xref: /openbmc/linux/fs/xfs/xfs_pnfs.c (revision b9b77222)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2014 Christoph Hellwig.
4  */
5 #include <linux/iomap.h>
6 #include "xfs.h"
7 #include "xfs_format.h"
8 #include "xfs_log_format.h"
9 #include "xfs_trans_resv.h"
10 #include "xfs_sb.h"
11 #include "xfs_mount.h"
12 #include "xfs_inode.h"
13 #include "xfs_trans.h"
14 #include "xfs_log.h"
15 #include "xfs_bmap.h"
16 #include "xfs_bmap_util.h"
17 #include "xfs_error.h"
18 #include "xfs_iomap.h"
19 #include "xfs_shared.h"
20 #include "xfs_bit.h"
21 #include "xfs_pnfs.h"
22 
23 /*
24  * Ensure that we do not have any outstanding pNFS layouts that can be used by
25  * clients to directly read from or write to this inode.  This must be called
26  * before every operation that can remove blocks from the extent map.
27  * Additionally we call it during the write operation, where aren't concerned
28  * about exposing unallocated blocks but just want to provide basic
29  * synchronization between a local writer and pNFS clients.  mmap writes would
30  * also benefit from this sort of synchronization, but due to the tricky locking
31  * rules in the page fault path we don't bother.
32  */
33 int
34 xfs_break_leased_layouts(
35 	struct inode		*inode,
36 	uint			*iolock,
37 	bool			*did_unlock)
38 {
39 	struct xfs_inode	*ip = XFS_I(inode);
40 	int			error;
41 
42 	while ((error = break_layout(inode, false) == -EWOULDBLOCK)) {
43 		xfs_iunlock(ip, *iolock);
44 		*did_unlock = true;
45 		error = break_layout(inode, true);
46 		*iolock &= ~XFS_IOLOCK_SHARED;
47 		*iolock |= XFS_IOLOCK_EXCL;
48 		xfs_ilock(ip, *iolock);
49 	}
50 
51 	return error;
52 }
53 
54 /*
55  * Get a unique ID including its location so that the client can identify
56  * the exported device.
57  */
58 int
59 xfs_fs_get_uuid(
60 	struct super_block	*sb,
61 	u8			*buf,
62 	u32			*len,
63 	u64			*offset)
64 {
65 	struct xfs_mount	*mp = XFS_M(sb);
66 
67 	printk_once(KERN_NOTICE
68 "XFS (%s): using experimental pNFS feature, use at your own risk!\n",
69 		mp->m_fsname);
70 
71 	if (*len < sizeof(uuid_t))
72 		return -EINVAL;
73 
74 	memcpy(buf, &mp->m_sb.sb_uuid, sizeof(uuid_t));
75 	*len = sizeof(uuid_t);
76 	*offset = offsetof(struct xfs_dsb, sb_uuid);
77 	return 0;
78 }
79 
80 /*
81  * Get a layout for the pNFS client.
82  */
83 int
84 xfs_fs_map_blocks(
85 	struct inode		*inode,
86 	loff_t			offset,
87 	u64			length,
88 	struct iomap		*iomap,
89 	bool			write,
90 	u32			*device_generation)
91 {
92 	struct xfs_inode	*ip = XFS_I(inode);
93 	struct xfs_mount	*mp = ip->i_mount;
94 	struct xfs_bmbt_irec	imap;
95 	xfs_fileoff_t		offset_fsb, end_fsb;
96 	loff_t			limit;
97 	int			bmapi_flags = XFS_BMAPI_ENTIRE;
98 	int			nimaps = 1;
99 	uint			lock_flags;
100 	int			error = 0;
101 
102 	if (XFS_FORCED_SHUTDOWN(mp))
103 		return -EIO;
104 
105 	/*
106 	 * We can't export inodes residing on the realtime device.  The realtime
107 	 * device doesn't have a UUID to identify it, so the client has no way
108 	 * to find it.
109 	 */
110 	if (XFS_IS_REALTIME_INODE(ip))
111 		return -ENXIO;
112 
113 	/*
114 	 * The pNFS block layout spec actually supports reflink like
115 	 * functionality, but the Linux pNFS server doesn't implement it yet.
116 	 */
117 	if (xfs_is_reflink_inode(ip))
118 		return -ENXIO;
119 
120 	/*
121 	 * Lock out any other I/O before we flush and invalidate the pagecache,
122 	 * and then hand out a layout to the remote system.  This is very
123 	 * similar to direct I/O, except that the synchronization is much more
124 	 * complicated.  See the comment near xfs_break_leased_layouts
125 	 * for a detailed explanation.
126 	 */
127 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
128 
129 	error = -EINVAL;
130 	limit = mp->m_super->s_maxbytes;
131 	if (!write)
132 		limit = max(limit, round_up(i_size_read(inode),
133 				     inode->i_sb->s_blocksize));
134 	if (offset > limit)
135 		goto out_unlock;
136 	if (offset > limit - length)
137 		length = limit - offset;
138 
139 	error = filemap_write_and_wait(inode->i_mapping);
140 	if (error)
141 		goto out_unlock;
142 	error = invalidate_inode_pages2(inode->i_mapping);
143 	if (WARN_ON_ONCE(error))
144 		return error;
145 
146 	end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + length);
147 	offset_fsb = XFS_B_TO_FSBT(mp, offset);
148 
149 	lock_flags = xfs_ilock_data_map_shared(ip);
150 	error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
151 				&imap, &nimaps, bmapi_flags);
152 	xfs_iunlock(ip, lock_flags);
153 
154 	if (error)
155 		goto out_unlock;
156 
157 	if (write) {
158 		enum xfs_prealloc_flags	flags = 0;
159 
160 		ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
161 
162 		if (!nimaps || imap.br_startblock == HOLESTARTBLOCK) {
163 			/*
164 			 * xfs_iomap_write_direct() expects to take ownership of
165 			 * the shared ilock.
