1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 */
26
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dbuf.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
33 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
34 #include <sys/dsl_pool.h>
35 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 #include <sys/spa.h>
37 #include <sys/sa.h>
38 #include <sys/sa_impl.h>
39 #include <sys/zfs_context.h>
40 #include <sys/varargs.h>
41
42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43 uint64_t arg1, uint64_t arg2);
44
45
46 dmu_tx_t *
47 dmu_tx_create_dd(dsl_dir_t *dd)
48 {
49 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
50 tx->tx_dir = dd;
51 if (dd != NULL)
52 tx->tx_pool = dd->dd_pool;
53 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
54 offsetof(dmu_tx_hold_t, txh_node));
55 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
56 offsetof(dmu_tx_callback_t, dcb_node));
57 tx->tx_start = gethrtime();
58 #ifdef ZFS_DEBUG
59 refcount_create(&tx->tx_space_written);
60 refcount_create(&tx->tx_space_freed);
61 #endif
62 return (tx);
63 }
64
65 dmu_tx_t *
66 dmu_tx_create(objset_t *os)
67 {
68 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
69 tx->tx_objset = os;
70 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
71 return (tx);
72 }
73
74 dmu_tx_t *
75 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
76 {
77 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
78
79 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
80 tx->tx_pool = dp;
81 tx->tx_txg = txg;
82 tx->tx_anyobj = TRUE;
83
84 return (tx);
85 }
86
87 int
88 dmu_tx_is_syncing(dmu_tx_t *tx)
89 {
90 return (tx->tx_anyobj);
91 }
92
93 int
94 dmu_tx_private_ok(dmu_tx_t *tx)
95 {
96 return (tx->tx_anyobj);
97 }
98
99 static dmu_tx_hold_t *
100 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
101 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
102 {
103 dmu_tx_hold_t *txh;
104 dnode_t *dn = NULL;
105 int err;
106
107 if (object != DMU_NEW_OBJECT) {
108 err = dnode_hold(os, object, tx, &dn);
109 if (err) {
110 tx->tx_err = err;
111 return (NULL);
112 }
113
114 if (err == 0 && tx->tx_txg != 0) {
115 mutex_enter(&dn->dn_mtx);
116 /*
117 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
118 * problem, but there's no way for it to happen (for
119 * now, at least).
120 */
121 ASSERT(dn->dn_assigned_txg == 0);
122 dn->dn_assigned_txg = tx->tx_txg;
123 (void) refcount_add(&dn->dn_tx_holds, tx);
124 mutex_exit(&dn->dn_mtx);
125 }
126 }
127
128 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
129 txh->txh_tx = tx;
130 txh->txh_dnode = dn;
131 #ifdef ZFS_DEBUG
132 txh->txh_type = type;
133 txh->txh_arg1 = arg1;
134 txh->txh_arg2 = arg2;
135 #endif
136 list_insert_tail(&tx->tx_holds, txh);
137
138 return (txh);
139 }
140
141 void
142 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
143 {
144 /*
145 * If we're syncing, they can manipulate any object anyhow, and
146 * the hold on the dnode_t can cause problems.
147 */
148 if (!dmu_tx_is_syncing(tx)) {
149 (void) dmu_tx_hold_object_impl(tx, os,
150 object, THT_NEWOBJECT, 0, 0);
151 }
152 }
153
154 static int
155 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
156 {
157 int err;
158 dmu_buf_impl_t *db;
159
160 rw_enter(&dn->dn_struct_rwlock, RW_READER);
161 db = dbuf_hold_level(dn, level, blkid, FTAG);
162 rw_exit(&dn->dn_struct_rwlock);
163 if (db == NULL)
164 return (SET_ERROR(EIO));
165 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
166 dbuf_rele(db, FTAG);
167 return (err);
168 }
169
170 static void
171 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
172 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
173 {
174 objset_t *os = dn->dn_objset;
175 dsl_dataset_t *ds = os->os_dsl_dataset;
176 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
177 dmu_buf_impl_t *parent = NULL;
178 blkptr_t *bp = NULL;
179 uint64_t space;
180
181 if (level >= dn->dn_nlevels || history[level] == blkid)
182 return;
183
184 history[level] = blkid;
185
186 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
187
188 if (db == NULL || db == dn->dn_dbuf) {
189 ASSERT(level != 0);
190 db = NULL;
191 } else {
192 ASSERT(DB_DNODE(db) == dn);
193 ASSERT(db->db_level == level);
194 ASSERT(db->db.db_size == space);
195 ASSERT(db->db_blkid == blkid);
196 bp = db->db_blkptr;
197 parent = db->db_parent;
198 }
199
200 freeable = (bp && (freeable ||
201 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
202
203 if (freeable)
204 txh->txh_space_tooverwrite += space;
205 else
206 txh->txh_space_towrite += space;
207 if (bp)
208 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
209
210 dmu_tx_count_twig(txh, dn, parent, level + 1,
211 blkid >> epbs, freeable, history);
212 }
213
214 /* ARGSUSED */
215 static void
216 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
217 {
218 dnode_t *dn = txh->txh_dnode;
219 uint64_t start, end, i;
220 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
221 int err = 0;
222
223 if (len == 0)
224 return;
225
226 min_bs = SPA_MINBLOCKSHIFT;
227 max_bs = SPA_MAXBLOCKSHIFT;
228 min_ibs = DN_MIN_INDBLKSHIFT;
229 max_ibs = DN_MAX_INDBLKSHIFT;
230
231 if (dn) {
232 uint64_t history[DN_MAX_LEVELS];
233 int nlvls = dn->dn_nlevels;
234 int delta;
235
236 /*
237 * For i/o error checking, read the first and last level-0
238 * blocks (if they are not aligned), and all the level-1 blocks.
