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
100 unsigned long max_mapnr;
101 EXPORT_SYMBOL(max_mapnr);
102
103 struct page *mem_map;
104 EXPORT_SYMBOL(mem_map);
105 #endif
106
107 static vm_fault_t do_fault(struct vm_fault *vmf);
108 static vm_fault_t do_anonymous_page(struct vm_fault *vmf);
109 static 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 */
vmf_orig_pte_uffd_wp(struct vm_fault * vmf)115 static __always_inline bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
116 {
117 if (!userfaultfd_wp(vmf->vma))
118 return false;
119 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
120 return false;
121
122 return pte_marker_uffd_wp(vmf->orig_pte);
123 }
124
125 /*
126 * A number of key systems in x86 including ioremap() rely on the assumption
127 * that high_memory defines the upper bound on direct map memory, then end
128 * of ZONE_NORMAL.
129 */
130 void *high_memory;
131 EXPORT_SYMBOL(high_memory);
132
133 /*
134 * Randomize the address space (stacks, mmaps, brk, etc.).
135 *
136 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
137 * as ancient (libc5 based) binaries can segfault. )
138 */
139 int randomize_va_space __read_mostly =
140 #ifdef CONFIG_COMPAT_BRK
141 1;
142 #else
143 2;
144 #endif
145
146 #ifndef arch_wants_old_prefaulted_pte
arch_wants_old_prefaulted_pte(void)147 static inline bool arch_wants_old_prefaulted_pte(void)
148 {
149 /*
150 * Transitioning a PTE from 'old' to 'young' can be expensive on
151 * some architectures, even if it's performed in hardware. By
152 * default, "false" means prefaulted entries will be 'young'.
153 */
154 return false;
155 }
156 #endif
157
disable_randmaps(char * s)158 static int __init disable_randmaps(char *s)
159 {
160 randomize_va_space = 0;
161 return 1;
162 }
163 __setup("norandmaps", disable_randmaps);
164
165 unsigned long zero_pfn __read_mostly;
166 EXPORT_SYMBOL(zero_pfn);
167
168 unsigned long highest_memmap_pfn __read_mostly;
169
170 /*
171 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
172 */
init_zero_pfn(void)173 static int __init init_zero_pfn(void)
174 {
175 zero_pfn = page_to_pfn(ZERO_PAGE(0));
176 return 0;
177 }
178 early_initcall(init_zero_pfn);
179
mm_trace_rss_stat(struct mm_struct * mm,int member)180 void mm_trace_rss_stat(struct mm_struct *mm, int member)
181 {
182 trace_rss_stat(mm, member);
183 }
184
185 /*
186 * Note: this doesn't free the actual pages themselves. That
187 * has been handled earlier when unmapping all the memory regions.
188 */
free_pte_range(struct mmu_gather * tlb,pmd_t * pmd,unsigned long addr)189 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
190 unsigned long addr)
191 {
192 pgtable_t token = pmd_pgtable(*pmd);
193 pmd_clear(pmd);
194 pte_free_tlb(tlb, token, addr);
195 mm_dec_nr_ptes(tlb->mm);
196 }
197
free_pmd_range(struct mmu_gather * tlb,pud_t * pud,unsigned long addr,unsigned long end,unsigned long floor,unsigned long ceiling)198 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
199 unsigned long addr, unsigned long end,
200 unsigned long floor, unsigned long ceiling)
201 {
202 pmd_t *pmd;
203 unsigned long next;
204 unsigned long start;
205
206 start = addr;
207 pmd = pmd_offset(pud, addr);
208 do {
209 next = pmd_addr_end(addr, end);
210 if (pmd_none_or_clear_bad(pmd))
211 continue;
212 free_pte_range(tlb, pmd, addr);
213 } while (pmd++, addr = next, addr != end);
214
215 start &= PUD_MASK;
216 if (start < floor)
217 return;
218 if (ceiling) {
219 ceiling &= PUD_MASK;
220 if (!ceiling)
221 return;
222 }
223 if (end - 1 > ceiling - 1)
224 return;
225
226 pmd = pmd_offset(pud, start);
227 pud_clear(pud);
228 pmd_free_tlb(tlb, pmd, start);
229 mm_dec_nr_pmds(tlb->mm);
230 }
231
free_pud_range(struct mmu_gather * tlb,p4d_t * p4d,unsigned long addr,unsigned long end,unsigned long floor,unsigned long ceiling)232 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
233 unsigned long addr, unsigned long end,
234 unsigned long floor, unsigned long ceiling)
235 {
236 pud_t *pud;
237 unsigned long next;
238 unsigned long start;
239
240 start = addr;
241 pud = pud_offset(p4d, addr);
242 do {
243 next = pud_addr_end(addr, end);
244 if (pud_none_or_clear_bad(pud))
245 continue;
246 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
247 } while (pud++, addr = next, addr != end);
248
249 start &= P4D_MASK;
250 if (start < floor)
251 return;
252 if (ceiling) {
253 ceiling &= P4D_MASK;
254 if (!ceiling)
255 return;
256 }
257 if (end - 1 > ceiling - 1)
258 return;
259
260 pud = pud_offset(p4d, start);
261 p4d_clear(p4d);
262 pud_free_tlb(tlb, pud, start);
263 mm_dec_nr_puds(tlb->mm);
264 }
265
free_p4d_range(struct mmu_gather * tlb,pgd_t * pgd,unsigned long addr,unsigned long end,unsigned long floor,unsigned long ceiling)266 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
267 unsigned long addr, unsigned long end,
268 unsigned long floor, unsigned long ceiling)
269 {
270 p4d_t *p4d;
271 unsigned long next;
272 unsigned long start;
273
274 start = addr;
275 p4d = p4d_offset(pgd, addr);
276 do {
277 next = p4d_addr_end(addr, end);
278 if (p4d_none_or_clear_bad(p4d))
279 continue;
280 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
281 } while (p4d++, addr = next, addr != end);
282
283 start &= PGDIR_MASK;
284 if (start < floor)
285 return;
286 if (ceiling) {
287 ceiling &= PGDIR_MASK;
288 if (!ceiling)
289 return;
290 }
291 if (end - 1 > ceiling - 1)
292 return;
293
294 p4d = p4d_offset(pgd, start);
295 pgd_clear(pgd);
296 p4d_free_tlb(tlb, p4d, start);
297 }
298
299 /*
300 * This function frees user-level page tables of a process.
301 */
free_pgd_range(struct mmu_gather * tlb,unsigned long addr,unsigned long end,unsigned long floor,unsigned long ceiling)302 void free_pgd_range(struct mmu_gather *tlb,
303 unsigned long addr, unsigned long end,
304 unsigned long floor, unsigned long ceiling)
305 {
306 pgd_t *pgd;
307 unsigned long next;
308
309 /*
310 * The next few lines have given us lots of grief...
311 *
312 * Why are we testing PMD* at this top level? Because often
313 * there will be no work to do at all, and we'd prefer not to
314 * go all the way down to the bottom just to discover that.
315 *
316 * Why all these "- 1"s? Because 0 represents both the bottom
317 * of the address space and the top of it (using -1 for the
318 * top wouldn't help much: the masks would do the wrong thing).
319 * The rule is that addr 0 and floor 0 refer to the bottom of
320 * the address space, but end 0 and ceiling 0 refer to the top
321 * Comparisons need to use "end - 1" and "ceiling - 1" (though
322 * that end 0 case should be mythical).
323 *
324 * Wherever addr is brought up or ceiling brought down, we must
325 * be careful to reject "the opposite 0" before it confuses the
326 * subsequent tests. But what about where end is brought down
327 * by PMD_SIZE below? no, end can't go down to 0 there.
328 *
329 * Whereas we round start (addr) and ceiling down, by different
330 * masks at different levels, in order to test whether a table
331 * now has no other vmas using it, so can be freed, we don't
332 * bother to round floor or end up - the tests don't need that.
333 */
334
335 addr &= PMD_MASK;
336 if (addr < floor) {
337 addr += PMD_SIZE;
338 if (!addr)
339 return;
340 }
341 if (ceiling) {
342 ceiling &= PMD_MASK;
343 if (!ceiling)
344 return;
345 }
346 if (end - 1 > ceiling - 1)
347 end -= PMD_SIZE;
348 if (addr > end - 1)
349 return;
350 /*
351 * We add page table cache pages with PAGE_SIZE,
352 * (see pte_free_tlb()), flush the tlb if we need
353 */
354 tlb_change_page_size(tlb, PAGE_SIZE);
355 pgd = pgd_offset(tlb->mm, addr);
356 do {
357 next = pgd_addr_end(addr, end);
358 if (pgd_none_or_clear_bad(pgd))
359 continue;
360 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
361 } while (pgd++, addr = next, addr != end);
362 }
363
free_pgtables(struct mmu_gather * tlb,struct ma_state * mas,struct vm_area_struct * vma,unsigned long floor,unsigned long ceiling,bool mm_wr_locked)364 void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
365 struct vm_area_struct *vma, unsigned long floor,
366 unsigned long ceiling, bool mm_wr_locked)
367 {
368 do {
369 unsigned long addr = vma->vm_start;
370 struct vm_area_struct *next;
371
372 /*
373 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
374 * be 0. This will underflow and is okay.
375 */
376 next = mas_find(mas, ceiling - 1);
377 if (unlikely(xa_is_zero(next)))
378 next = NULL;
379
380 /*
381 * Hide vma from rmap and truncate_pagecache before freeing
382 * pgtables
383 */
384 if (mm_wr_locked)
385 vma_start_write(vma);
386 unlink_anon_vmas(vma);
387 unlink_file_vma(vma);
388
389 if (is_vm_hugetlb_page(vma)) {
390 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
391 floor, next ? next->vm_start : ceiling);
392 } else {
393 /*
394 * Optimization: gather nearby vmas into one call down
395 */
396 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
397 && !is_vm_hugetlb_page(next)) {
398 vma = next;
399 next = mas_find(mas, ceiling - 1);
400 if (unlikely(xa_is_zero(next)))
401 next = NULL;
402 if (mm_wr_locked)
403 vma_start_write(vma);
404 unlink_anon_vmas(vma);
405 unlink_file_vma(vma);
406 }
407 free_pgd_range(tlb, addr, vma->vm_end,
408 floor, next ? next->vm_start : ceiling);
409 }
410 vma = next;
411 } while (vma);
412 }
413
pmd_install(struct mm_struct * mm,pmd_t * pmd,pgtable_t * pte)414 void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
415 {
416 spinlock_t *ptl = pmd_lock(mm, pmd);
417
418 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
419 mm_inc_nr_ptes(mm);
420 /*
421 * Ensure all pte setup (eg. pte page lock and page clearing) are
422 * visible before the pte is made visible to other CPUs by being
423 * put into page tables.
424 *
425 * The other side of the story is the pointer chasing in the page
426 * table walking code (when walking the page table without locking;
427 * ie. most of the time). Fortunately, these data accesses consist
428 * of a chain of data-dependent loads, meaning most CPUs (alpha
429 * being the notable exception) will already guarantee loads are
430 * seen in-order. See the alpha page table accessors for the
431 * smp_rmb() barriers in page table walking code.
432 */
433 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
434 pmd_populate(mm, pmd, *pte);
435 *pte = NULL;
436 }
437 spin_unlock(ptl);
438 }
439
__pte_alloc(struct mm_struct * mm,pmd_t * pmd)440 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
441 {
442 pgtable_t new = pte_alloc_one(mm);
443 if (!new)
444 return -ENOMEM;
445
446 pmd_install(mm, pmd, &new);
447 if (new)
448 pte_free(mm, new);
449 return 0;
450 }
451
__pte_alloc_kernel(pmd_t * pmd)452 int __pte_alloc_kernel(pmd_t *pmd)
453 {
454 pte_t *new = pte_alloc_one_kernel(&init_mm);
455 if (!new)
456 return -ENOMEM;
457
458 spin_lock(&init_mm.page_table_lock);
459 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
460 smp_wmb(); /* See comment in pmd_install() */
461 pmd_populate_kernel(&init_mm, pmd, new);
462 new = NULL;
463 }
464 spin_unlock(&init_mm.page_table_lock);
465 if (new)
466 pte_free_kernel(&init_mm, new);
467 return 0;
468 }
469
init_rss_vec(int * rss)470 static inline void init_rss_vec(int *rss)
471 {
472 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
473 }
474
add_mm_rss_vec(struct mm_struct * mm,int * rss)475 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
476 {
477 int i;
478
479 for (i = 0; i < NR_MM_COUNTERS; i++)
480 if (rss[i])
481 add_mm_counter(mm, i, rss[i]);
482 }
483
484 /*
485 * This function is called to print an error when a bad pte
486 * is found. For example, we might have a PFN-mapped pte in
487 * a region that doesn't allow it.
488 *
489 * The calling function must still handle the error.
490 */
print_bad_pte(struct vm_area_struct * vma,unsigned long addr,pte_t pte,struct page * page)491 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
492 pte_t pte, struct page *page)
493 {
494 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
495 p4d_t *p4d = p4d_offset(pgd, addr);
496 pud_t *pud = pud_offset(p4d, addr);
497 pmd_t *pmd = pmd_offset(pud, addr);
498 struct address_space *mapping;
499 pgoff_t index;
500 static unsigned long resume;
501 static unsigned long nr_shown;
502 static unsigned long nr_unshown;
503
504 /*
505 * Allow a burst of 60 reports, then keep quiet for that minute;
506 * or allow a steady drip of one report per second.
507 */
508 if (nr_shown == 60) {
509 if (time_before(jiffies, resume)) {
510 nr_unshown++;
511 return;
512 }
513 if (nr_unshown) {
514 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
515 nr_unshown);
516 nr_unshown = 0;
517 }
518 nr_shown = 0;
519 }
520 if (nr_shown++ == 0)
521 resume = jiffies + 60 * HZ;
522
523 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
524 index = linear_page_index(vma, addr);
525
526 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
527 current->comm,
528 (long long)pte_val(pte), (long long)pmd_val(*pmd));
529 if (page)
530 dump_page(page, "bad pte");
531 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
532 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
533 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
534 vma->vm_file,
535 vma->vm_ops ? vma->vm_ops->fault : NULL,
536 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
537 mapping ? mapping->a_ops->read_folio : NULL);
538 dump_stack();
539 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
540 }
541
542 /*
543 * vm_normal_page -- This function gets the "struct page" associated with a pte.
544 *
545 * "Special" mappings do not wish to be associated with a "struct page" (either
546 * it doesn't exist, or it exists but they don't want to touch it). In this
547 * case, NULL is returned here. "Normal" mappings do have a struct page.
548 *
549 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
550 * pte bit, in which case this function is trivial. Secondly, an architecture
551 * may not have a spare pte bit, which requires a more complicated scheme,
552 * described below.
553 *
554 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
555 * special mapping (even if there are underlying and valid "struct pages").
556 * COWed pages of a VM_PFNMAP are always normal.
557 *
558 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
559 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
560 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
561 * mapping will always honor the rule
562 *
563 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
564 *
565 * And for normal mappings this is false.
566 *
567 * This restricts such mappings to be a linear translation from virtual address
568 * to pfn. To get around this restriction, we allow arbitrary mappings so long
569 * as the vma is not a COW mapping; in that case, we know that all ptes are
570 * special (because none can have been COWed).
571 *
572 *
573 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
574 *
575 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
576 * page" backing, however the difference is that _all_ pages with a struct
577 * page (that is, those where pfn_valid is true) are refcounted and considered
578 * normal pages by the VM. The disadvantage is that pages are refcounted
579 * (which can be slower and simply not an option for some PFNMAP users). The
580 * advantage is that we don't have to follow the strict linearity rule of
581 * PFNMAP mappings in order to support COWable mappings.
582 *
583 */
vm_normal_page(struct vm_area_struct * vma,unsigned long addr,pte_t pte)584 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
585 pte_t pte)
586 {
587 unsigned long pfn = pte_pfn(pte);
588
589 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
590 if (likely(!pte_special(pte)))
591 goto check_pfn;
592 if (vma->vm_ops && vma->vm_ops->find_special_page)
593 return vma->vm_ops->find_special_page(vma, addr);
594 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
595 return NULL;
596 if (is_zero_pfn(pfn))
597 return NULL;
598 if (pte_devmap(pte))
599 /*
600 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
601 * and will have refcounts incremented on their struct pages
602 * when they are inserted into PTEs, thus they are safe to
603 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
604 * do not have refcounts. Example of legacy ZONE_DEVICE is
605 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
606 */
607 return NULL;
608
609 print_bad_pte(vma, addr, pte, NULL);
610 return NULL;
611 }
612
613 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
614
615 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
616 if (vma->vm_flags & VM_MIXEDMAP) {
617 if (!pfn_valid(pfn))
618 return NULL;
619 goto out;
620 } else {
621 unsigned long off;
622 off = (addr - vma->vm_start) >> PAGE_SHIFT;
623 if (pfn == vma->vm_pgoff + off)
624 return NULL;
625 if (!is_cow_mapping(vma->vm_flags))
626 return NULL;
627 }
628 }
629
630 if (is_zero_pfn(pfn))
631 return NULL;
632
633 check_pfn:
634 if (unlikely(pfn > highest_memmap_pfn)) {
635 print_bad_pte(vma, addr, pte, NULL);
636 return NULL;
637 }
638
639 /*
640 * NOTE! We still have PageReserved() pages in the page tables.
641 * eg. VDSO mappings can cause them to exist.
642 */
643 out:
644 return pfn_to_page(pfn);
645 }
646
vm_normal_folio(struct vm_area_struct * vma,unsigned long addr,pte_t pte)647 struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr,
648 pte_t pte)
649 {
650 struct page *page = vm_normal_page(vma, addr, pte);
651
652 if (page)
653 return page_folio(page);
654 return NULL;
655 }
656
657 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
vm_normal_page_pmd(struct vm_area_struct * vma,unsigned long addr,pmd_t pmd)658 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
659 pmd_t pmd)
660 {
661 unsigned long pfn = pmd_pfn(pmd);
662
663 /*
664 * There is no pmd_special() but there may be special pmds, e.g.
665 * in a direct-access (dax) mapping, so let's just replicate the
666 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
667 */
668 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
669 if (vma->vm_flags & VM_MIXEDMAP) {
670 if (!pfn_valid(pfn))
671 return NULL;
672 goto out;
673 } else {
674 unsigned long off;
675 off = (addr - vma->vm_start) >> PAGE_SHIFT;
676 if (pfn == vma->vm_pgoff + off)
677 return NULL;
678 if (!is_cow_mapping(vma->vm_flags))
679 return NULL;
680 }
681 }
682
683 if (pmd_devmap(pmd))
684 return NULL;
685 if (is_huge_zero_pmd(pmd))
686 return NULL;
687 if (unlikely(pfn > highest_memmap_pfn))
688 return NULL;
689
690 /*
691 * NOTE! We still have PageReserved() pages in the page tables.
692 * eg. VDSO mappings can cause them to exist.
693 */
694 out:
695 return pfn_to_page(pfn);
696 }
697
vm_normal_folio_pmd(struct vm_area_struct * vma,unsigned long addr,pmd_t pmd)698 struct folio *vm_normal_folio_pmd(struct vm_area_struct *vma,
699 unsigned long addr, pmd_t pmd)
700 {
701 struct page *page = vm_normal_page_pmd(vma, addr, pmd);
702
703 if (page)
704 return page_folio(page);
705 return NULL;
706 }
707 #endif
708
restore_exclusive_pte(struct vm_area_struct * vma,struct page * page,unsigned long address,pte_t * ptep)709 static void restore_exclusive_pte(struct vm_area_struct *vma,
710 struct page *page, unsigned long address,
711 pte_t *ptep)
712 {
713 struct folio *folio = page_folio(page);
714 pte_t orig_pte;
715 pte_t pte;
716 swp_entry_t entry;
717
718 orig_pte = ptep_get(ptep);
719 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
720 if (pte_swp_soft_dirty(orig_pte))
721 pte = pte_mksoft_dirty(pte);
722
723 entry = pte_to_swp_entry(orig_pte);
724 if (pte_swp_uffd_wp(orig_pte))
725 pte = pte_mkuffd_wp(pte);
726 else if (is_writable_device_exclusive_entry(entry))
727 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
728
729 VM_BUG_ON_FOLIO(pte_write(pte) && (!folio_test_anon(folio) &&
730 PageAnonExclusive(page)), folio);
731
732 /*
733 * No need to take a page reference as one was already
734 * created when the swap entry was made.
735 */
736 if (folio_test_anon(folio))
737 folio_add_anon_rmap_pte(folio, page, vma, address, RMAP_NONE);
738 else
739 /*
740 * Currently device exclusive access only supports anonymous
741 * memory so the entry shouldn't point to a filebacked page.
742 */
743 WARN_ON_ONCE(1);
744
745 set_pte_at(vma->vm_mm, address, ptep, pte);
746
747 /*
748 * No need to invalidate - it was non-present before. However
749 * secondary CPUs may have mappings that need invalidating.
750 */
751 update_mmu_cache(vma, address, ptep);
752 }
753
754 /*
755 * Tries to restore an exclusive pte if the page lock can be acquired without
756 * sleeping.
757 */
758 static int
try_restore_exclusive_pte(pte_t * src_pte,struct vm_area_struct * vma,unsigned long addr)759 try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
760 unsigned long addr)
761 {
762 swp_entry_t entry = pte_to_swp_entry(ptep_get(src_pte));
763 struct page *page = pfn_swap_entry_to_page(entry);
764
765 if (trylock_page(page)) {
766 restore_exclusive_pte(vma, page, addr, src_pte);
767 unlock_page(page);
768 return 0;
769 }
770
771 return -EBUSY;
772 }
773
774 /*
775 * copy one vm_area from one task to the other. Assumes the page tables
776 * already present in the new task to be cleared in the whole range
777 * covered by this vma.
778 */
779
780 static unsigned long
copy_nonpresent_pte(struct mm_struct * dst_mm,struct mm_struct * src_mm,pte_t * dst_pte,pte_t * src_pte,struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,unsigned long addr,int * rss)781 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
782 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
783 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
784 {
785 unsigned long vm_flags = dst_vma->vm_flags;
786 pte_t orig_pte = ptep_get(src_pte);
787 pte_t pte = orig_pte;
788 struct folio *folio;
789 struct page *page;
790 swp_entry_t entry = pte_to_swp_entry(orig_pte);
791
792 if (likely(!non_swap_entry(entry))) {
793 if (swap_duplicate(entry) < 0)
794 return -EIO;
795
796 /* make sure dst_mm is on swapoff's mmlist. */
797 if (unlikely(list_empty(&dst_mm->mmlist))) {
798 spin_lock(&mmlist_lock);
799 if (list_empty(&dst_mm->mmlist))
800 list_add(&dst_mm->mmlist,
801 &src_mm->mmlist);
802 spin_unlock(&mmlist_lock);
803 }
804 /* Mark the swap entry as shared. */
805 if (pte_swp_exclusive(orig_pte)) {
806 pte = pte_swp_clear_exclusive(orig_pte);
807 set_pte_at(src_mm, addr, src_pte, pte);
808 }
809 rss[MM_SWAPENTS]++;
810 } else if (is_migration_entry(entry)) {
811 folio = pfn_swap_entry_folio(entry);
812
813 rss[mm_counter(folio)]++;
814
815 if (!is_readable_migration_entry(entry) &&
816 is_cow_mapping(vm_flags)) {
817 /*
818 * COW mappings require pages in both parent and child
819 * to be set to read. A previously exclusive entry is
820 * now shared.
821 */
822 entry = make_readable_migration_entry(
823 swp_offset(entry));
824 pte = swp_entry_to_pte(entry);
825 if (pte_swp_soft_dirty(orig_pte))
826 pte = pte_swp_mksoft_dirty(pte);
827 if (pte_swp_uffd_wp(orig_pte))
828 pte = pte_swp_mkuffd_wp(pte);
829 set_pte_at(src_mm, addr, src_pte, pte);
830 }
831 } else if (is_device_private_entry(entry)) {
832 page = pfn_swap_entry_to_page(entry);
833 folio = page_folio(page);
834
835 /*
836 * Update rss count even for unaddressable pages, as
837 * they should treated just like normal pages in this
838 * respect.
839 *
840 * We will likely want to have some new rss counters
841 * for unaddressable pages, at some point. But for now
842 * keep things as they are.
843 */
844 folio_get(folio);
845 rss[mm_counter(folio)]++;
846 /* Cannot fail as these pages cannot get pinned. */
847 folio_try_dup_anon_rmap_pte(folio, page, src_vma);
848
849 /*
850 * We do not preserve soft-dirty information, because so
851 * far, checkpoint/restore is the only feature that
852 * requires that. And checkpoint/restore does not work
853 * when a device driver is involved (you cannot easily
854 * save and restore device driver state).
855 */
856 if (is_writable_device_private_entry(entry) &&
857 is_cow_mapping(vm_flags)) {
858 entry = make_readable_device_private_entry(
859 swp_offset(entry));
860 pte = swp_entry_to_pte(entry);
861 if (pte_swp_uffd_wp(orig_pte))
862 pte = pte_swp_mkuffd_wp(pte);
863 set_pte_at(src_mm, addr, src_pte, pte);
864 }
865 } else if (is_device_exclusive_entry(entry)) {
866 /*
867 * Make device exclusive entries present by restoring the
868 * original entry then copying as for a present pte. Device
869 * exclusive entries currently only support private writable
870 * (ie. COW) mappings.
871 */
872 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
873 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
874 return -EBUSY;
875 return -ENOENT;
876 } else if (is_pte_marker_entry(entry)) {
877 pte_marker marker = copy_pte_marker(entry, dst_vma);
878
879 if (marker)
880 set_pte_at(dst_mm, addr, dst_pte,
881 make_pte_marker(marker));
882 return 0;
883 }
884 if (!userfaultfd_wp(dst_vma))
885 pte = pte_swp_clear_uffd_wp(pte);
886 set_pte_at(dst_mm, addr, dst_pte, pte);
887 return 0;
888 }
889
890 /*
891 * Copy a present and normal page.
892 *
893 * NOTE! The usual case is that this isn't required;
894 * instead, the caller can just increase the page refcount
895 * and re-use the pte the traditional way.
896 *
897 * And if we need a pre-allocated page but don't yet have
898 * one, return a negative error to let the preallocation
899 * code know so that it can do so outside the page table
900 * lock.
901 */
902 static inline int
copy_present_page(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,pte_t * dst_pte,pte_t * src_pte,unsigned long addr,int * rss,struct folio ** prealloc,struct page * page)903 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
904 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
905 struct folio **prealloc, struct page *page)
906 {
907 struct folio *new_folio;
908 pte_t pte;
909
910 new_folio = *prealloc;
911 if (!new_folio)
912 return -EAGAIN;
913
914 /*
915 * We have a prealloc page, all good! Take it
916 * over and copy the page & arm it.
917 */
918 *prealloc = NULL;
919 copy_user_highpage(&new_folio->page, page, addr, src_vma);
920 __folio_mark_uptodate(new_folio);
921 folio_add_new_anon_rmap(new_folio, dst_vma, addr);
922 folio_add_lru_vma(new_folio, dst_vma);
923 rss[MM_ANONPAGES]++;
924
925 /* All done, just insert the new page copy in the child */
926 pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot);
927 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
928 if (userfaultfd_pte_wp(dst_vma, ptep_get(src_pte)))
929 /* Uffd-wp needs to be delivered to dest pte as well */
930 pte = pte_mkuffd_wp(pte);
931 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
932 return 0;
933 }
934
__copy_present_ptes(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,pte_t * dst_pte,pte_t * src_pte,pte_t pte,unsigned long addr,int nr)935 static __always_inline void __copy_present_ptes(struct vm_area_struct *dst_vma,
936 struct vm_area_struct *src_vma, pte_t *dst_pte, pte_t *src_pte,
937 pte_t pte, unsigned long addr, int nr)
938 {
939 struct mm_struct *src_mm = src_vma->vm_mm;
940
941 /* If it's a COW mapping, write protect it both processes. */
942 if (is_cow_mapping(src_vma->vm_flags) && pte_write(pte)) {
943 wrprotect_ptes(src_mm, addr, src_pte, nr);
944 pte = pte_wrprotect(pte);
945 }
946
947 /* If it's a shared mapping, mark it clean in the child. */
948 if (src_vma->vm_flags & VM_SHARED)
949 pte = pte_mkclean(pte);
950 pte = pte_mkold(pte);
951
952 if (!userfaultfd_wp(dst_vma))
953 pte = pte_clear_uffd_wp(pte);
954
955 set_ptes(dst_vma->vm_mm, addr, dst_pte, pte, nr);
956 }
957
958 /*
959 * Copy one present PTE, trying to batch-process subsequent PTEs that map
960 * consecutive pages of the same folio by copying them as well.
