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 <linux/sched/sysctl.h>
33#include <linux/userfaultfd_k.h>
34#include <linux/memory-tiers.h>
35#include <asm/cacheflush.h>
36#include <asm/mmu_context.h>
37#include <asm/tlbflush.h>
38#include <asm/tlb.h>
39
40#include "internal.h"
41
42bool can_change_pte_writable(struct vm_area_struct *vma, unsigned long addr,
43 pte_t pte)
44{
45 struct page *page;
46
47 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
48 return false;
49
50 /* Don't touch entries that are not even readable. */
51 if (pte_protnone(pte))
52 return false;
53
54 /* Do we need write faults for softdirty tracking? */
55 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
56 return false;
57
58 /* Do we need write faults for uffd-wp tracking? */
59 if (userfaultfd_pte_wp(vma, pte))
60 return false;
61
62 if (!(vma->vm_flags & VM_SHARED)) {
63 /*
64 * Writable MAP_PRIVATE mapping: We can only special-case on
65 * exclusive anonymous pages, because we know that our
66 * write-fault handler similarly would map them writable without
67 * any additional checks while holding the PT lock.
68 */
69 page = vm_normal_page(vma, addr, pte);
70 return page && PageAnon(page) && PageAnonExclusive(page);
71 }
72
73 /*
74 * Writable MAP_SHARED mapping: "clean" might indicate that the FS still
75 * needs a real write-fault for writenotify
76 * (see vma_wants_writenotify()). If "dirty", the assumption is that the
77 * FS was already notified and we can simply mark the PTE writable
78 * just like the write-fault handler would do.
79 */
80 return pte_dirty(pte);
81}
82
83static unsigned long change_pte_range(struct mmu_gather *tlb,
84 struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
85 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
86{
87 pte_t *pte, oldpte;
88 spinlock_t *ptl;
89 unsigned long pages = 0;
90 int target_node = NUMA_NO_NODE;
91 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
92 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
93 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
94
95 tlb_change_page_size(tlb, PAGE_SIZE);
96
97 /*
98 * Can be called with only the mmap_lock for reading by
99 * prot_numa so we must check the pmd isn't constantly
100 * changing from under us from pmd_none to pmd_trans_huge
101 * and/or the other way around.
102 */
103 if (pmd_trans_unstable(pmd))
104 return 0;
105
106 /*
107 * The pmd points to a regular pte so the pmd can't change
108 * from under us even if the mmap_lock is only hold for
109 * reading.
110 */
111 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
112
113 /* Get target node for single threaded private VMAs */
114 if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
115 atomic_read(&vma->vm_mm->mm_users) == 1)
116 target_node = numa_node_id();
117
118 flush_tlb_batched_pending(vma->vm_mm);
119 arch_enter_lazy_mmu_mode();
120 do {
121 oldpte = *pte;
122 if (pte_present(oldpte)) {
123 pte_t ptent;
124
125 /*
126 * Avoid trapping faults against the zero or KSM
127 * pages. See similar comment in change_huge_pmd.
128 */
129 if (prot_numa) {
130 struct page *page;
131 int nid;
132 bool toptier;
133
134 /* Avoid TLB flush if possible */
135 if (pte_protnone(oldpte))
136 continue;
137
138 page = vm_normal_page(vma, addr, oldpte);
139 if (!page || is_zone_device_page(page) || PageKsm(page))
140 continue;
141
142 /* Also skip shared copy-on-write pages */
143 if (is_cow_mapping(vma->vm_flags) &&
144 page_count(page) != 1)
145 continue;
146
147 /*
148 * While migration can move some dirty pages,
149 * it cannot move them all from MIGRATE_ASYNC
150 * context.
151 */
152 if (page_is_file_lru(page) && PageDirty(page))
153 continue;
154
155 /*
156 * Don't mess with PTEs if page is already on the node
157 * a single-threaded process is running on.
158 */
159 nid = page_to_nid(page);
160 if (target_node == nid)
161 continue;
162 toptier = node_is_toptier(nid);
163
164 /*
165 * Skip scanning top tier node if normal numa
166 * balancing is disabled
167 */
168 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
169 toptier)
170 continue;
171 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
172 !toptier)
173 xchg_page_access_time(page,
174 jiffies_to_msecs(jiffies));
175 }
176
177 oldpte = ptep_modify_prot_start(vma, addr, pte);
178 ptent = pte_modify(oldpte, newprot);
179
180 if (uffd_wp) {
181 ptent = pte_wrprotect(ptent);
182 ptent = pte_mkuffd_wp(ptent);
183 } else if (uffd_wp_resolve) {
184 ptent = pte_clear_uffd_wp(ptent);
185 }
186
187 /*
188 * In some writable, shared mappings, we might want
189 * to catch actual write access -- see
190 * vma_wants_writenotify().
191 *
192 * In all writable, private mappings, we have to
193 * properly handle COW.
194 *
195 * In both cases, we can sometimes still change PTEs
196 * writable and avoid the write-fault handler, for
197 * example, if a PTE is already dirty and no other
198 * COW or special handling is required.
