mmu-notifiers: core
[linux-2.6.git] / mm / mprotect.c
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@redhat.com>
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/slab.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 <asm/uaccess.h>
26 #include <asm/pgtable.h>
27 #include <asm/cacheflush.h>
28 #include <asm/tlbflush.h>
29
30 #ifndef pgprot_modify
31 static inline pgprot_t pgprot_modify(pgprot_t oldprot, pgprot_t newprot)
32 {
33         return newprot;
34 }
35 #endif
36
37 static void change_pte_range(struct mm_struct *mm, pmd_t *pmd,
38                 unsigned long addr, unsigned long end, pgprot_t newprot,
39                 int dirty_accountable)
40 {
41         pte_t *pte, oldpte;
42         spinlock_t *ptl;
43
44         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
45         arch_enter_lazy_mmu_mode();
46         do {
47                 oldpte = *pte;
48                 if (pte_present(oldpte)) {
49                         pte_t ptent;
50
51                         ptent = ptep_modify_prot_start(mm, addr, pte);
52                         ptent = pte_modify(ptent, newprot);
53
54                         /*
55                          * Avoid taking write faults for pages we know to be
56                          * dirty.
57                          */
58                         if (dirty_accountable && pte_dirty(ptent))
59                                 ptent = pte_mkwrite(ptent);
60
61                         ptep_modify_prot_commit(mm, addr, pte, ptent);
62 #ifdef CONFIG_MIGRATION
63                 } else if (!pte_file(oldpte)) {
64                         swp_entry_t entry = pte_to_swp_entry(oldpte);
65
66                         if (is_write_migration_entry(entry)) {
67                                 /*
68                                  * A protection check is difficult so
69                                  * just be safe and disable write
70                                  */
71                                 make_migration_entry_read(&entry);
72                                 set_pte_at(mm, addr, pte,
73                                         swp_entry_to_pte(entry));
74                         }
75 #endif
76                 }
77
78         } while (pte++, addr += PAGE_SIZE, addr != end);
79         arch_leave_lazy_mmu_mode();
80         pte_unmap_unlock(pte - 1, ptl);
81 }
82
83 static inline void change_pmd_range(struct mm_struct *mm, pud_t *pud,
84                 unsigned long addr, unsigned long end, pgprot_t newprot,
85                 int dirty_accountable)
86 {
87         pmd_t *pmd;
88         unsigned long next;
89
90         pmd = pmd_offset(pud, addr);
91         do {
92                 next = pmd_addr_end(addr, end);
93                 if (pmd_none_or_clear_bad(pmd))
94                         continue;
95                 change_pte_range(mm, pmd, addr, next, newprot, dirty_accountable);
96         } while (pmd++, addr = next, addr != end);
97 }
98
99 static inline void change_pud_range(struct mm_struct *mm, pgd_t *pgd,
100                 unsigned long addr, unsigned long end, pgprot_t newprot,
101                 int dirty_accountable)
102 {
103         pud_t *pud;
104         unsigned long next;
105
106         pud = pud_offset(pgd, addr);
107         do {
108                 next = pud_addr_end(addr, end);
109                 if (pud_none_or_clear_bad(pud))
110                         continue;
111                 change_pmd_range(mm, pud, addr, next, newprot, dirty_accountable);
112         } while (pud++, addr = next, addr != end);
113 }
114
115 static void change_protection(struct vm_area_struct *vma,
116                 unsigned long addr, unsigned long end, pgprot_t newprot,
117                 int dirty_accountable)
118 {
119         struct mm_struct *mm = vma->vm_mm;
120         pgd_t *pgd;
121         unsigned long next;
122         unsigned long start = addr;
123
124         BUG_ON(addr >= end);
125         pgd = pgd_offset(mm, addr);
126         flush_cache_range(vma, addr, end);
127         do {
128                 next = pgd_addr_end(addr, end);
129                 if (pgd_none_or_clear_bad(pgd))
130                         continue;
131                 change_pud_range(mm, pgd, addr, next, newprot, dirty_accountable);
132         } while (pgd++, addr = next, addr != end);
133         flush_tlb_range(vma, start, end);
134 }
135
136 int
137 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
138         unsigned long start, unsigned long end, unsigned long newflags)
139 {
140         struct mm_struct *mm = vma->vm_mm;
141         unsigned long oldflags = vma->vm_flags;
142         long nrpages = (end - start) >> PAGE_SHIFT;
143         unsigned long charged = 0;
144         pgoff_t pgoff;
145         int error;
146         int dirty_accountable = 0;
147
148         if (newflags == oldflags) {
149                 *pprev = vma;
150                 return 0;
151         }
152
153         /*
154          * If we make a private mapping writable we increase our commit;
155          * but (without finer accounting) cannot reduce our commit if we
156          * make it unwritable again.
