4 * (C) Copyright 1994 Linus Torvalds
5 * (C) Copyright 2002 Christoph Hellwig
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
12 #include <linux/hugetlb.h>
13 #include <linux/shm.h>
14 #include <linux/mman.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>
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.
43 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
44 unsigned long addr, int prot_numa, spinlock_t **ptl)
49 /* !prot_numa is protected by mmap_sem held for write */
51 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
53 pmdl = pmd_lock(vma->vm_mm, pmd);
54 if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
59 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
64 static 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)
68 struct mm_struct *mm = vma->vm_mm;
71 unsigned long pages = 0;
72 int target_node = NUMA_NO_NODE;
74 pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
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();
83 arch_enter_lazy_mmu_mode();
86 if (pte_present(oldpte)) {
88 bool preserve_write = prot_numa && pte_write(oldpte);
91 * Avoid trapping faults against the zero or KSM
92 * pages. See similar comment in change_huge_pmd.
97 page = vm_normal_page(vma, addr, oldpte);
98 if (!page || PageKsm(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_mkwrite(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 } while (pte++, addr += PAGE_SIZE, addr != end);
145 arch_leave_lazy_mmu_mode();
146 pte_unmap_unlock(pte - 1, ptl);
151 static 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)
156 struct mm_struct *mm = vma->vm_mm;
158 unsigned long pages = 0;
159 unsigned long nr_huge_updates = 0;
160 unsigned long mni_start = 0;
162 pmd = pmd_offset(pud, addr);
164 unsigned long this_pages;
166 next = pmd_addr_end(addr, end);
167 if (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
168 && pmd_none_or_clear_bad(pmd))
171 /* invoke the mmu notifier if the pmd is populated */
174 mmu_notifier_invalidate_range_start(mm, mni_start, end);
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))
183 int nr_ptes = change_huge_pmd(vma, pmd, addr,
187 if (nr_ptes == HPAGE_PMD_NR) {
188 pages += HPAGE_PMD_NR;
192 /* huge pmd was handled */
196 /* fall through, the trans huge pmd just split */
198 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
199 dirty_accountable, prot_numa);
201 } while (pmd++, addr = next, addr != end);
204 mmu_notifier_invalidate_range_end(mm, mni_start, end);
207 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
211 static 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)
217 unsigned long pages = 0;
219 pud = pud_offset(pgd, addr);
221 next = pud_addr_end(addr, end);
222 if (pud_none_or_clear_bad(pud))
224 pages += change_pmd_range(vma, pud, addr, next, newprot,
225 dirty_accountable, prot_numa);
226 } while (pud++, addr = next, addr != end);
231 static 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)
235 struct mm_struct *mm = vma->vm_mm;
238 unsigned long start = addr;
239 unsigned long pages = 0;
242 pgd = pgd_offset(mm, addr);
243 flush_cache_range(vma, addr, end);
244 set_tlb_flush_pending(mm);
246 next = pgd_addr_end(addr, end);
247 if (pgd_none_or_clear_bad(pgd))
249 pages += change_pud_range(vma, pgd, addr, next, newprot,
250 dirty_accountable, prot_numa);
251 } while (pgd++, addr = next, addr != end);
253 /* Only flush the TLB if we actually modified any entries: */
255 flush_tlb_range(vma, start, end);
256 clear_tlb_flush_pending(mm);
261 unsigned 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)
267 if (is_vm_hugetlb_page(vma))
268 pages = hugetlb_change_protection(vma, start, end, newprot);
270 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
276 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
277 unsigned long start, unsigned long end, unsigned long newflags)
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;
285 int dirty_accountable = 0;
287 if (newflags == oldflags) {
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
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))
303 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
304 VM_SHARED|VM_NORESERVE))) {
306 if (security_vm_enough_memory_mm(mm, charged))
308 newflags |= VM_ACCOUNT;
313 * First try to merge with previous and/or next vma.
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);
321 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
327 if (start != vma->vm_start) {
328 error = split_vma(mm, vma, start, 1);
333 if (end != vma->vm_end) {
334 error = split_vma(mm, vma, end, 0);
341 * vm_flags and vm_page_prot are protected by the mmap_sem
342 * held in write mode.
344 vma->vm_flags = newflags;
345 dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
346 vma_set_page_prot(vma);
348 change_protection(vma, start, end, vma->vm_page_prot,
349 dirty_accountable, 0);
352 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
355 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
356 (newflags & VM_WRITE)) {
357 populate_vma_page_range(vma, start, end, NULL);
360 vm_stat_account(mm, oldflags, -nrpages);
361 vm_stat_account(mm, newflags, nrpages);
362 perf_event_mmap(vma);
366 vm_unacct_memory(charged);
371 * pkey==-1 when doing a legacy mprotect()
373 static int do_mprotect_pkey(unsigned long start, size_t len,
374 unsigned long prot, int pkey)
376 unsigned long nstart, end, tmp, reqprot;
377 struct vm_area_struct *vma, *prev;
379 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
380 const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
383 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
384 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
387 if (start & ~PAGE_MASK)
391 len = PAGE_ALIGN(len);
395 if (!arch_validate_prot(prot))
400 if (down_write_killable(¤t->mm->mmap_sem))
404 * If userspace did not allocate the pkey, do not let
408 if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
411 vma = find_vma(current->mm, start);
416 if (unlikely(grows & PROT_GROWSDOWN)) {
417 if (vma->vm_start >= end)
419 start = vma->vm_start;
421 if (!(vma->vm_flags & VM_GROWSDOWN))
424 if (vma->vm_start > start)
426 if (unlikely(grows & PROT_GROWSUP)) {
429 if (!(vma->vm_flags & VM_GROWSUP))
433 if (start > vma->vm_start)
436 for (nstart = start ; ; ) {
437 unsigned long mask_off_old_flags;
438 unsigned long newflags;
441 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
443 /* Does the application expect PROT_READ to imply PROT_EXEC */
444 if (rier && (vma->vm_flags & VM_MAYEXEC))
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.
452 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
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);
459 /* newflags >> 4 shift VM_MAY% in place of VM_% */
460 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
465 error = security_file_mprotect(vma, reqprot, prot);
472 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
477 if (nstart < prev->vm_end)
478 nstart = prev->vm_end;
483 if (!vma || vma->vm_start != nstart) {
490 up_write(¤t->mm->mmap_sem);
494 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
497 return do_mprotect_pkey(start, len, prot, -1);
500 #ifdef CONFIG_ARCH_HAS_PKEYS
502 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
503 unsigned long, prot, int, pkey)
505 return do_mprotect_pkey(start, len, prot, pkey);
508 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
513 /* No flags supported yet. */
516 /* check for unsupported init values */
517 if (init_val & ~PKEY_ACCESS_MASK)
520 down_write(¤t->mm->mmap_sem);
521 pkey = mm_pkey_alloc(current->mm);
527 ret = arch_set_user_pkey_access(current, pkey, init_val);
529 mm_pkey_free(current->mm, pkey);
534 up_write(¤t->mm->mmap_sem);
538 SYSCALL_DEFINE1(pkey_free, int, pkey)
542 down_write(¤t->mm->mmap_sem);
543 ret = mm_pkey_free(current->mm, pkey);
544 up_write(¤t->mm->mmap_sem);
547 * We could provie warnings or errors if any VMA still
548 * has the pkey set here.
553 #endif /* CONFIG_ARCH_HAS_PKEYS */