idle: Remove GENERIC_IDLE_LOOP config switch
[linux-3.10.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 unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
39                 unsigned long addr, unsigned long end, pgprot_t newprot,
40                 int dirty_accountable, int prot_numa, bool *ret_all_same_node)
41 {
42         struct mm_struct *mm = vma->vm_mm;
43         pte_t *pte, oldpte;
44         spinlock_t *ptl;
45         unsigned long pages = 0;
46         bool all_same_node = true;
47         int last_nid = -1;
48
49         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
50         arch_enter_lazy_mmu_mode();
51         do {
52                 oldpte = *pte;
53                 if (pte_present(oldpte)) {
54                         pte_t ptent;
55                         bool updated = false;
56
57                         ptent = ptep_modify_prot_start(mm, addr, pte);
58                         if (!prot_numa) {
59                                 ptent = pte_modify(ptent, newprot);
60                                 updated = true;
61                         } else {
62                                 struct page *page;
63
64                                 page = vm_normal_page(vma, addr, oldpte);
65                                 if (page) {
66                                         int this_nid = page_to_nid(page);
67                                         if (last_nid == -1)
68                                                 last_nid = this_nid;
69                                         if (last_nid != this_nid)
70                                                 all_same_node = false;
71
72                                         /* only check non-shared pages */
73                                         if (!pte_numa(oldpte) &&
74                                             page_mapcount(page) == 1) {
75                                                 ptent = pte_mknuma(ptent);
76                                                 updated = true;
77                                         }
78                                 }
79                         }
80
81                         /*
82                          * Avoid taking write faults for pages we know to be
83                          * dirty.
84                          */
85                         if (dirty_accountable && pte_dirty(ptent)) {
86                                 ptent = pte_mkwrite(ptent);
87                                 updated = true;
88                         }
89
90                         if (updated)
91                                 pages++;
92                         ptep_modify_prot_commit(mm, addr, pte, ptent);
93                 } else if (IS_ENABLED(CONFIG_MIGRATION) && !pte_file(oldpte)) {
94                         swp_entry_t entry = pte_to_swp_entry(oldpte);
95
96                         if (is_write_migration_entry(entry)) {
97                                 /*
98                                  * A protection check is difficult so
99                                  * just be safe and disable write
100                                  */
101                                 make_migration_entry_read(&entry);
102                                 set_pte_at(mm, addr, pte,
103                                         swp_entry_to_pte(entry));
104                         }
105                         pages++;
106                 }
107         } while (pte++, addr += PAGE_SIZE, addr != end);
108         arch_leave_lazy_mmu_mode();
109         pte_unmap_unlock(pte - 1, ptl);
110
111         *ret_all_same_node = all_same_node;
112         return pages;
113 }
114
115 #ifdef CONFIG_NUMA_BALANCING
116 static inline void change_pmd_protnuma(struct mm_struct *mm, unsigned long addr,
117                                        pmd_t *pmd)
118 {
119         spin_lock(&mm->page_table_lock);
120         set_pmd_at(mm, addr & PMD_MASK, pmd, pmd_mknuma(*pmd));
121         spin_unlock(&mm->page_table_lock);
122 }
123 #else
124 static inline void change_pmd_protnuma(struct mm_struct *mm, unsigned long addr,
125                                        pmd_t *pmd)
126 {
127         BUG();
128 }
129 #endif /* CONFIG_NUMA_BALANCING */
130
131 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
132                 pud_t *pud, unsigned long addr, unsigned long end,
133                 pgprot_t newprot, int dirty_accountable, int prot_numa)
134 {
135         pmd_t *pmd;
136         unsigned long next;
137         unsigned long pages = 0;
138         bool all_same_node;
139
140         pmd = pmd_offset(pud, addr);
141         do {
142                 next = pmd_addr_end(addr, end);
143                 if (pmd_trans_huge(*pmd)) {
144                         if (next - addr != HPAGE_PMD_SIZE)
145                                 split_huge_page_pmd(vma, addr, pmd);
146                         else if (change_huge_pmd(vma, pmd, addr, newprot,
147                                                  prot_numa)) {
