Print this page
4082 zfs receive gets EFBIG from dmu_tx_hold_free()
Reviewed by: Eric Schrock <eric.schrock@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
| 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) 2013 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 tx->tx_start = gethrtime();
58 58 #ifdef ZFS_DEBUG
59 59 refcount_create(&tx->tx_space_written);
60 60 refcount_create(&tx->tx_space_freed);
61 61 #endif
62 62 return (tx);
63 63 }
64 64
65 65 dmu_tx_t *
66 66 dmu_tx_create(objset_t *os)
67 67 {
68 68 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
69 69 tx->tx_objset = os;
70 70 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
71 71 return (tx);
72 72 }
73 73
74 74 dmu_tx_t *
75 75 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
76 76 {
77 77 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
78 78
79 79 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
80 80 tx->tx_pool = dp;
81 81 tx->tx_txg = txg;
82 82 tx->tx_anyobj = TRUE;
83 83
84 84 return (tx);
85 85 }
86 86
87 87 int
88 88 dmu_tx_is_syncing(dmu_tx_t *tx)
89 89 {
90 90 return (tx->tx_anyobj);
91 91 }
92 92
93 93 int
94 94 dmu_tx_private_ok(dmu_tx_t *tx)
95 95 {
96 96 return (tx->tx_anyobj);
97 97 }
98 98
99 99 static dmu_tx_hold_t *
100 100 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
101 101 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
102 102 {
103 103 dmu_tx_hold_t *txh;
104 104 dnode_t *dn = NULL;
105 105 int err;
106 106
107 107 if (object != DMU_NEW_OBJECT) {
108 108 err = dnode_hold(os, object, tx, &dn);
109 109 if (err) {
110 110 tx->tx_err = err;
111 111 return (NULL);
112 112 }
113 113
114 114 if (err == 0 && tx->tx_txg != 0) {
115 115 mutex_enter(&dn->dn_mtx);
116 116 /*
117 117 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
118 118 * problem, but there's no way for it to happen (for
119 119 * now, at least).
120 120 */
121 121 ASSERT(dn->dn_assigned_txg == 0);
122 122 dn->dn_assigned_txg = tx->tx_txg;
123 123 (void) refcount_add(&dn->dn_tx_holds, tx);
124 124 mutex_exit(&dn->dn_mtx);
125 125 }
126 126 }
127 127
128 128 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
129 129 txh->txh_tx = tx;
130 130 txh->txh_dnode = dn;
131 131 #ifdef ZFS_DEBUG
132 132 txh->txh_type = type;
133 133 txh->txh_arg1 = arg1;
134 134 txh->txh_arg2 = arg2;
135 135 #endif
136 136 list_insert_tail(&tx->tx_holds, txh);
137 137
138 138 return (txh);
139 139 }
140 140
141 141 void
142 142 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
143 143 {
144 144 /*
145 145 * If we're syncing, they can manipulate any object anyhow, and
146 146 * the hold on the dnode_t can cause problems.
147 147 */
148 148 if (!dmu_tx_is_syncing(tx)) {
149 149 (void) dmu_tx_hold_object_impl(tx, os,
150 150 object, THT_NEWOBJECT, 0, 0);
151 151 }
152 152 }
153 153
154 154 static int
155 155 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
156 156 {
157 157 int err;
158 158 dmu_buf_impl_t *db;
159 159
160 160 rw_enter(&dn->dn_struct_rwlock, RW_READER);
161 161 db = dbuf_hold_level(dn, level, blkid, FTAG);
162 162 rw_exit(&dn->dn_struct_rwlock);
163 163 if (db == NULL)
164 164 return (SET_ERROR(EIO));
165 165 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
166 166 dbuf_rele(db, FTAG);
167 167 return (err);
168 168 }
169 169
170 170 static void
171 171 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
172 172 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
173 173 {
174 174 objset_t *os = dn->dn_objset;
175 175 dsl_dataset_t *ds = os->os_dsl_dataset;
176 176 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
177 177 dmu_buf_impl_t *parent = NULL;
178 178 blkptr_t *bp = NULL;
179 179 uint64_t space;
180 180
181 181 if (level >= dn->dn_nlevels || history[level] == blkid)
182 182 return;
183 183
184 184 history[level] = blkid;
185 185
186 186 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
187 187
188 188 if (db == NULL || db == dn->dn_dbuf) {
189 189 ASSERT(level != 0);
190 190 db = NULL;
191 191 } else {
192 192 ASSERT(DB_DNODE(db) == dn);
193 193 ASSERT(db->db_level == level);
194 194 ASSERT(db->db.db_size == space);
195 195 ASSERT(db->db_blkid == blkid);
196 196 bp = db->db_blkptr;
197 197 parent = db->db_parent;
198 198 }
199 199
200 200 freeable = (bp && (freeable ||
201 201 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
202 202
203 203 if (freeable)
204 204 txh->txh_space_tooverwrite += space;
205 205 else
206 206 txh->txh_space_towrite += space;
207 207 if (bp)
208 208 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
209 209
210 210 dmu_tx_count_twig(txh, dn, parent, level + 1,
211 211 blkid >> epbs, freeable, history);
212 212 }
213 213
214 214 /* ARGSUSED */
215 215 static void
216 216 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
217 217 {
218 218 dnode_t *dn = txh->txh_dnode;
219 219 uint64_t start, end, i;
220 220 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
221 221 int err = 0;
222 222
223 223 if (len == 0)
224 224 return;
225 225
226 226 min_bs = SPA_MINBLOCKSHIFT;
227 227 max_bs = SPA_MAXBLOCKSHIFT;
228 228 min_ibs = DN_MIN_INDBLKSHIFT;
229 229 max_ibs = DN_MAX_INDBLKSHIFT;
230 230
231 231 if (dn) {
232 232 uint64_t history[DN_MAX_LEVELS];
233 233 int nlvls = dn->dn_nlevels;
234 234 int delta;
235 235
236 236 /*
237 237 * For i/o error checking, read the first and last level-0
238 238 * blocks (if they are not aligned), and all the level-1 blocks.
239 239 */
240 240 if (dn->dn_maxblkid == 0) {
241 241 delta = dn->dn_datablksz;
242 242 start = (off < dn->dn_datablksz) ? 0 : 1;
243 243 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
244 244 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
245 245 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
246 246 if (err)
247 247 goto out;
248 248 delta -= off;
249 249 }
250 250 } else {
251 251 zio_t *zio = zio_root(dn->dn_objset->os_spa,
252 252 NULL, NULL, ZIO_FLAG_CANFAIL);
253 253
254 254 /* first level-0 block */
255 255 start = off >> dn->dn_datablkshift;
256 256 if (P2PHASE(off, dn->dn_datablksz) ||
257 257 len < dn->dn_datablksz) {
258 258 err = dmu_tx_check_ioerr(zio, dn, 0, start);
259 259 if (err)
260 260 goto out;
261 261 }
262 262
263 263 /* last level-0 block */
264 264 end = (off+len-1) >> dn->dn_datablkshift;
265 265 if (end != start && end <= dn->dn_maxblkid &&
266 266 P2PHASE(off+len, dn->dn_datablksz)) {
267 267 err = dmu_tx_check_ioerr(zio, dn, 0, end);
268 268 if (err)
269 269 goto out;
270 270 }
271 271
272 272 /* level-1 blocks */
273 273 if (nlvls > 1) {
274 274 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
275 275 for (i = (start>>shft)+1; i < end>>shft; i++) {
276 276 err = dmu_tx_check_ioerr(zio, dn, 1, i);
277 277 if (err)
278 278 goto out;
279 279 }
280 280 }
281 281
282 282 err = zio_wait(zio);
283 283 if (err)
284 284 goto out;
285 285 delta = P2NPHASE(off, dn->dn_datablksz);
286 286 }
287 287
288 288 min_ibs = max_ibs = dn->dn_indblkshift;
289 289 if (dn->dn_maxblkid > 0) {
290 290 /*
291 291 * The blocksize can't change,
292 292 * so we can make a more precise estimate.
293 293 */
294 294 ASSERT(dn->dn_datablkshift != 0);
295 295 min_bs = max_bs = dn->dn_datablkshift;
296 296 }
297 297
298 298 /*
299 299 * If this write is not off the end of the file
300 300 * we need to account for overwrites/unref.
301 301 */
302 302 if (start <= dn->dn_maxblkid) {
303 303 for (int l = 0; l < DN_MAX_LEVELS; l++)
304 304 history[l] = -1ULL;
305 305 }
306 306 while (start <= dn->dn_maxblkid) {
307 307 dmu_buf_impl_t *db;
308 308
309 309 rw_enter(&dn->dn_struct_rwlock, RW_READER);
310 310 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
311 311 rw_exit(&dn->dn_struct_rwlock);
312 312
313 313 if (err) {
314 314 txh->txh_tx->tx_err = err;
315 315 return;
316 316 }
317 317
318 318 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
319 319 history);
320 320 dbuf_rele(db, FTAG);
321 321 if (++start > end) {
322 322 /*
323 323 * Account for new indirects appearing
324 324 * before this IO gets assigned into a txg.
325 325 */
326 326 bits = 64 - min_bs;
327 327 epbs = min_ibs - SPA_BLKPTRSHIFT;
328 328 for (bits -= epbs * (nlvls - 1);
329 329 bits >= 0; bits -= epbs)
330 330 txh->txh_fudge += 1ULL << max_ibs;
331 331 goto out;
332 332 }
333 333 off += delta;
334 334 if (len >= delta)
335 335 len -= delta;
336 336 delta = dn->dn_datablksz;
337 337 }
338 338 }
339 339
340 340 /*
341 341 * 'end' is the last thing we will access, not one past.
342 342 * This way we won't overflow when accessing the last byte.
343 343 */
344 344 start = P2ALIGN(off, 1ULL << max_bs);
345 345 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
346 346 txh->txh_space_towrite += end - start + 1;
347 347
348 348 start >>= min_bs;
349 349 end >>= min_bs;
350 350
351 351 epbs = min_ibs - SPA_BLKPTRSHIFT;
352 352
353 353 /*
354 354 * The object contains at most 2^(64 - min_bs) blocks,
355 355 * and each indirect level maps 2^epbs.
356 356 */
357 357 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
358 358 start >>= epbs;
359 359 end >>= epbs;
360 360 ASSERT3U(end, >=, start);
361 361 txh->txh_space_towrite += (end - start + 1) << max_ibs;
362 362 if (start != 0) {
363 363 /*
364 364 * We also need a new blkid=0 indirect block
365 365 * to reference any existing file data.
366 366 */
367 367 txh->txh_space_towrite += 1ULL << max_ibs;
368 368 }
369 369 }
370 370
371 371 out:
372 372 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
373 373 2 * DMU_MAX_ACCESS)
374 374 err = SET_ERROR(EFBIG);
375 375
376 376 if (err)
377 377 txh->txh_tx->tx_err = err;
378 378 }
379 379
380 380 static void
381 381 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
382 382 {
383 383 dnode_t *dn = txh->txh_dnode;
384 384 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
385 385 uint64_t space = mdn->dn_datablksz +
386 386 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
387 387
388 388 if (dn && dn->dn_dbuf->db_blkptr &&
389 389 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
390 390 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
391 391 txh->txh_space_tooverwrite += space;
392 392 txh->txh_space_tounref += space;
393 393 } else {
394 394 txh->txh_space_towrite += space;
395 395 if (dn && dn->dn_dbuf->db_blkptr)
396 396 txh->txh_space_tounref += space;
397 397 }
398 398 }
399 399
400 400 void
401 401 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
402 402 {
403 403 dmu_tx_hold_t *txh;
404 404
405 405 ASSERT(tx->tx_txg == 0);
406 406 ASSERT(len < DMU_MAX_ACCESS);
407 407 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
408 408
409 409 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
410 410 object, THT_WRITE, off, len);
411 411 if (txh == NULL)
412 412 return;
413 413
414 414 dmu_tx_count_write(txh, off, len);
415 415 dmu_tx_count_dnode(txh);
416 416 }
417 417
418 418 static void
419 419 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
420 420 {
421 421 uint64_t blkid, nblks, lastblk;
422 422 uint64_t space = 0, unref = 0, skipped = 0;
423 423 dnode_t *dn = txh->txh_dnode;
424 424 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
425 425 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
426 426 int epbs;
427 427 uint64_t l0span = 0, nl1blks = 0;
428 428
429 429 if (dn->dn_nlevels == 0)
430 430 return;
431 431
432 432 /*
433 433 * The struct_rwlock protects us against dn_nlevels
434 434 * changing, in case (against all odds) we manage to dirty &
435 435 * sync out the changes after we check for being dirty.
436 436 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
437 437 */
438 438 rw_enter(&dn->dn_struct_rwlock, RW_READER);
439 439 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
440 440 if (dn->dn_maxblkid == 0) {
441 441 if (off == 0 && len >= dn->dn_datablksz) {
442 442 blkid = 0;
443 443 nblks = 1;
444 444 } else {
445 445 rw_exit(&dn->dn_struct_rwlock);
446 446 return;
447 447 }
448 448 } else {
449 449 blkid = off >> dn->dn_datablkshift;
450 450 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
451 451
452 452 if (blkid > dn->dn_maxblkid) {
453 453 rw_exit(&dn->dn_struct_rwlock);
454 454 return;
455 455 }
456 456 if (blkid + nblks > dn->dn_maxblkid)
457 457 nblks = dn->dn_maxblkid - blkid + 1;
458 458
459 459 }
460 460 l0span = nblks; /* save for later use to calc level > 1 overhead */
461 461 if (dn->dn_nlevels == 1) {
462 462 int i;
463 463 for (i = 0; i < nblks; i++) {
464 464 blkptr_t *bp = dn->dn_phys->dn_blkptr;
465 465 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
466 466 bp += blkid + i;
467 467 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
468 468 dprintf_bp(bp, "can free old%s", "");
469 469 space += bp_get_dsize(spa, bp);
470 470 }
471 471 unref += BP_GET_ASIZE(bp);
472 472 }
473 473 nl1blks = 1;
474 474 nblks = 0;
475 475 }
476 476
477 477 lastblk = blkid + nblks - 1;
478 478 while (nblks) {
479 479 dmu_buf_impl_t *dbuf;
480 480 uint64_t ibyte, new_blkid;
481 481 int epb = 1 << epbs;
482 482 int err, i, blkoff, tochk;
483 483 blkptr_t *bp;
484 484
485 485 ibyte = blkid << dn->dn_datablkshift;
486 486 err = dnode_next_offset(dn,
487 487 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
488 488 new_blkid = ibyte >> dn->dn_datablkshift;
489 489 if (err == ESRCH) {
490 490 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
491 491 break;
492 492 }
493 493 if (err) {
494 494 txh->txh_tx->tx_err = err;
495 495 break;
496 496 }
497 497 if (new_blkid > lastblk) {
498 498 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
499 499 break;
500 500 }
501 501
502 502 if (new_blkid > blkid) {
503 503 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
504 504 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
505 505 nblks -= new_blkid - blkid;
506 506 blkid = new_blkid;
507 507 }
508 508 blkoff = P2PHASE(blkid, epb);
509 509 tochk = MIN(epb - blkoff, nblks);
510 510
511 511 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
512 512 if (err) {
513 513 txh->txh_tx->tx_err = err;
514 514 break;
515 515 }
516 516
517 517 txh->txh_memory_tohold += dbuf->db.db_size;
518 518
519 519 /*
520 520 * We don't check memory_tohold against DMU_MAX_ACCESS because
521 521 * memory_tohold is an over-estimation (especially the >L1
522 522 * indirect blocks), so it could fail. Callers should have
523 523 * already verified that they will not be holding too much
524 524 * memory.
525 525 */
526 526
527 527 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
528 528 if (err != 0) {
529 529 txh->txh_tx->tx_err = err;
530 530 dbuf_rele(dbuf, FTAG);
531 531 break;
532 532 }
533 533
534 534 bp = dbuf->db.db_data;
535 535 bp += blkoff;
536 536
537 537 for (i = 0; i < tochk; i++) {
538 538 if (dsl_dataset_block_freeable(ds, &bp[i],
539 539 bp[i].blk_birth)) {
540 540 dprintf_bp(&bp[i], "can free old%s", "");
541 541 space += bp_get_dsize(spa, &bp[i]);
542 542 }
543 543 unref += BP_GET_ASIZE(bp);
544 544 }
545 545 dbuf_rele(dbuf, FTAG);
546 546
547 547 ++nl1blks;
548 548 blkid += tochk;
549 549 nblks -= tochk;
550 550 }
551 551 rw_exit(&dn->dn_struct_rwlock);
552 552
553 553 /*
554 554 * Add in memory requirements of higher-level indirects.
555 555 * This assumes a worst-possible scenario for dn_nlevels and a
556 556 * worst-possible distribution of l1-blocks over the region to free.
557 557 */
558 558 {
559 559 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
560 560 int level = 2;
561 561 /*
562 562 * Here we don't use DN_MAX_LEVEL, but calculate it with the
563 563 * given datablkshift and indblkshift. This makes the
564 564 * difference between 19 and 8 on large files.
565 565 */
566 566 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
567 567 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
568 568
569 569 while (level++ < maxlevel) {
570 570 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
571 571 << dn->dn_indblkshift;
572 572 blkcnt = 1 + (blkcnt >> epbs);
573 573 }
574 574 }
575 575
576 576 /* account for new level 1 indirect blocks that might show up */
577 577 if (skipped > 0) {
578 578 txh->txh_fudge += skipped << dn->dn_indblkshift;
579 579 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
580 580 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
581 581 }
582 582 txh->txh_space_tofree += space;
583 583 txh->txh_space_tounref += unref;
584 584 }
585 585
586 586 void
587 587 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
588 588 {
589 589 dmu_tx_hold_t *txh;
590 590 dnode_t *dn;
591 591 int err;
592 592 zio_t *zio;
593 593
594 594 ASSERT(tx->tx_txg == 0);
595 595
596 596 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
597 597 object, THT_FREE, off, len);
|
↓ open down ↓ |
597 lines elided |
↑ open up ↑ |
598 598 if (txh == NULL)
599 599 return;
600 600 dn = txh->txh_dnode;
601 601 dmu_tx_count_dnode(txh);
602 602
603 603 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
604 604 return;
605 605 if (len == DMU_OBJECT_END)
606 606 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
607 607
608 -
609 608 /*
610 609 * For i/o error checking, we read the first and last level-0
611 610 * blocks if they are not aligned, and all the level-1 blocks.
612 611 *
613 612 * Note: dbuf_free_range() assumes that we have not instantiated
614 613 * any level-0 dbufs that will be completely freed. Therefore we must
615 614 * exercise care to not read or count the first and last blocks
616 615 * if they are blocksize-aligned.
617 616 */
618 617 if (dn->dn_datablkshift == 0) {
619 618 if (off != 0 || len < dn->dn_datablksz)
620 - dmu_tx_count_write(txh, off, len);
619 + dmu_tx_count_write(txh, 0, dn->dn_datablksz);
621 620 } else {
622 621 /* first block will be modified if it is not aligned */
623 622 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
624 623 dmu_tx_count_write(txh, off, 1);
625 624 /* last block will be modified if it is not aligned */
626 625 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
627 626 dmu_tx_count_write(txh, off+len, 1);
628 627 }
629 628
630 629 /*
631 630 * Check level-1 blocks.
632 631 */
633 632 if (dn->dn_nlevels > 1) {
634 633 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
635 634 SPA_BLKPTRSHIFT;
636 635 uint64_t start = off >> shift;
637 636 uint64_t end = (off + len) >> shift;
638 637
639 638 ASSERT(dn->dn_datablkshift != 0);
640 639 ASSERT(dn->dn_indblkshift != 0);
641 640
642 641 zio = zio_root(tx->tx_pool->dp_spa,
643 642 NULL, NULL, ZIO_FLAG_CANFAIL);
644 643 for (uint64_t i = start; i <= end; i++) {
645 644 uint64_t ibyte = i << shift;
646 645 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
647 646 i = ibyte >> shift;
648 647 if (err == ESRCH)
649 648 break;
650 649 if (err) {
651 650 tx->tx_err = err;
652 651 return;
653 652 }
654 653
655 654 err = dmu_tx_check_ioerr(zio, dn, 1, i);
656 655 if (err) {
657 656 tx->tx_err = err;
658 657 return;
659 658 }
660 659 }
661 660 err = zio_wait(zio);
662 661 if (err) {
663 662 tx->tx_err = err;
664 663 return;
665 664 }
666 665 }
667 666
668 667 dmu_tx_count_free(txh, off, len);
669 668 }
670 669
671 670 void
672 671 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
673 672 {
674 673 dmu_tx_hold_t *txh;
675 674 dnode_t *dn;
676 675 uint64_t nblocks;
677 676 int epbs, err;
678 677
679 678 ASSERT(tx->tx_txg == 0);
680 679
681 680 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
682 681 object, THT_ZAP, add, (uintptr_t)name);
683 682 if (txh == NULL)
684 683 return;
685 684 dn = txh->txh_dnode;
686 685
687 686 dmu_tx_count_dnode(txh);
688 687
689 688 if (dn == NULL) {
690 689 /*
691 690 * We will be able to fit a new object's entries into one leaf
692 691 * block. So there will be at most 2 blocks total,
693 692 * including the header block.
694 693 */
695 694 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
696 695 return;
697 696 }
698 697
699 698 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
700 699
701 700 if (dn->dn_maxblkid == 0 && !add) {
702 701 blkptr_t *bp;
703 702
704 703 /*
705 704 * If there is only one block (i.e. this is a micro-zap)
706 705 * and we are not adding anything, the accounting is simple.
707 706 */
708 707 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
709 708 if (err) {
710 709 tx->tx_err = err;
711 710 return;
712 711 }
713 712
714 713 /*
715 714 * Use max block size here, since we don't know how much
716 715 * the size will change between now and the dbuf dirty call.
717 716 */
718 717 bp = &dn->dn_phys->dn_blkptr[0];
719 718 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
720 719 bp, bp->blk_birth))
721 720 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
722 721 else
723 722 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
724 723 if (!BP_IS_HOLE(bp))
725 724 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
726 725 return;
727 726 }
728 727
729 728 if (dn->dn_maxblkid > 0 && name) {
730 729 /*
731 730 * access the name in this fat-zap so that we'll check
732 731 * for i/o errors to the leaf blocks, etc.
733 732 */
734 733 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
735 734 8, 0, NULL);
736 735 if (err == EIO) {
737 736 tx->tx_err = err;
738 737 return;
739 738 }
740 739 }
741 740
742 741 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
743 742 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
744 743
745 744 /*
746 745 * If the modified blocks are scattered to the four winds,
747 746 * we'll have to modify an indirect twig for each.
748 747 */
749 748 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
750 749 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
751 750 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
752 751 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
753 752 else
754 753 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
755 754 }
756 755
757 756 void
758 757 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
759 758 {
760 759 dmu_tx_hold_t *txh;
761 760
762 761 ASSERT(tx->tx_txg == 0);
763 762
764 763 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
765 764 object, THT_BONUS, 0, 0);
766 765 if (txh)
767 766 dmu_tx_count_dnode(txh);
768 767 }
769 768
770 769 void
771 770 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
772 771 {
773 772 dmu_tx_hold_t *txh;
774 773 ASSERT(tx->tx_txg == 0);
775 774
776 775 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
777 776 DMU_NEW_OBJECT, THT_SPACE, space, 0);
778 777
779 778 txh->txh_space_towrite += space;
780 779 }
781 780
782 781 int
783 782 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
784 783 {
785 784 dmu_tx_hold_t *txh;
786 785 int holds = 0;
787 786
788 787 /*
789 788 * By asserting that the tx is assigned, we're counting the
790 789 * number of dn_tx_holds, which is the same as the number of
791 790 * dn_holds. Otherwise, we'd be counting dn_holds, but
792 791 * dn_tx_holds could be 0.
793 792 */
794 793 ASSERT(tx->tx_txg != 0);
795 794
796 795 /* if (tx->tx_anyobj == TRUE) */
797 796 /* return (0); */
798 797
799 798 for (txh = list_head(&tx->tx_holds); txh;
800 799 txh = list_next(&tx->tx_holds, txh)) {
801 800 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
802 801 holds++;
803 802 }
804 803
805 804 return (holds);
806 805 }
807 806
808 807 #ifdef ZFS_DEBUG
809 808 void
810 809 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
811 810 {
812 811 dmu_tx_hold_t *txh;
813 812 int match_object = FALSE, match_offset = FALSE;
814 813 dnode_t *dn;
815 814
816 815 DB_DNODE_ENTER(db);
817 816 dn = DB_DNODE(db);
818 817 ASSERT(tx->tx_txg != 0);
819 818 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
820 819 ASSERT3U(dn->dn_object, ==, db->db.db_object);
821 820
822 821 if (tx->tx_anyobj) {
823 822 DB_DNODE_EXIT(db);
824 823 return;
825 824 }
826 825
827 826 /* XXX No checking on the meta dnode for now */
828 827 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
829 828 DB_DNODE_EXIT(db);
830 829 return;
831 830 }
832 831
833 832 for (txh = list_head(&tx->tx_holds); txh;
834 833 txh = list_next(&tx->tx_holds, txh)) {
835 834 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
836 835 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
837 836 match_object = TRUE;
838 837 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
839 838 int datablkshift = dn->dn_datablkshift ?
840 839 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
841 840 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
842 841 int shift = datablkshift + epbs * db->db_level;
843 842 uint64_t beginblk = shift >= 64 ? 0 :
844 843 (txh->txh_arg1 >> shift);
845 844 uint64_t endblk = shift >= 64 ? 0 :
846 845 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
847 846 uint64_t blkid = db->db_blkid;
848 847
849 848 /* XXX txh_arg2 better not be zero... */
850 849
851 850 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
852 851 txh->txh_type, beginblk, endblk);
853 852
854 853 switch (txh->txh_type) {
855 854 case THT_WRITE:
856 855 if (blkid >= beginblk && blkid <= endblk)
857 856 match_offset = TRUE;
858 857 /*
859 858 * We will let this hold work for the bonus
860 859 * or spill buffer so that we don't need to
861 860 * hold it when creating a new object.
862 861 */
863 862 if (blkid == DMU_BONUS_BLKID ||
864 863 blkid == DMU_SPILL_BLKID)
865 864 match_offset = TRUE;
866 865 /*
867 866 * They might have to increase nlevels,
868 867 * thus dirtying the new TLIBs. Or the
869 868 * might have to change the block size,
870 869 * thus dirying the new lvl=0 blk=0.
871 870 */
872 871 if (blkid == 0)
873 872 match_offset = TRUE;
874 873 break;
875 874 case THT_FREE:
876 875 /*
877 876 * We will dirty all the level 1 blocks in
878 877 * the free range and perhaps the first and
879 878 * last level 0 block.
880 879 */
881 880 if (blkid >= beginblk && (blkid <= endblk ||
882 881 txh->txh_arg2 == DMU_OBJECT_END))
883 882 match_offset = TRUE;
884 883 break;
885 884 case THT_SPILL:
886 885 if (blkid == DMU_SPILL_BLKID)
887 886 match_offset = TRUE;
888 887 break;
889 888 case THT_BONUS:
890 889 if (blkid == DMU_BONUS_BLKID)
891 890 match_offset = TRUE;
892 891 break;
893 892 case THT_ZAP:
894 893 match_offset = TRUE;
895 894 break;
896 895 case THT_NEWOBJECT:
897 896 match_object = TRUE;
898 897 break;
899 898 default:
900 899 ASSERT(!"bad txh_type");
901 900 }
902 901 }
903 902 if (match_object && match_offset) {
904 903 DB_DNODE_EXIT(db);
905 904 return;
906 905 }
907 906 }
908 907 DB_DNODE_EXIT(db);
909 908 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
910 909 (u_longlong_t)db->db.db_object, db->db_level,
911 910 (u_longlong_t)db->db_blkid);
912 911 }
913 912 #endif
914 913
915 914 /*
916 915 * If we can't do 10 iops, something is wrong. Let us go ahead
917 916 * and hit zfs_dirty_data_max.
918 917 */
919 918 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
920 919 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
921 920
922 921 /*
923 922 * We delay transactions when we've determined that the backend storage
924 923 * isn't able to accommodate the rate of incoming writes.
925 924 *
926 925 * If there is already a transaction waiting, we delay relative to when
927 926 * that transaction finishes waiting. This way the calculated min_time
928 927 * is independent of the number of threads concurrently executing
929 928 * transactions.
930 929 *
931 930 * If we are the only waiter, wait relative to when the transaction
932 931 * started, rather than the current time. This credits the transaction for
933 932 * "time already served", e.g. reading indirect blocks.
934 933 *
935 934 * The minimum time for a transaction to take is calculated as:
936 935 * min_time = scale * (dirty - min) / (max - dirty)
937 936 * min_time is then capped at zfs_delay_max_ns.
938 937 *
939 938 * The delay has two degrees of freedom that can be adjusted via tunables.
940 939 * The percentage of dirty data at which we start to delay is defined by
941 940 * zfs_delay_min_dirty_percent. This should typically be at or above
942 941 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
943 942 * delay after writing at full speed has failed to keep up with the incoming
944 943 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
945 944 * speaking, this variable determines the amount of delay at the midpoint of
946 945 * the curve.
947 946 *
948 947 * delay
949 948 * 10ms +-------------------------------------------------------------*+
950 949 * | *|
951 950 * 9ms + *+
952 951 * | *|
953 952 * 8ms + *+
954 953 * | * |
955 954 * 7ms + * +
956 955 * | * |
957 956 * 6ms + * +
958 957 * | * |
959 958 * 5ms + * +
960 959 * | * |
961 960 * 4ms + * +
962 961 * | * |
963 962 * 3ms + * +
964 963 * | * |
965 964 * 2ms + (midpoint) * +
966 965 * | | ** |
967 966 * 1ms + v *** +
968 967 * | zfs_delay_scale ----------> ******** |
969 968 * 0 +-------------------------------------*********----------------+
970 969 * 0% <- zfs_dirty_data_max -> 100%
971 970 *
972 971 * Note that since the delay is added to the outstanding time remaining on the
973 972 * most recent transaction, the delay is effectively the inverse of IOPS.
974 973 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
975 974 * was chosen such that small changes in the amount of accumulated dirty data
976 975 * in the first 3/4 of the curve yield relatively small differences in the
977 976 * amount of delay.
978 977 *
979 978 * The effects can be easier to understand when the amount of delay is
980 979 * represented on a log scale:
981 980 *
982 981 * delay
983 982 * 100ms +-------------------------------------------------------------++
984 983 * + +
985 984 * | |
986 985 * + *+
987 986 * 10ms + *+
988 987 * + ** +
989 988 * | (midpoint) ** |
990 989 * + | ** +
991 990 * 1ms + v **** +
992 991 * + zfs_delay_scale ----------> ***** +
993 992 * | **** |
994 993 * + **** +
995 994 * 100us + ** +
996 995 * + * +
997 996 * | * |
998 997 * + * +
999 998 * 10us + * +
1000 999 * + +
1001 1000 * | |
1002 1001 * + +
1003 1002 * +--------------------------------------------------------------+
1004 1003 * 0% <- zfs_dirty_data_max -> 100%
1005 1004 *
1006 1005 * Note here that only as the amount of dirty data approaches its limit does
1007 1006 * the delay start to increase rapidly. The goal of a properly tuned system
1008 1007 * should be to keep the amount of dirty data out of that range by first
1009 1008 * ensuring that the appropriate limits are set for the I/O scheduler to reach
1010 1009 * optimal throughput on the backend storage, and then by changing the value
1011 1010 * of zfs_delay_scale to increase the steepness of the curve.
1012 1011 */
1013 1012 static void
1014 1013 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1015 1014 {
1016 1015 dsl_pool_t *dp = tx->tx_pool;
1017 1016 uint64_t delay_min_bytes =
1018 1017 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1019 1018 hrtime_t wakeup, min_tx_time, now;
1020 1019
1021 1020 if (dirty <= delay_min_bytes)
1022 1021 return;
1023 1022
1024 1023 /*
1025 1024 * The caller has already waited until we are under the max.
1026 1025 * We make them pass us the amount of dirty data so we don't
1027 1026 * have to handle the case of it being >= the max, which could
1028 1027 * cause a divide-by-zero if it's == the max.
1029 1028 */
1030 1029 ASSERT3U(dirty, <, zfs_dirty_data_max);
1031 1030
1032 1031 now = gethrtime();
1033 1032 min_tx_time = zfs_delay_scale *
1034 1033 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1035 1034 if (now > tx->tx_start + min_tx_time)
1036 1035 return;
1037 1036
1038 1037 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1039 1038
1040 1039 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1041 1040 uint64_t, min_tx_time);
1042 1041
1043 1042 mutex_enter(&dp->dp_lock);
1044 1043 wakeup = MAX(tx->tx_start + min_tx_time,
1045 1044 dp->dp_last_wakeup + min_tx_time);
1046 1045 dp->dp_last_wakeup = wakeup;
1047 1046 mutex_exit(&dp->dp_lock);
1048 1047
1049 1048 #ifdef _KERNEL
1050 1049 mutex_enter(&curthread->t_delay_lock);
1051 1050 while (cv_timedwait_hires(&curthread->t_delay_cv,
1052 1051 &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
1053 1052 CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
1054 1053 continue;
1055 1054 mutex_exit(&curthread->t_delay_lock);
1056 1055 #else
1057 1056 hrtime_t delta = wakeup - gethrtime();
1058 1057 struct timespec ts;
1059 1058 ts.tv_sec = delta / NANOSEC;
1060 1059 ts.tv_nsec = delta % NANOSEC;
1061 1060 (void) nanosleep(&ts, NULL);
1062 1061 #endif
1063 1062 }
1064 1063
1065 1064 static int
1066 1065 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1067 1066 {
1068 1067 dmu_tx_hold_t *txh;
1069 1068 spa_t *spa = tx->tx_pool->dp_spa;
1070 1069 uint64_t memory, asize, fsize, usize;
1071 1070 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1072 1071
1073 1072 ASSERT0(tx->tx_txg);
1074 1073
1075 1074 if (tx->tx_err)
1076 1075 return (tx->tx_err);
1077 1076
1078 1077 if (spa_suspended(spa)) {
1079 1078 /*
1080 1079 * If the user has indicated a blocking failure mode
1081 1080 * then return ERESTART which will block in dmu_tx_wait().
1082 1081 * Otherwise, return EIO so that an error can get
1083 1082 * propagated back to the VOP calls.
1084 1083 *
1085 1084 * Note that we always honor the txg_how flag regardless
1086 1085 * of the failuremode setting.
1087 1086 */
1088 1087 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1089 1088 txg_how != TXG_WAIT)
1090 1089 return (SET_ERROR(EIO));
1091 1090
1092 1091 return (SET_ERROR(ERESTART));
1093 1092 }
1094 1093
1095 1094 if (!tx->tx_waited &&
1096 1095 dsl_pool_need_dirty_delay(tx->tx_pool)) {
1097 1096 tx->tx_wait_dirty = B_TRUE;
1098 1097 return (SET_ERROR(ERESTART));
1099 1098 }
1100 1099
1101 1100 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1102 1101 tx->tx_needassign_txh = NULL;
1103 1102
1104 1103 /*
1105 1104 * NB: No error returns are allowed after txg_hold_open, but
1106 1105 * before processing the dnode holds, due to the
1107 1106 * dmu_tx_unassign() logic.
1108 1107 */
1109 1108
1110 1109 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1111 1110 for (txh = list_head(&tx->tx_holds); txh;
1112 1111 txh = list_next(&tx->tx_holds, txh)) {
1113 1112 dnode_t *dn = txh->txh_dnode;
1114 1113 if (dn != NULL) {
1115 1114 mutex_enter(&dn->dn_mtx);
1116 1115 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1117 1116 mutex_exit(&dn->dn_mtx);
1118 1117 tx->tx_needassign_txh = txh;
1119 1118 return (SET_ERROR(ERESTART));
1120 1119 }
1121 1120 if (dn->dn_assigned_txg == 0)
1122 1121 dn->dn_assigned_txg = tx->tx_txg;
1123 1122 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1124 1123 (void) refcount_add(&dn->dn_tx_holds, tx);
1125 1124 mutex_exit(&dn->dn_mtx);
1126 1125 }
1127 1126 towrite += txh->txh_space_towrite;
1128 1127 tofree += txh->txh_space_tofree;
1129 1128 tooverwrite += txh->txh_space_tooverwrite;
1130 1129 tounref += txh->txh_space_tounref;
1131 1130 tohold += txh->txh_memory_tohold;
1132 1131 fudge += txh->txh_fudge;
1133 1132 }
1134 1133
1135 1134 /*
1136 1135 * If a snapshot has been taken since we made our estimates,
1137 1136 * assume that we won't be able to free or overwrite anything.
1138 1137 */
1139 1138 if (tx->tx_objset &&
1140 1139 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1141 1140 tx->tx_lastsnap_txg) {
1142 1141 towrite += tooverwrite;
1143 1142 tooverwrite = tofree = 0;
1144 1143 }
1145 1144
1146 1145 /* needed allocation: worst-case estimate of write space */
1147 1146 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1148 1147 /* freed space estimate: worst-case overwrite + free estimate */
1149 1148 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1150 1149 /* convert unrefd space to worst-case estimate */
1151 1150 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1152 1151 /* calculate memory footprint estimate */
1153 1152 memory = towrite + tooverwrite + tohold;
1154 1153
1155 1154 #ifdef ZFS_DEBUG
1156 1155 /*
1157 1156 * Add in 'tohold' to account for our dirty holds on this memory
1158 1157 * XXX - the "fudge" factor is to account for skipped blocks that
1159 1158 * we missed because dnode_next_offset() misses in-core-only blocks.
1160 1159 */
1161 1160 tx->tx_space_towrite = asize +
1162 1161 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1163 1162 tx->tx_space_tofree = tofree;
1164 1163 tx->tx_space_tooverwrite = tooverwrite;
1165 1164 tx->tx_space_tounref = tounref;
1166 1165 #endif
1167 1166
1168 1167 if (tx->tx_dir && asize != 0) {
1169 1168 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1170 1169 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1171 1170 if (err)
1172 1171 return (err);
1173 1172 }
1174 1173
1175 1174 return (0);
1176 1175 }
1177 1176
1178 1177 static void
1179 1178 dmu_tx_unassign(dmu_tx_t *tx)
1180 1179 {
1181 1180 dmu_tx_hold_t *txh;
1182 1181
1183 1182 if (tx->tx_txg == 0)
1184 1183 return;
1185 1184
1186 1185 txg_rele_to_quiesce(&tx->tx_txgh);
1187 1186
1188 1187 /*
1189 1188 * Walk the transaction's hold list, removing the hold on the
1190 1189 * associated dnode, and notifying waiters if the refcount drops to 0.
1191 1190 */
1192 1191 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1193 1192 txh = list_next(&tx->tx_holds, txh)) {
1194 1193 dnode_t *dn = txh->txh_dnode;
1195 1194
1196 1195 if (dn == NULL)
1197 1196 continue;
1198 1197 mutex_enter(&dn->dn_mtx);
1199 1198 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1200 1199
1201 1200 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1202 1201 dn->dn_assigned_txg = 0;
1203 1202 cv_broadcast(&dn->dn_notxholds);
1204 1203 }
1205 1204 mutex_exit(&dn->dn_mtx);
1206 1205 }
1207 1206
1208 1207 txg_rele_to_sync(&tx->tx_txgh);
1209 1208
1210 1209 tx->tx_lasttried_txg = tx->tx_txg;
1211 1210 tx->tx_txg = 0;
1212 1211 }
1213 1212
1214 1213 /*
1215 1214 * Assign tx to a transaction group. txg_how can be one of:
1216 1215 *
1217 1216 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1218 1217 * a new one. This should be used when you're not holding locks.
1219 1218 * It will only fail if we're truly out of space (or over quota).
1220 1219 *
1221 1220 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1222 1221 * blocking, returns immediately with ERESTART. This should be used
1223 1222 * whenever you're holding locks. On an ERESTART error, the caller
1224 1223 * should drop locks, do a dmu_tx_wait(tx), and try again.
1225 1224 *
1226 1225 * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1227 1226 * has already been called on behalf of this operation (though
1228 1227 * most likely on a different tx).
1229 1228 */
1230 1229 int
1231 1230 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1232 1231 {
1233 1232 int err;
1234 1233
1235 1234 ASSERT(tx->tx_txg == 0);
1236 1235 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1237 1236 txg_how == TXG_WAITED);
1238 1237 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1239 1238
1240 1239 /* If we might wait, we must not hold the config lock. */
1241 1240 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1242 1241
1243 1242 if (txg_how == TXG_WAITED)
1244 1243 tx->tx_waited = B_TRUE;
1245 1244
1246 1245 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1247 1246 dmu_tx_unassign(tx);
1248 1247
1249 1248 if (err != ERESTART || txg_how != TXG_WAIT)
1250 1249 return (err);
1251 1250
1252 1251 dmu_tx_wait(tx);
1253 1252 }
1254 1253
1255 1254 txg_rele_to_quiesce(&tx->tx_txgh);
1256 1255
1257 1256 return (0);
1258 1257 }
1259 1258
1260 1259 void
1261 1260 dmu_tx_wait(dmu_tx_t *tx)
1262 1261 {
1263 1262 spa_t *spa = tx->tx_pool->dp_spa;
1264 1263 dsl_pool_t *dp = tx->tx_pool;
1265 1264
1266 1265 ASSERT(tx->tx_txg == 0);
1267 1266 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1268 1267
1269 1268 if (tx->tx_wait_dirty) {
1270 1269 /*
1271 1270 * dmu_tx_try_assign() has determined that we need to wait
1272 1271 * because we've consumed much or all of the dirty buffer
1273 1272 * space.
1274 1273 */
1275 1274 mutex_enter(&dp->dp_lock);
1276 1275 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1277 1276 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1278 1277 uint64_t dirty = dp->dp_dirty_total;
1279 1278 mutex_exit(&dp->dp_lock);
1280 1279
1281 1280 dmu_tx_delay(tx, dirty);
1282 1281
1283 1282 tx->tx_wait_dirty = B_FALSE;
1284 1283
1285 1284 /*
1286 1285 * Note: setting tx_waited only has effect if the caller
1287 1286 * used TX_WAIT. Otherwise they are going to destroy
1288 1287 * this tx and try again. The common case, zfs_write(),
1289 1288 * uses TX_WAIT.
1290 1289 */
1291 1290 tx->tx_waited = B_TRUE;
1292 1291 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1293 1292 /*
1294 1293 * If the pool is suspended we need to wait until it
1295 1294 * is resumed. Note that it's possible that the pool
1296 1295 * has become active after this thread has tried to
1297 1296 * obtain a tx. If that's the case then tx_lasttried_txg
1298 1297 * would not have been set.
1299 1298 */
1300 1299 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1301 1300 } else if (tx->tx_needassign_txh) {
1302 1301 /*
1303 1302 * A dnode is assigned to the quiescing txg. Wait for its
1304 1303 * transaction to complete.
1305 1304 */
1306 1305 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1307 1306
1308 1307 mutex_enter(&dn->dn_mtx);
1309 1308 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1310 1309 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1311 1310 mutex_exit(&dn->dn_mtx);
1312 1311 tx->tx_needassign_txh = NULL;
1313 1312 } else {
1314 1313 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1315 1314 }
1316 1315 }
1317 1316
1318 1317 void
1319 1318 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1320 1319 {
1321 1320 #ifdef ZFS_DEBUG
1322 1321 if (tx->tx_dir == NULL || delta == 0)
1323 1322 return;
1324 1323
1325 1324 if (delta > 0) {
1326 1325 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1327 1326 tx->tx_space_towrite);
1328 1327 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1329 1328 } else {
1330 1329 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1331 1330 }
1332 1331 #endif
1333 1332 }
1334 1333
1335 1334 void
1336 1335 dmu_tx_commit(dmu_tx_t *tx)
1337 1336 {
1338 1337 dmu_tx_hold_t *txh;
1339 1338
1340 1339 ASSERT(tx->tx_txg != 0);
1341 1340
1342 1341 /*
1343 1342 * Go through the transaction's hold list and remove holds on
1344 1343 * associated dnodes, notifying waiters if no holds remain.
1345 1344 */
1346 1345 while (txh = list_head(&tx->tx_holds)) {
1347 1346 dnode_t *dn = txh->txh_dnode;
1348 1347
1349 1348 list_remove(&tx->tx_holds, txh);
1350 1349 kmem_free(txh, sizeof (dmu_tx_hold_t));
1351 1350 if (dn == NULL)
1352 1351 continue;
1353 1352 mutex_enter(&dn->dn_mtx);
1354 1353 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1355 1354
1356 1355 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1357 1356 dn->dn_assigned_txg = 0;
1358 1357 cv_broadcast(&dn->dn_notxholds);
1359 1358 }
1360 1359 mutex_exit(&dn->dn_mtx);
1361 1360 dnode_rele(dn, tx);
1362 1361 }
1363 1362
1364 1363 if (tx->tx_tempreserve_cookie)
1365 1364 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1366 1365
1367 1366 if (!list_is_empty(&tx->tx_callbacks))
1368 1367 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1369 1368
1370 1369 if (tx->tx_anyobj == FALSE)
1371 1370 txg_rele_to_sync(&tx->tx_txgh);
1372 1371
1373 1372 list_destroy(&tx->tx_callbacks);
1374 1373 list_destroy(&tx->tx_holds);
1375 1374 #ifdef ZFS_DEBUG
1376 1375 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1377 1376 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1378 1377 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1379 1378 refcount_destroy_many(&tx->tx_space_written,
1380 1379 refcount_count(&tx->tx_space_written));
1381 1380 refcount_destroy_many(&tx->tx_space_freed,
1382 1381 refcount_count(&tx->tx_space_freed));
1383 1382 #endif
1384 1383 kmem_free(tx, sizeof (dmu_tx_t));
1385 1384 }
1386 1385
1387 1386 void
1388 1387 dmu_tx_abort(dmu_tx_t *tx)
1389 1388 {
1390 1389 dmu_tx_hold_t *txh;
1391 1390
1392 1391 ASSERT(tx->tx_txg == 0);
1393 1392
1394 1393 while (txh = list_head(&tx->tx_holds)) {
1395 1394 dnode_t *dn = txh->txh_dnode;
1396 1395
1397 1396 list_remove(&tx->tx_holds, txh);
1398 1397 kmem_free(txh, sizeof (dmu_tx_hold_t));
1399 1398 if (dn != NULL)
1400 1399 dnode_rele(dn, tx);
1401 1400 }
1402 1401
1403 1402 /*
1404 1403 * Call any registered callbacks with an error code.
