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 <asm/pgtable.h>
31 #include <asm/cacheflush.h>
32 #include <asm/mmu_context.h>
33 #include <asm/tlbflush.h>
37 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
38 unsigned long addr, unsigned long end, pgprot_t newprot,
39 int dirty_accountable, int prot_numa)
41 struct mm_struct *mm = vma->vm_mm;
44 unsigned long pages = 0;
45 int target_node = NUMA_NO_NODE;
48 * Can be called with only the mmap_sem for reading by
49 * prot_numa so we must check the pmd isn't constantly
50 * changing from under us from pmd_none to pmd_trans_huge
51 * and/or the other way around.
53 if (pmd_trans_unstable(pmd))
57 * The pmd points to a regular pte so the pmd can't change
58 * from under us even if the mmap_sem is only hold for
61 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
63 /* Get target node for single threaded private VMAs */
64 if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
65 atomic_read(&vma->vm_mm->mm_users) == 1)
66 target_node = numa_node_id();
68 flush_tlb_batched_pending(vma->vm_mm);
69 arch_enter_lazy_mmu_mode();
72 if (pte_present(oldpte)) {
74 bool preserve_write = prot_numa && pte_write(oldpte);
77 * Avoid trapping faults against the zero or KSM
78 * pages. See similar comment in change_huge_pmd.
83 page = vm_normal_page(vma, addr, oldpte);
84 if (!page || PageKsm(page))
87 /* Avoid TLB flush if possible */
88 if (pte_protnone(oldpte))
92 * Don't mess with PTEs if page is already on the node
93 * a single-threaded process is running on.
95 if (target_node == page_to_nid(page))
99 ptent = ptep_modify_prot_start(mm, addr, pte);
100 ptent = pte_modify(ptent, newprot);
102 ptent = pte_mk_savedwrite(ptent);
104 /* Avoid taking write faults for known dirty pages */
105 if (dirty_accountable && pte_dirty(ptent) &&
106 (pte_soft_dirty(ptent) ||
107 !(vma->vm_flags & VM_SOFTDIRTY))) {
108 ptent = pte_mkwrite(ptent);
110 ptep_modify_prot_commit(mm, addr, pte, ptent);
112 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
113 swp_entry_t entry = pte_to_swp_entry(oldpte);
115 if (is_write_migration_entry(entry)) {
118 * A protection check is difficult so
119 * just be safe and disable write
121 make_migration_entry_read(&entry);
122 newpte = swp_entry_to_pte(entry);
123 if (pte_swp_soft_dirty(oldpte))
124 newpte = pte_swp_mksoft_dirty(newpte);
125 set_pte_at(mm, addr, pte, newpte);
130 if (is_write_device_private_entry(entry)) {
134 * We do not preserve soft-dirtiness. See
135 * copy_one_pte() for explanation.
137 make_device_private_entry_read(&entry);
138 newpte = swp_entry_to_pte(entry);
139 set_pte_at(mm, addr, pte, newpte);
144 } while (pte++, addr += PAGE_SIZE, addr != end);
145 arch_leave_lazy_mmu_mode();
146 pte_unmap_unlock(pte - 1, ptl);
152 * Used when setting automatic NUMA hinting protection where it is
153 * critical that a numa hinting PMD is not confused with a bad PMD.
155 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
157 pmd_t pmdval = pmd_read_atomic(pmd);
159 /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
160 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
164 if (pmd_none(pmdval))
166 if (pmd_trans_huge(pmdval))
168 if (unlikely(pmd_bad(pmdval))) {
176 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
177 pud_t *pud, unsigned long addr, unsigned long end,
178 pgprot_t newprot, int dirty_accountable, int prot_numa)
181 struct mm_struct *mm = vma->vm_mm;
183 unsigned long pages = 0;
184 unsigned long nr_huge_updates = 0;
185 unsigned long mni_start = 0;
187 pmd = pmd_offset(pud, addr);
189 unsigned long this_pages;
191 next = pmd_addr_end(addr, end);
194 * Automatic NUMA balancing walks the tables with mmap_sem
195 * held for read. It's possible a parallel update to occur
196 * between pmd_trans_huge() and a pmd_none_or_clear_bad()
197 * check leading to a false positive and clearing.
198 * Hence, it's necessary to atomically read the PMD value
199 * for all the checks.
