4 * Copyright (C) 2002, Linus Torvalds
6 * 11Jan2003 Andrew Morton
10 #include <linux/kernel.h>
11 #include <linux/file.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>
22 #include <asm/unistd.h>
25 * POSIX_FADV_WILLNEED could set PG_Referenced, and POSIX_FADV_NOREUSE could
26 * deactivate the pages and clear PG_Referenced.
28 SYSCALL_DEFINE4(fadvise64_64
, int, fd
, loff_t
, offset
, loff_t
, len
, int, advice
)
30 struct fd f
= fdget(fd
);
32 struct address_space
*mapping
;
33 struct backing_dev_info
*bdi
;
34 loff_t endbyte
; /* inclusive */
37 unsigned long nrpages
;
43 inode
= file_inode(f
.file
);
44 if (S_ISFIFO(inode
->i_mode
)) {
49 mapping
= f
.file
->f_mapping
;
50 if (!mapping
|| len
< 0) {
55 bdi
= inode_to_bdi(mapping
->host
);
57 if (IS_DAX(inode
) || (bdi
== &noop_backing_dev_info
)) {
59 case POSIX_FADV_NORMAL
:
60 case POSIX_FADV_RANDOM
:
61 case POSIX_FADV_SEQUENTIAL
:
62 case POSIX_FADV_WILLNEED
:
63 case POSIX_FADV_NOREUSE
:
64 case POSIX_FADV_DONTNEED
:
65 /* no bad return value, but ignore advice */
73 /* Careful about overflows. Len == 0 means "as much as possible" */
74 endbyte
= offset
+ len
;
75 if (!len
|| endbyte
< len
)
78 endbyte
--; /* inclusive */
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
);
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
);
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
);
98 case POSIX_FADV_WILLNEED
:
99 /* First and last PARTIAL page! */
100 start_index
= offset
>> PAGE_SHIFT
;
101 end_index
= endbyte
>> PAGE_SHIFT
;
103 /* Careful about overflow on the "+1" */
104 nrpages
= end_index
- start_index
+ 1;
109 * Ignore return value because fadvise() shall return
110 * success even if filesystem can't retrieve a hint,
112 force_page_cache_readahead(mapping
, f
.file
, start_index
,
115 case POSIX_FADV_NOREUSE
:
117 case POSIX_FADV_DONTNEED
:
118 if (!inode_write_congested(mapping
->host
))
119 __filemap_fdatawrite_range(mapping
, offset
, endbyte
,
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.
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.
141 if (end_index
>= start_index
) {
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.
155 count
= invalidate_mapping_pages(mapping
,
156 start_index
, end_index
);
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.
164 if (count
< (end_index
- start_index
+ 1)) {
166 invalidate_mapping_pages(mapping
, start_index
,
179 #ifdef __ARCH_WANT_SYS_FADVISE64
181 SYSCALL_DEFINE4(fadvise64
, int, fd
, loff_t
, offset
, size_t, len
, int, advice
)
183 return sys_fadvise64_64(fd
, offset
, len
, advice
);