author | eschrock |
Fri, 03 Mar 2006 20:08:16 -0800 | |
changeset 1544 | 938876158511 |
parent 1491 | bdcb30e07e7d |
child 1596 | 2e2377ccbf85 |
permissions | -rw-r--r-- |
789 | 1 |
/* |
2 |
* CDDL HEADER START |
|
3 |
* |
|
4 |
* The contents of this file are subject to the terms of the |
|
1491
bdcb30e07e7d
6389368 fat zap should use 16k blocks (with backwards compatability)
ahrens
parents:
873
diff
changeset
|
5 |
* Common Development and Distribution License (the "License"). |
bdcb30e07e7d
6389368 fat zap should use 16k blocks (with backwards compatability)
ahrens
parents:
873
diff
changeset
|
6 |
* You may not use this file except in compliance with the License. |
789 | 7 |
* |
8 |
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE |
|
9 |
* or http://www.opensolaris.org/os/licensing. |
|
10 |
* See the License for the specific language governing permissions |
|
11 |
* and limitations under the License. |
|
12 |
* |
|
13 |
* When distributing Covered Code, include this CDDL HEADER in each |
|
14 |
* file and include the License file at usr/src/OPENSOLARIS.LICENSE. |
|
15 |
* If applicable, add the following below this CDDL HEADER, with the |
|
16 |
* fields enclosed by brackets "[]" replaced with your own identifying |
|
17 |
* information: Portions Copyright [yyyy] [name of copyright owner] |
|
18 |
* |
|
19 |
* CDDL HEADER END |
|
20 |
*/ |
|
21 |
/* |
|
1491
bdcb30e07e7d
6389368 fat zap should use 16k blocks (with backwards compatability)
ahrens
parents:
873
diff
changeset
|
22 |
* Copyright 2006 Sun Microsystems, Inc. All rights reserved. |
789 | 23 |
* Use is subject to license terms. |
24 |
*/ |
|
25 |
||
26 |
#pragma ident "%Z%%M% %I% %E% SMI" |
|
27 |
||
28 |
#include <sys/dmu.h> |
|
29 |
#include <sys/dmu_impl.h> |
|
30 |
#include <sys/dbuf.h> |
|
31 |
#include <sys/dmu_tx.h> |
|
32 |
#include <sys/dmu_objset.h> |
|
33 |
#include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */ |
|
34 |
#include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */ |
|
35 |
#include <sys/dsl_pool.h> |
|
36 |
#include <sys/zap_impl.h> /* for ZAP_BLOCK_SHIFT */ |
|
37 |
#include <sys/spa.h> |
|
38 |
#include <sys/zfs_context.h> |
|
39 |
||
1544 | 40 |
typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn, |
41 |
uint64_t arg1, uint64_t arg2); |
|
42 |
||
789 | 43 |
#ifdef ZFS_DEBUG |
44 |
int dmu_use_tx_debug_bufs = 1; |
|
45 |
#endif |
|
46 |
||
47 |
dmu_tx_t * |
|
48 |
dmu_tx_create_ds(dsl_dir_t *dd) |
|
49 |
{ |
|
50 |
dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP); |
|
51 |
tx->tx_dir = dd; |
|
52 |
if (dd) |
|
53 |
tx->tx_pool = dd->dd_pool; |
|
54 |
list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t), |
|
55 |
offsetof(dmu_tx_hold_t, dth_node)); |
|
56 |
refcount_create(&tx->tx_space_written); |
|
57 |
refcount_create(&tx->tx_space_freed); |
|
58 |
return (tx); |
|
59 |
} |
|
60 |
||
61 |
dmu_tx_t * |
|
62 |
dmu_tx_create(objset_t *os) |
|
63 |
{ |
|
64 |
dmu_tx_t *tx = dmu_tx_create_ds(os->os->os_dsl_dataset->ds_dir); |
|
65 |
tx->tx_objset = os; |
|
1544 | 66 |
tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os->os_dsl_dataset); |
789 | 67 |
return (tx); |
68 |
} |
|
69 |
||
70 |
dmu_tx_t * |
|
71 |
dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg) |
|
72 |
{ |
|
73 |
dmu_tx_t *tx = dmu_tx_create_ds(NULL); |
|
74 |
||
75 |
ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg); |
|
76 |
tx->tx_pool = dp; |
|
77 |
tx->tx_txg = txg; |
|
78 |
tx->tx_anyobj = TRUE; |
|
79 |
||
80 |
return (tx); |
|
81 |
} |
|
82 |
||
83 |
int |
|
84 |
dmu_tx_is_syncing(dmu_tx_t *tx) |
|
85 |
{ |
|
86 |
return (tx->tx_anyobj); |
|
87 |
} |
|
88 |
||
89 |
int |
|
90 |
dmu_tx_private_ok(dmu_tx_t *tx) |
|
91 |
{ |
|
1544 | 92 |
return (tx->tx_anyobj); |
789 | 93 |
} |
94 |
||
95 |
static void |
|
96 |
dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object, |
|
97 |
enum dmu_tx_hold_type type, dmu_tx_hold_func_t func, |
|
98 |
uint64_t arg1, uint64_t arg2) |
|
99 |
{ |
|
100 |
dmu_tx_hold_t *dth; |
|
101 |
dnode_t *dn = NULL; |
|
1544 | 102 |
int err; |
789 | 103 |
|
104 |
if (object != DMU_NEW_OBJECT) { |
