1 // SPDX-License-Identifier: GPL-2.0
5 * (C) Copyright 1994 Linus Torvalds
6 * (C) Copyright 2002 Christoph Hellwig
8 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
9 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
13 #include <linux/hugetlb.h>
14 #include <linux/shm.h>
15 #include <linux/mman.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 <asm/pgtable.h>
32 #include <asm/cacheflush.h>
33 #include <asm/mmu_context.h>
34 #include <asm/tlbflush.h>
38 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
39 unsigned long addr, unsigned long end, pgprot_t newprot,
40 int dirty_accountable, int prot_numa)
42 struct mm_struct *mm = vma->vm_mm;
45 unsigned long pages = 0;
46 int target_node = NUMA_NO_NODE;
49 * Can be called with only the mmap_sem for reading by
50 * prot_numa so we must check the pmd isn't constantly
51 * changing from under us from pmd_none to pmd_trans_huge
52 * and/or the other way around.
54 if (pmd_trans_unstable(pmd))
58 * The pmd points to a regular pte so the pmd can't change
59 * from under us even if the mmap_sem is only hold for
62 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
64 /* Get target node for single threaded private VMAs */
65 if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
66 atomic_read(&vma->vm_mm->mm_users) == 1)
67 target_node = numa_node_id();
69 flush_tlb_batched_pending(vma->vm_mm);
70 arch_enter_lazy_mmu_mode();
73 if (pte_present(oldpte)) {
75 bool preserve_write = prot_numa && pte_write(oldpte);
78 * Avoid trapping faults against the zero or KSM
79 * pages. See similar comment in change_huge_pmd.
84 page = vm_normal_page(vma, addr, oldpte);
85 if (!page || PageKsm(page))
88 /* Also skip shared copy-on-write pages */
89 if (is_cow_mapping(vma->vm_flags) &&
90 page_mapcount(page) != 1)
94 * While migration can move some dirty pages,
95 * it cannot move them all from MIGRATE_ASYNC
98 if (page_is_file_cache(page) && PageDirty(page))
101 /* Avoid TLB flush if possible */
102 if (pte_protnone(oldpte))
106 * Don't mess with PTEs if page is already on the node
107 * a single-threaded process is running on.
109 if (target_node == page_to_nid(page))
113 ptent = ptep_modify_prot_start(mm, addr, pte);
114 ptent = pte_modify(ptent, newprot);
116 ptent = pte_mk_savedwrite(ptent);
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);
124 ptep_modify_prot_commit(mm, addr, pte, ptent);
126 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
127 swp_entry_t entry = pte_to_swp_entry(oldpte);
129 if (is_write_migration_entry(entry)) {
132 * A protection check is difficult so
133 * just be safe and disable write
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);
144 if (is_write_device_private_entry(entry)) {
148 * We do not preserve soft-dirtiness. See
149 * copy_one_pte() for explanation.
151 make_device_private_entry_read(&entry);
152 newpte = swp_entry_to_pte(entry);
153 set_pte_at(mm, addr, pte, newpte);
158 } while (pte++, addr += PAGE_SIZE, addr != end);
159 arch_leave_lazy_mmu_mode();
160 pte_unmap_unlock(pte - 1, ptl);
166 * Used when setting automatic NUMA hinting protection where it is
167 * critical that a numa hinting PMD is not confused with a bad PMD.
169 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
171 pmd_t pmdval = pmd_read_atomic(pmd);
173 /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
174 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
178 if (pmd_none(pmdval))
180 if (pmd_trans_huge(pmdval))
182 if (unlikely(pmd_bad(pmdval))) {
190 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
191 pud_t *pud, unsigned long addr, unsigned long end,
192 pgprot_t newprot, int dirty_accountable, int prot_numa)
195 struct mm_struct *mm = vma->vm_mm;
197 unsigned long pages = 0;
198 unsigned long nr_huge_updates = 0;
199 unsigned long mni_start = 0;
201 pmd = pmd_offset(pud, addr);
203 unsigned long this_pages;
205 next = pmd_addr_end(addr, end);
208 * Automatic NUMA balancing walks the tables with mmap_sem
209 * held for read. It's possible a parallel update to occur
210 * between pmd_trans_huge() and a pmd_none_or_clear_bad()
211 * check leading to a false positive and clearing.
212 * Hence, it's necessary to atomically read the PMD value
213 * for all the checks.