961 *
962 * Returns -EAGAIN if one preallocated page is required to copy the next PTE.
963 * Otherwise, returns the number of copied PTEs (at least 1).
964 */
965 static inline int
copy_present_ptes(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,pte_t * dst_pte,pte_t * src_pte,pte_t pte,unsigned long addr,int max_nr,int * rss,struct folio ** prealloc)966 copy_present_ptes(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
967 pte_t *dst_pte, pte_t *src_pte, pte_t pte, unsigned long addr,
968 int max_nr, int *rss, struct folio **prealloc)
969 {
970 struct page *page;
971 struct folio *folio;
972 bool any_writable;
973 fpb_t flags = 0;
974 int err, nr;
975
976 page = vm_normal_page(src_vma, addr, pte);
977 if (unlikely(!page))
978 goto copy_pte;
979
980 folio = page_folio(page);
981
982 /*
983 * If we likely have to copy, just don't bother with batching. Make
984 * sure that the common "small folio" case is as fast as possible
985 * by keeping the batching logic separate.
986 */
987 if (unlikely(!*prealloc && folio_test_large(folio) && max_nr != 1)) {
988 if (src_vma->vm_flags & VM_SHARED)
989 flags |= FPB_IGNORE_DIRTY;
990 if (!vma_soft_dirty_enabled(src_vma))
991 flags |= FPB_IGNORE_SOFT_DIRTY;
992
993 nr = folio_pte_batch(folio, addr, src_pte, pte, max_nr, flags,
994 &any_writable, NULL, NULL);
995 folio_ref_add(folio, nr);
996 if (folio_test_anon(folio)) {
997 if (unlikely(folio_try_dup_anon_rmap_ptes(folio, page,
998 nr, src_vma))) {
999 folio_ref_sub(folio, nr);
1000 return -EAGAIN;
1001 }
1002 rss[MM_ANONPAGES] += nr;
1003 VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio);
1004 } else {
1005 folio_dup_file_rmap_ptes(folio, page, nr);
1006 rss[mm_counter_file(folio)] += nr;
1007 }
1008 if (any_writable)
1009 pte = pte_mkwrite(pte, src_vma);
1010 __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte,
1011 addr, nr);
1012 return nr;
1013 }
1014
1015 folio_get(folio);
1016 if (folio_test_anon(folio)) {
1017 /*
1018 * If this page may have been pinned by the parent process,
1019 * copy the page immediately for the child so that we'll always
1020 * guarantee the pinned page won't be randomly replaced in the
1021 * future.
1022 */
1023 if (unlikely(folio_try_dup_anon_rmap_pte(folio, page, src_vma))) {
1024 /* Page may be pinned, we have to copy. */
1025 folio_put(folio);
1026 err = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
1027 addr, rss, prealloc, page);
1028 return err ? err : 1;
1029 }
1030 rss[MM_ANONPAGES]++;
1031 VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio);
1032 } else {
1033 folio_dup_file_rmap_pte(folio, page);
1034 rss[mm_counter_file(folio)]++;
1035 }
1036
1037 copy_pte:
1038 __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte, addr, 1);
1039 return 1;
1040 }
1041
folio_prealloc(struct mm_struct * src_mm,struct vm_area_struct * vma,unsigned long addr,bool need_zero)1042 static inline struct folio *folio_prealloc(struct mm_struct *src_mm,
1043 struct vm_area_struct *vma, unsigned long addr, bool need_zero)
1044 {
1045 struct folio *new_folio;
1046
1047 if (need_zero)
1048 new_folio = vma_alloc_zeroed_movable_folio(vma, addr);
1049 else
1050 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma,
1051 addr, false);
1052
1053 if (!new_folio)
1054 return NULL;
1055
1056 if (mem_cgroup_charge(new_folio, src_mm, GFP_KERNEL)) {
1057 folio_put(new_folio);
1058 return NULL;
1059 }
1060 folio_throttle_swaprate(new_folio, GFP_KERNEL);
1061
1062 return new_folio;
1063 }
1064
1065 static int
copy_pte_range(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,pmd_t * dst_pmd,pmd_t * src_pmd,unsigned long addr,unsigned long end)1066 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1067 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1068 unsigned long end)
1069 {
1070 struct mm_struct *dst_mm = dst_vma->vm_mm;
1071 struct mm_struct *src_mm = src_vma->vm_mm;
1072 pte_t *orig_src_pte, *orig_dst_pte;
1073 pte_t *src_pte, *dst_pte;
1074 pte_t ptent;
1075 spinlock_t *src_ptl, *dst_ptl;
1076 int progress, max_nr, ret = 0;
1077 int rss[NR_MM_COUNTERS];
1078 swp_entry_t entry = (swp_entry_t){0};
1079 struct folio *prealloc = NULL;
1080 int nr;
1081
1082 again:
1083 progress = 0;
1084 init_rss_vec(rss);
1085
1086 /*
1087 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1088 * error handling here, assume that exclusive mmap_lock on dst and src
1089 * protects anon from unexpected THP transitions; with shmem and file
1090 * protected by mmap_lock-less collapse skipping areas with anon_vma
1091 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1092 * can remove such assumptions later, but this is good enough for now.
1093 */
1094 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1095 if (!dst_pte) {
1096 ret = -ENOMEM;
1097 goto out;
1098 }
1099 src_pte = pte_offset_map_nolock(src_mm, src_pmd, addr, &src_ptl);
1100 if (!src_pte) {
1101 pte_unmap_unlock(dst_pte, dst_ptl);
1102 /* ret == 0 */
1103 goto out;
1104 }
1105 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1106 orig_src_pte = src_pte;
1107 orig_dst_pte = dst_pte;
1108 arch_enter_lazy_mmu_mode();
1109
1110 do {
1111 nr = 1;
1112
1113 /*
1114 * We are holding two locks at this point - either of them
1115 * could generate latencies in another task on another CPU.
1116 */
1117 if (progress >= 32) {
1118 progress = 0;
1119 if (need_resched() ||
1120 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1121 break;
1122 }
1123 ptent = ptep_get(src_pte);
1124 if (pte_none(ptent)) {
1125 progress++;
1126 continue;
1127 }
1128 if (unlikely(!pte_present(ptent))) {
1129 ret = copy_nonpresent_pte(dst_mm, src_mm,
1130 dst_pte, src_pte,
1131 dst_vma, src_vma,
1132 addr, rss);
1133 if (ret == -EIO) {
1134 entry = pte_to_swp_entry(ptep_get(src_pte));
1135 break;
1136 } else if (ret == -EBUSY) {
1137 break;
1138 } else if (!ret) {
1139 progress += 8;
1140 continue;
1141 }
1142 ptent = ptep_get(src_pte);
1143 VM_WARN_ON_ONCE(!pte_present(ptent));
1144
1145 /*
1146 * Device exclusive entry restored, continue by copying
1147 * the now present pte.
1148 */
1149 WARN_ON_ONCE(ret != -ENOENT);
1150 }
1151 /* copy_present_ptes() will clear `*prealloc' if consumed */
1152 max_nr = (end - addr) / PAGE_SIZE;
1153 ret = copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte,
1154 ptent, addr, max_nr, rss, &prealloc);
1155 /*
1156 * If we need a pre-allocated page for this pte, drop the
1157 * locks, allocate, and try again.
1158 */
1159 if (unlikely(ret == -EAGAIN))
1160 break;
1161 if (unlikely(prealloc)) {
1162 /*
1163 * pre-alloc page cannot be reused by next time so as
1164 * to strictly follow mempolicy (e.g., alloc_page_vma()
1165 * will allocate page according to address). This
1166 * could only happen if one pinned pte changed.
1167 */
1168 folio_put(prealloc);
1169 prealloc = NULL;
1170 }
1171 nr = ret;
1172 progress += 8 * nr;
1173 } while (dst_pte += nr, src_pte += nr, addr += PAGE_SIZE * nr,
1174 addr != end);
1175
1176 arch_leave_lazy_mmu_mode();
1177 pte_unmap_unlock(orig_src_pte, src_ptl);
1178 add_mm_rss_vec(dst_mm, rss);
1179 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1180 cond_resched();
1181
1182 if (ret == -EIO) {
1183 VM_WARN_ON_ONCE(!entry.val);
1184 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1185 ret = -ENOMEM;
1186 goto out;
1187 }
1188 entry.val = 0;
1189 } else if (ret == -EBUSY) {
1190 goto out;
1191 } else if (ret == -EAGAIN) {
1192 prealloc = folio_prealloc(src_mm, src_vma, addr, false);
1193 if (!prealloc)
1194 return -ENOMEM;
1195 } else if (ret < 0) {
1196 VM_WARN_ON_ONCE(1);
1197 }
1198
1199 /* We've captured and resolved the error. Reset, try again. */
1200 ret = 0;
1201
1202 if (addr != end)
1203 goto again;
1204 out:
1205 if (unlikely(prealloc))
1206 folio_put(prealloc);
1207 return ret;
1208 }
1209
1210 static inline int
copy_pmd_range(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,pud_t * dst_pud,pud_t * src_pud,unsigned long addr,unsigned long end)1211 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1212 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1213 unsigned long end)
1214 {
1215 struct mm_struct *dst_mm = dst_vma->vm_mm;
1216 struct mm_struct *src_mm = src_vma->vm_mm;
1217 pmd_t *src_pmd, *dst_pmd;
1218 unsigned long next;
1219
1220 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1221 if (!dst_pmd)
1222 return -ENOMEM;
1223 src_pmd = pmd_offset(src_pud, addr);
1224 do {
1225 next = pmd_addr_end(addr, end);
1226 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1227 || pmd_devmap(*src_pmd)) {
1228 int err;
1229 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1230 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1231 addr, dst_vma, src_vma);
1232 if (err == -ENOMEM)
1233 return -ENOMEM;
1234 if (!err)
1235 continue;
1236 /* fall through */
1237 }
1238 if (pmd_none_or_clear_bad(src_pmd))
1239 continue;
1240 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1241 addr, next))
1242 return -ENOMEM;
1243 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1244 return 0;
1245 }
1246
1247 static inline int
copy_pud_range(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,p4d_t * dst_p4d,p4d_t * src_p4d,unsigned long addr,unsigned long end)1248 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1249 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1250 unsigned long end)
1251 {
1252 struct mm_struct *dst_mm = dst_vma->vm_mm;
1253 struct mm_struct *src_mm = src_vma->vm_mm;
1254 pud_t *src_pud, *dst_pud;
1255 unsigned long next;
1256
1257 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1258 if (!dst_pud)
1259 return -ENOMEM;
1260 src_pud = pud_offset(src_p4d, addr);
1261 do {
1262 next = pud_addr_end(addr, end);
1263 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1264 int err;
1265
1266 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1267 err = copy_huge_pud(dst_mm, src_mm,
1268 dst_pud, src_pud, addr, src_vma);
1269 if (err == -ENOMEM)
1270 return -ENOMEM;
1271 if (!err)
1272 continue;
1273 /* fall through */
1274 }
1275 if (pud_none_or_clear_bad(src_pud))
1276 continue;
1277 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1278 addr, next))
1279 return -ENOMEM;
1280 } while (dst_pud++, src_pud++, addr = next, addr != end);
1281 return 0;
1282 }
1283
1284 static inline int
copy_p4d_range(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,pgd_t * dst_pgd,pgd_t * src_pgd,unsigned long addr,unsigned long end)1285 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1286 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1287 unsigned long end)
1288 {
1289 struct mm_struct *dst_mm = dst_vma->vm_mm;
1290 p4d_t *src_p4d, *dst_p4d;
1291 unsigned long next;
1292
1293 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1294 if (!dst_p4d)
1295 return -ENOMEM;
1296 src_p4d = p4d_offset(src_pgd, addr);
1297 do {
1298 next = p4d_addr_end(addr, end);
1299 if (p4d_none_or_clear_bad(src_p4d))
1300 continue;
1301 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1302 addr, next))
1303 return -ENOMEM;
1304 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1305 return 0;
1306 }
1307
1308 /*
1309 * Return true if the vma needs to copy the pgtable during this fork(). Return
1310 * false when we can speed up fork() by allowing lazy page faults later until
1311 * when the child accesses the memory range.
1312 */
1313 static bool
vma_needs_copy(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma)1314 vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1315 {
1316 /*
1317 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1318 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1319 * contains uffd-wp protection information, that's something we can't
1320 * retrieve from page cache, and skip copying will lose those info.
1321 */
1322 if (userfaultfd_wp(dst_vma))
1323 return true;
1324
1325 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1326 return true;
1327
1328 if (src_vma->anon_vma)
1329 return true;
1330
1331 /*
1332 * Don't copy ptes where a page fault will fill them correctly. Fork
1333 * becomes much lighter when there are big shared or private readonly
1334 * mappings. The tradeoff is that copy_page_range is more efficient
1335 * than faulting.
1336 */
1337 return false;
1338 }
1339
1340 int
copy_page_range(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma)1341 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1342 {
1343 pgd_t *src_pgd, *dst_pgd;
1344 unsigned long next;
1345 unsigned long addr = src_vma->vm_start;
1346 unsigned long end = src_vma->vm_end;
1347 struct mm_struct *dst_mm = dst_vma->vm_mm;
1348 struct mm_struct *src_mm = src_vma->vm_mm;
1349 struct mmu_notifier_range range;
1350 bool is_cow;
1351 int ret;
1352
1353 if (!vma_needs_copy(dst_vma, src_vma))
1354 return 0;
1355
1356 if (is_vm_hugetlb_page(src_vma))
1357 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1358
1359 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1360 /*
1361 * We do not free on error cases below as remove_vma
1362 * gets called on error from higher level routine
1363 */
1364 ret = track_pfn_copy(src_vma);
1365 if (ret)
1366 return ret;
1367 }
1368
1369 /*
1370 * We need to invalidate the secondary MMU mappings only when
1371 * there could be a permission downgrade on the ptes of the
1372 * parent mm. And a permission downgrade will only happen if
1373 * is_cow_mapping() returns true.
1374 */
1375 is_cow = is_cow_mapping(src_vma->vm_flags);
1376
1377 if (is_cow) {
1378 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1379 0, src_mm, addr, end);
1380 mmu_notifier_invalidate_range_start(&range);
1381 /*
1382 * Disabling preemption is not needed for the write side, as
1383 * the read side doesn't spin, but goes to the mmap_lock.
1384 *
1385 * Use the raw variant of the seqcount_t write API to avoid
1386 * lockdep complaining about preemptibility.
1387 */
1388 vma_assert_write_locked(src_vma);
1389 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1390 }
1391
1392 ret = 0;
1393 dst_pgd = pgd_offset(dst_mm, addr);
1394 src_pgd = pgd_offset(src_mm, addr);
1395 do {
1396 next = pgd_addr_end(addr, end);
1397 if (pgd_none_or_clear_bad(src_pgd))
1398 continue;
1399 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1400 addr, next))) {
1401 untrack_pfn_clear(dst_vma);
1402 ret = -ENOMEM;
1403 break;
1404 }
1405 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1406
1407 if (is_cow) {
1408 raw_write_seqcount_end(&src_mm->write_protect_seq);
1409 mmu_notifier_invalidate_range_end(&range);
1410 }
1411 return ret;
1412 }
1413
1414 /* Whether we should zap all COWed (private) pages too */
should_zap_cows(struct zap_details * details)1415 static inline bool should_zap_cows(struct zap_details *details)
1416 {
1417 /* By default, zap all pages */
1418 if (!details)
1419 return true;
1420
1421 /* Or, we zap COWed pages only if the caller wants to */
1422 return details->even_cows;
1423 }
1424
1425 /* Decides whether we should zap this folio with the folio pointer specified */
should_zap_folio(struct zap_details * details,struct folio * folio)1426 static inline bool should_zap_folio(struct zap_details *details,
1427 struct folio *folio)
1428 {
1429 /* If we can make a decision without *folio.. */
1430 if (should_zap_cows(details))
1431 return true;
1432
1433 /* Otherwise we should only zap non-anon folios */
1434 return !folio_test_anon(folio);
1435 }
1436
zap_drop_file_uffd_wp(struct zap_details * details)1437 static inline bool zap_drop_file_uffd_wp(struct zap_details *details)
1438 {
1439 if (!details)
1440 return false;
1441
1442 return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1443 }
1444
1445 /*
1446 * This function makes sure that we'll replace the none pte with an uffd-wp
1447 * swap special pte marker when necessary. Must be with the pgtable lock held.
1448 */
1449 static inline void
zap_install_uffd_wp_if_needed(struct vm_area_struct * vma,unsigned long addr,pte_t * pte,int nr,struct zap_details * details,pte_t pteval)1450 zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1451 unsigned long addr, pte_t *pte, int nr,
1452 struct zap_details *details, pte_t pteval)
1453 {
1454 /* Zap on anonymous always means dropping everything */
1455 if (vma_is_anonymous(vma))
1456 return;
1457
1458 if (zap_drop_file_uffd_wp(details))
1459 return;
1460
1461 for (;;) {
1462 /* the PFN in the PTE is irrelevant. */
1463 pte_install_uffd_wp_if_needed(vma, addr, pte, pteval);
1464 if (--nr == 0)
1465 break;
1466 pte++;
1467 addr += PAGE_SIZE;
1468 }
1469 }
1470
zap_present_folio_ptes(struct mmu_gather * tlb,struct vm_area_struct * vma,struct folio * folio,struct page * page,pte_t * pte,pte_t ptent,unsigned int nr,unsigned long addr,struct zap_details * details,int * rss,bool * force_flush,bool * force_break)1471 static __always_inline void zap_present_folio_ptes(struct mmu_gather *tlb,
1472 struct vm_area_struct *vma, struct folio *folio,
1473 struct page *page, pte_t *pte, pte_t ptent, unsigned int nr,
1474 unsigned long addr, struct zap_details *details, int *rss,
1475 bool *force_flush, bool *force_break)
1476 {
1477 struct mm_struct *mm = tlb->mm;
1478 bool delay_rmap = false;
1479
1480 if (!folio_test_anon(folio)) {
1481 ptent = get_and_clear_full_ptes(mm, addr, pte, nr, tlb->fullmm);
1482 if (pte_dirty(ptent)) {
1483 folio_mark_dirty(folio);
1484 if (tlb_delay_rmap(tlb)) {
1485 delay_rmap = true;
1486 *force_flush = true;
1487 }
1488 }
1489 if (pte_young(ptent) && likely(vma_has_recency(vma)))
1490 folio_mark_accessed(folio);
1491 rss[mm_counter(folio)] -= nr;
1492 } else {
1493 /* We don't need up-to-date accessed/dirty bits. */
1494 clear_full_ptes(mm, addr, pte, nr, tlb->fullmm);
1495 rss[MM_ANONPAGES] -= nr;
1496 }
1497 /* Checking a single PTE in a batch is sufficient. */
1498 arch_check_zapped_pte(vma, ptent);
1499 tlb_remove_tlb_entries(tlb, pte, nr, addr);
1500 if (unlikely(userfaultfd_pte_wp(vma, ptent)))
1501 zap_install_uffd_wp_if_needed(vma, addr, pte, nr, details,
1502 ptent);
1503
1504 if (!delay_rmap) {
1505 folio_remove_rmap_ptes(folio, page, nr, vma);
1506
1507 if (unlikely(folio_mapcount(folio) < 0))
1508 print_bad_pte(vma, addr, ptent, page);
1509 }
1510 if (unlikely(__tlb_remove_folio_pages(tlb, page, nr, delay_rmap))) {
1511 *force_flush = true;
1512 *force_break = true;
1513 }
1514 }
1515
1516 /*
1517 * Zap or skip at least one present PTE, trying to batch-process subsequent
1518 * PTEs that map consecutive pages of the same folio.
1519 *
1520 * Returns the number of processed (skipped or zapped) PTEs (at least 1).
1521 */
zap_present_ptes(struct mmu_gather * tlb,struct vm_area_struct * vma,pte_t * pte,pte_t ptent,unsigned int max_nr,unsigned long addr,struct zap_details * details,int * rss,bool * force_flush,bool * force_break)1522 static inline int zap_present_ptes(struct mmu_gather *tlb,
1523 struct vm_area_struct *vma, pte_t *pte, pte_t ptent,
1524 unsigned int max_nr, unsigned long addr,
1525 struct zap_details *details, int *rss, bool *force_flush,
1526 bool *force_break)
1527 {
1528 const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
1529 struct mm_struct *mm = tlb->mm;
1530 struct folio *folio;
1531 struct page *page;
1532 int nr;
1533
1534 page = vm_normal_page(vma, addr, ptent);
1535 if (!page) {
1536 /* We don't need up-to-date accessed/dirty bits. */
1537 ptep_get_and_clear_full(mm, addr, pte, tlb->fullmm);
1538 arch_check_zapped_pte(vma, ptent);
1539 tlb_remove_tlb_entry(tlb, pte, addr);
1540 if (userfaultfd_pte_wp(vma, ptent))
1541 zap_install_uffd_wp_if_needed(vma, addr, pte, 1,
1542 details, ptent);
1543 ksm_might_unmap_zero_page(mm, ptent);
1544 return 1;
1545 }
1546
1547 folio = page_folio(page);
1548 if (unlikely(!should_zap_folio(details, folio)))
1549 return 1;
1550
1551 /*
1552 * Make sure that the common "small folio" case is as fast as possible
1553 * by keeping the batching logic separate.
1554 */
1555 if (unlikely(folio_test_large(folio) && max_nr != 1)) {
1556 nr = folio_pte_batch(folio, addr, pte, ptent, max_nr, fpb_flags,
1557 NULL, NULL, NULL);
1558
1559 zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, nr,
1560 addr, details, rss, force_flush,
1561 force_break);
1562 return nr;
1563 }
1564 zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, 1, addr,
1565 details, rss, force_flush, force_break);
1566 return 1;
1567 }
1568
zap_pte_range(struct mmu_gather * tlb,struct vm_area_struct * vma,pmd_t * pmd,unsigned long addr,unsigned long end,struct zap_details * details)1569 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1570 struct vm_area_struct *vma, pmd_t *pmd,
1571 unsigned long addr, unsigned long end,
1572 struct zap_details *details)
1573 {
1574 bool force_flush = false, force_break = false;
1575 struct mm_struct *mm = tlb->mm;
1576 int rss[NR_MM_COUNTERS];
1577 spinlock_t *ptl;
1578 pte_t *start_pte;
1579 pte_t *pte;
1580 swp_entry_t entry;
1581 int nr;
1582
1583 tlb_change_page_size(tlb, PAGE_SIZE);
1584 init_rss_vec(rss);
1585 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1586 if (!pte)
1587 return addr;
1588
1589 flush_tlb_batched_pending(mm);
1590 arch_enter_lazy_mmu_mode();
1591 do {
1592 pte_t ptent = ptep_get(pte);
1593 struct folio *folio;
1594 struct page *page;
1595 int max_nr;
1596
1597 nr = 1;
1598 if (pte_none(ptent))
1599 continue;
1600
1601 if (need_resched())
1602 break;
1603
1604 if (pte_present(ptent)) {
1605 max_nr = (end - addr) / PAGE_SIZE;
1606 nr = zap_present_ptes(tlb, vma, pte, ptent, max_nr,
1607 addr, details, rss, &force_flush,
1608 &force_break);
1609 if (unlikely(force_break)) {
1610 addr += nr * PAGE_SIZE;
1611 break;
1612 }
1613 continue;
1614 }
1615
1616 entry = pte_to_swp_entry(ptent);
1617 if (is_device_private_entry(entry) ||
1618 is_device_exclusive_entry(entry)) {
1619 page = pfn_swap_entry_to_page(entry);
1620 folio = page_folio(page);
1621 if (unlikely(!should_zap_folio(details, folio)))
1622 continue;
1623 /*
1624 * Both device private/exclusive mappings should only
1625 * work with anonymous page so far, so we don't need to
1626 * consider uffd-wp bit when zap. For more information,
1627 * see zap_install_uffd_wp_if_needed().
1628 */
1629 WARN_ON_ONCE(!vma_is_anonymous(vma));
1630 rss[mm_counter(folio)]--;
1631 if (is_device_private_entry(entry))
1632 folio_remove_rmap_pte(folio, page, vma);
1633 folio_put(folio);
1634 } else if (!non_swap_entry(entry)) {
1635 max_nr = (end - addr) / PAGE_SIZE;
1636 nr = swap_pte_batch(pte, max_nr, ptent);
1637 /* Genuine swap entries, hence a private anon pages */
1638 if (!should_zap_cows(details))
1639 continue;
1640 rss[MM_SWAPENTS] -= nr;
1641 free_swap_and_cache_nr(entry, nr);
1642 } else if (is_migration_entry(entry)) {
1643 folio = pfn_swap_entry_folio(entry);
1644 if (!should_zap_folio(details, folio))
1645 continue;
1646 rss[mm_counter(folio)]--;
1647 } else if (pte_marker_entry_uffd_wp(entry)) {
1648 /*
1649 * For anon: always drop the marker; for file: only
1650 * drop the marker if explicitly requested.
1651 */
1652 if (!vma_is_anonymous(vma) &&
1653 !zap_drop_file_uffd_wp(details))
1654 continue;
1655 } else if (is_hwpoison_entry(entry) ||
1656 is_poisoned_swp_entry(entry)) {
1657 if (!should_zap_cows(details))
1658 continue;
1659 } else {
1660 /* We should have covered all the swap entry types */
1661 pr_alert("unrecognized swap entry 0x%lx\n", entry.val);
1662 WARN_ON_ONCE(1);
1663 }
1664 clear_not_present_full_ptes(mm, addr, pte, nr, tlb->fullmm);
1665 zap_install_uffd_wp_if_needed(vma, addr, pte, nr, details, ptent);
1666 } while (pte += nr, addr += PAGE_SIZE * nr, addr != end);
1667
1668 add_mm_rss_vec(mm, rss);
1669 arch_leave_lazy_mmu_mode();
1670
1671 /* Do the actual TLB flush before dropping ptl */
1672 if (force_flush) {
1673 tlb_flush_mmu_tlbonly(tlb);
1674 tlb_flush_rmaps(tlb, vma);
1675 }
1676 pte_unmap_unlock(start_pte, ptl);
1677
1678 /*
1679 * If we forced a TLB flush (either due to running out of
1680 * batch buffers or because we needed to flush dirty TLB
1681 * entries before releasing the ptl), free the batched
1682 * memory too. Come back again if we didn't do everything.
