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         uint64_t start, end, i;
 591         int err, shift;
 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 
 602         /* first block */
 603         if (off != 0)
 604                 dmu_tx_count_write(txh, off, 1);
 605         /* last block */
 606         if (len != DMU_OBJECT_END)
 607                 dmu_tx_count_write(txh, off+len, 1);
 608 
 609         dmu_tx_count_dnode(txh);
 610 
 611         if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
 612                 return;
 613         if (len == DMU_OBJECT_END)
 614                 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
 615 
 616         /*
 617          * For i/o error checking, read the first and last level-0
 618          * blocks, and all the level-1 blocks.  The above count_write's
 619          * have already taken care of the level-0 blocks.
 620          */
 621         if (dn->dn_nlevels > 1) {
 622                 shift = dn->dn_datablkshift + dn->dn_indblkshift -
 623                     SPA_BLKPTRSHIFT;
 624                 start = off >> shift;
 625                 end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
 626 
 627                 zio = zio_root(tx->tx_pool->dp_spa,
 628                     NULL, NULL, ZIO_FLAG_CANFAIL);
 629                 for (i = start; i <= end; i++) {
 630                         uint64_t ibyte = i << shift;
 631                         err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
 632                         i = ibyte >> shift;
 633                         if (err == ESRCH)
 634                                 break;
 635                         if (err) {
 636                                 tx->tx_err = err;
 637                                 return;
 638                         }
 639 
 640                         err = dmu_tx_check_ioerr(zio, dn, 1, i);
 641                         if (err) {
 642                                 tx->tx_err = err;
 643                                 return;
 644                         }
 645                 }
 646                 err = zio_wait(zio);
 647                 if (err) {
 648                         tx->tx_err = err;
 649                         return;
 650                 }
 651         }
 652 
 653         dmu_tx_count_free(txh, off, len);
 654 }
 655 
 656 void
 657 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
 658 {
 659         dmu_tx_hold_t *txh;
 660         dnode_t *dn;
 661         uint64_t nblocks;
 662         int epbs, err;
 663 
 664         ASSERT(tx->tx_txg == 0);
 665 
 666         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
 667             object, THT_ZAP, add, (uintptr_t)name);
 668         if (txh == NULL)
 669                 return;
 670         dn = txh->txh_dnode;
 671 
 672         dmu_tx_count_dnode(txh);
 673 
 674         if (dn == NULL) {
 675                 /*
 676                  * We will be able to fit a new object's entries into one leaf
 677                  * block.  So there will be at most 2 blocks total,
 678                  * including the header block.
 679                  */
 680                 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
 681                 return;
 682         }
 683 
 684         ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
 685 
 686         if (dn->dn_maxblkid == 0 && !add) {
 687                 blkptr_t *bp;
 688 
 689                 /*
 690                  * If there is only one block  (i.e. this is a micro-zap)
 691                  * and we are not adding anything, the accounting is simple.
 692                  */
 693                 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
 694                 if (err) {
 695                         tx->tx_err = err;
 696                         return;
 697                 }
 698 
 699                 /*
 700                  * Use max block size here, since we don't know how much
 701                  * the size will change between now and the dbuf dirty call.
 702                  */
 703                 bp = &dn->dn_phys->dn_blkptr[0];
 704                 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
 705                     bp, bp->blk_birth))
 706                         txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
 707                 else
 708                         txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
 709                 if (!BP_IS_HOLE(bp))
 710                         txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
 711                 return;
 712         }
 713 
 714         if (dn->dn_maxblkid > 0 && name) {
 715                 /*
 716                  * access the name in this fat-zap so that we'll check
 717                  * for i/o errors to the leaf blocks, etc.
 718                  */
 719                 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
 720                     8, 0, NULL);
 721                 if (err == EIO) {
 722                         tx->tx_err = err;
 723                         return;
 724                 }
 725         }
 726 
 727         err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
 728             &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
 729 
 730         /*
 731          * If the modified blocks are scattered to the four winds,
 732          * we'll have to modify an indirect twig for each.
