Print this page
1862 incremental zfs receive fails for sparse file > 8PB
dmu_tx_count_free is doing a horrible over-estimation of used memory. It
assumes that the file is fully non-sparse and calculates a worst-case estimate
of how much memory is needed to hold all metadata for the file. If a large
hole needs to be freed, the estimation goes into the TB-range, which obviously
fails later on.
This patch tries to calculate a more realistic estimate by counting the l1
blocks (the loop for this is already present) and assumes a worst-case
distribution of those blocks over the full length given.
Reviewed by: Matt Ahrens <matthew.ahrens@delphix.com>
Reviewed by: Simon Klinkert <klinkert@webgods.de>
Split |
Close |
Expand all |
Collapse all |
--- old/usr/src/uts/common/fs/zfs/dmu_tx.c
+++ new/usr/src/uts/common/fs/zfs/dmu_tx.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 24 * Copyright (c) 2012 by Delphix. All rights reserved.
25 25 */
26 26
27 27 #include <sys/dmu.h>
28 28 #include <sys/dmu_impl.h>
29 29 #include <sys/dbuf.h>
30 30 #include <sys/dmu_tx.h>
31 31 #include <sys/dmu_objset.h>
32 32 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
33 33 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
34 34 #include <sys/dsl_pool.h>
35 35 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 36 #include <sys/spa.h>
37 37 #include <sys/sa.h>
38 38 #include <sys/sa_impl.h>
39 39 #include <sys/zfs_context.h>
40 40 #include <sys/varargs.h>
41 41
42 42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43 43 uint64_t arg1, uint64_t arg2);
44 44
45 45
46 46 dmu_tx_t *
47 47 dmu_tx_create_dd(dsl_dir_t *dd)
48 48 {
49 49 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
50 50 tx->tx_dir = dd;
51 51 if (dd != NULL)
52 52 tx->tx_pool = dd->dd_pool;
53 53 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
54 54 offsetof(dmu_tx_hold_t, txh_node));
55 55 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
56 56 offsetof(dmu_tx_callback_t, dcb_node));
57 57 #ifdef ZFS_DEBUG
58 58 refcount_create(&tx->tx_space_written);
59 59 refcount_create(&tx->tx_space_freed);
60 60 #endif
61 61 return (tx);
62 62 }
63 63
64 64 dmu_tx_t *
65 65 dmu_tx_create(objset_t *os)
66 66 {
67 67 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
68 68 tx->tx_objset = os;
69 69 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
70 70 return (tx);
71 71 }
72 72
73 73 dmu_tx_t *
74 74 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
75 75 {
76 76 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
77 77
78 78 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
79 79 tx->tx_pool = dp;
80 80 tx->tx_txg = txg;
81 81 tx->tx_anyobj = TRUE;
82 82
83 83 return (tx);
84 84 }
85 85
86 86 int
87 87 dmu_tx_is_syncing(dmu_tx_t *tx)
88 88 {
89 89 return (tx->tx_anyobj);
90 90 }
91 91
92 92 int
93 93 dmu_tx_private_ok(dmu_tx_t *tx)
94 94 {
95 95 return (tx->tx_anyobj);
96 96 }
97 97
98 98 static dmu_tx_hold_t *
99 99 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
100 100 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
101 101 {
102 102 dmu_tx_hold_t *txh;
103 103 dnode_t *dn = NULL;
104 104 int err;
105 105
106 106 if (object != DMU_NEW_OBJECT) {
107 107 err = dnode_hold(os, object, tx, &dn);
108 108 if (err) {
109 109 tx->tx_err = err;
110 110 return (NULL);
111 111 }
112 112
113 113 if (err == 0 && tx->tx_txg != 0) {
114 114 mutex_enter(&dn->dn_mtx);
115 115 /*
116 116 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
117 117 * problem, but there's no way for it to happen (for
118 118 * now, at least).
119 119 */
120 120 ASSERT(dn->dn_assigned_txg == 0);
121 121 dn->dn_assigned_txg = tx->tx_txg;
122 122 (void) refcount_add(&dn->dn_tx_holds, tx);
123 123 mutex_exit(&dn->dn_mtx);
124 124 }
125 125 }
126 126
127 127 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
128 128 txh->txh_tx = tx;
129 129 txh->txh_dnode = dn;
130 130 #ifdef ZFS_DEBUG
131 131 txh->txh_type = type;
132 132 txh->txh_arg1 = arg1;
133 133 txh->txh_arg2 = arg2;
134 134 #endif
135 135 list_insert_tail(&tx->tx_holds, txh);
136 136
137 137 return (txh);
138 138 }
139 139
140 140 void
141 141 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
142 142 {
143 143 /*
144 144 * If we're syncing, they can manipulate any object anyhow, and
145 145 * the hold on the dnode_t can cause problems.
