages
4288 */
4289int access_process_vm(struct task_struct *tsk, unsigned long addr,
4290 void *buf, int len, int write)
4291{
4292 struct mm_struct *mm;
4293 int ret;
4294
4295 mm = get_task_mm(tsk);
4296 if (!mm)
4297 return 0;
4298
4299 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4300 mmput(mm);
4301
4302 return ret;
4303}
4304
4305/*
4306 * Print the name of a VMA.
4307 */
4308void print_vma_addr(char *prefix, unsigned long ip)
4309{
4310 struct mm_struct *mm = current->mm;
4311 struct vm_area_struct *vma;
4312
4313 /*
4314 * Do not print if we are in atomic
4315 * contexts (in exception stacks, etc.):
4316 */
4317 if (preempt_count())
4318 return;
4319
4320 down_read(&mm->mmap_sem);
4321 vma = find_vma(mm, ip);
4322 if (vma && vma->vm_file) {
4323 struct file *f = vma->vm_file;
4324 char *buf = (char *)__get_free_page(GFP_KERNEL);
4325 if (buf) {
4326 char *p;
4327
4328 p = d_path(&f->f_path, buf, PAGE_SIZE);
4329 if (IS_ERR(p))
4330 p = "?";
4331 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4332 vma->vm_start,
4333 vma->vm_end - vma->vm_start);
4334 free_page((unsigned long)buf);
4335 }
4336 }
4337 up_read(&mm->mmap_sem);
4338}
4339
4340#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4341void might_fault(void)
4342{
4343 /*
4344 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4345 * holding the mmap_sem, this is safe because kernel memory doesn't
4346 * get paged out, therefore we'll never actually fault, and the
4347 * below annotations will generate false positives.
4348 */
4349 if (segment_eq(get_fs(), KERNEL_DS))
4350 return;
4351
4352 /*
4353 * it would be nicer only to annotate paths which are not under
4354 * pagefault_disable, however that requires a larger audit and
4355 * providing helpers like get_user_atomic.
4356 */
4357 if (in_atomic())
4358 return;
4359
4360 __might_sleep(__FILE__, __LINE__, 0);
4361
4362 if (current->mm)
4363 might_lock_read(¤t->mm->mmap_sem);
4364}
4365EXPORT_SYMBOL(might_fault);
4366#endif
4367
4368#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4369static void clear_gigantic_page(struct page *page,
4370 unsigned long addr,
4371 unsigned int pages_per_huge_page)
4372{
4373 int i;
4374 struct page *p = page;
4375
4376 might_sleep();
4377 for (i = 0; i < pages_per_huge_page;
4378 i++, p = mem_map_next(p, page, i)) {
4379 cond_resched();
4380 clear_user_highpage(p, addr + i * PAGE_SIZE);
4381 }
4382}
4383void clear_huge_page(struct page *page,
4384 unsigned long addr, unsigned int pages_per_huge_page)
4385{
4386 int i;
4387
4388 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4389 clear_gigantic_page(page, addr, pages_per_huge_page);
4390 return;
4391 }
4392
4393 might_sleep();
4394 for (i = 0; i < pages_per_huge_page; i++) {
4395 cond_resched();
4396 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4397 }
4398}
4399
4400static void copy_user_gigantic_page(struct page *dst, struct page *src,
4401 unsigned long addr,
4402 struct vm_area_struct *vma,
4403 unsigned int pages_per_huge_page)
4404{
4405 int i;
4406 struct page *dst_base = dst;
4407 struct page *src_base = src;
4408
4409 for (i = 0; i < pages_per_huge_page; ) {
4410 cond_resched();
4411 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4412
4413 i++;
4414 dst = mem_map_next(dst, dst_base, i);
4415 src = mem_map_next(src, src_base, i);
4416 }
4417}
4418
4419void copy_user_huge_page(struct page *dst, struct page *src,
4420 unsigned long addr, struct vm_area_struct *vma,
4421 unsigned int pages_per_huge_page)
4422{
4423 int i;
4424
4425 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4426 copy_user_gigantic_page(dst, src, addr, vma,
4427 pages_per_huge_page);
4428 return;
4429 }
4430
4431 might_sleep();
4432 for (i = 0; i < pages_per_huge_page; i++) {
4433 cond_resched();
4434 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4435 }
4436}
4437#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4438
4439#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4440
4441static struct kmem_cache *page_ptl_cachep;
4442
4443void __init ptlock_cache_init(void)
4444{
4445 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4446 SLAB_PANIC, NULL);
4447}
4448
4449bool ptlock_alloc(struct page *page)
4450{
4451 spinlock_t *ptl;
4452
4453 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4454 if (!ptl)
4455 return false;
4456 page->ptl = ptl;
4457 return true;
4458}
4459
4460void ptlock_free(struct page *page)
4461{
4462 kmem_cache_free(page_ptl_cachep, page->ptl);
4463}
4464#endif
1
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