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