146 146 */
147 147 if (!dmu_tx_is_syncing(tx)) {
148 148 (void) dmu_tx_hold_object_impl(tx, os,
149 149 object, THT_NEWOBJECT, 0, 0);
150 150 }
151 151 }
152 152
153 153 static int
154 154 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
155 155 {
156 156 int err;
157 157 dmu_buf_impl_t *db;
158 158
159 159 rw_enter(&dn->dn_struct_rwlock, RW_READER);
160 160 db = dbuf_hold_level(dn, level, blkid, FTAG);
161 161 rw_exit(&dn->dn_struct_rwlock);
162 162 if (db == NULL)
163 163 return (EIO);
164 164 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
165 165 dbuf_rele(db, FTAG);
166 166 return (err);
167 167 }
168 168
169 169 static void
170 170 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
171 171 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
172 172 {
173 173 objset_t *os = dn->dn_objset;
174 174 dsl_dataset_t *ds = os->os_dsl_dataset;
175 175 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
176 176 dmu_buf_impl_t *parent = NULL;
177 177 blkptr_t *bp = NULL;
178 178 uint64_t space;
179 179
180 180 if (level >= dn->dn_nlevels || history[level] == blkid)
181 181 return;
182 182
183 183 history[level] = blkid;
184 184
185 185 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
186 186
187 187 if (db == NULL || db == dn->dn_dbuf) {
188 188 ASSERT(level != 0);
189 189 db = NULL;
190 190 } else {
191 191 ASSERT(DB_DNODE(db) == dn);
192 192 ASSERT(db->db_level == level);
193 193 ASSERT(db->db.db_size == space);
194 194 ASSERT(db->db_blkid == blkid);
195 195 bp = db->db_blkptr;
196 196 parent = db->db_parent;
197 197 }
198 198
199 199 freeable = (bp && (freeable ||
200 200 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
201 201
202 202 if (freeable)
203 203 txh->txh_space_tooverwrite += space;
204 204 else
205 205 txh->txh_space_towrite += space;
206 206 if (bp)
207 207 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
208 208
209 209 dmu_tx_count_twig(txh, dn, parent, level + 1,
210 210 blkid >> epbs, freeable, history);
211 211 }
212 212
213 213 /* ARGSUSED */
214 214 static void
215 215 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
216 216 {
217 217 dnode_t *dn = txh->txh_dnode;
218 218 uint64_t start, end, i;
219 219 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
220 220 int err = 0;
221 221
222 222 if (len == 0)
223 223 return;
224 224
225 225 min_bs = SPA_MINBLOCKSHIFT;
226 226 max_bs = SPA_MAXBLOCKSHIFT;
227 227 min_ibs = DN_MIN_INDBLKSHIFT;
228 228 max_ibs = DN_MAX_INDBLKSHIFT;
229 229
230 230 if (dn) {
231 231 uint64_t history[DN_MAX_LEVELS];
232 232 int nlvls = dn->dn_nlevels;
233 233 int delta;
234 234
235 235 /*
236 236 * For i/o error checking, read the first and last level-0
237 237 * blocks (if they are not aligned), and all the level-1 blocks.
238 238 */
239 239 if (dn->dn_maxblkid == 0) {
240 240 delta = dn->dn_datablksz;
241 241 start = (off < dn->dn_datablksz) ? 0 : 1;
242 242 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
243 243 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
244 244 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
245 245 if (err)
246 246 goto out;
247 247 delta -= off;
248 248 }
249 249 } else {
250 250 zio_t *zio = zio_root(dn->dn_objset->os_spa,
251 251 NULL, NULL, ZIO_FLAG_CANFAIL);
252 252
253 253 /* first level-0 block */
254 254 start = off >> dn->dn_datablkshift;
255 255 if (P2PHASE(off, dn->dn_datablksz) ||
256 256 len < dn->dn_datablksz) {
257 257 err = dmu_tx_check_ioerr(zio, dn, 0, start);
258 258 if (err)
259 259 goto out;
260 260 }
261 261
262 262 /* last level-0 block */
263 263 end = (off+len-1) >> dn->dn_datablkshift;
264 264 if (end != start && end <= dn->dn_maxblkid &&
265 265 P2PHASE(off+len, dn->dn_datablksz)) {
266 266 err = dmu_tx_check_ioerr(zio, dn, 0, end);
267 267 if (err)
268 268 goto out;
269 269 }
270 270
271 271 /* level-1 blocks */
272 272 if (nlvls > 1) {
273 273 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
274 274 for (i = (start>>shft)+1; i < end>>shft; i++) {
275 275 err = dmu_tx_check_ioerr(zio, dn, 1, i);
276 276 if (err)
277 277 goto out;
278 278 }
279 279 }
280 280
281 281 err = zio_wait(zio);
282 282 if (err)
283 283 goto out;
284 284 delta = P2NPHASE(off, dn->dn_datablksz);
285 285 }
286 286
287 287 if (dn->dn_maxblkid > 0) {
288 288 /*
289 289 * The blocksize can't change,
290 290 * so we can make a more precise estimate.
291 291 */
292 292 ASSERT(dn->dn_datablkshift != 0);
293 293 min_bs = max_bs = dn->dn_datablkshift;
294 294 min_ibs = max_ibs = dn->dn_indblkshift;
295 295 } else if (dn->dn_indblkshift > max_ibs) {
296 296 /*
297 297 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
298 298 * the code will still work correctly on older pools.
299 299 */
300 300 min_ibs = max_ibs = dn->dn_indblkshift;
301 301 }
302 302
303 303 /*
304 304 * If this write is not off the end of the file
305 305 * we need to account for overwrites/unref.
306 306 */
307 307 if (start <= dn->dn_maxblkid) {
308 308 for (int l = 0; l < DN_MAX_LEVELS; l++)
309 309 history[l] = -1ULL;
310 310 }
311 311 while (start <= dn->dn_maxblkid) {
312 312 dmu_buf_impl_t *db;
313 313
314 314 rw_enter(&dn->dn_struct_rwlock, RW_READER);
315 315 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
316 316 rw_exit(&dn->dn_struct_rwlock);
317 317
318 318 if (err) {
319 319 txh->txh_tx->tx_err = err;
320 320 return;
321 321 }
322 322
323 323 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
324 324 history);
325 325 dbuf_rele(db, FTAG);
326 326 if (++start > end) {
327 327 /*
328 328 * Account for new indirects appearing
329 329 * before this IO gets assigned into a txg.
330 330 */
331 331 bits = 64 - min_bs;
332 332 epbs = min_ibs - SPA_BLKPTRSHIFT;
333 333 for (bits -= epbs * (nlvls - 1);
334 334 bits >= 0; bits -= epbs)
335 335 txh->txh_fudge += 1ULL << max_ibs;
336 336 goto out;
337 337 }
338 338 off += delta;
339 339 if (len >= delta)
340 340 len -= delta;
341 341 delta = dn->dn_datablksz;
342 342 }
343 343 }
344 344
345 345 /*
346 346 * 'end' is the last thing we will access, not one past.
347 347 * This way we won't overflow when accessing the last byte.
