Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2014 Christoph Hellwig.
4 */
5#include "xfs.h"
6#include "xfs_shared.h"
7#include "xfs_format.h"
8#include "xfs_log_format.h"
9#include "xfs_trans_resv.h"
10#include "xfs_mount.h"
11#include "xfs_inode.h"
12#include "xfs_trans.h"
13#include "xfs_bmap.h"
14#include "xfs_iomap.h"
15#include "xfs_pnfs.h"
16
17/*
18 * Ensure that we do not have any outstanding pNFS layouts that can be used by
19 * clients to directly read from or write to this inode. This must be called
20 * before every operation that can remove blocks from the extent map.
21 * Additionally we call it during the write operation, where aren't concerned
22 * about exposing unallocated blocks but just want to provide basic
23 * synchronization between a local writer and pNFS clients. mmap writes would
24 * also benefit from this sort of synchronization, but due to the tricky locking
25 * rules in the page fault path we don't bother.
26 */
27int
28xfs_break_leased_layouts(
29 struct inode *inode,
30 uint *iolock,
31 bool *did_unlock)
32{
33 struct xfs_inode *ip = XFS_I(inode);
34 int error;
35
36 while ((error = break_layout(inode, false)) == -EWOULDBLOCK) {
37 xfs_iunlock(ip, *iolock);
38 *did_unlock = true;
39 error = break_layout(inode, true);
40 *iolock &= ~XFS_IOLOCK_SHARED;
41 *iolock |= XFS_IOLOCK_EXCL;
42 xfs_ilock(ip, *iolock);
43 }
44
45 return error;
46}
47
48/*
49 * Get a unique ID including its location so that the client can identify
50 * the exported device.
51 */
52int
53xfs_fs_get_uuid(
54 struct super_block *sb,
55 u8 *buf,
56 u32 *len,
57 u64 *offset)
58{
59 struct xfs_mount *mp = XFS_M(sb);
60
61 xfs_notice_once(mp,
62"Using experimental pNFS feature, use at your own risk!");
63
64 if (*len < sizeof(uuid_t))
65 return -EINVAL;
66
67 memcpy(buf, &mp->m_sb.sb_uuid, sizeof(uuid_t));
68 *len = sizeof(uuid_t);
69 *offset = offsetof(struct xfs_dsb, sb_uuid);
70 return 0;
71}
72
73/*
74 * We cannot use file based VFS helpers such as file_modified() to update
75 * inode state as we modify the data/metadata in the inode here. Hence we have
76 * to open code the timestamp updates and SUID/SGID stripping. We also need
77 * to set the inode prealloc flag to ensure that the extents we allocate are not
78 * removed if the inode is reclaimed from memory before xfs_fs_block_commit()
79 * is from the client to indicate that data has been written and the file size
80 * can be extended.
81 */
82static int
83xfs_fs_map_update_inode(
84 struct xfs_inode *ip)
85{
86 struct xfs_trans *tp;
87 int error;
88
89 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_writeid,
90 0, 0, 0, &tp);
91 if (error)
92 return error;
93
94 xfs_ilock(ip, XFS_ILOCK_EXCL);
95 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
96
97 VFS_I(ip)->i_mode &= ~S_ISUID;
98 if (VFS_I(ip)->i_mode & S_IXGRP)
99 VFS_I(ip)->i_mode &= ~S_ISGID;
100 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
101 ip->i_diflags |= XFS_DIFLAG_PREALLOC;
102
103 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
104 return xfs_trans_commit(tp);
105}
106
107/*
108 * Get a layout for the pNFS client.
109 */
110int
111xfs_fs_map_blocks(
112 struct inode *inode,
113 loff_t offset,
114 u64 length,
115 struct iomap *iomap,
116 bool write,
117 u32 *device_generation)
118{
119 struct xfs_inode *ip = XFS_I(inode);
120 struct xfs_mount *mp = ip->i_mount;
121 struct xfs_bmbt_irec imap;
122 xfs_fileoff_t offset_fsb, end_fsb;
123 loff_t limit;
124 int bmapi_flags = XFS_BMAPI_ENTIRE;
125 int nimaps = 1;
126 uint lock_flags;
127 int error = 0;
128 u64 seq;
129
130 if (xfs_is_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 * The pNFS block layout spec actually supports reflink like
143 * functionality, but the Linux pNFS server doesn't implement it yet.
144 */
145 if (xfs_is_reflink_inode(ip))
146 return -ENXIO;
147
148 /*
149 * Lock out any other I/O before we flush and invalidate the pagecache,
150 * and then hand out a layout to the remote system. This is very
151 * similar to direct I/O, except that the synchronization is much more
152 * complicated. See the comment near xfs_break_leased_layouts
153 * for a detailed explanation.