|
1544 | 105 |
err = dnode_hold(os->os, object, tx, &dn); |
106 |
if (err) { |
|
107 |
tx->tx_err = err; |
|
108 |
return; |
|
109 |
} |
|
789 | 110 |
|
1544 | 111 |
if (err == 0 && tx->tx_txg != 0) { |
789 | 112 |
mutex_enter(&dn->dn_mtx); |
113 |
/* |
|
114 |
* dn->dn_assigned_txg == tx->tx_txg doesn't pose a |
|
115 |
* problem, but there's no way for it to happen (for |
|
116 |
* now, at least). |
|
117 |
*/ |
|
118 |
ASSERT(dn->dn_assigned_txg == 0); |
|
119 |
ASSERT(dn->dn_assigned_tx == NULL); |
|
120 |
dn->dn_assigned_txg = tx->tx_txg; |
|
121 |
dn->dn_assigned_tx = tx; |
|
122 |
(void) refcount_add(&dn->dn_tx_holds, tx); |
|
123 |
mutex_exit(&dn->dn_mtx); |
|
124 |
} |
|
125 |
} |
|
126 |
||
127 |
dth = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP); |
|
128 |
dth->dth_dnode = dn; |
|
129 |
dth->dth_type = type; |
|
130 |
dth->dth_arg1 = arg1; |
|
131 |
dth->dth_arg2 = arg2; |
|
132 |
list_insert_tail(&tx->tx_holds, dth); |
|
1544 | 133 |
|
134 |
if (func) |
|
135 |
func(tx, dn, arg1, arg2); |
|
789 | 136 |
} |
137 |
||
138 |
void |
|
139 |
dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object) |
|
140 |
{ |
|
141 |
/* |
|
142 |
* If we're syncing, they can manipulate any object anyhow, and |
|
143 |
* the hold on the dnode_t can cause problems. |
|
144 |
*/ |
|
145 |
if (!dmu_tx_is_syncing(tx)) { |
|
146 |
dmu_tx_hold_object_impl(tx, os, object, THT_NEWOBJECT, |
|
147 |
NULL, 0, 0); |
|
148 |
} |
|
149 |
} |
|
150 |
||
1544 | 151 |
static int |
152 |
dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid) |
|
153 |
{ |
|
154 |
int err; |
|
155 |
dmu_buf_impl_t *db; |
|
156 |
||
157 |
rw_enter(&dn->dn_struct_rwlock, RW_READER); |
|
158 |
db = dbuf_hold_level(dn, level, blkid, FTAG); |
|
159 |
rw_exit(&dn->dn_struct_rwlock); |
|
160 |
if (db == NULL) |
|
161 |
return (EIO); |
|
162 |
err = dbuf_read(db, zio, DB_RF_CANFAIL); |
|
163 |
dbuf_rele(db, FTAG); |
|
164 |
return (err); |
|
165 |
} |
|
166 |
||
789 | 167 |
/* ARGSUSED */ |
168 |
static void |
|
169 |
dmu_tx_count_write(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len) |
|
170 |
{ |
|
1544 | 171 |
uint64_t start, end, i, space; |
789 | 172 |
int min_bs, max_bs, min_ibs, max_ibs, epbs, bits; |
173 |
||
174 |
if (len == 0) |
|
175 |
return; |
|
176 |
||
177 |
min_bs = SPA_MINBLOCKSHIFT; |
|
178 |
max_bs = SPA_MAXBLOCKSHIFT; |
|
179 |
min_ibs = DN_MIN_INDBLKSHIFT; |
|
180 |
max_ibs = DN_MAX_INDBLKSHIFT; |
|
181 |
||
182 |
/* |
|
1544 | 183 |
* For i/o error checking, read the first and last level-0 |
184 |
* blocks, and all the level-1 blocks. We needn't do this on |
|
185 |
* the meta-dnode, because we've already read it in. |
|
186 |
*/ |
|
187 |
||
188 |
if (dn && dn->dn_object != DMU_META_DNODE_OBJECT) { |
|
189 |
int err; |
|
190 |
||
191 |
if (dn->dn_maxblkid == 0) { |
|
192 |
err = dmu_tx_check_ioerr(NULL, dn, 0, 0); |
|
193 |
if (err) { |
|
194 |
tx->tx_err = err; |
|
195 |
return; |
|
196 |
} |
|
197 |
} else { |
|
198 |
zio_t *zio = zio_root(tx->tx_pool->dp_spa, |
|
199 |
NULL, NULL, ZIO_FLAG_CANFAIL); |
|
200 |
||
201 |
/* first level-0 block */ |
|
202 |
start = off/dn->dn_datablksz; |
|
203 |
err = dmu_tx_check_ioerr(zio, dn, 0, start); |
|
204 |
if (err) { |
|
205 |
tx->tx_err = err; |
|
206 |
return; |
|
207 |
} |
|
208 |
||
209 |
/* last level-0 block */ |
|
210 |
end = (off+len)/dn->dn_datablksz; |
|
211 |
if (end != start) { |
|
212 |
err = dmu_tx_check_ioerr(zio, dn, 0, end); |
|
213 |
if (err) { |
|
214 |
tx->tx_err = err; |
|
215 |
return; |
|
216 |
} |
|
217 |
} |
|
218 |
||
219 |
/* level-1 blocks */ |
|
220 |
if (dn->dn_nlevels > 1) { |
|
221 |
start >>= dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
|
222 |
end >>= dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
|
223 |
for (i = start+1; i < end; i++) { |
|
224 |
err = dmu_tx_check_ioerr(zio, dn, 1, i); |
|
225 |
if (err) { |
|
226 |
tx->tx_err = err; |
|
227 |
return; |
|
228 |
} |
|
229 |
} |
|
230 |
} |
|
231 |
||
232 |
err = zio_wait(zio); |
|
233 |
if (err) { |
|
234 |
tx->tx_err = err; |
|
235 |
return; |
|
236 |
} |
|
237 |
} |
|
238 |
