xref: /linux/mm/fadvise.c (revision ca64d84e93762f4e587e040a44ad9f6089afc777)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * mm/fadvise.c
4  *
5  * Copyright (C) 2002, Linus Torvalds
6  *
7  * 11Jan2003	Andrew Morton
8  *		Initial version.
9  */
10 
11 #include <linux/kernel.h>
12 #include <linux/file.h>
13 #include <linux/fs.h>
14 #include <linux/mm.h>
15 #include <linux/pagemap.h>
16 #include <linux/backing-dev.h>
17 #include <linux/pagevec.h>
18 #include <linux/fadvise.h>
19 #include <linux/writeback.h>
20 #include <linux/syscalls.h>
21 #include <linux/swap.h>
22 
23 #include <asm/unistd.h>
24 
25 /*
26  * POSIX_FADV_WILLNEED could set PG_Referenced, and POSIX_FADV_NOREUSE could
27  * deactivate the pages and clear PG_Referenced.
28  */
29 
30 int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
31 {
32 	struct inode *inode;
33 	struct address_space *mapping;
34 	struct backing_dev_info *bdi;
35 	loff_t endbyte;			/* inclusive */
36 	pgoff_t start_index;
37 	pgoff_t end_index;
38 	unsigned long nrpages;
39 
40 	inode = file_inode(file);
41 	if (S_ISFIFO(inode->i_mode))
42 		return -ESPIPE;
43 
44 	mapping = file->f_mapping;
45 	if (!mapping || len < 0)
46 		return -EINVAL;
47 
48 	bdi = inode_to_bdi(mapping->host);
49 
50 	if (IS_DAX(inode) || (bdi == &noop_backing_dev_info)) {
51 		switch (advice) {
52 		case POSIX_FADV_NORMAL:
53 		case POSIX_FADV_RANDOM:
54 		case POSIX_FADV_SEQUENTIAL:
55 		case POSIX_FADV_WILLNEED:
56 		case POSIX_FADV_NOREUSE:
57 		case POSIX_FADV_DONTNEED:
58 			/* no bad return value, but ignore advice */
59 			break;
60 		default:
61 			return -EINVAL;
62 		}
63 		return 0;
64 	}
65 
66 	/*
67 	 * Careful about overflows. Len == 0 means "as much as possible".  Use
68 	 * unsigned math because signed overflows are undefined and UBSan
69 	 * complains.
70 	 */
71 	endbyte = (u64)offset + (u64)len;
72 	if (!len || endbyte < len)
73 		endbyte = -1;
74 	else
75 		endbyte--;		/* inclusive */
76 
77 	switch (advice) {
78 	case POSIX_FADV_NORMAL:
79 		file->f_ra.ra_pages = bdi->ra_pages;
80 		spin_lock(&file->f_lock);
81 		file->f_mode &= ~FMODE_RANDOM;
82 		spin_unlock(&file->f_lock);
83 		break;
84 	case POSIX_FADV_RANDOM:
85 		spin_lock(&file->f_lock);
86 		file->f_mode |= FMODE_RANDOM;
87 		spin_unlock(&file->f_lock);
88 		break;
89 	case POSIX_FADV_SEQUENTIAL:
90 		file->f_ra.ra_pages = bdi->ra_pages * 2;
91 		spin_lock(&file->f_lock);
92 		file->f_mode &= ~FMODE_RANDOM;
93 		spin_unlock(&file->f_lock);
94 		break;
95 	case POSIX_FADV_WILLNEED:
96 		/* First and last PARTIAL page! */
97 		start_index = offset >> PAGE_SHIFT;
98 		end_index = endbyte >> PAGE_SHIFT;
99 
100 		/* Careful about overflow on the "+1" */
101 		nrpages = end_index - start_index + 1;
102 		if (!nrpages)
103 			nrpages = ~0UL;
104 
105 		/*
106 		 * Ignore return value because fadvise() shall return
107 		 * success even if filesystem can't retrieve a hint,
108 		 */
109 		force_page_cache_readahead(mapping, file, start_index, nrpages);
110 		break;
111 	case POSIX_FADV_NOREUSE:
112 		break;
113 	case POSIX_FADV_DONTNEED:
114 		if (!inode_write_congested(mapping->host))
115 			__filemap_fdatawrite_range(mapping, offset, endbyte,
116 						   WB_SYNC_NONE);
117 
118 		/*
119 		 * First and last FULL page! Partial pages are deliberately
120 		 * preserved on the expectation that it is better to preserve
121 		 * needed memory than to discard unneeded memory.