239 */
240 if (dn->dn_maxblkid == 0) {
241 delta = dn->dn_datablksz;
242 start = (off < dn->dn_datablksz) ? 0 : 1;
243 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
244 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
245 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
246 if (err)
247 goto out;
248 delta -= off;
249 }
250 } else {
251 zio_t *zio = zio_root(dn->dn_objset->os_spa,
252 NULL, NULL, ZIO_FLAG_CANFAIL);
253
254 /* first level-0 block */
255 start = off >> dn->dn_datablkshift;
256 if (P2PHASE(off, dn->dn_datablksz) ||
257 len < dn->dn_datablksz) {
258 err = dmu_tx_check_ioerr(zio, dn, 0, start);
259 if (err)
260 goto out;
261 }
262
263 /* last level-0 block */
264 end = (off+len-1) >> dn->dn_datablkshift;
265 if (end != start && end <= dn->dn_maxblkid &&
266 P2PHASE(off+len, dn->dn_datablksz)) {
267 err = dmu_tx_check_ioerr(zio, dn, 0, end);
268 if (err)
269 goto out;
270 }
271
272 /* level-1 blocks */
273 if (nlvls > 1) {
274 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
275 for (i = (start>>shft)+1; i < end>>shft; i++) {
276 err = dmu_tx_check_ioerr(zio, dn, 1, i);
277 if (err)
278 goto out;
279 }
280 }
281
282 err = zio_wait(zio);
283 if (err)
284 goto out;
285 delta = P2NPHASE(off, dn->dn_datablksz);
286 }
287
288 min_ibs = max_ibs = dn->dn_indblkshift;
289 if (dn->dn_maxblkid > 0) {
290 /*
291 * The blocksize can't change,
292 * so we can make a more precise estimate.
293 */
294 ASSERT(dn->dn_datablkshift != 0);
295 min_bs = max_bs = dn->dn_datablkshift;
296 }
297
298 /*
299 * If this write is not off the end of the file
300 * we need to account for overwrites/unref.
301 */
302 if (start <= dn->dn_maxblkid) {
303 for (int l = 0; l < DN_MAX_LEVELS; l++)
304 history[l] = -1ULL;
305 }
306 while (start <= dn->dn_maxblkid) {
307 dmu_buf_impl_t *db;
308
309 rw_enter(&dn->dn_struct_rwlock, RW_READER);
310 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
311 rw_exit(&dn->dn_struct_rwlock);
312
313 if (err) {
314 txh->txh_tx->tx_err = err;
315 return;
316 }
317
318 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
319 history);
320 dbuf_rele(db, FTAG);
321 if (++start > end) {
322 /*
323 * Account for new indirects appearing
324 * before this IO gets assigned into a txg.
325 */
326 bits = 64 - min_bs;
327 epbs = min_ibs - SPA_BLKPTRSHIFT;
328 for (bits -= epbs * (nlvls - 1);
329 bits >= 0; bits -= epbs)
330 txh->txh_fudge += 1ULL << max_ibs;
331 goto out;
332 }
333 off += delta;
334 if (len >= delta)
335 len -= delta;
336 delta = dn->dn_datablksz;
337 }
338 }
339
340 /*
341 * 'end' is the last thing we will access, not one past.
342 * This way we won't overflow when accessing the last byte.
343 */
344 start = P2ALIGN(off, 1ULL << max_bs);
345 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
346 txh->txh_space_towrite += end - start + 1;
347
348 start >>= min_bs;
349 end >>= min_bs;
350
351 epbs = min_ibs - SPA_BLKPTRSHIFT;
352
353 /*
354 * The object contains at most 2^(64 - min_bs) blocks,
355 * and each indirect level maps 2^epbs.
356 */
357 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
358 start >>= epbs;
359 end >>= epbs;
360 ASSERT3U(end, >=, start);
361 txh->txh_space_towrite += (end - start + 1) << max_ibs;
362 if (start != 0) {
363 /*
364 * We also need a new blkid=0 indirect block
365 * to reference any existing file data.