348 348 */
349 349 start = P2ALIGN(off, 1ULL << max_bs);
350 350 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
351 351 txh->txh_space_towrite += end - start + 1;
352 352
353 353 start >>= min_bs;
354 354 end >>= min_bs;
355 355
356 356 epbs = min_ibs - SPA_BLKPTRSHIFT;
357 357
358 358 /*
359 359 * The object contains at most 2^(64 - min_bs) blocks,
360 360 * and each indirect level maps 2^epbs.
361 361 */
362 362 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
363 363 start >>= epbs;
364 364 end >>= epbs;
365 365 ASSERT3U(end, >=, start);
366 366 txh->txh_space_towrite += (end - start + 1) << max_ibs;
367 367 if (start != 0) {
368 368 /*
369 369 * We also need a new blkid=0 indirect block
370 370 * to reference any existing file data.
371 371 */
372 372 txh->txh_space_towrite += 1ULL << max_ibs;
373 373 }
374 374 }
375 375
376 376 out:
377 377 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
378 378 2 * DMU_MAX_ACCESS)
379 379 err = EFBIG;
380 380
381 381 if (err)
382 382 txh->txh_tx->tx_err = err;
383 383 }
384 384
385 385 static void
386 386 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
387 387 {
388 388 dnode_t *dn = txh->txh_dnode;
389 389 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
390 390 uint64_t space = mdn->dn_datablksz +
391 391 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
392 392
393 393 if (dn && dn->dn_dbuf->db_blkptr &&
394 394 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
395 395 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
396 396 txh->txh_space_tooverwrite += space;
397 397 txh->txh_space_tounref += space;
398 398 } else {
399 399 txh->txh_space_towrite += space;
400 400 if (dn && dn->dn_dbuf->db_blkptr)
401 401 txh->txh_space_tounref += space;
402 402 }
403 403 }
404 404
405 405 void
406 406 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
407 407 {
408 408 dmu_tx_hold_t *txh;
409 409
410 410 ASSERT(tx->tx_txg == 0);
411 411 ASSERT(len < DMU_MAX_ACCESS);
412 412 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
413 413
414 414 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
415 415 object, THT_WRITE, off, len);
416 416 if (txh == NULL)
417 417 return;
418 418
419 419 dmu_tx_count_write(txh, off, len);
420 420 dmu_tx_count_dnode(txh);
421 421 }
↓ open down ↓ |
421 lines elided |
↑ open up ↑ |
422 422
423 423 static void
424 424 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
425 425 {
426 426 uint64_t blkid, nblks, lastblk;
427 427 uint64_t space = 0, unref = 0, skipped = 0;
428 428 dnode_t *dn = txh->txh_dnode;
429 429 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
430 430 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
431 431 int epbs;
432 + uint64_t l0span = 0, nl1blks = 0;
432 433
433 434 if (dn->dn_nlevels == 0)
434 435 return;
435 436
436 437 /*
437 438 * The struct_rwlock protects us against dn_nlevels
438 439 * changing, in case (against all odds) we manage to dirty &
439 440 * sync out the changes after we check for being dirty.
440 441 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
441 442 */
442 443 rw_enter(&dn->dn_struct_rwlock, RW_READER);
443 444 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
444 445 if (dn->dn_maxblkid == 0) {
445 446 if (off == 0 && len >= dn->dn_datablksz) {
446 447 blkid = 0;
447 448 nblks = 1;
448 449 } else {
449 450 rw_exit(&dn->dn_struct_rwlock);
450 451 return;
451 452 }
452 453 } else {
453 454 blkid = off >> dn->dn_datablkshift;
↓ open down ↓ |
12 lines elided |
↑ open up ↑ |
454 455 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
455 456
456 457 if (blkid >= dn->dn_maxblkid) {
457 458 rw_exit(&dn->dn_struct_rwlock);
458 459 return;
459 460 }
460 461 if (blkid + nblks > dn->dn_maxblkid)
461 462 nblks = dn->dn_maxblkid - blkid;
462 463
463 464 }
465 + l0span = nblks; /* save for later use to calc level > 1 overhead */
464 466 if (dn->dn_nlevels == 1) {
465 467 int i;
466 468 for (i = 0; i < nblks; i++) {
467 469 blkptr_t *bp = dn->dn_phys->dn_blkptr;
468 470 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
469 471 bp += blkid + i;
470 472 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
471 473 dprintf_bp(bp, "can free old%s", "");
472 474 space += bp_get_dsize(spa, bp);
473 475 }
474 476 unref += BP_GET_ASIZE(bp);
475 477 }
478 + nl1blks = 1;
476 479 nblks = 0;
477 480 }
478 481
479 - /*
480 - * Add in memory requirements of higher-level indirects.
481 - * This assumes a worst-possible scenario for dn_nlevels.
482 - */
483 - {
484 - uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
485 - int level = (dn->dn_nlevels > 1) ? 2 : 1;
486 -
487 - while (level++ < DN_MAX_LEVELS) {
488 - txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
489 - blkcnt = 1 + (blkcnt >> epbs);
490 - }
491 - ASSERT(blkcnt <= dn->dn_nblkptr);
492 - }
493 -
494 482 lastblk = blkid + nblks - 1;
495 483 while (nblks) {
496 484 dmu_buf_impl_t *dbuf;
497 485 uint64_t ibyte, new_blkid;
498 486 int epb = 1 << epbs;
499 487 int err, i, blkoff, tochk;
500 488 blkptr_t *bp;
501 489
502 490 ibyte = blkid << dn->dn_datablkshift;
503 491 err = dnode_next_offset(dn,
504 492 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
505 493 new_blkid = ibyte >> dn->dn_datablkshift;
506 494 if (err == ESRCH) {
507 495 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
508 496 break;
509 497 }
510 498 if (err) {
511 499 txh->txh_tx->tx_err = err;
512 500 break;
513 501 }
514 502 if (new_blkid > lastblk) {
515 503 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
516 504 break;
517 505 }
518 506
519 507 if (new_blkid > blkid) {
520 508 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
521 509 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
522 510 nblks -= new_blkid - blkid;
523 511 blkid = new_blkid;
524 512 }
525 513 blkoff = P2PHASE(blkid, epb);
526 514 tochk = MIN(epb - blkoff, nblks);
527 515
528 516 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
529 517 if (err) {
530 518 txh->txh_tx->tx_err = err;
531 519 break;
532 520 }
533 521
534 522 txh->txh_memory_tohold += dbuf->db.db_size;
535 523
536 524 /*
537 525 * We don't check memory_tohold against DMU_MAX_ACCESS because
538 526 * memory_tohold is an over-estimation (especially the >L1
539 527 * indirect blocks), so it could fail. Callers should have
540 528 * already verified that they will not be holding too much
541 529 * memory.