148                                 pages += HPAGE_PMD_NR;
149                                 continue;
150                         }
151                         /* fall through */
152                 }
153                 if (pmd_none_or_clear_bad(pmd))
154                         continue;
155                 pages += change_pte_range(vma, pmd, addr, next, newprot,
156                                  dirty_accountable, prot_numa, &all_same_node);
157
158                 /*
159                  * If we are changing protections for NUMA hinting faults then
160                  * set pmd_numa if the examined pages were all on the same
161                  * node. This allows a regular PMD to be handled as one fault
162                  * and effectively batches the taking of the PTL
163                  */
164                 if (prot_numa && all_same_node)
165                         change_pmd_protnuma(vma->vm_mm, addr, pmd);
166         } while (pmd++, addr = next, addr != end);
167
168         return pages;
169 }
170
171 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
172                 pgd_t *pgd, unsigned long addr, unsigned long end,
173                 pgprot_t newprot, int dirty_accountable, int prot_numa)
174 {
175         pud_t *pud;
176         unsigned long next;
177         unsigned long pages = 0;
178
179         pud = pud_offset(pgd, addr);
180         do {
181                 next = pud_addr_end(addr, end);
182                 if (pud_none_or_clear_bad(pud))
183                         continue;
184                 pages += change_pmd_range(vma, pud, addr, next, newprot,
185                                  dirty_accountable, prot_numa);
186         } while (pud++, addr = next, addr != end);
187
188         return pages;
189 }
190
191 static unsigned long change_protection_range(struct vm_area_struct *vma,
192                 unsigned long addr, unsigned long end, pgprot_t newprot,
193                 int dirty_accountable, int prot_numa)
194 {
195         struct mm_struct *mm = vma->vm_mm;
196         pgd_t *pgd;
197         unsigned long next;
198         unsigned long start = addr;
199         unsigned long pages = 0;
200
201         BUG_ON(addr >= end);
202         pgd = pgd_offset(mm, addr);
203         flush_cache_range(vma, addr, end);
204         do {
205                 next = pgd_addr_end(addr, end);
206                 if (pgd_none_or_clear_bad(pgd))
207                         continue;
208                 pages += change_pud_range(vma, pgd, addr, next, newprot,
209                                  dirty_accountable, prot_numa);
210         } while (pgd++, addr = next, addr != end);
211
212         /* Only flush the TLB if we actually modified any entries: */
213         if (pages)
214                 flush_tlb_range(vma, start, end);
215
216         return pages;
217 }
218
219 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
220                        unsigned long end, pgprot_t newprot,
221                        int dirty_accountable, int prot_numa)
222 {
223         struct mm_struct *mm = vma->vm_mm;
224         unsigned long pages;
225
226         mmu_notifier_invalidate_range_start(mm, start, end);
227         if (is_vm_hugetlb_page(vma))
228                 pages = hugetlb_change_protection(vma, start, end, newprot);
229         else
230                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
231         mmu_notifier_invalidate_range_end(mm, start, end);
232
233         return pages;
234 }
235
236 int
237 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
238         unsigned long start, unsigned long end, unsigned long newflags)
239 {
240         struct mm_struct *mm = vma->vm_mm;
241         unsigned long oldflags = vma->vm_flags;
242         long nrpages = (end - start) >> PAGE_SHIFT;
243         unsigned long charged = 0;
244         pgoff_t pgoff;
245         int error;
246         int dirty_accountable = 0;
247
248         if (newflags == oldflags) {
249                 *pprev = vma;
250                 return 0;
251         }
252
253         /*
254          * If we make a private mapping writable we increase our commit;
255          * but (without finer accounting) cannot reduce our commit if we
256          * make it unwritable again. hugetlb mapping were accounted for
257          * even if read-only so there is no need to account for them here
258          */
259         if (newflags & VM_WRITE) {
260                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
261                                                 VM_SHARED|VM_NORESERVE))) {
262                         charged = nrpages;
263                         if (security_vm_enough_memory_mm(mm, charged))
264                                 return -ENOMEM;
265                         newflags |= VM_ACCOUNT;
266                 }
267         }
268
269         /*
270          * First try to merge with previous and/or next vma.
