Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * mm/mprotect.c
4 *
5 * (C) Copyright 1994 Linus Torvalds
6 * (C) Copyright 2002 Christoph Hellwig
7 *
8 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
10 */
11
12#include <linux/pagewalk.h>
13#include <linux/hugetlb.h>
14#include <linux/shm.h>
15#include <linux/mman.h>
16#include <linux/fs.h>
17#include <linux/highmem.h>
18#include <linux/security.h>
19#include <linux/mempolicy.h>
20#include <linux/personality.h>
21#include <linux/syscalls.h>
22#include <linux/swap.h>
23#include <linux/swapops.h>
24#include <linux/mmu_notifier.h>
25#include <linux/migrate.h>
26#include <linux/perf_event.h>
27#include <linux/pkeys.h>
28#include <linux/ksm.h>
29#include <linux/uaccess.h>
30#include <linux/mm_inline.h>
31#include <linux/pgtable.h>
32#include <asm/cacheflush.h>
33#include <asm/mmu_context.h>
34#include <asm/tlbflush.h>
35
36#include "internal.h"
37
38static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
39 unsigned long addr, unsigned long end, pgprot_t newprot,
40 unsigned long cp_flags)
41{
42 pte_t *pte, oldpte;
43 spinlock_t *ptl;
44 unsigned long pages = 0;
45 int target_node = NUMA_NO_NODE;
46 bool dirty_accountable = cp_flags & MM_CP_DIRTY_ACCT;
47 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
48 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
49 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
50
51 /*
52 * Can be called with only the mmap_lock for reading by
53 * prot_numa so we must check the pmd isn't constantly
54 * changing from under us from pmd_none to pmd_trans_huge
55 * and/or the other way around.
56 */
57 if (pmd_trans_unstable(pmd))
58 return 0;
59
60 /*
61 * The pmd points to a regular pte so the pmd can't change
62 * from under us even if the mmap_lock is only hold for
63 * reading.
64 */
65 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
66
67 /* Get target node for single threaded private VMAs */
68 if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
69 atomic_read(&vma->vm_mm->mm_users) == 1)
70 target_node = numa_node_id();
71
72 flush_tlb_batched_pending(vma->vm_mm);
73 arch_enter_lazy_mmu_mode();
74 do {
75 oldpte = *pte;
76 if (pte_present(oldpte)) {
77 pte_t ptent;
78 bool preserve_write = prot_numa && pte_write(oldpte);
79
80 /*
81 * Avoid trapping faults against the zero or KSM
82 * pages. See similar comment in change_huge_pmd.
83 */
84 if (prot_numa) {
85 struct page *page;
86
87 /* Avoid TLB flush if possible */
88 if (pte_protnone(oldpte))
89 continue;
90
91 page = vm_normal_page(vma, addr, oldpte);
92 if (!page || PageKsm(page))
93 continue;
94
95 /* Also skip shared copy-on-write pages */
96 if (is_cow_mapping(vma->vm_flags) &&
97 page_mapcount(page) != 1)
98 continue;
99
100 /*
101 * While migration can move some dirty pages,
102 * it cannot move them all from MIGRATE_ASYNC
103 * context.
104 */
105 if (page_is_file_lru(page) && PageDirty(page))
106 continue;
107
108 /*
109 * Don't mess with PTEs if page is already on the node
110 * a single-threaded process is running on.
111 */
112 if (target_node == page_to_nid(page))
113 continue;
114 }
115
116 oldpte = ptep_modify_prot_start(vma, addr, pte);
117 ptent = pte_modify(oldpte, newprot);
118 if (preserve_write)
119 ptent = pte_mk_savedwrite(ptent);
120
121 if (uffd_wp) {
122 ptent = pte_wrprotect(ptent);
123 ptent = pte_mkuffd_wp(ptent);
124 } else if (uffd_wp_resolve) {
125 /*
126 * Leave the write bit to be handled
127 * by PF interrupt handler, then
128 * things like COW could be properly
129 * handled.
130 */
131 ptent = pte_clear_uffd_wp(ptent);
132 }
133
134 /* Avoid taking write faults for known dirty pages */
135 if (dirty_accountable && pte_dirty(ptent) &&
136 (pte_soft_dirty(ptent) ||
137 !(vma->vm_flags & VM_SOFTDIRTY))) {
138 ptent = pte_mkwrite(ptent);
139 }
140 ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
141 pages++;
142 } else if (is_swap_pte(oldpte)) {
143 swp_entry_t entry = pte_to_swp_entry(oldpte);
144 pte_t newpte;
145
146 if (is_writable_migration_entry(entry)) {
147 /*
148 * A protection check is difficult so
149 * just be safe and disable write
150 */
151 entry = make_readable_migration_entry(
152 swp_offset(entry));
153 newpte = swp_entry_to_pte(entry);
154 if (pte_swp_soft_dirty(oldpte))
155 newpte = pte_swp_mksoft_dirty(newpte);
156 if (pte_swp_uffd_wp(oldpte))
157 newpte = pte_swp_mkuffd_wp(newpte);
158 } else if (is_writable_device_private_entry(entry)) {
159 /*
160 * We do not preserve soft-dirtiness. See
161 * copy_one_pte() for explanation.
