seccomp: Refactor the filter callback and the API
[pandora-kernel.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 <linux/ksm.h>
27 #include <asm/uaccess.h>
28 #include <asm/pgtable.h>
29 #include <asm/cacheflush.h>
30 #include <asm/tlbflush.h>
31
32 #ifndef pgprot_modify
33 static inline pgprot_t pgprot_modify(pgprot_t oldprot, pgprot_t newprot)
34 {
35         return newprot;
36 }
37 #endif
38
39 /*
40  * For a prot_numa update we only hold mmap_sem for read so there is a
41  * potential race with faulting where a pmd was temporarily none. This
42  * function checks for a transhuge pmd under the appropriate lock. It
43  * returns a pte if it was successfully locked or NULL if it raced with
44  * a transhuge insertion.
45  */
46 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
47                         unsigned long addr, int prot_numa, spinlock_t **ptl)
48 {
49         pte_t *pte;
50         spinlock_t *pmdl;
51
52         /* !prot_numa is protected by mmap_sem held for write */
53         if (!prot_numa)
54                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
55
56         pmdl = pmd_lock(vma->vm_mm, pmd);
57         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
58                 spin_unlock(pmdl);
59                 return NULL;
60         }
61
62         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
63         spin_unlock(pmdl);
64         return pte;
65 }
66
67 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
68                 unsigned long addr, unsigned long end, pgprot_t newprot,
69                 int dirty_accountable, int prot_numa)
70 {
71         struct mm_struct *mm = vma->vm_mm;
72         pte_t *pte, oldpte;
73         spinlock_t *ptl;
74         unsigned long pages = 0;
75
76         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
77         if (!pte)
78                 return 0;
79
80         arch_enter_lazy_mmu_mode();
81         do {
82                 oldpte = *pte;
83                 if (pte_present(oldpte)) {
84                         pte_t ptent;
85                         bool updated = false;
86
87                         if (!prot_numa) {
88                                 ptent = ptep_modify_prot_start(mm, addr, pte);
89                                 if (pte_numa(ptent))
90                                         ptent = pte_mknonnuma(ptent);
91                                 ptent = pte_modify(ptent, newprot);
92                                 /*
93                                  * Avoid taking write faults for pages we
94                                  * know to be dirty.
95                                  */
96                                 if (dirty_accountable && pte_dirty(ptent))
97                                         ptent = pte_mkwrite(ptent);
98                                 ptep_modify_prot_commit(mm, addr, pte, ptent);
99                                 updated = true;
100                         } else {
101                                 struct page *page;
102
103                                 page = vm_normal_page(vma, addr, oldpte);
104                                 if (page && !PageKsm(page)) {
105                                         if (!pte_numa(oldpte)) {
106                                                 ptep_set_numa(mm, addr, pte);
107                                                 updated = true;
108                                         }
109                                 }
110                         }
111                         if (updated)
112                                 pages++;
113                 } else if (IS_ENABLED(CONFIG_MIGRATION) && !pte_file(oldpte)) {
114                         swp_entry_t entry = pte_to_swp_entry(oldpte);
115
116                         if (is_write_migration_entry(entry)) {
117                                 pte_t newpte;
118                                 /*
119                                  * A protection check is difficult so
120                                  * just be safe and disable write
121                                  */
122                                 make_migration_entry_read(&entry);
123                                 newpte = swp_entry_to_pte(entry);
124                                 if (pte_swp_soft_dirty(oldpte))
125                                         newpte = pte_swp_mksoft_dirty(newpte);
126                                 set_pte_at(mm, addr, pte, newpte);
127
128                                 pages++;
129                         }
130                 }
131         } while (pte++, addr += PAGE_SIZE, addr != end);
132         arch_leave_lazy_mmu_mode();
133         pte_unmap_unlock(pte - 1, ptl);
134
135         return pages;
136 }
137
138 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
139                 pud_t *pud, unsigned long addr, unsigned long end,
140                 pgprot_t newprot, int dirty_accountable, int prot_numa)
141 {
142         pmd_t *pmd;
143         struct mm_struct *mm = vma->vm_mm;
144         unsigned long next;
145         unsigned long pages = 0;
146         unsigned long nr_huge_updates = 0;
147         unsigned long mni_start = 0;
148
149         pmd = pmd_offset(pud, addr);
150         do {
151                 unsigned long this_pages;
152
153                 next = pmd_addr_end(addr, end);
154                 if (!pmd_trans_huge(*pmd) && pmd_none_or_clear_bad(pmd))
155                         continue;
156
157                 /* invoke the mmu notifier if the pmd is populated */
