[PATCH] fadvise(): write commands
[safe/jmp/linux-2.6] / mm / fadvise.c
1 /*
2  * mm/fadvise.c
3  *
4  * Copyright (C) 2002, Linus Torvalds
5  *
6  * 11Jan2003    akpm@digeo.com
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
21 #include <asm/unistd.h>
22
23 /*
24  * POSIX_FADV_WILLNEED could set PG_Referenced, and POSIX_FADV_NOREUSE could
25  * deactivate the pages and clear PG_Referenced.
26  *
27  * LINUX_FADV_ASYNC_WRITE: start async writeout of any dirty pages between file
28  * offsets `offset' and `offset+len' inclusive.  Any pages which are currently
29  * under writeout are skipped, whether or not they are dirty.
30  *
31  * LINUX_FADV_WRITE_WAIT: wait upon writeout of any dirty pages between file
32  * offsets `offset' and `offset+len'.
33  *
34  * By combining these two operations the application may do several things:
35  *
36  * LINUX_FADV_ASYNC_WRITE: push some or all of the dirty pages at the disk.
37  *
38  * LINUX_FADV_WRITE_WAIT, LINUX_FADV_ASYNC_WRITE: push all of the currently
39  * dirty pages at the disk.
40  *
41  * LINUX_FADV_WRITE_WAIT, LINUX_FADV_ASYNC_WRITE, LINUX_FADV_WRITE_WAIT: push
42  * all of the currently dirty pages at the disk, wait until they have been
43  * written.
44  *
45  * It should be noted that none of these operations write out the file's
46  * metadata.  So unless the application is strictly performing overwrites of
47  * already-instantiated disk blocks, there are no guarantees here that the data
48  * will be available after a crash.
49  */
50 asmlinkage long sys_fadvise64_64(int fd, loff_t offset, loff_t len, int advice)
51 {
52         struct file *file = fget(fd);
53         struct address_space *mapping;
54         struct backing_dev_info *bdi;
55         loff_t endbyte;                 /* inclusive */
56         pgoff_t start_index;
57         pgoff_t end_index;
58         unsigned long nrpages;
59         int ret = 0;
60
61         if (!file)
62                 return -EBADF;
63
64         if (S_ISFIFO(file->f_dentry->d_inode->i_mode)) {
65                 ret = -ESPIPE;
66                 goto out;
67         }
68
69         mapping = file->f_mapping;
70         if (!mapping || len < 0) {
71                 ret = -EINVAL;
72                 goto out;
73         }
74
75         if (mapping->a_ops->get_xip_page)
76                 /* no bad return value, but ignore advice */
77                 goto out;
78
79         /* Careful about overflows. Len == 0 means "as much as possible" */
80         endbyte = offset + len;
81         if (!len || endbyte < len)
82                 endbyte = -1;
83         else
84                 endbyte--;              /* inclusive */
85
86         bdi = mapping->backing_dev_info;
87
88         switch (advice) {
89         case POSIX_FADV_NORMAL:
90                 file->f_ra.ra_pages = bdi->ra_pages;
91                 break;
92         case POSIX_FADV_RANDOM:
93                 file->f_ra.ra_pages = 0;
94                 break;
95         case POSIX_FADV_SEQUENTIAL:
96                 file->f_ra.ra_pages = bdi->ra_pages * 2;
97                 break;
98         case POSIX_FADV_WILLNEED:
99         case POSIX_FADV_NOREUSE:
100                 if (!mapping->a_ops->readpage) {
101                         ret = -EINVAL;
102                         break;
103                 }
104
105                 /* First and last PARTIAL page! */
106                 start_index = offset >> PAGE_CACHE_SHIFT;
107                 end_index = endbyte >> PAGE_CACHE_SHIFT;
108
109                 /* Careful about overflow on the "+1" */
110                 nrpages = end_index - start_index + 1;
111                 if (!nrpages)
112                         nrpages = ~0UL;
113                 
114                 ret = force_page_cache_readahead(mapping, file,
115                                 start_index,
116                                 max_sane_readahead(nrpages));
117                 if (ret > 0)
118                         ret = 0;
119                 break;
120         case POSIX_FADV_DONTNEED:
121                 if (!bdi_write_congested(mapping->backing_dev_info))
122                         filemap_flush(mapping);
123
124                 /* First and last FULL page! */
125                 start_index = (offset+(PAGE_CACHE_SIZE-1)) >> PAGE_CACHE_SHIFT;
126                 end_index = (endbyte >> PAGE_CACHE_SHIFT);
127
128                 if (end_index >= start_index)
129                         invalidate_mapping_pages(mapping, start_index,
130                                                 end_index);
131                 break;
132         case LINUX_FADV_ASYNC_WRITE:
133                 ret = __filemap_fdatawrite_range(mapping, offset, endbyte,
134                                                 WB_SYNC_NONE);
135                 break;
136         case LINUX_FADV_WRITE_WAIT:
137                 ret = wait_on_page_writeback_range(mapping,
138                                         offset >> PAGE_CACHE_SHIFT,
139                                         endbyte >> PAGE_CACHE_SHIFT);
140                 break;
141         default:
142                 ret = -EINVAL;
143         }
144 out:
145         fput(file);
146         return ret;
147 }
148
149 #ifdef __ARCH_WANT_SYS_FADVISE64
150
151 asmlinkage long sys_fadvise64(int fd, loff_t offset, size_t len, int advice)
152 {
153         return sys_fadvise64_64(fd, offset, len, advice);
154 }
155
156 #endif