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) 2013 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 *
  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 #ifdef ZFS_DEBUG
  58         refcount_create(&tx->tx_space_written);
  59         refcount_create(&tx->tx_space_freed);
  60 #endif
  61         return (tx);
  62 }
  63 
  64 dmu_tx_t *
  65 dmu_tx_create(objset_t *os)
  66 {
  67         dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
  68         tx->tx_objset = os;
  69         tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
  70         return (tx);
  71 }
  72 
  73 dmu_tx_t *
  74 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
  75 {
  76         dmu_tx_t *tx = dmu_tx_create_dd(NULL);
  77 
  78         ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
  79         tx->tx_pool = dp;
  80         tx->tx_txg = txg;
  81         tx->tx_anyobj = TRUE;
  82 
  83         return (tx);
  84 }
  85 
  86 int
  87 dmu_tx_is_syncing(dmu_tx_t *tx)
  88 {
  89         return (tx->tx_anyobj);
  90 }
  91 
  92 int
  93 dmu_tx_private_ok(dmu_tx_t *tx)
  94 {
  95         return (tx->tx_anyobj);
  96 }
  97 
  98 static dmu_tx_hold_t *
  99 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
 100     enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
 101 {
 102         dmu_tx_hold_t *txh;
 103         dnode_t *dn = NULL;
 104         int err;
 105 
 106         if (object != DMU_NEW_OBJECT) {
 107                 err = dnode_hold(os, object, tx, &dn);
 108                 if (err) {
 109                         tx->tx_err = err;
 110                         return (NULL);
 111                 }
 112 
 113                 if (err == 0 && tx->tx_txg != 0) {
 114                         mutex_enter(&dn->dn_mtx);
 115                         /*
 116                          * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
 117                          * problem, but there's no way for it to happen (for
 118                          * now, at least).
 119                          */
 120                         ASSERT(dn->dn_assigned_txg == 0);
 121                         dn->dn_assigned_txg = tx->tx_txg;
 122                         (void) refcount_add(&dn->dn_tx_holds, tx);
 123                         mutex_exit(&dn->dn_mtx);
 124                 }
 125         }
 126 
 127         txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
 128         txh->txh_tx = tx;
 129         txh->txh_dnode = dn;
 130 #ifdef ZFS_DEBUG
 131         txh->txh_type = type;
 132         txh->txh_arg1 = arg1;
 133         txh->txh_arg2 = arg2;
 134 #endif
 135         list_insert_tail(&tx->tx_holds, txh);
 136 
 137         return (txh);
 138 }
 139 
 140 void
 141 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
 142 {
 143         /*
 144          * If we're syncing, they can manipulate any object anyhow, and
 145          * the hold on the dnode_t can cause problems.
 146          */
 147         if (!dmu_tx_is_syncing(tx)) {
 148                 (void) dmu_tx_hold_object_impl(tx, os,
 149                     object, THT_NEWOBJECT, 0, 0);
 150         }
 151 }
 152 
 153 static int
 154 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
 155 {
 156         int err;
 157         dmu_buf_impl_t *db;
 158 
 159         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 160         db = dbuf_hold_level(dn, level, blkid, FTAG);
 161         rw_exit(&dn->dn_struct_rwlock);
 162         if (db == NULL)
 163                 return (SET_ERROR(EIO));
 164         err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
 165         dbuf_rele(db, FTAG);
 166         return (err);
 167 }
 168 
 169 static void
 170 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
 171     int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
 172 {
 173         objset_t *os = dn->dn_objset;
 174         dsl_dataset_t *ds = os->os_dsl_dataset;
 175         int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
 176         dmu_buf_impl_t *parent = NULL;
 177         blkptr_t *bp = NULL;
 178         uint64_t space;
 179 
 180         if (level >= dn->dn_nlevels || history[level] == blkid)
 181                 return;
 182 
 183         history[level] = blkid;
 184 
 185         space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
 186 
 187         if (db == NULL || db == dn->dn_dbuf) {
 188                 ASSERT(level != 0);
 189                 db = NULL;
 190         } else {
 191                 ASSERT(DB_DNODE(db) == dn);
 192                 ASSERT(db->db_level == level);
 193                 ASSERT(db->db.db_size == space);
 194                 ASSERT(db->db_blkid == blkid);
 195                 bp = db->db_blkptr;
 196                 parent = db->db_parent;
 197         }
 198 
 199         freeable = (bp && (freeable ||
 200             dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
 201 
 202         if (freeable)
 203                 txh->txh_space_tooverwrite += space;
 204         else
 205                 txh->txh_space_towrite += space;
 206         if (bp)
 207                 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
 208 
 209         dmu_tx_count_twig(txh, dn, parent, level + 1,
 210             blkid >> epbs, freeable, history);
 211 }
 212 
 213 /* ARGSUSED */
 214 static void
 215 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
 216 {
 217         dnode_t *dn = txh->txh_dnode;
 218         uint64_t start, end, i;
 219         int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
 220         int err = 0;
 221 
 222         if (len == 0)
 223                 return;
 224 
 225         min_bs = SPA_MINBLOCKSHIFT;
 226         max_bs = SPA_MAXBLOCKSHIFT;
 227         min_ibs = DN_MIN_INDBLKSHIFT;
 228         max_ibs = DN_MAX_INDBLKSHIFT;
 229 
 230         if (dn) {
 231                 uint64_t history[DN_MAX_LEVELS];
 232                 int nlvls = dn->dn_nlevels;
 233                 int delta;
 234 
 235                 /*
 236                  * For i/o error checking, read the first and last level-0
 237                  * blocks (if they are not aligned), and all the level-1 blocks.