162 */
163 entry = make_readable_device_private_entry(
164 swp_offset(entry));
165 newpte = swp_entry_to_pte(entry);
166 if (pte_swp_uffd_wp(oldpte))
167 newpte = pte_swp_mkuffd_wp(newpte);
168 } else if (is_writable_device_exclusive_entry(entry)) {
169 entry = make_readable_device_exclusive_entry(
170 swp_offset(entry));
171 newpte = swp_entry_to_pte(entry);
172 if (pte_swp_soft_dirty(oldpte))
173 newpte = pte_swp_mksoft_dirty(newpte);
174 if (pte_swp_uffd_wp(oldpte))
175 newpte = pte_swp_mkuffd_wp(newpte);
176 } else {
177 newpte = oldpte;
178 }
179
180 if (uffd_wp)
181 newpte = pte_swp_mkuffd_wp(newpte);
182 else if (uffd_wp_resolve)
183 newpte = pte_swp_clear_uffd_wp(newpte);
184
185 if (!pte_same(oldpte, newpte)) {
186 set_pte_at(vma->vm_mm, addr, pte, newpte);
187 pages++;
188 }
189 }
190 } while (pte++, addr += PAGE_SIZE, addr != end);
191 arch_leave_lazy_mmu_mode();
192 pte_unmap_unlock(pte - 1, ptl);
193
194 return pages;
195}
196
197/*
198 * Used when setting automatic NUMA hinting protection where it is
199 * critical that a numa hinting PMD is not confused with a bad PMD.
200 */
201static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
202{
203 pmd_t pmdval = pmd_read_atomic(pmd);
204
205 /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
206#ifdef CONFIG_TRANSPARENT_HUGEPAGE
207 barrier();
208#endif
209
210 if (pmd_none(pmdval))
211 return 1;
212 if (pmd_trans_huge(pmdval))
213 return 0;
214 if (unlikely(pmd_bad(pmdval))) {
215 pmd_clear_bad(pmd);
216 return 1;
217 }
218
219 return 0;
220}
221
222static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
223 pud_t *pud, unsigned long addr, unsigned long end,
224 pgprot_t newprot, unsigned long cp_flags)
225{
226 pmd_t *pmd;
227 unsigned long next;
228 unsigned long pages = 0;
229 unsigned long nr_huge_updates = 0;
230 struct mmu_notifier_range range;
231
232 range.start = 0;
233
234 pmd = pmd_offset(pud, addr);
235 do {
236 unsigned long this_pages;
237
238 next = pmd_addr_end(addr, end);
239
240 /*
241 * Automatic NUMA balancing walks the tables with mmap_lock
242 * held for read. It's possible a parallel update to occur
243 * between pmd_trans_huge() and a pmd_none_or_clear_bad()
244 * check leading to a false positive and clearing.
245 * Hence, it's necessary to atomically read the PMD value
246 * for all the checks.
247 */
248 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
249 pmd_none_or_clear_bad_unless_trans_huge(pmd))
250 goto next;
251
252 /* invoke the mmu notifier if the pmd is populated */
253 if (!range.start) {
254 mmu_notifier_range_init(&range,
255 MMU_NOTIFY_PROTECTION_VMA, 0,
256 vma, vma->vm_mm, addr, end);
257 mmu_notifier_invalidate_range_start(&range);
258 }
259
260 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
261 if (next - addr != HPAGE_PMD_SIZE) {
262 __split_huge_pmd(vma, pmd, addr, false, NULL);
263 } else {
264 int nr_ptes = change_huge_pmd(vma, pmd, addr,
265 newprot, cp_flags);
266
267 if (nr_ptes) {
268 if (nr_ptes == HPAGE_PMD_NR) {
269 pages += HPAGE_PMD_NR;
270 nr_huge_updates++;
271 }
272
273 /* huge pmd was handled */
274 goto next;
275 }
276 }
277 /* fall through, the trans huge pmd just split */
278 }
279 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
280 cp_flags);
281 pages += this_pages;
282next:
283 cond_resched();
284 } while (pmd++, addr = next, addr != end);
285
286 if (range.start)
287 mmu_notifier_invalidate_range_end(&range);
288
289 if (nr_huge_updates)
290 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
291 return pages;
292}
293
294static inline unsigned long change_pud_range(struct vm_area_struct *vma,
295 p4d_t *p4d, unsigned long addr, unsigned long end,
296 pgprot_t newprot, unsigned long cp_flags)
297{
298 pud_t *pud;
299 unsigned long next;
300 unsigned long pages = 0;
301
302 pud = pud_offset(p4d, addr);
303 do {
304 next = pud_addr_end(addr, end);
305 if (pud_none_or_clear_bad(pud))
306 continue;
307 pages += change_pmd_range(vma, pud, addr, next, newprot,
308 cp_flags);
309 } while (pud++, addr = next, addr != end);
310
311 return pages;
312}
313
314static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
315 pgd_t *pgd, unsigned long addr, unsigned long end,
316 pgprot_t newprot, unsigned long cp_flags)
317{
318 p4d_t *p4d;
319 unsigned long next;
320 unsigned long pages = 0;
321
322 p4d = p4d_offset(pgd, addr);
323 do {
324 next = p4d_addr_end(addr, end);
325 if (p4d_none_or_clear_bad(p4d))
326 continue;
327 pages += change_pud_range(vma, p4d, addr, next, newprot,
328 cp_flags);
329 } while (p4d++, addr = next, addr != end);
330
331 return pages;
332}
333
334static unsigned long change_protection_range(struct vm_area_struct *vma,