} |
|
239 |
||
240 |
/* |
|
789 | 241 |
* If there's more than one block, the blocksize can't change, |
242 |
* so we can make a more precise estimate. Alternatively, |
|
243 |
* if the dnode's ibs is larger than max_ibs, always use that. |
|
244 |
* This ensures that if we reduce DN_MAX_INDBLKSHIFT, |
|
245 |
* the code will still work correctly on existing pools. |
|
246 |
*/ |
|
247 |
if (dn && (dn->dn_maxblkid != 0 || dn->dn_indblkshift > max_ibs)) { |
|
248 |
min_ibs = max_ibs = dn->dn_indblkshift; |
|
249 |
if (dn->dn_datablkshift != 0) |
|
250 |
min_bs = max_bs = dn->dn_datablkshift; |
|
251 |
} |
|
252 |
||
253 |
/* |
|
254 |
* 'end' is the last thing we will access, not one past. |
|
255 |
* This way we won't overflow when accessing the last byte. |
|
256 |
*/ |
|
257 |
start = P2ALIGN(off, 1ULL << max_bs); |
|
258 |
end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1; |
|
259 |
space = end - start + 1; |
|
260 |
||
261 |
start >>= min_bs; |
|
262 |
end >>= min_bs; |
|
263 |
||
264 |
epbs = min_ibs - SPA_BLKPTRSHIFT; |
|
265 |
||
266 |
/* |
|
267 |
* The object contains at most 2^(64 - min_bs) blocks, |
|
268 |
* and each indirect level maps 2^epbs. |
|
269 |
*/ |
|
270 |
for (bits = 64 - min_bs; bits >= 0; bits -= epbs) { |
|
271 |
start >>= epbs; |
|
272 |
end >>= epbs; |
|
273 |
/* |
|
274 |
* If we increase the number of levels of indirection, |
|
275 |
* we'll need new blkid=0 indirect blocks. If start == 0, |
|
276 |
* we're already accounting for that blocks; and if end == 0, |
|
277 |
* we can't increase the number of levels beyond that. |
|
278 |
*/ |
|
279 |
if (start != 0 && end != 0) |
|
280 |
space += 1ULL << max_ibs; |
|
281 |
space += (end - start + 1) << max_ibs; |
|
282 |
} |
|
283 |
||
284 |
ASSERT(space < 2 * DMU_MAX_ACCESS); |
|
285 |
||
286 |
tx->tx_space_towrite += space; |
|
287 |
} |
|
288 |
||
289 |
static void |
|
290 |
dmu_tx_count_dnode(dmu_tx_t *tx, dnode_t *dn) |
|
291 |
{ |
|
292 |
dnode_t *mdn = tx->tx_objset->os->os_meta_dnode; |
|
293 |
uint64_t object = dn ? dn->dn_object : DN_MAX_OBJECT - 1; |
|
294 |
uint64_t pre_write_space; |
|
295 |
||
296 |
ASSERT(object < DN_MAX_OBJECT); |
|
297 |
pre_write_space = tx->tx_space_towrite; |
|
298 |
dmu_tx_count_write(tx, mdn, object << DNODE_SHIFT, 1 << DNODE_SHIFT); |
|
299 |
if (dn && dn->dn_dbuf->db_blkptr && |
|
300 |
dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset, |
|
1544 | 301 |
dn->dn_dbuf->db_blkptr->blk_birth)) { |
789 | 302 |
tx->tx_space_tooverwrite += |
303 |
tx->tx_space_towrite - pre_write_space; |
|
304 |
tx->tx_space_towrite = pre_write_space; |
|
305 |
} |
|
306 |
} |
|
307 |
||
308 |
/* ARGSUSED */ |
|
309 |
static void |
|
310 |
dmu_tx_hold_write_impl(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len) |
|
311 |
{ |
|
312 |
dmu_tx_count_write(tx, dn, off, len); |
|
313 |
dmu_tx_count_dnode(tx, dn); |
|
314 |
} |
|
315 |
||
316 |
void |
|
317 |
dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len) |
|
318 |
{ |
|
319 |
ASSERT(tx->tx_txg == 0); |
|
1544 | 320 |
ASSERT(len < DMU_MAX_ACCESS); |
789 | 321 |
ASSERT(UINT64_MAX - off >= len - 1); |
322 |
||
323 |
dmu_tx_hold_object_impl(tx, tx->tx_objset, object, THT_WRITE, |
|
324 |
dmu_tx_hold_write_impl, off, len); |
|
325 |
} |
|
326 |
||
327 |
static void |
|
328 |
dmu_tx_count_free(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len) |
|
329 |
{ |
|
330 |
uint64_t blkid, nblks; |
|
331 |
uint64_t space = 0; |
|
332 |
dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; |
|
333 |
||
334 |
if (dn->dn_datablkshift == 0) |
|
335 |
return; |
|
336 |
/* |
|
337 |
* not that the dnode can change, since it isn't dirty, but |
|
338 |
* dbuf_hold_impl() wants us to have the struct_rwlock. |
|
339 |
* also need it to protect dn_maxblkid. |
|
340 |
*/ |
|
341 |
rw_enter(&dn->dn_struct_rwlock, RW_READER); |
|
342 |
blkid = off >> dn->dn_datablkshift; |
|
343 |
nblks = (off + len) >> dn->dn_datablkshift; |
|
344 |
||
1544 | 345 |
if (blkid >= dn->dn_maxblkid) { |
346 |
rw_exit(&dn->dn_struct_rwlock); |
|
347 |
return; |
|
348 |
} |
|
789 | 349 |
