1
2// SPDX-License-Identifier: GPL-2.0-only
3/*
4 * linux/mm/memory.c
5 *
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 */
8
9/*
10 * demand-loading started 01.12.91 - seems it is high on the list of
11 * things wanted, and it should be easy to implement. - Linus
12 */
13
14/*
15 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
16 * pages started 02.12.91, seems to work. - Linus.
17 *
18 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
19 * would have taken more than the 6M I have free, but it worked well as
20 * far as I could see.
21 *
22 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
23 */
24
25/*
26 * Real VM (paging to/from disk) started 18.12.91. Much more work and
27 * thought has to go into this. Oh, well..
28 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
29 * Found it. Everything seems to work now.
30 * 20.12.91 - Ok, making the swap-device changeable like the root.
31 */
32
33/*
34 * 05.04.94 - Multi-page memory management added for v1.1.
35 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 *
37 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
38 * (Gerhard.Wichert@pdb.siemens.de)
39 *
40 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 */
42
43#include <linux/kernel_stat.h>
44#include <linux/mm.h>
45#include <linux/mm_inline.h>
46#include <linux/sched/mm.h>
47#include <linux/sched/coredump.h>
48#include <linux/sched/numa_balancing.h>
49#include <linux/sched/task.h>
50#include <linux/hugetlb.h>
51#include <linux/mman.h>
52#include <linux/swap.h>
53#include <linux/highmem.h>
54#include <linux/pagemap.h>
55#include <linux/memremap.h>
56#include <linux/kmsan.h>
57#include <linux/ksm.h>
58#include <linux/rmap.h>
59#include <linux/export.h>
60#include <linux/delayacct.h>
61#include <linux/init.h>
62#include <linux/pfn_t.h>
63#include <linux/writeback.h>
64#include <linux/memcontrol.h>
65#include <linux/mmu_notifier.h>
66#include <linux/swapops.h>
67#include <linux/elf.h>
68#include <linux/gfp.h>
69#include <linux/migrate.h>
70#include <linux/string.h>
71#include <linux/memory-tiers.h>
72#include <linux/debugfs.h>
73#include <linux/userfaultfd_k.h>
74#include <linux/dax.h>
75#include <linux/oom.h>
76#include <linux/numa.h>
77#include <linux/perf_event.h>
78#include <linux/ptrace.h>
79#include <linux/vmalloc.h>
80#include <linux/sched/sysctl.h>
81
82#include <trace/events/kmem.h>
83
84#include <asm/io.h>
85#include <asm/mmu_context.h>
86#include <asm/pgalloc.h>
87#include <linux/uaccess.h>
88#include <asm/tlb.h>
89#include <asm/tlbflush.h>
90
91#include "pgalloc-track.h"
92#include "internal.h"
93#include "swap.h"
94
95#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
96#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
97#endif
98
99#ifndef CONFIG_NUMA
100unsigned long max_mapnr;
101EXPORT_SYMBOL(max_mapnr);
102
103struct page *mem_map;
104EXPORT_SYMBOL(mem_map);
105#endif
106
107static vm_fault_t do_fault(struct vm_fault *vmf);
108static vm_fault_t do_anonymous_page(struct vm_fault *vmf);
109static bool vmf_pte_changed(struct vm_fault *vmf);
110
111/*
112 * Return true if the original pte was a uffd-wp pte marker (so the pte was
113 * wr-protected).
114 */
115static bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
116{
117 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
118 return false;
119
120 return pte_marker_uffd_wp(pte: vmf->orig_pte);
121}
122
123/*
124 * A number of key systems in x86 including ioremap() rely on the assumption
125 * that high_memory defines the upper bound on direct map memory, then end
126 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
127 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
128 * and ZONE_HIGHMEM.
129 */
130void *high_memory;
131EXPORT_SYMBOL(high_memory);
132
133/*
134 * Randomize the address space (stacks, mmaps, brk, etc.).
135 *
136 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
137 * as ancient (libc5 based) binaries can segfault. )
138 */
139int randomize_va_space __read_mostly =
140#ifdef CONFIG_COMPAT_BRK
141 1;
142#else
143 2;
144#endif
145
146#ifndef arch_wants_old_prefaulted_pte
147static inline bool arch_wants_old_prefaulted_pte(void)
148{
149 /*
150 * Transitioning a PTE from 'old' to 'young' can be expensive on
151 * some architectures, even if it's performed in hardware. By
152 * default, "false" means prefaulted entries will be 'young'.
153 */
154 return false;
155}
156#endif
157
158static int __init disable_randmaps(char *s)
159{
160 randomize_va_space = 0;
161 return 1;
162}
163__setup("norandmaps", disable_randmaps);
164
165unsigned long zero_pfn __read_mostly;
166EXPORT_SYMBOL(zero_pfn);
167
168unsigned long highest_memmap_pfn __read_mostly;
169
170/*
171 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
172 */
173static int __init init_zero_pfn(void)
174{
175 zero_pfn = page_to_pfn(ZERO_PAGE(0));
176 return 0;
177}
178early_initcall(init_zero_pfn);
179
180void mm_trace_rss_stat(struct mm_struct *mm, int member)
181{
182 trace_rss_stat(mm, member);
183}
184
185/*
186 * Note: this doesn't free the actual pages themselves. That
187 * has been handled earlier when unmapping all the memory regions.
188 */
189static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
190 unsigned long addr)
191{
192 pgtable_t token = pmd_pgtable(*pmd);
193 pmd_clear(pmdp: pmd);
194 pte_free_tlb(tlb, token, addr);
195 mm_dec_nr_ptes(mm: tlb->mm);
196}
197
198static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
199 unsigned long addr, unsigned long end,
200 unsigned long floor, unsigned long ceiling)
201{
202 pmd_t *pmd;
203 unsigned long next;
204 unsigned long start;
205
206 start = addr;
207 pmd = pmd_offset(pud, address: addr);
208 do {
209 next = pmd_addr_end(addr, end);
210 if (pmd_none_or_clear_bad(pmd))
211 continue;
212 free_pte_range(tlb, pmd, addr);
213 } while (pmd++, addr = next, addr != end);
214
215 start &= PUD_MASK;
216 if (start < floor)
217 return;
218 if (ceiling) {
219 ceiling &= PUD_MASK;
220 if (!ceiling)
221 return;
222 }
223 if (end - 1 > ceiling - 1)
224 return;
225
226 pmd = pmd_offset(pud, address: start);
227 pud_clear(pudp: pud);
228 pmd_free_tlb(tlb, pmd, start);
229 mm_dec_nr_pmds(mm: tlb->mm);
230}
231
232static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
233 unsigned long addr, unsigned long end,
234 unsigned long floor, unsigned long ceiling)
235{
236 pud_t *pud;
237 unsigned long next;
238 unsigned long start;
239
240 start = addr;
241 pud = pud_offset(p4d, address: addr);
242 do {
243 next = pud_addr_end(addr, end);
244 if (pud_none_or_clear_bad(pud))
245 continue;
246 free_pmd_range(tlb, pud, addr, end: next, floor, ceiling);
247 } while (pud++, addr = next, addr != end);
248
249 start &= P4D_MASK;
250 if (start < floor)
251 return;
252 if (ceiling) {
253 ceiling &= P4D_MASK;
254 if (!ceiling)
255 return;
256 }
257 if (end - 1 > ceiling - 1)
258 return;
259
260 pud = pud_offset(p4d, address: start);
261 p4d_clear(p4dp: p4d);
262 pud_free_tlb(tlb, pud, start);
263 mm_dec_nr_puds(mm: tlb->mm);
264}
265
266static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
267 unsigned long addr, unsigned long end,
268 unsigned long floor, unsigned long ceiling)
269{
270 p4d_t *p4d;
271 unsigned long next;
272 unsigned long start;
273
274 start = addr;
275 p4d = p4d_offset(pgd, address: addr);
276 do {
277 next = p4d_addr_end(addr, end);
278 if (p4d_none_or_clear_bad(p4d))
279 continue;
280 free_pud_range(tlb, p4d, addr, end: next, floor, ceiling);
281 } while (p4d++, addr = next, addr != end);
282
283 start &= PGDIR_MASK;
284 if (start < floor)
285 return;
286 if (ceiling) {
287 ceiling &= PGDIR_MASK;
288 if (!ceiling)
289 return;
290 }
291 if (end - 1 > ceiling - 1)
292 return;
293
294 p4d = p4d_offset(pgd, address: start);
295 pgd_clear(pgd);
296 p4d_free_tlb(tlb, p4d, start);
297}
298
299/*
300 * This function frees user-level page tables of a process.
301 */
302void free_pgd_range(struct mmu_gather *tlb,
303 unsigned long addr, unsigned long end,
304 unsigned long floor, unsigned long ceiling)
305{
306 pgd_t *pgd;
307 unsigned long next;
308
309 /*
310 * The next few lines have given us lots of grief...
311 *
312 * Why are we testing PMD* at this top level? Because often
313 * there will be no work to do at all, and we'd prefer not to
314 * go all the way down to the bottom just to discover that.
315 *
316 * Why all these "- 1"s? Because 0 represents both the bottom
317 * of the address space and the top of it (using -1 for the
318 * top wouldn't help much: the masks would do the wrong thing).
319 * The rule is that addr 0 and floor 0 refer to the bottom of
320 * the address space, but end 0 and ceiling 0 refer to the top
321 * Comparisons need to use "end - 1" and "ceiling - 1" (though
322 * that end 0 case should be mythical).
323 *
324 * Wherever addr is brought up or ceiling brought down, we must
325 * be careful to reject "the opposite 0" before it confuses the
326 * subsequent tests. But what about where end is brought down
327 * by PMD_SIZE below? no, end can't go down to 0 there.
328 *
329 * Whereas we round start (addr) and ceiling down, by different
330 * masks at different levels, in order to test whether a table
331 * now has no other vmas using it, so can be freed, we don't
332 * bother to round floor or end up - the tests don't need that.
333 */
334
335 addr &= PMD_MASK;
336 if (addr < floor) {
337 addr += PMD_SIZE;
338 if (!addr)
339 return;
340 }
341 if (ceiling) {
342 ceiling &= PMD_MASK;
343 if (!ceiling)
344 return;
345 }
346 if (end - 1 > ceiling - 1)
347 end -= PMD_SIZE;
348 if (addr > end - 1)
349 return;
350 /*
351 * We add page table cache pages with PAGE_SIZE,
352 * (see pte_free_tlb()), flush the tlb if we need
353 */
354 tlb_change_page_size(tlb, PAGE_SIZE);
355 pgd = pgd_offset(tlb->mm, addr);
356 do {
357 next = pgd_addr_end(addr, end);
358 if (pgd_none_or_clear_bad(pgd))
359 continue;
360 free_p4d_range(tlb, pgd, addr, end: next, floor, ceiling);
361 } while (pgd++, addr = next, addr != end);
362}
363
364void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
365 struct vm_area_struct *vma, unsigned long floor,
366 unsigned long ceiling, bool mm_wr_locked)
367{
368 do {
369 unsigned long addr = vma->vm_start;
370 struct vm_area_struct *next;
371
372 /*
373 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
374 * be 0. This will underflow and is okay.
375 */
376 next = mas_find(mas, max: ceiling - 1);
377
378 /*
379 * Hide vma from rmap and truncate_pagecache before freeing
380 * pgtables
381 */
382 if (mm_wr_locked)
383 vma_start_write(vma);
384 unlink_anon_vmas(vma);
385 unlink_file_vma(vma);
386
387 if (is_vm_hugetlb_page(vma)) {
388 hugetlb_free_pgd_range(tlb, addr, end: vma->vm_end,
389 floor, ceiling: next ? next->vm_start : ceiling);
390 } else {
391 /*
392 * Optimization: gather nearby vmas into one call down
393 */
394 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
395 && !is_vm_hugetlb_page(vma: next)) {
396 vma = next;
397 next = mas_find(mas, max: ceiling - 1);
398 if (mm_wr_locked)
399 vma_start_write(vma);
400 unlink_anon_vmas(vma);
401 unlink_file_vma(vma);
402 }
403 free_pgd_range(tlb, addr, end: vma->vm_end,
404 floor, ceiling: next ? next->vm_start : ceiling);
405 }
406 vma = next;
407 } while (vma);
408}
409
410void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
411{
412 spinlock_t *ptl = pmd_lock(mm, pmd);
413
414 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
415 mm_inc_nr_ptes(mm);
416 /*
417 * Ensure all pte setup (eg. pte page lock and page clearing) are
418 * visible before the pte is made visible to other CPUs by being
419 * put into page tables.
420 *
421 * The other side of the story is the pointer chasing in the page
422 * table walking code (when walking the page table without locking;
423 * ie. most of the time). Fortunately, these data accesses consist
424 * of a chain of data-dependent loads, meaning most CPUs (alpha
425 * being the notable exception) will already guarantee loads are
426 * seen in-order. See the alpha page table accessors for the
427 * smp_rmb() barriers in page table walking code.
428 */
429 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
430 pmd_populate(mm, pmd, pte: *pte);
431 *pte = NULL;
432 }
433 spin_unlock(lock: ptl);
434}
435
436int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
437{
438 pgtable_t new = pte_alloc_one(mm);
439 if (!new)
440 return -ENOMEM;
441
442 pmd_install(mm, pmd, pte: &new);
443 if (new)
444 pte_free(mm, pte_page: new);
445 return 0;
446}
447
448int __pte_alloc_kernel(pmd_t *pmd)
449{
450 pte_t *new = pte_alloc_one_kernel(mm: &init_mm);
451 if (!new)
452 return -ENOMEM;
453
454 spin_lock(lock: &init_mm.page_table_lock);
455 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
456 smp_wmb(); /* See comment in pmd_install() */
457 pmd_populate_kernel(mm: &init_mm, pmd, pte: new);
458 new = NULL;
459 }
460 spin_unlock(lock: &init_mm.page_table_lock);
461 if (new)
462 pte_free_kernel(mm: &init_mm, pte: new);
463 return 0;
464}
465
466static inline void init_rss_vec(int *rss)
467{
468 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
469}
470
471static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
472{
473 int i;
474
475 for (i = 0; i < NR_MM_COUNTERS; i++)
476 if (rss[i])
477 add_mm_counter(mm, member: i, value: rss[i]);
478}
479
480/*
481 * This function is called to print an error when a bad pte
482 * is found. For example, we might have a PFN-mapped pte in
483 * a region that doesn't allow it.
484 *
485 * The calling function must still handle the error.
486 */
487static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
488 pte_t pte, struct page *page)
489{
490 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
491 p4d_t *p4d = p4d_offset(pgd, address: addr);
492 pud_t *pud = pud_offset(p4d, address: addr);
493 pmd_t *pmd = pmd_offset(pud, address: addr);
494 struct address_space *mapping;
495 pgoff_t index;
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;
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
519 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
520 index = linear_page_index(vma, address: addr);
521
522 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
523 current->comm,
524 (long long)pte_val(pte), (long long)pmd_val(*pmd));
525 if (page)
526 dump_page(page, reason: "bad pte");
527 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
528 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
529 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
530 vma->vm_file,
531 vma->vm_ops ? vma->vm_ops->fault : NULL,
532 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
533 mapping ? mapping->a_ops->read_folio : NULL);
534 dump_stack();
535 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
536}
537
538/*
539 * vm_normal_page -- This function gets the "struct page" associated with a pte.
540 *
541 * "Special" mappings do not wish to be associated with a "struct page" (either
542 * it doesn't exist, or it exists but they don't want to touch it). In this
543 * case, NULL is returned here. "Normal" mappings do have a struct page.
544 *
545 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
546 * pte bit, in which case this function is trivial. Secondly, an architecture
547 * may not have a spare pte bit, which requires a more complicated scheme,
548 * described below.
549 *
550 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
551 * special mapping (even if there are underlying and valid "struct pages").
552 * COWed pages of a VM_PFNMAP are always normal.
553 *
554 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
555 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
556 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
557 * mapping will always honor the rule
558 *
559 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
560 *
561 * And for normal mappings this is false.
562 *
563 * This restricts such mappings to be a linear translation from virtual address
564 * to pfn. To get around this restriction, we allow arbitrary mappings so long
565 * as the vma is not a COW mapping; in that case, we know that all ptes are
566 * special (because none can have been COWed).
567 *
568 *
569 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
570 *
571 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
572 * page" backing, however the difference is that _all_ pages with a struct
573 * page (that is, those where pfn_valid is true) are refcounted and considered
574 * normal pages by the VM. The disadvantage is that pages are refcounted
575 * (which can be slower and simply not an option for some PFNMAP users). The
576 * advantage is that we don't have to follow the strict linearity rule of
577 * PFNMAP mappings in order to support COWable mappings.
578 *
579 */
580struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
581 pte_t pte)
582{
583 unsigned long pfn = pte_pfn(pte);
584
585 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
586 if (likely(!pte_special(pte)))
587 goto check_pfn;
588 if (vma->vm_ops && vma->vm_ops->find_special_page)
589 return vma->vm_ops->find_special_page(vma, addr);
590 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
591 return NULL;
592 if (is_zero_pfn(pfn))
593 return NULL;
594 if (pte_devmap(a: pte))
595 /*
596 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
597 * and will have refcounts incremented on their struct pages
598 * when they are inserted into PTEs, thus they are safe to
599 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
600 * do not have refcounts. Example of legacy ZONE_DEVICE is
601 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
602 */
603 return NULL;
604
605 print_bad_pte(vma, addr, pte, NULL);
606 return NULL;
607 }
608
609 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
610
611 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
612 if (vma->vm_flags & VM_MIXEDMAP) {
613 if (!pfn_valid(pfn))
614 return NULL;
615 goto out;
616 } else {
617 unsigned long off;
618 off = (addr - vma->vm_start) >> PAGE_SHIFT;
619 if (pfn == vma->vm_pgoff + off)
620 return NULL;
621 if (!is_cow_mapping(flags: vma->vm_flags))
622 return NULL;
623 }
624 }
625
626 if (is_zero_pfn(pfn))
627 return NULL;
628
629check_pfn:
630 if (unlikely(pfn > highest_memmap_pfn)) {
631 print_bad_pte(vma, addr, pte, NULL);
632 return NULL;
633 }
634
635 /*
636 * NOTE! We still have PageReserved() pages in the page tables.
637 * eg. VDSO mappings can cause them to exist.
638 */
639out:
640 return pfn_to_page(pfn);
641}
642
643struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr,
644 pte_t pte)
645{
646 struct page *page = vm_normal_page(vma, addr, pte);
647
648 if (page)
649 return page_folio(page);
650 return NULL;
651}
652
653#ifdef CONFIG_TRANSPARENT_HUGEPAGE
654struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
655 pmd_t pmd)
656{
657 unsigned long pfn = pmd_pfn(pmd);
658
659 /*
660 * There is no pmd_special() but there may be special pmds, e.g.
661 * in a direct-access (dax) mapping, so let's just replicate the
662 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
663 */
664 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
665 if (vma->vm_flags & VM_MIXEDMAP) {
666 if (!pfn_valid(pfn))
667 return NULL;
668 goto out;
669 } else {
670 unsigned long off;
671 off = (addr - vma->vm_start) >> PAGE_SHIFT;
672 if (pfn == vma->vm_pgoff + off)
673 return NULL;
674 if (!is_cow_mapping(flags: vma->vm_flags))
675 return NULL;
676 }
677 }
678
679 if (pmd_devmap(pmd))
680 return NULL;
681 if (is_huge_zero_pmd(pmd))
682 return NULL;
683 if (unlikely(pfn > highest_memmap_pfn))
684 return NULL;
685
686 /*
687 * NOTE! We still have PageReserved() pages in the page tables.
688 * eg. VDSO mappings can cause them to exist.
689 */
690out:
691 return pfn_to_page(pfn);
692}
693
694struct folio *vm_normal_folio_pmd(struct vm_area_struct *vma,
695 unsigned long addr, pmd_t pmd)
696{
697 struct page *page = vm_normal_page_pmd(vma, addr, pmd);
698
699 if (page)
700 return page_folio(page);
701 return NULL;
702}
703#endif
704
705static void restore_exclusive_pte(struct vm_area_struct *vma,
706 struct page *page, unsigned long address,
707 pte_t *ptep)
708{
709 pte_t orig_pte;
710 pte_t pte;
711 swp_entry_t entry;
712
713 orig_pte = ptep_get(ptep);
714 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
715 if (pte_swp_soft_dirty(pte: orig_pte))
716 pte = pte_mksoft_dirty(pte);
717
718 entry = pte_to_swp_entry(pte: orig_pte);
719 if (pte_swp_uffd_wp(pte: orig_pte))
720 pte = pte_mkuffd_wp(pte);
721 else if (is_writable_device_exclusive_entry(entry))
722 pte = maybe_mkwrite(pte: pte_mkdirty(pte), vma);
723
724 VM_BUG_ON(pte_write(pte) && !(PageAnon(page) && PageAnonExclusive(page)));
725
726 /*
727 * No need to take a page reference as one was already
728 * created when the swap entry was made.
729 */
730 if (PageAnon(page))
731 page_add_anon_rmap(page, vma, address, RMAP_NONE);
732 else
733 /*
734 * Currently device exclusive access only supports anonymous
735 * memory so the entry shouldn't point to a filebacked page.
736 */
737 WARN_ON_ONCE(1);
738
739 set_pte_at(vma->vm_mm, address, ptep, pte);
740
741 /*
742 * No need to invalidate - it was non-present before. However
743 * secondary CPUs may have mappings that need invalidating.
744 */
745 update_mmu_cache(vma, addr: address, ptep);
746}
747
748/*
749 * Tries to restore an exclusive pte if the page lock can be acquired without
750 * sleeping.
751 */
752static int
753try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
754 unsigned long addr)
755{
756 swp_entry_t entry = pte_to_swp_entry(pte: ptep_get(ptep: src_pte));
757 struct page *page = pfn_swap_entry_to_page(entry);
758
759 if (trylock_page(page)) {
760 restore_exclusive_pte(vma, page, address: addr, ptep: src_pte);
761 unlock_page(page);
762 return 0;
763 }
764
765 return -EBUSY;
766}
767
768/*
769 * copy one vm_area from one task to the other. Assumes the page tables
770 * already present in the new task to be cleared in the whole range
771 * covered by this vma.
772 */
773
774static unsigned long
775copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
776 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
777 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
778{
779 unsigned long vm_flags = dst_vma->vm_flags;
780 pte_t orig_pte = ptep_get(ptep: src_pte);
781 pte_t pte = orig_pte;
782 struct page *page;
783 swp_entry_t entry = pte_to_swp_entry(pte: orig_pte);
784
785 if (likely(!non_swap_entry(entry))) {
786 if (swap_duplicate(entry) < 0)
787 return -EIO;
788
789 /* make sure dst_mm is on swapoff's mmlist. */
790 if (unlikely(list_empty(&dst_mm->mmlist))) {
791 spin_lock(lock: &mmlist_lock);
792 if (list_empty(head: &dst_mm->mmlist))
793 list_add(new: &dst_mm->mmlist,
794 head: &src_mm->mmlist);
795 spin_unlock(lock: &mmlist_lock);
796 }
797 /* Mark the swap entry as shared. */
798 if (pte_swp_exclusive(pte: orig_pte)) {
799 pte = pte_swp_clear_exclusive(pte: orig_pte);
800 set_pte_at(src_mm, addr, src_pte, pte);
801 }
802 rss[MM_SWAPENTS]++;
803 } else if (is_migration_entry(entry)) {
804 page = pfn_swap_entry_to_page(entry);
805
806 rss[mm_counter(page)]++;
807
808 if (!is_readable_migration_entry(entry) &&
809 is_cow_mapping(flags: vm_flags)) {
810 /*
811 * COW mappings require pages in both parent and child
812 * to be set to read. A previously exclusive entry is
813 * now shared.
814 */
815 entry = make_readable_migration_entry(
816 offset: swp_offset(entry));
817 pte = swp_entry_to_pte(entry);
818 if (pte_swp_soft_dirty(pte: orig_pte))
819 pte = pte_swp_mksoft_dirty(pte);
820 if (pte_swp_uffd_wp(pte: orig_pte))
821 pte = pte_swp_mkuffd_wp(pte);
822 set_pte_at(src_mm, addr, src_pte, pte);
823 }
824 } else if (is_device_private_entry(entry)) {
825 page = pfn_swap_entry_to_page(entry);
826
827 /*
828 * Update rss count even for unaddressable pages, as
829 * they should treated just like normal pages in this
830 * respect.
831 *
832 * We will likely want to have some new rss counters
833 * for unaddressable pages, at some point. But for now
834 * keep things as they are.
835 */
836 get_page(page);
837 rss[mm_counter(page)]++;
838 /* Cannot fail as these pages cannot get pinned. */
839 BUG_ON(page_try_dup_anon_rmap(page, false, src_vma));
840
841 /*
842 * We do not preserve soft-dirty information, because so
843 * far, checkpoint/restore is the only feature that
844 * requires that. And checkpoint/restore does not work
845 * when a device driver is involved (you cannot easily
846 * save and restore device driver state).
847 */
848 if (is_writable_device_private_entry(entry) &&
849 is_cow_mapping(flags: vm_flags)) {
850 entry = make_readable_device_private_entry(
851 offset: swp_offset(entry));
852 pte = swp_entry_to_pte(entry);
853 if (pte_swp_uffd_wp(pte: orig_pte))
854 pte = pte_swp_mkuffd_wp(pte);
855 set_pte_at(src_mm, addr, src_pte, pte);
856 }
857 } else if (is_device_exclusive_entry(entry)) {
858 /*
859 * Make device exclusive entries present by restoring the
860 * original entry then copying as for a present pte. Device
861 * exclusive entries currently only support private writable
862 * (ie. COW) mappings.