335 unsigned long addr, unsigned long end, pgprot_t newprot,
336 unsigned long cp_flags)
337{
338 struct mm_struct *mm = vma->vm_mm;
339 pgd_t *pgd;
340 unsigned long next;
341 unsigned long start = addr;
342 unsigned long pages = 0;
343
344 BUG_ON(addr >= end);
345 pgd = pgd_offset(mm, addr);
346 flush_cache_range(vma, addr, end);
347 inc_tlb_flush_pending(mm);
348 do {
349 next = pgd_addr_end(addr, end);
350 if (pgd_none_or_clear_bad(pgd))
351 continue;
352 pages += change_p4d_range(vma, pgd, addr, next, newprot,
353 cp_flags);
354 } while (pgd++, addr = next, addr != end);
355
356 /* Only flush the TLB if we actually modified any entries: */
357 if (pages)
358 flush_tlb_range(vma, start, end);
359 dec_tlb_flush_pending(mm);
360
361 return pages;
362}
363
364unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
365 unsigned long end, pgprot_t newprot,
366 unsigned long cp_flags)
367{
368 unsigned long pages;
369
370 BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
371
372 if (is_vm_hugetlb_page(vma))
373 pages = hugetlb_change_protection(vma, start, end, newprot);
374 else
375 pages = change_protection_range(vma, start, end, newprot,
376 cp_flags);
377
378 return pages;
379}
380
381static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
382 unsigned long next, struct mm_walk *walk)
383{
384 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
385 0 : -EACCES;
386}
387
388static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
389 unsigned long addr, unsigned long next,
390 struct mm_walk *walk)
391{
392 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
393 0 : -EACCES;
394}
395
396static int prot_none_test(unsigned long addr, unsigned long next,
397 struct mm_walk *walk)
398{
399 return 0;
400}
401
402static const struct mm_walk_ops prot_none_walk_ops = {
403 .pte_entry = prot_none_pte_entry,
404 .hugetlb_entry = prot_none_hugetlb_entry,
405 .test_walk = prot_none_test,
406};
407
408int
409mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
410 unsigned long start, unsigned long end, unsigned long newflags)
411{
412 struct mm_struct *mm = vma->vm_mm;
413 unsigned long oldflags = vma->vm_flags;
414 long nrpages = (end - start) >> PAGE_SHIFT;
415 unsigned long charged = 0;
416 pgoff_t pgoff;
417 int error;
418 int dirty_accountable = 0;
419
420 if (newflags == oldflags) {
421 *pprev = vma;
422 return 0;
423 }
424
425 /*
426 * Do PROT_NONE PFN permission checks here when we can still
427 * bail out without undoing a lot of state. This is a rather
428 * uncommon case, so doesn't need to be very optimized.
429 */
430 if (arch_has_pfn_modify_check() &&
431 (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
432 (newflags & VM_ACCESS_FLAGS) == 0) {
433 pgprot_t new_pgprot = vm_get_page_prot(newflags);
434
435 error = walk_page_range(current->mm, start, end,
436 &prot_none_walk_ops, &new_pgprot);
437 if (error)
438 return error;
439 }
440
441 /*
442 * If we make a private mapping writable we increase our commit;
443 * but (without finer accounting) cannot reduce our commit if we
444 * make it unwritable again. hugetlb mapping were accounted for
445 * even if read-only so there is no need to account for them here
446 */
447 if (newflags & VM_WRITE) {
448 /* Check space limits when area turns into data. */
449 if (!may_expand_vm(mm, newflags, nrpages) &&
450 may_expand_vm(mm, oldflags, nrpages))
451 return -ENOMEM;
452 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
453 VM_SHARED|VM_NORESERVE))) {
454 charged = nrpages;
455 if (security_vm_enough_memory_mm(mm, charged))
456 return -ENOMEM;
457 newflags |= VM_ACCOUNT;
458 }
459 }
460
461 /*
462 * First try to merge with previous and/or next vma.
463 */
464 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
465 *pprev = vma_merge(mm, *pprev, start, end, newflags,
466 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
467 vma->vm_userfaultfd_ctx);
468 if (*pprev) {
469 vma = *pprev;
470 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
471 goto success;
472 }
473
474 *pprev = vma;
475
476 if (start != vma->vm_start) {
477 error = split_vma(mm, vma, start, 1);
478 if (error)
479 goto fail;
480 }
481
482 if (end != vma->vm_end) {
483 error = split_vma(mm, vma, end, 0);
484 if (error)
485 goto fail;
486 }
487
488success:
489 /*
490 * vm_flags and vm_page_prot are protected by the mmap_lock
491 * held in write mode.
492 */
493 vma->vm_flags = newflags;
494 dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
495 vma_set_page_prot(vma);
496
497 change_protection(vma, start, end, vma->vm_page_prot,
498 dirty_accountable ? MM_CP_DIRTY_ACCT : 0);
499
500 /*
501 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
502 * fault on access.
503 */
504 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
505 (newflags & VM_WRITE)) {
506 populate_vma_page_range(vma, start, end, NULL);
507 }
508
509 vm_stat_account(mm, oldflags, -nrpages);
510 vm_stat_account(mm, newflags, nrpages);
511 perf_event_mmap(vma);
512 return 0;
513
514fail:
515 vm_unacct_memory(charged);
516 return error;
517}
518
519/*
520 * pkey==-1 when doing a legacy mprotect()
521 */
522static int do_mprotect_pkey(unsigned long start, size_t len,
523 unsigned long prot, int pkey)
524{
525 unsigned long nstart, end, tmp, reqprot;
526 struct vm_area_struct *vma, *prev;
527 int error = -EINVAL;
528 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
529 const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
530 (prot & PROT_READ);
531
532 start = untagged_addr(start);
533
534 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
535 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
536 return -EINVAL;
537
538 if (start & ~PAGE_MASK)
539 return -EINVAL;
540 if (!len)
541 return 0;
542 len = PAGE_ALIGN(len);
543 end = start + len;
544 if (end <= start)
545 return -ENOMEM;
546 if (!arch_validate_prot(prot, start))
547 return -EINVAL;
548
549 reqprot = prot;
550
551 if (mmap_write_lock_killable(current->mm))
552 return -EINTR;
553
554 /*
555 * If userspace did not allocate the pkey, do not let
556 * them use it here.
557 */
558 error = -EINVAL;
559 if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
560 goto out;
561
562 vma = find_vma(current->mm, start);
563 error = -ENOMEM;
564 if (!vma)
565 goto out;
566 prev = vma->vm_prev;
567 if (unlikely(grows & PROT_GROWSDOWN)) {
568 if (vma->vm_start >= end)
569 goto out;
570 start = vma->vm_start;
571 error = -EINVAL;
572 if (!(vma->vm_flags & VM_GROWSDOWN))
573 goto out;
574 } else {
575 if (vma->vm_start > start)
576 goto out;
577 if (unlikely(grows & PROT_GROWSUP)) {
578 end = vma->vm_end;
579 error = -EINVAL;
580 if (!(vma->vm_flags & VM_GROWSUP))
581 goto out;
582 }
583 }
584 if (start > vma->vm_start)
585 prev = vma;
586
587 for (nstart = start ; ; ) {
588 unsigned long mask_off_old_flags;
589 unsigned long newflags;
590 int new_vma_pkey;
591
592 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
593
594 /* Does the application expect PROT_READ to imply PROT_EXEC */
595 if (rier && (vma->vm_flags & VM_MAYEXEC))
596 prot |= PROT_EXEC;
597
598 /*
599 * Each mprotect() call explicitly passes r/w/x permissions.
600 * If a permission is not passed to mprotect(), it must be
601 * cleared from the VMA.
602 */
603 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
604 VM_FLAGS_CLEAR;
605
606 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
607 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
608 newflags |= (vma->vm_flags & ~mask_off_old_flags);
609
610 /* newflags >> 4 shift VM_MAY% in place of VM_% */
611 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
612 error = -EACCES;
613 goto out;
614 }
615
616 /* Allow architectures to sanity-check the new flags */
617 if (!arch_validate_flags(newflags)) {
618 error = -EINVAL;
619 goto out;
620 }
621
622 error = security_file_mprotect(vma, reqprot, prot);
623 if (error)
624 goto out;
625
626 tmp = vma->vm_end;
627 if (tmp > end)
628 tmp = end;
629
630 if (vma->vm_ops && vma->vm_ops->mprotect) {
631 error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
632 if (error)
633 goto out;
634 }
635
636 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
637 if (error)
638 goto out;
639
640 nstart = tmp;
641
642 if (nstart < prev->vm_end)
643 nstart = prev->vm_end;
644 if (nstart >= end)
645 goto out;
646
647 vma = prev->vm_next;
648 if (!vma || vma->vm_start != nstart) {
649 error = -ENOMEM;
650 goto out;
651 }
652 prot = reqprot;
653 }
654out:
655 mmap_write_unlock(current->mm);
656 return error;
657}
658
659SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
660 unsigned long, prot)
661{
662 return do_mprotect_pkey(start, len, prot, -1);
663}
664
665#ifdef CONFIG_ARCH_HAS_PKEYS
666
667SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
668 unsigned long, prot, int, pkey)
669{
670 return do_mprotect_pkey(start, len, prot, pkey);
671}
672
673SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
674{
675 int pkey;
676 int ret;
677
678 /* No flags supported yet. */
679 if (flags)
680 return -EINVAL;
681 /* check for unsupported init values */
682 if (init_val & ~PKEY_ACCESS_MASK)
683 return -EINVAL;
684
685 mmap_write_lock(current->mm);
686 pkey = mm_pkey_alloc(current->mm);
687
688 ret = -ENOSPC;
689 if (pkey == -1)
690 goto out;
691
692 ret = arch_set_user_pkey_access(current, pkey, init_val);
693 if (ret) {
694 mm_pkey_free(current->mm, pkey);
695 goto out;
696 }
697 ret = pkey;
698out:
699 mmap_write_unlock(current->mm);
700 return ret;
701}
702
703SYSCALL_DEFINE1(pkey_free, int, pkey)
704{
705 int ret;
706
707 mmap_write_lock(current->mm);
708 ret = mm_pkey_free(current->mm, pkey);
709 mmap_write_unlock(current->mm);
710
711 /*
712 * We could provide warnings or errors if any VMA still
713 * has the pkey set here.