if (blkid + nblks > dn->dn_maxblkid) |
350 |
nblks = dn->dn_maxblkid - blkid; |
|
351 |
||
352 |
/* don't bother after the 100,000 blocks */ |
|
353 |
nblks = MIN(nblks, 128*1024); |
|
354 |
||
355 |
if (dn->dn_phys->dn_nlevels == 1) { |
|
356 |
int i; |
|
357 |
for (i = 0; i < nblks; i++) { |
|
358 |
blkptr_t *bp = dn->dn_phys->dn_blkptr; |
|
359 |
ASSERT3U(blkid + i, <, dn->dn_phys->dn_nblkptr); |
|
360 |
bp += blkid + i; |
|
1544 | 361 |
if (dsl_dataset_block_freeable(ds, bp->blk_birth)) { |
789 | 362 |
dprintf_bp(bp, "can free old%s", ""); |
363 |
space += BP_GET_ASIZE(bp); |
|
364 |
} |
|
365 |
} |
|
1544 | 366 |
nblks = 0; |
789 | 367 |
} |
368 |
||
369 |
while (nblks) { |
|
370 |
dmu_buf_impl_t *dbuf; |
|
371 |
int err, epbs, blkoff, tochk; |
|
372 |
||
373 |
epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
|
374 |
blkoff = P2PHASE(blkid, 1<<epbs); |
|
375 |
tochk = MIN((1<<epbs) - blkoff, nblks); |
|
376 |
||
377 |
err = dbuf_hold_impl(dn, 1, blkid >> epbs, TRUE, FTAG, &dbuf); |
|
378 |
if (err == 0) { |
|
379 |
int i; |
|
380 |
blkptr_t *bp; |
|
381 |
||
1544 | 382 |
err = dbuf_read(dbuf, NULL, |
383 |
DB_RF_HAVESTRUCT | DB_RF_CANFAIL); |
|
384 |
if (err != 0) { |
|
385 |
tx->tx_err = err; |
|
386 |
dbuf_rele(dbuf, FTAG); |
|
387 |
break; |
|
388 |
} |
|
789 | 389 |
|
390 |
bp = dbuf->db.db_data; |
|
391 |
bp += blkoff; |
|
392 |
||
393 |
for (i = 0; i < tochk; i++) { |
|
394 |
if (dsl_dataset_block_freeable(ds, |
|
1544 | 395 |
bp[i].blk_birth)) { |
789 | 396 |
dprintf_bp(&bp[i], |
397 |
"can free old%s", ""); |
|
398 |
space += BP_GET_ASIZE(&bp[i]); |
|
399 |
} |
|
400 |
} |
|
1544 | 401 |
dbuf_rele(dbuf, FTAG); |
789 | 402 |
} else { |
403 |
/* the indirect block is sparse */ |
|
404 |
ASSERT(err == ENOENT); |
|
405 |
} |
|
406 |
||
407 |
blkid += tochk; |
|
408 |
nblks -= tochk; |
|
409 |
} |
|
410 |
rw_exit(&dn->dn_struct_rwlock); |
|
411 |
||
412 |
tx->tx_space_tofree += space; |
|
413 |
} |
|
414 |
||
415 |
static void |
|
416 |
dmu_tx_hold_free_impl(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len) |
|
417 |
{ |
|
1544 | 418 |
uint64_t start, end, i; |
419 |
int dirty, err, shift; |
|
420 |
zio_t *zio; |
|
789 | 421 |
|
422 |
/* first block */ |
|
423 |
if (off != 0 /* || dn->dn_maxblkid == 0 */) |
|
424 |
dmu_tx_count_write(tx, dn, off, 1); |
|
425 |
/* last block */ |
|
426 |
if (len != DMU_OBJECT_END) |
|
427 |
dmu_tx_count_write(tx, dn, off+len, 1); |
|
428 |
||
429 |
if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz) |
|
430 |
return; |
|
431 |
if (len == DMU_OBJECT_END) |
|
432 |
len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off; |
|
433 |
||
1544 | 434 |
/* |
435 |
* For i/o error checking, read the first and last level-0 |
|
436 |
* blocks, and all the level-1 blocks. The above count_write's |
|
437 |
* will take care of the level-0 blocks. |
|
438 |
*/ |
|
439 |
shift = dn->dn_datablkshift + dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
|
440 |
start = off >> shift; |
|
441 |
end = dn->dn_datablkshift ? ((off+len) >> shift) : 0; |
|
442 |
||
443 |
zio = zio_root(tx->tx_pool->dp_spa, NULL, NULL, ZIO_FLAG_CANFAIL); |
|
444 |
for (i = start+1; i < end; i++) { |
|
445 |
uint64_t ibyte = i << shift; |
|
446 |
err = dnode_next_offset(dn, FALSE, &ibyte, 2, 1); |
|
447 |
i = ibyte >> shift; |
|
448 |
if (err == ESRCH) |
|
449 |
break; |
|
450 |
if (err) { |
|
451 |
tx->tx_err = err; |
|
452 |
return; |
|
453 |
} |
|
454 |
||
455 |
err = dmu_tx_check_ioerr(zio, dn, 1, i); |
|
456 |
if (err) { |
|
457 |
tx->tx_err = err; |
|
458 |
return; |
|
459 |
} |
|
460 |
} |
|
461 |
err = zio_wait(zio); |
|
462 |
if (err) { |
|
463 |
tx->tx_err = err; |
|
464 |
return; |
|
465 |
} |
|
466 |
||
467 |
dmu_tx_count_dnode(tx, dn); |
|
468 |
||
789 | 469 |
/* XXX locking */ |
470 |
dirty = dn->dn_dirtyblksz[0] | dn->dn_dirtyblksz[1] | |
|
471 |
dn->dn_dirtyblksz[2] | dn->dn_dirtyblksz[3]; |
|
472 |
if (dn->dn_assigned_tx != NULL && !dirty) |
|
473 |
dmu_tx_count_free(tx, dn, off, len); |
|
474 |
} |
|
475 |
||
476 |
void |
|
477 |
dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len) |
|
478 |