863 */
864 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
865 if (try_restore_exclusive_pte(src_pte, vma: src_vma, addr))
866 return -EBUSY;
867 return -ENOENT;
868 } else if (is_pte_marker_entry(entry)) {
869 pte_marker marker = copy_pte_marker(entry, dst_vma);
870
871 if (marker)
872 set_pte_at(dst_mm, addr, dst_pte,
873 make_pte_marker(marker));
874 return 0;
875 }
876 if (!userfaultfd_wp(vma: dst_vma))
877 pte = pte_swp_clear_uffd_wp(pte);
878 set_pte_at(dst_mm, addr, dst_pte, pte);
879 return 0;
880}
881
882/*
883 * Copy a present and normal page.
884 *
885 * NOTE! The usual case is that this isn't required;
886 * instead, the caller can just increase the page refcount
887 * and re-use the pte the traditional way.
888 *
889 * And if we need a pre-allocated page but don't yet have
890 * one, return a negative error to let the preallocation
891 * code know so that it can do so outside the page table
892 * lock.
893 */
894static inline int
895copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
896 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
897 struct folio **prealloc, struct page *page)
898{
899 struct folio *new_folio;
900 pte_t pte;
901
902 new_folio = *prealloc;
903 if (!new_folio)
904 return -EAGAIN;
905
906 /*
907 * We have a prealloc page, all good! Take it
908 * over and copy the page & arm it.
909 */
910 *prealloc = NULL;
911 copy_user_highpage(to: &new_folio->page, from: page, vaddr: addr, vma: src_vma);
912 __folio_mark_uptodate(folio: new_folio);
913 folio_add_new_anon_rmap(new_folio, dst_vma, address: addr);
914 folio_add_lru_vma(new_folio, dst_vma);
915 rss[MM_ANONPAGES]++;
916
917 /* All done, just insert the new page copy in the child */
918 pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot);
919 pte = maybe_mkwrite(pte: pte_mkdirty(pte), vma: dst_vma);
920 if (userfaultfd_pte_wp(vma: dst_vma, pte: ptep_get(ptep: src_pte)))
921 /* Uffd-wp needs to be delivered to dest pte as well */
922 pte = pte_mkuffd_wp(pte);
923 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
924 return 0;
925}
926
927/*
928 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
929 * is required to copy this pte.
930 */
931static inline int
932copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
933 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
934 struct folio **prealloc)
935{
936 struct mm_struct *src_mm = src_vma->vm_mm;
937 unsigned long vm_flags = src_vma->vm_flags;
938 pte_t pte = ptep_get(ptep: src_pte);
939 struct page *page;
940 struct folio *folio;
941
942 page = vm_normal_page(vma: src_vma, addr, pte);
943 if (page)
944 folio = page_folio(page);
945 if (page && folio_test_anon(folio)) {
946 /*
947 * If this page may have been pinned by the parent process,
948 * copy the page immediately for the child so that we'll always
949 * guarantee the pinned page won't be randomly replaced in the
950 * future.
951 */
952 folio_get(folio);
953 if (unlikely(page_try_dup_anon_rmap(page, false, src_vma))) {
954 /* Page may be pinned, we have to copy. */
955 folio_put(folio);
956 return copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
957 addr, rss, prealloc, page);
958 }
959 rss[MM_ANONPAGES]++;
960 } else if (page) {
961 folio_get(folio);
962 page_dup_file_rmap(page, compound: false);
963 rss[mm_counter_file(page)]++;
964 }
965
966 /*
967 * If it's a COW mapping, write protect it both
968 * in the parent and the child
969 */
970 if (is_cow_mapping(flags: vm_flags) && pte_write(pte)) {
971 ptep_set_wrprotect(mm: src_mm, addr, ptep: src_pte);
972 pte = pte_wrprotect(pte);
973 }
974 VM_BUG_ON(page && folio_test_anon(folio) && PageAnonExclusive(page));
975
976 /*
977 * If it's a shared mapping, mark it clean in
978 * the child
979 */
980 if (vm_flags & VM_SHARED)
981 pte = pte_mkclean(pte);
982 pte = pte_mkold(pte);
983
984 if (!userfaultfd_wp(vma: dst_vma))
985 pte = pte_clear_uffd_wp(pte);
986
987 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
988 return 0;
989}
990
991static inline struct folio *page_copy_prealloc(struct mm_struct *src_mm,
992 struct vm_area_struct *vma, unsigned long addr)
993{
994 struct folio *new_folio;
995
996 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, order: 0, vma, addr, hugepage: false);
997 if (!new_folio)
998 return NULL;
999
1000 if (mem_cgroup_charge(folio: new_folio, mm: src_mm, GFP_KERNEL)) {
1001 folio_put(folio: new_folio);
1002 return NULL;
1003 }
1004 folio_throttle_swaprate(folio: new_folio, GFP_KERNEL);
1005
1006 return new_folio;
1007}
1008
1009static int
1010copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1011 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1012 unsigned long end)
1013{
1014 struct mm_struct *dst_mm = dst_vma->vm_mm;
1015 struct mm_struct *src_mm = src_vma->vm_mm;
1016 pte_t *orig_src_pte, *orig_dst_pte;
1017 pte_t *src_pte, *dst_pte;
1018 pte_t ptent;
1019 spinlock_t *src_ptl, *dst_ptl;
1020 int progress, ret = 0;
1021 int rss[NR_MM_COUNTERS];
1022 swp_entry_t entry = (swp_entry_t){0};
1023 struct folio *prealloc = NULL;
1024
1025again:
1026 progress = 0;
1027 init_rss_vec(rss);
1028
1029 /*
1030 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1031 * error handling here, assume that exclusive mmap_lock on dst and src
1032 * protects anon from unexpected THP transitions; with shmem and file
1033 * protected by mmap_lock-less collapse skipping areas with anon_vma
1034 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1035 * can remove such assumptions later, but this is good enough for now.
1036 */
1037 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1038 if (!dst_pte) {
1039 ret = -ENOMEM;
1040 goto out;
1041 }
1042 src_pte = pte_offset_map_nolock(mm: src_mm, pmd: src_pmd, addr, ptlp: &src_ptl);
1043 if (!src_pte) {
1044 pte_unmap_unlock(dst_pte, dst_ptl);
1045 /* ret == 0 */
1046 goto out;
1047 }
1048 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1049 orig_src_pte = src_pte;
1050 orig_dst_pte = dst_pte;
1051 arch_enter_lazy_mmu_mode();
1052
1053 do {
1054 /*
1055 * We are holding two locks at this point - either of them
1056 * could generate latencies in another task on another CPU.
1057 */
1058 if (progress >= 32) {
1059 progress = 0;
1060 if (need_resched() ||
1061 spin_needbreak(lock: src_ptl) || spin_needbreak(lock: dst_ptl))
1062 break;
1063 }
1064 ptent = ptep_get(ptep: src_pte);
1065 if (pte_none(pte: ptent)) {
1066 progress++;
1067 continue;
1068 }
1069 if (unlikely(!pte_present(ptent))) {
1070 ret = copy_nonpresent_pte(dst_mm, src_mm,
1071 dst_pte, src_pte,
1072 dst_vma, src_vma,
1073 addr, rss);
1074 if (ret == -EIO) {
1075 entry = pte_to_swp_entry(pte: ptep_get(ptep: src_pte));
1076 break;
1077 } else if (ret == -EBUSY) {
1078 break;
1079 } else if (!ret) {
1080 progress += 8;
1081 continue;
1082 }
1083
1084 /*
1085 * Device exclusive entry restored, continue by copying
1086 * the now present pte.
1087 */
1088 WARN_ON_ONCE(ret != -ENOENT);
1089 }
1090 /* copy_present_pte() will clear `*prealloc' if consumed */
1091 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1092 addr, rss, prealloc: &prealloc);
1093 /*
1094 * If we need a pre-allocated page for this pte, drop the
1095 * locks, allocate, and try again.
1096 */
1097 if (unlikely(ret == -EAGAIN))
1098 break;
1099 if (unlikely(prealloc)) {
1100 /*
1101 * pre-alloc page cannot be reused by next time so as
1102 * to strictly follow mempolicy (e.g., alloc_page_vma()
1103 * will allocate page according to address). This
1104 * could only happen if one pinned pte changed.
1105 */
1106 folio_put(folio: prealloc);
1107 prealloc = NULL;
1108 }
1109 progress += 8;
1110 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1111
1112 arch_leave_lazy_mmu_mode();
1113 pte_unmap_unlock(orig_src_pte, src_ptl);
1114 add_mm_rss_vec(mm: dst_mm, rss);
1115 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1116 cond_resched();
1117
1118 if (ret == -EIO) {
1119 VM_WARN_ON_ONCE(!entry.val);
1120 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1121 ret = -ENOMEM;
1122 goto out;
1123 }
1124 entry.val = 0;
1125 } else if (ret == -EBUSY) {
1126 goto out;
1127 } else if (ret == -EAGAIN) {
1128 prealloc = page_copy_prealloc(src_mm, vma: src_vma, addr);
1129 if (!prealloc)
1130 return -ENOMEM;
1131 } else if (ret) {
1132 VM_WARN_ON_ONCE(1);
1133 }
1134
1135 /* We've captured and resolved the error. Reset, try again. */
1136 ret = 0;
1137
1138 if (addr != end)
1139 goto again;
1140out:
1141 if (unlikely(prealloc))
1142 folio_put(folio: prealloc);
1143 return ret;
1144}
1145
1146static inline int
1147copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1148 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1149 unsigned long end)
1150{
1151 struct mm_struct *dst_mm = dst_vma->vm_mm;
1152 struct mm_struct *src_mm = src_vma->vm_mm;
1153 pmd_t *src_pmd, *dst_pmd;
1154 unsigned long next;
1155
1156 dst_pmd = pmd_alloc(mm: dst_mm, pud: dst_pud, address: addr);
1157 if (!dst_pmd)
1158 return -ENOMEM;
1159 src_pmd = pmd_offset(pud: src_pud, address: addr);
1160 do {
1161 next = pmd_addr_end(addr, end);
1162 if (is_swap_pmd(pmd: *src_pmd) || pmd_trans_huge(pmd: *src_pmd)
1163 || pmd_devmap(pmd: *src_pmd)) {
1164 int err;
1165 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1166 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1167 addr, dst_vma, src_vma);
1168 if (err == -ENOMEM)
1169 return -ENOMEM;
1170 if (!err)
1171 continue;
1172 /* fall through */
1173 }
1174 if (pmd_none_or_clear_bad(pmd: src_pmd))
1175 continue;
1176 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1177 addr, end: next))
1178 return -ENOMEM;
1179 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1180 return 0;
1181}
1182
1183static inline int
1184copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1185 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1186 unsigned long end)
1187{
1188 struct mm_struct *dst_mm = dst_vma->vm_mm;
1189 struct mm_struct *src_mm = src_vma->vm_mm;
1190 pud_t *src_pud, *dst_pud;
1191 unsigned long next;
1192
1193 dst_pud = pud_alloc(mm: dst_mm, p4d: dst_p4d, address: addr);
1194 if (!dst_pud)
1195 return -ENOMEM;
1196 src_pud = pud_offset(p4d: src_p4d, address: addr);
1197 do {
1198 next = pud_addr_end(addr, end);
1199 if (pud_trans_huge(pud: *src_pud) || pud_devmap(pud: *src_pud)) {
1200 int err;
1201
1202 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1203 err = copy_huge_pud(dst_mm, src_mm,
1204 dst_pud, src_pud, addr, vma: src_vma);
1205 if (err == -ENOMEM)
1206 return -ENOMEM;
1207 if (!err)
1208 continue;
1209 /* fall through */
1210 }
1211 if (pud_none_or_clear_bad(pud: src_pud))
1212 continue;
1213 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1214 addr, end: next))
1215 return -ENOMEM;
1216 } while (dst_pud++, src_pud++, addr = next, addr != end);
1217 return 0;
1218}
1219
1220static inline int
1221copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1222 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1223 unsigned long end)
1224{
1225 struct mm_struct *dst_mm = dst_vma->vm_mm;
1226 p4d_t *src_p4d, *dst_p4d;
1227 unsigned long next;
1228
1229 dst_p4d = p4d_alloc(mm: dst_mm, pgd: dst_pgd, address: addr);
1230 if (!dst_p4d)
1231 return -ENOMEM;
1232 src_p4d = p4d_offset(pgd: src_pgd, address: addr);
1233 do {
1234 next = p4d_addr_end(addr, end);
1235 if (p4d_none_or_clear_bad(p4d: src_p4d))
1236 continue;
1237 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1238 addr, end: next))
1239 return -ENOMEM;
1240 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1241 return 0;
1242}
1243
1244/*
1245 * Return true if the vma needs to copy the pgtable during this fork(). Return
1246 * false when we can speed up fork() by allowing lazy page faults later until
1247 * when the child accesses the memory range.
1248 */
1249static bool
1250vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1251{
1252 /*
1253 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1254 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1255 * contains uffd-wp protection information, that's something we can't
1256 * retrieve from page cache, and skip copying will lose those info.
1257 */
1258 if (userfaultfd_wp(vma: dst_vma))
1259 return true;
1260
1261 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1262 return true;
1263
1264 if (src_vma->anon_vma)
1265 return true;
1266
1267 /*
1268 * Don't copy ptes where a page fault will fill them correctly. Fork
1269 * becomes much lighter when there are big shared or private readonly
1270 * mappings. The tradeoff is that copy_page_range is more efficient
1271 * than faulting.
1272 */
1273 return false;
1274}
1275
1276int
1277copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1278{
1279 pgd_t *src_pgd, *dst_pgd;
1280 unsigned long next;
1281 unsigned long addr = src_vma->vm_start;
1282 unsigned long end = src_vma->vm_end;
1283 struct mm_struct *dst_mm = dst_vma->vm_mm;
1284 struct mm_struct *src_mm = src_vma->vm_mm;
1285 struct mmu_notifier_range range;
1286 bool is_cow;
1287 int ret;
1288
1289 if (!vma_needs_copy(dst_vma, src_vma))
1290 return 0;
1291
1292 if (is_vm_hugetlb_page(vma: src_vma))
1293 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1294
1295 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1296 /*
1297 * We do not free on error cases below as remove_vma
1298 * gets called on error from higher level routine
1299 */
1300 ret = track_pfn_copy(vma: src_vma);
1301 if (ret)
1302 return ret;
1303 }
1304
1305 /*
1306 * We need to invalidate the secondary MMU mappings only when
1307 * there could be a permission downgrade on the ptes of the
1308 * parent mm. And a permission downgrade will only happen if
1309 * is_cow_mapping() returns true.
1310 */
1311 is_cow = is_cow_mapping(flags: src_vma->vm_flags);
1312
1313 if (is_cow) {
1314 mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_PROTECTION_PAGE,
1315 flags: 0, mm: src_mm, start: addr, end);
1316 mmu_notifier_invalidate_range_start(range: &range);
1317 /*
1318 * Disabling preemption is not needed for the write side, as
1319 * the read side doesn't spin, but goes to the mmap_lock.
1320 *
1321 * Use the raw variant of the seqcount_t write API to avoid
1322 * lockdep complaining about preemptibility.
1323 */
1324 vma_assert_write_locked(vma: src_vma);
1325 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1326 }
1327
1328 ret = 0;
1329 dst_pgd = pgd_offset(dst_mm, addr);
1330 src_pgd = pgd_offset(src_mm, addr);
1331 do {
1332 next = pgd_addr_end(addr, end);
1333 if (pgd_none_or_clear_bad(pgd: src_pgd))
1334 continue;
1335 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1336 addr, next))) {
1337 untrack_pfn_clear(vma: dst_vma);
1338 ret = -ENOMEM;
1339 break;
1340 }
1341 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1342
1343 if (is_cow) {
1344 raw_write_seqcount_end(&src_mm->write_protect_seq);
1345 mmu_notifier_invalidate_range_end(range: &range);
1346 }
1347 return ret;
1348}
1349
1350/* Whether we should zap all COWed (private) pages too */
1351static inline bool should_zap_cows(struct zap_details *details)
1352{
1353 /* By default, zap all pages */
1354 if (!details)
1355 return true;
1356
1357 /* Or, we zap COWed pages only if the caller wants to */
1358 return details->even_cows;
1359}
1360
1361/* Decides whether we should zap this page with the page pointer specified */
1362static inline bool should_zap_page(struct zap_details *details, struct page *page)
1363{
1364 /* If we can make a decision without *page.. */
1365 if (should_zap_cows(details))
1366 return true;
1367
1368 /* E.g. the caller passes NULL for the case of a zero page */
1369 if (!page)
1370 return true;
1371
1372 /* Otherwise we should only zap non-anon pages */
1373 return !PageAnon(page);
1374}
1375
1376static inline bool zap_drop_file_uffd_wp(struct zap_details *details)
1377{
1378 if (!details)
1379 return false;
1380
1381 return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1382}
1383
1384/*
1385 * This function makes sure that we'll replace the none pte with an uffd-wp
1386 * swap special pte marker when necessary. Must be with the pgtable lock held.
1387 */
1388static inline void
1389zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1390 unsigned long addr, pte_t *pte,
1391 struct zap_details *details, pte_t pteval)
1392{
1393 /* Zap on anonymous always means dropping everything */
1394 if (vma_is_anonymous(vma))
1395 return;
1396
1397 if (zap_drop_file_uffd_wp(details))
1398 return;
1399
1400 pte_install_uffd_wp_if_needed(vma, addr, pte, pteval);
1401}
1402
1403static unsigned long zap_pte_range(struct mmu_gather *tlb,
1404 struct vm_area_struct *vma, pmd_t *pmd,
1405 unsigned long addr, unsigned long end,
1406 struct zap_details *details)
1407{
1408 struct mm_struct *mm = tlb->mm;
1409 int force_flush = 0;
1410 int rss[NR_MM_COUNTERS];
1411 spinlock_t *ptl;
1412 pte_t *start_pte;
1413 pte_t *pte;
1414 swp_entry_t entry;
1415
1416 tlb_change_page_size(tlb, PAGE_SIZE);
1417 init_rss_vec(rss);
1418 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, ptlp: &ptl);
1419 if (!pte)
1420 return addr;
1421
1422 flush_tlb_batched_pending(mm);
1423 arch_enter_lazy_mmu_mode();
1424 do {
1425 pte_t ptent = ptep_get(ptep: pte);
1426 struct page *page;
1427
1428 if (pte_none(pte: ptent))
1429 continue;
1430
1431 if (need_resched())
1432 break;
1433
1434 if (pte_present(a: ptent)) {
1435 unsigned int delay_rmap;
1436
1437 page = vm_normal_page(vma, addr, pte: ptent);
1438 if (unlikely(!should_zap_page(details, page)))
1439 continue;
1440 ptent = ptep_get_and_clear_full(mm, addr, ptep: pte,
1441 full: tlb->fullmm);
1442 arch_check_zapped_pte(vma, pte: ptent);
1443 tlb_remove_tlb_entry(tlb, pte, addr);
1444 zap_install_uffd_wp_if_needed(vma, addr, pte, details,
1445 pteval: ptent);
1446 if (unlikely(!page)) {
1447 ksm_might_unmap_zero_page(mm, pte: ptent);
1448 continue;
1449 }
1450
1451 delay_rmap = 0;
1452 if (!PageAnon(page)) {
1453 if (pte_dirty(pte: ptent)) {
1454 set_page_dirty(page);
1455 if (tlb_delay_rmap(tlb)) {
1456 delay_rmap = 1;
1457 force_flush = 1;
1458 }
1459 }
1460 if (pte_young(pte: ptent) && likely(vma_has_recency(vma)))
1461 mark_page_accessed(page);
1462 }
1463 rss[mm_counter(page)]--;
1464 if (!delay_rmap) {
1465 page_remove_rmap(page, vma, compound: false);
1466 if (unlikely(page_mapcount(page) < 0))
1467 print_bad_pte(vma, addr, pte: ptent, page);
1468 }
1469 if (unlikely(__tlb_remove_page(tlb, page, delay_rmap))) {
1470 force_flush = 1;
1471 addr += PAGE_SIZE;
1472 break;
1473 }
1474 continue;
1475 }
1476
1477 entry = pte_to_swp_entry(pte: ptent);
1478 if (is_device_private_entry(entry) ||
1479 is_device_exclusive_entry(entry)) {
1480 page = pfn_swap_entry_to_page(entry);
1481 if (unlikely(!should_zap_page(details, page)))
1482 continue;
1483 /*
1484 * Both device private/exclusive mappings should only
1485 * work with anonymous page so far, so we don't need to
1486 * consider uffd-wp bit when zap. For more information,
1487 * see zap_install_uffd_wp_if_needed().
1488 */
1489 WARN_ON_ONCE(!vma_is_anonymous(vma));
1490 rss[mm_counter(page)]--;
1491 if (is_device_private_entry(entry))
1492 page_remove_rmap(page, vma, compound: false);
1493 put_page(page);
1494 } else if (!non_swap_entry(entry)) {
1495 /* Genuine swap entry, hence a private anon page */
1496 if (!should_zap_cows(details))
1497 continue;
1498 rss[MM_SWAPENTS]--;
1499 if (unlikely(!free_swap_and_cache(entry)))
1500 print_bad_pte(vma, addr, pte: ptent, NULL);
1501 } else if (is_migration_entry(entry)) {
1502 page = pfn_swap_entry_to_page(entry);
1503 if (!should_zap_page(details, page))
1504 continue;
1505 rss[mm_counter(page)]--;
1506 } else if (pte_marker_entry_uffd_wp(entry)) {
1507 /*
1508 * For anon: always drop the marker; for file: only
1509 * drop the marker if explicitly requested.
1510 */
1511 if (!vma_is_anonymous(vma) &&
1512 !zap_drop_file_uffd_wp(details))
1513 continue;
1514 } else if (is_hwpoison_entry(entry) ||
1515 is_poisoned_swp_entry(entry)) {
1516 if (!should_zap_cows(details))
1517 continue;
1518 } else {
1519 /* We should have covered all the swap entry types */
1520 WARN_ON_ONCE(1);
1521 }
1522 pte_clear_not_present_full(mm, address: addr, ptep: pte, full: tlb->fullmm);
1523 zap_install_uffd_wp_if_needed(vma, addr, pte, details, pteval: ptent);
1524 } while (pte++, addr += PAGE_SIZE, addr != end);
1525
1526 add_mm_rss_vec(mm, rss);
1527 arch_leave_lazy_mmu_mode();
1528
1529 /* Do the actual TLB flush before dropping ptl */
1530 if (force_flush) {
1531 tlb_flush_mmu_tlbonly(tlb);
1532 tlb_flush_rmaps(tlb, vma);
1533 }
1534 pte_unmap_unlock(start_pte, ptl);
1535
1536 /*
1537 * If we forced a TLB flush (either due to running out of
1538 * batch buffers or because we needed to flush dirty TLB
1539 * entries before releasing the ptl), free the batched
1540 * memory too. Come back again if we didn't do everything.
1541 */
1542 if (force_flush)
1543 tlb_flush_mmu(tlb);
1544
1545 return addr;
1546}
1547
1548static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1549 struct vm_area_struct *vma, pud_t *pud,
1550 unsigned long addr, unsigned long end,
1551 struct zap_details *details)
1552{
1553 pmd_t *pmd;
1554 unsigned long next;
1555
1556 pmd = pmd_offset(pud, address: addr);
1557 do {
1558 next = pmd_addr_end(addr, end);
1559 if (is_swap_pmd(pmd: *pmd) || pmd_trans_huge(pmd: *pmd) || pmd_devmap(pmd: *pmd)) {
1560 if (next - addr != HPAGE_PMD_SIZE)
1561 __split_huge_pmd(vma, pmd, address: addr, freeze: false, NULL);
1562 else if (zap_huge_pmd(tlb, vma, pmd, addr)) {
1563 addr = next;
1564 continue;
1565 }
1566 /* fall through */
1567 } else if (details && details->single_folio &&
1568 folio_test_pmd_mappable(folio: details->single_folio) &&
1569 next - addr == HPAGE_PMD_SIZE && pmd_none(pmd: *pmd)) {
1570 spinlock_t *ptl = pmd_lock(mm: tlb->mm, pmd);
1571 /*
1572 * Take and drop THP pmd lock so that we cannot return
1573 * prematurely, while zap_huge_pmd() has cleared *pmd,
1574 * but not yet decremented compound_mapcount().
1575 */
1576 spin_unlock(lock: ptl);
1577 }
1578 if (pmd_none(pmd: *pmd)) {
1579 addr = next;
1580 continue;
1581 }
1582 addr = zap_pte_range(tlb, vma, pmd, addr, end: next, details);
1583 if (addr != next)
1584 pmd--;
1585 } while (pmd++, cond_resched(), addr != end);
1586
1587 return addr;
1588}
1589
1590static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1591 struct vm_area_struct *vma, p4d_t *p4d,
1592 unsigned long addr, unsigned long end,
1593 struct zap_details *details)
1594{
1595 pud_t *pud;
1596 unsigned long next;
1597
1598 pud = pud_offset(p4d, address: addr);
1599 do {
1600 next = pud_addr_end(addr, end);
1601 if (pud_trans_huge(pud: *pud) || pud_devmap(pud: *pud)) {
1602 if (next - addr != HPAGE_PUD_SIZE) {
1603 mmap_assert_locked(mm: tlb->mm);
1604 split_huge_pud(vma, pud, addr);
1605 } else if (zap_huge_pud(tlb, vma, pud, addr))
1606 goto next;
1607 /* fall through */
1608 }
1609 if (pud_none_or_clear_bad(pud))
1610 continue;
1611 next = zap_pmd_range(tlb, vma, pud, addr, end: next, details);
1612next:
1613 cond_resched();
1614 } while (pud++, addr = next, addr != end);
1615
1616 return addr;
1617}
1618
1619static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1620 struct vm_area_struct *vma, pgd_t *pgd,
1621 unsigned long addr, unsigned long end,
1622 struct zap_details *details)
1623{
1624 p4d_t *p4d;
1625 unsigned long next;
1626
1627 p4d = p4d_offset(pgd, address: addr);
1628 do {
1629 next = p4d_addr_end(addr, end);
1630 if (p4d_none_or_clear_bad(p4d))
1631 continue;
1632 next = zap_pud_range(tlb, vma, p4d, addr, end: next, details);
1633 } while (p4d++, addr = next, addr != end);
1634
1635 return addr;
1636}
1637
1638void unmap_page_range(struct mmu_gather *tlb,
1639 struct vm_area_struct *vma,
1640 unsigned long addr, unsigned long end,
1641 struct zap_details *details)
1642{
1643 pgd_t *pgd;
1644 unsigned long next;
1645
1646 BUG_ON(addr >= end);
1647 tlb_start_vma(tlb, vma);
1648 pgd = pgd_offset(vma->vm_mm, addr);
1649 do {
1650 next = pgd_addr_end(addr, end);
1651 if (pgd_none_or_clear_bad(pgd))
1652 continue;
1653 next = zap_p4d_range(tlb, vma, pgd, addr, end: next, details);
1654 } while (pgd++, addr = next, addr != end);
1655 tlb_end_vma(tlb, vma);
1656}
1657
1658
1659static void unmap_single_vma(struct mmu_gather *tlb,
1660 struct vm_area_struct *vma, unsigned long start_addr,
1661 unsigned long end_addr,
1662 struct zap_details *details, bool mm_wr_locked)
1663{
1664 unsigned long start = max(vma->vm_start, start_addr);
1665 unsigned long end;
1666
1667 if (start >= vma->vm_end)
1668 return;
1669 end = min(vma->vm_end, end_addr);
1670 if (end <= vma->vm_start)
1671 return;
1672
1673 if (vma->vm_file)
1674 uprobe_munmap(vma, start, end);
1675
1676 if (unlikely(vma->vm_flags & VM_PFNMAP))
1677 untrack_pfn(vma, pfn: 0, size: 0, mm_wr_locked);
1678
1679 if (start != end) {
1680 if (unlikely(is_vm_hugetlb_page(vma))) {
1681 /*
1682 * It is undesirable to test vma->vm_file as it
1683 * should be non-null for valid hugetlb area.