366 */
367 txh->txh_space_towrite += 1ULL << max_ibs;
368 }
369 }
370
371 out:
372 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
373 2 * DMU_MAX_ACCESS)
374 err = SET_ERROR(EFBIG);
375
376 if (err)
377 txh->txh_tx->tx_err = err;
378 }
379
380 static void
381 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
382 {
383 dnode_t *dn = txh->txh_dnode;
384 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
385 uint64_t space = mdn->dn_datablksz +
386 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
387
388 if (dn && dn->dn_dbuf->db_blkptr &&
389 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
390 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
391 txh->txh_space_tooverwrite += space;
392 txh->txh_space_tounref += space;
393 } else {
394 txh->txh_space_towrite += space;
395 if (dn && dn->dn_dbuf->db_blkptr)
396 txh->txh_space_tounref += space;
397 }
398 }
399
400 void
401 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
402 {
403 dmu_tx_hold_t *txh;
404
405 ASSERT(tx->tx_txg == 0);
406 ASSERT(len < DMU_MAX_ACCESS);
407 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
408
409 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
410 object, THT_WRITE, off, len);
411 if (txh == NULL)
412 return;
413
414 dmu_tx_count_write(txh, off, len);
415 dmu_tx_count_dnode(txh);
416 }
417
418 static void
419 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
420 {
421 uint64_t blkid, nblks, lastblk;
422 uint64_t space = 0, unref = 0, skipped = 0;
423 dnode_t *dn = txh->txh_dnode;
424 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
425 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
426 int epbs;
427 uint64_t l0span = 0, nl1blks = 0;
428
429 if (dn->dn_nlevels == 0)
430 return;
431
432 /*
433 * The struct_rwlock protects us against dn_nlevels
434 * changing, in case (against all odds) we manage to dirty &
435 * sync out the changes after we check for being dirty.
436 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
437 */
438 rw_enter(&dn->dn_struct_rwlock, RW_READER);
439 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
440 if (dn->dn_maxblkid == 0) {
441 if (off == 0 && len >= dn->dn_datablksz) {
442 blkid = 0;
443 nblks = 1;
444 } else {
445 rw_exit(&dn->dn_struct_rwlock);
446 return;
447 }
448 } else {
449 blkid = off >> dn->dn_datablkshift;
450 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
451
452 if (blkid > dn->dn_maxblkid) {
453 rw_exit(&dn->dn_struct_rwlock);
454 return;
455 }
456 if (blkid + nblks > dn->dn_maxblkid)
457 nblks = dn->dn_maxblkid - blkid + 1;
458
459 }
460 l0span = nblks; /* save for later use to calc level > 1 overhead */
461 if (dn->dn_nlevels == 1) {
462 int i;
463 for (i = 0; i < nblks; i++) {
464 blkptr_t *bp = dn->dn_phys->dn_blkptr;
465 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
466 bp += blkid + i;
467 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
468 dprintf_bp(bp, "can free old%s", "");
469 space += bp_get_dsize(spa, bp);
470 }
471 unref += BP_GET_ASIZE(bp);
472 }
473 nl1blks = 1;
474 nblks = 0;
475 }
476
477 lastblk = blkid + nblks - 1;
478 while (nblks) {
479 dmu_buf_impl_t *dbuf;
480 uint64_t ibyte, new_blkid;
481 int epb = 1 << epbs;
482 int err, i, blkoff, tochk;
483 blkptr_t *bp;
484
485 ibyte = blkid << dn->dn_datablkshift;
486 err = dnode_next_offset(dn,
487 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
488 new_blkid = ibyte >> dn->dn_datablkshift;
489 if (err == ESRCH) {
490 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
491 break;
492 }
493 if (err) {
494 txh->txh_tx->tx_err = err;
495 break;
496 }
497 if (new_blkid > lastblk) {
498 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
499 break;
500 }
501
502 if (new_blkid > blkid) {
503 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
504 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
505 nblks -= new_blkid - blkid;
506 blkid = new_blkid;
507 }
508 blkoff = P2PHASE(blkid, epb);
509 tochk = MIN(epb - blkoff, nblks);
510
511 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
512 if (err) {
513 txh->txh_tx->tx_err = err;
514 break;
515 }
516
517 txh->txh_memory_tohold += dbuf->db.db_size;
518
519 /*
520 * We don't check memory_tohold against DMU_MAX_ACCESS because
521 * memory_tohold is an over-estimation (especially the >L1
522 * indirect blocks), so it could fail. Callers should have
523 * already verified that they will not be holding too much
524 * memory.