714 */
715 return ret;
716}
717
718#endif /* CONFIG_ARCH_HAS_PKEYS */
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * mm/mprotect.c
4 *
5 * (C) Copyright 1994 Linus Torvalds
6 * (C) Copyright 2002 Christoph Hellwig
7 *
8 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
10 */
11
12#include <linux/pagewalk.h>
13#include <linux/hugetlb.h>
14#include <linux/shm.h>
15#include <linux/mman.h>
16#include <linux/fs.h>
17#include <linux/highmem.h>
18#include <linux/security.h>
19#include <linux/mempolicy.h>
20#include <linux/personality.h>
21#include <linux/syscalls.h>
22#include <linux/swap.h>
23#include <linux/swapops.h>
24#include <linux/mmu_notifier.h>
25#include <linux/migrate.h>
26#include <linux/perf_event.h>
27#include <linux/pkeys.h>
28#include <linux/ksm.h>
29#include <linux/uaccess.h>
30#include <linux/mm_inline.h>
31#include <linux/pgtable.h>
32#include <asm/cacheflush.h>
33#include <asm/mmu_context.h>
34#include <asm/tlbflush.h>
35
36#include "internal.h"
37
38static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
39 unsigned long addr, unsigned long end, pgprot_t newprot,
40 unsigned long cp_flags)
41{
42 pte_t *pte, oldpte;
43 spinlock_t *ptl;
44 unsigned long pages = 0;
45 int target_node = NUMA_NO_NODE;
46 bool dirty_accountable = cp_flags & MM_CP_DIRTY_ACCT;
47 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
48 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
49 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
50
51 /*
52 * Can be called with only the mmap_lock for reading by
53 * prot_numa so we must check the pmd isn't constantly
54 * changing from under us from pmd_none to pmd_trans_huge
55 * and/or the other way around.
56 */
57 if (pmd_trans_unstable(pmd))
58 return 0;
59
60 /*
61 * The pmd points to a regular pte so the pmd can't change
62 * from under us even if the mmap_lock is only hold for
63 * reading.
64 */
65 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
66
67 /* Get target node for single threaded private VMAs */
68 if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
69 atomic_read(&vma->vm_mm->mm_users) == 1)
70 target_node = numa_node_id();
71
72 flush_tlb_batched_pending(vma->vm_mm);
73 arch_enter_lazy_mmu_mode();
74 do {
75 oldpte = *pte;
76 if (pte_present(oldpte)) {
77 pte_t ptent;
78 bool preserve_write = prot_numa && pte_write(oldpte);
79
80 /*
81 * Avoid trapping faults against the zero or KSM
82 * pages. See similar comment in change_huge_pmd.
83 */
84 if (prot_numa) {
85 struct page *page;
86
87 /* Avoid TLB flush if possible */
88 if (pte_protnone(oldpte))
89 continue;
90
91 page = vm_normal_page(vma, addr, oldpte);
92 if (!page || PageKsm(page))
93 continue;
94
95 /* Also skip shared copy-on-write pages */
96 if (is_cow_mapping(vma->vm_flags) &&
97 page_mapcount(page) != 1)
98 continue;
99
100 /*
101 * While migration can move some dirty pages,
102 * it cannot move them all from MIGRATE_ASYNC
103 * context.
104 */
105 if (page_is_file_lru(page) && PageDirty(page))
106 continue;
107
108 /*
109 * Don't mess with PTEs if page is already on the node
110 * a single-threaded process is running on.
111 */
112 if (target_node == page_to_nid(page))
113 continue;
114 }
115
116 oldpte = ptep_modify_prot_start(vma, addr, pte);
117 ptent = pte_modify(oldpte, newprot);
118 if (preserve_write)
119 ptent = pte_mk_savedwrite(ptent);
120
121 if (uffd_wp) {
122 ptent = pte_wrprotect(ptent);
123 ptent = pte_mkuffd_wp(ptent);
124 } else if (uffd_wp_resolve) {
125 /*
126 * Leave the write bit to be handled
127 * by PF interrupt handler, then
128 * things like COW could be properly
129 * handled.
130 */
131 ptent = pte_clear_uffd_wp(ptent);
132 }
133
134 /* Avoid taking write faults for known dirty pages */
135 if (dirty_accountable && pte_dirty(ptent) &&
136 (pte_soft_dirty(ptent) ||
137 !(vma->vm_flags & VM_SOFTDIRTY))) {
138 ptent = pte_mkwrite(ptent);
139 }
140 ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
141 pages++;
142 } else if (is_swap_pte(oldpte)) {
143 swp_entry_t entry = pte_to_swp_entry(oldpte);
144 pte_t newpte;
145
146 if (is_write_migration_entry(entry)) {
147 /*
148 * A protection check is difficult so
149 * just be safe and disable write
150 */
151 make_migration_entry_read(&entry);
152 newpte = swp_entry_to_pte(entry);
153 if (pte_swp_soft_dirty(oldpte))
154 newpte = pte_swp_mksoft_dirty(newpte);
155 if (pte_swp_uffd_wp(oldpte))
156 newpte = pte_swp_mkuffd_wp(newpte);
157 } else if (is_write_device_private_entry(entry)) {
158 /*
159 * We do not preserve soft-dirtiness. See
160 * copy_one_pte() for explanation.
