Print this page
4045 zfs write throttle & i/o scheduler performance work
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@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 *
|
↓ open down ↓ |
46 lines elided |
↑ open up ↑ |
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 + tx->tx_start = gethrtime();
57 58 #ifdef ZFS_DEBUG
58 59 refcount_create(&tx->tx_space_written);
59 60 refcount_create(&tx->tx_space_freed);
60 61 #endif
61 62 return (tx);
62 63 }
63 64
64 65 dmu_tx_t *
65 66 dmu_tx_create(objset_t *os)
66 67 {
67 68 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
68 69 tx->tx_objset = os;
69 70 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
70 71 return (tx);
71 72 }
72 73
73 74 dmu_tx_t *
74 75 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
75 76 {
76 77 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
77 78
78 79 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
79 80 tx->tx_pool = dp;
80 81 tx->tx_txg = txg;
81 82 tx->tx_anyobj = TRUE;
82 83
83 84 return (tx);
84 85 }
85 86
86 87 int
87 88 dmu_tx_is_syncing(dmu_tx_t *tx)
88 89 {
89 90 return (tx->tx_anyobj);
90 91 }
91 92
92 93 int
93 94 dmu_tx_private_ok(dmu_tx_t *tx)
94 95 {
95 96 return (tx->tx_anyobj);
96 97 }
97 98
98 99 static dmu_tx_hold_t *
99 100 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
100 101 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
101 102 {
102 103 dmu_tx_hold_t *txh;
103 104 dnode_t *dn = NULL;
104 105 int err;
105 106
106 107 if (object != DMU_NEW_OBJECT) {
107 108 err = dnode_hold(os, object, tx, &dn);
108 109 if (err) {
109 110 tx->tx_err = err;
110 111 return (NULL);
111 112 }
112 113
113 114 if (err == 0 && tx->tx_txg != 0) {
114 115 mutex_enter(&dn->dn_mtx);
115 116 /*
116 117 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
117 118 * problem, but there's no way for it to happen (for
118 119 * now, at least).
119 120 */
120 121 ASSERT(dn->dn_assigned_txg == 0);
121 122 dn->dn_assigned_txg = tx->tx_txg;
122 123 (void) refcount_add(&dn->dn_tx_holds, tx);
123 124 mutex_exit(&dn->dn_mtx);
124 125 }
125 126 }
126 127
127 128 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
128 129 txh->txh_tx = tx;
129 130 txh->txh_dnode = dn;
130 131 #ifdef ZFS_DEBUG
131 132 txh->txh_type = type;
132 133 txh->txh_arg1 = arg1;
133 134 txh->txh_arg2 = arg2;
134 135 #endif
135 136 list_insert_tail(&tx->tx_holds, txh);
136 137
137 138 return (txh);
138 139 }
139 140
140 141 void
141 142 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
142 143 {
143 144 /*
144 145 * If we're syncing, they can manipulate any object anyhow, and
145 146 * the hold on the dnode_t can cause problems.
146 147 */
147 148 if (!dmu_tx_is_syncing(tx)) {
148 149 (void) dmu_tx_hold_object_impl(tx, os,
149 150 object, THT_NEWOBJECT, 0, 0);
150 151 }
151 152 }
152 153
153 154 static int
154 155 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
155 156 {
156 157 int err;
157 158 dmu_buf_impl_t *db;
158 159
159 160 rw_enter(&dn->dn_struct_rwlock, RW_READER);
160 161 db = dbuf_hold_level(dn, level, blkid, FTAG);
161 162 rw_exit(&dn->dn_struct_rwlock);
162 163 if (db == NULL)
163 164 return (SET_ERROR(EIO));
164 165 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
165 166 dbuf_rele(db, FTAG);
166 167 return (err);
167 168 }
168 169
169 170 static void
170 171 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
171 172 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
172 173 {
173 174 objset_t *os = dn->dn_objset;
174 175 dsl_dataset_t *ds = os->os_dsl_dataset;
175 176 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
176 177 dmu_buf_impl_t *parent = NULL;
177 178 blkptr_t *bp = NULL;
178 179 uint64_t space;
179 180
180 181 if (level >= dn->dn_nlevels || history[level] == blkid)
181 182 return;
182 183
183 184 history[level] = blkid;
184 185
185 186 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
186 187
187 188 if (db == NULL || db == dn->dn_dbuf) {
188 189 ASSERT(level != 0);
189 190 db = NULL;
190 191 } else {
191 192 ASSERT(DB_DNODE(db) == dn);
192 193 ASSERT(db->db_level == level);
193 194 ASSERT(db->db.db_size == space);
194 195 ASSERT(db->db_blkid == blkid);
195 196 bp = db->db_blkptr;
196 197 parent = db->db_parent;
197 198 }
198 199
199 200 freeable = (bp && (freeable ||
200 201 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
201 202
202 203 if (freeable)
203 204 txh->txh_space_tooverwrite += space;
204 205 else
205 206 txh->txh_space_towrite += space;
206 207 if (bp)
207 208 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
208 209
209 210 dmu_tx_count_twig(txh, dn, parent, level + 1,
210 211 blkid >> epbs, freeable, history);
211 212 }
212 213
213 214 /* ARGSUSED */
214 215 static void
215 216 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
216 217 {
217 218 dnode_t *dn = txh->txh_dnode;
218 219 uint64_t start, end, i;
219 220 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
220 221 int err = 0;
221 222
222 223 if (len == 0)
223 224 return;
224 225
225 226 min_bs = SPA_MINBLOCKSHIFT;
226 227 max_bs = SPA_MAXBLOCKSHIFT;
227 228 min_ibs = DN_MIN_INDBLKSHIFT;
228 229 max_ibs = DN_MAX_INDBLKSHIFT;
229 230
230 231 if (dn) {
231 232 uint64_t history[DN_MAX_LEVELS];
232 233 int nlvls = dn->dn_nlevels;
233 234 int delta;
234 235
235 236 /*
236 237 * For i/o error checking, read the first and last level-0
237 238 * blocks (if they are not aligned), and all the level-1 blocks.