154 */
155 xfs_ilock(ip, XFS_IOLOCK_EXCL);
156
157 error = -EINVAL;
158 limit = mp->m_super->s_maxbytes;
159 if (!write)
160 limit = max(limit, round_up(i_size_read(inode),
161 inode->i_sb->s_blocksize));
162 if (offset > limit)
163 goto out_unlock;
164 if (offset > limit - length)
165 length = limit - offset;
166
167 error = filemap_write_and_wait(inode->i_mapping);
168 if (error)
169 goto out_unlock;
170 error = invalidate_inode_pages2(inode->i_mapping);
171 if (WARN_ON_ONCE(error))
172 goto out_unlock;
173
174 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + length);
175 offset_fsb = XFS_B_TO_FSBT(mp, offset);
176
177 lock_flags = xfs_ilock_data_map_shared(ip);
178 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
179 &imap, &nimaps, bmapi_flags);
180 seq = xfs_iomap_inode_sequence(ip, 0);
181
182 ASSERT(!nimaps || imap.br_startblock != DELAYSTARTBLOCK);
183
184 if (!error && write &&
185 (!nimaps || imap.br_startblock == HOLESTARTBLOCK)) {
186 if (offset + length > XFS_ISIZE(ip))
187 end_fsb = xfs_iomap_eof_align_last_fsb(ip, end_fsb);
188 else if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
189 end_fsb = min(end_fsb, imap.br_startoff +
190 imap.br_blockcount);
191 xfs_iunlock(ip, lock_flags);
192
193 error = xfs_iomap_write_direct(ip, offset_fsb,
194 end_fsb - offset_fsb, 0, &imap, &seq);
195 if (error)
196 goto out_unlock;
197
198 /*
199 * Ensure the next transaction is committed synchronously so
200 * that the blocks allocated and handed out to the client are
201 * guaranteed to be present even after a server crash.
202 */
203 error = xfs_fs_map_update_inode(ip);
204 if (!error)
205 error = xfs_log_force_inode(ip);
206 if (error)
207 goto out_unlock;
208
209 } else {
210 xfs_iunlock(ip, lock_flags);
211 }
212 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
213
214 error = xfs_bmbt_to_iomap(ip, iomap, &imap, 0, 0, seq);
215 *device_generation = mp->m_generation;
216 return error;
217out_unlock:
218 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
219 return error;
220}
221
222/*
223 * Ensure the size update falls into a valid allocated block.
224 */
225static int
226xfs_pnfs_validate_isize(
227 struct xfs_inode *ip,
228 xfs_off_t isize)
229{
230 struct xfs_bmbt_irec imap;
231 int nimaps = 1;
232 int error = 0;
233
234 xfs_ilock(ip, XFS_ILOCK_SHARED);
235 error = xfs_bmapi_read(ip, XFS_B_TO_FSBT(ip->i_mount, isize - 1), 1,
236 &imap, &nimaps, 0);
237 xfs_iunlock(ip, XFS_ILOCK_SHARED);
238 if (error)
239 return error;
240
241 if (imap.br_startblock == HOLESTARTBLOCK ||
242 imap.br_startblock == DELAYSTARTBLOCK ||
243 imap.br_state == XFS_EXT_UNWRITTEN)
244 return -EIO;
245 return 0;
246}
247
248/*
249 * Make sure the blocks described by maps are stable on disk. This includes
250 * converting any unwritten extents, flushing the disk cache and updating the
251 * time stamps.
252 *
253 * Note that we rely on the caller to always send us a timestamp update so that
254 * we always commit a transaction here. If that stops being true we will have
255 * to manually flush the cache here similar to what the fsync code path does
256 * for datasyncs on files that have no dirty metadata.
257 */
258int
259xfs_fs_commit_blocks(
260 struct inode *inode,
261 struct iomap *maps,
262 int nr_maps,
263 struct iattr *iattr)
264{
265 struct xfs_inode *ip = XFS_I(inode);
266 struct xfs_mount *mp = ip->i_mount;
267 struct xfs_trans *tp;
268 bool update_isize = false;
269 int error, i;
270 loff_t size;
271
272 ASSERT(iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME));
273
274 xfs_ilock(ip, XFS_IOLOCK_EXCL);
275
276 size = i_size_read(inode);
277 if ((iattr->ia_valid & ATTR_SIZE) && iattr->ia_size > size) {
278 update_isize = true;
279 size = iattr->ia_size;
280 }
281
282 for (i = 0; i < nr_maps; i++) {
283 u64 start, length, end;
284
285 start = maps[i].offset;
286 if (start > size)
287 continue;
288
289 end = start + maps[i].length;
290 if (end > size)
291 end = size;
292
293 length = end - start;
294 if (!length)
295 continue;
296
297 /*
298 * Make sure reads through the pagecache see the new data.