158                 if (!mni_start) {
159                         mni_start = addr;
160                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
161                 }
162
163                 if (pmd_trans_huge(*pmd)) {
164                         if (next - addr != HPAGE_PMD_SIZE)
165                                 split_huge_page_pmd(vma, addr, pmd);
166                         else {
167                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
168                                                 newprot, prot_numa);
169
170                                 if (nr_ptes) {
171                                         if (nr_ptes == HPAGE_PMD_NR) {
172                                                 pages += HPAGE_PMD_NR;
173                                                 nr_huge_updates++;
174                                         }
175
176                                         /* huge pmd was handled */
177                                         continue;
178                                 }
179                         }
180                         /* fall through, the trans huge pmd just split */
181                 }
182                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
183                                  dirty_accountable, prot_numa);
184                 pages += this_pages;
185         } while (pmd++, addr = next, addr != end);
186
187         if (mni_start)
188                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
189
190         if (nr_huge_updates)
191                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
192         return pages;
193 }
194
195 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
196                 pgd_t *pgd, unsigned long addr, unsigned long end,
197                 pgprot_t newprot, int dirty_accountable, int prot_numa)
198 {
199         pud_t *pud;
200         unsigned long next;
201         unsigned long pages = 0;
202
203         pud = pud_offset(pgd, addr);
204         do {
205                 next = pud_addr_end(addr, end);
206                 if (pud_none_or_clear_bad(pud))
207                         continue;
208                 pages += change_pmd_range(vma, pud, addr, next, newprot,
209                                  dirty_accountable, prot_numa);
210         } while (pud++, addr = next, addr != end);
211
212         return pages;
213 }
214
215 static unsigned long change_protection_range(struct vm_area_struct *vma,
216                 unsigned long addr, unsigned long end, pgprot_t newprot,
217                 int dirty_accountable, int prot_numa)
218 {
219         struct mm_struct *mm = vma->vm_mm;
220         pgd_t *pgd;
221         unsigned long next;
222         unsigned long start = addr;
223         unsigned long pages = 0;
224
225         BUG_ON(addr >= end);
226         pgd = pgd_offset(mm, addr);
227         flush_cache_range(vma, addr, end);
228         set_tlb_flush_pending(mm);
229         do {
230                 next = pgd_addr_end(addr, end);
231                 if (pgd_none_or_clear_bad(pgd))
232                         continue;
233                 pages += change_pud_range(vma, pgd, addr, next, newprot,
234                                  dirty_accountable, prot_numa);
235         } while (pgd++, addr = next, addr != end);
236
237         /* Only flush the TLB if we actually modified any entries: */
238         if (pages)
239                 flush_tlb_range(vma, start, end);
240         clear_tlb_flush_pending(mm);
241
242         return pages;
243 }
244
245 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
246                        unsigned long end, pgprot_t newprot,
247                        int dirty_accountable, int prot_numa)
248 {
249         unsigned long pages;
250
251         if (is_vm_hugetlb_page(vma))
252                 pages = hugetlb_change_protection(vma, start, end, newprot);
253         else
254                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
255
256         return pages;
257 }
258
259 int
260 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
261         unsigned long start, unsigned long end, unsigned long newflags)
262 {
263         struct mm_struct *mm = vma->vm_mm;
264         unsigned long oldflags = vma->vm_flags;
265         long nrpages = (end - start) >> PAGE_SHIFT;
266         unsigned long charged = 0;
267         pgoff_t pgoff;
268         int error;
269         int dirty_accountable = 0;
270
271         if (newflags == oldflags) {
272                 *pprev = vma;
273                 return 0;
274         }
275
276         /*
277          * If we make a private mapping writable we increase our commit;
278          * but (without finer accounting) cannot reduce our commit if we
279          * make it unwritable again. hugetlb mapping were accounted for
280          * even if read-only so there is no need to account for them here
281          */
282         if (newflags & VM_WRITE) {
283                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
284                                                 VM_SHARED|VM_NORESERVE))) {
285                         charged = nrpages;
286                         if (security_vm_enough_memory_mm(mm, charged))
287                                 return -ENOMEM;
288                         newflags |= VM_ACCOUNT;
289                 }
290         }
291
292         /*
293          * First try to merge with previous and/or next vma.
294          */
295         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
296         *pprev = vma_merge(mm, *pprev, start, end, newflags,
297                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
298         if (*pprev) {
299                 vma = *pprev;
300                 goto success;
301         }
302
303         *pprev = vma;
304
305         if (start != vma->vm_start) {
306                 error = split_vma(mm, vma, start, 1);
307                 if (error)
308                         goto fail;
309         }
310
311         if (end != vma->vm_end) {
312                 error = split_vma(mm, vma, end, 0);
313                 if (error)
314                         goto fail;
315         }
316
317 success:
318         /*
319          * vm_flags and vm_page_prot are protected by the mmap_sem
320          * held in write mode.
321          */
322         vma->vm_flags = newflags;
323         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
324                                           vm_get_page_prot(newflags));
325
326         if (vma_wants_writenotify(vma)) {
327                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
328                 dirty_accountable = 1;
329         }
330
331         change_protection(vma, start, end, vma->vm_page_prot,
332                           dirty_accountable, 0);
333
334         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
335         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
336         perf_event_mmap(vma);
337         return 0;
338
339 fail:
340         vm_unacct_memory(charged);
341         return error;
342 }
343
344 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
345                 unsigned long, prot)
346 {
347         unsigned long vm_flags, nstart, end, tmp, reqprot;
348         struct vm_area_struct *vma, *prev;
349         int error = -EINVAL;
350         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
351         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
352         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
353                 return -EINVAL;
354
355         if (start & ~PAGE_MASK)
356                 return -EINVAL;
357         if (!len)
358                 return 0;
359         len = PAGE_ALIGN(len);
360         end = start + len;
361         if (end <= start)
362                 return -ENOMEM;
363         if (!arch_validate_prot(prot))
364                 return -EINVAL;
365
366         reqprot = prot;
367         /*
368          * Does the application expect PROT_READ to imply PROT_EXEC:
369          */
370         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
371                 prot |= PROT_EXEC;
372
373         vm_flags = calc_vm_prot_bits(prot);
374
375         down_write(&current->mm->mmap_sem);
376
377         vma = find_vma(current->mm, start);
378         error = -ENOMEM;
379         if (!vma)
380                 goto out;
381         prev = vma->vm_prev;
382         if (unlikely(grows & PROT_GROWSDOWN)) {
383                 if (vma->vm_start >= end)
384                         goto out;
385                 start = vma->vm_start;
386                 error = -EINVAL;
387                 if (!(vma->vm_flags & VM_GROWSDOWN))
388                         goto out;
389         } else {
390                 if (vma->vm_start > start)
391                         goto out;
392                 if (unlikely(grows & PROT_GROWSUP)) {
393                         end = vma->vm_end;
394                         error = -EINVAL;
395                         if (!(vma->vm_flags & VM_GROWSUP))
396                                 goto out;
397                 }
398         }
399         if (start > vma->vm_start)
400                 prev = vma;
401
402         for (nstart = start ; ; ) {
403                 unsigned long newflags;
404
405                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
406
407                 newflags = vm_flags;
408                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
409
410                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
411                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
412                         error = -EACCES;
413                         goto out;
414                 }
415
416                 error = security_file_mprotect(vma, reqprot, prot);
417                 if (error)
418                         goto out;
419
420                 tmp = vma->vm_end;
421                 if (tmp > end)
422                         tmp = end;
423                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
424                 if (error)
425                         goto out;
426                 nstart = tmp;
427
428                 if (nstart < prev->vm_end)
429                         nstart = prev->vm_end;
430                 if (nstart >= end)
431                         goto out;
432
433                 vma = prev->vm_next;
434                 if (!vma || vma->vm_start != nstart) {
435                         error = -ENOMEM;
436                         goto out;
437                 }
438         }
439 out:
440         up_write(&current->mm->mmap_sem);
441         return error;
442 }