238 239 */
239 240 if (dn->dn_maxblkid == 0) {
240 241 delta = dn->dn_datablksz;
241 242 start = (off < dn->dn_datablksz) ? 0 : 1;
242 243 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
243 244 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
244 245 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
245 246 if (err)
246 247 goto out;
247 248 delta -= off;
248 249 }
249 250 } else {
250 251 zio_t *zio = zio_root(dn->dn_objset->os_spa,
251 252 NULL, NULL, ZIO_FLAG_CANFAIL);
252 253
253 254 /* first level-0 block */
254 255 start = off >> dn->dn_datablkshift;
255 256 if (P2PHASE(off, dn->dn_datablksz) ||
256 257 len < dn->dn_datablksz) {
257 258 err = dmu_tx_check_ioerr(zio, dn, 0, start);
258 259 if (err)
259 260 goto out;
260 261 }
261 262
262 263 /* last level-0 block */
263 264 end = (off+len-1) >> dn->dn_datablkshift;
264 265 if (end != start && end <= dn->dn_maxblkid &&
265 266 P2PHASE(off+len, dn->dn_datablksz)) {
266 267 err = dmu_tx_check_ioerr(zio, dn, 0, end);
267 268 if (err)
268 269 goto out;
269 270 }
270 271
271 272 /* level-1 blocks */
272 273 if (nlvls > 1) {
273 274 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
274 275 for (i = (start>>shft)+1; i < end>>shft; i++) {
275 276 err = dmu_tx_check_ioerr(zio, dn, 1, i);
276 277 if (err)
277 278 goto out;
278 279 }
279 280 }
280 281
281 282 err = zio_wait(zio);
282 283 if (err)
283 284 goto out;
284 285 delta = P2NPHASE(off, dn->dn_datablksz);
285 286 }
286 287
287 288 min_ibs = max_ibs = dn->dn_indblkshift;
288 289 if (dn->dn_maxblkid > 0) {
289 290 /*
290 291 * The blocksize can't change,
291 292 * so we can make a more precise estimate.
292 293 */
293 294 ASSERT(dn->dn_datablkshift != 0);
294 295 min_bs = max_bs = dn->dn_datablkshift;
295 296 }
296 297
297 298 /*
298 299 * If this write is not off the end of the file
299 300 * we need to account for overwrites/unref.
300 301 */
301 302 if (start <= dn->dn_maxblkid) {
302 303 for (int l = 0; l < DN_MAX_LEVELS; l++)
303 304 history[l] = -1ULL;
304 305 }
305 306 while (start <= dn->dn_maxblkid) {
306 307 dmu_buf_impl_t *db;
307 308
308 309 rw_enter(&dn->dn_struct_rwlock, RW_READER);
309 310 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
310 311 rw_exit(&dn->dn_struct_rwlock);
311 312
312 313 if (err) {
313 314 txh->txh_tx->tx_err = err;
314 315 return;
315 316 }
316 317
317 318 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
318 319 history);
319 320 dbuf_rele(db, FTAG);
320 321 if (++start > end) {
321 322 /*
322 323 * Account for new indirects appearing
323 324 * before this IO gets assigned into a txg.
324 325 */
325 326 bits = 64 - min_bs;
326 327 epbs = min_ibs - SPA_BLKPTRSHIFT;
327 328 for (bits -= epbs * (nlvls - 1);
328 329 bits >= 0; bits -= epbs)
329 330 txh->txh_fudge += 1ULL << max_ibs;
330 331 goto out;
331 332 }
332 333 off += delta;
333 334 if (len >= delta)
334 335 len -= delta;
335 336 delta = dn->dn_datablksz;
336 337 }
337 338 }
338 339
339 340 /*
340 341 * 'end' is the last thing we will access, not one past.
341 342 * This way we won't overflow when accessing the last byte.
342 343 */
343 344 start = P2ALIGN(off, 1ULL << max_bs);
344 345 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
345 346 txh->txh_space_towrite += end - start + 1;
346 347
347 348 start >>= min_bs;
348 349 end >>= min_bs;
349 350
350 351 epbs = min_ibs - SPA_BLKPTRSHIFT;
351 352
352 353 /*
353 354 * The object contains at most 2^(64 - min_bs) blocks,
354 355 * and each indirect level maps 2^epbs.
355 356 */
356 357 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
357 358 start >>= epbs;
358 359 end >>= epbs;
359 360 ASSERT3U(end, >=, start);
360 361 txh->txh_space_towrite += (end - start + 1) << max_ibs;
361 362 if (start != 0) {
362 363 /*
363 364 * We also need a new blkid=0 indirect block
364 365 * to reference any existing file data.
365 366 */
366 367 txh->txh_space_towrite += 1ULL << max_ibs;
367 368 }
368 369 }
369 370
370 371 out:
371 372 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
372 373 2 * DMU_MAX_ACCESS)
373 374 err = SET_ERROR(EFBIG);
374 375
375 376 if (err)
376 377 txh->txh_tx->tx_err = err;
377 378 }
378 379
379 380 static void
380 381 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
381 382 {
382 383 dnode_t *dn = txh->txh_dnode;
383 384 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
384 385 uint64_t space = mdn->dn_datablksz +
385 386 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
386 387
387 388 if (dn && dn->dn_dbuf->db_blkptr &&
388 389 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
389 390 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
390 391 txh->txh_space_tooverwrite += space;
391 392 txh->txh_space_tounref += space;
392 393 } else {
393 394 txh->txh_space_towrite += space;
394 395 if (dn && dn->dn_dbuf->db_blkptr)
395 396 txh->txh_space_tounref += space;
396 397 }
397 398 }
398 399
399 400 void
400 401 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
401 402 {
402 403 dmu_tx_hold_t *txh;
403 404
404 405 ASSERT(tx->tx_txg == 0);
405 406 ASSERT(len < DMU_MAX_ACCESS);
406 407 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
407 408
408 409 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
409 410 object, THT_WRITE, off, len);
410 411 if (txh == NULL)
411 412 return;
412 413
413 414 dmu_tx_count_write(txh, off, len);
414 415 dmu_tx_count_dnode(txh);
415 416 }
416 417
417 418 static void
418 419 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
419 420 {
420 421 uint64_t blkid, nblks, lastblk;
421 422 uint64_t space = 0, unref = 0, skipped = 0;
422 423 dnode_t *dn = txh->txh_dnode;
423 424 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
424 425 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
425 426 int epbs;
426 427 uint64_t l0span = 0, nl1blks = 0;
427 428
428 429 if (dn->dn_nlevels == 0)
429 430 return;
430 431
431 432 /*
432 433 * The struct_rwlock protects us against dn_nlevels
433 434 * changing, in case (against all odds) we manage to dirty &
434 435 * sync out the changes after we check for being dirty.
