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