525 */
526
527 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
528 if (err != 0) {
529 txh->txh_tx->tx_err = err;
530 dbuf_rele(dbuf, FTAG);
531 break;
532 }
533
534 bp = dbuf->db.db_data;
535 bp += blkoff;
536
537 for (i = 0; i < tochk; i++) {
538 if (dsl_dataset_block_freeable(ds, &bp[i],
539 bp[i].blk_birth)) {
540 dprintf_bp(&bp[i], "can free old%s", "");
541 space += bp_get_dsize(spa, &bp[i]);
542 }
543 unref += BP_GET_ASIZE(bp);
544 }
545 dbuf_rele(dbuf, FTAG);
546
547 ++nl1blks;
548 blkid += tochk;
549 nblks -= tochk;
550 }
551 rw_exit(&dn->dn_struct_rwlock);
552
553 /*
554 * Add in memory requirements of higher-level indirects.
555 * This assumes a worst-possible scenario for dn_nlevels and a
556 * worst-possible distribution of l1-blocks over the region to free.
557 */
558 {
559 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
560 int level = 2;
561 /*
562 * Here we don't use DN_MAX_LEVEL, but calculate it with the
563 * given datablkshift and indblkshift. This makes the
564 * difference between 19 and 8 on large files.
565 */
566 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
567 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
568
569 while (level++ < maxlevel) {
570 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
571 << dn->dn_indblkshift;
572 blkcnt = 1 + (blkcnt >> epbs);
573 }
574 }
575
576 /* account for new level 1 indirect blocks that might show up */
577 if (skipped > 0) {
578 txh->txh_fudge += skipped << dn->dn_indblkshift;
579 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
580 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
581 }
582 txh->txh_space_tofree += space;
583 txh->txh_space_tounref += unref;
584 }
585
586 void
587 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
588 {
589 dmu_tx_hold_t *txh;
590 dnode_t *dn;
591 int err;
592 zio_t *zio;
593
594 ASSERT(tx->tx_txg == 0);
595
596 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
597 object, THT_FREE, off, len);
598 if (txh == NULL)
599 return;
600 dn = txh->txh_dnode;
601 dmu_tx_count_dnode(txh);
602
603 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
604 return;
605 if (len == DMU_OBJECT_END)
606 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
607
608
609 /*
610 * For i/o error checking, we read the first and last level-0
611 * blocks if they are not aligned, and all the level-1 blocks.
612 *
613 * Note: dbuf_free_range() assumes that we have not instantiated
614 * any level-0 dbufs that will be completely freed. Therefore we must
615 * exercise care to not read or count the first and last blocks
616 * if they are blocksize-aligned.
617 */
618 if (dn->dn_datablkshift == 0) {
619 if (off != 0 || len < dn->dn_datablksz)
620 dmu_tx_count_write(txh, off, len);
621 } else {
622 /* first block will be modified if it is not aligned */
623 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
624 dmu_tx_count_write(txh, off, 1);
625 /* last block will be modified if it is not aligned */
626 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
627 dmu_tx_count_write(txh, off+len, 1);
628 }
629
630 /*
631 * Check level-1 blocks.
632 */
633 if (dn->dn_nlevels > 1) {
634 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
635 SPA_BLKPTRSHIFT;
636 uint64_t start = off >> shift;
637 uint64_t end = (off + len) >> shift;
638
639 ASSERT(dn->dn_datablkshift != 0);
640 ASSERT(dn->dn_indblkshift != 0);
641
642 zio = zio_root(tx->tx_pool->dp_spa,
643 NULL, NULL, ZIO_FLAG_CANFAIL);
644 for (uint64_t i = start; i <= end; i++) {
645 uint64_t ibyte = i << shift;
646 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
647 i = ibyte >> shift;
648 if (err == ESRCH)
649 break;
650 if (err) {
651 tx->tx_err = err;
652 return;
653 }
654
655 err = dmu_tx_check_ioerr(zio, dn, 1, i);
656 if (err) {
657 tx->tx_err = err;
658 return;
659 }
660 }
661 err = zio_wait(zio);
662 if (err) {
663 tx->tx_err = err;
664 return;
665 }
666 }
667
668 dmu_tx_count_free(txh, off, len);
669 }
670
671 void
672 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
673 {
674 dmu_tx_hold_t *txh;
675 dnode_t *dn;
676 uint64_t nblocks;
677 int epbs, err;
678
679 ASSERT(tx->tx_txg == 0);
680
681 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
682 object, THT_ZAP, add, (uintptr_t)name);
683 if (txh == NULL)
684 return;
685 dn = txh->txh_dnode;
686
687 dmu_tx_count_dnode(txh);
688
689 if (dn == NULL) {
690 /*
691 * We will be able to fit a new object's entries into one leaf
692 * block. So there will be at most 2 blocks total,
693 * including the header block.
694 */
695 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
696 return;
697 }
698
699 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
700
701 if (dn->dn_maxblkid == 0 && !add) {
702 blkptr_t *bp;
703
704 /*
705 * If there is only one block (i.e. this is a micro-zap)
706 * and we are not adding anything, the accounting is simple.
707 */
708 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
709 if (err) {
710 tx->tx_err = err;
711 return;
712 }
713
714 /*
715 * Use max block size here, since we don't know how much
716 * the size will change between now and the dbuf dirty call.
717 */
718 bp = &dn->dn_phys->dn_blkptr[0];
719 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
720 bp, bp->blk_birth))
721 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
722 else
723 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
724 if (!BP_IS_HOLE(bp))
725 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
726 return;
727 }
728
729 if (dn->dn_maxblkid > 0 && name) {
730 /*
731 * access the name in this fat-zap so that we'll check
732 * for i/o errors to the leaf blocks, etc.
733 */
734 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
735 8, 0, NULL);
736 if (err == EIO) {
737 tx->tx_err = err;
738 return;
739 }
740 }
741
742 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
743 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
744
745 /*
746 * If the modified blocks are scattered to the four winds,
747 * we'll have to modify an indirect twig for each.
748 */
749 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
750 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
751 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
752 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
753 else
754 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
755 }
756
757 void
758 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
759 {
760 dmu_tx_hold_t *txh;
761
762 ASSERT(tx->tx_txg == 0);
763
764 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
765 object, THT_BONUS, 0, 0);
766 if (txh)
767 dmu_tx_count_dnode(txh);
768 }
769
770 void
771 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
772 {
773 dmu_tx_hold_t *txh;
774 ASSERT(tx->tx_txg == 0);
775
776 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
777 DMU_NEW_OBJECT, THT_SPACE, space, 0);
778
779 txh->txh_space_towrite += space;
780 }
781
782 int
783 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
784 {
785 dmu_tx_hold_t *txh;
786 int holds = 0;
787
788 /*
789 * By asserting that the tx is assigned, we're counting the
790 * number of dn_tx_holds, which is the same as the number of
791 * dn_holds. Otherwise, we'd be counting dn_holds, but
792 * dn_tx_holds could be 0.
793 */
794 ASSERT(tx->tx_txg != 0);
795
796 /* if (tx->tx_anyobj == TRUE) */
797 /* return (0); */
798
799 for (txh = list_head(&tx->tx_holds); txh;
800 txh = list_next(&tx->tx_holds, txh)) {
801 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
802 holds++;
803 }
804
805 return (holds);
806 }
807
808 #ifdef ZFS_DEBUG
809 void
810 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
811 {
812 dmu_tx_hold_t *txh;
813 int match_object = FALSE, match_offset = FALSE;
814 dnode_t *dn;
815
816 DB_DNODE_ENTER(db);
817 dn = DB_DNODE(db);
818 ASSERT(tx->tx_txg != 0);
819 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
820 ASSERT3U(dn->dn_object, ==, db->db.db_object);
821
822 if (tx->tx_anyobj) {
823 DB_DNODE_EXIT(db);
824 return;
825 }
826
827 /* XXX No checking on the meta dnode for now */
828 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
829 DB_DNODE_EXIT(db);
830 return;
831 }
832
833 for (txh = list_head(&tx->tx_holds); txh;
834 txh = list_next(&tx->tx_holds, txh)) {
835 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
836 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
837 match_object = TRUE;
838 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
839 int datablkshift = dn->dn_datablkshift ?
840 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
841 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
842 int shift = datablkshift + epbs * db->db_level;
843 uint64_t beginblk = shift >= 64 ? 0 :
844 (txh->txh_arg1 >> shift);
845 uint64_t endblk = shift >= 64 ? 0 :
846 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
847 uint64_t blkid = db->db_blkid;
848
849 /* XXX txh_arg2 better not be zero... */
850
851 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
852 txh->txh_type, beginblk, endblk);
853
854 switch (txh->txh_type) {
855 case THT_WRITE:
856 if (blkid >= beginblk && blkid <= endblk)
857 match_offset = TRUE;
858 /*
859 * We will let this hold work for the bonus
860 * or spill buffer so that we don't need to
861 * hold it when creating a new object.
862 */
863 if (blkid == DMU_BONUS_BLKID ||
864 blkid == DMU_SPILL_BLKID)
865 match_offset = TRUE;
866 /*
867 * They might have to increase nlevels,
868 * thus dirtying the new TLIBs. Or the
869 * might have to change the block size,
870 * thus dirying the new lvl=0 blk=0.
871 */
872 if (blkid == 0)
873 match_offset = TRUE;
874 break;
875 case THT_FREE:
876 /*
877 * We will dirty all the level 1 blocks in
878 * the free range and perhaps the first and
879 * last level 0 block.
880 */
881 if (blkid >= beginblk && (blkid <= endblk ||
882 txh->txh_arg2 == DMU_OBJECT_END))
883 match_offset = TRUE;
884 break;
885 case THT_SPILL:
886 if (blkid == DMU_SPILL_BLKID)
887 match_offset = TRUE;
888 break;
889 case THT_BONUS:
890 if (blkid == DMU_BONUS_BLKID)
891 match_offset = TRUE;
892 break;
893 case THT_ZAP:
894 match_offset = TRUE;
895 break;
896 case THT_NEWOBJECT:
897 match_object = TRUE;
898 break;
899 default:
900 ASSERT(!"bad txh_type");
901 }
902 }
903 if (match_object && match_offset) {
904 DB_DNODE_EXIT(db);
905 return;
906 }
907 }
908 DB_DNODE_EXIT(db);
909 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
910 (u_longlong_t)db->db.db_object, db->db_level,
911 (u_longlong_t)db->db_blkid);
912 }
913 #endif
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
1065 static int
1066 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1067 {
1068 dmu_tx_hold_t *txh;
1069 spa_t *spa = tx->tx_pool->dp_spa;
1070 uint64_t memory, asize, fsize, usize;
1071 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1072
1073 ASSERT0(tx->tx_txg);
1074
1075 if (tx->tx_err)
1076 return (tx->tx_err);
1077
1078 if (spa_suspended(spa)) {
1079 /*
1080 * If the user has indicated a blocking failure mode
1081 * then return ERESTART which will block in dmu_tx_wait().
1082 * Otherwise, return EIO so that an error can get
1083 * propagated back to the VOP calls.
1084 *
1085 * Note that we always honor the txg_how flag regardless
1086 * of the failuremode setting.
1087 */
1088 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1089 txg_how != TXG_WAIT)
1090 return (SET_ERROR(EIO));
1091
1092 return (SET_ERROR(ERESTART));
1093 }
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
1101 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1102 tx->tx_needassign_txh = NULL;
1103
1104 /*
1105 * NB: No error returns are allowed after txg_hold_open, but
1106 * before processing the dnode holds, due to the
1107 * dmu_tx_unassign() logic.
1108 */
1109
1110 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1111 for (txh = list_head(&tx->tx_holds); txh;
1112 txh = list_next(&tx->tx_holds, txh)) {
1113 dnode_t *dn = txh->txh_dnode;
1114 if (dn != NULL) {
1115 mutex_enter(&dn->dn_mtx);
1116 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1117 mutex_exit(&dn->dn_mtx);
1118 tx->tx_needassign_txh = txh;
1119 return (SET_ERROR(ERESTART));
1120 }
1121 if (dn->dn_assigned_txg == 0)
1122 dn->dn_assigned_txg = tx->tx_txg;
1123 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1124 (void) refcount_add(&dn->dn_tx_holds, tx);
1125 mutex_exit(&dn->dn_mtx);
1126 }
1127 towrite += txh->txh_space_towrite;
1128 tofree += txh->txh_space_tofree;
1129 tooverwrite += txh->txh_space_tooverwrite;
1130 tounref += txh->txh_space_tounref;
1131 tohold += txh->txh_memory_tohold;
1132 fudge += txh->txh_fudge;
1133 }
1134
1135 /*
1136 * If a snapshot has been taken since we made our estimates,
1137 * assume that we won't be able to free or overwrite anything.
1138 */
1139 if (tx->tx_objset &&
1140 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1141 tx->tx_lastsnap_txg) {
1142 towrite += tooverwrite;
1143 tooverwrite = tofree = 0;
1144 }
1145
1146 /* needed allocation: worst-case estimate of write space */
1147 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1148 /* freed space estimate: worst-case overwrite + free estimate */
1149 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1150 /* convert unrefd space to worst-case estimate */
1151 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1152 /* calculate memory footprint estimate */
1153 memory = towrite + tooverwrite + tohold;
1154
1155 #ifdef ZFS_DEBUG
1156 /*
1157 * Add in 'tohold' to account for our dirty holds on this memory
1158 * XXX - the "fudge" factor is to account for skipped blocks that
1159 * we missed because dnode_next_offset() misses in-core-only blocks.
1160 */
1161 tx->tx_space_towrite = asize +
1162 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1163 tx->tx_space_tofree = tofree;
1164 tx->tx_space_tooverwrite = tooverwrite;
1165 tx->tx_space_tounref = tounref;
1166 #endif
1167
1168 if (tx->tx_dir && asize != 0) {
1169 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1170 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1171 if (err)
1172 return (err);
1173 }
1174
1175 return (0);
1176 }
1177
1178 static void
1179 dmu_tx_unassign(dmu_tx_t *tx)
1180 {
1181 dmu_tx_hold_t *txh;
1182
1183 if (tx->tx_txg == 0)
1184 return;
1185
1186 txg_rele_to_quiesce(&tx->tx_txgh);
1187
1188 /*
1189 * Walk the transaction's hold list, removing the hold on the
1190 * associated dnode, and notifying waiters if the refcount drops to 0.
1191 */
1192 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1193 txh = list_next(&tx->tx_holds, txh)) {
1194 dnode_t *dn = txh->txh_dnode;
1195
1196 if (dn == NULL)
1197 continue;
1198 mutex_enter(&dn->dn_mtx);
1199 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1200
1201 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1202 dn->dn_assigned_txg = 0;
1203 cv_broadcast(&dn->dn_notxholds);
1204 }
1205 mutex_exit(&dn->dn_mtx);
1206 }
1207
1208 txg_rele_to_sync(&tx->tx_txgh);
1209
1210 tx->tx_lasttried_txg = tx->tx_txg;
1211 tx->tx_txg = 0;
1212 }
1213
1214 /*
1215 * Assign tx to a transaction group. txg_how can be one of:
1216 *
1217 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1218 * a new one. This should be used when you're not holding locks.
1219 * It will only fail if we're truly out of space (or over quota).
1220 *
1221 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1222 * blocking, returns immediately with ERESTART. This should be used
1223 * whenever you're holding locks. On an ERESTART error, the caller
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).
1229 */
1230 int
1231 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1232 {
1233 int err;
1234
1235 ASSERT(tx->tx_txg == 0);
1236 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1237 txg_how == TXG_WAITED);
1238 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1239
1240 /* If we might wait, we must not hold the config lock. */
1241 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1242
1243 if (txg_how == TXG_WAITED)
1244 tx->tx_waited = B_TRUE;
1245
1246 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1247 dmu_tx_unassign(tx);
1248
1249 if (err != ERESTART || txg_how != TXG_WAIT)
1250 return (err);
1251
1252 dmu_tx_wait(tx);
1253 }
1254
1255 txg_rele_to_quiesce(&tx->tx_txgh);
1256
1257 return (0);
1258 }
1259
1260 void
1261 dmu_tx_wait(dmu_tx_t *tx)
1262 {
1263 spa_t *spa = tx->tx_pool->dp_spa;
1264 dsl_pool_t *dp = tx->tx_pool;
1265
1266 ASSERT(tx->tx_txg == 0);
1267 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1268
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);
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 */
1306 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1307
1308 mutex_enter(&dn->dn_mtx);
1309 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1310 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1311 mutex_exit(&dn->dn_mtx);
1312 tx->tx_needassign_txh = NULL;
1313 } else {
1314 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1315 }
1316 }
1317
1318 void
1319 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1320 {
1321 #ifdef ZFS_DEBUG
1322 if (tx->tx_dir == NULL || delta == 0)
1323 return;
1324
1325 if (delta > 0) {
1326 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1327 tx->tx_space_towrite);
1328 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1329 } else {
1330 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1331 }
1332 #endif
1333 }
1334
1335 void
1336 dmu_tx_commit(dmu_tx_t *tx)
1337 {
1338 dmu_tx_hold_t *txh;
1339
1340 ASSERT(tx->tx_txg != 0);
1341
1342 /*
1343 * Go through the transaction's hold list and remove holds on
1344 * associated dnodes, notifying waiters if no holds remain.
1345 */
1346 while (txh = list_head(&tx->tx_holds)) {
1347 dnode_t *dn = txh->txh_dnode;
1348
1349 list_remove(&tx->tx_holds, txh);
1350 kmem_free(txh, sizeof (dmu_tx_hold_t));
1351 if (dn == NULL)
1352 continue;
1353 mutex_enter(&dn->dn_mtx);
1354 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1355
1356 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1357 dn->dn_assigned_txg = 0;
1358 cv_broadcast(&dn->dn_notxholds);
1359 }
1360 mutex_exit(&dn->dn_mtx);
1361 dnode_rele(dn, tx);
1362 }
1363
1364 if (tx->tx_tempreserve_cookie)
1365 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1366
1367 if (!list_is_empty(&tx->tx_callbacks))
1368 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1369
1370 if (tx->tx_anyobj == FALSE)
1371 txg_rele_to_sync(&tx->tx_txgh);
1372
1373 list_destroy(&tx->tx_callbacks);
1374 list_destroy(&tx->tx_holds);
1375 #ifdef ZFS_DEBUG
1376 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1377 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1378 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1379 refcount_destroy_many(&tx->tx_space_written,
1380 refcount_count(&tx->tx_space_written));
1381 refcount_destroy_many(&tx->tx_space_freed,
1382 refcount_count(&tx->tx_space_freed));
1383 #endif
1384 kmem_free(tx, sizeof (dmu_tx_t));
1385 }
1386
1387 void
1388 dmu_tx_abort(dmu_tx_t *tx)
1389 {
1390 dmu_tx_hold_t *txh;
1391
1392 ASSERT(tx->tx_txg == 0);
1393
1394 while (txh = list_head(&tx->tx_holds)) {
1395 dnode_t *dn = txh->txh_dnode;
1396
1397 list_remove(&tx->tx_holds, txh);
1398 kmem_free(txh, sizeof (dmu_tx_hold_t));
1399 if (dn != NULL)
1400 dnode_rele(dn, tx);
1401 }
1402
1403 /*
1404 * Call any registered callbacks with an error code.