 238                  */
 239                 if (dn->dn_maxblkid == 0) {
 240                         delta = dn->dn_datablksz;
 241                         start = (off < dn->dn_datablksz) ? 0 : 1;
 242                         end = (off+len <= dn->dn_datablksz) ? 0 : 1;
 243                         if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
 244                                 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
 245                                 if (err)
 246                                         goto out;
 247                                 delta -= off;
 248                         }
 249                 } else {
 250                         zio_t *zio = zio_root(dn->dn_objset->os_spa,
 251                             NULL, NULL, ZIO_FLAG_CANFAIL);
 252 
 253                         /* first level-0 block */
 254                         start = off >> dn->dn_datablkshift;
 255                         if (P2PHASE(off, dn->dn_datablksz) ||
 256                             len < dn->dn_datablksz) {
 257                                 err = dmu_tx_check_ioerr(zio, dn, 0, start);
 258                                 if (err)
 259                                         goto out;
 260                         }
 261 
 262                         /* last level-0 block */
 263                         end = (off+len-1) >> dn->dn_datablkshift;
 264                         if (end != start && end <= dn->dn_maxblkid &&
 265                             P2PHASE(off+len, dn->dn_datablksz)) {
 266                                 err = dmu_tx_check_ioerr(zio, dn, 0, end);
 267                                 if (err)
 268                                         goto out;
 269                         }
 270 
 271                         /* level-1 blocks */
 272                         if (nlvls > 1) {
 273                                 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
 274                                 for (i = (start>>shft)+1; i < end>>shft; i++) {
 275                                         err = dmu_tx_check_ioerr(zio, dn, 1, i);
 276                                         if (err)
 277                                                 goto out;
 278                                 }
 279                         }
 280 
 281                         err = zio_wait(zio);
 282                         if (err)
 283                                 goto out;
 284                         delta = P2NPHASE(off, dn->dn_datablksz);
 285                 }
 286 
 287                 min_ibs = max_ibs = dn->dn_indblkshift;
 288                 if (dn->dn_maxblkid > 0) {
 289                         /*
 290                          * The blocksize can't change,
 291                          * so we can make a more precise estimate.
 292                          */
 293                         ASSERT(dn->dn_datablkshift != 0);
 294                         min_bs = max_bs = dn->dn_datablkshift;
 295                 }
 296 
 297                 /*
 298                  * If this write is not off the end of the file
 299                  * we need to account for overwrites/unref.
 300                  */
 301                 if (start <= dn->dn_maxblkid) {
 302                         for (int l = 0; l < DN_MAX_LEVELS; l++)
 303                                 history[l] = -1ULL;
 304                 }
 305                 while (start <= dn->dn_maxblkid) {
 306                         dmu_buf_impl_t *db;
 307 
 308                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 309                         err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
 310                         rw_exit(&dn->dn_struct_rwlock);
 311 
 312                         if (err) {
 313                                 txh->txh_tx->tx_err = err;
 314                                 return;
 315                         }
 316 
 317                         dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
 318                             history);
 319                         dbuf_rele(db, FTAG);
 320                         if (++start > end) {
 321                                 /*
 322                                  * Account for new indirects appearing
 323                                  * before this IO gets assigned into a txg.
 324                                  */
 325                                 bits = 64 - min_bs;
 326                                 epbs = min_ibs - SPA_BLKPTRSHIFT;
 327                                 for (bits -= epbs * (nlvls - 1);
 328                                     bits >= 0; bits -= epbs)
 329                                         txh->txh_fudge += 1ULL << max_ibs;
 330                                 goto out;
 331                         }
 332                         off += delta;
 333                         if (len >= delta)
 334                                 len -= delta;
 335                         delta = dn->dn_datablksz;
 336                 }
 337         }
 338 
 339         /*
 340          * 'end' is the last thing we will access, not one past.
 341          * This way we won't overflow when accessing the last byte.
 342          */
 343         start = P2ALIGN(off, 1ULL << max_bs);
 344         end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
 345         txh->txh_space_towrite += end - start + 1;
 346 
 347         start >>= min_bs;
 348         end >>= min_bs;
 349 
 350         epbs = min_ibs - SPA_BLKPTRSHIFT;
 351 
 352         /*
 353          * The object contains at most 2^(64 - min_bs) blocks,
 354          * and each indirect level maps 2^epbs.
 355          */
 356         for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
 357                 start >>= epbs;
 358                 end >>= epbs;
 359                 ASSERT3U(end, >=, start);
 360                 txh->txh_space_towrite += (end - start + 1) << max_ibs;
 361                 if (start != 0) {
 362                         /*
 363                          * We also need a new blkid=0 indirect block
 364                          * to reference any existing file data.