435 436 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
436 437 */
437 438 rw_enter(&dn->dn_struct_rwlock, RW_READER);
438 439 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
439 440 if (dn->dn_maxblkid == 0) {
440 441 if (off == 0 && len >= dn->dn_datablksz) {
441 442 blkid = 0;
442 443 nblks = 1;
443 444 } else {
444 445 rw_exit(&dn->dn_struct_rwlock);
445 446 return;
446 447 }
447 448 } else {
448 449 blkid = off >> dn->dn_datablkshift;
449 450 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
450 451
451 452 if (blkid > dn->dn_maxblkid) {
452 453 rw_exit(&dn->dn_struct_rwlock);
453 454 return;
454 455 }
455 456 if (blkid + nblks > dn->dn_maxblkid)
456 457 nblks = dn->dn_maxblkid - blkid + 1;
457 458
458 459 }
459 460 l0span = nblks; /* save for later use to calc level > 1 overhead */
460 461 if (dn->dn_nlevels == 1) {
461 462 int i;
462 463 for (i = 0; i < nblks; i++) {
463 464 blkptr_t *bp = dn->dn_phys->dn_blkptr;
464 465 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
465 466 bp += blkid + i;
466 467 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
467 468 dprintf_bp(bp, "can free old%s", "");
468 469 space += bp_get_dsize(spa, bp);
469 470 }
470 471 unref += BP_GET_ASIZE(bp);
471 472 }
472 473 nl1blks = 1;
473 474 nblks = 0;
474 475 }
475 476
476 477 lastblk = blkid + nblks - 1;
477 478 while (nblks) {
478 479 dmu_buf_impl_t *dbuf;
479 480 uint64_t ibyte, new_blkid;
480 481 int epb = 1 << epbs;
481 482 int err, i, blkoff, tochk;
482 483 blkptr_t *bp;
483 484
484 485 ibyte = blkid << dn->dn_datablkshift;
485 486 err = dnode_next_offset(dn,
486 487 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
487 488 new_blkid = ibyte >> dn->dn_datablkshift;
488 489 if (err == ESRCH) {
489 490 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
490 491 break;
491 492 }
492 493 if (err) {
493 494 txh->txh_tx->tx_err = err;
494 495 break;
495 496 }
496 497 if (new_blkid > lastblk) {
497 498 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
498 499 break;
499 500 }
500 501
501 502 if (new_blkid > blkid) {
502 503 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
503 504 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
504 505 nblks -= new_blkid - blkid;
505 506 blkid = new_blkid;
506 507 }
507 508 blkoff = P2PHASE(blkid, epb);
508 509 tochk = MIN(epb - blkoff, nblks);
509 510
510 511 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
511 512 if (err) {
512 513 txh->txh_tx->tx_err = err;
513 514 break;
514 515 }
515 516
516 517 txh->txh_memory_tohold += dbuf->db.db_size;
517 518
518 519 /*
519 520 * We don't check memory_tohold against DMU_MAX_ACCESS because
520 521 * memory_tohold is an over-estimation (especially the >L1
521 522 * indirect blocks), so it could fail. Callers should have
522 523 * already verified that they will not be holding too much
523 524 * memory.
524 525 */
525 526
526 527 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
527 528 if (err != 0) {
528 529 txh->txh_tx->tx_err = err;
529 530 dbuf_rele(dbuf, FTAG);
530 531 break;
531 532 }
532 533
533 534 bp = dbuf->db.db_data;
534 535 bp += blkoff;
535 536
536 537 for (i = 0; i < tochk; i++) {
537 538 if (dsl_dataset_block_freeable(ds, &bp[i],
538 539 bp[i].blk_birth)) {
539 540 dprintf_bp(&bp[i], "can free old%s", "");
540 541 space += bp_get_dsize(spa, &bp[i]);
541 542 }
542 543 unref += BP_GET_ASIZE(bp);
543 544 }
544 545 dbuf_rele(dbuf, FTAG);
545 546
546 547 ++nl1blks;
547 548 blkid += tochk;
548 549 nblks -= tochk;
549 550 }
550 551 rw_exit(&dn->dn_struct_rwlock);
551 552
552 553 /*
553 554 * Add in memory requirements of higher-level indirects.
554 555 * This assumes a worst-possible scenario for dn_nlevels and a
555 556 * worst-possible distribution of l1-blocks over the region to free.
556 557 */
557 558 {
558 559 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
559 560 int level = 2;
560 561 /*
561 562 * Here we don't use DN_MAX_LEVEL, but calculate it with the
562 563 * given datablkshift and indblkshift. This makes the
563 564 * difference between 19 and 8 on large files.
564 565 */
565 566 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
566 567 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
567 568
568 569 while (level++ < maxlevel) {
569 570 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
570 571 << dn->dn_indblkshift;
571 572 blkcnt = 1 + (blkcnt >> epbs);
572 573 }
573 574 }
574 575
575 576 /* account for new level 1 indirect blocks that might show up */
576 577 if (skipped > 0) {
577 578 txh->txh_fudge += skipped << dn->dn_indblkshift;
578 579 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
579 580 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
580 581 }
581 582 txh->txh_space_tofree += space;
582 583 txh->txh_space_tounref += unref;
583 584 }
584 585
585 586 void
586 587 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
587 588 {
588 589 dmu_tx_hold_t *txh;
589 590 dnode_t *dn;
|
↓ open down ↓ |
523 lines elided |
↑ open up ↑ |
590 591 int err;
591 592 zio_t *zio;
592 593
593 594 ASSERT(tx->tx_txg == 0);
594 595
595 596 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
596 597 object, THT_FREE, off, len);
597 598 if (txh == NULL)
598 599 return;
599 600 dn = txh->txh_dnode;
601 + dmu_tx_count_dnode(txh);
600 602
601 603 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
602 604 return;
603 605 if (len == DMU_OBJECT_END)
604 606 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
605 607
606 - dmu_tx_count_dnode(txh);
607 608
608 609 /*
609 610 * For i/o error checking, we read the first and last level-0
610 611 * blocks if they are not aligned, and all the level-1 blocks.
611 612 *
612 613 * Note: dbuf_free_range() assumes that we have not instantiated
613 614 * any level-0 dbufs that will be completely freed. Therefore we must
614 615 * exercise care to not read or count the first and last blocks
615 616 * if they are blocksize-aligned.
616 617 */
617 618 if (dn->dn_datablkshift == 0) {
618 619 if (off != 0 || len < dn->dn_datablksz)
619 620 dmu_tx_count_write(txh, off, len);
620 621 } else {
621 622 /* first block will be modified if it is not aligned */
622 623 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
623 624 dmu_tx_count_write(txh, off, 1);
624 625 /* last block will be modified if it is not aligned */
625 626 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
626 627 dmu_tx_count_write(txh, off+len, 1);
627 628 }
628 629
629 630 /*
630 631 * Check level-1 blocks.