122 		 */
123 		start_index = (offset+(PAGE_SIZE-1)) >> PAGE_SHIFT;
124 		end_index = (endbyte >> PAGE_SHIFT);
125 		/*
126 		 * The page at end_index will be inclusively discarded according
127 		 * by invalidate_mapping_pages(), so subtracting 1 from
128 		 * end_index means we will skip the last page.  But if endbyte
129 		 * is page aligned or is at the end of file, we should not skip
130 		 * that page - discarding the last page is safe enough.
131 		 */
132 		if ((endbyte & ~PAGE_MASK) != ~PAGE_MASK &&
133 				endbyte != inode->i_size - 1) {
134 			/* First page is tricky as 0 - 1 = -1, but pgoff_t
135 			 * is unsigned, so the end_index >= start_index
136 			 * check below would be true and we'll discard the whole
137 			 * file cache which is not what was asked.
138 			 */
139 			if (end_index == 0)
140 				break;
141 
142 			end_index--;
143 		}
144 
145 		if (end_index >= start_index) {
146 			unsigned long count;
147 
148 			/*
149 			 * It's common to FADV_DONTNEED right after
150 			 * the read or write that instantiates the
151 			 * pages, in which case there will be some
152 			 * sitting on the local LRU cache. Try to
153 			 * avoid the expensive remote drain and the
154 			 * second cache tree walk below by flushing
155 			 * them out right away.
156 			 */
157 			lru_add_drain();
158 
159 			count = invalidate_mapping_pages(mapping,
160 						start_index, end_index);
161 
162 			/*
163 			 * If fewer pages were invalidated than expected then
164 			 * it is possible that some of the pages were on
165 			 * a per-cpu pagevec for a remote CPU. Drain all
166 			 * pagevecs and try again.
167 			 */
168 			if (count < (end_index - start_index + 1)) {
169 				lru_add_drain_all();
170 				invalidate_mapping_pages(mapping, start_index,
171 						end_index);
172 			}
173 		}
174 		break;
175 	default:
176 		return -EINVAL;
177 	}
178 	return 0;
179 }
180 EXPORT_SYMBOL(generic_fadvise);
181 
182 int vfs_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
183 {
184 	if (file->f_op->fadvise)
185 		return file->f_op->fadvise(file, offset, len, advice);
186 
187 	return generic_fadvise(file, offset, len, advice);
188 }
189 EXPORT_SYMBOL(vfs_fadvise);
190 
191 #ifdef CONFIG_ADVISE_SYSCALLS
192 
193 int ksys_fadvise64_64(int fd, loff_t offset, loff_t len, int advice)
194 {
195 	struct fd f = fdget(fd);
196 	int ret;
197 
198 	if (!f.file)
199 		return -EBADF;
200 
201 	ret = vfs_fadvise(f.file, offset, len, advice);
202 
203 	fdput(f);
204 	return ret;
205 }
206 
207 SYSCALL_DEFINE4(fadvise64_64, int, fd, loff_t, offset, loff_t, len, int, advice)
208 {
209 	return ksys_fadvise64_64(fd, offset, len, advice);
210 }
211 
212 #ifdef __ARCH_WANT_SYS_FADVISE64
213 
214 SYSCALL_DEFINE4(fadvise64, int, fd, loff_t, offset, size_t, len, int, advice)
215 {
216 	return ksys_fadvise64_64(fd, offset, len, advice);
217 }
218 
219 #endif
220 #endif
221