 365                          */
 366                         txh->txh_space_towrite += 1ULL << max_ibs;
 367                 }
 368         }
 369 
 370 out:
 371         if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
 372             2 * DMU_MAX_ACCESS)
 373                 err = SET_ERROR(EFBIG);
 374 
 375         if (err)
 376                 txh->txh_tx->tx_err = err;
 377 }
 378 
 379 static void
 380 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
 381 {
 382         dnode_t *dn = txh->txh_dnode;
 383         dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
 384         uint64_t space = mdn->dn_datablksz +
 385             ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
 386 
 387         if (dn && dn->dn_dbuf->db_blkptr &&
 388             dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
 389             dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
 390                 txh->txh_space_tooverwrite += space;
 391                 txh->txh_space_tounref += space;
 392         } else {
 393                 txh->txh_space_towrite += space;
 394                 if (dn && dn->dn_dbuf->db_blkptr)
 395                         txh->txh_space_tounref += space;
 396         }
 397 }
 398 
 399 void
 400 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
 401 {
 402         dmu_tx_hold_t *txh;
 403 
 404         ASSERT(tx->tx_txg == 0);
 405         ASSERT(len < DMU_MAX_ACCESS);
 406         ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
 407 
 408         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
 409             object, THT_WRITE, off, len);
 410         if (txh == NULL)
 411                 return;
 412 
 413         dmu_tx_count_write(txh, off, len);
 414         dmu_tx_count_dnode(txh);
 415 }
 416 
 417 static void
 418 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
 419 {
 420         uint64_t blkid, nblks, lastblk;
 421         uint64_t space = 0, unref = 0, skipped = 0;
 422         dnode_t *dn = txh->txh_dnode;
 423         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
 424         spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
 425         int epbs;
 426         uint64_t l0span = 0, nl1blks = 0;
 427 
 428         if (dn->dn_nlevels == 0)
 429                 return;
 430 
 431         /*
 432          * The struct_rwlock protects us against dn_nlevels
 433          * changing, in case (against all odds) we manage to dirty &
 434          * sync out the changes after we check for being dirty.
 435          * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
 436          */
 437         rw_enter(&dn->dn_struct_rwlock, RW_READER);
 438         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
 439         if (dn->dn_maxblkid == 0) {
 440                 if (off == 0 && len >= dn->dn_datablksz) {
 441                         blkid = 0;
 442                         nblks = 1;
 443                 } else {
 444                         rw_exit(&dn->dn_struct_rwlock);
 445                         return;
 446                 }
 447         } else {
 448                 blkid = off >> dn->dn_datablkshift;
 449                 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
 450 
 451                 if (blkid >= dn->dn_maxblkid) {
 452                         rw_exit(&dn->dn_struct_rwlock);
 453                         return;
 454                 }
 455                 if (blkid + nblks > dn->dn_maxblkid)
 456                         nblks = dn->dn_maxblkid - blkid;
 457 
 458         }
 459         l0span = nblks;    /* save for later use to calc level > 1 overhead */
 460         if (dn->dn_nlevels == 1) {
 461                 int i;
 462                 for (i = 0; i < nblks; i++) {
 463                         blkptr_t *bp = dn->dn_phys->dn_blkptr;
 464                         ASSERT3U(blkid + i, <, dn->dn_nblkptr);
 465                         bp += blkid + i;
 466                         if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
 467                                 dprintf_bp(bp, "can free old%s", "");
 468                                 space += bp_get_dsize(spa, bp);
 469                         }
 470                         unref += BP_GET_ASIZE(bp);
 471                 }
 472                 nl1blks = 1;
 473                 nblks = 0;
 474         }
 475 
 476         lastblk = blkid + nblks - 1;
 477         while (nblks) {
 478                 dmu_buf_impl_t *dbuf;
 479                 uint64_t ibyte, new_blkid;
 480                 int epb = 1 << epbs;
 481                 int err, i, blkoff, tochk;
 482                 blkptr_t *bp;
 483 
 484                 ibyte = blkid << dn->dn_datablkshift;
 485                 err = dnode_next_offset(dn,
 486                     DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
 487                 new_blkid = ibyte >> dn->dn_datablkshift;
 488                 if (err == ESRCH) {
 489                         skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
 490                         break;
 491                 }
 492                 if (err) {
 493                         txh->txh_tx->tx_err = err;
 494                         break;
 495                 }
 496                 if (new_blkid > lastblk) {
 497                         skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
 498                         break;
 499                 }
 500 
 501                 if (new_blkid > blkid) {
 502                         ASSERT((new_blkid >> epbs) > (blkid >> epbs));
 503                         skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
 504                         nblks -= new_blkid - blkid;
 505                         blkid = new_blkid;
 506                 }
 507                 blkoff = P2PHASE(blkid, epb);
 508                 tochk = MIN(epb - blkoff, nblks);
 509 
 510                 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
 511                 if (err) {
 512                         txh->txh_tx->tx_err = err;
 513                         break;
 514                 }
 515 
 516                 txh->txh_memory_tohold += dbuf->db.db_size;
 517 
 518                 /*
 519                  * We don't check memory_tohold against DMU_MAX_ACCESS because
 520                  * memory_tohold is an over-estimation (especially the >L1
 521                  * indirect blocks), so it could fail.  Callers should have
 522                  * already verified that they will not be holding too much
 523                  * memory.
 524                  */
 525 
 526                 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
 527                 if (err != 0) {
 528                         txh->txh_tx->tx_err = err;
 529                         dbuf_rele(dbuf, FTAG);
 530                         break;
 531                 }
 532 
 533                 bp = dbuf->db.db_data;
 534                 bp += blkoff;
 535 
 536                 for (i = 0; i < tochk; i++) {
 537                         if (dsl_dataset_block_freeable(ds, &bp[i],
 538                             bp[i].blk_birth)) {
 539                                 dprintf_bp(&bp[i], "can free old%s", "");
 540                                 space += bp_get_dsize(spa, &bp[i]);
 541                         }
 542                         unref += BP_GET_ASIZE(bp);
 543                 }
 544                 dbuf_rele(dbuf, FTAG);
 545 
 546                 ++nl1blks;
 547                 blkid += tochk;
 548                 nblks -= tochk;
 549         }
 550         rw_exit(&dn->dn_struct_rwlock);
 551 
 552         /*
 553          * Add in memory requirements of higher-level indirects.