1683 */
1684 if (force_flush)
1685 tlb_flush_mmu(tlb);
1686
1687 return addr;
1688 }
1689
zap_pmd_range(struct mmu_gather * tlb,struct vm_area_struct * vma,pud_t * pud,unsigned long addr,unsigned long end,struct zap_details * details)1690 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1691 struct vm_area_struct *vma, pud_t *pud,
1692 unsigned long addr, unsigned long end,
1693 struct zap_details *details)
1694 {
1695 pmd_t *pmd;
1696 unsigned long next;
1697
1698 pmd = pmd_offset(pud, addr);
1699 do {
1700 next = pmd_addr_end(addr, end);
1701 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1702 if (next - addr != HPAGE_PMD_SIZE)
1703 __split_huge_pmd(vma, pmd, addr, false, NULL);
1704 else if (zap_huge_pmd(tlb, vma, pmd, addr)) {
1705 addr = next;
1706 continue;
1707 }
1708 /* fall through */
1709 } else if (details && details->single_folio &&
1710 folio_test_pmd_mappable(details->single_folio) &&
1711 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1712 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1713 /*
1714 * Take and drop THP pmd lock so that we cannot return
1715 * prematurely, while zap_huge_pmd() has cleared *pmd,
1716 * but not yet decremented compound_mapcount().
1717 */
1718 spin_unlock(ptl);
1719 }
1720 if (pmd_none(*pmd)) {
1721 addr = next;
1722 continue;
1723 }
1724 addr = zap_pte_range(tlb, vma, pmd, addr, next, details);
1725 if (addr != next)
1726 pmd--;
1727 } while (pmd++, cond_resched(), addr != end);
1728
1729 return addr;
1730 }
1731
zap_pud_range(struct mmu_gather * tlb,struct vm_area_struct * vma,p4d_t * p4d,unsigned long addr,unsigned long end,struct zap_details * details)1732 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1733 struct vm_area_struct *vma, p4d_t *p4d,
1734 unsigned long addr, unsigned long end,
1735 struct zap_details *details)
1736 {
1737 pud_t *pud;
1738 unsigned long next;
1739
1740 pud = pud_offset(p4d, addr);
1741 do {
1742 next = pud_addr_end(addr, end);
1743 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1744 if (next - addr != HPAGE_PUD_SIZE) {
1745 mmap_assert_locked(tlb->mm);
1746 split_huge_pud(vma, pud, addr);
1747 } else if (zap_huge_pud(tlb, vma, pud, addr))
1748 goto next;
1749 /* fall through */
1750 }
1751 if (pud_none_or_clear_bad(pud))
1752 continue;
1753 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1754 next:
1755 cond_resched();
1756 } while (pud++, addr = next, addr != end);
1757
1758 return addr;
1759 }
1760
zap_p4d_range(struct mmu_gather * tlb,struct vm_area_struct * vma,pgd_t * pgd,unsigned long addr,unsigned long end,struct zap_details * details)1761 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1762 struct vm_area_struct *vma, pgd_t *pgd,
1763 unsigned long addr, unsigned long end,
1764 struct zap_details *details)
1765 {
1766 p4d_t *p4d;
1767 unsigned long next;
1768
1769 p4d = p4d_offset(pgd, addr);
1770 do {
1771 next = p4d_addr_end(addr, end);
1772 if (p4d_none_or_clear_bad(p4d))
1773 continue;
1774 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1775 } while (p4d++, addr = next, addr != end);
1776
1777 return addr;
1778 }
1779
unmap_page_range(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long addr,unsigned long end,struct zap_details * details)1780 void unmap_page_range(struct mmu_gather *tlb,
1781 struct vm_area_struct *vma,
1782 unsigned long addr, unsigned long end,
1783 struct zap_details *details)
1784 {
1785 pgd_t *pgd;
1786 unsigned long next;
1787
1788 BUG_ON(addr >= end);
1789 tlb_start_vma(tlb, vma);
1790 pgd = pgd_offset(vma->vm_mm, addr);
1791 do {
1792 next = pgd_addr_end(addr, end);
1793 if (pgd_none_or_clear_bad(pgd))
1794 continue;
1795 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1796 } while (pgd++, addr = next, addr != end);
1797 tlb_end_vma(tlb, vma);
1798 }
1799
1800
unmap_single_vma(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long start_addr,unsigned long end_addr,struct zap_details * details,bool mm_wr_locked)1801 static void unmap_single_vma(struct mmu_gather *tlb,
1802 struct vm_area_struct *vma, unsigned long start_addr,
1803 unsigned long end_addr,
1804 struct zap_details *details, bool mm_wr_locked)
1805 {
1806 unsigned long start = max(vma->vm_start, start_addr);
1807 unsigned long end;
1808
1809 if (start >= vma->vm_end)
1810 return;
1811 end = min(vma->vm_end, end_addr);
1812 if (end <= vma->vm_start)
1813 return;
1814
1815 if (vma->vm_file)
1816 uprobe_munmap(vma, start, end);
1817
1818 if (unlikely(vma->vm_flags & VM_PFNMAP))
1819 untrack_pfn(vma, 0, 0, mm_wr_locked);
1820
1821 if (start != end) {
1822 if (unlikely(is_vm_hugetlb_page(vma))) {
1823 /*
1824 * It is undesirable to test vma->vm_file as it
1825 * should be non-null for valid hugetlb area.
1826 * However, vm_file will be NULL in the error
1827 * cleanup path of mmap_region. When
1828 * hugetlbfs ->mmap method fails,
1829 * mmap_region() nullifies vma->vm_file
1830 * before calling this function to clean up.
1831 * Since no pte has actually been setup, it is
1832 * safe to do nothing in this case.
1833 */
1834 if (vma->vm_file) {
1835 zap_flags_t zap_flags = details ?
1836 details->zap_flags : 0;
1837 __unmap_hugepage_range(tlb, vma, start, end,
1838 NULL, zap_flags);
1839 }
1840 } else
1841 unmap_page_range(tlb, vma, start, end, details);
1842 }
1843 }
1844
1845 /**
1846 * unmap_vmas - unmap a range of memory covered by a list of vma's
1847 * @tlb: address of the caller's struct mmu_gather
1848 * @mas: the maple state
1849 * @vma: the starting vma
1850 * @start_addr: virtual address at which to start unmapping
1851 * @end_addr: virtual address at which to end unmapping
1852 * @tree_end: The maximum index to check
1853 * @mm_wr_locked: lock flag
1854 *
1855 * Unmap all pages in the vma list.
1856 *
1857 * Only addresses between `start' and `end' will be unmapped.
1858 *
1859 * The VMA list must be sorted in ascending virtual address order.
1860 *
1861 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1862 * range after unmap_vmas() returns. So the only responsibility here is to
1863 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1864 * drops the lock and schedules.
1865 */
unmap_vmas(struct mmu_gather * tlb,struct ma_state * mas,struct vm_area_struct * vma,unsigned long start_addr,unsigned long end_addr,unsigned long tree_end,bool mm_wr_locked)1866 void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas,
1867 struct vm_area_struct *vma, unsigned long start_addr,
1868 unsigned long end_addr, unsigned long tree_end,
1869 bool mm_wr_locked)
1870 {
1871 struct mmu_notifier_range range;
1872 struct zap_details details = {
1873 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
1874 /* Careful - we need to zap private pages too! */
1875 .even_cows = true,
1876 };
1877
1878 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm,
1879 start_addr, end_addr);
1880 mmu_notifier_invalidate_range_start(&range);
1881 do {
1882 unsigned long start = start_addr;
1883 unsigned long end = end_addr;
1884 hugetlb_zap_begin(vma, &start, &end);
1885 unmap_single_vma(tlb, vma, start, end, &details,
1886 mm_wr_locked);
1887 hugetlb_zap_end(vma, &details);
1888 vma = mas_find(mas, tree_end - 1);
1889 } while (vma && likely(!xa_is_zero(vma)));
1890 mmu_notifier_invalidate_range_end(&range);
1891 }
1892
1893 /**
1894 * zap_page_range_single - remove user pages in a given range
1895 * @vma: vm_area_struct holding the applicable pages
1896 * @address: starting address of pages to zap
1897 * @size: number of bytes to zap
1898 * @details: details of shared cache invalidation
1899 *
1900 * The range must fit into one VMA.
1901 */
zap_page_range_single(struct vm_area_struct * vma,unsigned long address,unsigned long size,struct zap_details * details)1902 void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1903 unsigned long size, struct zap_details *details)
1904 {
1905 const unsigned long end = address + size;
1906 struct mmu_notifier_range range;
1907 struct mmu_gather tlb;
1908
1909 lru_add_drain();
1910 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1911 address, end);
1912 hugetlb_zap_begin(vma, &range.start, &range.end);
1913 tlb_gather_mmu(&tlb, vma->vm_mm);
1914 update_hiwater_rss(vma->vm_mm);
1915 mmu_notifier_invalidate_range_start(&range);
1916 /*
1917 * unmap 'address-end' not 'range.start-range.end' as range
1918 * could have been expanded for hugetlb pmd sharing.
1919 */
1920 unmap_single_vma(&tlb, vma, address, end, details, false);
1921 mmu_notifier_invalidate_range_end(&range);
1922 tlb_finish_mmu(&tlb);
1923 hugetlb_zap_end(vma, details);
1924 }
1925
1926 /**
1927 * zap_vma_ptes - remove ptes mapping the vma
1928 * @vma: vm_area_struct holding ptes to be zapped
1929 * @address: starting address of pages to zap
1930 * @size: number of bytes to zap
1931 *
1932 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1933 *
1934 * The entire address range must be fully contained within the vma.
1935 *
1936 */
zap_vma_ptes(struct vm_area_struct * vma,unsigned long address,unsigned long size)1937 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1938 unsigned long size)
1939 {
1940 if (!range_in_vma(vma, address, address + size) ||
1941 !(vma->vm_flags & VM_PFNMAP))
1942 return;
1943
1944 zap_page_range_single(vma, address, size, NULL);
1945 }
1946 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1947
walk_to_pmd(struct mm_struct * mm,unsigned long addr)1948 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1949 {
1950 pgd_t *pgd;
1951 p4d_t *p4d;
1952 pud_t *pud;
1953 pmd_t *pmd;
1954
1955 pgd = pgd_offset(mm, addr);
1956 p4d = p4d_alloc(mm, pgd, addr);
1957 if (!p4d)
1958 return NULL;
1959 pud = pud_alloc(mm, p4d, addr);
1960 if (!pud)
1961 return NULL;
1962 pmd = pmd_alloc(mm, pud, addr);
1963 if (!pmd)
1964 return NULL;
1965
1966 VM_BUG_ON(pmd_trans_huge(*pmd));
1967 return pmd;
1968 }
1969
__get_locked_pte(struct mm_struct * mm,unsigned long addr,spinlock_t ** ptl)1970 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1971 spinlock_t **ptl)
1972 {
1973 pmd_t *pmd = walk_to_pmd(mm, addr);
1974
1975 if (!pmd)
1976 return NULL;
1977 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1978 }
1979
validate_page_before_insert(struct page * page)1980 static int validate_page_before_insert(struct page *page)
1981 {
1982 struct folio *folio = page_folio(page);
1983
1984 if (folio_test_anon(folio) || folio_test_slab(folio) ||
1985 page_has_type(page))
1986 return -EINVAL;
1987 flush_dcache_folio(folio);
1988 return 0;
1989 }
1990
insert_page_into_pte_locked(struct vm_area_struct * vma,pte_t * pte,unsigned long addr,struct page * page,pgprot_t prot)1991 static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1992 unsigned long addr, struct page *page, pgprot_t prot)
1993 {
1994 struct folio *folio = page_folio(page);
1995
1996 if (!pte_none(ptep_get(pte)))
1997 return -EBUSY;
1998 /* Ok, finally just insert the thing.. */
1999 folio_get(folio);
2000 inc_mm_counter(vma->vm_mm, mm_counter_file(folio));
2001 folio_add_file_rmap_pte(folio, page, vma);
2002 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
2003 return 0;
2004 }
2005
2006 /*
2007 * This is the old fallback for page remapping.
2008 *
2009 * For historical reasons, it only allows reserved pages. Only
2010 * old drivers should use this, and they needed to mark their
2011 * pages reserved for the old functions anyway.
2012 */
insert_page(struct vm_area_struct * vma,unsigned long addr,struct page * page,pgprot_t prot)2013 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2014 struct page *page, pgprot_t prot)
2015 {
2016 int retval;
2017 pte_t *pte;
2018 spinlock_t *ptl;
2019
2020 retval = validate_page_before_insert(page);
2021 if (retval)
2022 goto out;
2023 retval = -ENOMEM;
2024 pte = get_locked_pte(vma->vm_mm, addr, &ptl);
2025 if (!pte)
2026 goto out;
2027 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
2028 pte_unmap_unlock(pte, ptl);
2029 out:
2030 return retval;
2031 }
2032
insert_page_in_batch_locked(struct vm_area_struct * vma,pte_t * pte,unsigned long addr,struct page * page,pgprot_t prot)2033 static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
2034 unsigned long addr, struct page *page, pgprot_t prot)
2035 {
2036 int err;
2037
2038 if (!page_count(page))
2039 return -EINVAL;
2040 err = validate_page_before_insert(page);
2041 if (err)
2042 return err;
2043 return insert_page_into_pte_locked(vma, pte, addr, page, prot);
2044 }
2045
2046 /* insert_pages() amortizes the cost of spinlock operations
2047 * when inserting pages in a loop.
2048 */
insert_pages(struct vm_area_struct * vma,unsigned long addr,struct page ** pages,unsigned long * num,pgprot_t prot)2049 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
2050 struct page **pages, unsigned long *num, pgprot_t prot)
2051 {
2052 pmd_t *pmd = NULL;
2053 pte_t *start_pte, *pte;
2054 spinlock_t *pte_lock;
2055 struct mm_struct *const mm = vma->vm_mm;
2056 unsigned long curr_page_idx = 0;
2057 unsigned long remaining_pages_total = *num;
2058 unsigned long pages_to_write_in_pmd;
2059 int ret;
2060 more:
2061 ret = -EFAULT;
2062 pmd = walk_to_pmd(mm, addr);
2063 if (!pmd)
2064 goto out;
2065
2066 pages_to_write_in_pmd = min_t(unsigned long,
2067 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
2068
2069 /* Allocate the PTE if necessary; takes PMD lock once only. */
2070 ret = -ENOMEM;
2071 if (pte_alloc(mm, pmd))
2072 goto out;
2073
2074 while (pages_to_write_in_pmd) {
2075 int pte_idx = 0;
2076 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
2077
2078 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
2079 if (!start_pte) {
2080 ret = -EFAULT;
2081 goto out;
2082 }
2083 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
2084 int err = insert_page_in_batch_locked(vma, pte,
2085 addr, pages[curr_page_idx], prot);
2086 if (unlikely(err)) {
2087 pte_unmap_unlock(start_pte, pte_lock);
2088 ret = err;
2089 remaining_pages_total -= pte_idx;
2090 goto out;
2091 }
2092 addr += PAGE_SIZE;
2093 ++curr_page_idx;
2094 }
2095 pte_unmap_unlock(start_pte, pte_lock);
2096 pages_to_write_in_pmd -= batch_size;
2097 remaining_pages_total -= batch_size;
2098 }
2099 if (remaining_pages_total)
2100 goto more;
2101 ret = 0;
2102 out:
2103 *num = remaining_pages_total;
2104 return ret;
2105 }
2106
2107 /**
2108 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
2109 * @vma: user vma to map to
2110 * @addr: target start user address of these pages
2111 * @pages: source kernel pages
2112 * @num: in: number of pages to map. out: number of pages that were *not*
2113 * mapped. (0 means all pages were successfully mapped).
2114 *
2115 * Preferred over vm_insert_page() when inserting multiple pages.
2116 *
2117 * In case of error, we may have mapped a subset of the provided
2118 * pages. It is the caller's responsibility to account for this case.
2119 *
2120 * The same restrictions apply as in vm_insert_page().
2121 */
vm_insert_pages(struct vm_area_struct * vma,unsigned long addr,struct page ** pages,unsigned long * num)2122 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
2123 struct page **pages, unsigned long *num)
2124 {
2125 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
2126
2127 if (addr < vma->vm_start || end_addr >= vma->vm_end)
2128 return -EFAULT;
2129 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2130 BUG_ON(mmap_read_trylock(vma->vm_mm));
2131 BUG_ON(vma->vm_flags & VM_PFNMAP);
2132 vm_flags_set(vma, VM_MIXEDMAP);
2133 }
2134 /* Defer page refcount checking till we're about to map that page. */
2135 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
2136 }
2137 EXPORT_SYMBOL(vm_insert_pages);
2138
2139 /**
2140 * vm_insert_page - insert single page into user vma
2141 * @vma: user vma to map to
2142 * @addr: target user address of this page
2143 * @page: source kernel page
2144 *
2145 * This allows drivers to insert individual pages they've allocated
2146 * into a user vma.
2147 *
2148 * The page has to be a nice clean _individual_ kernel allocation.
2149 * If you allocate a compound page, you need to have marked it as
2150 * such (__GFP_COMP), or manually just split the page up yourself
2151 * (see split_page()).
2152 *
2153 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2154 * took an arbitrary page protection parameter. This doesn't allow
2155 * that. Your vma protection will have to be set up correctly, which
2156 * means that if you want a shared writable mapping, you'd better
2157 * ask for a shared writable mapping!
2158 *
2159 * The page does not need to be reserved.
2160 *
2161 * Usually this function is called from f_op->mmap() handler
2162 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2163 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2164 * function from other places, for example from page-fault handler.
2165 *
2166 * Return: %0 on success, negative error code otherwise.
2167 */
vm_insert_page(struct vm_area_struct * vma,unsigned long addr,struct page * page)2168 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2169 struct page *page)
2170 {
2171 if (addr < vma->vm_start || addr >= vma->vm_end)
2172 return -EFAULT;
2173 if (!page_count(page))
2174 return -EINVAL;
2175 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2176 BUG_ON(mmap_read_trylock(vma->vm_mm));
2177 BUG_ON(vma->vm_flags & VM_PFNMAP);
2178 vm_flags_set(vma, VM_MIXEDMAP);
2179 }
2180 return insert_page(vma, addr, page, vma->vm_page_prot);
2181 }
2182 EXPORT_SYMBOL(vm_insert_page);
2183
2184 /*
2185 * __vm_map_pages - maps range of kernel pages into user vma
2186 * @vma: user vma to map to
2187 * @pages: pointer to array of source kernel pages
2188 * @num: number of pages in page array
2189 * @offset: user's requested vm_pgoff
2190 *
2191 * This allows drivers to map range of kernel pages into a user vma.
2192 *
2193 * Return: 0 on success and error code otherwise.
2194 */
__vm_map_pages(struct vm_area_struct * vma,struct page ** pages,unsigned long num,unsigned long offset)2195 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2196 unsigned long num, unsigned long offset)
2197 {
2198 unsigned long count = vma_pages(vma);
2199 unsigned long uaddr = vma->vm_start;
2200 int ret, i;
2201
2202 /* Fail if the user requested offset is beyond the end of the object */
2203 if (offset >= num)
2204 return -ENXIO;
2205
2206 /* Fail if the user requested size exceeds available object size */
2207 if (count > num - offset)
2208 return -ENXIO;
2209
2210 for (i = 0; i < count; i++) {
2211 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2212 if (ret < 0)
2213 return ret;
2214 uaddr += PAGE_SIZE;
2215 }
2216
2217 return 0;
2218 }
2219
2220 /**
2221 * vm_map_pages - maps range of kernel pages starts with non zero offset
2222 * @vma: user vma to map to
2223 * @pages: pointer to array of source kernel pages
2224 * @num: number of pages in page array
2225 *
2226 * Maps an object consisting of @num pages, catering for the user's
2227 * requested vm_pgoff
2228 *
2229 * If we fail to insert any page into the vma, the function will return
2230 * immediately leaving any previously inserted pages present. Callers
2231 * from the mmap handler may immediately return the error as their caller
2232 * will destroy the vma, removing any successfully inserted pages. Other
2233 * callers should make their own arrangements for calling unmap_region().
2234 *
2235 * Context: Process context. Called by mmap handlers.
2236 * Return: 0 on success and error code otherwise.
2237 */
vm_map_pages(struct vm_area_struct * vma,struct page ** pages,unsigned long num)2238 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2239 unsigned long num)
2240 {
2241 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2242 }
2243 EXPORT_SYMBOL(vm_map_pages);
2244
2245 /**
2246 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2247 * @vma: user vma to map to
2248 * @pages: pointer to array of source kernel pages
2249 * @num: number of pages in page array
2250 *
2251 * Similar to vm_map_pages(), except that it explicitly sets the offset
2252 * to 0. This function is intended for the drivers that did not consider
2253 * vm_pgoff.
2254 *
2255 * Context: Process context. Called by mmap handlers.
2256 * Return: 0 on success and error code otherwise.
2257 */
vm_map_pages_zero(struct vm_area_struct * vma,struct page ** pages,unsigned long num)2258 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2259 unsigned long num)
2260 {
2261 return __vm_map_pages(vma, pages, num, 0);
2262 }
2263 EXPORT_SYMBOL(vm_map_pages_zero);
2264
insert_pfn(struct vm_area_struct * vma,unsigned long addr,pfn_t pfn,pgprot_t prot,bool mkwrite)2265 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2266 pfn_t pfn, pgprot_t prot, bool mkwrite)
2267 {
2268 struct mm_struct *mm = vma->vm_mm;
2269 pte_t *pte, entry;
2270 spinlock_t *ptl;
2271
2272 pte = get_locked_pte(mm, addr, &ptl);
2273 if (!pte)
2274 return VM_FAULT_OOM;
2275 entry = ptep_get(pte);
2276 if (!pte_none(entry)) {
2277 if (mkwrite) {
2278 /*
2279 * For read faults on private mappings the PFN passed
2280 * in may not match the PFN we have mapped if the
2281 * mapped PFN is a writeable COW page. In the mkwrite
2282 * case we are creating a writable PTE for a shared
2283 * mapping and we expect the PFNs to match. If they
2284 * don't match, we are likely racing with block
2285 * allocation and mapping invalidation so just skip the
2286 * update.
2287 */
2288 if (pte_pfn(entry) != pfn_t_to_pfn(pfn)) {
2289 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
2290 goto out_unlock;
2291 }
2292 entry = pte_mkyoung(entry);
2293 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2294 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2295 update_mmu_cache(vma, addr, pte);
2296 }
2297 goto out_unlock;
2298 }
2299
2300 /* Ok, finally just insert the thing.. */
2301 if (pfn_t_devmap(pfn))
2302 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2303 else
2304 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2305
2306 if (mkwrite) {
2307 entry = pte_mkyoung(entry);
2308 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2309 }
2310
2311 set_pte_at(mm, addr, pte, entry);
2312 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2313
2314 out_unlock:
2315 pte_unmap_unlock(pte, ptl);
2316 return VM_FAULT_NOPAGE;
2317 }
2318
2319 /**
2320 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2321 * @vma: user vma to map to
2322 * @addr: target user address of this page
2323 * @pfn: source kernel pfn
2324 * @pgprot: pgprot flags for the inserted page
2325 *
2326 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2327 * to override pgprot on a per-page basis.
2328 *
2329 * This only makes sense for IO mappings, and it makes no sense for
2330 * COW mappings. In general, using multiple vmas is preferable;
2331 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2332 * impractical.
2333 *
2334 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2335 * caching- and encryption bits different than those of @vma->vm_page_prot,
2336 * because the caching- or encryption mode may not be known at mmap() time.
2337 *
2338 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2339 * to set caching and encryption bits for those vmas (except for COW pages).
2340 * This is ensured by core vm only modifying these page table entries using
2341 * functions that don't touch caching- or encryption bits, using pte_modify()
2342 * if needed. (See for example mprotect()).
2343 *
2344 * Also when new page-table entries are created, this is only done using the
2345 * fault() callback, and never using the value of vma->vm_page_prot,
2346 * except for page-table entries that point to anonymous pages as the result
2347 * of COW.
2348 *
2349 * Context: Process context. May allocate using %GFP_KERNEL.
2350 * Return: vm_fault_t value.
2351 */
vmf_insert_pfn_prot(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn,pgprot_t pgprot)2352 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2353 unsigned long pfn, pgprot_t pgprot)
2354 {
2355 /*
2356 * Technically, architectures with pte_special can avoid all these
2357 * restrictions (same for remap_pfn_range). However we would like
2358 * consistency in testing and feature parity among all, so we should
2359 * try to keep these invariants in place for everybody.
2360 */
2361 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2362 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2363 (VM_PFNMAP|VM_MIXEDMAP));
2364 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2365 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2366
2367 if (addr < vma->vm_start || addr >= vma->vm_end)
2368 return VM_FAULT_SIGBUS;
2369
2370 if (!pfn_modify_allowed(pfn, pgprot))
2371 return VM_FAULT_SIGBUS;
2372
2373 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2374
2375 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2376 false);
2377 }
2378 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2379
2380 /**
2381 * vmf_insert_pfn - insert single pfn into user vma
2382 * @vma: user vma to map to
2383 * @addr: target user address of this page
2384 * @pfn: source kernel pfn
2385 *
2386 * Similar to vm_insert_page, this allows drivers to insert individual pages
2387 * they've allocated into a user vma. Same comments apply.
2388 *
2389 * This function should only be called from a vm_ops->fault handler, and
2390 * in that case the handler should return the result of this function.
2391 *
2392 * vma cannot be a COW mapping.
2393 *
2394 * As this is called only for pages that do not currently exist, we
2395 * do not need to flush old virtual caches or the TLB.
2396 *
2397 * Context: Process context. May allocate using %GFP_KERNEL.
2398 * Return: vm_fault_t value.
2399 */
vmf_insert_pfn(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn)2400 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2401 unsigned long pfn)
2402 {
2403 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2404 }
2405 EXPORT_SYMBOL(vmf_insert_pfn);
2406
vm_mixed_ok(struct vm_area_struct * vma,pfn_t pfn)2407 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2408 {
2409 /* these checks mirror the abort conditions in vm_normal_page */
2410 if (vma->vm_flags & VM_MIXEDMAP)
2411 return true;
2412 if (pfn_t_devmap(pfn))
2413 return true;
2414 if (pfn_t_special(pfn))
2415 return true;
2416 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2417 return true;
2418 return false;
2419 }
2420
__vm_insert_mixed(struct vm_area_struct * vma,unsigned long addr,pfn_t pfn,bool mkwrite)2421 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2422 unsigned long addr, pfn_t pfn, bool mkwrite)
2423 {
2424 pgprot_t pgprot = vma->vm_page_prot;
2425 int err;
2426
2427 BUG_ON(!vm_mixed_ok(vma, pfn));
2428
2429 if (addr < vma->vm_start || addr >= vma->vm_end)
2430 return VM_FAULT_SIGBUS;
2431
2432 track_pfn_insert(vma, &pgprot, pfn);
2433
2434 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2435 return VM_FAULT_SIGBUS;
2436
2437 /*
2438 * If we don't have pte special, then we have to use the pfn_valid()
2439 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2440 * refcount the page if pfn_valid is true (hence insert_page rather
2441 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2442 * without pte special, it would there be refcounted as a normal page.
2443 */
2444 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2445 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2446 struct page *page;
2447
2448 /*
2449 * At this point we are committed to insert_page()
2450 * regardless of whether the caller specified flags that
2451 * result in pfn_t_has_page() == false.
2452 */
2453 page = pfn_to_page(pfn_t_to_pfn(pfn));
2454 err = insert_page(vma, addr, page, pgprot);
2455 } else {
2456 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2457 }
2458
2459 if (err == -ENOMEM)
2460 return VM_FAULT_OOM;
2461 if (err < 0 && err != -EBUSY)
2462 return VM_FAULT_SIGBUS;
2463
2464 return VM_FAULT_NOPAGE;
2465 }
2466
vmf_insert_mixed(struct vm_area_struct * vma,unsigned long addr,pfn_t pfn)2467 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2468 pfn_t pfn)
2469 {
2470 return __vm_insert_mixed(vma, addr, pfn, false);
2471 }
2472 EXPORT_SYMBOL(vmf_insert_mixed);
2473
2474 /*
2475 * If the insertion of PTE failed because someone else already added a
2476 * different entry in the mean time, we treat that as success as we assume
2477 * the same entry was actually inserted.
