xref: /linux/mm/madvise.c (revision ba6e8564f459211117ce300eae2c7fdd23befe34)
1 /*
2  *	linux/mm/madvise.c
3  *
4  * Copyright (C) 1999  Linus Torvalds
5  * Copyright (C) 2002  Christoph Hellwig
6  */
7 
8 #include <linux/mman.h>
9 #include <linux/pagemap.h>
10 #include <linux/syscalls.h>
11 #include <linux/mempolicy.h>
12 #include <linux/hugetlb.h>
13 
14 /*
15  * We can potentially split a vm area into separate
16  * areas, each area with its own behavior.
17  */
18 static long madvise_behavior(struct vm_area_struct * vma,
19 		     struct vm_area_struct **prev,
20 		     unsigned long start, unsigned long end, int behavior)
21 {
22 	struct mm_struct * mm = vma->vm_mm;
23 	int error = 0;
24 	pgoff_t pgoff;
25 	int new_flags = vma->vm_flags;
26 
27 	switch (behavior) {
28 	case MADV_NORMAL:
29 		new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
30 		break;
31 	case MADV_SEQUENTIAL:
32 		new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
33 		break;
34 	case MADV_RANDOM:
35 		new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
36 		break;
37 	case MADV_DONTFORK:
38 		new_flags |= VM_DONTCOPY;
39 		break;
40 	case MADV_DOFORK:
41 		new_flags &= ~VM_DONTCOPY;
42 		break;
43 	}
44 
45 	if (new_flags == vma->vm_flags) {
46 		*prev = vma;
47 		goto out;
48 	}
49 
50 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
51 	*prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
52 				vma->vm_file, pgoff, vma_policy(vma));
53 	if (*prev) {
54 		vma = *prev;
55 		goto success;
56 	}
57 
58 	*prev = vma;
59 
60 	if (start != vma->vm_start) {
61 		error = split_vma(mm, vma, start, 1);
62 		if (error)
63 			goto out;
64 	}
65 
66 	if (end != vma->vm_end) {
67 		error = split_vma(mm, vma, end, 0);
68 		if (error)
69 			goto out;
70 	}
71 
72 success:
73 	/*
74 	 * vm_flags is protected by the mmap_sem held in write mode.
75 	 */
76 	vma->vm_flags = new_flags;
77 
78 out:
79 	if (error == -ENOMEM)
80 		error = -EAGAIN;
81 	return error;
82 }
83 
84 /*
85  * Schedule all required I/O operations.  Do not wait for completion.
86  */
87 static long madvise_willneed(struct vm_area_struct * vma,
88 			     struct vm_area_struct ** prev,
89 			     unsigned long start, unsigned long end)
90 {
91 	struct file *file = vma->vm_file;
92 
93 	if (!file)
94 		return -EBADF;
95 
96 	if (file->f_mapping->a_ops->get_xip_page) {
97 		/* no bad return value, but ignore advice */
98 		return 0;
99 	}
100 
101 	*prev = vma;
102 	start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
103 	if (end > vma->vm_end)
104 		end = vma->vm_end;
105 	end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
106 
107 	force_page_cache_readahead(file->f_mapping,
108 			file, start, max_sane_readahead(end - start));
109 	return 0;
110 }
111 
112 /*
113  * Application no longer needs these pages.  If the pages are dirty,
114  * it's OK to just throw them away.  The app will be more careful about
115  * data it wants to keep.  Be sure to free swap resources too.  The
116  * zap_page_range call sets things up for refill_inactive to actually free
117  * these pages later if no one else has touched them in the meantime,
118  * although we could add these pages to a global reuse list for
119  * refill_inactive to pick up before reclaiming other pages.
120  *
121  * NB: This interface discards data rather than pushes it out to swap,
122  * as some implementations do.  This has performance implications for
123  * applications like large transactional databases which want to discard
124  * pages in anonymous maps after committing to backing store the data
125  * that was kept in them.  There is no reason to write this data out to
126  * the swap area if the application is discarding it.
127  *
128  * An interface that causes the system to free clean pages and flush
129  * dirty pages is already available as msync(MS_INVALIDATE).
130  */
131 static long madvise_dontneed(struct vm_area_struct * vma,
132 			     struct vm_area_struct ** prev,
133 			     unsigned long start, unsigned long end)
134 {
135 	*prev = vma;
136 	if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
137 		return -EINVAL;
138 
139 	if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
140 		struct zap_details details = {
141 			.nonlinear_vma = vma,
142 			.last_index = ULONG_MAX,
143 		};
144 		zap_page_range(vma, start, end - start, &details);
145 	} else
146 		zap_page_range(vma, start, end - start, NULL);
147 	return 0;
148 }
149 
150 /*
151  * Application wants to free up the pages and associated backing store.
152  * This is effectively punching a hole into the middle of a file.
153  *
154  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
155  * Other filesystems return -ENOSYS.