199 */
200 if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
201 !pte_write(ptent) &&
202 can_change_pte_writable(vma, addr, ptent))
203 ptent = pte_mkwrite(ptent);
204
205 ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
206 if (pte_needs_flush(oldpte, ptent))
207 tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
208 pages++;
209 } else if (is_swap_pte(oldpte)) {
210 swp_entry_t entry = pte_to_swp_entry(oldpte);
211 pte_t newpte;
212
213 if (is_writable_migration_entry(entry)) {
214 struct page *page = pfn_swap_entry_to_page(entry);
215
216 /*
217 * A protection check is difficult so
218 * just be safe and disable write
219 */
220 if (PageAnon(page))
221 entry = make_readable_exclusive_migration_entry(
222 swp_offset(entry));
223 else
224 entry = make_readable_migration_entry(swp_offset(entry));
225 newpte = swp_entry_to_pte(entry);
226 if (pte_swp_soft_dirty(oldpte))
227 newpte = pte_swp_mksoft_dirty(newpte);
228 if (pte_swp_uffd_wp(oldpte))
229 newpte = pte_swp_mkuffd_wp(newpte);
230 } else if (is_writable_device_private_entry(entry)) {
231 /*
232 * We do not preserve soft-dirtiness. See
233 * copy_one_pte() for explanation.
234 */
235 entry = make_readable_device_private_entry(
236 swp_offset(entry));
237 newpte = swp_entry_to_pte(entry);
238 if (pte_swp_uffd_wp(oldpte))
239 newpte = pte_swp_mkuffd_wp(newpte);
240 } else if (is_writable_device_exclusive_entry(entry)) {
241 entry = make_readable_device_exclusive_entry(
242 swp_offset(entry));
243 newpte = swp_entry_to_pte(entry);
244 if (pte_swp_soft_dirty(oldpte))
245 newpte = pte_swp_mksoft_dirty(newpte);
246 if (pte_swp_uffd_wp(oldpte))
247 newpte = pte_swp_mkuffd_wp(newpte);
248 } else if (is_pte_marker_entry(entry)) {
249 /*
250 * Ignore swapin errors unconditionally,
251 * because any access should sigbus anyway.
252 */
253 if (is_swapin_error_entry(entry))
254 continue;
255 /*
256 * If this is uffd-wp pte marker and we'd like
257 * to unprotect it, drop it; the next page
258 * fault will trigger without uffd trapping.
259 */
260 if (uffd_wp_resolve) {
261 pte_clear(vma->vm_mm, addr, pte);
262 pages++;
263 }
264 continue;
265 } else {
266 newpte = oldpte;
267 }
268
269 if (uffd_wp)
270 newpte = pte_swp_mkuffd_wp(newpte);
271 else if (uffd_wp_resolve)
272 newpte = pte_swp_clear_uffd_wp(newpte);
273
274 if (!pte_same(oldpte, newpte)) {
275 set_pte_at(vma->vm_mm, addr, pte, newpte);
276 pages++;
277 }
278 } else {
279 /* It must be an none page, or what else?.. */
280 WARN_ON_ONCE(!pte_none(oldpte));
281 if (unlikely(uffd_wp && !vma_is_anonymous(vma))) {
282 /*
283 * For file-backed mem, we need to be able to
284 * wr-protect a none pte, because even if the
285 * pte is none, the page/swap cache could
286 * exist. Doing that by install a marker.
287 */
288 set_pte_at(vma->vm_mm, addr, pte,
289 make_pte_marker(PTE_MARKER_UFFD_WP));
290 pages++;
291 }
292 }
293 } while (pte++, addr += PAGE_SIZE, addr != end);
294 arch_leave_lazy_mmu_mode();
295 pte_unmap_unlock(pte - 1, ptl);
296
297 return pages;
298}
299
300/*
301 * Used when setting automatic NUMA hinting protection where it is
302 * critical that a numa hinting PMD is not confused with a bad PMD.
303 */
304static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
305{
306 pmd_t pmdval = pmdp_get_lockless(pmd);
307
308 /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
309#ifdef CONFIG_TRANSPARENT_HUGEPAGE
310 barrier();
311#endif
312
313 if (pmd_none(pmdval))
314 return 1;
315 if (pmd_trans_huge(pmdval))
316 return 0;
317 if (unlikely(pmd_bad(pmdval))) {
318 pmd_clear_bad(pmd);
319 return 1;
320 }
321
322 return 0;
323}
324
325/* Return true if we're uffd wr-protecting file-backed memory, or false */
326static inline bool
327uffd_wp_protect_file(struct vm_area_struct *vma, unsigned long cp_flags)
328{
329 return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
330}
331
332/*
333 * If wr-protecting the range for file-backed, populate pgtable for the case
334 * when pgtable is empty but page cache exists. When {pte|pmd|...}_alloc()
335 * failed it means no memory, we don't have a better option but stop.
