xref: /openbmc/linux/mm/mprotect.c (revision cd5d5810)
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 				pte_t newpte;
98 				/*
99 				 * A protection check is difficult so
100 				 * just be safe and disable write
101 				 */
102 				make_migration_entry_read(&entry);
103 				newpte = swp_entry_to_pte(entry);
104 				if (pte_swp_soft_dirty(oldpte))
105 					newpte = pte_swp_mksoft_dirty(newpte);
106 				set_pte_at(mm, addr, pte, newpte);
107 			}
108 			pages++;
109 		}
110 	} while (pte++, addr += PAGE_SIZE, addr != end);
111 	arch_leave_lazy_mmu_mode();
112 	pte_unmap_unlock(pte - 1, ptl);
113 
114 	*ret_all_same_node = all_same_node;
115 	return pages;
116 }
117 
118 #ifdef CONFIG_NUMA_BALANCING
119 static inline void change_pmd_protnuma(struct mm_struct *mm, unsigned long addr,
120 				       pmd_t *pmd)
121 {
122 	spin_lock(&mm->page_table_lock);
123 	set_pmd_at(mm, addr & PMD_MASK, pmd, pmd_mknuma(*pmd));
124 	spin_unlock(&mm->page_table_lock);
125 }
126 #else
127 static inline void change_pmd_protnuma(struct mm_struct *mm, unsigned long addr,
128 				       pmd_t *pmd)
129 {
130 	BUG();
131 }
132 #endif /* CONFIG_NUMA_BALANCING */
133 
134 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
135 		pud_t *pud, unsigned long addr, unsigned long end,
136 		pgprot_t newprot, int dirty_accountable, int prot_numa)
137 {
138 	pmd_t *pmd;
139 	unsigned long next;
140 	unsigned long pages = 0;
141 	bool all_same_node;
142 
143 	pmd = pmd_offset(pud, addr);
144 	do {
145 		next = pmd_addr_end(addr, end);
146 		if (pmd_trans_huge(*pmd)) {
147 			if (next - addr != HPAGE_PMD_SIZE)
148 				split_huge_page_pmd(vma, addr, pmd);
149 			else if (change_huge_pmd(vma, pmd, addr, newprot,
150 						 prot_numa)) {
151 				pages += HPAGE_PMD_NR;
152 				continue;
153 			}
154 			/* fall through */
155 		}
156 		if (pmd_none_or_clear_bad(pmd))
157 			continue;
158 		pages += change_pte_range(vma, pmd, addr, next, newprot,
159 				 dirty_accountable, prot_numa, &all_same_node);
160 
161 		/*
162 		 * If we are changing protections for NUMA hinting faults then
163 		 * set pmd_numa if the examined pages were all on the same
164 		 * node. This allows a regular PMD to be handled as one fault
165 		 * and effectively batches the taking of the PTL
166 		 */
167 		if (prot_numa && all_same_node)
168 			change_pmd_protnuma(vma->vm_mm, addr, pmd);
169 	} while (pmd++, addr = next, addr != end);
170 
171 	return pages;
172 }
173 
174 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
175 		pgd_t *pgd, unsigned long addr, unsigned long end,
176 		pgprot_t newprot, int dirty_accountable, int prot_numa)
177 {
178 	pud_t *pud;
179 	unsigned long next;
180 	unsigned long pages = 0;
181 
182 	pud = pud_offset(pgd, addr);
183 	do {
184 		next = pud_addr_end(addr, end);
185 		if (pud_none_or_clear_bad(pud))
186 			continue;
187 		pages += change_pmd_range(vma, pud, addr, next, newprot,
188 				 dirty_accountable, prot_numa);
189 	} while (pud++, addr = next, addr != end);
190 
191 	return pages;
192 }
193 
194 static unsigned long change_protection_range(struct vm_area_struct *vma,
195 		unsigned long addr, unsigned long end, pgprot_t newprot,
196 		int dirty_accountable, int prot_numa)
197 {
198 	struct mm_struct *mm = vma->vm_mm;
199 	pgd_t *pgd;
200 	unsigned long next;
201 	unsigned long start = addr;
202 	unsigned long pages = 0;
203 
204 	BUG_ON(addr >= end);
205 	pgd = pgd_offset(mm, addr);
206 	flush_cache_range(vma, addr, end);
207 	do {
208 		next = pgd_addr_end(addr, end);
209 		if (pgd_none_or_clear_bad(pgd))
210 			continue;
211 		pages += change_pud_range(vma, pgd, addr, next, newprot,
212 				 dirty_accountable, prot_numa);
213 	} while (pgd++, addr = next, addr != end);
214 
215 	/* Only flush the TLB if we actually modified any entries: */
216 	if (pages)
217 		flush_tlb_range(vma, start, end);
218 
219 	return pages;
220 }
221 
222 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
223 		       unsigned long end, pgprot_t newprot,
224 		       int dirty_accountable, int prot_numa)
225 {
226 	struct mm_struct *mm = vma->vm_mm;
227 	unsigned long pages;
228 
229 	mmu_notifier_invalidate_range_start(mm, start, end);
230 	if (is_vm_hugetlb_page(vma))
231 		pages = hugetlb_change_protection(vma, start, end, newprot);
232 	else
233 		pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
234 	mmu_notifier_invalidate_range_end(mm, start, end);
235 
236 	return pages;
237 }
238 
239 int
240 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
241 	unsigned long start, unsigned long end, unsigned long newflags)
242 {
243 	struct mm_struct *mm = vma->vm_mm;
244 	unsigned long oldflags = vma->vm_flags;
245 	long nrpages = (end - start) >> PAGE_SHIFT;
246 	unsigned long charged = 0;
247 	pgoff_t pgoff;
248 	int error;
249 	int dirty_accountable = 0;
250 
251 	if (newflags == oldflags) {
252 		*pprev = vma;
253 		return 0;
254 	}
255 
256 	/*
257 	 * If we make a private mapping writable we increase our commit;
258 	 * but (without finer accounting) cannot reduce our commit if we
259 	 * make it unwritable again. hugetlb mapping were accounted for
260 	 * even if read-only so there is no need to account for them here
261 	 */
262 	if (newflags & VM_WRITE) {
263 		if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
264 						VM_SHARED|VM_NORESERVE))) {
265 			charged = nrpages;
266 			if (security_vm_enough_memory_mm(mm, charged))
267 				return -ENOMEM;
268 			newflags |= VM_ACCOUNT;
269 		}
270 	}
271 
272 	/*
273 	 * First try to merge with previous and/or next vma.
