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/ksm.h>
27 #include <asm/uaccess.h>
28 #include <asm/pgtable.h>
29 #include <asm/cacheflush.h>
30 #include <asm/tlbflush.h>
35 * For a prot_numa update we only hold mmap_sem for read so there is a
36 * potential race with faulting where a pmd was temporarily none. This
37 * function checks for a transhuge pmd under the appropriate lock. It
38 * returns a pte if it was successfully locked or NULL if it raced with
39 * a transhuge insertion.
41 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
42 unsigned long addr, int prot_numa, spinlock_t **ptl)
47 /* !prot_numa is protected by mmap_sem held for write */
49 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
51 pmdl = pmd_lock(vma->vm_mm, pmd);
52 if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
57 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
62 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
63 unsigned long addr, unsigned long end, pgprot_t newprot,
64 int dirty_accountable, int prot_numa)
66 struct mm_struct *mm = vma->vm_mm;
69 unsigned long pages = 0;
71 pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
75 flush_tlb_batched_pending(vma->vm_mm);
76 arch_enter_lazy_mmu_mode();
79 if (pte_present(oldpte)) {
81 bool preserve_write = prot_numa && pte_write(oldpte);
84 * Avoid trapping faults against the zero or KSM
85 * pages. See similar comment in change_huge_pmd.
90 page = vm_normal_page(vma, addr, oldpte);
91 if (!page || PageKsm(page))
94 /* Avoid TLB flush if possible */
95 if (pte_protnone(oldpte))
99 ptent = ptep_modify_prot_start(mm, addr, pte);
100 ptent = pte_modify(ptent, newprot);
102 ptent = pte_mkwrite(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 } while (pte++, addr += PAGE_SIZE, addr != end);
131 arch_leave_lazy_mmu_mode();
132 pte_unmap_unlock(pte - 1, ptl);
137 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
138 pud_t *pud, unsigned long addr, unsigned long end,
139 pgprot_t newprot, int dirty_accountable, int prot_numa)
142 struct mm_struct *mm = vma->vm_mm;
144 unsigned long pages = 0;
145 unsigned long nr_huge_updates = 0;
146 unsigned long mni_start = 0;
148 pmd = pmd_offset(pud, addr);
150 unsigned long this_pages;
152 next = pmd_addr_end(addr, end);
153 if (!pmd_trans_huge(*pmd) && pmd_none_or_clear_bad(pmd))
156 /* invoke the mmu notifier if the pmd is populated */
159 mmu_notifier_invalidate_range_start(mm, mni_start, end);
162 if (pmd_trans_huge(*pmd)) {
163 if (next - addr != HPAGE_PMD_SIZE)
164 split_huge_page_pmd(vma, addr, pmd);
166 int nr_ptes = change_huge_pmd(vma, pmd, addr,
170 if (nr_ptes == HPAGE_PMD_NR) {
171 pages += HPAGE_PMD_NR;
175 /* huge pmd was handled */
179 /* fall through, the trans huge pmd just split */
181 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
182 dirty_accountable, prot_numa);
184 } while (pmd++, addr = next, addr != end);
187 mmu_notifier_invalidate_range_end(mm, mni_start, end);
190 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
194 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
195 pgd_t *pgd, unsigned long addr, unsigned long end,
196 pgprot_t newprot, int dirty_accountable, int prot_numa)
200 unsigned long pages = 0;
202 pud = pud_offset(pgd, addr);
204 next = pud_addr_end(addr, end);
205 if (pud_none_or_clear_bad(pud))
207 pages += change_pmd_range(vma, pud, addr, next, newprot,
208 dirty_accountable, prot_numa);
209 } while (pud++, addr = next, addr != end);
214 static unsigned long change_protection_range(struct vm_area_struct *vma,
215 unsigned long addr, unsigned long end, pgprot_t newprot,
216 int dirty_accountable, int prot_numa)
218 struct mm_struct *mm = vma->vm_mm;
221 unsigned long start = addr;
222 unsigned long pages = 0;
225 pgd = pgd_offset(mm, addr);
226 flush_cache_range(vma, addr, end);
227 set_tlb_flush_pending(mm);
229 next = pgd_addr_end(addr, end);
230 if (pgd_none_or_clear_bad(pgd))
232 pages += change_pud_range(vma, pgd, addr, next, newprot,
233 dirty_accountable, prot_numa);
234 } while (pgd++, addr = next, addr != end);
236 /* Only flush the TLB if we actually modified any entries: */
238 flush_tlb_range(vma, start, end);
239 clear_tlb_flush_pending(mm);
244 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
245 unsigned long end, pgprot_t newprot,
246 int dirty_accountable, int prot_numa)
250 if (is_vm_hugetlb_page(vma))
251 pages = hugetlb_change_protection(vma, start, end, newprot);
253 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
258 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
