xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_tx.c (revision c3d26abc9ee97b4f60233556aadeb57e0bd30bb9)
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  * Copyright (c) 2014 Integros [integros.com]
26  */
27 
28 #include <sys/dmu.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dbuf.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
34 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
35 #include <sys/dsl_pool.h>
36 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
37 #include <sys/spa.h>
38 #include <sys/sa.h>
39 #include <sys/sa_impl.h>
40 #include <sys/zfs_context.h>
41 #include <sys/varargs.h>
42 
43 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
44     uint64_t arg1, uint64_t arg2);
45 
46 
47 dmu_tx_t *
48 dmu_tx_create_dd(dsl_dir_t *dd)
49 {
50 	dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
51 	tx->tx_dir = dd;
52 	if (dd != NULL)
53 		tx->tx_pool = dd->dd_pool;
54 	list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
55 	    offsetof(dmu_tx_hold_t, txh_node));
56 	list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
57 	    offsetof(dmu_tx_callback_t, dcb_node));
58 	tx->tx_start = gethrtime();
59 #ifdef ZFS_DEBUG
60 	refcount_create(&tx->tx_space_written);
61 	refcount_create(&tx->tx_space_freed);
62 #endif
63 	return (tx);
64 }
65 
66 dmu_tx_t *
67 dmu_tx_create(objset_t *os)
68 {
69 	dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
70 	tx->tx_objset = os;
71 	tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
72 	return (tx);
73 }
74 
75 dmu_tx_t *
76 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
77 {
78 	dmu_tx_t *tx = dmu_tx_create_dd(NULL);
79 
80 	ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
81 	tx->tx_pool = dp;
82 	tx->tx_txg = txg;
83 	tx->tx_anyobj = TRUE;
84 
85 	return (tx);
86 }
87 
88 int
89 dmu_tx_is_syncing(dmu_tx_t *tx)
90 {
91 	return (tx->tx_anyobj);
92 }
93 
94 int
95 dmu_tx_private_ok(dmu_tx_t *tx)
96 {
97 	return (tx->tx_anyobj);
98 }
99 
100 static dmu_tx_hold_t *
101 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
102     enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
103 {
104 	dmu_tx_hold_t *txh;
105 	dnode_t *dn = NULL;
106 	int err;
107 
108 	if (object != DMU_NEW_OBJECT) {
109 		err = dnode_hold(os, object, tx, &dn);
110 		if (err) {
111 			tx->tx_err = err;
112 			return (NULL);
113 		}
114 
115 		if (err == 0 && tx->tx_txg != 0) {
116 			mutex_enter(&dn->dn_mtx);
117 			/*
118 			 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
119 			 * problem, but there's no way for it to happen (for
120 			 * now, at least).
121 			 */
122 			ASSERT(dn->dn_assigned_txg == 0);
123 			dn->dn_assigned_txg = tx->tx_txg;
124 			(void) refcount_add(&dn->dn_tx_holds, tx);
125 			mutex_exit(&dn->dn_mtx);
126 		}
127 	}
128 
129 	txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
130 	txh->txh_tx = tx;
131 	txh->txh_dnode = dn;
132 #ifdef ZFS_DEBUG
133 	txh->txh_type = type;
134 	txh->txh_arg1 = arg1;
135 	txh->txh_arg2 = arg2;
136 #endif
137 	list_insert_tail(&tx->tx_holds, txh);
138 
139 	return (txh);
140 }
141 
142 void
143 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
144 {
145 	/*
146 	 * If we're syncing, they can manipulate any object anyhow, and
147 	 * the hold on the dnode_t can cause problems.