274 	 */
275 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
276 	*pprev = vma_merge(mm, *pprev, start, end, newflags,
277 			vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
278 	if (*pprev) {
279 		vma = *pprev;
280 		goto success;
281 	}
282 
283 	*pprev = vma;
284 
285 	if (start != vma->vm_start) {
286 		error = split_vma(mm, vma, start, 1);
287 		if (error)
288 			goto fail;
289 	}
290 
291 	if (end != vma->vm_end) {
292 		error = split_vma(mm, vma, end, 0);
293 		if (error)
294 			goto fail;
295 	}
296 
297 success:
298 	/*
299 	 * vm_flags and vm_page_prot are protected by the mmap_sem
300 	 * held in write mode.
301 	 */
302 	vma->vm_flags = newflags;
303 	vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
304 					  vm_get_page_prot(newflags));
305 
306 	if (vma_wants_writenotify(vma)) {
307 		vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
308 		dirty_accountable = 1;
309 	}
310 
311 	change_protection(vma, start, end, vma->vm_page_prot,
312 			  dirty_accountable, 0);
313 
314 	vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
315 	vm_stat_account(mm, newflags, vma->vm_file, nrpages);
316 	perf_event_mmap(vma);
317 	return 0;
318 
319 fail:
320 	vm_unacct_memory(charged);
321 	return error;
322 }
323 
324 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
325 		unsigned long, prot)
326 {
327 	unsigned long vm_flags, nstart, end, tmp, reqprot;
328 	struct vm_area_struct *vma, *prev;
329 	int error = -EINVAL;
330 	const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
331 	prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
332 	if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
333 		return -EINVAL;
334 
335 	if (start & ~PAGE_MASK)
336 		return -EINVAL;
337 	if (!len)
338 		return 0;
339 	len = PAGE_ALIGN(len);
340 	end = start + len;
341 	if (end <= start)
342 		return -ENOMEM;
343 	if (!arch_validate_prot(prot))
344 		return -EINVAL;
345 
346 	reqprot = prot;
347 	/*
348 	 * Does the application expect PROT_READ to imply PROT_EXEC:
349 	 */
350 	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
351 		prot |= PROT_EXEC;
352 
353 	vm_flags = calc_vm_prot_bits(prot);
354 
355 	down_write(&current->mm->mmap_sem);
356 
357 	vma = find_vma(current->mm, start);
358 	error = -ENOMEM;
359 	if (!vma)
360 		goto out;
361 	prev = vma->vm_prev;
362 	if (unlikely(grows & PROT_GROWSDOWN)) {
363 		if (vma->vm_start >= end)
364 			goto out;
365 		start = vma->vm_start;
366 		error = -EINVAL;
367 		if (!(vma->vm_flags & VM_GROWSDOWN))
368 			goto out;
369 	} else {
370 		if (vma->vm_start > start)
371 			goto out;
372 		if (unlikely(grows & PROT_GROWSUP)) {
373 			end = vma->vm_end;
374 			error = -EINVAL;
375 			if (!(vma->vm_flags & VM_GROWSUP))
376 				goto out;
377 		}
378 	}
379 	if (start > vma->vm_start)
380 		prev = vma;
381 
382 	for (nstart = start ; ; ) {
383 		unsigned long newflags;
384 
385 		/* Here we know that vma->vm_start <= nstart < vma->vm_end. */
386 
387 		newflags = vm_flags;
388 		newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
389 
390 		/* newflags >> 4 shift VM_MAY% in place of VM_% */
391 		if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
392 			error = -EACCES;
393 			goto out;
394 		}
395 
396 		error = security_file_mprotect(vma, reqprot, prot);
397 		if (error)
398 			goto out;
399 
400 		tmp = vma->vm_end;
401 		if (tmp > end)
402 			tmp = end;
403 		error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
404 		if (error)
405 			goto out;
406 		nstart = tmp;
407 
408 		if (nstart < prev->vm_end)
409 			nstart = prev->vm_end;
410 		if (nstart >= end)
411 			goto out;
412 
413 		vma = prev->vm_next;
414 		if (!vma || vma->vm_start != nstart) {
415 			error = -ENOMEM;
416 			goto out;
417 		}
418 	}
419 out:
420 	up_write(&current->mm->mmap_sem);
421 	return error;
422 }
423