xref: /openbmc/linux/mm/fadvise.c (revision e53134fe)
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 ksys_fadvise64_64(int fd, loff_t offset, loff_t len, int advice)
31 {
32 	struct fd f = fdget(fd);
33 	struct inode *inode;
34 	struct address_space *mapping;
35 	struct backing_dev_info *bdi;
36 	loff_t endbyte;			/* inclusive */
37 	pgoff_t start_index;
38 	pgoff_t end_index;
39 	unsigned long nrpages;
40 	int ret = 0;
41 
42 	if (!f.file)
43 		return -EBADF;
44 
45 	inode = file_inode(f.file);
46 	if (S_ISFIFO(inode->i_mode)) {
47 		ret = -ESPIPE;
48 		goto out;
49 	}
50 
51 	mapping = f.file->f_mapping;
52 	if (!mapping || len < 0) {
53 		ret = -EINVAL;
54 		goto out;
55 	}
56 
57 	bdi = inode_to_bdi(mapping->host);
58 
59 	if (IS_DAX(inode) || (bdi == &noop_backing_dev_info)) {
60 		switch (advice) {
61 		case POSIX_FADV_NORMAL:
62 		case POSIX_FADV_RANDOM:
63 		case POSIX_FADV_SEQUENTIAL:
64 		case POSIX_FADV_WILLNEED:
65 		case POSIX_FADV_NOREUSE:
66 		case POSIX_FADV_DONTNEED:
67 			/* no bad return value, but ignore advice */
68 			break;
69 		default:
70 			ret = -EINVAL;
71 		}
72 		goto out;
73 	}
74 
75 	/*
76 	 * Careful about overflows. Len == 0 means "as much as possible".  Use
77 	 * unsigned math because signed overflows are undefined and UBSan
78 	 * complains.
79 	 */
80 	endbyte = (u64)offset + (u64)len;
81 	if (!len || endbyte < len)
82 		endbyte = -1;
83 	else
84 		endbyte--;		/* inclusive */
85 
86 	switch (advice) {
87 	case POSIX_FADV_NORMAL:
88 		f.file->f_ra.ra_pages = bdi->ra_pages;
89 		spin_lock(&f.file->f_lock);
90 		f.file->f_mode &= ~FMODE_RANDOM;
91 		spin_unlock(&f.file->f_lock);
92 		break;
93 	case POSIX_FADV_RANDOM:
94 		spin_lock(&f.file->f_lock);
95 		f.file->f_mode |= FMODE_RANDOM;
96 		spin_unlock(&f.file->f_lock);
97 		break;
98 	case POSIX_FADV_SEQUENTIAL:
99 		f.file->f_ra.ra_pages = bdi->ra_pages * 2;
100 		spin_lock(&f.file->f_lock);
101 		f.file->f_mode &= ~FMODE_RANDOM;
102 		spin_unlock(&f.file->f_lock);
103 		break;
104 	case POSIX_FADV_WILLNEED:
105 		/* First and last PARTIAL page! */
106 		start_index = offset >> PAGE_SHIFT;
107 		end_index = endbyte >> PAGE_SHIFT;
108 
109 		/* Careful about overflow on the "+1" */
110 		nrpages = end_index - start_index + 1;
111 		if (!nrpages)
112 			nrpages = ~0UL;
113 
114 		/*
115 		 * Ignore return value because fadvise() shall return
116 		 * success even if filesystem can't retrieve a hint,
117 		 */
118 		force_page_cache_readahead(mapping, f.file, start_index,
119 					   nrpages);
120 		break;
121 	case POSIX_FADV_NOREUSE:
122 		break;
123 	case POSIX_FADV_DONTNEED:
124 		if (!inode_write_congested(mapping->host))
125 			__filemap_fdatawrite_range(mapping, offset, endbyte,
126 						   WB_SYNC_NONE);
127 
128 		/*
129 		 * First and last FULL page! Partial pages are deliberately
130 		 * preserved on the expectation that it is better to preserve
131 		 * needed memory than to discard unneeded memory.
132 		 */
133 		start_index = (offset+(PAGE_SIZE-1)) >> PAGE_SHIFT;
134 		end_index = (endbyte >> PAGE_SHIFT);
135 		/*
136 		 * The page at end_index will be inclusively discarded according
137 		 * by invalidate_mapping_pages(), so subtracting 1 from
138 		 * end_index means we will skip the last page.  But if endbyte
139 		 * is page aligned or is at the end of file, we should not skip
140 		 * that page - discarding the last page is safe enough.
141 		 */
142 		if ((endbyte & ~PAGE_MASK) != ~PAGE_MASK &&
143 				endbyte != inode->i_size - 1) {
144 			/* First page is tricky as 0 - 1 = -1, but pgoff_t
145 			 * is unsigned, so the end_index >= start_index
146 			 * check below would be true and we'll discard the whole
147 			 * file cache which is not what was asked.
148 			 */
149 			if (end_index == 0)
150 				break;
151 
152 			end_index--;
153 		}
154 
155 		if (end_index >= start_index) {
156 			unsigned long count;
157 
158 			/*
159 			 * It's common to FADV_DONTNEED right after
160 			 * the read or write that instantiates the
161 			 * pages, in which case there will be some
162 			 * sitting on the local LRU cache. Try to
163 			 * avoid the expensive remote drain and the
164 			 * second cache tree walk below by flushing
165 			 * them out right away.
166 			 */
167 			lru_add_drain();
168 
169 			count = invalidate_mapping_pages(mapping,
170 						start_index, end_index);
171 
172 			/*
173 			 * If fewer pages were invalidated than expected then
174 			 * it is possible that some of the pages were on
175 			 * a per-cpu pagevec for a remote CPU. Drain all
176 			 * pagevecs and try again.
177 			 */
178 			if (count < (end_index - start_index + 1)) {
179 				lru_add_drain_all();
180 				invalidate_mapping_pages(mapping, start_index,
181 						end_index);
182 			}
183 		}
184 		break;
185 	default:
186 		ret = -EINVAL;
187 	}
188 out:
189 	fdput(f);
190 	return ret;
191 }
192 
193 SYSCALL_DEFINE4(fadvise64_64, int, fd, loff_t, offset, loff_t, len, int, advice)
194 {
195 	return ksys_fadvise64_64(fd, offset, len, advice);
196 }
197 
198 #ifdef __ARCH_WANT_SYS_FADVISE64
199 
200 SYSCALL_DEFINE4(fadvise64, int, fd, loff_t, offset, size_t, len, int, advice)
201 {
202 	return ksys_fadvise64_64(fd, offset, len, advice);
203 }
204 
205 #endif
206