2478 */
vmf_insert_mixed_mkwrite(struct vm_area_struct * vma,unsigned long addr,pfn_t pfn)2479 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2480 unsigned long addr, pfn_t pfn)
2481 {
2482 return __vm_insert_mixed(vma, addr, pfn, true);
2483 }
2484 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2485
2486 /*
2487 * maps a range of physical memory into the requested pages. the old
2488 * mappings are removed. any references to nonexistent pages results
2489 * in null mappings (currently treated as "copy-on-access")
2490 */
remap_pte_range(struct mm_struct * mm,pmd_t * pmd,unsigned long addr,unsigned long end,unsigned long pfn,pgprot_t prot)2491 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2492 unsigned long addr, unsigned long end,
2493 unsigned long pfn, pgprot_t prot)
2494 {
2495 pte_t *pte, *mapped_pte;
2496 spinlock_t *ptl;
2497 int err = 0;
2498
2499 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2500 if (!pte)
2501 return -ENOMEM;
2502 arch_enter_lazy_mmu_mode();
2503 do {
2504 BUG_ON(!pte_none(ptep_get(pte)));
2505 if (!pfn_modify_allowed(pfn, prot)) {
2506 err = -EACCES;
2507 break;
2508 }
2509 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2510 pfn++;
2511 } while (pte++, addr += PAGE_SIZE, addr != end);
2512 arch_leave_lazy_mmu_mode();
2513 pte_unmap_unlock(mapped_pte, ptl);
2514 return err;
2515 }
2516
remap_pmd_range(struct mm_struct * mm,pud_t * pud,unsigned long addr,unsigned long end,unsigned long pfn,pgprot_t prot)2517 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2518 unsigned long addr, unsigned long end,
2519 unsigned long pfn, pgprot_t prot)
2520 {
2521 pmd_t *pmd;
2522 unsigned long next;
2523 int err;
2524
2525 pfn -= addr >> PAGE_SHIFT;
2526 pmd = pmd_alloc(mm, pud, addr);
2527 if (!pmd)
2528 return -ENOMEM;
2529 VM_BUG_ON(pmd_trans_huge(*pmd));
2530 do {
2531 next = pmd_addr_end(addr, end);
2532 err = remap_pte_range(mm, pmd, addr, next,
2533 pfn + (addr >> PAGE_SHIFT), prot);
2534 if (err)
2535 return err;
2536 } while (pmd++, addr = next, addr != end);
2537 return 0;
2538 }
2539
remap_pud_range(struct mm_struct * mm,p4d_t * p4d,unsigned long addr,unsigned long end,unsigned long pfn,pgprot_t prot)2540 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2541 unsigned long addr, unsigned long end,
2542 unsigned long pfn, pgprot_t prot)
2543 {
2544 pud_t *pud;
2545 unsigned long next;
2546 int err;
2547
2548 pfn -= addr >> PAGE_SHIFT;
2549 pud = pud_alloc(mm, p4d, addr);
2550 if (!pud)
2551 return -ENOMEM;
2552 do {
2553 next = pud_addr_end(addr, end);
2554 err = remap_pmd_range(mm, pud, addr, next,
2555 pfn + (addr >> PAGE_SHIFT), prot);
2556 if (err)
2557 return err;
2558 } while (pud++, addr = next, addr != end);
2559 return 0;
2560 }
2561
remap_p4d_range(struct mm_struct * mm,pgd_t * pgd,unsigned long addr,unsigned long end,unsigned long pfn,pgprot_t prot)2562 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2563 unsigned long addr, unsigned long end,
2564 unsigned long pfn, pgprot_t prot)
2565 {
2566 p4d_t *p4d;
2567 unsigned long next;
2568 int err;
2569
2570 pfn -= addr >> PAGE_SHIFT;
2571 p4d = p4d_alloc(mm, pgd, addr);
2572 if (!p4d)
2573 return -ENOMEM;
2574 do {
2575 next = p4d_addr_end(addr, end);
2576 err = remap_pud_range(mm, p4d, addr, next,
2577 pfn + (addr >> PAGE_SHIFT), prot);
2578 if (err)
2579 return err;
2580 } while (p4d++, addr = next, addr != end);
2581 return 0;
2582 }
2583
2584 /*
2585 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2586 * must have pre-validated the caching bits of the pgprot_t.
2587 */
remap_pfn_range_notrack(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn,unsigned long size,pgprot_t prot)2588 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2589 unsigned long pfn, unsigned long size, pgprot_t prot)
2590 {
2591 pgd_t *pgd;
2592 unsigned long next;
2593 unsigned long end = addr + PAGE_ALIGN(size);
2594 struct mm_struct *mm = vma->vm_mm;
2595 int err;
2596
2597 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2598 return -EINVAL;
2599
2600 /*
2601 * Physically remapped pages are special. Tell the
2602 * rest of the world about it:
2603 * VM_IO tells people not to look at these pages
2604 * (accesses can have side effects).
2605 * VM_PFNMAP tells the core MM that the base pages are just
2606 * raw PFN mappings, and do not have a "struct page" associated
2607 * with them.
2608 * VM_DONTEXPAND
2609 * Disable vma merging and expanding with mremap().
2610 * VM_DONTDUMP
2611 * Omit vma from core dump, even when VM_IO turned off.
2612 *
2613 * There's a horrible special case to handle copy-on-write
2614 * behaviour that some programs depend on. We mark the "original"
2615 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2616 * See vm_normal_page() for details.
2617 */
2618 if (is_cow_mapping(vma->vm_flags)) {
2619 if (addr != vma->vm_start || end != vma->vm_end)
2620 return -EINVAL;
2621 vma->vm_pgoff = pfn;
2622 }
2623
2624 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
2625
2626 BUG_ON(addr >= end);
2627 pfn -= addr >> PAGE_SHIFT;
2628 pgd = pgd_offset(mm, addr);
2629 flush_cache_range(vma, addr, end);
2630 do {
2631 next = pgd_addr_end(addr, end);
2632 err = remap_p4d_range(mm, pgd, addr, next,
2633 pfn + (addr >> PAGE_SHIFT), prot);
2634 if (err)
2635 return err;
2636 } while (pgd++, addr = next, addr != end);
2637
2638 return 0;
2639 }
2640
2641 /**
2642 * remap_pfn_range - remap kernel memory to userspace
2643 * @vma: user vma to map to
2644 * @addr: target page aligned user address to start at
2645 * @pfn: page frame number of kernel physical memory address
2646 * @size: size of mapping area
2647 * @prot: page protection flags for this mapping
2648 *
2649 * Note: this is only safe if the mm semaphore is held when called.
2650 *
2651 * Return: %0 on success, negative error code otherwise.
2652 */
remap_pfn_range(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn,unsigned long size,pgprot_t prot)2653 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2654 unsigned long pfn, unsigned long size, pgprot_t prot)
2655 {
2656 int err;
2657
2658 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2659 if (err)
2660 return -EINVAL;
2661
2662 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2663 if (err)
2664 untrack_pfn(vma, pfn, PAGE_ALIGN(size), true);
2665 return err;
2666 }
2667 EXPORT_SYMBOL(remap_pfn_range);
2668
2669 /**
2670 * vm_iomap_memory - remap memory to userspace
2671 * @vma: user vma to map to
2672 * @start: start of the physical memory to be mapped
2673 * @len: size of area
2674 *
2675 * This is a simplified io_remap_pfn_range() for common driver use. The
2676 * driver just needs to give us the physical memory range to be mapped,
2677 * we'll figure out the rest from the vma information.
2678 *
2679 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2680 * whatever write-combining details or similar.
2681 *
2682 * Return: %0 on success, negative error code otherwise.
2683 */
vm_iomap_memory(struct vm_area_struct * vma,phys_addr_t start,unsigned long len)2684 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2685 {
2686 unsigned long vm_len, pfn, pages;
2687
2688 /* Check that the physical memory area passed in looks valid */
2689 if (start + len < start)
2690 return -EINVAL;
2691 /*
2692 * You *really* shouldn't map things that aren't page-aligned,
2693 * but we've historically allowed it because IO memory might
2694 * just have smaller alignment.
2695 */
2696 len += start & ~PAGE_MASK;
2697 pfn = start >> PAGE_SHIFT;
2698 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2699 if (pfn + pages < pfn)
2700 return -EINVAL;
2701
2702 /* We start the mapping 'vm_pgoff' pages into the area */
2703 if (vma->vm_pgoff > pages)
2704 return -EINVAL;
2705 pfn += vma->vm_pgoff;
2706 pages -= vma->vm_pgoff;
2707
2708 /* Can we fit all of the mapping? */
2709 vm_len = vma->vm_end - vma->vm_start;
2710 if (vm_len >> PAGE_SHIFT > pages)
2711 return -EINVAL;
2712
2713 /* Ok, let it rip */
2714 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2715 }
2716 EXPORT_SYMBOL(vm_iomap_memory);
2717
apply_to_pte_range(struct mm_struct * mm,pmd_t * pmd,unsigned long addr,unsigned long end,pte_fn_t fn,void * data,bool create,pgtbl_mod_mask * mask)2718 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2719 unsigned long addr, unsigned long end,
2720 pte_fn_t fn, void *data, bool create,
2721 pgtbl_mod_mask *mask)
2722 {
2723 pte_t *pte, *mapped_pte;
2724 int err = 0;
2725 spinlock_t *ptl;
2726
2727 if (create) {
2728 mapped_pte = pte = (mm == &init_mm) ?
2729 pte_alloc_kernel_track(pmd, addr, mask) :
2730 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2731 if (!pte)
2732 return -ENOMEM;
2733 } else {
2734 mapped_pte = pte = (mm == &init_mm) ?
2735 pte_offset_kernel(pmd, addr) :
2736 pte_offset_map_lock(mm, pmd, addr, &ptl);
2737 if (!pte)
2738 return -EINVAL;
2739 }
2740
2741 arch_enter_lazy_mmu_mode();
2742
2743 if (fn) {
2744 do {
2745 if (create || !pte_none(ptep_get(pte))) {
2746 err = fn(pte++, addr, data);
2747 if (err)
2748 break;
2749 }
2750 } while (addr += PAGE_SIZE, addr != end);
2751 }
2752 *mask |= PGTBL_PTE_MODIFIED;
2753
2754 arch_leave_lazy_mmu_mode();
2755
2756 if (mm != &init_mm)
2757 pte_unmap_unlock(mapped_pte, ptl);
2758 return err;
2759 }
2760
apply_to_pmd_range(struct mm_struct * mm,pud_t * pud,unsigned long addr,unsigned long end,pte_fn_t fn,void * data,bool create,pgtbl_mod_mask * mask)2761 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2762 unsigned long addr, unsigned long end,
2763 pte_fn_t fn, void *data, bool create,
2764 pgtbl_mod_mask *mask)
2765 {
2766 pmd_t *pmd;
2767 unsigned long next;
2768 int err = 0;
2769
2770 BUG_ON(pud_leaf(*pud));
2771
2772 if (create) {
2773 pmd = pmd_alloc_track(mm, pud, addr, mask);
2774 if (!pmd)
2775 return -ENOMEM;
2776 } else {
2777 pmd = pmd_offset(pud, addr);
2778 }
2779 do {
2780 next = pmd_addr_end(addr, end);
2781 if (pmd_none(*pmd) && !create)
2782 continue;
2783 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2784 return -EINVAL;
2785 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2786 if (!create)
2787 continue;
2788 pmd_clear_bad(pmd);
2789 }
2790 err = apply_to_pte_range(mm, pmd, addr, next,
2791 fn, data, create, mask);
2792 if (err)
2793 break;
2794 } while (pmd++, addr = next, addr != end);
2795
2796 return err;
2797 }
2798
apply_to_pud_range(struct mm_struct * mm,p4d_t * p4d,unsigned long addr,unsigned long end,pte_fn_t fn,void * data,bool create,pgtbl_mod_mask * mask)2799 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2800 unsigned long addr, unsigned long end,
2801 pte_fn_t fn, void *data, bool create,
2802 pgtbl_mod_mask *mask)
2803 {
2804 pud_t *pud;
2805 unsigned long next;
2806 int err = 0;
2807
2808 if (create) {
2809 pud = pud_alloc_track(mm, p4d, addr, mask);
2810 if (!pud)
2811 return -ENOMEM;
2812 } else {
2813 pud = pud_offset(p4d, addr);
2814 }
2815 do {
2816 next = pud_addr_end(addr, end);
2817 if (pud_none(*pud) && !create)
2818 continue;
2819 if (WARN_ON_ONCE(pud_leaf(*pud)))
2820 return -EINVAL;
2821 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2822 if (!create)
2823 continue;
2824 pud_clear_bad(pud);
2825 }
2826 err = apply_to_pmd_range(mm, pud, addr, next,
2827 fn, data, create, mask);
2828 if (err)
2829 break;
2830 } while (pud++, addr = next, addr != end);
2831
2832 return err;
2833 }
2834
apply_to_p4d_range(struct mm_struct * mm,pgd_t * pgd,unsigned long addr,unsigned long end,pte_fn_t fn,void * data,bool create,pgtbl_mod_mask * mask)2835 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2836 unsigned long addr, unsigned long end,
2837 pte_fn_t fn, void *data, bool create,
2838 pgtbl_mod_mask *mask)
2839 {
2840 p4d_t *p4d;
2841 unsigned long next;
2842 int err = 0;
2843
2844 if (create) {
2845 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2846 if (!p4d)
2847 return -ENOMEM;
2848 } else {
2849 p4d = p4d_offset(pgd, addr);
2850 }
2851 do {
2852 next = p4d_addr_end(addr, end);
2853 if (p4d_none(*p4d) && !create)
2854 continue;
2855 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2856 return -EINVAL;
2857 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2858 if (!create)
2859 continue;
2860 p4d_clear_bad(p4d);
2861 }
2862 err = apply_to_pud_range(mm, p4d, addr, next,
2863 fn, data, create, mask);
2864 if (err)
2865 break;
2866 } while (p4d++, addr = next, addr != end);
2867
2868 return err;
2869 }
2870
__apply_to_page_range(struct mm_struct * mm,unsigned long addr,unsigned long size,pte_fn_t fn,void * data,bool create)2871 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2872 unsigned long size, pte_fn_t fn,
2873 void *data, bool create)
2874 {
2875 pgd_t *pgd;
2876 unsigned long start = addr, next;
2877 unsigned long end = addr + size;
2878 pgtbl_mod_mask mask = 0;
2879 int err = 0;
2880
2881 if (WARN_ON(addr >= end))
2882 return -EINVAL;
2883
2884 pgd = pgd_offset(mm, addr);
2885 do {
2886 next = pgd_addr_end(addr, end);
2887 if (pgd_none(*pgd) && !create)
2888 continue;
2889 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2890 return -EINVAL;
2891 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2892 if (!create)
2893 continue;
2894 pgd_clear_bad(pgd);
2895 }
2896 err = apply_to_p4d_range(mm, pgd, addr, next,
2897 fn, data, create, &mask);
2898 if (err)
2899 break;
2900 } while (pgd++, addr = next, addr != end);
2901
2902 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2903 arch_sync_kernel_mappings(start, start + size);
2904
2905 return err;
2906 }
2907
2908 /*
2909 * Scan a region of virtual memory, filling in page tables as necessary
2910 * and calling a provided function on each leaf page table.
2911 */
apply_to_page_range(struct mm_struct * mm,unsigned long addr,unsigned long size,pte_fn_t fn,void * data)2912 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2913 unsigned long size, pte_fn_t fn, void *data)
2914 {
2915 return __apply_to_page_range(mm, addr, size, fn, data, true);
2916 }
2917 EXPORT_SYMBOL_GPL(apply_to_page_range);
2918
2919 /*
2920 * Scan a region of virtual memory, calling a provided function on
2921 * each leaf page table where it exists.
2922 *
2923 * Unlike apply_to_page_range, this does _not_ fill in page tables
2924 * where they are absent.
2925 */
apply_to_existing_page_range(struct mm_struct * mm,unsigned long addr,unsigned long size,pte_fn_t fn,void * data)2926 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2927 unsigned long size, pte_fn_t fn, void *data)
2928 {
2929 return __apply_to_page_range(mm, addr, size, fn, data, false);
2930 }
2931 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2932
2933 /*
2934 * handle_pte_fault chooses page fault handler according to an entry which was
2935 * read non-atomically. Before making any commitment, on those architectures
2936 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2937 * parts, do_swap_page must check under lock before unmapping the pte and
2938 * proceeding (but do_wp_page is only called after already making such a check;
2939 * and do_anonymous_page can safely check later on).
2940 */
pte_unmap_same(struct vm_fault * vmf)2941 static inline int pte_unmap_same(struct vm_fault *vmf)
2942 {
2943 int same = 1;
2944 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2945 if (sizeof(pte_t) > sizeof(unsigned long)) {
2946 spin_lock(vmf->ptl);
2947 same = pte_same(ptep_get(vmf->pte), vmf->orig_pte);
2948 spin_unlock(vmf->ptl);
2949 }
2950 #endif
2951 pte_unmap(vmf->pte);
2952 vmf->pte = NULL;
2953 return same;
2954 }
2955
2956 /*
2957 * Return:
2958 * 0: copied succeeded
2959 * -EHWPOISON: copy failed due to hwpoison in source page
2960 * -EAGAIN: copied failed (some other reason)
2961 */
__wp_page_copy_user(struct page * dst,struct page * src,struct vm_fault * vmf)2962 static inline int __wp_page_copy_user(struct page *dst, struct page *src,
2963 struct vm_fault *vmf)
2964 {
2965 int ret;
2966 void *kaddr;
2967 void __user *uaddr;
2968 struct vm_area_struct *vma = vmf->vma;
2969 struct mm_struct *mm = vma->vm_mm;
2970 unsigned long addr = vmf->address;
2971
2972 if (likely(src)) {
2973 if (copy_mc_user_highpage(dst, src, addr, vma)) {
2974 memory_failure_queue(page_to_pfn(src), 0);
2975 return -EHWPOISON;
2976 }
2977 return 0;
2978 }
2979
2980 /*
2981 * If the source page was a PFN mapping, we don't have
2982 * a "struct page" for it. We do a best-effort copy by
2983 * just copying from the original user address. If that
2984 * fails, we just zero-fill it. Live with it.
2985 */
2986 kaddr = kmap_local_page(dst);
2987 pagefault_disable();
2988 uaddr = (void __user *)(addr & PAGE_MASK);
2989
2990 /*
2991 * On architectures with software "accessed" bits, we would
2992 * take a double page fault, so mark it accessed here.
2993 */
2994 vmf->pte = NULL;
2995 if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) {
2996 pte_t entry;
2997
2998 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2999 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3000 /*
3001 * Other thread has already handled the fault
3002 * and update local tlb only
3003 */
3004 if (vmf->pte)
3005 update_mmu_tlb(vma, addr, vmf->pte);
3006 ret = -EAGAIN;
3007 goto pte_unlock;
3008 }
3009
3010 entry = pte_mkyoung(vmf->orig_pte);
3011 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
3012 update_mmu_cache_range(vmf, vma, addr, vmf->pte, 1);
3013 }
3014
3015 /*
3016 * This really shouldn't fail, because the page is there
3017 * in the page tables. But it might just be unreadable,
3018 * in which case we just give up and fill the result with
3019 * zeroes.
3020 */
3021 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
3022 if (vmf->pte)
3023 goto warn;
3024
3025 /* Re-validate under PTL if the page is still mapped */
3026 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
3027 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3028 /* The PTE changed under us, update local tlb */
3029 if (vmf->pte)
3030 update_mmu_tlb(vma, addr, vmf->pte);
3031 ret = -EAGAIN;
3032 goto pte_unlock;
3033 }
3034
3035 /*
3036 * The same page can be mapped back since last copy attempt.
3037 * Try to copy again under PTL.
3038 */
3039 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
3040 /*
3041 * Give a warn in case there can be some obscure
3042 * use-case
3043 */
3044 warn:
3045 WARN_ON_ONCE(1);
3046 clear_page(kaddr);
3047 }
3048 }
3049
3050 ret = 0;
3051
3052 pte_unlock:
3053 if (vmf->pte)
3054 pte_unmap_unlock(vmf->pte, vmf->ptl);
3055 pagefault_enable();
3056 kunmap_local(kaddr);
3057 flush_dcache_page(dst);
3058
3059 return ret;
3060 }
3061
__get_fault_gfp_mask(struct vm_area_struct * vma)3062 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
3063 {
3064 struct file *vm_file = vma->vm_file;
3065
3066 if (vm_file)
3067 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
3068
3069 /*
3070 * Special mappings (e.g. VDSO) do not have any file so fake
3071 * a default GFP_KERNEL for them.
3072 */
3073 return GFP_KERNEL;
3074 }
3075
3076 /*
3077 * Notify the address space that the page is about to become writable so that
3078 * it can prohibit this or wait for the page to get into an appropriate state.
3079 *
3080 * We do this without the lock held, so that it can sleep if it needs to.
3081 */
do_page_mkwrite(struct vm_fault * vmf,struct folio * folio)3082 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio)
3083 {
3084 vm_fault_t ret;
3085 unsigned int old_flags = vmf->flags;
3086
3087 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3088
3089 if (vmf->vma->vm_file &&
3090 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
3091 return VM_FAULT_SIGBUS;
3092
3093 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
3094 /* Restore original flags so that caller is not surprised */
3095 vmf->flags = old_flags;
3096 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
3097 return ret;
3098 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
3099 folio_lock(folio);
3100 if (!folio->mapping) {
3101 folio_unlock(folio);
3102 return 0; /* retry */
3103 }
3104 ret |= VM_FAULT_LOCKED;
3105 } else
3106 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3107 return ret;
3108 }
3109
3110 /*
3111 * Handle dirtying of a page in shared file mapping on a write fault.
3112 *
3113 * The function expects the page to be locked and unlocks it.
3114 */
fault_dirty_shared_page(struct vm_fault * vmf)3115 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
3116 {
3117 struct vm_area_struct *vma = vmf->vma;
3118 struct address_space *mapping;
3119 struct folio *folio = page_folio(vmf->page);
3120 bool dirtied;
3121 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
3122
3123 dirtied = folio_mark_dirty(folio);
3124 VM_BUG_ON_FOLIO(folio_test_anon(folio), folio);
3125 /*
3126 * Take a local copy of the address_space - folio.mapping may be zeroed
3127 * by truncate after folio_unlock(). The address_space itself remains
3128 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
3129 * release semantics to prevent the compiler from undoing this copying.
3130 */
3131 mapping = folio_raw_mapping(folio);
3132 folio_unlock(folio);
3133
3134 if (!page_mkwrite)
3135 file_update_time(vma->vm_file);
3136
3137 /*
3138 * Throttle page dirtying rate down to writeback speed.
3139 *
3140 * mapping may be NULL here because some device drivers do not
3141 * set page.mapping but still dirty their pages
3142 *
3143 * Drop the mmap_lock before waiting on IO, if we can. The file
3144 * is pinning the mapping, as per above.
3145 */
3146 if ((dirtied || page_mkwrite) && mapping) {
3147 struct file *fpin;
3148
3149 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
3150 balance_dirty_pages_ratelimited(mapping);
3151 if (fpin) {
3152 fput(fpin);
3153 return VM_FAULT_COMPLETED;
3154 }
3155 }
3156
3157 return 0;
3158 }
3159
3160 /*
3161 * Handle write page faults for pages that can be reused in the current vma
3162 *
3163 * This can happen either due to the mapping being with the VM_SHARED flag,
3164 * or due to us being the last reference standing to the page. In either
3165 * case, all we need to do here is to mark the page as writable and update
3166 * any related book-keeping.
3167 */
wp_page_reuse(struct vm_fault * vmf,struct folio * folio)3168 static inline void wp_page_reuse(struct vm_fault *vmf, struct folio *folio)
3169 __releases(vmf->ptl)
3170 {
3171 struct vm_area_struct *vma = vmf->vma;
3172 pte_t entry;
3173
3174 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3175
3176 if (folio) {
3177 VM_BUG_ON(folio_test_anon(folio) &&
3178 !PageAnonExclusive(vmf->page));
3179 /*
3180 * Clear the folio's cpupid information as the existing
3181 * information potentially belongs to a now completely
3182 * unrelated process.
3183 */
3184 folio_xchg_last_cpupid(folio, (1 << LAST_CPUPID_SHIFT) - 1);
3185 }
3186
3187 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3188 entry = pte_mkyoung(vmf->orig_pte);
3189 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3190 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
3191 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3192 pte_unmap_unlock(vmf->pte, vmf->ptl);
3193 count_vm_event(PGREUSE);
3194 }
3195
3196 /*
3197 * We could add a bitflag somewhere, but for now, we know that all
3198 * vm_ops that have a ->map_pages have been audited and don't need
3199 * the mmap_lock to be held.
3200 */
vmf_can_call_fault(const struct vm_fault * vmf)3201 static inline vm_fault_t vmf_can_call_fault(const struct vm_fault *vmf)
3202 {
3203 struct vm_area_struct *vma = vmf->vma;
3204
3205 if (vma->vm_ops->map_pages || !(vmf->flags & FAULT_FLAG_VMA_LOCK))
3206 return 0;
3207 vma_end_read(vma);
3208 return VM_FAULT_RETRY;
3209 }
3210
3211 /**
3212 * vmf_anon_prepare - Prepare to handle an anonymous fault.
3213 * @vmf: The vm_fault descriptor passed from the fault handler.
3214 *
3215 * When preparing to insert an anonymous page into a VMA from a
3216 * fault handler, call this function rather than anon_vma_prepare().
3217 * If this vma does not already have an associated anon_vma and we are
3218 * only protected by the per-VMA lock, the caller must retry with the
3219 * mmap_lock held. __anon_vma_prepare() will look at adjacent VMAs to
3220 * determine if this VMA can share its anon_vma, and that's not safe to
3221 * do with only the per-VMA lock held for this VMA.
3222 *
3223 * Return: 0 if fault handling can proceed. Any other value should be
3224 * returned to the caller.
3225 */
vmf_anon_prepare(struct vm_fault * vmf)3226 vm_fault_t vmf_anon_prepare(struct vm_fault *vmf)
3227 {
3228 struct vm_area_struct *vma = vmf->vma;
3229 vm_fault_t ret = 0;
3230
3231 if (likely(vma->anon_vma))
3232 return 0;
3233 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3234 if (!mmap_read_trylock(vma->vm_mm)) {
3235 vma_end_read(vma);
3236 return VM_FAULT_RETRY;
3237 }
3238 }
3239 if (__anon_vma_prepare(vma))
3240 ret = VM_FAULT_OOM;
3241 if (vmf->flags & FAULT_FLAG_VMA_LOCK)
3242 mmap_read_unlock(vma->vm_mm);
3243 return ret;
3244 }
3245
3246 /*
3247 * Handle the case of a page which we actually need to copy to a new page,
3248 * either due to COW or unsharing.
3249 *
3250 * Called with mmap_lock locked and the old page referenced, but
3251 * without the ptl held.
3252 *
3253 * High level logic flow:
3254 *
3255 * - Allocate a page, copy the content of the old page to the new one.
3256 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3257 * - Take the PTL. If the pte changed, bail out and release the allocated page
3258 * - If the pte is still the way we remember it, update the page table and all
3259 * relevant references. This includes dropping the reference the page-table
3260 * held to the old page, as well as updating the rmap.
3261 * - In any case, unlock the PTL and drop the reference we took to the old page.