1684 * However, vm_file will be NULL in the error
1685 * cleanup path of mmap_region. When
1686 * hugetlbfs ->mmap method fails,
1687 * mmap_region() nullifies vma->vm_file
1688 * before calling this function to clean up.
1689 * Since no pte has actually been setup, it is
1690 * safe to do nothing in this case.
1691 */
1692 if (vma->vm_file) {
1693 zap_flags_t zap_flags = details ?
1694 details->zap_flags : 0;
1695 __unmap_hugepage_range(tlb, vma, start, end,
1696 NULL, zap_flags);
1697 }
1698 } else
1699 unmap_page_range(tlb, vma, addr: start, end, details);
1700 }
1701}
1702
1703/**
1704 * unmap_vmas - unmap a range of memory covered by a list of vma's
1705 * @tlb: address of the caller's struct mmu_gather
1706 * @mas: the maple state
1707 * @vma: the starting vma
1708 * @start_addr: virtual address at which to start unmapping
1709 * @end_addr: virtual address at which to end unmapping
1710 * @tree_end: The maximum index to check
1711 * @mm_wr_locked: lock flag
1712 *
1713 * Unmap all pages in the vma list.
1714 *
1715 * Only addresses between `start' and `end' will be unmapped.
1716 *
1717 * The VMA list must be sorted in ascending virtual address order.
1718 *
1719 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1720 * range after unmap_vmas() returns. So the only responsibility here is to
1721 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1722 * drops the lock and schedules.
1723 */
1724void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas,
1725 struct vm_area_struct *vma, unsigned long start_addr,
1726 unsigned long end_addr, unsigned long tree_end,
1727 bool mm_wr_locked)
1728{
1729 struct mmu_notifier_range range;
1730 struct zap_details details = {
1731 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
1732 /* Careful - we need to zap private pages too! */
1733 .even_cows = true,
1734 };
1735
1736 mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_UNMAP, flags: 0, mm: vma->vm_mm,
1737 start: start_addr, end: end_addr);
1738 mmu_notifier_invalidate_range_start(range: &range);
1739 do {
1740 unsigned long start = start_addr;
1741 unsigned long end = end_addr;
1742 hugetlb_zap_begin(vma, start: &start, end: &end);
1743 unmap_single_vma(tlb, vma, start_addr: start, end_addr: end, details: &details,
1744 mm_wr_locked);
1745 hugetlb_zap_end(vma, details: &details);
1746 } while ((vma = mas_find(mas, max: tree_end - 1)) != NULL);
1747 mmu_notifier_invalidate_range_end(range: &range);
1748}
1749
1750/**
1751 * zap_page_range_single - remove user pages in a given range
1752 * @vma: vm_area_struct holding the applicable pages
1753 * @address: starting address of pages to zap
1754 * @size: number of bytes to zap
1755 * @details: details of shared cache invalidation
1756 *
1757 * The range must fit into one VMA.
1758 */
1759void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1760 unsigned long size, struct zap_details *details)
1761{
1762 const unsigned long end = address + size;
1763 struct mmu_notifier_range range;
1764 struct mmu_gather tlb;
1765
1766 lru_add_drain();
1767 mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_CLEAR, flags: 0, mm: vma->vm_mm,
1768 start: address, end);
1769 hugetlb_zap_begin(vma, start: &range.start, end: &range.end);
1770 tlb_gather_mmu(tlb: &tlb, mm: vma->vm_mm);
1771 update_hiwater_rss(mm: vma->vm_mm);
1772 mmu_notifier_invalidate_range_start(range: &range);
1773 /*
1774 * unmap 'address-end' not 'range.start-range.end' as range
1775 * could have been expanded for hugetlb pmd sharing.
1776 */
1777 unmap_single_vma(tlb: &tlb, vma, start_addr: address, end_addr: end, details, mm_wr_locked: false);
1778 mmu_notifier_invalidate_range_end(range: &range);
1779 tlb_finish_mmu(tlb: &tlb);
1780 hugetlb_zap_end(vma, details);
1781}
1782
1783/**
1784 * zap_vma_ptes - remove ptes mapping the vma
1785 * @vma: vm_area_struct holding ptes to be zapped
1786 * @address: starting address of pages to zap
1787 * @size: number of bytes to zap
1788 *
1789 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1790 *
1791 * The entire address range must be fully contained within the vma.
1792 *
1793 */
1794void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1795 unsigned long size)
1796{
1797 if (!range_in_vma(vma, start: address, end: address + size) ||
1798 !(vma->vm_flags & VM_PFNMAP))
1799 return;
1800
1801 zap_page_range_single(vma, address, size, NULL);
1802}
1803EXPORT_SYMBOL_GPL(zap_vma_ptes);
1804
1805static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1806{
1807 pgd_t *pgd;
1808 p4d_t *p4d;
1809 pud_t *pud;
1810 pmd_t *pmd;
1811
1812 pgd = pgd_offset(mm, addr);
1813 p4d = p4d_alloc(mm, pgd, address: addr);
1814 if (!p4d)
1815 return NULL;
1816 pud = pud_alloc(mm, p4d, address: addr);
1817 if (!pud)
1818 return NULL;
1819 pmd = pmd_alloc(mm, pud, address: addr);
1820 if (!pmd)
1821 return NULL;
1822
1823 VM_BUG_ON(pmd_trans_huge(*pmd));
1824 return pmd;
1825}
1826
1827pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1828 spinlock_t **ptl)
1829{
1830 pmd_t *pmd = walk_to_pmd(mm, addr);
1831
1832 if (!pmd)
1833 return NULL;
1834 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1835}
1836
1837static int validate_page_before_insert(struct page *page)
1838{
1839 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1840 return -EINVAL;
1841 flush_dcache_page(page);
1842 return 0;
1843}
1844
1845static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1846 unsigned long addr, struct page *page, pgprot_t prot)
1847{
1848 if (!pte_none(pte: ptep_get(ptep: pte)))
1849 return -EBUSY;
1850 /* Ok, finally just insert the thing.. */
1851 get_page(page);
1852 inc_mm_counter(mm: vma->vm_mm, member: mm_counter_file(page));
1853 page_add_file_rmap(page, vma, compound: false);
1854 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
1855 return 0;
1856}
1857
1858/*
1859 * This is the old fallback for page remapping.
1860 *
1861 * For historical reasons, it only allows reserved pages. Only
1862 * old drivers should use this, and they needed to mark their
1863 * pages reserved for the old functions anyway.
1864 */
1865static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1866 struct page *page, pgprot_t prot)
1867{
1868 int retval;
1869 pte_t *pte;
1870 spinlock_t *ptl;
1871
1872 retval = validate_page_before_insert(page);
1873 if (retval)
1874 goto out;
1875 retval = -ENOMEM;
1876 pte = get_locked_pte(mm: vma->vm_mm, addr, ptl: &ptl);
1877 if (!pte)
1878 goto out;
1879 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
1880 pte_unmap_unlock(pte, ptl);
1881out:
1882 return retval;
1883}
1884
1885static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
1886 unsigned long addr, struct page *page, pgprot_t prot)
1887{
1888 int err;
1889
1890 if (!page_count(page))
1891 return -EINVAL;
1892 err = validate_page_before_insert(page);
1893 if (err)
1894 return err;
1895 return insert_page_into_pte_locked(vma, pte, addr, page, prot);
1896}
1897
1898/* insert_pages() amortizes the cost of spinlock operations
1899 * when inserting pages in a loop.
1900 */
1901static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1902 struct page **pages, unsigned long *num, pgprot_t prot)
1903{
1904 pmd_t *pmd = NULL;
1905 pte_t *start_pte, *pte;
1906 spinlock_t *pte_lock;
1907 struct mm_struct *const mm = vma->vm_mm;
1908 unsigned long curr_page_idx = 0;
1909 unsigned long remaining_pages_total = *num;
1910 unsigned long pages_to_write_in_pmd;
1911 int ret;
1912more:
1913 ret = -EFAULT;
1914 pmd = walk_to_pmd(mm, addr);
1915 if (!pmd)
1916 goto out;
1917
1918 pages_to_write_in_pmd = min_t(unsigned long,
1919 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1920
1921 /* Allocate the PTE if necessary; takes PMD lock once only. */
1922 ret = -ENOMEM;
1923 if (pte_alloc(mm, pmd))
1924 goto out;
1925
1926 while (pages_to_write_in_pmd) {
1927 int pte_idx = 0;
1928 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1929
1930 start_pte = pte_offset_map_lock(mm, pmd, addr, ptlp: &pte_lock);
1931 if (!start_pte) {
1932 ret = -EFAULT;
1933 goto out;
1934 }
1935 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1936 int err = insert_page_in_batch_locked(vma, pte,
1937 addr, page: pages[curr_page_idx], prot);
1938 if (unlikely(err)) {
1939 pte_unmap_unlock(start_pte, pte_lock);
1940 ret = err;
1941 remaining_pages_total -= pte_idx;
1942 goto out;
1943 }
1944 addr += PAGE_SIZE;
1945 ++curr_page_idx;
1946 }
1947 pte_unmap_unlock(start_pte, pte_lock);
1948 pages_to_write_in_pmd -= batch_size;
1949 remaining_pages_total -= batch_size;
1950 }
1951 if (remaining_pages_total)
1952 goto more;
1953 ret = 0;
1954out:
1955 *num = remaining_pages_total;
1956 return ret;
1957}
1958
1959/**
1960 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1961 * @vma: user vma to map to
1962 * @addr: target start user address of these pages
1963 * @pages: source kernel pages
1964 * @num: in: number of pages to map. out: number of pages that were *not*
1965 * mapped. (0 means all pages were successfully mapped).
1966 *
1967 * Preferred over vm_insert_page() when inserting multiple pages.
1968 *
1969 * In case of error, we may have mapped a subset of the provided
1970 * pages. It is the caller's responsibility to account for this case.
1971 *
1972 * The same restrictions apply as in vm_insert_page().
1973 */
1974int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1975 struct page **pages, unsigned long *num)
1976{
1977 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1978
1979 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1980 return -EFAULT;
1981 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1982 BUG_ON(mmap_read_trylock(vma->vm_mm));
1983 BUG_ON(vma->vm_flags & VM_PFNMAP);
1984 vm_flags_set(vma, VM_MIXEDMAP);
1985 }
1986 /* Defer page refcount checking till we're about to map that page. */
1987 return insert_pages(vma, addr, pages, num, prot: vma->vm_page_prot);
1988}
1989EXPORT_SYMBOL(vm_insert_pages);
1990
1991/**
1992 * vm_insert_page - insert single page into user vma
1993 * @vma: user vma to map to
1994 * @addr: target user address of this page
1995 * @page: source kernel page
1996 *
1997 * This allows drivers to insert individual pages they've allocated
1998 * into a user vma.
1999 *
2000 * The page has to be a nice clean _individual_ kernel allocation.
2001 * If you allocate a compound page, you need to have marked it as
2002 * such (__GFP_COMP), or manually just split the page up yourself
2003 * (see split_page()).
2004 *
2005 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2006 * took an arbitrary page protection parameter. This doesn't allow
2007 * that. Your vma protection will have to be set up correctly, which
2008 * means that if you want a shared writable mapping, you'd better
2009 * ask for a shared writable mapping!
2010 *
2011 * The page does not need to be reserved.
2012 *
2013 * Usually this function is called from f_op->mmap() handler
2014 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2015 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2016 * function from other places, for example from page-fault handler.
2017 *
2018 * Return: %0 on success, negative error code otherwise.
2019 */
2020int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2021 struct page *page)
2022{
2023 if (addr < vma->vm_start || addr >= vma->vm_end)
2024 return -EFAULT;
2025 if (!page_count(page))
2026 return -EINVAL;
2027 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2028 BUG_ON(mmap_read_trylock(vma->vm_mm));
2029 BUG_ON(vma->vm_flags & VM_PFNMAP);
2030 vm_flags_set(vma, VM_MIXEDMAP);
2031 }
2032 return insert_page(vma, addr, page, prot: vma->vm_page_prot);
2033}
2034EXPORT_SYMBOL(vm_insert_page);
2035
2036/*
2037 * __vm_map_pages - maps range of kernel pages into user vma
2038 * @vma: user vma to map to
2039 * @pages: pointer to array of source kernel pages
2040 * @num: number of pages in page array
2041 * @offset: user's requested vm_pgoff
2042 *
2043 * This allows drivers to map range of kernel pages into a user vma.
2044 *
2045 * Return: 0 on success and error code otherwise.
2046 */
2047static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2048 unsigned long num, unsigned long offset)
2049{
2050 unsigned long count = vma_pages(vma);
2051 unsigned long uaddr = vma->vm_start;
2052 int ret, i;
2053
2054 /* Fail if the user requested offset is beyond the end of the object */
2055 if (offset >= num)
2056 return -ENXIO;
2057
2058 /* Fail if the user requested size exceeds available object size */
2059 if (count > num - offset)
2060 return -ENXIO;
2061
2062 for (i = 0; i < count; i++) {
2063 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2064 if (ret < 0)
2065 return ret;
2066 uaddr += PAGE_SIZE;
2067 }
2068
2069 return 0;
2070}
2071
2072/**
2073 * vm_map_pages - maps range of kernel pages starts with non zero offset
2074 * @vma: user vma to map to
2075 * @pages: pointer to array of source kernel pages
2076 * @num: number of pages in page array
2077 *
2078 * Maps an object consisting of @num pages, catering for the user's
2079 * requested vm_pgoff
2080 *
2081 * If we fail to insert any page into the vma, the function will return
2082 * immediately leaving any previously inserted pages present. Callers
2083 * from the mmap handler may immediately return the error as their caller
2084 * will destroy the vma, removing any successfully inserted pages. Other
2085 * callers should make their own arrangements for calling unmap_region().
2086 *
2087 * Context: Process context. Called by mmap handlers.
2088 * Return: 0 on success and error code otherwise.
2089 */
2090int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2091 unsigned long num)
2092{
2093 return __vm_map_pages(vma, pages, num, offset: vma->vm_pgoff);
2094}
2095EXPORT_SYMBOL(vm_map_pages);
2096
2097/**
2098 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2099 * @vma: user vma to map to
2100 * @pages: pointer to array of source kernel pages
2101 * @num: number of pages in page array
2102 *
2103 * Similar to vm_map_pages(), except that it explicitly sets the offset
2104 * to 0. This function is intended for the drivers that did not consider
2105 * vm_pgoff.
2106 *
2107 * Context: Process context. Called by mmap handlers.
2108 * Return: 0 on success and error code otherwise.
2109 */
2110int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2111 unsigned long num)
2112{
2113 return __vm_map_pages(vma, pages, num, offset: 0);
2114}
2115EXPORT_SYMBOL(vm_map_pages_zero);
2116
2117static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2118 pfn_t pfn, pgprot_t prot, bool mkwrite)
2119{
2120 struct mm_struct *mm = vma->vm_mm;
2121 pte_t *pte, entry;
2122 spinlock_t *ptl;
2123
2124 pte = get_locked_pte(mm, addr, ptl: &ptl);
2125 if (!pte)
2126 return VM_FAULT_OOM;
2127 entry = ptep_get(ptep: pte);
2128 if (!pte_none(pte: entry)) {
2129 if (mkwrite) {
2130 /*
2131 * For read faults on private mappings the PFN passed
2132 * in may not match the PFN we have mapped if the
2133 * mapped PFN is a writeable COW page. In the mkwrite
2134 * case we are creating a writable PTE for a shared
2135 * mapping and we expect the PFNs to match. If they
2136 * don't match, we are likely racing with block
2137 * allocation and mapping invalidation so just skip the
2138 * update.
2139 */
2140 if (pte_pfn(pte: entry) != pfn_t_to_pfn(pfn)) {
2141 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
2142 goto out_unlock;
2143 }
2144 entry = pte_mkyoung(pte: entry);
2145 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
2146 if (ptep_set_access_flags(vma, address: addr, ptep: pte, entry, dirty: 1))
2147 update_mmu_cache(vma, addr, ptep: pte);
2148 }
2149 goto out_unlock;
2150 }
2151
2152 /* Ok, finally just insert the thing.. */
2153 if (pfn_t_devmap(pfn))
2154 entry = pte_mkdevmap(pte: pfn_t_pte(pfn, pgprot: prot));
2155 else
2156 entry = pte_mkspecial(pte: pfn_t_pte(pfn, pgprot: prot));
2157
2158 if (mkwrite) {
2159 entry = pte_mkyoung(pte: entry);
2160 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
2161 }
2162
2163 set_pte_at(mm, addr, pte, entry);
2164 update_mmu_cache(vma, addr, ptep: pte); /* XXX: why not for insert_page? */
2165
2166out_unlock:
2167 pte_unmap_unlock(pte, ptl);
2168 return VM_FAULT_NOPAGE;
2169}
2170
2171/**
2172 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2173 * @vma: user vma to map to
2174 * @addr: target user address of this page
2175 * @pfn: source kernel pfn
2176 * @pgprot: pgprot flags for the inserted page
2177 *
2178 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2179 * to override pgprot on a per-page basis.
2180 *
2181 * This only makes sense for IO mappings, and it makes no sense for
2182 * COW mappings. In general, using multiple vmas is preferable;
2183 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2184 * impractical.
2185 *
2186 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2187 * caching- and encryption bits different than those of @vma->vm_page_prot,
2188 * because the caching- or encryption mode may not be known at mmap() time.
2189 *
2190 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2191 * to set caching and encryption bits for those vmas (except for COW pages).
2192 * This is ensured by core vm only modifying these page table entries using
2193 * functions that don't touch caching- or encryption bits, using pte_modify()
2194 * if needed. (See for example mprotect()).
2195 *
2196 * Also when new page-table entries are created, this is only done using the
2197 * fault() callback, and never using the value of vma->vm_page_prot,
2198 * except for page-table entries that point to anonymous pages as the result
2199 * of COW.
2200 *
2201 * Context: Process context. May allocate using %GFP_KERNEL.
2202 * Return: vm_fault_t value.
2203 */
2204vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2205 unsigned long pfn, pgprot_t pgprot)
2206{
2207 /*
2208 * Technically, architectures with pte_special can avoid all these
2209 * restrictions (same for remap_pfn_range). However we would like
2210 * consistency in testing and feature parity among all, so we should
2211 * try to keep these invariants in place for everybody.
2212 */
2213 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2214 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2215 (VM_PFNMAP|VM_MIXEDMAP));
2216 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2217 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2218
2219 if (addr < vma->vm_start || addr >= vma->vm_end)
2220 return VM_FAULT_SIGBUS;
2221
2222 if (!pfn_modify_allowed(pfn, prot: pgprot))
2223 return VM_FAULT_SIGBUS;
2224
2225 track_pfn_insert(vma, prot: &pgprot, pfn: __pfn_to_pfn_t(pfn, PFN_DEV));
2226
2227 return insert_pfn(vma, addr, pfn: __pfn_to_pfn_t(pfn, PFN_DEV), prot: pgprot,
2228 mkwrite: false);
2229}
2230EXPORT_SYMBOL(vmf_insert_pfn_prot);
2231
2232/**
2233 * vmf_insert_pfn - insert single pfn into user vma
2234 * @vma: user vma to map to
2235 * @addr: target user address of this page
2236 * @pfn: source kernel pfn
2237 *
2238 * Similar to vm_insert_page, this allows drivers to insert individual pages
2239 * they've allocated into a user vma. Same comments apply.
2240 *
2241 * This function should only be called from a vm_ops->fault handler, and
2242 * in that case the handler should return the result of this function.
2243 *
2244 * vma cannot be a COW mapping.
2245 *
2246 * As this is called only for pages that do not currently exist, we
2247 * do not need to flush old virtual caches or the TLB.
2248 *
2249 * Context: Process context. May allocate using %GFP_KERNEL.
2250 * Return: vm_fault_t value.
2251 */
2252vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2253 unsigned long pfn)
2254{
2255 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2256}
2257EXPORT_SYMBOL(vmf_insert_pfn);
2258
2259static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2260{
2261 /* these checks mirror the abort conditions in vm_normal_page */
2262 if (vma->vm_flags & VM_MIXEDMAP)
2263 return true;
2264 if (pfn_t_devmap(pfn))
2265 return true;
2266 if (pfn_t_special(pfn))
2267 return true;
2268 if (is_zero_pfn(pfn: pfn_t_to_pfn(pfn)))
2269 return true;
2270 return false;
2271}
2272
2273static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2274 unsigned long addr, pfn_t pfn, bool mkwrite)
2275{
2276 pgprot_t pgprot = vma->vm_page_prot;
2277 int err;
2278
2279 BUG_ON(!vm_mixed_ok(vma, pfn));
2280
2281 if (addr < vma->vm_start || addr >= vma->vm_end)
2282 return VM_FAULT_SIGBUS;
2283
2284 track_pfn_insert(vma, prot: &pgprot, pfn);
2285
2286 if (!pfn_modify_allowed(pfn: pfn_t_to_pfn(pfn), prot: pgprot))
2287 return VM_FAULT_SIGBUS;
2288
2289 /*
2290 * If we don't have pte special, then we have to use the pfn_valid()
2291 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2292 * refcount the page if pfn_valid is true (hence insert_page rather
2293 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2294 * without pte special, it would there be refcounted as a normal page.
2295 */
2296 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2297 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2298 struct page *page;
2299
2300 /*
2301 * At this point we are committed to insert_page()
2302 * regardless of whether the caller specified flags that
2303 * result in pfn_t_has_page() == false.
2304 */
2305 page = pfn_to_page(pfn_t_to_pfn(pfn));
2306 err = insert_page(vma, addr, page, prot: pgprot);
2307 } else {
2308 return insert_pfn(vma, addr, pfn, prot: pgprot, mkwrite);
2309 }
2310
2311 if (err == -ENOMEM)
2312 return VM_FAULT_OOM;
2313 if (err < 0 && err != -EBUSY)
2314 return VM_FAULT_SIGBUS;
2315
2316 return VM_FAULT_NOPAGE;
2317}
2318
2319vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2320 pfn_t pfn)
2321{
2322 return __vm_insert_mixed(vma, addr, pfn, mkwrite: false);
2323}
2324EXPORT_SYMBOL(vmf_insert_mixed);
2325
2326/*
2327 * If the insertion of PTE failed because someone else already added a
2328 * different entry in the mean time, we treat that as success as we assume
2329 * the same entry was actually inserted.
2330 */
2331vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2332 unsigned long addr, pfn_t pfn)
2333{
2334 return __vm_insert_mixed(vma, addr, pfn, mkwrite: true);
2335}
2336EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2337
2338/*
2339 * maps a range of physical memory into the requested pages. the old
2340 * mappings are removed. any references to nonexistent pages results
2341 * in null mappings (currently treated as "copy-on-access")
2342 */
2343static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2344 unsigned long addr, unsigned long end,
2345 unsigned long pfn, pgprot_t prot)
2346{
2347 pte_t *pte, *mapped_pte;
2348 spinlock_t *ptl;
2349 int err = 0;
2350
2351 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2352 if (!pte)
2353 return -ENOMEM;
2354 arch_enter_lazy_mmu_mode();
2355 do {
2356 BUG_ON(!pte_none(ptep_get(pte)));
2357 if (!pfn_modify_allowed(pfn, prot)) {
2358 err = -EACCES;
2359 break;
2360 }
2361 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2362 pfn++;
2363 } while (pte++, addr += PAGE_SIZE, addr != end);
2364 arch_leave_lazy_mmu_mode();
2365 pte_unmap_unlock(mapped_pte, ptl);
2366 return err;
2367}
2368
2369static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2370 unsigned long addr, unsigned long end,
2371 unsigned long pfn, pgprot_t prot)
2372{
2373 pmd_t *pmd;
2374 unsigned long next;
2375 int err;
2376
2377 pfn -= addr >> PAGE_SHIFT;
2378 pmd = pmd_alloc(mm, pud, address: addr);
2379 if (!pmd)
2380 return -ENOMEM;
2381 VM_BUG_ON(pmd_trans_huge(*pmd));
2382 do {
2383 next = pmd_addr_end(addr, end);
2384 err = remap_pte_range(mm, pmd, addr, end: next,
2385 pfn: pfn + (addr >> PAGE_SHIFT), prot);
2386 if (err)
2387 return err;
2388 } while (pmd++, addr = next, addr != end);
2389 return 0;
2390}
2391
2392static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2393 unsigned long addr, unsigned long end,
2394 unsigned long pfn, pgprot_t prot)
2395{
2396 pud_t *pud;
2397 unsigned long next;
2398 int err;
2399
2400 pfn -= addr >> PAGE_SHIFT;
2401 pud = pud_alloc(mm, p4d, address: addr);
2402 if (!pud)
2403 return -ENOMEM;
2404 do {
2405 next = pud_addr_end(addr, end);
2406 err = remap_pmd_range(mm, pud, addr, end: next,
2407 pfn: pfn + (addr >> PAGE_SHIFT), prot);
2408 if (err)
2409 return err;
2410 } while (pud++, addr = next, addr != end);
2411 return 0;
2412}
2413
2414static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2415 unsigned long addr, unsigned long end,
2416 unsigned long pfn, pgprot_t prot)
2417{
2418 p4d_t *p4d;
2419 unsigned long next;
2420 int err;
2421
2422 pfn -= addr >> PAGE_SHIFT;
2423 p4d = p4d_alloc(mm, pgd, address: addr);
2424 if (!p4d)
2425 return -ENOMEM;
2426 do {
2427 next = p4d_addr_end(addr, end);
2428 err = remap_pud_range(mm, p4d, addr, end: next,
2429 pfn: pfn + (addr >> PAGE_SHIFT), prot);
2430 if (err)
2431 return err;
2432 } while (p4d++, addr = next, addr != end);
2433 return 0;
2434}
2435
2436/*
2437 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2438 * must have pre-validated the caching bits of the pgprot_t.