631 632 */
632 633 if (dn->dn_nlevels > 1) {
633 634 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
634 635 SPA_BLKPTRSHIFT;
635 636 uint64_t start = off >> shift;
636 637 uint64_t end = (off + len) >> shift;
637 638
638 639 ASSERT(dn->dn_datablkshift != 0);
639 640 ASSERT(dn->dn_indblkshift != 0);
640 641
641 642 zio = zio_root(tx->tx_pool->dp_spa,
642 643 NULL, NULL, ZIO_FLAG_CANFAIL);
643 644 for (uint64_t i = start; i <= end; i++) {
644 645 uint64_t ibyte = i << shift;
645 646 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
646 647 i = ibyte >> shift;
647 648 if (err == ESRCH)
648 649 break;
649 650 if (err) {
650 651 tx->tx_err = err;
651 652 return;
652 653 }
653 654
654 655 err = dmu_tx_check_ioerr(zio, dn, 1, i);
655 656 if (err) {
656 657 tx->tx_err = err;
657 658 return;
658 659 }
659 660 }
660 661 err = zio_wait(zio);
661 662 if (err) {
662 663 tx->tx_err = err;
663 664 return;
664 665 }
665 666 }
666 667
667 668 dmu_tx_count_free(txh, off, len);
668 669 }
669 670
670 671 void
671 672 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
672 673 {
673 674 dmu_tx_hold_t *txh;
674 675 dnode_t *dn;
675 676 uint64_t nblocks;
676 677 int epbs, err;
677 678
678 679 ASSERT(tx->tx_txg == 0);
679 680
680 681 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
681 682 object, THT_ZAP, add, (uintptr_t)name);
682 683 if (txh == NULL)
683 684 return;
684 685 dn = txh->txh_dnode;
685 686
686 687 dmu_tx_count_dnode(txh);
687 688
688 689 if (dn == NULL) {
689 690 /*
690 691 * We will be able to fit a new object's entries into one leaf
691 692 * block. So there will be at most 2 blocks total,
692 693 * including the header block.
693 694 */
694 695 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
695 696 return;
696 697 }
697 698
698 699 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
699 700
700 701 if (dn->dn_maxblkid == 0 && !add) {
701 702 blkptr_t *bp;
702 703
703 704 /*
704 705 * If there is only one block (i.e. this is a micro-zap)
705 706 * and we are not adding anything, the accounting is simple.
706 707 */
707 708 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
708 709 if (err) {
709 710 tx->tx_err = err;
710 711 return;
711 712 }
712 713
713 714 /*
714 715 * Use max block size here, since we don't know how much
715 716 * the size will change between now and the dbuf dirty call.
716 717 */
717 718 bp = &dn->dn_phys->dn_blkptr[0];
718 719 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
719 720 bp, bp->blk_birth))
720 721 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
721 722 else
722 723 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
723 724 if (!BP_IS_HOLE(bp))
724 725 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
725 726 return;
726 727 }
727 728
728 729 if (dn->dn_maxblkid > 0 && name) {
729 730 /*
730 731 * access the name in this fat-zap so that we'll check
731 732 * for i/o errors to the leaf blocks, etc.
732 733 */
733 734 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
734 735 8, 0, NULL);
735 736 if (err == EIO) {
736 737 tx->tx_err = err;
737 738 return;
738 739 }
739 740 }
740 741
741 742 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
742 743 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
743 744
744 745 /*
745 746 * If the modified blocks are scattered to the four winds,
746 747 * we'll have to modify an indirect twig for each.
747 748 */
748 749 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
749 750 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
750 751 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
751 752 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
752 753 else
753 754 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
754 755 }
755 756
756 757 void
757 758 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
758 759 {
759 760 dmu_tx_hold_t *txh;
760 761
761 762 ASSERT(tx->tx_txg == 0);
762 763
763 764 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
764 765 object, THT_BONUS, 0, 0);
765 766 if (txh)
766 767 dmu_tx_count_dnode(txh);
767 768 }
768 769
769 770 void
770 771 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
771 772 {
772 773 dmu_tx_hold_t *txh;
773 774 ASSERT(tx->tx_txg == 0);
774 775
775 776 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
776 777 DMU_NEW_OBJECT, THT_SPACE, space, 0);
777 778
778 779 txh->txh_space_towrite += space;
779 780 }
780 781
781 782 int
782 783 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
783 784 {
784 785 dmu_tx_hold_t *txh;
785 786 int holds = 0;
786 787
787 788 /*
788 789 * By asserting that the tx is assigned, we're counting the
789 790 * number of dn_tx_holds, which is the same as the number of
790 791 * dn_holds. Otherwise, we'd be counting dn_holds, but
791 792 * dn_tx_holds could be 0.
792 793 */
793 794 ASSERT(tx->tx_txg != 0);
794 795
795 796 /* if (tx->tx_anyobj == TRUE) */
796 797 /* return (0); */
797 798
798 799 for (txh = list_head(&tx->tx_holds); txh;
799 800 txh = list_next(&tx->tx_holds, txh)) {
800 801 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
801 802 holds++;
802 803 }
803 804
804 805 return (holds);
805 806 }
806 807
807 808 #ifdef ZFS_DEBUG
808 809 void
809 810 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
810 811 {
811 812 dmu_tx_hold_t *txh;
812 813 int match_object = FALSE, match_offset = FALSE;
813 814 dnode_t *dn;
814 815
815 816 DB_DNODE_ENTER(db);
816 817 dn = DB_DNODE(db);
817 818 ASSERT(tx->tx_txg != 0);
818 819 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
819 820 ASSERT3U(dn->dn_object, ==, db->db.db_object);
820 821
821 822 if (tx->tx_anyobj) {
822 823 DB_DNODE_EXIT(db);
823 824 return;
824 825 }
825 826
826 827 /* XXX No checking on the meta dnode for now */
827 828 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
828 829 DB_DNODE_EXIT(db);
829 830 return;
830 831 }
831 832
832 833 for (txh = list_head(&tx->tx_holds); txh;
833 834 txh = list_next(&tx->tx_holds, txh)) {
834 835 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
835 836 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
836 837 match_object = TRUE;
837 838 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
838 839 int datablkshift = dn->dn_datablkshift ?
839 840 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
840 841 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
841 842 int shift = datablkshift + epbs * db->db_level;
842 843 uint64_t beginblk = shift >= 64 ? 0 :
843 844 (txh->txh_arg1 >> shift);
844 845 uint64_t endblk = shift >= 64 ? 0 :
845 846 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
846 847 uint64_t blkid = db->db_blkid;
847 848
848 849 /* XXX txh_arg2 better not be zero... */
849 850
850 851 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
851 852 txh->txh_type, beginblk, endblk);
852 853
853 854 switch (txh->txh_type) {
854 855 case THT_WRITE:
855 856 if (blkid >= beginblk && blkid <= endblk)
856 857 match_offset = TRUE;
857 858 /*
858 859 * We will let this hold work for the bonus
859 860 * or spill buffer so that we don't need to
860 861 * hold it when creating a new object.
861 862 */
862 863 if (blkid == DMU_BONUS_BLKID ||
863 864 blkid == DMU_SPILL_BLKID)
864 865 match_offset = TRUE;
865 866 /*
866 867 * They might have to increase nlevels,
867 868 * thus dirtying the new TLIBs. Or the
868 869 * might have to change the block size,
869 870 * thus dirying the new lvl=0 blk=0.
870 871 */
871 872 if (blkid == 0)
872 873 match_offset = TRUE;
873 874 break;
874 875 case THT_FREE:
875 876 /*
876 877 * We will dirty all the level 1 blocks in
877 878 * the free range and perhaps the first and
878 879 * last level 0 block.
879 880 */
880 881 if (blkid >= beginblk && (blkid <= endblk ||
881 882 txh->txh_arg2 == DMU_OBJECT_END))
882 883 match_offset = TRUE;
883 884 break;
884 885 case THT_SPILL:
885 886 if (blkid == DMU_SPILL_BLKID)
886 887 match_offset = TRUE;
887 888 break;
888 889 case THT_BONUS:
889 890 if (blkid == DMU_BONUS_BLKID)
890 891 match_offset = TRUE;
891 892 break;
892 893 case THT_ZAP:
893 894 match_offset = TRUE;
894 895 break;
895 896 case THT_NEWOBJECT:
896 897 match_object = TRUE;
897 898 break;
898 899 default:
899 900 ASSERT(!"bad txh_type");
900 901 }
901 902 }
902 903 if (match_object && match_offset) {
903 904 DB_DNODE_EXIT(db);
|
↓ open down ↓ |
287 lines elided |
↑ open up ↑ |
904 905 return;
905 906 }
906 907 }
907 908 DB_DNODE_EXIT(db);
908 909 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
909 910 (u_longlong_t)db->db.db_object, db->db_level,
910 911 (u_longlong_t)db->db_blkid);
911 912 }
912 913 #endif
913 914
915 +/*
916 + * If we can't do 10 iops, something is wrong. Let us go ahead
917 + * and hit zfs_dirty_data_max.
918 + */
919 +hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
920 +int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
921 +
922 +/*
923 + * We delay transactions when we've determined that the backend storage
924 + * isn't able to accommodate the rate of incoming writes.
925 + *
926 + * If there is already a transaction waiting, we delay relative to when
927 + * that transaction finishes waiting. This way the calculated min_time
928 + * is independent of the number of threads concurrently executing
929 + * transactions.
930 + *
931 + * If we are the only waiter, wait relative to when the transaction
932 + * started, rather than the current time. This credits the transaction for
933 + * "time already served", e.g. reading indirect blocks.
934 + *
935 + * The minimum time for a transaction to take is calculated as:
936 + * min_time = scale * (dirty - min) / (max - dirty)
937 + * min_time is then capped at zfs_delay_max_ns.
938 + *
939 + * The delay has two degrees of freedom that can be adjusted via tunables.
940 + * The percentage of dirty data at which we start to delay is defined by
941 + * zfs_delay_min_dirty_percent. This should typically be at or above
942 + * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
943 + * delay after writing at full speed has failed to keep up with the incoming
944 + * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
945 + * speaking, this variable determines the amount of delay at the midpoint of
946 + * the curve.
947 + *
948 + * delay
949 + * 10ms +-------------------------------------------------------------*+
950 + * | *|
951 + * 9ms + *+
952 + * | *|
953 + * 8ms + *+
954 + * | * |
955 + * 7ms + * +
956 + * | * |
957 + * 6ms + * +
958 + * | * |
959 + * 5ms + * +
960 + * | * |
961 + * 4ms + * +
962 + * | * |
963 + * 3ms + * +
964 + * | * |
965 + * 2ms + (midpoint) * +
966 + * | | ** |
967 + * 1ms + v *** +
968 + * | zfs_delay_scale ----------> ******** |
969 + * 0 +-------------------------------------*********----------------+
970 + * 0% <- zfs_dirty_data_max -> 100%
971 + *
972 + * Note that since the delay is added to the outstanding time remaining on the
973 + * most recent transaction, the delay is effectively the inverse of IOPS.
974 + * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
975 + * was chosen such that small changes in the amount of accumulated dirty data
976 + * in the first 3/4 of the curve yield relatively small differences in the
977 + * amount of delay.
978 + *
979 + * The effects can be easier to understand when the amount of delay is
980 + * represented on a log scale:
981 + *
982 + * delay
983 + * 100ms +-------------------------------------------------------------++
984 + * + +
985 + * | |
986 + * + *+
987 + * 10ms + *+
988 + * + ** +
989 + * | (midpoint) ** |
990 + * + | ** +
991 + * 1ms + v **** +
992 + * + zfs_delay_scale ----------> ***** +
993 + * | **** |
994 + * + **** +
995 + * 100us + ** +
996 + * + * +
997 + * | * |
998 + * + * +
999 + * 10us + * +
1000 + * + +
1001 + * | |
1002 + * + +
1003 + * +--------------------------------------------------------------+
1004 + * 0% <- zfs_dirty_data_max -> 100%
1005 + *
1006 + * Note here that only as the amount of dirty data approaches its limit does
1007 + * the delay start to increase rapidly. The goal of a properly tuned system
1008 + * should be to keep the amount of dirty data out of that range by first
1009 + * ensuring that the appropriate limits are set for the I/O scheduler to reach
1010 + * optimal throughput on the backend storage, and then by changing the value
1011 + * of zfs_delay_scale to increase the steepness of the curve.
1012 + */
1013 +static void
1014 +dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1015 +{
1016 + dsl_pool_t *dp = tx->tx_pool;
1017 + uint64_t delay_min_bytes =
1018 + zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1019 + hrtime_t wakeup, min_tx_time, now;
1020 +
1021 + if (dirty <= delay_min_bytes)
1022 + return;
1023 +
1024 + /*
1025 + * The caller has already waited until we are under the max.
1026 + * We make them pass us the amount of dirty data so we don't
1027 + * have to handle the case of it being >= the max, which could
1028 + * cause a divide-by-zero if it's == the max.
1029 + */
1030 + ASSERT3U(dirty, <, zfs_dirty_data_max);
1031 +
1032 + now = gethrtime();
1033 + min_tx_time = zfs_delay_scale *
1034 + (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1035 + if (now > tx->tx_start + min_tx_time)
1036 + return;
1037 +
1038 + min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1039 +
1040 + DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1041 + uint64_t, min_tx_time);
1042 +
1043 + mutex_enter(&dp->dp_lock);
1044 + wakeup = MAX(tx->tx_start + min_tx_time,
1045 + dp->dp_last_wakeup + min_tx_time);
1046 + dp->dp_last_wakeup = wakeup;
1047 + mutex_exit(&dp->dp_lock);
1048 +
1049 +#ifdef _KERNEL
1050 + mutex_enter(&curthread->t_delay_lock);
1051 + while (cv_timedwait_hires(&curthread->t_delay_cv,
1052 + &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
1053 + CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
1054 + continue;
1055 + mutex_exit(&curthread->t_delay_lock);
1056 +#else
1057 + hrtime_t delta = wakeup - gethrtime();
1058 + struct timespec ts;
1059 + ts.tv_sec = delta / NANOSEC;
1060 + ts.tv_nsec = delta % NANOSEC;
1061 + (void) nanosleep(&ts, NULL);
1062 +#endif
1063 +}
1064 +
914 1065 static int
915 1066 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
916 1067 {
917 1068 dmu_tx_hold_t *txh;
918 1069 spa_t *spa = tx->tx_pool->dp_spa;
919 1070 uint64_t memory, asize, fsize, usize;
920 1071 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
921 1072
922 1073 ASSERT0(tx->tx_txg);
923 1074
924 1075 if (tx->tx_err)
925 1076 return (tx->tx_err);
926 1077
927 1078 if (spa_suspended(spa)) {
928 1079 /*
929 1080 * If the user has indicated a blocking failure mode
930 1081 * then return ERESTART which will block in dmu_tx_wait().
