1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/memory.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 */
7
8/*
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
11 */
12
13/*
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
16 *
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
19 * far as I could see.
20 *
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
22 */
23
24/*
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
30 */
31
32/*
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 *
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
38 *
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
40 */
41
42#include <linux/kernel_stat.h>
43#include <linux/mm.h>
44#include <linux/mm_inline.h>
45#include <linux/sched/mm.h>
46#include <linux/sched/numa_balancing.h>
47#include <linux/sched/task.h>
48#include <linux/hugetlb.h>
49#include <linux/mman.h>
50#include <linux/swap.h>
51#include <linux/highmem.h>
52#include <linux/pagemap.h>
53#include <linux/memremap.h>
54#include <linux/kmsan.h>
55#include <linux/ksm.h>
56#include <linux/rmap.h>
57#include <linux/export.h>
58#include <linux/delayacct.h>
59#include <linux/init.h>
60#include <linux/writeback.h>
61#include <linux/memcontrol.h>
62#include <linux/mmu_notifier.h>
63#include <linux/swapops.h>
64#include <linux/elf.h>
65#include <linux/gfp.h>
66#include <linux/migrate.h>
67#include <linux/string.h>
68#include <linux/memory-tiers.h>
69#include <linux/debugfs.h>
70#include <linux/userfaultfd_k.h>
71#include <linux/dax.h>
72#include <linux/oom.h>
73#include <linux/numa.h>
74#include <linux/perf_event.h>
75#include <linux/ptrace.h>
76#include <linux/vmalloc.h>
77#include <linux/sched/sysctl.h>
78
79#include <trace/events/kmem.h>
80
81#include <asm/io.h>
82#include <asm/mmu_context.h>
83#include <asm/pgalloc.h>
84#include <linux/uaccess.h>
85#include <asm/tlb.h>
86#include <asm/tlbflush.h>
87
88#include "pgalloc-track.h"
89#include "internal.h"
90#include "swap.h"
91
92#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
93#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
94#endif
95
96static vm_fault_t do_fault(struct vm_fault *vmf);
97static vm_fault_t do_anonymous_page(struct vm_fault *vmf);
98static bool vmf_pte_changed(struct vm_fault *vmf);
99
100/*
101 * Return true if the original pte was a uffd-wp pte marker (so the pte was
102 * wr-protected).
103 */
104static __always_inline bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
105{
106 if (!userfaultfd_wp(vma: vmf->vma))
107 return false;
108 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
109 return false;
110
111 return pte_marker_uffd_wp(pte: vmf->orig_pte);
112}
113
114/*
115 * Randomize the address space (stacks, mmaps, brk, etc.).
116 *
117 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
118 * as ancient (libc5 based) binaries can segfault. )
119 */
120int randomize_va_space __read_mostly =
121#ifdef CONFIG_COMPAT_BRK
122 1;
123#else
124 2;
125#endif
126
127static const struct ctl_table mmu_sysctl_table[] = {
128 {
129 .procname = "randomize_va_space",
130 .data = &randomize_va_space,
131 .maxlen = sizeof(int),
132 .mode = 0644,
133 .proc_handler = proc_dointvec,
134 },
135};
136
137static int __init init_mm_sysctl(void)
138{
139 register_sysctl_init("kernel", mmu_sysctl_table);
140 return 0;
141}
142
143subsys_initcall(init_mm_sysctl);
144
145#ifndef arch_wants_old_prefaulted_pte
146static inline bool arch_wants_old_prefaulted_pte(void)
147{
148 /*
149 * Transitioning a PTE from 'old' to 'young' can be expensive on
150 * some architectures, even if it's performed in hardware. By
151 * default, "false" means prefaulted entries will be 'young'.
152 */
153 return false;
154}
155#endif
156
157static int __init disable_randmaps(char *s)
158{
159 randomize_va_space = 0;
160 return 1;
161}
162__setup("norandmaps", disable_randmaps);
163
164unsigned long zero_pfn __read_mostly;
165EXPORT_SYMBOL(zero_pfn);
166
167unsigned long highest_memmap_pfn __read_mostly;
168
169/*
170 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
171 */
172static int __init init_zero_pfn(void)
173{
174 zero_pfn = page_to_pfn(ZERO_PAGE(0));
175 return 0;
176}
177early_initcall(init_zero_pfn);
178
179void mm_trace_rss_stat(struct mm_struct *mm, int member)
180{
181 trace_rss_stat(mm, member);
182}
183
184/*
185 * Note: this doesn't free the actual pages themselves. That
186 * has been handled earlier when unmapping all the memory regions.
187 */
188static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
189 unsigned long addr)
190{
191 pgtable_t token = pmd_pgtable(*pmd);
192 pmd_clear(pmd);
193 pte_free_tlb(tlb, token, addr);
194 mm_dec_nr_ptes(mm: tlb->mm);
195}
196
197static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
198 unsigned long addr, unsigned long end,
199 unsigned long floor, unsigned long ceiling)
200{
201 pmd_t *pmd;
202 unsigned long next;
203 unsigned long start;
204
205 start = addr;
206 pmd = pmd_offset(pud, address: addr);
207 do {
208 next = pmd_addr_end(addr, end);
209 if (pmd_none_or_clear_bad(pmd))
210 continue;
211 free_pte_range(tlb, pmd, addr);
212 } while (pmd++, addr = next, addr != end);
213
214 start &= PUD_MASK;
215 if (start < floor)
216 return;
217 if (ceiling) {
218 ceiling &= PUD_MASK;
219 if (!ceiling)
220 return;
221 }
222 if (end - 1 > ceiling - 1)
223 return;
224
225 pmd = pmd_offset(pud, address: start);
226 pud_clear(pud);
227 pmd_free_tlb(tlb, pmd, start);
228 mm_dec_nr_pmds(mm: tlb->mm);
229}
230
231static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
232 unsigned long addr, unsigned long end,
233 unsigned long floor, unsigned long ceiling)
234{
235 pud_t *pud;
236 unsigned long next;
237 unsigned long start;
238
239 start = addr;
240 pud = pud_offset(p4d, address: addr);
241 do {
242 next = pud_addr_end(addr, end);
243 if (pud_none_or_clear_bad(pud))
244 continue;
245 free_pmd_range(tlb, pud, addr, end: next, floor, ceiling);
246 } while (pud++, addr = next, addr != end);
247
248 start &= P4D_MASK;
249 if (start < floor)
250 return;
251 if (ceiling) {
252 ceiling &= P4D_MASK;
253 if (!ceiling)
254 return;
255 }
256 if (end - 1 > ceiling - 1)
257 return;
258
259 pud = pud_offset(p4d, address: start);
260 p4d_clear(p4d);
261 pud_free_tlb(tlb, pud, start);
262 mm_dec_nr_puds(mm: tlb->mm);
263}
264
265static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
266 unsigned long addr, unsigned long end,
267 unsigned long floor, unsigned long ceiling)
268{
269 p4d_t *p4d;
270 unsigned long next;
271 unsigned long start;
272
273 start = addr;
274 p4d = p4d_offset(pgd, address: addr);
275 do {
276 next = p4d_addr_end(addr, end);
277 if (p4d_none_or_clear_bad(p4d))
278 continue;
279 free_pud_range(tlb, p4d, addr, end: next, floor, ceiling);
280 } while (p4d++, addr = next, addr != end);
281
282 start &= PGDIR_MASK;
283 if (start < floor)
284 return;
285 if (ceiling) {
286 ceiling &= PGDIR_MASK;
287 if (!ceiling)
288 return;
289 }
290 if (end - 1 > ceiling - 1)
291 return;
292
293 p4d = p4d_offset(pgd, address: start);
294 pgd_clear(pgd);
295 p4d_free_tlb(tlb, p4d, start);
296}
297
298/**
299 * free_pgd_range - Unmap and free page tables in the range
300 * @tlb: the mmu_gather containing pending TLB flush info
301 * @addr: virtual address start
302 * @end: virtual address end
303 * @floor: lowest address boundary
304 * @ceiling: highest address boundary
305 *
306 * This function tears down all user-level page tables in the
307 * specified virtual address range [@addr..@end). It is part of
308 * the memory unmap flow.
309 */
310void free_pgd_range(struct mmu_gather *tlb,
311 unsigned long addr, unsigned long end,
312 unsigned long floor, unsigned long ceiling)
313{
314 pgd_t *pgd;
315 unsigned long next;
316
317 /*
318 * The next few lines have given us lots of grief...
319 *
320 * Why are we testing PMD* at this top level? Because often
321 * there will be no work to do at all, and we'd prefer not to
322 * go all the way down to the bottom just to discover that.
323 *
324 * Why all these "- 1"s? Because 0 represents both the bottom
325 * of the address space and the top of it (using -1 for the
326 * top wouldn't help much: the masks would do the wrong thing).
327 * The rule is that addr 0 and floor 0 refer to the bottom of
328 * the address space, but end 0 and ceiling 0 refer to the top
329 * Comparisons need to use "end - 1" and "ceiling - 1" (though
330 * that end 0 case should be mythical).
331 *
332 * Wherever addr is brought up or ceiling brought down, we must
333 * be careful to reject "the opposite 0" before it confuses the
334 * subsequent tests. But what about where end is brought down
335 * by PMD_SIZE below? no, end can't go down to 0 there.
336 *
337 * Whereas we round start (addr) and ceiling down, by different
338 * masks at different levels, in order to test whether a table
339 * now has no other vmas using it, so can be freed, we don't
340 * bother to round floor or end up - the tests don't need that.
341 */
342
343 addr &= PMD_MASK;
344 if (addr < floor) {
345 addr += PMD_SIZE;
346 if (!addr)
347 return;
348 }
349 if (ceiling) {
350 ceiling &= PMD_MASK;
351 if (!ceiling)
352 return;
353 }
354 if (end - 1 > ceiling - 1)
355 end -= PMD_SIZE;
356 if (addr > end - 1)
357 return;
358 /*
359 * We add page table cache pages with PAGE_SIZE,
360 * (see pte_free_tlb()), flush the tlb if we need
361 */
362 tlb_change_page_size(tlb, PAGE_SIZE);
363 pgd = pgd_offset(tlb->mm, addr);
364 do {
365 next = pgd_addr_end(addr, end);
366 if (pgd_none_or_clear_bad(pgd))
367 continue;
368 free_p4d_range(tlb, pgd, addr, end: next, floor, ceiling);
369 } while (pgd++, addr = next, addr != end);
370}
371
372void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
373 struct vm_area_struct *vma, unsigned long floor,
374 unsigned long ceiling, bool mm_wr_locked)
375{
376 struct unlink_vma_file_batch vb;
377
378 tlb_free_vmas(tlb);
379
380 do {
381 unsigned long addr = vma->vm_start;
382 struct vm_area_struct *next;
383
384 /*
385 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
386 * be 0. This will underflow and is okay.
387 */
388 next = mas_find(mas, max: ceiling - 1);
389 if (unlikely(xa_is_zero(next)))
390 next = NULL;
391
392 /*
393 * Hide vma from rmap and truncate_pagecache before freeing
394 * pgtables
395 */
396 if (mm_wr_locked)
397 vma_start_write(vma);
398 unlink_anon_vmas(vma);
399
400 unlink_file_vma_batch_init(vb: &vb);
401 unlink_file_vma_batch_add(vb: &vb, vma);
402
403 /*
404 * Optimization: gather nearby vmas into one call down
405 */
406 while (next && next->vm_start <= vma->vm_end + PMD_SIZE) {
407 vma = next;
408 next = mas_find(mas, max: ceiling - 1);
409 if (unlikely(xa_is_zero(next)))
410 next = NULL;
411 if (mm_wr_locked)
412 vma_start_write(vma);
413 unlink_anon_vmas(vma);
414 unlink_file_vma_batch_add(vb: &vb, vma);
415 }
416 unlink_file_vma_batch_final(vb: &vb);
417
418 free_pgd_range(tlb, addr, end: vma->vm_end,
419 floor, ceiling: next ? next->vm_start : ceiling);
420 vma = next;
421 } while (vma);
422}
423
424void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
425{
426 spinlock_t *ptl = pmd_lock(mm, pmd);
427
428 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
429 mm_inc_nr_ptes(mm);
430 /*
431 * Ensure all pte setup (eg. pte page lock and page clearing) are
432 * visible before the pte is made visible to other CPUs by being
433 * put into page tables.
434 *
435 * The other side of the story is the pointer chasing in the page
436 * table walking code (when walking the page table without locking;
437 * ie. most of the time). Fortunately, these data accesses consist
438 * of a chain of data-dependent loads, meaning most CPUs (alpha
439 * being the notable exception) will already guarantee loads are
440 * seen in-order. See the alpha page table accessors for the
441 * smp_rmb() barriers in page table walking code.
442 */
443 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
444 pmd_populate(mm, pmd, pte: *pte);
445 *pte = NULL;
446 }
447 spin_unlock(lock: ptl);
448}
449
450int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
451{
452 pgtable_t new = pte_alloc_one(mm);
453 if (!new)
454 return -ENOMEM;
455
456 pmd_install(mm, pmd, pte: &new);
457 if (new)
458 pte_free(mm, pte_page: new);
459 return 0;
460}
461
462int __pte_alloc_kernel(pmd_t *pmd)
463{
464 pte_t *new = pte_alloc_one_kernel(&init_mm);
465 if (!new)
466 return -ENOMEM;
467
468 spin_lock(lock: &init_mm.page_table_lock);
469 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
470 smp_wmb(); /* See comment in pmd_install() */
471 pmd_populate_kernel(mm: &init_mm, pmd, pte: new);
472 new = NULL;
473 }
474 spin_unlock(lock: &init_mm.page_table_lock);
475 if (new)
476 pte_free_kernel(mm: &init_mm, pte: new);
477 return 0;
478}
479
480static inline void init_rss_vec(int *rss)
481{
482 memset(s: rss, c: 0, n: sizeof(int) * NR_MM_COUNTERS);
483}
484
485static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
486{
487 int i;
488
489 for (i = 0; i < NR_MM_COUNTERS; i++)
490 if (rss[i])
491 add_mm_counter(mm, member: i, value: rss[i]);
492}
493
494static bool is_bad_page_map_ratelimited(void)
495{
496 static unsigned long resume;
497 static unsigned long nr_shown;
498 static unsigned long nr_unshown;
499
500 /*
501 * Allow a burst of 60 reports, then keep quiet for that minute;
502 * or allow a steady drip of one report per second.
503 */
504 if (nr_shown == 60) {
505 if (time_before(jiffies, resume)) {
506 nr_unshown++;
507 return true;
508 }
509 if (nr_unshown) {
510 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
511 nr_unshown);
512 nr_unshown = 0;
513 }
514 nr_shown = 0;
515 }
516 if (nr_shown++ == 0)
517 resume = jiffies + 60 * HZ;
518 return false;
519}
520
521static void __print_bad_page_map_pgtable(struct mm_struct *mm, unsigned long addr)
522{
523 unsigned long long pgdv, p4dv, pudv, pmdv;
524 p4d_t p4d, *p4dp;
525 pud_t pud, *pudp;
526 pmd_t pmd, *pmdp;
527 pgd_t *pgdp;
528
529 /*
530 * Although this looks like a fully lockless pgtable walk, it is not:
531 * see locking requirements for print_bad_page_map().
532 */
533 pgdp = pgd_offset(mm, addr);
534 pgdv = pgd_val(*pgdp);
535
536 if (!pgd_present(pgd: *pgdp) || pgd_leaf(*pgdp)) {
537 pr_alert("pgd:%08llx\n", pgdv);
538 return;
539 }
540
541 p4dp = p4d_offset(pgd: pgdp, address: addr);
542 p4d = p4dp_get(p4dp);
543 p4dv = p4d_val(p4d);
544
545 if (!p4d_present(p4d) || p4d_leaf(p4d)) {
546 pr_alert("pgd:%08llx p4d:%08llx\n", pgdv, p4dv);
547 return;
548 }
549
550 pudp = pud_offset(p4d: p4dp, address: addr);
551 pud = pudp_get(pudp);
552 pudv = pud_val(pud);
553
554 if (!pud_present(pud) || pud_leaf(pud)) {
555 pr_alert("pgd:%08llx p4d:%08llx pud:%08llx\n", pgdv, p4dv, pudv);
556 return;
557 }
558
559 pmdp = pmd_offset(pud: pudp, address: addr);
560 pmd = pmdp_get(pmdp);
561 pmdv = pmd_val(pmd);
562
563 /*
564 * Dumping the PTE would be nice, but it's tricky with CONFIG_HIGHPTE,
565 * because the table should already be mapped by the caller and
566 * doing another map would be bad. print_bad_page_map() should
567 * already take care of printing the PTE.
568 */
569 pr_alert("pgd:%08llx p4d:%08llx pud:%08llx pmd:%08llx\n", pgdv,
570 p4dv, pudv, pmdv);
571}
572
573/*
574 * This function is called to print an error when a bad page table entry (e.g.,
575 * corrupted page table entry) is found. For example, we might have a
576 * PFN-mapped pte in a region that doesn't allow it.
577 *
578 * The calling function must still handle the error.
579 *
580 * This function must be called during a proper page table walk, as it will
581 * re-walk the page table to dump information: the caller MUST prevent page
582 * table teardown (by holding mmap, vma or rmap lock) and MUST hold the leaf
583 * page table lock.
584 */
585static void print_bad_page_map(struct vm_area_struct *vma,
586 unsigned long addr, unsigned long long entry, struct page *page,
587 enum pgtable_level level)
588{
589 struct address_space *mapping;
590 pgoff_t index;
591
592 if (is_bad_page_map_ratelimited())
593 return;
594
595 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
596 index = linear_page_index(vma, address: addr);
597
598 pr_alert("BUG: Bad page map in process %s %s:%08llx", current->comm,
599 pgtable_level_to_str(level), entry);
600 __print_bad_page_map_pgtable(mm: vma->vm_mm, addr);
601 if (page)
602 dump_page(page, reason: "bad page map");
603 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
604 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
605 pr_alert("file:%pD fault:%ps mmap:%ps mmap_prepare: %ps read_folio:%ps\n",
606 vma->vm_file,
607 vma->vm_ops ? vma->vm_ops->fault : NULL,
608 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
609 vma->vm_file ? vma->vm_file->f_op->mmap_prepare : NULL,
610 mapping ? mapping->a_ops->read_folio : NULL);
611 dump_stack();
612 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
613}
614#define print_bad_pte(vma, addr, pte, page) \
615 print_bad_page_map(vma, addr, pte_val(pte), page, PGTABLE_LEVEL_PTE)
616
617/**
618 * __vm_normal_page() - Get the "struct page" associated with a page table entry.
619 * @vma: The VMA mapping the page table entry.
620 * @addr: The address where the page table entry is mapped.
621 * @pfn: The PFN stored in the page table entry.
622 * @special: Whether the page table entry is marked "special".
623 * @level: The page table level for error reporting purposes only.
624 * @entry: The page table entry value for error reporting purposes only.
625 *
626 * "Special" mappings do not wish to be associated with a "struct page" (either
627 * it doesn't exist, or it exists but they don't want to touch it). In this
628 * case, NULL is returned here. "Normal" mappings do have a struct page and
629 * are ordinarily refcounted.
630 *
631 * Page mappings of the shared zero folios are always considered "special", as
632 * they are not ordinarily refcounted: neither the refcount nor the mapcount
633 * of these folios is adjusted when mapping them into user page tables.
634 * Selected page table walkers (such as GUP) can still identify mappings of the
635 * shared zero folios and work with the underlying "struct page".
636 *
637 * There are 2 broad cases. Firstly, an architecture may define a "special"
638 * page table entry bit, such as pte_special(), in which case this function is
639 * trivial. Secondly, an architecture may not have a spare page table
640 * entry bit, which requires a more complicated scheme, described below.
641 *
642 * With CONFIG_FIND_NORMAL_PAGE, we might have the "special" bit set on
643 * page table entries that actually map "normal" pages: however, that page
644 * cannot be looked up through the PFN stored in the page table entry, but
645 * instead will be looked up through vm_ops->find_normal_page(). So far, this
646 * only applies to PTEs.
647 *
648 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
649 * special mapping (even if there are underlying and valid "struct pages").
650 * COWed pages of a VM_PFNMAP are always normal.
651 *
652 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
653 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
654 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
655 * mapping will always honor the rule
656 *
657 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
658 *
659 * And for normal mappings this is false.
660 *
661 * This restricts such mappings to be a linear translation from virtual address
662 * to pfn. To get around this restriction, we allow arbitrary mappings so long
663 * as the vma is not a COW mapping; in that case, we know that all ptes are
664 * special (because none can have been COWed).
665 *
666 *
667 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
668 *
669 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
670 * page" backing, however the difference is that _all_ pages with a struct
671 * page (that is, those where pfn_valid is true, except the shared zero
672 * folios) are refcounted and considered normal pages by the VM.
673 *
674 * The disadvantage is that pages are refcounted (which can be slower and
675 * simply not an option for some PFNMAP users). The advantage is that we
676 * don't have to follow the strict linearity rule of PFNMAP mappings in
677 * order to support COWable mappings.
678 *
679 * Return: Returns the "struct page" if this is a "normal" mapping. Returns
680 * NULL if this is a "special" mapping.
681 */
682static inline struct page *__vm_normal_page(struct vm_area_struct *vma,
683 unsigned long addr, unsigned long pfn, bool special,
684 unsigned long long entry, enum pgtable_level level)
685{
686 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
687 if (unlikely(special)) {
688#ifdef CONFIG_FIND_NORMAL_PAGE
689 if (vma->vm_ops && vma->vm_ops->find_normal_page)
690 return vma->vm_ops->find_normal_page(vma, addr);
691#endif /* CONFIG_FIND_NORMAL_PAGE */
692 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
693 return NULL;
694 if (is_zero_pfn(pfn) || is_huge_zero_pfn(pfn))
695 return NULL;
696
697 print_bad_page_map(vma, addr, entry, NULL, level);
698 return NULL;
699 }
700 /*
701 * With CONFIG_ARCH_HAS_PTE_SPECIAL, any special page table
702 * mappings (incl. shared zero folios) are marked accordingly.
703 */
704 } else {
705 if (unlikely(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))) {
706 if (vma->vm_flags & VM_MIXEDMAP) {
707 /* If it has a "struct page", it's "normal". */
708 if (!pfn_valid(pfn))
709 return NULL;
710 } else {
711 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
712
713 /* Only CoW'ed anon folios are "normal". */
714 if (pfn == vma->vm_pgoff + off)
715 return NULL;
716 if (!is_cow_mapping(flags: vma->vm_flags))
717 return NULL;
718 }
719 }
720
721 if (is_zero_pfn(pfn) || is_huge_zero_pfn(pfn))
722 return NULL;
723 }
724
725 if (unlikely(pfn > highest_memmap_pfn)) {
726 /* Corrupted page table entry. */
727 print_bad_page_map(vma, addr, entry, NULL, level);
728 return NULL;
729 }
730 /*
731 * NOTE! We still have PageReserved() pages in the page tables.
732 * For example, VDSO mappings can cause them to exist.
733 */
734 VM_WARN_ON_ONCE(is_zero_pfn(pfn) || is_huge_zero_pfn(pfn));
735 return pfn_to_page(pfn);
736}
737
738/**
739 * vm_normal_page() - Get the "struct page" associated with a PTE
740 * @vma: The VMA mapping the @pte.
741 * @addr: The address where the @pte is mapped.
742 * @pte: The PTE.
743 *
744 * Get the "struct page" associated with a PTE. See __vm_normal_page()
745 * for details on "normal" and "special" mappings.
746 *
747 * Return: Returns the "struct page" if this is a "normal" mapping. Returns
748 * NULL if this is a "special" mapping.
749 */
750struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
751 pte_t pte)
752{
753 return __vm_normal_page(vma, addr, pfn: pte_pfn(pte), special: pte_special(pte),
754 pte_val(pte), level: PGTABLE_LEVEL_PTE);
755}
756
757/**
758 * vm_normal_folio() - Get the "struct folio" associated with a PTE
759 * @vma: The VMA mapping the @pte.
760 * @addr: The address where the @pte is mapped.
761 * @pte: The PTE.
762 *
763 * Get the "struct folio" associated with a PTE. See __vm_normal_page()
764 * for details on "normal" and "special" mappings.
765 *
766 * Return: Returns the "struct folio" if this is a "normal" mapping. Returns
767 * NULL if this is a "special" mapping.
768 */
769struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr,
770 pte_t pte)
771{
772 struct page *page = vm_normal_page(vma, addr, pte);
773
774 if (page)
775 return page_folio(page);
776 return NULL;
777}
778
779#ifdef CONFIG_PGTABLE_HAS_HUGE_LEAVES
780/**
781 * vm_normal_page_pmd() - Get the "struct page" associated with a PMD
782 * @vma: The VMA mapping the @pmd.
783 * @addr: The address where the @pmd is mapped.
784 * @pmd: The PMD.
785 *
786 * Get the "struct page" associated with a PTE. See __vm_normal_page()
787 * for details on "normal" and "special" mappings.
788 *
789 * Return: Returns the "struct page" if this is a "normal" mapping. Returns
790 * NULL if this is a "special" mapping.
791 */
792struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
793 pmd_t pmd)
794{
795 return __vm_normal_page(vma, addr, pfn: pmd_pfn(pmd), special: pmd_special(pmd),
796 pmd_val(pmd), level: PGTABLE_LEVEL_PMD);
797}
798
799/**
800 * vm_normal_folio_pmd() - Get the "struct folio" associated with a PMD
801 * @vma: The VMA mapping the @pmd.
802 * @addr: The address where the @pmd is mapped.
803 * @pmd: The PMD.
804 *
805 * Get the "struct folio" associated with a PTE. See __vm_normal_page()
806 * for details on "normal" and "special" mappings.
807 *
808 * Return: Returns the "struct folio" if this is a "normal" mapping. Returns
809 * NULL if this is a "special" mapping.
810 */
811struct folio *vm_normal_folio_pmd(struct vm_area_struct *vma,
812 unsigned long addr, pmd_t pmd)
813{
814 struct page *page = vm_normal_page_pmd(vma, addr, pmd);
815
816 if (page)
817 return page_folio(page);
818 return NULL;
819}
820
821/**
822 * vm_normal_page_pud() - Get the "struct page" associated with a PUD
823 * @vma: The VMA mapping the @pud.
824 * @addr: The address where the @pud is mapped.
825 * @pud: The PUD.
826 *
827 * Get the "struct page" associated with a PUD. See __vm_normal_page()
828 * for details on "normal" and "special" mappings.
829 *
830 * Return: Returns the "struct page" if this is a "normal" mapping. Returns
831 * NULL if this is a "special" mapping.
832 */
833struct page *vm_normal_page_pud(struct vm_area_struct *vma,
834 unsigned long addr, pud_t pud)
835{
836 return __vm_normal_page(vma, addr, pud_pfn(pud), special: pud_special(pud),
837 pud_val(pud), level: PGTABLE_LEVEL_PUD);
838}
839#endif
840
841/**
842 * restore_exclusive_pte - Restore a device-exclusive entry
843 * @vma: VMA covering @address
844 * @folio: the mapped folio
845 * @page: the mapped folio page
846 * @address: the virtual address
847 * @ptep: pte pointer into the locked page table mapping the folio page
848 * @orig_pte: pte value at @ptep
849 *
850 * Restore a device-exclusive non-swap entry to an ordinary present pte.
851 *
852 * The folio and the page table must be locked, and MMU notifiers must have
853 * been called to invalidate any (exclusive) device mappings.
854 *
855 * Locking the folio makes sure that anybody who just converted the pte to
856 * a device-exclusive entry can map it into the device to make forward
857 * progress without others converting it back until the folio was unlocked.
858 *
859 * If the folio lock ever becomes an issue, we can stop relying on the folio
860 * lock; it might make some scenarios with heavy thrashing less likely to
861 * make forward progress, but these scenarios might not be valid use cases.
862 *
863 * Note that the folio lock does not protect against all cases of concurrent
864 * page table modifications (e.g., MADV_DONTNEED, mprotect), so device drivers
865 * must use MMU notifiers to sync against any concurrent changes.
866 */
867static void restore_exclusive_pte(struct vm_area_struct *vma,
868 struct folio *folio, struct page *page, unsigned long address,
869 pte_t *ptep, pte_t orig_pte)
870{
871 pte_t pte;
872
873 VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
874
875 pte = pte_mkold(pte: mk_pte(page, READ_ONCE(vma->vm_page_prot)));
876 if (pte_swp_soft_dirty(pte: orig_pte))
877 pte = pte_mksoft_dirty(pte);
878
879 if (pte_swp_uffd_wp(pte: orig_pte))
880 pte = pte_mkuffd_wp(pte);
881
882 if ((vma->vm_flags & VM_WRITE) &&
883 can_change_pte_writable(vma, addr: address, pte)) {
884 if (folio_test_dirty(folio))
885 pte = pte_mkdirty(pte);
886 pte = pte_mkwrite(pte, vma);
887 }
888 set_pte_at(vma->vm_mm, address, ptep, pte);
889
890 /*
891 * No need to invalidate - it was non-present before. However
892 * secondary CPUs may have mappings that need invalidating.
893 */
894 update_mmu_cache(vma, addr: address, ptep);
895}
896
897/*
898 * Tries to restore an exclusive pte if the page lock can be acquired without
899 * sleeping.
900 */
901static int try_restore_exclusive_pte(struct vm_area_struct *vma,
902 unsigned long addr, pte_t *ptep, pte_t orig_pte)
903{
904 struct page *page = pfn_swap_entry_to_page(entry: pte_to_swp_entry(pte: orig_pte));
905 struct folio *folio = page_folio(page);
906
907 if (folio_trylock(folio)) {
908 restore_exclusive_pte(vma, folio, page, address: addr, ptep, orig_pte);
909 folio_unlock(folio);
910 return 0;
911 }
912
913 return -EBUSY;
914}
915
916/*
917 * copy one vm_area from one task to the other. Assumes the page tables
918 * already present in the new task to be cleared in the whole range
919 * covered by this vma.
920 */
921
922static unsigned long
923copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
924 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
925 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
926{
927 vm_flags_t vm_flags = dst_vma->vm_flags;
928 pte_t orig_pte = ptep_get(ptep: src_pte);
929 pte_t pte = orig_pte;
930 struct folio *folio;
931 struct page *page;
932 swp_entry_t entry = pte_to_swp_entry(pte: orig_pte);
933
934 if (likely(!non_swap_entry(entry))) {
935 if (swap_duplicate(entry) < 0)
936 return -EIO;
937
938 /* make sure dst_mm is on swapoff's mmlist. */
939 if (unlikely(list_empty(&dst_mm->mmlist))) {
940 spin_lock(lock: &mmlist_lock);
941 if (list_empty(head: &dst_mm->mmlist))
942 list_add(new: &dst_mm->mmlist,
943 head: &src_mm->mmlist);
944 spin_unlock(lock: &mmlist_lock);
945 }
946 /* Mark the swap entry as shared. */
947 if (pte_swp_exclusive(pte: orig_pte)) {
948 pte = pte_swp_clear_exclusive(pte: orig_pte);
949 set_pte_at(src_mm, addr, src_pte, pte);
950 }
951 rss[MM_SWAPENTS]++;
952 } else if (is_migration_entry(entry)) {
953 folio = pfn_swap_entry_folio(entry);
954
955 rss[mm_counter(folio)]++;
956
957 if (!is_readable_migration_entry(entry) &&
958 is_cow_mapping(flags: vm_flags)) {
959 /*
960 * COW mappings require pages in both parent and child
961 * to be set to read. A previously exclusive entry is
962 * now shared.
963 */
964 entry = make_readable_migration_entry(
965 offset: swp_offset(entry));
966 pte = swp_entry_to_pte(entry);
967 if (pte_swp_soft_dirty(pte: orig_pte))
968 pte = pte_swp_mksoft_dirty(pte);
969 if (pte_swp_uffd_wp(pte: orig_pte))
970 pte = pte_swp_mkuffd_wp(pte);
971 set_pte_at(src_mm, addr, src_pte, pte);
972 }
973 } else if (is_device_private_entry(entry)) {
974 page = pfn_swap_entry_to_page(entry);
975 folio = page_folio(page);
976
977 /*
978 * Update rss count even for unaddressable pages, as
979 * they should treated just like normal pages in this
980 * respect.
981 *
982 * We will likely want to have some new rss counters
983 * for unaddressable pages, at some point. But for now
984 * keep things as they are.
985 */
986 folio_get(folio);
987 rss[mm_counter(folio)]++;
988 /* Cannot fail as these pages cannot get pinned. */
989 folio_try_dup_anon_rmap_pte(folio, page, dst_vma, src_vma);
990
991 /*
992 * We do not preserve soft-dirty information, because so
993 * far, checkpoint/restore is the only feature that
994 * requires that. And checkpoint/restore does not work
995 * when a device driver is involved (you cannot easily
996 * save and restore device driver state).