166 			 */
167 			xfs_ilock(ip, XFS_ILOCK_SHARED);
168 			error = xfs_iomap_write_direct(ip, offset, length,
169 						       &imap, nimaps);
170 			if (error)
171 				goto out_unlock;
172 
173 			/*
174 			 * Ensure the next transaction is committed
175 			 * synchronously so that the blocks allocated and
176 			 * handed out to the client are guaranteed to be
177 			 * present even after a server crash.
178 			 */
179 			flags |= XFS_PREALLOC_SET | XFS_PREALLOC_SYNC;
180 		}
181 
182 		error = xfs_update_prealloc_flags(ip, flags);
183 		if (error)
184 			goto out_unlock;
185 	}
186 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
187 
188 	xfs_bmbt_to_iomap(ip, iomap, &imap);
189 	*device_generation = mp->m_generation;
190 	return error;
191 out_unlock:
192 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
193 	return error;
194 }
195 
196 /*
197  * Ensure the size update falls into a valid allocated block.
198  */
199 static int
200 xfs_pnfs_validate_isize(
201 	struct xfs_inode	*ip,
202 	xfs_off_t		isize)
203 {
204 	struct xfs_bmbt_irec	imap;
205 	int			nimaps = 1;
206 	int			error = 0;
207 
208 	xfs_ilock(ip, XFS_ILOCK_SHARED);
209 	error = xfs_bmapi_read(ip, XFS_B_TO_FSBT(ip->i_mount, isize - 1), 1,
210 				&imap, &nimaps, 0);
211 	xfs_iunlock(ip, XFS_ILOCK_SHARED);
212 	if (error)
213 		return error;
214 
215 	if (imap.br_startblock == HOLESTARTBLOCK ||
216 	    imap.br_startblock == DELAYSTARTBLOCK ||
217 	    imap.br_state == XFS_EXT_UNWRITTEN)
218 		return -EIO;
219 	return 0;
220 }
221 
222 /*
223  * Make sure the blocks described by maps are stable on disk.  This includes
224  * converting any unwritten extents, flushing the disk cache and updating the
225  * time stamps.
226  *
227  * Note that we rely on the caller to always send us a timestamp update so that
228  * we always commit a transaction here.  If that stops being true we will have
229  * to manually flush the cache here similar to what the fsync code path does
230  * for datasyncs on files that have no dirty metadata.
231  */
232 int
233 xfs_fs_commit_blocks(
234 	struct inode		*inode,
235 	struct iomap		*maps,
236 	int			nr_maps,
237 	struct iattr		*iattr)
238 {
239 	struct xfs_inode	*ip = XFS_I(inode);
240 	struct xfs_mount	*mp = ip->i_mount;
241 	struct xfs_trans	*tp;
242 	bool			update_isize = false;
243 	int			error, i;
244 	loff_t			size;
245 
246 	ASSERT(iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME));
247 
248 	xfs_ilock(ip, XFS_IOLOCK_EXCL);
249 
250 	size = i_size_read(inode);
251 	if ((iattr->ia_valid & ATTR_SIZE) && iattr->ia_size > size) {
252 		update_isize = true;
253 		size = iattr->ia_size;
254 	}
255 
256 	for (i = 0; i < nr_maps; i++) {
257 		u64 start, length, end;
258 
259 		start = maps[i].offset;
260 		if (start > size)
261 			continue;
262 
263 		end = start + maps[i].length;
264 		if (end > size)
265 			end = size;
266 
267 		length = end - start;
268 		if (!length)
269 			continue;
270 
271 		/*
272 		 * Make sure reads through the pagecache see the new data.
273 		 */
274 		error = invalidate_inode_pages2_range(inode->i_mapping,
275 					start >> PAGE_SHIFT,
276 					(end - 1) >> PAGE_SHIFT);
277 		WARN_ON_ONCE(error);
278 
279 		error = xfs_iomap_write_unwritten(ip, start, length, false);
280 		if (error)
281 			goto out_drop_iolock;
282 	}
283 
284 	if (update_isize) {
285 		error = xfs_pnfs_validate_isize(ip, size);
286 		if (error)
287 			goto out_drop_iolock;
288 	}
289 
290 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
291 	if (error)
292 		goto out_drop_iolock;
293 
294 	xfs_ilock(ip, XFS_ILOCK_EXCL);
295 	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
296 	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
297 
298 	xfs_setattr_time(ip, iattr);
299 	if (update_isize) {
300 		i_size_write(inode, iattr->ia_size);
301 		ip->i_d.di_size = iattr->ia_size;
302 	}
303 
304 	xfs_trans_set_sync(tp);
305 	error = xfs_trans_commit(tp);
306 
307 out_drop_iolock:
308 	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
309 	return error;
310 }
311