542 530 */
543 531
544 532 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
545 533 if (err != 0) {
546 534 txh->txh_tx->tx_err = err;
547 535 dbuf_rele(dbuf, FTAG);
548 536 break;
549 537 }
550 538
551 539 bp = dbuf->db.db_data;
552 540 bp += blkoff;
553 541
↓ open down ↓ |
50 lines elided |
↑ open up ↑ |
554 542 for (i = 0; i < tochk; i++) {
555 543 if (dsl_dataset_block_freeable(ds, &bp[i],
556 544 bp[i].blk_birth)) {
557 545 dprintf_bp(&bp[i], "can free old%s", "");
558 546 space += bp_get_dsize(spa, &bp[i]);
559 547 }
560 548 unref += BP_GET_ASIZE(bp);
561 549 }
562 550 dbuf_rele(dbuf, FTAG);
563 551
552 + ++nl1blks;
564 553 blkid += tochk;
565 554 nblks -= tochk;
566 555 }
567 556 rw_exit(&dn->dn_struct_rwlock);
568 557
558 + /*
559 + * Add in memory requirements of higher-level indirects.
560 + * This assumes a worst-possible scenario for dn_nlevels and a
561 + * worst-possible distribution of l1-blocks over the region to free.
562 + */
563 + {
564 + uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
565 + int level = 2;
566 + /*
567 + * Here we don't use DN_MAX_LEVEL, but calculate it with the
568 + * given datablkshift and indblkshift. This makes the
569 + * difference between 19 and 8 on large files.
570 + */
571 + int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
572 + (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
573 +
574 + while (level++ < maxlevel) {
575 + txh->txh_memory_tohold += MIN(blkcnt, (nl1blks >> epbs))
576 + << dn->dn_indblkshift;
577 + blkcnt = 1 + (blkcnt >> epbs);
578 + }
579 + }
580 +
569 581 /* account for new level 1 indirect blocks that might show up */
570 582 if (skipped > 0) {
571 583 txh->txh_fudge += skipped << dn->dn_indblkshift;
572 584 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
573 585 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
574 586 }
575 587 txh->txh_space_tofree += space;
576 588 txh->txh_space_tounref += unref;
577 589 }
578 590
579 591 void
580 592 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
581 593 {
582 594 dmu_tx_hold_t *txh;
583 595 dnode_t *dn;
584 596 uint64_t start, end, i;
585 597 int err, shift;
586 598 zio_t *zio;
587 599
588 600 ASSERT(tx->tx_txg == 0);
589 601
590 602 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
591 603 object, THT_FREE, off, len);
592 604 if (txh == NULL)
593 605 return;
594 606 dn = txh->txh_dnode;
595 607
596 608 /* first block */
597 609 if (off != 0)
598 610 dmu_tx_count_write(txh, off, 1);
599 611 /* last block */
600 612 if (len != DMU_OBJECT_END)
601 613 dmu_tx_count_write(txh, off+len, 1);
602 614
603 615 dmu_tx_count_dnode(txh);
604 616
605 617 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
606 618 return;
607 619 if (len == DMU_OBJECT_END)
608 620 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
609 621
610 622 /*
611 623 * For i/o error checking, read the first and last level-0
612 624 * blocks, and all the level-1 blocks. The above count_write's
613 625 * have already taken care of the level-0 blocks.
614 626 */
615 627 if (dn->dn_nlevels > 1) {
616 628 shift = dn->dn_datablkshift + dn->dn_indblkshift -
617 629 SPA_BLKPTRSHIFT;
618 630 start = off >> shift;
619 631 end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
620 632
621 633 zio = zio_root(tx->tx_pool->dp_spa,
622 634 NULL, NULL, ZIO_FLAG_CANFAIL);
623 635 for (i = start; i <= end; i++) {
624 636 uint64_t ibyte = i << shift;
625 637 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
626 638 i = ibyte >> shift;
627 639 if (err == ESRCH)
628 640 break;
629 641 if (err) {
630 642 tx->tx_err = err;
631 643 return;
632 644 }
633 645
634 646 err = dmu_tx_check_ioerr(zio, dn, 1, i);
635 647 if (err) {
636 648 tx->tx_err = err;
637 649 return;
638 650 }
639 651 }
640 652 err = zio_wait(zio);
641 653 if (err) {
642 654 tx->tx_err = err;
643 655 return;
644 656 }
645 657 }
646 658
647 659 dmu_tx_count_free(txh, off, len);
648 660 }
649 661
650 662 void
651 663 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
652 664 {
653 665 dmu_tx_hold_t *txh;
654 666 dnode_t *dn;
655 667 uint64_t nblocks;
656 668 int epbs, err;
657 669
658 670 ASSERT(tx->tx_txg == 0);
659 671
660 672 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
661 673 object, THT_ZAP, add, (uintptr_t)name);
662 674 if (txh == NULL)
663 675 return;
664 676 dn = txh->txh_dnode;
665 677
666 678 dmu_tx_count_dnode(txh);
667 679
668 680 if (dn == NULL) {
669 681 /*
670 682 * We will be able to fit a new object's entries into one leaf
671 683 * block. So there will be at most 2 blocks total,
672 684 * including the header block.
673 685 */
674 686 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
675 687 return;
676 688 }
677 689
678 690 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
679 691
680 692 if (dn->dn_maxblkid == 0 && !add) {
681 693 blkptr_t *bp;
682 694
683 695 /*
684 696 * If there is only one block (i.e. this is a micro-zap)
685 697 * and we are not adding anything, the accounting is simple.