259 unsigned long next, struct mm_walk *walk)
261 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
265 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
266 unsigned long addr, unsigned long next,
267 struct mm_walk *walk)
269 return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
273 static int prot_none_test(unsigned long addr, unsigned long next,
274 struct mm_walk *walk)
279 static int prot_none_walk(struct vm_area_struct *vma, unsigned long start,
280 unsigned long end, unsigned long newflags)
282 pgprot_t new_pgprot = vm_get_page_prot(newflags);
283 struct mm_walk prot_none_walk = {
284 .pte_entry = prot_none_pte_entry,
285 .hugetlb_entry = prot_none_hugetlb_entry,
286 .test_walk = prot_none_test,
288 .private = &new_pgprot,
291 return walk_page_range(start, end, &prot_none_walk);
295 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
296 unsigned long start, unsigned long end, unsigned long newflags)
298 struct mm_struct *mm = vma->vm_mm;
299 unsigned long oldflags = vma->vm_flags;
300 long nrpages = (end - start) >> PAGE_SHIFT;
301 unsigned long charged = 0;
304 int dirty_accountable = 0;
306 if (newflags == oldflags) {
312 * Do PROT_NONE PFN permission checks here when we can still
313 * bail out without undoing a lot of state. This is a rather
314 * uncommon case, so doesn't need to be very optimized.
316 if (arch_has_pfn_modify_check() &&
317 (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
318 (newflags & (VM_READ|VM_WRITE|VM_EXEC)) == 0) {
319 error = prot_none_walk(vma, start, end, newflags);
325 * If we make a private mapping writable we increase our commit;
326 * but (without finer accounting) cannot reduce our commit if we
327 * make it unwritable again. hugetlb mapping were accounted for
328 * even if read-only so there is no need to account for them here
330 if (newflags & VM_WRITE) {
331 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
332 VM_SHARED|VM_NORESERVE))) {
334 if (security_vm_enough_memory_mm(mm, charged))
336 newflags |= VM_ACCOUNT;
341 * First try to merge with previous and/or next vma.
343 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
344 *pprev = vma_merge(mm, *pprev, start, end, newflags,
345 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
346 vma->vm_userfaultfd_ctx, vma_get_anon_name(vma));
354 if (start != vma->vm_start) {
355 error = split_vma(mm, vma, start, 1);
360 if (end != vma->vm_end) {
361 error = split_vma(mm, vma, end, 0);
368 * vm_flags and vm_page_prot are protected by the mmap_sem
369 * held in write mode.
371 vma->vm_flags = newflags;
372 dirty_accountable = vma_wants_writenotify(vma);
373 vma_set_page_prot(vma);
375 change_protection(vma, start, end, vma->vm_page_prot,
376 dirty_accountable, 0);
379 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
382 if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
383 (newflags & VM_WRITE)) {
384 populate_vma_page_range(vma, start, end, NULL);
387 vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
388 vm_stat_account(mm, newflags, vma->vm_file, nrpages);
389 perf_event_mmap(vma);
393 vm_unacct_memory(charged);
397 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
400 unsigned long vm_flags, nstart, end, tmp, reqprot;
401 struct vm_area_struct *vma, *prev;
403 const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
404 prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
405 if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
408 if (start & ~PAGE_MASK)
412 len = PAGE_ALIGN(len);
416 if (!arch_validate_prot(prot))
421 * Does the application expect PROT_READ to imply PROT_EXEC:
423 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
426 vm_flags = calc_vm_prot_bits(prot);
428 down_write(¤t->mm->mmap_sem);
430 vma = find_vma(current->mm, start);
435 if (unlikely(grows & PROT_GROWSDOWN)) {
436 if (vma->vm_start >= end)
438 start = vma->vm_start;
440 if (!(vma->vm_flags & VM_GROWSDOWN))
443 if (vma->vm_start > start)
445 if (unlikely(grows & PROT_GROWSUP)) {
448 if (!(vma->vm_flags & VM_GROWSUP))
452 if (start > vma->vm_start)
455 for (nstart = start ; ; ) {
456 unsigned long newflags;
458 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
461 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
463 /* newflags >> 4 shift VM_MAY% in place of VM_% */
464 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
469 error = security_file_mprotect(vma, reqprot, prot);
476 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
481 if (nstart < prev->vm_end)
482 nstart = prev->vm_end;
487 if (!vma || vma->vm_start != nstart) {
493 up_write(¤t->mm->mmap_sem);