 554          * This assumes a worst-possible scenario for dn_nlevels and a
 555          * worst-possible distribution of l1-blocks over the region to free.
 556          */
 557         {
 558                 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
 559                 int level = 2;
 560                 /*
 561                  * Here we don't use DN_MAX_LEVEL, but calculate it with the
 562                  * given datablkshift and indblkshift. This makes the
 563                  * difference between 19 and 8 on large files.
 564                  */
 565                 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
 566                     (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
 567 
 568                 while (level++ < maxlevel) {
 569                         txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
 570                             << dn->dn_indblkshift;
 571                         blkcnt = 1 + (blkcnt >> epbs);
 572                 }
 573         }
 574 
 575         /* account for new level 1 indirect blocks that might show up */
 576         if (skipped > 0) {
 577                 txh->txh_fudge += skipped << dn->dn_indblkshift;
 578                 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
 579                 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
 580         }
 581         txh->txh_space_tofree += space;
 582         txh->txh_space_tounref += unref;
 583 }
 584 
 585 void
 586 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
 587 {
 588         dmu_tx_hold_t *txh;
 589         dnode_t *dn;
 590         int err;
 591         zio_t *zio;
 592 
 593         ASSERT(tx->tx_txg == 0);
 594 
 595         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
 596             object, THT_FREE, off, len);
 597         if (txh == NULL)
 598                 return;
 599         dn = txh->txh_dnode;
 600 
 601         if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
 602                 return;
 603         if (len == DMU_OBJECT_END)
 604                 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
 605 
 606         dmu_tx_count_dnode(txh);
 607 
 608         /*
 609          * For i/o error checking, we read the first and last level-0
 610          * blocks if they are not aligned, and all the level-1 blocks.
 611          *
 612          * Note:  dbuf_free_range() assumes that we have not instantiated
 613          * any level-0 dbufs that will be completely freed.  Therefore we must
 614          * exercise care to not read or count the first and last blocks
 615          * if they are blocksize-aligned.
 616          */
 617         if (dn->dn_datablkshift == 0) {
 618                 dmu_tx_count_write(txh, off, len);
 619         } else {
 620                 /* first block will be modified if it is not aligned */
 621                 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
 622                         dmu_tx_count_write(txh, off, 1);
 623                 /* last block will be modified if it is not aligned */
 624                 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
 625                         dmu_tx_count_write(txh, off+len, 1);
 626         }
 627 
 628         /*
 629          * Check level-1 blocks.
 630          */
 631         if (dn->dn_nlevels > 1) {
 632                 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
 633                     SPA_BLKPTRSHIFT;
 634                 uint64_t start = off >> shift;
 635                 uint64_t end = (off + len) >> shift;
 636 
 637                 ASSERT(dn->dn_datablkshift != 0);
 638                 ASSERT(dn->dn_indblkshift != 0);
 639 
 640                 zio = zio_root(tx->tx_pool->dp_spa,
 641                     NULL, NULL, ZIO_FLAG_CANFAIL);
 642                 for (uint64_t i = start; i <= end; i++) {
 643                         uint64_t ibyte = i << shift;
 644                         err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
 645                         i = ibyte >> shift;
 646                         if (err == ESRCH)
 647                                 break;
 648                         if (err) {
 649                                 tx->tx_err = err;
 650                                 return;
 651                         }
 652 
 653                         err = dmu_tx_check_ioerr(zio, dn, 1, i);
 654                         if (err) {
 655                                 tx->tx_err = err;
 656                                 return;
 657                         }
 658                 }
 659                 err = zio_wait(zio);
 660                 if (err) {
 661                         tx->tx_err = err;
 662                         return;
 663                 }
 664         }
 665 
 666         dmu_tx_count_free(txh, off, len);
 667 }
 668 
 669 void
 670 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
 671 {
 672         dmu_tx_hold_t *txh;
 673         dnode_t *dn;
 674         uint64_t nblocks;
 675         int epbs, err;
 676 
 677         ASSERT(tx->tx_txg == 0);
 678 
 679         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
 680             object, THT_ZAP, add, (uintptr_t)name);
 681         if (txh == NULL)
 682                 return;
 683         dn = txh->txh_dnode;
 684 
 685         dmu_tx_count_dnode(txh);
 686 
 687         if (dn == NULL) {
 688                 /*
 689                  * We will be able to fit a new object's entries into one leaf
 690                  * block.  So there will be at most 2 blocks total,
 691                  * including the header block.
 692                  */
 693                 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
 694                 return;
 695         }
 696 
 697         ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
 698 
 699         if (dn->dn_maxblkid == 0 && !add) {
 700                 blkptr_t *bp;
 701 
 702                 /*
 703                  * If there is only one block  (i.e. this is a micro-zap)
 704                  * and we are not adding anything, the accounting is simple.
 705                  */
 706                 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
 707                 if (err) {
 708                         tx->tx_err = err;
 709                         return;
 710                 }
 711 
 712                 /*
 713                  * Use max block size here, since we don't know how much
 714                  * the size will change between now and the dbuf dirty call.