931 1082 * Otherwise, return EIO so that an error can get
932 1083 * propagated back to the VOP calls.
933 1084 *
|
↓ open down ↓ |
10 lines elided |
↑ open up ↑ |
934 1085 * Note that we always honor the txg_how flag regardless
935 1086 * of the failuremode setting.
936 1087 */
937 1088 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
938 1089 txg_how != TXG_WAIT)
939 1090 return (SET_ERROR(EIO));
940 1091
941 1092 return (SET_ERROR(ERESTART));
942 1093 }
943 1094
1095 + if (!tx->tx_waited &&
1096 + dsl_pool_need_dirty_delay(tx->tx_pool)) {
1097 + tx->tx_wait_dirty = B_TRUE;
1098 + return (SET_ERROR(ERESTART));
1099 + }
1100 +
944 1101 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
945 1102 tx->tx_needassign_txh = NULL;
946 1103
947 1104 /*
948 1105 * NB: No error returns are allowed after txg_hold_open, but
949 1106 * before processing the dnode holds, due to the
950 1107 * dmu_tx_unassign() logic.
951 1108 */
952 1109
953 1110 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
954 1111 for (txh = list_head(&tx->tx_holds); txh;
955 1112 txh = list_next(&tx->tx_holds, txh)) {
956 1113 dnode_t *dn = txh->txh_dnode;
957 1114 if (dn != NULL) {
958 1115 mutex_enter(&dn->dn_mtx);
959 1116 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
960 1117 mutex_exit(&dn->dn_mtx);
961 1118 tx->tx_needassign_txh = txh;
962 1119 return (SET_ERROR(ERESTART));
963 1120 }
964 1121 if (dn->dn_assigned_txg == 0)
965 1122 dn->dn_assigned_txg = tx->tx_txg;
966 1123 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
967 1124 (void) refcount_add(&dn->dn_tx_holds, tx);
968 1125 mutex_exit(&dn->dn_mtx);
969 1126 }
970 1127 towrite += txh->txh_space_towrite;
971 1128 tofree += txh->txh_space_tofree;
972 1129 tooverwrite += txh->txh_space_tooverwrite;
973 1130 tounref += txh->txh_space_tounref;
974 1131 tohold += txh->txh_memory_tohold;
975 1132 fudge += txh->txh_fudge;
976 1133 }
977 1134
978 1135 /*
979 1136 * If a snapshot has been taken since we made our estimates,
980 1137 * assume that we won't be able to free or overwrite anything.
981 1138 */
982 1139 if (tx->tx_objset &&
983 1140 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
984 1141 tx->tx_lastsnap_txg) {
985 1142 towrite += tooverwrite;
986 1143 tooverwrite = tofree = 0;
987 1144 }
988 1145
989 1146 /* needed allocation: worst-case estimate of write space */
990 1147 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
991 1148 /* freed space estimate: worst-case overwrite + free estimate */
992 1149 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
993 1150 /* convert unrefd space to worst-case estimate */
994 1151 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
995 1152 /* calculate memory footprint estimate */
996 1153 memory = towrite + tooverwrite + tohold;
997 1154
998 1155 #ifdef ZFS_DEBUG
999 1156 /*
1000 1157 * Add in 'tohold' to account for our dirty holds on this memory
1001 1158 * XXX - the "fudge" factor is to account for skipped blocks that
1002 1159 * we missed because dnode_next_offset() misses in-core-only blocks.
1003 1160 */
1004 1161 tx->tx_space_towrite = asize +
1005 1162 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1006 1163 tx->tx_space_tofree = tofree;
1007 1164 tx->tx_space_tooverwrite = tooverwrite;
1008 1165 tx->tx_space_tounref = tounref;
1009 1166 #endif
1010 1167
1011 1168 if (tx->tx_dir && asize != 0) {
1012 1169 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1013 1170 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1014 1171 if (err)
1015 1172 return (err);
1016 1173 }
1017 1174
1018 1175 return (0);
1019 1176 }
1020 1177
1021 1178 static void
1022 1179 dmu_tx_unassign(dmu_tx_t *tx)
1023 1180 {
1024 1181 dmu_tx_hold_t *txh;
1025 1182
1026 1183 if (tx->tx_txg == 0)
1027 1184 return;
1028 1185
1029 1186 txg_rele_to_quiesce(&tx->tx_txgh);
1030 1187
1031 1188 /*
1032 1189 * Walk the transaction's hold list, removing the hold on the
1033 1190 * associated dnode, and notifying waiters if the refcount drops to 0.
1034 1191 */
1035 1192 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1036 1193 txh = list_next(&tx->tx_holds, txh)) {
1037 1194 dnode_t *dn = txh->txh_dnode;
1038 1195
1039 1196 if (dn == NULL)
1040 1197 continue;
1041 1198 mutex_enter(&dn->dn_mtx);
1042 1199 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1043 1200
1044 1201 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1045 1202 dn->dn_assigned_txg = 0;
1046 1203 cv_broadcast(&dn->dn_notxholds);
1047 1204 }
1048 1205 mutex_exit(&dn->dn_mtx);
1049 1206 }
1050 1207
1051 1208 txg_rele_to_sync(&tx->tx_txgh);
1052 1209
1053 1210 tx->tx_lasttried_txg = tx->tx_txg;
1054 1211 tx->tx_txg = 0;
1055 1212 }
1056 1213
1057 1214 /*
|
↓ open down ↓ |
104 lines elided |
↑ open up ↑ |
1058 1215 * Assign tx to a transaction group. txg_how can be one of:
1059 1216 *
1060 1217 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1061 1218 * a new one. This should be used when you're not holding locks.
1062 1219 * It will only fail if we're truly out of space (or over quota).
1063 1220 *
1064 1221 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1065 1222 * blocking, returns immediately with ERESTART. This should be used
1066 1223 * whenever you're holding locks. On an ERESTART error, the caller
1067 1224 * should drop locks, do a dmu_tx_wait(tx), and try again.