148 	 */
149 	if (!dmu_tx_is_syncing(tx)) {
150 		(void) dmu_tx_hold_object_impl(tx, os,
151 		    object, THT_NEWOBJECT, 0, 0);
152 	}
153 }
154 
155 static int
156 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
157 {
158 	int err;
159 	dmu_buf_impl_t *db;
160 
161 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
162 	db = dbuf_hold_level(dn, level, blkid, FTAG);
163 	rw_exit(&dn->dn_struct_rwlock);
164 	if (db == NULL)
165 		return (SET_ERROR(EIO));
166 	err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
167 	dbuf_rele(db, FTAG);
168 	return (err);
169 }
170 
171 static void
172 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
173     int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
174 {
175 	objset_t *os = dn->dn_objset;
176 	dsl_dataset_t *ds = os->os_dsl_dataset;
177 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
178 	dmu_buf_impl_t *parent = NULL;
179 	blkptr_t *bp = NULL;
180 	uint64_t space;
181 
182 	if (level >= dn->dn_nlevels || history[level] == blkid)
183 		return;
184 
185 	history[level] = blkid;
186 
187 	space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
188 
189 	if (db == NULL || db == dn->dn_dbuf) {
190 		ASSERT(level != 0);
191 		db = NULL;
192 	} else {
193 		ASSERT(DB_DNODE(db) == dn);
194 		ASSERT(db->db_level == level);
195 		ASSERT(db->db.db_size == space);
196 		ASSERT(db->db_blkid == blkid);
197 		bp = db->db_blkptr;
198 		parent = db->db_parent;
199 	}
200 
201 	freeable = (bp && (freeable ||
202 	    dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
203 
204 	if (freeable)
205 		txh->txh_space_tooverwrite += space;
206 	else
207 		txh->txh_space_towrite += space;
208 	if (bp)
209 		txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
210 
211 	dmu_tx_count_twig(txh, dn, parent, level + 1,
212 	    blkid >> epbs, freeable, history);
213 }
214 
215 /* ARGSUSED */
216 static void
217 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
218 {
219 	dnode_t *dn = txh->txh_dnode;
220 	uint64_t start, end, i;
221 	int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
222 	int err = 0;
223 
224 	if (len == 0)
225 		return;
226 
227 	min_bs = SPA_MINBLOCKSHIFT;
228 	max_bs = highbit64(txh->txh_tx->tx_objset->os_recordsize) - 1;
229 	min_ibs = DN_MIN_INDBLKSHIFT;
230 	max_ibs = DN_MAX_INDBLKSHIFT;
231 
232 	if (dn) {
233 		uint64_t history[DN_MAX_LEVELS];
234 		int nlvls = dn->dn_nlevels;
235 		int delta;
236 
237 		/*
238 		 * For i/o error checking, read the first and last level-0
239 		 * blocks (if they are not aligned), and all the level-1 blocks.
240 		 */
241 		if (dn->dn_maxblkid == 0) {
242 			delta = dn->dn_datablksz;
243 			start = (off < dn->dn_datablksz) ? 0 : 1;
244 			end = (off+len <= dn->dn_datablksz) ? 0 : 1;
245 			if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
246 				err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
247 				if (err)
248 					goto out;
249 				delta -= off;
250 			}
251 		} else {
252 			zio_t *zio = zio_root(dn->dn_objset->os_spa,
253 			    NULL, NULL, ZIO_FLAG_CANFAIL);
254 
255 			/* first level-0 block */
256 			start = off >> dn->dn_datablkshift;
257 			if (P2PHASE(off, dn->dn_datablksz) ||
258 			    len < dn->dn_datablksz) {
259 				err = dmu_tx_check_ioerr(zio, dn, 0, start);
260 				if (err)
261 					goto out;
262 			}
263 
264 			/* last level-0 block */
265 			end = (off+len-1) >> dn->dn_datablkshift;
266 			if (end != start && end <= dn->dn_maxblkid &&
267 			    P2PHASE(off+len, dn->dn_datablksz)) {
268 				err = dmu_tx_check_ioerr(zio, dn, 0, end);
269 				if (err)
270 					goto out;
271 			}
272 
273 			/* level-1 blocks */
274 			if (nlvls > 1) {
275 				int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
276 				for (i = (start>>shft)+1; i < end>>shft; i++) {
277 					err = dmu_tx_check_ioerr(zio, dn, 1, i);
278 					if (err)
279 						goto out;
280 				}
281 			}
282 
283 			err = zio_wait(zio);
284 			if (err)
285 				goto out;
286 			delta = P2NPHASE(off, dn->dn_datablksz);
287 		}
288 
289 		min_ibs = max_ibs = dn->dn_indblkshift;
290 		if (dn->dn_maxblkid > 0) {
291 			/*
292 			 * The blocksize can't change,
293 			 * so we can make a more precise estimate.
294 			 */
295 			ASSERT(dn->dn_datablkshift != 0);
296 			min_bs = max_bs = dn->dn_datablkshift;
297 		} else {
298 			/*
299 			 * The blocksize can increase up to the recordsize,
300 			 * or if it is already more than the recordsize,
301 			 * up to the next power of 2.
302 			 */
303 			min_bs = highbit64(dn->dn_datablksz - 1);
304 			max_bs = MAX(max_bs, highbit64(dn->dn_datablksz - 1));
305 		}
306 
307 		/*
308 		 * If this write is not off the end of the file
309 		 * we need to account for overwrites/unref.
310 		 */
311 		if (start <= dn->dn_maxblkid) {
312 			for (int l = 0; l < DN_MAX_LEVELS; l++)
313 				history[l] = -1ULL;
314 		}
315 		while (start <= dn->dn_maxblkid) {
316 			dmu_buf_impl_t *db;
317 
318 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
319 			err = dbuf_hold_impl(dn, 0, start,
320 			    FALSE, FALSE, FTAG, &db);
321 			rw_exit(&dn->dn_struct_rwlock);
322 
323 			if (err) {
324 				txh->txh_tx->tx_err = err;
325 				return;
326 			}
327 
328 			dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
329 			    history);
330 			dbuf_rele(db, FTAG);
331 			if (++start > end) {
332 				/*
333 				 * Account for new indirects appearing
334 				 * before this IO gets assigned into a txg.