 715                  */
 716                 bp = &dn->dn_phys->dn_blkptr[0];
 717                 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
 718                     bp, bp->blk_birth))
 719                         txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
 720                 else
 721                         txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
 722                 if (!BP_IS_HOLE(bp))
 723                         txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
 724                 return;
 725         }
 726 
 727         if (dn->dn_maxblkid > 0 && name) {
 728                 /*
 729                  * access the name in this fat-zap so that we'll check
 730                  * for i/o errors to the leaf blocks, etc.
 731                  */
 732                 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
 733                     8, 0, NULL);
 734                 if (err == EIO) {
 735                         tx->tx_err = err;
 736                         return;
 737                 }
 738         }
 739 
 740         err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
 741             &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
 742 
 743         /*
 744          * If the modified blocks are scattered to the four winds,
 745          * we'll have to modify an indirect twig for each.
 746          */
 747         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
 748         for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
 749                 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
 750                         txh->txh_space_towrite += 3 << dn->dn_indblkshift;
 751                 else
 752                         txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
 753 }
 754 
 755 void
 756 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
 757 {
 758         dmu_tx_hold_t *txh;
 759 
 760         ASSERT(tx->tx_txg == 0);
 761 
 762         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
 763             object, THT_BONUS, 0, 0);
 764         if (txh)
 765                 dmu_tx_count_dnode(txh);
 766 }
 767 
 768 void
 769 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
 770 {
 771         dmu_tx_hold_t *txh;
 772         ASSERT(tx->tx_txg == 0);
 773 
 774         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
 775             DMU_NEW_OBJECT, THT_SPACE, space, 0);
 776 
 777         txh->txh_space_towrite += space;
 778 }
 779 
 780 int
 781 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
 782 {
 783         dmu_tx_hold_t *txh;
 784         int holds = 0;
 785 
 786         /*
 787          * By asserting that the tx is assigned, we're counting the
 788          * number of dn_tx_holds, which is the same as the number of
 789          * dn_holds.  Otherwise, we'd be counting dn_holds, but
 790          * dn_tx_holds could be 0.
 791          */
 792         ASSERT(tx->tx_txg != 0);
 793 
 794         /* if (tx->tx_anyobj == TRUE) */
 795                 /* return (0); */
 796 
 797         for (txh = list_head(&tx->tx_holds); txh;
 798             txh = list_next(&tx->tx_holds, txh)) {
 799                 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
 800                         holds++;
 801         }
 802 
 803         return (holds);
 804 }
 805 
 806 #ifdef ZFS_DEBUG
 807 void
 808 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
 809 {
 810         dmu_tx_hold_t *txh;
 811         int match_object = FALSE, match_offset = FALSE;
 812         dnode_t *dn;
 813 
 814         DB_DNODE_ENTER(db);
 815         dn = DB_DNODE(db);
 816         ASSERT(tx->tx_txg != 0);
 817         ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
 818         ASSERT3U(dn->dn_object, ==, db->db.db_object);
 819 
 820         if (tx->tx_anyobj) {
 821                 DB_DNODE_EXIT(db);
 822                 return;
 823         }
 824 
 825         /* XXX No checking on the meta dnode for now */
 826         if (db->db.db_object == DMU_META_DNODE_OBJECT) {
 827                 DB_DNODE_EXIT(db);
 828                 return;
 829         }
 830 
 831         for (txh = list_head(&tx->tx_holds); txh;
 832             txh = list_next(&tx->tx_holds, txh)) {
 833                 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
 834                 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
 835                         match_object = TRUE;
 836                 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
 837                         int datablkshift = dn->dn_datablkshift ?
 838                             dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
 839                         int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
 840                         int shift = datablkshift + epbs * db->db_level;
 841                         uint64_t beginblk = shift >= 64 ? 0 :
 842                             (txh->txh_arg1 >> shift);
 843                         uint64_t endblk = shift >= 64 ? 0 :
 844                             ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
 845                         uint64_t blkid = db->db_blkid;
 846 
 847                         /* XXX txh_arg2 better not be zero... */
 848 
 849                         dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
 850                             txh->txh_type, beginblk, endblk);
 851 
 852                         switch (txh->txh_type) {
 853                         case THT_WRITE:
 854                                 if (blkid >= beginblk && blkid <= endblk)
 855                                         match_offset = TRUE;
 856                                 /*
 857                                  * We will let this hold work for the bonus
 858                                  * or spill buffer so that we don't need to
 859                                  * hold it when creating a new object.
 860                                  */
 861                                 if (blkid == DMU_BONUS_BLKID ||
 862                                     blkid == DMU_SPILL_BLKID)
 863                                         match_offset = TRUE;
 864                                 /*
 865                                  * They might have to increase nlevels,
 866                                  * thus dirtying the new TLIBs.  Or the
 867                                  * might have to change the block size,
 868                                  * thus dirying the new lvl=0 blk=0.
 869                                  */
 870                                 if (blkid == 0)
 871                                         match_offset = TRUE;
 872                                 break;
 873                         case THT_FREE:
 874                                 /*
 875                                  * We will dirty all the level 1 blocks in
 876                                  * the free range and perhaps the first and
 877                                  * last level 0 block.