997 */
998 if (is_writable_device_private_entry(entry) &&
999 is_cow_mapping(flags: vm_flags)) {
1000 entry = make_readable_device_private_entry(
1001 offset: swp_offset(entry));
1002 pte = swp_entry_to_pte(entry);
1003 if (pte_swp_uffd_wp(pte: orig_pte))
1004 pte = pte_swp_mkuffd_wp(pte);
1005 set_pte_at(src_mm, addr, src_pte, pte);
1006 }
1007 } else if (is_device_exclusive_entry(entry)) {
1008 /*
1009 * Make device exclusive entries present by restoring the
1010 * original entry then copying as for a present pte. Device
1011 * exclusive entries currently only support private writable
1012 * (ie. COW) mappings.
1013 */
1014 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
1015 if (try_restore_exclusive_pte(vma: src_vma, addr, ptep: src_pte, orig_pte))
1016 return -EBUSY;
1017 return -ENOENT;
1018 } else if (is_pte_marker_entry(entry)) {
1019 pte_marker marker = copy_pte_marker(entry, dst_vma);
1020
1021 if (marker)
1022 set_pte_at(dst_mm, addr, dst_pte,
1023 make_pte_marker(marker));
1024 return 0;
1025 }
1026 if (!userfaultfd_wp(vma: dst_vma))
1027 pte = pte_swp_clear_uffd_wp(pte);
1028 set_pte_at(dst_mm, addr, dst_pte, pte);
1029 return 0;
1030}
1031
1032/*
1033 * Copy a present and normal page.
1034 *
1035 * NOTE! The usual case is that this isn't required;
1036 * instead, the caller can just increase the page refcount
1037 * and re-use the pte the traditional way.
1038 *
1039 * And if we need a pre-allocated page but don't yet have
1040 * one, return a negative error to let the preallocation
1041 * code know so that it can do so outside the page table
1042 * lock.
1043 */
1044static inline int
1045copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1046 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
1047 struct folio **prealloc, struct page *page)
1048{
1049 struct folio *new_folio;
1050 pte_t pte;
1051
1052 new_folio = *prealloc;
1053 if (!new_folio)
1054 return -EAGAIN;
1055
1056 /*
1057 * We have a prealloc page, all good! Take it
1058 * over and copy the page & arm it.
1059 */
1060
1061 if (copy_mc_user_highpage(to: &new_folio->page, from: page, vaddr: addr, vma: src_vma))
1062 return -EHWPOISON;
1063
1064 *prealloc = NULL;
1065 __folio_mark_uptodate(folio: new_folio);
1066 folio_add_new_anon_rmap(new_folio, dst_vma, address: addr, RMAP_EXCLUSIVE);
1067 folio_add_lru_vma(new_folio, dst_vma);
1068 rss[MM_ANONPAGES]++;
1069
1070 /* All done, just insert the new page copy in the child */
1071 pte = folio_mk_pte(folio: new_folio, pgprot: dst_vma->vm_page_prot);
1072 pte = maybe_mkwrite(pte: pte_mkdirty(pte), vma: dst_vma);
1073 if (userfaultfd_pte_wp(vma: dst_vma, pte: ptep_get(ptep: src_pte)))
1074 /* Uffd-wp needs to be delivered to dest pte as well */
1075 pte = pte_mkuffd_wp(pte);
1076 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
1077 return 0;
1078}
1079
1080static __always_inline void __copy_present_ptes(struct vm_area_struct *dst_vma,
1081 struct vm_area_struct *src_vma, pte_t *dst_pte, pte_t *src_pte,
1082 pte_t pte, unsigned long addr, int nr)
1083{
1084 struct mm_struct *src_mm = src_vma->vm_mm;
1085
1086 /* If it's a COW mapping, write protect it both processes. */
1087 if (is_cow_mapping(flags: src_vma->vm_flags) && pte_write(pte)) {
1088 wrprotect_ptes(mm: src_mm, addr, ptep: src_pte, nr);
1089 pte = pte_wrprotect(pte);
1090 }
1091
1092 /* If it's a shared mapping, mark it clean in the child. */
1093 if (src_vma->vm_flags & VM_SHARED)
1094 pte = pte_mkclean(pte);
1095 pte = pte_mkold(pte);
1096
1097 if (!userfaultfd_wp(vma: dst_vma))
1098 pte = pte_clear_uffd_wp(pte);
1099
1100 set_ptes(mm: dst_vma->vm_mm, addr, ptep: dst_pte, pte, nr);
1101}
1102
1103/*
1104 * Copy one present PTE, trying to batch-process subsequent PTEs that map
1105 * consecutive pages of the same folio by copying them as well.
1106 *
1107 * Returns -EAGAIN if one preallocated page is required to copy the next PTE.
1108 * Otherwise, returns the number of copied PTEs (at least 1).
1109 */
1110static inline int
1111copy_present_ptes(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1112 pte_t *dst_pte, pte_t *src_pte, pte_t pte, unsigned long addr,
1113 int max_nr, int *rss, struct folio **prealloc)
1114{
1115 fpb_t flags = FPB_MERGE_WRITE;
1116 struct page *page;
1117 struct folio *folio;
1118 int err, nr;
1119
1120 page = vm_normal_page(vma: src_vma, addr, pte);
1121 if (unlikely(!page))
1122 goto copy_pte;
1123
1124 folio = page_folio(page);
1125
1126 /*
1127 * If we likely have to copy, just don't bother with batching. Make
1128 * sure that the common "small folio" case is as fast as possible
1129 * by keeping the batching logic separate.
1130 */
1131 if (unlikely(!*prealloc && folio_test_large(folio) && max_nr != 1)) {
1132 if (!(src_vma->vm_flags & VM_SHARED))
1133 flags |= FPB_RESPECT_DIRTY;
1134 if (vma_soft_dirty_enabled(vma: src_vma))
1135 flags |= FPB_RESPECT_SOFT_DIRTY;
1136
1137 nr = folio_pte_batch_flags(folio, vma: src_vma, ptep: src_pte, ptentp: &pte, max_nr, flags);
1138 folio_ref_add(folio, nr);
1139 if (folio_test_anon(folio)) {
1140 if (unlikely(folio_try_dup_anon_rmap_ptes(folio, page,
1141 nr, dst_vma, src_vma))) {
1142 folio_ref_sub(folio, nr);
1143 return -EAGAIN;
1144 }
1145 rss[MM_ANONPAGES] += nr;
1146 VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio);
1147 } else {
1148 folio_dup_file_rmap_ptes(folio, page, nr_pages: nr, dst_vma);
1149 rss[mm_counter_file(folio)] += nr;
1150 }
1151 __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte,
1152 addr, nr);
1153 return nr;
1154 }
1155
1156 folio_get(folio);
1157 if (folio_test_anon(folio)) {
1158 /*
1159 * If this page may have been pinned by the parent process,
1160 * copy the page immediately for the child so that we'll always
1161 * guarantee the pinned page won't be randomly replaced in the
1162 * future.
1163 */
1164 if (unlikely(folio_try_dup_anon_rmap_pte(folio, page, dst_vma, src_vma))) {
1165 /* Page may be pinned, we have to copy. */
1166 folio_put(folio);
1167 err = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
1168 addr, rss, prealloc, page);
1169 return err ? err : 1;
1170 }
1171 rss[MM_ANONPAGES]++;
1172 VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio);
1173 } else {
1174 folio_dup_file_rmap_pte(folio, page, dst_vma);
1175 rss[mm_counter_file(folio)]++;
1176 }
1177
1178copy_pte:
1179 __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte, addr, nr: 1);
1180 return 1;
1181}
1182
1183static inline struct folio *folio_prealloc(struct mm_struct *src_mm,
1184 struct vm_area_struct *vma, unsigned long addr, bool need_zero)
1185{
1186 struct folio *new_folio;
1187
1188 if (need_zero)
1189 new_folio = vma_alloc_zeroed_movable_folio(vma, addr);
1190 else
1191 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma, addr);
1192
1193 if (!new_folio)
1194 return NULL;
1195
1196 if (mem_cgroup_charge(folio: new_folio, mm: src_mm, GFP_KERNEL)) {
1197 folio_put(folio: new_folio);
1198 return NULL;
1199 }
1200 folio_throttle_swaprate(folio: new_folio, GFP_KERNEL);
1201
1202 return new_folio;
1203}
1204
1205static int
1206copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1207 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1208 unsigned long end)
1209{
1210 struct mm_struct *dst_mm = dst_vma->vm_mm;
1211 struct mm_struct *src_mm = src_vma->vm_mm;
1212 pte_t *orig_src_pte, *orig_dst_pte;
1213 pte_t *src_pte, *dst_pte;
1214 pmd_t dummy_pmdval;
1215 pte_t ptent;
1216 spinlock_t *src_ptl, *dst_ptl;
1217 int progress, max_nr, ret = 0;
1218 int rss[NR_MM_COUNTERS];
1219 swp_entry_t entry = (swp_entry_t){0};
1220 struct folio *prealloc = NULL;
1221 int nr;
1222
1223again:
1224 progress = 0;
1225 init_rss_vec(rss);
1226
1227 /*
1228 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1229 * error handling here, assume that exclusive mmap_lock on dst and src
1230 * protects anon from unexpected THP transitions; with shmem and file
1231 * protected by mmap_lock-less collapse skipping areas with anon_vma
1232 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1233 * can remove such assumptions later, but this is good enough for now.
1234 */
1235 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1236 if (!dst_pte) {
1237 ret = -ENOMEM;
1238 goto out;
1239 }
1240
1241 /*
1242 * We already hold the exclusive mmap_lock, the copy_pte_range() and
1243 * retract_page_tables() are using vma->anon_vma to be exclusive, so
1244 * the PTE page is stable, and there is no need to get pmdval and do
1245 * pmd_same() check.
1246 */
1247 src_pte = pte_offset_map_rw_nolock(mm: src_mm, pmd: src_pmd, addr, pmdvalp: &dummy_pmdval,
1248 ptlp: &src_ptl);
1249 if (!src_pte) {
1250 pte_unmap_unlock(dst_pte, dst_ptl);
1251 /* ret == 0 */
1252 goto out;
1253 }
1254 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1255 orig_src_pte = src_pte;
1256 orig_dst_pte = dst_pte;
1257 arch_enter_lazy_mmu_mode();
1258
1259 do {
1260 nr = 1;
1261
1262 /*
1263 * We are holding two locks at this point - either of them
1264 * could generate latencies in another task on another CPU.
1265 */
1266 if (progress >= 32) {
1267 progress = 0;
1268 if (need_resched() ||
1269 spin_needbreak(lock: src_ptl) || spin_needbreak(lock: dst_ptl))
1270 break;
1271 }
1272 ptent = ptep_get(ptep: src_pte);
1273 if (pte_none(pte: ptent)) {
1274 progress++;
1275 continue;
1276 }
1277 if (unlikely(!pte_present(ptent))) {
1278 ret = copy_nonpresent_pte(dst_mm, src_mm,
1279 dst_pte, src_pte,
1280 dst_vma, src_vma,
1281 addr, rss);
1282 if (ret == -EIO) {
1283 entry = pte_to_swp_entry(pte: ptep_get(ptep: src_pte));
1284 break;
1285 } else if (ret == -EBUSY) {
1286 break;
1287 } else if (!ret) {
1288 progress += 8;
1289 continue;
1290 }
1291 ptent = ptep_get(ptep: src_pte);
1292 VM_WARN_ON_ONCE(!pte_present(ptent));
1293
1294 /*
1295 * Device exclusive entry restored, continue by copying
1296 * the now present pte.
1297 */
1298 WARN_ON_ONCE(ret != -ENOENT);
1299 }
1300 /* copy_present_ptes() will clear `*prealloc' if consumed */
1301 max_nr = (end - addr) / PAGE_SIZE;
1302 ret = copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte,
1303 pte: ptent, addr, max_nr, rss, prealloc: &prealloc);
1304 /*
1305 * If we need a pre-allocated page for this pte, drop the
1306 * locks, allocate, and try again.
1307 * If copy failed due to hwpoison in source page, break out.
1308 */
1309 if (unlikely(ret == -EAGAIN || ret == -EHWPOISON))
1310 break;
1311 if (unlikely(prealloc)) {
1312 /*
1313 * pre-alloc page cannot be reused by next time so as
1314 * to strictly follow mempolicy (e.g., alloc_page_vma()
1315 * will allocate page according to address). This
1316 * could only happen if one pinned pte changed.
1317 */
1318 folio_put(folio: prealloc);
1319 prealloc = NULL;
1320 }
1321 nr = ret;
1322 progress += 8 * nr;
1323 } while (dst_pte += nr, src_pte += nr, addr += PAGE_SIZE * nr,
1324 addr != end);
1325
1326 arch_leave_lazy_mmu_mode();
1327 pte_unmap_unlock(orig_src_pte, src_ptl);
1328 add_mm_rss_vec(mm: dst_mm, rss);
1329 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1330 cond_resched();
1331
1332 if (ret == -EIO) {
1333 VM_WARN_ON_ONCE(!entry.val);
1334 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1335 ret = -ENOMEM;
1336 goto out;
1337 }
1338 entry.val = 0;
1339 } else if (ret == -EBUSY || unlikely(ret == -EHWPOISON)) {
1340 goto out;
1341 } else if (ret == -EAGAIN) {
1342 prealloc = folio_prealloc(src_mm, vma: src_vma, addr, need_zero: false);
1343 if (!prealloc)
1344 return -ENOMEM;
1345 } else if (ret < 0) {
1346 VM_WARN_ON_ONCE(1);
1347 }
1348
1349 /* We've captured and resolved the error. Reset, try again. */
1350 ret = 0;
1351
1352 if (addr != end)
1353 goto again;
1354out:
1355 if (unlikely(prealloc))
1356 folio_put(folio: prealloc);
1357 return ret;
1358}
1359
1360static inline int
1361copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1362 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1363 unsigned long end)
1364{
1365 struct mm_struct *dst_mm = dst_vma->vm_mm;
1366 struct mm_struct *src_mm = src_vma->vm_mm;
1367 pmd_t *src_pmd, *dst_pmd;
1368 unsigned long next;
1369
1370 dst_pmd = pmd_alloc(mm: dst_mm, pud: dst_pud, address: addr);
1371 if (!dst_pmd)
1372 return -ENOMEM;
1373 src_pmd = pmd_offset(pud: src_pud, address: addr);
1374 do {
1375 next = pmd_addr_end(addr, end);
1376 if (is_swap_pmd(pmd: *src_pmd) || pmd_trans_huge(pmd: *src_pmd)) {
1377 int err;
1378 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1379 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1380 addr, dst_vma, src_vma);
1381 if (err == -ENOMEM)
1382 return -ENOMEM;
1383 if (!err)
1384 continue;
1385 /* fall through */
1386 }
1387 if (pmd_none_or_clear_bad(pmd: src_pmd))
1388 continue;
1389 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1390 addr, end: next))
1391 return -ENOMEM;
1392 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1393 return 0;
1394}
1395
1396static inline int
1397copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1398 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1399 unsigned long end)
1400{
1401 struct mm_struct *dst_mm = dst_vma->vm_mm;
1402 struct mm_struct *src_mm = src_vma->vm_mm;
1403 pud_t *src_pud, *dst_pud;
1404 unsigned long next;
1405
1406 dst_pud = pud_alloc(mm: dst_mm, p4d: dst_p4d, address: addr);
1407 if (!dst_pud)
1408 return -ENOMEM;
1409 src_pud = pud_offset(p4d: src_p4d, address: addr);
1410 do {
1411 next = pud_addr_end(addr, end);
1412 if (pud_trans_huge(pud: *src_pud)) {
1413 int err;
1414
1415 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1416 err = copy_huge_pud(dst_mm, src_mm,
1417 dst_pud, src_pud, addr, vma: src_vma);
1418 if (err == -ENOMEM)
1419 return -ENOMEM;
1420 if (!err)
1421 continue;
1422 /* fall through */
1423 }
1424 if (pud_none_or_clear_bad(pud: src_pud))
1425 continue;
1426 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1427 addr, end: next))
1428 return -ENOMEM;
1429 } while (dst_pud++, src_pud++, addr = next, addr != end);
1430 return 0;
1431}
1432
1433static inline int
1434copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1435 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1436 unsigned long end)
1437{
1438 struct mm_struct *dst_mm = dst_vma->vm_mm;
1439 p4d_t *src_p4d, *dst_p4d;
1440 unsigned long next;
1441
1442 dst_p4d = p4d_alloc(mm: dst_mm, pgd: dst_pgd, address: addr);
1443 if (!dst_p4d)
1444 return -ENOMEM;
1445 src_p4d = p4d_offset(pgd: src_pgd, address: addr);
1446 do {
1447 next = p4d_addr_end(addr, end);
1448 if (p4d_none_or_clear_bad(p4d: src_p4d))
1449 continue;
1450 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1451 addr, end: next))
1452 return -ENOMEM;
1453 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1454 return 0;
1455}
1456
1457/*
1458 * Return true if the vma needs to copy the pgtable during this fork(). Return
1459 * false when we can speed up fork() by allowing lazy page faults later until
1460 * when the child accesses the memory range.
1461 */
1462static bool
1463vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1464{
1465 /*
1466 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1467 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1468 * contains uffd-wp protection information, that's something we can't
1469 * retrieve from page cache, and skip copying will lose those info.
1470 */
1471 if (userfaultfd_wp(vma: dst_vma))
1472 return true;
1473
1474 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1475 return true;
1476
1477 if (src_vma->anon_vma)
1478 return true;
1479
1480 /*
1481 * Don't copy ptes where a page fault will fill them correctly. Fork
1482 * becomes much lighter when there are big shared or private readonly
1483 * mappings. The tradeoff is that copy_page_range is more efficient
1484 * than faulting.
1485 */
1486 return false;
1487}
1488
1489int
1490copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1491{
1492 pgd_t *src_pgd, *dst_pgd;
1493 unsigned long addr = src_vma->vm_start;
1494 unsigned long end = src_vma->vm_end;
1495 struct mm_struct *dst_mm = dst_vma->vm_mm;
1496 struct mm_struct *src_mm = src_vma->vm_mm;
1497 struct mmu_notifier_range range;
1498 unsigned long next;
1499 bool is_cow;
1500 int ret;
1501
1502 if (!vma_needs_copy(dst_vma, src_vma))
1503 return 0;
1504
1505 if (is_vm_hugetlb_page(vma: src_vma))
1506 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1507
1508 /*
1509 * We need to invalidate the secondary MMU mappings only when
1510 * there could be a permission downgrade on the ptes of the
1511 * parent mm. And a permission downgrade will only happen if
1512 * is_cow_mapping() returns true.
1513 */
1514 is_cow = is_cow_mapping(flags: src_vma->vm_flags);
1515
1516 if (is_cow) {
1517 mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_PROTECTION_PAGE,
1518 flags: 0, mm: src_mm, start: addr, end);
1519 mmu_notifier_invalidate_range_start(range: &range);
1520 /*
1521 * Disabling preemption is not needed for the write side, as
1522 * the read side doesn't spin, but goes to the mmap_lock.
1523 *
1524 * Use the raw variant of the seqcount_t write API to avoid
1525 * lockdep complaining about preemptibility.
1526 */
1527 vma_assert_write_locked(vma: src_vma);
1528 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1529 }
1530
1531 ret = 0;
1532 dst_pgd = pgd_offset(dst_mm, addr);
1533 src_pgd = pgd_offset(src_mm, addr);
1534 do {
1535 next = pgd_addr_end(addr, end);
1536 if (pgd_none_or_clear_bad(pgd: src_pgd))
1537 continue;
1538 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1539 addr, next))) {
1540 ret = -ENOMEM;
1541 break;
1542 }
1543 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1544
1545 if (is_cow) {
1546 raw_write_seqcount_end(&src_mm->write_protect_seq);
1547 mmu_notifier_invalidate_range_end(range: &range);
1548 }
1549 return ret;
1550}
1551
1552/* Whether we should zap all COWed (private) pages too */
1553static inline bool should_zap_cows(struct zap_details *details)
1554{
1555 /* By default, zap all pages */
1556 if (!details || details->reclaim_pt)
1557 return true;
1558
1559 /* Or, we zap COWed pages only if the caller wants to */
1560 return details->even_cows;
1561}
1562
1563/* Decides whether we should zap this folio with the folio pointer specified */
1564static inline bool should_zap_folio(struct zap_details *details,
1565 struct folio *folio)
1566{
1567 /* If we can make a decision without *folio.. */
1568 if (should_zap_cows(details))
1569 return true;
1570
1571 /* Otherwise we should only zap non-anon folios */
1572 return !folio_test_anon(folio);
1573}
1574
1575static inline bool zap_drop_markers(struct zap_details *details)
1576{
1577 if (!details)
1578 return false;
1579
1580 return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1581}
1582
1583/*
1584 * This function makes sure that we'll replace the none pte with an uffd-wp
1585 * swap special pte marker when necessary. Must be with the pgtable lock held.
1586 *
1587 * Returns true if uffd-wp ptes was installed, false otherwise.
1588 */
1589static inline bool
1590zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1591 unsigned long addr, pte_t *pte, int nr,
1592 struct zap_details *details, pte_t pteval)
1593{
1594 bool was_installed = false;
1595
1596#ifdef CONFIG_PTE_MARKER_UFFD_WP
1597 /* Zap on anonymous always means dropping everything */
1598 if (vma_is_anonymous(vma))
1599 return false;
1600
1601 if (zap_drop_markers(details))
1602 return false;
1603
1604 for (;;) {
1605 /* the PFN in the PTE is irrelevant. */
1606 if (pte_install_uffd_wp_if_needed(vma, addr, pte, pteval))
1607 was_installed = true;
1608 if (--nr == 0)
1609 break;
1610 pte++;
1611 addr += PAGE_SIZE;
1612 }
1613#endif
1614 return was_installed;
1615}
1616
1617static __always_inline void zap_present_folio_ptes(struct mmu_gather *tlb,
1618 struct vm_area_struct *vma, struct folio *folio,
1619 struct page *page, pte_t *pte, pte_t ptent, unsigned int nr,
1620 unsigned long addr, struct zap_details *details, int *rss,
1621 bool *force_flush, bool *force_break, bool *any_skipped)
1622{
1623 struct mm_struct *mm = tlb->mm;
1624 bool delay_rmap = false;
1625
1626 if (!folio_test_anon(folio)) {
1627 ptent = get_and_clear_full_ptes(mm, addr, ptep: pte, nr, full: tlb->fullmm);
1628 if (pte_dirty(pte: ptent)) {
1629 folio_mark_dirty(folio);
1630 if (tlb_delay_rmap(tlb)) {
1631 delay_rmap = true;
1632 *force_flush = true;
1633 }
1634 }
1635 if (pte_young(pte: ptent) && likely(vma_has_recency(vma)))
1636 folio_mark_accessed(folio);
1637 rss[mm_counter(folio)] -= nr;
1638 } else {
1639 /* We don't need up-to-date accessed/dirty bits. */
1640 clear_full_ptes(mm, addr, ptep: pte, nr, full: tlb->fullmm);
1641 rss[MM_ANONPAGES] -= nr;
1642 }
1643 /* Checking a single PTE in a batch is sufficient. */
1644 arch_check_zapped_pte(vma, pte: ptent);
1645 tlb_remove_tlb_entries(tlb, ptep: pte, nr, address: addr);
1646 if (unlikely(userfaultfd_pte_wp(vma, ptent)))
1647 *any_skipped = zap_install_uffd_wp_if_needed(vma, addr, pte,
1648 nr, details, pteval: ptent);
1649
1650 if (!delay_rmap) {
1651 folio_remove_rmap_ptes(folio, page, nr_pages: nr, vma);
1652
1653 if (unlikely(folio_mapcount(folio) < 0))
1654 print_bad_pte(vma, addr, ptent, page);
1655 }
1656 if (unlikely(__tlb_remove_folio_pages(tlb, page, nr, delay_rmap))) {
1657 *force_flush = true;
1658 *force_break = true;
1659 }
1660}
1661
1662/*
1663 * Zap or skip at least one present PTE, trying to batch-process subsequent
1664 * PTEs that map consecutive pages of the same folio.
1665 *
1666 * Returns the number of processed (skipped or zapped) PTEs (at least 1).
1667 */
1668static inline int zap_present_ptes(struct mmu_gather *tlb,
1669 struct vm_area_struct *vma, pte_t *pte, pte_t ptent,
1670 unsigned int max_nr, unsigned long addr,
1671 struct zap_details *details, int *rss, bool *force_flush,
1672 bool *force_break, bool *any_skipped)
1673{
1674 struct mm_struct *mm = tlb->mm;
1675 struct folio *folio;
1676 struct page *page;
1677 int nr;
1678
1679 page = vm_normal_page(vma, addr, pte: ptent);
1680 if (!page) {
1681 /* We don't need up-to-date accessed/dirty bits. */
1682 ptep_get_and_clear_full(mm, addr, ptep: pte, full: tlb->fullmm);
1683 arch_check_zapped_pte(vma, pte: ptent);
1684 tlb_remove_tlb_entry(tlb, pte, addr);
1685 if (userfaultfd_pte_wp(vma, pte: ptent))
1686 *any_skipped = zap_install_uffd_wp_if_needed(vma, addr,
1687 pte, nr: 1, details, pteval: ptent);
1688 ksm_might_unmap_zero_page(mm, pte: ptent);
1689 return 1;
1690 }
1691
1692 folio = page_folio(page);
1693 if (unlikely(!should_zap_folio(details, folio))) {
1694 *any_skipped = true;
1695 return 1;
1696 }
1697
1698 /*
1699 * Make sure that the common "small folio" case is as fast as possible
1700 * by keeping the batching logic separate.
1701 */
1702 if (unlikely(folio_test_large(folio) && max_nr != 1)) {
1703 nr = folio_pte_batch(folio, ptep: pte, pte: ptent, max_nr);
1704 zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, nr,
1705 addr, details, rss, force_flush,
1706 force_break, any_skipped);
1707 return nr;
1708 }
1709 zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, nr: 1, addr,
1710 details, rss, force_flush, force_break, any_skipped);
1711 return 1;
1712}
1713
1714static inline int zap_nonpresent_ptes(struct mmu_gather *tlb,
1715 struct vm_area_struct *vma, pte_t *pte, pte_t ptent,
1716 unsigned int max_nr, unsigned long addr,
1717 struct zap_details *details, int *rss, bool *any_skipped)
1718{
1719 swp_entry_t entry;
1720 int nr = 1;
1721
1722 *any_skipped = true;
1723 entry = pte_to_swp_entry(pte: ptent);
1724 if (is_device_private_entry(entry) ||
1725 is_device_exclusive_entry(entry)) {
1726 struct page *page = pfn_swap_entry_to_page(entry);
1727 struct folio *folio = page_folio(page);
1728
1729 if (unlikely(!should_zap_folio(details, folio)))
1730 return 1;
1731 /*
1732 * Both device private/exclusive mappings should only
1733 * work with anonymous page so far, so we don't need to
1734 * consider uffd-wp bit when zap. For more information,
1735 * see zap_install_uffd_wp_if_needed().
1736 */
1737 WARN_ON_ONCE(!vma_is_anonymous(vma));
1738 rss[mm_counter(folio)]--;
1739 folio_remove_rmap_pte(folio, page, vma);
1740 folio_put(folio);
1741 } else if (!non_swap_entry(entry)) {
1742 /* Genuine swap entries, hence a private anon pages */
1743 if (!should_zap_cows(details))
1744 return 1;
1745
1746 nr = swap_pte_batch(start_ptep: pte, max_nr, pte: ptent);
1747 rss[MM_SWAPENTS] -= nr;
1748 free_swap_and_cache_nr(entry, nr);
1749 } else if (is_migration_entry(entry)) {
1750 struct folio *folio = pfn_swap_entry_folio(entry);
1751
1752 if (!should_zap_folio(details, folio))
1753 return 1;
1754 rss[mm_counter(folio)]--;
1755 } else if (pte_marker_entry_uffd_wp(entry)) {
1756 /*
1757 * For anon: always drop the marker; for file: only
1758 * drop the marker if explicitly requested.
1759 */
1760 if (!vma_is_anonymous(vma) && !zap_drop_markers(details))
1761 return 1;
1762 } else if (is_guard_swp_entry(entry)) {
1763 /*
1764 * Ordinary zapping should not remove guard PTE
1765 * markers. Only do so if we should remove PTE markers
1766 * in general.
1767 */
1768 if (!zap_drop_markers(details))
1769 return 1;
1770 } else if (is_hwpoison_entry(swp: entry) || is_poisoned_swp_entry(entry)) {
1771 if (!should_zap_cows(details))
1772 return 1;
1773 } else {
1774 /* We should have covered all the swap entry types */
1775 pr_alert("unrecognized swap entry 0x%lx\n", entry.val);
1776 WARN_ON_ONCE(1);
1777 }
1778 clear_not_present_full_ptes(mm: vma->vm_mm, addr, ptep: pte, nr, full: tlb->fullmm);
1779 *any_skipped = zap_install_uffd_wp_if_needed(vma, addr, pte, nr, details, pteval: ptent);
1780
1781 return nr;
1782}
1783
1784static inline int do_zap_pte_range(struct mmu_gather *tlb,
1785 struct vm_area_struct *vma, pte_t *pte,
1786 unsigned long addr, unsigned long end,
1787 struct zap_details *details, int *rss,
1788 bool *force_flush, bool *force_break,
1789 bool *any_skipped)
1790{
1791 pte_t ptent = ptep_get(ptep: pte);
1792 int max_nr = (end - addr) / PAGE_SIZE;
1793 int nr = 0;
1794
1795 /* Skip all consecutive none ptes */
1796 if (pte_none(pte: ptent)) {
1797 for (nr = 1; nr < max_nr; nr++) {
1798 ptent = ptep_get(ptep: pte + nr);
1799 if (!pte_none(pte: ptent))
1800 break;
1801 }
1802 max_nr -= nr;
1803 if (!max_nr)
1804 return nr;
1805 pte += nr;
1806 addr += nr * PAGE_SIZE;
1807 }
1808
1809 if (pte_present(a: ptent))
1810 nr += zap_present_ptes(tlb, vma, pte, ptent, max_nr, addr,
1811 details, rss, force_flush, force_break,
1812 any_skipped);
1813 else
1814 nr += zap_nonpresent_ptes(tlb, vma, pte, ptent, max_nr, addr,
1815 details, rss, any_skipped);
1816
1817 return nr;
1818}
1819
1820static unsigned long zap_pte_range(struct mmu_gather *tlb,
1821 struct vm_area_struct *vma, pmd_t *pmd,
1822 unsigned long addr, unsigned long end,
1823 struct zap_details *details)
1824{
1825 bool force_flush = false, force_break = false;
1826 struct mm_struct *mm = tlb->mm;
1827 int rss[NR_MM_COUNTERS];
1828 spinlock_t *ptl;
1829 pte_t *start_pte;
1830 pte_t *pte;
1831 pmd_t pmdval;
1832 unsigned long start = addr;
1833 bool can_reclaim_pt = reclaim_pt_is_enabled(start, end, details);
1834 bool direct_reclaim = true;
1835 int nr;
1836
1837retry:
1838 tlb_change_page_size(tlb, PAGE_SIZE);
1839 init_rss_vec(rss);
1840 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, ptlp: &ptl);
1841 if (!pte)
1842 return addr;
1843
1844 flush_tlb_batched_pending(mm);
1845 arch_enter_lazy_mmu_mode();
1846 do {
1847 bool any_skipped = false;
1848
1849 if (need_resched()) {
1850 direct_reclaim = false;
1851 break;
1852 }
1853
1854 nr = do_zap_pte_range(tlb, vma, pte, addr, end, details, rss,
1855 force_flush: &force_flush, force_break: &force_break, any_skipped: &any_skipped);
1856 if (any_skipped)
1857 can_reclaim_pt = false;
1858 if (unlikely(force_break)) {
1859 addr += nr * PAGE_SIZE;
1860 direct_reclaim = false;
1861 break;
1862 }
1863 } while (pte += nr, addr += PAGE_SIZE * nr, addr != end);
1864
1865 /*
1866 * Fast path: try to hold the pmd lock and unmap the PTE page.
1867 *
1868 * If the pte lock was released midway (retry case), or if the attempt
1869 * to hold the pmd lock failed, then we need to recheck all pte entries
1870 * to ensure they are still none, thereby preventing the pte entries
1871 * from being repopulated by another thread.