299 */
300 error = invalidate_inode_pages2_range(inode->i_mapping,
301 start >> PAGE_SHIFT,
302 (end - 1) >> PAGE_SHIFT);
303 WARN_ON_ONCE(error);
304
305 error = xfs_iomap_write_unwritten(ip, start, length, false);
306 if (error)
307 goto out_drop_iolock;
308 }
309
310 if (update_isize) {
311 error = xfs_pnfs_validate_isize(ip, size);
312 if (error)
313 goto out_drop_iolock;
314 }
315
316 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
317 if (error)
318 goto out_drop_iolock;
319
320 xfs_ilock(ip, XFS_ILOCK_EXCL);
321 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
322 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
323
324 ASSERT(!(iattr->ia_valid & (ATTR_UID | ATTR_GID)));
325 setattr_copy(&init_user_ns, inode, iattr);
326 if (update_isize) {
327 i_size_write(inode, iattr->ia_size);
328 ip->i_disk_size = iattr->ia_size;
329 }
330
331 xfs_trans_set_sync(tp);
332 error = xfs_trans_commit(tp);
333
334out_drop_iolock:
335 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
336 return error;
337}
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 */
32int
33xfs_break_layouts(
34 struct inode *inode,
35 uint *iolock)
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 error = break_layout(inode, true);
45 *iolock = XFS_IOLOCK_EXCL;
46 xfs_ilock(ip, *iolock);
47 }
48
49 return error;
50}
51
52/*
53 * Get a unique ID including its location so that the client can identify
54 * the exported device.
55 */
56int
57xfs_fs_get_uuid(
58 struct super_block *sb,
59 u8 *buf,
60 u32 *len,
61 u64 *offset)
62{
63 struct xfs_mount *mp = XFS_M(sb);
64
65 printk_once(KERN_NOTICE
66"XFS (%s): using experimental pNFS feature, use at your own risk!\n",
67 mp->m_fsname);
68
69 if (*len < sizeof(uuid_t))
70 return -EINVAL;
71
72 memcpy(buf, &mp->m_sb.sb_uuid, sizeof(uuid_t));
73 *len = sizeof(uuid_t);
74 *offset = offsetof(struct xfs_dsb, sb_uuid);
75 return 0;
76}
77
78/*
79 * Get a layout for the pNFS client.
80 */
81int
82xfs_fs_map_blocks(
83 struct inode *inode,
84 loff_t offset,
85 u64 length,
86 struct iomap *iomap,
87 bool write,
88 u32 *device_generation)
89{
90 struct xfs_inode *ip = XFS_I(inode);
91 struct xfs_mount *mp = ip->i_mount;
92 struct xfs_bmbt_irec imap;
93 xfs_fileoff_t offset_fsb, end_fsb;
94 loff_t limit;
95 int bmapi_flags = XFS_BMAPI_ENTIRE;
96 int nimaps = 1;
97 uint lock_flags;
98 int error = 0;
99
100 if (XFS_FORCED_SHUTDOWN(mp))
101 return -EIO;
102
103 /*
104 * We can't export inodes residing on the realtime device. The realtime
105 * device doesn't have a UUID to identify it, so the client has no way
106 * to find it.
107 */
108 if (XFS_IS_REALTIME_INODE(ip))
109 return -ENXIO;
110
111 /*
112 * The pNFS block layout spec actually supports reflink like
113 * functionality, but the Linux pNFS server doesn't implement it yet.
114 */
115 if (xfs_is_reflink_inode(ip))
116 return -ENXIO;
117
118 /*
119 * Lock out any other I/O before we flush and invalidate the pagecache,
120 * and then hand out a layout to the remote system. This is very
121 * similar to direct I/O, except that the synchronization is much more
122 * complicated. See the comment near xfs_break_layouts for a detailed
123 * explanation.
124 */
125 xfs_ilock(ip, XFS_IOLOCK_EXCL);
126
127 error = -EINVAL;
128 limit = mp->m_super->s_maxbytes;
129 if (!write)
130 limit = max(limit, round_up(i_size_read(inode),
131 inode->i_sb->s_blocksize));
132 if (offset > limit)
133 goto out_unlock;
134 if (offset > limit - length)
135 length = limit - offset;
136
137 error = filemap_write_and_wait(inode->i_mapping);
138 if (error)
139 goto out_unlock;
140 error = invalidate_inode_pages2(inode->i_mapping);
141 if (WARN_ON_ONCE(error))
142 return error;
143
144 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + length);
145 offset_fsb = XFS_B_TO_FSBT(mp, offset);
146
147 lock_flags = xfs_ilock_data_map_shared(ip);
148 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
149 &imap, &nimaps, bmapi_flags);
150 xfs_iunlock(ip, lock_flags);
151
152 if (error)
153 goto out_unlock;
154
155 if (write) {
156 enum xfs_prealloc_flags flags = 0;
157
158 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
159
160 if (!nimaps || imap.br_startblock == HOLESTARTBLOCK) {
161 /*
162 * xfs_iomap_write_direct() expects to take ownership of
163 * the shared ilock.
164 */
165 xfs_ilock(ip, XFS_ILOCK_SHARED);
166 error = xfs_iomap_write_direct(ip, offset, length,
167 &imap, nimaps);
168 if (error)
169 goto out_unlock;
170
171 /*
172 * Ensure the next transaction is committed
173 * synchronously so that the blocks allocated and
174 * handed out to the client are guaranteed to be
175 * present even after a server crash.
176 */
177 flags |= XFS_PREALLOC_SET | XFS_PREALLOC_SYNC;
178 }
179
180 error = xfs_update_prealloc_flags(ip, flags);
181 if (error)
182 goto out_unlock;
183 }
184 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
185
186 xfs_bmbt_to_iomap(ip, iomap, &imap);
187 *device_generation = mp->m_generation;
188 return error;
189out_unlock:
190 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
191 return error;
192}
193
194/*
195 * Ensure the size update falls into a valid allocated block.
196 */
197static int
198xfs_pnfs_validate_isize(
199 struct xfs_inode *ip,
200 xfs_off_t isize)
201{
202 struct xfs_bmbt_irec imap;
203 int nimaps = 1;
204 int error = 0;
205
206 xfs_ilock(ip, XFS_ILOCK_SHARED);
207 error = xfs_bmapi_read(ip, XFS_B_TO_FSBT(ip->i_mount, isize - 1), 1,
208 &imap, &nimaps, 0);
209 xfs_iunlock(ip, XFS_ILOCK_SHARED);
210 if (error)
211 return error;
212
213 if (imap.br_startblock == HOLESTARTBLOCK ||
214 imap.br_startblock == DELAYSTARTBLOCK ||
215 imap.br_state == XFS_EXT_UNWRITTEN)
216 return -EIO;
217 return 0;
218}
219
220/*
221 * Make sure the blocks described by maps are stable on disk. This includes
222 * converting any unwritten extents, flushing the disk cache and updating the
223 * time stamps.
224 *
225 * Note that we rely on the caller to always send us a timestamp update so that
226 * we always commit a transaction here. If that stops being true we will have
227 * to manually flush the cache here similar to what the fsync code path does
228 * for datasyncs on files that have no dirty metadata.
229 */
230int
231xfs_fs_commit_blocks(
232 struct inode *inode,
233 struct iomap *maps,
234 int nr_maps,
235 struct iattr *iattr)
236{
237 struct xfs_inode *ip = XFS_I(inode);
238 struct xfs_mount *mp = ip->i_mount;
239 struct xfs_trans *tp;
240 bool update_isize = false;
241 int error, i;
242 loff_t size;
243
244 ASSERT(iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME));
245
246 xfs_ilock(ip, XFS_IOLOCK_EXCL);
247
248 size = i_size_read(inode);
249 if ((iattr->ia_valid & ATTR_SIZE) && iattr->ia_size > size) {
250 update_isize = true;
251 size = iattr->ia_size;
252 }
253
254 for (i = 0; i < nr_maps; i++) {
255 u64 start, length, end;
256
257 start = maps[i].offset;
258 if (start > size)
259 continue;
260
261 end = start + maps[i].length;
262 if (end > size)
263 end = size;
264
265 length = end - start;
266 if (!length)
267 continue;
268
269 /*
270 * Make sure reads through the pagecache see the new data.
271 */
272 error = invalidate_inode_pages2_range(inode->i_mapping,
273 start >> PAGE_SHIFT,
274 (end - 1) >> PAGE_SHIFT);
275 WARN_ON_ONCE(error);
276
277 error = xfs_iomap_write_unwritten(ip, start, length);
278 if (error)
279 goto out_drop_iolock;
280 }
281
282 if (update_isize) {
283 error = xfs_pnfs_validate_isize(ip, size);
284 if (error)
285 goto out_drop_iolock;
286 }
287
288 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
289 if (error)
290 goto out_drop_iolock;
291
292 xfs_ilock(ip, XFS_ILOCK_EXCL);
293 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
294 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
295
296 xfs_setattr_time(ip, iattr);
297 if (update_isize) {
298 i_size_write(inode, iattr->ia_size);
299 ip->i_d.di_size = iattr->ia_size;
300 }
301
302 xfs_trans_set_sync(tp);
303 error = xfs_trans_commit(tp);
304
305out_drop_iolock:
306 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
307 return error;
308}