1405 1404 */
1406 1405 if (!list_is_empty(&tx->tx_callbacks))
1407 1406 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1408 1407
1409 1408 list_destroy(&tx->tx_callbacks);
1410 1409 list_destroy(&tx->tx_holds);
1411 1410 #ifdef ZFS_DEBUG
1412 1411 refcount_destroy_many(&tx->tx_space_written,
1413 1412 refcount_count(&tx->tx_space_written));
1414 1413 refcount_destroy_many(&tx->tx_space_freed,
1415 1414 refcount_count(&tx->tx_space_freed));
1416 1415 #endif
1417 1416 kmem_free(tx, sizeof (dmu_tx_t));
1418 1417 }
1419 1418
1420 1419 uint64_t
1421 1420 dmu_tx_get_txg(dmu_tx_t *tx)
1422 1421 {
1423 1422 ASSERT(tx->tx_txg != 0);
1424 1423 return (tx->tx_txg);
1425 1424 }
1426 1425
1427 1426 dsl_pool_t *
1428 1427 dmu_tx_pool(dmu_tx_t *tx)
1429 1428 {
1430 1429 ASSERT(tx->tx_pool != NULL);
1431 1430 return (tx->tx_pool);
1432 1431 }
1433 1432
1434 1433
1435 1434 void
1436 1435 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1437 1436 {
1438 1437 dmu_tx_callback_t *dcb;
1439 1438
1440 1439 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1441 1440
1442 1441 dcb->dcb_func = func;
1443 1442 dcb->dcb_data = data;
1444 1443
1445 1444 list_insert_tail(&tx->tx_callbacks, dcb);
1446 1445 }
1447 1446
1448 1447 /*
1449 1448 * Call all the commit callbacks on a list, with a given error code.
1450 1449 */
1451 1450 void
1452 1451 dmu_tx_do_callbacks(list_t *cb_list, int error)
1453 1452 {
1454 1453 dmu_tx_callback_t *dcb;
1455 1454
1456 1455 while (dcb = list_head(cb_list)) {
1457 1456 list_remove(cb_list, dcb);
1458 1457 dcb->dcb_func(dcb->dcb_data, error);
1459 1458 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1460 1459 }
1461 1460 }
1462 1461
1463 1462 /*
1464 1463 * Interface to hold a bunch of attributes.
1465 1464 * used for creating new files.
1466 1465 * attrsize is the total size of all attributes
1467 1466 * to be added during object creation
1468 1467 *
1469 1468 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1470 1469 */
1471 1470
1472 1471 /*
1473 1472 * hold necessary attribute name for attribute registration.
1474 1473 * should be a very rare case where this is needed. If it does
1475 1474 * happen it would only happen on the first write to the file system.
1476 1475 */
1477 1476 static void
1478 1477 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1479 1478 {
1480 1479 int i;
1481 1480
1482 1481 if (!sa->sa_need_attr_registration)
1483 1482 return;
1484 1483
1485 1484 for (i = 0; i != sa->sa_num_attrs; i++) {
1486 1485 if (!sa->sa_attr_table[i].sa_registered) {
1487 1486 if (sa->sa_reg_attr_obj)
1488 1487 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1489 1488 B_TRUE, sa->sa_attr_table[i].sa_name);
1490 1489 else
1491 1490 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1492 1491 B_TRUE, sa->sa_attr_table[i].sa_name);
1493 1492 }
1494 1493 }
1495 1494 }
1496 1495
1497 1496
1498 1497 void
1499 1498 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1500 1499 {
1501 1500 dnode_t *dn;
1502 1501 dmu_tx_hold_t *txh;
1503 1502
1504 1503 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1505 1504 THT_SPILL, 0, 0);
1506 1505
1507 1506 dn = txh->txh_dnode;
1508 1507
1509 1508 if (dn == NULL)
1510 1509 return;
1511 1510
1512 1511 /* If blkptr doesn't exist then add space to towrite */
1513 1512 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1514 1513 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1515 1514 } else {
1516 1515 blkptr_t *bp;
1517 1516
1518 1517 bp = &dn->dn_phys->dn_spill;
1519 1518 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1520 1519 bp, bp->blk_birth))
1521 1520 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1522 1521 else
1523 1522 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1524 1523 if (!BP_IS_HOLE(bp))
1525 1524 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1526 1525 }
1527 1526 }
1528 1527
1529 1528 void
1530 1529 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1531 1530 {
1532 1531 sa_os_t *sa = tx->tx_objset->os_sa;
1533 1532
1534 1533 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1535 1534
1536 1535 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1537 1536 return;
1538 1537
1539 1538 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1540 1539 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1541 1540 else {
1542 1541 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1543 1542 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1544 1543 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1545 1544 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1546 1545 }
1547 1546
1548 1547 dmu_tx_sa_registration_hold(sa, tx);
1549 1548
1550 1549 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1551 1550 return;
1552 1551
1553 1552 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1554 1553 THT_SPILL, 0, 0);
1555 1554 }
1556 1555
1557 1556 /*
1558 1557 * Hold SA attribute
1559 1558 *
1560 1559 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1561 1560 *
1562 1561 * variable_size is the total size of all variable sized attributes
1563 1562 * passed to this function. It is not the total size of all
1564 1563 * variable size attributes that *may* exist on this object.
1565 1564 */
1566 1565 void
1567 1566 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1568 1567 {
1569 1568 uint64_t object;
1570 1569 sa_os_t *sa = tx->tx_objset->os_sa;
1571 1570
1572 1571 ASSERT(hdl != NULL);
1573 1572
1574 1573 object = sa_handle_object(hdl);
1575 1574
1576 1575 dmu_tx_hold_bonus(tx, object);
1577 1576
1578 1577 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1579 1578 return;
1580 1579
1581 1580 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1582 1581 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1583 1582 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1584 1583 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1585 1584 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1586 1585 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1587 1586 }
1588 1587
1589 1588 dmu_tx_sa_registration_hold(sa, tx);
1590 1589
1591 1590 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1592 1591 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1593 1592
1594 1593 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1595 1594 ASSERT(tx->tx_txg == 0);
1596 1595 dmu_tx_hold_spill(tx, object);
1597 1596 } else {
1598 1597 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1599 1598 dnode_t *dn;
1600 1599
1601 1600 DB_DNODE_ENTER(db);
1602 1601 dn = DB_DNODE(db);
1603 1602 if (dn->dn_have_spill) {
1604 1603 ASSERT(tx->tx_txg == 0);
1605 1604 dmu_tx_hold_spill(tx, object);
1606 1605 }
1607 1606 DB_DNODE_EXIT(db);
1608 1607 }
1609 1608 }
|
↓ open down ↓ |
979 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX