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