xref: /titanic_41/usr/src/uts/common/fs/zfs/dmu_tx.c (revision 5f9bb2f350e19a276d3aff33afb5fc94fb47f02a)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
25  */
26 
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dbuf.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
33 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
34 #include <sys/dsl_pool.h>
35 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 #include <sys/spa.h>
37 #include <sys/sa.h>
38 #include <sys/sa_impl.h>
39 #include <sys/zfs_context.h>
40 #include <sys/varargs.h>
41 
42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43     uint64_t arg1, uint64_t arg2);
44 
45 
46 dmu_tx_t *
dmu_tx_create_dd(dsl_dir_t * dd)47 dmu_tx_create_dd(dsl_dir_t *dd)
48 {
49 	dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
50 	tx->tx_dir = dd;
51 	if (dd != NULL)
52 		tx->tx_pool = dd->dd_pool;
53 	list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
54 	    offsetof(dmu_tx_hold_t, txh_node));
55 	list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
56 	    offsetof(dmu_tx_callback_t, dcb_node));
57 	tx->tx_start = gethrtime();
58 #ifdef ZFS_DEBUG
59 	refcount_create(&tx->tx_space_written);
60 	refcount_create(&tx->tx_space_freed);
61 #endif
62 	return (tx);
63 }
64 
65 dmu_tx_t *
dmu_tx_create(objset_t * os)66 dmu_tx_create(objset_t *os)
67 {
68 	dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
69 	tx->tx_objset = os;
70 	tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
71 	return (tx);
72 }
73 
74 dmu_tx_t *
dmu_tx_create_assigned(struct dsl_pool * dp,uint64_t txg)75 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
76 {
77 	dmu_tx_t *tx = dmu_tx_create_dd(NULL);
78 
79 	ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
80 	tx->tx_pool = dp;
81 	tx->tx_txg = txg;
82 	tx->tx_anyobj = TRUE;
83 
84 	return (tx);
85 }
86 
87 int
dmu_tx_is_syncing(dmu_tx_t * tx)88 dmu_tx_is_syncing(dmu_tx_t *tx)
89 {
90 	return (tx->tx_anyobj);
91 }
92 
93 int
dmu_tx_private_ok(dmu_tx_t * tx)94 dmu_tx_private_ok(dmu_tx_t *tx)
95 {
96 	return (tx->tx_anyobj);
97 }
98 
99 static dmu_tx_hold_t *
dmu_tx_hold_object_impl(dmu_tx_t * tx,objset_t * os,uint64_t object,enum dmu_tx_hold_type type,uint64_t arg1,uint64_t arg2)100 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
101     enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
102 {
103 	dmu_tx_hold_t *txh;
104 	dnode_t *dn = NULL;
105 	int err;
106 
107 	if (object != DMU_NEW_OBJECT) {
108 		err = dnode_hold(os, object, tx, &dn);
109 		if (err) {
110 			tx->tx_err = err;
111 			return (NULL);
112 		}
113 
114 		if (err == 0 && tx->tx_txg != 0) {
115 			mutex_enter(&dn->dn_mtx);
116 			/*
117 			 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
118 			 * problem, but there's no way for it to happen (for
119 			 * now, at least).
120 			 */
121 			ASSERT(dn->dn_assigned_txg == 0);
122 			dn->dn_assigned_txg = tx->tx_txg;
123 			(void) refcount_add(&dn->dn_tx_holds, tx);
124 			mutex_exit(&dn->dn_mtx);
125 		}
126 	}
127 
128 	txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
129 	txh->txh_tx = tx;
130 	txh->txh_dnode = dn;
131 #ifdef ZFS_DEBUG
132 	txh->txh_type = type;
133 	txh->txh_arg1 = arg1;
134 	txh->txh_arg2 = arg2;
135 #endif
136 	list_insert_tail(&tx->tx_holds, txh);
137 
138 	return (txh);
139 }
140 
141 void
dmu_tx_add_new_object(dmu_tx_t * tx,objset_t * os,uint64_t object)142 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
143 {
144 	/*
145 	 * If we're syncing, they can manipulate any object anyhow, and
146 	 * the hold on the dnode_t can cause problems.
147 	 */
148 	if (!dmu_tx_is_syncing(tx)) {
149 		(void) dmu_tx_hold_object_impl(tx, os,
150 		    object, THT_NEWOBJECT, 0, 0);
151 	}
152 }
153 
154 static int
dmu_tx_check_ioerr(zio_t * zio,dnode_t * dn,int level,uint64_t blkid)155 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
156 {
157 	int err;
158 	dmu_buf_impl_t *db;
159 
160 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
161 	db = dbuf_hold_level(dn, level, blkid, FTAG);
162 	rw_exit(&dn->dn_struct_rwlock);
163 	if (db == NULL)
164 		return (SET_ERROR(EIO));
165 	err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
166 	dbuf_rele(db, FTAG);
167 	return (err);
168 }
169 
170 static void
dmu_tx_count_twig(dmu_tx_hold_t * txh,dnode_t * dn,dmu_buf_impl_t * db,int level,uint64_t blkid,boolean_t freeable,uint64_t * history)171 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
172     int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
173 {
174 	objset_t *os = dn->dn_objset;
175 	dsl_dataset_t *ds = os->os_dsl_dataset;
176 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
177 	dmu_buf_impl_t *parent = NULL;
178 	blkptr_t *bp = NULL;
179 	uint64_t space;
180 
181 	if (level >= dn->dn_nlevels || history[level] == blkid)
182 		return;
183 
184 	history[level] = blkid;
185 
186 	space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
187 
188 	if (db == NULL || db == dn->dn_dbuf) {
189 		ASSERT(level != 0);
190 		db = NULL;
191 	} else {
192 		ASSERT(DB_DNODE(db) == dn);
193 		ASSERT(db->db_level == level);
194 		ASSERT(db->db.db_size == space);
195 		ASSERT(db->db_blkid == blkid);
196 		bp = db->db_blkptr;
197 		parent = db->db_parent;
198 	}
199 
200 	freeable = (bp && (freeable ||
201 	    dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
202 
203 	if (freeable)
204 		txh->txh_space_tooverwrite += space;
205 	else
206 		txh->txh_space_towrite += space;
207 	if (bp)
208 		txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
209 
210 	dmu_tx_count_twig(txh, dn, parent, level + 1,
211 	    blkid >> epbs, freeable, history);
212 }
213 
214 /* ARGSUSED */
215 static void
dmu_tx_count_write(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)216 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
217 {
218 	dnode_t *dn = txh->txh_dnode;
219 	uint64_t start, end, i;
220 	int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
221 	int err = 0;
222 
223 	if (len == 0)
224 		return;
225 
226 	min_bs = SPA_MINBLOCKSHIFT;
227 	max_bs = highbit64(txh->txh_tx->tx_objset->os_recordsize) - 1;
228 	min_ibs = DN_MIN_INDBLKSHIFT;
229 	max_ibs = DN_MAX_INDBLKSHIFT;
230 
231 	if (dn) {
232 		uint64_t history[DN_MAX_LEVELS];
233 		int nlvls = dn->dn_nlevels;
234 		int delta;
235 
236 		/*
237 		 * For i/o error checking, read the first and last level-0
238 		 * blocks (if they are not aligned), and all the level-1 blocks.
239 		 */
240 		if (dn->dn_maxblkid == 0) {
241 			delta = dn->dn_datablksz;
242 			start = (off < dn->dn_datablksz) ? 0 : 1;
243 			end = (off+len <= dn->dn_datablksz) ? 0 : 1;
244 			if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
245 				err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
246 				if (err)
247 					goto out;
248 				delta -= off;
249 			}
250 		} else {
251 			zio_t *zio = zio_root(dn->dn_objset->os_spa,
252 			    NULL, NULL, ZIO_FLAG_CANFAIL);
253 
254 			/* first level-0 block */
255 			start = off >> dn->dn_datablkshift;
256 			if (P2PHASE(off, dn->dn_datablksz) ||
257 			    len < dn->dn_datablksz) {
258 				err = dmu_tx_check_ioerr(zio, dn, 0, start);
259 				if (err)
260 					goto out;
261 			}
262 
263 			/* last level-0 block */
264 			end = (off+len-1) >> dn->dn_datablkshift;
265 			if (end != start && end <= dn->dn_maxblkid &&
266 			    P2PHASE(off+len, dn->dn_datablksz)) {
267 				err = dmu_tx_check_ioerr(zio, dn, 0, end);
268 				if (err)
269 					goto out;
270 			}
271 
272 			/* level-1 blocks */
273 			if (nlvls > 1) {
274 				int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
275 				for (i = (start>>shft)+1; i < end>>shft; i++) {
276 					err = dmu_tx_check_ioerr(zio, dn, 1, i);
277 					if (err)
278 						goto out;
279 				}
280 			}
281 
282 			err = zio_wait(zio);
283 			if (err)
284 				goto out;
285 			delta = P2NPHASE(off, dn->dn_datablksz);
286 		}
287 
288 		min_ibs = max_ibs = dn->dn_indblkshift;
289 		if (dn->dn_maxblkid > 0) {
290 			/*
291 			 * The blocksize can't change,
292 			 * so we can make a more precise estimate.
293 			 */
294 			ASSERT(dn->dn_datablkshift != 0);
295 			min_bs = max_bs = dn->dn_datablkshift;
296 		} else {
297 			/*
298 			 * The blocksize can increase up to the recordsize,
299 			 * or if it is already more than the recordsize,
300 			 * up to the next power of 2.
301 			 */
302 			min_bs = highbit64(dn->dn_datablksz - 1);
303 			max_bs = MAX(max_bs, highbit64(dn->dn_datablksz - 1));
304 		}
305 
306 		/*
307 		 * If this write is not off the end of the file
308 		 * we need to account for overwrites/unref.
309 		 */
310 		if (start <= dn->dn_maxblkid) {
311 			for (int l = 0; l < DN_MAX_LEVELS; l++)
312 				history[l] = -1ULL;
313 		}
314 		while (start <= dn->dn_maxblkid) {
315 			dmu_buf_impl_t *db;
316 
317 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
318 			err = dbuf_hold_impl(dn, 0, start,
319 			    FALSE, FALSE, FTAG, &db);
320 			rw_exit(&dn->dn_struct_rwlock);
321 
322 			if (err) {
323 				txh->txh_tx->tx_err = err;
324 				return;
325 			}
326 
327 			dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
328 			    history);
329 			dbuf_rele(db, FTAG);
330 			if (++start > end) {
331 				/*
332 				 * Account for new indirects appearing
333 				 * before this IO gets assigned into a txg.
334 				 */
335 				bits = 64 - min_bs;
336 				epbs = min_ibs - SPA_BLKPTRSHIFT;
337 				for (bits -= epbs * (nlvls - 1);
338 				    bits >= 0; bits -= epbs)
339 					txh->txh_fudge += 1ULL << max_ibs;
340 				goto out;
341 			}
342 			off += delta;
343 			if (len >= delta)
344 				len -= delta;
345 			delta = dn->dn_datablksz;
346 		}
347 	}
348 
349 	/*
350 	 * 'end' is the last thing we will access, not one past.
351 	 * This way we won't overflow when accessing the last byte.
352 	 */
353 	start = P2ALIGN(off, 1ULL << max_bs);
354 	end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
355 	txh->txh_space_towrite += end - start + 1;
356 
357 	start >>= min_bs;
358 	end >>= min_bs;
359 
360 	epbs = min_ibs - SPA_BLKPTRSHIFT;
361 
362 	/*
363 	 * The object contains at most 2^(64 - min_bs) blocks,
364 	 * and each indirect level maps 2^epbs.
365 	 */
366 	for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
367 		start >>= epbs;
368 		end >>= epbs;
369 		ASSERT3U(end, >=, start);
370 		txh->txh_space_towrite += (end - start + 1) << max_ibs;
371 		if (start != 0) {
372 			/*
373 			 * We also need a new blkid=0 indirect block
374 			 * to reference any existing file data.
375 			 */
376 			txh->txh_space_towrite += 1ULL << max_ibs;
377 		}
378 	}
379 
380 out:
381 	if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
382 	    2 * DMU_MAX_ACCESS)
383 		err = SET_ERROR(EFBIG);
384 
385 	if (err)
386 		txh->txh_tx->tx_err = err;
387 }
388 
389 static void
dmu_tx_count_dnode(dmu_tx_hold_t * txh)390 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
391 {
392 	dnode_t *dn = txh->txh_dnode;
393 	dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
394 	uint64_t space = mdn->dn_datablksz +
395 	    ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
396 
397 	if (dn && dn->dn_dbuf->db_blkptr &&
398 	    dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
399 	    dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
400 		txh->txh_space_tooverwrite += space;
401 		txh->txh_space_tounref += space;
402 	} else {
403 		txh->txh_space_towrite += space;
404 		if (dn && dn->dn_dbuf->db_blkptr)
405 			txh->txh_space_tounref += space;
406 	}
407 }
408 
409 void
dmu_tx_hold_write(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)410 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
411 {
412 	dmu_tx_hold_t *txh;
413 
414 	ASSERT(tx->tx_txg == 0);
415 	ASSERT(len < DMU_MAX_ACCESS);
416 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
417 
418 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
419 	    object, THT_WRITE, off, len);
420 	if (txh == NULL)
421 		return;
422 
423 	dmu_tx_count_write(txh, off, len);
424 	dmu_tx_count_dnode(txh);
425 }
426 
427 static void
dmu_tx_count_free(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)428 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
429 {
430 	uint64_t blkid, nblks, lastblk;
431 	uint64_t space = 0, unref = 0, skipped = 0;
432 	dnode_t *dn = txh->txh_dnode;
433 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
434 	spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
435 	int epbs;
436 	uint64_t l0span = 0, nl1blks = 0;
437 
438 	if (dn->dn_nlevels == 0)
439 		return;
440 
441 	/*
442 	 * The struct_rwlock protects us against dn_nlevels
443 	 * changing, in case (against all odds) we manage to dirty &
444 	 * sync out the changes after we check for being dirty.
445 	 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
446 	 */
447 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
448 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
449 	if (dn->dn_maxblkid == 0) {
450 		if (off == 0 && len >= dn->dn_datablksz) {
451 			blkid = 0;
452 			nblks = 1;
453 		} else {
454 			rw_exit(&dn->dn_struct_rwlock);
455 			return;
456 		}
457 	} else {
458 		blkid = off >> dn->dn_datablkshift;
459 		nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
460 
461 		if (blkid > dn->dn_maxblkid) {
462 			rw_exit(&dn->dn_struct_rwlock);
463 			return;
464 		}
465 		if (blkid + nblks > dn->dn_maxblkid)
466 			nblks = dn->dn_maxblkid - blkid + 1;
467 
468 	}
469 	l0span = nblks;    /* save for later use to calc level > 1 overhead */
470 	if (dn->dn_nlevels == 1) {
471 		int i;
472 		for (i = 0; i < nblks; i++) {
473 			blkptr_t *bp = dn->dn_phys->dn_blkptr;
474 			ASSERT3U(blkid + i, <, dn->dn_nblkptr);
475 			bp += blkid + i;
476 			if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
477 				dprintf_bp(bp, "can free old%s", "");
478 				space += bp_get_dsize(spa, bp);
479 			}
480 			unref += BP_GET_ASIZE(bp);
481 		}
482 		nl1blks = 1;
483 		nblks = 0;
484 	}
485 
486 	lastblk = blkid + nblks - 1;
487 	while (nblks) {
488 		dmu_buf_impl_t *dbuf;
489 		uint64_t ibyte, new_blkid;
490 		int epb = 1 << epbs;
491 		int err, i, blkoff, tochk;
492 		blkptr_t *bp;
493 
494 		ibyte = blkid << dn->dn_datablkshift;
495 		err = dnode_next_offset(dn,
496 		    DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
497 		new_blkid = ibyte >> dn->dn_datablkshift;
498 		if (err == ESRCH) {
499 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
500 			break;
501 		}
502 		if (err) {
503 			txh->txh_tx->tx_err = err;
504 			break;
505 		}
506 		if (new_blkid > lastblk) {
507 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
508 			break;
509 		}
510 
511 		if (new_blkid > blkid) {
512 			ASSERT((new_blkid >> epbs) > (blkid >> epbs));
513 			skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
514 			nblks -= new_blkid - blkid;
515 			blkid = new_blkid;
516 		}
517 		blkoff = P2PHASE(blkid, epb);
518 		tochk = MIN(epb - blkoff, nblks);
519 
520 		err = dbuf_hold_impl(dn, 1, blkid >> epbs,
521 		    FALSE, FALSE, FTAG, &dbuf);
522 		if (err) {
523 			txh->txh_tx->tx_err = err;
524 			break;
525 		}
526 
527 		txh->txh_memory_tohold += dbuf->db.db_size;
528 
529 		/*
530 		 * We don't check memory_tohold against DMU_MAX_ACCESS because
531 		 * memory_tohold is an over-estimation (especially the >L1
532 		 * indirect blocks), so it could fail.  Callers should have
533 		 * already verified that they will not be holding too much
534 		 * memory.
535 		 */
536 
537 		err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
538 		if (err != 0) {
539 			txh->txh_tx->tx_err = err;
540 			dbuf_rele(dbuf, FTAG);
541 			break;
542 		}
543 
544 		bp = dbuf->db.db_data;
545 		bp += blkoff;
546 
547 		for (i = 0; i < tochk; i++) {
548 			if (dsl_dataset_block_freeable(ds, &bp[i],
549 			    bp[i].blk_birth)) {
550 				dprintf_bp(&bp[i], "can free old%s", "");
551 				space += bp_get_dsize(spa, &bp[i]);
552 			}
553 			unref += BP_GET_ASIZE(bp);
554 		}
555 		dbuf_rele(dbuf, FTAG);
556 
557 		++nl1blks;
558 		blkid += tochk;
559 		nblks -= tochk;
560 	}
561 	rw_exit(&dn->dn_struct_rwlock);
562 
563 	/*
564 	 * Add in memory requirements of higher-level indirects.
565 	 * This assumes a worst-possible scenario for dn_nlevels and a
566 	 * worst-possible distribution of l1-blocks over the region to free.
567 	 */
568 	{
569 		uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
570 		int level = 2;
571 		/*
572 		 * Here we don't use DN_MAX_LEVEL, but calculate it with the
573 		 * given datablkshift and indblkshift. This makes the
574 		 * difference between 19 and 8 on large files.
575 		 */
576 		int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
577 		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
578 
579 		while (level++ < maxlevel) {
580 			txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
581 			    << dn->dn_indblkshift;
582 			blkcnt = 1 + (blkcnt >> epbs);
583 		}
584 	}
585 
586 	/* account for new level 1 indirect blocks that might show up */
587 	if (skipped > 0) {
588 		txh->txh_fudge += skipped << dn->dn_indblkshift;
589 		skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
590 		txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
591 	}
592 	txh->txh_space_tofree += space;
593 	txh->txh_space_tounref += unref;
594 }
595 
596 /*
597  * This function marks the transaction as being a "net free".  The end
598  * result is that refquotas will be disabled for this transaction, and
599  * this transaction will be able to use half of the pool space overhead
600  * (see dsl_pool_adjustedsize()).  Therefore this function should only
601  * be called for transactions that we expect will not cause a net increase
602  * in the amount of space used (but it's OK if that is occasionally not true).
603  */
604 void
dmu_tx_mark_netfree(dmu_tx_t * tx)605 dmu_tx_mark_netfree(dmu_tx_t *tx)
606 {
607 	dmu_tx_hold_t *txh;
608 
609 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
610 	    DMU_NEW_OBJECT, THT_FREE, 0, 0);
611 
612 	/*
613 	 * Pretend that this operation will free 1GB of space.  This
614 	 * should be large enough to cancel out the largest write.
615 	 * We don't want to use something like UINT64_MAX, because that would
616 	 * cause overflows when doing math with these values (e.g. in
617 	 * dmu_tx_try_assign()).
618 	 */
619 	txh->txh_space_tofree = txh->txh_space_tounref = 1024 * 1024 * 1024;
620 }
621 
622 void
dmu_tx_hold_free(dmu_tx_t * tx,uint64_t object,uint64_t off,uint64_t len)623 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
624 {
625 	dmu_tx_hold_t *txh;
626 	dnode_t *dn;
627 	int err;
628 	zio_t *zio;
629 
630 	ASSERT(tx->tx_txg == 0);
631 
632 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
633 	    object, THT_FREE, off, len);
634 	if (txh == NULL)
635 		return;
636 	dn = txh->txh_dnode;
637 	dmu_tx_count_dnode(txh);
638 
639 	if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
640 		return;
641 	if (len == DMU_OBJECT_END)
642 		len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
643 
644 	/*
645 	 * For i/o error checking, we read the first and last level-0
646 	 * blocks if they are not aligned, and all the level-1 blocks.
647 	 *
648 	 * Note:  dbuf_free_range() assumes that we have not instantiated
649 	 * any level-0 dbufs that will be completely freed.  Therefore we must
650 	 * exercise care to not read or count the first and last blocks
651 	 * if they are blocksize-aligned.
652 	 */
653 	if (dn->dn_datablkshift == 0) {
654 		if (off != 0 || len < dn->dn_datablksz)
655 			dmu_tx_count_write(txh, 0, dn->dn_datablksz);
656 	} else {
657 		/* first block will be modified if it is not aligned */
658 		if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
659 			dmu_tx_count_write(txh, off, 1);
660 		/* last block will be modified if it is not aligned */
661 		if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
662 			dmu_tx_count_write(txh, off+len, 1);
663 	}
664 
665 	/*
666 	 * Check level-1 blocks.
667 	 */
668 	if (dn->dn_nlevels > 1) {
669 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
670 		    SPA_BLKPTRSHIFT;
671 		uint64_t start = off >> shift;
672 		uint64_t end = (off + len) >> shift;
673 
674 		ASSERT(dn->dn_indblkshift != 0);
675 
676 		/*
677 		 * dnode_reallocate() can result in an object with indirect
678 		 * blocks having an odd data block size.  In this case,
679 		 * just check the single block.
680 		 */
681 		if (dn->dn_datablkshift == 0)
682 			start = end = 0;
683 
684 		zio = zio_root(tx->tx_pool->dp_spa,
685 		    NULL, NULL, ZIO_FLAG_CANFAIL);
686 		for (uint64_t i = start; i <= end; i++) {
687 			uint64_t ibyte = i << shift;
688 			err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
689 			i = ibyte >> shift;
690 			if (err == ESRCH || i > end)
691 				break;
692 			if (err) {
693 				tx->tx_err = err;
694 				return;
695 			}
696 
697 			err = dmu_tx_check_ioerr(zio, dn, 1, i);
698 			if (err) {
699 				tx->tx_err = err;
700 				return;
701 			}
702 		}
703 		err = zio_wait(zio);
704 		if (err) {
705 			tx->tx_err = err;
706 			return;
707 		}
708 	}
709 
710 	dmu_tx_count_free(txh, off, len);
711 }
712 
713 void
dmu_tx_hold_zap(dmu_tx_t * tx,uint64_t object,int add,const char * name)714 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
715 {
716 	dmu_tx_hold_t *txh;
717 	dnode_t *dn;
718 	dsl_dataset_phys_t *ds_phys;
719 	uint64_t nblocks;
720 	int epbs, err;
721 
722 	ASSERT(tx->tx_txg == 0);
723 
724 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
725 	    object, THT_ZAP, add, (uintptr_t)name);
726 	if (txh == NULL)
727 		return;
728 	dn = txh->txh_dnode;
729 
730 	dmu_tx_count_dnode(txh);
731 
732 	if (dn == NULL) {
733 		/*
734 		 * We will be able to fit a new object's entries into one leaf
735 		 * block.  So there will be at most 2 blocks total,
736 		 * including the header block.
737 		 */
738 		dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
739 		return;
740 	}
741 
742 	ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
743 
744 	if (dn->dn_maxblkid == 0 && !add) {
745 		blkptr_t *bp;
746 
747 		/*
748 		 * If there is only one block  (i.e. this is a micro-zap)
749 		 * and we are not adding anything, the accounting is simple.
750 		 */
751 		err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
752 		if (err) {
753 			tx->tx_err = err;
754 			return;
755 		}
756 
757 		/*
758 		 * Use max block size here, since we don't know how much
759 		 * the size will change between now and the dbuf dirty call.
760 		 */
761 		bp = &dn->dn_phys->dn_blkptr[0];
762 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
763 		    bp, bp->blk_birth))
764 			txh->txh_space_tooverwrite += MZAP_MAX_BLKSZ;
765 		else
766 			txh->txh_space_towrite += MZAP_MAX_BLKSZ;
767 		if (!BP_IS_HOLE(bp))
768 			txh->txh_space_tounref += MZAP_MAX_BLKSZ;
769 		return;
770 	}
771 
772 	if (dn->dn_maxblkid > 0 && name) {
773 		/*
774 		 * access the name in this fat-zap so that we'll check
775 		 * for i/o errors to the leaf blocks, etc.
776 		 */
777 		err = zap_lookup(dn->dn_objset, dn->dn_object, name,
778 		    8, 0, NULL);
779 		if (err == EIO) {
780 			tx->tx_err = err;
781 			return;
782 		}
783 	}
784 
785 	err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
786 	    &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
787 
788 	/*
789 	 * If the modified blocks are scattered to the four winds,
790 	 * we'll have to modify an indirect twig for each.
791 	 */
792 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
793 	ds_phys = dsl_dataset_phys(dn->dn_objset->os_dsl_dataset);
794 	for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
795 		if (ds_phys->ds_prev_snap_obj)
796 			txh->txh_space_towrite += 3 << dn->dn_indblkshift;
797 		else
798 			txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
799 }
800 
801 void
dmu_tx_hold_bonus(dmu_tx_t * tx,uint64_t object)802 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
803 {
804 	dmu_tx_hold_t *txh;
805 
806 	ASSERT(tx->tx_txg == 0);
807 
808 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
809 	    object, THT_BONUS, 0, 0);
810 	if (txh)
811 		dmu_tx_count_dnode(txh);
812 }
813 
814 void
dmu_tx_hold_space(dmu_tx_t * tx,uint64_t space)815 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
816 {
817 	dmu_tx_hold_t *txh;
818 	ASSERT(tx->tx_txg == 0);
819 
820 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
821 	    DMU_NEW_OBJECT, THT_SPACE, space, 0);
822 
823 	txh->txh_space_towrite += space;
824 }
825 
826 int
dmu_tx_holds(dmu_tx_t * tx,uint64_t object)827 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
828 {
829 	dmu_tx_hold_t *txh;
830 	int holds = 0;
831 
832 	/*
833 	 * By asserting that the tx is assigned, we're counting the
834 	 * number of dn_tx_holds, which is the same as the number of
835 	 * dn_holds.  Otherwise, we'd be counting dn_holds, but
836 	 * dn_tx_holds could be 0.
837 	 */
838 	ASSERT(tx->tx_txg != 0);
839 
840 	/* if (tx->tx_anyobj == TRUE) */
841 		/* return (0); */
842 
843 	for (txh = list_head(&tx->tx_holds); txh;
844 	    txh = list_next(&tx->tx_holds, txh)) {
845 		if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
846 			holds++;
847 	}
848 
849 	return (holds);
850 }
851 
852 #ifdef ZFS_DEBUG
853 void
dmu_tx_dirty_buf(dmu_tx_t * tx,dmu_buf_impl_t * db)854 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
855 {
856 	dmu_tx_hold_t *txh;
857 	int match_object = FALSE, match_offset = FALSE;
858 	dnode_t *dn;
859 
860 	DB_DNODE_ENTER(db);
861 	dn = DB_DNODE(db);
862 	ASSERT(tx->tx_txg != 0);
863 	ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
864 	ASSERT3U(dn->dn_object, ==, db->db.db_object);
865 
866 	if (tx->tx_anyobj) {
867 		DB_DNODE_EXIT(db);
868 		return;
869 	}
870 
871 	/* XXX No checking on the meta dnode for now */
872 	if (db->db.db_object == DMU_META_DNODE_OBJECT) {
873 		DB_DNODE_EXIT(db);
874 		return;
875 	}
876 
877 	for (txh = list_head(&tx->tx_holds); txh;
878 	    txh = list_next(&tx->tx_holds, txh)) {
879 		ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
880 		if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
881 			match_object = TRUE;
882 		if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
883 			int datablkshift = dn->dn_datablkshift ?
884 			    dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
885 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
886 			int shift = datablkshift + epbs * db->db_level;
887 			uint64_t beginblk = shift >= 64 ? 0 :
888 			    (txh->txh_arg1 >> shift);
889 			uint64_t endblk = shift >= 64 ? 0 :
890 			    ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
891 			uint64_t blkid = db->db_blkid;
892 
893 			/* XXX txh_arg2 better not be zero... */
894 
895 			dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
896 			    txh->txh_type, beginblk, endblk);
897 
898 			switch (txh->txh_type) {
899 			case THT_WRITE:
900 				if (blkid >= beginblk && blkid <= endblk)
901 					match_offset = TRUE;
902 				/*
903 				 * We will let this hold work for the bonus
904 				 * or spill buffer so that we don't need to
905 				 * hold it when creating a new object.
906 				 */
907 				if (blkid == DMU_BONUS_BLKID ||
908 				    blkid == DMU_SPILL_BLKID)
909 					match_offset = TRUE;
910 				/*
911 				 * They might have to increase nlevels,
912 				 * thus dirtying the new TLIBs.  Or the
913 				 * might have to change the block size,
914 				 * thus dirying the new lvl=0 blk=0.
915 				 */
916 				if (blkid == 0)
917 					match_offset = TRUE;
918 				break;
919 			case THT_FREE:
920 				/*
921 				 * We will dirty all the level 1 blocks in
922 				 * the free range and perhaps the first and
923 				 * last level 0 block.
924 				 */
925 				if (blkid >= beginblk && (blkid <= endblk ||
926 				    txh->txh_arg2 == DMU_OBJECT_END))
927 					match_offset = TRUE;
928 				break;
929 			case THT_SPILL:
930 				if (blkid == DMU_SPILL_BLKID)
931 					match_offset = TRUE;
932 				break;
933 			case THT_BONUS:
934 				if (blkid == DMU_BONUS_BLKID)
935 					match_offset = TRUE;
936 				break;
937 			case THT_ZAP:
938 				match_offset = TRUE;
939 				break;
940 			case THT_NEWOBJECT:
941 				match_object = TRUE;
942 				break;
943 			default:
944 				ASSERT(!"bad txh_type");
945 			}
946 		}
947 		if (match_object && match_offset) {
948 			DB_DNODE_EXIT(db);
949 			return;
950 		}
951 	}
952 	DB_DNODE_EXIT(db);
953 	panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
954 	    (u_longlong_t)db->db.db_object, db->db_level,
955 	    (u_longlong_t)db->db_blkid);
956 }
957 #endif
958 
959 /*
960  * If we can't do 10 iops, something is wrong.  Let us go ahead
961  * and hit zfs_dirty_data_max.
962  */
963 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
964 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
965 
966 /*
967  * We delay transactions when we've determined that the backend storage
968  * isn't able to accommodate the rate of incoming writes.
969  *
970  * If there is already a transaction waiting, we delay relative to when
971  * that transaction finishes waiting.  This way the calculated min_time
972  * is independent of the number of threads concurrently executing
973  * transactions.
974  *
975  * If we are the only waiter, wait relative to when the transaction
976  * started, rather than the current time.  This credits the transaction for
977  * "time already served", e.g. reading indirect blocks.
978  *
979  * The minimum time for a transaction to take is calculated as:
980  *     min_time = scale * (dirty - min) / (max - dirty)
981  *     min_time is then capped at zfs_delay_max_ns.
982  *
983  * The delay has two degrees of freedom that can be adjusted via tunables.
984  * The percentage of dirty data at which we start to delay is defined by
985  * zfs_delay_min_dirty_percent. This should typically be at or above
986  * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
987  * delay after writing at full speed has failed to keep up with the incoming
988  * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
989  * speaking, this variable determines the amount of delay at the midpoint of
990  * the curve.
991  *
992  * delay
993  *  10ms +-------------------------------------------------------------*+
994  *       |                                                             *|
995  *   9ms +                                                             *+
996  *       |                                                             *|
997  *   8ms +                                                             *+
998  *       |                                                            * |
999  *   7ms +                                                            * +
1000  *       |                                                            * |
1001  *   6ms +                                                            * +
1002  *       |                                                            * |
1003  *   5ms +                                                           *  +
1004  *       |                                                           *  |
1005  *   4ms +                                                           *  +
1006  *       |                                                           *  |
1007  *   3ms +                                                          *   +
1008  *       |                                                          *   |
1009  *   2ms +                                              (midpoint) *    +
1010  *       |                                                  |    **     |
1011  *   1ms +                                                  v ***       +
1012  *       |             zfs_delay_scale ---------->     ********         |
1013  *     0 +-------------------------------------*********----------------+
1014  *       0%                    <- zfs_dirty_data_max ->               100%
1015  *
1016  * Note that since the delay is added to the outstanding time remaining on the
1017  * most recent transaction, the delay is effectively the inverse of IOPS.
1018  * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
1019  * was chosen such that small changes in the amount of accumulated dirty data
1020  * in the first 3/4 of the curve yield relatively small differences in the
1021  * amount of delay.
1022  *
1023  * The effects can be easier to understand when the amount of delay is
1024  * represented on a log scale:
1025  *
1026  * delay
1027  * 100ms +-------------------------------------------------------------++
1028  *       +                                                              +
1029  *       |                                                              |
1030  *       +                                                             *+
1031  *  10ms +                                                             *+
1032  *       +                                                           ** +
1033  *       |                                              (midpoint)  **  |
1034  *       +                                                  |     **    +
1035  *   1ms +                                                  v ****      +
1036  *       +             zfs_delay_scale ---------->        *****         +
1037  *       |                                             ****             |
1038  *       +                                          ****                +
1039  * 100us +                                        **                    +
1040  *       +                                       *                      +
1041  *       |                                      *                       |
1042  *       +                                     *                        +
1043  *  10us +                                     *                        +
1044  *       +                                                              +
1045  *       |                                                              |
1046  *       +                                                              +
1047  *       +--------------------------------------------------------------+
1048  *       0%                    <- zfs_dirty_data_max ->               100%
1049  *
1050  * Note here that only as the amount of dirty data approaches its limit does
1051  * the delay start to increase rapidly. The goal of a properly tuned system
1052  * should be to keep the amount of dirty data out of that range by first
1053  * ensuring that the appropriate limits are set for the I/O scheduler to reach
1054  * optimal throughput on the backend storage, and then by changing the value
1055  * of zfs_delay_scale to increase the steepness of the curve.
1056  */
1057 static void
dmu_tx_delay(dmu_tx_t * tx,uint64_t dirty)1058 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1059 {
1060 	dsl_pool_t *dp = tx->tx_pool;
1061 	uint64_t delay_min_bytes =
1062 	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1063 	hrtime_t wakeup, min_tx_time, now;
1064 
1065 	if (dirty <= delay_min_bytes)
1066 		return;
1067 
1068 	/*
1069 	 * The caller has already waited until we are under the max.
1070 	 * We make them pass us the amount of dirty data so we don't
1071 	 * have to handle the case of it being >= the max, which could
1072 	 * cause a divide-by-zero if it's == the max.
1073 	 */
1074 	ASSERT3U(dirty, <, zfs_dirty_data_max);
1075 
1076 	now = gethrtime();
1077 	min_tx_time = zfs_delay_scale *
1078 	    (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1079 	if (now > tx->tx_start + min_tx_time)
1080 		return;
1081 
1082 	min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1083 
1084 	DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1085 	    uint64_t, min_tx_time);
1086 
1087 	mutex_enter(&dp->dp_lock);
1088 	wakeup = MAX(tx->tx_start + min_tx_time,
1089 	    dp->dp_last_wakeup + min_tx_time);
1090 	dp->dp_last_wakeup = wakeup;
1091 	mutex_exit(&dp->dp_lock);
1092 
1093 #ifdef _KERNEL
1094 	mutex_enter(&curthread->t_delay_lock);
1095 	while (cv_timedwait_hires(&curthread->t_delay_cv,
1096 	    &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
1097 	    CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
1098 		continue;
1099 	mutex_exit(&curthread->t_delay_lock);
1100 #else
1101 	hrtime_t delta = wakeup - gethrtime();
1102 	struct timespec ts;
1103 	ts.tv_sec = delta / NANOSEC;
1104 	ts.tv_nsec = delta % NANOSEC;
1105 	(void) nanosleep(&ts, NULL);
1106 #endif
1107 }
1108 
1109 static int
dmu_tx_try_assign(dmu_tx_t * tx,txg_how_t txg_how)1110 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1111 {
1112 	dmu_tx_hold_t *txh;
1113 	spa_t *spa = tx->tx_pool->dp_spa;
1114 	uint64_t memory, asize, fsize, usize;
1115 	uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1116 
1117 	ASSERT0(tx->tx_txg);
1118 
1119 	if (tx->tx_err)
1120 		return (tx->tx_err);
1121 
1122 	if (spa_suspended(spa)) {
1123 		/*
1124 		 * If the user has indicated a blocking failure mode
1125 		 * then return ERESTART which will block in dmu_tx_wait().
1126 		 * Otherwise, return EIO so that an error can get
1127 		 * propagated back to the VOP calls.
1128 		 *
1129 		 * Note that we always honor the txg_how flag regardless
1130 		 * of the failuremode setting.
1131 		 */
1132 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1133 		    txg_how != TXG_WAIT)
1134 			return (SET_ERROR(EIO));
1135 
1136 		return (SET_ERROR(ERESTART));
1137 	}
1138 
1139 	if (!tx->tx_waited &&
1140 	    dsl_pool_need_dirty_delay(tx->tx_pool)) {
1141 		tx->tx_wait_dirty = B_TRUE;
1142 		return (SET_ERROR(ERESTART));
1143 	}
1144 
1145 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1146 	tx->tx_needassign_txh = NULL;
1147 
1148 	/*
1149 	 * NB: No error returns are allowed after txg_hold_open, but
1150 	 * before processing the dnode holds, due to the
1151 	 * dmu_tx_unassign() logic.
1152 	 */
1153 
1154 	towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1155 	for (txh = list_head(&tx->tx_holds); txh;
1156 	    txh = list_next(&tx->tx_holds, txh)) {
1157 		dnode_t *dn = txh->txh_dnode;
1158 		if (dn != NULL) {
1159 			mutex_enter(&dn->dn_mtx);
1160 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1161 				mutex_exit(&dn->dn_mtx);
1162 				tx->tx_needassign_txh = txh;
1163 				return (SET_ERROR(ERESTART));
1164 			}
1165 			if (dn->dn_assigned_txg == 0)
1166 				dn->dn_assigned_txg = tx->tx_txg;
1167 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1168 			(void) refcount_add(&dn->dn_tx_holds, tx);
1169 			mutex_exit(&dn->dn_mtx);
1170 		}
1171 		towrite += txh->txh_space_towrite;
1172 		tofree += txh->txh_space_tofree;
1173 		tooverwrite += txh->txh_space_tooverwrite;
1174 		tounref += txh->txh_space_tounref;
1175 		tohold += txh->txh_memory_tohold;
1176 		fudge += txh->txh_fudge;
1177 	}
1178 
1179 	/*
1180 	 * If a snapshot has been taken since we made our estimates,
1181 	 * assume that we won't be able to free or overwrite anything.
1182 	 */
1183 	if (tx->tx_objset &&
1184 	    dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1185 	    tx->tx_lastsnap_txg) {
1186 		towrite += tooverwrite;
1187 		tooverwrite = tofree = 0;
1188 	}
1189 
1190 	/* needed allocation: worst-case estimate of write space */
1191 	asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1192 	/* freed space estimate: worst-case overwrite + free estimate */
1193 	fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1194 	/* convert unrefd space to worst-case estimate */
1195 	usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1196 	/* calculate memory footprint estimate */
1197 	memory = towrite + tooverwrite + tohold;
1198 
1199 #ifdef ZFS_DEBUG
1200 	/*
1201 	 * Add in 'tohold' to account for our dirty holds on this memory
1202 	 * XXX - the "fudge" factor is to account for skipped blocks that
1203 	 * we missed because dnode_next_offset() misses in-core-only blocks.
1204 	 */
1205 	tx->tx_space_towrite = asize +
1206 	    spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1207 	tx->tx_space_tofree = tofree;
1208 	tx->tx_space_tooverwrite = tooverwrite;
1209 	tx->tx_space_tounref = tounref;
1210 #endif
1211 
1212 	if (tx->tx_dir && asize != 0) {
1213 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1214 		    asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1215 		if (err)
1216 			return (err);
1217 	}
1218 
1219 	return (0);
1220 }
1221 
1222 static void
dmu_tx_unassign(dmu_tx_t * tx)1223 dmu_tx_unassign(dmu_tx_t *tx)
1224 {
1225 	dmu_tx_hold_t *txh;
1226 
1227 	if (tx->tx_txg == 0)
1228 		return;
1229 
1230 	txg_rele_to_quiesce(&tx->tx_txgh);
1231 
1232 	/*
1233 	 * Walk the transaction's hold list, removing the hold on the
1234 	 * associated dnode, and notifying waiters if the refcount drops to 0.
1235 	 */
1236 	for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1237 	    txh = list_next(&tx->tx_holds, txh)) {
1238 		dnode_t *dn = txh->txh_dnode;
1239 
1240 		if (dn == NULL)
1241 			continue;
1242 		mutex_enter(&dn->dn_mtx);
1243 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1244 
1245 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1246 			dn->dn_assigned_txg = 0;
1247 			cv_broadcast(&dn->dn_notxholds);
1248 		}
1249 		mutex_exit(&dn->dn_mtx);
1250 	}
1251 
1252 	txg_rele_to_sync(&tx->tx_txgh);
1253 
1254 	tx->tx_lasttried_txg = tx->tx_txg;
1255 	tx->tx_txg = 0;
1256 }
1257 
1258 /*
1259  * Assign tx to a transaction group.  txg_how can be one of:
1260  *
1261  * (1)	TXG_WAIT.  If the current open txg is full, waits until there's
1262  *	a new one.  This should be used when you're not holding locks.
1263  *	It will only fail if we're truly out of space (or over quota).
1264  *
1265  * (2)	TXG_NOWAIT.  If we can't assign into the current open txg without
1266  *	blocking, returns immediately with ERESTART.  This should be used
1267  *	whenever you're holding locks.  On an ERESTART error, the caller
1268  *	should drop locks, do a dmu_tx_wait(tx), and try again.
1269  *
1270  * (3)  TXG_WAITED.  Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1271  *      has already been called on behalf of this operation (though
1272  *      most likely on a different tx).
1273  */
1274 int
dmu_tx_assign(dmu_tx_t * tx,txg_how_t txg_how)1275 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1276 {
1277 	int err;
1278 
1279 	ASSERT(tx->tx_txg == 0);
1280 	ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1281 	    txg_how == TXG_WAITED);
1282 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1283 
1284 	/* If we might wait, we must not hold the config lock. */
1285 	ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1286 
1287 	if (txg_how == TXG_WAITED)
1288 		tx->tx_waited = B_TRUE;
1289 
1290 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1291 		dmu_tx_unassign(tx);
1292 
1293 		if (err != ERESTART || txg_how != TXG_WAIT)
1294 			return (err);
1295 
1296 		dmu_tx_wait(tx);
1297 	}
1298 
1299 	txg_rele_to_quiesce(&tx->tx_txgh);
1300 
1301 	return (0);
1302 }
1303 
1304 void
dmu_tx_wait(dmu_tx_t * tx)1305 dmu_tx_wait(dmu_tx_t *tx)
1306 {
1307 	spa_t *spa = tx->tx_pool->dp_spa;
1308 	dsl_pool_t *dp = tx->tx_pool;
1309 
1310 	ASSERT(tx->tx_txg == 0);
1311 	ASSERT(!dsl_pool_config_held(tx->tx_pool));
1312 
1313 	if (tx->tx_wait_dirty) {
1314 		/*
1315 		 * dmu_tx_try_assign() has determined that we need to wait
1316 		 * because we've consumed much or all of the dirty buffer
1317 		 * space.
1318 		 */
1319 		mutex_enter(&dp->dp_lock);
1320 		while (dp->dp_dirty_total >= zfs_dirty_data_max)
1321 			cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1322 		uint64_t dirty = dp->dp_dirty_total;
1323 		mutex_exit(&dp->dp_lock);
1324 
1325 		dmu_tx_delay(tx, dirty);
1326 
1327 		tx->tx_wait_dirty = B_FALSE;
1328 
1329 		/*
1330 		 * Note: setting tx_waited only has effect if the caller
1331 		 * used TX_WAIT.  Otherwise they are going to destroy
1332 		 * this tx and try again.  The common case, zfs_write(),
1333 		 * uses TX_WAIT.
1334 		 */
1335 		tx->tx_waited = B_TRUE;
1336 	} else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1337 		/*
1338 		 * If the pool is suspended we need to wait until it
1339 		 * is resumed.  Note that it's possible that the pool
1340 		 * has become active after this thread has tried to
1341 		 * obtain a tx.  If that's the case then tx_lasttried_txg
1342 		 * would not have been set.
1343 		 */
1344 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1345 	} else if (tx->tx_needassign_txh) {
1346 		/*
1347 		 * A dnode is assigned to the quiescing txg.  Wait for its
1348 		 * transaction to complete.
1349 		 */
1350 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1351 
1352 		mutex_enter(&dn->dn_mtx);
1353 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1354 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1355 		mutex_exit(&dn->dn_mtx);
1356 		tx->tx_needassign_txh = NULL;
1357 	} else {
1358 		txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1359 	}
1360 }
1361 
1362 void
dmu_tx_willuse_space(dmu_tx_t * tx,int64_t delta)1363 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1364 {
1365 #ifdef ZFS_DEBUG
1366 	if (tx->tx_dir == NULL || delta == 0)
1367 		return;
1368 
1369 	if (delta > 0) {
1370 		ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1371 		    tx->tx_space_towrite);
1372 		(void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1373 	} else {
1374 		(void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1375 	}
1376 #endif
1377 }
1378 
1379 void
dmu_tx_commit(dmu_tx_t * tx)1380 dmu_tx_commit(dmu_tx_t *tx)
1381 {
1382 	dmu_tx_hold_t *txh;
1383 
1384 	ASSERT(tx->tx_txg != 0);
1385 
1386 	/*
1387 	 * Go through the transaction's hold list and remove holds on
1388 	 * associated dnodes, notifying waiters if no holds remain.
1389 	 */
1390 	while (txh = list_head(&tx->tx_holds)) {
1391 		dnode_t *dn = txh->txh_dnode;
1392 
1393 		list_remove(&tx->tx_holds, txh);
1394 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1395 		if (dn == NULL)
1396 			continue;
1397 		mutex_enter(&dn->dn_mtx);
1398 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1399 
1400 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1401 			dn->dn_assigned_txg = 0;
1402 			cv_broadcast(&dn->dn_notxholds);
1403 		}
1404 		mutex_exit(&dn->dn_mtx);
1405 		dnode_rele(dn, tx);
1406 	}
1407 
1408 	if (tx->tx_tempreserve_cookie)
1409 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1410 
1411 	if (!list_is_empty(&tx->tx_callbacks))
1412 		txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1413 
1414 	if (tx->tx_anyobj == FALSE)
1415 		txg_rele_to_sync(&tx->tx_txgh);
1416 
1417 	list_destroy(&tx->tx_callbacks);
1418 	list_destroy(&tx->tx_holds);
1419 #ifdef ZFS_DEBUG
1420 	dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1421 	    tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1422 	    tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1423 	refcount_destroy_many(&tx->tx_space_written,
1424 	    refcount_count(&tx->tx_space_written));
1425 	refcount_destroy_many(&tx->tx_space_freed,
1426 	    refcount_count(&tx->tx_space_freed));
1427 #endif
1428 	kmem_free(tx, sizeof (dmu_tx_t));
1429 }
1430 
1431 void
dmu_tx_abort(dmu_tx_t * tx)1432 dmu_tx_abort(dmu_tx_t *tx)
1433 {
1434 	dmu_tx_hold_t *txh;
1435 
1436 	ASSERT(tx->tx_txg == 0);
1437 
1438 	while (txh = list_head(&tx->tx_holds)) {
1439 		dnode_t *dn = txh->txh_dnode;
1440 
1441 		list_remove(&tx->tx_holds, txh);
1442 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1443 		if (dn != NULL)
1444 			dnode_rele(dn, tx);
1445 	}
1446 
1447 	/*
1448 	 * Call any registered callbacks with an error code.
1449 	 */
1450 	if (!list_is_empty(&tx->tx_callbacks))
1451 		dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1452 
1453 	list_destroy(&tx->tx_callbacks);
1454 	list_destroy(&tx->tx_holds);
1455 #ifdef ZFS_DEBUG
1456 	refcount_destroy_many(&tx->tx_space_written,
1457 	    refcount_count(&tx->tx_space_written));
1458 	refcount_destroy_many(&tx->tx_space_freed,
1459 	    refcount_count(&tx->tx_space_freed));
1460 #endif
1461 	kmem_free(tx, sizeof (dmu_tx_t));
1462 }
1463 
1464 uint64_t
dmu_tx_get_txg(dmu_tx_t * tx)1465 dmu_tx_get_txg(dmu_tx_t *tx)
1466 {
1467 	ASSERT(tx->tx_txg != 0);
1468 	return (tx->tx_txg);
1469 }
1470 
1471 dsl_pool_t *
dmu_tx_pool(dmu_tx_t * tx)1472 dmu_tx_pool(dmu_tx_t *tx)
1473 {
1474 	ASSERT(tx->tx_pool != NULL);
1475 	return (tx->tx_pool);
1476 }
1477 
1478 
1479 void
dmu_tx_callback_register(dmu_tx_t * tx,dmu_tx_callback_func_t * func,void * data)1480 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1481 {
1482 	dmu_tx_callback_t *dcb;
1483 
1484 	dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1485 
1486 	dcb->dcb_func = func;
1487 	dcb->dcb_data = data;
1488 
1489 	list_insert_tail(&tx->tx_callbacks, dcb);
1490 }
1491 
1492 /*
1493  * Call all the commit callbacks on a list, with a given error code.
1494  */
1495 void
dmu_tx_do_callbacks(list_t * cb_list,int error)1496 dmu_tx_do_callbacks(list_t *cb_list, int error)
1497 {
1498 	dmu_tx_callback_t *dcb;
1499 
1500 	while (dcb = list_head(cb_list)) {
1501 		list_remove(cb_list, dcb);
1502 		dcb->dcb_func(dcb->dcb_data, error);
1503 		kmem_free(dcb, sizeof (dmu_tx_callback_t));
1504 	}
1505 }
1506 
1507 /*
1508  * Interface to hold a bunch of attributes.
1509  * used for creating new files.
1510  * attrsize is the total size of all attributes
1511  * to be added during object creation
1512  *
1513  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1514  */
1515 
1516 /*
1517  * hold necessary attribute name for attribute registration.
1518  * should be a very rare case where this is needed.  If it does
1519  * happen it would only happen on the first write to the file system.
1520  */
1521 static void
dmu_tx_sa_registration_hold(sa_os_t * sa,dmu_tx_t * tx)1522 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1523 {
1524 	int i;
1525 
1526 	if (!sa->sa_need_attr_registration)
1527 		return;
1528 
1529 	for (i = 0; i != sa->sa_num_attrs; i++) {
1530 		if (!sa->sa_attr_table[i].sa_registered) {
1531 			if (sa->sa_reg_attr_obj)
1532 				dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1533 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1534 			else
1535 				dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1536 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1537 		}
1538 	}
1539 }
1540 
1541 
1542 void
dmu_tx_hold_spill(dmu_tx_t * tx,uint64_t object)1543 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1544 {
1545 	dnode_t *dn;
1546 	dmu_tx_hold_t *txh;
1547 
1548 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1549 	    THT_SPILL, 0, 0);
1550 
1551 	dn = txh->txh_dnode;
1552 
1553 	if (dn == NULL)
1554 		return;
1555 
1556 	/* If blkptr doesn't exist then add space to towrite */
1557 	if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1558 		txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1559 	} else {
1560 		blkptr_t *bp;
1561 
1562 		bp = &dn->dn_phys->dn_spill;
1563 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1564 		    bp, bp->blk_birth))
1565 			txh->txh_space_tooverwrite += SPA_OLD_MAXBLOCKSIZE;
1566 		else
1567 			txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1568 		if (!BP_IS_HOLE(bp))
1569 			txh->txh_space_tounref += SPA_OLD_MAXBLOCKSIZE;
1570 	}
1571 }
1572 
1573 void
dmu_tx_hold_sa_create(dmu_tx_t * tx,int attrsize)1574 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1575 {
1576 	sa_os_t *sa = tx->tx_objset->os_sa;
1577 
1578 	dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1579 
1580 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1581 		return;
1582 
1583 	if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1584 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1585 	else {
1586 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1587 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1588 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1589 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1590 	}
1591 
1592 	dmu_tx_sa_registration_hold(sa, tx);
1593 
1594 	if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1595 		return;
1596 
1597 	(void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1598 	    THT_SPILL, 0, 0);
1599 }
1600 
1601 /*
1602  * Hold SA attribute
1603  *
1604  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1605  *
1606  * variable_size is the total size of all variable sized attributes
1607  * passed to this function.  It is not the total size of all
1608  * variable size attributes that *may* exist on this object.
1609  */
1610 void
dmu_tx_hold_sa(dmu_tx_t * tx,sa_handle_t * hdl,boolean_t may_grow)1611 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1612 {
1613 	uint64_t object;
1614 	sa_os_t *sa = tx->tx_objset->os_sa;
1615 
1616 	ASSERT(hdl != NULL);
1617 
1618 	object = sa_handle_object(hdl);
1619 
1620 	dmu_tx_hold_bonus(tx, object);
1621 
1622 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1623 		return;
1624 
1625 	if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1626 	    tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1627 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1628 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1629 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1630 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1631 	}
1632 
1633 	dmu_tx_sa_registration_hold(sa, tx);
1634 
1635 	if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1636 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1637 
1638 	if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1639 		ASSERT(tx->tx_txg == 0);
1640 		dmu_tx_hold_spill(tx, object);
1641 	} else {
1642 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1643 		dnode_t *dn;
1644 
1645 		DB_DNODE_ENTER(db);
1646 		dn = DB_DNODE(db);
1647 		if (dn->dn_have_spill) {
1648 			ASSERT(tx->tx_txg == 0);
1649 			dmu_tx_hold_spill(tx, object);
1650 		}
1651 		DB_DNODE_EXIT(db);
1652 	}
1653 }
1654