1405 */
1406 if (!list_is_empty(&tx->tx_callbacks))
1407 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1408
1409 list_destroy(&tx->tx_callbacks);
1410 list_destroy(&tx->tx_holds);
1411 #ifdef ZFS_DEBUG
1412 refcount_destroy_many(&tx->tx_space_written,
1413 refcount_count(&tx->tx_space_written));
1414 refcount_destroy_many(&tx->tx_space_freed,
1415 refcount_count(&tx->tx_space_freed));
1416 #endif
1417 kmem_free(tx, sizeof (dmu_tx_t));
1418 }
1419
1420 uint64_t
1421 dmu_tx_get_txg(dmu_tx_t *tx)
1422 {
1423 ASSERT(tx->tx_txg != 0);
1424 return (tx->tx_txg);
1425 }
1426
1427 dsl_pool_t *
1428 dmu_tx_pool(dmu_tx_t *tx)
1429 {
1430 ASSERT(tx->tx_pool != NULL);
1431 return (tx->tx_pool);
1432 }
1433
1434
1435 void
1436 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1437 {
1438 dmu_tx_callback_t *dcb;
1439
1440 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1441
1442 dcb->dcb_func = func;
1443 dcb->dcb_data = data;
1444
1445 list_insert_tail(&tx->tx_callbacks, dcb);
1446 }
1447
1448 /*
1449 * Call all the commit callbacks on a list, with a given error code.
1450 */
1451 void
1452 dmu_tx_do_callbacks(list_t *cb_list, int error)
1453 {
1454 dmu_tx_callback_t *dcb;
1455
1456 while (dcb = list_head(cb_list)) {
1457 list_remove(cb_list, dcb);
1458 dcb->dcb_func(dcb->dcb_data, error);
1459 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1460 }
1461 }
1462
1463 /*
1464 * Interface to hold a bunch of attributes.
1465 * used for creating new files.
1466 * attrsize is the total size of all attributes
1467 * to be added during object creation
1468 *
1469 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1470 */
1471
1472 /*
1473 * hold necessary attribute name for attribute registration.
1474 * should be a very rare case where this is needed. If it does
1475 * happen it would only happen on the first write to the file system.
1476 */
1477 static void
1478 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1479 {
1480 int i;
1481
1482 if (!sa->sa_need_attr_registration)
1483 return;
1484
1485 for (i = 0; i != sa->sa_num_attrs; i++) {
1486 if (!sa->sa_attr_table[i].sa_registered) {
1487 if (sa->sa_reg_attr_obj)
1488 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1489 B_TRUE, sa->sa_attr_table[i].sa_name);
1490 else
1491 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1492 B_TRUE, sa->sa_attr_table[i].sa_name);
1493 }
1494 }
1495 }
1496
1497
1498 void
1499 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1500 {
1501 dnode_t *dn;
1502 dmu_tx_hold_t *txh;
1503
1504 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1505 THT_SPILL, 0, 0);
1506
1507 dn = txh->txh_dnode;
1508
1509 if (dn == NULL)
1510 return;
1511
1512 /* If blkptr doesn't exist then add space to towrite */
1513 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1514 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1515 } else {
1516 blkptr_t *bp;
1517
1518 bp = &dn->dn_phys->dn_spill;
1519 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1520 bp, bp->blk_birth))
1521 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1522 else
1523 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1524 if (!BP_IS_HOLE(bp))
1525 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1526 }
1527 }
1528
1529 void
1530 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1531 {
1532 sa_os_t *sa = tx->tx_objset->os_sa;
1533
1534 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1535
1536 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1537 return;
1538
1539 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1540 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1541 else {
1542 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1543 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1544 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1545 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1546 }
1547
1548 dmu_tx_sa_registration_hold(sa, tx);
1549
1550 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1551 return;
1552
1553 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1554 THT_SPILL, 0, 0);
1555 }
1556
1557 /*
1558 * Hold SA attribute
1559 *
1560 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1561 *
1562 * variable_size is the total size of all variable sized attributes
1563 * passed to this function. It is not the total size of all
1564 * variable size attributes that *may* exist on this object.
1565 */
1566 void
1567 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1568 {
1569 uint64_t object;
1570 sa_os_t *sa = tx->tx_objset->os_sa;
1571
1572 ASSERT(hdl != NULL);
1573
1574 object = sa_handle_object(hdl);
1575
1576 dmu_tx_hold_bonus(tx, object);
1577
1578 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1579 return;
1580
1581 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1582 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1583 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1584 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1585 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1586 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1587 }
1588
1589 dmu_tx_sa_registration_hold(sa, tx);
1590
1591 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1592 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1593
1594 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1595 ASSERT(tx->tx_txg == 0);
1596 dmu_tx_hold_spill(tx, object);
1597 } else {
1598 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1599 dnode_t *dn;
1600
1601 DB_DNODE_ENTER(db);
1602 dn = DB_DNODE(db);
1603 if (dn->dn_have_spill) {
1604 ASSERT(tx->tx_txg == 0);
1605 dmu_tx_hold_spill(tx, object);
1606 }
1607 DB_DNODE_EXIT(db);
1608 }
1609 }