336 */
337#define change_pmd_prepare(vma, pmd, cp_flags) \
338 do { \
339 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) { \
340 if (WARN_ON_ONCE(pte_alloc(vma->vm_mm, pmd))) \
341 break; \
342 } \
343 } while (0)
344/*
345 * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
346 * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
347 * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
348 */
349#define change_prepare(vma, high, low, addr, cp_flags) \
350 do { \
351 if (unlikely(uffd_wp_protect_file(vma, cp_flags))) { \
352 low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
353 if (WARN_ON_ONCE(p == NULL)) \
354 break; \
355 } \
356 } while (0)
357
358static inline unsigned long change_pmd_range(struct mmu_gather *tlb,
359 struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
360 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
361{
362 pmd_t *pmd;
363 unsigned long next;
364 unsigned long pages = 0;
365 unsigned long nr_huge_updates = 0;
366 struct mmu_notifier_range range;
367
368 range.start = 0;
369
370 pmd = pmd_offset(pud, addr);
371 do {
372 unsigned long this_pages;
373
374 next = pmd_addr_end(addr, end);
375
376 change_pmd_prepare(vma, pmd, cp_flags);
377 /*
378 * Automatic NUMA balancing walks the tables with mmap_lock
379 * held for read. It's possible a parallel update to occur
380 * between pmd_trans_huge() and a pmd_none_or_clear_bad()
381 * check leading to a false positive and clearing.
382 * Hence, it's necessary to atomically read the PMD value
383 * for all the checks.
384 */
385 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
386 pmd_none_or_clear_bad_unless_trans_huge(pmd))
387 goto next;
388
389 /* invoke the mmu notifier if the pmd is populated */
390 if (!range.start) {
391 mmu_notifier_range_init(&range,
392 MMU_NOTIFY_PROTECTION_VMA, 0,
393 vma, vma->vm_mm, addr, end);
394 mmu_notifier_invalidate_range_start(&range);
395 }
396
397 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
398 if ((next - addr != HPAGE_PMD_SIZE) ||
399 uffd_wp_protect_file(vma, cp_flags)) {
400 __split_huge_pmd(vma, pmd, addr, false, NULL);
401 /*
402 * For file-backed, the pmd could have been
403 * cleared; make sure pmd populated if
404 * necessary, then fall-through to pte level.
405 */
406 change_pmd_prepare(vma, pmd, cp_flags);
407 } else {
408 /*
409 * change_huge_pmd() does not defer TLB flushes,
410 * so no need to propagate the tlb argument.
411 */
412 int nr_ptes = change_huge_pmd(tlb, vma, pmd,
413 addr, newprot, cp_flags);
414
415 if (nr_ptes) {
416 if (nr_ptes == HPAGE_PMD_NR) {
417 pages += HPAGE_PMD_NR;
418 nr_huge_updates++;
419 }
420
421 /* huge pmd was handled */
422 goto next;
423 }
424 }
425 /* fall through, the trans huge pmd just split */
426 }
427 this_pages = change_pte_range(tlb, vma, pmd, addr, next,
428 newprot, cp_flags);
429 pages += this_pages;
430next:
431 cond_resched();
432 } while (pmd++, addr = next, addr != end);
433
434 if (range.start)
435 mmu_notifier_invalidate_range_end(&range);
436
437 if (nr_huge_updates)
438 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
439 return pages;
440}
441
442static inline unsigned long change_pud_range(struct mmu_gather *tlb,
443 struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
444 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
445{
446 pud_t *pud;
447 unsigned long next;
448 unsigned long pages = 0;
449
450 pud = pud_offset(p4d, addr);
451 do {
452 next = pud_addr_end(addr, end);
453 change_prepare(vma, pud, pmd, addr, cp_flags);
454 if (pud_none_or_clear_bad(pud))
455 continue;
456 pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
457 cp_flags);
458 } while (pud++, addr = next, addr != end);
459
460 return pages;
461}
462
463static inline unsigned long change_p4d_range(struct mmu_gather *tlb,
464 struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
465 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
466{
467 p4d_t *p4d;
468 unsigned long next;
469 unsigned long pages = 0;
470
471 p4d = p4d_offset(pgd, addr);
472 do {
473 next = p4d_addr_end(addr, end);
474 change_prepare(vma, p4d, pud, addr, cp_flags);
475 if (p4d_none_or_clear_bad(p4d))
476 continue;
477 pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
478 cp_flags);
479 } while (p4d++, addr = next, addr != end);
480
481 return pages;
482}
483
484static unsigned long change_protection_range(struct mmu_gather *tlb,
485 struct vm_area_struct *vma, unsigned long addr,
486 unsigned long end, pgprot_t newprot, unsigned long cp_flags)
487{
488 struct mm_struct *mm = vma->vm_mm;
489 pgd_t *pgd;
490 unsigned long next;
491 unsigned long pages = 0;
492
493 BUG_ON(addr >= end);
494 pgd = pgd_offset(mm, addr);
495 tlb_start_vma(tlb, vma);
496 do {
497 next = pgd_addr_end(addr, end);
498 change_prepare(vma, pgd, p4d, addr, cp_flags);
499 if (pgd_none_or_clear_bad(pgd))
500 continue;
501 pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
502 cp_flags);
503 } while (pgd++, addr = next, addr != end);
504
505 tlb_end_vma(tlb, vma);
506
507 return pages;
508}
509
510unsigned long change_protection(struct mmu_gather *tlb,
511 struct vm_area_struct *vma, unsigned long start,
512 unsigned long end, pgprot_t newprot,
513 unsigned long cp_flags)
514{
515 unsigned long pages;
516
517 BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