161 */
162 make_device_private_entry_read(&entry);
163 newpte = swp_entry_to_pte(entry);
164 if (pte_swp_uffd_wp(oldpte))
165 newpte = pte_swp_mkuffd_wp(newpte);
166 } else {
167 newpte = oldpte;
168 }
169
170 if (uffd_wp)
171 newpte = pte_swp_mkuffd_wp(newpte);
172 else if (uffd_wp_resolve)
173 newpte = pte_swp_clear_uffd_wp(newpte);
174
175 if (!pte_same(oldpte, newpte)) {
176 set_pte_at(vma->vm_mm, addr, pte, newpte);
177 pages++;
178 }
179 }
180 } while (pte++, addr += PAGE_SIZE, addr != end);
181 arch_leave_lazy_mmu_mode();
182 pte_unmap_unlock(pte - 1, ptl);
183
184 return pages;
185}
186
187/*
188 * Used when setting automatic NUMA hinting protection where it is
189 * critical that a numa hinting PMD is not confused with a bad PMD.
190 */
191static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
192{
193 pmd_t pmdval = pmd_read_atomic(pmd);
194
195 /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
196#ifdef CONFIG_TRANSPARENT_HUGEPAGE
197 barrier();
198#endif
199
200 if (pmd_none(pmdval))
201 return 1;
202 if (pmd_trans_huge(pmdval))
203 return 0;
204 if (unlikely(pmd_bad(pmdval))) {
205 pmd_clear_bad(pmd);
206 return 1;
207 }
208
209 return 0;
210}
211
212static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
213 pud_t *pud, unsigned long addr, unsigned long end,
214 pgprot_t newprot, unsigned long cp_flags)
215{
216 pmd_t *pmd;
217 unsigned long next;
218 unsigned long pages = 0;
219 unsigned long nr_huge_updates = 0;
220 struct mmu_notifier_range range;
221
222 range.start = 0;
223
224 pmd = pmd_offset(pud, addr);
225 do {
226 unsigned long this_pages;
227
228 next = pmd_addr_end(addr, end);
229
230 /*
231 * Automatic NUMA balancing walks the tables with mmap_lock
232 * held for read. It's possible a parallel update to occur
233 * between pmd_trans_huge() and a pmd_none_or_clear_bad()
234 * check leading to a false positive and clearing.
235 * Hence, it's necessary to atomically read the PMD value
236 * for all the checks.
237 */
238 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
239 pmd_none_or_clear_bad_unless_trans_huge(pmd))
240 goto next;
241
242 /* invoke the mmu notifier if the pmd is populated */
243 if (!range.start) {
244 mmu_notifier_range_init(&range,
245 MMU_NOTIFY_PROTECTION_VMA, 0,
246 vma, vma->vm_mm, addr, end);
247 mmu_notifier_invalidate_range_start(&range);
248 }
249
250 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
251 if (next - addr != HPAGE_PMD_SIZE) {
252 __split_huge_pmd(vma, pmd, addr, false, NULL);
253 } else {
254 int nr_ptes = change_huge_pmd(vma, pmd, addr,
255 newprot, cp_flags);
256
257 if (nr_ptes) {
258 if (nr_ptes == HPAGE_PMD_NR) {
259 pages += HPAGE_PMD_NR;
260 nr_huge_updates++;
261 }
262
263 /* huge pmd was handled */
264 goto next;
265 }
266 }
267 /* fall through, the trans huge pmd just split */
268 }
269 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
270 cp_flags);
271 pages += this_pages;
272next:
273 cond_resched();
274 } while (pmd++, addr = next, addr != end);
275
276 if (range.start)
277 mmu_notifier_invalidate_range_end(&range);
278
279 if (nr_huge_updates)
280 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
281 return pages;
282}
283
284static inline unsigned long change_pud_range(struct vm_area_struct *vma,
285 p4d_t *p4d, unsigned long addr, unsigned long end,
286 pgprot_t newprot, unsigned long cp_flags)
287{
288 pud_t *pud;
289 unsigned long next;
290 unsigned long pages = 0;
291
292 pud = pud_offset(p4d, addr);
293 do {
294 next = pud_addr_end(addr, end);
295 if (pud_none_or_clear_bad(pud))
296 continue;
297 pages += change_pmd_range(vma, pud, addr, next, newprot,
298 cp_flags);
299 } while (pud++, addr = next, addr != end);
300
301 return pages;
302}
303
304static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
305 pgd_t *pgd, unsigned long addr, unsigned long end,
306 pgprot_t newprot, unsigned long cp_flags)
307{
308 p4d_t *p4d;
309 unsigned long next;
310 unsigned long pages = 0;
311
312 p4d = p4d_offset(pgd, addr);
313 do {
314 next = p4d_addr_end(addr, end);
315 if (p4d_none_or_clear_bad(p4d))
316 continue;
317 pages += change_pud_range(vma, p4d, addr, next, newprot,
318 cp_flags);
319 } while (p4d++, addr = next, addr != end);
320
321 return pages;
322}
323
324static unsigned long change_protection_range(struct vm_area_struct *vma,
325 unsigned long addr, unsigned long end, pgprot_t newprot,
326 unsigned long cp_flags)
327{
328 struct mm_struct *mm = vma->vm_mm;
329 pgd_t *pgd;
330 unsigned long next;