1872 */
1873 if (can_reclaim_pt && direct_reclaim && addr == end)
1874 direct_reclaim = try_get_and_clear_pmd(mm, pmd, pmdval: &pmdval);
1875
1876 add_mm_rss_vec(mm, rss);
1877 arch_leave_lazy_mmu_mode();
1878
1879 /* Do the actual TLB flush before dropping ptl */
1880 if (force_flush) {
1881 tlb_flush_mmu_tlbonly(tlb);
1882 tlb_flush_rmaps(tlb, vma);
1883 }
1884 pte_unmap_unlock(start_pte, ptl);
1885
1886 /*
1887 * If we forced a TLB flush (either due to running out of
1888 * batch buffers or because we needed to flush dirty TLB
1889 * entries before releasing the ptl), free the batched
1890 * memory too. Come back again if we didn't do everything.
1891 */
1892 if (force_flush)
1893 tlb_flush_mmu(tlb);
1894
1895 if (addr != end) {
1896 cond_resched();
1897 force_flush = false;
1898 force_break = false;
1899 goto retry;
1900 }
1901
1902 if (can_reclaim_pt) {
1903 if (direct_reclaim)
1904 free_pte(mm, addr: start, tlb, pmdval);
1905 else
1906 try_to_free_pte(mm, pmd, addr: start, tlb);
1907 }
1908
1909 return addr;
1910}
1911
1912static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1913 struct vm_area_struct *vma, pud_t *pud,
1914 unsigned long addr, unsigned long end,
1915 struct zap_details *details)
1916{
1917 pmd_t *pmd;
1918 unsigned long next;
1919
1920 pmd = pmd_offset(pud, address: addr);
1921 do {
1922 next = pmd_addr_end(addr, end);
1923 if (is_swap_pmd(pmd: *pmd) || pmd_trans_huge(pmd: *pmd)) {
1924 if (next - addr != HPAGE_PMD_SIZE)
1925 __split_huge_pmd(vma, pmd, address: addr, freeze: false);
1926 else if (zap_huge_pmd(tlb, vma, pmd, addr)) {
1927 addr = next;
1928 continue;
1929 }
1930 /* fall through */
1931 } else if (details && details->single_folio &&
1932 folio_test_pmd_mappable(folio: details->single_folio) &&
1933 next - addr == HPAGE_PMD_SIZE && pmd_none(pmd: *pmd)) {
1934 spinlock_t *ptl = pmd_lock(mm: tlb->mm, pmd);
1935 /*
1936 * Take and drop THP pmd lock so that we cannot return
1937 * prematurely, while zap_huge_pmd() has cleared *pmd,
1938 * but not yet decremented compound_mapcount().
1939 */
1940 spin_unlock(lock: ptl);
1941 }
1942 if (pmd_none(pmd: *pmd)) {
1943 addr = next;
1944 continue;
1945 }
1946 addr = zap_pte_range(tlb, vma, pmd, addr, end: next, details);
1947 if (addr != next)
1948 pmd--;
1949 } while (pmd++, cond_resched(), addr != end);
1950
1951 return addr;
1952}
1953
1954static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1955 struct vm_area_struct *vma, p4d_t *p4d,
1956 unsigned long addr, unsigned long end,
1957 struct zap_details *details)
1958{
1959 pud_t *pud;
1960 unsigned long next;
1961
1962 pud = pud_offset(p4d, address: addr);
1963 do {
1964 next = pud_addr_end(addr, end);
1965 if (pud_trans_huge(pud: *pud)) {
1966 if (next - addr != HPAGE_PUD_SIZE) {
1967 mmap_assert_locked(mm: tlb->mm);
1968 split_huge_pud(vma, pud, addr);
1969 } else if (zap_huge_pud(tlb, vma, pud, addr))
1970 goto next;
1971 /* fall through */
1972 }
1973 if (pud_none_or_clear_bad(pud))
1974 continue;
1975 next = zap_pmd_range(tlb, vma, pud, addr, end: next, details);
1976next:
1977 cond_resched();
1978 } while (pud++, addr = next, addr != end);
1979
1980 return addr;
1981}
1982
1983static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1984 struct vm_area_struct *vma, pgd_t *pgd,
1985 unsigned long addr, unsigned long end,
1986 struct zap_details *details)
1987{
1988 p4d_t *p4d;
1989 unsigned long next;
1990
1991 p4d = p4d_offset(pgd, address: addr);
1992 do {
1993 next = p4d_addr_end(addr, end);
1994 if (p4d_none_or_clear_bad(p4d))
1995 continue;
1996 next = zap_pud_range(tlb, vma, p4d, addr, end: next, details);
1997 } while (p4d++, addr = next, addr != end);
1998
1999 return addr;
2000}
2001
2002void unmap_page_range(struct mmu_gather *tlb,
2003 struct vm_area_struct *vma,
2004 unsigned long addr, unsigned long end,
2005 struct zap_details *details)
2006{
2007 pgd_t *pgd;
2008 unsigned long next;
2009
2010 BUG_ON(addr >= end);
2011 tlb_start_vma(tlb, vma);
2012 pgd = pgd_offset(vma->vm_mm, addr);
2013 do {
2014 next = pgd_addr_end(addr, end);
2015 if (pgd_none_or_clear_bad(pgd))
2016 continue;
2017 next = zap_p4d_range(tlb, vma, pgd, addr, end: next, details);
2018 } while (pgd++, addr = next, addr != end);
2019 tlb_end_vma(tlb, vma);
2020}
2021
2022
2023static void unmap_single_vma(struct mmu_gather *tlb,
2024 struct vm_area_struct *vma, unsigned long start_addr,
2025 unsigned long end_addr,
2026 struct zap_details *details, bool mm_wr_locked)
2027{
2028 unsigned long start = max(vma->vm_start, start_addr);
2029 unsigned long end;
2030
2031 if (start >= vma->vm_end)
2032 return;
2033 end = min(vma->vm_end, end_addr);
2034 if (end <= vma->vm_start)
2035 return;
2036
2037 if (vma->vm_file)
2038 uprobe_munmap(vma, start, end);
2039
2040 if (start != end) {
2041 if (unlikely(is_vm_hugetlb_page(vma))) {
2042 /*
2043 * It is undesirable to test vma->vm_file as it
2044 * should be non-null for valid hugetlb area.
2045 * However, vm_file will be NULL in the error
2046 * cleanup path of mmap_region. When
2047 * hugetlbfs ->mmap method fails,
2048 * mmap_region() nullifies vma->vm_file
2049 * before calling this function to clean up.
2050 * Since no pte has actually been setup, it is
2051 * safe to do nothing in this case.
2052 */
2053 if (vma->vm_file) {
2054 zap_flags_t zap_flags = details ?
2055 details->zap_flags : 0;
2056 __unmap_hugepage_range(tlb, vma, start, end,
2057 NULL, zap_flags);
2058 }
2059 } else
2060 unmap_page_range(tlb, vma, addr: start, end, details);
2061 }
2062}
2063
2064/**
2065 * unmap_vmas - unmap a range of memory covered by a list of vma's
2066 * @tlb: address of the caller's struct mmu_gather
2067 * @mas: the maple state
2068 * @vma: the starting vma
2069 * @start_addr: virtual address at which to start unmapping
2070 * @end_addr: virtual address at which to end unmapping
2071 * @tree_end: The maximum index to check
2072 * @mm_wr_locked: lock flag
2073 *
2074 * Unmap all pages in the vma list.
2075 *
2076 * Only addresses between `start' and `end' will be unmapped.
2077 *
2078 * The VMA list must be sorted in ascending virtual address order.
2079 *
2080 * unmap_vmas() assumes that the caller will flush the whole unmapped address
2081 * range after unmap_vmas() returns. So the only responsibility here is to
2082 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
2083 * drops the lock and schedules.
2084 */
2085void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas,
2086 struct vm_area_struct *vma, unsigned long start_addr,
2087 unsigned long end_addr, unsigned long tree_end,
2088 bool mm_wr_locked)
2089{
2090 struct mmu_notifier_range range;
2091 struct zap_details details = {
2092 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
2093 /* Careful - we need to zap private pages too! */
2094 .even_cows = true,
2095 };
2096
2097 mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_UNMAP, flags: 0, mm: vma->vm_mm,
2098 start: start_addr, end: end_addr);
2099 mmu_notifier_invalidate_range_start(range: &range);
2100 do {
2101 unsigned long start = start_addr;
2102 unsigned long end = end_addr;
2103 hugetlb_zap_begin(vma, start: &start, end: &end);
2104 unmap_single_vma(tlb, vma, start_addr: start, end_addr: end, details: &details,
2105 mm_wr_locked);
2106 hugetlb_zap_end(vma, details: &details);
2107 vma = mas_find(mas, max: tree_end - 1);
2108 } while (vma && likely(!xa_is_zero(vma)));
2109 mmu_notifier_invalidate_range_end(range: &range);
2110}
2111
2112/**
2113 * zap_page_range_single_batched - remove user pages in a given range
2114 * @tlb: pointer to the caller's struct mmu_gather
2115 * @vma: vm_area_struct holding the applicable pages
2116 * @address: starting address of pages to remove
2117 * @size: number of bytes to remove
2118 * @details: details of shared cache invalidation
2119 *
2120 * @tlb shouldn't be NULL. The range must fit into one VMA. If @vma is for
2121 * hugetlb, @tlb is flushed and re-initialized by this function.
2122 */
2123void zap_page_range_single_batched(struct mmu_gather *tlb,
2124 struct vm_area_struct *vma, unsigned long address,
2125 unsigned long size, struct zap_details *details)
2126{
2127 const unsigned long end = address + size;
2128 struct mmu_notifier_range range;
2129
2130 VM_WARN_ON_ONCE(!tlb || tlb->mm != vma->vm_mm);
2131
2132 mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_CLEAR, flags: 0, mm: vma->vm_mm,
2133 start: address, end);
2134 hugetlb_zap_begin(vma, start: &range.start, end: &range.end);
2135 update_hiwater_rss(mm: vma->vm_mm);
2136 mmu_notifier_invalidate_range_start(range: &range);
2137 /*
2138 * unmap 'address-end' not 'range.start-range.end' as range
2139 * could have been expanded for hugetlb pmd sharing.
2140 */
2141 unmap_single_vma(tlb, vma, start_addr: address, end_addr: end, details, mm_wr_locked: false);
2142 mmu_notifier_invalidate_range_end(range: &range);
2143 if (is_vm_hugetlb_page(vma)) {
2144 /*
2145 * flush tlb and free resources before hugetlb_zap_end(), to
2146 * avoid concurrent page faults' allocation failure.
2147 */
2148 tlb_finish_mmu(tlb);
2149 hugetlb_zap_end(vma, details);
2150 tlb_gather_mmu(tlb, mm: vma->vm_mm);
2151 }
2152}
2153
2154/**
2155 * zap_page_range_single - remove user pages in a given range
2156 * @vma: vm_area_struct holding the applicable pages
2157 * @address: starting address of pages to zap
2158 * @size: number of bytes to zap
2159 * @details: details of shared cache invalidation
2160 *
2161 * The range must fit into one VMA.
2162 */
2163void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
2164 unsigned long size, struct zap_details *details)
2165{
2166 struct mmu_gather tlb;
2167
2168 tlb_gather_mmu(tlb: &tlb, mm: vma->vm_mm);
2169 zap_page_range_single_batched(tlb: &tlb, vma, address, size, details);
2170 tlb_finish_mmu(tlb: &tlb);
2171}
2172
2173/**
2174 * zap_vma_ptes - remove ptes mapping the vma
2175 * @vma: vm_area_struct holding ptes to be zapped
2176 * @address: starting address of pages to zap
2177 * @size: number of bytes to zap
2178 *
2179 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
2180 *
2181 * The entire address range must be fully contained within the vma.
2182 *
2183 */
2184void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
2185 unsigned long size)
2186{
2187 if (!range_in_vma(vma, start: address, end: address + size) ||
2188 !(vma->vm_flags & VM_PFNMAP))
2189 return;
2190
2191 zap_page_range_single(vma, address, size, NULL);
2192}
2193EXPORT_SYMBOL_GPL(zap_vma_ptes);
2194
2195static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
2196{
2197 pgd_t *pgd;
2198 p4d_t *p4d;
2199 pud_t *pud;
2200 pmd_t *pmd;
2201
2202 pgd = pgd_offset(mm, addr);
2203 p4d = p4d_alloc(mm, pgd, address: addr);
2204 if (!p4d)
2205 return NULL;
2206 pud = pud_alloc(mm, p4d, address: addr);
2207 if (!pud)
2208 return NULL;
2209 pmd = pmd_alloc(mm, pud, address: addr);
2210 if (!pmd)
2211 return NULL;
2212
2213 VM_BUG_ON(pmd_trans_huge(*pmd));
2214 return pmd;
2215}
2216
2217pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2218 spinlock_t **ptl)
2219{
2220 pmd_t *pmd = walk_to_pmd(mm, addr);
2221
2222 if (!pmd)
2223 return NULL;
2224 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2225}
2226
2227static bool vm_mixed_zeropage_allowed(struct vm_area_struct *vma)
2228{
2229 VM_WARN_ON_ONCE(vma->vm_flags & VM_PFNMAP);
2230 /*
2231 * Whoever wants to forbid the zeropage after some zeropages
2232 * might already have been mapped has to scan the page tables and
2233 * bail out on any zeropages. Zeropages in COW mappings can
2234 * be unshared using FAULT_FLAG_UNSHARE faults.
2235 */
2236 if (mm_forbids_zeropage(vma->vm_mm))
2237 return false;
2238 /* zeropages in COW mappings are common and unproblematic. */
2239 if (is_cow_mapping(flags: vma->vm_flags))
2240 return true;
2241 /* Mappings that do not allow for writable PTEs are unproblematic. */
2242 if (!(vma->vm_flags & (VM_WRITE | VM_MAYWRITE)))
2243 return true;
2244 /*
2245 * Why not allow any VMA that has vm_ops->pfn_mkwrite? GUP could
2246 * find the shared zeropage and longterm-pin it, which would
2247 * be problematic as soon as the zeropage gets replaced by a different
2248 * page due to vma->vm_ops->pfn_mkwrite, because what's mapped would
2249 * now differ to what GUP looked up. FSDAX is incompatible to
2250 * FOLL_LONGTERM and VM_IO is incompatible to GUP completely (see
2251 * check_vma_flags).
2252 */
2253 return vma->vm_ops && vma->vm_ops->pfn_mkwrite &&
2254 (vma_is_fsdax(vma) || vma->vm_flags & VM_IO);
2255}
2256
2257static int validate_page_before_insert(struct vm_area_struct *vma,
2258 struct page *page)
2259{
2260 struct folio *folio = page_folio(page);
2261
2262 if (!folio_ref_count(folio))
2263 return -EINVAL;
2264 if (unlikely(is_zero_folio(folio))) {
2265 if (!vm_mixed_zeropage_allowed(vma))
2266 return -EINVAL;
2267 return 0;
2268 }
2269 if (folio_test_anon(folio) || page_has_type(page))
2270 return -EINVAL;
2271 flush_dcache_folio(folio);
2272 return 0;
2273}
2274
2275static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
2276 unsigned long addr, struct page *page,
2277 pgprot_t prot, bool mkwrite)
2278{
2279 struct folio *folio = page_folio(page);
2280 pte_t pteval = ptep_get(ptep: pte);
2281
2282 if (!pte_none(pte: pteval)) {
2283 if (!mkwrite)
2284 return -EBUSY;
2285
2286 /* see insert_pfn(). */
2287 if (pte_pfn(pte: pteval) != page_to_pfn(page)) {
2288 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(pteval)));
2289 return -EFAULT;
2290 }
2291 pteval = maybe_mkwrite(pte: pteval, vma);
2292 pteval = pte_mkyoung(pte: pteval);
2293 if (ptep_set_access_flags(vma, address: addr, ptep: pte, entry: pteval, dirty: 1))
2294 update_mmu_cache(vma, addr, ptep: pte);
2295 return 0;
2296 }
2297
2298 /* Ok, finally just insert the thing.. */
2299 pteval = mk_pte(page, pgprot: prot);
2300 if (unlikely(is_zero_folio(folio))) {
2301 pteval = pte_mkspecial(pte: pteval);
2302 } else {
2303 folio_get(folio);
2304 pteval = mk_pte(page, pgprot: prot);
2305 if (mkwrite) {
2306 pteval = pte_mkyoung(pte: pteval);
2307 pteval = maybe_mkwrite(pte: pte_mkdirty(pte: pteval), vma);
2308 }
2309 inc_mm_counter(mm: vma->vm_mm, member: mm_counter_file(folio));
2310 folio_add_file_rmap_pte(folio, page, vma);
2311 }
2312 set_pte_at(vma->vm_mm, addr, pte, pteval);
2313 return 0;
2314}
2315
2316static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2317 struct page *page, pgprot_t prot, bool mkwrite)
2318{
2319 int retval;
2320 pte_t *pte;
2321 spinlock_t *ptl;
2322
2323 retval = validate_page_before_insert(vma, page);
2324 if (retval)
2325 goto out;
2326 retval = -ENOMEM;
2327 pte = get_locked_pte(mm: vma->vm_mm, addr, ptl: &ptl);
2328 if (!pte)
2329 goto out;
2330 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot,
2331 mkwrite);
2332 pte_unmap_unlock(pte, ptl);
2333out:
2334 return retval;
2335}
2336
2337static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
2338 unsigned long addr, struct page *page, pgprot_t prot)
2339{
2340 int err;
2341
2342 err = validate_page_before_insert(vma, page);
2343 if (err)
2344 return err;
2345 return insert_page_into_pte_locked(vma, pte, addr, page, prot, mkwrite: false);
2346}
2347
2348/* insert_pages() amortizes the cost of spinlock operations
2349 * when inserting pages in a loop.
2350 */
2351static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
2352 struct page **pages, unsigned long *num, pgprot_t prot)
2353{
2354 pmd_t *pmd = NULL;
2355 pte_t *start_pte, *pte;
2356 spinlock_t *pte_lock;
2357 struct mm_struct *const mm = vma->vm_mm;
2358 unsigned long curr_page_idx = 0;
2359 unsigned long remaining_pages_total = *num;
2360 unsigned long pages_to_write_in_pmd;
2361 int ret;
2362more:
2363 ret = -EFAULT;
2364 pmd = walk_to_pmd(mm, addr);
2365 if (!pmd)
2366 goto out;
2367
2368 pages_to_write_in_pmd = min_t(unsigned long,
2369 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
2370
2371 /* Allocate the PTE if necessary; takes PMD lock once only. */
2372 ret = -ENOMEM;
2373 if (pte_alloc(mm, pmd))
2374 goto out;
2375
2376 while (pages_to_write_in_pmd) {
2377 int pte_idx = 0;
2378 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
2379
2380 start_pte = pte_offset_map_lock(mm, pmd, addr, ptlp: &pte_lock);
2381 if (!start_pte) {
2382 ret = -EFAULT;
2383 goto out;
2384 }
2385 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
2386 int err = insert_page_in_batch_locked(vma, pte,
2387 addr, page: pages[curr_page_idx], prot);
2388 if (unlikely(err)) {
2389 pte_unmap_unlock(start_pte, pte_lock);
2390 ret = err;
2391 remaining_pages_total -= pte_idx;
2392 goto out;
2393 }
2394 addr += PAGE_SIZE;
2395 ++curr_page_idx;
2396 }
2397 pte_unmap_unlock(start_pte, pte_lock);
2398 pages_to_write_in_pmd -= batch_size;
2399 remaining_pages_total -= batch_size;
2400 }
2401 if (remaining_pages_total)
2402 goto more;
2403 ret = 0;
2404out:
2405 *num = remaining_pages_total;
2406 return ret;
2407}
2408
2409/**
2410 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
2411 * @vma: user vma to map to
2412 * @addr: target start user address of these pages
2413 * @pages: source kernel pages
2414 * @num: in: number of pages to map. out: number of pages that were *not*
2415 * mapped. (0 means all pages were successfully mapped).
2416 *
2417 * Preferred over vm_insert_page() when inserting multiple pages.
2418 *
2419 * In case of error, we may have mapped a subset of the provided
2420 * pages. It is the caller's responsibility to account for this case.
2421 *
2422 * The same restrictions apply as in vm_insert_page().
2423 */
2424int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
2425 struct page **pages, unsigned long *num)
2426{
2427 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
2428
2429 if (addr < vma->vm_start || end_addr >= vma->vm_end)
2430 return -EFAULT;
2431 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2432 BUG_ON(mmap_read_trylock(vma->vm_mm));
2433 BUG_ON(vma->vm_flags & VM_PFNMAP);
2434 vm_flags_set(vma, VM_MIXEDMAP);
2435 }
2436 /* Defer page refcount checking till we're about to map that page. */
2437 return insert_pages(vma, addr, pages, num, prot: vma->vm_page_prot);
2438}
2439EXPORT_SYMBOL(vm_insert_pages);
2440
2441/**
2442 * vm_insert_page - insert single page into user vma
2443 * @vma: user vma to map to
2444 * @addr: target user address of this page
2445 * @page: source kernel page
2446 *
2447 * This allows drivers to insert individual pages they've allocated
2448 * into a user vma. The zeropage is supported in some VMAs,
2449 * see vm_mixed_zeropage_allowed().
2450 *
2451 * The page has to be a nice clean _individual_ kernel allocation.
2452 * If you allocate a compound page, you need to have marked it as
2453 * such (__GFP_COMP), or manually just split the page up yourself
2454 * (see split_page()).
2455 *
2456 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2457 * took an arbitrary page protection parameter. This doesn't allow
2458 * that. Your vma protection will have to be set up correctly, which
2459 * means that if you want a shared writable mapping, you'd better
2460 * ask for a shared writable mapping!
2461 *
2462 * The page does not need to be reserved.
2463 *
2464 * Usually this function is called from f_op->mmap() handler
2465 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2466 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2467 * function from other places, for example from page-fault handler.
2468 *
2469 * Return: %0 on success, negative error code otherwise.
2470 */
2471int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2472 struct page *page)
2473{
2474 if (addr < vma->vm_start || addr >= vma->vm_end)
2475 return -EFAULT;
2476 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2477 BUG_ON(mmap_read_trylock(vma->vm_mm));
2478 BUG_ON(vma->vm_flags & VM_PFNMAP);
2479 vm_flags_set(vma, VM_MIXEDMAP);
2480 }
2481 return insert_page(vma, addr, page, prot: vma->vm_page_prot, mkwrite: false);
2482}
2483EXPORT_SYMBOL(vm_insert_page);
2484
2485/*
2486 * __vm_map_pages - maps range of kernel pages into user vma
2487 * @vma: user vma to map to
2488 * @pages: pointer to array of source kernel pages
2489 * @num: number of pages in page array
2490 * @offset: user's requested vm_pgoff
2491 *
2492 * This allows drivers to map range of kernel pages into a user vma.
2493 * The zeropage is supported in some VMAs, see
2494 * vm_mixed_zeropage_allowed().
2495 *
2496 * Return: 0 on success and error code otherwise.
2497 */
2498static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2499 unsigned long num, unsigned long offset)
2500{
2501 unsigned long count = vma_pages(vma);
2502 unsigned long uaddr = vma->vm_start;
2503 int ret, i;
2504
2505 /* Fail if the user requested offset is beyond the end of the object */
2506 if (offset >= num)
2507 return -ENXIO;
2508
2509 /* Fail if the user requested size exceeds available object size */
2510 if (count > num - offset)
2511 return -ENXIO;
2512
2513 for (i = 0; i < count; i++) {
2514 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2515 if (ret < 0)
2516 return ret;
2517 uaddr += PAGE_SIZE;
2518 }
2519
2520 return 0;
2521}
2522
2523/**
2524 * vm_map_pages - maps range of kernel pages starts with non zero offset
2525 * @vma: user vma to map to
2526 * @pages: pointer to array of source kernel pages
2527 * @num: number of pages in page array
2528 *
2529 * Maps an object consisting of @num pages, catering for the user's
2530 * requested vm_pgoff
2531 *
2532 * If we fail to insert any page into the vma, the function will return
2533 * immediately leaving any previously inserted pages present. Callers
2534 * from the mmap handler may immediately return the error as their caller
2535 * will destroy the vma, removing any successfully inserted pages. Other
2536 * callers should make their own arrangements for calling unmap_region().
2537 *
2538 * Context: Process context. Called by mmap handlers.
2539 * Return: 0 on success and error code otherwise.
2540 */
2541int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2542 unsigned long num)
2543{
2544 return __vm_map_pages(vma, pages, num, offset: vma->vm_pgoff);
2545}
2546EXPORT_SYMBOL(vm_map_pages);
2547
2548/**
2549 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2550 * @vma: user vma to map to
2551 * @pages: pointer to array of source kernel pages
2552 * @num: number of pages in page array
2553 *
2554 * Similar to vm_map_pages(), except that it explicitly sets the offset
2555 * to 0. This function is intended for the drivers that did not consider
2556 * vm_pgoff.
2557 *
2558 * Context: Process context. Called by mmap handlers.
2559 * Return: 0 on success and error code otherwise.
2560 */
2561int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2562 unsigned long num)
2563{
2564 return __vm_map_pages(vma, pages, num, offset: 0);
2565}
2566EXPORT_SYMBOL(vm_map_pages_zero);
2567
2568static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2569 unsigned long pfn, pgprot_t prot, bool mkwrite)
2570{
2571 struct mm_struct *mm = vma->vm_mm;
2572 pte_t *pte, entry;
2573 spinlock_t *ptl;
2574
2575 pte = get_locked_pte(mm, addr, ptl: &ptl);
2576 if (!pte)
2577 return VM_FAULT_OOM;
2578 entry = ptep_get(ptep: pte);
2579 if (!pte_none(pte: entry)) {
2580 if (mkwrite) {
2581 /*
2582 * For read faults on private mappings the PFN passed
2583 * in may not match the PFN we have mapped if the
2584 * mapped PFN is a writeable COW page. In the mkwrite
2585 * case we are creating a writable PTE for a shared
2586 * mapping and we expect the PFNs to match. If they
2587 * don't match, we are likely racing with block
2588 * allocation and mapping invalidation so just skip the
2589 * update.
2590 */
2591 if (pte_pfn(pte: entry) != pfn) {
2592 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
2593 goto out_unlock;
2594 }
2595 entry = pte_mkyoung(pte: entry);
2596 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
2597 if (ptep_set_access_flags(vma, address: addr, ptep: pte, entry, dirty: 1))
2598 update_mmu_cache(vma, addr, ptep: pte);
2599 }
2600 goto out_unlock;
2601 }
2602
2603 /* Ok, finally just insert the thing.. */
2604 entry = pte_mkspecial(pte: pfn_pte(page_nr: pfn, pgprot: prot));
2605
2606 if (mkwrite) {
2607 entry = pte_mkyoung(pte: entry);
2608 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
2609 }
2610
2611 set_pte_at(mm, addr, pte, entry);
2612 update_mmu_cache(vma, addr, ptep: pte); /* XXX: why not for insert_page? */
2613
2614out_unlock:
2615 pte_unmap_unlock(pte, ptl);
2616 return VM_FAULT_NOPAGE;
2617}
2618
2619/**
2620 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2621 * @vma: user vma to map to
2622 * @addr: target user address of this page
2623 * @pfn: source kernel pfn
2624 * @pgprot: pgprot flags for the inserted page
2625 *
2626 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2627 * to override pgprot on a per-page basis.
2628 *
2629 * This only makes sense for IO mappings, and it makes no sense for
2630 * COW mappings. In general, using multiple vmas is preferable;
2631 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2632 * impractical.
2633 *
2634 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2635 * caching- and encryption bits different than those of @vma->vm_page_prot,
2636 * because the caching- or encryption mode may not be known at mmap() time.
2637 *
2638 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2639 * to set caching and encryption bits for those vmas (except for COW pages).
2640 * This is ensured by core vm only modifying these page table entries using
2641 * functions that don't touch caching- or encryption bits, using pte_modify()
2642 * if needed. (See for example mprotect()).
2643 *
2644 * Also when new page-table entries are created, this is only done using the
2645 * fault() callback, and never using the value of vma->vm_page_prot,
2646 * except for page-table entries that point to anonymous pages as the result
2647 * of COW.
2648 *
2649 * Context: Process context. May allocate using %GFP_KERNEL.
2650 * Return: vm_fault_t value.
2651 */
2652vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2653 unsigned long pfn, pgprot_t pgprot)
2654{
2655 /*
2656 * Technically, architectures with pte_special can avoid all these
2657 * restrictions (same for remap_pfn_range). However we would like
2658 * consistency in testing and feature parity among all, so we should
2659 * try to keep these invariants in place for everybody.
2660 */
2661 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2662 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2663 (VM_PFNMAP|VM_MIXEDMAP));
2664 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2665 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2666
2667 if (addr < vma->vm_start || addr >= vma->vm_end)
2668 return VM_FAULT_SIGBUS;
2669
2670 if (!pfn_modify_allowed(pfn, prot: pgprot))
2671 return VM_FAULT_SIGBUS;
2672
2673 pfnmap_setup_cachemode_pfn(pfn, prot: &pgprot);
2674
2675 return insert_pfn(vma, addr, pfn, prot: pgprot, mkwrite: false);
2676}
2677EXPORT_SYMBOL(vmf_insert_pfn_prot);
2678
2679/**
2680 * vmf_insert_pfn - insert single pfn into user vma
2681 * @vma: user vma to map to
2682 * @addr: target user address of this page
2683 * @pfn: source kernel pfn
2684 *
2685 * Similar to vm_insert_page, this allows drivers to insert individual pages
2686 * they've allocated into a user vma. Same comments apply.
2687 *
2688 * This function should only be called from a vm_ops->fault handler, and
2689 * in that case the handler should return the result of this function.
2690 *
2691 * vma cannot be a COW mapping.
2692 *
2693 * As this is called only for pages that do not currently exist, we
2694 * do not need to flush old virtual caches or the TLB.
2695 *
2696 * Context: Process context. May allocate using %GFP_KERNEL.
2697 * Return: vm_fault_t value.
2698 */
2699vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2700 unsigned long pfn)
2701{
2702 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2703}
2704EXPORT_SYMBOL(vmf_insert_pfn);
2705
2706static bool vm_mixed_ok(struct vm_area_struct *vma, unsigned long pfn,
2707 bool mkwrite)
2708{
2709 if (unlikely(is_zero_pfn(pfn)) &&
2710 (mkwrite || !vm_mixed_zeropage_allowed(vma)))
2711 return false;
2712 /* these checks mirror the abort conditions in vm_normal_page */
2713 if (vma->vm_flags & VM_MIXEDMAP)
2714 return true;
2715 if (is_zero_pfn(pfn))
2716 return true;
2717 return false;
2718}
2719
2720static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2721 unsigned long addr, unsigned long pfn, bool mkwrite)
2722{
2723 pgprot_t pgprot = vma->vm_page_prot;
2724 int err;
2725
2726 if (!vm_mixed_ok(vma, pfn, mkwrite))
2727 return VM_FAULT_SIGBUS;
2728
2729 if (addr < vma->vm_start || addr >= vma->vm_end)
2730 return VM_FAULT_SIGBUS;
2731
2732 pfnmap_setup_cachemode_pfn(pfn, prot: &pgprot);
2733
2734 if (!pfn_modify_allowed(pfn, prot: pgprot))
2735 return VM_FAULT_SIGBUS;
2736
2737 /*
2738 * If we don't have pte special, then we have to use the pfn_valid()
2739 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2740 * refcount the page if pfn_valid is true (hence insert_page rather
2741 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2742 * without pte special, it would there be refcounted as a normal page.
2743 */
2744 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pfn_valid(pfn)) {
2745 struct page *page;
2746
2747 /*
2748 * At this point we are committed to insert_page()
2749 * regardless of whether the caller specified flags that
2750 * result in pfn_t_has_page() == false.
2751 */
2752 page = pfn_to_page(pfn);
2753 err = insert_page(vma, addr, page, prot: pgprot, mkwrite);
2754 } else {
2755 return insert_pfn(vma, addr, pfn, prot: pgprot, mkwrite);
2756 }
2757
2758 if (err == -ENOMEM)
2759 return VM_FAULT_OOM;
2760 if (err < 0 && err != -EBUSY)
2761 return VM_FAULT_SIGBUS;
2762
2763 return VM_FAULT_NOPAGE;
2764}
2765
2766vm_fault_t vmf_insert_page_mkwrite(struct vm_fault *vmf, struct page *page,
2767 bool write)
2768{
2769 pgprot_t pgprot = vmf->vma->vm_page_prot;
2770 unsigned long addr = vmf->address;
2771 int err;
2772
2773 if (addr < vmf->vma->vm_start || addr >= vmf->vma->vm_end)
2774 return VM_FAULT_SIGBUS;
2775
2776 err = insert_page(vma: vmf->vma, addr, page, prot: pgprot, mkwrite: write);
2777 if (err == -ENOMEM)
2778 return VM_FAULT_OOM;
2779 if (err < 0 && err != -EBUSY)
2780 return VM_FAULT_SIGBUS;
2781
2782 return VM_FAULT_NOPAGE;
2783}
2784EXPORT_SYMBOL_GPL(vmf_insert_page_mkwrite);
2785
2786vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2787 unsigned long pfn)
2788{
2789 return __vm_insert_mixed(vma, addr, pfn, mkwrite: false);
2790}
2791EXPORT_SYMBOL(vmf_insert_mixed);
2792
2793/*
2794 * If the insertion of PTE failed because someone else already added a
2795 * different entry in the mean time, we treat that as success as we assume
2796 * the same entry was actually inserted.