201 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
202 pmd_none_or_clear_bad_unless_trans_huge(pmd))
205 /* invoke the mmu notifier if the pmd is populated */
208 mmu_notifier_invalidate_range_start(mm, mni_start, end);
211 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
212 if (next - addr != HPAGE_PMD_SIZE) {
213 __split_huge_pmd(vma, pmd, addr, false, NULL);
215 int nr_ptes = change_huge_pmd(vma, pmd, addr,
219 if (nr_ptes == HPAGE_PMD_NR) {
220 pages += HPAGE_PMD_NR;
224 /* huge pmd was handled */
228 /* fall through, the trans huge pmd just split */
230 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
231 dirty_accountable, prot_numa);
235 } while (pmd++, addr = next, addr != end);
238 mmu_notifier_invalidate_range_end(mm, mni_start, end);
241 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
245 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
246 p4d_t *p4d, unsigned long addr, unsigned long end,
247 pgprot_t newprot, int dirty_accountable, int prot_numa)
251 unsigned long pages = 0;
253 pud = pud_offset(p4d, addr);
255 next = pud_addr_end(addr, end);
256 if (pud_none_or_clear_bad(pud))
258 pages += change_pmd_range(vma, pud, addr, next, newprot,
259 dirty_accountable, prot_numa);
260 } while (pud++, addr = next, addr != end);
265 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
266 pgd_t *pgd, unsigned long addr, unsigned long end,
267 pgprot_t newprot, int dirty_accountable, int prot_numa)
271 unsigned long pages = 0;
273 p4d = p4d_offset(pgd, addr);
275 next = p4d_addr_end(addr, end);
276 if (p4d_none_or_clear_bad(p4d))
278 pages += change_pud_range(vma, p4d, addr, next, newprot,
279 dirty_accountable, prot_numa);
280 } while (p4d++, addr = next, addr != end);
285 static unsigned long change_protection_range(struct vm_area_struct *vma,
286 unsigned long addr, unsigned long end, pgprot_t newprot,
287 int dirty_accountable, int prot_numa)
289 struct mm_struct *mm = vma->vm_mm;
292 unsigned long start = addr;
293 unsigned long pages = 0;
296 pgd = pgd_offset(mm, addr);
297 flush_cache_range(vma, addr, end);
298 inc_tlb_flush_pending(mm);
300 next = pgd_addr_end(addr, end);
301 if (pgd_none_or_clear_bad(pgd))
303 pages += change_p4d_range(vma, pgd, addr, next, newprot,
304 dirty_accountable, prot_numa);
305 } while (pgd++, addr = next, addr != end);
307 /* Only flush the TLB if we actually modified any entries: */
309 flush_tlb_range(vma, start, end);
310 dec_tlb_flush_pending(mm);
315 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
316 unsigned long end, pgprot_t newprot,
317 int dirty_accountable, int prot_numa)
321 if (is_vm_hugetlb_page(vma))
322 pages = hugetlb_change_protection(vma, start, end, newprot);
324 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
329 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
330 unsigned long next, struct mm_walk *walk)
332 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
336 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
337 unsigned long addr, unsigned long next,
338 struct mm_walk *walk)
340 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
344 static int prot_none_test(unsigned long addr, unsigned long next,
345 struct mm_walk *walk)
350 static int prot_none_walk(struct vm_area_struct *vma, unsigned long start,
351 unsigned long end, unsigned long newflags)
353 pgprot_t new_pgprot = vm_get_page_prot(newflags);
354 struct mm_walk prot_none_walk = {
355 .pte_entry = prot_none_pte_entry,
356 .hugetlb_entry = prot_none_hugetlb_entry,
357 .test_walk = prot_none_test,
359 .private = &new_pgprot,
362 return walk_page_range(start, end, &prot_none_walk);
366 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
367 unsigned long start, unsigned long end, unsigned long newflags)
369 struct mm_struct *mm = vma->vm_mm;
370 unsigned long oldflags = vma->vm_flags;
371 long nrpages = (end - start) >> PAGE_SHIFT;
372 unsigned long charged = 0;
375 int dirty_accountable = 0;
377 if (newflags == oldflags) {
383 * Do PROT_NONE PFN permission checks here when we can still
384 * bail out without undoing a lot of state. This is a rather
385 * uncommon case, so doesn't need to be very optimized.
