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