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