387 if (arch_has_pfn_modify_check() &&
388 (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
389 (newflags & (VM_READ|VM_WRITE|VM_EXEC)) == 0) {
390 error = prot_none_walk(vma, start, end, newflags);
396 * If we make a private mapping writable we increase our commit;
397 * but (without finer accounting) cannot reduce our commit if we
398 * make it unwritable again. hugetlb mapping were accounted for
399 * even if read-only so there is no need to account for them here
401 if (newflags & VM_WRITE) {
402 /* Check space limits when area turns into data. */
403 if (!may_expand_vm(mm, newflags, nrpages) &&
404 may_expand_vm(mm, oldflags, nrpages))
406 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
407 VM_SHARED|VM_NORESERVE))) {
409 if (security_vm_enough_memory_mm(mm, charged))
411 newflags |= VM_ACCOUNT;
416 * First try to merge with previous and/or next vma.
418 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
419 *pprev = vma_merge(mm, *pprev, start, end, newflags,
420 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
421 vma->vm_userfaultfd_ctx);
424 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
430 if (start != vma->vm_start) {
431 error = split_vma(mm, vma, start, 1);
436 if (end != vma->vm_end) {
437 error = split_vma(mm, vma, end, 0);
444 * vm_flags and vm_page_prot are protected by the mmap_sem
445 * held in write mode.
447 vma->vm_flags = newflags;
448 dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
449 vma_set_page_prot(vma);
451 change_protection(vma, start, end, vma->vm_page_prot,
452 dirty_accountable, 0);
455 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
458 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
459 (newflags & VM_WRITE)) {
460 populate_vma_page_range(vma, start, end, NULL);
463 vm_stat_account(mm, oldflags, -nrpages);
464 vm_stat_account(mm, newflags, nrpages);
465 perf_event_mmap(vma);
469 vm_unacct_memory(charged);
474 * pkey==-1 when doing a legacy mprotect()
476 static int do_mprotect_pkey(unsigned long start, size_t len,
477 unsigned long prot, int pkey)
479 unsigned long nstart, end, tmp, reqprot;
480 struct vm_area_struct *vma, *prev;
482 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
483 const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
486 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
487 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
490 if (start & ~PAGE_MASK)
494 len = PAGE_ALIGN(len);
498 if (!arch_validate_prot(prot))
503 if (down_write_killable(¤t->mm->mmap_sem))
507 * If userspace did not allocate the pkey, do not let
511 if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
514 vma = find_vma(current->mm, start);
519 if (unlikely(grows & PROT_GROWSDOWN)) {
520 if (vma->vm_start >= end)
522 start = vma->vm_start;
524 if (!(vma->vm_flags & VM_GROWSDOWN))
527 if (vma->vm_start > start)
529 if (unlikely(grows & PROT_GROWSUP)) {
532 if (!(vma->vm_flags & VM_GROWSUP))
536 if (start > vma->vm_start)
539 for (nstart = start ; ; ) {
540 unsigned long mask_off_old_flags;
541 unsigned long newflags;
544 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
546 /* Does the application expect PROT_READ to imply PROT_EXEC */
547 if (rier && (vma->vm_flags & VM_MAYEXEC))
551 * Each mprotect() call explicitly passes r/w/x permissions.
552 * If a permission is not passed to mprotect(), it must be
553 * cleared from the VMA.
555 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
558 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
559 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
560 newflags |= (vma->vm_flags & ~mask_off_old_flags);
562 /* newflags >> 4 shift VM_MAY% in place of VM_% */
563 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
568 error = security_file_mprotect(vma, reqprot, prot);
575 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
580 if (nstart < prev->vm_end)
581 nstart = prev->vm_end;
586 if (!vma || vma->vm_start != nstart) {
593 up_write(¤t->mm->mmap_sem);
597 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
600 return do_mprotect_pkey(start, len, prot, -1);
603 #ifdef CONFIG_ARCH_HAS_PKEYS
605 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
606 unsigned long, prot, int, pkey)
608 return do_mprotect_pkey(start, len, prot, pkey);
611 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
616 /* No flags supported yet. */
619 /* check for unsupported init values */
620 if (init_val & ~PKEY_ACCESS_MASK)
623 down_write(¤t->mm->mmap_sem);
624 pkey = mm_pkey_alloc(current->mm);
630 ret = arch_set_user_pkey_access(current, pkey, init_val);
632 mm_pkey_free(current->mm, pkey);
637 up_write(¤t->mm->mmap_sem);
641 SYSCALL_DEFINE1(pkey_free, int, pkey)
645 down_write(¤t->mm->mmap_sem);
646 ret = mm_pkey_free(current->mm, pkey);
647 up_write(¤t->mm->mmap_sem);
650 * We could provie warnings or errors if any VMA still
651 * has the pkey set here.
656 #endif /* CONFIG_ARCH_HAS_PKEYS */