{ |
|
479 |
ASSERT(tx->tx_txg == 0); |
|
480 |
||
481 |
dmu_tx_hold_object_impl(tx, tx->tx_objset, object, THT_FREE, |
|
482 |
dmu_tx_hold_free_impl, off, len); |
|
483 |
} |
|
484 |
||
485 |
/* ARGSUSED */ |
|
486 |
static void |
|
1544 | 487 |
dmu_tx_hold_zap_impl(dmu_tx_t *tx, dnode_t *dn, uint64_t add, uint64_t iname) |
789 | 488 |
{ |
489 |
uint64_t nblocks; |
|
1544 | 490 |
int epbs, err; |
491 |
char *name = (char *)(uintptr_t)iname; |
|
789 | 492 |
|
493 |
dmu_tx_count_dnode(tx, dn); |
|
494 |
||
495 |
if (dn == NULL) { |
|
496 |
/* |
|
1544 | 497 |
* We will be able to fit a new object's entries into one leaf |
789 | 498 |
* block. So there will be at most 2 blocks total, |
499 |
* including the header block. |
|
500 |
*/ |
|
1491
bdcb30e07e7d
6389368 fat zap should use 16k blocks (with backwards compatability)
ahrens
parents:
873
diff
changeset
|
501 |
dmu_tx_count_write(tx, dn, 0, 2 << fzap_default_block_shift); |
789 | 502 |
return; |
503 |
} |
|
504 |
||
505 |
ASSERT3P(dmu_ot[dn->dn_type].ot_byteswap, ==, zap_byteswap); |
|
506 |
||
1544 | 507 |
if (dn->dn_maxblkid == 0 && !add) { |
789 | 508 |
/* |
509 |
* If there is only one block (i.e. this is a micro-zap) |
|
1544 | 510 |
* and we are not adding anything, the accounting is simple. |
789 | 511 |
*/ |
1544 | 512 |
err = dmu_tx_check_ioerr(NULL, dn, 0, 0); |
513 |
if (err) { |
|
514 |
tx->tx_err = err; |
|
515 |
return; |
|
516 |
} |
|
517 |
||
789 | 518 |
if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset, |
1544 | 519 |
dn->dn_phys->dn_blkptr[0].blk_birth)) |
789 | 520 |
tx->tx_space_tooverwrite += dn->dn_datablksz; |
521 |
else |
|
522 |
tx->tx_space_towrite += dn->dn_datablksz; |
|
523 |
return; |
|
524 |
} |
|
525 |
||
1544 | 526 |
if (dn->dn_maxblkid > 0 && name) { |
527 |
/* |
|
528 |
* access the name in this fat-zap so that we'll check |
|
529 |
* for i/o errors to the leaf blocks, etc. |
|
530 |
*/ |
|
531 |
err = zap_lookup(&dn->dn_objset->os, dn->dn_object, name, |
|
532 |
8, 0, NULL); |
|
533 |
if (err == EIO) { |
|
534 |
tx->tx_err = err; |
|
535 |
return; |
|
536 |
} |
|
537 |
} |
|
538 |
||
789 | 539 |
/* |
1544 | 540 |
* 3 blocks overwritten: target leaf, ptrtbl block, header block |
541 |
* 3 new blocks written if adding: new split leaf, 2 grown ptrtbl blocks |
|
789 | 542 |
*/ |
543 |
dmu_tx_count_write(tx, dn, dn->dn_maxblkid * dn->dn_datablksz, |
|
1544 | 544 |
(3 + add ? 3 : 0) << dn->dn_datablkshift); |
789 | 545 |
|
546 |
/* |
|
547 |
* If the modified blocks are scattered to the four winds, |
|
548 |
* we'll have to modify an indirect twig for each. |
|
549 |
*/ |
|
550 |
epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
|
551 |
for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs) |
|
1544 | 552 |
tx->tx_space_towrite += 3 << dn->dn_indblkshift; |
789 | 553 |
} |
554 |
||
555 |
void |
|
1544 | 556 |
dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, char *name) |
789 | 557 |
{ |
558 |
ASSERT(tx->tx_txg == 0); |
|
559 |
||
560 |
dmu_tx_hold_object_impl(tx, tx->tx_objset, object, THT_ZAP, |
|
1544 | 561 |
dmu_tx_hold_zap_impl, add, (uintptr_t)name); |
789 | 562 |
} |
563 |
||
564 |
void |
|
565 |
dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object) |
|
566 |
{ |
|
567 |
ASSERT(tx->tx_txg == 0); |
|
568 |
||
569 |
dmu_tx_hold_object_impl(tx, tx->tx_objset, object, THT_BONUS, |
|
570 |
dmu_tx_hold_write_impl, 0, 0); |
|
571 |
} |
|
572 |
||
573 |
||
574 |
/* ARGSUSED */ |
|
575 |
static void |
|
576 |
dmu_tx_hold_space_impl(dmu_tx_t *tx, dnode_t *dn, |
|
577 |
uint64_t space, uint64_t unused) |
|
578 |
{ |
|
579 |
tx->tx_space_towrite += space; |
|
580 |
} |
|
581 |
||
582 |
void |
|
583 |
dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space) |
|
584 |
{ |
|
585 |
ASSERT(tx->tx_txg == 0); |
|
586 |
||
587 |
dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT, THT_SPACE, |
|
588 |
dmu_tx_hold_space_impl, space, 0); |
|
589 |
} |
|
590 |
||
591 |
int |
|
592 |
dmu_tx_holds(dmu_tx_t *tx, uint64_t object) |
|
593 |
{ |
|
594 |
dmu_tx_hold_t *dth; |
|
595 |
int holds = 0; |
|
596 |
||
597 |
/* |
|
598 |
