xref: /trueos/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu_tx.c (revision b878ae0412195c252dc69da54b764deab388e210)
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, 2014 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, FALSE, FTAG, &db);
319 			rw_exit(&dn->dn_struct_rwlock);
320 
321 			if (err) {
322 				txh->txh_tx->tx_err = err;
323 				return;
324 			}
325 
326 			dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
327 			    history);
328 			dbuf_rele(db, FTAG);
329 			if (++start > end) {
330 				/*
331 				 * Account for new indirects appearing
332 				 * before this IO gets assigned into a txg.
333 				 */
334 				bits = 64 - min_bs;
335 				epbs = min_ibs - SPA_BLKPTRSHIFT;
336 				for (bits -= epbs * (nlvls - 1);
337 				    bits >= 0; bits -= epbs)
338 					txh->txh_fudge += 1ULL << max_ibs;
339 				goto out;
340 			}
341 			off += delta;
342 			if (len >= delta)
343 				len -= delta;
344 			delta = dn->dn_datablksz;
345 		}
346 	}
347 
348 	/*
349 	 * 'end' is the last thing we will access, not one past.
350 	 * This way we won't overflow when accessing the last byte.
351 	 */
352 	start = P2ALIGN(off, 1ULL << max_bs);
353 	end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
354 	txh->txh_space_towrite += end - start + 1;
355 
356 	start >>= min_bs;
357 	end >>= min_bs;
358 
359 	epbs = min_ibs - SPA_BLKPTRSHIFT;
360 
361 	/*
362 	 * The object contains at most 2^(64 - min_bs) blocks,
363 	 * and each indirect level maps 2^epbs.
364 	 */
365 	for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
366 		start >>= epbs;
367 		end >>= epbs;
368 		ASSERT3U(end, >=, start);
369 		txh->txh_space_towrite += (end - start + 1) << max_ibs;
370 		if (start != 0) {
371 			/*
372 			 * We also need a new blkid=0 indirect block
373 			 * to reference any existing file data.
374 			 */
375 			txh->txh_space_towrite += 1ULL << max_ibs;
376 		}
377 	}
378 
379 out:
380 	if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
381 	    2 * DMU_MAX_ACCESS)
382 		err = SET_ERROR(EFBIG);
383 
384 	if (err)
385 		txh->txh_tx->tx_err = err;
386 }
387 
388 static void
dmu_tx_count_dnode(dmu_tx_hold_t * txh)389 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
390 {
391 	dnode_t *dn = txh->txh_dnode;
392 	dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
393 	uint64_t space = mdn->dn_datablksz +
394 	    ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
395 
396 	if (dn && dn->dn_dbuf->db_blkptr &&
397 	    dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
398 	    dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
399 		txh->txh_space_tooverwrite += space;
400 		txh->txh_space_tounref += space;
401 	} else {
402 		txh->txh_space_towrite += space;
403 		if (dn && dn->dn_dbuf->db_blkptr)
404 			txh->txh_space_tounref += space;
405 	}
406 }
407 
408 void
dmu_tx_hold_write(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)409 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
410 {
411 	dmu_tx_hold_t *txh;
412 
413 	ASSERT(tx->tx_txg == 0);
414 	ASSERT(len < DMU_MAX_ACCESS);
415 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
416 
417 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
418 	    object, THT_WRITE, off, len);
419 	if (txh == NULL)
420 		return;
421 
422 	dmu_tx_count_write(txh, off, len);
423 	dmu_tx_count_dnode(txh);
424 }
425 
426 static void
dmu_tx_count_free(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)427 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
428 {
429 	uint64_t blkid, nblks, lastblk;
430 	uint64_t space = 0, unref = 0, skipped = 0;
431 	dnode_t *dn = txh->txh_dnode;
432 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
433 	spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
434 	int epbs;
435 	uint64_t l0span = 0, nl1blks = 0;
436 
437 	if (dn->dn_nlevels == 0)
438 		return;
439 
440 	/*
441 	 * The struct_rwlock protects us against dn_nlevels
442 	 * changing, in case (against all odds) we manage to dirty &
443 	 * sync out the changes after we check for being dirty.
444 	 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
445 	 */
446 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
447 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
448 	if (dn->dn_maxblkid == 0) {
449 		if (off == 0 && len >= dn->dn_datablksz) {
450 			blkid = 0;
451 			nblks = 1;
452 		} else {
453 			rw_exit(&dn->dn_struct_rwlock);
454 			return;
455 		}
456 	} else {
457 		blkid = off >> dn->dn_datablkshift;
458 		nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
459 
460 		if (blkid > dn->dn_maxblkid) {
461 			rw_exit(&dn->dn_struct_rwlock);
462 			return;
463 		}
464 		if (blkid + nblks > dn->dn_maxblkid)
465 			nblks = dn->dn_maxblkid - blkid + 1;
466 
467 	}
468 	l0span = nblks;    /* save for later use to calc level > 1 overhead */
469 	if (dn->dn_nlevels == 1) {
470 		int i;
471 		for (i = 0; i < nblks; i++) {
472 			blkptr_t *bp = dn->dn_phys->dn_blkptr;
473 			ASSERT3U(blkid + i, <, dn->dn_nblkptr);
474 			bp += blkid + i;
475 			if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
476 				dprintf_bp(bp, "can free old%s", "");
477 				space += bp_get_dsize(spa, bp);
478 			}
479 			unref += BP_GET_ASIZE(bp);
480 		}
481 		nl1blks = 1;
482 		nblks = 0;
483 	}
484 
485 	lastblk = blkid + nblks - 1;
486 	while (nblks) {
487 		dmu_buf_impl_t *dbuf;
488 		uint64_t ibyte, new_blkid;
489 		int epb = 1 << epbs;
490 		int err, i, blkoff, tochk;
491 		blkptr_t *bp;
492 
493 		ibyte = blkid << dn->dn_datablkshift;
494 		err = dnode_next_offset(dn,
495 		    DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
496 		new_blkid = ibyte >> dn->dn_datablkshift;
497 		if (err == ESRCH) {
498 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
499 			break;
500 		}
501 		if (err) {
502 			txh->txh_tx->tx_err = err;
503 			break;
504 		}
505 		if (new_blkid > lastblk) {
506 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
507 			break;
508 		}
509 
510 		if (new_blkid > blkid) {
511 			ASSERT((new_blkid >> epbs) > (blkid >> epbs));
512 			skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
513 			nblks -= new_blkid - blkid;
514 			blkid = new_blkid;
515 		}
516 		blkoff = P2PHASE(blkid, epb);
517 		tochk = MIN(epb - blkoff, nblks);
518 
519 		err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
520 		if (err) {
521 			txh->txh_tx->tx_err = err;
522 			break;
523 		}
524 
525 		txh->txh_memory_tohold += dbuf->db.db_size;
526 
527 		/*
528 		 * We don't check memory_tohold against DMU_MAX_ACCESS because
529 		 * memory_tohold is an over-estimation (especially the >L1
530 		 * indirect blocks), so it could fail.  Callers should have
531 		 * already verified that they will not be holding too much
532 		 * memory.