1225 + *
1226 + * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1227 + * has already been called on behalf of this operation (though
1228 + * most likely on a different tx).
1068 1229 */
1069 1230 int
1070 1231 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1071 1232 {
1072 1233 int err;
1073 1234
1074 1235 ASSERT(tx->tx_txg == 0);
1075 - ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT);
1236 + ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1237 + txg_how == TXG_WAITED);
1076 1238 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1077 1239
1078 1240 /* If we might wait, we must not hold the config lock. */
1079 1241 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1080 1242
1243 + if (txg_how == TXG_WAITED)
1244 + tx->tx_waited = B_TRUE;
1245 +
1081 1246 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1082 1247 dmu_tx_unassign(tx);
1083 1248
1084 1249 if (err != ERESTART || txg_how != TXG_WAIT)
1085 1250 return (err);
1086 1251
1087 1252 dmu_tx_wait(tx);
1088 1253 }
1089 1254
1090 1255 txg_rele_to_quiesce(&tx->tx_txgh);
1091 1256
1092 1257 return (0);
1093 1258 }
1094 1259
1095 1260 void
1096 1261 dmu_tx_wait(dmu_tx_t *tx)
1097 1262 {
1098 1263 spa_t *spa = tx->tx_pool->dp_spa;
1264 + dsl_pool_t *dp = tx->tx_pool;
1099 1265
1100 1266 ASSERT(tx->tx_txg == 0);
1101 1267 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1102 1268
1103 - /*
1104 - * It's possible that the pool has become active after this thread
1105 - * has tried to obtain a tx. If that's the case then his
1106 - * tx_lasttried_txg would not have been assigned.
1107 - */
1108 - if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1109 - txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1269 + if (tx->tx_wait_dirty) {
1270 + /*
1271 + * dmu_tx_try_assign() has determined that we need to wait
1272 + * because we've consumed much or all of the dirty buffer
1273 + * space.
1274 + */
1275 + mutex_enter(&dp->dp_lock);
1276 + while (dp->dp_dirty_total >= zfs_dirty_data_max)
1277 + cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1278 + uint64_t dirty = dp->dp_dirty_total;
1279 + mutex_exit(&dp->dp_lock);
1280 +
1281 + dmu_tx_delay(tx, dirty);
1282 +
1283 + tx->tx_wait_dirty = B_FALSE;
1284 +
1285 + /*
1286 + * Note: setting tx_waited only has effect if the caller
1287 + * used TX_WAIT. Otherwise they are going to destroy
1288 + * this tx and try again. The common case, zfs_write(),
1289 + * uses TX_WAIT.
1290 + */
1291 + tx->tx_waited = B_TRUE;
1292 + } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1293 + /*
1294 + * If the pool is suspended we need to wait until it
1295 + * is resumed. Note that it's possible that the pool
1296 + * has become active after this thread has tried to
1297 + * obtain a tx. If that's the case then tx_lasttried_txg
1298 + * would not have been set.
1299 + */
1300 + txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1110 1301 } else if (tx->tx_needassign_txh) {
1302 + /*
1303 + * A dnode is assigned to the quiescing txg. Wait for its
1304 + * transaction to complete.
1305 + */
1111 1306 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1112 1307
1113 1308 mutex_enter(&dn->dn_mtx);
1114 1309 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1115 1310 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1116 1311 mutex_exit(&dn->dn_mtx);
1117 1312 tx->tx_needassign_txh = NULL;
1118 1313 } else {
1119 1314 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1120 1315 }
1121 1316 }
1122 1317
1123 1318 void
1124 1319 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1125 1320 {
1126 1321 #ifdef ZFS_DEBUG
1127 1322 if (tx->tx_dir == NULL || delta == 0)
1128 1323 return;
1129 1324
1130 1325 if (delta > 0) {
1131 1326 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1132 1327 tx->tx_space_towrite);
1133 1328 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1134 1329 } else {
1135 1330 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1136 1331 }
1137 1332 #endif
1138 1333 }
1139 1334
1140 1335 void
1141 1336 dmu_tx_commit(dmu_tx_t *tx)
1142 1337 {
1143 1338 dmu_tx_hold_t *txh;
1144 1339
1145 1340 ASSERT(tx->tx_txg != 0);
1146 1341
1147 1342 /*
1148 1343 * Go through the transaction's hold list and remove holds on
1149 1344 * associated dnodes, notifying waiters if no holds remain.
1150 1345 */
1151 1346 while (txh = list_head(&tx->tx_holds)) {
1152 1347 dnode_t *dn = txh->txh_dnode;
1153 1348
1154 1349 list_remove(&tx->tx_holds, txh);
1155 1350 kmem_free(txh, sizeof (dmu_tx_hold_t));
1156 1351 if (dn == NULL)
1157 1352 continue;
1158 1353 mutex_enter(&dn->dn_mtx);
1159 1354 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1160 1355
1161 1356 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1162 1357 dn->dn_assigned_txg = 0;
1163 1358 cv_broadcast(&dn->dn_notxholds);
1164 1359 }
1165 1360 mutex_exit(&dn->dn_mtx);
1166 1361 dnode_rele(dn, tx);
1167 1362 }
1168 1363
1169 1364 if (tx->tx_tempreserve_cookie)
1170 1365 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1171 1366
1172 1367 if (!list_is_empty(&tx->tx_callbacks))
1173 1368 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1174 1369
1175 1370 if (tx->tx_anyobj == FALSE)
1176 1371 txg_rele_to_sync(&tx->tx_txgh);
1177 1372
1178 1373 list_destroy(&tx->tx_callbacks);
1179 1374 list_destroy(&tx->tx_holds);
1180 1375 #ifdef ZFS_DEBUG
1181 1376 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1182 1377 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1183 1378 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1184 1379 refcount_destroy_many(&tx->tx_space_written,
1185 1380 refcount_count(&tx->tx_space_written));
1186 1381 refcount_destroy_many(&tx->tx_space_freed,
1187 1382 refcount_count(&tx->tx_space_freed));
1188 1383 #endif
1189 1384 kmem_free(tx, sizeof (dmu_tx_t));
1190 1385 }
1191 1386
1192 1387 void
1193 1388 dmu_tx_abort(dmu_tx_t *tx)
1194 1389 {
1195 1390 dmu_tx_hold_t *txh;
1196 1391
1197 1392 ASSERT(tx->tx_txg == 0);
1198 1393
1199 1394 while (txh = list_head(&tx->tx_holds)) {
1200 1395 dnode_t *dn = txh->txh_dnode;
1201 1396
1202 1397 list_remove(&tx->tx_holds, txh);
1203 1398 kmem_free(txh, sizeof (dmu_tx_hold_t));
1204 1399 if (dn != NULL)
1205 1400 dnode_rele(dn, tx);
1206 1401 }
1207 1402
1208 1403 /*
1209 1404 * Call any registered callbacks with an error code.