686 698 */
687 699 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
688 700 if (err) {
689 701 tx->tx_err = err;
690 702 return;
691 703 }
692 704
693 705 /*
694 706 * Use max block size here, since we don't know how much
695 707 * the size will change between now and the dbuf dirty call.
696 708 */
697 709 bp = &dn->dn_phys->dn_blkptr[0];
698 710 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
699 711 bp, bp->blk_birth))
700 712 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
701 713 else
702 714 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
703 715 if (!BP_IS_HOLE(bp))
704 716 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
705 717 return;
706 718 }
707 719
708 720 if (dn->dn_maxblkid > 0 && name) {
709 721 /*
710 722 * access the name in this fat-zap so that we'll check
711 723 * for i/o errors to the leaf blocks, etc.
712 724 */
713 725 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
714 726 8, 0, NULL);
715 727 if (err == EIO) {
716 728 tx->tx_err = err;
717 729 return;
718 730 }
719 731 }
720 732
721 733 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
722 734 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
723 735
724 736 /*
725 737 * If the modified blocks are scattered to the four winds,
726 738 * we'll have to modify an indirect twig for each.
727 739 */
728 740 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
729 741 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
730 742 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
731 743 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
732 744 else
733 745 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
734 746 }
735 747
736 748 void
737 749 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
738 750 {
739 751 dmu_tx_hold_t *txh;
740 752
741 753 ASSERT(tx->tx_txg == 0);
742 754
743 755 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
744 756 object, THT_BONUS, 0, 0);
745 757 if (txh)
746 758 dmu_tx_count_dnode(txh);
747 759 }
748 760
749 761 void
750 762 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
751 763 {
752 764 dmu_tx_hold_t *txh;
753 765 ASSERT(tx->tx_txg == 0);
754 766
755 767 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
756 768 DMU_NEW_OBJECT, THT_SPACE, space, 0);
757 769
758 770 txh->txh_space_towrite += space;
759 771 }
760 772
761 773 int
762 774 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
763 775 {
764 776 dmu_tx_hold_t *txh;
765 777 int holds = 0;
766 778
767 779 /*
768 780 * By asserting that the tx is assigned, we're counting the
769 781 * number of dn_tx_holds, which is the same as the number of
770 782 * dn_holds. Otherwise, we'd be counting dn_holds, but
771 783 * dn_tx_holds could be 0.
772 784 */
773 785 ASSERT(tx->tx_txg != 0);
774 786
775 787 /* if (tx->tx_anyobj == TRUE) */
776 788 /* return (0); */
777 789
778 790 for (txh = list_head(&tx->tx_holds); txh;
779 791 txh = list_next(&tx->tx_holds, txh)) {
780 792 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
781 793 holds++;
782 794 }
783 795
784 796 return (holds);
785 797 }
786 798
787 799 #ifdef ZFS_DEBUG
788 800 void
789 801 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
790 802 {
791 803 dmu_tx_hold_t *txh;
792 804 int match_object = FALSE, match_offset = FALSE;
793 805 dnode_t *dn;
794 806
795 807 DB_DNODE_ENTER(db);
796 808 dn = DB_DNODE(db);
797 809 ASSERT(tx->tx_txg != 0);
798 810 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
799 811 ASSERT3U(dn->dn_object, ==, db->db.db_object);
800 812
801 813 if (tx->tx_anyobj) {
802 814 DB_DNODE_EXIT(db);
803 815 return;
804 816 }
805 817
806 818 /* XXX No checking on the meta dnode for now */
807 819 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
808 820 DB_DNODE_EXIT(db);
809 821 return;
810 822 }
811 823
812 824 for (txh = list_head(&tx->tx_holds); txh;
813 825 txh = list_next(&tx->tx_holds, txh)) {
814 826 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
815 827 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
816 828 match_object = TRUE;
817 829 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
818 830 int datablkshift = dn->dn_datablkshift ?
819 831 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
820 832 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
821 833 int shift = datablkshift + epbs * db->db_level;
822 834 uint64_t beginblk = shift >= 64 ? 0 :
823 835 (txh->txh_arg1 >> shift);
824 836 uint64_t endblk = shift >= 64 ? 0 :
825 837 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
826 838 uint64_t blkid = db->db_blkid;
827 839
828 840 /* XXX txh_arg2 better not be zero... */
829 841
830 842 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
831 843 txh->txh_type, beginblk, endblk);
832 844
833 845 switch (txh->txh_type) {
834 846 case THT_WRITE:
835 847 if (blkid >= beginblk && blkid <= endblk)
836 848 match_offset = TRUE;
837 849 /*
838 850 * We will let this hold work for the bonus
839 851 * or spill buffer so that we don't need to
840 852 * hold it when creating a new object.
841 853 */
842 854 if (blkid == DMU_BONUS_BLKID ||
843 855 blkid == DMU_SPILL_BLKID)
844 856 match_offset = TRUE;
845 857 /*
846 858 * They might have to increase nlevels,
847 859 * thus dirtying the new TLIBs. Or the
848 860 * might have to change the block size,
849 861 * thus dirying the new lvl=0 blk=0.
850 862 */
851 863 if (blkid == 0)
852 864 match_offset = TRUE;
853 865 break;
854 866 case THT_FREE:
855 867 /*
856 868 * We will dirty all the level 1 blocks in
857 869 * the free range and perhaps the first and
858 870 * last level 0 block.
