blob: eb3269e590024c1708cf686692abf1de800cfbfc
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 | /* |
72 | * Careful about overflows. Len == 0 means "as much as possible". Use |
73 | * unsigned math because signed overflows are undefined and UBSan |
74 | * complains. |
75 | */ |
76 | endbyte = (u64)offset + (u64)len; |
77 | if (!len || endbyte < len) |
78 | endbyte = -1; |
79 | else |
80 | endbyte--; /* inclusive */ |
81 | |
82 | bdi = inode_to_bdi(mapping->host); |
83 | |
84 | switch (advice) { |
85 | case POSIX_FADV_NORMAL: |
86 | f.file->f_ra.ra_pages = bdi->ra_pages; |
87 | spin_lock(&f.file->f_lock); |
88 | f.file->f_mode &= ~FMODE_RANDOM; |
89 | spin_unlock(&f.file->f_lock); |
90 | break; |
91 | case POSIX_FADV_RANDOM: |
92 | spin_lock(&f.file->f_lock); |
93 | f.file->f_mode |= FMODE_RANDOM; |
94 | spin_unlock(&f.file->f_lock); |
95 | break; |
96 | case POSIX_FADV_SEQUENTIAL: |
97 | f.file->f_ra.ra_pages = bdi->ra_pages * 2; |
98 | spin_lock(&f.file->f_lock); |
99 | f.file->f_mode &= ~FMODE_RANDOM; |
100 | spin_unlock(&f.file->f_lock); |
101 | break; |
102 | case POSIX_FADV_WILLNEED: |
103 | /* First and last PARTIAL page! */ |
104 | start_index = offset >> PAGE_SHIFT; |
105 | end_index = endbyte >> PAGE_SHIFT; |
106 | |
107 | /* Careful about overflow on the "+1" */ |
108 | nrpages = end_index - start_index + 1; |
109 | if (!nrpages) |
110 | nrpages = ~0UL; |
111 | |
112 | /* |
113 | * Ignore return value because fadvise() shall return |
114 | * success even if filesystem can't retrieve a hint, |
115 | */ |
116 | force_page_cache_readahead(mapping, f.file, start_index, |
117 | nrpages); |
118 | break; |
119 | case POSIX_FADV_NOREUSE: |
120 | break; |
121 | case POSIX_FADV_DONTNEED: |
122 | if (!inode_write_congested(mapping->host)) |
123 | __filemap_fdatawrite_range(mapping, offset, endbyte, |
124 | WB_SYNC_NONE); |
125 | |
126 | /* |
127 | * First and last FULL page! Partial pages are deliberately |
128 | * preserved on the expectation that it is better to preserve |
129 | * needed memory than to discard unneeded memory. |
130 | */ |
131 | start_index = (offset+(PAGE_SIZE-1)) >> PAGE_SHIFT; |
132 | end_index = (endbyte >> PAGE_SHIFT); |
133 | /* |
134 | * The page at end_index will be inclusively discarded according |
135 | * by invalidate_mapping_pages(), so subtracting 1 from |
136 | * end_index means we will skip the last page. But if endbyte |
137 | * is page aligned or is at the end of file, we should not skip |
138 | * that page - discarding the last page is safe enough. |
139 | */ |
140 | if ((endbyte & ~PAGE_MASK) != ~PAGE_MASK && |
141 | endbyte != inode->i_size - 1) { |
142 | /* First page is tricky as 0 - 1 = -1, but pgoff_t |
143 | * is unsigned, so the end_index >= start_index |
144 | * check below would be true and we'll discard the whole |
145 | * file cache which is not what was asked. |
146 | */ |
147 | if (end_index == 0) |
148 | break; |
149 | |
150 | end_index--; |
151 | } |
152 | |
153 | if (end_index >= start_index) { |
154 | unsigned long count = invalidate_mapping_pages(mapping, |
155 | start_index, end_index); |
156 | |
157 | /* |
158 | * If fewer pages were invalidated than expected then |
159 | * it is possible that some of the pages were on |
160 | * a per-cpu pagevec for a remote CPU. Drain all |
161 | * pagevecs and try again. |
162 | */ |
163 | if (count < (end_index - start_index + 1)) { |
164 | lru_add_drain_all(); |
165 | invalidate_mapping_pages(mapping, start_index, |
166 | end_index); |
167 | } |
168 | } |
169 | break; |
170 | default: |
171 | ret = -EINVAL; |
172 | } |
173 | out: |
174 | fdput(f); |
175 | return ret; |
176 | } |
177 | |
178 | #ifdef __ARCH_WANT_SYS_FADVISE64 |
179 | |
180 | SYSCALL_DEFINE4(fadvise64, int, fd, loff_t, offset, size_t, len, int, advice) |
181 | { |
182 | return sys_fadvise64_64(fd, offset, len, advice); |
183 | } |
184 | |
185 | #endif |
186 |