thp: mprotect: pass vma down to page table walkers
[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@lxorguk.ukuu.org.uk>
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/shm.h>
14 #include <linux/mman.h>
15 #include <linux/fs.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 <asm/uaccess.h>
27 #include <asm/pgtable.h>
28 #include <asm/cacheflush.h>
29 #include <asm/tlbflush.h>
30
31 #ifndef pgprot_modify
32 static inline pgprot_t pgprot_modify(pgprot_t oldprot, pgprot_t newprot)
33 {
34         return newprot;
35 }
36 #endif
37
38 static void change_pte_range(struct mm_struct *mm, pmd_t *pmd,
39                 unsigned long addr, unsigned long end, pgprot_t newprot,
40                 int dirty_accountable)
41 {
42         pte_t *pte, oldpte;
43         spinlock_t *ptl;
44
45         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
46         arch_enter_lazy_mmu_mode();
47         do {
48                 oldpte = *pte;
49                 if (pte_present(oldpte)) {
50                         pte_t ptent;
51
52                         ptent = ptep_modify_prot_start(mm, addr, pte);
53                         ptent = pte_modify(ptent, newprot);
54
55                         /*
56                          * Avoid taking write faults for pages we know to be
57                          * dirty.
58                          */
59                         if (dirty_accountable && pte_dirty(ptent))
60                                 ptent = pte_mkwrite(ptent);
61
62                         ptep_modify_prot_commit(mm, addr, pte, ptent);
63                 } else if (PAGE_MIGRATION && !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                 }
76         } while (pte++, addr += PAGE_SIZE, addr != end);
77         arch_leave_lazy_mmu_mode();
78         pte_unmap_unlock(pte - 1, ptl);
79 }
80
81 static inline void change_pmd_range(struct vm_area_struct *vma, pud_t *pud,
82                 unsigned long addr, unsigned long end, pgprot_t newprot,
83                 int dirty_accountable)
84 {
85         pmd_t *pmd;
86         unsigned long next;
87
88         pmd = pmd_offset(pud, addr);
89         do {
90                 next = pmd_addr_end(addr, end);
91                 split_huge_page_pmd(vma->vm_mm, pmd);
92                 if (pmd_none_or_clear_bad(pmd))
93                         continue;
94                 change_pte_range(vma->vm_mm, pmd, addr, next, newprot,
95                                  dirty_accountable);
96         } while (pmd++, addr = next, addr != end);
97 }
98
99 static inline void change_pud_range(struct vm_area_struct *vma, 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(vma, pud, addr, next, newprot,
112                                  dirty_accountable);
113         } while (pud++, addr = next, addr != end);
114 }
115
116 static void change_protection(struct vm_area_struct *vma,
117                 unsigned long addr, unsigned long end, pgprot_t newprot,
118                 int dirty_accountable)
119 {
120         struct mm_struct *mm = vma->vm_mm;
121         pgd_t *pgd;
122         unsigned long next;
123         unsigned long start = addr;
124
125         BUG_ON(addr >= end);
126         pgd = pgd_offset(mm, addr);
127         flush_cache_range(vma, addr, end);
128         do {
129                 next = pgd_addr_end(addr, end);
130                 if (pgd_none_or_clear_bad(pgd))
131                         continue;
132                 change_pud_range(vma, pgd, addr, next, newprot,
133                                  dirty_accountable);
134         } while (pgd++, addr = next, addr != end);
135         flush_tlb_range(vma, start, end);
136 }
137
138 int
139 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
140         unsigned long start, unsigned long end, unsigned long newflags)
141 {
142         struct mm_struct *mm = vma->vm_mm;
143         unsigned long oldflags = vma->vm_flags;
144         long nrpages = (end - start) >> PAGE_SHIFT;
145         unsigned long charged = 0;
146         pgoff_t pgoff;
147         int error;
148         int dirty_accountable = 0;
149
150         if (newflags == oldflags) {
151                 *pprev = vma;
152                 return 0;
153         }
154
155         /*
156          * If we make a private mapping writable we increase our commit;
157          * but (without finer accounting) cannot reduce our commit if we
158          * make it unwritable again. hugetlb mapping were accounted for
159          * even if read-only so there is no need to account for them here
160          */
161         if (newflags & VM_WRITE) {
162                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
163                                                 VM_SHARED|VM_NORESERVE))) {
164                         charged = nrpages;
165                         if (security_vm_enough_memory(charged))
166                                 return -ENOMEM;
167                         newflags |= VM_ACCOUNT;
168                 }
169         }
170
171         /*
172          * First try to merge with previous and/or next vma.