859 871 */
860 872 if (blkid >= beginblk && (blkid <= endblk ||
861 873 txh->txh_arg2 == DMU_OBJECT_END))
862 874 match_offset = TRUE;
863 875 break;
864 876 case THT_SPILL:
865 877 if (blkid == DMU_SPILL_BLKID)
866 878 match_offset = TRUE;
867 879 break;
868 880 case THT_BONUS:
869 881 if (blkid == DMU_BONUS_BLKID)
870 882 match_offset = TRUE;
871 883 break;
872 884 case THT_ZAP:
873 885 match_offset = TRUE;
874 886 break;
875 887 case THT_NEWOBJECT:
876 888 match_object = TRUE;
877 889 break;
878 890 default:
879 891 ASSERT(!"bad txh_type");
880 892 }
881 893 }
882 894 if (match_object && match_offset) {
883 895 DB_DNODE_EXIT(db);
884 896 return;
885 897 }
886 898 }
887 899 DB_DNODE_EXIT(db);
888 900 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
889 901 (u_longlong_t)db->db.db_object, db->db_level,
890 902 (u_longlong_t)db->db_blkid);
891 903 }
892 904 #endif
893 905
894 906 static int
895 907 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
896 908 {
897 909 dmu_tx_hold_t *txh;
898 910 spa_t *spa = tx->tx_pool->dp_spa;
899 911 uint64_t memory, asize, fsize, usize;
900 912 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
901 913
902 914 ASSERT3U(tx->tx_txg, ==, 0);
903 915
904 916 if (tx->tx_err)
905 917 return (tx->tx_err);
906 918
907 919 if (spa_suspended(spa)) {
908 920 /*
909 921 * If the user has indicated a blocking failure mode
910 922 * then return ERESTART which will block in dmu_tx_wait().
911 923 * Otherwise, return EIO so that an error can get
912 924 * propagated back to the VOP calls.
913 925 *
914 926 * Note that we always honor the txg_how flag regardless
915 927 * of the failuremode setting.
916 928 */
917 929 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
918 930 txg_how != TXG_WAIT)
919 931 return (EIO);
920 932
921 933 return (ERESTART);
922 934 }
923 935
924 936 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
925 937 tx->tx_needassign_txh = NULL;
926 938
927 939 /*
928 940 * NB: No error returns are allowed after txg_hold_open, but
929 941 * before processing the dnode holds, due to the
930 942 * dmu_tx_unassign() logic.
931 943 */
932 944
933 945 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
934 946 for (txh = list_head(&tx->tx_holds); txh;
935 947 txh = list_next(&tx->tx_holds, txh)) {
936 948 dnode_t *dn = txh->txh_dnode;
937 949 if (dn != NULL) {
938 950 mutex_enter(&dn->dn_mtx);
939 951 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
940 952 mutex_exit(&dn->dn_mtx);
941 953 tx->tx_needassign_txh = txh;
942 954 return (ERESTART);
943 955 }
944 956 if (dn->dn_assigned_txg == 0)
945 957 dn->dn_assigned_txg = tx->tx_txg;
946 958 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
947 959 (void) refcount_add(&dn->dn_tx_holds, tx);
948 960 mutex_exit(&dn->dn_mtx);
949 961 }
950 962 towrite += txh->txh_space_towrite;
951 963 tofree += txh->txh_space_tofree;
952 964 tooverwrite += txh->txh_space_tooverwrite;
953 965 tounref += txh->txh_space_tounref;
954 966 tohold += txh->txh_memory_tohold;
955 967 fudge += txh->txh_fudge;
956 968 }
957 969
958 970 /*
959 971 * NB: This check must be after we've held the dnodes, so that
960 972 * the dmu_tx_unassign() logic will work properly
961 973 */
962 974 if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
963 975 return (ERESTART);
964 976
965 977 /*
966 978 * If a snapshot has been taken since we made our estimates,
967 979 * assume that we won't be able to free or overwrite anything.
968 980 */
969 981 if (tx->tx_objset &&
970 982 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
971 983 tx->tx_lastsnap_txg) {
972 984 towrite += tooverwrite;
973 985 tooverwrite = tofree = 0;
974 986 }
975 987
976 988 /* needed allocation: worst-case estimate of write space */
977 989 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
978 990 /* freed space estimate: worst-case overwrite + free estimate */
979 991 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
980 992 /* convert unrefd space to worst-case estimate */
981 993 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
982 994 /* calculate memory footprint estimate */
983 995 memory = towrite + tooverwrite + tohold;
984 996
985 997 #ifdef ZFS_DEBUG
986 998 /*
987 999 * Add in 'tohold' to account for our dirty holds on this memory
988 1000 * XXX - the "fudge" factor is to account for skipped blocks that
989 1001 * we missed because dnode_next_offset() misses in-core-only blocks.
990 1002 */
991 1003 tx->tx_space_towrite = asize +
992 1004 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
993 1005 tx->tx_space_tofree = tofree;
994 1006 tx->tx_space_tooverwrite = tooverwrite;
995 1007 tx->tx_space_tounref = tounref;
996 1008 #endif
997 1009
998 1010 if (tx->tx_dir && asize != 0) {
999 1011 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1000 1012 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1001 1013 if (err)
1002 1014 return (err);
1003 1015 }
1004 1016
1005 1017 return (0);
1006 1018 }
1007 1019
1008 1020 static void
1009 1021 dmu_tx_unassign(dmu_tx_t *tx)
1010 1022 {
1011 1023 dmu_tx_hold_t *txh;
1012 1024
1013 1025 if (tx->tx_txg == 0)
1014 1026 return;
1015 1027
1016 1028 txg_rele_to_quiesce(&tx->tx_txgh);
1017 1029
1018 1030 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1019 1031 txh = list_next(&tx->tx_holds, txh)) {
1020 1032 dnode_t *dn = txh->txh_dnode;
1021 1033
1022 1034 if (dn == NULL)
1023 1035 continue;
1024 1036 mutex_enter(&dn->dn_mtx);
1025 1037 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1026 1038
1027 1039 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1028 1040 dn->dn_assigned_txg = 0;
1029 1041 cv_broadcast(&dn->dn_notxholds);
1030 1042 }
1031 1043 mutex_exit(&dn->dn_mtx);
1032 1044 }
1033 1045
1034 1046 txg_rele_to_sync(&tx->tx_txgh);
1035 1047
1036 1048 tx->tx_lasttried_txg = tx->tx_txg;
1037 1049 tx->tx_txg = 0;
1038 1050 }
1039 1051
1040 1052 /*
1041 1053 * Assign tx to a transaction group. txg_how can be one of:
1042 1054 *
1043 1055 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1044 1056 * a new one. This should be used when you're not holding locks.
