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/*
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 */
31int
32xfs_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 inode_unlock(inode);
46 error = break_layout(inode, true);
47 *iolock = XFS_IOLOCK_EXCL;
48 if (with_imutex)
49 inode_lock(inode);
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 */
60int
61xfs_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
82static void
83xfs_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 */
111int
112xfs_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 /*
185 * xfs_iomap_write_direct() expects to take ownership of
186 * the shared ilock.
187 */
188 xfs_ilock(ip, XFS_ILOCK_SHARED);
189 error = xfs_iomap_write_direct(ip, offset, length,
190 &imap, nimaps);
191 if (error)
192 goto out_unlock;
193
194 /*
195 * Ensure the next transaction is committed
196 * synchronously so that the blocks allocated and
197 * handed out to the client are guaranteed to be
198 * present even after a server crash.
199 */
200 flags |= XFS_PREALLOC_SET | XFS_PREALLOC_SYNC;
201 }
202
203 error = xfs_update_prealloc_flags(ip, flags);
204 if (error)
205 goto out_unlock;
206 }
207 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
208
209 xfs_bmbt_to_iomap(ip, iomap, &imap);
210 *device_generation = mp->m_generation;
211 return error;
212out_unlock:
213 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
214 return error;
215}
216
217/*
218 * Ensure the size update falls into a valid allocated block.
219 */
220static int
221xfs_pnfs_validate_isize(
222 struct xfs_inode *ip,
223 xfs_off_t isize)
224{
225 struct xfs_bmbt_irec imap;
226 int nimaps = 1;
227 int error = 0;
228
229 xfs_ilock(ip, XFS_ILOCK_SHARED);
230 error = xfs_bmapi_read(ip, XFS_B_TO_FSBT(ip->i_mount, isize - 1), 1,
231 &imap, &nimaps, 0);
232 xfs_iunlock(ip, XFS_ILOCK_SHARED);
233 if (error)
234 return error;
235
236 if (imap.br_startblock == HOLESTARTBLOCK ||
237 imap.br_startblock == DELAYSTARTBLOCK ||
238 imap.br_state == XFS_EXT_UNWRITTEN)
239 return -EIO;
240 return 0;
241}
242
243/*
244 * Make sure the blocks described by maps are stable on disk. This includes
245 * converting any unwritten extents, flushing the disk cache and updating the
246 * time stamps.
247 *
248 * Note that we rely on the caller to always send us a timestamp update so that
249 * we always commit a transaction here. If that stops being true we will have
250 * to manually flush the cache here similar to what the fsync code path does
251 * for datasyncs on files that have no dirty metadata.
252 */
253int
254xfs_fs_commit_blocks(
255 struct inode *inode,
256 struct iomap *maps,
257 int nr_maps,
258 struct iattr *iattr)
259{
260 struct xfs_inode *ip = XFS_I(inode);
261 struct xfs_mount *mp = ip->i_mount;
262 struct xfs_trans *tp;
263 bool update_isize = false;
264 int error, i;
265 loff_t size;
266
267 ASSERT(iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME));
268
269 xfs_ilock(ip, XFS_IOLOCK_EXCL);
270
271 size = i_size_read(inode);
272 if ((iattr->ia_valid & ATTR_SIZE) && iattr->ia_size > size) {
273 update_isize = true;
274 size = iattr->ia_size;
275 }
276
277 for (i = 0; i < nr_maps; i++) {
278 u64 start, length, end;
279
280 start = maps[i].offset;
281 if (start > size)
282 continue;
283
284 end = start + maps[i].length;
285 if (end > size)
286 end = size;
287
288 length = end - start;
289 if (!length)
290 continue;
291
292 /*
293 * Make sure reads through the pagecache see the new data.
294 */
295 error = invalidate_inode_pages2_range(inode->i_mapping,
296 start >> PAGE_SHIFT,
297 (end - 1) >> PAGE_SHIFT);
298 WARN_ON_ONCE(error);
299
300 error = xfs_iomap_write_unwritten(ip, start, length);
301 if (error)
302 goto out_drop_iolock;
303 }
304
305 if (update_isize) {
306 error = xfs_pnfs_validate_isize(ip, size);
307 if (error)
308 goto out_drop_iolock;
309 }
310
311 tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_NOT_SIZE);
312 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_ichange, 0, 0);
313 if (error) {
314 xfs_trans_cancel(tp);
315 goto out_drop_iolock;
316 }
317
318 xfs_ilock(ip, XFS_ILOCK_EXCL);
319 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
320 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
321
322 xfs_setattr_time(ip, iattr);
323 if (update_isize) {
324 i_size_write(inode, iattr->ia_size);
325 ip->i_d.di_size = iattr->ia_size;
326 }
327
328 xfs_trans_set_sync(tp);
329 error = xfs_trans_commit(tp);
330
331out_drop_iolock:
332 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
333 return error;
334}