2439 */
2440int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2441 unsigned long pfn, unsigned long size, pgprot_t prot)
2442{
2443 pgd_t *pgd;
2444 unsigned long next;
2445 unsigned long end = addr + PAGE_ALIGN(size);
2446 struct mm_struct *mm = vma->vm_mm;
2447 int err;
2448
2449 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2450 return -EINVAL;
2451
2452 /*
2453 * Physically remapped pages are special. Tell the
2454 * rest of the world about it:
2455 * VM_IO tells people not to look at these pages
2456 * (accesses can have side effects).
2457 * VM_PFNMAP tells the core MM that the base pages are just
2458 * raw PFN mappings, and do not have a "struct page" associated
2459 * with them.
2460 * VM_DONTEXPAND
2461 * Disable vma merging and expanding with mremap().
2462 * VM_DONTDUMP
2463 * Omit vma from core dump, even when VM_IO turned off.
2464 *
2465 * There's a horrible special case to handle copy-on-write
2466 * behaviour that some programs depend on. We mark the "original"
2467 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2468 * See vm_normal_page() for details.
2469 */
2470 if (is_cow_mapping(flags: vma->vm_flags)) {
2471 if (addr != vma->vm_start || end != vma->vm_end)
2472 return -EINVAL;
2473 vma->vm_pgoff = pfn;
2474 }
2475
2476 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
2477
2478 BUG_ON(addr >= end);
2479 pfn -= addr >> PAGE_SHIFT;
2480 pgd = pgd_offset(mm, addr);
2481 flush_cache_range(vma, start: addr, end);
2482 do {
2483 next = pgd_addr_end(addr, end);
2484 err = remap_p4d_range(mm, pgd, addr, end: next,
2485 pfn: pfn + (addr >> PAGE_SHIFT), prot);
2486 if (err)
2487 return err;
2488 } while (pgd++, addr = next, addr != end);
2489
2490 return 0;
2491}
2492
2493/**
2494 * remap_pfn_range - remap kernel memory to userspace
2495 * @vma: user vma to map to
2496 * @addr: target page aligned user address to start at
2497 * @pfn: page frame number of kernel physical memory address
2498 * @size: size of mapping area
2499 * @prot: page protection flags for this mapping
2500 *
2501 * Note: this is only safe if the mm semaphore is held when called.
2502 *
2503 * Return: %0 on success, negative error code otherwise.
2504 */
2505int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2506 unsigned long pfn, unsigned long size, pgprot_t prot)
2507{
2508 int err;
2509
2510 err = track_pfn_remap(vma, prot: &prot, pfn, addr, PAGE_ALIGN(size));
2511 if (err)
2512 return -EINVAL;
2513
2514 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2515 if (err)
2516 untrack_pfn(vma, pfn, PAGE_ALIGN(size), mm_wr_locked: true);
2517 return err;
2518}
2519EXPORT_SYMBOL(remap_pfn_range);
2520
2521/**
2522 * vm_iomap_memory - remap memory to userspace
2523 * @vma: user vma to map to
2524 * @start: start of the physical memory to be mapped
2525 * @len: size of area
2526 *
2527 * This is a simplified io_remap_pfn_range() for common driver use. The
2528 * driver just needs to give us the physical memory range to be mapped,
2529 * we'll figure out the rest from the vma information.
2530 *
2531 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2532 * whatever write-combining details or similar.
2533 *
2534 * Return: %0 on success, negative error code otherwise.
2535 */
2536int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2537{
2538 unsigned long vm_len, pfn, pages;
2539
2540 /* Check that the physical memory area passed in looks valid */
2541 if (start + len < start)
2542 return -EINVAL;
2543 /*
2544 * You *really* shouldn't map things that aren't page-aligned,
2545 * but we've historically allowed it because IO memory might
2546 * just have smaller alignment.
2547 */
2548 len += start & ~PAGE_MASK;
2549 pfn = start >> PAGE_SHIFT;
2550 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2551 if (pfn + pages < pfn)
2552 return -EINVAL;
2553
2554 /* We start the mapping 'vm_pgoff' pages into the area */
2555 if (vma->vm_pgoff > pages)
2556 return -EINVAL;
2557 pfn += vma->vm_pgoff;
2558 pages -= vma->vm_pgoff;
2559
2560 /* Can we fit all of the mapping? */
2561 vm_len = vma->vm_end - vma->vm_start;
2562 if (vm_len >> PAGE_SHIFT > pages)
2563 return -EINVAL;
2564
2565 /* Ok, let it rip */
2566 return io_remap_pfn_range(vma, addr: vma->vm_start, pfn, size: vm_len, prot: vma->vm_page_prot);
2567}
2568EXPORT_SYMBOL(vm_iomap_memory);
2569
2570static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2571 unsigned long addr, unsigned long end,
2572 pte_fn_t fn, void *data, bool create,
2573 pgtbl_mod_mask *mask)
2574{
2575 pte_t *pte, *mapped_pte;
2576 int err = 0;
2577 spinlock_t *ptl;
2578
2579 if (create) {
2580 mapped_pte = pte = (mm == &init_mm) ?
2581 pte_alloc_kernel_track(pmd, addr, mask) :
2582 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2583 if (!pte)
2584 return -ENOMEM;
2585 } else {
2586 mapped_pte = pte = (mm == &init_mm) ?
2587 pte_offset_kernel(pmd, address: addr) :
2588 pte_offset_map_lock(mm, pmd, addr, ptlp: &ptl);
2589 if (!pte)
2590 return -EINVAL;
2591 }
2592
2593 arch_enter_lazy_mmu_mode();
2594
2595 if (fn) {
2596 do {
2597 if (create || !pte_none(pte: ptep_get(ptep: pte))) {
2598 err = fn(pte++, addr, data);
2599 if (err)
2600 break;
2601 }
2602 } while (addr += PAGE_SIZE, addr != end);
2603 }
2604 *mask |= PGTBL_PTE_MODIFIED;
2605
2606 arch_leave_lazy_mmu_mode();
2607
2608 if (mm != &init_mm)
2609 pte_unmap_unlock(mapped_pte, ptl);
2610 return err;
2611}
2612
2613static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2614 unsigned long addr, unsigned long end,
2615 pte_fn_t fn, void *data, bool create,
2616 pgtbl_mod_mask *mask)
2617{
2618 pmd_t *pmd;
2619 unsigned long next;
2620 int err = 0;
2621
2622 BUG_ON(pud_huge(*pud));
2623
2624 if (create) {
2625 pmd = pmd_alloc_track(mm, pud, address: addr, mod_mask: mask);
2626 if (!pmd)
2627 return -ENOMEM;
2628 } else {
2629 pmd = pmd_offset(pud, address: addr);
2630 }
2631 do {
2632 next = pmd_addr_end(addr, end);
2633 if (pmd_none(pmd: *pmd) && !create)
2634 continue;
2635 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2636 return -EINVAL;
2637 if (!pmd_none(pmd: *pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2638 if (!create)
2639 continue;
2640 pmd_clear_bad(pmd);
2641 }
2642 err = apply_to_pte_range(mm, pmd, addr, end: next,
2643 fn, data, create, mask);
2644 if (err)
2645 break;
2646 } while (pmd++, addr = next, addr != end);
2647
2648 return err;
2649}
2650
2651static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2652 unsigned long addr, unsigned long end,
2653 pte_fn_t fn, void *data, bool create,
2654 pgtbl_mod_mask *mask)
2655{
2656 pud_t *pud;
2657 unsigned long next;
2658 int err = 0;
2659
2660 if (create) {
2661 pud = pud_alloc_track(mm, p4d, address: addr, mod_mask: mask);
2662 if (!pud)
2663 return -ENOMEM;
2664 } else {
2665 pud = pud_offset(p4d, address: addr);
2666 }
2667 do {
2668 next = pud_addr_end(addr, end);
2669 if (pud_none(pud: *pud) && !create)
2670 continue;
2671 if (WARN_ON_ONCE(pud_leaf(*pud)))
2672 return -EINVAL;
2673 if (!pud_none(pud: *pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2674 if (!create)
2675 continue;
2676 pud_clear_bad(pud);
2677 }
2678 err = apply_to_pmd_range(mm, pud, addr, end: next,
2679 fn, data, create, mask);
2680 if (err)
2681 break;
2682 } while (pud++, addr = next, addr != end);
2683
2684 return err;
2685}
2686
2687static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2688 unsigned long addr, unsigned long end,
2689 pte_fn_t fn, void *data, bool create,
2690 pgtbl_mod_mask *mask)
2691{
2692 p4d_t *p4d;
2693 unsigned long next;
2694 int err = 0;
2695
2696 if (create) {
2697 p4d = p4d_alloc_track(mm, pgd, address: addr, mod_mask: mask);
2698 if (!p4d)
2699 return -ENOMEM;
2700 } else {
2701 p4d = p4d_offset(pgd, address: addr);
2702 }
2703 do {
2704 next = p4d_addr_end(addr, end);
2705 if (p4d_none(p4d: *p4d) && !create)
2706 continue;
2707 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2708 return -EINVAL;
2709 if (!p4d_none(p4d: *p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2710 if (!create)
2711 continue;
2712 p4d_clear_bad(p4d);
2713 }
2714 err = apply_to_pud_range(mm, p4d, addr, end: next,
2715 fn, data, create, mask);
2716 if (err)
2717 break;
2718 } while (p4d++, addr = next, addr != end);
2719
2720 return err;
2721}
2722
2723static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2724 unsigned long size, pte_fn_t fn,
2725 void *data, bool create)
2726{
2727 pgd_t *pgd;
2728 unsigned long start = addr, next;
2729 unsigned long end = addr + size;
2730 pgtbl_mod_mask mask = 0;
2731 int err = 0;
2732
2733 if (WARN_ON(addr >= end))
2734 return -EINVAL;
2735
2736 pgd = pgd_offset(mm, addr);
2737 do {
2738 next = pgd_addr_end(addr, end);
2739 if (pgd_none(pgd: *pgd) && !create)
2740 continue;
2741 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2742 return -EINVAL;
2743 if (!pgd_none(pgd: *pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2744 if (!create)
2745 continue;
2746 pgd_clear_bad(pgd);
2747 }
2748 err = apply_to_p4d_range(mm, pgd, addr, end: next,
2749 fn, data, create, mask: &mask);
2750 if (err)
2751 break;
2752 } while (pgd++, addr = next, addr != end);
2753
2754 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2755 arch_sync_kernel_mappings(start, end: start + size);
2756
2757 return err;
2758}
2759
2760/*
2761 * Scan a region of virtual memory, filling in page tables as necessary
2762 * and calling a provided function on each leaf page table.
2763 */
2764int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2765 unsigned long size, pte_fn_t fn, void *data)
2766{
2767 return __apply_to_page_range(mm, addr, size, fn, data, create: true);
2768}
2769EXPORT_SYMBOL_GPL(apply_to_page_range);
2770
2771/*
2772 * Scan a region of virtual memory, calling a provided function on
2773 * each leaf page table where it exists.
2774 *
2775 * Unlike apply_to_page_range, this does _not_ fill in page tables
2776 * where they are absent.
2777 */
2778int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2779 unsigned long size, pte_fn_t fn, void *data)
2780{
2781 return __apply_to_page_range(mm, addr, size, fn, data, create: false);
2782}
2783EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2784
2785/*
2786 * handle_pte_fault chooses page fault handler according to an entry which was
2787 * read non-atomically. Before making any commitment, on those architectures
2788 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2789 * parts, do_swap_page must check under lock before unmapping the pte and
2790 * proceeding (but do_wp_page is only called after already making such a check;
2791 * and do_anonymous_page can safely check later on).
2792 */
2793static inline int pte_unmap_same(struct vm_fault *vmf)
2794{
2795 int same = 1;
2796#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2797 if (sizeof(pte_t) > sizeof(unsigned long)) {
2798 spin_lock(lock: vmf->ptl);
2799 same = pte_same(a: ptep_get(ptep: vmf->pte), b: vmf->orig_pte);
2800 spin_unlock(lock: vmf->ptl);
2801 }
2802#endif
2803 pte_unmap(pte: vmf->pte);
2804 vmf->pte = NULL;
2805 return same;
2806}
2807
2808/*
2809 * Return:
2810 * 0: copied succeeded
2811 * -EHWPOISON: copy failed due to hwpoison in source page
2812 * -EAGAIN: copied failed (some other reason)
2813 */
2814static inline int __wp_page_copy_user(struct page *dst, struct page *src,
2815 struct vm_fault *vmf)
2816{
2817 int ret;
2818 void *kaddr;
2819 void __user *uaddr;
2820 struct vm_area_struct *vma = vmf->vma;
2821 struct mm_struct *mm = vma->vm_mm;
2822 unsigned long addr = vmf->address;
2823
2824 if (likely(src)) {
2825 if (copy_mc_user_highpage(to: dst, from: src, vaddr: addr, vma)) {
2826 memory_failure_queue(page_to_pfn(src), flags: 0);
2827 return -EHWPOISON;
2828 }
2829 return 0;
2830 }
2831
2832 /*
2833 * If the source page was a PFN mapping, we don't have
2834 * a "struct page" for it. We do a best-effort copy by
2835 * just copying from the original user address. If that
2836 * fails, we just zero-fill it. Live with it.
2837 */
2838 kaddr = kmap_atomic(page: dst);
2839 uaddr = (void __user *)(addr & PAGE_MASK);
2840
2841 /*
2842 * On architectures with software "accessed" bits, we would
2843 * take a double page fault, so mark it accessed here.
2844 */
2845 vmf->pte = NULL;
2846 if (!arch_has_hw_pte_young() && !pte_young(pte: vmf->orig_pte)) {
2847 pte_t entry;
2848
2849 vmf->pte = pte_offset_map_lock(mm, pmd: vmf->pmd, addr, ptlp: &vmf->ptl);
2850 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2851 /*
2852 * Other thread has already handled the fault
2853 * and update local tlb only
2854 */
2855 if (vmf->pte)
2856 update_mmu_tlb(vma, address: addr, ptep: vmf->pte);
2857 ret = -EAGAIN;
2858 goto pte_unlock;
2859 }
2860
2861 entry = pte_mkyoung(pte: vmf->orig_pte);
2862 if (ptep_set_access_flags(vma, address: addr, ptep: vmf->pte, entry, dirty: 0))
2863 update_mmu_cache_range(vmf, vma, addr, ptep: vmf->pte, nr: 1);
2864 }
2865
2866 /*
2867 * This really shouldn't fail, because the page is there
2868 * in the page tables. But it might just be unreadable,
2869 * in which case we just give up and fill the result with
2870 * zeroes.
2871 */
2872 if (__copy_from_user_inatomic(to: kaddr, from: uaddr, PAGE_SIZE)) {
2873 if (vmf->pte)
2874 goto warn;
2875
2876 /* Re-validate under PTL if the page is still mapped */
2877 vmf->pte = pte_offset_map_lock(mm, pmd: vmf->pmd, addr, ptlp: &vmf->ptl);
2878 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2879 /* The PTE changed under us, update local tlb */
2880 if (vmf->pte)
2881 update_mmu_tlb(vma, address: addr, ptep: vmf->pte);
2882 ret = -EAGAIN;
2883 goto pte_unlock;
2884 }
2885
2886 /*
2887 * The same page can be mapped back since last copy attempt.
2888 * Try to copy again under PTL.
2889 */
2890 if (__copy_from_user_inatomic(to: kaddr, from: uaddr, PAGE_SIZE)) {
2891 /*
2892 * Give a warn in case there can be some obscure
2893 * use-case
2894 */
2895warn:
2896 WARN_ON_ONCE(1);
2897 clear_page(page: kaddr);
2898 }
2899 }
2900
2901 ret = 0;
2902
2903pte_unlock:
2904 if (vmf->pte)
2905 pte_unmap_unlock(vmf->pte, vmf->ptl);
2906 kunmap_atomic(kaddr);
2907 flush_dcache_page(page: dst);
2908
2909 return ret;
2910}
2911
2912static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2913{
2914 struct file *vm_file = vma->vm_file;
2915
2916 if (vm_file)
2917 return mapping_gfp_mask(mapping: vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2918
2919 /*
2920 * Special mappings (e.g. VDSO) do not have any file so fake
2921 * a default GFP_KERNEL for them.
2922 */
2923 return GFP_KERNEL;
2924}
2925
2926/*
2927 * Notify the address space that the page is about to become writable so that
2928 * it can prohibit this or wait for the page to get into an appropriate state.
2929 *
2930 * We do this without the lock held, so that it can sleep if it needs to.
2931 */
2932static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio)
2933{
2934 vm_fault_t ret;
2935 unsigned int old_flags = vmf->flags;
2936
2937 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2938
2939 if (vmf->vma->vm_file &&
2940 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2941 return VM_FAULT_SIGBUS;
2942
2943 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2944 /* Restore original flags so that caller is not surprised */
2945 vmf->flags = old_flags;
2946 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2947 return ret;
2948 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2949 folio_lock(folio);
2950 if (!folio->mapping) {
2951 folio_unlock(folio);
2952 return 0; /* retry */
2953 }
2954 ret |= VM_FAULT_LOCKED;
2955 } else
2956 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2957 return ret;
2958}
2959
2960/*
2961 * Handle dirtying of a page in shared file mapping on a write fault.
2962 *
2963 * The function expects the page to be locked and unlocks it.
2964 */
2965static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2966{
2967 struct vm_area_struct *vma = vmf->vma;
2968 struct address_space *mapping;
2969 struct folio *folio = page_folio(vmf->page);
2970 bool dirtied;
2971 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2972
2973 dirtied = folio_mark_dirty(folio);
2974 VM_BUG_ON_FOLIO(folio_test_anon(folio), folio);
2975 /*
2976 * Take a local copy of the address_space - folio.mapping may be zeroed
2977 * by truncate after folio_unlock(). The address_space itself remains
2978 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
2979 * release semantics to prevent the compiler from undoing this copying.
2980 */
2981 mapping = folio_raw_mapping(folio);
2982 folio_unlock(folio);
2983
2984 if (!page_mkwrite)
2985 file_update_time(file: vma->vm_file);
2986
2987 /*
2988 * Throttle page dirtying rate down to writeback speed.
2989 *
2990 * mapping may be NULL here because some device drivers do not
2991 * set page.mapping but still dirty their pages
2992 *
2993 * Drop the mmap_lock before waiting on IO, if we can. The file
2994 * is pinning the mapping, as per above.
2995 */
2996 if ((dirtied || page_mkwrite) && mapping) {
2997 struct file *fpin;
2998
2999 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
3000 balance_dirty_pages_ratelimited(mapping);
3001 if (fpin) {
3002 fput(fpin);
3003 return VM_FAULT_COMPLETED;
3004 }
3005 }
3006
3007 return 0;
3008}
3009
3010/*
3011 * Handle write page faults for pages that can be reused in the current vma
3012 *
3013 * This can happen either due to the mapping being with the VM_SHARED flag,
3014 * or due to us being the last reference standing to the page. In either
3015 * case, all we need to do here is to mark the page as writable and update
3016 * any related book-keeping.
3017 */
3018static inline void wp_page_reuse(struct vm_fault *vmf, struct folio *folio)
3019 __releases(vmf->ptl)
3020{
3021 struct vm_area_struct *vma = vmf->vma;
3022 pte_t entry;
3023
3024 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3025
3026 if (folio) {
3027 VM_BUG_ON(folio_test_anon(folio) &&
3028 !PageAnonExclusive(vmf->page));
3029 /*
3030 * Clear the folio's cpupid information as the existing
3031 * information potentially belongs to a now completely
3032 * unrelated process.
3033 */
3034 folio_xchg_last_cpupid(folio, (1 << LAST_CPUPID_SHIFT) - 1);
3035 }
3036
3037 flush_cache_page(vma, vmaddr: vmf->address, pfn: pte_pfn(pte: vmf->orig_pte));
3038 entry = pte_mkyoung(pte: vmf->orig_pte);
3039 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
3040 if (ptep_set_access_flags(vma, address: vmf->address, ptep: vmf->pte, entry, dirty: 1))
3041 update_mmu_cache_range(vmf, vma, addr: vmf->address, ptep: vmf->pte, nr: 1);
3042 pte_unmap_unlock(vmf->pte, vmf->ptl);
3043 count_vm_event(item: PGREUSE);
3044}
3045
3046/*
3047 * We could add a bitflag somewhere, but for now, we know that all
3048 * vm_ops that have a ->map_pages have been audited and don't need
3049 * the mmap_lock to be held.
3050 */
3051static inline vm_fault_t vmf_can_call_fault(const struct vm_fault *vmf)
3052{
3053 struct vm_area_struct *vma = vmf->vma;
3054
3055 if (vma->vm_ops->map_pages || !(vmf->flags & FAULT_FLAG_VMA_LOCK))
3056 return 0;
3057 vma_end_read(vma);
3058 return VM_FAULT_RETRY;
3059}
3060
3061static vm_fault_t vmf_anon_prepare(struct vm_fault *vmf)
3062{
3063 struct vm_area_struct *vma = vmf->vma;
3064
3065 if (likely(vma->anon_vma))
3066 return 0;
3067 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3068 vma_end_read(vma);
3069 return VM_FAULT_RETRY;
3070 }
3071 if (__anon_vma_prepare(vma))
3072 return VM_FAULT_OOM;
3073 return 0;
3074}
3075
3076/*
3077 * Handle the case of a page which we actually need to copy to a new page,
3078 * either due to COW or unsharing.
3079 *
3080 * Called with mmap_lock locked and the old page referenced, but
3081 * without the ptl held.
3082 *
3083 * High level logic flow:
3084 *
3085 * - Allocate a page, copy the content of the old page to the new one.
3086 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3087 * - Take the PTL. If the pte changed, bail out and release the allocated page
3088 * - If the pte is still the way we remember it, update the page table and all
3089 * relevant references. This includes dropping the reference the page-table
3090 * held to the old page, as well as updating the rmap.
3091 * - In any case, unlock the PTL and drop the reference we took to the old page.
3092 */
3093static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3094{
3095 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3096 struct vm_area_struct *vma = vmf->vma;
3097 struct mm_struct *mm = vma->vm_mm;
3098 struct folio *old_folio = NULL;
3099 struct folio *new_folio = NULL;
3100 pte_t entry;
3101 int page_copied = 0;
3102 struct mmu_notifier_range range;
3103 vm_fault_t ret;
3104
3105 delayacct_wpcopy_start();
3106
3107 if (vmf->page)
3108 old_folio = page_folio(vmf->page);
3109 ret = vmf_anon_prepare(vmf);
3110 if (unlikely(ret))
3111 goto out;
3112
3113 if (is_zero_pfn(pfn: pte_pfn(pte: vmf->orig_pte))) {
3114 new_folio = vma_alloc_zeroed_movable_folio(vma, vmf->address);
3115 if (!new_folio)
3116 goto oom;
3117 } else {
3118 int err;
3119 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, order: 0, vma,
3120 addr: vmf->address, hugepage: false);
3121 if (!new_folio)
3122 goto oom;
3123
3124 err = __wp_page_copy_user(dst: &new_folio->page, src: vmf->page, vmf);
3125 if (err) {
3126 /*
3127 * COW failed, if the fault was solved by other,
3128 * it's fine. If not, userspace would re-fault on
3129 * the same address and we will handle the fault
3130 * from the second attempt.
3131 * The -EHWPOISON case will not be retried.
