xref: /openbmc/linux/mm/fadvise.c (revision d3741027)
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 #include "internal.h"
26 
27 /*
28  * POSIX_FADV_WILLNEED could set PG_Referenced, and POSIX_FADV_NOREUSE could
29  * deactivate the pages and clear PG_Referenced.
30  */
31 
32 int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
33 {
34 	struct inode *inode;
35 	struct address_space *mapping;
36 	struct backing_dev_info *bdi;
37 	loff_t endbyte;			/* inclusive */
38 	pgoff_t start_index;
39 	pgoff_t end_index;
40 	unsigned long nrpages;
41 
42 	inode = file_inode(file);
43 	if (S_ISFIFO(inode->i_mode))
44 		return -ESPIPE;
45 
46 	mapping = file->f_mapping;
47 	if (!mapping || len < 0)
48 		return -EINVAL;
49 
50 	bdi = inode_to_bdi(mapping->host);
51 
52 	if (IS_DAX(inode) || (bdi == &noop_backing_dev_info)) {
53 		switch (advice) {
54 		case POSIX_FADV_NORMAL:
55 		case POSIX_FADV_RANDOM:
56 		case POSIX_FADV_SEQUENTIAL:
57 		case POSIX_FADV_WILLNEED:
58 		case POSIX_FADV_NOREUSE:
59 		case POSIX_FADV_DONTNEED:
60 			/* no bad return value, but ignore advice */
61 			break;
62 		default:
63 			return -EINVAL;
64 		}
65 		return 0;
66 	}
67 
68 	/*
69 	 * Careful about overflows. Len == 0 means "as much as possible".  Use
70 	 * unsigned math because signed overflows are undefined and UBSan
71 	 * complains.
72 	 */
73 	endbyte = (u64)offset + (u64)len;
74 	if (!len || endbyte < len)
75 		endbyte = -1;
76 	else
77 		endbyte--;		/* inclusive */
78 
79 	switch (advice) {
80 	case POSIX_FADV_NORMAL:
81 		file->f_ra.ra_pages = bdi->ra_pages;
82 		spin_lock(&file->f_lock);
83 		file->f_mode &= ~FMODE_RANDOM;
84 		spin_unlock(&file->f_lock);
85 		break;
86 	case POSIX_FADV_RANDOM:
87 		spin_lock(&file->f_lock);
88 		file->f_mode |= FMODE_RANDOM;
89 		spin_unlock(&file->f_lock);
90 		break;
91 	case POSIX_FADV_SEQUENTIAL:
92 		file->f_ra.ra_pages = bdi->ra_pages * 2;
93 		spin_lock(&file->f_lock);
94 		file->f_mode &= ~FMODE_RANDOM;
95 		spin_unlock(&file->f_lock);
96 		break;
97 	case POSIX_FADV_WILLNEED:
98 		/* First and last PARTIAL page! */
99 		start_index = offset >> PAGE_SHIFT;
100 		end_index = endbyte >> PAGE_SHIFT;
101 
102 		/* Careful about overflow on the "+1" */
103 		nrpages = end_index - start_index + 1;
104 		if (!nrpages)
105 			nrpages = ~0UL;
106 
107 		force_page_cache_readahead(mapping, file, start_index, nrpages);
108 		break;
109 	case POSIX_FADV_NOREUSE:
110 		break;
111 	case POSIX_FADV_DONTNEED:
112 		__filemap_fdatawrite_range(mapping, offset, endbyte,
113 					   WB_SYNC_NONE);
114 
115 		/*
116 		 * First and last FULL page! Partial pages are deliberately
117 		 * preserved on the expectation that it is better to preserve
118 		 * needed memory than to discard unneeded memory.
119 		 */
120 		start_index = (offset+(PAGE_SIZE-1)) >> PAGE_SHIFT;
121 		end_index = (endbyte >> PAGE_SHIFT);
122 		/*
123 		 * The page at end_index will be inclusively discarded according
124 		 * by invalidate_mapping_pages(), so subtracting 1 from
125 		 * end_index means we will skip the last page.  But if endbyte
126 		 * is page aligned or is at the end of file, we should not skip
127 		 * that page - discarding the last page is safe enough.
128 		 */
129 		if ((endbyte & ~PAGE_MASK) != ~PAGE_MASK &&
130 				endbyte != inode->i_size - 1) {
131 			/* First page is tricky as 0 - 1 = -1, but pgoff_t
132 			 * is unsigned, so the end_index >= start_index
133 			 * check below would be true and we'll discard the whole
134 			 * file cache which is not what was asked.
135 			 */
136 			if (end_index == 0)
137 				break;
138 
139 			end_index--;
140 		}
141 
142 		if (end_index >= start_index) {
143 			unsigned long nr_pagevec = 0;
144 
145 			/*
146 			 * It's common to FADV_DONTNEED right after
147 			 * the read or write that instantiates the
148 			 * pages, in which case there will be some
149 			 * sitting on the local LRU cache. Try to
150 			 * avoid the expensive remote drain and the
151 			 * second cache tree walk below by flushing
152 			 * them out right away.
153 			 */
154 			lru_add_drain();
155 
156 			invalidate_mapping_pagevec(mapping,
157 						start_index, end_index,
158 						&nr_pagevec);
159 
160 			/*
161 			 * If fewer pages were invalidated than expected then
162 			 * it is possible that some of the pages were on
163 			 * a per-cpu pagevec for a remote CPU. Drain all
164 			 * pagevecs and try again.
165 			 */
166 			if (nr_pagevec) {
167 				lru_add_drain_all();
168 				invalidate_mapping_pages(mapping, start_index,
169 						end_index);
170 			}
171 		}
172 		break;
173 	default:
174 		return -EINVAL;
175 	}
176 	return 0;
177 }
178 EXPORT_SYMBOL(generic_fadvise);
179 
180 int vfs_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
181 {
182 	if (file->f_op->fadvise)
183 		return file->f_op->fadvise(file, offset, len, advice);
184 
185 	return generic_fadvise(file, offset, len, advice);
186 }
187 EXPORT_SYMBOL(vfs_fadvise);
188 
189 #ifdef CONFIG_ADVISE_SYSCALLS
190 
191 int ksys_fadvise64_64(int fd, loff_t offset, loff_t len, int advice)
192 {
193 	struct fd f = fdget(fd);
194 	int ret;
195 
196 	if (!f.file)
197 		return -EBADF;
198 
199 	ret = vfs_fadvise(f.file, offset, len, advice);
200 
201 	fdput(f);
202 	return ret;
203 }
204 
205 SYSCALL_DEFINE4(fadvise64_64, int, fd, loff_t, offset, loff_t, len, int, advice)
206 {
207 	return ksys_fadvise64_64(fd, offset, len, advice);
208 }
209 
210 #ifdef __ARCH_WANT_SYS_FADVISE64
211 
212 SYSCALL_DEFINE4(fadvise64, int, fd, loff_t, offset, size_t, len, int, advice)
213 {
214 	return ksys_fadvise64_64(fd, offset, len, advice);
215 }
216 
217 #endif
218 
219 #if defined(CONFIG_COMPAT) && defined(__ARCH_WANT_COMPAT_FADVISE64_64)
220 
221 COMPAT_SYSCALL_DEFINE6(fadvise64_64, int, fd, compat_arg_u64_dual(offset),
222 		       compat_arg_u64_dual(len), int, advice)
223 {
224 	return ksys_fadvise64_64(fd, compat_arg_u64_glue(offset),
225 				 compat_arg_u64_glue(len), advice);
226 }
227 
228 #endif
229 #endif
230