3262 */
wp_page_copy(struct vm_fault * vmf)3263 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3264 {
3265 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3266 struct vm_area_struct *vma = vmf->vma;
3267 struct mm_struct *mm = vma->vm_mm;
3268 struct folio *old_folio = NULL;
3269 struct folio *new_folio = NULL;
3270 pte_t entry;
3271 int page_copied = 0;
3272 struct mmu_notifier_range range;
3273 vm_fault_t ret;
3274 bool pfn_is_zero;
3275
3276 delayacct_wpcopy_start();
3277
3278 if (vmf->page)
3279 old_folio = page_folio(vmf->page);
3280 ret = vmf_anon_prepare(vmf);
3281 if (unlikely(ret))
3282 goto out;
3283
3284 pfn_is_zero = is_zero_pfn(pte_pfn(vmf->orig_pte));
3285 new_folio = folio_prealloc(mm, vma, vmf->address, pfn_is_zero);
3286 if (!new_folio)
3287 goto oom;
3288
3289 if (!pfn_is_zero) {
3290 int err;
3291
3292 err = __wp_page_copy_user(&new_folio->page, vmf->page, vmf);
3293 if (err) {
3294 /*
3295 * COW failed, if the fault was solved by other,
3296 * it's fine. If not, userspace would re-fault on
3297 * the same address and we will handle the fault
3298 * from the second attempt.
3299 * The -EHWPOISON case will not be retried.
3300 */
3301 folio_put(new_folio);
3302 if (old_folio)
3303 folio_put(old_folio);
3304
3305 delayacct_wpcopy_end();
3306 return err == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3307 }
3308 kmsan_copy_page_meta(&new_folio->page, vmf->page);
3309 }
3310
3311 __folio_mark_uptodate(new_folio);
3312
3313 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
3314 vmf->address & PAGE_MASK,
3315 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3316 mmu_notifier_invalidate_range_start(&range);
3317
3318 /*
3319 * Re-check the pte - we dropped the lock
3320 */
3321 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3322 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3323 if (old_folio) {
3324 if (!folio_test_anon(old_folio)) {
3325 dec_mm_counter(mm, mm_counter_file(old_folio));
3326 inc_mm_counter(mm, MM_ANONPAGES);
3327 }
3328 } else {
3329 ksm_might_unmap_zero_page(mm, vmf->orig_pte);
3330 inc_mm_counter(mm, MM_ANONPAGES);
3331 }
3332 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3333 entry = mk_pte(&new_folio->page, vma->vm_page_prot);
3334 entry = pte_sw_mkyoung(entry);
3335 if (unlikely(unshare)) {
3336 if (pte_soft_dirty(vmf->orig_pte))
3337 entry = pte_mksoft_dirty(entry);
3338 if (pte_uffd_wp(vmf->orig_pte))
3339 entry = pte_mkuffd_wp(entry);
3340 } else {
3341 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3342 }
3343
3344 /*
3345 * Clear the pte entry and flush it first, before updating the
3346 * pte with the new entry, to keep TLBs on different CPUs in
3347 * sync. This code used to set the new PTE then flush TLBs, but
3348 * that left a window where the new PTE could be loaded into
3349 * some TLBs while the old PTE remains in others.
3350 */
3351 ptep_clear_flush(vma, vmf->address, vmf->pte);
3352 folio_add_new_anon_rmap(new_folio, vma, vmf->address);
3353 folio_add_lru_vma(new_folio, vma);
3354 BUG_ON(unshare && pte_write(entry));
3355 set_pte_at(mm, vmf->address, vmf->pte, entry);
3356 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3357 if (old_folio) {
3358 /*
3359 * Only after switching the pte to the new page may
3360 * we remove the mapcount here. Otherwise another
3361 * process may come and find the rmap count decremented
3362 * before the pte is switched to the new page, and
3363 * "reuse" the old page writing into it while our pte
3364 * here still points into it and can be read by other
3365 * threads.
3366 *
3367 * The critical issue is to order this
3368 * folio_remove_rmap_pte() with the ptp_clear_flush
3369 * above. Those stores are ordered by (if nothing else,)
3370 * the barrier present in the atomic_add_negative
3371 * in folio_remove_rmap_pte();
3372 *
3373 * Then the TLB flush in ptep_clear_flush ensures that
3374 * no process can access the old page before the
3375 * decremented mapcount is visible. And the old page
3376 * cannot be reused until after the decremented
3377 * mapcount is visible. So transitively, TLBs to
3378 * old page will be flushed before it can be reused.
3379 */
3380 folio_remove_rmap_pte(old_folio, vmf->page, vma);
3381 }
3382
3383 /* Free the old page.. */
3384 new_folio = old_folio;
3385 page_copied = 1;
3386 pte_unmap_unlock(vmf->pte, vmf->ptl);
3387 } else if (vmf->pte) {
3388 update_mmu_tlb(vma, vmf->address, vmf->pte);
3389 pte_unmap_unlock(vmf->pte, vmf->ptl);
3390 }
3391
3392 mmu_notifier_invalidate_range_end(&range);
3393
3394 if (new_folio)
3395 folio_put(new_folio);
3396 if (old_folio) {
3397 if (page_copied)
3398 free_swap_cache(old_folio);
3399 folio_put(old_folio);
3400 }
3401
3402 delayacct_wpcopy_end();
3403 return 0;
3404 oom:
3405 ret = VM_FAULT_OOM;
3406 out:
3407 if (old_folio)
3408 folio_put(old_folio);
3409
3410 delayacct_wpcopy_end();
3411 return ret;
3412 }
3413
3414 /**
3415 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3416 * writeable once the page is prepared
3417 *
3418 * @vmf: structure describing the fault
3419 * @folio: the folio of vmf->page
3420 *
3421 * This function handles all that is needed to finish a write page fault in a
3422 * shared mapping due to PTE being read-only once the mapped page is prepared.
3423 * It handles locking of PTE and modifying it.
3424 *
3425 * The function expects the page to be locked or other protection against
3426 * concurrent faults / writeback (such as DAX radix tree locks).
3427 *
3428 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3429 * we acquired PTE lock.
3430 */
finish_mkwrite_fault(struct vm_fault * vmf,struct folio * folio)3431 static vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf, struct folio *folio)
3432 {
3433 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3434 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3435 &vmf->ptl);
3436 if (!vmf->pte)
3437 return VM_FAULT_NOPAGE;
3438 /*
3439 * We might have raced with another page fault while we released the
3440 * pte_offset_map_lock.
3441 */
3442 if (!pte_same(ptep_get(vmf->pte), vmf->orig_pte)) {
3443 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3444 pte_unmap_unlock(vmf->pte, vmf->ptl);
3445 return VM_FAULT_NOPAGE;
3446 }
3447 wp_page_reuse(vmf, folio);
3448 return 0;
3449 }
3450
3451 /*
3452 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3453 * mapping
3454 */
wp_pfn_shared(struct vm_fault * vmf)3455 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3456 {
3457 struct vm_area_struct *vma = vmf->vma;
3458
3459 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3460 vm_fault_t ret;
3461
3462 pte_unmap_unlock(vmf->pte, vmf->ptl);
3463 ret = vmf_can_call_fault(vmf);
3464 if (ret)
3465 return ret;
3466
3467 vmf->flags |= FAULT_FLAG_MKWRITE;
3468 ret = vma->vm_ops->pfn_mkwrite(vmf);
3469 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3470 return ret;
3471 return finish_mkwrite_fault(vmf, NULL);
3472 }
3473 wp_page_reuse(vmf, NULL);
3474 return 0;
3475 }
3476
wp_page_shared(struct vm_fault * vmf,struct folio * folio)3477 static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio)
3478 __releases(vmf->ptl)
3479 {
3480 struct vm_area_struct *vma = vmf->vma;
3481 vm_fault_t ret = 0;
3482
3483 folio_get(folio);
3484
3485 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3486 vm_fault_t tmp;
3487
3488 pte_unmap_unlock(vmf->pte, vmf->ptl);
3489 tmp = vmf_can_call_fault(vmf);
3490 if (tmp) {
3491 folio_put(folio);
3492 return tmp;
3493 }
3494
3495 tmp = do_page_mkwrite(vmf, folio);
3496 if (unlikely(!tmp || (tmp &
3497 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3498 folio_put(folio);
3499 return tmp;
3500 }
3501 tmp = finish_mkwrite_fault(vmf, folio);
3502 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3503 folio_unlock(folio);
3504 folio_put(folio);
3505 return tmp;
3506 }
3507 } else {
3508 wp_page_reuse(vmf, folio);
3509 folio_lock(folio);
3510 }
3511 ret |= fault_dirty_shared_page(vmf);
3512 folio_put(folio);
3513
3514 return ret;
3515 }
3516
wp_can_reuse_anon_folio(struct folio * folio,struct vm_area_struct * vma)3517 static bool wp_can_reuse_anon_folio(struct folio *folio,
3518 struct vm_area_struct *vma)
3519 {
3520 /*
3521 * We could currently only reuse a subpage of a large folio if no
3522 * other subpages of the large folios are still mapped. However,
3523 * let's just consistently not reuse subpages even if we could
3524 * reuse in that scenario, and give back a large folio a bit
3525 * sooner.
3526 */
3527 if (folio_test_large(folio))
3528 return false;
3529
3530 /*
3531 * We have to verify under folio lock: these early checks are
3532 * just an optimization to avoid locking the folio and freeing
3533 * the swapcache if there is little hope that we can reuse.
3534 *
3535 * KSM doesn't necessarily raise the folio refcount.
3536 */
3537 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3538 return false;
3539 if (!folio_test_lru(folio))
3540 /*
3541 * We cannot easily detect+handle references from
3542 * remote LRU caches or references to LRU folios.
3543 */
3544 lru_add_drain();
3545 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
3546 return false;
3547 if (!folio_trylock(folio))
3548 return false;
3549 if (folio_test_swapcache(folio))
3550 folio_free_swap(folio);
3551 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
3552 folio_unlock(folio);
3553 return false;
3554 }
3555 /*
3556 * Ok, we've got the only folio reference from our mapping
3557 * and the folio is locked, it's dark out, and we're wearing
3558 * sunglasses. Hit it.
3559 */
3560 folio_move_anon_rmap(folio, vma);
3561 folio_unlock(folio);
3562 return true;
3563 }
3564
3565 /*
3566 * This routine handles present pages, when
3567 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3568 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3569 * (FAULT_FLAG_UNSHARE)
3570 *
3571 * It is done by copying the page to a new address and decrementing the
3572 * shared-page counter for the old page.
3573 *
3574 * Note that this routine assumes that the protection checks have been
3575 * done by the caller (the low-level page fault routine in most cases).
3576 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3577 * done any necessary COW.
3578 *
3579 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3580 * though the page will change only once the write actually happens. This
3581 * avoids a few races, and potentially makes it more efficient.
3582 *
3583 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3584 * but allow concurrent faults), with pte both mapped and locked.
3585 * We return with mmap_lock still held, but pte unmapped and unlocked.
3586 */
do_wp_page(struct vm_fault * vmf)3587 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3588 __releases(vmf->ptl)
3589 {
3590 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3591 struct vm_area_struct *vma = vmf->vma;
3592 struct folio *folio = NULL;
3593 pte_t pte;
3594
3595 if (likely(!unshare)) {
3596 if (userfaultfd_pte_wp(vma, ptep_get(vmf->pte))) {
3597 if (!userfaultfd_wp_async(vma)) {
3598 pte_unmap_unlock(vmf->pte, vmf->ptl);
3599 return handle_userfault(vmf, VM_UFFD_WP);
3600 }
3601
3602 /*
3603 * Nothing needed (cache flush, TLB invalidations,
3604 * etc.) because we're only removing the uffd-wp bit,
3605 * which is completely invisible to the user.
3606 */
3607 pte = pte_clear_uffd_wp(ptep_get(vmf->pte));
3608
3609 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3610 /*
3611 * Update this to be prepared for following up CoW
3612 * handling
3613 */
3614 vmf->orig_pte = pte;
3615 }
3616
3617 /*
3618 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3619 * is flushed in this case before copying.
3620 */
3621 if (unlikely(userfaultfd_wp(vmf->vma) &&
3622 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3623 flush_tlb_page(vmf->vma, vmf->address);
3624 }
3625
3626 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3627
3628 if (vmf->page)
3629 folio = page_folio(vmf->page);
3630
3631 /*
3632 * Shared mapping: we are guaranteed to have VM_WRITE and
3633 * FAULT_FLAG_WRITE set at this point.
3634 */
3635 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
3636 /*
3637 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3638 * VM_PFNMAP VMA.
3639 *
3640 * We should not cow pages in a shared writeable mapping.
3641 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3642 */
3643 if (!vmf->page)
3644 return wp_pfn_shared(vmf);
3645 return wp_page_shared(vmf, folio);
3646 }
3647
3648 /*
3649 * Private mapping: create an exclusive anonymous page copy if reuse
3650 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3651 *
3652 * If we encounter a page that is marked exclusive, we must reuse
3653 * the page without further checks.
3654 */
3655 if (folio && folio_test_anon(folio) &&
3656 (PageAnonExclusive(vmf->page) || wp_can_reuse_anon_folio(folio, vma))) {
3657 if (!PageAnonExclusive(vmf->page))
3658 SetPageAnonExclusive(vmf->page);
3659 if (unlikely(unshare)) {
3660 pte_unmap_unlock(vmf->pte, vmf->ptl);
3661 return 0;
3662 }
3663 wp_page_reuse(vmf, folio);
3664 return 0;
3665 }
3666 /*
3667 * Ok, we need to copy. Oh, well..
3668 */
3669 if (folio)
3670 folio_get(folio);
3671
3672 pte_unmap_unlock(vmf->pte, vmf->ptl);
3673 #ifdef CONFIG_KSM
3674 if (folio && folio_test_ksm(folio))
3675 count_vm_event(COW_KSM);
3676 #endif
3677 return wp_page_copy(vmf);
3678 }
3679
unmap_mapping_range_vma(struct vm_area_struct * vma,unsigned long start_addr,unsigned long end_addr,struct zap_details * details)3680 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3681 unsigned long start_addr, unsigned long end_addr,
3682 struct zap_details *details)
3683 {
3684 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3685 }
3686
unmap_mapping_range_tree(struct rb_root_cached * root,pgoff_t first_index,pgoff_t last_index,struct zap_details * details)3687 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3688 pgoff_t first_index,
3689 pgoff_t last_index,
3690 struct zap_details *details)
3691 {
3692 struct vm_area_struct *vma;
3693 pgoff_t vba, vea, zba, zea;
3694
3695 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3696 vba = vma->vm_pgoff;
3697 vea = vba + vma_pages(vma) - 1;
3698 zba = max(first_index, vba);
3699 zea = min(last_index, vea);
3700
3701 unmap_mapping_range_vma(vma,
3702 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3703 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3704 details);
3705 }
3706 }
3707
3708 /**
3709 * unmap_mapping_folio() - Unmap single folio from processes.
3710 * @folio: The locked folio to be unmapped.
3711 *
3712 * Unmap this folio from any userspace process which still has it mmaped.
3713 * Typically, for efficiency, the range of nearby pages has already been
3714 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3715 * truncation or invalidation holds the lock on a folio, it may find that
3716 * the page has been remapped again: and then uses unmap_mapping_folio()
3717 * to unmap it finally.
3718 */
unmap_mapping_folio(struct folio * folio)3719 void unmap_mapping_folio(struct folio *folio)
3720 {
3721 struct address_space *mapping = folio->mapping;
3722 struct zap_details details = { };
3723 pgoff_t first_index;
3724 pgoff_t last_index;
3725
3726 VM_BUG_ON(!folio_test_locked(folio));
3727
3728 first_index = folio->index;
3729 last_index = folio_next_index(folio) - 1;
3730
3731 details.even_cows = false;
3732 details.single_folio = folio;
3733 details.zap_flags = ZAP_FLAG_DROP_MARKER;
3734
3735 i_mmap_lock_read(mapping);
3736 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3737 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3738 last_index, &details);
3739 i_mmap_unlock_read(mapping);
3740 }
3741
3742 /**
3743 * unmap_mapping_pages() - Unmap pages from processes.
3744 * @mapping: The address space containing pages to be unmapped.
3745 * @start: Index of first page to be unmapped.
3746 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3747 * @even_cows: Whether to unmap even private COWed pages.
3748 *
3749 * Unmap the pages in this address space from any userspace process which
3750 * has them mmaped. Generally, you want to remove COWed pages as well when
3751 * a file is being truncated, but not when invalidating pages from the page
3752 * cache.
3753 */
unmap_mapping_pages(struct address_space * mapping,pgoff_t start,pgoff_t nr,bool even_cows)3754 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3755 pgoff_t nr, bool even_cows)
3756 {
3757 struct zap_details details = { };
3758 pgoff_t first_index = start;
3759 pgoff_t last_index = start + nr - 1;
3760
3761 details.even_cows = even_cows;
3762 if (last_index < first_index)
3763 last_index = ULONG_MAX;
3764
3765 i_mmap_lock_read(mapping);
3766 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3767 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3768 last_index, &details);
3769 i_mmap_unlock_read(mapping);
3770 }
3771 EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3772
3773 /**
3774 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3775 * address_space corresponding to the specified byte range in the underlying
3776 * file.
3777 *
3778 * @mapping: the address space containing mmaps to be unmapped.
3779 * @holebegin: byte in first page to unmap, relative to the start of
3780 * the underlying file. This will be rounded down to a PAGE_SIZE
3781 * boundary. Note that this is different from truncate_pagecache(), which
3782 * must keep the partial page. In contrast, we must get rid of
3783 * partial pages.
3784 * @holelen: size of prospective hole in bytes. This will be rounded
3785 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3786 * end of the file.
3787 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3788 * but 0 when invalidating pagecache, don't throw away private data.
3789 */
unmap_mapping_range(struct address_space * mapping,loff_t const holebegin,loff_t const holelen,int even_cows)3790 void unmap_mapping_range(struct address_space *mapping,
3791 loff_t const holebegin, loff_t const holelen, int even_cows)
3792 {
3793 pgoff_t hba = (pgoff_t)(holebegin) >> PAGE_SHIFT;
3794 pgoff_t hlen = ((pgoff_t)(holelen) + PAGE_SIZE - 1) >> PAGE_SHIFT;
3795
3796 /* Check for overflow. */
3797 if (sizeof(holelen) > sizeof(hlen)) {
3798 long long holeend =
3799 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3800 if (holeend & ~(long long)ULONG_MAX)
3801 hlen = ULONG_MAX - hba + 1;
3802 }
3803
3804 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3805 }
3806 EXPORT_SYMBOL(unmap_mapping_range);
3807
3808 /*
3809 * Restore a potential device exclusive pte to a working pte entry
3810 */
remove_device_exclusive_entry(struct vm_fault * vmf)3811 static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3812 {
3813 struct folio *folio = page_folio(vmf->page);
3814 struct vm_area_struct *vma = vmf->vma;
3815 struct mmu_notifier_range range;
3816 vm_fault_t ret;
3817
3818 /*
3819 * We need a reference to lock the folio because we don't hold
3820 * the PTL so a racing thread can remove the device-exclusive
3821 * entry and unmap it. If the folio is free the entry must
3822 * have been removed already. If it happens to have already
3823 * been re-allocated after being freed all we do is lock and
3824 * unlock it.
3825 */
3826 if (!folio_try_get(folio))
3827 return 0;
3828
3829 ret = folio_lock_or_retry(folio, vmf);
3830 if (ret) {
3831 folio_put(folio);
3832 return ret;
3833 }
3834 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
3835 vma->vm_mm, vmf->address & PAGE_MASK,
3836 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3837 mmu_notifier_invalidate_range_start(&range);
3838
3839 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3840 &vmf->ptl);
3841 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3842 restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte);
3843
3844 if (vmf->pte)
3845 pte_unmap_unlock(vmf->pte, vmf->ptl);
3846 folio_unlock(folio);
3847 folio_put(folio);
3848
3849 mmu_notifier_invalidate_range_end(&range);
3850 return 0;
3851 }
3852
should_try_to_free_swap(struct folio * folio,struct vm_area_struct * vma,unsigned int fault_flags)3853 static inline bool should_try_to_free_swap(struct folio *folio,
3854 struct vm_area_struct *vma,
3855 unsigned int fault_flags)
3856 {
3857 if (!folio_test_swapcache(folio))
3858 return false;
3859 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
3860 folio_test_mlocked(folio))
3861 return true;
3862 /*
3863 * If we want to map a page that's in the swapcache writable, we
3864 * have to detect via the refcount if we're really the exclusive
3865 * user. Try freeing the swapcache to get rid of the swapcache
3866 * reference only in case it's likely that we'll be the exlusive user.
3867 */
3868 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
3869 folio_ref_count(folio) == 2;
3870 }
3871
pte_marker_clear(struct vm_fault * vmf)3872 static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
3873 {
3874 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
3875 vmf->address, &vmf->ptl);
3876 if (!vmf->pte)
3877 return 0;
3878 /*
3879 * Be careful so that we will only recover a special uffd-wp pte into a
3880 * none pte. Otherwise it means the pte could have changed, so retry.
3881 *
3882 * This should also cover the case where e.g. the pte changed
3883 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3884 * So is_pte_marker() check is not enough to safely drop the pte.
3885 */
3886 if (pte_same(vmf->orig_pte, ptep_get(vmf->pte)))
3887 pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
3888 pte_unmap_unlock(vmf->pte, vmf->ptl);
3889 return 0;
3890 }
3891
do_pte_missing(struct vm_fault * vmf)3892 static vm_fault_t do_pte_missing(struct vm_fault *vmf)
3893 {
3894 if (vma_is_anonymous(vmf->vma))
3895 return do_anonymous_page(vmf);
3896 else
3897 return do_fault(vmf);
3898 }
3899
3900 /*
3901 * This is actually a page-missing access, but with uffd-wp special pte
3902 * installed. It means this pte was wr-protected before being unmapped.
3903 */
pte_marker_handle_uffd_wp(struct vm_fault * vmf)3904 static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
3905 {
3906 /*
3907 * Just in case there're leftover special ptes even after the region
3908 * got unregistered - we can simply clear them.
3909 */
3910 if (unlikely(!userfaultfd_wp(vmf->vma)))
3911 return pte_marker_clear(vmf);
3912
3913 return do_pte_missing(vmf);
3914 }
3915
handle_pte_marker(struct vm_fault * vmf)3916 static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
3917 {
3918 swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte);
3919 unsigned long marker = pte_marker_get(entry);
3920
3921 /*
3922 * PTE markers should never be empty. If anything weird happened,
3923 * the best thing to do is to kill the process along with its mm.
3924 */
3925 if (WARN_ON_ONCE(!marker))
3926 return VM_FAULT_SIGBUS;
3927
3928 /* Higher priority than uffd-wp when data corrupted */
3929 if (marker & PTE_MARKER_POISONED)
3930 return VM_FAULT_HWPOISON;
3931
3932 if (pte_marker_entry_uffd_wp(entry))
3933 return pte_marker_handle_uffd_wp(vmf);
3934
3935 /* This is an unknown pte marker */
3936 return VM_FAULT_SIGBUS;
3937 }
3938
3939 /*
3940 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3941 * but allow concurrent faults), and pte mapped but not yet locked.
3942 * We return with pte unmapped and unlocked.
3943 *
3944 * We return with the mmap_lock locked or unlocked in the same cases
3945 * as does filemap_fault().
3946 */
do_swap_page(struct vm_fault * vmf)3947 vm_fault_t do_swap_page(struct vm_fault *vmf)
3948 {
3949 struct vm_area_struct *vma = vmf->vma;
3950 struct folio *swapcache, *folio = NULL;
3951 struct page *page;
3952 struct swap_info_struct *si = NULL;
3953 rmap_t rmap_flags = RMAP_NONE;
3954 bool need_clear_cache = false;
3955 bool exclusive = false;
3956 swp_entry_t entry;
3957 pte_t pte;
3958 vm_fault_t ret = 0;
3959 void *shadow = NULL;
3960
3961 if (!pte_unmap_same(vmf))
3962 goto out;
3963
3964 entry = pte_to_swp_entry(vmf->orig_pte);
3965 if (unlikely(non_swap_entry(entry))) {
3966 if (is_migration_entry(entry)) {
3967 migration_entry_wait(vma->vm_mm, vmf->pmd,
3968 vmf->address);
3969 } else if (is_device_exclusive_entry(entry)) {
3970 vmf->page = pfn_swap_entry_to_page(entry);
3971 ret = remove_device_exclusive_entry(vmf);
3972 } else if (is_device_private_entry(entry)) {
3973 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3974 /*
3975 * migrate_to_ram is not yet ready to operate
3976 * under VMA lock.
3977 */
3978 vma_end_read(vma);
3979 ret = VM_FAULT_RETRY;
3980 goto out;
3981 }
3982
3983 vmf->page = pfn_swap_entry_to_page(entry);
3984 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3985 vmf->address, &vmf->ptl);
3986 if (unlikely(!vmf->pte ||
3987 !pte_same(ptep_get(vmf->pte),
3988 vmf->orig_pte)))
3989 goto unlock;
3990
3991 /*
3992 * Get a page reference while we know the page can't be
3993 * freed.
3994 */
3995 get_page(vmf->page);
3996 pte_unmap_unlock(vmf->pte, vmf->ptl);
3997 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3998 put_page(vmf->page);
3999 } else if (is_hwpoison_entry(entry)) {
4000 ret = VM_FAULT_HWPOISON;
4001 } else if (is_pte_marker_entry(entry)) {
4002 ret = handle_pte_marker(vmf);
4003 } else {
4004 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
4005 ret = VM_FAULT_SIGBUS;
4006 }
4007 goto out;
4008 }
4009
4010 /* Prevent swapoff from happening to us. */
4011 si = get_swap_device(entry);
4012 if (unlikely(!si))
4013 goto out;
4014
4015 folio = swap_cache_get_folio(entry, vma, vmf->address);
4016 if (folio)
4017 page = folio_file_page(folio, swp_offset(entry));
4018 swapcache = folio;
4019
4020 if (!folio) {
4021 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
4022 __swap_count(entry) == 1) {
4023 /*
4024 * Prevent parallel swapin from proceeding with
4025 * the cache flag. Otherwise, another thread may
4026 * finish swapin first, free the entry, and swapout
4027 * reusing the same entry. It's undetectable as
4028 * pte_same() returns true due to entry reuse.
4029 */
4030 if (swapcache_prepare(entry)) {
4031 /* Relax a bit to prevent rapid repeated page faults */
4032 schedule_timeout_uninterruptible(1);
4033 goto out;
4034 }
4035 need_clear_cache = true;
4036
4037 /* skip swapcache */
4038 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0,
4039 vma, vmf->address, false);
4040 page = &folio->page;
4041 if (folio) {
4042 __folio_set_locked(folio);
4043 __folio_set_swapbacked(folio);
4044
4045 if (mem_cgroup_swapin_charge_folio(folio,
4046 vma->vm_mm, GFP_KERNEL,
4047 entry)) {
4048 ret = VM_FAULT_OOM;
4049 goto out_page;
4050 }
4051 mem_cgroup_swapin_uncharge_swap(entry);
4052
4053 shadow = get_shadow_from_swap_cache(entry);
4054 if (shadow)
4055 workingset_refault(folio, shadow);
4056
4057 folio_add_lru(folio);
4058
4059 /* To provide entry to swap_read_folio() */
4060 folio->swap = entry;
4061 swap_read_folio(folio, true, NULL);
4062 folio->private = NULL;
4063 }
4064 } else {
4065 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
4066 vmf);
4067 if (page)
4068 folio = page_folio(page);
4069 swapcache = folio;
4070 }
4071
4072 if (!folio) {
4073 /*
4074 * Back out if somebody else faulted in this pte
4075 * while we released the pte lock.
4076 */
4077 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4078 vmf->address, &vmf->ptl);
4079 if (likely(vmf->pte &&
4080 pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
4081 ret = VM_FAULT_OOM;
4082 goto unlock;
4083 }
4084
4085 /* Had to read the page from swap area: Major fault */
4086 ret = VM_FAULT_MAJOR;
4087 count_vm_event(PGMAJFAULT);
4088 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
4089 } else if (PageHWPoison(page)) {
4090 /*
4091 * hwpoisoned dirty swapcache pages are kept for killing
4092 * owner processes (which may be unknown at hwpoison time)
4093 */
4094 ret = VM_FAULT_HWPOISON;
4095 goto out_release;
4096 }
4097
4098 ret |= folio_lock_or_retry(folio, vmf);
4099 if (ret & VM_FAULT_RETRY)
4100 goto out_release;
4101
4102 if (swapcache) {
4103 /*
4104 * Make sure folio_free_swap() or swapoff did not release the
4105 * swapcache from under us. The page pin, and pte_same test
4106 * below, are not enough to exclude that. Even if it is still
4107 * swapcache, we need to check that the page's swap has not
4108 * changed.
4109 */
4110 if (unlikely(!folio_test_swapcache(folio) ||
4111 page_swap_entry(page).val != entry.val))
4112 goto out_page;
4113
4114 /*
4115 * KSM sometimes has to copy on read faults, for example, if
4116 * page->index of !PageKSM() pages would be nonlinear inside the
4117 * anon VMA -- PageKSM() is lost on actual swapout.
4118 */
4119 folio = ksm_might_need_to_copy(folio, vma, vmf->address);
4120 if (unlikely(!folio)) {
4121 ret = VM_FAULT_OOM;
4122 folio = swapcache;
4123 goto out_page;
4124 } else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
4125 ret = VM_FAULT_HWPOISON;
4126 folio = swapcache;
4127 goto out_page;
4128 }
4129 if (folio != swapcache)
4130 page = folio_page(folio, 0);
4131
4132 /*
4133 * If we want to map a page that's in the swapcache writable, we
4134 * have to detect via the refcount if we're really the exclusive
4135 * owner. Try removing the extra reference from the local LRU
4136 * caches if required.
4137 */
4138 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
4139 !folio_test_ksm(folio) && !folio_test_lru(folio))
4140 lru_add_drain();
4141 }
4142
4143 folio_throttle_swaprate(folio, GFP_KERNEL);
4144
4145 /*
4146 * Back out if somebody else already faulted in this pte.
4147 */
4148 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
4149 &vmf->ptl);
4150 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
4151 goto out_nomap;
4152
4153 if (unlikely(!folio_test_uptodate(folio))) {
4154 ret = VM_FAULT_SIGBUS;
4155 goto out_nomap;
4156 }
4157
4158 /*
4159 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4160 * must never point at an anonymous page in the swapcache that is
4161 * PG_anon_exclusive. Sanity check that this holds and especially, that
4162 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4163 * check after taking the PT lock and making sure that nobody
4164 * concurrently faulted in this page and set PG_anon_exclusive.
4165 */
4166 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
4167 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
4168
4169 /*
4170 * Check under PT lock (to protect against concurrent fork() sharing
4171 * the swap entry concurrently) for certainly exclusive pages.
4172 */
4173 if (!folio_test_ksm(folio)) {
4174 exclusive = pte_swp_exclusive(vmf->orig_pte);
4175 if (folio != swapcache) {
4176 /*
4177 * We have a fresh page that is not exposed to the
4178 * swapcache -> certainly exclusive.
4179 */
4180 exclusive = true;
4181 } else if (exclusive && folio_test_writeback(folio) &&
4182 data_race(si->flags & SWP_STABLE_WRITES)) {
4183 /*
4184 * This is tricky: not all swap backends support
4185 * concurrent page modifications while under writeback.
4186 *
4187 * So if we stumble over such a page in the swapcache
4188 * we must not set the page exclusive, otherwise we can
4189 * map it writable without further checks and modify it
4190 * while still under writeback.
4191 *
4192 * For these problematic swap backends, simply drop the
4193 * exclusive marker: this is perfectly fine as we start
4194 * writeback only if we fully unmapped the page and
4195 * there are no unexpected references on the page after
4196 * unmapping succeeded. After fully unmapped, no
4197 * further GUP references (FOLL_GET and FOLL_PIN) can
4198 * appear, so dropping the exclusive marker and mapping
4199 * it only R/O is fine.
4200 */
4201 exclusive = false;
4202 }
4203 }
4204
4205 /*
4206 * Some architectures may have to restore extra metadata to the page
4207 * when reading from swap. This metadata may be indexed by swap entry
4208 * so this must be called before swap_free().
4209 */
4210 arch_swap_restore(folio_swap(entry, folio), folio);
4211
4212 /*
4213 * Remove the swap entry and conditionally try to free up the swapcache.
4214 * We're already holding a reference on the page but haven't mapped it
4215 * yet.
4216 */
4217 swap_free(entry);
4218 if (should_try_to_free_swap(folio, vma, vmf->flags))
4219 folio_free_swap(folio);
4220
4221 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4222 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
4223 pte = mk_pte(page, vma->vm_page_prot);
4224
4225 /*
4226 * Same logic as in do_wp_page(); however, optimize for pages that are
4227 * certainly not shared either because we just allocated them without
4228 * exposing them to the swapcache or because the swap entry indicates
4229 * exclusivity.
4230 */
4231 if (!folio_test_ksm(folio) &&
4232 (exclusive || folio_ref_count(folio) == 1)) {
4233 if (vmf->flags & FAULT_FLAG_WRITE) {
4234 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
4235 vmf->flags &= ~FAULT_FLAG_WRITE;
4236 }
4237 rmap_flags |= RMAP_EXCLUSIVE;
4238 }
4239 flush_icache_page(vma, page);
4240 if (pte_swp_soft_dirty(vmf->orig_pte))
4241 pte = pte_mksoft_dirty(pte);
4242 if (pte_swp_uffd_wp(vmf->orig_pte))
4243 pte = pte_mkuffd_wp(pte);
4244 vmf->orig_pte = pte;
4245
4246 /* ksm created a completely new copy */
4247 if (unlikely(folio != swapcache && swapcache)) {
4248 folio_add_new_anon_rmap(folio, vma, vmf->address);
4249 folio_add_lru_vma(folio, vma);
4250 } else {
4251 folio_add_anon_rmap_pte(folio, page, vma, vmf->address,
4252 rmap_flags);
4253 }
4254
4255 VM_BUG_ON(!folio_test_anon(folio) ||
4256 (pte_write(pte) && !PageAnonExclusive(page)));
4257 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
4258 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
4259
4260 folio_unlock(folio);
4261 if (folio != swapcache && swapcache) {
4262 /*
4263 * Hold the lock to avoid the swap entry to be reused
4264 * until we take the PT lock for the pte_same() check
4265 * (to avoid false positives from pte_same). For
4266 * further safety release the lock after the swap_free
4267 * so that the swap count won't change under a
4268 * parallel locked swapcache.
4269 */
4270 folio_unlock(swapcache);
4271 folio_put(swapcache);
4272 }
4273
4274 if (vmf->flags & FAULT_FLAG_WRITE) {
4275 ret |= do_wp_page(vmf);
4276 if (ret & VM_FAULT_ERROR)
4277 ret &= VM_FAULT_ERROR;
4278 goto out;
4279 }
4280
4281 /* No need to invalidate - it was non-present before */
4282 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
4283 unlock:
4284 if (vmf->pte)
4285 pte_unmap_unlock(vmf->pte, vmf->ptl);
4286 out:
4287 /* Clear the swap cache pin for direct swapin after PTL unlock */
4288 if (need_clear_cache)
4289 swapcache_clear(si, entry);
4290 if (si)
4291 put_swap_device(si);
4292 return ret;
4293 out_nomap:
4294 if (vmf->pte)
4295 pte_unmap_unlock(vmf->pte, vmf->ptl);
4296 out_page:
4297 folio_unlock(folio);
4298 out_release:
4299 folio_put(folio);
4300 if (folio != swapcache && swapcache) {
4301 folio_unlock(swapcache);
4302 folio_put(swapcache);
4303 }
4304 if (need_clear_cache)
4305 swapcache_clear(si, entry);
4306 if (si)
4307 put_swap_device(si);
4308 return ret;
4309 }
4310
pte_range_none(pte_t * pte,int nr_pages)4311 static bool pte_range_none(pte_t *pte, int nr_pages)
4312 {
4313 int i;
4314
4315 for (i = 0; i < nr_pages; i++) {
4316 if (!pte_none(ptep_get_lockless(pte + i)))
4317 return false;
4318 }
4319
4320 return true;
4321 }
4322
alloc_anon_folio(struct vm_fault * vmf)4323 static struct folio *alloc_anon_folio(struct vm_fault *vmf)
4324 {
4325 struct vm_area_struct *vma = vmf->vma;
4326 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4327 unsigned long orders;
4328 struct folio *folio;
4329 unsigned long addr;
4330 pte_t *pte;
4331 gfp_t gfp;
4332 int order;
4333
4334 /*
4335 * If uffd is active for the vma we need per-page fault fidelity to
4336 * maintain the uffd semantics.
4337 */
4338 if (unlikely(userfaultfd_armed(vma)))
4339 goto fallback;
4340
4341 /*
4342 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4343 * for this vma. Then filter out the orders that can't be allocated over
4344 * the faulting address and still be fully contained in the vma.
4345 */
4346 orders = thp_vma_allowable_orders(vma, vma->vm_flags,
4347 TVA_IN_PF | TVA_ENFORCE_SYSFS, BIT(PMD_ORDER) - 1);
4348 orders = thp_vma_suitable_orders(vma, vmf->address, orders);
4349
4350 if (!orders)
4351 goto fallback;
4352
4353 pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK);
4354 if (!pte)
4355 return ERR_PTR(-EAGAIN);
4356
4357 /*
4358 * Find the highest order where the aligned range is completely
4359 * pte_none(). Note that all remaining orders will be completely
4360 * pte_none().
4361 */
4362 order = highest_order(orders);
4363 while (orders) {
4364 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4365 if (pte_range_none(pte + pte_index(addr), 1 << order))
4366 break;
4367 order = next_order(&orders, order);
4368 }
4369
4370 pte_unmap(pte);
4371
4372 if (!orders)
4373 goto fallback;
4374
4375 /* Try allocating the highest of the remaining orders. */
4376 gfp = vma_thp_gfp_mask(vma);
4377 while (orders) {
4378 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4379 folio = vma_alloc_folio(gfp, order, vma, addr, true);
4380 if (folio) {
4381 if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
4382 count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
4383 folio_put(folio);
4384 goto next;
4385 }
4386 folio_throttle_swaprate(folio, gfp);
4387 clear_huge_page(&folio->page, vmf->address, 1 << order);
4388 return folio;
4389 }
4390 next:
4391 count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK);
4392 order = next_order(&orders, order);
4393 }
4394
4395 fallback:
4396 #endif
4397 return folio_prealloc(vma->vm_mm, vma, vmf->address, true);
4398 }
4399
4400 /*
4401 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4402 * but allow concurrent faults), and pte mapped but not yet locked.
4403 * We return with mmap_lock still held, but pte unmapped and unlocked.
4404 */
do_anonymous_page(struct vm_fault * vmf)4405 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
4406 {
4407 struct vm_area_struct *vma = vmf->vma;
4408 unsigned long addr = vmf->address;
4409 struct folio *folio;
4410 vm_fault_t ret = 0;
4411 int nr_pages = 1;
4412 pte_t entry;
4413 int i;
4414
4415 /* File mapping without ->vm_ops ? */
4416 if (vma->vm_flags & VM_SHARED)
4417 return VM_FAULT_SIGBUS;
4418
4419 /*
4420 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4421 * be distinguished from a transient failure of pte_offset_map().
4422 */
4423 if (pte_alloc(vma->vm_mm, vmf->pmd))
4424 return VM_FAULT_OOM;
4425
4426 /* Use the zero-page for reads */
4427 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
4428 !mm_forbids_zeropage(vma->vm_mm)) {
4429 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
4430 vma->vm_page_prot));
4431 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4432 vmf->address, &vmf->ptl);
4433 if (!vmf->pte)
4434 goto unlock;
4435 if (vmf_pte_changed(vmf)) {
4436 update_mmu_tlb(vma, vmf->address, vmf->pte);
4437 goto unlock;
4438 }
4439 ret = check_stable_address_space(vma->vm_mm);
4440 if (ret)
4441 goto unlock;
4442 /* Deliver the page fault to userland, check inside PT lock */
4443 if (userfaultfd_missing(vma)) {
4444 pte_unmap_unlock(vmf->pte, vmf->ptl);
4445 return handle_userfault(vmf, VM_UFFD_MISSING);
4446 }
4447 goto setpte;
4448 }
4449
4450 /* Allocate our own private page. */
4451 ret = vmf_anon_prepare(vmf);
4452 if (ret)
4453 return ret;
4454 /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
4455 folio = alloc_anon_folio(vmf);
4456 if (IS_ERR(folio))
4457 return 0;
4458 if (!folio)
4459 goto oom;
4460
4461 nr_pages = folio_nr_pages(folio);
4462 addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE);
4463
4464 /*
4465 * The memory barrier inside __folio_mark_uptodate makes sure that
4466 * preceding stores to the page contents become visible before
4467 * the set_pte_at() write.
4468 */
4469 __folio_mark_uptodate(folio);
4470
4471 entry = mk_pte(&folio->page, vma->vm_page_prot);
4472 entry = pte_sw_mkyoung(entry);
4473 if (vma->vm_flags & VM_WRITE)
4474 entry = pte_mkwrite(pte_mkdirty(entry), vma);
4475
4476 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
4477 if (!vmf->pte)
4478 goto release;
4479 if (nr_pages == 1 && vmf_pte_changed(vmf)) {
4480 update_mmu_tlb(vma, addr, vmf->pte);
4481 goto release;
4482 } else if (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages)) {
4483 for (i = 0; i < nr_pages; i++)
4484 update_mmu_tlb(vma, addr + PAGE_SIZE * i, vmf->pte + i);
4485 goto release;
4486 }
4487
4488 ret = check_stable_address_space(vma->vm_mm);
4489 if (ret)
4490 goto release;
4491
4492 /* Deliver the page fault to userland, check inside PT lock */
4493 if (userfaultfd_missing(vma)) {
4494 pte_unmap_unlock(vmf->pte, vmf->ptl);
4495 folio_put(folio);
4496 return handle_userfault(vmf, VM_UFFD_MISSING);
4497 }
4498
4499 folio_ref_add(folio, nr_pages - 1);
4500 add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages);
4501 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4502 count_mthp_stat(folio_order(folio), MTHP_STAT_ANON_FAULT_ALLOC);
4503 #endif
4504 folio_add_new_anon_rmap(folio, vma, addr);
4505 folio_add_lru_vma(folio, vma);
4506 setpte:
4507 if (vmf_orig_pte_uffd_wp(vmf))
4508 entry = pte_mkuffd_wp(entry);
4509 set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr_pages);
4510
4511 /* No need to invalidate - it was non-present before */
4512 update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr_pages);
4513 unlock:
4514 if (vmf->pte)
4515 pte_unmap_unlock(vmf->pte, vmf->ptl);
4516 return ret;
4517 release:
4518 folio_put(folio);
4519 goto unlock;
4520 oom:
4521 return VM_FAULT_OOM;
4522 }
4523
4524 /*
4525 * The mmap_lock must have been held on entry, and may have been
4526 * released depending on flags and vma->vm_ops->fault() return value.
4527 * See filemap_fault() and __lock_page_retry().
4528 */
__do_fault(struct vm_fault * vmf)4529 static vm_fault_t __do_fault(struct vm_fault *vmf)
4530 {
4531 struct vm_area_struct *vma = vmf->vma;
4532 struct folio *folio;
4533 vm_fault_t ret;
4534
4535 /*
4536 * Preallocate pte before we take page_lock because this might lead to
4537 * deadlocks for memcg reclaim which waits for pages under writeback:
4538 * lock_page(A)
4539 * SetPageWriteback(A)
4540 * unlock_page(A)
4541 * lock_page(B)
4542 * lock_page(B)
4543 * pte_alloc_one
4544 * shrink_page_list
4545 * wait_on_page_writeback(A)
4546 * SetPageWriteback(B)
4547 * unlock_page(B)
4548 * # flush A, B to clear the writeback
4549 */
4550 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
4551 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4552 if (!vmf->prealloc_pte)
4553 return VM_FAULT_OOM;
4554 }
4555
4556 ret = vma->vm_ops->fault(vmf);
4557 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
4558 VM_FAULT_DONE_COW)))
4559 return ret;
4560
4561 folio = page_folio(vmf->page);
4562 if (unlikely(PageHWPoison(vmf->page))) {
4563 vm_fault_t poisonret = VM_FAULT_HWPOISON;
4564 if (ret & VM_FAULT_LOCKED) {
4565 if (page_mapped(vmf->page))
4566 unmap_mapping_folio(folio);
4567 /* Retry if a clean folio was removed from the cache. */
4568 if (mapping_evict_folio(folio->mapping, folio))
4569 poisonret = VM_FAULT_NOPAGE;
4570 folio_unlock(folio);
4571 }
4572 folio_put(folio);
4573 vmf->page = NULL;
4574 return poisonret;
4575 }
4576
4577 if (unlikely(!(ret & VM_FAULT_LOCKED)))
4578 folio_lock(folio);
4579 else
4580 VM_BUG_ON_PAGE(!folio_test_locked(folio), vmf->page);
4581
4582 return ret;
4583 }
4584
4585 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
deposit_prealloc_pte(struct vm_fault * vmf)4586 static void deposit_prealloc_pte(struct vm_fault *vmf)
4587 {
4588 struct vm_area_struct *vma = vmf->vma;
4589
4590 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4591 /*
4592 * We are going to consume the prealloc table,
4593 * count that as nr_ptes.
4594 */
4595 mm_inc_nr_ptes(vma->vm_mm);
4596 vmf->prealloc_pte = NULL;
4597 }
4598
do_set_pmd(struct vm_fault * vmf,struct page * page)4599 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4600 {
4601 struct folio *folio = page_folio(page);
4602 struct vm_area_struct *vma = vmf->vma;
4603 bool write = vmf->flags & FAULT_FLAG_WRITE;
4604 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
4605 pmd_t entry;
4606 vm_fault_t ret = VM_FAULT_FALLBACK;
4607
4608 if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
4609 return ret;
4610
4611 if (folio_order(folio) != HPAGE_PMD_ORDER)
4612 return ret;
4613 page = &folio->page;
4614
4615 /*
4616 * Just backoff if any subpage of a THP is corrupted otherwise
4617 * the corrupted page may mapped by PMD silently to escape the
4618 * check. This kind of THP just can be PTE mapped. Access to
4619 * the corrupted subpage should trigger SIGBUS as expected.
4620 */
4621 if (unlikely(folio_test_has_hwpoisoned(folio)))
4622 return ret;
4623
4624 /*
4625 * Archs like ppc64 need additional space to store information
4626 * related to pte entry. Use the preallocated table for that.
4627 */
4628 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
4629 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4630 if (!vmf->prealloc_pte)
4631 return VM_FAULT_OOM;
4632 }
4633
4634 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4635 if (unlikely(!pmd_none(*vmf->pmd)))
4636 goto out;
4637
4638 flush_icache_pages(vma, page, HPAGE_PMD_NR);
4639
4640 entry = mk_huge_pmd(page, vma->vm_page_prot);
4641 if (write)
4642 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
4643
4644 add_mm_counter(vma->vm_mm, mm_counter_file(folio), HPAGE_PMD_NR);
4645 folio_add_file_rmap_pmd(folio, page, vma);
4646
4647 /*
4648 * deposit and withdraw with pmd lock held
4649 */
4650 if (arch_needs_pgtable_deposit())
4651 deposit_prealloc_pte(vmf);
4652
4653 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
4654
4655 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
4656
4657 /* fault is handled */
4658 ret = 0;
4659 count_vm_event(THP_FILE_MAPPED);
4660 out:
4661 spin_unlock(vmf->ptl);
4662 return ret;
4663 }
4664 #else
do_set_pmd(struct vm_fault * vmf,struct page * page)4665 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4666 {
4667 return VM_FAULT_FALLBACK;
4668 }
4669 #endif
4670
4671 /**
4672 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4673 * @vmf: Fault decription.
4674 * @folio: The folio that contains @page.
4675 * @page: The first page to create a PTE for.
4676 * @nr: The number of PTEs to create.
4677 * @addr: The first address to create a PTE for.
4678 */
set_pte_range(struct vm_fault * vmf,struct folio * folio,struct page * page,unsigned int nr,unsigned long addr)4679 void set_pte_range(struct vm_fault *vmf, struct folio *folio,
4680 struct page *page, unsigned int nr, unsigned long addr)
4681 {
4682 struct vm_area_struct *vma = vmf->vma;
4683 bool write = vmf->flags & FAULT_FLAG_WRITE;
4684 bool prefault = in_range(vmf->address, addr, nr * PAGE_SIZE);
4685 pte_t entry;
4686
4687 flush_icache_pages(vma, page, nr);
4688 entry = mk_pte(page, vma->vm_page_prot);
4689
4690 if (prefault && arch_wants_old_prefaulted_pte())
4691 entry = pte_mkold(entry);
4692 else
4693 entry = pte_sw_mkyoung(entry);
4694
4695 if (write)
4696 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4697 if (unlikely(vmf_orig_pte_uffd_wp(vmf)))
4698 entry = pte_mkuffd_wp(entry);
4699 /* copy-on-write page */
4700 if (write && !(vma->vm_flags & VM_SHARED)) {
4701 VM_BUG_ON_FOLIO(nr != 1, folio);
4702 folio_add_new_anon_rmap(folio, vma, addr);
4703 folio_add_lru_vma(folio, vma);
4704 } else {
4705 folio_add_file_rmap_ptes(folio, page, nr, vma);
4706 }
4707 set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr);
4708
4709 /* no need to invalidate: a not-present page won't be cached */
4710 update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr);
4711 }
4712
vmf_pte_changed(struct vm_fault * vmf)4713 static bool vmf_pte_changed(struct vm_fault *vmf)
4714 {
4715 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
4716 return !pte_same(ptep_get(vmf->pte), vmf->orig_pte);
4717
4718 return !pte_none(ptep_get(vmf->pte));
4719 }
4720
4721 /**
4722 * finish_fault - finish page fault once we have prepared the page to fault
4723 *
4724 * @vmf: structure describing the fault
4725 *
4726 * This function handles all that is needed to finish a page fault once the
4727 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4728 * given page, adds reverse page mapping, handles memcg charges and LRU
4729 * addition.
4730 *
4731 * The function expects the page to be locked and on success it consumes a
4732 * reference of a page being mapped (for the PTE which maps it).
4733 *
4734 * Return: %0 on success, %VM_FAULT_ code in case of error.
4735 */
finish_fault(struct vm_fault * vmf)4736 vm_fault_t finish_fault(struct vm_fault *vmf)
4737 {
4738 struct vm_area_struct *vma = vmf->vma;
4739 struct page *page;
4740 vm_fault_t ret;
4741 bool is_cow = (vmf->flags & FAULT_FLAG_WRITE) &&
4742 !(vma->vm_flags & VM_SHARED);
4743
4744 /* Did we COW the page? */
4745 if (is_cow)
4746 page = vmf->cow_page;
4747 else
4748 page = vmf->page;
4749
4750 /*
4751 * check even for read faults because we might have lost our CoWed
4752 * page
4753 */
4754 if (!(vma->vm_flags & VM_SHARED)) {
4755 ret = check_stable_address_space(vma->vm_mm);
4756 if (ret)
4757 return ret;
4758 }
4759
4760 if (pmd_none(*vmf->pmd)) {
4761 if (PageTransCompound(page)) {
4762 ret = do_set_pmd(vmf, page);
4763 if (ret != VM_FAULT_FALLBACK)
4764 return ret;
4765 }
4766
4767 if (vmf->prealloc_pte)
4768 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4769 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4770 return VM_FAULT_OOM;
4771 }
4772
4773 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4774 vmf->address, &vmf->ptl);
4775 if (!vmf->pte)
4776 return VM_FAULT_NOPAGE;
4777
4778 /* Re-check under ptl */
4779 if (likely(!vmf_pte_changed(vmf))) {
4780 struct folio *folio = page_folio(page);
4781 int type = is_cow ? MM_ANONPAGES : mm_counter_file(folio);
4782
4783 set_pte_range(vmf, folio, page, 1, vmf->address);
4784 add_mm_counter(vma->vm_mm, type, 1);
4785 ret = 0;
4786 } else {
4787 update_mmu_tlb(vma, vmf->address, vmf->pte);
4788 ret = VM_FAULT_NOPAGE;
4789 }
4790
4791 pte_unmap_unlock(vmf->pte, vmf->ptl);
4792 return ret;
4793 }
4794
4795 static unsigned long fault_around_pages __read_mostly =
4796 65536 >> PAGE_SHIFT;
4797
4798 #ifdef CONFIG_DEBUG_FS
fault_around_bytes_get(void * data,u64 * val)4799 static int fault_around_bytes_get(void *data, u64 *val)
4800 {
4801 *val = fault_around_pages << PAGE_SHIFT;
4802 return 0;
4803 }
4804
4805 /*
4806 * fault_around_bytes must be rounded down to the nearest page order as it's
4807 * what do_fault_around() expects to see.
4808 */
fault_around_bytes_set(void * data,u64 val)4809 static int fault_around_bytes_set(void *data, u64 val)
4810 {
4811 if (val / PAGE_SIZE > PTRS_PER_PTE)
4812 return -EINVAL;
4813
4814 /*
4815 * The minimum value is 1 page, however this results in no fault-around
4816 * at all. See should_fault_around().
4817 */
4818 val = max(val, PAGE_SIZE);
4819 fault_around_pages = rounddown_pow_of_two(val) >> PAGE_SHIFT;
4820
4821 return 0;
4822 }
4823 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4824 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4825
fault_around_debugfs(void)4826 static int __init fault_around_debugfs(void)
4827 {
4828 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4829 &fault_around_bytes_fops);
4830 return 0;
4831 }
4832 late_initcall(fault_around_debugfs);
4833 #endif
4834
4835 /*
4836 * do_fault_around() tries to map few pages around the fault address. The hope
4837 * is that the pages will be needed soon and this will lower the number of
4838 * faults to handle.
4839 *
4840 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4841 * not ready to be mapped: not up-to-date, locked, etc.
4842 *
4843 * This function doesn't cross VMA or page table boundaries, in order to call
4844 * map_pages() and acquire a PTE lock only once.
4845 *
4846 * fault_around_pages defines how many pages we'll try to map.
4847 * do_fault_around() expects it to be set to a power of two less than or equal
4848 * to PTRS_PER_PTE.
4849 *
4850 * The virtual address of the area that we map is naturally aligned to
4851 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4852 * (and therefore to page order). This way it's easier to guarantee
4853 * that we don't cross page table boundaries.
4854 */
do_fault_around(struct vm_fault * vmf)4855 static vm_fault_t do_fault_around(struct vm_fault *vmf)
4856 {
4857 pgoff_t nr_pages = READ_ONCE(fault_around_pages);
4858 pgoff_t pte_off = pte_index(vmf->address);
4859 /* The page offset of vmf->address within the VMA. */
4860 pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
4861 pgoff_t from_pte, to_pte;
4862 vm_fault_t ret;
4863
4864 /* The PTE offset of the start address, clamped to the VMA. */
4865 from_pte = max(ALIGN_DOWN(pte_off, nr_pages),
4866 pte_off - min(pte_off, vma_off));
4867
4868 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4869 to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE,
4870 pte_off + vma_pages(vmf->vma) - vma_off) - 1;
4871
4872 if (pmd_none(*vmf->pmd)) {
4873 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4874 if (!vmf->prealloc_pte)
4875 return VM_FAULT_OOM;
4876 }
4877
4878 rcu_read_lock();
4879 ret = vmf->vma->vm_ops->map_pages(vmf,
4880 vmf->pgoff + from_pte - pte_off,
4881 vmf->pgoff + to_pte - pte_off);
4882 rcu_read_unlock();
4883
4884 return ret;
4885 }
4886
4887 /* Return true if we should do read fault-around, false otherwise */
should_fault_around(struct vm_fault * vmf)4888 static inline bool should_fault_around(struct vm_fault *vmf)
4889 {
4890 /* No ->map_pages? No way to fault around... */
4891 if (!vmf->vma->vm_ops->map_pages)
4892 return false;
4893
4894 if (uffd_disable_fault_around(vmf->vma))
4895 return false;
4896
4897 /* A single page implies no faulting 'around' at all. */
4898 return fault_around_pages > 1;
4899 }
4900
do_read_fault(struct vm_fault * vmf)4901 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4902 {
4903 vm_fault_t ret = 0;
4904 struct folio *folio;
4905
4906 /*
4907 * Let's call ->map_pages() first and use ->fault() as fallback
4908 * if page by the offset is not ready to be mapped (cold cache or
4909 * something).
4910 */
4911 if (should_fault_around(vmf)) {
4912 ret = do_fault_around(vmf);
4913 if (ret)
4914 return ret;
4915 }
4916
4917 ret = vmf_can_call_fault(vmf);
4918 if (ret)
4919 return ret;
4920
4921 ret = __do_fault(vmf);
4922 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4923 return ret;
4924
4925 ret |= finish_fault(vmf);
4926 folio = page_folio(vmf->page);
4927 folio_unlock(folio);
4928 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4929 folio_put(folio);
4930 return ret;
4931 }
4932
do_cow_fault(struct vm_fault * vmf)4933 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4934 {
4935 struct vm_area_struct *vma = vmf->vma;
4936 struct folio *folio;
4937 vm_fault_t ret;
4938
4939 ret = vmf_can_call_fault(vmf);
4940 if (!ret)
4941 ret = vmf_anon_prepare(vmf);
4942 if (ret)
4943 return ret;
4944
4945 folio = folio_prealloc(vma->vm_mm, vma, vmf->address, false);
4946 if (!folio)
4947 return VM_FAULT_OOM;
4948
4949 vmf->cow_page = &folio->page;
4950
4951 ret = __do_fault(vmf);
4952 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4953 goto uncharge_out;
4954 if (ret & VM_FAULT_DONE_COW)
4955 return ret;
4956
4957 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4958 __folio_mark_uptodate(folio);
4959
4960 ret |= finish_fault(vmf);
4961 unlock_page(vmf->page);
4962 put_page(vmf->page);
4963 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4964 goto uncharge_out;
4965 return ret;
4966 uncharge_out:
4967 folio_put(folio);
4968 return ret;
4969 }
4970
do_shared_fault(struct vm_fault * vmf)4971 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4972 {
4973 struct vm_area_struct *vma = vmf->vma;
4974 vm_fault_t ret, tmp;
4975 struct folio *folio;
4976
4977 ret = vmf_can_call_fault(vmf);
4978 if (ret)
4979 return ret;
4980
4981 ret = __do_fault(vmf);
4982 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4983 return ret;
4984
4985 folio = page_folio(vmf->page);
4986
4987 /*
4988 * Check if the backing address space wants to know that the page is
4989 * about to become writable
4990 */
4991 if (vma->vm_ops->page_mkwrite) {
4992 folio_unlock(folio);
4993 tmp = do_page_mkwrite(vmf, folio);
4994 if (unlikely(!tmp ||
4995 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4996 folio_put(folio);
4997 return tmp;
4998 }
4999 }
5000
5001 ret |= finish_fault(vmf);
5002 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
5003 VM_FAULT_RETRY))) {
5004 folio_unlock(folio);
5005 folio_put(folio);
5006 return ret;
5007 }
5008
5009 ret |= fault_dirty_shared_page(vmf);
5010 return ret;
5011 }
5012
5013 /*
5014 * We enter with non-exclusive mmap_lock (to exclude vma changes,
5015 * but allow concurrent faults).
5016 * The mmap_lock may have been released depending on flags and our
5017 * return value. See filemap_fault() and __folio_lock_or_retry().
5018 * If mmap_lock is released, vma may become invalid (for example
5019 * by other thread calling munmap()).
5020 */
do_fault(struct vm_fault * vmf)5021 static vm_fault_t do_fault(struct vm_fault *vmf)
5022 {
5023 struct vm_area_struct *vma = vmf->vma;
5024 struct mm_struct *vm_mm = vma->vm_mm;
5025 vm_fault_t ret;
5026
5027 /*
5028 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
5029 */
5030 if (!vma->vm_ops->fault) {
5031 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
5032 vmf->address, &vmf->ptl);
5033 if (unlikely(!vmf->pte))
5034 ret = VM_FAULT_SIGBUS;
5035 else {
5036 /*
5037 * Make sure this is not a temporary clearing of pte
5038 * by holding ptl and checking again. A R/M/W update
5039 * of pte involves: take ptl, clearing the pte so that
5040 * we don't have concurrent modification by hardware
5041 * followed by an update.
5042 */
5043 if (unlikely(pte_none(ptep_get(vmf->pte))))
5044 ret = VM_FAULT_SIGBUS;
5045 else
5046 ret = VM_FAULT_NOPAGE;
5047
5048 pte_unmap_unlock(vmf->pte, vmf->ptl);
5049 }
5050 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
5051 ret = do_read_fault(vmf);
5052 else if (!(vma->vm_flags & VM_SHARED))
5053 ret = do_cow_fault(vmf);
5054 else
5055 ret = do_shared_fault(vmf);
5056
5057 /* preallocated pagetable is unused: free it */
5058 if (vmf->prealloc_pte) {
5059 pte_free(vm_mm, vmf->prealloc_pte);
5060 vmf->prealloc_pte = NULL;
5061 }
5062 return ret;
5063 }
5064
numa_migrate_prep(struct folio * folio,struct vm_fault * vmf,unsigned long addr,int page_nid,int * flags)5065 int numa_migrate_prep(struct folio *folio, struct vm_fault *vmf,
5066 unsigned long addr, int page_nid, int *flags)
5067 {
5068 struct vm_area_struct *vma = vmf->vma;
5069
5070 folio_get(folio);
5071
5072 /* Record the current PID acceesing VMA */
5073 vma_set_access_pid_bit(vma);
5074
5075 count_vm_numa_event(NUMA_HINT_FAULTS);
5076 if (page_nid == numa_node_id()) {
5077 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
5078 *flags |= TNF_FAULT_LOCAL;
5079 }
5080
5081 return mpol_misplaced(folio, vmf, addr);
5082 }
5083
numa_rebuild_single_mapping(struct vm_fault * vmf,struct vm_area_struct * vma,unsigned long fault_addr,pte_t * fault_pte,bool writable)5084 static void numa_rebuild_single_mapping(struct vm_fault *vmf, struct vm_area_struct *vma,
5085 unsigned long fault_addr, pte_t *fault_pte,
5086 bool writable)
5087 {
5088 pte_t pte, old_pte;
5089
5090 old_pte = ptep_modify_prot_start(vma, fault_addr, fault_pte);
5091 pte = pte_modify(old_pte, vma->vm_page_prot);
5092 pte = pte_mkyoung(pte);
5093 if (writable)
5094 pte = pte_mkwrite(pte, vma);
5095 ptep_modify_prot_commit(vma, fault_addr, fault_pte, old_pte, pte);
5096 update_mmu_cache_range(vmf, vma, fault_addr, fault_pte, 1);
5097 }
5098
numa_rebuild_large_mapping(struct vm_fault * vmf,struct vm_area_struct * vma,struct folio * folio,pte_t fault_pte,bool ignore_writable,bool pte_write_upgrade)5099 static void numa_rebuild_large_mapping(struct vm_fault *vmf, struct vm_area_struct *vma,
5100 struct folio *folio, pte_t fault_pte,
5101 bool ignore_writable, bool pte_write_upgrade)
5102 {
5103 int nr = pte_pfn(fault_pte) - folio_pfn(folio);
5104 unsigned long start, end, addr = vmf->address;
5105 unsigned long addr_start = addr - (nr << PAGE_SHIFT);
5106 unsigned long pt_start = ALIGN_DOWN(addr, PMD_SIZE);
5107 pte_t *start_ptep;
5108
5109 /* Stay within the VMA and within the page table. */
5110 start = max3(addr_start, pt_start, vma->vm_start);
5111 end = min3(addr_start + folio_size(folio), pt_start + PMD_SIZE,
5112 vma->vm_end);
5113 start_ptep = vmf->pte - ((addr - start) >> PAGE_SHIFT);
5114
5115 /* Restore all PTEs' mapping of the large folio */
5116 for (addr = start; addr != end; start_ptep++, addr += PAGE_SIZE) {
5117 pte_t ptent = ptep_get(start_ptep);
5118 bool writable = false;
5119
5120 if (!pte_present(ptent) || !pte_protnone(ptent))
5121 continue;
5122
5123 if (pfn_folio(pte_pfn(ptent)) != folio)
5124 continue;
5125
5126 if (!ignore_writable) {
5127 ptent = pte_modify(ptent, vma->vm_page_prot);
5128 writable = pte_write(ptent);
5129 if (!writable && pte_write_upgrade &&
5130 can_change_pte_writable(vma, addr, ptent))
5131 writable = true;
5132 }
5133
5134 numa_rebuild_single_mapping(vmf, vma, addr, start_ptep, writable);
5135 }
5136 }
5137
do_numa_page(struct vm_fault * vmf)5138 static vm_fault_t do_numa_page(struct vm_fault *vmf)
5139 {
5140 struct vm_area_struct *vma = vmf->vma;
5141 struct folio *folio = NULL;
5142 int nid = NUMA_NO_NODE;
5143 bool writable = false, ignore_writable = false;
5144 bool pte_write_upgrade = vma_wants_manual_pte_write_upgrade(vma);
5145 int last_cpupid;
5146 int target_nid;
5147 pte_t pte, old_pte;
5148 int flags = 0, nr_pages;
5149
5150 /*
5151 * The pte cannot be used safely until we verify, while holding the page
5152 * table lock, that its contents have not changed during fault handling.
5153 */
5154 spin_lock(vmf->ptl);
5155 /* Read the live PTE from the page tables: */
5156 old_pte = ptep_get(vmf->pte);
5157
5158 if (unlikely(!pte_same(old_pte, vmf->orig_pte))) {
5159 pte_unmap_unlock(vmf->pte, vmf->ptl);
5160 goto out;
5161 }
5162
5163 pte = pte_modify(old_pte, vma->vm_page_prot);
5164
5165 /*
5166 * Detect now whether the PTE could be writable; this information
5167 * is only valid while holding the PT lock.
5168 */
5169 writable = pte_write(pte);
5170 if (!writable && pte_write_upgrade &&
5171 can_change_pte_writable(vma, vmf->address, pte))
5172 writable = true;
5173
5174 folio = vm_normal_folio(vma, vmf->address, pte);
5175 if (!folio || folio_is_zone_device(folio))
5176 goto out_map;
5177
5178 /*
5179 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
5180 * much anyway since they can be in shared cache state. This misses
5181 * the case where a mapping is writable but the process never writes
5182 * to it but pte_write gets cleared during protection updates and
5183 * pte_dirty has unpredictable behaviour between PTE scan updates,
5184 * background writeback, dirty balancing and application behaviour.
5185 */
5186 if (!writable)
5187 flags |= TNF_NO_GROUP;
5188
5189 /*
5190 * Flag if the folio is shared between multiple address spaces. This
5191 * is later used when determining whether to group tasks together
5192 */
5193 if (folio_likely_mapped_shared(folio) && (vma->vm_flags & VM_SHARED))
5194 flags |= TNF_SHARED;
5195
5196 nid = folio_nid(folio);
5197 nr_pages = folio_nr_pages(folio);
5198 /*
5199 * For memory tiering mode, cpupid of slow memory page is used
5200 * to record page access time. So use default value.
5201 */
5202 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
5203 !node_is_toptier(nid))
5204 last_cpupid = (-1 & LAST_CPUPID_MASK);
5205 else
5206 last_cpupid = folio_last_cpupid(folio);
5207 target_nid = numa_migrate_prep(folio, vmf, vmf->address, nid, &flags);
5208 if (target_nid == NUMA_NO_NODE) {
5209 folio_put(folio);
5210 goto out_map;
5211 }
5212 pte_unmap_unlock(vmf->pte, vmf->ptl);
5213 writable = false;
5214 ignore_writable = true;
5215
5216 /* Migrate to the requested node */
5217 if (migrate_misplaced_folio(folio, vma, target_nid)) {
5218 nid = target_nid;
5219 flags |= TNF_MIGRATED;
5220 } else {
5221 flags |= TNF_MIGRATE_FAIL;
5222 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
5223 vmf->address, &vmf->ptl);
5224 if (unlikely(!vmf->pte))
5225 goto out;
5226 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
5227 pte_unmap_unlock(vmf->pte, vmf->ptl);
5228 goto out;
5229 }
5230 goto out_map;
5231 }
5232
5233 out:
5234 if (nid != NUMA_NO_NODE)
5235 task_numa_fault(last_cpupid, nid, nr_pages, flags);
5236 return 0;
5237 out_map:
5238 /*
5239 * Make it present again, depending on how arch implements
5240 * non-accessible ptes, some can allow access by kernel mode.
5241 */
5242 if (folio && folio_test_large(folio))
5243 numa_rebuild_large_mapping(vmf, vma, folio, pte, ignore_writable,
5244 pte_write_upgrade);
5245 else
5246 numa_rebuild_single_mapping(vmf, vma, vmf->address, vmf->pte,
5247 writable);
5248 pte_unmap_unlock(vmf->pte, vmf->ptl);
5249 goto out;
5250 }
5251
create_huge_pmd(struct vm_fault * vmf)5252 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
5253 {
5254 struct vm_area_struct *vma = vmf->vma;
5255 if (vma_is_anonymous(vma))
5256 return do_huge_pmd_anonymous_page(vmf);
5257 if (vma->vm_ops->huge_fault)
5258 return vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5259 return VM_FAULT_FALLBACK;
5260 }
5261
5262 /* `inline' is required to avoid gcc 4.1.2 build error */
wp_huge_pmd(struct vm_fault * vmf)5263 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
5264 {
5265 struct vm_area_struct *vma = vmf->vma;
5266 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
5267 vm_fault_t ret;
5268
5269 if (vma_is_anonymous(vma)) {
5270 if (likely(!unshare) &&
5271 userfaultfd_huge_pmd_wp(vma, vmf->orig_pmd)) {
5272 if (userfaultfd_wp_async(vmf->vma))
5273 goto split;
5274 return handle_userfault(vmf, VM_UFFD_WP);
5275 }
5276 return do_huge_pmd_wp_page(vmf);
5277 }
5278
5279 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5280 if (vma->vm_ops->huge_fault) {
5281 ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5282 if (!(ret & VM_FAULT_FALLBACK))
5283 return ret;
5284 }
5285 }
5286
5287 split:
5288 /* COW or write-notify handled on pte level: split pmd. */
5289 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
5290
5291 return VM_FAULT_FALLBACK;
5292 }
5293
create_huge_pud(struct vm_fault * vmf)5294 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
5295 {
5296 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5297 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5298 struct vm_area_struct *vma = vmf->vma;
5299 /* No support for anonymous transparent PUD pages yet */
5300 if (vma_is_anonymous(vma))
5301 return VM_FAULT_FALLBACK;
5302 if (vma->vm_ops->huge_fault)
5303 return vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5304 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5305 return VM_FAULT_FALLBACK;
5306 }
5307
wp_huge_pud(struct vm_fault * vmf,pud_t orig_pud)5308 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
5309 {
5310 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5311 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5312 struct vm_area_struct *vma = vmf->vma;
5313 vm_fault_t ret;
5314
5315 /* No support for anonymous transparent PUD pages yet */
5316 if (vma_is_anonymous(vma))
5317 goto split;
5318 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5319 if (vma->vm_ops->huge_fault) {
5320 ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5321 if (!(ret & VM_FAULT_FALLBACK))
5322 return ret;
5323 }
5324 }
5325 split:
5326 /* COW or write-notify not handled on PUD level: split pud.*/
5327 __split_huge_pud(vma, vmf->pud, vmf->address);
5328 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
5329 return VM_FAULT_FALLBACK;
5330 }
5331
5332 /*
5333 * These routines also need to handle stuff like marking pages dirty
5334 * and/or accessed for architectures that don't do it in hardware (most
5335 * RISC architectures). The early dirtying is also good on the i386.
5336 *
5337 * There is also a hook called "update_mmu_cache()" that architectures
5338 * with external mmu caches can use to update those (ie the Sparc or
5339 * PowerPC hashed page tables that act as extended TLBs).
5340 *
5341 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
5342 * concurrent faults).
5343 *
5344 * The mmap_lock may have been released depending on flags and our return value.
5345 * See filemap_fault() and __folio_lock_or_retry().
5346 */
handle_pte_fault(struct vm_fault * vmf)5347 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
5348 {
5349 pte_t entry;
5350
5351 if (unlikely(pmd_none(*vmf->pmd))) {
5352 /*
5353 * Leave __pte_alloc() until later: because vm_ops->fault may
5354 * want to allocate huge page, and if we expose page table
5355 * for an instant, it will be difficult to retract from
5356 * concurrent faults and from rmap lookups.
5357 */
5358 vmf->pte = NULL;
5359 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
5360 } else {
5361 /*
5362 * A regular pmd is established and it can't morph into a huge
5363 * pmd by anon khugepaged, since that takes mmap_lock in write
5364 * mode; but shmem or file collapse to THP could still morph
5365 * it into a huge pmd: just retry later if so.
5366 */
5367 vmf->pte = pte_offset_map_nolock(vmf->vma->vm_mm, vmf->pmd,
5368 vmf->address, &vmf->ptl);
5369 if (unlikely(!vmf->pte))
5370 return 0;
5371 vmf->orig_pte = ptep_get_lockless(vmf->pte);
5372 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
5373
5374 if (pte_none(vmf->orig_pte)) {
5375 pte_unmap(vmf->pte);
5376 vmf->pte = NULL;
5377 }
5378 }
5379
5380 if (!vmf->pte)
5381 return do_pte_missing(vmf);
5382
5383 if (!pte_present(vmf->orig_pte))
5384 return do_swap_page(vmf);
5385
5386 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
5387 return do_numa_page(vmf);
5388
5389 spin_lock(vmf->ptl);
5390 entry = vmf->orig_pte;
5391 if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) {
5392 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
5393 goto unlock;
5394 }
5395 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
5396 if (!pte_write(entry))
5397 return do_wp_page(vmf);
5398 else if (likely(vmf->flags & FAULT_FLAG_WRITE))
5399 entry = pte_mkdirty(entry);
5400 }
5401 entry = pte_mkyoung(entry);
5402 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
5403 vmf->flags & FAULT_FLAG_WRITE)) {
5404 update_mmu_cache_range(vmf, vmf->vma, vmf->address,
5405 vmf->pte, 1);
5406 } else {
5407 /* Skip spurious TLB flush for retried page fault */
5408 if (vmf->flags & FAULT_FLAG_TRIED)
5409 goto unlock;
5410 /*
5411 * This is needed only for protection faults but the arch code
5412 * is not yet telling us if this is a protection fault or not.
5413 * This still avoids useless tlb flushes for .text page faults
5414 * with threads.
5415 */
5416 if (vmf->flags & FAULT_FLAG_WRITE)
5417 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address,
5418 vmf->pte);
5419 }
5420 unlock:
5421 pte_unmap_unlock(vmf->pte, vmf->ptl);
5422 return 0;
5423 }
5424
5425 /*
5426 * On entry, we hold either the VMA lock or the mmap_lock
5427 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5428 * the result, the mmap_lock is not held on exit. See filemap_fault()
5429 * and __folio_lock_or_retry().
5430 */
__handle_mm_fault(struct vm_area_struct * vma,unsigned long address,unsigned int flags)5431 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
5432 unsigned long address, unsigned int flags)
5433 {
5434 struct vm_fault vmf = {
5435 .vma = vma,
5436 .address = address & PAGE_MASK,
5437 .real_address = address,
5438 .flags = flags,
5439 .pgoff = linear_page_index(vma, address),
5440 .gfp_mask = __get_fault_gfp_mask(vma),
5441 };
5442 struct mm_struct *mm = vma->vm_mm;
5443 unsigned long vm_flags = vma->vm_flags;
5444 pgd_t *pgd;
5445 p4d_t *p4d;
5446 vm_fault_t ret;
5447
5448 pgd = pgd_offset(mm, address);
5449 p4d = p4d_alloc(mm, pgd, address);
5450 if (!p4d)
5451 return VM_FAULT_OOM;
5452
5453 vmf.pud = pud_alloc(mm, p4d, address);
5454 if (!vmf.pud)
5455 return VM_FAULT_OOM;
5456 retry_pud:
5457 if (pud_none(*vmf.pud) &&
5458 thp_vma_allowable_order(vma, vm_flags,
5459 TVA_IN_PF | TVA_ENFORCE_SYSFS, PUD_ORDER)) {
5460 ret = create_huge_pud(&vmf);
5461 if (!(ret & VM_FAULT_FALLBACK))
5462 return ret;
5463 } else {
5464 pud_t orig_pud = *vmf.pud;
5465
5466 barrier();
5467 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
5468
5469 /*
5470 * TODO once we support anonymous PUDs: NUMA case and
5471 * FAULT_FLAG_UNSHARE handling.
5472 */
5473 if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) {
5474 ret = wp_huge_pud(&vmf, orig_pud);
5475 if (!(ret & VM_FAULT_FALLBACK))
5476 return ret;
5477 } else {
5478 huge_pud_set_accessed(&vmf, orig_pud);
5479 return 0;
5480 }
5481 }
5482 }
5483
5484 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
5485 if (!vmf.pmd)
5486 return VM_FAULT_OOM;
5487
5488 /* Huge pud page fault raced with pmd_alloc? */
5489 if (pud_trans_unstable(vmf.pud))
5490 goto retry_pud;
5491
5492 if (pmd_none(*vmf.pmd) &&
5493 thp_vma_allowable_order(vma, vm_flags,
5494 TVA_IN_PF | TVA_ENFORCE_SYSFS, PMD_ORDER)) {
5495 ret = create_huge_pmd(&vmf);
5496 if (!(ret & VM_FAULT_FALLBACK))
5497 return ret;
5498 } else {
5499 vmf.orig_pmd = pmdp_get_lockless(vmf.pmd);
5500
5501 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
5502 VM_BUG_ON(thp_migration_supported() &&
5503 !is_pmd_migration_entry(vmf.orig_pmd));
5504 if (is_pmd_migration_entry(vmf.orig_pmd))
5505 pmd_migration_entry_wait(mm, vmf.pmd);
5506 return 0;
5507 }
5508 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
5509 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
5510 return do_huge_pmd_numa_page(&vmf);
5511
5512 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5513 !pmd_write(vmf.orig_pmd)) {
5514 ret = wp_huge_pmd(&vmf);
5515 if (!(ret & VM_FAULT_FALLBACK))
5516 return ret;
5517 } else {
5518 huge_pmd_set_accessed(&vmf);
5519 return 0;
5520 }
5521 }
5522 }
5523
5524 return handle_pte_fault(&vmf);
5525 }
5526
5527 /**
5528 * mm_account_fault - Do page fault accounting
5529 * @mm: mm from which memcg should be extracted. It can be NULL.
5530 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5531 * of perf event counters, but we'll still do the per-task accounting to
5532 * the task who triggered this page fault.
5533 * @address: the faulted address.
5534 * @flags: the fault flags.
5535 * @ret: the fault retcode.
5536 *
5537 * This will take care of most of the page fault accounting. Meanwhile, it
5538 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5539 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5540 * still be in per-arch page fault handlers at the entry of page fault.
5541 */
mm_account_fault(struct mm_struct * mm,struct pt_regs * regs,unsigned long address,unsigned int flags,vm_fault_t ret)5542 static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs,
5543 unsigned long address, unsigned int flags,
5544 vm_fault_t ret)
5545 {
5546 bool major;
5547
5548 /* Incomplete faults will be accounted upon completion. */
5549 if (ret & VM_FAULT_RETRY)
5550 return;
5551
5552 /*
5553 * To preserve the behavior of older kernels, PGFAULT counters record
5554 * both successful and failed faults, as opposed to perf counters,
5555 * which ignore failed cases.
5556 */
5557 count_vm_event(PGFAULT);
5558 count_memcg_event_mm(mm, PGFAULT);
5559
5560 /*
5561 * Do not account for unsuccessful faults (e.g. when the address wasn't
5562 * valid). That includes arch_vma_access_permitted() failing before
5563 * reaching here. So this is not a "this many hardware page faults"
5564 * counter. We should use the hw profiling for that.
5565 */
5566 if (ret & VM_FAULT_ERROR)
5567 return;
5568
5569 /*
5570 * We define the fault as a major fault when the final successful fault
5571 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5572 * handle it immediately previously).
5573 */
5574 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
5575
5576 if (major)
5577 current->maj_flt++;
5578 else
5579 current->min_flt++;
5580
5581 /*
5582 * If the fault is done for GUP, regs will be NULL. We only do the
5583 * accounting for the per thread fault counters who triggered the
5584 * fault, and we skip the perf event updates.
5585 */
5586 if (!regs)
5587 return;
5588
5589 if (major)
5590 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
5591 else
5592 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
5593 }
5594
5595 #ifdef CONFIG_LRU_GEN
lru_gen_enter_fault(struct vm_area_struct * vma)5596 static void lru_gen_enter_fault(struct vm_area_struct *vma)
5597 {
5598 /* the LRU algorithm only applies to accesses with recency */
5599 current->in_lru_fault = vma_has_recency(vma);
5600 }
5601
lru_gen_exit_fault(void)5602 static void lru_gen_exit_fault(void)
5603 {
5604 current->in_lru_fault = false;
5605 }
5606 #else
lru_gen_enter_fault(struct vm_area_struct * vma)5607 static void lru_gen_enter_fault(struct vm_area_struct *vma)
5608 {
5609 }
5610
lru_gen_exit_fault(void)5611 static void lru_gen_exit_fault(void)
5612 {
5613 }
5614 #endif /* CONFIG_LRU_GEN */
5615
sanitize_fault_flags(struct vm_area_struct * vma,unsigned int * flags)5616 static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
5617 unsigned int *flags)
5618 {
5619 if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
5620 if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
5621 return VM_FAULT_SIGSEGV;
5622 /*
5623 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5624 * just treat it like an ordinary read-fault otherwise.
5625 */
5626 if (!is_cow_mapping(vma->vm_flags))
5627 *flags &= ~FAULT_FLAG_UNSHARE;
5628 } else if (*flags & FAULT_FLAG_WRITE) {
5629 /* Write faults on read-only mappings are impossible ... */
5630 if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
5631 return VM_FAULT_SIGSEGV;
5632 /* ... and FOLL_FORCE only applies to COW mappings. */
5633 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
5634 !is_cow_mapping(vma->vm_flags)))
5635 return VM_FAULT_SIGSEGV;
5636 }
5637 #ifdef CONFIG_PER_VMA_LOCK
5638 /*
5639 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5640 * the assumption that lock is dropped on VM_FAULT_RETRY.
5641 */
5642 if (WARN_ON_ONCE((*flags &
5643 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) ==
5644 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)))
5645 return VM_FAULT_SIGSEGV;
5646 #endif
5647
5648 return 0;
5649 }
5650
5651 /*
5652 * By the time we get here, we already hold the mm semaphore
5653 *
5654 * The mmap_lock may have been released depending on flags and our
5655 * return value. See filemap_fault() and __folio_lock_or_retry().
5656 */
handle_mm_fault(struct vm_area_struct * vma,unsigned long address,unsigned int flags,struct pt_regs * regs)5657 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
5658 unsigned int flags, struct pt_regs *regs)
5659 {
5660 /* If the fault handler drops the mmap_lock, vma may be freed */
5661 struct mm_struct *mm = vma->vm_mm;
5662 vm_fault_t ret;
5663
5664 __set_current_state(TASK_RUNNING);
5665
5666 ret = sanitize_fault_flags(vma, &flags);
5667 if (ret)
5668 goto out;
5669
5670 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
5671 flags & FAULT_FLAG_INSTRUCTION,
5672 flags & FAULT_FLAG_REMOTE)) {
5673 ret = VM_FAULT_SIGSEGV;
5674 goto out;
5675 }
5676
5677 /*
5678 * Enable the memcg OOM handling for faults triggered in user
5679 * space. Kernel faults are handled more gracefully.
5680 */
5681 if (flags & FAULT_FLAG_USER)
5682 mem_cgroup_enter_user_fault();
5683
5684 lru_gen_enter_fault(vma);
5685
5686 if (unlikely(is_vm_hugetlb_page(vma)))
5687 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
5688 else
5689 ret = __handle_mm_fault(vma, address, flags);
5690
5691 lru_gen_exit_fault();
5692
5693 if (flags & FAULT_FLAG_USER) {
5694 mem_cgroup_exit_user_fault();
5695 /*
5696 * The task may have entered a memcg OOM situation but
5697 * if the allocation error was handled gracefully (no
5698 * VM_FAULT_OOM), there is no need to kill anything.
5699 * Just clean up the OOM state peacefully.
5700 */
5701 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
5702 mem_cgroup_oom_synchronize(false);
5703 }
5704 out:
5705 mm_account_fault(mm, regs, address, flags, ret);
5706
5707 return ret;
5708 }
5709 EXPORT_SYMBOL_GPL(handle_mm_fault);
5710
5711 #ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5712 #include <linux/extable.h>
5713
get_mmap_lock_carefully(struct mm_struct * mm,struct pt_regs * regs)5714 static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5715 {
5716 if (likely(mmap_read_trylock(mm)))
5717 return true;
5718
5719 if (regs && !user_mode(regs)) {
5720 unsigned long ip = exception_ip(regs);
5721 if (!search_exception_tables(ip))
5722 return false;
5723 }
5724
5725 return !mmap_read_lock_killable(mm);
5726 }
5727
mmap_upgrade_trylock(struct mm_struct * mm)5728 static inline bool mmap_upgrade_trylock(struct mm_struct *mm)
5729 {
5730 /*
5731 * We don't have this operation yet.
5732 *
5733 * It should be easy enough to do: it's basically a
5734 * atomic_long_try_cmpxchg_acquire()
5735 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5736 * it also needs the proper lockdep magic etc.
5737 */
5738 return false;
5739 }
5740
upgrade_mmap_lock_carefully(struct mm_struct * mm,struct pt_regs * regs)5741 static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5742 {
5743 mmap_read_unlock(mm);
5744 if (regs && !user_mode(regs)) {
5745 unsigned long ip = exception_ip(regs);
5746 if (!search_exception_tables(ip))
5747 return false;
5748 }
5749 return !mmap_write_lock_killable(mm);
5750 }
5751
5752 /*
5753 * Helper for page fault handling.
5754 *
5755 * This is kind of equivalend to "mmap_read_lock()" followed
5756 * by "find_extend_vma()", except it's a lot more careful about
5757 * the locking (and will drop the lock on failure).
5758 *
5759 * For example, if we have a kernel bug that causes a page
5760 * fault, we don't want to just use mmap_read_lock() to get
5761 * the mm lock, because that would deadlock if the bug were
5762 * to happen while we're holding the mm lock for writing.
5763 *
5764 * So this checks the exception tables on kernel faults in
5765 * order to only do this all for instructions that are actually
5766 * expected to fault.
5767 *
5768 * We can also actually take the mm lock for writing if we
5769 * need to extend the vma, which helps the VM layer a lot.
5770 */
lock_mm_and_find_vma(struct mm_struct * mm,unsigned long addr,struct pt_regs * regs)5771 struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm,
5772 unsigned long addr, struct pt_regs *regs)
5773 {
5774 struct vm_area_struct *vma;
5775
5776 if (!get_mmap_lock_carefully(mm, regs))
5777 return NULL;
5778
5779 vma = find_vma(mm, addr);
5780 if (likely(vma && (vma->vm_start <= addr)))
5781 return vma;
5782
5783 /*
5784 * Well, dang. We might still be successful, but only
5785 * if we can extend a vma to do so.
5786 */
5787 if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) {
5788 mmap_read_unlock(mm);
5789 return NULL;
5790 }
5791
5792 /*
5793 * We can try to upgrade the mmap lock atomically,
5794 * in which case we can continue to use the vma
5795 * we already looked up.
5796 *
5797 * Otherwise we'll have to drop the mmap lock and
5798 * re-take it, and also look up the vma again,
5799 * re-checking it.
5800 */
5801 if (!mmap_upgrade_trylock(mm)) {
5802 if (!upgrade_mmap_lock_carefully(mm, regs))
5803 return NULL;
5804
5805 vma = find_vma(mm, addr);
5806 if (!vma)
5807 goto fail;
5808 if (vma->vm_start <= addr)
5809 goto success;
5810 if (!(vma->vm_flags & VM_GROWSDOWN))
5811 goto fail;
5812 }
5813
5814 if (expand_stack_locked(vma, addr))
5815 goto fail;
5816
5817 success:
5818 mmap_write_downgrade(mm);
5819 return vma;
5820
5821 fail:
5822 mmap_write_unlock(mm);
5823 return NULL;
5824 }
5825 #endif
5826
5827 #ifdef CONFIG_PER_VMA_LOCK
5828 /*
5829 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5830 * stable and not isolated. If the VMA is not found or is being modified the
5831 * function returns NULL.
5832 */
lock_vma_under_rcu(struct mm_struct * mm,unsigned long address)5833 struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
5834 unsigned long address)
5835 {
5836 MA_STATE(mas, &mm->mm_mt, address, address);
5837 struct vm_area_struct *vma;
5838
5839 rcu_read_lock();
5840 retry:
5841 vma = mas_walk(&mas);
5842 if (!vma)
5843 goto inval;
5844
5845 if (!vma_start_read(vma))
5846 goto inval;
5847
5848 /* Check since vm_start/vm_end might change before we lock the VMA */
5849 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5850 goto inval_end_read;
5851
5852 /* Check if the VMA got isolated after we found it */
5853 if (vma->detached) {
5854 vma_end_read(vma);
5855 count_vm_vma_lock_event(VMA_LOCK_MISS);
5856 /* The area was replaced with another one */
5857 goto retry;
5858 }
5859
5860 rcu_read_unlock();
5861 return vma;
5862
5863 inval_end_read:
5864 vma_end_read(vma);
5865 inval:
5866 rcu_read_unlock();
5867 count_vm_vma_lock_event(VMA_LOCK_ABORT);
5868 return NULL;
5869 }
5870 #endif /* CONFIG_PER_VMA_LOCK */
5871
5872 #ifndef __PAGETABLE_P4D_FOLDED
5873 /*
5874 * Allocate p4d page table.
5875 * We've already handled the fast-path in-line.
5876 */
__p4d_alloc(struct mm_struct * mm,pgd_t * pgd,unsigned long address)5877 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
5878 {
5879 p4d_t *new = p4d_alloc_one(mm, address);
5880 if (!new)
5881 return -ENOMEM;
5882
5883 spin_lock(&mm->page_table_lock);
5884 if (pgd_present(*pgd)) { /* Another has populated it */
5885 p4d_free(mm, new);
5886 } else {
5887 smp_wmb(); /* See comment in pmd_install() */
5888 pgd_populate(mm, pgd, new);
5889 }
5890 spin_unlock(&mm->page_table_lock);
5891 return 0;
5892 }
5893 #endif /* __PAGETABLE_P4D_FOLDED */
5894
5895 #ifndef __PAGETABLE_PUD_FOLDED
5896 /*
5897 * Allocate page upper directory.
5898 * We've already handled the fast-path in-line.
5899 */
__pud_alloc(struct mm_struct * mm,p4d_t * p4d,unsigned long address)5900 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
5901 {
5902 pud_t *new = pud_alloc_one(mm, address);
5903 if (!new)
5904 return -ENOMEM;
5905
5906 spin_lock(&mm->page_table_lock);
5907 if (!p4d_present(*p4d)) {
5908 mm_inc_nr_puds(mm);
5909 smp_wmb(); /* See comment in pmd_install() */
5910 p4d_populate(mm, p4d, new);
5911 } else /* Another has populated it */
5912 pud_free(mm, new);
5913 spin_unlock(&mm->page_table_lock);
5914 return 0;
5915 }
5916 #endif /* __PAGETABLE_PUD_FOLDED */
5917
5918 #ifndef __PAGETABLE_PMD_FOLDED
5919 /*
5920 * Allocate page middle directory.
5921 * We've already handled the fast-path in-line.
5922 */
__pmd_alloc(struct mm_struct * mm,pud_t * pud,unsigned long address)5923 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
5924 {
5925 spinlock_t *ptl;
5926 pmd_t *new = pmd_alloc_one(mm, address);
5927 if (!new)
5928 return -ENOMEM;
5929
5930 ptl = pud_lock(mm, pud);
5931 if (!pud_present(*pud)) {
5932 mm_inc_nr_pmds(mm);
5933 smp_wmb(); /* See comment in pmd_install() */
5934 pud_populate(mm, pud, new);
5935 } else { /* Another has populated it */
5936 pmd_free(mm, new);
5937 }
5938 spin_unlock(ptl);
5939 return 0;
5940 }
5941 #endif /* __PAGETABLE_PMD_FOLDED */
5942
5943 /**
5944 * follow_pte - look up PTE at a user virtual address
5945 * @vma: the memory mapping
5946 * @address: user virtual address
5947 * @ptepp: location to store found PTE
5948 * @ptlp: location to store the lock for the PTE
5949 *
5950 * On a successful return, the pointer to the PTE is stored in @ptepp;
5951 * the corresponding lock is taken and its location is stored in @ptlp.
5952 *
5953 * The contents of the PTE are only stable until @ptlp is released using
5954 * pte_unmap_unlock(). This function will fail if the PTE is non-present.
5955 * Present PTEs may include PTEs that map refcounted pages, such as
5956 * anonymous folios in COW mappings.
5957 *
5958 * Callers must be careful when relying on PTE content after
5959 * pte_unmap_unlock(). Especially if the PTE maps a refcounted page,
5960 * callers must protect against invalidation with MMU notifiers; otherwise
5961 * access to the PFN at a later point in time can trigger use-after-free.
5962 *
5963 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5964 * should be taken for read.
5965 *
5966 * This function must not be used to modify PTE content.
5967 *
5968 * Return: zero on success, -ve otherwise.
5969 */
follow_pte(struct vm_area_struct * vma,unsigned long address,pte_t ** ptepp,spinlock_t ** ptlp)5970 int follow_pte(struct vm_area_struct *vma, unsigned long address,
5971 pte_t **ptepp, spinlock_t **ptlp)
5972 {
5973 struct mm_struct *mm = vma->vm_mm;
5974 pgd_t *pgd;
5975 p4d_t *p4d;
5976 pud_t *pud;
5977 pmd_t *pmd;
5978 pte_t *ptep;
5979
5980 mmap_assert_locked(mm);
5981 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5982 goto out;
5983
5984 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5985 goto out;
5986
5987 pgd = pgd_offset(mm, address);
5988 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
5989 goto out;
5990
5991 p4d = p4d_offset(pgd, address);
5992 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
5993 goto out;
5994
5995 pud = pud_offset(p4d, address);
5996 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
5997 goto out;
5998
5999 pmd = pmd_offset(pud, address);
6000 VM_BUG_ON(pmd_trans_huge(*pmd));
6001
6002 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
6003 if (!ptep)
6004 goto out;
6005 if (!pte_present(ptep_get(ptep)))
6006 goto unlock;
6007 *ptepp = ptep;
6008 return 0;
6009 unlock:
6010 pte_unmap_unlock(ptep, *ptlp);
6011 out:
6012 return -EINVAL;
6013 }
6014 EXPORT_SYMBOL_GPL(follow_pte);
6015
6016 #ifdef CONFIG_HAVE_IOREMAP_PROT
6017 /**
6018 * generic_access_phys - generic implementation for iomem mmap access
6019 * @vma: the vma to access
6020 * @addr: userspace address, not relative offset within @vma
6021 * @buf: buffer to read/write
6022 * @len: length of transfer
6023 * @write: set to FOLL_WRITE when writing, otherwise reading
6024 *
6025 * This is a generic implementation for &vm_operations_struct.access for an
6026 * iomem mapping. This callback is used by access_process_vm() when the @vma is
6027 * not page based.
6028 */
generic_access_phys(struct vm_area_struct * vma,unsigned long addr,void * buf,int len,int write)6029 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
6030 void *buf, int len, int write)
6031 {
6032 resource_size_t phys_addr;
6033 unsigned long prot = 0;
6034 void __iomem *maddr;
6035 pte_t *ptep, pte;
6036 spinlock_t *ptl;
6037 int offset = offset_in_page(addr);
6038 int ret = -EINVAL;
6039
6040 retry:
6041 if (follow_pte(vma, addr, &ptep, &ptl))
6042 return -EINVAL;
6043 pte = ptep_get(ptep);
6044 pte_unmap_unlock(ptep, ptl);
6045
6046 prot = pgprot_val(pte_pgprot(pte));
6047 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
6048
6049 if ((write & FOLL_WRITE) && !pte_write(pte))
6050 return -EINVAL;
6051
6052 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
6053 if (!maddr)
6054 return -ENOMEM;
6055
6056 if (follow_pte(vma, addr, &ptep, &ptl))
6057 goto out_unmap;
6058
6059 if (!pte_same(pte, ptep_get(ptep))) {
6060 pte_unmap_unlock(ptep, ptl);
6061 iounmap(maddr);
6062
6063 goto retry;
6064 }
6065
6066 if (write)
6067 memcpy_toio(maddr + offset, buf, len);
6068 else
6069 memcpy_fromio(buf, maddr + offset, len);
6070 ret = len;
6071 pte_unmap_unlock(ptep, ptl);
6072 out_unmap:
6073 iounmap(maddr);
6074
6075 return ret;
6076 }
6077 EXPORT_SYMBOL_GPL(generic_access_phys);
6078 #endif
6079
6080 /*
6081 * Access another process' address space as given in mm.
6082 */
__access_remote_vm(struct mm_struct * mm,unsigned long addr,void * buf,int len,unsigned int gup_flags)6083 static int __access_remote_vm(struct mm_struct *mm, unsigned long addr,
6084 void *buf, int len, unsigned int gup_flags)
6085 {
6086 void *old_buf = buf;
6087 int write = gup_flags & FOLL_WRITE;
6088
6089 if (mmap_read_lock_killable(mm))
6090 return 0;
6091
6092 /* Untag the address before looking up the VMA */
6093 addr = untagged_addr_remote(mm, addr);
6094
6095 /* Avoid triggering the temporary warning in __get_user_pages */
6096 if (!vma_lookup(mm, addr) && !expand_stack(mm, addr))
6097 return 0;
6098
6099 /* ignore errors, just check how much was successfully transferred */
6100 while (len) {
6101 int bytes, offset;
6102 void *maddr;
6103 struct vm_area_struct *vma = NULL;
6104 struct page *page = get_user_page_vma_remote(mm, addr,
6105 gup_flags, &vma);
6106
6107 if (IS_ERR(page)) {
6108 /* We might need to expand the stack to access it */
6109 vma = vma_lookup(mm, addr);
6110 if (!vma) {
6111 vma = expand_stack(mm, addr);
6112
6113 /* mmap_lock was dropped on failure */
6114 if (!vma)
6115 return buf - old_buf;
6116
6117 /* Try again if stack expansion worked */
6118 continue;
6119 }
6120
6121 /*
6122 * Check if this is a VM_IO | VM_PFNMAP VMA, which
6123 * we can access using slightly different code.
6124 */
6125 bytes = 0;
6126 #ifdef CONFIG_HAVE_IOREMAP_PROT
6127 if (vma->vm_ops && vma->vm_ops->access)
6128 bytes = vma->vm_ops->access(vma, addr, buf,
6129 len, write);
6130 #endif
6131 if (bytes <= 0)
6132 break;
6133 } else {
6134 bytes = len;
6135 offset = addr & (PAGE_SIZE-1);
6136 if (bytes > PAGE_SIZE-offset)
6137 bytes = PAGE_SIZE-offset;
6138
6139 maddr = kmap_local_page(page);
6140 if (write) {
6141 copy_to_user_page(vma, page, addr,
6142 maddr + offset, buf, bytes);
6143 set_page_dirty_lock(page);
6144 } else {
6145 copy_from_user_page(vma, page, addr,
6146 buf, maddr + offset, bytes);
6147 }
6148 unmap_and_put_page(page, maddr);
6149 }
6150 len -= bytes;
6151 buf += bytes;
6152 addr += bytes;
6153 }
6154 mmap_read_unlock(mm);
6155
6156 return buf - old_buf;
6157 }
6158
6159 /**
6160 * access_remote_vm - access another process' address space
6161 * @mm: the mm_struct of the target address space
6162 * @addr: start address to access
6163 * @buf: source or destination buffer
6164 * @len: number of bytes to transfer
6165 * @gup_flags: flags modifying lookup behaviour
6166 *
6167 * The caller must hold a reference on @mm.
6168 *
6169 * Return: number of bytes copied from source to destination.
6170 */
access_remote_vm(struct mm_struct * mm,unsigned long addr,void * buf,int len,unsigned int gup_flags)6171 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
6172 void *buf, int len, unsigned int gup_flags)
6173 {
6174 return __access_remote_vm(mm, addr, buf, len, gup_flags);
6175 }
6176
6177 /*
6178 * Access another process' address space.
6179 * Source/target buffer must be kernel space,
6180 * Do not walk the page table directly, use get_user_pages
6181 */
access_process_vm(struct task_struct * tsk,unsigned long addr,void * buf,int len,unsigned int gup_flags)6182 int access_process_vm(struct task_struct *tsk, unsigned long addr,
6183 void *buf, int len, unsigned int gup_flags)
6184 {
6185 struct mm_struct *mm;
6186 int ret;
6187
6188 mm = get_task_mm(tsk);
6189 if (!mm)
6190 return 0;
6191
6192 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
6193
6194 mmput(mm);
6195
6196 return ret;
6197 }
6198 EXPORT_SYMBOL_GPL(access_process_vm);
6199
6200 /*
6201 * Print the name of a VMA.
6202 */
print_vma_addr(char * prefix,unsigned long ip)6203 void print_vma_addr(char *prefix, unsigned long ip)
6204 {
6205 struct mm_struct *mm = current->mm;
6206 struct vm_area_struct *vma;
6207
6208 /*
6209 * we might be running from an atomic context so we cannot sleep
6210 */
6211 if (!mmap_read_trylock(mm))
6212 return;
6213
6214 vma = vma_lookup(mm, ip);
6215 if (vma && vma->vm_file) {
6216 struct file *f = vma->vm_file;
6217 ip -= vma->vm_start;
6218 ip += vma->vm_pgoff << PAGE_SHIFT;
6219 printk("%s%pD[%lx,%lx+%lx]", prefix, f, ip,
6220 vma->vm_start,
6221 vma->vm_end - vma->vm_start);
6222 }
6223 mmap_read_unlock(mm);
6224 }
6225
6226 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
__might_fault(const char * file,int line)6227 void __might_fault(const char *file, int line)
6228 {
6229 if (pagefault_disabled())
6230 return;
6231 __might_sleep(file, line);
6232 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6233 if (current->mm)
6234 might_lock_read(¤t->mm->mmap_lock);
6235 #endif
6236 }
6237 EXPORT_SYMBOL(__might_fault);
6238 #endif
6239
6240 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
6241 /*
6242 * Process all subpages of the specified huge page with the specified
6243 * operation. The target subpage will be processed last to keep its
6244 * cache lines hot.
6245 */
process_huge_page(unsigned long addr_hint,unsigned int pages_per_huge_page,int (* process_subpage)(unsigned long addr,int idx,void * arg),void * arg)6246 static inline int process_huge_page(
6247 unsigned long addr_hint, unsigned int pages_per_huge_page,
6248 int (*process_subpage)(unsigned long addr, int idx, void *arg),
6249 void *arg)
6250 {
6251 int i, n, base, l, ret;
6252 unsigned long addr = addr_hint &
6253 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6254
6255 /* Process target subpage last to keep its cache lines hot */
6256 might_sleep();
6257 n = (addr_hint - addr) / PAGE_SIZE;
6258 if (2 * n <= pages_per_huge_page) {
6259 /* If target subpage in first half of huge page */
6260 base = 0;
6261 l = n;
6262 /* Process subpages at the end of huge page */
6263 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
6264 cond_resched();
6265 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6266 if (ret)
6267 return ret;
6268 }
6269 } else {
6270 /* If target subpage in second half of huge page */
6271 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
6272 l = pages_per_huge_page - n;
6273 /* Process subpages at the begin of huge page */
6274 for (i = 0; i < base; i++) {
6275 cond_resched();
6276 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6277 if (ret)
6278 return ret;
6279 }
6280 }
6281 /*
6282 * Process remaining subpages in left-right-left-right pattern
6283 * towards the target subpage
6284 */
6285 for (i = 0; i < l; i++) {
6286 int left_idx = base + i;
6287 int right_idx = base + 2 * l - 1 - i;
6288
6289 cond_resched();
6290 ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
6291 if (ret)
6292 return ret;
6293 cond_resched();
6294 ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
6295 if (ret)
6296 return ret;
6297 }
6298 return 0;
6299 }
6300
clear_gigantic_page(struct page * page,unsigned long addr,unsigned int pages_per_huge_page)6301 static void clear_gigantic_page(struct page *page,
6302 unsigned long addr,
6303 unsigned int pages_per_huge_page)
6304 {
6305 int i;
6306 struct page *p;
6307
6308 might_sleep();
6309 for (i = 0; i < pages_per_huge_page; i++) {
6310 p = nth_page(page, i);
6311 cond_resched();
6312 clear_user_highpage(p, addr + i * PAGE_SIZE);
6313 }
6314 }
6315
clear_subpage(unsigned long addr,int idx,void * arg)6316 static int clear_subpage(unsigned long addr, int idx, void *arg)
6317 {
6318 struct page *page = arg;
6319
6320 clear_user_highpage(nth_page(page, idx), addr);
6321 return 0;
6322 }
6323
clear_huge_page(struct page * page,unsigned long addr_hint,unsigned int pages_per_huge_page)6324 void clear_huge_page(struct page *page,
6325 unsigned long addr_hint, unsigned int pages_per_huge_page)
6326 {
6327 unsigned long addr = addr_hint &
6328 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6329
6330 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
6331 clear_gigantic_page(page, addr, pages_per_huge_page);
6332 return;
6333 }
6334
6335 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
6336 }
6337
copy_user_gigantic_page(struct folio * dst,struct folio * src,unsigned long addr,struct vm_area_struct * vma,unsigned int pages_per_huge_page)6338 static int copy_user_gigantic_page(struct folio *dst, struct folio *src,
6339 unsigned long addr,
6340 struct vm_area_struct *vma,
6341 unsigned int pages_per_huge_page)
6342 {
6343 int i;
6344 struct page *dst_page;
6345 struct page *src_page;
6346
6347 for (i = 0; i < pages_per_huge_page; i++) {
6348 dst_page = folio_page(dst, i);
6349 src_page = folio_page(src, i);
6350
6351 cond_resched();
6352 if (copy_mc_user_highpage(dst_page, src_page,
6353 addr + i*PAGE_SIZE, vma)) {
6354 memory_failure_queue(page_to_pfn(src_page), 0);
6355 return -EHWPOISON;
6356 }
6357 }
6358 return 0;
6359 }
6360
6361 struct copy_subpage_arg {
6362 struct page *dst;
6363 struct page *src;
6364 struct vm_area_struct *vma;
6365 };
6366
copy_subpage(unsigned long addr,int idx,void * arg)6367 static int copy_subpage(unsigned long addr, int idx, void *arg)
6368 {
6369 struct copy_subpage_arg *copy_arg = arg;
6370 struct page *dst = nth_page(copy_arg->dst, idx);
6371 struct page *src = nth_page(copy_arg->src, idx);
6372
6373 if (copy_mc_user_highpage(dst, src, addr, copy_arg->vma)) {
6374 memory_failure_queue(page_to_pfn(src), 0);
6375 return -EHWPOISON;
6376 }
6377 return 0;
6378 }
6379
copy_user_large_folio(struct folio * dst,struct folio * src,unsigned long addr_hint,struct vm_area_struct * vma)6380 int copy_user_large_folio(struct folio *dst, struct folio *src,
6381 unsigned long addr_hint, struct vm_area_struct *vma)
6382 {
6383 unsigned int pages_per_huge_page = folio_nr_pages(dst);
6384 unsigned long addr = addr_hint &
6385 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6386 struct copy_subpage_arg arg = {
6387 .dst = &dst->page,
6388 .src = &src->page,
6389 .vma = vma,
6390 };
6391
6392 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES))
6393 return copy_user_gigantic_page(dst, src, addr, vma,
6394 pages_per_huge_page);
6395
6396 return process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
6397 }
6398
copy_folio_from_user(struct folio * dst_folio,const void __user * usr_src,bool allow_pagefault)6399 long copy_folio_from_user(struct folio *dst_folio,
6400 const void __user *usr_src,
6401 bool allow_pagefault)
6402 {
6403 void *kaddr;
6404 unsigned long i, rc = 0;
6405 unsigned int nr_pages = folio_nr_pages(dst_folio);
6406 unsigned long ret_val = nr_pages * PAGE_SIZE;
6407 struct page *subpage;
6408
6409 for (i = 0; i < nr_pages; i++) {
6410 subpage = folio_page(dst_folio, i);
6411 kaddr = kmap_local_page(subpage);
6412 if (!allow_pagefault)
6413 pagefault_disable();
6414 rc = copy_from_user(kaddr, usr_src + i * PAGE_SIZE, PAGE_SIZE);
6415 if (!allow_pagefault)
6416 pagefault_enable();
6417 kunmap_local(kaddr);
6418
6419 ret_val -= (PAGE_SIZE - rc);
6420 if (rc)
6421 break;
6422
6423 flush_dcache_page(subpage);
6424
6425 cond_resched();
6426 }
6427 return ret_val;
6428 }
6429 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6430
6431 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6432
6433 static struct kmem_cache *page_ptl_cachep;
6434
ptlock_cache_init(void)6435 void __init ptlock_cache_init(void)
6436 {
6437 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
6438 SLAB_PANIC, NULL);
6439 }
6440
ptlock_alloc(struct ptdesc * ptdesc)6441 bool ptlock_alloc(struct ptdesc *ptdesc)
6442 {
6443 spinlock_t *ptl;
6444
6445 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
6446 if (!ptl)
6447 return false;
6448 ptdesc->ptl = ptl;
6449 return true;
6450 }
6451
ptlock_free(struct ptdesc * ptdesc)6452 void ptlock_free(struct ptdesc *ptdesc)
6453 {
6454 kmem_cache_free(page_ptl_cachep, ptdesc->ptl);
6455 }
6456 #endif
6457
vma_pgtable_walk_begin(struct vm_area_struct * vma)6458 void vma_pgtable_walk_begin(struct vm_area_struct *vma)
6459 {
6460 if (is_vm_hugetlb_page(vma))
6461 hugetlb_vma_lock_read(vma);
6462 }
6463
vma_pgtable_walk_end(struct vm_area_struct * vma)6464 void vma_pgtable_walk_end(struct vm_area_struct *vma)
6465 {
6466 if (is_vm_hugetlb_page(vma))
6467 hugetlb_vma_unlock_read(vma);
6468 }
6469