1210 1405 */
1211 1406 if (!list_is_empty(&tx->tx_callbacks))
1212 1407 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1213 1408
1214 1409 list_destroy(&tx->tx_callbacks);
1215 1410 list_destroy(&tx->tx_holds);
1216 1411 #ifdef ZFS_DEBUG
1217 1412 refcount_destroy_many(&tx->tx_space_written,
1218 1413 refcount_count(&tx->tx_space_written));
1219 1414 refcount_destroy_many(&tx->tx_space_freed,
1220 1415 refcount_count(&tx->tx_space_freed));
1221 1416 #endif
1222 1417 kmem_free(tx, sizeof (dmu_tx_t));
1223 1418 }
1224 1419
1225 1420 uint64_t
1226 1421 dmu_tx_get_txg(dmu_tx_t *tx)
1227 1422 {
1228 1423 ASSERT(tx->tx_txg != 0);
1229 1424 return (tx->tx_txg);
1230 1425 }
1231 1426
1232 1427 dsl_pool_t *
1233 1428 dmu_tx_pool(dmu_tx_t *tx)
1234 1429 {
1235 1430 ASSERT(tx->tx_pool != NULL);
1236 1431 return (tx->tx_pool);
1237 1432 }
1238 1433
1239 1434
1240 1435 void
1241 1436 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1242 1437 {
1243 1438 dmu_tx_callback_t *dcb;
1244 1439
1245 1440 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1246 1441
1247 1442 dcb->dcb_func = func;
1248 1443 dcb->dcb_data = data;
1249 1444
1250 1445 list_insert_tail(&tx->tx_callbacks, dcb);
1251 1446 }
1252 1447
1253 1448 /*
1254 1449 * Call all the commit callbacks on a list, with a given error code.
1255 1450 */
1256 1451 void
1257 1452 dmu_tx_do_callbacks(list_t *cb_list, int error)
1258 1453 {
1259 1454 dmu_tx_callback_t *dcb;
1260 1455
1261 1456 while (dcb = list_head(cb_list)) {
1262 1457 list_remove(cb_list, dcb);
1263 1458 dcb->dcb_func(dcb->dcb_data, error);
1264 1459 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1265 1460 }
1266 1461 }
1267 1462
1268 1463 /*
1269 1464 * Interface to hold a bunch of attributes.
1270 1465 * used for creating new files.
1271 1466 * attrsize is the total size of all attributes
1272 1467 * to be added during object creation
1273 1468 *
1274 1469 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1275 1470 */
1276 1471
1277 1472 /*
1278 1473 * hold necessary attribute name for attribute registration.
1279 1474 * should be a very rare case where this is needed. If it does
1280 1475 * happen it would only happen on the first write to the file system.
1281 1476 */
1282 1477 static void
1283 1478 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1284 1479 {
1285 1480 int i;
1286 1481
1287 1482 if (!sa->sa_need_attr_registration)
1288 1483 return;
1289 1484
1290 1485 for (i = 0; i != sa->sa_num_attrs; i++) {
1291 1486 if (!sa->sa_attr_table[i].sa_registered) {
1292 1487 if (sa->sa_reg_attr_obj)
1293 1488 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1294 1489 B_TRUE, sa->sa_attr_table[i].sa_name);
1295 1490 else
1296 1491 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1297 1492 B_TRUE, sa->sa_attr_table[i].sa_name);
1298 1493 }
1299 1494 }
1300 1495 }
1301 1496
1302 1497
1303 1498 void
1304 1499 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1305 1500 {
1306 1501 dnode_t *dn;
1307 1502 dmu_tx_hold_t *txh;
1308 1503
1309 1504 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1310 1505 THT_SPILL, 0, 0);
1311 1506
1312 1507 dn = txh->txh_dnode;
1313 1508
1314 1509 if (dn == NULL)
1315 1510 return;
1316 1511
1317 1512 /* If blkptr doesn't exist then add space to towrite */
1318 1513 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1319 1514 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1320 1515 } else {
1321 1516 blkptr_t *bp;
1322 1517
1323 1518 bp = &dn->dn_phys->dn_spill;
1324 1519 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1325 1520 bp, bp->blk_birth))
1326 1521 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1327 1522 else
1328 1523 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1329 1524 if (!BP_IS_HOLE(bp))
1330 1525 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1331 1526 }
1332 1527 }
1333 1528
1334 1529 void
1335 1530 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1336 1531 {
1337 1532 sa_os_t *sa = tx->tx_objset->os_sa;
1338 1533
1339 1534 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1340 1535
1341 1536 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1342 1537 return;
1343 1538
1344 1539 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1345 1540 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1346 1541 else {
1347 1542 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1348 1543 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1349 1544 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1350 1545 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1351 1546 }
1352 1547
1353 1548 dmu_tx_sa_registration_hold(sa, tx);
1354 1549
1355 1550 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1356 1551 return;
1357 1552
1358 1553 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1359 1554 THT_SPILL, 0, 0);
1360 1555 }
1361 1556
1362 1557 /*
1363 1558 * Hold SA attribute
1364 1559 *
1365 1560 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1366 1561 *
1367 1562 * variable_size is the total size of all variable sized attributes
1368 1563 * passed to this function. It is not the total size of all
1369 1564 * variable size attributes that *may* exist on this object.
1370 1565 */
1371 1566 void
1372 1567 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1373 1568 {
1374 1569 uint64_t object;
1375 1570 sa_os_t *sa = tx->tx_objset->os_sa;
1376 1571
1377 1572 ASSERT(hdl != NULL);
1378 1573
1379 1574 object = sa_handle_object(hdl);
1380 1575
1381 1576 dmu_tx_hold_bonus(tx, object);
1382 1577
1383 1578 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1384 1579 return;
1385 1580
1386 1581 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1387 1582 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1388 1583 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1389 1584 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1390 1585 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1391 1586 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1392 1587 }
1393 1588
1394 1589 dmu_tx_sa_registration_hold(sa, tx);
1395 1590
1396 1591 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1397 1592 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1398 1593
1399 1594 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1400 1595 ASSERT(tx->tx_txg == 0);
1401 1596 dmu_tx_hold_spill(tx, object);
1402 1597 } else {
1403 1598 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1404 1599 dnode_t *dn;
1405 1600
1406 1601 DB_DNODE_ENTER(db);
1407 1602 dn = DB_DNODE(db);
1408 1603 if (dn->dn_have_spill) {
1409 1604 ASSERT(tx->tx_txg == 0);
1410 1605 dmu_tx_hold_spill(tx, object);
1411 1606 }
1412 1607 DB_DNODE_EXIT(db);
1413 1608 }
1414 1609 }
|
↓ open down ↓ |
294 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX