flush icache before set_pte() on ia64: flush icache at set_pte
[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 <asm/uaccess.h>
25 #include <asm/pgtable.h>
26 #include <asm/cacheflush.h>
27 #include <asm/tlbflush.h>
28
29 static void change_pte_range(struct mm_struct *mm, pmd_t *pmd,
30                 unsigned long addr, unsigned long end, pgprot_t newprot,
31                 int dirty_accountable)
32 {
33         pte_t *pte, oldpte;
34         spinlock_t *ptl;
35
36         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
37         arch_enter_lazy_mmu_mode();
38         do {
39                 oldpte = *pte;
40                 if (pte_present(oldpte)) {
41                         pte_t ptent;
42
43                         /* Avoid an SMP race with hardware updated dirty/clean
44                          * bits by wiping the pte and then setting the new pte
45                          * into place.
46                          */
47                         ptent = ptep_get_and_clear(mm, addr, pte);
48                         ptent = pte_modify(ptent, newprot);
49                         /*
50                          * Avoid taking write faults for pages we know to be
51                          * dirty.
52                          */
53                         if (dirty_accountable && pte_dirty(ptent))
54                                 ptent = pte_mkwrite(ptent);
55                         set_pte_at(mm, addr, pte, ptent);
56 #ifdef CONFIG_MIGRATION
57                 } else if (!pte_file(oldpte)) {
58                         swp_entry_t entry = pte_to_swp_entry(oldpte);
59
60                         if (is_write_migration_entry(entry)) {
61                                 /*
62                                  * A protection check is difficult so
63                                  * just be safe and disable write
64                                  */
65                                 make_migration_entry_read(&entry);
66                                 set_pte_at(mm, addr, pte,
67                                         swp_entry_to_pte(entry));
68                         }
69 #endif
70                 }
71
72         } while (pte++, addr += PAGE_SIZE, addr != end);
73         arch_leave_lazy_mmu_mode();
74         pte_unmap_unlock(pte - 1, ptl);
75 }
76
77 static inline void change_pmd_range(struct mm_struct *mm, pud_t *pud,
78                 unsigned long addr, unsigned long end, pgprot_t newprot,
79                 int dirty_accountable)
80 {
81         pmd_t *pmd;
82         unsigned long next;
83
84         pmd = pmd_offset(pud, addr);
85         do {
86                 next = pmd_addr_end(addr, end);
87                 if (pmd_none_or_clear_bad(pmd))
88                         continue;
89                 change_pte_range(mm, pmd, addr, next, newprot, dirty_accountable);
90         } while (pmd++, addr = next, addr != end);
91 }
92
93 static inline void change_pud_range(struct mm_struct *mm, pgd_t *pgd,
94                 unsigned long addr, unsigned long end, pgprot_t newprot,
95                 int dirty_accountable)
96 {
97         pud_t *pud;
98         unsigned long next;
99
100         pud = pud_offset(pgd, addr);
101         do {
102                 next = pud_addr_end(addr, end);
103                 if (pud_none_or_clear_bad(pud))
104                         continue;
105                 change_pmd_range(mm, pud, addr, next, newprot, dirty_accountable);
106         } while (pud++, addr = next, addr != end);
107 }
108
109 static void change_protection(struct vm_area_struct *vma,
110                 unsigned long addr, unsigned long end, pgprot_t newprot,
111                 int dirty_accountable)
112 {
113         struct mm_struct *mm = vma->vm_mm;
114         pgd_t *pgd;
115         unsigned long next;
116         unsigned long start = addr;
117
118         BUG_ON(addr >= end);
119         pgd = pgd_offset(mm, addr);
120         flush_cache_range(vma, addr, end);
121         do {
122                 next = pgd_addr_end(addr, end);
123                 if (pgd_none_or_clear_bad(pgd))
124                         continue;
125                 change_pud_range(mm, pgd, addr, next, newprot, dirty_accountable);
126         } while (pgd++, addr = next, addr != end);
127         flush_tlb_range(vma, start, end);
128 }
129
130 int
131 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
132         unsigned long start, unsigned long end, unsigned long newflags)
133 {
134         struct mm_struct *mm = vma->vm_mm;
135         unsigned long oldflags = vma->vm_flags;
136         long nrpages = (end - start) >> PAGE_SHIFT;
137         unsigned long charged = 0;
138         pgoff_t pgoff;
139         int error;
140         int dirty_accountable = 0;
141
142         if (newflags == oldflags) {
143                 *pprev = vma;
144                 return 0;
145         }
146
147         /*
148          * If we make a private mapping writable we increase our commit;
149          * but (without finer accounting) cannot reduce our commit if we
150          * make it unwritable again.
151          *
152          * FIXME? We haven't defined a VM_NORESERVE flag, so mprotecting
153          * a MAP_NORESERVE private mapping to writable will now reserve.
154          */
155         if (newflags & VM_WRITE) {
156                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_SHARED))) {
157                         charged = nrpages;
158                         if (security_vm_enough_memory(charged))
159                                 return -ENOMEM;
160                         newflags |= VM_ACCOUNT;
161                 }
162         }
163
164         /*
165          * First try to merge with previous and/or next vma.
166          */
167         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
168         *pprev = vma_merge(mm, *pprev, start, end, newflags,
169                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
170         if (*pprev) {
171                 vma = *pprev;
172                 goto success;
173         }
174
175         *pprev = vma;
176
177         if (start != vma->vm_start) {
178                 error = split_vma(mm, vma, start, 1);
179                 if (error)
180                         goto fail;
181         }
182
183         if (end != vma->vm_end) {
184                 error = split_vma(mm, vma, end, 0);
185                 if (error)
186                         goto fail;
187         }
188
189 success:
190         /*
191          * vm_flags and vm_page_prot are protected by the mmap_sem
192          * held in write mode.
193          */
194         vma->vm_flags = newflags;
195         vma->vm_page_prot = protection_map[newflags &
196                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
197         if (vma_wants_writenotify(vma)) {
198                 vma->vm_page_prot = protection_map[newflags &
199                         (VM_READ|VM_WRITE|VM_EXEC)];
200                 dirty_accountable = 1;
201         }
202
203         if (is_vm_hugetlb_page(vma))
204                 hugetlb_change_protection(vma, start, end, vma->vm_page_prot);
205         else
206                 change_protection(vma, start, end, vma->vm_page_prot, dirty_accountable);
207         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
208         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
209         return 0;
210
211 fail:
212         vm_unacct_memory(charged);
213         return error;
214 }
215
216 asmlinkage long
217 sys_mprotect(unsigned long start, size_t len, unsigned long prot)
218 {
219         unsigned long vm_flags, nstart, end, tmp, reqprot;
220         struct vm_area_struct *vma, *prev;
221         int error = -EINVAL;
222         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
223         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
224         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
225                 return -EINVAL;
226
227         if (start & ~PAGE_MASK)
228                 return -EINVAL;
229         if (!len)
230                 return 0;
231         len = PAGE_ALIGN(len);
232         end = start + len;
233         if (end <= start)
234                 return -ENOMEM;
235         if (prot & ~(PROT_READ | PROT_WRITE | PROT_EXEC | PROT_SEM))
236                 return -EINVAL;
237
238         reqprot = prot;
239         /*
240          * Does the application expect PROT_READ to imply PROT_EXEC:
241          */
242         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
243                 prot |= PROT_EXEC;
244
245         vm_flags = calc_vm_prot_bits(prot);
246
247         down_write(&current->mm->mmap_sem);
248
249         vma = find_vma_prev(current->mm, start, &prev);
250         error = -ENOMEM;
251         if (!vma)
252                 goto out;
253         if (unlikely(grows & PROT_GROWSDOWN)) {
254                 if (vma->vm_start >= end)
255                         goto out;
256                 start = vma->vm_start;
257                 error = -EINVAL;
258                 if (!(vma->vm_flags & VM_GROWSDOWN))
259                         goto out;
260         }
261         else {
262                 if (vma->vm_start > start)
263                         goto out;
264                 if (unlikely(grows & PROT_GROWSUP)) {
265                         end = vma->vm_end;
266                         error = -EINVAL;
267                         if (!(vma->vm_flags & VM_GROWSUP))
268                                 goto out;
269                 }
270         }
271         if (start > vma->vm_start)
272                 prev = vma;
273
274         for (nstart = start ; ; ) {
275                 unsigned long newflags;
276
277                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
278
279                 newflags = vm_flags | (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
280
281                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
282                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
283                         error = -EACCES;
284                         goto out;
285                 }
286
287                 error = security_file_mprotect(vma, reqprot, prot);
288                 if (error)
289                         goto out;
290
291                 tmp = vma->vm_end;
292                 if (tmp > end)
293                         tmp = end;
294                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
295                 if (error)
296                         goto out;
297                 nstart = tmp;
298
299                 if (nstart < prev->vm_end)
300                         nstart = prev->vm_end;
301                 if (nstart >= end)
302                         goto out;
303
304                 vma = prev->vm_next;
305                 if (!vma || vma->vm_start != nstart) {
306                         error = -ENOMEM;
307                         goto out;
308                 }
309         }
310 out:
311         up_write(&current->mm->mmap_sem);
312         return error;
313 }