xref: /openbmc/linux/fs/xfs/xfs_pnfs.c (revision 4bce6fce)
1 /*
2  * Copyright (c) 2014 Christoph Hellwig.
3  */
4 #include "xfs.h"
5 #include "xfs_format.h"
6 #include "xfs_log_format.h"
7 #include "xfs_trans_resv.h"
8 #include "xfs_sb.h"
9 #include "xfs_mount.h"
10 #include "xfs_inode.h"
11 #include "xfs_trans.h"
12 #include "xfs_log.h"
13 #include "xfs_bmap.h"
14 #include "xfs_bmap_util.h"
15 #include "xfs_error.h"
16 #include "xfs_iomap.h"
17 #include "xfs_shared.h"
18 #include "xfs_bit.h"
19 #include "xfs_pnfs.h"
20 
21 /*
22  * Ensure that we do not have any outstanding pNFS layouts that can be used by
23  * clients to directly read from or write to this inode.  This must be called
24  * before every operation that can remove blocks from the extent map.
25  * Additionally we call it during the write operation, where aren't concerned
26  * about exposing unallocated blocks but just want to provide basic
27  * synchronization between a local writer and pNFS clients.  mmap writes would
28  * also benefit from this sort of synchronization, but due to the tricky locking
29  * rules in the page fault path we don't bother.
30  */
31 int
32 xfs_break_layouts(
33 	struct inode		*inode,
34 	uint			*iolock,
35 	bool			with_imutex)
36 {
37 	struct xfs_inode	*ip = XFS_I(inode);
38 	int			error;
39 
40 	ASSERT(xfs_isilocked(ip, XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
41 
42 	while ((error = break_layout(inode, false) == -EWOULDBLOCK)) {
43 		xfs_iunlock(ip, *iolock);
44 		if (with_imutex && (*iolock & XFS_IOLOCK_EXCL))
45 			mutex_unlock(&inode->i_mutex);
46 		error = break_layout(inode, true);
47 		*iolock = XFS_IOLOCK_EXCL;
48 		if (with_imutex)
49 			mutex_lock(&inode->i_mutex);
50 		xfs_ilock(ip, *iolock);
51 	}
52 
53 	return error;
54 }
55 
56 /*
57  * Get a unique ID including its location so that the client can identify
58  * the exported device.
59  */
60 int
61 xfs_fs_get_uuid(
62 	struct super_block	*sb,
63 	u8			*buf,
64 	u32			*len,
65 	u64			*offset)
66 {
67 	struct xfs_mount	*mp = XFS_M(sb);
68 
69 	printk_once(KERN_NOTICE
70 "XFS (%s): using experimental pNFS feature, use at your own risk!\n",
71 		mp->m_fsname);
72 
73 	if (*len < sizeof(uuid_t))
74 		return -EINVAL;
75 
76 	memcpy(buf, &mp->m_sb.sb_uuid, sizeof(uuid_t));
77 	*len = sizeof(uuid_t);
78 	*offset = offsetof(struct xfs_dsb, sb_uuid);
79 	return 0;
80 }
81 
82 static void
83 xfs_bmbt_to_iomap(
84 	struct xfs_inode	*ip,
85 	struct iomap		*iomap,
86 	struct xfs_bmbt_irec	*imap)
87 {
88 	struct xfs_mount	*mp = ip->i_mount;
89 
90 	if (imap->br_startblock == HOLESTARTBLOCK) {
91 		iomap->blkno = IOMAP_NULL_BLOCK;
92 		iomap->type = IOMAP_HOLE;
93 	} else if (imap->br_startblock == DELAYSTARTBLOCK) {
94 		iomap->blkno = IOMAP_NULL_BLOCK;
95 		iomap->type = IOMAP_DELALLOC;
96 	} else {
97 		iomap->blkno =
98 			XFS_FSB_TO_DADDR(ip->i_mount, imap->br_startblock);
99 		if (imap->br_state == XFS_EXT_UNWRITTEN)
100 			iomap->type = IOMAP_UNWRITTEN;
101 		else
102 			iomap->type = IOMAP_MAPPED;
103 	}
104 	iomap->offset = XFS_FSB_TO_B(mp, imap->br_startoff);
105 	iomap->length = XFS_FSB_TO_B(mp, imap->br_blockcount);
106 }
107 
108 /*
109  * Get a layout for the pNFS client.
110  */
111 int
112 xfs_fs_map_blocks(
113 	struct inode		*inode,
114 	loff_t			offset,
115 	u64			length,
116 	struct iomap		*iomap,
117 	bool			write,
118 	u32			*device_generation)
119 {
120 	struct xfs_inode	*ip = XFS_I(inode);
121 	struct xfs_mount	*mp = ip->i_mount;
122 	struct xfs_bmbt_irec	imap;
123 	xfs_fileoff_t		offset_fsb, end_fsb;
124 	loff_t			limit;
125 	int			bmapi_flags = XFS_BMAPI_ENTIRE;
126 	int			nimaps = 1;
127 	uint			lock_flags;
128 	int			error = 0;
129 
130 	if (XFS_FORCED_SHUTDOWN(mp))
131 		return -EIO;
132 
133 	/*
134 	 * We can't export inodes residing on the realtime device.  The realtime
135 	 * device doesn't have a UUID to identify it, so the client has no way
136 	 * to find it.
137 	 */
138 	if (XFS_IS_REALTIME_INODE(ip))
139 		return -ENXIO;
140 
141 	/*
142 	 * Lock out any other I/O before we flush and invalidate the pagecache,
143 	 * and then hand out a layout to the remote system.  This is very
144 	 * similar to direct I/O, except that the synchronization is much more
145 	 * complicated.  See the comment near xfs_break_layouts for a detailed
146 	 * explanation.
147 	 */
148 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
149 
150 	error = -EINVAL;
151 	limit = mp->m_super->s_maxbytes;
152 	if (!write)
153 		limit = max(limit, round_up(i_size_read(inode),
154 				     inode->i_sb->s_blocksize));
155 	if (offset > limit)
156 		goto out_unlock;
157 	if (offset > limit - length)
158 		length = limit - offset;
159 
160 	error = filemap_write_and_wait(inode->i_mapping);
161 	if (error)
162 		goto out_unlock;
163 	error = invalidate_inode_pages2(inode->i_mapping);
164 	if (WARN_ON_ONCE(error))
165 		return error;
166 
167 	end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + length);
168 	offset_fsb = XFS_B_TO_FSBT(mp, offset);
169 
170 	lock_flags = xfs_ilock_data_map_shared(ip);
171 	error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
172 				&imap, &nimaps, bmapi_flags);
173 	xfs_iunlock(ip, lock_flags);
174 
175 	if (error)
176 		goto out_unlock;
177 
178 	if (write) {
179 		enum xfs_prealloc_flags	flags = 0;
180 
181 		ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
182 
183 		if (!nimaps || imap.br_startblock == HOLESTARTBLOCK) {
184 			error = xfs_iomap_write_direct(ip, offset, length,
185 						       &imap, nimaps);
186 			if (error)
187 				goto out_unlock;
188 
189 			/*
190 			 * Ensure the next transaction is committed
191 			 * synchronously so that the blocks allocated and
192 			 * handed out to the client are guaranteed to be
193 			 * present even after a server crash.
194 			 */
195 			flags |= XFS_PREALLOC_SET | XFS_PREALLOC_SYNC;
196 		}
197 
198 		error = xfs_update_prealloc_flags(ip, flags);
199 		if (error)
200 			goto out_unlock;
201 	}
202 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
203 
204 	xfs_bmbt_to_iomap(ip, iomap, &imap);
205 	*device_generation = mp->m_generation;
206 	return error;
207 out_unlock:
208 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
209 	return error;
210 }
211 
212 /*
213  * Ensure the size update falls into a valid allocated block.