335 				 */
336 				bits = 64 - min_bs;
337 				epbs = min_ibs - SPA_BLKPTRSHIFT;
338 				for (bits -= epbs * (nlvls - 1);
339 				    bits >= 0; bits -= epbs)
340 					txh->txh_fudge += 1ULL << max_ibs;
341 				goto out;
342 			}
343 			off += delta;
344 			if (len >= delta)
345 				len -= delta;
346 			delta = dn->dn_datablksz;
347 		}
348 	}
349 
350 	/*
351 	 * 'end' is the last thing we will access, not one past.
352 	 * This way we won't overflow when accessing the last byte.
353 	 */
354 	start = P2ALIGN(off, 1ULL << max_bs);
355 	end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
356 	txh->txh_space_towrite += end - start + 1;
357 
358 	start >>= min_bs;
359 	end >>= min_bs;
360 
361 	epbs = min_ibs - SPA_BLKPTRSHIFT;
362 
363 	/*
364 	 * The object contains at most 2^(64 - min_bs) blocks,
365 	 * and each indirect level maps 2^epbs.
366 	 */
367 	for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
368 		start >>= epbs;
369 		end >>= epbs;
370 		ASSERT3U(end, >=, start);
371 		txh->txh_space_towrite += (end - start + 1) << max_ibs;
372 		if (start != 0) {
373 			/*
374 			 * We also need a new blkid=0 indirect block
375 			 * to reference any existing file data.
376 			 */
377 			txh->txh_space_towrite += 1ULL << max_ibs;
378 		}
379 	}
380 
381 out:
382 	if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
383 	    2 * DMU_MAX_ACCESS)
384 		err = SET_ERROR(EFBIG);
385 
386 	if (err)
387 		txh->txh_tx->tx_err = err;
388 }
389 
390 static void
391 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
392 {
393 	dnode_t *dn = txh->txh_dnode;
394 	dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
395 	uint64_t space = mdn->dn_datablksz +
396 	    ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
397 
398 	if (dn && dn->dn_dbuf->db_blkptr &&
399 	    dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
400 	    dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
401 		txh->txh_space_tooverwrite += space;
402 		txh->txh_space_tounref += space;
403 	} else {
404 		txh->txh_space_towrite += space;
405 		if (dn && dn->dn_dbuf->db_blkptr)
406 			txh->txh_space_tounref += space;
407 	}
408 }
409 
410 void
411 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
412 {
413 	dmu_tx_hold_t *txh;
414 
415 	ASSERT(tx->tx_txg == 0);
416 	ASSERT(len < DMU_MAX_ACCESS);
417 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
418 
419 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
420 	    object, THT_WRITE, off, len);
421 	if (txh == NULL)
422 		return;
423 
424 	dmu_tx_count_write(txh, off, len);
425 	dmu_tx_count_dnode(txh);
426 }
427 
428 static void
429 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
430 {
431 	uint64_t blkid, nblks, lastblk;
432 	uint64_t space = 0, unref = 0, skipped = 0;
433 	dnode_t *dn = txh->txh_dnode;
434 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
435 	spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
436 	int epbs;
437 	uint64_t l0span = 0, nl1blks = 0;
438 
439 	if (dn->dn_nlevels == 0)
440 		return;
441 
442 	/*
443 	 * The struct_rwlock protects us against dn_nlevels
444 	 * changing, in case (against all odds) we manage to dirty &
445 	 * sync out the changes after we check for being dirty.
446 	 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
447 	 */
448 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
449 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
450 	if (dn->dn_maxblkid == 0) {
451 		if (off == 0 && len >= dn->dn_datablksz) {
452 			blkid = 0;
453 			nblks = 1;
454 		} else {
455 			rw_exit(&dn->dn_struct_rwlock);
456 			return;
457 		}
458 	} else {
459 		blkid = off >> dn->dn_datablkshift;
460 		nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
461 
462 		if (blkid > dn->dn_maxblkid) {
463 			rw_exit(&dn->dn_struct_rwlock);
464 			return;
465 		}
466 		if (blkid + nblks > dn->dn_maxblkid)
467 			nblks = dn->dn_maxblkid - blkid + 1;
468 
469 	}
470 	l0span = nblks;    /* save for later use to calc level > 1 overhead */
471 	if (dn->dn_nlevels == 1) {
472 		int i;
473 		for (i = 0; i < nblks; i++) {
474 			blkptr_t *bp = dn->dn_phys->dn_blkptr;
475 			ASSERT3U(blkid + i, <, dn->dn_nblkptr);
476 			bp += blkid + i;
477 			if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
478 				dprintf_bp(bp, "can free old%s", "");
479 				space += bp_get_dsize(spa, bp);
480 			}
481 			unref += BP_GET_ASIZE(bp);
482 		}
483 		nl1blks = 1;
484 		nblks = 0;
485 	}
486 
487 	lastblk = blkid + nblks - 1;
488 	while (nblks) {
489 		dmu_buf_impl_t *dbuf;
490 		uint64_t ibyte, new_blkid;
491 		int epb = 1 << epbs;
492 		int err, i, blkoff, tochk;
493 		blkptr_t *bp;
494 
495 		ibyte = blkid << dn->dn_datablkshift;
496 		err = dnode_next_offset(dn,
497 		    DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
498 		new_blkid = ibyte >> dn->dn_datablkshift;
499 		if (err == ESRCH) {
500 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
501 			break;
502 		}
503 		if (err) {
504 			txh->txh_tx->tx_err = err;
505 			break;
506 		}
507 		if (new_blkid > lastblk) {
508 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
509 			break;
510 		}
511 
512 		if (new_blkid > blkid) {
513 			ASSERT((new_blkid >> epbs) > (blkid >> epbs));
514 			skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
515 			nblks -= new_blkid - blkid;
516 			blkid = new_blkid;
517 		}
518 		blkoff = P2PHASE(blkid, epb);
519 		tochk = MIN(epb - blkoff, nblks);
520 
521 		err = dbuf_hold_impl(dn, 1, blkid >> epbs,
522 		    FALSE, FALSE, FTAG, &dbuf);
523 		if (err) {
524 			txh->txh_tx->tx_err = err;
525 			break;
526 		}
527 
528 		txh->txh_memory_tohold += dbuf->db.db_size;
529 
530 		/*
531 		 * We don't check memory_tohold against DMU_MAX_ACCESS because
532 		 * memory_tohold is an over-estimation (especially the >L1
533 		 * indirect blocks), so it could fail.  Callers should have
534 		 * already verified that they will not be holding too much
535 		 * memory.
536 		 */
537 
538 		err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
539 		if (err != 0) {
540 			txh->txh_tx->tx_err = err;
541 			dbuf_rele(dbuf, FTAG);
542 			break;
543 		}
544 
545 		bp = dbuf->db.db_data;
546 		bp += blkoff;
547 
548 		for (i = 0; i < tochk; i++) {
549 			if (dsl_dataset_block_freeable(ds, &bp[i],
550 			    bp[i].blk_birth)) {
551 				dprintf_bp(&bp[i], "can free old%s", "");
552 				space += bp_get_dsize(spa, &bp[i]);
553 			}
554 			unref += BP_GET_ASIZE(bp);
555 		}
556 		dbuf_rele(dbuf, FTAG);
557 
558 		++nl1blks;
559 		blkid += tochk;
560 		nblks -= tochk;
561 	}
562 	rw_exit(&dn->dn_struct_rwlock);
563 
564 	/*
565 	 * Add in memory requirements of higher-level indirects.
566 	 * This assumes a worst-possible scenario for dn_nlevels and a
567 	 * worst-possible distribution of l1-blocks over the region to free.
568 	 */
569 	{
570 		uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
571 		int level = 2;
572 		/*
573 		 * Here we don't use DN_MAX_LEVEL, but calculate it with the
574 		 * given datablkshift and indblkshift. This makes the
575 		 * difference between 19 and 8 on large files.
576 		 */
577 		int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
578 		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
579 
580 		while (level++ < maxlevel) {
581 			txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
582 			    << dn->dn_indblkshift;
583 			blkcnt = 1 + (blkcnt >> epbs);
584 		}
585 	}
586 
587 	/* account for new level 1 indirect blocks that might show up */
588 	if (skipped > 0) {
589 		txh->txh_fudge += skipped << dn->dn_indblkshift;
590 		skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
591 		txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
592 	}
593 	txh->txh_space_tofree += space;
594 	txh->txh_space_tounref += unref;
595 }
596 
597 /*
598  * This function marks the transaction as being a "net free".  The end
599  * result is that refquotas will be disabled for this transaction, and
600  * this transaction will be able to use half of the pool space overhead
601  * (see dsl_pool_adjustedsize()).  Therefore this function should only
602  * be called for transactions that we expect will not cause a net increase
603  * in the amount of space used (but it's OK if that is occasionally not true).
604  */
605 void
606 dmu_tx_mark_netfree(dmu_tx_t *tx)
607 {
608 	dmu_tx_hold_t *txh;
609 
610 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
611 	    DMU_NEW_OBJECT, THT_FREE, 0, 0);
612 
613 	/*
614 	 * Pretend that this operation will free 1GB of space.  This
615 	 * should be large enough to cancel out the largest write.
616 	 * We don't want to use something like UINT64_MAX, because that would
617 	 * cause overflows when doing math with these values (e.g. in
618 	 * dmu_tx_try_assign()).
619 	 */
620 	txh->txh_space_tofree = txh->txh_space_tounref = 1024 * 1024 * 1024;
621 }
622 
623 void
624 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
625 {
626 	dmu_tx_hold_t *txh;
627 	dnode_t *dn;
628 	int err;
629 	zio_t *zio;
630 
631 	ASSERT(tx->tx_txg == 0);
632 
633 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
634 	    object, THT_FREE, off, len);
635 	if (txh == NULL)
636 		return;
637 	dn = txh->txh_dnode;
638 	dmu_tx_count_dnode(txh);
639 
640 	if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
641 		return;
642 	if (len == DMU_OBJECT_END)
643 		len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
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
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
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
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
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
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
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 	mutex_enter(&curthread->t_delay_lock);
1096 	while (cv_timedwait_hires(&curthread->t_delay_cv,
1097 	    &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
1098 	    CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
1099 		continue;
1100 	mutex_exit(&curthread->t_delay_lock);
1101 #else
1102 	hrtime_t delta = wakeup - gethrtime();
1103 	struct timespec ts;
1104 	ts.tv_sec = delta / NANOSEC;
1105 	ts.tv_nsec = delta % NANOSEC;
1106 	(void) nanosleep(&ts, NULL);
1107 #endif
1108 }
1109 
1110 static int
1111 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1112 {
1113 	dmu_tx_hold_t *txh;
1114 	spa_t *spa = tx->tx_pool->dp_spa;
1115 	uint64_t memory, asize, fsize, usize;
1116 	uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1117 
1118 	ASSERT0(tx->tx_txg);
1119 
1120 	if (tx->tx_err)
1121 		return (tx->tx_err);
1122 
1123 	if (spa_suspended(spa)) {
1124 		/*
1125 		 * If the user has indicated a blocking failure mode
1126 		 * then return ERESTART which will block in dmu_tx_wait().
1127 		 * Otherwise, return EIO so that an error can get
1128 		 * propagated back to the VOP calls.
1129 		 *
1130 		 * Note that we always honor the txg_how flag regardless
1131 		 * of the failuremode setting.
1132 		 */
1133 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1134 		    txg_how != TXG_WAIT)
1135 			return (SET_ERROR(EIO));
1136 
1137 		return (SET_ERROR(ERESTART));
1138 	}
1139 
1140 	if (!tx->tx_waited &&
1141 	    dsl_pool_need_dirty_delay(tx->tx_pool)) {
1142 		tx->tx_wait_dirty = B_TRUE;
1143 		return (SET_ERROR(ERESTART));
1144 	}
1145 
1146 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1147 	tx->tx_needassign_txh = NULL;
1148 
1149 	/*
1150 	 * NB: No error returns are allowed after txg_hold_open, but
1151 	 * before processing the dnode holds, due to the
1152 	 * dmu_tx_unassign() logic.
1153 	 */
1154 
1155 	towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1156 	for (txh = list_head(&tx->tx_holds); txh;
1157 	    txh = list_next(&tx->tx_holds, txh)) {
1158 		dnode_t *dn = txh->txh_dnode;
1159 		if (dn != NULL) {
1160 			mutex_enter(&dn->dn_mtx);
1161 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1162 				mutex_exit(&dn->dn_mtx);
1163 				tx->tx_needassign_txh = txh;
1164 				return (SET_ERROR(ERESTART));
1165 			}
1166 			if (dn->dn_assigned_txg == 0)
1167 				dn->dn_assigned_txg = tx->tx_txg;
1168 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1169 			(void) refcount_add(&dn->dn_tx_holds, tx);
1170 			mutex_exit(&dn->dn_mtx);
1171 		}
1172 		towrite += txh->txh_space_towrite;
1173 		tofree += txh->txh_space_tofree;
1174 		tooverwrite += txh->txh_space_tooverwrite;
1175 		tounref += txh->txh_space_tounref;
1176 		tohold += txh->txh_memory_tohold;
1177 		fudge += txh->txh_fudge;
1178 	}
1179 
1180 	/*
1181 	 * If a snapshot has been taken since we made our estimates,
1182 	 * assume that we won't be able to free or overwrite anything.
1183 	 */
1184 	if (tx->tx_objset &&
1185 	    dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1186 	    tx->tx_lastsnap_txg) {
1187 		towrite += tooverwrite;
1188 		tooverwrite = tofree = 0;
1189 	}
1190 
1191 	/* needed allocation: worst-case estimate of write space */
1192 	asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1193 	/* freed space estimate: worst-case overwrite + free estimate */
1194 	fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1195 	/* convert unrefd space to worst-case estimate */
1196 	usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1197 	/* calculate memory footprint estimate */
1198 	memory = towrite + tooverwrite + tohold;
1199 
1200 #ifdef ZFS_DEBUG
1201 	/*
1202 	 * Add in 'tohold' to account for our dirty holds on this memory
1203 	 * XXX - the "fudge" factor is to account for skipped blocks that
1204 	 * we missed because dnode_next_offset() misses in-core-only blocks.