331 unsigned long start = addr;
332 unsigned long pages = 0;
333
334 BUG_ON(addr >= end);
335 pgd = pgd_offset(mm, addr);
336 flush_cache_range(vma, addr, end);
337 inc_tlb_flush_pending(mm);
338 do {
339 next = pgd_addr_end(addr, end);
340 if (pgd_none_or_clear_bad(pgd))
341 continue;
342 pages += change_p4d_range(vma, pgd, addr, next, newprot,
343 cp_flags);
344 } while (pgd++, addr = next, addr != end);
345
346 /* Only flush the TLB if we actually modified any entries: */
347 if (pages)
348 flush_tlb_range(vma, start, end);
349 dec_tlb_flush_pending(mm);
350
351 return pages;
352}
353
354unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
355 unsigned long end, pgprot_t newprot,
356 unsigned long cp_flags)
357{
358 unsigned long pages;
359
360 BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
361
362 if (is_vm_hugetlb_page(vma))
363 pages = hugetlb_change_protection(vma, start, end, newprot);
364 else
365 pages = change_protection_range(vma, start, end, newprot,
366 cp_flags);
367
368 return pages;
369}
370
371static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
372 unsigned long next, struct mm_walk *walk)
373{
374 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
375 0 : -EACCES;
376}
377
378static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
379 unsigned long addr, unsigned long next,
380 struct mm_walk *walk)
381{
382 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
383 0 : -EACCES;
384}
385
386static int prot_none_test(unsigned long addr, unsigned long next,
387 struct mm_walk *walk)
388{
389 return 0;
390}
391
392static const struct mm_walk_ops prot_none_walk_ops = {
393 .pte_entry = prot_none_pte_entry,
394 .hugetlb_entry = prot_none_hugetlb_entry,
395 .test_walk = prot_none_test,
396};
397
398int
399mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
400 unsigned long start, unsigned long end, unsigned long newflags)
401{
402 struct mm_struct *mm = vma->vm_mm;
403 unsigned long oldflags = vma->vm_flags;
404 long nrpages = (end - start) >> PAGE_SHIFT;
405 unsigned long charged = 0;
406 pgoff_t pgoff;
407 int error;
408 int dirty_accountable = 0;
409
410 if (newflags == oldflags) {
411 *pprev = vma;
412 return 0;
413 }
414
415 /*
416 * Do PROT_NONE PFN permission checks here when we can still
417 * bail out without undoing a lot of state. This is a rather
418 * uncommon case, so doesn't need to be very optimized.
419 */
420 if (arch_has_pfn_modify_check() &&
421 (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
422 (newflags & VM_ACCESS_FLAGS) == 0) {
423 pgprot_t new_pgprot = vm_get_page_prot(newflags);
424
425 error = walk_page_range(current->mm, start, end,
426 &prot_none_walk_ops, &new_pgprot);
427 if (error)
428 return error;
429 }
430
431 /*
432 * If we make a private mapping writable we increase our commit;
433 * but (without finer accounting) cannot reduce our commit if we
434 * make it unwritable again. hugetlb mapping were accounted for
435 * even if read-only so there is no need to account for them here
436 */
437 if (newflags & VM_WRITE) {
438 /* Check space limits when area turns into data. */
439 if (!may_expand_vm(mm, newflags, nrpages) &&
440 may_expand_vm(mm, oldflags, nrpages))
441 return -ENOMEM;
442 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
443 VM_SHARED|VM_NORESERVE))) {
444 charged = nrpages;
445 if (security_vm_enough_memory_mm(mm, charged))
446 return -ENOMEM;
447 newflags |= VM_ACCOUNT;
448 }
449 }
450
451 /*
452 * First try to merge with previous and/or next vma.
453 */
454 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
455 *pprev = vma_merge(mm, *pprev, start, end, newflags,
456 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
457 vma->vm_userfaultfd_ctx);
458 if (*pprev) {
459 vma = *pprev;
460 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
461 goto success;
462 }
463
464 *pprev = vma;
465
466 if (start != vma->vm_start) {
467 error = split_vma(mm, vma, start, 1);
468 if (error)
469 goto fail;
470 }
471
472 if (end != vma->vm_end) {
473 error = split_vma(mm, vma, end, 0);
474 if (error)
475 goto fail;
476 }
477
478success:
479 /*
480 * vm_flags and vm_page_prot are protected by the mmap_lock
481 * held in write mode.
482 */
483 vma->vm_flags = newflags;
484 dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
485 vma_set_page_prot(vma);
486
487 change_protection(vma, start, end, vma->vm_page_prot,
488 dirty_accountable ? MM_CP_DIRTY_ACCT : 0);
489
490 /*
491 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
492 * fault on access.
493 */
494 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
495 (newflags & VM_WRITE)) {
496 populate_vma_page_range(vma, start, end, NULL);
497 }
498
499 vm_stat_account(mm, oldflags, -nrpages);
500 vm_stat_account(mm, newflags, nrpages);
501 perf_event_mmap(vma);
502 return 0;
503
504fail:
505 vm_unacct_memory(charged);
506 return error;
507}
508
509/*
510 * pkey==-1 when doing a legacy mprotect()
511 */
512static int do_mprotect_pkey(unsigned long start, size_t len,
513 unsigned long prot, int pkey)
514{
515 unsigned long nstart, end, tmp, reqprot;
516 struct vm_area_struct *vma, *prev;
517 int error = -EINVAL;
518 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
519 const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
520 (prot & PROT_READ);
521
522 start = untagged_addr(start);
523
524 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
525 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
526 return -EINVAL;
527
528 if (start & ~PAGE_MASK)
529 return -EINVAL;
530 if (!len)
531 return 0;
532 len = PAGE_ALIGN(len);
533 end = start + len;
534 if (end <= start)
535 return -ENOMEM;
536 if (!arch_validate_prot(prot, start))
537 return -EINVAL;
538
539 reqprot = prot;
540
541 if (mmap_write_lock_killable(current->mm))
542 return -EINTR;
543
544 /*
545 * If userspace did not allocate the pkey, do not let
546 * them use it here.
547 */
548 error = -EINVAL;
549 if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
550 goto out;
551
552 vma = find_vma(current->mm, start);
553 error = -ENOMEM;
554 if (!vma)
555 goto out;
556 prev = vma->vm_prev;
557 if (unlikely(grows & PROT_GROWSDOWN)) {
558 if (vma->vm_start >= end)
559 goto out;
560 start = vma->vm_start;
561 error = -EINVAL;
562 if (!(vma->vm_flags & VM_GROWSDOWN))
563 goto out;
564 } else {
565 if (vma->vm_start > start)
566 goto out;
567 if (unlikely(grows & PROT_GROWSUP)) {
568 end = vma->vm_end;
569 error = -EINVAL;
570 if (!(vma->vm_flags & VM_GROWSUP))
571 goto out;
572 }
573 }
574 if (start > vma->vm_start)
575 prev = vma;
576
577 for (nstart = start ; ; ) {
578 unsigned long mask_off_old_flags;
579 unsigned long newflags;
580 int new_vma_pkey;
581
582 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
583
584 /* Does the application expect PROT_READ to imply PROT_EXEC */
585 if (rier && (vma->vm_flags & VM_MAYEXEC))
586 prot |= PROT_EXEC;
587
588 /*
589 * Each mprotect() call explicitly passes r/w/x permissions.
590 * If a permission is not passed to mprotect(), it must be
591 * cleared from the VMA.
592 */
593 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
594 VM_FLAGS_CLEAR;
595
596 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
597 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
598 newflags |= (vma->vm_flags & ~mask_off_old_flags);
599
600 /* newflags >> 4 shift VM_MAY% in place of VM_% */
601 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
602 error = -EACCES;
603 goto out;
604 }
605
606 error = security_file_mprotect(vma, reqprot, prot);
607 if (error)
608 goto out;
609
610 tmp = vma->vm_end;
611 if (tmp > end)
612 tmp = end;
613 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
614 if (error)
615 goto out;
616 nstart = tmp;
617
618 if (nstart < prev->vm_end)
619 nstart = prev->vm_end;
620 if (nstart >= end)
621 goto out;
622
623 vma = prev->vm_next;
624 if (!vma || vma->vm_start != nstart) {
625 error = -ENOMEM;
626 goto out;
627 }
628 prot = reqprot;
629 }
630out:
631 mmap_write_unlock(current->mm);
632 return error;
633}
634
635SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
636 unsigned long, prot)
637{
638 return do_mprotect_pkey(start, len, prot, -1);
639}
640
641#ifdef CONFIG_ARCH_HAS_PKEYS
642
643SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
644 unsigned long, prot, int, pkey)
645{
646 return do_mprotect_pkey(start, len, prot, pkey);
647}
648
649SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
650{
651 int pkey;
652 int ret;
653
654 /* No flags supported yet. */
655 if (flags)
656 return -EINVAL;
657 /* check for unsupported init values */
658 if (init_val & ~PKEY_ACCESS_MASK)
659 return -EINVAL;
660
661 mmap_write_lock(current->mm);
662 pkey = mm_pkey_alloc(current->mm);
663
664 ret = -ENOSPC;
665 if (pkey == -1)
666 goto out;
667
668 ret = arch_set_user_pkey_access(current, pkey, init_val);
669 if (ret) {
670 mm_pkey_free(current->mm, pkey);
671 goto out;
672 }
673 ret = pkey;
674out:
675 mmap_write_unlock(current->mm);
676 return ret;
677}
678
679SYSCALL_DEFINE1(pkey_free, int, pkey)
680{
681 int ret;
682
683 mmap_write_lock(current->mm);
684 ret = mm_pkey_free(current->mm, pkey);
685 mmap_write_unlock(current->mm);
686
687 /*
688 * We could provie warnings or errors if any VMA still
689 * has the pkey set here.
690 */
691 return ret;
692}
693
694#endif /* CONFIG_ARCH_HAS_PKEYS */