 733          */
 734         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
 735         for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
 736                 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
 737                         txh->txh_space_towrite += 3 << dn->dn_indblkshift;
 738                 else
 739                         txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
 740 }
 741 
 742 void
 743 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
 744 {
 745         dmu_tx_hold_t *txh;
 746 
 747         ASSERT(tx->tx_txg == 0);
 748 
 749         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
 750             object, THT_BONUS, 0, 0);
 751         if (txh)
 752                 dmu_tx_count_dnode(txh);
 753 }
 754 
 755 void
 756 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
 757 {
 758         dmu_tx_hold_t *txh;
 759         ASSERT(tx->tx_txg == 0);
 760 
 761         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
 762             DMU_NEW_OBJECT, THT_SPACE, space, 0);
 763 
 764         txh->txh_space_towrite += space;
 765 }
 766 
 767 int
 768 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
 769 {
 770         dmu_tx_hold_t *txh;
 771         int holds = 0;
 772 
 773         /*
 774          * By asserting that the tx is assigned, we're counting the
 775          * number of dn_tx_holds, which is the same as the number of
 776          * dn_holds.  Otherwise, we'd be counting dn_holds, but
 777          * dn_tx_holds could be 0.
 778          */
 779         ASSERT(tx->tx_txg != 0);
 780 
 781         /* if (tx->tx_anyobj == TRUE) */
 782                 /* return (0); */
 783 
 784         for (txh = list_head(&tx->tx_holds); txh;
 785             txh = list_next(&tx->tx_holds, txh)) {
 786                 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
 787                         holds++;
 788         }
 789 
 790         return (holds);
 791 }
 792 
 793 #ifdef ZFS_DEBUG
 794 void
 795 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
 796 {
 797         dmu_tx_hold_t *txh;
 798         int match_object = FALSE, match_offset = FALSE;
 799         dnode_t *dn;
 800 
 801         DB_DNODE_ENTER(db);
 802         dn = DB_DNODE(db);
 803         ASSERT(tx->tx_txg != 0);
 804         ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
 805         ASSERT3U(dn->dn_object, ==, db->db.db_object);
 806 
 807         if (tx->tx_anyobj) {
 808                 DB_DNODE_EXIT(db);
 809                 return;
 810         }
 811 
 812         /* XXX No checking on the meta dnode for now */
 813         if (db->db.db_object == DMU_META_DNODE_OBJECT) {
 814                 DB_DNODE_EXIT(db);
 815                 return;
 816         }
 817 
 818         for (txh = list_head(&tx->tx_holds); txh;
 819             txh = list_next(&tx->tx_holds, txh)) {
 820                 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
 821                 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
 822                         match_object = TRUE;
 823                 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
 824                         int datablkshift = dn->dn_datablkshift ?
 825                             dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
 826                         int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
 827                         int shift = datablkshift + epbs * db->db_level;
 828                         uint64_t beginblk = shift >= 64 ? 0 :
 829                             (txh->txh_arg1 >> shift);
 830                         uint64_t endblk = shift >= 64 ? 0 :
 831                             ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
 832                         uint64_t blkid = db->db_blkid;
 833 
 834                         /* XXX txh_arg2 better not be zero... */
 835 
 836                         dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
 837                             txh->txh_type, beginblk, endblk);
 838 
 839                         switch (txh->txh_type) {
 840                         case THT_WRITE:
 841                                 if (blkid >= beginblk && blkid <= endblk)
 842                                         match_offset = TRUE;
 843                                 /*
 844                                  * We will let this hold work for the bonus
 845                                  * or spill buffer so that we don't need to
 846                                  * hold it when creating a new object.
 847                                  */
 848                                 if (blkid == DMU_BONUS_BLKID ||
 849                                     blkid == DMU_SPILL_BLKID)
 850                                         match_offset = TRUE;
 851                                 /*
 852                                  * They might have to increase nlevels,
 853                                  * thus dirtying the new TLIBs.  Or the
 854                                  * might have to change the block size,
 855                                  * thus dirying the new lvl=0 blk=0.