157          */
158         if (newflags & VM_WRITE) {
159                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|
160                                                 VM_SHARED|VM_NORESERVE))) {
161                         charged = nrpages;
162                         if (security_vm_enough_memory(charged))
163                                 return -ENOMEM;
164                         newflags |= VM_ACCOUNT;
165                 }
166         }
167
168         /*
169          * First try to merge with previous and/or next vma.
170          */
171         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
172         *pprev = vma_merge(mm, *pprev, start, end, newflags,
173                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
174         if (*pprev) {
175                 vma = *pprev;
176                 goto success;
177         }
178
179         *pprev = vma;
180
181         if (start != vma->vm_start) {
182                 error = split_vma(mm, vma, start, 1);
183                 if (error)
184                         goto fail;
185         }
186
187         if (end != vma->vm_end) {
188                 error = split_vma(mm, vma, end, 0);
189                 if (error)
190                         goto fail;
191         }
192
193 success:
194         /*
195          * vm_flags and vm_page_prot are protected by the mmap_sem
196          * held in write mode.
197          */
198         vma->vm_flags = newflags;
199         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
200                                           vm_get_page_prot(newflags));
201
202         if (vma_wants_writenotify(vma)) {
203                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
204                 dirty_accountable = 1;
205         }
206
207         mmu_notifier_invalidate_range_start(mm, start, end);
208         if (is_vm_hugetlb_page(vma))
209                 hugetlb_change_protection(vma, start, end, vma->vm_page_prot);
210         else
211                 change_protection(vma, start, end, vma->vm_page_prot, dirty_accountable);
212         mmu_notifier_invalidate_range_end(mm, start, end);
213         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
214         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
215         return 0;
216
217 fail:
218         vm_unacct_memory(charged);
219         return error;
220 }
221
222 asmlinkage long
223 sys_mprotect(unsigned long start, size_t len, unsigned long prot)
224 {
225         unsigned long vm_flags, nstart, end, tmp, reqprot;
226         struct vm_area_struct *vma, *prev;
227         int error = -EINVAL;
228         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
229         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
230         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
231                 return -EINVAL;
232
233         if (start & ~PAGE_MASK)
234                 return -EINVAL;
235         if (!len)
236                 return 0;
237         len = PAGE_ALIGN(len);
238         end = start + len;
239         if (end <= start)
240                 return -ENOMEM;
241         if (!arch_validate_prot(prot))
242                 return -EINVAL;
243
244         reqprot = prot;
245         /*
246          * Does the application expect PROT_READ to imply PROT_EXEC:
247          */
248         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
249                 prot |= PROT_EXEC;
250
251         vm_flags = calc_vm_prot_bits(prot);
252
253         down_write(&current->mm->mmap_sem);
254
255         vma = find_vma_prev(current->mm, start, &prev);
256         error = -ENOMEM;
257         if (!vma)
258                 goto out;
259         if (unlikely(grows & PROT_GROWSDOWN)) {
260                 if (vma->vm_start >= end)
261                         goto out;
262                 start = vma->vm_start;
263                 error = -EINVAL;
264                 if (!(vma->vm_flags & VM_GROWSDOWN))
265                         goto out;
266         }
267         else {
268                 if (vma->vm_start > start)
269                         goto out;
270                 if (unlikely(grows & PROT_GROWSUP)) {
271                         end = vma->vm_end;
272                         error = -EINVAL;
273                         if (!(vma->vm_flags & VM_GROWSUP))
274                                 goto out;
275                 }
276         }
277         if (start > vma->vm_start)
278                 prev = vma;
279
280         for (nstart = start ; ; ) {
281                 unsigned long newflags;
282
283                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
284
285                 newflags = vm_flags | (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
286
287                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
288                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
289                         error = -EACCES;
290                         goto out;
291                 }
292
293                 error = security_file_mprotect(vma, reqprot, prot);
294                 if (error)
295                         goto out;
296
297                 tmp = vma->vm_end;
298                 if (tmp > end)
299                         tmp = end;
300                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
301                 if (error)
302                         goto out;
303                 nstart = tmp;
304
305                 if (nstart < prev->vm_end)
306                         nstart = prev->vm_end;
307                 if (nstart >= end)
308                         goto out;
309
310                 vma = prev->vm_next;
311                 if (!vma || vma->vm_start != nstart) {
312                         error = -ENOMEM;
313                         goto out;
314                 }
315         }
316 out:
317         up_write(&current->mm->mmap_sem);
318         return error;
319 }