156  */
157 static long madvise_remove(struct vm_area_struct *vma,
158 				struct vm_area_struct **prev,
159 				unsigned long start, unsigned long end)
160 {
161 	struct address_space *mapping;
162 	loff_t offset, endoff;
163 	int error;
164 
165 	*prev = NULL;	/* tell sys_madvise we drop mmap_sem */
166 
167 	if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
168 		return -EINVAL;
169 
170 	if (!vma->vm_file || !vma->vm_file->f_mapping
171 		|| !vma->vm_file->f_mapping->host) {
172 			return -EINVAL;
173 	}
174 
175 	if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
176 		return -EACCES;
177 
178 	mapping = vma->vm_file->f_mapping;
179 
180 	offset = (loff_t)(start - vma->vm_start)
181 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
182 	endoff = (loff_t)(end - vma->vm_start - 1)
183 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
184 
185 	/* vmtruncate_range needs to take i_mutex and i_alloc_sem */
186 	up_write(&current->mm->mmap_sem);
187 	error = vmtruncate_range(mapping->host, offset, endoff);
188 	down_write(&current->mm->mmap_sem);
189 	return error;
190 }
191 
192 static long
193 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
194 		unsigned long start, unsigned long end, int behavior)
195 {
196 	long error;
197 
198 	switch (behavior) {
199 	case MADV_DOFORK:
200 		if (vma->vm_flags & VM_IO) {
201 			error = -EINVAL;
202 			break;
203 		}
204 	case MADV_DONTFORK:
205 	case MADV_NORMAL:
206 	case MADV_SEQUENTIAL:
207 	case MADV_RANDOM:
208 		error = madvise_behavior(vma, prev, start, end, behavior);
209 		break;
210 	case MADV_REMOVE:
211 		error = madvise_remove(vma, prev, start, end);
212 		break;
213 
214 	case MADV_WILLNEED:
215 		error = madvise_willneed(vma, prev, start, end);
216 		break;
217 
218 	case MADV_DONTNEED:
219 		error = madvise_dontneed(vma, prev, start, end);
220 		break;
221 
222 	default:
223 		error = -EINVAL;
224 		break;
225 	}
226 	return error;
227 }
228 
229 /*
230  * The madvise(2) system call.
231  *
232  * Applications can use madvise() to advise the kernel how it should
233  * handle paging I/O in this VM area.  The idea is to help the kernel
234  * use appropriate read-ahead and caching techniques.  The information
235  * provided is advisory only, and can be safely disregarded by the
236  * kernel without affecting the correct operation of the application.
237  *
238  * behavior values:
239  *  MADV_NORMAL - the default behavior is to read clusters.  This
240  *		results in some read-ahead and read-behind.
241  *  MADV_RANDOM - the system should read the minimum amount of data
242  *		on any access, since it is unlikely that the appli-
243  *		cation will need more than what it asks for.
244  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
245  *		once, so they can be aggressively read ahead, and
246  *		can be freed soon after they are accessed.
247  *  MADV_WILLNEED - the application is notifying the system to read
248  *		some pages ahead.
249  *  MADV_DONTNEED - the application is finished with the given range,
250  *		so the kernel can free resources associated with it.
251  *  MADV_REMOVE - the application wants to free up the given range of
252  *		pages and associated backing store.
253  *
254  * return values:
255  *  zero    - success
256  *  -EINVAL - start + len < 0, start is not page-aligned,
257  *		"behavior" is not a valid value, or application
258  *		is attempting to release locked or shared pages.
259  *  -ENOMEM - addresses in the specified range are not currently
260  *		mapped, or are outside the AS of the process.
261  *  -EIO    - an I/O error occurred while paging in data.
262  *  -EBADF  - map exists, but area maps something that isn't a file.
263  *  -EAGAIN - a kernel resource was temporarily unavailable.
264  */
265 asmlinkage long sys_madvise(unsigned long start, size_t len_in, int behavior)
266 {
267 	unsigned long end, tmp;
268 	struct vm_area_struct * vma, *prev;
269 	int unmapped_error = 0;
270 	int error = -EINVAL;
271 	size_t len;
272 
273 	down_write(&current->mm->mmap_sem);
274 
275 	if (start & ~PAGE_MASK)
276 		goto out;
277 	len = (len_in + ~PAGE_MASK) & PAGE_MASK;
278 
279 	/* Check to see whether len was rounded up from small -ve to zero */
280 	if (len_in && !len)
281 		goto out;
282 
283 	end = start + len;
284 	if (end < start)
285 		goto out;
286 
287 	error = 0;
288 	if (end == start)
289 		goto out;
290 
291 	/*
292 	 * If the interval [start,end) covers some unmapped address
293 	 * ranges, just ignore them, but return -ENOMEM at the end.
294 	 * - different from the way of handling in mlock etc.
295 	 */
296 	vma = find_vma_prev(current->mm, start, &prev);
297 	if (vma && start > vma->vm_start)
298 		prev = vma;
299 
300 	for (;;) {
301 		/* Still start < end. */
302 		error = -ENOMEM;
303 		if (!vma)
304 			goto out;
305 
306 		/* Here start < (end|vma->vm_end). */
307 		if (start < vma->vm_start) {
308 			unmapped_error = -ENOMEM;
309 			start = vma->vm_start;
310 			if (start >= end)
311 				goto out;
312 		}
313 
314 		/* Here vma->vm_start <= start < (end|vma->vm_end) */
315 		tmp = vma->vm_end;
316 		if (end < tmp)
317 			tmp = end;
318 
319 		/* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
320 		error = madvise_vma(vma, &prev, start, tmp, behavior);
321 		if (error)
322 			goto out;
323 		start = tmp;
324 		if (prev && start < prev->vm_end)
325 			start = prev->vm_end;
326 		error = unmapped_error;
327 		if (start >= end)
328 			goto out;
329 		if (prev)
330 			vma = prev->vm_next;
331 		else	/* madvise_remove dropped mmap_sem */
332 			vma = find_vma(current->mm, start);
333 	}
334 out:
335 	up_write(&current->mm->mmap_sem);
336 	return error;
337 }
338