 856                                  */
 857                                 if (blkid == 0)
 858                                         match_offset = TRUE;
 859                                 break;
 860                         case THT_FREE:
 861                                 /*
 862                                  * We will dirty all the level 1 blocks in
 863                                  * the free range and perhaps the first and
 864                                  * last level 0 block.
 865                                  */
 866                                 if (blkid >= beginblk && (blkid <= endblk ||
 867                                     txh->txh_arg2 == DMU_OBJECT_END))
 868                                         match_offset = TRUE;
 869                                 break;
 870                         case THT_SPILL:
 871                                 if (blkid == DMU_SPILL_BLKID)
 872                                         match_offset = TRUE;
 873                                 break;
 874                         case THT_BONUS:
 875                                 if (blkid == DMU_BONUS_BLKID)
 876                                         match_offset = TRUE;
 877                                 break;
 878                         case THT_ZAP:
 879                                 match_offset = TRUE;
 880                                 break;
 881                         case THT_NEWOBJECT:
 882                                 match_object = TRUE;
 883                                 break;
 884                         default:
 885                                 ASSERT(!"bad txh_type");
 886                         }
 887                 }
 888                 if (match_object && match_offset) {
 889                         DB_DNODE_EXIT(db);
 890                         return;
 891                 }
 892         }
 893         DB_DNODE_EXIT(db);
 894         panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
 895             (u_longlong_t)db->db.db_object, db->db_level,
 896             (u_longlong_t)db->db_blkid);
 897 }
 898 #endif
 899 
 900 static int
 901 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
 902 {
 903         dmu_tx_hold_t *txh;
 904         spa_t *spa = tx->tx_pool->dp_spa;
 905         uint64_t memory, asize, fsize, usize;
 906         uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
 907 
 908         ASSERT0(tx->tx_txg);
 909 
 910         if (tx->tx_err)
 911                 return (tx->tx_err);
 912 
 913         if (spa_suspended(spa)) {
 914                 /*
 915                  * If the user has indicated a blocking failure mode
 916                  * then return ERESTART which will block in dmu_tx_wait().
 917                  * Otherwise, return EIO so that an error can get
 918                  * propagated back to the VOP calls.
 919                  *
 920                  * Note that we always honor the txg_how flag regardless
 921                  * of the failuremode setting.
 922                  */
 923                 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
 924                     txg_how != TXG_WAIT)
 925                         return (SET_ERROR(EIO));
 926 
 927                 return (SET_ERROR(ERESTART));
 928         }
 929 
 930         tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
 931         tx->tx_needassign_txh = NULL;
 932 
 933         /*
 934          * NB: No error returns are allowed after txg_hold_open, but
 935          * before processing the dnode holds, due to the
 936          * dmu_tx_unassign() logic.
 937          */
 938 
 939         towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
 940         for (txh = list_head(&tx->tx_holds); txh;
 941             txh = list_next(&tx->tx_holds, txh)) {
 942                 dnode_t *dn = txh->txh_dnode;
 943                 if (dn != NULL) {
 944                         mutex_enter(&dn->dn_mtx);
 945                         if (dn->dn_assigned_txg == tx->tx_txg - 1) {
 946                                 mutex_exit(&dn->dn_mtx);
 947                                 tx->tx_needassign_txh = txh;
 948                                 return (SET_ERROR(ERESTART));
 949                         }
 950                         if (dn->dn_assigned_txg == 0)
 951                                 dn->dn_assigned_txg = tx->tx_txg;
 952                         ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
 953                         (void) refcount_add(&dn->dn_tx_holds, tx);
 954                         mutex_exit(&dn->dn_mtx);
 955                 }
 956                 towrite += txh->txh_space_towrite;
 957                 tofree += txh->txh_space_tofree;
 958                 tooverwrite += txh->txh_space_tooverwrite;
 959                 tounref += txh->txh_space_tounref;
 960                 tohold += txh->txh_memory_tohold;
 961                 fudge += txh->txh_fudge;
 962         }
 963 
 964         /*
 965          * If a snapshot has been taken since we made our estimates,
 966          * assume that we won't be able to free or overwrite anything.