533 		 */
534 
535 		err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
536 		if (err != 0) {
537 			txh->txh_tx->tx_err = err;
538 			dbuf_rele(dbuf, FTAG);
539 			break;
540 		}
541 
542 		bp = dbuf->db.db_data;
543 		bp += blkoff;
544 
545 		for (i = 0; i < tochk; i++) {
546 			if (dsl_dataset_block_freeable(ds, &bp[i],
547 			    bp[i].blk_birth)) {
548 				dprintf_bp(&bp[i], "can free old%s", "");
549 				space += bp_get_dsize(spa, &bp[i]);
550 			}
551 			unref += BP_GET_ASIZE(bp);
552 		}
553 		dbuf_rele(dbuf, FTAG);
554 
555 		++nl1blks;
556 		blkid += tochk;
557 		nblks -= tochk;
558 	}
559 	rw_exit(&dn->dn_struct_rwlock);
560 
561 	/*
562 	 * Add in memory requirements of higher-level indirects.
563 	 * This assumes a worst-possible scenario for dn_nlevels and a
564 	 * worst-possible distribution of l1-blocks over the region to free.
565 	 */
566 	{
567 		uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
568 		int level = 2;
569 		/*
570 		 * Here we don't use DN_MAX_LEVEL, but calculate it with the
571 		 * given datablkshift and indblkshift. This makes the
572 		 * difference between 19 and 8 on large files.
573 		 */
574 		int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
575 		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
576 
577 		while (level++ < maxlevel) {
578 			txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
579 			    << dn->dn_indblkshift;
580 			blkcnt = 1 + (blkcnt >> epbs);
581 		}
582 	}
583 
584 	/* account for new level 1 indirect blocks that might show up */
585 	if (skipped > 0) {
586 		txh->txh_fudge += skipped << dn->dn_indblkshift;
587 		skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
588 		txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
589 	}
590 	txh->txh_space_tofree += space;
591 	txh->txh_space_tounref += unref;
592 }
593 
594 /*
595  * This function marks the transaction as being a "net free".  The end
596  * result is that refquotas will be disabled for this transaction, and
597  * this transaction will be able to use half of the pool space overhead
598  * (see dsl_pool_adjustedsize()).  Therefore this function should only
599  * be called for transactions that we expect will not cause a net increase
600  * in the amount of space used (but it's OK if that is occasionally not true).
601  */
602 void
dmu_tx_mark_netfree(dmu_tx_t * tx)603 dmu_tx_mark_netfree(dmu_tx_t *tx)
604 {
605 	dmu_tx_hold_t *txh;
606 
607 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
608 	    DMU_NEW_OBJECT, THT_FREE, 0, 0);
609 
610 	/*
611 	 * Pretend that this operation will free 1GB of space.  This
612 	 * should be large enough to cancel out the largest write.
613 	 * We don't want to use something like UINT64_MAX, because that would
614 	 * cause overflows when doing math with these values (e.g. in
615 	 * dmu_tx_try_assign()).
616 	 */
617 	txh->txh_space_tofree = txh->txh_space_tounref = 1024 * 1024 * 1024;
618 }
619 
620 void
dmu_tx_hold_free(dmu_tx_t * tx,uint64_t object,uint64_t off,uint64_t len)621 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
622 {
623 	dmu_tx_hold_t *txh;
624 	dnode_t *dn;
625 	int err;
626 	zio_t *zio;
627 
628 	ASSERT(tx->tx_txg == 0);
629 
630 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
631 	    object, THT_FREE, off, len);
632 	if (txh == NULL)
633 		return;
634 	dn = txh->txh_dnode;
635 	dmu_tx_count_dnode(txh);
636 
637 	if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
638 		return;
639 	if (len == DMU_OBJECT_END)
640 		len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
641 
642 	/*
643 	 * For i/o error checking, we read the first and last level-0
644 	 * blocks if they are not aligned, and all the level-1 blocks.
645 	 *
646 	 * Note:  dbuf_free_range() assumes that we have not instantiated
647 	 * any level-0 dbufs that will be completely freed.  Therefore we must
648 	 * exercise care to not read or count the first and last blocks
649 	 * if they are blocksize-aligned.
650 	 */
651 	if (dn->dn_datablkshift == 0) {
652 		if (off != 0 || len < dn->dn_datablksz)
653 			dmu_tx_count_write(txh, 0, dn->dn_datablksz);
654 	} else {
655 		/* first block will be modified if it is not aligned */
656 		if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
657 			dmu_tx_count_write(txh, off, 1);
658 		/* last block will be modified if it is not aligned */
659 		if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
660 			dmu_tx_count_write(txh, off+len, 1);
661 	}
662 
663 	/*
664 	 * Check level-1 blocks.
665 	 */
666 	if (dn->dn_nlevels > 1) {
667 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
668 		    SPA_BLKPTRSHIFT;
669 		uint64_t start = off >> shift;
670 		uint64_t end = (off + len) >> shift;
671 
672 		ASSERT(dn->dn_indblkshift != 0);
673 
674 		/*
675 		 * dnode_reallocate() can result in an object with indirect
676 		 * blocks having an odd data block size.  In this case,
677 		 * just check the single block.
678 		 */
679 		if (dn->dn_datablkshift == 0)
680 			start = end = 0;
681 
682 		zio = zio_root(tx->tx_pool->dp_spa,
683 		    NULL, NULL, ZIO_FLAG_CANFAIL);
684 		for (uint64_t i = start; i <= end; i++) {
685 			uint64_t ibyte = i << shift;
686 			err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
687 			i = ibyte >> shift;
688 			if (err == ESRCH)
689 				break;
690 			if (err) {
691 				tx->tx_err = err;
692 				return;
693 			}
694 
695 			err = dmu_tx_check_ioerr(zio, dn, 1, i);
696 			if (err) {
697 				tx->tx_err = err;
698 				return;
699 			}
700 		}
701 		err = zio_wait(zio);
702 		if (err) {
703 			tx->tx_err = err;
704 			return;
705 		}
706 	}
707 
708 	dmu_tx_count_free(txh, off, len);
709 }
710 
711 void
dmu_tx_hold_zap(dmu_tx_t * tx,uint64_t object,int add,const char * name)712 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
713 {
714 	dmu_tx_hold_t *txh;
715 	dnode_t *dn;
716 	dsl_dataset_phys_t *ds_phys;
717 	uint64_t nblocks;
718 	int epbs, err;
719 
720 	ASSERT(tx->tx_txg == 0);
721 
722 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
723 	    object, THT_ZAP, add, (uintptr_t)name);
724 	if (txh == NULL)
725 		return;
726 	dn = txh->txh_dnode;
727 
728 	dmu_tx_count_dnode(txh);
729 
730 	if (dn == NULL) {
731 		/*
732 		 * We will be able to fit a new object's entries into one leaf
733 		 * block.  So there will be at most 2 blocks total,
734 		 * including the header block.
735 		 */
736 		dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
737 		return;
738 	}
739 
740 	ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
741 
742 	if (dn->dn_maxblkid == 0 && !add) {
743 		blkptr_t *bp;
744 
745 		/*
746 		 * If there is only one block  (i.e. this is a micro-zap)
747 		 * and we are not adding anything, the accounting is simple.
748 		 */
749 		err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
750 		if (err) {
751 			tx->tx_err = err;
752 			return;
753 		}
754 
755 		/*
756 		 * Use max block size here, since we don't know how much
757 		 * the size will change between now and the dbuf dirty call.
758 		 */
759 		bp = &dn->dn_phys->dn_blkptr[0];
760 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
761 		    bp, bp->blk_birth))
762 			txh->txh_space_tooverwrite += MZAP_MAX_BLKSZ;
763 		else
764 			txh->txh_space_towrite += MZAP_MAX_BLKSZ;
765 		if (!BP_IS_HOLE(bp))
766 			txh->txh_space_tounref += MZAP_MAX_BLKSZ;
767 		return;
768 	}
769 
770 	if (dn->dn_maxblkid > 0 && name) {
771 		/*
772 		 * access the name in this fat-zap so that we'll check
773 		 * for i/o errors to the leaf blocks, etc.
774 		 */
775 		err = zap_lookup(dn->dn_objset, dn->dn_object, name,
776 		    8, 0, NULL);
777 		if (err == EIO) {
778 			tx->tx_err = err;
779 			return;
780 		}
781 	}
782 
783 	err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
784 	    &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
785 
786 	/*
787 	 * If the modified blocks are scattered to the four winds,
788 	 * we'll have to modify an indirect twig for each.
789 	 */
790 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
791 	ds_phys = dsl_dataset_phys(dn->dn_objset->os_dsl_dataset);
792 	for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
793 		if (ds_phys->ds_prev_snap_obj)
794 			txh->txh_space_towrite += 3 << dn->dn_indblkshift;
795 		else
796 			txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
797 }
798 
799 void
dmu_tx_hold_bonus(dmu_tx_t * tx,uint64_t object)800 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
801 {
802 	dmu_tx_hold_t *txh;
803 
804 	ASSERT(tx->tx_txg == 0);
805 
806 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
807 	    object, THT_BONUS, 0, 0);
808 	if (txh)
809 		dmu_tx_count_dnode(txh);
810 }
811 
812 void
dmu_tx_hold_space(dmu_tx_t * tx,uint64_t space)813 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
814 {
815 	dmu_tx_hold_t *txh;
816 	ASSERT(tx->tx_txg == 0);
817 
818 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
819 	    DMU_NEW_OBJECT, THT_SPACE, space, 0);
820 
821 	txh->txh_space_towrite += space;
822 }
823 
824 int
dmu_tx_holds(dmu_tx_t * tx,uint64_t object)825 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
826 {
827 	dmu_tx_hold_t *txh;
828 	int holds = 0;
829 
830 	/*
831 	 * By asserting that the tx is assigned, we're counting the
832 	 * number of dn_tx_holds, which is the same as the number of
833 	 * dn_holds.  Otherwise, we'd be counting dn_holds, but
834 	 * dn_tx_holds could be 0.
835 	 */
836 	ASSERT(tx->tx_txg != 0);
837 
838 	/* if (tx->tx_anyobj == TRUE) */
839 		/* return (0); */
840 
841 	for (txh = list_head(&tx->tx_holds); txh;
842 	    txh = list_next(&tx->tx_holds, txh)) {
843 		if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
844 			holds++;
845 	}
846 
847 	return (holds);
848 }
849 
850 #ifdef ZFS_DEBUG
851 void
dmu_tx_dirty_buf(dmu_tx_t * tx,dmu_buf_impl_t * db)852 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
853 {
854 	dmu_tx_hold_t *txh;
855 	int match_object = FALSE, match_offset = FALSE;
856 	dnode_t *dn;
857 
858 	DB_DNODE_ENTER(db);
859 	dn = DB_DNODE(db);
860 	ASSERT(tx->tx_txg != 0);
861 	ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
862 	ASSERT3U(dn->dn_object, ==, db->db.db_object);
863 
864 	if (tx->tx_anyobj) {
865 		DB_DNODE_EXIT(db);
866 		return;
867 	}
868 
869 	/* XXX No checking on the meta dnode for now */
870 	if (db->db.db_object == DMU_META_DNODE_OBJECT) {
871 		DB_DNODE_EXIT(db);
872 		return;
873 	}
874 
875 	for (txh = list_head(&tx->tx_holds); txh;
876 	    txh = list_next(&tx->tx_holds, txh)) {
877 		ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
878 		if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
879 			match_object = TRUE;
880 		if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
881 			int datablkshift = dn->dn_datablkshift ?
882 			    dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
883 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
884 			int shift = datablkshift + epbs * db->db_level;
885 			uint64_t beginblk = shift >= 64 ? 0 :
886 			    (txh->txh_arg1 >> shift);
887 			uint64_t endblk = shift >= 64 ? 0 :
888 			    ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
889 			uint64_t blkid = db->db_blkid;
890 
891 			/* XXX txh_arg2 better not be zero... */
892 
893 			dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
894 			    txh->txh_type, beginblk, endblk);
895 
896 			switch (txh->txh_type) {
897 			case THT_WRITE:
898 				if (blkid >= beginblk && blkid <= endblk)
899 					match_offset = TRUE;
900 				/*
901 				 * We will let this hold work for the bonus
902 				 * or spill buffer so that we don't need to
903 				 * hold it when creating a new object.
904 				 */
905 				if (blkid == DMU_BONUS_BLKID ||
906 				    blkid == DMU_SPILL_BLKID)
907 					match_offset = TRUE;
908 				/*
909 				 * They might have to increase nlevels,
910 				 * thus dirtying the new TLIBs.  Or the
911 				 * might have to change the block size,
912 				 * thus dirying the new lvl=0 blk=0.
913 				 */
914 				if (blkid == 0)
915 					match_offset = TRUE;
916 				break;
917 			case THT_FREE:
918 				/*
919 				 * We will dirty all the level 1 blocks in
920 				 * the free range and perhaps the first and
921 				 * last level 0 block.
922 				 */
923 				if (blkid >= beginblk && (blkid <= endblk ||
924 				    txh->txh_arg2 == DMU_OBJECT_END))
925 					match_offset = TRUE;
926 				break;
927 			case THT_SPILL:
928 				if (blkid == DMU_SPILL_BLKID)
929 					match_offset = TRUE;
930 				break;
931 			case THT_BONUS:
932 				if (blkid == DMU_BONUS_BLKID)
933 					match_offset = TRUE;
934 				break;
935 			case THT_ZAP:
936 				match_offset = TRUE;
937 				break;
938 			case THT_NEWOBJECT:
939 				match_object = TRUE;
940 				break;
941 			default:
942 				ASSERT(!"bad txh_type");
943 			}
944 		}
945 		if (match_object && match_offset) {
946 			DB_DNODE_EXIT(db);
947 			return;
948 		}
949 	}
950 	DB_DNODE_EXIT(db);
951 	panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
952 	    (u_longlong_t)db->db.db_object, db->db_level,
953 	    (u_longlong_t)db->db_blkid);
954 }
955 #endif
956 
957 /*
958  * If we can't do 10 iops, something is wrong.  Let us go ahead
959  * and hit zfs_dirty_data_max.
960  */
961 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
962 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
963 
964 /*
965  * We delay transactions when we've determined that the backend storage
966  * isn't able to accommodate the rate of incoming writes.
967  *
968  * If there is already a transaction waiting, we delay relative to when
969  * that transaction finishes waiting.  This way the calculated min_time
970  * is independent of the number of threads concurrently executing
971  * transactions.
972  *
973  * If we are the only waiter, wait relative to when the transaction
974  * started, rather than the current time.  This credits the transaction for
975  * "time already served", e.g. reading indirect blocks.
976  *
977  * The minimum time for a transaction to take is calculated as:
978  *     min_time = scale * (dirty - min) / (max - dirty)
979  *     min_time is then capped at zfs_delay_max_ns.
980  *
981  * The delay has two degrees of freedom that can be adjusted via tunables.
982  * The percentage of dirty data at which we start to delay is defined by
983  * zfs_delay_min_dirty_percent. This should typically be at or above
984  * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
985  * delay after writing at full speed has failed to keep up with the incoming
986  * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
987  * speaking, this variable determines the amount of delay at the midpoint of
988  * the curve.
989  *
990  * delay
991  *  10ms +-------------------------------------------------------------*+
992  *       |                                                             *|
993  *   9ms +                                                             *+
994  *       |                                                             *|
995  *   8ms +                                                             *+
996  *       |                                                            * |
997  *   7ms +                                                            * +
998  *       |                                                            * |
999  *   6ms +                                                            * +
1000  *       |                                                            * |
1001  *   5ms +                                                           *  +
1002  *       |                                                           *  |
1003  *   4ms +                                                           *  +
1004  *       |                                                           *  |
1005  *   3ms +                                                          *   +
1006  *       |                                                          *   |
1007  *   2ms +                                              (midpoint) *    +
1008  *       |                                                  |    **     |
1009  *   1ms +                                                  v ***       +
1010  *       |             zfs_delay_scale ---------->     ********         |
1011  *     0 +-------------------------------------*********----------------+
1012  *       0%                    <- zfs_dirty_data_max ->               100%
1013  *
1014  * Note that since the delay is added to the outstanding time remaining on the
1015  * most recent transaction, the delay is effectively the inverse of IOPS.
1016  * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
1017  * was chosen such that small changes in the amount of accumulated dirty data
1018  * in the first 3/4 of the curve yield relatively small differences in the
1019  * amount of delay.
1020  *
1021  * The effects can be easier to understand when the amount of delay is
1022  * represented on a log scale:
1023  *
1024  * delay
1025  * 100ms +-------------------------------------------------------------++
1026  *       +                                                              +
1027  *       |                                                              |
1028  *       +                                                             *+
1029  *  10ms +                                                             *+
1030  *       +                                                           ** +
1031  *       |                                              (midpoint)  **  |
1032  *       +                                                  |     **    +
1033  *   1ms +                                                  v ****      +
1034  *       +             zfs_delay_scale ---------->        *****         +
1035  *       |                                             ****             |
1036  *       +                                          ****                +
1037  * 100us +                                        **                    +
1038  *       +                                       *                      +
1039  *       |                                      *                       |
1040  *       +                                     *                        +
1041  *  10us +                                     *                        +
1042  *       +                                                              +
1043  *       |                                                              |
1044  *       +                                                              +
1045  *       +--------------------------------------------------------------+
1046  *       0%                    <- zfs_dirty_data_max ->               100%
1047  *
1048  * Note here that only as the amount of dirty data approaches its limit does
1049  * the delay start to increase rapidly. The goal of a properly tuned system
1050  * should be to keep the amount of dirty data out of that range by first
1051  * ensuring that the appropriate limits are set for the I/O scheduler to reach
1052  * optimal throughput on the backend storage, and then by changing the value
1053  * of zfs_delay_scale to increase the steepness of the curve.
1054  */
1055 static void
dmu_tx_delay(dmu_tx_t * tx,uint64_t dirty)1056 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1057 {
1058 	dsl_pool_t *dp = tx->tx_pool;
1059 	uint64_t delay_min_bytes =
1060 	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1061 	hrtime_t wakeup, min_tx_time, now;
1062 
1063 	if (dirty <= delay_min_bytes)
1064 		return;
1065 
1066 	/*
1067 	 * The caller has already waited until we are under the max.
1068 	 * We make them pass us the amount of dirty data so we don't
1069 	 * have to handle the case of it being >= the max, which could
1070 	 * cause a divide-by-zero if it's == the max.
1071 	 */
1072 	ASSERT3U(dirty, <, zfs_dirty_data_max);
1073 
1074 	now = gethrtime();
1075 	min_tx_time = zfs_delay_scale *
1076 	    (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1077 	if (now > tx->tx_start + min_tx_time)
1078 		return;
1079 
1080 	min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1081 
1082 	DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1083 	    uint64_t, min_tx_time);
1084 
1085 	mutex_enter(&dp->dp_lock);
1086 	wakeup = MAX(tx->tx_start + min_tx_time,
1087 	    dp->dp_last_wakeup + min_tx_time);
1088 	dp->dp_last_wakeup = wakeup;
1089 	mutex_exit(&dp->dp_lock);
1090 
1091 #ifdef _KERNEL
1092 #ifdef illumos
1093 	mutex_enter(&curthread->t_delay_lock);
1094 	while (cv_timedwait_hires(&curthread->t_delay_cv,
1095 	    &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
1096 	    CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
1097 		continue;
1098 	mutex_exit(&curthread->t_delay_lock);
1099 #else
1100 	pause_sbt("dmu_tx_delay", wakeup * SBT_1NS,
1101 	    zfs_delay_resolution_ns * SBT_1NS, C_ABSOLUTE);
1102 #endif
1103 #else
1104 	hrtime_t delta = wakeup - gethrtime();
1105 	struct timespec ts;
1106 	ts.tv_sec = delta / NANOSEC;
1107 	ts.tv_nsec = delta % NANOSEC;
1108 	(void) nanosleep(&ts, NULL);
1109 #endif
1110 }
1111 
1112 static int
dmu_tx_try_assign(dmu_tx_t * tx,txg_how_t txg_how)1113 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1114 {
1115 	dmu_tx_hold_t *txh;
1116 	spa_t *spa = tx->tx_pool->dp_spa;
1117 	uint64_t memory, asize, fsize, usize;
1118 	uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1119 
1120 	ASSERT0(tx->tx_txg);
1121 
1122 	if (tx->tx_err)
1123 		return (tx->tx_err);
1124 
1125 	if (spa_suspended(spa)) {
1126 		/*
1127 		 * If the user has indicated a blocking failure mode
1128 		 * then return ERESTART which will block in dmu_tx_wait().
1129 		 * Otherwise, return EIO so that an error can get
1130 		 * propagated back to the VOP calls.
1131 		 *
1132 		 * Note that we always honor the txg_how flag regardless
1133 		 * of the failuremode setting.
1134 		 */
1135 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1136 		    txg_how != TXG_WAIT)
1137 			return (SET_ERROR(EIO));
1138 
1139 		return (SET_ERROR(ERESTART));
1140 	}
1141 
1142 	if (!tx->tx_waited &&
1143 	    dsl_pool_need_dirty_delay(tx->tx_pool)) {
1144 		tx->tx_wait_dirty = B_TRUE;
1145 		return (SET_ERROR(ERESTART));
1146 	}
1147 
1148 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1149 	tx->tx_needassign_txh = NULL;
1150 
1151 	/*
1152 	 * NB: No error returns are allowed after txg_hold_open, but
1153 	 * before processing the dnode holds, due to the
1154 	 * dmu_tx_unassign() logic.
1155 	 */
1156 
1157 	towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1158 	for (txh = list_head(&tx->tx_holds); txh;
1159 	    txh = list_next(&tx->tx_holds, txh)) {
1160 		dnode_t *dn = txh->txh_dnode;
1161 		if (dn != NULL) {
1162 			mutex_enter(&dn->dn_mtx);
1163 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1164 				mutex_exit(&dn->dn_mtx);
1165 				tx->tx_needassign_txh = txh;
1166 				return (SET_ERROR(ERESTART));
1167 			}
1168 			if (dn->dn_assigned_txg == 0)
1169 				dn->dn_assigned_txg = tx->tx_txg;
1170 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1171 			(void) refcount_add(&dn->dn_tx_holds, tx);
1172 			mutex_exit(&dn->dn_mtx);
1173 		}
1174 		towrite += txh->txh_space_towrite;
1175 		tofree += txh->txh_space_tofree;
1176 		tooverwrite += txh->txh_space_tooverwrite;
1177 		tounref += txh->txh_space_tounref;
1178 		tohold += txh->txh_memory_tohold;
1179 		fudge += txh->txh_fudge;
1180 	}
1181 
1182 	/*
1183 	 * If a snapshot has been taken since we made our estimates,
1184 	 * assume that we won't be able to free or overwrite anything.
1185 	 */
1186 	if (tx->tx_objset &&
1187 	    dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1188 	    tx->tx_lastsnap_txg) {
1189 		towrite += tooverwrite;
1190 		tooverwrite = tofree = 0;
1191 	}
1192 
1193 	/* needed allocation: worst-case estimate of write space */
1194 	asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1195 	/* freed space estimate: worst-case overwrite + free estimate */
1196 	fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1197 	/* convert unrefd space to worst-case estimate */
1198 	usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1199 	/* calculate memory footprint estimate */
1200 	memory = towrite + tooverwrite + tohold;
1201 
1202 #ifdef ZFS_DEBUG
1203 	/*
1204 	 * Add in 'tohold' to account for our dirty holds on this memory
1205 	 * XXX - the "fudge" factor is to account for skipped blocks that
1206 	 * we missed because dnode_next_offset() misses in-core-only blocks.
1207 	 */
1208 	tx->tx_space_towrite = asize +
1209 	    spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1210 	tx->tx_space_tofree = tofree;
1211 	tx->tx_space_tooverwrite = tooverwrite;
1212 	tx->tx_space_tounref = tounref;
1213 #endif
1214 
1215 	if (tx->tx_dir && asize != 0) {
1216 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1217 		    asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1218 		if (err)
1219 			return (err);
1220 	}
1221 
1222 	return (0);
1223 }
1224 
1225 static void
dmu_tx_unassign(dmu_tx_t * tx)1226 dmu_tx_unassign(dmu_tx_t *tx)
1227 {
1228 	dmu_tx_hold_t *txh;
1229 
1230 	if (tx->tx_txg == 0)
1231 		return;
1232 
1233 	txg_rele_to_quiesce(&tx->tx_txgh);
1234 
1235 	/*
1236 	 * Walk the transaction's hold list, removing the hold on the
1237 	 * associated dnode, and notifying waiters if the refcount drops to 0.
1238 	 */
1239 	for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1240 	    txh = list_next(&tx->tx_holds, txh)) {
1241 		dnode_t *dn = txh->txh_dnode;
1242 
1243 		if (dn == NULL)
1244 			continue;
1245 		mutex_enter(&dn->dn_mtx);
1246 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1247 
1248 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1249 			dn->dn_assigned_txg = 0;
1250 			cv_broadcast(&dn->dn_notxholds);
1251 		}
1252 		mutex_exit(&dn->dn_mtx);
1253 	}
1254 
1255 	txg_rele_to_sync(&tx->tx_txgh);
1256 
1257 	tx->tx_lasttried_txg = tx->tx_txg;
1258 	tx->tx_txg = 0;
1259 }
1260 
1261 /*
1262  * Assign tx to a transaction group.  txg_how can be one of:
1263  *
1264  * (1)	TXG_WAIT.  If the current open txg is full, waits until there's
1265  *	a new one.  This should be used when you're not holding locks.
1266  *	It will only fail if we're truly out of space (or over quota).
1267  *
1268  * (2)	TXG_NOWAIT.  If we can't assign into the current open txg without
1269  *	blocking, returns immediately with ERESTART.  This should be used
1270  *	whenever you're holding locks.  On an ERESTART error, the caller
1271  *	should drop locks, do a dmu_tx_wait(tx), and try again.
1272  *
1273  * (3)  TXG_WAITED.  Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1274  *      has already been called on behalf of this operation (though
1275  *      most likely on a different tx).
1276  */
1277 int
dmu_tx_assign(dmu_tx_t * tx,txg_how_t txg_how)1278 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1279 {
1280 	int err;
1281 
1282 	ASSERT(tx->tx_txg == 0);
1283 	ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1284 	    txg_how == TXG_WAITED);
1285 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1286 
1287 	/* If we might wait, we must not hold the config lock. */
1288 	ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1289 
1290 	if (txg_how == TXG_WAITED)
1291 		tx->tx_waited = B_TRUE;
1292 
1293 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1294 		dmu_tx_unassign(tx);
1295 
1296 		if (err != ERESTART || txg_how != TXG_WAIT)
1297 			return (err);
1298 
1299 		dmu_tx_wait(tx);
1300 	}
1301 
1302 	txg_rele_to_quiesce(&tx->tx_txgh);
1303 
1304 	return (0);
1305 }
1306 
1307 void
dmu_tx_wait(dmu_tx_t * tx)1308 dmu_tx_wait(dmu_tx_t *tx)
1309 {
1310 	spa_t *spa = tx->tx_pool->dp_spa;
1311 	dsl_pool_t *dp = tx->tx_pool;
1312 
1313 	ASSERT(tx->tx_txg == 0);
1314 	ASSERT(!dsl_pool_config_held(tx->tx_pool));
1315 
1316 	if (tx->tx_wait_dirty) {
1317 		/*
1318 		 * dmu_tx_try_assign() has determined that we need to wait
1319 		 * because we've consumed much or all of the dirty buffer
1320 		 * space.
1321 		 */
1322 		mutex_enter(&dp->dp_lock);
1323 		while (dp->dp_dirty_total >= zfs_dirty_data_max)
1324 			cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1325 		uint64_t dirty = dp->dp_dirty_total;
1326 		mutex_exit(&dp->dp_lock);
1327 
1328 		dmu_tx_delay(tx, dirty);
1329 
1330 		tx->tx_wait_dirty = B_FALSE;
1331 
1332 		/*
1333 		 * Note: setting tx_waited only has effect if the caller
1334 		 * used TX_WAIT.  Otherwise they are going to destroy
1335 		 * this tx and try again.  The common case, zfs_write(),
1336 		 * uses TX_WAIT.
1337 		 */
1338 		tx->tx_waited = B_TRUE;
1339 	} else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1340 		/*
1341 		 * If the pool is suspended we need to wait until it
1342 		 * is resumed.  Note that it's possible that the pool
1343 		 * has become active after this thread has tried to
1344 		 * obtain a tx.  If that's the case then tx_lasttried_txg
1345 		 * would not have been set.
1346 		 */
1347 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1348 	} else if (tx->tx_needassign_txh) {
1349 		/*
1350 		 * A dnode is assigned to the quiescing txg.  Wait for its
1351 		 * transaction to complete.
1352 		 */
1353 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1354 
1355 		mutex_enter(&dn->dn_mtx);
1356 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1357 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1358 		mutex_exit(&dn->dn_mtx);
1359 		tx->tx_needassign_txh = NULL;
1360 	} else {
1361 		txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1362 	}
1363 }
1364 
1365 void
dmu_tx_willuse_space(dmu_tx_t * tx,int64_t delta)1366 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1367 {
1368 #ifdef ZFS_DEBUG
1369 	if (tx->tx_dir == NULL || delta == 0)
1370 		return;
1371 
1372 	if (delta > 0) {
1373 		ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1374 		    tx->tx_space_towrite);
1375 		(void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1376 	} else {
1377 		(void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1378 	}
1379 #endif
1380 }
1381 
1382 void
dmu_tx_commit(dmu_tx_t * tx)1383 dmu_tx_commit(dmu_tx_t *tx)
1384 {
1385 	dmu_tx_hold_t *txh;
1386 
1387 	ASSERT(tx->tx_txg != 0);
1388 
1389 	/*
1390 	 * Go through the transaction's hold list and remove holds on
1391 	 * associated dnodes, notifying waiters if no holds remain.
1392 	 */
1393 	while (txh = list_head(&tx->tx_holds)) {
1394 		dnode_t *dn = txh->txh_dnode;
1395 
1396 		list_remove(&tx->tx_holds, txh);
1397 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1398 		if (dn == NULL)
1399 			continue;
1400 		mutex_enter(&dn->dn_mtx);
1401 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1402 
1403 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1404 			dn->dn_assigned_txg = 0;
1405 			cv_broadcast(&dn->dn_notxholds);
1406 		}
1407 		mutex_exit(&dn->dn_mtx);
1408 		dnode_rele(dn, tx);
1409 	}
1410 
1411 	if (tx->tx_tempreserve_cookie)
1412 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1413 
1414 	if (!list_is_empty(&tx->tx_callbacks))
1415 		txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1416 
1417 	if (tx->tx_anyobj == FALSE)
1418 		txg_rele_to_sync(&tx->tx_txgh);
1419 
1420 	list_destroy(&tx->tx_callbacks);
1421 	list_destroy(&tx->tx_holds);
1422 #ifdef ZFS_DEBUG
1423 	dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1424 	    tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1425 	    tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1426 	refcount_destroy_many(&tx->tx_space_written,
1427 	    refcount_count(&tx->tx_space_written));
1428 	refcount_destroy_many(&tx->tx_space_freed,
1429 	    refcount_count(&tx->tx_space_freed));
1430 #endif
1431 	kmem_free(tx, sizeof (dmu_tx_t));
1432 }
1433 
1434 void
dmu_tx_abort(dmu_tx_t * tx)1435 dmu_tx_abort(dmu_tx_t *tx)
1436 {
1437 	dmu_tx_hold_t *txh;
1438 
1439 	ASSERT(tx->tx_txg == 0);
1440 
1441 	while (txh = list_head(&tx->tx_holds)) {
1442 		dnode_t *dn = txh->txh_dnode;
1443 
1444 		list_remove(&tx->tx_holds, txh);
1445 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1446 		if (dn != NULL)
1447 			dnode_rele(dn, tx);
1448 	}
1449 
1450 	/*
1451 	 * Call any registered callbacks with an error code.
1452 	 */
1453 	if (!list_is_empty(&tx->tx_callbacks))
1454 		dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1455 
1456 	list_destroy(&tx->tx_callbacks);
1457 	list_destroy(&tx->tx_holds);
1458 #ifdef ZFS_DEBUG
1459 	refcount_destroy_many(&tx->tx_space_written,
1460 	    refcount_count(&tx->tx_space_written));
1461 	refcount_destroy_many(&tx->tx_space_freed,
1462 	    refcount_count(&tx->tx_space_freed));
1463 #endif
1464 	kmem_free(tx, sizeof (dmu_tx_t));
1465 }
1466 
1467 uint64_t
dmu_tx_get_txg(dmu_tx_t * tx)1468 dmu_tx_get_txg(dmu_tx_t *tx)
1469 {
1470 	ASSERT(tx->tx_txg != 0);
1471 	return (tx->tx_txg);
1472 }
1473 
1474 dsl_pool_t *
dmu_tx_pool(dmu_tx_t * tx)1475 dmu_tx_pool(dmu_tx_t *tx)
1476 {
1477 	ASSERT(tx->tx_pool != NULL);
1478 	return (tx->tx_pool);
1479 }
1480 
1481 
1482 void
dmu_tx_callback_register(dmu_tx_t * tx,dmu_tx_callback_func_t * func,void * data)1483 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1484 {
1485 	dmu_tx_callback_t *dcb;
1486 
1487 	dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1488 
1489 	dcb->dcb_func = func;
1490 	dcb->dcb_data = data;
1491 
1492 	list_insert_tail(&tx->tx_callbacks, dcb);
1493 }
1494 
1495 /*
1496  * Call all the commit callbacks on a list, with a given error code.
1497  */
1498 void
dmu_tx_do_callbacks(list_t * cb_list,int error)1499 dmu_tx_do_callbacks(list_t *cb_list, int error)
1500 {
1501 	dmu_tx_callback_t *dcb;
1502 
1503 	while (dcb = list_head(cb_list)) {
1504 		list_remove(cb_list, dcb);
1505 		dcb->dcb_func(dcb->dcb_data, error);
1506 		kmem_free(dcb, sizeof (dmu_tx_callback_t));
1507 	}
1508 }
1509 
1510 /*
1511  * Interface to hold a bunch of attributes.
1512  * used for creating new files.
1513  * attrsize is the total size of all attributes
1514  * to be added during object creation
1515  *
1516  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1517  */
1518 
1519 /*
1520  * hold necessary attribute name for attribute registration.
1521  * should be a very rare case where this is needed.  If it does
1522  * happen it would only happen on the first write to the file system.
1523  */
1524 static void
dmu_tx_sa_registration_hold(sa_os_t * sa,dmu_tx_t * tx)1525 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1526 {
1527 	int i;
1528 
1529 	if (!sa->sa_need_attr_registration)
1530 		return;
1531 
1532 	for (i = 0; i != sa->sa_num_attrs; i++) {
1533 		if (!sa->sa_attr_table[i].sa_registered) {
1534 			if (sa->sa_reg_attr_obj)
1535 				dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1536 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1537 			else
1538 				dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1539 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1540 		}
1541 	}
1542 }
1543 
1544 
1545 void
dmu_tx_hold_spill(dmu_tx_t * tx,uint64_t object)1546 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1547 {
1548 	dnode_t *dn;
1549 	dmu_tx_hold_t *txh;
1550 
1551 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1552 	    THT_SPILL, 0, 0);
1553 
1554 	dn = txh->txh_dnode;
1555 
1556 	if (dn == NULL)
1557 		return;
1558 
1559 	/* If blkptr doesn't exist then add space to towrite */
1560 	if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1561 		txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1562 	} else {
1563 		blkptr_t *bp;
1564 
1565 		bp = &dn->dn_phys->dn_spill;
1566 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1567 		    bp, bp->blk_birth))
1568 			txh->txh_space_tooverwrite += SPA_OLD_MAXBLOCKSIZE;
1569 		else
1570 			txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1571 		if (!BP_IS_HOLE(bp))
1572 			txh->txh_space_tounref += SPA_OLD_MAXBLOCKSIZE;
1573 	}
1574 }
1575 
1576 void
dmu_tx_hold_sa_create(dmu_tx_t * tx,int attrsize)1577 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1578 {
1579 	sa_os_t *sa = tx->tx_objset->os_sa;
1580 
1581 	dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1582 
1583 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1584 		return;
1585 
1586 	if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1587 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1588 	else {
1589 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1590 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1591 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1592 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1593 	}
1594 
1595 	dmu_tx_sa_registration_hold(sa, tx);
1596 
1597 	if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1598 		return;
1599 
1600 	(void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1601 	    THT_SPILL, 0, 0);
1602 }
1603 
1604 /*
1605  * Hold SA attribute
1606  *
1607  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1608  *
1609  * variable_size is the total size of all variable sized attributes
1610  * passed to this function.  It is not the total size of all
1611  * variable size attributes that *may* exist on this object.
1612  */
1613 void
dmu_tx_hold_sa(dmu_tx_t * tx,sa_handle_t * hdl,boolean_t may_grow)1614 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1615 {
1616 	uint64_t object;
1617 	sa_os_t *sa = tx->tx_objset->os_sa;
1618 
1619 	ASSERT(hdl != NULL);
1620 
1621 	object = sa_handle_object(hdl);
1622 
1623 	dmu_tx_hold_bonus(tx, object);
1624 
1625 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1626 		return;
1627 
1628 	if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1629 	    tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1630 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1631 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1632 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1633 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1634 	}
1635 
1636 	dmu_tx_sa_registration_hold(sa, tx);
1637 
1638 	if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1639 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1640 
1641 	if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1642 		ASSERT(tx->tx_txg == 0);
1643 		dmu_tx_hold_spill(tx, object);
1644 	} else {
1645 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1646 		dnode_t *dn;
1647 
1648 		DB_DNODE_ENTER(db);
1649 		dn = DB_DNODE(db);
1650 		if (dn->dn_have_spill) {
1651 			ASSERT(tx->tx_txg == 0);
1652 			dmu_tx_hold_spill(tx, object);
1653 		}
1654 		DB_DNODE_EXIT(db);
1655 	}
1656 }
1657