1205 	 */
1206 	tx->tx_space_towrite = asize +
1207 	    spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1208 	tx->tx_space_tofree = tofree;
1209 	tx->tx_space_tooverwrite = tooverwrite;
1210 	tx->tx_space_tounref = tounref;
1211 #endif
1212 
1213 	if (tx->tx_dir && asize != 0) {
1214 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1215 		    asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1216 		if (err)
1217 			return (err);
1218 	}
1219 
1220 	return (0);
1221 }
1222 
1223 static void
1224 dmu_tx_unassign(dmu_tx_t *tx)
1225 {
1226 	dmu_tx_hold_t *txh;
1227 
1228 	if (tx->tx_txg == 0)
1229 		return;
1230 
1231 	txg_rele_to_quiesce(&tx->tx_txgh);
1232 
1233 	/*
1234 	 * Walk the transaction's hold list, removing the hold on the
1235 	 * associated dnode, and notifying waiters if the refcount drops to 0.
1236 	 */
1237 	for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1238 	    txh = list_next(&tx->tx_holds, txh)) {
1239 		dnode_t *dn = txh->txh_dnode;
1240 
1241 		if (dn == NULL)
1242 			continue;
1243 		mutex_enter(&dn->dn_mtx);
1244 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1245 
1246 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1247 			dn->dn_assigned_txg = 0;
1248 			cv_broadcast(&dn->dn_notxholds);
1249 		}
1250 		mutex_exit(&dn->dn_mtx);
1251 	}
1252 
1253 	txg_rele_to_sync(&tx->tx_txgh);
1254 
1255 	tx->tx_lasttried_txg = tx->tx_txg;
1256 	tx->tx_txg = 0;
1257 }
1258 
1259 /*
1260  * Assign tx to a transaction group.  txg_how can be one of:
1261  *
1262  * (1)	TXG_WAIT.  If the current open txg is full, waits until there's
1263  *	a new one.  This should be used when you're not holding locks.
1264  *	It will only fail if we're truly out of space (or over quota).
1265  *
1266  * (2)	TXG_NOWAIT.  If we can't assign into the current open txg without
1267  *	blocking, returns immediately with ERESTART.  This should be used
1268  *	whenever you're holding locks.  On an ERESTART error, the caller
1269  *	should drop locks, do a dmu_tx_wait(tx), and try again.
1270  *
1271  * (3)  TXG_WAITED.  Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1272  *      has already been called on behalf of this operation (though
1273  *      most likely on a different tx).
1274  */
1275 int
1276 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1277 {
1278 	int err;
1279 
1280 	ASSERT(tx->tx_txg == 0);
1281 	ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1282 	    txg_how == TXG_WAITED);
1283 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1284 
1285 	/* If we might wait, we must not hold the config lock. */
1286 	ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1287 
1288 	if (txg_how == TXG_WAITED)
1289 		tx->tx_waited = B_TRUE;
1290 
1291 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1292 		dmu_tx_unassign(tx);
1293 
1294 		if (err != ERESTART || txg_how != TXG_WAIT)
1295 			return (err);
1296 
1297 		dmu_tx_wait(tx);
1298 	}
1299 
1300 	txg_rele_to_quiesce(&tx->tx_txgh);
1301 
1302 	return (0);
1303 }
1304 
1305 void
1306 dmu_tx_wait(dmu_tx_t *tx)
1307 {
1308 	spa_t *spa = tx->tx_pool->dp_spa;
1309 	dsl_pool_t *dp = tx->tx_pool;
1310 
1311 	ASSERT(tx->tx_txg == 0);
1312 	ASSERT(!dsl_pool_config_held(tx->tx_pool));
1313 
1314 	if (tx->tx_wait_dirty) {
1315 		/*
1316 		 * dmu_tx_try_assign() has determined that we need to wait
1317 		 * because we've consumed much or all of the dirty buffer
1318 		 * space.
1319 		 */
1320 		mutex_enter(&dp->dp_lock);
1321 		while (dp->dp_dirty_total >= zfs_dirty_data_max)
1322 			cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1323 		uint64_t dirty = dp->dp_dirty_total;
1324 		mutex_exit(&dp->dp_lock);
1325 
1326 		dmu_tx_delay(tx, dirty);
1327 
1328 		tx->tx_wait_dirty = B_FALSE;
1329 
1330 		/*
1331 		 * Note: setting tx_waited only has effect if the caller
1332 		 * used TX_WAIT.  Otherwise they are going to destroy
1333 		 * this tx and try again.  The common case, zfs_write(),
1334 		 * uses TX_WAIT.
1335 		 */
1336 		tx->tx_waited = B_TRUE;
1337 	} else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1338 		/*
1339 		 * If the pool is suspended we need to wait until it
1340 		 * is resumed.  Note that it's possible that the pool
1341 		 * has become active after this thread has tried to
1342 		 * obtain a tx.  If that's the case then tx_lasttried_txg
1343 		 * would not have been set.
1344 		 */
1345 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1346 	} else if (tx->tx_needassign_txh) {
1347 		/*
1348 		 * A dnode is assigned to the quiescing txg.  Wait for its
1349 		 * transaction to complete.
1350 		 */
1351 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1352 
1353 		mutex_enter(&dn->dn_mtx);
1354 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1355 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1356 		mutex_exit(&dn->dn_mtx);
1357 		tx->tx_needassign_txh = NULL;
1358 	} else {
1359 		txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1360 	}
1361 }
1362 
1363 void
1364 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1365 {
1366 #ifdef ZFS_DEBUG
1367 	if (tx->tx_dir == NULL || delta == 0)
1368 		return;
1369 
1370 	if (delta > 0) {
1371 		ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1372 		    tx->tx_space_towrite);
1373 		(void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1374 	} else {
1375 		(void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1376 	}
1377 #endif
1378 }
1379 
1380 void
1381 dmu_tx_commit(dmu_tx_t *tx)
1382 {
1383 	dmu_tx_hold_t *txh;
1384 
1385 	ASSERT(tx->tx_txg != 0);
1386 
1387 	/*
1388 	 * Go through the transaction's hold list and remove holds on
1389 	 * associated dnodes, notifying waiters if no holds remain.
1390 	 */
1391 	while (txh = list_head(&tx->tx_holds)) {
1392 		dnode_t *dn = txh->txh_dnode;
1393 
1394 		list_remove(&tx->tx_holds, txh);
1395 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1396 		if (dn == NULL)
1397 			continue;
1398 		mutex_enter(&dn->dn_mtx);
1399 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1400 
1401 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1402 			dn->dn_assigned_txg = 0;
1403 			cv_broadcast(&dn->dn_notxholds);
1404 		}
1405 		mutex_exit(&dn->dn_mtx);
1406 		dnode_rele(dn, tx);
1407 	}
1408 
1409 	if (tx->tx_tempreserve_cookie)
1410 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1411 
1412 	if (!list_is_empty(&tx->tx_callbacks))
1413 		txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1414 
1415 	if (tx->tx_anyobj == FALSE)
1416 		txg_rele_to_sync(&tx->tx_txgh);
1417 
1418 	list_destroy(&tx->tx_callbacks);
1419 	list_destroy(&tx->tx_holds);
1420 #ifdef ZFS_DEBUG
1421 	dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1422 	    tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1423 	    tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1424 	refcount_destroy_many(&tx->tx_space_written,
1425 	    refcount_count(&tx->tx_space_written));
1426 	refcount_destroy_many(&tx->tx_space_freed,
1427 	    refcount_count(&tx->tx_space_freed));
1428 #endif
1429 	kmem_free(tx, sizeof (dmu_tx_t));
1430 }
1431 
1432 void
1433 dmu_tx_abort(dmu_tx_t *tx)
1434 {
1435 	dmu_tx_hold_t *txh;
1436 
1437 	ASSERT(tx->tx_txg == 0);
1438 
1439 	while (txh = list_head(&tx->tx_holds)) {
1440 		dnode_t *dn = txh->txh_dnode;
1441 
1442 		list_remove(&tx->tx_holds, txh);
1443 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1444 		if (dn != NULL)
1445 			dnode_rele(dn, tx);
1446 	}
1447 
1448 	/*
1449 	 * Call any registered callbacks with an error code.
1450 	 */
1451 	if (!list_is_empty(&tx->tx_callbacks))
1452 		dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1453 
1454 	list_destroy(&tx->tx_callbacks);
1455 	list_destroy(&tx->tx_holds);
1456 #ifdef ZFS_DEBUG
1457 	refcount_destroy_many(&tx->tx_space_written,
1458 	    refcount_count(&tx->tx_space_written));
1459 	refcount_destroy_many(&tx->tx_space_freed,
1460 	    refcount_count(&tx->tx_space_freed));
1461 #endif
1462 	kmem_free(tx, sizeof (dmu_tx_t));
1463 }
1464 
1465 uint64_t
1466 dmu_tx_get_txg(dmu_tx_t *tx)
1467 {
1468 	ASSERT(tx->tx_txg != 0);
1469 	return (tx->tx_txg);
1470 }
1471 
1472 dsl_pool_t *
1473 dmu_tx_pool(dmu_tx_t *tx)
1474 {
1475 	ASSERT(tx->tx_pool != NULL);
1476 	return (tx->tx_pool);
1477 }
1478 
1479 
1480 void
1481 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1482 {
1483 	dmu_tx_callback_t *dcb;
1484 
1485 	dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1486 
1487 	dcb->dcb_func = func;
1488 	dcb->dcb_data = data;
1489 
1490 	list_insert_tail(&tx->tx_callbacks, dcb);
1491 }
1492 
1493 /*
1494  * Call all the commit callbacks on a list, with a given error code.
1495  */
1496 void
1497 dmu_tx_do_callbacks(list_t *cb_list, int error)
1498 {
1499 	dmu_tx_callback_t *dcb;
1500 
1501 	while (dcb = list_head(cb_list)) {
1502 		list_remove(cb_list, dcb);
1503 		dcb->dcb_func(dcb->dcb_data, error);
1504 		kmem_free(dcb, sizeof (dmu_tx_callback_t));
1505 	}
1506 }
1507 
1508 /*
1509  * Interface to hold a bunch of attributes.
1510  * used for creating new files.
1511  * attrsize is the total size of all attributes
1512  * to be added during object creation
1513  *
1514  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1515  */
1516 
1517 /*
1518  * hold necessary attribute name for attribute registration.
1519  * should be a very rare case where this is needed.  If it does
1520  * happen it would only happen on the first write to the file system.
1521  */
1522 static void
1523 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1524 {
1525 	int i;
1526 
1527 	if (!sa->sa_need_attr_registration)
1528 		return;
1529 
1530 	for (i = 0; i != sa->sa_num_attrs; i++) {
1531 		if (!sa->sa_attr_table[i].sa_registered) {
1532 			if (sa->sa_reg_attr_obj)
1533 				dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1534 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1535 			else
1536 				dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1537 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1538 		}
1539 	}
1540 }
1541 
1542 
1543 void
1544 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1545 {
1546 	dnode_t *dn;
1547 	dmu_tx_hold_t *txh;
1548 
1549 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1550 	    THT_SPILL, 0, 0);
1551 
1552 	dn = txh->txh_dnode;
1553 
1554 	if (dn == NULL)
1555 		return;
1556 
1557 	/* If blkptr doesn't exist then add space to towrite */
1558 	if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1559 		txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1560 	} else {
1561 		blkptr_t *bp;
1562 
1563 		bp = &dn->dn_phys->dn_spill;
1564 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1565 		    bp, bp->blk_birth))
1566 			txh->txh_space_tooverwrite += SPA_OLD_MAXBLOCKSIZE;
1567 		else
1568 			txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1569 		if (!BP_IS_HOLE(bp))
1570 			txh->txh_space_tounref += SPA_OLD_MAXBLOCKSIZE;
1571 	}
1572 }
1573 
1574 void
1575 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1576 {
1577 	sa_os_t *sa = tx->tx_objset->os_sa;
1578 
1579 	dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1580 
1581 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1582 		return;
1583 
1584 	if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1585 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1586 	else {
1587 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1588 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1589 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1590 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1591 	}
1592 
1593 	dmu_tx_sa_registration_hold(sa, tx);
1594 
1595 	if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1596 		return;
1597 
1598 	(void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1599 	    THT_SPILL, 0, 0);
1600 }
1601 
1602 /*
1603  * Hold SA attribute
1604  *
1605  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1606  *
1607  * variable_size is the total size of all variable sized attributes
1608  * passed to this function.  It is not the total size of all
1609  * variable size attributes that *may* exist on this object.
1610  */
1611 void
1612 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1613 {
1614 	uint64_t object;
1615 	sa_os_t *sa = tx->tx_objset->os_sa;
1616 
1617 	ASSERT(hdl != NULL);
1618 
1619 	object = sa_handle_object(hdl);
1620 
1621 	dmu_tx_hold_bonus(tx, object);
1622 
1623 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1624 		return;
1625 
1626 	if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1627 	    tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1628 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1629 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1630 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1631 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1632 	}
1633 
1634 	dmu_tx_sa_registration_hold(sa, tx);
1635 
1636 	if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1637 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1638 
1639 	if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1640 		ASSERT(tx->tx_txg == 0);
1641 		dmu_tx_hold_spill(tx, object);
1642 	} else {
1643 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1644 		dnode_t *dn;
1645 
1646 		DB_DNODE_ENTER(db);
1647 		dn = DB_DNODE(db);
1648 		if (dn->dn_have_spill) {
1649 			ASSERT(tx->tx_txg == 0);
1650 			dmu_tx_hold_spill(tx, object);
1651 		}
1652 		DB_DNODE_EXIT(db);
1653 	}
1654 }
1655