* By asserting that the tx is assigned, we're counting the |
|
599 |
* number of dn_tx_holds, which is the same as the number of |
|
600 |
* dn_holds. Otherwise, we'd be counting dn_holds, but |
|
601 |
* dn_tx_holds could be 0. |
|
602 |
*/ |
|
603 |
ASSERT(tx->tx_txg != 0); |
|
604 |
||
605 |
/* if (tx->tx_anyobj == TRUE) */ |
|
606 |
/* return (0); */ |
|
607 |
||
608 |
for (dth = list_head(&tx->tx_holds); dth; |
|
609 |
dth = list_next(&tx->tx_holds, dth)) { |
|
610 |
if (dth->dth_dnode && dth->dth_dnode->dn_object == object) |
|
611 |
holds++; |
|
612 |
} |
|
613 |
||
614 |
return (holds); |
|
615 |
} |
|
616 |
||
873
adefbfa5f42d
6347448 non ZFS_DEBUG kernels shouldn't call empty verify functions
ek110237
parents:
789
diff
changeset
|
617 |
#ifdef ZFS_DEBUG |
789 | 618 |
void |
619 |
dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db) |
|
620 |
{ |
|
621 |
dmu_tx_hold_t *dth; |
|
622 |
int match_object = FALSE, match_offset = FALSE; |
|
623 |
dnode_t *dn = db->db_dnode; |
|
624 |
||
625 |
ASSERT(tx->tx_txg != 0); |
|
626 |
ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset->os); |
|
627 |
ASSERT3U(dn->dn_object, ==, db->db.db_object); |
|
628 |
||
629 |
if (tx->tx_anyobj) |
|
630 |
return; |
|
631 |
||
632 |
/* XXX No checking on the meta dnode for now */ |
|
1544 | 633 |
if (db->db.db_object == DMU_META_DNODE_OBJECT) |
789 | 634 |
return; |
635 |
||
636 |
for (dth = list_head(&tx->tx_holds); dth; |
|
637 |
dth = list_next(&tx->tx_holds, dth)) { |
|
638 |
ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg); |
|
639 |
if (dth->dth_dnode == dn && dth->dth_type != THT_NEWOBJECT) |
|
640 |
match_object = TRUE; |
|
641 |
if (dth->dth_dnode == NULL || dth->dth_dnode == dn) { |
|
642 |
int datablkshift = dn->dn_datablkshift ? |
|
643 |
dn->dn_datablkshift : SPA_MAXBLOCKSHIFT; |
|
644 |
int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
|
645 |
int shift = datablkshift + epbs * db->db_level; |
|
646 |
uint64_t beginblk = shift >= 64 ? 0 : |
|
647 |
(dth->dth_arg1 >> shift); |
|
648 |
uint64_t endblk = shift >= 64 ? 0 : |
|
649 |
((dth->dth_arg1 + dth->dth_arg2 - 1) >> shift); |
|
650 |
uint64_t blkid = db->db_blkid; |
|
651 |
||
652 |
/* XXX dth_arg2 better not be zero... */ |
|
653 |
||
654 |
dprintf("found dth type %x beginblk=%llx endblk=%llx\n", |
|
655 |
dth->dth_type, beginblk, endblk); |
|
656 |
||
657 |
switch (dth->dth_type) { |
|
658 |
case THT_WRITE: |
|
659 |
if (blkid >= beginblk && blkid <= endblk) |
|
660 |
match_offset = TRUE; |
|
661 |
/* |
|
662 |
* We will let this hold work for the bonus |
|
663 |
* buffer so that we don't need to hold it |
|
664 |
* when creating a new object. |
|
665 |
*/ |
|
666 |
if (blkid == DB_BONUS_BLKID) |
|
667 |
match_offset = TRUE; |
|
668 |
/* |
|
669 |
* They might have to increase nlevels, |
|
670 |
* thus dirtying the new TLIBs. Or the |
|
671 |
* might have to change the block size, |
|
672 |
* thus dirying the new lvl=0 blk=0. |
|
673 |
*/ |
|
674 |
if (blkid == 0) |
|
675 |
match_offset = TRUE; |
|
676 |
break; |
|
677 |
case THT_FREE: |
|
678 |
if (blkid == beginblk && |
|
679 |
(dth->dth_arg1 != 0 || |
|
680 |
dn->dn_maxblkid == 0)) |
|
681 |
match_offset = TRUE; |
|
682 |
if (blkid == endblk && |
|
683 |
dth->dth_arg2 != DMU_OBJECT_END) |
|
684 |
match_offset = TRUE; |
|
685 |
break; |
|
686 |
case THT_BONUS: |
|
687 |
if (blkid == DB_BONUS_BLKID) |
|
688 |
match_offset = TRUE; |
|
689 |
break; |
|
690 |
case THT_ZAP: |
|
691 |
match_offset = TRUE; |
|
692 |
break; |
|
693 |
case THT_NEWOBJECT: |
|
694 |
match_object = TRUE; |
|
695 |
break; |
|
696 |
default: |
|
697 |
ASSERT(!"bad dth_type"); |
|
698 |
} |
|
699 |
} |
|
700 |
if (match_object && match_offset) |
|
701 |
return; |
|
702 |
} |
|
703 |
panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n", |
|
704 |
(u_longlong_t)db->db.db_object, db->db_level, |
|
705 |
(u_longlong_t)db->db_blkid); |
|
873
adefbfa5f42d
6347448 non ZFS_DEBUG kernels shouldn't call empty verify functions
ek110237
parents:
789
diff
changeset
|
706 |
} |
789 | 707 |
#endif |
708 |
||
709 |
static int |
|
710 |
dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how, dmu_tx_hold_t **last_dth) |
|
711 |
{ |
|
712 |
dmu_tx_hold_t *dth; |
|
1544 | 713 |
uint64_t lsize, asize, fsize, towrite; |
789 | 714 |
|
715 |
*last_dth = NULL; |
|
716 |
||
717 |
tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh); |
|
718 |
||
719 |
if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg) |
|
720 |
return (ERESTART); |
|
1544 | 721 |
if (tx->tx_err) |
722 |
return (tx->tx_err); |
|
789 | 723 |
|
724 |
for (dth = list_head(&tx->tx_holds); dth; |
|
1544 | 725 |
dth = list_next(&tx->tx_holds, dth)) { |
789 | 726 |
dnode_t *dn = dth->dth_dnode; |
727 |
if (dn != NULL) { |
|
728 |
mutex_enter(&dn->dn_mtx); |
|
729 |
while (dn->dn_assigned_txg == tx->tx_txg - 1) { |
|
730 |
if (txg_how != TXG_WAIT) { |
|
731 |
mutex_exit(&dn->dn_mtx); |
|
732 |
return (ERESTART); |
|
733 |
} |
|
734 |
cv_wait(&dn->dn_notxholds, &dn->dn_mtx); |
|
735 |
} |
|
736 |
if (dn->dn_assigned_txg == 0) { |
|
737 |
ASSERT(dn->dn_assigned_tx == NULL); |
|
738 |
dn->dn_assigned_txg = tx->tx_txg; |
|
739 |
dn->dn_assigned_tx = tx; |
|
740 |
} else { |
|
741 |
ASSERT(dn->dn_assigned_txg == tx->tx_txg); |
|
742 |
if (dn->dn_assigned_tx != tx) |
|
743 |
dn->dn_assigned_tx = NULL; |
|
744 |
} |
|
745 |
(void) refcount_add(&dn->dn_tx_holds, tx); |
|
746 |
mutex_exit(&dn->dn_mtx); |
|
747 |
} |
|
1544 | 748 |
*last_dth = dth; |
749 |
if (tx->tx_err) |
|
750 |
return (tx->tx_err); |
|
751 |
} |
|
752 |
||
753 |
/* |
|
754 |
* If a snapshot has been taken since we made our estimates, |
|
755 |
* assume that we won't be able to free or overwrite anything. |
|
756 |
*/ |
|
757 |
if (tx->tx_objset && |
|
758 |
dsl_dataset_prev_snap_txg(tx->tx_objset->os->os_dsl_dataset) > |
|
759 |
tx->tx_lastsnap_txg) { |
|
760 |
tx->tx_space_towrite += tx->tx_space_tooverwrite; |
|
761 |
tx->tx_space_tooverwrite = 0; |
|
762 |
tx->tx_space_tofree = 0; |
|
789 | 763 |
} |
764 |
||
765 |
/* |
|
766 |
* Convert logical size to worst-case allocated size. |
|
767 |
*/ |
|
768 |
fsize = spa_get_asize(tx->tx_pool->dp_spa, tx->tx_space_tooverwrite) + |
|
769 |
tx->tx_space_tofree; |
|
770 |
lsize = tx->tx_space_towrite + tx->tx_space_tooverwrite; |
|
771 |
asize = spa_get_asize(tx->tx_pool->dp_spa, lsize); |
|
1544 | 772 |
towrite = tx->tx_space_towrite; |
789 | 773 |
tx->tx_space_towrite = asize; |
774 |
||
775 |
if (tx->tx_dir && asize != 0) { |
|
776 |
int err = dsl_dir_tempreserve_space(tx->tx_dir, |
|
777 |
lsize, asize, fsize, &tx->tx_tempreserve_cookie, tx); |
|
1544 | 778 |
if (err) { |
779 |
tx->tx_space_towrite = towrite; |
|
789 | 780 |
return (err); |
1544 | 781 |
} |
789 | 782 |
} |
783 |
||
784 |
return (0); |
|
785 |
} |
|
786 |
||
787 |
static uint64_t |
|
788 |
dmu_tx_unassign(dmu_tx_t *tx, dmu_tx_hold_t *last_dth) |
|
789 |
{ |
|
790 |
uint64_t txg = tx->tx_txg; |
|
791 |
dmu_tx_hold_t *dth; |
|
792 |
||
793 |
ASSERT(txg != 0); |
|
794 |
||
795 |
txg_rele_to_quiesce(&tx->tx_txgh); |
|
796 |
||
797 |
for (dth = last_dth; dth; dth = list_prev(&tx->tx_holds, dth)) { |
|
798 |
dnode_t *dn = dth->dth_dnode; |
|
799 |
||
800 |
if (dn == NULL) |
|
801 |
continue; |
|
802 |
mutex_enter(&dn->dn_mtx); |
|
803 |
ASSERT3U(dn->dn_assigned_txg, ==, txg); |
|
804 |
||
805 |
if (refcount_remove(&dn->dn_tx_holds, tx) == 0) { |
|
806 |
dn->dn_assigned_txg = 0; |
|
807 |
dn->dn_assigned_tx = NULL; |
|
808 |
cv_broadcast(&dn->dn_notxholds); |
|
809 |
} |
|
810 |
mutex_exit(&dn->dn_mtx); |
|
811 |
} |
|
812 |
||
813 |
txg_rele_to_sync(&tx->tx_txgh); |
|
814 |
||
815 |
tx->tx_txg = 0; |
|
816 |
return (txg); |
|
817 |
} |
|
818 |
||
819 |
/* |
|
820 |
* Assign tx to a transaction group. txg_how can be one of: |
|
821 |
* |
|
822 |
* (1) TXG_WAIT. If the current open txg is full, waits until there's |
|
823 |
* a new one. This should be used when you're not holding locks. |
|
824 |
* If will only fail if we're truly out of space (or over quota). |
|
825 |
* |
|
826 |
* (2) TXG_NOWAIT. If we can't assign into the current open txg without |
|
827 |
* blocking, returns immediately with ERESTART. This should be used |
|
828 |