 967          */
 968         if (tx->tx_objset &&
 969             dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
 970             tx->tx_lastsnap_txg) {
 971                 towrite += tooverwrite;
 972                 tooverwrite = tofree = 0;
 973         }
 974 
 975         /* needed allocation: worst-case estimate of write space */
 976         asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
 977         /* freed space estimate: worst-case overwrite + free estimate */
 978         fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
 979         /* convert unrefd space to worst-case estimate */
 980         usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
 981         /* calculate memory footprint estimate */
 982         memory = towrite + tooverwrite + tohold;
 983 
 984 #ifdef ZFS_DEBUG
 985         /*
 986          * Add in 'tohold' to account for our dirty holds on this memory
 987          * XXX - the "fudge" factor is to account for skipped blocks that
 988          * we missed because dnode_next_offset() misses in-core-only blocks.
 989          */
 990         tx->tx_space_towrite = asize +
 991             spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
 992         tx->tx_space_tofree = tofree;
 993         tx->tx_space_tooverwrite = tooverwrite;
 994         tx->tx_space_tounref = tounref;
 995 #endif
 996 
 997         if (tx->tx_dir && asize != 0) {
 998                 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
 999                     asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1000                 if (err)
1001                         return (err);
1002         }
1003 
1004         return (0);
1005 }
1006 
1007 static void
1008 dmu_tx_unassign(dmu_tx_t *tx)
1009 {
1010         dmu_tx_hold_t *txh;
1011 
1012         if (tx->tx_txg == 0)
1013                 return;
1014 
1015         txg_rele_to_quiesce(&tx->tx_txgh);
1016 
1017         for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1018             txh = list_next(&tx->tx_holds, txh)) {
1019                 dnode_t *dn = txh->txh_dnode;
1020 
1021                 if (dn == NULL)
1022                         continue;
1023                 mutex_enter(&dn->dn_mtx);
1024                 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1025 
1026                 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1027                         dn->dn_assigned_txg = 0;
1028                         cv_broadcast(&dn->dn_notxholds);
1029                 }
1030                 mutex_exit(&dn->dn_mtx);
1031         }
1032 
1033         txg_rele_to_sync(&tx->tx_txgh);
1034 
1035         tx->tx_lasttried_txg = tx->tx_txg;
1036         tx->tx_txg = 0;
1037 }
1038 
1039 /*
1040  * Assign tx to a transaction group.  txg_how can be one of:
1041  *
1042  * (1)  TXG_WAIT.  If the current open txg is full, waits until there's
1043  *      a new one.  This should be used when you're not holding locks.
1044  *      It will only fail if we're truly out of space (or over quota).
1045  *
1046  * (2)  TXG_NOWAIT.  If we can't assign into the current open txg without
1047  *      blocking, returns immediately with ERESTART.  This should be used
1048  *      whenever you're holding locks.  On an ERESTART error, the caller
1049  *      should drop locks, do a dmu_tx_wait(tx), and try again.
1050  */
1051 int
1052 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1053 {
1054         int err;
1055 
1056         ASSERT(tx->tx_txg == 0);
1057         ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT);
1058         ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1059 
1060         /* If we might wait, we must not hold the config lock. */
1061         ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1062 
1063         while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1064                 dmu_tx_unassign(tx);
1065 
1066                 if (err != ERESTART || txg_how != TXG_WAIT)
1067                         return (err);
1068 
1069                 dmu_tx_wait(tx);
1070         }
1071 
1072         txg_rele_to_quiesce(&tx->tx_txgh);
1073 
1074         return (0);
1075 }
1076 
1077 void
1078 dmu_tx_wait(dmu_tx_t *tx)
1079 {
1080         spa_t *spa = tx->tx_pool->dp_spa;
1081 
1082         ASSERT(tx->tx_txg == 0);
1083         ASSERT(!dsl_pool_config_held(tx->tx_pool));
1084 
1085         /*
1086          * It's possible that the pool has become active after this thread
1087          * has tried to obtain a tx. If that's the case then his
1088          * tx_lasttried_txg would not have been assigned.
1089          */
1090         if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1091                 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1092         } else if (tx->tx_needassign_txh) {
1093                 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1094 
1095                 mutex_enter(&dn->dn_mtx);
1096                 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1097                         cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1098                 mutex_exit(&dn->dn_mtx);
1099                 tx->tx_needassign_txh = NULL;
1100         } else {
1101                 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1102         }
1103 }
1104 
1105 void
1106 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1107 {
1108 #ifdef ZFS_DEBUG
1109         if (tx->tx_dir == NULL || delta == 0)
1110                 return;
1111 
1112         if (delta > 0) {
1113                 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1114                     tx->tx_space_towrite);
1115                 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1116         } else {
1117                 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1118         }
1119 #endif
1120 }
1121 
1122 void
1123 dmu_tx_commit(dmu_tx_t *tx)
1124 {
1125         dmu_tx_hold_t *txh;
1126 
1127         ASSERT(tx->tx_txg != 0);
1128 
1129         while (txh = list_head(&tx->tx_holds)) {
1130                 dnode_t *dn = txh->txh_dnode;
1131 
1132                 list_remove(&tx->tx_holds, txh);
1133                 kmem_free(txh, sizeof (dmu_tx_hold_t));
1134                 if (dn == NULL)
1135                         continue;
1136                 mutex_enter(&dn->dn_mtx);
1137                 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1138 
1139                 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1140                         dn->dn_assigned_txg = 0;
1141                         cv_broadcast(&dn->dn_notxholds);
1142                 }
1143                 mutex_exit(&dn->dn_mtx);
1144                 dnode_rele(dn, tx);
1145         }
1146 
1147         if (tx->tx_tempreserve_cookie)
1148                 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1149 
1150         if (!list_is_empty(&tx->tx_callbacks))
1151                 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1152 
1153         if (tx->tx_anyobj == FALSE)
1154                 txg_rele_to_sync(&tx->tx_txgh);
1155 
1156         list_destroy(&tx->tx_callbacks);
1157         list_destroy(&tx->tx_holds);
1158 #ifdef ZFS_DEBUG
1159         dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1160             tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1161             tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1162         refcount_destroy_many(&tx->tx_space_written,
1163             refcount_count(&tx->tx_space_written));
1164         refcount_destroy_many(&tx->tx_space_freed,
1165             refcount_count(&tx->tx_space_freed));
1166 #endif
1167         kmem_free(tx, sizeof (dmu_tx_t));
1168 }
1169 
1170 void
1171 dmu_tx_abort(dmu_tx_t *tx)
1172 {
1173         dmu_tx_hold_t *txh;
1174 
1175         ASSERT(tx->tx_txg == 0);
1176 
1177         while (txh = list_head(&tx->tx_holds)) {
1178                 dnode_t *dn = txh->txh_dnode;
1179 
1180                 list_remove(&tx->tx_holds, txh);
1181                 kmem_free(txh, sizeof (dmu_tx_hold_t));
1182                 if (dn != NULL)
1183                         dnode_rele(dn, tx);
1184         }
1185 
1186         /*
1187          * Call any registered callbacks with an error code.