2797 */
2798vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2799 unsigned long addr, unsigned long pfn)
2800{
2801 return __vm_insert_mixed(vma, addr, pfn, mkwrite: true);
2802}
2803
2804/*
2805 * maps a range of physical memory into the requested pages. the old
2806 * mappings are removed. any references to nonexistent pages results
2807 * in null mappings (currently treated as "copy-on-access")
2808 */
2809static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2810 unsigned long addr, unsigned long end,
2811 unsigned long pfn, pgprot_t prot)
2812{
2813 pte_t *pte, *mapped_pte;
2814 spinlock_t *ptl;
2815 int err = 0;
2816
2817 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2818 if (!pte)
2819 return -ENOMEM;
2820 arch_enter_lazy_mmu_mode();
2821 do {
2822 BUG_ON(!pte_none(ptep_get(pte)));
2823 if (!pfn_modify_allowed(pfn, prot)) {
2824 err = -EACCES;
2825 break;
2826 }
2827 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2828 pfn++;
2829 } while (pte++, addr += PAGE_SIZE, addr != end);
2830 arch_leave_lazy_mmu_mode();
2831 pte_unmap_unlock(mapped_pte, ptl);
2832 return err;
2833}
2834
2835static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2836 unsigned long addr, unsigned long end,
2837 unsigned long pfn, pgprot_t prot)
2838{
2839 pmd_t *pmd;
2840 unsigned long next;
2841 int err;
2842
2843 pfn -= addr >> PAGE_SHIFT;
2844 pmd = pmd_alloc(mm, pud, address: addr);
2845 if (!pmd)
2846 return -ENOMEM;
2847 VM_BUG_ON(pmd_trans_huge(*pmd));
2848 do {
2849 next = pmd_addr_end(addr, end);
2850 err = remap_pte_range(mm, pmd, addr, end: next,
2851 pfn: pfn + (addr >> PAGE_SHIFT), prot);
2852 if (err)
2853 return err;
2854 } while (pmd++, addr = next, addr != end);
2855 return 0;
2856}
2857
2858static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2859 unsigned long addr, unsigned long end,
2860 unsigned long pfn, pgprot_t prot)
2861{
2862 pud_t *pud;
2863 unsigned long next;
2864 int err;
2865
2866 pfn -= addr >> PAGE_SHIFT;
2867 pud = pud_alloc(mm, p4d, address: addr);
2868 if (!pud)
2869 return -ENOMEM;
2870 do {
2871 next = pud_addr_end(addr, end);
2872 err = remap_pmd_range(mm, pud, addr, end: next,
2873 pfn: pfn + (addr >> PAGE_SHIFT), prot);
2874 if (err)
2875 return err;
2876 } while (pud++, addr = next, addr != end);
2877 return 0;
2878}
2879
2880static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2881 unsigned long addr, unsigned long end,
2882 unsigned long pfn, pgprot_t prot)
2883{
2884 p4d_t *p4d;
2885 unsigned long next;
2886 int err;
2887
2888 pfn -= addr >> PAGE_SHIFT;
2889 p4d = p4d_alloc(mm, pgd, address: addr);
2890 if (!p4d)
2891 return -ENOMEM;
2892 do {
2893 next = p4d_addr_end(addr, end);
2894 err = remap_pud_range(mm, p4d, addr, end: next,
2895 pfn: pfn + (addr >> PAGE_SHIFT), prot);
2896 if (err)
2897 return err;
2898 } while (p4d++, addr = next, addr != end);
2899 return 0;
2900}
2901
2902static int remap_pfn_range_internal(struct vm_area_struct *vma, unsigned long addr,
2903 unsigned long pfn, unsigned long size, pgprot_t prot)
2904{
2905 pgd_t *pgd;
2906 unsigned long next;
2907 unsigned long end = addr + PAGE_ALIGN(size);
2908 struct mm_struct *mm = vma->vm_mm;
2909 int err;
2910
2911 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2912 return -EINVAL;
2913
2914 /*
2915 * Physically remapped pages are special. Tell the
2916 * rest of the world about it:
2917 * VM_IO tells people not to look at these pages
2918 * (accesses can have side effects).
2919 * VM_PFNMAP tells the core MM that the base pages are just
2920 * raw PFN mappings, and do not have a "struct page" associated
2921 * with them.
2922 * VM_DONTEXPAND
2923 * Disable vma merging and expanding with mremap().
2924 * VM_DONTDUMP
2925 * Omit vma from core dump, even when VM_IO turned off.
2926 *
2927 * There's a horrible special case to handle copy-on-write
2928 * behaviour that some programs depend on. We mark the "original"
2929 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2930 * See vm_normal_page() for details.
2931 */
2932 if (is_cow_mapping(flags: vma->vm_flags)) {
2933 if (addr != vma->vm_start || end != vma->vm_end)
2934 return -EINVAL;
2935 vma->vm_pgoff = pfn;
2936 }
2937
2938 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
2939
2940 BUG_ON(addr >= end);
2941 pfn -= addr >> PAGE_SHIFT;
2942 pgd = pgd_offset(mm, addr);
2943 flush_cache_range(vma, start: addr, end);
2944 do {
2945 next = pgd_addr_end(addr, end);
2946 err = remap_p4d_range(mm, pgd, addr, end: next,
2947 pfn: pfn + (addr >> PAGE_SHIFT), prot);
2948 if (err)
2949 return err;
2950 } while (pgd++, addr = next, addr != end);
2951
2952 return 0;
2953}
2954
2955/*
2956 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2957 * must have pre-validated the caching bits of the pgprot_t.
2958 */
2959int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2960 unsigned long pfn, unsigned long size, pgprot_t prot)
2961{
2962 int error = remap_pfn_range_internal(vma, addr, pfn, size, prot);
2963
2964 if (!error)
2965 return 0;
2966
2967 /*
2968 * A partial pfn range mapping is dangerous: it does not
2969 * maintain page reference counts, and callers may free
2970 * pages due to the error. So zap it early.
2971 */
2972 zap_page_range_single(vma, address: addr, size, NULL);
2973 return error;
2974}
2975
2976#ifdef __HAVE_PFNMAP_TRACKING
2977static inline struct pfnmap_track_ctx *pfnmap_track_ctx_alloc(unsigned long pfn,
2978 unsigned long size, pgprot_t *prot)
2979{
2980 struct pfnmap_track_ctx *ctx;
2981
2982 if (pfnmap_track(pfn, size, prot))
2983 return ERR_PTR(error: -EINVAL);
2984
2985 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
2986 if (unlikely(!ctx)) {
2987 pfnmap_untrack(pfn, size);
2988 return ERR_PTR(error: -ENOMEM);
2989 }
2990
2991 ctx->pfn = pfn;
2992 ctx->size = size;
2993 kref_init(kref: &ctx->kref);
2994 return ctx;
2995}
2996
2997void pfnmap_track_ctx_release(struct kref *ref)
2998{
2999 struct pfnmap_track_ctx *ctx = container_of(ref, struct pfnmap_track_ctx, kref);
3000
3001 pfnmap_untrack(pfn: ctx->pfn, size: ctx->size);
3002 kfree(objp: ctx);
3003}
3004#endif /* __HAVE_PFNMAP_TRACKING */
3005
3006/**
3007 * remap_pfn_range - remap kernel memory to userspace
3008 * @vma: user vma to map to
3009 * @addr: target page aligned user address to start at
3010 * @pfn: page frame number of kernel physical memory address
3011 * @size: size of mapping area
3012 * @prot: page protection flags for this mapping
3013 *
3014 * Note: this is only safe if the mm semaphore is held when called.
3015 *
3016 * Return: %0 on success, negative error code otherwise.
3017 */
3018#ifdef __HAVE_PFNMAP_TRACKING
3019int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
3020 unsigned long pfn, unsigned long size, pgprot_t prot)
3021{
3022 struct pfnmap_track_ctx *ctx = NULL;
3023 int err;
3024
3025 size = PAGE_ALIGN(size);
3026
3027 /*
3028 * If we cover the full VMA, we'll perform actual tracking, and
3029 * remember to untrack when the last reference to our tracking
3030 * context from a VMA goes away. We'll keep tracking the whole pfn
3031 * range even during VMA splits and partial unmapping.
3032 *
3033 * If we only cover parts of the VMA, we'll only setup the cachemode
3034 * in the pgprot for the pfn range.
3035 */
3036 if (addr == vma->vm_start && addr + size == vma->vm_end) {
3037 if (vma->pfnmap_track_ctx)
3038 return -EINVAL;
3039 ctx = pfnmap_track_ctx_alloc(pfn, size, prot: &prot);
3040 if (IS_ERR(ptr: ctx))
3041 return PTR_ERR(ptr: ctx);
3042 } else if (pfnmap_setup_cachemode(pfn, size, prot: &prot)) {
3043 return -EINVAL;
3044 }
3045
3046 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
3047 if (ctx) {
3048 if (err)
3049 kref_put(kref: &ctx->kref, release: pfnmap_track_ctx_release);
3050 else
3051 vma->pfnmap_track_ctx = ctx;
3052 }
3053 return err;
3054}
3055
3056#else
3057int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
3058 unsigned long pfn, unsigned long size, pgprot_t prot)
3059{
3060 return remap_pfn_range_notrack(vma, addr, pfn, size, prot);
3061}
3062#endif
3063EXPORT_SYMBOL(remap_pfn_range);
3064
3065/**
3066 * vm_iomap_memory - remap memory to userspace
3067 * @vma: user vma to map to
3068 * @start: start of the physical memory to be mapped
3069 * @len: size of area
3070 *
3071 * This is a simplified io_remap_pfn_range() for common driver use. The
3072 * driver just needs to give us the physical memory range to be mapped,
3073 * we'll figure out the rest from the vma information.
3074 *
3075 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
3076 * whatever write-combining details or similar.
3077 *
3078 * Return: %0 on success, negative error code otherwise.
3079 */
3080int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
3081{
3082 unsigned long vm_len, pfn, pages;
3083
3084 /* Check that the physical memory area passed in looks valid */
3085 if (start + len < start)
3086 return -EINVAL;
3087 /*
3088 * You *really* shouldn't map things that aren't page-aligned,
3089 * but we've historically allowed it because IO memory might
3090 * just have smaller alignment.
3091 */
3092 len += start & ~PAGE_MASK;
3093 pfn = start >> PAGE_SHIFT;
3094 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
3095 if (pfn + pages < pfn)
3096 return -EINVAL;
3097
3098 /* We start the mapping 'vm_pgoff' pages into the area */
3099 if (vma->vm_pgoff > pages)
3100 return -EINVAL;
3101 pfn += vma->vm_pgoff;
3102 pages -= vma->vm_pgoff;
3103
3104 /* Can we fit all of the mapping? */
3105 vm_len = vma->vm_end - vma->vm_start;
3106 if (vm_len >> PAGE_SHIFT > pages)
3107 return -EINVAL;
3108
3109 /* Ok, let it rip */
3110 return io_remap_pfn_range(vma, addr: vma->vm_start, pfn, size: vm_len, prot: vma->vm_page_prot);
3111}
3112EXPORT_SYMBOL(vm_iomap_memory);
3113
3114static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
3115 unsigned long addr, unsigned long end,
3116 pte_fn_t fn, void *data, bool create,
3117 pgtbl_mod_mask *mask)
3118{
3119 pte_t *pte, *mapped_pte;
3120 int err = 0;
3121 spinlock_t *ptl;
3122
3123 if (create) {
3124 mapped_pte = pte = (mm == &init_mm) ?
3125 pte_alloc_kernel_track(pmd, addr, mask) :
3126 pte_alloc_map_lock(mm, pmd, addr, &ptl);
3127 if (!pte)
3128 return -ENOMEM;
3129 } else {
3130 mapped_pte = pte = (mm == &init_mm) ?
3131 pte_offset_kernel(pmd, address: addr) :
3132 pte_offset_map_lock(mm, pmd, addr, ptlp: &ptl);
3133 if (!pte)
3134 return -EINVAL;
3135 }
3136
3137 arch_enter_lazy_mmu_mode();
3138
3139 if (fn) {
3140 do {
3141 if (create || !pte_none(pte: ptep_get(ptep: pte))) {
3142 err = fn(pte, addr, data);
3143 if (err)
3144 break;
3145 }
3146 } while (pte++, addr += PAGE_SIZE, addr != end);
3147 }
3148 *mask |= PGTBL_PTE_MODIFIED;
3149
3150 arch_leave_lazy_mmu_mode();
3151
3152 if (mm != &init_mm)
3153 pte_unmap_unlock(mapped_pte, ptl);
3154 return err;
3155}
3156
3157static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
3158 unsigned long addr, unsigned long end,
3159 pte_fn_t fn, void *data, bool create,
3160 pgtbl_mod_mask *mask)
3161{
3162 pmd_t *pmd;
3163 unsigned long next;
3164 int err = 0;
3165
3166 BUG_ON(pud_leaf(*pud));
3167
3168 if (create) {
3169 pmd = pmd_alloc_track(mm, pud, address: addr, mod_mask: mask);
3170 if (!pmd)
3171 return -ENOMEM;
3172 } else {
3173 pmd = pmd_offset(pud, address: addr);
3174 }
3175 do {
3176 next = pmd_addr_end(addr, end);
3177 if (pmd_none(pmd: *pmd) && !create)
3178 continue;
3179 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
3180 return -EINVAL;
3181 if (!pmd_none(pmd: *pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
3182 if (!create)
3183 continue;
3184 pmd_clear_bad(pmd);
3185 }
3186 err = apply_to_pte_range(mm, pmd, addr, end: next,
3187 fn, data, create, mask);
3188 if (err)
3189 break;
3190 } while (pmd++, addr = next, addr != end);
3191
3192 return err;
3193}
3194
3195static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
3196 unsigned long addr, unsigned long end,
3197 pte_fn_t fn, void *data, bool create,
3198 pgtbl_mod_mask *mask)
3199{
3200 pud_t *pud;
3201 unsigned long next;
3202 int err = 0;
3203
3204 if (create) {
3205 pud = pud_alloc_track(mm, p4d, address: addr, mod_mask: mask);
3206 if (!pud)
3207 return -ENOMEM;
3208 } else {
3209 pud = pud_offset(p4d, address: addr);
3210 }
3211 do {
3212 next = pud_addr_end(addr, end);
3213 if (pud_none(pud: *pud) && !create)
3214 continue;
3215 if (WARN_ON_ONCE(pud_leaf(*pud)))
3216 return -EINVAL;
3217 if (!pud_none(pud: *pud) && WARN_ON_ONCE(pud_bad(*pud))) {
3218 if (!create)
3219 continue;
3220 pud_clear_bad(pud);
3221 }
3222 err = apply_to_pmd_range(mm, pud, addr, end: next,
3223 fn, data, create, mask);
3224 if (err)
3225 break;
3226 } while (pud++, addr = next, addr != end);
3227
3228 return err;
3229}
3230
3231static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
3232 unsigned long addr, unsigned long end,
3233 pte_fn_t fn, void *data, bool create,
3234 pgtbl_mod_mask *mask)
3235{
3236 p4d_t *p4d;
3237 unsigned long next;
3238 int err = 0;
3239
3240 if (create) {
3241 p4d = p4d_alloc_track(mm, pgd, address: addr, mod_mask: mask);
3242 if (!p4d)
3243 return -ENOMEM;
3244 } else {
3245 p4d = p4d_offset(pgd, address: addr);
3246 }
3247 do {
3248 next = p4d_addr_end(addr, end);
3249 if (p4d_none(p4d: *p4d) && !create)
3250 continue;
3251 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
3252 return -EINVAL;
3253 if (!p4d_none(p4d: *p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
3254 if (!create)
3255 continue;
3256 p4d_clear_bad(p4d);
3257 }
3258 err = apply_to_pud_range(mm, p4d, addr, end: next,
3259 fn, data, create, mask);
3260 if (err)
3261 break;
3262 } while (p4d++, addr = next, addr != end);
3263
3264 return err;
3265}
3266
3267static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
3268 unsigned long size, pte_fn_t fn,
3269 void *data, bool create)
3270{
3271 pgd_t *pgd;
3272 unsigned long start = addr, next;
3273 unsigned long end = addr + size;
3274 pgtbl_mod_mask mask = 0;
3275 int err = 0;
3276
3277 if (WARN_ON(addr >= end))
3278 return -EINVAL;
3279
3280 pgd = pgd_offset(mm, addr);
3281 do {
3282 next = pgd_addr_end(addr, end);
3283 if (pgd_none(pgd: *pgd) && !create)
3284 continue;
3285 if (WARN_ON_ONCE(pgd_leaf(*pgd))) {
3286 err = -EINVAL;
3287 break;
3288 }
3289 if (!pgd_none(pgd: *pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
3290 if (!create)
3291 continue;
3292 pgd_clear_bad(pgd);
3293 }
3294 err = apply_to_p4d_range(mm, pgd, addr, end: next,
3295 fn, data, create, mask: &mask);
3296 if (err)
3297 break;
3298 } while (pgd++, addr = next, addr != end);
3299
3300 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
3301 arch_sync_kernel_mappings(start, end: start + size);
3302
3303 return err;
3304}
3305
3306/*
3307 * Scan a region of virtual memory, filling in page tables as necessary
3308 * and calling a provided function on each leaf page table.
3309 */
3310int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
3311 unsigned long size, pte_fn_t fn, void *data)
3312{
3313 return __apply_to_page_range(mm, addr, size, fn, data, create: true);
3314}
3315EXPORT_SYMBOL_GPL(apply_to_page_range);
3316
3317/*
3318 * Scan a region of virtual memory, calling a provided function on
3319 * each leaf page table where it exists.
3320 *
3321 * Unlike apply_to_page_range, this does _not_ fill in page tables
3322 * where they are absent.
3323 */
3324int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
3325 unsigned long size, pte_fn_t fn, void *data)
3326{
3327 return __apply_to_page_range(mm, addr, size, fn, data, create: false);
3328}
3329
3330/*
3331 * handle_pte_fault chooses page fault handler according to an entry which was
3332 * read non-atomically. Before making any commitment, on those architectures
3333 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
3334 * parts, do_swap_page must check under lock before unmapping the pte and
3335 * proceeding (but do_wp_page is only called after already making such a check;
3336 * and do_anonymous_page can safely check later on).
3337 */
3338static inline int pte_unmap_same(struct vm_fault *vmf)
3339{
3340 int same = 1;
3341#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
3342 if (sizeof(pte_t) > sizeof(unsigned long)) {
3343 spin_lock(lock: vmf->ptl);
3344 same = pte_same(a: ptep_get(ptep: vmf->pte), b: vmf->orig_pte);
3345 spin_unlock(lock: vmf->ptl);
3346 }
3347#endif
3348 pte_unmap(pte: vmf->pte);
3349 vmf->pte = NULL;
3350 return same;
3351}
3352
3353/*
3354 * Return:
3355 * 0: copied succeeded
3356 * -EHWPOISON: copy failed due to hwpoison in source page
3357 * -EAGAIN: copied failed (some other reason)
3358 */
3359static inline int __wp_page_copy_user(struct page *dst, struct page *src,
3360 struct vm_fault *vmf)
3361{
3362 int ret;
3363 void *kaddr;
3364 void __user *uaddr;
3365 struct vm_area_struct *vma = vmf->vma;
3366 struct mm_struct *mm = vma->vm_mm;
3367 unsigned long addr = vmf->address;
3368
3369 if (likely(src)) {
3370 if (copy_mc_user_highpage(to: dst, from: src, vaddr: addr, vma))
3371 return -EHWPOISON;
3372 return 0;
3373 }
3374
3375 /*
3376 * If the source page was a PFN mapping, we don't have
3377 * a "struct page" for it. We do a best-effort copy by
3378 * just copying from the original user address. If that
3379 * fails, we just zero-fill it. Live with it.
3380 */
3381 kaddr = kmap_local_page(page: dst);
3382 pagefault_disable();
3383 uaddr = (void __user *)(addr & PAGE_MASK);
3384
3385 /*
3386 * On architectures with software "accessed" bits, we would
3387 * take a double page fault, so mark it accessed here.
3388 */
3389 vmf->pte = NULL;
3390 if (!arch_has_hw_pte_young() && !pte_young(pte: vmf->orig_pte)) {
3391 pte_t entry;
3392
3393 vmf->pte = pte_offset_map_lock(mm, pmd: vmf->pmd, addr, ptlp: &vmf->ptl);
3394 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3395 /*
3396 * Other thread has already handled the fault
3397 * and update local tlb only
3398 */
3399 if (vmf->pte)
3400 update_mmu_tlb(vma, address: addr, ptep: vmf->pte);
3401 ret = -EAGAIN;
3402 goto pte_unlock;
3403 }
3404
3405 entry = pte_mkyoung(pte: vmf->orig_pte);
3406 if (ptep_set_access_flags(vma, address: addr, ptep: vmf->pte, entry, dirty: 0))
3407 update_mmu_cache_range(vmf, vma, addr, ptep: vmf->pte, nr: 1);
3408 }
3409
3410 /*
3411 * This really shouldn't fail, because the page is there
3412 * in the page tables. But it might just be unreadable,
3413 * in which case we just give up and fill the result with
3414 * zeroes.
3415 */
3416 if (__copy_from_user_inatomic(to: kaddr, from: uaddr, PAGE_SIZE)) {
3417 if (vmf->pte)
3418 goto warn;
3419
3420 /* Re-validate under PTL if the page is still mapped */
3421 vmf->pte = pte_offset_map_lock(mm, pmd: vmf->pmd, addr, ptlp: &vmf->ptl);
3422 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3423 /* The PTE changed under us, update local tlb */
3424 if (vmf->pte)
3425 update_mmu_tlb(vma, address: addr, ptep: vmf->pte);
3426 ret = -EAGAIN;
3427 goto pte_unlock;
3428 }
3429
3430 /*
3431 * The same page can be mapped back since last copy attempt.
3432 * Try to copy again under PTL.
3433 */
3434 if (__copy_from_user_inatomic(to: kaddr, from: uaddr, PAGE_SIZE)) {
3435 /*
3436 * Give a warn in case there can be some obscure
3437 * use-case
3438 */
3439warn:
3440 WARN_ON_ONCE(1);
3441 clear_page(page: kaddr);
3442 }
3443 }
3444
3445 ret = 0;
3446
3447pte_unlock:
3448 if (vmf->pte)
3449 pte_unmap_unlock(vmf->pte, vmf->ptl);
3450 pagefault_enable();
3451 kunmap_local(kaddr);
3452 flush_dcache_page(page: dst);
3453
3454 return ret;
3455}
3456
3457static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
3458{
3459 struct file *vm_file = vma->vm_file;
3460
3461 if (vm_file)
3462 return mapping_gfp_mask(mapping: vm_file->f_mapping) | __GFP_FS | __GFP_IO;
3463
3464 /*
3465 * Special mappings (e.g. VDSO) do not have any file so fake
3466 * a default GFP_KERNEL for them.
3467 */
3468 return GFP_KERNEL;
3469}
3470
3471/*
3472 * Notify the address space that the page is about to become writable so that
3473 * it can prohibit this or wait for the page to get into an appropriate state.
3474 *
3475 * We do this without the lock held, so that it can sleep if it needs to.
3476 */
3477static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio)
3478{
3479 vm_fault_t ret;
3480 unsigned int old_flags = vmf->flags;
3481
3482 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3483
3484 if (vmf->vma->vm_file &&
3485 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
3486 return VM_FAULT_SIGBUS;
3487
3488 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
3489 /* Restore original flags so that caller is not surprised */
3490 vmf->flags = old_flags;
3491 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
3492 return ret;
3493 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
3494 folio_lock(folio);
3495 if (!folio->mapping) {
3496 folio_unlock(folio);
3497 return 0; /* retry */
3498 }
3499 ret |= VM_FAULT_LOCKED;
3500 } else
3501 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3502 return ret;
3503}
3504
3505/*
3506 * Handle dirtying of a page in shared file mapping on a write fault.
3507 *
3508 * The function expects the page to be locked and unlocks it.
3509 */
3510static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
3511{
3512 struct vm_area_struct *vma = vmf->vma;
3513 struct address_space *mapping;
3514 struct folio *folio = page_folio(vmf->page);
3515 bool dirtied;
3516 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
3517
3518 dirtied = folio_mark_dirty(folio);
3519 VM_BUG_ON_FOLIO(folio_test_anon(folio), folio);
3520 /*
3521 * Take a local copy of the address_space - folio.mapping may be zeroed
3522 * by truncate after folio_unlock(). The address_space itself remains
3523 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
3524 * release semantics to prevent the compiler from undoing this copying.
3525 */
3526 mapping = folio_raw_mapping(folio);
3527 folio_unlock(folio);
3528
3529 if (!page_mkwrite)
3530 file_update_time(file: vma->vm_file);
3531
3532 /*
3533 * Throttle page dirtying rate down to writeback speed.
3534 *
3535 * mapping may be NULL here because some device drivers do not
3536 * set page.mapping but still dirty their pages
3537 *
3538 * Drop the mmap_lock before waiting on IO, if we can. The file
3539 * is pinning the mapping, as per above.
3540 */
3541 if ((dirtied || page_mkwrite) && mapping) {
3542 struct file *fpin;
3543
3544 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
3545 balance_dirty_pages_ratelimited(mapping);
3546 if (fpin) {
3547 fput(fpin);
3548 return VM_FAULT_COMPLETED;
3549 }
3550 }
3551
3552 return 0;
3553}
3554
3555/*
3556 * Handle write page faults for pages that can be reused in the current vma
3557 *
3558 * This can happen either due to the mapping being with the VM_SHARED flag,
3559 * or due to us being the last reference standing to the page. In either
3560 * case, all we need to do here is to mark the page as writable and update
3561 * any related book-keeping.
3562 */
3563static inline void wp_page_reuse(struct vm_fault *vmf, struct folio *folio)
3564 __releases(vmf->ptl)
3565{
3566 struct vm_area_struct *vma = vmf->vma;
3567 pte_t entry;
3568
3569 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3570 VM_WARN_ON(is_zero_pfn(pte_pfn(vmf->orig_pte)));
3571
3572 if (folio) {
3573 VM_BUG_ON(folio_test_anon(folio) &&
3574 !PageAnonExclusive(vmf->page));
3575 /*
3576 * Clear the folio's cpupid information as the existing
3577 * information potentially belongs to a now completely
3578 * unrelated process.
3579 */
3580 folio_xchg_last_cpupid(folio, cpupid: (1 << LAST_CPUPID_SHIFT) - 1);
3581 }
3582
3583 flush_cache_page(vma, vmaddr: vmf->address, pfn: pte_pfn(pte: vmf->orig_pte));
3584 entry = pte_mkyoung(pte: vmf->orig_pte);
3585 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
3586 if (ptep_set_access_flags(vma, address: vmf->address, ptep: vmf->pte, entry, dirty: 1))
3587 update_mmu_cache_range(vmf, vma, addr: vmf->address, ptep: vmf->pte, nr: 1);
3588 pte_unmap_unlock(vmf->pte, vmf->ptl);
3589 count_vm_event(item: PGREUSE);
3590}
3591
3592/*
3593 * We could add a bitflag somewhere, but for now, we know that all
3594 * vm_ops that have a ->map_pages have been audited and don't need
3595 * the mmap_lock to be held.
3596 */
3597static inline vm_fault_t vmf_can_call_fault(const struct vm_fault *vmf)
3598{
3599 struct vm_area_struct *vma = vmf->vma;
3600
3601 if (vma->vm_ops->map_pages || !(vmf->flags & FAULT_FLAG_VMA_LOCK))
3602 return 0;
3603 vma_end_read(vma);
3604 return VM_FAULT_RETRY;
3605}
3606
3607/**
3608 * __vmf_anon_prepare - Prepare to handle an anonymous fault.
3609 * @vmf: The vm_fault descriptor passed from the fault handler.
3610 *
3611 * When preparing to insert an anonymous page into a VMA from a
3612 * fault handler, call this function rather than anon_vma_prepare().
3613 * If this vma does not already have an associated anon_vma and we are
3614 * only protected by the per-VMA lock, the caller must retry with the
3615 * mmap_lock held. __anon_vma_prepare() will look at adjacent VMAs to
3616 * determine if this VMA can share its anon_vma, and that's not safe to
3617 * do with only the per-VMA lock held for this VMA.
3618 *
3619 * Return: 0 if fault handling can proceed. Any other value should be
3620 * returned to the caller.
3621 */
3622vm_fault_t __vmf_anon_prepare(struct vm_fault *vmf)
3623{
3624 struct vm_area_struct *vma = vmf->vma;
3625 vm_fault_t ret = 0;
3626
3627 if (likely(vma->anon_vma))
3628 return 0;
3629 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3630 if (!mmap_read_trylock(mm: vma->vm_mm))
3631 return VM_FAULT_RETRY;
3632 }
3633 if (__anon_vma_prepare(vma))
3634 ret = VM_FAULT_OOM;
3635 if (vmf->flags & FAULT_FLAG_VMA_LOCK)
3636 mmap_read_unlock(mm: vma->vm_mm);
3637 return ret;
3638}
3639
3640/*
3641 * Handle the case of a page which we actually need to copy to a new page,
3642 * either due to COW or unsharing.
3643 *
3644 * Called with mmap_lock locked and the old page referenced, but
3645 * without the ptl held.
3646 *
3647 * High level logic flow:
3648 *
3649 * - Allocate a page, copy the content of the old page to the new one.
3650 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3651 * - Take the PTL. If the pte changed, bail out and release the allocated page
3652 * - If the pte is still the way we remember it, update the page table and all
3653 * relevant references. This includes dropping the reference the page-table
3654 * held to the old page, as well as updating the rmap.
3655 * - In any case, unlock the PTL and drop the reference we took to the old page.
3656 */
3657static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3658{
3659 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3660 struct vm_area_struct *vma = vmf->vma;
3661 struct mm_struct *mm = vma->vm_mm;
3662 struct folio *old_folio = NULL;
3663 struct folio *new_folio = NULL;
3664 pte_t entry;
3665 int page_copied = 0;
3666 struct mmu_notifier_range range;
3667 vm_fault_t ret;
3668 bool pfn_is_zero;
3669
3670 delayacct_wpcopy_start();
3671
3672 if (vmf->page)
3673 old_folio = page_folio(vmf->page);
3674 ret = vmf_anon_prepare(vmf);
3675 if (unlikely(ret))
3676 goto out;
3677
3678 pfn_is_zero = is_zero_pfn(pfn: pte_pfn(pte: vmf->orig_pte));
3679 new_folio = folio_prealloc(src_mm: mm, vma, addr: vmf->address, need_zero: pfn_is_zero);
3680 if (!new_folio)
3681 goto oom;
3682
3683 if (!pfn_is_zero) {
3684 int err;
3685
3686 err = __wp_page_copy_user(dst: &new_folio->page, src: vmf->page, vmf);
3687 if (err) {
3688 /*
3689 * COW failed, if the fault was solved by other,
3690 * it's fine. If not, userspace would re-fault on
3691 * the same address and we will handle the fault
3692 * from the second attempt.
3693 * The -EHWPOISON case will not be retried.
3694 */
3695 folio_put(folio: new_folio);
3696 if (old_folio)
3697 folio_put(folio: old_folio);
3698
3699 delayacct_wpcopy_end();
3700 return err == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3701 }
3702 kmsan_copy_page_meta(dst: &new_folio->page, src: vmf->page);
3703 }
3704
3705 __folio_mark_uptodate(folio: new_folio);
3706
3707 mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_CLEAR, flags: 0, mm,
3708 start: vmf->address & PAGE_MASK,
3709 end: (vmf->address & PAGE_MASK) + PAGE_SIZE);
3710 mmu_notifier_invalidate_range_start(range: &range);
3711
3712 /*
3713 * Re-check the pte - we dropped the lock
3714 */
3715 vmf->pte = pte_offset_map_lock(mm, pmd: vmf->pmd, addr: vmf->address, ptlp: &vmf->ptl);
3716 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3717 if (old_folio) {
3718 if (!folio_test_anon(folio: old_folio)) {
3719 dec_mm_counter(mm, member: mm_counter_file(folio: old_folio));
3720 inc_mm_counter(mm, member: MM_ANONPAGES);
3721 }
3722 } else {
3723 ksm_might_unmap_zero_page(mm, pte: vmf->orig_pte);
3724 inc_mm_counter(mm, member: MM_ANONPAGES);
3725 }
3726 flush_cache_page(vma, vmaddr: vmf->address, pfn: pte_pfn(pte: vmf->orig_pte));
3727 entry = folio_mk_pte(folio: new_folio, pgprot: vma->vm_page_prot);
3728 entry = pte_sw_mkyoung(pte: entry);
3729 if (unlikely(unshare)) {
3730 if (pte_soft_dirty(pte: vmf->orig_pte))
3731 entry = pte_mksoft_dirty(pte: entry);
3732 if (pte_uffd_wp(pte: vmf->orig_pte))
3733 entry = pte_mkuffd_wp(pte: entry);
3734 } else {
3735 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
3736 }
3737
3738 /*
3739 * Clear the pte entry and flush it first, before updating the
3740 * pte with the new entry, to keep TLBs on different CPUs in
3741 * sync. This code used to set the new PTE then flush TLBs, but
3742 * that left a window where the new PTE could be loaded into
3743 * some TLBs while the old PTE remains in others.
3744 */
3745 ptep_clear_flush(vma, address: vmf->address, ptep: vmf->pte);
3746 folio_add_new_anon_rmap(new_folio, vma, address: vmf->address, RMAP_EXCLUSIVE);
3747 folio_add_lru_vma(new_folio, vma);
3748 BUG_ON(unshare && pte_write(entry));
3749 set_pte_at(mm, vmf->address, vmf->pte, entry);
3750 update_mmu_cache_range(vmf, vma, addr: vmf->address, ptep: vmf->pte, nr: 1);
3751 if (old_folio) {
3752 /*
3753 * Only after switching the pte to the new page may
3754 * we remove the mapcount here. Otherwise another
3755 * process may come and find the rmap count decremented
3756 * before the pte is switched to the new page, and
3757 * "reuse" the old page writing into it while our pte
3758 * here still points into it and can be read by other
3759 * threads.
3760 *
3761 * The critical issue is to order this
3762 * folio_remove_rmap_pte() with the ptp_clear_flush
3763 * above. Those stores are ordered by (if nothing else,)
3764 * the barrier present in the atomic_add_negative
3765 * in folio_remove_rmap_pte();
3766 *
3767 * Then the TLB flush in ptep_clear_flush ensures that
3768 * no process can access the old page before the
3769 * decremented mapcount is visible. And the old page
3770 * cannot be reused until after the decremented
3771 * mapcount is visible. So transitively, TLBs to
3772 * old page will be flushed before it can be reused.
3773 */
3774 folio_remove_rmap_pte(old_folio, vmf->page, vma);
3775 }
3776
3777 /* Free the old page.. */
3778 new_folio = old_folio;
3779 page_copied = 1;
3780 pte_unmap_unlock(vmf->pte, vmf->ptl);
3781 } else if (vmf->pte) {
3782 update_mmu_tlb(vma, address: vmf->address, ptep: vmf->pte);
3783 pte_unmap_unlock(vmf->pte, vmf->ptl);
3784 }
3785
3786 mmu_notifier_invalidate_range_end(range: &range);
3787
3788 if (new_folio)
3789 folio_put(folio: new_folio);
3790 if (old_folio) {
3791 if (page_copied)
3792 free_swap_cache(folio: old_folio);
3793 folio_put(folio: old_folio);
3794 }
3795
3796 delayacct_wpcopy_end();
3797 return 0;
3798oom:
3799 ret = VM_FAULT_OOM;
3800out:
3801 if (old_folio)
3802 folio_put(folio: old_folio);
3803
3804 delayacct_wpcopy_end();
3805 return ret;
3806}
3807
3808/**
3809 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3810 * writeable once the page is prepared
3811 *
3812 * @vmf: structure describing the fault
3813 * @folio: the folio of vmf->page
3814 *
3815 * This function handles all that is needed to finish a write page fault in a
3816 * shared mapping due to PTE being read-only once the mapped page is prepared.
3817 * It handles locking of PTE and modifying it.
3818 *
3819 * The function expects the page to be locked or other protection against
3820 * concurrent faults / writeback (such as DAX radix tree locks).
3821 *
3822 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3823 * we acquired PTE lock.
3824 */
3825static vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf, struct folio *folio)
3826{
3827 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3828 vmf->pte = pte_offset_map_lock(mm: vmf->vma->vm_mm, pmd: vmf->pmd, addr: vmf->address,
3829 ptlp: &vmf->ptl);
3830 if (!vmf->pte)
3831 return VM_FAULT_NOPAGE;
3832 /*
3833 * We might have raced with another page fault while we released the
3834 * pte_offset_map_lock.
3835 */
3836 if (!pte_same(a: ptep_get(ptep: vmf->pte), b: vmf->orig_pte)) {
3837 update_mmu_tlb(vma: vmf->vma, address: vmf->address, ptep: vmf->pte);
3838 pte_unmap_unlock(vmf->pte, vmf->ptl);
3839 return VM_FAULT_NOPAGE;
3840 }
3841 wp_page_reuse(vmf, folio);
3842 return 0;
3843}
3844
3845/*
3846 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3847 * mapping
3848 */
3849static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3850{
3851 struct vm_area_struct *vma = vmf->vma;
3852
3853 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3854 vm_fault_t ret;
3855
3856 pte_unmap_unlock(vmf->pte, vmf->ptl);
3857 ret = vmf_can_call_fault(vmf);
3858 if (ret)
3859 return ret;
3860
3861 vmf->flags |= FAULT_FLAG_MKWRITE;
3862 ret = vma->vm_ops->pfn_mkwrite(vmf);
3863 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3864 return ret;
3865 return finish_mkwrite_fault(vmf, NULL);
3866 }
3867 wp_page_reuse(vmf, NULL);
3868 return 0;
3869}
3870
3871static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio)
3872 __releases(vmf->ptl)
3873{
3874 struct vm_area_struct *vma = vmf->vma;
3875 vm_fault_t ret = 0;
3876
3877 folio_get(folio);
3878
3879 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3880 vm_fault_t tmp;
3881
3882 pte_unmap_unlock(vmf->pte, vmf->ptl);
3883 tmp = vmf_can_call_fault(vmf);
3884 if (tmp) {
3885 folio_put(folio);
3886 return tmp;
3887 }
3888
3889 tmp = do_page_mkwrite(vmf, folio);
3890 if (unlikely(!tmp || (tmp &
3891 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3892 folio_put(folio);
3893 return tmp;
3894 }
3895 tmp = finish_mkwrite_fault(vmf, folio);
3896 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3897 folio_unlock(folio);
3898 folio_put(folio);
3899 return tmp;
3900 }
3901 } else {
3902 wp_page_reuse(vmf, folio);
3903 folio_lock(folio);
3904 }
3905 ret |= fault_dirty_shared_page(vmf);
3906 folio_put(folio);
3907
3908 return ret;
3909}
3910
3911#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3912static bool __wp_can_reuse_large_anon_folio(struct folio *folio,
3913 struct vm_area_struct *vma)
3914{
3915 bool exclusive = false;
3916
3917 /* Let's just free up a large folio if only a single page is mapped. */
3918 if (folio_large_mapcount(folio) <= 1)
3919 return false;
3920
3921 /*
3922 * The assumption for anonymous folios is that each page can only get
3923 * mapped once into each MM. The only exception are KSM folios, which
3924 * are always small.
3925 *
3926 * Each taken mapcount must be paired with exactly one taken reference,
3927 * whereby the refcount must be incremented before the mapcount when
3928 * mapping a page, and the refcount must be decremented after the
3929 * mapcount when unmapping a page.
3930 *
3931 * If all folio references are from mappings, and all mappings are in
3932 * the page tables of this MM, then this folio is exclusive to this MM.
3933 */
3934 if (test_bit(FOLIO_MM_IDS_SHARED_BITNUM, &folio->_mm_ids))
3935 return false;
3936
3937 VM_WARN_ON_ONCE(folio_test_ksm(folio));
3938
3939 if (unlikely(folio_test_swapcache(folio))) {
3940 /*
3941 * Note: freeing up the swapcache will fail if some PTEs are
3942 * still swap entries.
3943 */
3944 if (!folio_trylock(folio))
3945 return false;
3946 folio_free_swap(folio);
3947 folio_unlock(folio);
3948 }
3949
3950 if (folio_large_mapcount(folio) != folio_ref_count(folio))
3951 return false;
3952
3953 /* Stabilize the mapcount vs. refcount and recheck. */
3954 folio_lock_large_mapcount(folio);
3955 VM_WARN_ON_ONCE_FOLIO(folio_large_mapcount(folio) > folio_ref_count(folio), folio);
3956
3957 if (test_bit(FOLIO_MM_IDS_SHARED_BITNUM, &folio->_mm_ids))
3958 goto unlock;
3959 if (folio_large_mapcount(folio) != folio_ref_count(folio))
3960 goto unlock;
3961
3962 VM_WARN_ON_ONCE_FOLIO(folio_large_mapcount(folio) > folio_nr_pages(folio), folio);
3963 VM_WARN_ON_ONCE_FOLIO(folio_entire_mapcount(folio), folio);
3964 VM_WARN_ON_ONCE(folio_mm_id(folio, 0) != vma->vm_mm->mm_id &&
3965 folio_mm_id(folio, 1) != vma->vm_mm->mm_id);
3966
3967 /*
3968 * Do we need the folio lock? Likely not. If there would have been
3969 * references from page migration/swapout, we would have detected
3970 * an additional folio reference and never ended up here.
3971 */
3972 exclusive = true;
3973unlock:
3974 folio_unlock_large_mapcount(folio);
3975 return exclusive;
3976}
3977#else /* !CONFIG_TRANSPARENT_HUGEPAGE */
3978static bool __wp_can_reuse_large_anon_folio(struct folio *folio,
3979 struct vm_area_struct *vma)
3980{
3981 BUILD_BUG();
3982}
3983#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3984
3985static bool wp_can_reuse_anon_folio(struct folio *folio,
3986 struct vm_area_struct *vma)
3987{
3988 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && folio_test_large(folio))
3989 return __wp_can_reuse_large_anon_folio(folio, vma);
3990
3991 /*
3992 * We have to verify under folio lock: these early checks are
3993 * just an optimization to avoid locking the folio and freeing
3994 * the swapcache if there is little hope that we can reuse.
3995 *
3996 * KSM doesn't necessarily raise the folio refcount.
3997 */
3998 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3999 return false;
4000 if (!folio_test_lru(folio))
4001 /*
4002 * We cannot easily detect+handle references from
4003 * remote LRU caches or references to LRU folios.
4004 */
4005 lru_add_drain();
4006 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
4007 return false;
4008 if (!folio_trylock(folio))
4009 return false;
4010 if (folio_test_swapcache(folio))
4011 folio_free_swap(folio);
4012 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
4013 folio_unlock(folio);
4014 return false;
4015 }
4016 /*
4017 * Ok, we've got the only folio reference from our mapping
4018 * and the folio is locked, it's dark out, and we're wearing
4019 * sunglasses. Hit it.
4020 */
4021 folio_move_anon_rmap(folio, vma);
4022 folio_unlock(folio);
4023 return true;
4024}
4025
4026/*
4027 * This routine handles present pages, when
4028 * * users try to write to a shared page (FAULT_FLAG_WRITE)
4029 * * GUP wants to take a R/O pin on a possibly shared anonymous page
4030 * (FAULT_FLAG_UNSHARE)
4031 *
4032 * It is done by copying the page to a new address and decrementing the
4033 * shared-page counter for the old page.
4034 *
4035 * Note that this routine assumes that the protection checks have been
4036 * done by the caller (the low-level page fault routine in most cases).
4037 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
4038 * done any necessary COW.
4039 *
4040 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
4041 * though the page will change only once the write actually happens. This
4042 * avoids a few races, and potentially makes it more efficient.
4043 *
4044 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4045 * but allow concurrent faults), with pte both mapped and locked.
4046 * We return with mmap_lock still held, but pte unmapped and unlocked.
4047 */
4048static vm_fault_t do_wp_page(struct vm_fault *vmf)
4049 __releases(vmf->ptl)
4050{
4051 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
4052 struct vm_area_struct *vma = vmf->vma;
4053 struct folio *folio = NULL;
4054 pte_t pte;
4055
4056 if (likely(!unshare)) {
4057 if (userfaultfd_pte_wp(vma, pte: ptep_get(ptep: vmf->pte))) {
4058 if (!userfaultfd_wp_async(vma)) {
4059 pte_unmap_unlock(vmf->pte, vmf->ptl);
4060 return handle_userfault(vmf, VM_UFFD_WP);
4061 }
4062
4063 /*
4064 * Nothing needed (cache flush, TLB invalidations,
4065 * etc.) because we're only removing the uffd-wp bit,
4066 * which is completely invisible to the user.
4067 */
4068 pte = pte_clear_uffd_wp(pte: ptep_get(ptep: vmf->pte));
4069
4070 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
4071 /*
4072 * Update this to be prepared for following up CoW
4073 * handling
4074 */
4075 vmf->orig_pte = pte;
4076 }
4077
4078 /*
4079 * Userfaultfd write-protect can defer flushes. Ensure the TLB
4080 * is flushed in this case before copying.
4081 */
4082 if (unlikely(userfaultfd_wp(vmf->vma) &&
4083 mm_tlb_flush_pending(vmf->vma->vm_mm)))
4084 flush_tlb_page(vma: vmf->vma, a: vmf->address);
4085 }
4086
4087 vmf->page = vm_normal_page(vma, addr: vmf->address, pte: vmf->orig_pte);
4088
4089 if (vmf->page)
4090 folio = page_folio(vmf->page);
4091
4092 /*
4093 * Shared mapping: we are guaranteed to have VM_WRITE and
4094 * FAULT_FLAG_WRITE set at this point.
4095 */
4096 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
4097 /*
4098 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
4099 * VM_PFNMAP VMA. FS DAX also wants ops->pfn_mkwrite called.
4100 *
4101 * We should not cow pages in a shared writeable mapping.
4102 * Just mark the pages writable and/or call ops->pfn_mkwrite.
4103 */
4104 if (!vmf->page || is_fsdax_page(page: vmf->page)) {
4105 vmf->page = NULL;
4106 return wp_pfn_shared(vmf);
4107 }
4108 return wp_page_shared(vmf, folio);
4109 }
4110
4111 /*
4112 * Private mapping: create an exclusive anonymous page copy if reuse
4113 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
4114 *
4115 * If we encounter a page that is marked exclusive, we must reuse
4116 * the page without further checks.
4117 */
4118 if (folio && folio_test_anon(folio) &&
4119 (PageAnonExclusive(page: vmf->page) || wp_can_reuse_anon_folio(folio, vma))) {
4120 if (!PageAnonExclusive(page: vmf->page))
4121 SetPageAnonExclusive(vmf->page);
4122 if (unlikely(unshare)) {
4123 pte_unmap_unlock(vmf->pte, vmf->ptl);
4124 return 0;
4125 }
4126 wp_page_reuse(vmf, folio);
4127 return 0;
4128 }
4129 /*
4130 * Ok, we need to copy. Oh, well..
4131 */
4132 if (folio)
4133 folio_get(folio);
4134
4135 pte_unmap_unlock(vmf->pte, vmf->ptl);
4136#ifdef CONFIG_KSM
4137 if (folio && folio_test_ksm(folio))
4138 count_vm_event(COW_KSM);
4139#endif
4140 return wp_page_copy(vmf);
4141}
4142
4143static void unmap_mapping_range_vma(struct vm_area_struct *vma,
4144 unsigned long start_addr, unsigned long end_addr,
4145 struct zap_details *details)
4146{
4147 zap_page_range_single(vma, address: start_addr, size: end_addr - start_addr, details);
4148}
4149
4150static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
4151 pgoff_t first_index,
4152 pgoff_t last_index,
4153 struct zap_details *details)
4154{
4155 struct vm_area_struct *vma;
4156 pgoff_t vba, vea, zba, zea;
4157
4158 vma_interval_tree_foreach(vma, root, first_index, last_index) {
4159 vba = vma->vm_pgoff;
4160 vea = vba + vma_pages(vma) - 1;
4161 zba = max(first_index, vba);
4162 zea = min(last_index, vea);
4163
4164 unmap_mapping_range_vma(vma,
4165 start_addr: ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
4166 end_addr: ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
4167 details);
4168 }
4169}
4170
4171/**
4172 * unmap_mapping_folio() - Unmap single folio from processes.
4173 * @folio: The locked folio to be unmapped.
4174 *
4175 * Unmap this folio from any userspace process which still has it mmaped.
4176 * Typically, for efficiency, the range of nearby pages has already been
4177 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
4178 * truncation or invalidation holds the lock on a folio, it may find that
4179 * the page has been remapped again: and then uses unmap_mapping_folio()
4180 * to unmap it finally.
4181 */
4182void unmap_mapping_folio(struct folio *folio)
4183{
4184 struct address_space *mapping = folio->mapping;
4185 struct zap_details details = { };
4186 pgoff_t first_index;
4187 pgoff_t last_index;
4188
4189 VM_BUG_ON(!folio_test_locked(folio));
4190
4191 first_index = folio->index;
4192 last_index = folio_next_index(folio) - 1;
4193
4194 details.even_cows = false;
4195 details.single_folio = folio;
4196 details.zap_flags = ZAP_FLAG_DROP_MARKER;
4197
4198 i_mmap_lock_read(mapping);
4199 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
4200 unmap_mapping_range_tree(root: &mapping->i_mmap, first_index,
4201 last_index, details: &details);
4202 i_mmap_unlock_read(mapping);
4203}
4204
4205/**
4206 * unmap_mapping_pages() - Unmap pages from processes.
4207 * @mapping: The address space containing pages to be unmapped.
4208 * @start: Index of first page to be unmapped.
4209 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
4210 * @even_cows: Whether to unmap even private COWed pages.
4211 *
4212 * Unmap the pages in this address space from any userspace process which
4213 * has them mmaped. Generally, you want to remove COWed pages as well when
4214 * a file is being truncated, but not when invalidating pages from the page
4215 * cache.
4216 */
4217void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
4218 pgoff_t nr, bool even_cows)
4219{
4220 struct zap_details details = { };
4221 pgoff_t first_index = start;
4222 pgoff_t last_index = start + nr - 1;
4223
4224 details.even_cows = even_cows;
4225 if (last_index < first_index)
4226 last_index = ULONG_MAX;
4227
4228 i_mmap_lock_read(mapping);
4229 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
4230 unmap_mapping_range_tree(root: &mapping->i_mmap, first_index,
4231 last_index, details: &details);
4232 i_mmap_unlock_read(mapping);
4233}
4234EXPORT_SYMBOL_GPL(unmap_mapping_pages);
4235
4236/**
4237 * unmap_mapping_range - unmap the portion of all mmaps in the specified
4238 * address_space corresponding to the specified byte range in the underlying
4239 * file.
4240 *
4241 * @mapping: the address space containing mmaps to be unmapped.
4242 * @holebegin: byte in first page to unmap, relative to the start of
4243 * the underlying file. This will be rounded down to a PAGE_SIZE
4244 * boundary. Note that this is different from truncate_pagecache(), which
4245 * must keep the partial page. In contrast, we must get rid of
4246 * partial pages.
4247 * @holelen: size of prospective hole in bytes. This will be rounded
4248 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
4249 * end of the file.
4250 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
4251 * but 0 when invalidating pagecache, don't throw away private data.
4252 */
4253void unmap_mapping_range(struct address_space *mapping,
4254 loff_t const holebegin, loff_t const holelen, int even_cows)
4255{
4256 pgoff_t hba = (pgoff_t)(holebegin) >> PAGE_SHIFT;
4257 pgoff_t hlen = ((pgoff_t)(holelen) + PAGE_SIZE - 1) >> PAGE_SHIFT;
4258
4259 /* Check for overflow. */
4260 if (sizeof(holelen) > sizeof(hlen)) {
4261 long long holeend =
4262 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
4263 if (holeend & ~(long long)ULONG_MAX)
4264 hlen = ULONG_MAX - hba + 1;
4265 }
4266
4267 unmap_mapping_pages(mapping, hba, hlen, even_cows);
4268}
4269EXPORT_SYMBOL(unmap_mapping_range);
4270
4271/*
4272 * Restore a potential device exclusive pte to a working pte entry
4273 */
4274static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
4275{
4276 struct folio *folio = page_folio(vmf->page);
4277 struct vm_area_struct *vma = vmf->vma;
4278 struct mmu_notifier_range range;
4279 vm_fault_t ret;
4280
4281 /*
4282 * We need a reference to lock the folio because we don't hold
4283 * the PTL so a racing thread can remove the device-exclusive
4284 * entry and unmap it. If the folio is free the entry must
4285 * have been removed already. If it happens to have already
4286 * been re-allocated after being freed all we do is lock and
4287 * unlock it.
4288 */
4289 if (!folio_try_get(folio))
4290 return 0;
4291
4292 ret = folio_lock_or_retry(folio, vmf);
4293 if (ret) {
4294 folio_put(folio);
4295 return ret;
4296 }
4297 mmu_notifier_range_init_owner(range: &range, event: MMU_NOTIFY_CLEAR, flags: 0,
4298 mm: vma->vm_mm, start: vmf->address & PAGE_MASK,
4299 end: (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
4300 mmu_notifier_invalidate_range_start(range: &range);
4301
4302 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd, addr: vmf->address,
4303 ptlp: &vmf->ptl);
4304 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
4305 restore_exclusive_pte(vma, folio, page: vmf->page, address: vmf->address,
4306 ptep: vmf->pte, orig_pte: vmf->orig_pte);
4307
4308 if (vmf->pte)
4309 pte_unmap_unlock(vmf->pte, vmf->ptl);
4310 folio_unlock(folio);
4311 folio_put(folio);
4312
4313 mmu_notifier_invalidate_range_end(range: &range);
4314 return 0;
4315}
4316
4317static inline bool should_try_to_free_swap(struct folio *folio,
4318 struct vm_area_struct *vma,
4319 unsigned int fault_flags)
4320{
4321 if (!folio_test_swapcache(folio))
4322 return false;
4323 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
4324 folio_test_mlocked(folio))
4325 return true;
4326 /*
4327 * If we want to map a page that's in the swapcache writable, we
4328 * have to detect via the refcount if we're really the exclusive
4329 * user. Try freeing the swapcache to get rid of the swapcache
4330 * reference only in case it's likely that we'll be the exlusive user.
4331 */
4332 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
4333 folio_ref_count(folio) == (1 + folio_nr_pages(folio));
4334}
4335
4336static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
4337{
4338 vmf->pte = pte_offset_map_lock(mm: vmf->vma->vm_mm, pmd: vmf->pmd,
4339 addr: vmf->address, ptlp: &vmf->ptl);
4340 if (!vmf->pte)
4341 return 0;
4342 /*
4343 * Be careful so that we will only recover a special uffd-wp pte into a
4344 * none pte. Otherwise it means the pte could have changed, so retry.
4345 *
4346 * This should also cover the case where e.g. the pte changed
4347 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
4348 * So is_pte_marker() check is not enough to safely drop the pte.
4349 */
4350 if (pte_same(a: vmf->orig_pte, b: ptep_get(ptep: vmf->pte)))
4351 pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
4352 pte_unmap_unlock(vmf->pte, vmf->ptl);
4353 return 0;
4354}
4355
4356static vm_fault_t do_pte_missing(struct vm_fault *vmf)
4357{
4358 if (vma_is_anonymous(vma: vmf->vma))
4359 return do_anonymous_page(vmf);
4360 else
4361 return do_fault(vmf);
4362}
4363
4364/*
4365 * This is actually a page-missing access, but with uffd-wp special pte
4366 * installed. It means this pte was wr-protected before being unmapped.
4367 */
4368static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
4369{
4370 /*
4371 * Just in case there're leftover special ptes even after the region
4372 * got unregistered - we can simply clear them.
4373 */
4374 if (unlikely(!userfaultfd_wp(vmf->vma)))
4375 return pte_marker_clear(vmf);
4376
4377 return do_pte_missing(vmf);
4378}
4379
4380static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
4381{
4382 swp_entry_t entry = pte_to_swp_entry(pte: vmf->orig_pte);
4383 unsigned long marker = pte_marker_get(entry);
4384
4385 /*
4386 * PTE markers should never be empty. If anything weird happened,
4387 * the best thing to do is to kill the process along with its mm.
4388 */
4389 if (WARN_ON_ONCE(!marker))
4390 return VM_FAULT_SIGBUS;
4391
4392 /* Higher priority than uffd-wp when data corrupted */
4393 if (marker & PTE_MARKER_POISONED)
4394 return VM_FAULT_HWPOISON;
4395
4396 /* Hitting a guard page is always a fatal condition. */
4397 if (marker & PTE_MARKER_GUARD)
4398 return VM_FAULT_SIGSEGV;
4399
4400 if (pte_marker_entry_uffd_wp(entry))
4401 return pte_marker_handle_uffd_wp(vmf);
4402
4403 /* This is an unknown pte marker */
4404 return VM_FAULT_SIGBUS;
4405}
4406
4407static struct folio *__alloc_swap_folio(struct vm_fault *vmf)
4408{
4409 struct vm_area_struct *vma = vmf->vma;
4410 struct folio *folio;
4411 swp_entry_t entry;
4412
4413 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma, vmf->address);
4414 if (!folio)
4415 return NULL;
4416
4417 entry = pte_to_swp_entry(pte: vmf->orig_pte);
4418 if (mem_cgroup_swapin_charge_folio(folio, mm: vma->vm_mm,
4419 GFP_KERNEL, entry)) {
4420 folio_put(folio);
4421 return NULL;
4422 }
4423
4424 return folio;
4425}
4426
4427#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4428/*
4429 * Check if the PTEs within a range are contiguous swap entries
4430 * and have consistent swapcache, zeromap.
4431 */
4432static bool can_swapin_thp(struct vm_fault *vmf, pte_t *ptep, int nr_pages)
4433{
4434 unsigned long addr;
4435 swp_entry_t entry;
4436 int idx;
4437 pte_t pte;
4438
4439 addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE);
4440 idx = (vmf->address - addr) / PAGE_SIZE;
4441 pte = ptep_get(ptep);
4442
4443 if (!pte_same(pte, pte_move_swp_offset(vmf->orig_pte, -idx)))
4444 return false;
4445 entry = pte_to_swp_entry(pte);
4446 if (swap_pte_batch(ptep, nr_pages, pte) != nr_pages)
4447 return false;
4448
4449 /*
4450 * swap_read_folio() can't handle the case a large folio is hybridly
4451 * from different backends. And they are likely corner cases. Similar
4452 * things might be added once zswap support large folios.
4453 */
4454 if (unlikely(swap_zeromap_batch(entry, nr_pages, NULL) != nr_pages))
4455 return false;
4456 if (unlikely(non_swapcache_batch(entry, nr_pages) != nr_pages))
4457 return false;
4458
4459 return true;
4460}
4461
4462static inline unsigned long thp_swap_suitable_orders(pgoff_t swp_offset,
4463 unsigned long addr,
4464 unsigned long orders)
4465{
4466 int order, nr;
4467
4468 order = highest_order(orders);
4469
4470 /*
4471 * To swap in a THP with nr pages, we require that its first swap_offset
4472 * is aligned with that number, as it was when the THP was swapped out.
4473 * This helps filter out most invalid entries.
4474 */
4475 while (orders) {
4476 nr = 1 << order;
4477 if ((addr >> PAGE_SHIFT) % nr == swp_offset % nr)
4478 break;
4479 order = next_order(&orders, order);
4480 }
4481
4482 return orders;
4483}
4484
4485static struct folio *alloc_swap_folio(struct vm_fault *vmf)
4486{
4487 struct vm_area_struct *vma = vmf->vma;
4488 unsigned long orders;
4489 struct folio *folio;
4490 unsigned long addr;
4491 swp_entry_t entry;
4492 spinlock_t *ptl;
4493 pte_t *pte;
4494 gfp_t gfp;
4495 int order;
4496
4497 /*
4498 * If uffd is active for the vma we need per-page fault fidelity to
4499 * maintain the uffd semantics.
4500 */
4501 if (unlikely(userfaultfd_armed(vma)))
4502 goto fallback;
4503
4504 /*
4505 * A large swapped out folio could be partially or fully in zswap. We
4506 * lack handling for such cases, so fallback to swapping in order-0
4507 * folio.
4508 */
4509 if (!zswap_never_enabled())
4510 goto fallback;
4511
4512 entry = pte_to_swp_entry(vmf->orig_pte);
4513 /*
4514 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4515 * and suitable for swapping THP.
4516 */
4517 orders = thp_vma_allowable_orders(vma, vma->vm_flags, TVA_PAGEFAULT,
4518 BIT(PMD_ORDER) - 1);
4519 orders = thp_vma_suitable_orders(vma, vmf->address, orders);
4520 orders = thp_swap_suitable_orders(swp_offset(entry),
4521 vmf->address, orders);
4522
4523 if (!orders)
4524 goto fallback;
4525
4526 pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
4527 vmf->address & PMD_MASK, &ptl);
4528 if (unlikely(!pte))
4529 goto fallback;
4530
4531 /*
4532 * For do_swap_page, find the highest order where the aligned range is
4533 * completely swap entries with contiguous swap offsets.
4534 */
4535 order = highest_order(orders);
4536 while (orders) {
4537 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4538 if (can_swapin_thp(vmf, pte + pte_index(addr), 1 << order))
4539 break;
4540 order = next_order(&orders, order);
4541 }
4542
4543 pte_unmap_unlock(pte, ptl);
4544
4545 /* Try allocating the highest of the remaining orders. */
4546 gfp = vma_thp_gfp_mask(vma);
4547 while (orders) {
4548 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4549 folio = vma_alloc_folio(gfp, order, vma, addr);
4550 if (folio) {
4551 if (!mem_cgroup_swapin_charge_folio(folio, vma->vm_mm,
4552 gfp, entry))
4553 return folio;
4554 count_mthp_stat(order, MTHP_STAT_SWPIN_FALLBACK_CHARGE);
4555 folio_put(folio);
4556 }
4557 count_mthp_stat(order, MTHP_STAT_SWPIN_FALLBACK);
4558 order = next_order(&orders, order);
4559 }
4560
4561fallback:
4562 return __alloc_swap_folio(vmf);
4563}
4564#else /* !CONFIG_TRANSPARENT_HUGEPAGE */
4565static struct folio *alloc_swap_folio(struct vm_fault *vmf)
4566{
4567 return __alloc_swap_folio(vmf);
4568}
4569#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4570
4571static DECLARE_WAIT_QUEUE_HEAD(swapcache_wq);
4572
4573/*
4574 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4575 * but allow concurrent faults), and pte mapped but not yet locked.
4576 * We return with pte unmapped and unlocked.
4577 *
4578 * We return with the mmap_lock locked or unlocked in the same cases
4579 * as does filemap_fault().
4580 */
4581vm_fault_t do_swap_page(struct vm_fault *vmf)
4582{
4583 struct vm_area_struct *vma = vmf->vma;
4584 struct folio *swapcache, *folio = NULL;
4585 DECLARE_WAITQUEUE(wait, current);
4586 struct page *page;
4587 struct swap_info_struct *si = NULL;
4588 rmap_t rmap_flags = RMAP_NONE;
4589 bool need_clear_cache = false;
4590 bool exclusive = false;
4591 swp_entry_t entry;
4592 pte_t pte;
4593 vm_fault_t ret = 0;
4594 void *shadow = NULL;
4595 int nr_pages;
4596 unsigned long page_idx;
4597 unsigned long address;
4598 pte_t *ptep;
4599
4600 if (!pte_unmap_same(vmf))
4601 goto out;
4602
4603 entry = pte_to_swp_entry(pte: vmf->orig_pte);
4604 if (unlikely(non_swap_entry(entry))) {
4605 if (is_migration_entry(entry)) {
4606 migration_entry_wait(mm: vma->vm_mm, pmd: vmf->pmd,
4607 address: vmf->address);
4608 } else if (is_device_exclusive_entry(entry)) {
4609 vmf->page = pfn_swap_entry_to_page(entry);
4610 ret = remove_device_exclusive_entry(vmf);
4611 } else if (is_device_private_entry(entry)) {
4612 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
4613 /*
4614 * migrate_to_ram is not yet ready to operate
4615 * under VMA lock.
4616 */
4617 vma_end_read(vma);
4618 ret = VM_FAULT_RETRY;
4619 goto out;
4620 }
4621
4622 vmf->page = pfn_swap_entry_to_page(entry);
4623 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
4624 addr: vmf->address, ptlp: &vmf->ptl);
4625 if (unlikely(!vmf->pte ||
4626 !pte_same(ptep_get(vmf->pte),
4627 vmf->orig_pte)))
4628 goto unlock;
4629
4630 /*
4631 * Get a page reference while we know the page can't be
4632 * freed.
4633 */
4634 if (trylock_page(page: vmf->page)) {
4635 struct dev_pagemap *pgmap;
4636
4637 get_page(page: vmf->page);
4638 pte_unmap_unlock(vmf->pte, vmf->ptl);
4639 pgmap = page_pgmap(page: vmf->page);
4640 ret = pgmap->ops->migrate_to_ram(vmf);
4641 unlock_page(page: vmf->page);
4642 put_page(page: vmf->page);
4643 } else {
4644 pte_unmap_unlock(vmf->pte, vmf->ptl);
4645 }
4646 } else if (is_hwpoison_entry(swp: entry)) {
4647 ret = VM_FAULT_HWPOISON;
4648 } else if (is_pte_marker_entry(entry)) {
4649 ret = handle_pte_marker(vmf);
4650 } else {
4651 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
4652 ret = VM_FAULT_SIGBUS;
4653 }
4654 goto out;
4655 }
4656
4657 /* Prevent swapoff from happening to us. */
4658 si = get_swap_device(entry);
4659 if (unlikely(!si))
4660 goto out;
4661
4662 folio = swap_cache_get_folio(entry);
4663 if (folio)
4664 swap_update_readahead(folio, vma, addr: vmf->address);
4665 swapcache = folio;
4666
4667 if (!folio) {
4668 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
4669 __swap_count(entry) == 1) {
4670 /* skip swapcache */
4671 folio = alloc_swap_folio(vmf);
4672 if (folio) {
4673 __folio_set_locked(folio);
4674 __folio_set_swapbacked(folio);
4675
4676 nr_pages = folio_nr_pages(folio);
4677 if (folio_test_large(folio))
4678 entry.val = ALIGN_DOWN(entry.val, nr_pages);
4679 /*
4680 * Prevent parallel swapin from proceeding with
4681 * the cache flag. Otherwise, another thread
4682 * may finish swapin first, free the entry, and
4683 * swapout reusing the same entry. It's
4684 * undetectable as pte_same() returns true due
4685 * to entry reuse.
4686 */
4687 if (swapcache_prepare(entry, nr: nr_pages)) {
4688 /*
4689 * Relax a bit to prevent rapid
4690 * repeated page faults.
4691 */
4692 add_wait_queue(wq_head: &swapcache_wq, wq_entry: &wait);
4693 schedule_timeout_uninterruptible(timeout: 1);
4694 remove_wait_queue(wq_head: &swapcache_wq, wq_entry: &wait);
4695 goto out_page;
4696 }
4697 need_clear_cache = true;
4698
4699 memcg1_swapin(entry, nr_pages);
4700
4701 shadow = swap_cache_get_shadow(entry);
4702 if (shadow)
4703 workingset_refault(folio, shadow);
4704
4705 folio_add_lru(folio);
4706
4707 /* To provide entry to swap_read_folio() */
4708 folio->swap = entry;
4709 swap_read_folio(folio, NULL);
4710 folio->private = NULL;
4711 }
4712 } else {
4713 folio = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
4714 vmf);
4715 swapcache = folio;
4716 }
4717
4718 if (!folio) {
4719 /*
4720 * Back out if somebody else faulted in this pte
4721 * while we released the pte lock.
4722 */
4723 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
4724 addr: vmf->address, ptlp: &vmf->ptl);
4725 if (likely(vmf->pte &&
4726 pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
4727 ret = VM_FAULT_OOM;
4728 goto unlock;
4729 }
4730
4731 /* Had to read the page from swap area: Major fault */
4732 ret = VM_FAULT_MAJOR;
4733 count_vm_event(item: PGMAJFAULT);
4734 count_memcg_event_mm(mm: vma->vm_mm, idx: PGMAJFAULT);
4735 }
4736
4737 ret |= folio_lock_or_retry(folio, vmf);
4738 if (ret & VM_FAULT_RETRY)
4739 goto out_release;
4740
4741 page = folio_file_page(folio, index: swp_offset(entry));
4742 if (swapcache) {
4743 /*
4744 * Make sure folio_free_swap() or swapoff did not release the
4745 * swapcache from under us. The page pin, and pte_same test
4746 * below, are not enough to exclude that. Even if it is still
4747 * swapcache, we need to check that the page's swap has not
4748 * changed.
4749 */
4750 if (unlikely(!folio_matches_swap_entry(folio, entry)))
4751 goto out_page;
4752
4753 if (unlikely(PageHWPoison(page))) {
4754 /*
4755 * hwpoisoned dirty swapcache pages are kept for killing
4756 * owner processes (which may be unknown at hwpoison time)
4757 */
4758 ret = VM_FAULT_HWPOISON;
4759 goto out_page;
4760 }
4761
4762 /*
4763 * KSM sometimes has to copy on read faults, for example, if
4764 * folio->index of non-ksm folios would be nonlinear inside the
4765 * anon VMA -- the ksm flag is lost on actual swapout.
4766 */
4767 folio = ksm_might_need_to_copy(folio, vma, addr: vmf->address);
4768 if (unlikely(!folio)) {
4769 ret = VM_FAULT_OOM;
4770 folio = swapcache;
4771 goto out_page;
4772 } else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
4773 ret = VM_FAULT_HWPOISON;
4774 folio = swapcache;
4775 goto out_page;
4776 }
4777 if (folio != swapcache)
4778 page = folio_page(folio, 0);
4779
4780 /*
4781 * If we want to map a page that's in the swapcache writable, we
4782 * have to detect via the refcount if we're really the exclusive
4783 * owner. Try removing the extra reference from the local LRU
4784 * caches if required.
4785 */
4786 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
4787 !folio_test_ksm(folio) && !folio_test_lru(folio))
4788 lru_add_drain();
4789 }
4790
4791 folio_throttle_swaprate(folio, GFP_KERNEL);
4792
4793 /*
4794 * Back out if somebody else already faulted in this pte.
4795 */
4796 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd, addr: vmf->address,
4797 ptlp: &vmf->ptl);
4798 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
4799 goto out_nomap;
4800
4801 if (unlikely(!folio_test_uptodate(folio))) {
4802 ret = VM_FAULT_SIGBUS;
4803 goto out_nomap;
4804 }
4805
4806 /* allocated large folios for SWP_SYNCHRONOUS_IO */
4807 if (folio_test_large(folio) && !folio_test_swapcache(folio)) {
4808 unsigned long nr = folio_nr_pages(folio);
4809 unsigned long folio_start = ALIGN_DOWN(vmf->address, nr * PAGE_SIZE);
4810 unsigned long idx = (vmf->address - folio_start) / PAGE_SIZE;
4811 pte_t *folio_ptep = vmf->pte - idx;
4812 pte_t folio_pte = ptep_get(ptep: folio_ptep);
4813
4814 if (!pte_same(a: folio_pte, b: pte_move_swp_offset(pte: vmf->orig_pte, delta: -idx)) ||
4815 swap_pte_batch(start_ptep: folio_ptep, max_nr: nr, pte: folio_pte) != nr)
4816 goto out_nomap;
4817
4818 page_idx = idx;
4819 address = folio_start;
4820 ptep = folio_ptep;
4821 goto check_folio;
4822 }
4823
4824 nr_pages = 1;
4825 page_idx = 0;
4826 address = vmf->address;
4827 ptep = vmf->pte;
4828 if (folio_test_large(folio) && folio_test_swapcache(folio)) {
4829 int nr = folio_nr_pages(folio);
4830 unsigned long idx = folio_page_idx(folio, page);
4831 unsigned long folio_start = address - idx * PAGE_SIZE;
4832 unsigned long folio_end = folio_start + nr * PAGE_SIZE;
4833 pte_t *folio_ptep;
4834 pte_t folio_pte;
4835
4836 if (unlikely(folio_start < max(address & PMD_MASK, vma->vm_start)))
4837 goto check_folio;
4838 if (unlikely(folio_end > pmd_addr_end(address, vma->vm_end)))
4839 goto check_folio;
4840
4841 folio_ptep = vmf->pte - idx;
4842 folio_pte = ptep_get(ptep: folio_ptep);
4843 if (!pte_same(a: folio_pte, b: pte_move_swp_offset(pte: vmf->orig_pte, delta: -idx)) ||
4844 swap_pte_batch(start_ptep: folio_ptep, max_nr: nr, pte: folio_pte) != nr)
4845 goto check_folio;
4846
4847 page_idx = idx;
4848 address = folio_start;
4849 ptep = folio_ptep;
4850 nr_pages = nr;
4851 entry = folio->swap;
4852 page = &folio->page;
4853 }
4854
4855check_folio:
4856 /*
4857 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4858 * must never point at an anonymous page in the swapcache that is
4859 * PG_anon_exclusive. Sanity check that this holds and especially, that
4860 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4861 * check after taking the PT lock and making sure that nobody
4862 * concurrently faulted in this page and set PG_anon_exclusive.
4863 */
4864 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
4865 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
4866
4867 /*
4868 * Check under PT lock (to protect against concurrent fork() sharing
4869 * the swap entry concurrently) for certainly exclusive pages.
4870 */
4871 if (!folio_test_ksm(folio)) {
4872 exclusive = pte_swp_exclusive(pte: vmf->orig_pte);
4873 if (folio != swapcache) {
4874 /*
4875 * We have a fresh page that is not exposed to the
4876 * swapcache -> certainly exclusive.
4877 */
4878 exclusive = true;
4879 } else if (exclusive && folio_test_writeback(folio) &&
4880 data_race(si->flags & SWP_STABLE_WRITES)) {
4881 /*
4882 * This is tricky: not all swap backends support
4883 * concurrent page modifications while under writeback.
4884 *
4885 * So if we stumble over such a page in the swapcache
4886 * we must not set the page exclusive, otherwise we can
4887 * map it writable without further checks and modify it
4888 * while still under writeback.
4889 *
4890 * For these problematic swap backends, simply drop the
4891 * exclusive marker: this is perfectly fine as we start
4892 * writeback only if we fully unmapped the page and
4893 * there are no unexpected references on the page after
4894 * unmapping succeeded. After fully unmapped, no
4895 * further GUP references (FOLL_GET and FOLL_PIN) can
4896 * appear, so dropping the exclusive marker and mapping
4897 * it only R/O is fine.
4898 */
4899 exclusive = false;
4900 }
4901 }
4902
4903 /*
4904 * Some architectures may have to restore extra metadata to the page
4905 * when reading from swap. This metadata may be indexed by swap entry
4906 * so this must be called before swap_free().
4907 */
4908 arch_swap_restore(entry: folio_swap(entry, folio), folio);
4909
4910 /*
4911 * Remove the swap entry and conditionally try to free up the swapcache.
4912 * We're already holding a reference on the page but haven't mapped it
4913 * yet.
4914 */
4915 swap_free_nr(entry, nr_pages);
4916 if (should_try_to_free_swap(folio, vma, fault_flags: vmf->flags))
4917 folio_free_swap(folio);
4918
4919 add_mm_counter(mm: vma->vm_mm, member: MM_ANONPAGES, value: nr_pages);
4920 add_mm_counter(mm: vma->vm_mm, member: MM_SWAPENTS, value: -nr_pages);
4921 pte = mk_pte(page, pgprot: vma->vm_page_prot);
4922 if (pte_swp_soft_dirty(pte: vmf->orig_pte))
4923 pte = pte_mksoft_dirty(pte);
4924 if (pte_swp_uffd_wp(pte: vmf->orig_pte))
4925 pte = pte_mkuffd_wp(pte);
4926
4927 /*
4928 * Same logic as in do_wp_page(); however, optimize for pages that are
4929 * certainly not shared either because we just allocated them without
4930 * exposing them to the swapcache or because the swap entry indicates
4931 * exclusivity.
4932 */
4933 if (!folio_test_ksm(folio) &&
4934 (exclusive || folio_ref_count(folio) == 1)) {
4935 if ((vma->vm_flags & VM_WRITE) && !userfaultfd_pte_wp(vma, pte) &&
4936 !pte_needs_soft_dirty_wp(vma, pte)) {
4937 pte = pte_mkwrite(pte, vma);
4938 if (vmf->flags & FAULT_FLAG_WRITE) {
4939 pte = pte_mkdirty(pte);
4940 vmf->flags &= ~FAULT_FLAG_WRITE;
4941 }
4942 }
4943 rmap_flags |= RMAP_EXCLUSIVE;
4944 }
4945 folio_ref_add(folio, nr: nr_pages - 1);
4946 flush_icache_pages(vma, page, nr: nr_pages);
4947 vmf->orig_pte = pte_advance_pfn(pte, nr: page_idx);
4948
4949 /* ksm created a completely new copy */
4950 if (unlikely(folio != swapcache && swapcache)) {
4951 folio_add_new_anon_rmap(folio, vma, address, RMAP_EXCLUSIVE);
4952 folio_add_lru_vma(folio, vma);
4953 } else if (!folio_test_anon(folio)) {
4954 /*
4955 * We currently only expect small !anon folios which are either
4956 * fully exclusive or fully shared, or new allocated large
4957 * folios which are fully exclusive. If we ever get large
4958 * folios within swapcache here, we have to be careful.
4959 */
4960 VM_WARN_ON_ONCE(folio_test_large(folio) && folio_test_swapcache(folio));
4961 VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
4962 folio_add_new_anon_rmap(folio, vma, address, flags: rmap_flags);
4963 } else {
4964 folio_add_anon_rmap_ptes(folio, page, nr_pages, vma, address,
4965 flags: rmap_flags);
4966 }
4967
4968 VM_BUG_ON(!folio_test_anon(folio) ||
4969 (pte_write(pte) && !PageAnonExclusive(page)));
4970 set_ptes(mm: vma->vm_mm, addr: address, ptep, pte, nr: nr_pages);
4971 arch_do_swap_page_nr(mm: vma->vm_mm, vma, addr: address,
4972 pte, oldpte: pte, nr: nr_pages);
4973
4974 folio_unlock(folio);
4975 if (folio != swapcache && swapcache) {
4976 /*
4977 * Hold the lock to avoid the swap entry to be reused
4978 * until we take the PT lock for the pte_same() check
4979 * (to avoid false positives from pte_same). For
4980 * further safety release the lock after the swap_free
4981 * so that the swap count won't change under a
4982 * parallel locked swapcache.
4983 */
4984 folio_unlock(folio: swapcache);
4985 folio_put(folio: swapcache);
4986 }
4987
4988 if (vmf->flags & FAULT_FLAG_WRITE) {
4989 ret |= do_wp_page(vmf);
4990 if (ret & VM_FAULT_ERROR)
4991 ret &= VM_FAULT_ERROR;
4992 goto out;
4993 }
4994
4995 /* No need to invalidate - it was non-present before */
4996 update_mmu_cache_range(vmf, vma, addr: address, ptep, nr: nr_pages);
4997unlock:
4998 if (vmf->pte)
4999 pte_unmap_unlock(vmf->pte, vmf->ptl);
5000out:
5001 /* Clear the swap cache pin for direct swapin after PTL unlock */
5002 if (need_clear_cache) {
5003 swapcache_clear(si, entry, nr: nr_pages);
5004 if (waitqueue_active(wq_head: &swapcache_wq))
5005 wake_up(&swapcache_wq);
5006 }
5007 if (si)
5008 put_swap_device(si);
5009 return ret;
5010out_nomap:
5011 if (vmf->pte)
5012 pte_unmap_unlock(vmf->pte, vmf->ptl);
5013out_page:
5014 folio_unlock(folio);
5015out_release:
5016 folio_put(folio);
5017 if (folio != swapcache && swapcache) {
5018 folio_unlock(folio: swapcache);
5019 folio_put(folio: swapcache);
5020 }
5021 if (need_clear_cache) {
5022 swapcache_clear(si, entry, nr: nr_pages);
5023 if (waitqueue_active(wq_head: &swapcache_wq))
5024 wake_up(&swapcache_wq);
5025 }
5026 if (si)
5027 put_swap_device(si);
5028 return ret;
5029}
5030
5031static bool pte_range_none(pte_t *pte, int nr_pages)
5032{
5033 int i;
5034
5035 for (i = 0; i < nr_pages; i++) {
5036 if (!pte_none(pte: ptep_get_lockless(ptep: pte + i)))
5037 return false;
5038 }
5039
5040 return true;
5041}
5042
5043static struct folio *alloc_anon_folio(struct vm_fault *vmf)
5044{
5045 struct vm_area_struct *vma = vmf->vma;
5046#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5047 unsigned long orders;
5048 struct folio *folio;
5049 unsigned long addr;
5050 pte_t *pte;
5051 gfp_t gfp;
5052 int order;
5053
5054 /*
5055 * If uffd is active for the vma we need per-page fault fidelity to
5056 * maintain the uffd semantics.
5057 */
5058 if (unlikely(userfaultfd_armed(vma)))
5059 goto fallback;
5060
5061 /*
5062 * Get a list of all the (large) orders below PMD_ORDER that are enabled
5063 * for this vma. Then filter out the orders that can't be allocated over
5064 * the faulting address and still be fully contained in the vma.
5065 */
5066 orders = thp_vma_allowable_orders(vma, vma->vm_flags, TVA_PAGEFAULT,
5067 BIT(PMD_ORDER) - 1);
5068 orders = thp_vma_suitable_orders(vma, vmf->address, orders);
5069
5070 if (!orders)
5071 goto fallback;
5072
5073 pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK);
5074 if (!pte)
5075 return ERR_PTR(-EAGAIN);
5076
5077 /*
5078 * Find the highest order where the aligned range is completely
5079 * pte_none(). Note that all remaining orders will be completely
5080 * pte_none().
5081 */
5082 order = highest_order(orders);
5083 while (orders) {
5084 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
5085 if (pte_range_none(pte + pte_index(addr), 1 << order))
5086 break;
5087 order = next_order(&orders, order);
5088 }
5089
5090 pte_unmap(pte);
5091
5092 if (!orders)
5093 goto fallback;
5094
5095 /* Try allocating the highest of the remaining orders. */
5096 gfp = vma_thp_gfp_mask(vma);
5097 while (orders) {
5098 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
5099 folio = vma_alloc_folio(gfp, order, vma, addr);
5100 if (folio) {
5101 if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
5102 count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
5103 folio_put(folio);
5104 goto next;
5105 }
5106 folio_throttle_swaprate(folio, gfp);
5107 /*
5108 * When a folio is not zeroed during allocation
5109 * (__GFP_ZERO not used) or user folios require special
5110 * handling, folio_zero_user() is used to make sure
5111 * that the page corresponding to the faulting address
5112 * will be hot in the cache after zeroing.
5113 */
5114 if (user_alloc_needs_zeroing())
5115 folio_zero_user(folio, vmf->address);
5116 return folio;
5117 }
5118next:
5119 count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK);
5120 order = next_order(&orders, order);
5121 }
5122
5123fallback:
5124#endif
5125 return folio_prealloc(src_mm: vma->vm_mm, vma, addr: vmf->address, need_zero: true);
5126}
5127
5128/*
5129 * We enter with non-exclusive mmap_lock (to exclude vma changes,
5130 * but allow concurrent faults), and pte mapped but not yet locked.
5131 * We return with mmap_lock still held, but pte unmapped and unlocked.
5132 */
5133static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
5134{
5135 struct vm_area_struct *vma = vmf->vma;
5136 unsigned long addr = vmf->address;
5137 struct folio *folio;
5138 vm_fault_t ret = 0;
5139 int nr_pages = 1;
5140 pte_t entry;
5141
5142 /* File mapping without ->vm_ops ? */
5143 if (vma->vm_flags & VM_SHARED)
5144 return VM_FAULT_SIGBUS;
5145
5146 /*
5147 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
5148 * be distinguished from a transient failure of pte_offset_map().
5149 */
5150 if (pte_alloc(vma->vm_mm, vmf->pmd))
5151 return VM_FAULT_OOM;
5152
5153 /* Use the zero-page for reads */
5154 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
5155 !mm_forbids_zeropage(vma->vm_mm)) {
5156 entry = pte_mkspecial(pte: pfn_pte(page_nr: my_zero_pfn(addr: vmf->address),
5157 pgprot: vma->vm_page_prot));
5158 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
5159 addr: vmf->address, ptlp: &vmf->ptl);
5160 if (!vmf->pte)
5161 goto unlock;
5162 if (vmf_pte_changed(vmf)) {
5163 update_mmu_tlb(vma, address: vmf->address, ptep: vmf->pte);
5164 goto unlock;
5165 }
5166 ret = check_stable_address_space(mm: vma->vm_mm);
5167 if (ret)
5168 goto unlock;
5169 /* Deliver the page fault to userland, check inside PT lock */
5170 if (userfaultfd_missing(vma)) {
5171 pte_unmap_unlock(vmf->pte, vmf->ptl);
5172 return handle_userfault(vmf, VM_UFFD_MISSING);
5173 }
5174 goto setpte;
5175 }
5176
5177 /* Allocate our own private page. */
5178 ret = vmf_anon_prepare(vmf);
5179 if (ret)
5180 return ret;
5181 /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
5182 folio = alloc_anon_folio(vmf);
5183 if (IS_ERR(ptr: folio))
5184 return 0;
5185 if (!folio)
5186 goto oom;
5187
5188 nr_pages = folio_nr_pages(folio);
5189 addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE);
5190
5191 /*
5192 * The memory barrier inside __folio_mark_uptodate makes sure that
5193 * preceding stores to the page contents become visible before
5194 * the set_pte_at() write.
5195 */
5196 __folio_mark_uptodate(folio);
5197
5198 entry = folio_mk_pte(folio, pgprot: vma->vm_page_prot);
5199 entry = pte_sw_mkyoung(pte: entry);
5200 if (vma->vm_flags & VM_WRITE)
5201 entry = pte_mkwrite(pte: pte_mkdirty(pte: entry), vma);
5202
5203 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd, addr, ptlp: &vmf->ptl);
5204 if (!vmf->pte)
5205 goto release;
5206 if (nr_pages == 1 && vmf_pte_changed(vmf)) {
5207 update_mmu_tlb(vma, address: addr, ptep: vmf->pte);
5208 goto release;
5209 } else if (nr_pages > 1 && !pte_range_none(pte: vmf->pte, nr_pages)) {
5210 update_mmu_tlb_range(vma, address: addr, ptep: vmf->pte, nr: nr_pages);
5211 goto release;
5212 }
5213
5214 ret = check_stable_address_space(mm: vma->vm_mm);
5215 if (ret)
5216 goto release;
5217
5218 /* Deliver the page fault to userland, check inside PT lock */
5219 if (userfaultfd_missing(vma)) {
5220 pte_unmap_unlock(vmf->pte, vmf->ptl);
5221 folio_put(folio);
5222 return handle_userfault(vmf, VM_UFFD_MISSING);
5223 }
5224
5225 folio_ref_add(folio, nr: nr_pages - 1);
5226 add_mm_counter(mm: vma->vm_mm, member: MM_ANONPAGES, value: nr_pages);
5227 count_mthp_stat(order: folio_order(folio), item: MTHP_STAT_ANON_FAULT_ALLOC);
5228 folio_add_new_anon_rmap(folio, vma, address: addr, RMAP_EXCLUSIVE);
5229 folio_add_lru_vma(folio, vma);
5230setpte:
5231 if (vmf_orig_pte_uffd_wp(vmf))
5232 entry = pte_mkuffd_wp(pte: entry);
5233 set_ptes(mm: vma->vm_mm, addr, ptep: vmf->pte, pte: entry, nr: nr_pages);
5234
5235 /* No need to invalidate - it was non-present before */
5236 update_mmu_cache_range(vmf, vma, addr, ptep: vmf->pte, nr: nr_pages);
5237unlock:
5238 if (vmf->pte)
5239 pte_unmap_unlock(vmf->pte, vmf->ptl);
5240 return ret;
5241release:
5242 folio_put(folio);
5243 goto unlock;
5244oom:
5245 return VM_FAULT_OOM;
5246}
5247
5248/*
5249 * The mmap_lock must have been held on entry, and may have been
5250 * released depending on flags and vma->vm_ops->fault() return value.
5251 * See filemap_fault() and __lock_page_retry().
5252 */
5253static vm_fault_t __do_fault(struct vm_fault *vmf)
5254{
5255 struct vm_area_struct *vma = vmf->vma;
5256 struct folio *folio;
5257 vm_fault_t ret;
5258
5259 /*
5260 * Preallocate pte before we take page_lock because this might lead to
5261 * deadlocks for memcg reclaim which waits for pages under writeback:
5262 * lock_page(A)
5263 * SetPageWriteback(A)
5264 * unlock_page(A)
5265 * lock_page(B)
5266 * lock_page(B)
5267 * pte_alloc_one
5268 * shrink_folio_list
5269 * wait_on_page_writeback(A)
5270 * SetPageWriteback(B)
5271 * unlock_page(B)
5272 * # flush A, B to clear the writeback
5273 */
5274 if (pmd_none(pmd: *vmf->pmd) && !vmf->prealloc_pte) {
5275 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
5276 if (!vmf->prealloc_pte)
5277 return VM_FAULT_OOM;
5278 }
5279
5280 ret = vma->vm_ops->fault(vmf);
5281 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
5282 VM_FAULT_DONE_COW)))
5283 return ret;
5284
5285 folio = page_folio(vmf->page);
5286 if (unlikely(PageHWPoison(vmf->page))) {
5287 vm_fault_t poisonret = VM_FAULT_HWPOISON;
5288 if (ret & VM_FAULT_LOCKED) {
5289 if (page_mapped(page: vmf->page))
5290 unmap_mapping_folio(folio);
5291 /* Retry if a clean folio was removed from the cache. */
5292 if (mapping_evict_folio(mapping: folio->mapping, folio))
5293 poisonret = VM_FAULT_NOPAGE;
5294 folio_unlock(folio);
5295 }
5296 folio_put(folio);
5297 vmf->page = NULL;
5298 return poisonret;
5299 }
5300
5301 if (unlikely(!(ret & VM_FAULT_LOCKED)))
5302 folio_lock(folio);
5303 else
5304 VM_BUG_ON_PAGE(!folio_test_locked(folio), vmf->page);
5305
5306 return ret;
5307}
5308
5309#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5310static void deposit_prealloc_pte(struct vm_fault *vmf)
5311{
5312 struct vm_area_struct *vma = vmf->vma;
5313
5314 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
5315 /*
5316 * We are going to consume the prealloc table,
5317 * count that as nr_ptes.
5318 */
5319 mm_inc_nr_ptes(vma->vm_mm);
5320 vmf->prealloc_pte = NULL;
5321}
5322
5323vm_fault_t do_set_pmd(struct vm_fault *vmf, struct folio *folio, struct page *page)
5324{
5325 struct vm_area_struct *vma = vmf->vma;
5326 bool write = vmf->flags & FAULT_FLAG_WRITE;
5327 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
5328 pmd_t entry;
5329 vm_fault_t ret = VM_FAULT_FALLBACK;
5330
5331 /*
5332 * It is too late to allocate a small folio, we already have a large
5333 * folio in the pagecache: especially s390 KVM cannot tolerate any
5334 * PMD mappings, but PTE-mapped THP are fine. So let's simply refuse any
5335 * PMD mappings if THPs are disabled. As we already have a THP,
5336 * behave as if we are forcing a collapse.
5337 */
5338 if (thp_disabled_by_hw() || vma_thp_disabled(vma, vma->vm_flags,
5339 /* forced_collapse=*/ true))
5340 return ret;
5341
5342 if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
5343 return ret;
5344
5345 if (folio_order(folio) != HPAGE_PMD_ORDER)
5346 return ret;
5347 page = &folio->page;
5348
5349 /*
5350 * Just backoff if any subpage of a THP is corrupted otherwise
5351 * the corrupted page may mapped by PMD silently to escape the
5352 * check. This kind of THP just can be PTE mapped. Access to
5353 * the corrupted subpage should trigger SIGBUS as expected.
5354 */
5355 if (unlikely(folio_test_has_hwpoisoned(folio)))
5356 return ret;
5357
5358 /*
5359 * Archs like ppc64 need additional space to store information
5360 * related to pte entry. Use the preallocated table for that.
5361 */
5362 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
5363 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
5364 if (!vmf->prealloc_pte)
5365 return VM_FAULT_OOM;
5366 }
5367
5368 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
5369 if (unlikely(!pmd_none(*vmf->pmd)))
5370 goto out;
5371
5372 flush_icache_pages(vma, page, HPAGE_PMD_NR);
5373
5374 entry = folio_mk_pmd(folio, vma->vm_page_prot);
5375 if (write)
5376 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
5377
5378 add_mm_counter(vma->vm_mm, mm_counter_file(folio), HPAGE_PMD_NR);
5379 folio_add_file_rmap_pmd(folio, page, vma);
5380
5381 /*
5382 * deposit and withdraw with pmd lock held
5383 */
5384 if (arch_needs_pgtable_deposit())
5385 deposit_prealloc_pte(vmf);
5386
5387 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
5388
5389 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
5390
5391 /* fault is handled */
5392 ret = 0;
5393 count_vm_event(THP_FILE_MAPPED);
5394out:
5395 spin_unlock(vmf->ptl);
5396 return ret;
5397}
5398#else
5399vm_fault_t do_set_pmd(struct vm_fault *vmf, struct folio *folio, struct page *page)
5400{
5401 return VM_FAULT_FALLBACK;
5402}
5403#endif
5404
5405/**
5406 * set_pte_range - Set a range of PTEs to point to pages in a folio.
5407 * @vmf: Fault decription.
5408 * @folio: The folio that contains @page.
5409 * @page: The first page to create a PTE for.
5410 * @nr: The number of PTEs to create.
5411 * @addr: The first address to create a PTE for.
5412 */
5413void set_pte_range(struct vm_fault *vmf, struct folio *folio,
5414 struct page *page, unsigned int nr, unsigned long addr)
5415{
5416 struct vm_area_struct *vma = vmf->vma;
5417 bool write = vmf->flags & FAULT_FLAG_WRITE;
5418 bool prefault = !in_range(vmf->address, addr, nr * PAGE_SIZE);
5419 pte_t entry;
5420
5421 flush_icache_pages(vma, page, nr);
5422 entry = mk_pte(page, pgprot: vma->vm_page_prot);
5423
5424 if (prefault && arch_wants_old_prefaulted_pte())
5425 entry = pte_mkold(pte: entry);
5426 else
5427 entry = pte_sw_mkyoung(pte: entry);
5428
5429 if (write)
5430 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
5431 else if (pte_write(pte: entry) && folio_test_dirty(folio))
5432 entry = pte_mkdirty(pte: entry);
5433 if (unlikely(vmf_orig_pte_uffd_wp(vmf)))
5434 entry = pte_mkuffd_wp(pte: entry);
5435 /* copy-on-write page */
5436 if (write && !(vma->vm_flags & VM_SHARED)) {
5437 VM_BUG_ON_FOLIO(nr != 1, folio);
5438 folio_add_new_anon_rmap(folio, vma, address: addr, RMAP_EXCLUSIVE);
5439 folio_add_lru_vma(folio, vma);
5440 } else {
5441 folio_add_file_rmap_ptes(folio, page, nr_pages: nr, vma);
5442 }
5443 set_ptes(mm: vma->vm_mm, addr, ptep: vmf->pte, pte: entry, nr);
5444
5445 /* no need to invalidate: a not-present page won't be cached */
5446 update_mmu_cache_range(vmf, vma, addr, ptep: vmf->pte, nr);
5447}
5448
5449static bool vmf_pte_changed(struct vm_fault *vmf)
5450{
5451 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
5452 return !pte_same(a: ptep_get(ptep: vmf->pte), b: vmf->orig_pte);
5453
5454 return !pte_none(pte: ptep_get(ptep: vmf->pte));
5455}
5456
5457/**
5458 * finish_fault - finish page fault once we have prepared the page to fault
5459 *
5460 * @vmf: structure describing the fault
5461 *
5462 * This function handles all that is needed to finish a page fault once the
5463 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
5464 * given page, adds reverse page mapping, handles memcg charges and LRU
5465 * addition.
5466 *
5467 * The function expects the page to be locked and on success it consumes a
5468 * reference of a page being mapped (for the PTE which maps it).
5469 *
5470 * Return: %0 on success, %VM_FAULT_ code in case of error.