214  */
215 static int
216 xfs_pnfs_validate_isize(
217 	struct xfs_inode	*ip,
218 	xfs_off_t		isize)
219 {
220 	struct xfs_bmbt_irec	imap;
221 	int			nimaps = 1;
222 	int			error = 0;
223 
224 	xfs_ilock(ip, XFS_ILOCK_SHARED);
225 	error = xfs_bmapi_read(ip, XFS_B_TO_FSBT(ip->i_mount, isize - 1), 1,
226 				&imap, &nimaps, 0);
227 	xfs_iunlock(ip, XFS_ILOCK_SHARED);
228 	if (error)
229 		return error;
230 
231 	if (imap.br_startblock == HOLESTARTBLOCK ||
232 	    imap.br_startblock == DELAYSTARTBLOCK ||
233 	    imap.br_state == XFS_EXT_UNWRITTEN)
234 		return -EIO;
235 	return 0;
236 }
237 
238 /*
239  * Make sure the blocks described by maps are stable on disk.  This includes
240  * converting any unwritten extents, flushing the disk cache and updating the
241  * time stamps.
242  *
243  * Note that we rely on the caller to always send us a timestamp update so that
244  * we always commit a transaction here.  If that stops being true we will have
245  * to manually flush the cache here similar to what the fsync code path does
246  * for datasyncs on files that have no dirty metadata.
247  */
248 int
249 xfs_fs_commit_blocks(
250 	struct inode		*inode,
251 	struct iomap		*maps,
252 	int			nr_maps,
253 	struct iattr		*iattr)
254 {
255 	struct xfs_inode	*ip = XFS_I(inode);
256 	struct xfs_mount	*mp = ip->i_mount;
257 	struct xfs_trans	*tp;
258 	bool			update_isize = false;
259 	int			error, i;
260 	loff_t			size;
261 
262 	ASSERT(iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME));
263 
264 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
265 
266 	size = i_size_read(inode);
267 	if ((iattr->ia_valid & ATTR_SIZE) && iattr->ia_size > size) {
268 		update_isize = true;
269 		size = iattr->ia_size;
270 	}
271 
272 	for (i = 0; i < nr_maps; i++) {
273 		u64 start, length, end;
274 
275 		start = maps[i].offset;
276 		if (start > size)
277 			continue;
278 
279 		end = start + maps[i].length;
280 		if (end > size)
281 			end = size;
282 
283 		length = end - start;
284 		if (!length)
285 			continue;
286 
287 		/*
288 		 * Make sure reads through the pagecache see the new data.
289 		 */
290 		error = invalidate_inode_pages2_range(inode->i_mapping,
291 					start >> PAGE_CACHE_SHIFT,
292 					(end - 1) >> PAGE_CACHE_SHIFT);
293 		WARN_ON_ONCE(error);
294 
295 		error = xfs_iomap_write_unwritten(ip, start, length);
296 		if (error)
297 			goto out_drop_iolock;
298 	}
299 
300 	if (update_isize) {
301 		error = xfs_pnfs_validate_isize(ip, size);
302 		if (error)
303 			goto out_drop_iolock;
304 	}
305 
306 	tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_NOT_SIZE);
307 	error = xfs_trans_reserve(tp, &M_RES(mp)->tr_ichange, 0, 0);
308 	if (error) {
309 		xfs_trans_cancel(tp, 0);
310 		goto out_drop_iolock;
311 	}
312 
313 	xfs_ilock(ip, XFS_ILOCK_EXCL);
314 	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
315 	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
316 
317 	xfs_setattr_time(ip, iattr);
318 	if (update_isize) {
319 		i_size_write(inode, iattr->ia_size);
320 		ip->i_d.di_size = iattr->ia_size;
321 	}
322 
323 	xfs_trans_set_sync(tp);
324 	error = xfs_trans_commit(tp, 0);
325 
326 out_drop_iolock:
327 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
328 	return error;
329 }
330