518
519 if (is_vm_hugetlb_page(vma))
520 pages = hugetlb_change_protection(vma, start, end, newprot,
521 cp_flags);
522 else
523 pages = change_protection_range(tlb, vma, start, end, newprot,
524 cp_flags);
525
526 return pages;
527}
528
529static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
530 unsigned long next, struct mm_walk *walk)
531{
532 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
533 0 : -EACCES;
534}
535
536static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
537 unsigned long addr, unsigned long next,
538 struct mm_walk *walk)
539{
540 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
541 0 : -EACCES;
542}
543
544static int prot_none_test(unsigned long addr, unsigned long next,
545 struct mm_walk *walk)
546{
547 return 0;
548}
549
550static const struct mm_walk_ops prot_none_walk_ops = {
551 .pte_entry = prot_none_pte_entry,
552 .hugetlb_entry = prot_none_hugetlb_entry,
553 .test_walk = prot_none_test,
554};
555
556int
557mprotect_fixup(struct mmu_gather *tlb, struct vm_area_struct *vma,
558 struct vm_area_struct **pprev, unsigned long start,
559 unsigned long end, unsigned long newflags)
560{
561 struct mm_struct *mm = vma->vm_mm;
562 unsigned long oldflags = vma->vm_flags;
563 long nrpages = (end - start) >> PAGE_SHIFT;
564 unsigned int mm_cp_flags = 0;
565 unsigned long charged = 0;
566 pgoff_t pgoff;
567 int error;
568
569 if (newflags == oldflags) {
570 *pprev = vma;
571 return 0;
572 }
573
574 /*
575 * Do PROT_NONE PFN permission checks here when we can still
576 * bail out without undoing a lot of state. This is a rather
577 * uncommon case, so doesn't need to be very optimized.
578 */
579 if (arch_has_pfn_modify_check() &&
580 (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
581 (newflags & VM_ACCESS_FLAGS) == 0) {
582 pgprot_t new_pgprot = vm_get_page_prot(newflags);
583
584 error = walk_page_range(current->mm, start, end,
585 &prot_none_walk_ops, &new_pgprot);
586 if (error)
587 return error;
588 }
589
590 /*
591 * If we make a private mapping writable we increase our commit;
592 * but (without finer accounting) cannot reduce our commit if we
593 * make it unwritable again. hugetlb mapping were accounted for
594 * even if read-only so there is no need to account for them here
595 */
596 if (newflags & VM_WRITE) {
597 /* Check space limits when area turns into data. */
598 if (!may_expand_vm(mm, newflags, nrpages) &&
599 may_expand_vm(mm, oldflags, nrpages))
600 return -ENOMEM;
601 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
602 VM_SHARED|VM_NORESERVE))) {
603 charged = nrpages;
604 if (security_vm_enough_memory_mm(mm, charged))
605 return -ENOMEM;
606 newflags |= VM_ACCOUNT;
607 }
608 }
609
610 /*
611 * First try to merge with previous and/or next vma.
612 */
613 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
614 *pprev = vma_merge(mm, *pprev, start, end, newflags,
615 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
616 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
617 if (*pprev) {
618 vma = *pprev;
619 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
620 goto success;
621 }
622
623 *pprev = vma;
624
625 if (start != vma->vm_start) {
626 error = split_vma(mm, vma, start, 1);
627 if (error)
628 goto fail;
629 }
630
631 if (end != vma->vm_end) {
632 error = split_vma(mm, vma, end, 0);
633 if (error)
634 goto fail;
635 }
636
637success:
638 /*
639 * vm_flags and vm_page_prot are protected by the mmap_lock
640 * held in write mode.
641 */
642 vma->vm_flags = newflags;
643 if (vma_wants_manual_pte_write_upgrade(vma))
644 mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
645 vma_set_page_prot(vma);
646
647 change_protection(tlb, vma, start, end, vma->vm_page_prot, mm_cp_flags);
648
649 /*
650 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
651 * fault on access.
652 */
653 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
654 (newflags & VM_WRITE)) {
655 populate_vma_page_range(vma, start, end, NULL);
656 }
657
658 vm_stat_account(mm, oldflags, -nrpages);
659 vm_stat_account(mm, newflags, nrpages);
660 perf_event_mmap(vma);
661 return 0;
662
663fail:
664 vm_unacct_memory(charged);
665 return error;
666}
667
668/*
669 * pkey==-1 when doing a legacy mprotect()
670 */
671static int do_mprotect_pkey(unsigned long start, size_t len,
672 unsigned long prot, int pkey)
673{
674 unsigned long nstart, end, tmp, reqprot;
675 struct vm_area_struct *vma, *prev;
676 int error;
677 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
678 const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
679 (prot & PROT_READ);
680 struct mmu_gather tlb;
681 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
682
683 start = untagged_addr(start);
684
685 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
686 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
687 return -EINVAL;
688
689 if (start & ~PAGE_MASK)
690 return -EINVAL;
691 if (!len)
692 return 0;
693 len = PAGE_ALIGN(len);
694 end = start + len;
695 if (end <= start)
696 return -ENOMEM;
697 if (!arch_validate_prot(prot, start))
698 return -EINVAL;
699
700 reqprot = prot;
701
702 if (mmap_write_lock_killable(current->mm))
703 return -EINTR;
704
705 /*
706 * If userspace did not allocate the pkey, do not let
707 * them use it here.