1045 1057 * If will only fail if we're truly out of space (or over quota).
1046 1058 *
1047 1059 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1048 1060 * blocking, returns immediately with ERESTART. This should be used
1049 1061 * whenever you're holding locks. On an ERESTART error, the caller
1050 1062 * should drop locks, do a dmu_tx_wait(tx), and try again.
1051 1063 *
1052 1064 * (3) A specific txg. Use this if you need to ensure that multiple
1053 1065 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1054 1066 * returns ERESTART if it can't assign you into the requested txg.
1055 1067 */
1056 1068 int
1057 1069 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1058 1070 {
1059 1071 int err;
1060 1072
1061 1073 ASSERT(tx->tx_txg == 0);
1062 1074 ASSERT(txg_how != 0);
1063 1075 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1064 1076
1065 1077 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1066 1078 dmu_tx_unassign(tx);
1067 1079
1068 1080 if (err != ERESTART || txg_how != TXG_WAIT)
1069 1081 return (err);
1070 1082
1071 1083 dmu_tx_wait(tx);
1072 1084 }
1073 1085
1074 1086 txg_rele_to_quiesce(&tx->tx_txgh);
1075 1087
1076 1088 return (0);
1077 1089 }
1078 1090
1079 1091 void
1080 1092 dmu_tx_wait(dmu_tx_t *tx)
1081 1093 {
1082 1094 spa_t *spa = tx->tx_pool->dp_spa;
1083 1095
1084 1096 ASSERT(tx->tx_txg == 0);
1085 1097
1086 1098 /*
1087 1099 * It's possible that the pool has become active after this thread
1088 1100 * has tried to obtain a tx. If that's the case then his
1089 1101 * tx_lasttried_txg would not have been assigned.
1090 1102 */
1091 1103 if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1092 1104 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1093 1105 } else if (tx->tx_needassign_txh) {
1094 1106 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1095 1107
1096 1108 mutex_enter(&dn->dn_mtx);
1097 1109 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1098 1110 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1099 1111 mutex_exit(&dn->dn_mtx);
1100 1112 tx->tx_needassign_txh = NULL;
1101 1113 } else {
1102 1114 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1103 1115 }
1104 1116 }
1105 1117
1106 1118 void
1107 1119 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1108 1120 {
1109 1121 #ifdef ZFS_DEBUG
1110 1122 if (tx->tx_dir == NULL || delta == 0)
1111 1123 return;
1112 1124
1113 1125 if (delta > 0) {
1114 1126 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1115 1127 tx->tx_space_towrite);
1116 1128 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1117 1129 } else {
1118 1130 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1119 1131 }
1120 1132 #endif
1121 1133 }
1122 1134
1123 1135 void
1124 1136 dmu_tx_commit(dmu_tx_t *tx)
1125 1137 {
1126 1138 dmu_tx_hold_t *txh;
1127 1139
1128 1140 ASSERT(tx->tx_txg != 0);
1129 1141
1130 1142 while (txh = list_head(&tx->tx_holds)) {
1131 1143 dnode_t *dn = txh->txh_dnode;
1132 1144
1133 1145 list_remove(&tx->tx_holds, txh);
1134 1146 kmem_free(txh, sizeof (dmu_tx_hold_t));
1135 1147 if (dn == NULL)
1136 1148 continue;
1137 1149 mutex_enter(&dn->dn_mtx);
1138 1150 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1139 1151
1140 1152 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1141 1153 dn->dn_assigned_txg = 0;
1142 1154 cv_broadcast(&dn->dn_notxholds);
1143 1155 }
1144 1156 mutex_exit(&dn->dn_mtx);
1145 1157 dnode_rele(dn, tx);
1146 1158 }
1147 1159
1148 1160 if (tx->tx_tempreserve_cookie)
1149 1161 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1150 1162
1151 1163 if (!list_is_empty(&tx->tx_callbacks))
1152 1164 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1153 1165
1154 1166 if (tx->tx_anyobj == FALSE)
1155 1167 txg_rele_to_sync(&tx->tx_txgh);
1156 1168
1157 1169 list_destroy(&tx->tx_callbacks);
1158 1170 list_destroy(&tx->tx_holds);
1159 1171 #ifdef ZFS_DEBUG
1160 1172 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1161 1173 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1162 1174 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1163 1175 refcount_destroy_many(&tx->tx_space_written,
1164 1176 refcount_count(&tx->tx_space_written));
1165 1177 refcount_destroy_many(&tx->tx_space_freed,
1166 1178 refcount_count(&tx->tx_space_freed));
1167 1179 #endif
1168 1180 kmem_free(tx, sizeof (dmu_tx_t));
1169 1181 }
1170 1182
1171 1183 void
1172 1184 dmu_tx_abort(dmu_tx_t *tx)
1173 1185 {
1174 1186 dmu_tx_hold_t *txh;
1175 1187
1176 1188 ASSERT(tx->tx_txg == 0);
1177 1189
1178 1190 while (txh = list_head(&tx->tx_holds)) {
1179 1191 dnode_t *dn = txh->txh_dnode;
1180 1192
1181 1193 list_remove(&tx->tx_holds, txh);
1182 1194 kmem_free(txh, sizeof (dmu_tx_hold_t));
1183 1195 if (dn != NULL)
1184 1196 dnode_rele(dn, tx);
1185 1197 }
1186 1198
1187 1199 /*
1188 1200 * Call any registered callbacks with an error code.