5471 */
5472vm_fault_t finish_fault(struct vm_fault *vmf)
5473{
5474 struct vm_area_struct *vma = vmf->vma;
5475 struct page *page;
5476 struct folio *folio;
5477 vm_fault_t ret;
5478 bool is_cow = (vmf->flags & FAULT_FLAG_WRITE) &&
5479 !(vma->vm_flags & VM_SHARED);
5480 int type, nr_pages;
5481 unsigned long addr;
5482 bool needs_fallback = false;
5483
5484fallback:
5485 addr = vmf->address;
5486
5487 /* Did we COW the page? */
5488 if (is_cow)
5489 page = vmf->cow_page;
5490 else
5491 page = vmf->page;
5492
5493 folio = page_folio(page);
5494 /*
5495 * check even for read faults because we might have lost our CoWed
5496 * page
5497 */
5498 if (!(vma->vm_flags & VM_SHARED)) {
5499 ret = check_stable_address_space(mm: vma->vm_mm);
5500 if (ret)
5501 return ret;
5502 }
5503
5504 if (pmd_none(pmd: *vmf->pmd)) {
5505 if (folio_test_pmd_mappable(folio)) {
5506 ret = do_set_pmd(vmf, folio, page);
5507 if (ret != VM_FAULT_FALLBACK)
5508 return ret;
5509 }
5510
5511 if (vmf->prealloc_pte)
5512 pmd_install(mm: vma->vm_mm, pmd: vmf->pmd, pte: &vmf->prealloc_pte);
5513 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
5514 return VM_FAULT_OOM;
5515 }
5516
5517 nr_pages = folio_nr_pages(folio);
5518
5519 /* Using per-page fault to maintain the uffd semantics */
5520 if (unlikely(userfaultfd_armed(vma)) || unlikely(needs_fallback)) {
5521 nr_pages = 1;
5522 } else if (nr_pages > 1) {
5523 pgoff_t idx = folio_page_idx(folio, page);
5524 /* The page offset of vmf->address within the VMA. */
5525 pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
5526 /* The index of the entry in the pagetable for fault page. */
5527 pgoff_t pte_off = pte_index(address: vmf->address);
5528
5529 /*
5530 * Fallback to per-page fault in case the folio size in page
5531 * cache beyond the VMA limits and PMD pagetable limits.
5532 */
5533 if (unlikely(vma_off < idx ||
5534 vma_off + (nr_pages - idx) > vma_pages(vma) ||
5535 pte_off < idx ||
5536 pte_off + (nr_pages - idx) > PTRS_PER_PTE)) {
5537 nr_pages = 1;
5538 } else {
5539 /* Now we can set mappings for the whole large folio. */
5540 addr = vmf->address - idx * PAGE_SIZE;
5541 page = &folio->page;
5542 }
5543 }
5544
5545 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
5546 addr, ptlp: &vmf->ptl);
5547 if (!vmf->pte)
5548 return VM_FAULT_NOPAGE;
5549
5550 /* Re-check under ptl */
5551 if (nr_pages == 1 && unlikely(vmf_pte_changed(vmf))) {
5552 update_mmu_tlb(vma, address: addr, ptep: vmf->pte);
5553 ret = VM_FAULT_NOPAGE;
5554 goto unlock;
5555 } else if (nr_pages > 1 && !pte_range_none(pte: vmf->pte, nr_pages)) {
5556 needs_fallback = true;
5557 pte_unmap_unlock(vmf->pte, vmf->ptl);
5558 goto fallback;
5559 }
5560
5561 folio_ref_add(folio, nr: nr_pages - 1);
5562 set_pte_range(vmf, folio, page, nr: nr_pages, addr);
5563 type = is_cow ? MM_ANONPAGES : mm_counter_file(folio);
5564 add_mm_counter(mm: vma->vm_mm, member: type, value: nr_pages);
5565 ret = 0;
5566
5567unlock:
5568 pte_unmap_unlock(vmf->pte, vmf->ptl);
5569 return ret;
5570}
5571
5572static unsigned long fault_around_pages __read_mostly =
5573 65536 >> PAGE_SHIFT;
5574
5575#ifdef CONFIG_DEBUG_FS
5576static int fault_around_bytes_get(void *data, u64 *val)
5577{
5578 *val = fault_around_pages << PAGE_SHIFT;
5579 return 0;
5580}
5581
5582/*
5583 * fault_around_bytes must be rounded down to the nearest page order as it's
5584 * what do_fault_around() expects to see.
5585 */
5586static int fault_around_bytes_set(void *data, u64 val)
5587{
5588 if (val / PAGE_SIZE > PTRS_PER_PTE)
5589 return -EINVAL;
5590
5591 /*
5592 * The minimum value is 1 page, however this results in no fault-around
5593 * at all. See should_fault_around().
5594 */
5595 val = max(val, PAGE_SIZE);
5596 fault_around_pages = rounddown_pow_of_two(val) >> PAGE_SHIFT;
5597
5598 return 0;
5599}
5600DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
5601 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
5602
5603static int __init fault_around_debugfs(void)
5604{
5605 debugfs_create_file_unsafe(name: "fault_around_bytes", mode: 0644, NULL, NULL,
5606 fops: &fault_around_bytes_fops);
5607 return 0;
5608}
5609late_initcall(fault_around_debugfs);
5610#endif
5611
5612/*
5613 * do_fault_around() tries to map few pages around the fault address. The hope
5614 * is that the pages will be needed soon and this will lower the number of
5615 * faults to handle.
5616 *
5617 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
5618 * not ready to be mapped: not up-to-date, locked, etc.
5619 *
5620 * This function doesn't cross VMA or page table boundaries, in order to call
5621 * map_pages() and acquire a PTE lock only once.
5622 *
5623 * fault_around_pages defines how many pages we'll try to map.
5624 * do_fault_around() expects it to be set to a power of two less than or equal
5625 * to PTRS_PER_PTE.
5626 *
5627 * The virtual address of the area that we map is naturally aligned to
5628 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
5629 * (and therefore to page order). This way it's easier to guarantee
5630 * that we don't cross page table boundaries.
5631 */
5632static vm_fault_t do_fault_around(struct vm_fault *vmf)
5633{
5634 pgoff_t nr_pages = READ_ONCE(fault_around_pages);
5635 pgoff_t pte_off = pte_index(address: vmf->address);
5636 /* The page offset of vmf->address within the VMA. */
5637 pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
5638 pgoff_t from_pte, to_pte;
5639 vm_fault_t ret;
5640
5641 /* The PTE offset of the start address, clamped to the VMA. */
5642 from_pte = max(ALIGN_DOWN(pte_off, nr_pages),
5643 pte_off - min(pte_off, vma_off));
5644
5645 /* The PTE offset of the end address, clamped to the VMA and PTE. */
5646 to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE,
5647 pte_off + vma_pages(vmf->vma) - vma_off) - 1;
5648
5649 if (pmd_none(pmd: *vmf->pmd)) {
5650 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
5651 if (!vmf->prealloc_pte)
5652 return VM_FAULT_OOM;
5653 }
5654
5655 rcu_read_lock();
5656 ret = vmf->vma->vm_ops->map_pages(vmf,
5657 vmf->pgoff + from_pte - pte_off,
5658 vmf->pgoff + to_pte - pte_off);
5659 rcu_read_unlock();
5660
5661 return ret;
5662}
5663
5664/* Return true if we should do read fault-around, false otherwise */
5665static inline bool should_fault_around(struct vm_fault *vmf)
5666{
5667 /* No ->map_pages? No way to fault around... */
5668 if (!vmf->vma->vm_ops->map_pages)
5669 return false;
5670
5671 if (uffd_disable_fault_around(vma: vmf->vma))
5672 return false;
5673
5674 /* A single page implies no faulting 'around' at all. */
5675 return fault_around_pages > 1;
5676}
5677
5678static vm_fault_t do_read_fault(struct vm_fault *vmf)
5679{
5680 vm_fault_t ret = 0;
5681 struct folio *folio;
5682
5683 /*
5684 * Let's call ->map_pages() first and use ->fault() as fallback
5685 * if page by the offset is not ready to be mapped (cold cache or
5686 * something).
5687 */
5688 if (should_fault_around(vmf)) {
5689 ret = do_fault_around(vmf);
5690 if (ret)
5691 return ret;
5692 }
5693
5694 ret = vmf_can_call_fault(vmf);
5695 if (ret)
5696 return ret;
5697
5698 ret = __do_fault(vmf);
5699 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
5700 return ret;
5701
5702 ret |= finish_fault(vmf);
5703 folio = page_folio(vmf->page);
5704 folio_unlock(folio);
5705 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
5706 folio_put(folio);
5707 return ret;
5708}
5709
5710static vm_fault_t do_cow_fault(struct vm_fault *vmf)
5711{
5712 struct vm_area_struct *vma = vmf->vma;
5713 struct folio *folio;
5714 vm_fault_t ret;
5715
5716 ret = vmf_can_call_fault(vmf);
5717 if (!ret)
5718 ret = vmf_anon_prepare(vmf);
5719 if (ret)
5720 return ret;
5721
5722 folio = folio_prealloc(src_mm: vma->vm_mm, vma, addr: vmf->address, need_zero: false);
5723 if (!folio)
5724 return VM_FAULT_OOM;
5725
5726 vmf->cow_page = &folio->page;
5727
5728 ret = __do_fault(vmf);
5729 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
5730 goto uncharge_out;
5731 if (ret & VM_FAULT_DONE_COW)
5732 return ret;
5733
5734 if (copy_mc_user_highpage(to: vmf->cow_page, from: vmf->page, vaddr: vmf->address, vma)) {
5735 ret = VM_FAULT_HWPOISON;
5736 goto unlock;
5737 }
5738 __folio_mark_uptodate(folio);
5739
5740 ret |= finish_fault(vmf);
5741unlock:
5742 unlock_page(page: vmf->page);
5743 put_page(page: vmf->page);
5744 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
5745 goto uncharge_out;
5746 return ret;
5747uncharge_out:
5748 folio_put(folio);
5749 return ret;
5750}
5751
5752static vm_fault_t do_shared_fault(struct vm_fault *vmf)
5753{
5754 struct vm_area_struct *vma = vmf->vma;
5755 vm_fault_t ret, tmp;
5756 struct folio *folio;
5757
5758 ret = vmf_can_call_fault(vmf);
5759 if (ret)
5760 return ret;
5761
5762 ret = __do_fault(vmf);
5763 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
5764 return ret;
5765
5766 folio = page_folio(vmf->page);
5767
5768 /*
5769 * Check if the backing address space wants to know that the page is
5770 * about to become writable
5771 */
5772 if (vma->vm_ops->page_mkwrite) {
5773 folio_unlock(folio);
5774 tmp = do_page_mkwrite(vmf, folio);
5775 if (unlikely(!tmp ||
5776 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
5777 folio_put(folio);
5778 return tmp;
5779 }
5780 }
5781
5782 ret |= finish_fault(vmf);
5783 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
5784 VM_FAULT_RETRY))) {
5785 folio_unlock(folio);
5786 folio_put(folio);
5787 return ret;
5788 }
5789
5790 ret |= fault_dirty_shared_page(vmf);
5791 return ret;
5792}
5793
5794/*
5795 * We enter with non-exclusive mmap_lock (to exclude vma changes,
5796 * but allow concurrent faults).
5797 * The mmap_lock may have been released depending on flags and our
5798 * return value. See filemap_fault() and __folio_lock_or_retry().
5799 * If mmap_lock is released, vma may become invalid (for example
5800 * by other thread calling munmap()).
5801 */
5802static vm_fault_t do_fault(struct vm_fault *vmf)
5803{
5804 struct vm_area_struct *vma = vmf->vma;
5805 struct mm_struct *vm_mm = vma->vm_mm;
5806 vm_fault_t ret;
5807
5808 /*
5809 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
5810 */
5811 if (!vma->vm_ops->fault) {
5812 vmf->pte = pte_offset_map_lock(mm: vmf->vma->vm_mm, pmd: vmf->pmd,
5813 addr: vmf->address, ptlp: &vmf->ptl);
5814 if (unlikely(!vmf->pte))
5815 ret = VM_FAULT_SIGBUS;
5816 else {
5817 /*
5818 * Make sure this is not a temporary clearing of pte
5819 * by holding ptl and checking again. A R/M/W update
5820 * of pte involves: take ptl, clearing the pte so that
5821 * we don't have concurrent modification by hardware
5822 * followed by an update.
5823 */
5824 if (unlikely(pte_none(ptep_get(vmf->pte))))
5825 ret = VM_FAULT_SIGBUS;
5826 else
5827 ret = VM_FAULT_NOPAGE;
5828
5829 pte_unmap_unlock(vmf->pte, vmf->ptl);
5830 }
5831 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
5832 ret = do_read_fault(vmf);
5833 else if (!(vma->vm_flags & VM_SHARED))
5834 ret = do_cow_fault(vmf);
5835 else
5836 ret = do_shared_fault(vmf);
5837
5838 /* preallocated pagetable is unused: free it */
5839 if (vmf->prealloc_pte) {
5840 pte_free(mm: vm_mm, pte_page: vmf->prealloc_pte);
5841 vmf->prealloc_pte = NULL;
5842 }
5843 return ret;
5844}
5845
5846int numa_migrate_check(struct folio *folio, struct vm_fault *vmf,
5847 unsigned long addr, int *flags,
5848 bool writable, int *last_cpupid)
5849{
5850 struct vm_area_struct *vma = vmf->vma;
5851
5852 /*
5853 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
5854 * much anyway since they can be in shared cache state. This misses
5855 * the case where a mapping is writable but the process never writes
5856 * to it but pte_write gets cleared during protection updates and
5857 * pte_dirty has unpredictable behaviour between PTE scan updates,
5858 * background writeback, dirty balancing and application behaviour.
5859 */
5860 if (!writable)
5861 *flags |= TNF_NO_GROUP;
5862
5863 /*
5864 * Flag if the folio is shared between multiple address spaces. This
5865 * is later used when determining whether to group tasks together
5866 */
5867 if (folio_maybe_mapped_shared(folio) && (vma->vm_flags & VM_SHARED))
5868 *flags |= TNF_SHARED;
5869 /*
5870 * For memory tiering mode, cpupid of slow memory page is used
5871 * to record page access time. So use default value.
5872 */
5873 if (folio_use_access_time(folio))
5874 *last_cpupid = (-1 & LAST_CPUPID_MASK);
5875 else
5876 *last_cpupid = folio_last_cpupid(folio);
5877
5878 /* Record the current PID acceesing VMA */
5879 vma_set_access_pid_bit(vma);
5880
5881 count_vm_numa_event(NUMA_HINT_FAULTS);
5882#ifdef CONFIG_NUMA_BALANCING
5883 count_memcg_folio_events(folio, NUMA_HINT_FAULTS, 1);
5884#endif
5885 if (folio_nid(folio) == numa_node_id()) {
5886 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
5887 *flags |= TNF_FAULT_LOCAL;
5888 }
5889
5890 return mpol_misplaced(folio, vmf, addr);
5891}
5892
5893static void numa_rebuild_single_mapping(struct vm_fault *vmf, struct vm_area_struct *vma,
5894 unsigned long fault_addr, pte_t *fault_pte,
5895 bool writable)
5896{
5897 pte_t pte, old_pte;
5898
5899 old_pte = ptep_modify_prot_start(vma, addr: fault_addr, ptep: fault_pte);
5900 pte = pte_modify(pte: old_pte, newprot: vma->vm_page_prot);
5901 pte = pte_mkyoung(pte);
5902 if (writable)
5903 pte = pte_mkwrite(pte, vma);
5904 ptep_modify_prot_commit(vma, addr: fault_addr, ptep: fault_pte, old_pte, pte);
5905 update_mmu_cache_range(vmf, vma, addr: fault_addr, ptep: fault_pte, nr: 1);
5906}
5907
5908static void numa_rebuild_large_mapping(struct vm_fault *vmf, struct vm_area_struct *vma,
5909 struct folio *folio, pte_t fault_pte,
5910 bool ignore_writable, bool pte_write_upgrade)
5911{
5912 int nr = pte_pfn(pte: fault_pte) - folio_pfn(folio);
5913 unsigned long start, end, addr = vmf->address;
5914 unsigned long addr_start = addr - (nr << PAGE_SHIFT);
5915 unsigned long pt_start = ALIGN_DOWN(addr, PMD_SIZE);
5916 pte_t *start_ptep;
5917
5918 /* Stay within the VMA and within the page table. */
5919 start = max3(addr_start, pt_start, vma->vm_start);
5920 end = min3(addr_start + folio_size(folio), pt_start + PMD_SIZE,
5921 vma->vm_end);
5922 start_ptep = vmf->pte - ((addr - start) >> PAGE_SHIFT);
5923
5924 /* Restore all PTEs' mapping of the large folio */
5925 for (addr = start; addr != end; start_ptep++, addr += PAGE_SIZE) {
5926 pte_t ptent = ptep_get(ptep: start_ptep);
5927 bool writable = false;
5928
5929 if (!pte_present(a: ptent) || !pte_protnone(pte: ptent))
5930 continue;
5931
5932 if (pfn_folio(pfn: pte_pfn(pte: ptent)) != folio)
5933 continue;
5934
5935 if (!ignore_writable) {
5936 ptent = pte_modify(pte: ptent, newprot: vma->vm_page_prot);
5937 writable = pte_write(pte: ptent);
5938 if (!writable && pte_write_upgrade &&
5939 can_change_pte_writable(vma, addr, pte: ptent))
5940 writable = true;
5941 }
5942
5943 numa_rebuild_single_mapping(vmf, vma, fault_addr: addr, fault_pte: start_ptep, writable);
5944 }
5945}
5946
5947static vm_fault_t do_numa_page(struct vm_fault *vmf)
5948{
5949 struct vm_area_struct *vma = vmf->vma;
5950 struct folio *folio = NULL;
5951 int nid = NUMA_NO_NODE;
5952 bool writable = false, ignore_writable = false;
5953 bool pte_write_upgrade = vma_wants_manual_pte_write_upgrade(vma);
5954 int last_cpupid;
5955 int target_nid;
5956 pte_t pte, old_pte;
5957 int flags = 0, nr_pages;
5958
5959 /*
5960 * The pte cannot be used safely until we verify, while holding the page
5961 * table lock, that its contents have not changed during fault handling.
5962 */
5963 spin_lock(lock: vmf->ptl);
5964 /* Read the live PTE from the page tables: */
5965 old_pte = ptep_get(ptep: vmf->pte);
5966
5967 if (unlikely(!pte_same(old_pte, vmf->orig_pte))) {
5968 pte_unmap_unlock(vmf->pte, vmf->ptl);
5969 return 0;
5970 }
5971
5972 pte = pte_modify(pte: old_pte, newprot: vma->vm_page_prot);
5973
5974 /*
5975 * Detect now whether the PTE could be writable; this information
5976 * is only valid while holding the PT lock.
5977 */
5978 writable = pte_write(pte);
5979 if (!writable && pte_write_upgrade &&
5980 can_change_pte_writable(vma, addr: vmf->address, pte))
5981 writable = true;
5982
5983 folio = vm_normal_folio(vma, addr: vmf->address, pte);
5984 if (!folio || folio_is_zone_device(folio))
5985 goto out_map;
5986
5987 nid = folio_nid(folio);
5988 nr_pages = folio_nr_pages(folio);
5989
5990 target_nid = numa_migrate_check(folio, vmf, addr: vmf->address, flags: &flags,
5991 writable, last_cpupid: &last_cpupid);
5992 if (target_nid == NUMA_NO_NODE)
5993 goto out_map;
5994 if (migrate_misplaced_folio_prepare(folio, vma, node: target_nid)) {
5995 flags |= TNF_MIGRATE_FAIL;
5996 goto out_map;
5997 }
5998 /* The folio is isolated and isolation code holds a folio reference. */
5999 pte_unmap_unlock(vmf->pte, vmf->ptl);
6000 writable = false;
6001 ignore_writable = true;
6002
6003 /* Migrate to the requested node */
6004 if (!migrate_misplaced_folio(folio, node: target_nid)) {
6005 nid = target_nid;
6006 flags |= TNF_MIGRATED;
6007 task_numa_fault(last_node: last_cpupid, node: nid, pages: nr_pages, flags);
6008 return 0;
6009 }
6010
6011 flags |= TNF_MIGRATE_FAIL;
6012 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
6013 addr: vmf->address, ptlp: &vmf->ptl);
6014 if (unlikely(!vmf->pte))
6015 return 0;
6016 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
6017 pte_unmap_unlock(vmf->pte, vmf->ptl);
6018 return 0;
6019 }
6020out_map:
6021 /*
6022 * Make it present again, depending on how arch implements
6023 * non-accessible ptes, some can allow access by kernel mode.
6024 */
6025 if (folio && folio_test_large(folio))
6026 numa_rebuild_large_mapping(vmf, vma, folio, fault_pte: pte, ignore_writable,
6027 pte_write_upgrade);
6028 else
6029 numa_rebuild_single_mapping(vmf, vma, fault_addr: vmf->address, fault_pte: vmf->pte,
6030 writable);
6031 pte_unmap_unlock(vmf->pte, vmf->ptl);
6032
6033 if (nid != NUMA_NO_NODE)
6034 task_numa_fault(last_node: last_cpupid, node: nid, pages: nr_pages, flags);
6035 return 0;
6036}
6037
6038static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
6039{
6040 struct vm_area_struct *vma = vmf->vma;
6041 if (vma_is_anonymous(vma))
6042 return do_huge_pmd_anonymous_page(vmf);
6043 if (vma->vm_ops->huge_fault)
6044 return vma->vm_ops->huge_fault(vmf, PMD_ORDER);
6045 return VM_FAULT_FALLBACK;
6046}
6047
6048/* `inline' is required to avoid gcc 4.1.2 build error */
6049static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
6050{
6051 struct vm_area_struct *vma = vmf->vma;
6052 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
6053 vm_fault_t ret;
6054
6055 if (vma_is_anonymous(vma)) {
6056 if (likely(!unshare) &&
6057 userfaultfd_huge_pmd_wp(vma, pmd: vmf->orig_pmd)) {
6058 if (userfaultfd_wp_async(vma: vmf->vma))
6059 goto split;
6060 return handle_userfault(vmf, VM_UFFD_WP);
6061 }
6062 return do_huge_pmd_wp_page(vmf);
6063 }
6064
6065 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
6066 if (vma->vm_ops->huge_fault) {
6067 ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER);
6068 if (!(ret & VM_FAULT_FALLBACK))
6069 return ret;
6070 }
6071 }
6072
6073split:
6074 /* COW or write-notify handled on pte level: split pmd. */
6075 __split_huge_pmd(vma, pmd: vmf->pmd, address: vmf->address, freeze: false);
6076
6077 return VM_FAULT_FALLBACK;
6078}
6079
6080static vm_fault_t create_huge_pud(struct vm_fault *vmf)
6081{
6082#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
6083 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
6084 struct vm_area_struct *vma = vmf->vma;
6085 /* No support for anonymous transparent PUD pages yet */
6086 if (vma_is_anonymous(vma))
6087 return VM_FAULT_FALLBACK;
6088 if (vma->vm_ops->huge_fault)
6089 return vma->vm_ops->huge_fault(vmf, PUD_ORDER);
6090#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
6091 return VM_FAULT_FALLBACK;
6092}
6093
6094static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
6095{
6096#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
6097 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
6098 struct vm_area_struct *vma = vmf->vma;
6099 vm_fault_t ret;
6100
6101 /* No support for anonymous transparent PUD pages yet */
6102 if (vma_is_anonymous(vma))
6103 goto split;
6104 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
6105 if (vma->vm_ops->huge_fault) {
6106 ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER);
6107 if (!(ret & VM_FAULT_FALLBACK))
6108 return ret;
6109 }
6110 }
6111split:
6112 /* COW or write-notify not handled on PUD level: split pud.*/
6113 __split_huge_pud(vma, vmf->pud, vmf->address);
6114#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
6115 return VM_FAULT_FALLBACK;
6116}
6117
6118/*
6119 * These routines also need to handle stuff like marking pages dirty
6120 * and/or accessed for architectures that don't do it in hardware (most
6121 * RISC architectures). The early dirtying is also good on the i386.
6122 *
6123 * There is also a hook called "update_mmu_cache()" that architectures
6124 * with external mmu caches can use to update those (ie the Sparc or
6125 * PowerPC hashed page tables that act as extended TLBs).
6126 *
6127 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
6128 * concurrent faults).
6129 *
6130 * The mmap_lock may have been released depending on flags and our return value.
6131 * See filemap_fault() and __folio_lock_or_retry().
6132 */
6133static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
6134{
6135 pte_t entry;
6136
6137 if (unlikely(pmd_none(*vmf->pmd))) {
6138 /*
6139 * Leave __pte_alloc() until later: because vm_ops->fault may
6140 * want to allocate huge page, and if we expose page table
6141 * for an instant, it will be difficult to retract from
6142 * concurrent faults and from rmap lookups.
6143 */
6144 vmf->pte = NULL;
6145 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
6146 } else {
6147 pmd_t dummy_pmdval;
6148
6149 /*
6150 * A regular pmd is established and it can't morph into a huge
6151 * pmd by anon khugepaged, since that takes mmap_lock in write
6152 * mode; but shmem or file collapse to THP could still morph
6153 * it into a huge pmd: just retry later if so.
6154 *
6155 * Use the maywrite version to indicate that vmf->pte may be
6156 * modified, but since we will use pte_same() to detect the
6157 * change of the !pte_none() entry, there is no need to recheck
6158 * the pmdval. Here we chooes to pass a dummy variable instead
6159 * of NULL, which helps new user think about why this place is
6160 * special.
6161 */
6162 vmf->pte = pte_offset_map_rw_nolock(mm: vmf->vma->vm_mm, pmd: vmf->pmd,
6163 addr: vmf->address, pmdvalp: &dummy_pmdval,
6164 ptlp: &vmf->ptl);
6165 if (unlikely(!vmf->pte))
6166 return 0;
6167 vmf->orig_pte = ptep_get_lockless(ptep: vmf->pte);
6168 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
6169
6170 if (pte_none(pte: vmf->orig_pte)) {
6171 pte_unmap(pte: vmf->pte);
6172 vmf->pte = NULL;
6173 }
6174 }
6175
6176 if (!vmf->pte)
6177 return do_pte_missing(vmf);
6178
6179 if (!pte_present(a: vmf->orig_pte))
6180 return do_swap_page(vmf);
6181
6182 if (pte_protnone(pte: vmf->orig_pte) && vma_is_accessible(vma: vmf->vma))
6183 return do_numa_page(vmf);
6184
6185 spin_lock(lock: vmf->ptl);
6186 entry = vmf->orig_pte;
6187 if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) {
6188 update_mmu_tlb(vma: vmf->vma, address: vmf->address, ptep: vmf->pte);
6189 goto unlock;
6190 }
6191 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6192 if (!pte_write(pte: entry))
6193 return do_wp_page(vmf);
6194 else if (likely(vmf->flags & FAULT_FLAG_WRITE))
6195 entry = pte_mkdirty(pte: entry);
6196 }
6197 entry = pte_mkyoung(pte: entry);
6198 if (ptep_set_access_flags(vma: vmf->vma, address: vmf->address, ptep: vmf->pte, entry,
6199 dirty: vmf->flags & FAULT_FLAG_WRITE)) {
6200 update_mmu_cache_range(vmf, vma: vmf->vma, addr: vmf->address,
6201 ptep: vmf->pte, nr: 1);
6202 } else {
6203 /* Skip spurious TLB flush for retried page fault */
6204 if (vmf->flags & FAULT_FLAG_TRIED)
6205 goto unlock;
6206 /*
6207 * This is needed only for protection faults but the arch code
6208 * is not yet telling us if this is a protection fault or not.
6209 * This still avoids useless tlb flushes for .text page faults
6210 * with threads.
6211 */
6212 if (vmf->flags & FAULT_FLAG_WRITE)
6213 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address,
6214 vmf->pte);
6215 }
6216unlock:
6217 pte_unmap_unlock(vmf->pte, vmf->ptl);
6218 return 0;
6219}
6220
6221/*
6222 * On entry, we hold either the VMA lock or the mmap_lock
6223 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
6224 * the result, the mmap_lock is not held on exit. See filemap_fault()
6225 * and __folio_lock_or_retry().
6226 */
6227static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
6228 unsigned long address, unsigned int flags)
6229{
6230 struct vm_fault vmf = {
6231 .vma = vma,
6232 .address = address & PAGE_MASK,
6233 .real_address = address,
6234 .flags = flags,
6235 .pgoff = linear_page_index(vma, address),
6236 .gfp_mask = __get_fault_gfp_mask(vma),
6237 };
6238 struct mm_struct *mm = vma->vm_mm;
6239 vm_flags_t vm_flags = vma->vm_flags;
6240 pgd_t *pgd;
6241 p4d_t *p4d;
6242 vm_fault_t ret;
6243
6244 pgd = pgd_offset(mm, address);
6245 p4d = p4d_alloc(mm, pgd, address);
6246 if (!p4d)
6247 return VM_FAULT_OOM;
6248
6249 vmf.pud = pud_alloc(mm, p4d, address);
6250 if (!vmf.pud)
6251 return VM_FAULT_OOM;
6252retry_pud:
6253 if (pud_none(pud: *vmf.pud) &&
6254 thp_vma_allowable_order(vma, vm_flags, TVA_PAGEFAULT, PUD_ORDER)) {
6255 ret = create_huge_pud(vmf: &vmf);
6256 if (!(ret & VM_FAULT_FALLBACK))
6257 return ret;
6258 } else {
6259 pud_t orig_pud = *vmf.pud;
6260
6261 barrier();
6262 if (pud_trans_huge(pud: orig_pud)) {
6263
6264 /*
6265 * TODO once we support anonymous PUDs: NUMA case and
6266 * FAULT_FLAG_UNSHARE handling.
6267 */
6268 if ((flags & FAULT_FLAG_WRITE) && !pud_write(pud: orig_pud)) {
6269 ret = wp_huge_pud(vmf: &vmf, orig_pud);
6270 if (!(ret & VM_FAULT_FALLBACK))
6271 return ret;
6272 } else {
6273 huge_pud_set_accessed(vmf: &vmf, orig_pud);
6274 return 0;
6275 }
6276 }
6277 }
6278
6279 vmf.pmd = pmd_alloc(mm, pud: vmf.pud, address);
6280 if (!vmf.pmd)
6281 return VM_FAULT_OOM;
6282
6283 /* Huge pud page fault raced with pmd_alloc? */
6284 if (pud_trans_unstable(pud: vmf.pud))
6285 goto retry_pud;
6286
6287 if (pmd_none(pmd: *vmf.pmd) &&
6288 thp_vma_allowable_order(vma, vm_flags, TVA_PAGEFAULT, PMD_ORDER)) {
6289 ret = create_huge_pmd(vmf: &vmf);
6290 if (!(ret & VM_FAULT_FALLBACK))
6291 return ret;
6292 } else {
6293 vmf.orig_pmd = pmdp_get_lockless(pmdp: vmf.pmd);
6294
6295 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
6296 VM_BUG_ON(thp_migration_supported() &&
6297 !is_pmd_migration_entry(vmf.orig_pmd));
6298 if (is_pmd_migration_entry(pmd: vmf.orig_pmd))
6299 pmd_migration_entry_wait(m: mm, p: vmf.pmd);
6300 return 0;
6301 }
6302 if (pmd_trans_huge(pmd: vmf.orig_pmd)) {
6303 if (pmd_protnone(pmd: vmf.orig_pmd) && vma_is_accessible(vma))
6304 return do_huge_pmd_numa_page(vmf: &vmf);
6305
6306 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6307 !pmd_write(pmd: vmf.orig_pmd)) {
6308 ret = wp_huge_pmd(vmf: &vmf);
6309 if (!(ret & VM_FAULT_FALLBACK))
6310 return ret;
6311 } else {
6312 huge_pmd_set_accessed(vmf: &vmf);
6313 return 0;
6314 }
6315 }
6316 }
6317
6318 return handle_pte_fault(vmf: &vmf);
6319}
6320
6321/**
6322 * mm_account_fault - Do page fault accounting
6323 * @mm: mm from which memcg should be extracted. It can be NULL.
6324 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
6325 * of perf event counters, but we'll still do the per-task accounting to
6326 * the task who triggered this page fault.
6327 * @address: the faulted address.
6328 * @flags: the fault flags.
6329 * @ret: the fault retcode.
6330 *
6331 * This will take care of most of the page fault accounting. Meanwhile, it
6332 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
6333 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
6334 * still be in per-arch page fault handlers at the entry of page fault.
6335 */
6336static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs,
6337 unsigned long address, unsigned int flags,
6338 vm_fault_t ret)
6339{
6340 bool major;
6341
6342 /* Incomplete faults will be accounted upon completion. */
6343 if (ret & VM_FAULT_RETRY)
6344 return;
6345
6346 /*
6347 * To preserve the behavior of older kernels, PGFAULT counters record
6348 * both successful and failed faults, as opposed to perf counters,
6349 * which ignore failed cases.
6350 */
6351 count_vm_event(item: PGFAULT);
6352 count_memcg_event_mm(mm, idx: PGFAULT);
6353
6354 /*
6355 * Do not account for unsuccessful faults (e.g. when the address wasn't
6356 * valid). That includes arch_vma_access_permitted() failing before
6357 * reaching here. So this is not a "this many hardware page faults"
6358 * counter. We should use the hw profiling for that.