215 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
216 pmd_none_or_clear_bad_unless_trans_huge(pmd))
219 /* invoke the mmu notifier if the pmd is populated */
222 mmu_notifier_invalidate_range_start(mm, mni_start, end);
225 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
226 if (next - addr != HPAGE_PMD_SIZE) {
227 __split_huge_pmd(vma, pmd, addr, false, NULL);
229 int nr_ptes = change_huge_pmd(vma, pmd, addr,
233 if (nr_ptes == HPAGE_PMD_NR) {
234 pages += HPAGE_PMD_NR;
238 /* huge pmd was handled */
242 /* fall through, the trans huge pmd just split */
244 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
245 dirty_accountable, prot_numa);
249 } while (pmd++, addr = next, addr != end);
252 mmu_notifier_invalidate_range_end(mm, mni_start, end);
255 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
259 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
260 p4d_t *p4d, unsigned long addr, unsigned long end,
261 pgprot_t newprot, int dirty_accountable, int prot_numa)
265 unsigned long pages = 0;
267 pud = pud_offset(p4d, addr);
269 next = pud_addr_end(addr, end);
270 if (pud_none_or_clear_bad(pud))
272 pages += change_pmd_range(vma, pud, addr, next, newprot,
273 dirty_accountable, prot_numa);
274 } while (pud++, addr = next, addr != end);
279 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
280 pgd_t *pgd, unsigned long addr, unsigned long end,
281 pgprot_t newprot, int dirty_accountable, int prot_numa)
285 unsigned long pages = 0;
287 p4d = p4d_offset(pgd, addr);
289 next = p4d_addr_end(addr, end);
290 if (p4d_none_or_clear_bad(p4d))
292 pages += change_pud_range(vma, p4d, addr, next, newprot,
293 dirty_accountable, prot_numa);
294 } while (p4d++, addr = next, addr != end);
299 static unsigned long change_protection_range(struct vm_area_struct *vma,
300 unsigned long addr, unsigned long end, pgprot_t newprot,
301 int dirty_accountable, int prot_numa)
303 struct mm_struct *mm = vma->vm_mm;
306 unsigned long start = addr;
307 unsigned long pages = 0;
310 pgd = pgd_offset(mm, addr);
311 flush_cache_range(vma, addr, end);
312 inc_tlb_flush_pending(mm);
314 next = pgd_addr_end(addr, end);
315 if (pgd_none_or_clear_bad(pgd))
317 pages += change_p4d_range(vma, pgd, addr, next, newprot,
318 dirty_accountable, prot_numa);
319 } while (pgd++, addr = next, addr != end);
321 /* Only flush the TLB if we actually modified any entries: */
323 flush_tlb_range(vma, start, end);
324 dec_tlb_flush_pending(mm);
329 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
330 unsigned long end, pgprot_t newprot,
331 int dirty_accountable, int prot_numa)
335 if (is_vm_hugetlb_page(vma))
336 pages = hugetlb_change_protection(vma, start, end, newprot);
338 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
343 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
344 unsigned long next, struct mm_walk *walk)
346 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
350 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
351 unsigned long addr, unsigned long next,
352 struct mm_walk *walk)
354 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
358 static int prot_none_test(unsigned long addr, unsigned long next,
359 struct mm_walk *walk)
364 static int prot_none_walk(struct vm_area_struct *vma, unsigned long start,
365 unsigned long end, unsigned long newflags)
367 pgprot_t new_pgprot = vm_get_page_prot(newflags);
368 struct mm_walk prot_none_walk = {
369 .pte_entry = prot_none_pte_entry,
370 .hugetlb_entry = prot_none_hugetlb_entry,
371 .test_walk = prot_none_test,
373 .private = &new_pgprot,
376 return walk_page_range(start, end, &prot_none_walk);
380 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
381 unsigned long start, unsigned long end, unsigned long newflags)
383 struct mm_struct *mm = vma->vm_mm;
384 unsigned long oldflags = vma->vm_flags;
385 long nrpages = (end - start) >> PAGE_SHIFT;
386 unsigned long charged = 0;
389 int dirty_accountable = 0;
391 if (newflags == oldflags) {
397 * Do PROT_NONE PFN permission checks here when we can still
398 * bail out without undoing a lot of state. This is a rather
399 * uncommon case, so doesn't need to be very optimized.