1189 1201 */
1190 1202 if (!list_is_empty(&tx->tx_callbacks))
1191 1203 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1192 1204
1193 1205 list_destroy(&tx->tx_callbacks);
1194 1206 list_destroy(&tx->tx_holds);
1195 1207 #ifdef ZFS_DEBUG
1196 1208 refcount_destroy_many(&tx->tx_space_written,
1197 1209 refcount_count(&tx->tx_space_written));
1198 1210 refcount_destroy_many(&tx->tx_space_freed,
1199 1211 refcount_count(&tx->tx_space_freed));
1200 1212 #endif
1201 1213 kmem_free(tx, sizeof (dmu_tx_t));
1202 1214 }
1203 1215
1204 1216 uint64_t
1205 1217 dmu_tx_get_txg(dmu_tx_t *tx)
1206 1218 {
1207 1219 ASSERT(tx->tx_txg != 0);
1208 1220 return (tx->tx_txg);
1209 1221 }
1210 1222
1211 1223 void
1212 1224 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1213 1225 {
1214 1226 dmu_tx_callback_t *dcb;
1215 1227
1216 1228 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1217 1229
1218 1230 dcb->dcb_func = func;
1219 1231 dcb->dcb_data = data;
1220 1232
1221 1233 list_insert_tail(&tx->tx_callbacks, dcb);
1222 1234 }
1223 1235
1224 1236 /*
1225 1237 * Call all the commit callbacks on a list, with a given error code.
1226 1238 */
1227 1239 void
1228 1240 dmu_tx_do_callbacks(list_t *cb_list, int error)
1229 1241 {
1230 1242 dmu_tx_callback_t *dcb;
1231 1243
1232 1244 while (dcb = list_head(cb_list)) {
1233 1245 list_remove(cb_list, dcb);
1234 1246 dcb->dcb_func(dcb->dcb_data, error);
1235 1247 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1236 1248 }
1237 1249 }
1238 1250
1239 1251 /*
1240 1252 * Interface to hold a bunch of attributes.
1241 1253 * used for creating new files.
1242 1254 * attrsize is the total size of all attributes
1243 1255 * to be added during object creation
1244 1256 *
1245 1257 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1246 1258 */
1247 1259
1248 1260 /*
1249 1261 * hold necessary attribute name for attribute registration.
1250 1262 * should be a very rare case where this is needed. If it does
1251 1263 * happen it would only happen on the first write to the file system.
1252 1264 */
1253 1265 static void
1254 1266 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1255 1267 {
1256 1268 int i;
1257 1269
1258 1270 if (!sa->sa_need_attr_registration)
1259 1271 return;
1260 1272
1261 1273 for (i = 0; i != sa->sa_num_attrs; i++) {
1262 1274 if (!sa->sa_attr_table[i].sa_registered) {
1263 1275 if (sa->sa_reg_attr_obj)
1264 1276 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1265 1277 B_TRUE, sa->sa_attr_table[i].sa_name);
1266 1278 else
1267 1279 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1268 1280 B_TRUE, sa->sa_attr_table[i].sa_name);
1269 1281 }
1270 1282 }
1271 1283 }
1272 1284
1273 1285
1274 1286 void
1275 1287 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1276 1288 {
1277 1289 dnode_t *dn;
1278 1290 dmu_tx_hold_t *txh;
1279 1291
1280 1292 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1281 1293 THT_SPILL, 0, 0);
1282 1294
1283 1295 dn = txh->txh_dnode;
1284 1296
1285 1297 if (dn == NULL)
1286 1298 return;
1287 1299
1288 1300 /* If blkptr doesn't exist then add space to towrite */
1289 1301 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1290 1302 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1291 1303 } else {
1292 1304 blkptr_t *bp;
1293 1305
1294 1306 bp = &dn->dn_phys->dn_spill;
1295 1307 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1296 1308 bp, bp->blk_birth))
1297 1309 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1298 1310 else
1299 1311 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1300 1312 if (!BP_IS_HOLE(bp))
1301 1313 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1302 1314 }
1303 1315 }
1304 1316
1305 1317 void
1306 1318 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1307 1319 {
1308 1320 sa_os_t *sa = tx->tx_objset->os_sa;
1309 1321
1310 1322 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1311 1323
1312 1324 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1313 1325 return;
1314 1326
1315 1327 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1316 1328 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1317 1329 else {
1318 1330 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1319 1331 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1320 1332 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1321 1333 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1322 1334 }
1323 1335
1324 1336 dmu_tx_sa_registration_hold(sa, tx);
1325 1337
1326 1338 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1327 1339 return;
1328 1340
1329 1341 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1330 1342 THT_SPILL, 0, 0);
1331 1343 }
1332 1344
1333 1345 /*
1334 1346 * Hold SA attribute
1335 1347 *
1336 1348 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1337 1349 *
1338 1350 * variable_size is the total size of all variable sized attributes
1339 1351 * passed to this function. It is not the total size of all
1340 1352 * variable size attributes that *may* exist on this object.
1341 1353 */
1342 1354 void
1343 1355 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1344 1356 {
1345 1357 uint64_t object;
1346 1358 sa_os_t *sa = tx->tx_objset->os_sa;
1347 1359
1348 1360 ASSERT(hdl != NULL);
1349 1361
1350 1362 object = sa_handle_object(hdl);
1351 1363
1352 1364 dmu_tx_hold_bonus(tx, object);
1353 1365
1354 1366 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1355 1367 return;
1356 1368
1357 1369 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1358 1370 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1359 1371 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1360 1372 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1361 1373 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1362 1374 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1363 1375 }
1364 1376
1365 1377 dmu_tx_sa_registration_hold(sa, tx);
1366 1378
1367 1379 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1368 1380 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1369 1381
1370 1382 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1371 1383 ASSERT(tx->tx_txg == 0);
1372 1384 dmu_tx_hold_spill(tx, object);
1373 1385 } else {
1374 1386 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1375 1387 dnode_t *dn;
1376 1388
1377 1389 DB_DNODE_ENTER(db);
1378 1390 dn = DB_DNODE(db);
1379 1391 if (dn->dn_have_spill) {
1380 1392 ASSERT(tx->tx_txg == 0);
1381 1393 dmu_tx_hold_spill(tx, object);
1382 1394 }
1383 1395 DB_DNODE_EXIT(db);
1384 1396 }
1385 1397 }
↓ open down ↓ |
807 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX