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