401 if (arch_has_pfn_modify_check() &&
402 (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
403 (newflags & (VM_READ|VM_WRITE|VM_EXEC)) == 0) {
404 error = prot_none_walk(vma, start, end, newflags);
410 * If we make a private mapping writable we increase our commit;
411 * but (without finer accounting) cannot reduce our commit if we
412 * make it unwritable again. hugetlb mapping were accounted for
413 * even if read-only so there is no need to account for them here
415 if (newflags & VM_WRITE) {
416 /* Check space limits when area turns into data. */
417 if (!may_expand_vm(mm, newflags, nrpages) &&
418 may_expand_vm(mm, oldflags, nrpages))
420 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
421 VM_SHARED|VM_NORESERVE))) {
423 if (security_vm_enough_memory_mm(mm, charged))
425 newflags |= VM_ACCOUNT;
430 * First try to merge with previous and/or next vma.
432 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
433 *pprev = vma_merge(mm, *pprev, start, end, newflags,
434 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
435 vma->vm_userfaultfd_ctx);
438 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
444 if (start != vma->vm_start) {
445 error = split_vma(mm, vma, start, 1);
450 if (end != vma->vm_end) {
451 error = split_vma(mm, vma, end, 0);
458 * vm_flags and vm_page_prot are protected by the mmap_sem
459 * held in write mode.
461 vma->vm_flags = newflags;
462 dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
463 vma_set_page_prot(vma);
465 change_protection(vma, start, end, vma->vm_page_prot,
466 dirty_accountable, 0);
469 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
472 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
473 (newflags & VM_WRITE)) {
474 populate_vma_page_range(vma, start, end, NULL);
477 vm_stat_account(mm, oldflags, -nrpages);
478 vm_stat_account(mm, newflags, nrpages);
479 perf_event_mmap(vma);
483 vm_unacct_memory(charged);
488 * pkey==-1 when doing a legacy mprotect()
490 static int do_mprotect_pkey(unsigned long start, size_t len,
491 unsigned long prot, int pkey)
493 unsigned long nstart, end, tmp, reqprot;
494 struct vm_area_struct *vma, *prev;
496 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
497 const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
500 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
501 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
504 if (start & ~PAGE_MASK)
508 len = PAGE_ALIGN(len);
512 if (!arch_validate_prot(prot, start))
517 if (down_write_killable(¤t->mm->mmap_sem))
521 * If userspace did not allocate the pkey, do not let
525 if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
528 vma = find_vma(current->mm, start);
533 if (unlikely(grows & PROT_GROWSDOWN)) {
534 if (vma->vm_start >= end)
536 start = vma->vm_start;
538 if (!(vma->vm_flags & VM_GROWSDOWN))
541 if (vma->vm_start > start)
543 if (unlikely(grows & PROT_GROWSUP)) {
546 if (!(vma->vm_flags & VM_GROWSUP))
550 if (start > vma->vm_start)
553 for (nstart = start ; ; ) {
554 unsigned long mask_off_old_flags;
555 unsigned long newflags;
558 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
560 /* Does the application expect PROT_READ to imply PROT_EXEC */
561 if (rier && (vma->vm_flags & VM_MAYEXEC))
565 * Each mprotect() call explicitly passes r/w/x permissions.
566 * If a permission is not passed to mprotect(), it must be
567 * cleared from the VMA.
569 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
572 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
573 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
574 newflags |= (vma->vm_flags & ~mask_off_old_flags);
576 /* newflags >> 4 shift VM_MAY% in place of VM_% */
577 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
582 error = security_file_mprotect(vma, reqprot, prot);
589 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
594 if (nstart < prev->vm_end)
595 nstart = prev->vm_end;
600 if (!vma || vma->vm_start != nstart) {
607 up_write(¤t->mm->mmap_sem);
611 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
614 return do_mprotect_pkey(start, len, prot, -1);
617 #ifdef CONFIG_ARCH_HAS_PKEYS
619 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
620 unsigned long, prot, int, pkey)
622 return do_mprotect_pkey(start, len, prot, pkey);
625 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
630 /* No flags supported yet. */
633 /* check for unsupported init values */
634 if (init_val & ~PKEY_ACCESS_MASK)
637 down_write(¤t->mm->mmap_sem);
638 pkey = mm_pkey_alloc(current->mm);
644 ret = arch_set_user_pkey_access(current, pkey, init_val);
646 mm_pkey_free(current->mm, pkey);
651 up_write(¤t->mm->mmap_sem);
655 SYSCALL_DEFINE1(pkey_free, int, pkey)
659 down_write(¤t->mm->mmap_sem);
660 ret = mm_pkey_free(current->mm, pkey);
661 up_write(¤t->mm->mmap_sem);
664 * We could provie warnings or errors if any VMA still
665 * has the pkey set here.
670 #endif /* CONFIG_ARCH_HAS_PKEYS */