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) 2012, 2015 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 *
dmu_tx_create_dd(dsl_dir_t * dd)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 *
dmu_tx_create(objset_t * os)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 *
dmu_tx_create_assigned(struct dsl_pool * dp,uint64_t txg)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
dmu_tx_is_syncing(dmu_tx_t * tx)88 dmu_tx_is_syncing(dmu_tx_t *tx)
89 {
90 return (tx->tx_anyobj);
91 }
92
93 int
dmu_tx_private_ok(dmu_tx_t * tx)94 dmu_tx_private_ok(dmu_tx_t *tx)
95 {
96 return (tx->tx_anyobj);
97 }
98
99 static dmu_tx_hold_t *
dmu_tx_hold_object_impl(dmu_tx_t * tx,objset_t * os,uint64_t object,enum dmu_tx_hold_type type,uint64_t arg1,uint64_t arg2)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
dmu_tx_add_new_object(dmu_tx_t * tx,objset_t * os,uint64_t object)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
dmu_tx_check_ioerr(zio_t * zio,dnode_t * dn,int level,uint64_t blkid)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
dmu_tx_count_twig(dmu_tx_hold_t * txh,dnode_t * dn,dmu_buf_impl_t * db,int level,uint64_t blkid,boolean_t freeable,uint64_t * history)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
dmu_tx_count_write(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)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 = highbit64(txh->txh_tx->tx_objset->os_recordsize) - 1;
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 } else {
297 /*
298 * The blocksize can increase up to the recordsize,
299 * or if it is already more than the recordsize,
300 * up to the next power of 2.
301 */
302 min_bs = highbit64(dn->dn_datablksz - 1);
303 max_bs = MAX(max_bs, highbit64(dn->dn_datablksz - 1));
304 }
305
306 /*
307 * If this write is not off the end of the file
308 * we need to account for overwrites/unref.
309 */
310 if (start <= dn->dn_maxblkid) {
311 for (int l = 0; l < DN_MAX_LEVELS; l++)
312 history[l] = -1ULL;
313 }
314 while (start <= dn->dn_maxblkid) {
315 dmu_buf_impl_t *db;
316
317 rw_enter(&dn->dn_struct_rwlock, RW_READER);
318 err = dbuf_hold_impl(dn, 0, start,
319 FALSE, FALSE, FTAG, &db);
320 rw_exit(&dn->dn_struct_rwlock);
321
322 if (err) {
323 txh->txh_tx->tx_err = err;
324 return;
325 }
326
327 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
328 history);
329 dbuf_rele(db, FTAG);
330 if (++start > end) {
331 /*
332 * Account for new indirects appearing
333 * before this IO gets assigned into a txg.
334 */
335 bits = 64 - min_bs;
336 epbs = min_ibs - SPA_BLKPTRSHIFT;
337 for (bits -= epbs * (nlvls - 1);
338 bits >= 0; bits -= epbs)
339 txh->txh_fudge += 1ULL << max_ibs;
340 goto out;
341 }
342 off += delta;
343 if (len >= delta)
344 len -= delta;
345 delta = dn->dn_datablksz;
346 }
347 }
348
349 /*
350 * 'end' is the last thing we will access, not one past.
351 * This way we won't overflow when accessing the last byte.
352 */
353 start = P2ALIGN(off, 1ULL << max_bs);
354 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
355 txh->txh_space_towrite += end - start + 1;
356
357 start >>= min_bs;
358 end >>= min_bs;
359
360 epbs = min_ibs - SPA_BLKPTRSHIFT;
361
362 /*
363 * The object contains at most 2^(64 - min_bs) blocks,
364 * and each indirect level maps 2^epbs.
365 */
366 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
367 start >>= epbs;
368 end >>= epbs;
369 ASSERT3U(end, >=, start);
370 txh->txh_space_towrite += (end - start + 1) << max_ibs;
371 if (start != 0) {
372 /*
373 * We also need a new blkid=0 indirect block
374 * to reference any existing file data.
375 */
376 txh->txh_space_towrite += 1ULL << max_ibs;
377 }
378 }
379
380 out:
381 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
382 2 * DMU_MAX_ACCESS)
383 err = SET_ERROR(EFBIG);
384
385 if (err)
386 txh->txh_tx->tx_err = err;
387 }
388
389 static void
dmu_tx_count_dnode(dmu_tx_hold_t * txh)390 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
391 {
392 dnode_t *dn = txh->txh_dnode;
393 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
394 uint64_t space = mdn->dn_datablksz +
395 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
396
397 if (dn && dn->dn_dbuf->db_blkptr &&
398 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
399 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
400 txh->txh_space_tooverwrite += space;
401 txh->txh_space_tounref += space;
402 } else {
403 txh->txh_space_towrite += space;
404 if (dn && dn->dn_dbuf->db_blkptr)
405 txh->txh_space_tounref += space;
406 }
407 }
408
409 void
dmu_tx_hold_write(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)410 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
411 {
412 dmu_tx_hold_t *txh;
413
414 ASSERT(tx->tx_txg == 0);
415 ASSERT(len < DMU_MAX_ACCESS);
416 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
417
418 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
419 object, THT_WRITE, off, len);
420 if (txh == NULL)
421 return;
422
423 dmu_tx_count_write(txh, off, len);
424 dmu_tx_count_dnode(txh);
425 }
426
427 static void
dmu_tx_count_free(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)428 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
429 {
430 uint64_t blkid, nblks, lastblk;
431 uint64_t space = 0, unref = 0, skipped = 0;
432 dnode_t *dn = txh->txh_dnode;
433 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
434 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
435 int epbs;
436 uint64_t l0span = 0, nl1blks = 0;
437
438 if (dn->dn_nlevels == 0)
439 return;
440
441 /*
442 * The struct_rwlock protects us against dn_nlevels
443 * changing, in case (against all odds) we manage to dirty &
444 * sync out the changes after we check for being dirty.
445 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
446 */
447 rw_enter(&dn->dn_struct_rwlock, RW_READER);
448 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
449 if (dn->dn_maxblkid == 0) {
450 if (off == 0 && len >= dn->dn_datablksz) {
451 blkid = 0;
452 nblks = 1;
453 } else {
454 rw_exit(&dn->dn_struct_rwlock);
455 return;
456 }
457 } else {
458 blkid = off >> dn->dn_datablkshift;
459 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
460
461 if (blkid > dn->dn_maxblkid) {
462 rw_exit(&dn->dn_struct_rwlock);
463 return;
464 }
465 if (blkid + nblks > dn->dn_maxblkid)
466 nblks = dn->dn_maxblkid - blkid + 1;
467
468 }
469 l0span = nblks; /* save for later use to calc level > 1 overhead */
470 if (dn->dn_nlevels == 1) {
471 int i;
472 for (i = 0; i < nblks; i++) {
473 blkptr_t *bp = dn->dn_phys->dn_blkptr;
474 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
475 bp += blkid + i;
476 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
477 dprintf_bp(bp, "can free old%s", "");
478 space += bp_get_dsize(spa, bp);
479 }
480 unref += BP_GET_ASIZE(bp);
481 }
482 nl1blks = 1;
483 nblks = 0;
484 }
485
486 lastblk = blkid + nblks - 1;
487 while (nblks) {
488 dmu_buf_impl_t *dbuf;
489 uint64_t ibyte, new_blkid;
490 int epb = 1 << epbs;
491 int err, i, blkoff, tochk;
492 blkptr_t *bp;
493
494 ibyte = blkid << dn->dn_datablkshift;
495 err = dnode_next_offset(dn,
496 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
497 new_blkid = ibyte >> dn->dn_datablkshift;
498 if (err == ESRCH) {
499 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
500 break;
501 }
502 if (err) {
503 txh->txh_tx->tx_err = err;
504 break;
505 }
506 if (new_blkid > lastblk) {
507 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
508 break;
509 }
510
511 if (new_blkid > blkid) {
512 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
513 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
514 nblks -= new_blkid - blkid;
515 blkid = new_blkid;
516 }
517 blkoff = P2PHASE(blkid, epb);
518 tochk = MIN(epb - blkoff, nblks);
519
520 err = dbuf_hold_impl(dn, 1, blkid >> epbs,
521 FALSE, FALSE, FTAG, &dbuf);
522 if (err) {
523 txh->txh_tx->tx_err = err;
524 break;
525 }
526
527 txh->txh_memory_tohold += dbuf->db.db_size;
528
529 /*
530 * We don't check memory_tohold against DMU_MAX_ACCESS because
531 * memory_tohold is an over-estimation (especially the >L1
532 * indirect blocks), so it could fail. Callers should have
533 * already verified that they will not be holding too much
534 * memory.
535 */
536
537 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
538 if (err != 0) {
539 txh->txh_tx->tx_err = err;
540 dbuf_rele(dbuf, FTAG);
541 break;
542 }
543
544 bp = dbuf->db.db_data;
545 bp += blkoff;
546
547 for (i = 0; i < tochk; i++) {
548 if (dsl_dataset_block_freeable(ds, &bp[i],
549 bp[i].blk_birth)) {
550 dprintf_bp(&bp[i], "can free old%s", "");
551 space += bp_get_dsize(spa, &bp[i]);
552 }
553 unref += BP_GET_ASIZE(bp);
554 }
555 dbuf_rele(dbuf, FTAG);
556
557 ++nl1blks;
558 blkid += tochk;
559 nblks -= tochk;
560 }
561 rw_exit(&dn->dn_struct_rwlock);
562
563 /*
564 * Add in memory requirements of higher-level indirects.
565 * This assumes a worst-possible scenario for dn_nlevels and a
566 * worst-possible distribution of l1-blocks over the region to free.
567 */
568 {
569 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
570 int level = 2;
571 /*
572 * Here we don't use DN_MAX_LEVEL, but calculate it with the
573 * given datablkshift and indblkshift. This makes the
574 * difference between 19 and 8 on large files.
575 */
576 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
577 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
578
579 while (level++ < maxlevel) {
580 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
581 << dn->dn_indblkshift;
582 blkcnt = 1 + (blkcnt >> epbs);
583 }
584 }
585
586 /* account for new level 1 indirect blocks that might show up */
587 if (skipped > 0) {
588 txh->txh_fudge += skipped << dn->dn_indblkshift;
589 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
590 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
591 }
592 txh->txh_space_tofree += space;
593 txh->txh_space_tounref += unref;
594 }
595
596 /*
597 * This function marks the transaction as being a "net free". The end
598 * result is that refquotas will be disabled for this transaction, and
599 * this transaction will be able to use half of the pool space overhead
600 * (see dsl_pool_adjustedsize()). Therefore this function should only
601 * be called for transactions that we expect will not cause a net increase
602 * in the amount of space used (but it's OK if that is occasionally not true).
603 */
604 void
dmu_tx_mark_netfree(dmu_tx_t * tx)605 dmu_tx_mark_netfree(dmu_tx_t *tx)
606 {
607 dmu_tx_hold_t *txh;
608
609 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
610 DMU_NEW_OBJECT, THT_FREE, 0, 0);
611
612 /*
613 * Pretend that this operation will free 1GB of space. This
614 * should be large enough to cancel out the largest write.
615 * We don't want to use something like UINT64_MAX, because that would
616 * cause overflows when doing math with these values (e.g. in
617 * dmu_tx_try_assign()).
618 */
619 txh->txh_space_tofree = txh->txh_space_tounref = 1024 * 1024 * 1024;
620 }
621
622 void
dmu_tx_hold_free(dmu_tx_t * tx,uint64_t object,uint64_t off,uint64_t len)623 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
624 {
625 dmu_tx_hold_t *txh;
626 dnode_t *dn;
627 int err;
628 zio_t *zio;
629
630 ASSERT(tx->tx_txg == 0);
631
632 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
633 object, THT_FREE, off, len);
634 if (txh == NULL)
635 return;
636 dn = txh->txh_dnode;
637 dmu_tx_count_dnode(txh);
638
639 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
640 return;
641 if (len == DMU_OBJECT_END)
642 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
643
644
645 /*
646 * For i/o error checking, we read the first and last level-0
647 * blocks if they are not aligned, and all the level-1 blocks.
648 *
649 * Note: dbuf_free_range() assumes that we have not instantiated
650 * any level-0 dbufs that will be completely freed. Therefore we must
651 * exercise care to not read or count the first and last blocks
652 * if they are blocksize-aligned.
653 */
654 if (dn->dn_datablkshift == 0) {
655 if (off != 0 || len < dn->dn_datablksz)
656 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
657 } else {
658 /* first block will be modified if it is not aligned */
659 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
660 dmu_tx_count_write(txh, off, 1);
661 /* last block will be modified if it is not aligned */
662 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
663 dmu_tx_count_write(txh, off+len, 1);
664 }
665
666 /*
667 * Check level-1 blocks.
668 */
669 if (dn->dn_nlevels > 1) {
670 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
671 SPA_BLKPTRSHIFT;
672 uint64_t start = off >> shift;
673 uint64_t end = (off + len) >> shift;
674
675 ASSERT(dn->dn_indblkshift != 0);
676
677 /*
678 * dnode_reallocate() can result in an object with indirect
679 * blocks having an odd data block size. In this case,
680 * just check the single block.
681 */
682 if (dn->dn_datablkshift == 0)
683 start = end = 0;
684
685 zio = zio_root(tx->tx_pool->dp_spa,
686 NULL, NULL, ZIO_FLAG_CANFAIL);
687 for (uint64_t i = start; i <= end; i++) {
688 uint64_t ibyte = i << shift;
689 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
690 i = ibyte >> shift;
691 if (err == ESRCH || i > end)
692 break;
693 if (err) {
694 tx->tx_err = err;
695 return;
696 }
697
698 err = dmu_tx_check_ioerr(zio, dn, 1, i);
699 if (err) {
700 tx->tx_err = err;
701 return;
702 }
703 }
704 err = zio_wait(zio);
705 if (err) {
706 tx->tx_err = err;
707 return;
708 }
709 }
710
711 dmu_tx_count_free(txh, off, len);
712 }
713
714 void
dmu_tx_hold_zap(dmu_tx_t * tx,uint64_t object,int add,const char * name)715 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
716 {
717 dmu_tx_hold_t *txh;
718 dnode_t *dn;
719 dsl_dataset_phys_t *ds_phys;
720 uint64_t nblocks;
721 int epbs, err;
722
723 ASSERT(tx->tx_txg == 0);
724
725 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
726 object, THT_ZAP, add, (uintptr_t)name);
727 if (txh == NULL)
728 return;
729 dn = txh->txh_dnode;
730
731 dmu_tx_count_dnode(txh);
732
733 if (dn == NULL) {
734 /*
735 * We will be able to fit a new object's entries into one leaf
736 * block. So there will be at most 2 blocks total,
737 * including the header block.
738 */
739 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
740 return;
741 }
742
743 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
744
745 if (dn->dn_maxblkid == 0 && !add) {
746 blkptr_t *bp;
747
748 /*
749 * If there is only one block (i.e. this is a micro-zap)
750 * and we are not adding anything, the accounting is simple.
751 */
752 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
753 if (err) {
754 tx->tx_err = err;
755 return;
756 }
757
758 /*
759 * Use max block size here, since we don't know how much
760 * the size will change between now and the dbuf dirty call.
761 */
762 bp = &dn->dn_phys->dn_blkptr[0];
763 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
764 bp, bp->blk_birth))
765 txh->txh_space_tooverwrite += MZAP_MAX_BLKSZ;
766 else
767 txh->txh_space_towrite += MZAP_MAX_BLKSZ;
768 if (!BP_IS_HOLE(bp))
769 txh->txh_space_tounref += MZAP_MAX_BLKSZ;
770 return;
771 }
772
773 if (dn->dn_maxblkid > 0 && name) {
774 /*
775 * access the name in this fat-zap so that we'll check
776 * for i/o errors to the leaf blocks, etc.
777 */
778 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
779 8, 0, NULL);
780 if (err == EIO) {
781 tx->tx_err = err;
782 return;
783 }
784 }
785
786 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
787 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
788
789 /*
790 * If the modified blocks are scattered to the four winds,
791 * we'll have to modify an indirect twig for each.
792 */
793 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
794 ds_phys = dsl_dataset_phys(dn->dn_objset->os_dsl_dataset);
795 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
796 if (ds_phys->ds_prev_snap_obj)
797 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
798 else
799 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
800 }
801
802 void
dmu_tx_hold_bonus(dmu_tx_t * tx,uint64_t object)803 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
804 {
805 dmu_tx_hold_t *txh;
806
807 ASSERT(tx->tx_txg == 0);
808
809 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
810 object, THT_BONUS, 0, 0);
811 if (txh)
812 dmu_tx_count_dnode(txh);
813 }
814
815 void
dmu_tx_hold_space(dmu_tx_t * tx,uint64_t space)816 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
817 {
818 dmu_tx_hold_t *txh;
819 ASSERT(tx->tx_txg == 0);
820
821 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
822 DMU_NEW_OBJECT, THT_SPACE, space, 0);
823
824 txh->txh_space_towrite += space;
825 }
826
827 int
dmu_tx_holds(dmu_tx_t * tx,uint64_t object)828 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
829 {
830 dmu_tx_hold_t *txh;
831 int holds = 0;
832
833 /*
834 * By asserting that the tx is assigned, we're counting the
835 * number of dn_tx_holds, which is the same as the number of
836 * dn_holds. Otherwise, we'd be counting dn_holds, but
837 * dn_tx_holds could be 0.
838 */
839 ASSERT(tx->tx_txg != 0);
840
841 /* if (tx->tx_anyobj == TRUE) */
842 /* return (0); */
843
844 for (txh = list_head(&tx->tx_holds); txh;
845 txh = list_next(&tx->tx_holds, txh)) {
846 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
847 holds++;
848 }
849
850 return (holds);
851 }
852
853 #ifdef ZFS_DEBUG
854 void
dmu_tx_dirty_buf(dmu_tx_t * tx,dmu_buf_impl_t * db)855 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
856 {
857 dmu_tx_hold_t *txh;
858 int match_object = FALSE, match_offset = FALSE;
859 dnode_t *dn;
860
861 DB_DNODE_ENTER(db);
862 dn = DB_DNODE(db);
863 ASSERT(tx->tx_txg != 0);
864 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
865 ASSERT3U(dn->dn_object, ==, db->db.db_object);
866
867 if (tx->tx_anyobj) {
868 DB_DNODE_EXIT(db);
869 return;
870 }
871
872 /* XXX No checking on the meta dnode for now */
873 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
874 DB_DNODE_EXIT(db);
875 return;
876 }
877
878 for (txh = list_head(&tx->tx_holds); txh;
879 txh = list_next(&tx->tx_holds, txh)) {
880 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
881 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
882 match_object = TRUE;
883 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
884 int datablkshift = dn->dn_datablkshift ?
885 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
886 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
887 int shift = datablkshift + epbs * db->db_level;
888 uint64_t beginblk = shift >= 64 ? 0 :
889 (txh->txh_arg1 >> shift);
890 uint64_t endblk = shift >= 64 ? 0 :
891 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
892 uint64_t blkid = db->db_blkid;
893
894 /* XXX txh_arg2 better not be zero... */
895
896 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
897 txh->txh_type, beginblk, endblk);
898
899 switch (txh->txh_type) {
900 case THT_WRITE:
901 if (blkid >= beginblk && blkid <= endblk)
902 match_offset = TRUE;
903 /*
904 * We will let this hold work for the bonus
905 * or spill buffer so that we don't need to
906 * hold it when creating a new object.
907 */
908 if (blkid == DMU_BONUS_BLKID ||
909 blkid == DMU_SPILL_BLKID)
910 match_offset = TRUE;
911 /*
912 * They might have to increase nlevels,
913 * thus dirtying the new TLIBs. Or the
914 * might have to change the block size,
915 * thus dirying the new lvl=0 blk=0.
916 */
917 if (blkid == 0)
918 match_offset = TRUE;
919 break;
920 case THT_FREE:
921 /*
922 * We will dirty all the level 1 blocks in
923 * the free range and perhaps the first and
924 * last level 0 block.
925 */
926 if (blkid >= beginblk && (blkid <= endblk ||
927 txh->txh_arg2 == DMU_OBJECT_END))
928 match_offset = TRUE;
929 break;
930 case THT_SPILL:
931 if (blkid == DMU_SPILL_BLKID)
932 match_offset = TRUE;
933 break;
934 case THT_BONUS:
935 if (blkid == DMU_BONUS_BLKID)
936 match_offset = TRUE;
937 break;
938 case THT_ZAP:
939 match_offset = TRUE;
940 break;
941 case THT_NEWOBJECT:
942 match_object = TRUE;
943 break;
944 default:
945 ASSERT(!"bad txh_type");
946 }
947 }
948 if (match_object && match_offset) {
949 DB_DNODE_EXIT(db);
950 return;
951 }
952 }
953 DB_DNODE_EXIT(db);
954 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
955 (u_longlong_t)db->db.db_object, db->db_level,
956 (u_longlong_t)db->db_blkid);
957 }
958 #endif
959
960 /*
961 * If we can't do 10 iops, something is wrong. Let us go ahead
962 * and hit zfs_dirty_data_max.
963 */
964 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
965 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
966
967 /*
968 * We delay transactions when we've determined that the backend storage
969 * isn't able to accommodate the rate of incoming writes.
970 *
971 * If there is already a transaction waiting, we delay relative to when
972 * that transaction finishes waiting. This way the calculated min_time
973 * is independent of the number of threads concurrently executing
974 * transactions.
975 *
976 * If we are the only waiter, wait relative to when the transaction
977 * started, rather than the current time. This credits the transaction for
978 * "time already served", e.g. reading indirect blocks.
979 *
980 * The minimum time for a transaction to take is calculated as:
981 * min_time = scale * (dirty - min) / (max - dirty)
982 * min_time is then capped at zfs_delay_max_ns.
983 *
984 * The delay has two degrees of freedom that can be adjusted via tunables.
985 * The percentage of dirty data at which we start to delay is defined by
986 * zfs_delay_min_dirty_percent. This should typically be at or above
987 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
988 * delay after writing at full speed has failed to keep up with the incoming
989 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
990 * speaking, this variable determines the amount of delay at the midpoint of
991 * the curve.
992 *
993 * delay
994 * 10ms +-------------------------------------------------------------*+
995 * | *|
996 * 9ms + *+
997 * | *|
998 * 8ms + *+
999 * | * |
1000 * 7ms + * +
1001 * | * |
1002 * 6ms + * +
1003 * | * |
1004 * 5ms + * +
1005 * | * |
1006 * 4ms + * +
1007 * | * |
1008 * 3ms + * +
1009 * | * |
1010 * 2ms + (midpoint) * +
1011 * | | ** |
1012 * 1ms + v *** +
1013 * | zfs_delay_scale ----------> ******** |
1014 * 0 +-------------------------------------*********----------------+
1015 * 0% <- zfs_dirty_data_max -> 100%
1016 *
1017 * Note that since the delay is added to the outstanding time remaining on the
1018 * most recent transaction, the delay is effectively the inverse of IOPS.
1019 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
1020 * was chosen such that small changes in the amount of accumulated dirty data
1021 * in the first 3/4 of the curve yield relatively small differences in the
1022 * amount of delay.
1023 *
1024 * The effects can be easier to understand when the amount of delay is
1025 * represented on a log scale:
1026 *
1027 * delay
1028 * 100ms +-------------------------------------------------------------++
1029 * + +
1030 * | |
1031 * + *+
1032 * 10ms + *+
1033 * + ** +
1034 * | (midpoint) ** |
1035 * + | ** +
1036 * 1ms + v **** +
1037 * + zfs_delay_scale ----------> ***** +
1038 * | **** |
1039 * + **** +
1040 * 100us + ** +
1041 * + * +
1042 * | * |
1043 * + * +
1044 * 10us + * +
1045 * + +
1046 * | |
1047 * + +
1048 * +--------------------------------------------------------------+
1049 * 0% <- zfs_dirty_data_max -> 100%
1050 *
1051 * Note here that only as the amount of dirty data approaches its limit does
1052 * the delay start to increase rapidly. The goal of a properly tuned system
1053 * should be to keep the amount of dirty data out of that range by first
1054 * ensuring that the appropriate limits are set for the I/O scheduler to reach
1055 * optimal throughput on the backend storage, and then by changing the value
1056 * of zfs_delay_scale to increase the steepness of the curve.
1057 */
1058 static void
dmu_tx_delay(dmu_tx_t * tx,uint64_t dirty)1059 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1060 {
1061 dsl_pool_t *dp = tx->tx_pool;
1062 uint64_t delay_min_bytes =
1063 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1064 hrtime_t wakeup, min_tx_time, now;
1065
1066 if (dirty <= delay_min_bytes)
1067 return;
1068
1069 /*
1070 * The caller has already waited until we are under the max.
1071 * We make them pass us the amount of dirty data so we don't
1072 * have to handle the case of it being >= the max, which could
1073 * cause a divide-by-zero if it's == the max.
1074 */
1075 ASSERT3U(dirty, <, zfs_dirty_data_max);
1076
1077 now = gethrtime();
1078 min_tx_time = zfs_delay_scale *
1079 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1080 if (now > tx->tx_start + min_tx_time)
1081 return;
1082
1083 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1084
1085 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1086 uint64_t, min_tx_time);
1087
1088 mutex_enter(&dp->dp_lock);
1089 wakeup = MAX(tx->tx_start + min_tx_time,
1090 dp->dp_last_wakeup + min_tx_time);
1091 dp->dp_last_wakeup = wakeup;
1092 mutex_exit(&dp->dp_lock);
1093
1094 #ifdef _KERNEL
1095 #ifdef illumos
1096 mutex_enter(&curthread->t_delay_lock);
1097 while (cv_timedwait_hires(&curthread->t_delay_cv,
1098 &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
1099 CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
1100 continue;
1101 mutex_exit(&curthread->t_delay_lock);
1102 #else
1103 pause_sbt("dmu_tx_delay", wakeup * SBT_1NS,
1104 zfs_delay_resolution_ns * SBT_1NS, C_ABSOLUTE);
1105 #endif
1106 #else
1107 hrtime_t delta = wakeup - gethrtime();
1108 struct timespec ts;
1109 ts.tv_sec = delta / NANOSEC;
1110 ts.tv_nsec = delta % NANOSEC;
1111 (void) nanosleep(&ts, NULL);
1112 #endif
1113 }
1114
1115 static int
dmu_tx_try_assign(dmu_tx_t * tx,txg_how_t txg_how)1116 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1117 {
1118 dmu_tx_hold_t *txh;
1119 spa_t *spa = tx->tx_pool->dp_spa;
1120 uint64_t memory, asize, fsize, usize;
1121 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1122
1123 ASSERT0(tx->tx_txg);
1124
1125 if (tx->tx_err)
1126 return (tx->tx_err);
1127
1128 if (spa_suspended(spa)) {
1129 /*
1130 * If the user has indicated a blocking failure mode
1131 * then return ERESTART which will block in dmu_tx_wait().
1132 * Otherwise, return EIO so that an error can get
1133 * propagated back to the VOP calls.
1134 *
1135 * Note that we always honor the txg_how flag regardless
1136 * of the failuremode setting.
1137 */
1138 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1139 txg_how != TXG_WAIT)
1140 return (SET_ERROR(EIO));
1141
1142 return (SET_ERROR(ERESTART));
1143 }
1144
1145 if (!tx->tx_waited &&
1146 dsl_pool_need_dirty_delay(tx->tx_pool)) {
1147 tx->tx_wait_dirty = B_TRUE;
1148 return (SET_ERROR(ERESTART));
1149 }
1150
1151 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1152 tx->tx_needassign_txh = NULL;
1153
1154 /*
1155 * NB: No error returns are allowed after txg_hold_open, but
1156 * before processing the dnode holds, due to the
1157 * dmu_tx_unassign() logic.
1158 */
1159
1160 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1161 for (txh = list_head(&tx->tx_holds); txh;
1162 txh = list_next(&tx->tx_holds, txh)) {
1163 dnode_t *dn = txh->txh_dnode;
1164 if (dn != NULL) {
1165 mutex_enter(&dn->dn_mtx);
1166 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1167 mutex_exit(&dn->dn_mtx);
1168 tx->tx_needassign_txh = txh;
1169 return (SET_ERROR(ERESTART));
1170 }
1171 if (dn->dn_assigned_txg == 0)
1172 dn->dn_assigned_txg = tx->tx_txg;
1173 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1174 (void) refcount_add(&dn->dn_tx_holds, tx);
1175 mutex_exit(&dn->dn_mtx);
1176 }
1177 towrite += txh->txh_space_towrite;
1178 tofree += txh->txh_space_tofree;
1179 tooverwrite += txh->txh_space_tooverwrite;
1180 tounref += txh->txh_space_tounref;
1181 tohold += txh->txh_memory_tohold;
1182 fudge += txh->txh_fudge;
1183 }
1184
1185 /*
1186 * If a snapshot has been taken since we made our estimates,
1187 * assume that we won't be able to free or overwrite anything.
1188 */
1189 if (tx->tx_objset &&
1190 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1191 tx->tx_lastsnap_txg) {
1192 towrite += tooverwrite;
1193 tooverwrite = tofree = 0;
1194 }
1195
1196 /* needed allocation: worst-case estimate of write space */
1197 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1198 /* freed space estimate: worst-case overwrite + free estimate */
1199 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1200 /* convert unrefd space to worst-case estimate */
1201 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1202 /* calculate memory footprint estimate */
1203 memory = towrite + tooverwrite + tohold;
1204
1205 #ifdef ZFS_DEBUG
1206 /*
1207 * Add in 'tohold' to account for our dirty holds on this memory
1208 * XXX - the "fudge" factor is to account for skipped blocks that
1209 * we missed because dnode_next_offset() misses in-core-only blocks.
1210 */
1211 tx->tx_space_towrite = asize +
1212 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1213 tx->tx_space_tofree = tofree;
1214 tx->tx_space_tooverwrite = tooverwrite;
1215 tx->tx_space_tounref = tounref;
1216 #endif
1217
1218 if (tx->tx_dir && asize != 0) {
1219 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1220 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1221 if (err)
1222 return (err);
1223 }
1224
1225 return (0);
1226 }
1227
1228 static void
dmu_tx_unassign(dmu_tx_t * tx)1229 dmu_tx_unassign(dmu_tx_t *tx)
1230 {
1231 dmu_tx_hold_t *txh;
1232
1233 if (tx->tx_txg == 0)
1234 return;
1235
1236 txg_rele_to_quiesce(&tx->tx_txgh);
1237
1238 /*
1239 * Walk the transaction's hold list, removing the hold on the
1240 * associated dnode, and notifying waiters if the refcount drops to 0.
1241 */
1242 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1243 txh = list_next(&tx->tx_holds, txh)) {
1244 dnode_t *dn = txh->txh_dnode;
1245
1246 if (dn == NULL)
1247 continue;
1248 mutex_enter(&dn->dn_mtx);
1249 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1250
1251 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1252 dn->dn_assigned_txg = 0;
1253 cv_broadcast(&dn->dn_notxholds);
1254 }
1255 mutex_exit(&dn->dn_mtx);
1256 }
1257
1258 txg_rele_to_sync(&tx->tx_txgh);
1259
1260 tx->tx_lasttried_txg = tx->tx_txg;
1261 tx->tx_txg = 0;
1262 }
1263
1264 /*
1265 * Assign tx to a transaction group. txg_how can be one of:
1266 *
1267 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1268 * a new one. This should be used when you're not holding locks.
1269 * It will only fail if we're truly out of space (or over quota).
1270 *
1271 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1272 * blocking, returns immediately with ERESTART. This should be used
1273 * whenever you're holding locks. On an ERESTART error, the caller
1274 * should drop locks, do a dmu_tx_wait(tx), and try again.
1275 *
1276 * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1277 * has already been called on behalf of this operation (though
1278 * most likely on a different tx).
1279 */
1280 int
dmu_tx_assign(dmu_tx_t * tx,txg_how_t txg_how)1281 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1282 {
1283 int err;
1284
1285 ASSERT(tx->tx_txg == 0);
1286 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1287 txg_how == TXG_WAITED);
1288 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1289
1290 /* If we might wait, we must not hold the config lock. */
1291 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1292
1293 if (txg_how == TXG_WAITED)
1294 tx->tx_waited = B_TRUE;
1295
1296 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1297 dmu_tx_unassign(tx);
1298
1299 if (err != ERESTART || txg_how != TXG_WAIT)
1300 return (err);
1301
1302 dmu_tx_wait(tx);
1303 }
1304
1305 txg_rele_to_quiesce(&tx->tx_txgh);
1306
1307 return (0);
1308 }
1309
1310 void
dmu_tx_wait(dmu_tx_t * tx)1311 dmu_tx_wait(dmu_tx_t *tx)
1312 {
1313 spa_t *spa = tx->tx_pool->dp_spa;
1314 dsl_pool_t *dp = tx->tx_pool;
1315
1316 ASSERT(tx->tx_txg == 0);
1317 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1318
1319 if (tx->tx_wait_dirty) {
1320 /*
1321 * dmu_tx_try_assign() has determined that we need to wait
1322 * because we've consumed much or all of the dirty buffer
1323 * space.
1324 */
1325 mutex_enter(&dp->dp_lock);
1326 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1327 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1328 uint64_t dirty = dp->dp_dirty_total;
1329 mutex_exit(&dp->dp_lock);
1330
1331 dmu_tx_delay(tx, dirty);
1332
1333 tx->tx_wait_dirty = B_FALSE;
1334
1335 /*
1336 * Note: setting tx_waited only has effect if the caller
1337 * used TX_WAIT. Otherwise they are going to destroy
1338 * this tx and try again. The common case, zfs_write(),
1339 * uses TX_WAIT.
1340 */
1341 tx->tx_waited = B_TRUE;
1342 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1343 /*
1344 * If the pool is suspended we need to wait until it
1345 * is resumed. Note that it's possible that the pool
1346 * has become active after this thread has tried to
1347 * obtain a tx. If that's the case then tx_lasttried_txg
1348 * would not have been set.
1349 */
1350 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1351 } else if (tx->tx_needassign_txh) {
1352 /*
1353 * A dnode is assigned to the quiescing txg. Wait for its
1354 * transaction to complete.
1355 */
1356 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1357
1358 mutex_enter(&dn->dn_mtx);
1359 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1360 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1361 mutex_exit(&dn->dn_mtx);
1362 tx->tx_needassign_txh = NULL;
1363 } else {
1364 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1365 }
1366 }
1367
1368 void
dmu_tx_willuse_space(dmu_tx_t * tx,int64_t delta)1369 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1370 {
1371 #ifdef ZFS_DEBUG
1372 if (tx->tx_dir == NULL || delta == 0)
1373 return;
1374
1375 if (delta > 0) {
1376 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1377 tx->tx_space_towrite);
1378 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1379 } else {
1380 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1381 }
1382 #endif
1383 }
1384
1385 void
dmu_tx_commit(dmu_tx_t * tx)1386 dmu_tx_commit(dmu_tx_t *tx)
1387 {
1388 dmu_tx_hold_t *txh;
1389
1390 ASSERT(tx->tx_txg != 0);
1391
1392 /*
1393 * Go through the transaction's hold list and remove holds on
1394 * associated dnodes, notifying waiters if no holds remain.
1395 */
1396 while (txh = list_head(&tx->tx_holds)) {
1397 dnode_t *dn = txh->txh_dnode;
1398
1399 list_remove(&tx->tx_holds, txh);
1400 kmem_free(txh, sizeof (dmu_tx_hold_t));
1401 if (dn == NULL)
1402 continue;
1403 mutex_enter(&dn->dn_mtx);
1404 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1405
1406 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1407 dn->dn_assigned_txg = 0;
1408 cv_broadcast(&dn->dn_notxholds);
1409 }
1410 mutex_exit(&dn->dn_mtx);
1411 dnode_rele(dn, tx);
1412 }
1413
1414 if (tx->tx_tempreserve_cookie)
1415 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1416
1417 if (!list_is_empty(&tx->tx_callbacks))
1418 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1419
1420 if (tx->tx_anyobj == FALSE)
1421 txg_rele_to_sync(&tx->tx_txgh);
1422
1423 list_destroy(&tx->tx_callbacks);
1424 list_destroy(&tx->tx_holds);
1425 #ifdef ZFS_DEBUG
1426 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1427 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1428 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1429 refcount_destroy_many(&tx->tx_space_written,
1430 refcount_count(&tx->tx_space_written));
1431 refcount_destroy_many(&tx->tx_space_freed,
1432 refcount_count(&tx->tx_space_freed));
1433 #endif
1434 kmem_free(tx, sizeof (dmu_tx_t));
1435 }
1436
1437 void
dmu_tx_abort(dmu_tx_t * tx)1438 dmu_tx_abort(dmu_tx_t *tx)
1439 {
1440 dmu_tx_hold_t *txh;
1441
1442 ASSERT(tx->tx_txg == 0);
1443
1444 while (txh = list_head(&tx->tx_holds)) {
1445 dnode_t *dn = txh->txh_dnode;
1446
1447 list_remove(&tx->tx_holds, txh);
1448 kmem_free(txh, sizeof (dmu_tx_hold_t));
1449 if (dn != NULL)
1450 dnode_rele(dn, tx);
1451 }
1452
1453 /*
1454 * Call any registered callbacks with an error code.
1455 */
1456 if (!list_is_empty(&tx->tx_callbacks))
1457 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1458
1459 list_destroy(&tx->tx_callbacks);
1460 list_destroy(&tx->tx_holds);
1461 #ifdef ZFS_DEBUG
1462 refcount_destroy_many(&tx->tx_space_written,
1463 refcount_count(&tx->tx_space_written));
1464 refcount_destroy_many(&tx->tx_space_freed,
1465 refcount_count(&tx->tx_space_freed));
1466 #endif
1467 kmem_free(tx, sizeof (dmu_tx_t));
1468 }
1469
1470 uint64_t
dmu_tx_get_txg(dmu_tx_t * tx)1471 dmu_tx_get_txg(dmu_tx_t *tx)
1472 {
1473 ASSERT(tx->tx_txg != 0);
1474 return (tx->tx_txg);
1475 }
1476
1477 dsl_pool_t *
dmu_tx_pool(dmu_tx_t * tx)1478 dmu_tx_pool(dmu_tx_t *tx)
1479 {
1480 ASSERT(tx->tx_pool != NULL);
1481 return (tx->tx_pool);
1482 }
1483
1484
1485 void
dmu_tx_callback_register(dmu_tx_t * tx,dmu_tx_callback_func_t * func,void * data)1486 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1487 {
1488 dmu_tx_callback_t *dcb;
1489
1490 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1491
1492 dcb->dcb_func = func;
1493 dcb->dcb_data = data;
1494
1495 list_insert_tail(&tx->tx_callbacks, dcb);
1496 }
1497
1498 /*
1499 * Call all the commit callbacks on a list, with a given error code.
1500 */
1501 void
dmu_tx_do_callbacks(list_t * cb_list,int error)1502 dmu_tx_do_callbacks(list_t *cb_list, int error)
1503 {
1504 dmu_tx_callback_t *dcb;
1505
1506 while (dcb = list_head(cb_list)) {
1507 list_remove(cb_list, dcb);
1508 dcb->dcb_func(dcb->dcb_data, error);
1509 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1510 }
1511 }
1512
1513 /*
1514 * Interface to hold a bunch of attributes.
1515 * used for creating new files.
1516 * attrsize is the total size of all attributes
1517 * to be added during object creation
1518 *
1519 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1520 */
1521
1522 /*
1523 * hold necessary attribute name for attribute registration.
1524 * should be a very rare case where this is needed. If it does
1525 * happen it would only happen on the first write to the file system.
1526 */
1527 static void
dmu_tx_sa_registration_hold(sa_os_t * sa,dmu_tx_t * tx)1528 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1529 {
1530 int i;
1531
1532 if (!sa->sa_need_attr_registration)
1533 return;
1534
1535 for (i = 0; i != sa->sa_num_attrs; i++) {
1536 if (!sa->sa_attr_table[i].sa_registered) {
1537 if (sa->sa_reg_attr_obj)
1538 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1539 B_TRUE, sa->sa_attr_table[i].sa_name);
1540 else
1541 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1542 B_TRUE, sa->sa_attr_table[i].sa_name);
1543 }
1544 }
1545 }
1546
1547
1548 void
dmu_tx_hold_spill(dmu_tx_t * tx,uint64_t object)1549 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1550 {
1551 dnode_t *dn;
1552 dmu_tx_hold_t *txh;
1553
1554 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1555 THT_SPILL, 0, 0);
1556
1557 dn = txh->txh_dnode;
1558
1559 if (dn == NULL)
1560 return;
1561
1562 /* If blkptr doesn't exist then add space to towrite */
1563 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1564 txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1565 } else {
1566 blkptr_t *bp;
1567
1568 bp = &dn->dn_phys->dn_spill;
1569 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1570 bp, bp->blk_birth))
1571 txh->txh_space_tooverwrite += SPA_OLD_MAXBLOCKSIZE;
1572 else
1573 txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1574 if (!BP_IS_HOLE(bp))
1575 txh->txh_space_tounref += SPA_OLD_MAXBLOCKSIZE;
1576 }
1577 }
1578
1579 void
dmu_tx_hold_sa_create(dmu_tx_t * tx,int attrsize)1580 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1581 {
1582 sa_os_t *sa = tx->tx_objset->os_sa;
1583
1584 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1585
1586 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1587 return;
1588
1589 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1590 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1591 else {
1592 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1593 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1594 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1595 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1596 }
1597
1598 dmu_tx_sa_registration_hold(sa, tx);
1599
1600 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1601 return;
1602
1603 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1604 THT_SPILL, 0, 0);
1605 }
1606
1607 /*
1608 * Hold SA attribute
1609 *
1610 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1611 *
1612 * variable_size is the total size of all variable sized attributes
1613 * passed to this function. It is not the total size of all
1614 * variable size attributes that *may* exist on this object.
1615 */
1616 void
dmu_tx_hold_sa(dmu_tx_t * tx,sa_handle_t * hdl,boolean_t may_grow)1617 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1618 {
1619 uint64_t object;
1620 sa_os_t *sa = tx->tx_objset->os_sa;
1621
1622 ASSERT(hdl != NULL);
1623
1624 object = sa_handle_object(hdl);
1625
1626 dmu_tx_hold_bonus(tx, object);
1627
1628 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1629 return;
1630
1631 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1632 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1633 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1634 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1635 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1636 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1637 }
1638
1639 dmu_tx_sa_registration_hold(sa, tx);
1640
1641 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1642 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1643
1644 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1645 ASSERT(tx->tx_txg == 0);
1646 dmu_tx_hold_spill(tx, object);
1647 } else {
1648 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1649 dnode_t *dn;
1650
1651 DB_DNODE_ENTER(db);
1652 dn = DB_DNODE(db);
1653 if (dn->dn_have_spill) {
1654 ASSERT(tx->tx_txg == 0);
1655 dmu_tx_hold_spill(tx, object);
1656 }
1657 DB_DNODE_EXIT(db);
1658 }
1659 }
1660