271          */
272         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
273         *pprev = vma_merge(mm, *pprev, start, end, newflags,
274                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
275         if (*pprev) {
276                 vma = *pprev;
277                 goto success;
278         }
279
280         *pprev = vma;
281
282         if (start != vma->vm_start) {
283                 error = split_vma(mm, vma, start, 1);
284                 if (error)
285                         goto fail;
286         }
287
288         if (end != vma->vm_end) {
289                 error = split_vma(mm, vma, end, 0);
290                 if (error)
291                         goto fail;
292         }
293
294 success:
295         /*
296          * vm_flags and vm_page_prot are protected by the mmap_sem
297          * held in write mode.
298          */
299         vma->vm_flags = newflags;
300         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
301                                           vm_get_page_prot(newflags));
302
303         if (vma_wants_writenotify(vma)) {
304                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
305                 dirty_accountable = 1;
306         }
307
308         change_protection(vma, start, end, vma->vm_page_prot,
309                           dirty_accountable, 0);
310
311         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
312         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
313         perf_event_mmap(vma);
314         return 0;
315
316 fail:
317         vm_unacct_memory(charged);
318         return error;
319 }
320
321 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
322                 unsigned long, prot)
323 {
324         unsigned long vm_flags, nstart, end, tmp, reqprot;
325         struct vm_area_struct *vma, *prev;
326         int error = -EINVAL;
327         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
328         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
329         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
330                 return -EINVAL;
331
332         if (start & ~PAGE_MASK)
333                 return -EINVAL;
334         if (!len)
335                 return 0;
336         len = PAGE_ALIGN(len);
337         end = start + len;
338         if (end <= start)
339                 return -ENOMEM;
340         if (!arch_validate_prot(prot))
341                 return -EINVAL;
342
343         reqprot = prot;
344         /*
345          * Does the application expect PROT_READ to imply PROT_EXEC:
346          */
347         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
348                 prot |= PROT_EXEC;
349
350         vm_flags = calc_vm_prot_bits(prot);
351
352         down_write(&current->mm->mmap_sem);
353
354         vma = find_vma(current->mm, start);
355         error = -ENOMEM;
356         if (!vma)
357                 goto out;
358         prev = vma->vm_prev;
359         if (unlikely(grows & PROT_GROWSDOWN)) {
360                 if (vma->vm_start >= end)
361                         goto out;
362                 start = vma->vm_start;
363                 error = -EINVAL;
364                 if (!(vma->vm_flags & VM_GROWSDOWN))
365                         goto out;
366         } else {
367                 if (vma->vm_start > start)
368                         goto out;
369                 if (unlikely(grows & PROT_GROWSUP)) {
370                         end = vma->vm_end;
371                         error = -EINVAL;
372                         if (!(vma->vm_flags & VM_GROWSUP))
373                                 goto out;
374                 }
375         }
376         if (start > vma->vm_start)
377                 prev = vma;
378
379         for (nstart = start ; ; ) {
380                 unsigned long newflags;
381
382                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
383
384                 newflags = vm_flags;
385                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
386
387                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
388                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
389                         error = -EACCES;
390                         goto out;
391                 }
392
393                 error = security_file_mprotect(vma, reqprot, prot);
394                 if (error)
395                         goto out;
396
397                 tmp = vma->vm_end;
398                 if (tmp > end)
399                         tmp = end;
400                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
401                 if (error)
402                         goto out;
403                 nstart = tmp;
404
405                 if (nstart < prev->vm_end)
406                         nstart = prev->vm_end;
407                 if (nstart >= end)
408                         goto out;
409
410                 vma = prev->vm_next;
411                 if (!vma || vma->vm_start != nstart) {
412                         error = -ENOMEM;
413                         goto out;
414                 }
415         }
416 out:
417         up_write(&current->mm->mmap_sem);
418         return error;
419 }