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 */
SYSCALL_DEFINE4(fadvise64_64,int,fd,loff_t,offset,loff_t,len,int,advice)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_CACHE_SHIFT;
105 end_index = endbyte >> PAGE_CACHE_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_CACHE_SIZE-1)) >> PAGE_CACHE_SHIFT;
132 end_index = (endbyte >> PAGE_CACHE_SHIFT);
133
134 if (end_index >= start_index) {
135 unsigned long count = invalidate_mapping_pages(mapping,
136 start_index, end_index);
137
138 /*
139 * If fewer pages were invalidated than expected then
140 * it is possible that some of the pages were on
141 * a per-cpu pagevec for a remote CPU. Drain all
142 * pagevecs and try again.
143 */
144 if (count < (end_index - start_index + 1)) {
145 lru_add_drain_all();
146 invalidate_mapping_pages(mapping, start_index,
147 end_index);
148 }
149 }
150 break;
151 default:
152 ret = -EINVAL;
153 }
154 out:
155 fdput(f);
156 return ret;
157 }
158
159 #ifdef __ARCH_WANT_SYS_FADVISE64
160
SYSCALL_DEFINE4(fadvise64,int,fd,loff_t,offset,size_t,len,int,advice)161 SYSCALL_DEFINE4(fadvise64, int, fd, loff_t, offset, size_t, len, int, advice)
162 {
163 return sys_fadvise64_64(fd, offset, len, advice);
164 }
165
166 #endif
167