3132 */
3133 folio_put(folio: new_folio);
3134 if (old_folio)
3135 folio_put(folio: old_folio);
3136
3137 delayacct_wpcopy_end();
3138 return err == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3139 }
3140 kmsan_copy_page_meta(dst: &new_folio->page, src: vmf->page);
3141 }
3142
3143 if (mem_cgroup_charge(folio: new_folio, mm, GFP_KERNEL))
3144 goto oom_free_new;
3145 folio_throttle_swaprate(folio: new_folio, GFP_KERNEL);
3146
3147 __folio_mark_uptodate(folio: new_folio);
3148
3149 mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_CLEAR, flags: 0, mm,
3150 start: vmf->address & PAGE_MASK,
3151 end: (vmf->address & PAGE_MASK) + PAGE_SIZE);
3152 mmu_notifier_invalidate_range_start(range: &range);
3153
3154 /*
3155 * Re-check the pte - we dropped the lock
3156 */
3157 vmf->pte = pte_offset_map_lock(mm, pmd: vmf->pmd, addr: vmf->address, ptlp: &vmf->ptl);
3158 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3159 if (old_folio) {
3160 if (!folio_test_anon(folio: old_folio)) {
3161 dec_mm_counter(mm, member: mm_counter_file(page: &old_folio->page));
3162 inc_mm_counter(mm, member: MM_ANONPAGES);
3163 }
3164 } else {
3165 ksm_might_unmap_zero_page(mm, pte: vmf->orig_pte);
3166 inc_mm_counter(mm, member: MM_ANONPAGES);
3167 }
3168 flush_cache_page(vma, vmaddr: vmf->address, pfn: pte_pfn(pte: vmf->orig_pte));
3169 entry = mk_pte(&new_folio->page, vma->vm_page_prot);
3170 entry = pte_sw_mkyoung(pte: entry);
3171 if (unlikely(unshare)) {
3172 if (pte_soft_dirty(pte: vmf->orig_pte))
3173 entry = pte_mksoft_dirty(pte: entry);
3174 if (pte_uffd_wp(pte: vmf->orig_pte))
3175 entry = pte_mkuffd_wp(pte: entry);
3176 } else {
3177 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
3178 }
3179
3180 /*
3181 * Clear the pte entry and flush it first, before updating the
3182 * pte with the new entry, to keep TLBs on different CPUs in
3183 * sync. This code used to set the new PTE then flush TLBs, but
3184 * that left a window where the new PTE could be loaded into
3185 * some TLBs while the old PTE remains in others.
3186 */
3187 ptep_clear_flush(vma, address: vmf->address, ptep: vmf->pte);
3188 folio_add_new_anon_rmap(new_folio, vma, address: vmf->address);
3189 folio_add_lru_vma(new_folio, vma);
3190 /*
3191 * We call the notify macro here because, when using secondary
3192 * mmu page tables (such as kvm shadow page tables), we want the
3193 * new page to be mapped directly into the secondary page table.
3194 */
3195 BUG_ON(unshare && pte_write(entry));
3196 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3197 update_mmu_cache_range(vmf, vma, addr: vmf->address, ptep: vmf->pte, nr: 1);
3198 if (old_folio) {
3199 /*
3200 * Only after switching the pte to the new page may
3201 * we remove the mapcount here. Otherwise another
3202 * process may come and find the rmap count decremented
3203 * before the pte is switched to the new page, and
3204 * "reuse" the old page writing into it while our pte
3205 * here still points into it and can be read by other
3206 * threads.
3207 *
3208 * The critical issue is to order this
3209 * page_remove_rmap with the ptp_clear_flush above.
3210 * Those stores are ordered by (if nothing else,)
3211 * the barrier present in the atomic_add_negative
3212 * in page_remove_rmap.
3213 *
3214 * Then the TLB flush in ptep_clear_flush ensures that
3215 * no process can access the old page before the
3216 * decremented mapcount is visible. And the old page
3217 * cannot be reused until after the decremented
3218 * mapcount is visible. So transitively, TLBs to
3219 * old page will be flushed before it can be reused.
3220 */
3221 page_remove_rmap(vmf->page, vma, compound: false);
3222 }
3223
3224 /* Free the old page.. */
3225 new_folio = old_folio;
3226 page_copied = 1;
3227 pte_unmap_unlock(vmf->pte, vmf->ptl);
3228 } else if (vmf->pte) {
3229 update_mmu_tlb(vma, address: vmf->address, ptep: vmf->pte);
3230 pte_unmap_unlock(vmf->pte, vmf->ptl);
3231 }
3232
3233 mmu_notifier_invalidate_range_end(range: &range);
3234
3235 if (new_folio)
3236 folio_put(folio: new_folio);
3237 if (old_folio) {
3238 if (page_copied)
3239 free_swap_cache(page: &old_folio->page);
3240 folio_put(folio: old_folio);
3241 }
3242
3243 delayacct_wpcopy_end();
3244 return 0;
3245oom_free_new:
3246 folio_put(folio: new_folio);
3247oom:
3248 ret = VM_FAULT_OOM;
3249out:
3250 if (old_folio)
3251 folio_put(folio: old_folio);
3252
3253 delayacct_wpcopy_end();
3254 return ret;
3255}
3256
3257/**
3258 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3259 * writeable once the page is prepared
3260 *
3261 * @vmf: structure describing the fault
3262 * @folio: the folio of vmf->page
3263 *
3264 * This function handles all that is needed to finish a write page fault in a
3265 * shared mapping due to PTE being read-only once the mapped page is prepared.
3266 * It handles locking of PTE and modifying it.
3267 *
3268 * The function expects the page to be locked or other protection against
3269 * concurrent faults / writeback (such as DAX radix tree locks).
3270 *
3271 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3272 * we acquired PTE lock.
3273 */
3274static vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf, struct folio *folio)
3275{
3276 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3277 vmf->pte = pte_offset_map_lock(mm: vmf->vma->vm_mm, pmd: vmf->pmd, addr: vmf->address,
3278 ptlp: &vmf->ptl);
3279 if (!vmf->pte)
3280 return VM_FAULT_NOPAGE;
3281 /*
3282 * We might have raced with another page fault while we released the
3283 * pte_offset_map_lock.
3284 */
3285 if (!pte_same(a: ptep_get(ptep: vmf->pte), b: vmf->orig_pte)) {
3286 update_mmu_tlb(vma: vmf->vma, address: vmf->address, ptep: vmf->pte);
3287 pte_unmap_unlock(vmf->pte, vmf->ptl);
3288 return VM_FAULT_NOPAGE;
3289 }
3290 wp_page_reuse(vmf, folio);
3291 return 0;
3292}
3293
3294/*
3295 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3296 * mapping
3297 */
3298static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3299{
3300 struct vm_area_struct *vma = vmf->vma;
3301
3302 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3303 vm_fault_t ret;
3304
3305 pte_unmap_unlock(vmf->pte, vmf->ptl);
3306 ret = vmf_can_call_fault(vmf);
3307 if (ret)
3308 return ret;
3309
3310 vmf->flags |= FAULT_FLAG_MKWRITE;
3311 ret = vma->vm_ops->pfn_mkwrite(vmf);
3312 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3313 return ret;
3314 return finish_mkwrite_fault(vmf, NULL);
3315 }
3316 wp_page_reuse(vmf, NULL);
3317 return 0;
3318}
3319
3320static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio)
3321 __releases(vmf->ptl)
3322{
3323 struct vm_area_struct *vma = vmf->vma;
3324 vm_fault_t ret = 0;
3325
3326 folio_get(folio);
3327
3328 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3329 vm_fault_t tmp;
3330
3331 pte_unmap_unlock(vmf->pte, vmf->ptl);
3332 tmp = vmf_can_call_fault(vmf);
3333 if (tmp) {
3334 folio_put(folio);
3335 return tmp;
3336 }
3337
3338 tmp = do_page_mkwrite(vmf, folio);
3339 if (unlikely(!tmp || (tmp &
3340 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3341 folio_put(folio);
3342 return tmp;
3343 }
3344 tmp = finish_mkwrite_fault(vmf, folio);
3345 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3346 folio_unlock(folio);
3347 folio_put(folio);
3348 return tmp;
3349 }
3350 } else {
3351 wp_page_reuse(vmf, folio);
3352 folio_lock(folio);
3353 }
3354 ret |= fault_dirty_shared_page(vmf);
3355 folio_put(folio);
3356
3357 return ret;
3358}
3359
3360static bool wp_can_reuse_anon_folio(struct folio *folio,
3361 struct vm_area_struct *vma)
3362{
3363 /*
3364 * We have to verify under folio lock: these early checks are
3365 * just an optimization to avoid locking the folio and freeing
3366 * the swapcache if there is little hope that we can reuse.
3367 *
3368 * KSM doesn't necessarily raise the folio refcount.
3369 */
3370 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3371 return false;
3372 if (!folio_test_lru(folio))
3373 /*
3374 * We cannot easily detect+handle references from
3375 * remote LRU caches or references to LRU folios.
3376 */
3377 lru_add_drain();
3378 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
3379 return false;
3380 if (!folio_trylock(folio))
3381 return false;
3382 if (folio_test_swapcache(folio))
3383 folio_free_swap(folio);
3384 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
3385 folio_unlock(folio);
3386 return false;
3387 }
3388 /*
3389 * Ok, we've got the only folio reference from our mapping
3390 * and the folio is locked, it's dark out, and we're wearing
3391 * sunglasses. Hit it.
3392 */
3393 folio_move_anon_rmap(folio, vma);
3394 folio_unlock(folio);
3395 return true;
3396}
3397
3398/*
3399 * This routine handles present pages, when
3400 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3401 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3402 * (FAULT_FLAG_UNSHARE)
3403 *
3404 * It is done by copying the page to a new address and decrementing the
3405 * shared-page counter for the old page.
3406 *
3407 * Note that this routine assumes that the protection checks have been
3408 * done by the caller (the low-level page fault routine in most cases).
3409 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3410 * done any necessary COW.
3411 *
3412 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3413 * though the page will change only once the write actually happens. This
3414 * avoids a few races, and potentially makes it more efficient.
3415 *
3416 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3417 * but allow concurrent faults), with pte both mapped and locked.
3418 * We return with mmap_lock still held, but pte unmapped and unlocked.
3419 */
3420static vm_fault_t do_wp_page(struct vm_fault *vmf)
3421 __releases(vmf->ptl)
3422{
3423 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3424 struct vm_area_struct *vma = vmf->vma;
3425 struct folio *folio = NULL;
3426 pte_t pte;
3427
3428 if (likely(!unshare)) {
3429 if (userfaultfd_pte_wp(vma, pte: ptep_get(ptep: vmf->pte))) {
3430 if (!userfaultfd_wp_async(vma)) {
3431 pte_unmap_unlock(vmf->pte, vmf->ptl);
3432 return handle_userfault(vmf, VM_UFFD_WP);
3433 }
3434
3435 /*
3436 * Nothing needed (cache flush, TLB invalidations,
3437 * etc.) because we're only removing the uffd-wp bit,
3438 * which is completely invisible to the user.
3439 */
3440 pte = pte_clear_uffd_wp(pte: ptep_get(ptep: vmf->pte));
3441
3442 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3443 /*
3444 * Update this to be prepared for following up CoW
3445 * handling
3446 */
3447 vmf->orig_pte = pte;
3448 }
3449
3450 /*
3451 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3452 * is flushed in this case before copying.
3453 */
3454 if (unlikely(userfaultfd_wp(vmf->vma) &&
3455 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3456 flush_tlb_page(vma: vmf->vma, a: vmf->address);
3457 }
3458
3459 vmf->page = vm_normal_page(vma, addr: vmf->address, pte: vmf->orig_pte);
3460
3461 if (vmf->page)
3462 folio = page_folio(vmf->page);
3463
3464 /*
3465 * Shared mapping: we are guaranteed to have VM_WRITE and
3466 * FAULT_FLAG_WRITE set at this point.
3467 */
3468 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
3469 /*
3470 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3471 * VM_PFNMAP VMA.
3472 *
3473 * We should not cow pages in a shared writeable mapping.
3474 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3475 */
3476 if (!vmf->page)
3477 return wp_pfn_shared(vmf);
3478 return wp_page_shared(vmf, folio);
3479 }
3480
3481 /*
3482 * Private mapping: create an exclusive anonymous page copy if reuse
3483 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3484 *
3485 * If we encounter a page that is marked exclusive, we must reuse
3486 * the page without further checks.
3487 */
3488 if (folio && folio_test_anon(folio) &&
3489 (PageAnonExclusive(page: vmf->page) || wp_can_reuse_anon_folio(folio, vma))) {
3490 if (!PageAnonExclusive(page: vmf->page))
3491 SetPageAnonExclusive(vmf->page);
3492 if (unlikely(unshare)) {
3493 pte_unmap_unlock(vmf->pte, vmf->ptl);
3494 return 0;
3495 }
3496 wp_page_reuse(vmf, folio);
3497 return 0;
3498 }
3499 /*
3500 * Ok, we need to copy. Oh, well..
3501 */
3502 if (folio)
3503 folio_get(folio);
3504
3505 pte_unmap_unlock(vmf->pte, vmf->ptl);
3506#ifdef CONFIG_KSM
3507 if (folio && folio_test_ksm(folio))
3508 count_vm_event(item: COW_KSM);
3509#endif
3510 return wp_page_copy(vmf);
3511}
3512
3513static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3514 unsigned long start_addr, unsigned long end_addr,
3515 struct zap_details *details)
3516{
3517 zap_page_range_single(vma, address: start_addr, size: end_addr - start_addr, details);
3518}
3519
3520static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3521 pgoff_t first_index,
3522 pgoff_t last_index,
3523 struct zap_details *details)
3524{
3525 struct vm_area_struct *vma;
3526 pgoff_t vba, vea, zba, zea;
3527
3528 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3529 vba = vma->vm_pgoff;
3530 vea = vba + vma_pages(vma) - 1;
3531 zba = max(first_index, vba);
3532 zea = min(last_index, vea);
3533
3534 unmap_mapping_range_vma(vma,
3535 start_addr: ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3536 end_addr: ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3537 details);
3538 }
3539}
3540
3541/**
3542 * unmap_mapping_folio() - Unmap single folio from processes.
3543 * @folio: The locked folio to be unmapped.
3544 *
3545 * Unmap this folio from any userspace process which still has it mmaped.
3546 * Typically, for efficiency, the range of nearby pages has already been
3547 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3548 * truncation or invalidation holds the lock on a folio, it may find that
3549 * the page has been remapped again: and then uses unmap_mapping_folio()
3550 * to unmap it finally.
3551 */
3552void unmap_mapping_folio(struct folio *folio)
3553{
3554 struct address_space *mapping = folio->mapping;
3555 struct zap_details details = { };
3556 pgoff_t first_index;
3557 pgoff_t last_index;
3558
3559 VM_BUG_ON(!folio_test_locked(folio));
3560
3561 first_index = folio->index;
3562 last_index = folio_next_index(folio) - 1;
3563
3564 details.even_cows = false;
3565 details.single_folio = folio;
3566 details.zap_flags = ZAP_FLAG_DROP_MARKER;
3567
3568 i_mmap_lock_read(mapping);
3569 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3570 unmap_mapping_range_tree(root: &mapping->i_mmap, first_index,
3571 last_index, details: &details);
3572 i_mmap_unlock_read(mapping);
3573}
3574
3575/**
3576 * unmap_mapping_pages() - Unmap pages from processes.
3577 * @mapping: The address space containing pages to be unmapped.
3578 * @start: Index of first page to be unmapped.
3579 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3580 * @even_cows: Whether to unmap even private COWed pages.
3581 *
3582 * Unmap the pages in this address space from any userspace process which
3583 * has them mmaped. Generally, you want to remove COWed pages as well when
3584 * a file is being truncated, but not when invalidating pages from the page
3585 * cache.
3586 */
3587void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3588 pgoff_t nr, bool even_cows)
3589{
3590 struct zap_details details = { };
3591 pgoff_t first_index = start;
3592 pgoff_t last_index = start + nr - 1;
3593
3594 details.even_cows = even_cows;
3595 if (last_index < first_index)
3596 last_index = ULONG_MAX;
3597
3598 i_mmap_lock_read(mapping);
3599 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3600 unmap_mapping_range_tree(root: &mapping->i_mmap, first_index,
3601 last_index, details: &details);
3602 i_mmap_unlock_read(mapping);
3603}
3604EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3605
3606/**
3607 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3608 * address_space corresponding to the specified byte range in the underlying
3609 * file.
3610 *
3611 * @mapping: the address space containing mmaps to be unmapped.
3612 * @holebegin: byte in first page to unmap, relative to the start of
3613 * the underlying file. This will be rounded down to a PAGE_SIZE
3614 * boundary. Note that this is different from truncate_pagecache(), which
3615 * must keep the partial page. In contrast, we must get rid of
3616 * partial pages.
3617 * @holelen: size of prospective hole in bytes. This will be rounded
3618 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3619 * end of the file.
3620 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3621 * but 0 when invalidating pagecache, don't throw away private data.
3622 */
3623void unmap_mapping_range(struct address_space *mapping,
3624 loff_t const holebegin, loff_t const holelen, int even_cows)
3625{
3626 pgoff_t hba = holebegin >> PAGE_SHIFT;
3627 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3628
3629 /* Check for overflow. */
3630 if (sizeof(holelen) > sizeof(hlen)) {
3631 long long holeend =
3632 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3633 if (holeend & ~(long long)ULONG_MAX)
3634 hlen = ULONG_MAX - hba + 1;
3635 }
3636
3637 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3638}
3639EXPORT_SYMBOL(unmap_mapping_range);
3640
3641/*
3642 * Restore a potential device exclusive pte to a working pte entry
3643 */
3644static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3645{
3646 struct folio *folio = page_folio(vmf->page);
3647 struct vm_area_struct *vma = vmf->vma;
3648 struct mmu_notifier_range range;
3649 vm_fault_t ret;
3650
3651 /*
3652 * We need a reference to lock the folio because we don't hold
3653 * the PTL so a racing thread can remove the device-exclusive
3654 * entry and unmap it. If the folio is free the entry must
3655 * have been removed already. If it happens to have already
3656 * been re-allocated after being freed all we do is lock and
3657 * unlock it.
3658 */
3659 if (!folio_try_get(folio))
3660 return 0;
3661
3662 ret = folio_lock_or_retry(folio, vmf);
3663 if (ret) {
3664 folio_put(folio);
3665 return ret;
3666 }
3667 mmu_notifier_range_init_owner(range: &range, event: MMU_NOTIFY_EXCLUSIVE, flags: 0,
3668 mm: vma->vm_mm, start: vmf->address & PAGE_MASK,
3669 end: (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3670 mmu_notifier_invalidate_range_start(range: &range);
3671
3672 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd, addr: vmf->address,
3673 ptlp: &vmf->ptl);
3674 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3675 restore_exclusive_pte(vma, page: vmf->page, address: vmf->address, ptep: vmf->pte);
3676
3677 if (vmf->pte)
3678 pte_unmap_unlock(vmf->pte, vmf->ptl);
3679 folio_unlock(folio);
3680 folio_put(folio);
3681
3682 mmu_notifier_invalidate_range_end(range: &range);
3683 return 0;
3684}
3685
3686static inline bool should_try_to_free_swap(struct folio *folio,
3687 struct vm_area_struct *vma,
3688 unsigned int fault_flags)
3689{
3690 if (!folio_test_swapcache(folio))
3691 return false;
3692 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
3693 folio_test_mlocked(folio))
3694 return true;
3695 /*
3696 * If we want to map a page that's in the swapcache writable, we
3697 * have to detect via the refcount if we're really the exclusive
3698 * user. Try freeing the swapcache to get rid of the swapcache
3699 * reference only in case it's likely that we'll be the exlusive user.
3700 */
3701 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
3702 folio_ref_count(folio) == 2;
3703}
3704
3705static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
3706{
3707 vmf->pte = pte_offset_map_lock(mm: vmf->vma->vm_mm, pmd: vmf->pmd,
3708 addr: vmf->address, ptlp: &vmf->ptl);
3709 if (!vmf->pte)
3710 return 0;
3711 /*
3712 * Be careful so that we will only recover a special uffd-wp pte into a
3713 * none pte. Otherwise it means the pte could have changed, so retry.
3714 *
3715 * This should also cover the case where e.g. the pte changed
3716 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3717 * So is_pte_marker() check is not enough to safely drop the pte.
3718 */
3719 if (pte_same(a: vmf->orig_pte, b: ptep_get(ptep: vmf->pte)))
3720 pte_clear(mm: vmf->vma->vm_mm, addr: vmf->address, ptep: vmf->pte);
3721 pte_unmap_unlock(vmf->pte, vmf->ptl);
3722 return 0;
3723}
3724
3725static vm_fault_t do_pte_missing(struct vm_fault *vmf)
3726{
3727 if (vma_is_anonymous(vma: vmf->vma))
3728 return do_anonymous_page(vmf);
3729 else
3730 return do_fault(vmf);
3731}
3732
3733/*
3734 * This is actually a page-missing access, but with uffd-wp special pte
3735 * installed. It means this pte was wr-protected before being unmapped.
3736 */
3737static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
3738{
3739 /*
3740 * Just in case there're leftover special ptes even after the region
3741 * got unregistered - we can simply clear them.
3742 */
3743 if (unlikely(!userfaultfd_wp(vmf->vma)))
3744 return pte_marker_clear(vmf);
3745
3746 return do_pte_missing(vmf);
3747}
3748
3749static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
3750{
3751 swp_entry_t entry = pte_to_swp_entry(pte: vmf->orig_pte);
3752 unsigned long marker = pte_marker_get(entry);
3753
3754 /*
3755 * PTE markers should never be empty. If anything weird happened,
3756 * the best thing to do is to kill the process along with its mm.
3757 */
3758 if (WARN_ON_ONCE(!marker))
3759 return VM_FAULT_SIGBUS;
3760
3761 /* Higher priority than uffd-wp when data corrupted */
3762 if (marker & PTE_MARKER_POISONED)
3763 return VM_FAULT_HWPOISON;
3764
3765 if (pte_marker_entry_uffd_wp(entry))
3766 return pte_marker_handle_uffd_wp(vmf);
3767
3768 /* This is an unknown pte marker */
3769 return VM_FAULT_SIGBUS;
3770}
3771
3772/*
3773 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3774 * but allow concurrent faults), and pte mapped but not yet locked.
3775 * We return with pte unmapped and unlocked.
3776 *
3777 * We return with the mmap_lock locked or unlocked in the same cases
3778 * as does filemap_fault().
3779 */
3780vm_fault_t do_swap_page(struct vm_fault *vmf)
3781{
3782 struct vm_area_struct *vma = vmf->vma;
3783 struct folio *swapcache, *folio = NULL;
3784 struct page *page;
3785 struct swap_info_struct *si = NULL;
3786 rmap_t rmap_flags = RMAP_NONE;
3787 bool exclusive = false;
3788 swp_entry_t entry;
3789 pte_t pte;
3790 vm_fault_t ret = 0;
3791 void *shadow = NULL;
3792
3793 if (!pte_unmap_same(vmf))
3794 goto out;
3795
3796 entry = pte_to_swp_entry(pte: vmf->orig_pte);
3797 if (unlikely(non_swap_entry(entry))) {
3798 if (is_migration_entry(entry)) {
3799 migration_entry_wait(mm: vma->vm_mm, pmd: vmf->pmd,
3800 address: vmf->address);
3801 } else if (is_device_exclusive_entry(entry)) {
3802 vmf->page = pfn_swap_entry_to_page(entry);
3803 ret = remove_device_exclusive_entry(vmf);
3804 } else if (is_device_private_entry(entry)) {
3805 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3806 /*
3807 * migrate_to_ram is not yet ready to operate
3808 * under VMA lock.
3809 */
3810 vma_end_read(vma);
3811 ret = VM_FAULT_RETRY;
3812 goto out;
3813 }
3814
3815 vmf->page = pfn_swap_entry_to_page(entry);
3816 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
3817 addr: vmf->address, ptlp: &vmf->ptl);
3818 if (unlikely(!vmf->pte ||
3819 !pte_same(ptep_get(vmf->pte),
3820 vmf->orig_pte)))
3821 goto unlock;
3822
3823 /*
3824 * Get a page reference while we know the page can't be
3825 * freed.
3826 */
3827 get_page(page: vmf->page);
3828 pte_unmap_unlock(vmf->pte, vmf->ptl);
3829 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3830 put_page(page: vmf->page);
3831 } else if (is_hwpoison_entry(entry)) {
3832 ret = VM_FAULT_HWPOISON;
3833 } else if (is_pte_marker_entry(entry)) {
3834 ret = handle_pte_marker(vmf);
3835 } else {
3836 print_bad_pte(vma, addr: vmf->address, pte: vmf->orig_pte, NULL);
3837 ret = VM_FAULT_SIGBUS;
3838 }
3839 goto out;
3840 }
3841
3842 /* Prevent swapoff from happening to us. */
3843 si = get_swap_device(entry);
3844 if (unlikely(!si))
3845 goto out;
3846
3847 folio = swap_cache_get_folio(entry, vma, addr: vmf->address);
3848 if (folio)
3849 page = folio_file_page(folio, index: swp_offset(entry));
3850 swapcache = folio;
3851
3852 if (!folio) {
3853 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3854 __swap_count(entry) == 1) {
3855 /* skip swapcache */
3856 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, order: 0,
3857 vma, addr: vmf->address, hugepage: false);
3858 page = &folio->page;
3859 if (folio) {
3860 __folio_set_locked(folio);
3861 __folio_set_swapbacked(folio);
3862
3863 if (mem_cgroup_swapin_charge_folio(folio,
3864 mm: vma->vm_mm, GFP_KERNEL,
3865 entry)) {
3866 ret = VM_FAULT_OOM;
3867 goto out_page;
3868 }
3869 mem_cgroup_swapin_uncharge_swap(entry);
3870
3871 shadow = get_shadow_from_swap_cache(entry);
3872 if (shadow)
3873 workingset_refault(folio, shadow);
3874
3875 folio_add_lru(folio);
3876
3877 /* To provide entry to swap_readpage() */
3878 folio->swap = entry;
3879 swap_readpage(page, do_poll: true, NULL);
3880 folio->private = NULL;
3881 }
3882 } else {
3883 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3884 vmf);
3885 if (page)
3886 folio = page_folio(page);
3887 swapcache = folio;
3888 }
3889
3890 if (!folio) {
3891 /*
3892 * Back out if somebody else faulted in this pte
3893 * while we released the pte lock.
3894 */
3895 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
3896 addr: vmf->address, ptlp: &vmf->ptl);
3897 if (likely(vmf->pte &&
3898 pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3899 ret = VM_FAULT_OOM;
3900 goto unlock;
3901 }
3902
3903 /* Had to read the page from swap area: Major fault */
3904 ret = VM_FAULT_MAJOR;
3905 count_vm_event(item: PGMAJFAULT);
3906 count_memcg_event_mm(mm: vma->vm_mm, idx: PGMAJFAULT);
3907 } else if (PageHWPoison(page)) {
3908 /*
3909 * hwpoisoned dirty swapcache pages are kept for killing
3910 * owner processes (which may be unknown at hwpoison time)
3911 */
3912 ret = VM_FAULT_HWPOISON;
3913 goto out_release;
3914 }
3915
3916 ret |= folio_lock_or_retry(folio, vmf);
3917 if (ret & VM_FAULT_RETRY)
3918 goto out_release;
3919
3920 if (swapcache) {
3921 /*
3922 * Make sure folio_free_swap() or swapoff did not release the
3923 * swapcache from under us. The page pin, and pte_same test
3924 * below, are not enough to exclude that. Even if it is still
3925 * swapcache, we need to check that the page's swap has not
3926 * changed.