708 */
709 error = -EINVAL;
710 if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
711 goto out;
712
713 mas_set(&mas, start);
714 vma = mas_find(&mas, ULONG_MAX);
715 error = -ENOMEM;
716 if (!vma)
717 goto out;
718
719 if (unlikely(grows & PROT_GROWSDOWN)) {
720 if (vma->vm_start >= end)
721 goto out;
722 start = vma->vm_start;
723 error = -EINVAL;
724 if (!(vma->vm_flags & VM_GROWSDOWN))
725 goto out;
726 } else {
727 if (vma->vm_start > start)
728 goto out;
729 if (unlikely(grows & PROT_GROWSUP)) {
730 end = vma->vm_end;
731 error = -EINVAL;
732 if (!(vma->vm_flags & VM_GROWSUP))
733 goto out;
734 }
735 }
736
737 if (start > vma->vm_start)
738 prev = vma;
739 else
740 prev = mas_prev(&mas, 0);
741
742 tlb_gather_mmu(&tlb, current->mm);
743 for (nstart = start ; ; ) {
744 unsigned long mask_off_old_flags;
745 unsigned long newflags;
746 int new_vma_pkey;
747
748 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
749
750 /* Does the application expect PROT_READ to imply PROT_EXEC */
751 if (rier && (vma->vm_flags & VM_MAYEXEC))
752 prot |= PROT_EXEC;
753
754 /*
755 * Each mprotect() call explicitly passes r/w/x permissions.
756 * If a permission is not passed to mprotect(), it must be
757 * cleared from the VMA.
758 */
759 mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
760
761 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
762 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
763 newflags |= (vma->vm_flags & ~mask_off_old_flags);
764
765 /* newflags >> 4 shift VM_MAY% in place of VM_% */
766 if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
767 error = -EACCES;
768 break;
769 }
770
771 /* Allow architectures to sanity-check the new flags */
772 if (!arch_validate_flags(newflags)) {
773 error = -EINVAL;
774 break;
775 }
776
777 error = security_file_mprotect(vma, reqprot, prot);
778 if (error)
779 break;
780
781 tmp = vma->vm_end;
782 if (tmp > end)
783 tmp = end;
784
785 if (vma->vm_ops && vma->vm_ops->mprotect) {
786 error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
787 if (error)
788 break;
789 }
790
791 error = mprotect_fixup(&tlb, vma, &prev, nstart, tmp, newflags);
792 if (error)
793 break;
794
795 nstart = tmp;
796
797 if (nstart < prev->vm_end)
798 nstart = prev->vm_end;
799 if (nstart >= end)
800 break;
801
802 vma = find_vma(current->mm, prev->vm_end);
803 if (!vma || vma->vm_start != nstart) {
804 error = -ENOMEM;
805 break;
806 }
807 prot = reqprot;
808 }
809 tlb_finish_mmu(&tlb);
810out:
811 mmap_write_unlock(current->mm);
812 return error;
813}
814
815SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
816 unsigned long, prot)
817{
818 return do_mprotect_pkey(start, len, prot, -1);
819}
820
821#ifdef CONFIG_ARCH_HAS_PKEYS
822
823SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
824 unsigned long, prot, int, pkey)
825{
826 return do_mprotect_pkey(start, len, prot, pkey);
827}
828
829SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
830{
831 int pkey;
832 int ret;
833
834 /* No flags supported yet. */
835 if (flags)
836 return -EINVAL;
837 /* check for unsupported init values */
838 if (init_val & ~PKEY_ACCESS_MASK)
839 return -EINVAL;
840
841 mmap_write_lock(current->mm);
842 pkey = mm_pkey_alloc(current->mm);
843
844 ret = -ENOSPC;
845 if (pkey == -1)
846 goto out;
847
848 ret = arch_set_user_pkey_access(current, pkey, init_val);
849 if (ret) {
850 mm_pkey_free(current->mm, pkey);
851 goto out;
852 }
853 ret = pkey;
854out:
855 mmap_write_unlock(current->mm);
856 return ret;
857}
858
859SYSCALL_DEFINE1(pkey_free, int, pkey)
860{
861 int ret;
862
863 mmap_write_lock(current->mm);
864 ret = mm_pkey_free(current->mm, pkey);
865 mmap_write_unlock(current->mm);
866
867 /*
868 * We could provide warnings or errors if any VMA still
869 * has the pkey set here.