6359 */
6360 if (ret & VM_FAULT_ERROR)
6361 return;
6362
6363 /*
6364 * We define the fault as a major fault when the final successful fault
6365 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
6366 * handle it immediately previously).
6367 */
6368 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
6369
6370 if (major)
6371 current->maj_flt++;
6372 else
6373 current->min_flt++;
6374
6375 /*
6376 * If the fault is done for GUP, regs will be NULL. We only do the
6377 * accounting for the per thread fault counters who triggered the
6378 * fault, and we skip the perf event updates.
6379 */
6380 if (!regs)
6381 return;
6382
6383 if (major)
6384 perf_sw_event(event_id: PERF_COUNT_SW_PAGE_FAULTS_MAJ, nr: 1, regs, addr: address);
6385 else
6386 perf_sw_event(event_id: PERF_COUNT_SW_PAGE_FAULTS_MIN, nr: 1, regs, addr: address);
6387}
6388
6389#ifdef CONFIG_LRU_GEN
6390static void lru_gen_enter_fault(struct vm_area_struct *vma)
6391{
6392 /* the LRU algorithm only applies to accesses with recency */
6393 current->in_lru_fault = vma_has_recency(vma);
6394}
6395
6396static void lru_gen_exit_fault(void)
6397{
6398 current->in_lru_fault = false;
6399}
6400#else
6401static void lru_gen_enter_fault(struct vm_area_struct *vma)
6402{
6403}
6404
6405static void lru_gen_exit_fault(void)
6406{
6407}
6408#endif /* CONFIG_LRU_GEN */
6409
6410static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
6411 unsigned int *flags)
6412{
6413 if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
6414 if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
6415 return VM_FAULT_SIGSEGV;
6416 /*
6417 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
6418 * just treat it like an ordinary read-fault otherwise.
6419 */
6420 if (!is_cow_mapping(flags: vma->vm_flags))
6421 *flags &= ~FAULT_FLAG_UNSHARE;
6422 } else if (*flags & FAULT_FLAG_WRITE) {
6423 /* Write faults on read-only mappings are impossible ... */
6424 if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
6425 return VM_FAULT_SIGSEGV;
6426 /* ... and FOLL_FORCE only applies to COW mappings. */
6427 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
6428 !is_cow_mapping(vma->vm_flags)))
6429 return VM_FAULT_SIGSEGV;
6430 }
6431#ifdef CONFIG_PER_VMA_LOCK
6432 /*
6433 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
6434 * the assumption that lock is dropped on VM_FAULT_RETRY.
6435 */
6436 if (WARN_ON_ONCE((*flags &
6437 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) ==
6438 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)))
6439 return VM_FAULT_SIGSEGV;
6440#endif
6441
6442 return 0;
6443}
6444
6445/*
6446 * By the time we get here, we already hold either the VMA lock or the
6447 * mmap_lock (FAULT_FLAG_VMA_LOCK tells you which).
6448 *
6449 * The mmap_lock may have been released depending on flags and our
6450 * return value. See filemap_fault() and __folio_lock_or_retry().
6451 */
6452vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
6453 unsigned int flags, struct pt_regs *regs)
6454{
6455 /* If the fault handler drops the mmap_lock, vma may be freed */
6456 struct mm_struct *mm = vma->vm_mm;
6457 vm_fault_t ret;
6458 bool is_droppable;
6459
6460 __set_current_state(TASK_RUNNING);
6461
6462 ret = sanitize_fault_flags(vma, flags: &flags);
6463 if (ret)
6464 goto out;
6465
6466 if (!arch_vma_access_permitted(vma, write: flags & FAULT_FLAG_WRITE,
6467 execute: flags & FAULT_FLAG_INSTRUCTION,
6468 foreign: flags & FAULT_FLAG_REMOTE)) {
6469 ret = VM_FAULT_SIGSEGV;
6470 goto out;
6471 }
6472
6473 is_droppable = !!(vma->vm_flags & VM_DROPPABLE);
6474
6475 /*
6476 * Enable the memcg OOM handling for faults triggered in user
6477 * space. Kernel faults are handled more gracefully.
6478 */
6479 if (flags & FAULT_FLAG_USER)
6480 mem_cgroup_enter_user_fault();
6481
6482 lru_gen_enter_fault(vma);
6483
6484 if (unlikely(is_vm_hugetlb_page(vma)))
6485 ret = hugetlb_fault(mm: vma->vm_mm, vma, address, flags);
6486 else
6487 ret = __handle_mm_fault(vma, address, flags);
6488
6489 /*
6490 * Warning: It is no longer safe to dereference vma-> after this point,
6491 * because mmap_lock might have been dropped by __handle_mm_fault(), so
6492 * vma might be destroyed from underneath us.
6493 */
6494
6495 lru_gen_exit_fault();
6496
6497 /* If the mapping is droppable, then errors due to OOM aren't fatal. */
6498 if (is_droppable)
6499 ret &= ~VM_FAULT_OOM;
6500
6501 if (flags & FAULT_FLAG_USER) {
6502 mem_cgroup_exit_user_fault();
6503 /*
6504 * The task may have entered a memcg OOM situation but
6505 * if the allocation error was handled gracefully (no
6506 * VM_FAULT_OOM), there is no need to kill anything.
6507 * Just clean up the OOM state peacefully.
6508 */
6509 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
6510 mem_cgroup_oom_synchronize(wait: false);
6511 }
6512out:
6513 mm_account_fault(mm, regs, address, flags, ret);
6514
6515 return ret;
6516}
6517EXPORT_SYMBOL_GPL(handle_mm_fault);
6518
6519#ifndef __PAGETABLE_P4D_FOLDED
6520/*
6521 * Allocate p4d page table.
6522 * We've already handled the fast-path in-line.
6523 */
6524int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
6525{
6526 p4d_t *new = p4d_alloc_one(mm, address);
6527 if (!new)
6528 return -ENOMEM;
6529
6530 spin_lock(lock: &mm->page_table_lock);
6531 if (pgd_present(pgd: *pgd)) { /* Another has populated it */
6532 p4d_free(mm, p4d: new);
6533 } else {
6534 smp_wmb(); /* See comment in pmd_install() */
6535 pgd_populate(mm, pgd, p4d: new);
6536 }
6537 spin_unlock(lock: &mm->page_table_lock);
6538 return 0;
6539}
6540#endif /* __PAGETABLE_P4D_FOLDED */
6541
6542#ifndef __PAGETABLE_PUD_FOLDED
6543/*
6544 * Allocate page upper directory.
6545 * We've already handled the fast-path in-line.
6546 */
6547int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
6548{
6549 pud_t *new = pud_alloc_one(mm, address);
6550 if (!new)
6551 return -ENOMEM;
6552
6553 spin_lock(lock: &mm->page_table_lock);
6554 if (!p4d_present(p4d: *p4d)) {
6555 mm_inc_nr_puds(mm);
6556 smp_wmb(); /* See comment in pmd_install() */
6557 p4d_populate(mm, p4d, pud: new);
6558 } else /* Another has populated it */
6559 pud_free(mm, pud: new);
6560 spin_unlock(lock: &mm->page_table_lock);
6561 return 0;
6562}
6563#endif /* __PAGETABLE_PUD_FOLDED */
6564
6565#ifndef __PAGETABLE_PMD_FOLDED
6566/*
6567 * Allocate page middle directory.
6568 * We've already handled the fast-path in-line.
6569 */
6570int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
6571{
6572 spinlock_t *ptl;
6573 pmd_t *new = pmd_alloc_one(mm, address);
6574 if (!new)
6575 return -ENOMEM;
6576
6577 ptl = pud_lock(mm, pud);
6578 if (!pud_present(pud: *pud)) {
6579 mm_inc_nr_pmds(mm);
6580 smp_wmb(); /* See comment in pmd_install() */
6581 pud_populate(mm, pud, pmd: new);
6582 } else { /* Another has populated it */
6583 pmd_free(mm, pmd: new);
6584 }
6585 spin_unlock(lock: ptl);
6586 return 0;
6587}
6588#endif /* __PAGETABLE_PMD_FOLDED */
6589
6590static inline void pfnmap_args_setup(struct follow_pfnmap_args *args,
6591 spinlock_t *lock, pte_t *ptep,
6592 pgprot_t pgprot, unsigned long pfn_base,
6593 unsigned long addr_mask, bool writable,
6594 bool special)
6595{
6596 args->lock = lock;
6597 args->ptep = ptep;
6598 args->pfn = pfn_base + ((args->address & ~addr_mask) >> PAGE_SHIFT);
6599 args->addr_mask = addr_mask;
6600 args->pgprot = pgprot;
6601 args->writable = writable;
6602 args->special = special;
6603}
6604
6605static inline void pfnmap_lockdep_assert(struct vm_area_struct *vma)
6606{
6607#ifdef CONFIG_LOCKDEP
6608 struct file *file = vma->vm_file;
6609 struct address_space *mapping = file ? file->f_mapping : NULL;
6610
6611 if (mapping)
6612 lockdep_assert(lockdep_is_held(&mapping->i_mmap_rwsem) ||
6613 lockdep_is_held(&vma->vm_mm->mmap_lock));
6614 else
6615 lockdep_assert(lockdep_is_held(&vma->vm_mm->mmap_lock));
6616#endif
6617}
6618
6619/**
6620 * follow_pfnmap_start() - Look up a pfn mapping at a user virtual address
6621 * @args: Pointer to struct @follow_pfnmap_args
6622 *
6623 * The caller needs to setup args->vma and args->address to point to the
6624 * virtual address as the target of such lookup. On a successful return,
6625 * the results will be put into other output fields.
6626 *
6627 * After the caller finished using the fields, the caller must invoke
6628 * another follow_pfnmap_end() to proper releases the locks and resources
6629 * of such look up request.
6630 *
6631 * During the start() and end() calls, the results in @args will be valid
6632 * as proper locks will be held. After the end() is called, all the fields
6633 * in @follow_pfnmap_args will be invalid to be further accessed. Further
6634 * use of such information after end() may require proper synchronizations
6635 * by the caller with page table updates, otherwise it can create a
6636 * security bug.
6637 *
6638 * If the PTE maps a refcounted page, callers are responsible to protect
6639 * against invalidation with MMU notifiers; otherwise access to the PFN at
6640 * a later point in time can trigger use-after-free.
6641 *
6642 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
6643 * should be taken for read, and the mmap semaphore cannot be released
6644 * before the end() is invoked.
6645 *
6646 * This function must not be used to modify PTE content.
6647 *
6648 * Return: zero on success, negative otherwise.
6649 */
6650int follow_pfnmap_start(struct follow_pfnmap_args *args)
6651{
6652 struct vm_area_struct *vma = args->vma;
6653 unsigned long address = args->address;
6654 struct mm_struct *mm = vma->vm_mm;
6655 spinlock_t *lock;
6656 pgd_t *pgdp;
6657 p4d_t *p4dp, p4d;
6658 pud_t *pudp, pud;
6659 pmd_t *pmdp, pmd;
6660 pte_t *ptep, pte;
6661
6662 pfnmap_lockdep_assert(vma);
6663
6664 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
6665 goto out;
6666
6667 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
6668 goto out;
6669retry:
6670 pgdp = pgd_offset(mm, address);
6671 if (pgd_none(pgd: *pgdp) || unlikely(pgd_bad(*pgdp)))
6672 goto out;
6673
6674 p4dp = p4d_offset(pgd: pgdp, address);
6675 p4d = READ_ONCE(*p4dp);
6676 if (p4d_none(p4d) || unlikely(p4d_bad(p4d)))
6677 goto out;
6678
6679 pudp = pud_offset(p4d: p4dp, address);
6680 pud = READ_ONCE(*pudp);
6681 if (pud_none(pud))
6682 goto out;
6683 if (pud_leaf(pud)) {
6684 lock = pud_lock(mm, pud: pudp);
6685 if (!unlikely(pud_leaf(pud))) {
6686 spin_unlock(lock);
6687 goto retry;
6688 }
6689 pfnmap_args_setup(args, lock, NULL, pud_pgprot(pud),
6690 pud_pfn(pud), PUD_MASK, pud_write(pud),
6691 special: pud_special(pud));
6692 return 0;
6693 }
6694
6695 pmdp = pmd_offset(pud: pudp, address);
6696 pmd = pmdp_get_lockless(pmdp);
6697 if (pmd_leaf(pte: pmd)) {
6698 lock = pmd_lock(mm, pmd: pmdp);
6699 if (!unlikely(pmd_leaf(pmd))) {
6700 spin_unlock(lock);
6701 goto retry;
6702 }
6703 pfnmap_args_setup(args, lock, NULL, pmd_pgprot(pmd),
6704 pfn_base: pmd_pfn(pmd), PMD_MASK, pmd_write(pmd),
6705 special: pmd_special(pmd));
6706 return 0;
6707 }
6708
6709 ptep = pte_offset_map_lock(mm, pmd: pmdp, addr: address, ptlp: &lock);
6710 if (!ptep)
6711 goto out;
6712 pte = ptep_get(ptep);
6713 if (!pte_present(a: pte))
6714 goto unlock;
6715 pfnmap_args_setup(args, lock, ptep, pte_pgprot(pte),
6716 pfn_base: pte_pfn(pte), PAGE_MASK, writable: pte_write(pte),
6717 special: pte_special(pte));
6718 return 0;
6719unlock:
6720 pte_unmap_unlock(ptep, lock);
6721out:
6722 return -EINVAL;
6723}
6724EXPORT_SYMBOL_GPL(follow_pfnmap_start);
6725
6726/**
6727 * follow_pfnmap_end(): End a follow_pfnmap_start() process
6728 * @args: Pointer to struct @follow_pfnmap_args
6729 *
6730 * Must be used in pair of follow_pfnmap_start(). See the start() function
6731 * above for more information.
6732 */
6733void follow_pfnmap_end(struct follow_pfnmap_args *args)
6734{
6735 if (args->lock)
6736 spin_unlock(lock: args->lock);
6737 if (args->ptep)
6738 pte_unmap(pte: args->ptep);
6739}
6740EXPORT_SYMBOL_GPL(follow_pfnmap_end);
6741
6742#ifdef CONFIG_HAVE_IOREMAP_PROT
6743/**
6744 * generic_access_phys - generic implementation for iomem mmap access
6745 * @vma: the vma to access
6746 * @addr: userspace address, not relative offset within @vma
6747 * @buf: buffer to read/write
6748 * @len: length of transfer
6749 * @write: set to FOLL_WRITE when writing, otherwise reading
6750 *
6751 * This is a generic implementation for &vm_operations_struct.access for an
6752 * iomem mapping. This callback is used by access_process_vm() when the @vma is
6753 * not page based.
6754 */
6755int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
6756 void *buf, int len, int write)
6757{
6758 resource_size_t phys_addr;
6759 pgprot_t prot = __pgprot(0);
6760 void __iomem *maddr;
6761 int offset = offset_in_page(addr);
6762 int ret = -EINVAL;
6763 bool writable;
6764 struct follow_pfnmap_args args = { .vma = vma, .address = addr };
6765
6766retry:
6767 if (follow_pfnmap_start(&args))
6768 return -EINVAL;
6769 prot = args.pgprot;
6770 phys_addr = (resource_size_t)args.pfn << PAGE_SHIFT;
6771 writable = args.writable;
6772 follow_pfnmap_end(&args);
6773
6774 if ((write & FOLL_WRITE) && !writable)
6775 return -EINVAL;
6776
6777 maddr = ioremap_prot(offset: phys_addr, PAGE_ALIGN(len + offset), prot);
6778 if (!maddr)
6779 return -ENOMEM;
6780
6781 if (follow_pfnmap_start(&args))
6782 goto out_unmap;
6783
6784 if ((pgprot_val(prot) != pgprot_val(args.pgprot)) ||
6785 (phys_addr != (args.pfn << PAGE_SHIFT)) ||
6786 (writable != args.writable)) {
6787 follow_pfnmap_end(&args);
6788 iounmap(addr: maddr);
6789 goto retry;
6790 }
6791
6792 if (write)
6793 memcpy_toio(maddr + offset, buf, len);
6794 else
6795 memcpy_fromio(buf, maddr + offset, len);
6796 ret = len;
6797 follow_pfnmap_end(&args);
6798out_unmap:
6799 iounmap(addr: maddr);
6800
6801 return ret;
6802}
6803EXPORT_SYMBOL_GPL(generic_access_phys);
6804#endif
6805
6806/*
6807 * Access another process' address space as given in mm.
6808 */
6809static int __access_remote_vm(struct mm_struct *mm, unsigned long addr,
6810 void *buf, int len, unsigned int gup_flags)
6811{
6812 void *old_buf = buf;
6813 int write = gup_flags & FOLL_WRITE;
6814
6815 if (mmap_read_lock_killable(mm))
6816 return 0;
6817
6818 /* Untag the address before looking up the VMA */
6819 addr = untagged_addr_remote(mm, addr);
6820
6821 /* Avoid triggering the temporary warning in __get_user_pages */
6822 if (!vma_lookup(mm, addr) && !expand_stack(mm, addr))
6823 return 0;
6824
6825 /* ignore errors, just check how much was successfully transferred */
6826 while (len) {
6827 int bytes, offset;
6828 void *maddr;
6829 struct folio *folio;
6830 struct vm_area_struct *vma = NULL;
6831 struct page *page = get_user_page_vma_remote(mm, addr,
6832 gup_flags, vmap: &vma);
6833
6834 if (IS_ERR(ptr: page)) {
6835 /* We might need to expand the stack to access it */
6836 vma = vma_lookup(mm, addr);
6837 if (!vma) {
6838 vma = expand_stack(mm, addr);
6839
6840 /* mmap_lock was dropped on failure */
6841 if (!vma)
6842 return buf - old_buf;
6843
6844 /* Try again if stack expansion worked */
6845 continue;
6846 }
6847
6848 /*
6849 * Check if this is a VM_IO | VM_PFNMAP VMA, which
6850 * we can access using slightly different code.
6851 */
6852 bytes = 0;
6853#ifdef CONFIG_HAVE_IOREMAP_PROT
6854 if (vma->vm_ops && vma->vm_ops->access)
6855 bytes = vma->vm_ops->access(vma, addr, buf,
6856 len, write);
6857#endif
6858 if (bytes <= 0)
6859 break;
6860 } else {
6861 folio = page_folio(page);
6862 bytes = len;
6863 offset = addr & (PAGE_SIZE-1);
6864 if (bytes > PAGE_SIZE-offset)
6865 bytes = PAGE_SIZE-offset;
6866
6867 maddr = kmap_local_folio(folio, offset: folio_page_idx(folio, page) * PAGE_SIZE);
6868 if (write) {
6869 copy_to_user_page(vma, page, addr,
6870 maddr + offset, buf, bytes);
6871 folio_mark_dirty_lock(folio);
6872 } else {
6873 copy_from_user_page(vma, page, addr,
6874 buf, maddr + offset, bytes);
6875 }
6876 folio_release_kmap(folio, addr: maddr);
6877 }
6878 len -= bytes;
6879 buf += bytes;
6880 addr += bytes;
6881 }
6882 mmap_read_unlock(mm);
6883
6884 return buf - old_buf;
6885}
6886
6887/**
6888 * access_remote_vm - access another process' address space
6889 * @mm: the mm_struct of the target address space
6890 * @addr: start address to access
6891 * @buf: source or destination buffer
6892 * @len: number of bytes to transfer
6893 * @gup_flags: flags modifying lookup behaviour
6894 *
6895 * The caller must hold a reference on @mm.
6896 *
6897 * Return: number of bytes copied from source to destination.
6898 */
6899int access_remote_vm(struct mm_struct *mm, unsigned long addr,
6900 void *buf, int len, unsigned int gup_flags)
6901{
6902 return __access_remote_vm(mm, addr, buf, len, gup_flags);
6903}
6904
6905/*
6906 * Access another process' address space.
6907 * Source/target buffer must be kernel space,
6908 * Do not walk the page table directly, use get_user_pages
6909 */
6910int access_process_vm(struct task_struct *tsk, unsigned long addr,
6911 void *buf, int len, unsigned int gup_flags)
6912{
6913 struct mm_struct *mm;
6914 int ret;
6915
6916 mm = get_task_mm(task: tsk);
6917 if (!mm)
6918 return 0;
6919
6920 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
6921
6922 mmput(mm);
6923
6924 return ret;
6925}
6926EXPORT_SYMBOL_GPL(access_process_vm);
6927
6928#ifdef CONFIG_BPF_SYSCALL
6929/*
6930 * Copy a string from another process's address space as given in mm.
6931 * If there is any error return -EFAULT.
6932 */
6933static int __copy_remote_vm_str(struct mm_struct *mm, unsigned long addr,
6934 void *buf, int len, unsigned int gup_flags)
6935{
6936 void *old_buf = buf;
6937 int err = 0;
6938
6939 *(char *)buf = '\0';
6940
6941 if (mmap_read_lock_killable(mm))
6942 return -EFAULT;
6943
6944 addr = untagged_addr_remote(mm, addr);
6945
6946 /* Avoid triggering the temporary warning in __get_user_pages */
6947 if (!vma_lookup(mm, addr)) {
6948 err = -EFAULT;
6949 goto out;
6950 }
6951
6952 while (len) {
6953 int bytes, offset, retval;
6954 void *maddr;
6955 struct folio *folio;
6956 struct page *page;
6957 struct vm_area_struct *vma = NULL;
6958
6959 page = get_user_page_vma_remote(mm, addr, gup_flags, &vma);
6960 if (IS_ERR(page)) {
6961 /*
6962 * Treat as a total failure for now until we decide how
6963 * to handle the CONFIG_HAVE_IOREMAP_PROT case and
6964 * stack expansion.
6965 */
6966 *(char *)buf = '\0';
6967 err = -EFAULT;
6968 goto out;
6969 }
6970
6971 folio = page_folio(page);
6972 bytes = len;
6973 offset = addr & (PAGE_SIZE - 1);
6974 if (bytes > PAGE_SIZE - offset)
6975 bytes = PAGE_SIZE - offset;
6976
6977 maddr = kmap_local_folio(folio, folio_page_idx(folio, page) * PAGE_SIZE);
6978 retval = strscpy(buf, maddr + offset, bytes);
6979 if (retval >= 0) {
6980 /* Found the end of the string */
6981 buf += retval;
6982 folio_release_kmap(folio, maddr);
6983 break;
6984 }
6985
6986 buf += bytes - 1;
6987 /*
6988 * Because strscpy always NUL terminates we need to
6989 * copy the last byte in the page if we are going to
6990 * load more pages
6991 */
6992 if (bytes != len) {
6993 addr += bytes - 1;
6994 copy_from_user_page(vma, page, addr, buf, maddr + (PAGE_SIZE - 1), 1);
6995 buf += 1;
6996 addr += 1;
6997 }
6998 len -= bytes;
6999
7000 folio_release_kmap(folio, maddr);
7001 }
7002
7003out:
7004 mmap_read_unlock(mm);
7005 if (err)
7006 return err;
7007 return buf - old_buf;
7008}
7009
7010/**
7011 * copy_remote_vm_str - copy a string from another process's address space.
7012 * @tsk: the task of the target address space
7013 * @addr: start address to read from
7014 * @buf: destination buffer
7015 * @len: number of bytes to copy
7016 * @gup_flags: flags modifying lookup behaviour
7017 *
7018 * The caller must hold a reference on @mm.
7019 *
7020 * Return: number of bytes copied from @addr (source) to @buf (destination);
7021 * not including the trailing NUL. Always guaranteed to leave NUL-terminated
7022 * buffer. On any error, return -EFAULT.
7023 */
7024int copy_remote_vm_str(struct task_struct *tsk, unsigned long addr,
7025 void *buf, int len, unsigned int gup_flags)
7026{
7027 struct mm_struct *mm;
7028 int ret;
7029
7030 if (unlikely(len == 0))
7031 return 0;
7032
7033 mm = get_task_mm(tsk);
7034 if (!mm) {
7035 *(char *)buf = '\0';
7036 return -EFAULT;
7037 }
7038
7039 ret = __copy_remote_vm_str(mm, addr, buf, len, gup_flags);
7040
7041 mmput(mm);
7042
7043 return ret;
7044}
7045EXPORT_SYMBOL_GPL(copy_remote_vm_str);
7046#endif /* CONFIG_BPF_SYSCALL */
7047
7048/*
7049 * Print the name of a VMA.
7050 */
7051void print_vma_addr(char *prefix, unsigned long ip)
7052{
7053 struct mm_struct *mm = current->mm;
7054 struct vm_area_struct *vma;
7055
7056 /*
7057 * we might be running from an atomic context so we cannot sleep
7058 */
7059 if (!mmap_read_trylock(mm))
7060 return;
7061
7062 vma = vma_lookup(mm, addr: ip);
7063 if (vma && vma->vm_file) {
7064 struct file *f = vma->vm_file;
7065 ip -= vma->vm_start;
7066 ip += vma->vm_pgoff << PAGE_SHIFT;
7067 printk("%s%pD[%lx,%lx+%lx]", prefix, f, ip,
7068 vma->vm_start,
7069 vma->vm_end - vma->vm_start);
7070 }
7071 mmap_read_unlock(mm);
7072}
7073
7074#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
7075void __might_fault(const char *file, int line)
7076{
7077 if (pagefault_disabled())
7078 return;
7079 __might_sleep(file, line);
7080 if (current->mm)
7081 might_lock_read(&current->mm->mmap_lock);
7082}
7083EXPORT_SYMBOL(__might_fault);
7084#endif
7085
7086#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
7087/*
7088 * Process all subpages of the specified huge page with the specified
7089 * operation. The target subpage will be processed last to keep its
7090 * cache lines hot.
7091 */
7092static inline int process_huge_page(
7093 unsigned long addr_hint, unsigned int nr_pages,
7094 int (*process_subpage)(unsigned long addr, int idx, void *arg),
7095 void *arg)
7096{
7097 int i, n, base, l, ret;
7098 unsigned long addr = addr_hint &
7099 ~(((unsigned long)nr_pages << PAGE_SHIFT) - 1);
7100
7101 /* Process target subpage last to keep its cache lines hot */
7102 might_sleep();
7103 n = (addr_hint - addr) / PAGE_SIZE;
7104 if (2 * n <= nr_pages) {
7105 /* If target subpage in first half of huge page */
7106 base = 0;
7107 l = n;
7108 /* Process subpages at the end of huge page */
7109 for (i = nr_pages - 1; i >= 2 * n; i--) {
7110 cond_resched();
7111 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
7112 if (ret)
7113 return ret;
7114 }
7115 } else {
7116 /* If target subpage in second half of huge page */
7117 base = nr_pages - 2 * (nr_pages - n);
7118 l = nr_pages - n;
7119 /* Process subpages at the begin of huge page */
7120 for (i = 0; i < base; i++) {
7121 cond_resched();
7122 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
7123 if (ret)
7124 return ret;
7125 }
7126 }
7127 /*
7128 * Process remaining subpages in left-right-left-right pattern
7129 * towards the target subpage
7130 */
7131 for (i = 0; i < l; i++) {
7132 int left_idx = base + i;
7133 int right_idx = base + 2 * l - 1 - i;
7134
7135 cond_resched();
7136 ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
7137 if (ret)
7138 return ret;
7139 cond_resched();
7140 ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
7141 if (ret)
7142 return ret;
7143 }
7144 return 0;
7145}
7146
7147static void clear_gigantic_page(struct folio *folio, unsigned long addr_hint,
7148 unsigned int nr_pages)
7149{
7150 unsigned long addr = ALIGN_DOWN(addr_hint, folio_size(folio));
7151 int i;
7152
7153 might_sleep();
7154 for (i = 0; i < nr_pages; i++) {
7155 cond_resched();
7156 clear_user_highpage(folio_page(folio, i), vaddr: addr + i * PAGE_SIZE);
7157 }
7158}
7159
7160static int clear_subpage(unsigned long addr, int idx, void *arg)
7161{
7162 struct folio *folio = arg;
7163
7164 clear_user_highpage(folio_page(folio, idx), vaddr: addr);
7165 return 0;
7166}
7167
7168/**
7169 * folio_zero_user - Zero a folio which will be mapped to userspace.
7170 * @folio: The folio to zero.
7171 * @addr_hint: The address will be accessed or the base address if uncelar.
7172 */
7173void folio_zero_user(struct folio *folio, unsigned long addr_hint)
7174{
7175 unsigned int nr_pages = folio_nr_pages(folio);
7176
7177 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES))
7178 clear_gigantic_page(folio, addr_hint, nr_pages);
7179 else
7180 process_huge_page(addr_hint, nr_pages, process_subpage: clear_subpage, arg: folio);
7181}
7182
7183static int copy_user_gigantic_page(struct folio *dst, struct folio *src,
7184 unsigned long addr_hint,
7185 struct vm_area_struct *vma,
7186 unsigned int nr_pages)
7187{
7188 unsigned long addr = ALIGN_DOWN(addr_hint, folio_size(dst));
7189 struct page *dst_page;
7190 struct page *src_page;
7191 int i;
7192
7193 for (i = 0; i < nr_pages; i++) {
7194 dst_page = folio_page(dst, i);
7195 src_page = folio_page(src, i);
7196
7197 cond_resched();
7198 if (copy_mc_user_highpage(to: dst_page, from: src_page,
7199 vaddr: addr + i*PAGE_SIZE, vma))
7200 return -EHWPOISON;
7201 }
7202 return 0;
7203}
7204
7205struct copy_subpage_arg {
7206 struct folio *dst;
7207 struct folio *src;
7208 struct vm_area_struct *vma;
7209};
7210
7211static int copy_subpage(unsigned long addr, int idx, void *arg)
7212{
7213 struct copy_subpage_arg *copy_arg = arg;
7214 struct page *dst = folio_page(copy_arg->dst, idx);
7215 struct page *src = folio_page(copy_arg->src, idx);
7216
7217 if (copy_mc_user_highpage(to: dst, from: src, vaddr: addr, vma: copy_arg->vma))
7218 return -EHWPOISON;
7219 return 0;
7220}
7221
7222int copy_user_large_folio(struct folio *dst, struct folio *src,
7223 unsigned long addr_hint, struct vm_area_struct *vma)
7224{
7225 unsigned int nr_pages = folio_nr_pages(folio: dst);
7226 struct copy_subpage_arg arg = {
7227 .dst = dst,
7228 .src = src,
7229 .vma = vma,
7230 };
7231
7232 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES))
7233 return copy_user_gigantic_page(dst, src, addr_hint, vma, nr_pages);
7234
7235 return process_huge_page(addr_hint, nr_pages, process_subpage: copy_subpage, arg: &arg);
7236}
7237
7238long copy_folio_from_user(struct folio *dst_folio,
7239 const void __user *usr_src,
7240 bool allow_pagefault)
7241{
7242 void *kaddr;
7243 unsigned long i, rc = 0;
7244 unsigned int nr_pages = folio_nr_pages(folio: dst_folio);
7245 unsigned long ret_val = nr_pages * PAGE_SIZE;
7246 struct page *subpage;
7247
7248 for (i = 0; i < nr_pages; i++) {
7249 subpage = folio_page(dst_folio, i);
7250 kaddr = kmap_local_page(page: subpage);
7251 if (!allow_pagefault)
7252 pagefault_disable();
7253 rc = copy_from_user(to: kaddr, from: usr_src + i * PAGE_SIZE, PAGE_SIZE);
7254 if (!allow_pagefault)
7255 pagefault_enable();
7256 kunmap_local(kaddr);
7257
7258 ret_val -= (PAGE_SIZE - rc);
7259 if (rc)
7260 break;
7261
7262 flush_dcache_page(page: subpage);
7263
7264 cond_resched();
7265 }
7266 return ret_val;
7267}
7268#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
7269
7270#if defined(CONFIG_SPLIT_PTE_PTLOCKS) && ALLOC_SPLIT_PTLOCKS
7271
7272static struct kmem_cache *page_ptl_cachep;
7273
7274void __init ptlock_cache_init(void)
7275{
7276 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
7277 SLAB_PANIC, NULL);
7278}
7279
7280bool ptlock_alloc(struct ptdesc *ptdesc)
7281{
7282 spinlock_t *ptl;
7283
7284 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
7285 if (!ptl)
7286 return false;
7287 ptdesc->ptl = ptl;
7288 return true;
7289}
7290
7291void ptlock_free(struct ptdesc *ptdesc)
7292{
7293 if (ptdesc->ptl)
7294 kmem_cache_free(page_ptl_cachep, ptdesc->ptl);
7295}
7296#endif
7297
7298void vma_pgtable_walk_begin(struct vm_area_struct *vma)
7299{
7300 if (is_vm_hugetlb_page(vma))
7301 hugetlb_vma_lock_read(vma);
7302}
7303
7304void vma_pgtable_walk_end(struct vm_area_struct *vma)
7305{
7306 if (is_vm_hugetlb_page(vma))
7307 hugetlb_vma_unlock_read(vma);
7308}
7309