xref: /openbmc/linux/fs/xfs/xfs_pnfs.c (revision 6dfcd296)
1 /*
2  * Copyright (c) 2014 Christoph Hellwig.
3  */
4 #include <linux/iomap.h>
5 #include "xfs.h"
6 #include "xfs_format.h"
7 #include "xfs_log_format.h"
8 #include "xfs_trans_resv.h"
9 #include "xfs_sb.h"
10 #include "xfs_mount.h"
11 #include "xfs_inode.h"
12 #include "xfs_trans.h"
13 #include "xfs_log.h"
14 #include "xfs_bmap.h"
15 #include "xfs_bmap_util.h"
16 #include "xfs_error.h"
17 #include "xfs_iomap.h"
18 #include "xfs_shared.h"
19 #include "xfs_bit.h"
20 #include "xfs_pnfs.h"
21 
22 /*
23  * Ensure that we do not have any outstanding pNFS layouts that can be used by
24  * clients to directly read from or write to this inode.  This must be called
25  * before every operation that can remove blocks from the extent map.
26  * Additionally we call it during the write operation, where aren't concerned
27  * about exposing unallocated blocks but just want to provide basic
28  * synchronization between a local writer and pNFS clients.  mmap writes would
29  * also benefit from this sort of synchronization, but due to the tricky locking
30  * rules in the page fault path we don't bother.
31  */
32 int
33 xfs_break_layouts(
34 	struct inode		*inode,
35 	uint			*iolock,
36 	bool			with_imutex)
37 {
38 	struct xfs_inode	*ip = XFS_I(inode);
39 	int			error;
40 
41 	ASSERT(xfs_isilocked(ip, XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
42 
43 	while ((error = break_layout(inode, false) == -EWOULDBLOCK)) {
44 		xfs_iunlock(ip, *iolock);
45 		if (with_imutex && (*iolock & XFS_IOLOCK_EXCL))
46 			inode_unlock(inode);
47 		error = break_layout(inode, true);
48 		*iolock = XFS_IOLOCK_EXCL;
49 		if (with_imutex)
50 			inode_lock(inode);
51 		xfs_ilock(ip, *iolock);
52 	}
53 
54 	return error;
55 }
56 
57 /*
58  * Get a unique ID including its location so that the client can identify
59  * the exported device.
60  */
61 int
62 xfs_fs_get_uuid(
63 	struct super_block	*sb,
64 	u8			*buf,
65 	u32			*len,
66 	u64			*offset)
67 {
68 	struct xfs_mount	*mp = XFS_M(sb);
69 
70 	printk_once(KERN_NOTICE
71 "XFS (%s): using experimental pNFS feature, use at your own risk!\n",
72 		mp->m_fsname);
73 
74 	if (*len < sizeof(uuid_t))
75 		return -EINVAL;
76 
77 	memcpy(buf, &mp->m_sb.sb_uuid, sizeof(uuid_t));
78 	*len = sizeof(uuid_t);
79 	*offset = offsetof(struct xfs_dsb, sb_uuid);
80 	return 0;
81 }
82 
83 /*
84  * Get a layout for the pNFS client.
85  */
86 int
87 xfs_fs_map_blocks(
88 	struct inode		*inode,
89 	loff_t			offset,
90 	u64			length,
91 	struct iomap		*iomap,
92 	bool			write,
93 	u32			*device_generation)
94 {
95 	struct xfs_inode	*ip = XFS_I(inode);
96 	struct xfs_mount	*mp = ip->i_mount;
97 	struct xfs_bmbt_irec	imap;
98 	xfs_fileoff_t		offset_fsb, end_fsb;
99 	loff_t			limit;
100 	int			bmapi_flags = XFS_BMAPI_ENTIRE;
101 	int			nimaps = 1;
102 	uint			lock_flags;
103 	int			error = 0;
104 
105 	if (XFS_FORCED_SHUTDOWN(mp))
106 		return -EIO;
107 
108 	/*
109 	 * We can't export inodes residing on the realtime device.  The realtime
110 	 * device doesn't have a UUID to identify it, so the client has no way
111 	 * to find it.
112 	 */
113 	if (XFS_IS_REALTIME_INODE(ip))
114 		return -ENXIO;
115 
116 	/*
117 	 * The pNFS block layout spec actually supports reflink like
118 	 * functionality, but the Linux pNFS server doesn't implement it yet.
119 	 */
120 	if (xfs_is_reflink_inode(ip))
121 		return -ENXIO;
122 
123 	/*
124 	 * Lock out any other I/O before we flush and invalidate the pagecache,
125 	 * and then hand out a layout to the remote system.  This is very
126 	 * similar to direct I/O, except that the synchronization is much more
127 	 * complicated.  See the comment near xfs_break_layouts for a detailed
128 	 * explanation.
129 	 */
130 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
131 
132 	error = -EINVAL;
133 	limit = mp->m_super->s_maxbytes;
134 	if (!write)
135 		limit = max(limit, round_up(i_size_read(inode),
136 				     inode->i_sb->s_blocksize));
137 	if (offset > limit)
138 		goto out_unlock;
139 	if (offset > limit - length)
140 		length = limit - offset;
141 
142 	error = filemap_write_and_wait(inode->i_mapping);
143 	if (error)
144 		goto out_unlock;
145 	error = invalidate_inode_pages2(inode->i_mapping);
146 	if (WARN_ON_ONCE(error))
147 		return error;
148 
149 	end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + length);
150 	offset_fsb = XFS_B_TO_FSBT(mp, offset);
151 
152 	lock_flags = xfs_ilock_data_map_shared(ip);
153 	error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
154 				&imap, &nimaps, bmapi_flags);
155 	xfs_iunlock(ip, lock_flags);
156 
157 	if (error)
158 		goto out_unlock;
159 
160 	if (write) {
161 		enum xfs_prealloc_flags	flags = 0;
162 
163 		ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
164 
165 		if (!nimaps || imap.br_startblock == HOLESTARTBLOCK) {
166 			/*
167 			 * xfs_iomap_write_direct() expects to take ownership of
168 			 * the shared ilock.
169 			 */
170 			xfs_ilock(ip, XFS_ILOCK_SHARED);
171 			error = xfs_iomap_write_direct(ip, offset, length,
172 						       &imap, nimaps);
173 			if (error)
174 				goto out_unlock;
175 
176 			/*
177 			 * Ensure the next transaction is committed
178 			 * synchronously so that the blocks allocated and
179 			 * handed out to the client are guaranteed to be
180 			 * present even after a server crash.
181 			 */
182 			flags |= XFS_PREALLOC_SET | XFS_PREALLOC_SYNC;
183 		}
184 
185 		error = xfs_update_prealloc_flags(ip, flags);
186 		if (error)
187 			goto out_unlock;
188 	}
189 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
190 
191 	xfs_bmbt_to_iomap(ip, iomap, &imap);
192 	*device_generation = mp->m_generation;
193 	return error;
194 out_unlock:
195 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
196 	return error;
197 }
198 
199 /*
200  * Ensure the size update falls into a valid allocated block.
201  */
202 static int
203 xfs_pnfs_validate_isize(
204 	struct xfs_inode	*ip,
205 	xfs_off_t		isize)
206 {
207 	struct xfs_bmbt_irec	imap;
208 	int			nimaps = 1;
209 	int			error = 0;
210 
211 	xfs_ilock(ip, XFS_ILOCK_SHARED);
212 	error = xfs_bmapi_read(ip, XFS_B_TO_FSBT(ip->i_mount, isize - 1), 1,
213 				&imap, &nimaps, 0);
214 	xfs_iunlock(ip, XFS_ILOCK_SHARED);
215 	if (error)
216 		return error;
217 
218 	if (imap.br_startblock == HOLESTARTBLOCK ||
219 	    imap.br_startblock == DELAYSTARTBLOCK ||
220 	    imap.br_state == XFS_EXT_UNWRITTEN)
221 		return -EIO;
222 	return 0;
223 }
224 
225 /*
226  * Make sure the blocks described by maps are stable on disk.  This includes
227  * converting any unwritten extents, flushing the disk cache and updating the
228  * time stamps.
229  *
230  * Note that we rely on the caller to always send us a timestamp update so that
231  * we always commit a transaction here.  If that stops being true we will have
232  * to manually flush the cache here similar to what the fsync code path does
233  * for datasyncs on files that have no dirty metadata.
234  */
235 int
236 xfs_fs_commit_blocks(
237 	struct inode		*inode,
238 	struct iomap		*maps,
239 	int			nr_maps,
240 	struct iattr		*iattr)
241 {
242 	struct xfs_inode	*ip = XFS_I(inode);
243 	struct xfs_mount	*mp = ip->i_mount;
244 	struct xfs_trans	*tp;
245 	bool			update_isize = false;
246 	int			error, i;
247 	loff_t			size;
248 
249 	ASSERT(iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME));
250 
251 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
252 
253 	size = i_size_read(inode);
254 	if ((iattr->ia_valid & ATTR_SIZE) && iattr->ia_size > size) {
255 		update_isize = true;
256 		size = iattr->ia_size;
257 	}
258 
259 	for (i = 0; i < nr_maps; i++) {
260 		u64 start, length, end;
261 
262 		start = maps[i].offset;
263 		if (start > size)
264 			continue;
265 
266 		end = start + maps[i].length;
267 		if (end > size)
268 			end = size;
269 
270 		length = end - start;
271 		if (!length)
272 			continue;
273 
274 		/*
275 		 * Make sure reads through the pagecache see the new data.
276 		 */
277 		error = invalidate_inode_pages2_range(inode->i_mapping,
278 					start >> PAGE_SHIFT,
279 					(end - 1) >> PAGE_SHIFT);
280 		WARN_ON_ONCE(error);
281 
282 		error = xfs_iomap_write_unwritten(ip, start, length);
283 		if (error)
284 			goto out_drop_iolock;
285 	}
286 
287 	if (update_isize) {
288 		error = xfs_pnfs_validate_isize(ip, size);
289 		if (error)
290 			goto out_drop_iolock;
291 	}
292 
293 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
294 	if (error)
295 		goto out_drop_iolock;
296 
297 	xfs_ilock(ip, XFS_ILOCK_EXCL);
298 	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
299 	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
300 
301 	xfs_setattr_time(ip, iattr);
302 	if (update_isize) {
303 		i_size_write(inode, iattr->ia_size);
304 		ip->i_d.di_size = iattr->ia_size;
305 	}
306 
307 	xfs_trans_set_sync(tp);
308 	error = xfs_trans_commit(tp);
309 
310 out_drop_iolock:
311 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
312 	return error;
313 }
314