870 */
871 return ret;
872}
873
874#endif /* CONFIG_ARCH_HAS_PKEYS */
1/*
2 * mm/mprotect.c
3 *
4 * (C) Copyright 1994 Linus Torvalds
5 * (C) Copyright 2002 Christoph Hellwig
6 *
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
9 */
10
11#include <linux/mm.h>
12#include <linux/hugetlb.h>
13#include <linux/shm.h>
14#include <linux/mman.h>
15#include <linux/fs.h>
16#include <linux/highmem.h>
17#include <linux/security.h>
18#include <linux/mempolicy.h>
19#include <linux/personality.h>
20#include <linux/syscalls.h>
21#include <linux/swap.h>
22#include <linux/swapops.h>
23#include <linux/mmu_notifier.h>
24#include <linux/migrate.h>
25#include <linux/perf_event.h>
26#include <linux/pkeys.h>
27#include <linux/ksm.h>
28#include <linux/uaccess.h>
29#include <asm/pgtable.h>
30#include <asm/cacheflush.h>
31#include <asm/mmu_context.h>
32#include <asm/tlbflush.h>
33
34#include "internal.h"
35
36/*
37 * For a prot_numa update we only hold mmap_sem for read so there is a
38 * potential race with faulting where a pmd was temporarily none. This
39 * function checks for a transhuge pmd under the appropriate lock. It
40 * returns a pte if it was successfully locked or NULL if it raced with
41 * a transhuge insertion.
42 */
43static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
44 unsigned long addr, int prot_numa, spinlock_t **ptl)
45{
46 pte_t *pte;
47 spinlock_t *pmdl;
48
49 /* !prot_numa is protected by mmap_sem held for write */
50 if (!prot_numa)
51 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
52
53 pmdl = pmd_lock(vma->vm_mm, pmd);
54 if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
55 spin_unlock(pmdl);
56 return NULL;
57 }
58
59 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
60 spin_unlock(pmdl);
61 return pte;
62}
63
64static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
65 unsigned long addr, unsigned long end, pgprot_t newprot,
66 int dirty_accountable, int prot_numa)
67{
68 struct mm_struct *mm = vma->vm_mm;
69 pte_t *pte, oldpte;
70 spinlock_t *ptl;
71 unsigned long pages = 0;
72 int target_node = NUMA_NO_NODE;
73
74 pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
75 if (!pte)
76 return 0;
77
78 /* Get target node for single threaded private VMAs */
79 if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
80 atomic_read(&vma->vm_mm->mm_users) == 1)
81 target_node = numa_node_id();
82
83 arch_enter_lazy_mmu_mode();
84 do {
85 oldpte = *pte;
86 if (pte_present(oldpte)) {
87 pte_t ptent;
88 bool preserve_write = prot_numa && pte_write(oldpte);
89
90 /*
91 * Avoid trapping faults against the zero or KSM
92 * pages. See similar comment in change_huge_pmd.
93 */
94 if (prot_numa) {
95 struct page *page;
96
97 page = vm_normal_page(vma, addr, oldpte);
98 if (!page || PageKsm(page))
99 continue;
100
101 /* Avoid TLB flush if possible */
102 if (pte_protnone(oldpte))
103 continue;
104
105 /*
106 * Don't mess with PTEs if page is already on the node
107 * a single-threaded process is running on.
108 */
109 if (target_node == page_to_nid(page))
110 continue;
111 }
112
113 ptent = ptep_modify_prot_start(mm, addr, pte);
114 ptent = pte_modify(ptent, newprot);
115 if (preserve_write)
116 ptent = pte_mkwrite(ptent);
117
118 /* Avoid taking write faults for known dirty pages */
119 if (dirty_accountable && pte_dirty(ptent) &&
120 (pte_soft_dirty(ptent) ||
121 !(vma->vm_flags & VM_SOFTDIRTY))) {
122 ptent = pte_mkwrite(ptent);
123 }
124 ptep_modify_prot_commit(mm, addr, pte, ptent);
125 pages++;
126 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
127 swp_entry_t entry = pte_to_swp_entry(oldpte);
128
129 if (is_write_migration_entry(entry)) {
130 pte_t newpte;
131 /*
132 * A protection check is difficult so
133 * just be safe and disable write
134 */
135 make_migration_entry_read(&entry);
136 newpte = swp_entry_to_pte(entry);
137 if (pte_swp_soft_dirty(oldpte))
138 newpte = pte_swp_mksoft_dirty(newpte);
139 set_pte_at(mm, addr, pte, newpte);
140
141 pages++;
142 }
143 }
144 } while (pte++, addr += PAGE_SIZE, addr != end);
145 arch_leave_lazy_mmu_mode();
146 pte_unmap_unlock(pte - 1, ptl);
147
148 return pages;
149}
150
151static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
152 pud_t *pud, unsigned long addr, unsigned long end,
153 pgprot_t newprot, int dirty_accountable, int prot_numa)
154{
155 pmd_t *pmd;
156 struct mm_struct *mm = vma->vm_mm;
157 unsigned long next;
158 unsigned long pages = 0;
159 unsigned long nr_huge_updates = 0;
160 unsigned long mni_start = 0;
161
162 pmd = pmd_offset(pud, addr);
163 do {
164 unsigned long this_pages;
165
166 next = pmd_addr_end(addr, end);
167 if (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
168 && pmd_none_or_clear_bad(pmd))
169 continue;
170
171 /* invoke the mmu notifier if the pmd is populated */
172 if (!mni_start) {
173 mni_start = addr;
174 mmu_notifier_invalidate_range_start(mm, mni_start, end);
175 }
176
177 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
178 if (next - addr != HPAGE_PMD_SIZE) {
179 __split_huge_pmd(vma, pmd, addr, false, NULL);
180 if (pmd_trans_unstable(pmd))
181 continue;
182 } else {
183 int nr_ptes = change_huge_pmd(vma, pmd, addr,
184 newprot, prot_numa);
185
186 if (nr_ptes) {
187 if (nr_ptes == HPAGE_PMD_NR) {
188 pages += HPAGE_PMD_NR;
189 nr_huge_updates++;
190 }
191
192 /* huge pmd was handled */
193 continue;
194 }
195 }
196 /* fall through, the trans huge pmd just split */
197 }
198 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
199 dirty_accountable, prot_numa);
200 pages += this_pages;
201 } while (pmd++, addr = next, addr != end);
202
203 if (mni_start)
204 mmu_notifier_invalidate_range_end(mm, mni_start, end);
205
206 if (nr_huge_updates)
207 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
208 return pages;
209}
210
211static inline unsigned long change_pud_range(struct vm_area_struct *vma,
212 pgd_t *pgd, unsigned long addr, unsigned long end,
213 pgprot_t newprot, int dirty_accountable, int prot_numa)
214{
215 pud_t *pud;
216 unsigned long next;
217 unsigned long pages = 0;
218
219 pud = pud_offset(pgd, addr);
220 do {
221 next = pud_addr_end(addr, end);
222 if (pud_none_or_clear_bad(pud))
223 continue;
224 pages += change_pmd_range(vma, pud, addr, next, newprot,
225 dirty_accountable, prot_numa);
226 } while (pud++, addr = next, addr != end);
227
228 return pages;
229}
230
231static unsigned long change_protection_range(struct vm_area_struct *vma,
232 unsigned long addr, unsigned long end, pgprot_t newprot,
233 int dirty_accountable, int prot_numa)
234{
235 struct mm_struct *mm = vma->vm_mm;
236 pgd_t *pgd;
237 unsigned long next;
238 unsigned long start = addr;
239 unsigned long pages = 0;
240
241 BUG_ON(addr >= end);
242 pgd = pgd_offset(mm, addr);
243 flush_cache_range(vma, addr, end);
244 set_tlb_flush_pending(mm);
245 do {
246 next = pgd_addr_end(addr, end);
247 if (pgd_none_or_clear_bad(pgd))
248 continue;
249 pages += change_pud_range(vma, pgd, addr, next, newprot,
250 dirty_accountable, prot_numa);
251 } while (pgd++, addr = next, addr != end);
252
253 /* Only flush the TLB if we actually modified any entries: */
254 if (pages)
255 flush_tlb_range(vma, start, end);
256 clear_tlb_flush_pending(mm);
257
258 return pages;
259}
260
261unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
262 unsigned long end, pgprot_t newprot,
263 int dirty_accountable, int prot_numa)
264{
265 unsigned long pages;
266
267 if (is_vm_hugetlb_page(vma))
268 pages = hugetlb_change_protection(vma, start, end, newprot);
269 else
270 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
271
272 return pages;
273}
274
275int
276mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
277 unsigned long start, unsigned long end, unsigned long newflags)
278{
279 struct mm_struct *mm = vma->vm_mm;
280 unsigned long oldflags = vma->vm_flags;
281 long nrpages = (end - start) >> PAGE_SHIFT;
282 unsigned long charged = 0;
283 pgoff_t pgoff;
284 int error;
285 int dirty_accountable = 0;
286
287 if (newflags == oldflags) {
288 *pprev = vma;
289 return 0;
290 }
291
292 /*
293 * If we make a private mapping writable we increase our commit;
294 * but (without finer accounting) cannot reduce our commit if we
295 * make it unwritable again. hugetlb mapping were accounted for
296 * even if read-only so there is no need to account for them here
297 */
298 if (newflags & VM_WRITE) {
299 /* Check space limits when area turns into data. */
300 if (!may_expand_vm(mm, newflags, nrpages) &&
301 may_expand_vm(mm, oldflags, nrpages))
302 return -ENOMEM;
303 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
304 VM_SHARED|VM_NORESERVE))) {
305 charged = nrpages;
306 if (security_vm_enough_memory_mm(mm, charged))
307 return -ENOMEM;
308 newflags |= VM_ACCOUNT;
309 }
310 }
311
312 /*
313 * First try to merge with previous and/or next vma.
314 */
315 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
316 *pprev = vma_merge(mm, *pprev, start, end, newflags,
317 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
318 vma->vm_userfaultfd_ctx);
319 if (*pprev) {
320 vma = *pprev;
321 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
322 goto success;
323 }
324
325 *pprev = vma;
326
327 if (start != vma->vm_start) {
328 error = split_vma(mm, vma, start, 1);
329 if (error)
330 goto fail;
331 }
332
333 if (end != vma->vm_end) {
334 error = split_vma(mm, vma, end, 0);
335 if (error)
336 goto fail;
337 }
338
339success:
340 /*
341 * vm_flags and vm_page_prot are protected by the mmap_sem
342 * held in write mode.
343 */
344 vma->vm_flags = newflags;
345 dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
346 vma_set_page_prot(vma);
347
348 change_protection(vma, start, end, vma->vm_page_prot,
349 dirty_accountable, 0);
350
351 /*
352 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
353 * fault on access.
354 */
355 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
356 (newflags & VM_WRITE)) {
357 populate_vma_page_range(vma, start, end, NULL);
358 }
359
360 vm_stat_account(mm, oldflags, -nrpages);
361 vm_stat_account(mm, newflags, nrpages);
362 perf_event_mmap(vma);
363 return 0;
364
365fail:
366 vm_unacct_memory(charged);
367 return error;
368}
369
370/*
371 * pkey==-1 when doing a legacy mprotect()
372 */
373static int do_mprotect_pkey(unsigned long start, size_t len,
374 unsigned long prot, int pkey)
375{
376 unsigned long nstart, end, tmp, reqprot;
377 struct vm_area_struct *vma, *prev;
378 int error = -EINVAL;
379 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
380 const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
381 (prot & PROT_READ);
382
383 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
384 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
385 return -EINVAL;
386
387 if (start & ~PAGE_MASK)
388 return -EINVAL;
389 if (!len)
390 return 0;
391 len = PAGE_ALIGN(len);
392 end = start + len;
393 if (end <= start)
394 return -ENOMEM;
395 if (!arch_validate_prot(prot))
396 return -EINVAL;
397
398 reqprot = prot;
399
400 if (down_write_killable(¤t->mm->mmap_sem))
401 return -EINTR;
402
403 /*
404 * If userspace did not allocate the pkey, do not let
405 * them use it here.
406 */
407 error = -EINVAL;
408 if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
409 goto out;
410
411 vma = find_vma(current->mm, start);
412 error = -ENOMEM;
413 if (!vma)
414 goto out;
415 prev = vma->vm_prev;
416 if (unlikely(grows & PROT_GROWSDOWN)) {
417 if (vma->vm_start >= end)
418 goto out;
419 start = vma->vm_start;
420 error = -EINVAL;
421 if (!(vma->vm_flags & VM_GROWSDOWN))
422 goto out;
423 } else {
424 if (vma->vm_start > start)
425 goto out;
426 if (unlikely(grows & PROT_GROWSUP)) {
427 end = vma->vm_end;
428 error = -EINVAL;
429 if (!(vma->vm_flags & VM_GROWSUP))
430 goto out;
431 }
432 }
433 if (start > vma->vm_start)
434 prev = vma;
435
436 for (nstart = start ; ; ) {
437 unsigned long mask_off_old_flags;
438 unsigned long newflags;
439 int new_vma_pkey;
440
441 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
442
443 /* Does the application expect PROT_READ to imply PROT_EXEC */
444 if (rier && (vma->vm_flags & VM_MAYEXEC))
445 prot |= PROT_EXEC;
446
447 /*
448 * Each mprotect() call explicitly passes r/w/x permissions.
449 * If a permission is not passed to mprotect(), it must be
450 * cleared from the VMA.
451 */
452 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
453 ARCH_VM_PKEY_FLAGS;
454
455 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
456 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
457 newflags |= (vma->vm_flags & ~mask_off_old_flags);
458
459 /* newflags >> 4 shift VM_MAY% in place of VM_% */
460 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
461 error = -EACCES;
462 goto out;
463 }
464
465 error = security_file_mprotect(vma, reqprot, prot);
466 if (error)
467 goto out;
468
469 tmp = vma->vm_end;
470 if (tmp > end)
471 tmp = end;
472 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
473 if (error)
474 goto out;
475 nstart = tmp;
476
477 if (nstart < prev->vm_end)
478 nstart = prev->vm_end;
479 if (nstart >= end)
480 goto out;
481
482 vma = prev->vm_next;
483 if (!vma || vma->vm_start != nstart) {
484 error = -ENOMEM;
485 goto out;
486 }
487 prot = reqprot;
488 }
489out:
490 up_write(¤t->mm->mmap_sem);
491 return error;
492}
493
494SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
495 unsigned long, prot)
496{
497 return do_mprotect_pkey(start, len, prot, -1);
498}
499
500#ifdef CONFIG_ARCH_HAS_PKEYS
501
502SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
503 unsigned long, prot, int, pkey)
504{
505 return do_mprotect_pkey(start, len, prot, pkey);
506}
507
508SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
509{
510 int pkey;
511 int ret;
512
513 /* No flags supported yet. */
514 if (flags)
515 return -EINVAL;
516 /* check for unsupported init values */
517 if (init_val & ~PKEY_ACCESS_MASK)
518 return -EINVAL;
519
520 down_write(¤t->mm->mmap_sem);
521 pkey = mm_pkey_alloc(current->mm);
522
523 ret = -ENOSPC;
524 if (pkey == -1)
525 goto out;
526
527 ret = arch_set_user_pkey_access(current, pkey, init_val);
528 if (ret) {
529 mm_pkey_free(current->mm, pkey);
530 goto out;
531 }
532 ret = pkey;
533out:
534 up_write(¤t->mm->mmap_sem);
535 return ret;
536}
537
538SYSCALL_DEFINE1(pkey_free, int, pkey)
539{
540 int ret;
541
542 down_write(¤t->mm->mmap_sem);
543 ret = mm_pkey_free(current->mm, pkey);
544 up_write(¤t->mm->mmap_sem);
545
546 /*
547 * We could provie warnings or errors if any VMA still
548 * has the pkey set here.
549 */
550 return ret;
551}
552
553#endif /* CONFIG_ARCH_HAS_PKEYS */