173          */
174         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
175         *pprev = vma_merge(mm, *pprev, start, end, newflags,
176                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
177         if (*pprev) {
178                 vma = *pprev;
179                 goto success;
180         }
181
182         *pprev = vma;
183
184         if (start != vma->vm_start) {
185                 error = split_vma(mm, vma, start, 1);
186                 if (error)
187                         goto fail;
188         }
189
190         if (end != vma->vm_end) {
191                 error = split_vma(mm, vma, end, 0);
192                 if (error)
193                         goto fail;
194         }
195
196 success:
197         /*
198          * vm_flags and vm_page_prot are protected by the mmap_sem
199          * held in write mode.
200          */
201         vma->vm_flags = newflags;
202         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
203                                           vm_get_page_prot(newflags));
204
205         if (vma_wants_writenotify(vma)) {
206                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
207                 dirty_accountable = 1;
208         }
209
210         mmu_notifier_invalidate_range_start(mm, start, end);
211         if (is_vm_hugetlb_page(vma))
212                 hugetlb_change_protection(vma, start, end, vma->vm_page_prot);
213         else
214                 change_protection(vma, start, end, vma->vm_page_prot, dirty_accountable);
215         mmu_notifier_invalidate_range_end(mm, start, end);
216         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
217         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
218         perf_event_mmap(vma);
219         return 0;
220
221 fail:
222         vm_unacct_memory(charged);
223         return error;
224 }
225
226 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
227                 unsigned long, prot)
228 {
229         unsigned long vm_flags, nstart, end, tmp, reqprot;
230         struct vm_area_struct *vma, *prev;
231         int error = -EINVAL;
232         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
233         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
234         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
235                 return -EINVAL;
236
237         if (start & ~PAGE_MASK)
238                 return -EINVAL;
239         if (!len)
240                 return 0;
241         len = PAGE_ALIGN(len);
242         end = start + len;
243         if (end <= start)
244                 return -ENOMEM;
245         if (!arch_validate_prot(prot))
246                 return -EINVAL;
247
248         reqprot = prot;
249         /*
250          * Does the application expect PROT_READ to imply PROT_EXEC:
251          */
252         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
253                 prot |= PROT_EXEC;
254
255         vm_flags = calc_vm_prot_bits(prot);
256
257         down_write(&current->mm->mmap_sem);
258
259         vma = find_vma_prev(current->mm, start, &prev);
260         error = -ENOMEM;
261         if (!vma)
262                 goto out;
263         if (unlikely(grows & PROT_GROWSDOWN)) {
264                 if (vma->vm_start >= end)
265                         goto out;
266                 start = vma->vm_start;
267                 error = -EINVAL;
268                 if (!(vma->vm_flags & VM_GROWSDOWN))
269                         goto out;
270         }
271         else {
272                 if (vma->vm_start > start)
273                         goto out;
274                 if (unlikely(grows & PROT_GROWSUP)) {
275                         end = vma->vm_end;
276                         error = -EINVAL;
277                         if (!(vma->vm_flags & VM_GROWSUP))
278                                 goto out;
279                 }
280         }
281         if (start > vma->vm_start)
282                 prev = vma;
283
284         for (nstart = start ; ; ) {
285                 unsigned long newflags;
286
287                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
288
289                 newflags = vm_flags | (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
290
291                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
292                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
293                         error = -EACCES;
294                         goto out;
295                 }
296
297                 error = security_file_mprotect(vma, reqprot, prot);
298                 if (error)
299                         goto out;
300
301                 tmp = vma->vm_end;
302                 if (tmp > end)
303                         tmp = end;
304                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
305                 if (error)
306                         goto out;
307                 nstart = tmp;
308
309                 if (nstart < prev->vm_end)
310                         nstart = prev->vm_end;
311                 if (nstart >= end)
312                         goto out;
313
314                 vma = prev->vm_next;
315                 if (!vma || vma->vm_start != nstart) {
316                         error = -ENOMEM;
317                         goto out;
318                 }
319         }
320 out:
321         up_write(&current->mm->mmap_sem);
322         return error;
323 }