3927 */
3928 if (unlikely(!folio_test_swapcache(folio) ||
3929 page_swap_entry(page).val != entry.val))
3930 goto out_page;
3931
3932 /*
3933 * KSM sometimes has to copy on read faults, for example, if
3934 * page->index of !PageKSM() pages would be nonlinear inside the
3935 * anon VMA -- PageKSM() is lost on actual swapout.
3936 */
3937 page = ksm_might_need_to_copy(page, vma, address: vmf->address);
3938 if (unlikely(!page)) {
3939 ret = VM_FAULT_OOM;
3940 goto out_page;
3941 } else if (unlikely(PTR_ERR(page) == -EHWPOISON)) {
3942 ret = VM_FAULT_HWPOISON;
3943 goto out_page;
3944 }
3945 folio = page_folio(page);
3946
3947 /*
3948 * If we want to map a page that's in the swapcache writable, we
3949 * have to detect via the refcount if we're really the exclusive
3950 * owner. Try removing the extra reference from the local LRU
3951 * caches if required.
3952 */
3953 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
3954 !folio_test_ksm(folio) && !folio_test_lru(folio))
3955 lru_add_drain();
3956 }
3957
3958 folio_throttle_swaprate(folio, GFP_KERNEL);
3959
3960 /*
3961 * Back out if somebody else already faulted in this pte.
3962 */
3963 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd, addr: vmf->address,
3964 ptlp: &vmf->ptl);
3965 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3966 goto out_nomap;
3967
3968 if (unlikely(!folio_test_uptodate(folio))) {
3969 ret = VM_FAULT_SIGBUS;
3970 goto out_nomap;
3971 }
3972
3973 /*
3974 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
3975 * must never point at an anonymous page in the swapcache that is
3976 * PG_anon_exclusive. Sanity check that this holds and especially, that
3977 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
3978 * check after taking the PT lock and making sure that nobody
3979 * concurrently faulted in this page and set PG_anon_exclusive.
3980 */
3981 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
3982 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
3983
3984 /*
3985 * Check under PT lock (to protect against concurrent fork() sharing
3986 * the swap entry concurrently) for certainly exclusive pages.
3987 */
3988 if (!folio_test_ksm(folio)) {
3989 exclusive = pte_swp_exclusive(pte: vmf->orig_pte);
3990 if (folio != swapcache) {
3991 /*
3992 * We have a fresh page that is not exposed to the
3993 * swapcache -> certainly exclusive.
3994 */
3995 exclusive = true;
3996 } else if (exclusive && folio_test_writeback(folio) &&
3997 data_race(si->flags & SWP_STABLE_WRITES)) {
3998 /*
3999 * This is tricky: not all swap backends support
4000 * concurrent page modifications while under writeback.
4001 *
4002 * So if we stumble over such a page in the swapcache
4003 * we must not set the page exclusive, otherwise we can
4004 * map it writable without further checks and modify it
4005 * while still under writeback.
4006 *
4007 * For these problematic swap backends, simply drop the
4008 * exclusive marker: this is perfectly fine as we start
4009 * writeback only if we fully unmapped the page and
4010 * there are no unexpected references on the page after
4011 * unmapping succeeded. After fully unmapped, no
4012 * further GUP references (FOLL_GET and FOLL_PIN) can
4013 * appear, so dropping the exclusive marker and mapping
4014 * it only R/O is fine.
4015 */
4016 exclusive = false;
4017 }
4018 }
4019
4020 /*
4021 * Some architectures may have to restore extra metadata to the page
4022 * when reading from swap. This metadata may be indexed by swap entry
4023 * so this must be called before swap_free().
4024 */
4025 arch_swap_restore(entry, folio);
4026
4027 /*
4028 * Remove the swap entry and conditionally try to free up the swapcache.
4029 * We're already holding a reference on the page but haven't mapped it
4030 * yet.
4031 */
4032 swap_free(entry);
4033 if (should_try_to_free_swap(folio, vma, fault_flags: vmf->flags))
4034 folio_free_swap(folio);
4035
4036 inc_mm_counter(mm: vma->vm_mm, member: MM_ANONPAGES);
4037 dec_mm_counter(mm: vma->vm_mm, member: MM_SWAPENTS);
4038 pte = mk_pte(page, vma->vm_page_prot);
4039
4040 /*
4041 * Same logic as in do_wp_page(); however, optimize for pages that are
4042 * certainly not shared either because we just allocated them without
4043 * exposing them to the swapcache or because the swap entry indicates
4044 * exclusivity.
4045 */
4046 if (!folio_test_ksm(folio) &&
4047 (exclusive || folio_ref_count(folio) == 1)) {
4048 if (vmf->flags & FAULT_FLAG_WRITE) {
4049 pte = maybe_mkwrite(pte: pte_mkdirty(pte), vma);
4050 vmf->flags &= ~FAULT_FLAG_WRITE;
4051 }
4052 rmap_flags |= RMAP_EXCLUSIVE;
4053 }
4054 flush_icache_page(vma, page);
4055 if (pte_swp_soft_dirty(pte: vmf->orig_pte))
4056 pte = pte_mksoft_dirty(pte);
4057 if (pte_swp_uffd_wp(pte: vmf->orig_pte))
4058 pte = pte_mkuffd_wp(pte);
4059 vmf->orig_pte = pte;
4060
4061 /* ksm created a completely new copy */
4062 if (unlikely(folio != swapcache && swapcache)) {
4063 page_add_new_anon_rmap(page, vma, address: vmf->address);
4064 folio_add_lru_vma(folio, vma);
4065 } else {
4066 page_add_anon_rmap(page, vma, address: vmf->address, flags: rmap_flags);
4067 }
4068
4069 VM_BUG_ON(!folio_test_anon(folio) ||
4070 (pte_write(pte) && !PageAnonExclusive(page)));
4071 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
4072 arch_do_swap_page(mm: vma->vm_mm, vma, addr: vmf->address, pte, oldpte: vmf->orig_pte);
4073
4074 folio_unlock(folio);
4075 if (folio != swapcache && swapcache) {
4076 /*
4077 * Hold the lock to avoid the swap entry to be reused
4078 * until we take the PT lock for the pte_same() check
4079 * (to avoid false positives from pte_same). For
4080 * further safety release the lock after the swap_free
4081 * so that the swap count won't change under a
4082 * parallel locked swapcache.
4083 */
4084 folio_unlock(folio: swapcache);
4085 folio_put(folio: swapcache);
4086 }
4087
4088 if (vmf->flags & FAULT_FLAG_WRITE) {
4089 ret |= do_wp_page(vmf);
4090 if (ret & VM_FAULT_ERROR)
4091 ret &= VM_FAULT_ERROR;
4092 goto out;
4093 }
4094
4095 /* No need to invalidate - it was non-present before */
4096 update_mmu_cache_range(vmf, vma, addr: vmf->address, ptep: vmf->pte, nr: 1);
4097unlock:
4098 if (vmf->pte)
4099 pte_unmap_unlock(vmf->pte, vmf->ptl);
4100out:
4101 if (si)
4102 put_swap_device(si);
4103 return ret;
4104out_nomap:
4105 if (vmf->pte)
4106 pte_unmap_unlock(vmf->pte, vmf->ptl);
4107out_page:
4108 folio_unlock(folio);
4109out_release:
4110 folio_put(folio);
4111 if (folio != swapcache && swapcache) {
4112 folio_unlock(folio: swapcache);
4113 folio_put(folio: swapcache);
4114 }
4115 if (si)
4116 put_swap_device(si);
4117 return ret;
4118}
4119
4120/*
4121 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4122 * but allow concurrent faults), and pte mapped but not yet locked.
4123 * We return with mmap_lock still held, but pte unmapped and unlocked.
4124 */
4125static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
4126{
4127 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4128 struct vm_area_struct *vma = vmf->vma;
4129 struct folio *folio;
4130 vm_fault_t ret = 0;
4131 pte_t entry;
4132
4133 /* File mapping without ->vm_ops ? */
4134 if (vma->vm_flags & VM_SHARED)
4135 return VM_FAULT_SIGBUS;
4136
4137 /*
4138 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4139 * be distinguished from a transient failure of pte_offset_map().
4140 */
4141 if (pte_alloc(vma->vm_mm, vmf->pmd))
4142 return VM_FAULT_OOM;
4143
4144 /* Use the zero-page for reads */
4145 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
4146 !mm_forbids_zeropage(vma->vm_mm)) {
4147 entry = pte_mkspecial(pte: pfn_pte(page_nr: my_zero_pfn(addr: vmf->address),
4148 pgprot: vma->vm_page_prot));
4149 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
4150 addr: vmf->address, ptlp: &vmf->ptl);
4151 if (!vmf->pte)
4152 goto unlock;
4153 if (vmf_pte_changed(vmf)) {
4154 update_mmu_tlb(vma, address: vmf->address, ptep: vmf->pte);
4155 goto unlock;
4156 }
4157 ret = check_stable_address_space(mm: vma->vm_mm);
4158 if (ret)
4159 goto unlock;
4160 /* Deliver the page fault to userland, check inside PT lock */
4161 if (userfaultfd_missing(vma)) {
4162 pte_unmap_unlock(vmf->pte, vmf->ptl);
4163 return handle_userfault(vmf, VM_UFFD_MISSING);
4164 }
4165 goto setpte;
4166 }
4167
4168 /* Allocate our own private page. */
4169 if (unlikely(anon_vma_prepare(vma)))
4170 goto oom;
4171 folio = vma_alloc_zeroed_movable_folio(vma, vmf->address);
4172 if (!folio)
4173 goto oom;
4174
4175 if (mem_cgroup_charge(folio, mm: vma->vm_mm, GFP_KERNEL))
4176 goto oom_free_page;
4177 folio_throttle_swaprate(folio, GFP_KERNEL);
4178
4179 /*
4180 * The memory barrier inside __folio_mark_uptodate makes sure that
4181 * preceding stores to the page contents become visible before
4182 * the set_pte_at() write.
4183 */
4184 __folio_mark_uptodate(folio);
4185
4186 entry = mk_pte(&folio->page, vma->vm_page_prot);
4187 entry = pte_sw_mkyoung(pte: entry);
4188 if (vma->vm_flags & VM_WRITE)
4189 entry = pte_mkwrite(pte: pte_mkdirty(pte: entry), vma);
4190
4191 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd, addr: vmf->address,
4192 ptlp: &vmf->ptl);
4193 if (!vmf->pte)
4194 goto release;
4195 if (vmf_pte_changed(vmf)) {
4196 update_mmu_tlb(vma, address: vmf->address, ptep: vmf->pte);
4197 goto release;
4198 }
4199
4200 ret = check_stable_address_space(mm: vma->vm_mm);
4201 if (ret)
4202 goto release;
4203
4204 /* Deliver the page fault to userland, check inside PT lock */
4205 if (userfaultfd_missing(vma)) {
4206 pte_unmap_unlock(vmf->pte, vmf->ptl);
4207 folio_put(folio);
4208 return handle_userfault(vmf, VM_UFFD_MISSING);
4209 }
4210
4211 inc_mm_counter(mm: vma->vm_mm, member: MM_ANONPAGES);
4212 folio_add_new_anon_rmap(folio, vma, address: vmf->address);
4213 folio_add_lru_vma(folio, vma);
4214setpte:
4215 if (uffd_wp)
4216 entry = pte_mkuffd_wp(pte: entry);
4217 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
4218
4219 /* No need to invalidate - it was non-present before */
4220 update_mmu_cache_range(vmf, vma, addr: vmf->address, ptep: vmf->pte, nr: 1);
4221unlock:
4222 if (vmf->pte)
4223 pte_unmap_unlock(vmf->pte, vmf->ptl);
4224 return ret;
4225release:
4226 folio_put(folio);
4227 goto unlock;
4228oom_free_page:
4229 folio_put(folio);
4230oom:
4231 return VM_FAULT_OOM;
4232}
4233
4234/*
4235 * The mmap_lock must have been held on entry, and may have been
4236 * released depending on flags and vma->vm_ops->fault() return value.
4237 * See filemap_fault() and __lock_page_retry().
4238 */
4239static vm_fault_t __do_fault(struct vm_fault *vmf)
4240{
4241 struct vm_area_struct *vma = vmf->vma;
4242 vm_fault_t ret;
4243
4244 /*
4245 * Preallocate pte before we take page_lock because this might lead to
4246 * deadlocks for memcg reclaim which waits for pages under writeback:
4247 * lock_page(A)
4248 * SetPageWriteback(A)
4249 * unlock_page(A)
4250 * lock_page(B)
4251 * lock_page(B)
4252 * pte_alloc_one
4253 * shrink_page_list
4254 * wait_on_page_writeback(A)
4255 * SetPageWriteback(B)
4256 * unlock_page(B)
4257 * # flush A, B to clear the writeback
4258 */
4259 if (pmd_none(pmd: *vmf->pmd) && !vmf->prealloc_pte) {
4260 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4261 if (!vmf->prealloc_pte)
4262 return VM_FAULT_OOM;
4263 }
4264
4265 ret = vma->vm_ops->fault(vmf);
4266 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
4267 VM_FAULT_DONE_COW)))
4268 return ret;
4269
4270 if (unlikely(PageHWPoison(vmf->page))) {
4271 struct page *page = vmf->page;
4272 vm_fault_t poisonret = VM_FAULT_HWPOISON;
4273 if (ret & VM_FAULT_LOCKED) {
4274 if (page_mapped(page))
4275 unmap_mapping_pages(page_mapping(page),
4276 page->index, 1, false);
4277 /* Retry if a clean page was removed from the cache. */
4278 if (invalidate_inode_page(page))
4279 poisonret = VM_FAULT_NOPAGE;
4280 unlock_page(page);
4281 }
4282 put_page(page);
4283 vmf->page = NULL;
4284 return poisonret;
4285 }
4286
4287 if (unlikely(!(ret & VM_FAULT_LOCKED)))
4288 lock_page(page: vmf->page);
4289 else
4290 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
4291
4292 return ret;
4293}
4294
4295#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4296static void deposit_prealloc_pte(struct vm_fault *vmf)
4297{
4298 struct vm_area_struct *vma = vmf->vma;
4299
4300 pgtable_trans_huge_deposit(mm: vma->vm_mm, pmdp: vmf->pmd, pgtable: vmf->prealloc_pte);
4301 /*
4302 * We are going to consume the prealloc table,
4303 * count that as nr_ptes.
4304 */
4305 mm_inc_nr_ptes(mm: vma->vm_mm);
4306 vmf->prealloc_pte = NULL;
4307}
4308
4309vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4310{
4311 struct vm_area_struct *vma = vmf->vma;
4312 bool write = vmf->flags & FAULT_FLAG_WRITE;
4313 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
4314 pmd_t entry;
4315 vm_fault_t ret = VM_FAULT_FALLBACK;
4316
4317 if (!transhuge_vma_suitable(vma, addr: haddr))
4318 return ret;
4319
4320 page = compound_head(page);
4321 if (compound_order(page) != HPAGE_PMD_ORDER)
4322 return ret;
4323
4324 /*
4325 * Just backoff if any subpage of a THP is corrupted otherwise
4326 * the corrupted page may mapped by PMD silently to escape the
4327 * check. This kind of THP just can be PTE mapped. Access to
4328 * the corrupted subpage should trigger SIGBUS as expected.
4329 */
4330 if (unlikely(PageHasHWPoisoned(page)))
4331 return ret;
4332
4333 /*
4334 * Archs like ppc64 need additional space to store information
4335 * related to pte entry. Use the preallocated table for that.
4336 */
4337 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
4338 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4339 if (!vmf->prealloc_pte)
4340 return VM_FAULT_OOM;
4341 }
4342
4343 vmf->ptl = pmd_lock(mm: vma->vm_mm, pmd: vmf->pmd);
4344 if (unlikely(!pmd_none(*vmf->pmd)))
4345 goto out;
4346
4347 flush_icache_pages(vma, page, HPAGE_PMD_NR);
4348
4349 entry = mk_huge_pmd(page, vma->vm_page_prot);
4350 if (write)
4351 entry = maybe_pmd_mkwrite(pmd: pmd_mkdirty(pmd: entry), vma);
4352
4353 add_mm_counter(mm: vma->vm_mm, member: mm_counter_file(page), HPAGE_PMD_NR);
4354 page_add_file_rmap(page, vma, compound: true);
4355
4356 /*
4357 * deposit and withdraw with pmd lock held
4358 */
4359 if (arch_needs_pgtable_deposit())
4360 deposit_prealloc_pte(vmf);
4361
4362 set_pmd_at(mm: vma->vm_mm, addr: haddr, pmdp: vmf->pmd, pmd: entry);
4363
4364 update_mmu_cache_pmd(vma, addr: haddr, pmd: vmf->pmd);
4365
4366 /* fault is handled */
4367 ret = 0;
4368 count_vm_event(item: THP_FILE_MAPPED);
4369out:
4370 spin_unlock(lock: vmf->ptl);
4371 return ret;
4372}
4373#else
4374vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4375{
4376 return VM_FAULT_FALLBACK;
4377}
4378#endif
4379
4380/**
4381 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4382 * @vmf: Fault decription.
4383 * @folio: The folio that contains @page.
4384 * @page: The first page to create a PTE for.
4385 * @nr: The number of PTEs to create.
4386 * @addr: The first address to create a PTE for.
4387 */
4388void set_pte_range(struct vm_fault *vmf, struct folio *folio,
4389 struct page *page, unsigned int nr, unsigned long addr)
4390{
4391 struct vm_area_struct *vma = vmf->vma;
4392 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4393 bool write = vmf->flags & FAULT_FLAG_WRITE;
4394 bool prefault = in_range(vmf->address, addr, nr * PAGE_SIZE);
4395 pte_t entry;
4396
4397 flush_icache_pages(vma, page, nr);
4398 entry = mk_pte(page, vma->vm_page_prot);
4399
4400 if (prefault && arch_wants_old_prefaulted_pte())
4401 entry = pte_mkold(pte: entry);
4402 else
4403 entry = pte_sw_mkyoung(pte: entry);
4404
4405 if (write)
4406 entry = maybe_mkwrite(pte: pte_mkdirty(pte: entry), vma);
4407 if (unlikely(uffd_wp))
4408 entry = pte_mkuffd_wp(pte: entry);
4409 /* copy-on-write page */
4410 if (write && !(vma->vm_flags & VM_SHARED)) {
4411 add_mm_counter(mm: vma->vm_mm, member: MM_ANONPAGES, value: nr);
4412 VM_BUG_ON_FOLIO(nr != 1, folio);
4413 folio_add_new_anon_rmap(folio, vma, address: addr);
4414 folio_add_lru_vma(folio, vma);
4415 } else {
4416 add_mm_counter(mm: vma->vm_mm, member: mm_counter_file(page), value: nr);
4417 folio_add_file_rmap_range(folio, page, nr, vma, compound: false);
4418 }
4419 set_ptes(mm: vma->vm_mm, addr, ptep: vmf->pte, pte: entry, nr);
4420
4421 /* no need to invalidate: a not-present page won't be cached */
4422 update_mmu_cache_range(vmf, vma, addr, ptep: vmf->pte, nr);
4423}
4424
4425static bool vmf_pte_changed(struct vm_fault *vmf)
4426{
4427 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
4428 return !pte_same(a: ptep_get(ptep: vmf->pte), b: vmf->orig_pte);
4429
4430 return !pte_none(pte: ptep_get(ptep: vmf->pte));
4431}
4432
4433/**
4434 * finish_fault - finish page fault once we have prepared the page to fault
4435 *
4436 * @vmf: structure describing the fault
4437 *
4438 * This function handles all that is needed to finish a page fault once the
4439 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4440 * given page, adds reverse page mapping, handles memcg charges and LRU
4441 * addition.
4442 *
4443 * The function expects the page to be locked and on success it consumes a
4444 * reference of a page being mapped (for the PTE which maps it).
4445 *
4446 * Return: %0 on success, %VM_FAULT_ code in case of error.
4447 */
4448vm_fault_t finish_fault(struct vm_fault *vmf)
4449{
4450 struct vm_area_struct *vma = vmf->vma;
4451 struct page *page;
4452 vm_fault_t ret;
4453
4454 /* Did we COW the page? */
4455 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4456 page = vmf->cow_page;
4457 else
4458 page = vmf->page;
4459
4460 /*
4461 * check even for read faults because we might have lost our CoWed
4462 * page
4463 */
4464 if (!(vma->vm_flags & VM_SHARED)) {
4465 ret = check_stable_address_space(mm: vma->vm_mm);
4466 if (ret)
4467 return ret;
4468 }
4469
4470 if (pmd_none(pmd: *vmf->pmd)) {
4471 if (PageTransCompound(page)) {
4472 ret = do_set_pmd(vmf, page);
4473 if (ret != VM_FAULT_FALLBACK)
4474 return ret;
4475 }
4476
4477 if (vmf->prealloc_pte)
4478 pmd_install(mm: vma->vm_mm, pmd: vmf->pmd, pte: &vmf->prealloc_pte);
4479 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4480 return VM_FAULT_OOM;
4481 }
4482
4483 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
4484 addr: vmf->address, ptlp: &vmf->ptl);
4485 if (!vmf->pte)
4486 return VM_FAULT_NOPAGE;
4487
4488 /* Re-check under ptl */
4489 if (likely(!vmf_pte_changed(vmf))) {
4490 struct folio *folio = page_folio(page);
4491
4492 set_pte_range(vmf, folio, page, nr: 1, addr: vmf->address);
4493 ret = 0;
4494 } else {
4495 update_mmu_tlb(vma, address: vmf->address, ptep: vmf->pte);
4496 ret = VM_FAULT_NOPAGE;
4497 }
4498
4499 pte_unmap_unlock(vmf->pte, vmf->ptl);
4500 return ret;
4501}
4502
4503static unsigned long fault_around_pages __read_mostly =
4504 65536 >> PAGE_SHIFT;
4505
4506#ifdef CONFIG_DEBUG_FS
4507static int fault_around_bytes_get(void *data, u64 *val)
4508{
4509 *val = fault_around_pages << PAGE_SHIFT;
4510 return 0;
4511}
4512
4513/*
4514 * fault_around_bytes must be rounded down to the nearest page order as it's
4515 * what do_fault_around() expects to see.
4516 */
4517static int fault_around_bytes_set(void *data, u64 val)
4518{
4519 if (val / PAGE_SIZE > PTRS_PER_PTE)
4520 return -EINVAL;
4521
4522 /*
4523 * The minimum value is 1 page, however this results in no fault-around
4524 * at all. See should_fault_around().
4525 */
4526 fault_around_pages = max(rounddown_pow_of_two(val) >> PAGE_SHIFT, 1UL);
4527
4528 return 0;
4529}
4530DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4531 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4532
4533static int __init fault_around_debugfs(void)
4534{
4535 debugfs_create_file_unsafe(name: "fault_around_bytes", mode: 0644, NULL, NULL,
4536 fops: &fault_around_bytes_fops);
4537 return 0;
4538}
4539late_initcall(fault_around_debugfs);
4540#endif
4541
4542/*
4543 * do_fault_around() tries to map few pages around the fault address. The hope
4544 * is that the pages will be needed soon and this will lower the number of
4545 * faults to handle.
4546 *
4547 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4548 * not ready to be mapped: not up-to-date, locked, etc.
4549 *
4550 * This function doesn't cross VMA or page table boundaries, in order to call
4551 * map_pages() and acquire a PTE lock only once.
4552 *
4553 * fault_around_pages defines how many pages we'll try to map.
4554 * do_fault_around() expects it to be set to a power of two less than or equal
4555 * to PTRS_PER_PTE.
4556 *
4557 * The virtual address of the area that we map is naturally aligned to
4558 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4559 * (and therefore to page order). This way it's easier to guarantee
4560 * that we don't cross page table boundaries.
4561 */
4562static vm_fault_t do_fault_around(struct vm_fault *vmf)
4563{
4564 pgoff_t nr_pages = READ_ONCE(fault_around_pages);
4565 pgoff_t pte_off = pte_index(address: vmf->address);
4566 /* The page offset of vmf->address within the VMA. */
4567 pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
4568 pgoff_t from_pte, to_pte;
4569 vm_fault_t ret;
4570
4571 /* The PTE offset of the start address, clamped to the VMA. */
4572 from_pte = max(ALIGN_DOWN(pte_off, nr_pages),
4573 pte_off - min(pte_off, vma_off));
4574
4575 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4576 to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE,
4577 pte_off + vma_pages(vmf->vma) - vma_off) - 1;
4578
4579 if (pmd_none(pmd: *vmf->pmd)) {
4580 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4581 if (!vmf->prealloc_pte)
4582 return VM_FAULT_OOM;
4583 }
4584
4585 rcu_read_lock();
4586 ret = vmf->vma->vm_ops->map_pages(vmf,
4587 vmf->pgoff + from_pte - pte_off,
4588 vmf->pgoff + to_pte - pte_off);
4589 rcu_read_unlock();
4590
4591 return ret;
4592}
4593
4594/* Return true if we should do read fault-around, false otherwise */
4595static inline bool should_fault_around(struct vm_fault *vmf)
4596{
4597 /* No ->map_pages? No way to fault around... */
4598 if (!vmf->vma->vm_ops->map_pages)
4599 return false;
4600
4601 if (uffd_disable_fault_around(vma: vmf->vma))
4602 return false;
4603
4604 /* A single page implies no faulting 'around' at all. */
4605 return fault_around_pages > 1;
4606}
4607
4608static vm_fault_t do_read_fault(struct vm_fault *vmf)
4609{
4610 vm_fault_t ret = 0;
4611 struct folio *folio;
4612
4613 /*
4614 * Let's call ->map_pages() first and use ->fault() as fallback
4615 * if page by the offset is not ready to be mapped (cold cache or
4616 * something).