 878                                  */
 879                                 if (blkid >= beginblk && (blkid <= endblk ||
 880                                     txh->txh_arg2 == DMU_OBJECT_END))
 881                                         match_offset = TRUE;
 882                                 break;
 883                         case THT_SPILL:
 884                                 if (blkid == DMU_SPILL_BLKID)
 885                                         match_offset = TRUE;
 886                                 break;
 887                         case THT_BONUS:
 888                                 if (blkid == DMU_BONUS_BLKID)
 889                                         match_offset = TRUE;
 890                                 break;
 891                         case THT_ZAP:
 892                                 match_offset = TRUE;
 893                                 break;
 894                         case THT_NEWOBJECT:
 895                                 match_object = TRUE;
 896                                 break;
 897                         default:
 898                                 ASSERT(!"bad txh_type");
 899                         }
 900                 }
 901                 if (match_object && match_offset) {
 902                         DB_DNODE_EXIT(db);
 903                         return;
 904                 }
 905         }
 906         DB_DNODE_EXIT(db);
 907         panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
 908             (u_longlong_t)db->db.db_object, db->db_level,
 909             (u_longlong_t)db->db_blkid);
 910 }
 911 #endif
 912 
 913 static int
 914 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
 915 {
 916         dmu_tx_hold_t *txh;
 917         spa_t *spa = tx->tx_pool->dp_spa;
 918         uint64_t memory, asize, fsize, usize;
 919         uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
 920 
 921         ASSERT0(tx->tx_txg);
 922 
 923         if (tx->tx_err)
 924                 return (tx->tx_err);
 925 
 926         if (spa_suspended(spa)) {
 927                 /*
 928                  * If the user has indicated a blocking failure mode
 929                  * then return ERESTART which will block in dmu_tx_wait().
 930                  * Otherwise, return EIO so that an error can get
 931                  * propagated back to the VOP calls.
 932                  *
 933                  * Note that we always honor the txg_how flag regardless
 934                  * of the failuremode setting.
 935                  */
 936                 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
 937                     txg_how != TXG_WAIT)
 938                         return (SET_ERROR(EIO));
 939 
 940                 return (SET_ERROR(ERESTART));
 941         }
 942 
 943         tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
 944         tx->tx_needassign_txh = NULL;
 945 
 946         /*
 947          * NB: No error returns are allowed after txg_hold_open, but
 948          * before processing the dnode holds, due to the
 949          * dmu_tx_unassign() logic.
 950          */
 951 
 952         towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
 953         for (txh = list_head(&tx->tx_holds); txh;
 954             txh = list_next(&tx->tx_holds, txh)) {
 955                 dnode_t *dn = txh->txh_dnode;
 956                 if (dn != NULL) {
 957                         mutex_enter(&dn->dn_mtx);
 958                         if (dn->dn_assigned_txg == tx->tx_txg - 1) {
 959                                 mutex_exit(&dn->dn_mtx);
 960                                 tx->tx_needassign_txh = txh;
 961                                 return (SET_ERROR(ERESTART));
 962                         }
 963                         if (dn->dn_assigned_txg == 0)
 964                                 dn->dn_assigned_txg = tx->tx_txg;
 965                         ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
 966                         (void) refcount_add(&dn->dn_tx_holds, tx);
 967                         mutex_exit(&dn->dn_mtx);
 968                 }
 969                 towrite += txh->txh_space_towrite;
 970                 tofree += txh->txh_space_tofree;
 971                 tooverwrite += txh->txh_space_tooverwrite;
 972                 tounref += txh->txh_space_tounref;
 973                 tohold += txh->txh_memory_tohold;
 974                 fudge += txh->txh_fudge;
 975         }
 976 
 977         /*
 978          * If a snapshot has been taken since we made our estimates,
 979          * assume that we won't be able to free or overwrite anything.
 980          */
 981         if (tx->tx_objset &&
 982             dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
 983             tx->tx_lastsnap_txg) {
 984                 towrite += tooverwrite;
 985                 tooverwrite = tofree = 0;
 986         }
 987 
 988         /* needed allocation: worst-case estimate of write space */
 989         asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
 990         /* freed space estimate: worst-case overwrite + free estimate */
 991         fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
 992         /* convert unrefd space to worst-case estimate */
 993         usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
 994         /* calculate memory footprint estimate */
 995         memory = towrite + tooverwrite + tohold;
 996 
 997 #ifdef ZFS_DEBUG
 998         /*
 999          * Add in 'tohold' to account for our dirty holds on this memory
1000          * XXX - the "fudge" factor is to account for skipped blocks that
1001          * we missed because dnode_next_offset() misses in-core-only blocks.
1002          */
1003         tx->tx_space_towrite = asize +
1004             spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1005         tx->tx_space_tofree = tofree;
1006         tx->tx_space_tooverwrite = tooverwrite;
1007         tx->tx_space_tounref = tounref;
1008 #endif
1009 
1010         if (tx->tx_dir && asize != 0) {
1011                 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1012                     asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1013                 if (err)
1014                         return (err);
1015         }
1016 
1017         return (0);
1018 }
1019 
1020 static void
1021 dmu_tx_unassign(dmu_tx_t *tx)
1022 {
1023         dmu_tx_hold_t *txh;
1024 
1025         if (tx->tx_txg == 0)
1026                 return;
1027 
1028         txg_rele_to_quiesce(&tx->tx_txgh);
1029 
1030         /*
1031          * Walk the transaction's hold list, removing the hold on the
1032          * associated dnode, and notifying waiters if the refcount drops to 0.
1033          */
1034         for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1035             txh = list_next(&tx->tx_holds, txh)) {
1036                 dnode_t *dn = txh->txh_dnode;
1037 
1038                 if (dn == NULL)
1039                         continue;
1040                 mutex_enter(&dn->dn_mtx);
1041                 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1042 
1043                 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1044                         dn->dn_assigned_txg = 0;
1045                         cv_broadcast(&dn->dn_notxholds);
1046                 }
1047                 mutex_exit(&dn->dn_mtx);
1048         }
1049 
1050         txg_rele_to_sync(&tx->tx_txgh);
1051 
1052         tx->tx_lasttried_txg = tx->tx_txg;
1053         tx->tx_txg = 0;
1054 }
1055 
1056 /*
1057  * Assign tx to a transaction group.  txg_how can be one of:
1058  *
1059  * (1)  TXG_WAIT.  If the current open txg is full, waits until there's
1060  *      a new one.  This should be used when you're not holding locks.
1061  *      It will only fail if we're truly out of space (or over quota).
1062  *
1063  * (2)  TXG_NOWAIT.  If we can't assign into the current open txg without
1064  *      blocking, returns immediately with ERESTART.  This should be used
1065  *      whenever you're holding locks.  On an ERESTART error, the caller
1066  *      should drop locks, do a dmu_tx_wait(tx), and try again.
1067  */
1068 int
1069 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1070 {
1071         int err;
1072 
1073         ASSERT(tx->tx_txg == 0);
1074         ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT);
1075         ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1076 
1077         /* If we might wait, we must not hold the config lock. */
1078         ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1079 
1080         while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1081                 dmu_tx_unassign(tx);
1082 
1083                 if (err != ERESTART || txg_how != TXG_WAIT)
1084                         return (err);
1085 
1086                 dmu_tx_wait(tx);
1087         }
1088 
1089         txg_rele_to_quiesce(&tx->tx_txgh);
1090 
1091         return (0);
1092 }
1093 
1094 void
1095 dmu_tx_wait(dmu_tx_t *tx)
1096 {
1097         spa_t *spa = tx->tx_pool->dp_spa;
1098 
1099         ASSERT(tx->tx_txg == 0);
1100         ASSERT(!dsl_pool_config_held(tx->tx_pool));
1101 
1102         /*
1103          * It's possible that the pool has become active after this thread
1104          * has tried to obtain a tx. If that's the case then his
1105          * tx_lasttried_txg would not have been assigned.
1106          */
1107         if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1108                 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1109         } else if (tx->tx_needassign_txh) {
1110                 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1111 
1112                 mutex_enter(&dn->dn_mtx);
1113                 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1114                         cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1115                 mutex_exit(&dn->dn_mtx);
1116                 tx->tx_needassign_txh = NULL;
1117         } else {
1118                 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1119         }
1120 }
1121 
1122 void
1123 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1124 {
1125 #ifdef ZFS_DEBUG
1126         if (tx->tx_dir == NULL || delta == 0)
1127                 return;
1128 
1129         if (delta > 0) {
1130                 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1131                     tx->tx_space_towrite);
1132                 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1133         } else {
1134                 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1135         }
1136 #endif
1137 }
1138 
1139 void
1140 dmu_tx_commit(dmu_tx_t *tx)
1141 {
1142         dmu_tx_hold_t *txh;
1143 
1144         ASSERT(tx->tx_txg != 0);
1145 
1146         /*
1147          * Go through the transaction's hold list and remove holds on
1148          * associated dnodes, notifying waiters if no holds remain.
1149          */
1150         while (txh = list_head(&tx->tx_holds)) {
1151                 dnode_t *dn = txh->txh_dnode;
1152 
1153                 list_remove(&tx->tx_holds, txh);
1154                 kmem_free(txh, sizeof (dmu_tx_hold_t));
1155                 if (dn == NULL)
1156                         continue;
1157                 mutex_enter(&dn->dn_mtx);
1158                 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1159 
1160                 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1161                         dn->dn_assigned_txg = 0;
1162                         cv_broadcast(&dn->dn_notxholds);
1163                 }
1164                 mutex_exit(&dn->dn_mtx);
1165                 dnode_rele(dn, tx);
1166         }
1167 
1168         if (tx->tx_tempreserve_cookie)
1169                 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1170 
1171         if (!list_is_empty(&tx->tx_callbacks))
1172                 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1173 
1174         if (tx->tx_anyobj == FALSE)
1175                 txg_rele_to_sync(&tx->tx_txgh);
1176 
1177         list_destroy(&tx->tx_callbacks);
1178         list_destroy(&tx->tx_holds);
1179 #ifdef ZFS_DEBUG
1180         dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1181             tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1182             tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1183         refcount_destroy_many(&tx->tx_space_written,
1184             refcount_count(&tx->tx_space_written));
1185         refcount_destroy_many(&tx->tx_space_freed,
1186             refcount_count(&tx->tx_space_freed));
1187 #endif
1188         kmem_free(tx, sizeof (dmu_tx_t));
1189 }
1190 
1191 void
1192 dmu_tx_abort(dmu_tx_t *tx)
1193 {
1194         dmu_tx_hold_t *txh;
1195 
1196         ASSERT(tx->tx_txg == 0);
1197 
1198         while (txh = list_head(&tx->tx_holds)) {
1199                 dnode_t *dn = txh->txh_dnode;
1200 
1201                 list_remove(&tx->tx_holds, txh);
1202                 kmem_free(txh, sizeof (dmu_tx_hold_t));
1203                 if (dn != NULL)
1204                         dnode_rele(dn, tx);
1205         }
1206 
1207         /*
1208          * Call any registered callbacks with an error code.
1209          */
1210         if (!list_is_empty(&tx->tx_callbacks))
1211                 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1212 
1213         list_destroy(&tx->tx_callbacks);
1214         list_destroy(&tx->tx_holds);
1215 #ifdef ZFS_DEBUG
1216         refcount_destroy_many(&tx->tx_space_written,
1217             refcount_count(&tx->tx_space_written));
1218         refcount_destroy_many(&tx->tx_space_freed,
1219             refcount_count(&tx->tx_space_freed));
1220 #endif
1221         kmem_free(tx, sizeof (dmu_tx_t));
1222 }
1223 
1224 uint64_t
1225 dmu_tx_get_txg(dmu_tx_t *tx)
1226 {
1227         ASSERT(tx->tx_txg != 0);
1228         return (tx->tx_txg);
1229 }
1230 
1231 dsl_pool_t *
1232 dmu_tx_pool(dmu_tx_t *tx)
1233 {
1234         ASSERT(tx->tx_pool != NULL);
1235         return (tx->tx_pool);
1236 }
1237 
1238 
1239 void
1240 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1241 {
1242         dmu_tx_callback_t *dcb;
1243 
1244         dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1245 
1246         dcb->dcb_func = func;
1247         dcb->dcb_data = data;
1248 
1249         list_insert_tail(&tx->tx_callbacks, dcb);
1250 }
1251 
1252 /*
1253  * Call all the commit callbacks on a list, with a given error code.
1254  */
1255 void
1256 dmu_tx_do_callbacks(list_t *cb_list, int error)
1257 {
1258         dmu_tx_callback_t *dcb;
1259 
1260         while (dcb = list_head(cb_list)) {
1261                 list_remove(cb_list, dcb);
1262                 dcb->dcb_func(dcb->dcb_data, error);
1263                 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1264         }
1265 }
1266 
1267 /*
1268  * Interface to hold a bunch of attributes.
1269  * used for creating new files.
1270  * attrsize is the total size of all attributes
1271  * to be added during object creation
1272  *
1273  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1274  */
1275 
1276 /*
1277  * hold necessary attribute name for attribute registration.
1278  * should be a very rare case where this is needed.  If it does
1279  * happen it would only happen on the first write to the file system.
1280  */
1281 static void
1282 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1283 {
1284         int i;
1285 
1286         if (!sa->sa_need_attr_registration)
1287                 return;
1288 
1289         for (i = 0; i != sa->sa_num_attrs; i++) {
1290                 if (!sa->sa_attr_table[i].sa_registered) {
1291                         if (sa->sa_reg_attr_obj)
1292                                 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1293                                     B_TRUE, sa->sa_attr_table[i].sa_name);
1294                         else
1295                                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1296                                     B_TRUE, sa->sa_attr_table[i].sa_name);
1297                 }
1298         }
1299 }
1300 
1301 
1302 void
1303 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1304 {
1305         dnode_t *dn;
1306         dmu_tx_hold_t *txh;
1307 
1308         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1309             THT_SPILL, 0, 0);
1310 
1311         dn = txh->txh_dnode;
1312 
1313         if (dn == NULL)
1314                 return;
1315 
1316         /* If blkptr doesn't exist then add space to towrite */
1317         if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1318                 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1319         } else {
1320                 blkptr_t *bp;
1321 
1322                 bp = &dn->dn_phys->dn_spill;
1323                 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1324                     bp, bp->blk_birth))
1325                         txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1326                 else
1327                         txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1328                 if (!BP_IS_HOLE(bp))
1329                         txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1330         }
1331 }
1332 
1333 void
1334 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1335 {
1336         sa_os_t *sa = tx->tx_objset->os_sa;
1337 
1338         dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1339 
1340         if (tx->tx_objset->os_sa->sa_master_obj == 0)
1341                 return;
1342 
1343         if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1344                 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1345         else {
1346                 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1347                 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1348                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1349                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1350         }
1351 
1352         dmu_tx_sa_registration_hold(sa, tx);
1353 
1354         if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1355                 return;
1356 
1357         (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1358             THT_SPILL, 0, 0);
1359 }
1360 
1361 /*
1362  * Hold SA attribute
1363  *
1364  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1365  *
1366  * variable_size is the total size of all variable sized attributes
1367  * passed to this function.  It is not the total size of all
1368  * variable size attributes that *may* exist on this object.
1369  */
1370 void
1371 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1372 {
1373         uint64_t object;
1374         sa_os_t *sa = tx->tx_objset->os_sa;
1375 
1376         ASSERT(hdl != NULL);
1377 
1378         object = sa_handle_object(hdl);
1379 
1380         dmu_tx_hold_bonus(tx, object);
1381 
1382         if (tx->tx_objset->os_sa->sa_master_obj == 0)
1383                 return;
1384 
1385         if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1386             tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1387                 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1388                 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1389                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1390                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1391         }
1392 
1393         dmu_tx_sa_registration_hold(sa, tx);
1394 
1395         if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1396                 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1397 
1398         if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1399                 ASSERT(tx->tx_txg == 0);
1400                 dmu_tx_hold_spill(tx, object);
1401         } else {
1402                 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1403                 dnode_t *dn;
1404 
1405                 DB_DNODE_ENTER(db);
1406                 dn = DB_DNODE(db);
1407                 if (dn->dn_have_spill) {
1408                         ASSERT(tx->tx_txg == 0);
1409                         dmu_tx_hold_spill(tx, object);
1410                 }
1411                 DB_DNODE_EXIT(db);
1412         }
1413 }