* whenever you're holding locks. On an ERESTART error, the caller |
|
829 |
* should drop locks, do a txg_wait_open(dp, 0), and try again. |
|
830 |
* |
|
831 |
* (3) A specific txg. Use this if you need to ensure that multiple |
|
832 |
* transactions all sync in the same txg. Like TXG_NOWAIT, it |
|
833 |
* returns ERESTART if it can't assign you into the requested txg. |
|
834 |
*/ |
|
835 |
int |
|
836 |
dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how) |
|
837 |
{ |
|
838 |
dmu_tx_hold_t *last_dth; |
|
839 |
int err; |
|
840 |
||
841 |
ASSERT(tx->tx_txg == 0); |
|
842 |
ASSERT(txg_how != 0); |
|
843 |
ASSERT(!dsl_pool_sync_context(tx->tx_pool)); |
|
844 |
||
845 |
while ((err = dmu_tx_try_assign(tx, txg_how, &last_dth)) != 0) { |
|
846 |
uint64_t txg = dmu_tx_unassign(tx, last_dth); |
|
847 |
||
848 |
if (err != ERESTART || txg_how != TXG_WAIT) |
|
849 |
return (err); |
|
850 |
||
851 |
txg_wait_open(tx->tx_pool, txg + 1); |
|
852 |
} |
|
853 |
||
854 |
txg_rele_to_quiesce(&tx->tx_txgh); |
|
855 |
||
856 |
return (0); |
|
857 |
} |
|
858 |
||
859 |
void |
|
860 |
dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta) |
|
861 |
{ |
|
862 |
if (tx->tx_dir == NULL || delta == 0) |
|
863 |
return; |
|
864 |
||
865 |
if (delta > 0) { |
|
866 |
ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=, |
|
867 |
tx->tx_space_towrite); |
|
868 |
(void) refcount_add_many(&tx->tx_space_written, delta, NULL); |
|
869 |
} else { |
|
870 |
(void) refcount_add_many(&tx->tx_space_freed, -delta, NULL); |
|
871 |
} |
|
872 |
} |
|
873 |
||
874 |
void |
|
875 |
dmu_tx_commit(dmu_tx_t *tx) |
|
876 |
{ |
|
877 |
dmu_tx_hold_t *dth; |
|
878 |
||
879 |
ASSERT(tx->tx_txg != 0); |
|
880 |
||
881 |
while (dth = list_head(&tx->tx_holds)) { |
|
882 |
dnode_t *dn = dth->dth_dnode; |
|
883 |
||
884 |
list_remove(&tx->tx_holds, dth); |
|
885 |
kmem_free(dth, sizeof (dmu_tx_hold_t)); |
|
886 |
if (dn == NULL) |
|
887 |
continue; |
|
888 |
mutex_enter(&dn->dn_mtx); |
|
889 |
ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg); |
|
890 |
||
891 |
if (refcount_remove(&dn->dn_tx_holds, tx) == 0) { |
|
892 |
dn->dn_assigned_txg = 0; |
|
893 |
dn->dn_assigned_tx = NULL; |
|
894 |
cv_broadcast(&dn->dn_notxholds); |
|
895 |
} |
|
896 |
mutex_exit(&dn->dn_mtx); |
|
897 |
dnode_rele(dn, tx); |
|
898 |
} |
|
899 |
||
900 |
if (tx->tx_dir && tx->tx_space_towrite > 0) { |
|
901 |
dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx); |
|
902 |
} |
|
903 |
||
904 |
if (tx->tx_anyobj == FALSE) |
|
905 |
txg_rele_to_sync(&tx->tx_txgh); |
|
906 |
dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n", |
|
907 |
tx->tx_space_towrite, refcount_count(&tx->tx_space_written), |
|
908 |
tx->tx_space_tofree, refcount_count(&tx->tx_space_freed)); |
|
909 |
refcount_destroy_many(&tx->tx_space_written, |
|
910 |
refcount_count(&tx->tx_space_written)); |
|
911 |
refcount_destroy_many(&tx->tx_space_freed, |
|
912 |
refcount_count(&tx->tx_space_freed)); |
|
913 |
#ifdef ZFS_DEBUG |
|
914 |
if (tx->tx_debug_buf) |
|
915 |
kmem_free(tx->tx_debug_buf, 4096); |
|
916 |
#endif |
|
917 |
kmem_free(tx, sizeof (dmu_tx_t)); |
|
918 |
} |
|
919 |
||
920 |
void |
|
921 |
dmu_tx_abort(dmu_tx_t *tx) |
|
922 |
{ |
|
923 |
dmu_tx_hold_t *dth; |
|
924 |
||
925 |
ASSERT(tx->tx_txg == 0); |
|
926 |
||
927 |
while (dth = list_head(&tx->tx_holds)) { |
|
928 |
dnode_t *dn = dth->dth_dnode; |
|
929 |
||
930 |
list_remove(&tx->tx_holds, dth); |
|
931 |
kmem_free(dth, sizeof (dmu_tx_hold_t)); |
|
932 |
if (dn != NULL) |
|
933 |
dnode_rele(dn, tx); |
|
934 |
} |
|
935 |
refcount_destroy_many(&tx->tx_space_written, |
|
936 |
refcount_count(&tx->tx_space_written)); |
|
937 |
refcount_destroy_many(&tx->tx_space_freed, |
|
938 |
refcount_count(&tx->tx_space_freed)); |
|
939 |
#ifdef ZFS_DEBUG |
|
940 |
if (tx->tx_debug_buf) |
|
941 |
kmem_free(tx->tx_debug_buf, 4096); |
|
942 |
#endif |
|
943 |
kmem_free(tx, sizeof (dmu_tx_t)); |
|
944 |
} |
|
945 |
||
946 |
uint64_t |
|
947 |
dmu_tx_get_txg(dmu_tx_t *tx) |
|
948 |
{ |
|
949 |
ASSERT(tx->tx_txg != 0); |
|
950 |
return (tx->tx_txg); |
|
951 |
} |