xref: /openbmc/linux/mm/fadvise.c (revision aac5987a)
1 /*
2  * mm/fadvise.c
3  *
4  * Copyright (C) 2002, Linus Torvalds
5  *
6  * 11Jan2003	Andrew Morton
7  *		Initial version.
8  */
9 
10 #include <linux/kernel.h>
11 #include <linux/file.h>
12 #include <linux/fs.h>
13 #include <linux/mm.h>
14 #include <linux/pagemap.h>
15 #include <linux/backing-dev.h>
16 #include <linux/pagevec.h>
17 #include <linux/fadvise.h>
18 #include <linux/writeback.h>
19 #include <linux/syscalls.h>
20 #include <linux/swap.h>
21 
22 #include <asm/unistd.h>
23 
24 /*
25  * POSIX_FADV_WILLNEED could set PG_Referenced, and POSIX_FADV_NOREUSE could
26  * deactivate the pages and clear PG_Referenced.
27  */
28 SYSCALL_DEFINE4(fadvise64_64, int, fd, loff_t, offset, loff_t, len, int, advice)
29 {
30 	struct fd f = fdget(fd);
31 	struct inode *inode;
32 	struct address_space *mapping;
33 	struct backing_dev_info *bdi;
34 	loff_t endbyte;			/* inclusive */
35 	pgoff_t start_index;
36 	pgoff_t end_index;
37 	unsigned long nrpages;
38 	int ret = 0;
39 
40 	if (!f.file)
41 		return -EBADF;
42 
43 	inode = file_inode(f.file);
44 	if (S_ISFIFO(inode->i_mode)) {
45 		ret = -ESPIPE;
46 		goto out;
47 	}
48 
49 	mapping = f.file->f_mapping;
50 	if (!mapping || len < 0) {
51 		ret = -EINVAL;
52 		goto out;
53 	}
54 
55 	if (IS_DAX(inode)) {
56 		switch (advice) {
57 		case POSIX_FADV_NORMAL:
58 		case POSIX_FADV_RANDOM:
59 		case POSIX_FADV_SEQUENTIAL:
60 		case POSIX_FADV_WILLNEED:
61 		case POSIX_FADV_NOREUSE:
62 		case POSIX_FADV_DONTNEED:
63 			/* no bad return value, but ignore advice */
64 			break;
65 		default:
66 			ret = -EINVAL;
67 		}
68 		goto out;
69 	}
70 
71 	/* Careful about overflows. Len == 0 means "as much as possible" */
72 	endbyte = offset + len;
73 	if (!len || endbyte < len)
74 		endbyte = -1;
75 	else
76 		endbyte--;		/* inclusive */
77 
78 	bdi = inode_to_bdi(mapping->host);
79 
80 	switch (advice) {
81 	case POSIX_FADV_NORMAL:
82 		f.file->f_ra.ra_pages = bdi->ra_pages;
83 		spin_lock(&f.file->f_lock);
84 		f.file->f_mode &= ~FMODE_RANDOM;
85 		spin_unlock(&f.file->f_lock);
86 		break;
87 	case POSIX_FADV_RANDOM:
88 		spin_lock(&f.file->f_lock);
89 		f.file->f_mode |= FMODE_RANDOM;
90 		spin_unlock(&f.file->f_lock);
91 		break;
92 	case POSIX_FADV_SEQUENTIAL:
93 		f.file->f_ra.ra_pages = bdi->ra_pages * 2;
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_WILLNEED:
99 		/* First and last PARTIAL page! */
100 		start_index = offset >> PAGE_SHIFT;
101 		end_index = endbyte >> PAGE_SHIFT;
102 
103 		/* Careful about overflow on the "+1" */
104 		nrpages = end_index - start_index + 1;
105 		if (!nrpages)
106 			nrpages = ~0UL;
107 
108 		/*
109 		 * Ignore return value because fadvise() shall return
110 		 * success even if filesystem can't retrieve a hint,
111 		 */
112 		force_page_cache_readahead(mapping, f.file, start_index,
113 					   nrpages);
114 		break;
115 	case POSIX_FADV_NOREUSE:
116 		break;
117 	case POSIX_FADV_DONTNEED:
118 		if (!inode_write_congested(mapping->host))
119 			__filemap_fdatawrite_range(mapping, offset, endbyte,
120 						   WB_SYNC_NONE);
121 
122 		/*
123 		 * First and last FULL page! Partial pages are deliberately
124 		 * preserved on the expectation that it is better to preserve
125 		 * needed memory than to discard unneeded memory.
126 		 */
127 		start_index = (offset+(PAGE_SIZE-1)) >> PAGE_SHIFT;
128 		end_index = (endbyte >> PAGE_SHIFT);
129 		if ((endbyte & ~PAGE_MASK) != ~PAGE_MASK) {
130 			/* First page is tricky as 0 - 1 = -1, but pgoff_t
131 			 * is unsigned, so the end_index >= start_index
132 			 * check below would be true and we'll discard the whole
133 			 * file cache which is not what was asked.
134 			 */
135 			if (end_index == 0)
136 				break;
137 
138 			end_index--;
139 		}
140 
141 		if (end_index >= start_index) {
142 			unsigned long count;
143 
144 			/*
145 			 * It's common to FADV_DONTNEED right after
146 			 * the read or write that instantiates the
147 			 * pages, in which case there will be some
148 			 * sitting on the local LRU cache. Try to
149 			 * avoid the expensive remote drain and the
150 			 * second cache tree walk below by flushing
151 			 * them out right away.
152 			 */
153 			lru_add_drain();
154 
155 			count = invalidate_mapping_pages(mapping,
156 						start_index, end_index);
157 
158 			/*
159 			 * If fewer pages were invalidated than expected then
160 			 * it is possible that some of the pages were on
161 			 * a per-cpu pagevec for a remote CPU. Drain all
162 			 * pagevecs and try again.
163 			 */
164 			if (count < (end_index - start_index + 1)) {
165 				lru_add_drain_all();
166 				invalidate_mapping_pages(mapping, start_index,
167 						end_index);
168 			}
169 		}
170 		break;
171 	default:
172 		ret = -EINVAL;
173 	}
174 out:
175 	fdput(f);
176 	return ret;
177 }
178 
179 #ifdef __ARCH_WANT_SYS_FADVISE64
180 
181 SYSCALL_DEFINE4(fadvise64, int, fd, loff_t, offset, size_t, len, int, advice)
182 {
183 	return sys_fadvise64_64(fd, offset, len, advice);
184 }
185 
186 #endif
187