4617 */
4618 if (should_fault_around(vmf)) {
4619 ret = do_fault_around(vmf);
4620 if (ret)
4621 return ret;
4622 }
4623
4624 ret = vmf_can_call_fault(vmf);
4625 if (ret)
4626 return ret;
4627
4628 ret = __do_fault(vmf);
4629 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4630 return ret;
4631
4632 ret |= finish_fault(vmf);
4633 folio = page_folio(vmf->page);
4634 folio_unlock(folio);
4635 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4636 folio_put(folio);
4637 return ret;
4638}
4639
4640static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4641{
4642 struct vm_area_struct *vma = vmf->vma;
4643 vm_fault_t ret;
4644
4645 ret = vmf_can_call_fault(vmf);
4646 if (!ret)
4647 ret = vmf_anon_prepare(vmf);
4648 if (ret)
4649 return ret;
4650
4651 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr: vmf->address);
4652 if (!vmf->cow_page)
4653 return VM_FAULT_OOM;
4654
4655 if (mem_cgroup_charge(page_folio(vmf->cow_page), mm: vma->vm_mm,
4656 GFP_KERNEL)) {
4657 put_page(page: vmf->cow_page);
4658 return VM_FAULT_OOM;
4659 }
4660 folio_throttle_swaprate(page_folio(vmf->cow_page), GFP_KERNEL);
4661
4662 ret = __do_fault(vmf);
4663 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4664 goto uncharge_out;
4665 if (ret & VM_FAULT_DONE_COW)
4666 return ret;
4667
4668 copy_user_highpage(to: vmf->cow_page, from: vmf->page, vaddr: vmf->address, vma);
4669 __SetPageUptodate(page: vmf->cow_page);
4670
4671 ret |= finish_fault(vmf);
4672 unlock_page(page: vmf->page);
4673 put_page(page: vmf->page);
4674 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4675 goto uncharge_out;
4676 return ret;
4677uncharge_out:
4678 put_page(page: vmf->cow_page);
4679 return ret;
4680}
4681
4682static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4683{
4684 struct vm_area_struct *vma = vmf->vma;
4685 vm_fault_t ret, tmp;
4686 struct folio *folio;
4687
4688 ret = vmf_can_call_fault(vmf);
4689 if (ret)
4690 return ret;
4691
4692 ret = __do_fault(vmf);
4693 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4694 return ret;
4695
4696 folio = page_folio(vmf->page);
4697
4698 /*
4699 * Check if the backing address space wants to know that the page is
4700 * about to become writable
4701 */
4702 if (vma->vm_ops->page_mkwrite) {
4703 folio_unlock(folio);
4704 tmp = do_page_mkwrite(vmf, folio);
4705 if (unlikely(!tmp ||
4706 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4707 folio_put(folio);
4708 return tmp;
4709 }
4710 }
4711
4712 ret |= finish_fault(vmf);
4713 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4714 VM_FAULT_RETRY))) {
4715 folio_unlock(folio);
4716 folio_put(folio);
4717 return ret;
4718 }
4719
4720 ret |= fault_dirty_shared_page(vmf);
4721 return ret;
4722}
4723
4724/*
4725 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4726 * but allow concurrent faults).
4727 * The mmap_lock may have been released depending on flags and our
4728 * return value. See filemap_fault() and __folio_lock_or_retry().
4729 * If mmap_lock is released, vma may become invalid (for example
4730 * by other thread calling munmap()).
4731 */
4732static vm_fault_t do_fault(struct vm_fault *vmf)
4733{
4734 struct vm_area_struct *vma = vmf->vma;
4735 struct mm_struct *vm_mm = vma->vm_mm;
4736 vm_fault_t ret;
4737
4738 /*
4739 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4740 */
4741 if (!vma->vm_ops->fault) {
4742 vmf->pte = pte_offset_map_lock(mm: vmf->vma->vm_mm, pmd: vmf->pmd,
4743 addr: vmf->address, ptlp: &vmf->ptl);
4744 if (unlikely(!vmf->pte))
4745 ret = VM_FAULT_SIGBUS;
4746 else {
4747 /*
4748 * Make sure this is not a temporary clearing of pte
4749 * by holding ptl and checking again. A R/M/W update
4750 * of pte involves: take ptl, clearing the pte so that
4751 * we don't have concurrent modification by hardware
4752 * followed by an update.
4753 */
4754 if (unlikely(pte_none(ptep_get(vmf->pte))))
4755 ret = VM_FAULT_SIGBUS;
4756 else
4757 ret = VM_FAULT_NOPAGE;
4758
4759 pte_unmap_unlock(vmf->pte, vmf->ptl);
4760 }
4761 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4762 ret = do_read_fault(vmf);
4763 else if (!(vma->vm_flags & VM_SHARED))
4764 ret = do_cow_fault(vmf);
4765 else
4766 ret = do_shared_fault(vmf);
4767
4768 /* preallocated pagetable is unused: free it */
4769 if (vmf->prealloc_pte) {
4770 pte_free(mm: vm_mm, pte_page: vmf->prealloc_pte);
4771 vmf->prealloc_pte = NULL;
4772 }
4773 return ret;
4774}
4775
4776int numa_migrate_prep(struct folio *folio, struct vm_area_struct *vma,
4777 unsigned long addr, int page_nid, int *flags)
4778{
4779 folio_get(folio);
4780
4781 /* Record the current PID acceesing VMA */
4782 vma_set_access_pid_bit(vma);
4783
4784 count_vm_numa_event(NUMA_HINT_FAULTS);
4785 if (page_nid == numa_node_id()) {
4786 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4787 *flags |= TNF_FAULT_LOCAL;
4788 }
4789
4790 return mpol_misplaced(folio, vma, addr);
4791}
4792
4793static vm_fault_t do_numa_page(struct vm_fault *vmf)
4794{
4795 struct vm_area_struct *vma = vmf->vma;
4796 struct folio *folio = NULL;
4797 int nid = NUMA_NO_NODE;
4798 bool writable = false;
4799 int last_cpupid;
4800 int target_nid;
4801 pte_t pte, old_pte;
4802 int flags = 0;
4803
4804 /*
4805 * The "pte" at this point cannot be used safely without
4806 * validation through pte_unmap_same(). It's of NUMA type but
4807 * the pfn may be screwed if the read is non atomic.
4808 */
4809 spin_lock(lock: vmf->ptl);
4810 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
4811 pte_unmap_unlock(vmf->pte, vmf->ptl);
4812 goto out;
4813 }
4814
4815 /* Get the normal PTE */
4816 old_pte = ptep_get(ptep: vmf->pte);
4817 pte = pte_modify(pte: old_pte, newprot: vma->vm_page_prot);
4818
4819 /*
4820 * Detect now whether the PTE could be writable; this information
4821 * is only valid while holding the PT lock.
4822 */
4823 writable = pte_write(pte);
4824 if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
4825 can_change_pte_writable(vma, addr: vmf->address, pte))
4826 writable = true;
4827
4828 folio = vm_normal_folio(vma, addr: vmf->address, pte);
4829 if (!folio || folio_is_zone_device(folio))
4830 goto out_map;
4831
4832 /* TODO: handle PTE-mapped THP */
4833 if (folio_test_large(folio))
4834 goto out_map;
4835
4836 /*
4837 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4838 * much anyway since they can be in shared cache state. This misses
4839 * the case where a mapping is writable but the process never writes
4840 * to it but pte_write gets cleared during protection updates and
4841 * pte_dirty has unpredictable behaviour between PTE scan updates,
4842 * background writeback, dirty balancing and application behaviour.
4843 */
4844 if (!writable)
4845 flags |= TNF_NO_GROUP;
4846
4847 /*
4848 * Flag if the folio is shared between multiple address spaces. This
4849 * is later used when determining whether to group tasks together
4850 */
4851 if (folio_estimated_sharers(folio) > 1 && (vma->vm_flags & VM_SHARED))
4852 flags |= TNF_SHARED;
4853
4854 nid = folio_nid(folio);
4855 /*
4856 * For memory tiering mode, cpupid of slow memory page is used
4857 * to record page access time. So use default value.
4858 */
4859 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4860 !node_is_toptier(node: nid))
4861 last_cpupid = (-1 & LAST_CPUPID_MASK);
4862 else
4863 last_cpupid = folio_last_cpupid(folio);
4864 target_nid = numa_migrate_prep(folio, vma, addr: vmf->address, page_nid: nid, flags: &flags);
4865 if (target_nid == NUMA_NO_NODE) {
4866 folio_put(folio);
4867 goto out_map;
4868 }
4869 pte_unmap_unlock(vmf->pte, vmf->ptl);
4870 writable = false;
4871
4872 /* Migrate to the requested node */
4873 if (migrate_misplaced_folio(folio, vma, node: target_nid)) {
4874 nid = target_nid;
4875 flags |= TNF_MIGRATED;
4876 } else {
4877 flags |= TNF_MIGRATE_FAIL;
4878 vmf->pte = pte_offset_map_lock(mm: vma->vm_mm, pmd: vmf->pmd,
4879 addr: vmf->address, ptlp: &vmf->ptl);
4880 if (unlikely(!vmf->pte))
4881 goto out;
4882 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
4883 pte_unmap_unlock(vmf->pte, vmf->ptl);
4884 goto out;
4885 }
4886 goto out_map;
4887 }
4888
4889out:
4890 if (nid != NUMA_NO_NODE)
4891 task_numa_fault(last_node: last_cpupid, node: nid, pages: 1, flags);
4892 return 0;
4893out_map:
4894 /*
4895 * Make it present again, depending on how arch implements
4896 * non-accessible ptes, some can allow access by kernel mode.
4897 */
4898 old_pte = ptep_modify_prot_start(vma, addr: vmf->address, ptep: vmf->pte);
4899 pte = pte_modify(pte: old_pte, newprot: vma->vm_page_prot);
4900 pte = pte_mkyoung(pte);
4901 if (writable)
4902 pte = pte_mkwrite(pte, vma);
4903 ptep_modify_prot_commit(vma, addr: vmf->address, ptep: vmf->pte, old_pte, pte);
4904 update_mmu_cache_range(vmf, vma, addr: vmf->address, ptep: vmf->pte, nr: 1);
4905 pte_unmap_unlock(vmf->pte, vmf->ptl);
4906 goto out;
4907}
4908
4909static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4910{
4911 struct vm_area_struct *vma = vmf->vma;
4912 if (vma_is_anonymous(vma))
4913 return do_huge_pmd_anonymous_page(vmf);
4914 if (vma->vm_ops->huge_fault)
4915 return vma->vm_ops->huge_fault(vmf, PMD_ORDER);
4916 return VM_FAULT_FALLBACK;
4917}
4918
4919/* `inline' is required to avoid gcc 4.1.2 build error */
4920static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4921{
4922 struct vm_area_struct *vma = vmf->vma;
4923 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
4924 vm_fault_t ret;
4925
4926 if (vma_is_anonymous(vma)) {
4927 if (likely(!unshare) &&
4928 userfaultfd_huge_pmd_wp(vma, pmd: vmf->orig_pmd)) {
4929 if (userfaultfd_wp_async(vma: vmf->vma))
4930 goto split;
4931 return handle_userfault(vmf, VM_UFFD_WP);
4932 }
4933 return do_huge_pmd_wp_page(vmf);
4934 }
4935
4936 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
4937 if (vma->vm_ops->huge_fault) {
4938 ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER);
4939 if (!(ret & VM_FAULT_FALLBACK))
4940 return ret;
4941 }
4942 }
4943
4944split:
4945 /* COW or write-notify handled on pte level: split pmd. */
4946 __split_huge_pmd(vma, pmd: vmf->pmd, address: vmf->address, freeze: false, NULL);
4947
4948 return VM_FAULT_FALLBACK;
4949}
4950
4951static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4952{
4953#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4954 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4955 struct vm_area_struct *vma = vmf->vma;
4956 /* No support for anonymous transparent PUD pages yet */
4957 if (vma_is_anonymous(vma))
4958 return VM_FAULT_FALLBACK;
4959 if (vma->vm_ops->huge_fault)
4960 return vma->vm_ops->huge_fault(vmf, PUD_ORDER);
4961#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4962 return VM_FAULT_FALLBACK;
4963}
4964
4965static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4966{
4967#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4968 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4969 struct vm_area_struct *vma = vmf->vma;
4970 vm_fault_t ret;
4971
4972 /* No support for anonymous transparent PUD pages yet */
4973 if (vma_is_anonymous(vma))
4974 goto split;
4975 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
4976 if (vma->vm_ops->huge_fault) {
4977 ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER);
4978 if (!(ret & VM_FAULT_FALLBACK))
4979 return ret;
4980 }
4981 }
4982split:
4983 /* COW or write-notify not handled on PUD level: split pud.*/
4984 __split_huge_pud(vma, pud: vmf->pud, address: vmf->address);
4985#endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
4986 return VM_FAULT_FALLBACK;
4987}
4988
4989/*
4990 * These routines also need to handle stuff like marking pages dirty
4991 * and/or accessed for architectures that don't do it in hardware (most
4992 * RISC architectures). The early dirtying is also good on the i386.
4993 *
4994 * There is also a hook called "update_mmu_cache()" that architectures
4995 * with external mmu caches can use to update those (ie the Sparc or
4996 * PowerPC hashed page tables that act as extended TLBs).
4997 *
4998 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4999 * concurrent faults).
5000 *
5001 * The mmap_lock may have been released depending on flags and our return value.
5002 * See filemap_fault() and __folio_lock_or_retry().
5003 */
5004static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
5005{
5006 pte_t entry;
5007
5008 if (unlikely(pmd_none(*vmf->pmd))) {
5009 /*
5010 * Leave __pte_alloc() until later: because vm_ops->fault may
5011 * want to allocate huge page, and if we expose page table
5012 * for an instant, it will be difficult to retract from
5013 * concurrent faults and from rmap lookups.
5014 */
5015 vmf->pte = NULL;
5016 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
5017 } else {
5018 /*
5019 * A regular pmd is established and it can't morph into a huge
5020 * pmd by anon khugepaged, since that takes mmap_lock in write
5021 * mode; but shmem or file collapse to THP could still morph
5022 * it into a huge pmd: just retry later if so.
5023 */
5024 vmf->pte = pte_offset_map_nolock(mm: vmf->vma->vm_mm, pmd: vmf->pmd,
5025 addr: vmf->address, ptlp: &vmf->ptl);
5026 if (unlikely(!vmf->pte))
5027 return 0;
5028 vmf->orig_pte = ptep_get_lockless(ptep: vmf->pte);
5029 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
5030
5031 if (pte_none(pte: vmf->orig_pte)) {
5032 pte_unmap(pte: vmf->pte);
5033 vmf->pte = NULL;
5034 }
5035 }
5036
5037 if (!vmf->pte)
5038 return do_pte_missing(vmf);
5039
5040 if (!pte_present(a: vmf->orig_pte))
5041 return do_swap_page(vmf);
5042
5043 if (pte_protnone(pte: vmf->orig_pte) && vma_is_accessible(vma: vmf->vma))
5044 return do_numa_page(vmf);
5045
5046 spin_lock(lock: vmf->ptl);
5047 entry = vmf->orig_pte;
5048 if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) {
5049 update_mmu_tlb(vma: vmf->vma, address: vmf->address, ptep: vmf->pte);
5050 goto unlock;
5051 }
5052 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
5053 if (!pte_write(pte: entry))
5054 return do_wp_page(vmf);
5055 else if (likely(vmf->flags & FAULT_FLAG_WRITE))
5056 entry = pte_mkdirty(pte: entry);
5057 }
5058 entry = pte_mkyoung(pte: entry);
5059 if (ptep_set_access_flags(vma: vmf->vma, address: vmf->address, ptep: vmf->pte, entry,
5060 dirty: vmf->flags & FAULT_FLAG_WRITE)) {
5061 update_mmu_cache_range(vmf, vma: vmf->vma, addr: vmf->address,
5062 ptep: vmf->pte, nr: 1);
5063 } else {
5064 /* Skip spurious TLB flush for retried page fault */
5065 if (vmf->flags & FAULT_FLAG_TRIED)
5066 goto unlock;
5067 /*
5068 * This is needed only for protection faults but the arch code
5069 * is not yet telling us if this is a protection fault or not.
5070 * This still avoids useless tlb flushes for .text page faults
5071 * with threads.
5072 */
5073 if (vmf->flags & FAULT_FLAG_WRITE)
5074 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address,
5075 vmf->pte);
5076 }
5077unlock:
5078 pte_unmap_unlock(vmf->pte, vmf->ptl);
5079 return 0;
5080}
5081
5082/*
5083 * On entry, we hold either the VMA lock or the mmap_lock
5084 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5085 * the result, the mmap_lock is not held on exit. See filemap_fault()
5086 * and __folio_lock_or_retry().
5087 */
5088static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
5089 unsigned long address, unsigned int flags)
5090{
5091 struct vm_fault vmf = {
5092 .vma = vma,
5093 .address = address & PAGE_MASK,
5094 .real_address = address,
5095 .flags = flags,
5096 .pgoff = linear_page_index(vma, address),
5097 .gfp_mask = __get_fault_gfp_mask(vma),
5098 };
5099 struct mm_struct *mm = vma->vm_mm;
5100 unsigned long vm_flags = vma->vm_flags;
5101 pgd_t *pgd;
5102 p4d_t *p4d;
5103 vm_fault_t ret;
5104
5105 pgd = pgd_offset(mm, address);
5106 p4d = p4d_alloc(mm, pgd, address);
5107 if (!p4d)
5108 return VM_FAULT_OOM;
5109
5110 vmf.pud = pud_alloc(mm, p4d, address);
5111 if (!vmf.pud)
5112 return VM_FAULT_OOM;
5113retry_pud:
5114 if (pud_none(pud: *vmf.pud) &&
5115 hugepage_vma_check(vma, vm_flags, smaps: false, in_pf: true, enforce_sysfs: true)) {
5116 ret = create_huge_pud(vmf: &vmf);
5117 if (!(ret & VM_FAULT_FALLBACK))
5118 return ret;
5119 } else {
5120 pud_t orig_pud = *vmf.pud;
5121
5122 barrier();
5123 if (pud_trans_huge(pud: orig_pud) || pud_devmap(pud: orig_pud)) {
5124
5125 /*
5126 * TODO once we support anonymous PUDs: NUMA case and
5127 * FAULT_FLAG_UNSHARE handling.
5128 */
5129 if ((flags & FAULT_FLAG_WRITE) && !pud_write(pud: orig_pud)) {
5130 ret = wp_huge_pud(vmf: &vmf, orig_pud);
5131 if (!(ret & VM_FAULT_FALLBACK))
5132 return ret;
5133 } else {
5134 huge_pud_set_accessed(vmf: &vmf, orig_pud);
5135 return 0;
5136 }
5137 }
5138 }
5139
5140 vmf.pmd = pmd_alloc(mm, pud: vmf.pud, address);
5141 if (!vmf.pmd)
5142 return VM_FAULT_OOM;
5143
5144 /* Huge pud page fault raced with pmd_alloc? */
5145 if (pud_trans_unstable(pud: vmf.pud))
5146 goto retry_pud;
5147
5148 if (pmd_none(pmd: *vmf.pmd) &&
5149 hugepage_vma_check(vma, vm_flags, smaps: false, in_pf: true, enforce_sysfs: true)) {
5150 ret = create_huge_pmd(vmf: &vmf);
5151 if (!(ret & VM_FAULT_FALLBACK))
5152 return ret;
5153 } else {
5154 vmf.orig_pmd = pmdp_get_lockless(pmdp: vmf.pmd);
5155
5156 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
5157 VM_BUG_ON(thp_migration_supported() &&
5158 !is_pmd_migration_entry(vmf.orig_pmd));
5159 if (is_pmd_migration_entry(pmd: vmf.orig_pmd))
5160 pmd_migration_entry_wait(mm, pmd: vmf.pmd);
5161 return 0;
5162 }
5163 if (pmd_trans_huge(pmd: vmf.orig_pmd) || pmd_devmap(pmd: vmf.orig_pmd)) {
5164 if (pmd_protnone(pmd: vmf.orig_pmd) && vma_is_accessible(vma))
5165 return do_huge_pmd_numa_page(vmf: &vmf);
5166
5167 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5168 !pmd_write(pmd: vmf.orig_pmd)) {
5169 ret = wp_huge_pmd(vmf: &vmf);
5170 if (!(ret & VM_FAULT_FALLBACK))
5171 return ret;
5172 } else {
5173 huge_pmd_set_accessed(vmf: &vmf);
5174 return 0;
5175 }
5176 }
5177 }
5178
5179 return handle_pte_fault(vmf: &vmf);
5180}
5181
5182/**
5183 * mm_account_fault - Do page fault accounting
5184 * @mm: mm from which memcg should be extracted. It can be NULL.
5185 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5186 * of perf event counters, but we'll still do the per-task accounting to
5187 * the task who triggered this page fault.
5188 * @address: the faulted address.
5189 * @flags: the fault flags.
5190 * @ret: the fault retcode.
5191 *
5192 * This will take care of most of the page fault accounting. Meanwhile, it
5193 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5194 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5195 * still be in per-arch page fault handlers at the entry of page fault.
5196 */
5197static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs,
5198 unsigned long address, unsigned int flags,
5199 vm_fault_t ret)
5200{
5201 bool major;
5202
5203 /* Incomplete faults will be accounted upon completion. */
5204 if (ret & VM_FAULT_RETRY)
5205 return;
5206
5207 /*
5208 * To preserve the behavior of older kernels, PGFAULT counters record
5209 * both successful and failed faults, as opposed to perf counters,
5210 * which ignore failed cases.
5211 */
5212 count_vm_event(item: PGFAULT);
5213 count_memcg_event_mm(mm, idx: PGFAULT);
5214
5215 /*
5216 * Do not account for unsuccessful faults (e.g. when the address wasn't
5217 * valid). That includes arch_vma_access_permitted() failing before
5218 * reaching here. So this is not a "this many hardware page faults"
5219 * counter. We should use the hw profiling for that.
5220 */
5221 if (ret & VM_FAULT_ERROR)
5222 return;
5223
5224 /*
5225 * We define the fault as a major fault when the final successful fault
5226 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5227 * handle it immediately previously).
5228 */
5229 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
5230
5231 if (major)
5232 current->maj_flt++;
5233 else
5234 current->min_flt++;
5235
5236 /*
5237 * If the fault is done for GUP, regs will be NULL. We only do the
5238 * accounting for the per thread fault counters who triggered the
5239 * fault, and we skip the perf event updates.
5240 */
5241 if (!regs)
5242 return;
5243
5244 if (major)
5245 perf_sw_event(event_id: PERF_COUNT_SW_PAGE_FAULTS_MAJ, nr: 1, regs, addr: address);
5246 else
5247 perf_sw_event(event_id: PERF_COUNT_SW_PAGE_FAULTS_MIN, nr: 1, regs, addr: address);
5248}
5249
5250#ifdef CONFIG_LRU_GEN
5251static void lru_gen_enter_fault(struct vm_area_struct *vma)
5252{
5253 /* the LRU algorithm only applies to accesses with recency */
5254 current->in_lru_fault = vma_has_recency(vma);
5255}
5256
5257static void lru_gen_exit_fault(void)
5258{
5259 current->in_lru_fault = false;
5260}
5261#else
5262static void lru_gen_enter_fault(struct vm_area_struct *vma)
5263{
5264}
5265
5266static void lru_gen_exit_fault(void)
5267{
5268}
5269#endif /* CONFIG_LRU_GEN */
5270
5271static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
5272 unsigned int *flags)
5273{
5274 if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
5275 if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
5276 return VM_FAULT_SIGSEGV;
5277 /*
5278 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5279 * just treat it like an ordinary read-fault otherwise.
5280 */
5281 if (!is_cow_mapping(flags: vma->vm_flags))
5282 *flags &= ~FAULT_FLAG_UNSHARE;
5283 } else if (*flags & FAULT_FLAG_WRITE) {
5284 /* Write faults on read-only mappings are impossible ... */
5285 if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
5286 return VM_FAULT_SIGSEGV;
5287 /* ... and FOLL_FORCE only applies to COW mappings. */
5288 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
5289 !is_cow_mapping(vma->vm_flags)))
5290 return VM_FAULT_SIGSEGV;
5291 }
5292#ifdef CONFIG_PER_VMA_LOCK
5293 /*
5294 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5295 * the assumption that lock is dropped on VM_FAULT_RETRY.
5296 */
5297 if (WARN_ON_ONCE((*flags &
5298 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) ==
5299 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)))
5300 return VM_FAULT_SIGSEGV;
5301#endif
5302
5303 return 0;
5304}
5305
5306/*
5307 * By the time we get here, we already hold the mm semaphore
5308 *
5309 * The mmap_lock may have been released depending on flags and our
5310 * return value. See filemap_fault() and __folio_lock_or_retry().
5311 */
5312vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
5313 unsigned int flags, struct pt_regs *regs)
5314{
5315 /* If the fault handler drops the mmap_lock, vma may be freed */
5316 struct mm_struct *mm = vma->vm_mm;
5317 vm_fault_t ret;
5318
5319 __set_current_state(TASK_RUNNING);
5320
5321 ret = sanitize_fault_flags(vma, flags: &flags);
5322 if (ret)
5323 goto out;
5324
5325 if (!arch_vma_access_permitted(vma, write: flags & FAULT_FLAG_WRITE,
5326 execute: flags & FAULT_FLAG_INSTRUCTION,
5327 foreign: flags & FAULT_FLAG_REMOTE)) {
5328 ret = VM_FAULT_SIGSEGV;
5329 goto out;
5330 }
5331
5332 /*
5333 * Enable the memcg OOM handling for faults triggered in user
5334 * space. Kernel faults are handled more gracefully.
5335 */
5336 if (flags & FAULT_FLAG_USER)
5337 mem_cgroup_enter_user_fault();
5338
5339 lru_gen_enter_fault(vma);
5340
5341 if (unlikely(is_vm_hugetlb_page(vma)))
5342 ret = hugetlb_fault(mm: vma->vm_mm, vma, address, flags);
5343 else
5344 ret = __handle_mm_fault(vma, address, flags);
5345
5346 lru_gen_exit_fault();
5347
5348 if (flags & FAULT_FLAG_USER) {
5349 mem_cgroup_exit_user_fault();
5350 /*
5351 * The task may have entered a memcg OOM situation but
5352 * if the allocation error was handled gracefully (no
5353 * VM_FAULT_OOM), there is no need to kill anything.
5354 * Just clean up the OOM state peacefully.
5355 */
5356 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
5357 mem_cgroup_oom_synchronize(wait: false);
5358 }
5359out:
5360 mm_account_fault(mm, regs, address, flags, ret);
5361
5362 return ret;
5363}
5364EXPORT_SYMBOL_GPL(handle_mm_fault);
5365
5366#ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5367#include <linux/extable.h>
5368
5369static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5370{
5371 if (likely(mmap_read_trylock(mm)))
5372 return true;
5373
5374 if (regs && !user_mode(regs)) {
5375 unsigned long ip = instruction_pointer(regs);
5376 if (!search_exception_tables(add: ip))
5377 return false;
5378 }
5379
5380 return !mmap_read_lock_killable(mm);
5381}
5382
5383static inline bool mmap_upgrade_trylock(struct mm_struct *mm)
5384{
5385 /*
5386 * We don't have this operation yet.
5387 *
5388 * It should be easy enough to do: it's basically a
5389 * atomic_long_try_cmpxchg_acquire()
5390 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5391 * it also needs the proper lockdep magic etc.
5392 */
5393 return false;
5394}
5395
5396static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5397{
5398 mmap_read_unlock(mm);
5399 if (regs && !user_mode(regs)) {
5400 unsigned long ip = instruction_pointer(regs);
5401 if (!search_exception_tables(add: ip))
5402 return false;
5403 }
5404 return !mmap_write_lock_killable(mm);
5405}
5406
5407/*
5408 * Helper for page fault handling.
5409 *
5410 * This is kind of equivalend to "mmap_read_lock()" followed
5411 * by "find_extend_vma()", except it's a lot more careful about
5412 * the locking (and will drop the lock on failure).
5413 *
5414 * For example, if we have a kernel bug that causes a page
5415 * fault, we don't want to just use mmap_read_lock() to get
5416 * the mm lock, because that would deadlock if the bug were
5417 * to happen while we're holding the mm lock for writing.
5418 *
5419 * So this checks the exception tables on kernel faults in
5420 * order to only do this all for instructions that are actually
5421 * expected to fault.
5422 *
5423 * We can also actually take the mm lock for writing if we
5424 * need to extend the vma, which helps the VM layer a lot.
5425 */
5426struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm,
5427 unsigned long addr, struct pt_regs *regs)
5428{
5429 struct vm_area_struct *vma;
5430
5431 if (!get_mmap_lock_carefully(mm, regs))
5432 return NULL;
5433
5434 vma = find_vma(mm, addr);
5435 if (likely(vma && (vma->vm_start <= addr)))
5436 return vma;
5437
5438 /*
5439 * Well, dang. We might still be successful, but only
5440 * if we can extend a vma to do so.
5441 */
5442 if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) {
5443 mmap_read_unlock(mm);
5444 return NULL;
5445 }
5446
5447 /*
5448 * We can try to upgrade the mmap lock atomically,
5449 * in which case we can continue to use the vma
5450 * we already looked up.
5451 *
5452 * Otherwise we'll have to drop the mmap lock and
5453 * re-take it, and also look up the vma again,
5454 * re-checking it.
5455 */
5456 if (!mmap_upgrade_trylock(mm)) {
5457 if (!upgrade_mmap_lock_carefully(mm, regs))
5458 return NULL;
5459
5460 vma = find_vma(mm, addr);
5461 if (!vma)
5462 goto fail;
5463 if (vma->vm_start <= addr)
5464 goto success;
5465 if (!(vma->vm_flags & VM_GROWSDOWN))
5466 goto fail;
5467 }
5468
5469 if (expand_stack_locked(vma, address: addr))
5470 goto fail;
5471
5472success:
5473 mmap_write_downgrade(mm);
5474 return vma;
5475
5476fail:
5477 mmap_write_unlock(mm);
5478 return NULL;
5479}
5480#endif
5481
5482#ifdef CONFIG_PER_VMA_LOCK
5483/*
5484 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5485 * stable and not isolated. If the VMA is not found or is being modified the
5486 * function returns NULL.
5487 */
5488struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
5489 unsigned long address)
5490{
5491 MA_STATE(mas, &mm->mm_mt, address, address);
5492 struct vm_area_struct *vma;
5493
5494 rcu_read_lock();
5495retry:
5496 vma = mas_walk(mas: &mas);
5497 if (!vma)
5498 goto inval;
5499
5500 if (!vma_start_read(vma))
5501 goto inval;
5502
5503 /*
5504 * find_mergeable_anon_vma uses adjacent vmas which are not locked.
5505 * This check must happen after vma_start_read(); otherwise, a
5506 * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA
5507 * from its anon_vma.
5508 */
5509 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma))
5510 goto inval_end_read;
5511
5512 /* Check since vm_start/vm_end might change before we lock the VMA */
5513 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5514 goto inval_end_read;
5515
5516 /* Check if the VMA got isolated after we found it */
5517 if (vma->detached) {
5518 vma_end_read(vma);
5519 count_vm_vma_lock_event(VMA_LOCK_MISS);
5520 /* The area was replaced with another one */
5521 goto retry;
5522 }
5523
5524 rcu_read_unlock();
5525 return vma;
5526
5527inval_end_read:
5528 vma_end_read(vma);
5529inval:
5530 rcu_read_unlock();
5531 count_vm_vma_lock_event(VMA_LOCK_ABORT);
5532 return NULL;
5533}
5534#endif /* CONFIG_PER_VMA_LOCK */
5535
5536#ifndef __PAGETABLE_P4D_FOLDED
5537/*
5538 * Allocate p4d page table.
5539 * We've already handled the fast-path in-line.
5540 */
5541int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
5542{
5543 p4d_t *new = p4d_alloc_one(mm, addr: address);
5544 if (!new)
5545 return -ENOMEM;
5546
5547 spin_lock(lock: &mm->page_table_lock);
5548 if (pgd_present(pgd: *pgd)) { /* Another has populated it */
5549 p4d_free(mm, p4d: new);
5550 } else {
5551 smp_wmb(); /* See comment in pmd_install() */
5552 pgd_populate(mm, pgd, p4d: new);
5553 }
5554 spin_unlock(lock: &mm->page_table_lock);
5555 return 0;
5556}
5557#endif /* __PAGETABLE_P4D_FOLDED */
5558
5559#ifndef __PAGETABLE_PUD_FOLDED
5560/*
5561 * Allocate page upper directory.
5562 * We've already handled the fast-path in-line.
5563 */
5564int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
5565{
5566 pud_t *new = pud_alloc_one(mm, addr: address);
5567 if (!new)
5568 return -ENOMEM;
5569
5570 spin_lock(lock: &mm->page_table_lock);
5571 if (!p4d_present(p4d: *p4d)) {
5572 mm_inc_nr_puds(mm);
5573 smp_wmb(); /* See comment in pmd_install() */
5574 p4d_populate(mm, p4d, pud: new);
5575 } else /* Another has populated it */
5576 pud_free(mm, pud: new);
5577 spin_unlock(lock: &mm->page_table_lock);
5578 return 0;
5579}
5580#endif /* __PAGETABLE_PUD_FOLDED */
5581
5582#ifndef __PAGETABLE_PMD_FOLDED
5583/*
5584 * Allocate page middle directory.
5585 * We've already handled the fast-path in-line.
5586 */
5587int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
5588{
5589 spinlock_t *ptl;
5590 pmd_t *new = pmd_alloc_one(mm, addr: address);
5591 if (!new)
5592 return -ENOMEM;
5593
5594 ptl = pud_lock(mm, pud);
5595 if (!pud_present(pud: *pud)) {
5596 mm_inc_nr_pmds(mm);
5597 smp_wmb(); /* See comment in pmd_install() */
5598 pud_populate(mm, pud, pmd: new);
5599 } else { /* Another has populated it */
5600 pmd_free(mm, pmd: new);
5601 }
5602 spin_unlock(lock: ptl);
5603 return 0;
5604}
5605#endif /* __PAGETABLE_PMD_FOLDED */
5606
5607/**
5608 * follow_pte - look up PTE at a user virtual address
5609 * @mm: the mm_struct of the target address space
5610 * @address: user virtual address
5611 * @ptepp: location to store found PTE
5612 * @ptlp: location to store the lock for the PTE
5613 *
5614 * On a successful return, the pointer to the PTE is stored in @ptepp;
5615 * the corresponding lock is taken and its location is stored in @ptlp.
5616 * The contents of the PTE are only stable until @ptlp is released;
5617 * any further use, if any, must be protected against invalidation
5618 * with MMU notifiers.
5619 *
5620 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5621 * should be taken for read.
5622 *
5623 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5624 * it is not a good general-purpose API.
5625 *
5626 * Return: zero on success, -ve otherwise.
5627 */
5628int follow_pte(struct mm_struct *mm, unsigned long address,
5629 pte_t **ptepp, spinlock_t **ptlp)
5630{
5631 pgd_t *pgd;
5632 p4d_t *p4d;
5633 pud_t *pud;
5634 pmd_t *pmd;
5635 pte_t *ptep;
5636
5637 pgd = pgd_offset(mm, address);
5638 if (pgd_none(pgd: *pgd) || unlikely(pgd_bad(*pgd)))
5639 goto out;
5640
5641 p4d = p4d_offset(pgd, address);
5642 if (p4d_none(p4d: *p4d) || unlikely(p4d_bad(*p4d)))
5643 goto out;
5644
5645 pud = pud_offset(p4d, address);
5646 if (pud_none(pud: *pud) || unlikely(pud_bad(*pud)))
5647 goto out;
5648
5649 pmd = pmd_offset(pud, address);
5650 VM_BUG_ON(pmd_trans_huge(*pmd));
5651
5652 ptep = pte_offset_map_lock(mm, pmd, addr: address, ptlp);
5653 if (!ptep)
5654 goto out;
5655 if (!pte_present(a: ptep_get(ptep)))
5656 goto unlock;
5657 *ptepp = ptep;
5658 return 0;
5659unlock:
5660 pte_unmap_unlock(ptep, *ptlp);
5661out:
5662 return -EINVAL;
5663}
5664EXPORT_SYMBOL_GPL(follow_pte);
5665
5666/**
5667 * follow_pfn - look up PFN at a user virtual address
5668 * @vma: memory mapping
5669 * @address: user virtual address
5670 * @pfn: location to store found PFN
5671 *
5672 * Only IO mappings and raw PFN mappings are allowed.
5673 *
5674 * This function does not allow the caller to read the permissions
5675 * of the PTE. Do not use it.
5676 *
5677 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5678 */
5679int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5680 unsigned long *pfn)
5681{
5682 int ret = -EINVAL;
5683 spinlock_t *ptl;
5684 pte_t *ptep;
5685
5686 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5687 return ret;
5688
5689 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5690 if (ret)
5691 return ret;
5692 *pfn = pte_pfn(pte: ptep_get(ptep));
5693 pte_unmap_unlock(ptep, ptl);
5694 return 0;
5695}
5696EXPORT_SYMBOL(follow_pfn);
5697
5698#ifdef CONFIG_HAVE_IOREMAP_PROT
5699int follow_phys(struct vm_area_struct *vma,
5700 unsigned long address, unsigned int flags,
5701 unsigned long *prot, resource_size_t *phys)
5702{
5703 int ret = -EINVAL;
5704 pte_t *ptep, pte;
5705 spinlock_t *ptl;
5706
5707 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5708 goto out;
5709
5710 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5711 goto out;
5712 pte = ptep_get(ptep);
5713
5714 if ((flags & FOLL_WRITE) && !pte_write(pte))
5715 goto unlock;
5716
5717 *prot = pgprot_val(pte_pgprot(pte));
5718 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5719
5720 ret = 0;
5721unlock:
5722 pte_unmap_unlock(ptep, ptl);
5723out:
5724 return ret;
5725}
5726
5727/**
5728 * generic_access_phys - generic implementation for iomem mmap access
5729 * @vma: the vma to access
5730 * @addr: userspace address, not relative offset within @vma
5731 * @buf: buffer to read/write
5732 * @len: length of transfer
5733 * @write: set to FOLL_WRITE when writing, otherwise reading
5734 *
5735 * This is a generic implementation for &vm_operations_struct.access for an
5736 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5737 * not page based.
5738 */
5739int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5740 void *buf, int len, int write)
5741{
5742 resource_size_t phys_addr;
5743 unsigned long prot = 0;
5744 void __iomem *maddr;
5745 pte_t *ptep, pte;
5746 spinlock_t *ptl;
5747 int offset = offset_in_page(addr);
5748 int ret = -EINVAL;
5749
5750 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5751 return -EINVAL;
5752
5753retry:
5754 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5755 return -EINVAL;
5756 pte = ptep_get(ptep);
5757 pte_unmap_unlock(ptep, ptl);
5758
5759 prot = pgprot_val(pte_pgprot(pte));
5760 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5761
5762 if ((write & FOLL_WRITE) && !pte_write(pte))
5763 return -EINVAL;
5764
5765 maddr = ioremap_prot(offset: phys_addr, PAGE_ALIGN(len + offset), prot_val: prot);
5766 if (!maddr)
5767 return -ENOMEM;
5768
5769 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5770 goto out_unmap;
5771
5772 if (!pte_same(a: pte, b: ptep_get(ptep))) {
5773 pte_unmap_unlock(ptep, ptl);
5774 iounmap(addr: maddr);
5775
5776 goto retry;
5777 }
5778
5779 if (write)
5780 memcpy_toio(maddr + offset, buf, len);
5781 else
5782 memcpy_fromio(buf, maddr + offset, len);
5783 ret = len;
5784 pte_unmap_unlock(ptep, ptl);
5785out_unmap:
5786 iounmap(addr: maddr);
5787
5788 return ret;
5789}
5790EXPORT_SYMBOL_GPL(generic_access_phys);
5791#endif
5792
5793/*
5794 * Access another process' address space as given in mm.
5795 */
5796static int __access_remote_vm(struct mm_struct *mm, unsigned long addr,
5797 void *buf, int len, unsigned int gup_flags)
5798{
5799 void *old_buf = buf;
5800 int write = gup_flags & FOLL_WRITE;
5801
5802 if (mmap_read_lock_killable(mm))
5803 return 0;
5804
5805 /* Untag the address before looking up the VMA */
5806 addr = untagged_addr_remote(mm, addr);
5807
5808 /* Avoid triggering the temporary warning in __get_user_pages */
5809 if (!vma_lookup(mm, addr) && !expand_stack(mm, addr))
5810 return 0;
5811
5812 /* ignore errors, just check how much was successfully transferred */
5813 while (len) {
5814 int bytes, offset;
5815 void *maddr;
5816 struct vm_area_struct *vma = NULL;
5817 struct page *page = get_user_page_vma_remote(mm, addr,
5818 gup_flags, vmap: &vma);
5819
5820 if (IS_ERR(ptr: page)) {
5821 /* We might need to expand the stack to access it */
5822 vma = vma_lookup(mm, addr);
5823 if (!vma) {
5824 vma = expand_stack(mm, addr);
5825
5826 /* mmap_lock was dropped on failure */
5827 if (!vma)
5828 return buf - old_buf;
5829
5830 /* Try again if stack expansion worked */
5831 continue;
5832 }
5833
5834 /*
5835 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5836 * we can access using slightly different code.
5837 */
5838 bytes = 0;
5839#ifdef CONFIG_HAVE_IOREMAP_PROT
5840 if (vma->vm_ops && vma->vm_ops->access)
5841 bytes = vma->vm_ops->access(vma, addr, buf,
5842 len, write);
5843#endif
5844 if (bytes <= 0)
5845 break;
5846 } else {
5847 bytes = len;
5848 offset = addr & (PAGE_SIZE-1);
5849 if (bytes > PAGE_SIZE-offset)
5850 bytes = PAGE_SIZE-offset;
5851
5852 maddr = kmap(page);
5853 if (write) {
5854 copy_to_user_page(vma, page, addr,
5855 maddr + offset, buf, bytes);
5856 set_page_dirty_lock(page);
5857 } else {
5858 copy_from_user_page(vma, page, addr,
5859 buf, maddr + offset, bytes);
5860 }
5861 kunmap(page);
5862 put_page(page);
5863 }
5864 len -= bytes;
5865 buf += bytes;
5866 addr += bytes;
5867 }
5868 mmap_read_unlock(mm);
5869
5870 return buf - old_buf;
5871}
5872
5873/**
5874 * access_remote_vm - access another process' address space
5875 * @mm: the mm_struct of the target address space
5876 * @addr: start address to access
5877 * @buf: source or destination buffer
5878 * @len: number of bytes to transfer
5879 * @gup_flags: flags modifying lookup behaviour
5880 *
5881 * The caller must hold a reference on @mm.
5882 *
5883 * Return: number of bytes copied from source to destination.
5884 */
5885int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5886 void *buf, int len, unsigned int gup_flags)
5887{
5888 return __access_remote_vm(mm, addr, buf, len, gup_flags);
5889}
5890
5891/*
5892 * Access another process' address space.
5893 * Source/target buffer must be kernel space,
5894 * Do not walk the page table directly, use get_user_pages
5895 */
5896int access_process_vm(struct task_struct *tsk, unsigned long addr,
5897 void *buf, int len, unsigned int gup_flags)
5898{
5899 struct mm_struct *mm;
5900 int ret;
5901
5902 mm = get_task_mm(task: tsk);
5903 if (!mm)
5904 return 0;
5905
5906 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5907
5908 mmput(mm);
5909
5910 return ret;
5911}
5912EXPORT_SYMBOL_GPL(access_process_vm);
5913
5914/*
5915 * Print the name of a VMA.
5916 */
5917void print_vma_addr(char *prefix, unsigned long ip)
5918{
5919 struct mm_struct *mm = current->mm;
5920 struct vm_area_struct *vma;
5921
5922 /*
5923 * we might be running from an atomic context so we cannot sleep
5924 */
5925 if (!mmap_read_trylock(mm))
5926 return;
5927
5928 vma = find_vma(mm, addr: ip);
5929 if (vma && vma->vm_file) {
5930 struct file *f = vma->vm_file;
5931 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5932 if (buf) {
5933 char *p;
5934
5935 p = file_path(f, buf, PAGE_SIZE);
5936 if (IS_ERR(ptr: p))
5937 p = "?";
5938 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5939 vma->vm_start,
5940 vma->vm_end - vma->vm_start);
5941 free_page((unsigned long)buf);
5942 }
5943 }
5944 mmap_read_unlock(mm);
5945}
5946
5947#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5948void __might_fault(const char *file, int line)
5949{
5950 if (pagefault_disabled())
5951 return;
5952 __might_sleep(file, line);
5953#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5954 if (current->mm)
5955 might_lock_read(&current->mm->mmap_lock);
5956#endif
5957}
5958EXPORT_SYMBOL(__might_fault);
5959#endif
5960
5961#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5962/*
5963 * Process all subpages of the specified huge page with the specified
5964 * operation. The target subpage will be processed last to keep its
5965 * cache lines hot.
5966 */
5967static inline int process_huge_page(
5968 unsigned long addr_hint, unsigned int pages_per_huge_page,
5969 int (*process_subpage)(unsigned long addr, int idx, void *arg),
5970 void *arg)
5971{
5972 int i, n, base, l, ret;
5973 unsigned long addr = addr_hint &
5974 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5975
5976 /* Process target subpage last to keep its cache lines hot */
5977 might_sleep();
5978 n = (addr_hint - addr) / PAGE_SIZE;
5979 if (2 * n <= pages_per_huge_page) {
5980 /* If target subpage in first half of huge page */
5981 base = 0;
5982 l = n;
5983 /* Process subpages at the end of huge page */
5984 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5985 cond_resched();
5986 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
5987 if (ret)
5988 return ret;
5989 }
5990 } else {
5991 /* If target subpage in second half of huge page */
5992 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5993 l = pages_per_huge_page - n;
5994 /* Process subpages at the begin of huge page */
5995 for (i = 0; i < base; i++) {
5996 cond_resched();
5997 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
5998 if (ret)
5999 return ret;
6000 }
6001 }
6002 /*
6003 * Process remaining subpages in left-right-left-right pattern
6004 * towards the target subpage
6005 */
6006 for (i = 0; i < l; i++) {
6007 int left_idx = base + i;
6008 int right_idx = base + 2 * l - 1 - i;
6009
6010 cond_resched();
6011 ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
6012 if (ret)
6013 return ret;
6014 cond_resched();
6015 ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
6016 if (ret)
6017 return ret;
6018 }
6019 return 0;
6020}
6021
6022static void clear_gigantic_page(struct page *page,
6023 unsigned long addr,
6024 unsigned int pages_per_huge_page)
6025{
6026 int i;
6027 struct page *p;
6028
6029 might_sleep();
6030 for (i = 0; i < pages_per_huge_page; i++) {
6031 p = nth_page(page, i);
6032 cond_resched();
6033 clear_user_highpage(page: p, vaddr: addr + i * PAGE_SIZE);
6034 }
6035}
6036
6037static int clear_subpage(unsigned long addr, int idx, void *arg)
6038{
6039 struct page *page = arg;
6040
6041 clear_user_highpage(page: page + idx, vaddr: addr);
6042 return 0;
6043}
6044
6045void clear_huge_page(struct page *page,
6046 unsigned long addr_hint, unsigned int pages_per_huge_page)
6047{
6048 unsigned long addr = addr_hint &
6049 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6050
6051 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
6052 clear_gigantic_page(page, addr, pages_per_huge_page);
6053 return;
6054 }
6055
6056 process_huge_page(addr_hint, pages_per_huge_page, process_subpage: clear_subpage, arg: page);
6057}
6058
6059static int copy_user_gigantic_page(struct folio *dst, struct folio *src,
6060 unsigned long addr,
6061 struct vm_area_struct *vma,
6062 unsigned int pages_per_huge_page)
6063{
6064 int i;
6065 struct page *dst_page;
6066 struct page *src_page;
6067
6068 for (i = 0; i < pages_per_huge_page; i++) {
6069 dst_page = folio_page(dst, i);
6070 src_page = folio_page(src, i);
6071
6072 cond_resched();
6073 if (copy_mc_user_highpage(to: dst_page, from: src_page,
6074 vaddr: addr + i*PAGE_SIZE, vma)) {
6075 memory_failure_queue(page_to_pfn(src_page), flags: 0);
6076 return -EHWPOISON;
6077 }
6078 }
6079 return 0;
6080}
6081
6082struct copy_subpage_arg {
6083 struct page *dst;
6084 struct page *src;
6085 struct vm_area_struct *vma;
6086};
6087
6088static int copy_subpage(unsigned long addr, int idx, void *arg)
6089{
6090 struct copy_subpage_arg *copy_arg = arg;
6091
6092 if (copy_mc_user_highpage(to: copy_arg->dst + idx, from: copy_arg->src + idx,
6093 vaddr: addr, vma: copy_arg->vma)) {
6094 memory_failure_queue(page_to_pfn(copy_arg->src + idx), flags: 0);
6095 return -EHWPOISON;
6096 }
6097 return 0;
6098}
6099
6100int copy_user_large_folio(struct folio *dst, struct folio *src,
6101 unsigned long addr_hint, struct vm_area_struct *vma)
6102{
6103 unsigned int pages_per_huge_page = folio_nr_pages(folio: dst);
6104 unsigned long addr = addr_hint &
6105 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6106 struct copy_subpage_arg arg = {
6107 .dst = &dst->page,
6108 .src = &src->page,
6109 .vma = vma,
6110 };
6111
6112 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES))
6113 return copy_user_gigantic_page(dst, src, addr, vma,
6114 pages_per_huge_page);
6115
6116 return process_huge_page(addr_hint, pages_per_huge_page, process_subpage: copy_subpage, arg: &arg);
6117}
6118
6119long copy_folio_from_user(struct folio *dst_folio,
6120 const void __user *usr_src,
6121 bool allow_pagefault)
6122{
6123 void *kaddr;
6124 unsigned long i, rc = 0;
6125 unsigned int nr_pages = folio_nr_pages(folio: dst_folio);
6126 unsigned long ret_val = nr_pages * PAGE_SIZE;
6127 struct page *subpage;
6128
6129 for (i = 0; i < nr_pages; i++) {
6130 subpage = folio_page(dst_folio, i);
6131 kaddr = kmap_local_page(page: subpage);
6132 if (!allow_pagefault)
6133 pagefault_disable();
6134 rc = copy_from_user(to: kaddr, from: usr_src + i * PAGE_SIZE, PAGE_SIZE);
6135 if (!allow_pagefault)
6136 pagefault_enable();
6137 kunmap_local(kaddr);
6138
6139 ret_val -= (PAGE_SIZE - rc);
6140 if (rc)
6141 break;
6142
6143 flush_dcache_page(page: subpage);
6144
6145 cond_resched();
6146 }
6147 return ret_val;
6148}
6149#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6150
6151#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6152
6153static struct kmem_cache *page_ptl_cachep;
6154
6155void __init ptlock_cache_init(void)
6156{
6157 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
6158 SLAB_PANIC, NULL);
6159}
6160
6161bool ptlock_alloc(struct ptdesc *ptdesc)
6162{
6163 spinlock_t *ptl;
6164
6165 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
6166 if (!ptl)
6167 return false;
6168 ptdesc->ptl = ptl;
6169 return true;
6170}
6171
6172void ptlock_free(struct ptdesc *ptdesc)
6173{
6174 kmem_cache_free(page_ptl_cachep, ptdesc->ptl);
6175}
6176#endif
6177

source code of linux/mm/memory.c