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