1188          */
1189         if (!list_is_empty(&tx->tx_callbacks))
1190                 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1191 
1192         list_destroy(&tx->tx_callbacks);
1193         list_destroy(&tx->tx_holds);
1194 #ifdef ZFS_DEBUG
1195         refcount_destroy_many(&tx->tx_space_written,
1196             refcount_count(&tx->tx_space_written));
1197         refcount_destroy_many(&tx->tx_space_freed,
1198             refcount_count(&tx->tx_space_freed));
1199 #endif
1200         kmem_free(tx, sizeof (dmu_tx_t));
1201 }
1202 
1203 uint64_t
1204 dmu_tx_get_txg(dmu_tx_t *tx)
1205 {
1206         ASSERT(tx->tx_txg != 0);
1207         return (tx->tx_txg);
1208 }
1209 
1210 dsl_pool_t *
1211 dmu_tx_pool(dmu_tx_t *tx)
1212 {
1213         ASSERT(tx->tx_pool != NULL);
1214         return (tx->tx_pool);
1215 }
1216 
1217 
1218 void
1219 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1220 {
1221         dmu_tx_callback_t *dcb;
1222 
1223         dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1224 
1225         dcb->dcb_func = func;
1226         dcb->dcb_data = data;
1227 
1228         list_insert_tail(&tx->tx_callbacks, dcb);
1229 }
1230 
1231 /*
1232  * Call all the commit callbacks on a list, with a given error code.
1233  */
1234 void
1235 dmu_tx_do_callbacks(list_t *cb_list, int error)
1236 {
1237         dmu_tx_callback_t *dcb;
1238 
1239         while (dcb = list_head(cb_list)) {
1240                 list_remove(cb_list, dcb);
1241                 dcb->dcb_func(dcb->dcb_data, error);
1242                 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1243         }
1244 }
1245 
1246 /*
1247  * Interface to hold a bunch of attributes.
1248  * used for creating new files.
1249  * attrsize is the total size of all attributes
1250  * to be added during object creation
1251  *
1252  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1253  */
1254 
1255 /*
1256  * hold necessary attribute name for attribute registration.
1257  * should be a very rare case where this is needed.  If it does
1258  * happen it would only happen on the first write to the file system.
1259  */
1260 static void
1261 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1262 {
1263         int i;
1264 
1265         if (!sa->sa_need_attr_registration)
1266                 return;
1267 
1268         for (i = 0; i != sa->sa_num_attrs; i++) {
1269                 if (!sa->sa_attr_table[i].sa_registered) {
1270                         if (sa->sa_reg_attr_obj)
1271                                 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1272                                     B_TRUE, sa->sa_attr_table[i].sa_name);
1273                         else
1274                                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1275                                     B_TRUE, sa->sa_attr_table[i].sa_name);
1276                 }
1277         }
1278 }
1279 
1280 
1281 void
1282 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1283 {
1284         dnode_t *dn;
1285         dmu_tx_hold_t *txh;
1286 
1287         txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1288             THT_SPILL, 0, 0);
1289 
1290         dn = txh->txh_dnode;
1291 
1292         if (dn == NULL)
1293                 return;
1294 
1295         /* If blkptr doesn't exist then add space to towrite */
1296         if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1297                 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1298         } else {
1299                 blkptr_t *bp;
1300 
1301                 bp = &dn->dn_phys->dn_spill;
1302                 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1303                     bp, bp->blk_birth))
1304                         txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1305                 else
1306                         txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1307                 if (!BP_IS_HOLE(bp))
1308                         txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1309         }
1310 }
1311 
1312 void
1313 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1314 {
1315         sa_os_t *sa = tx->tx_objset->os_sa;
1316 
1317         dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1318 
1319         if (tx->tx_objset->os_sa->sa_master_obj == 0)
1320                 return;
1321 
1322         if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1323                 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1324         else {
1325                 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1326                 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1327                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1328                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1329         }
1330 
1331         dmu_tx_sa_registration_hold(sa, tx);
1332 
1333         if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1334                 return;
1335 
1336         (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1337             THT_SPILL, 0, 0);
1338 }
1339 
1340 /*
1341  * Hold SA attribute
1342  *
1343  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1344  *
1345  * variable_size is the total size of all variable sized attributes
1346  * passed to this function.  It is not the total size of all
1347  * variable size attributes that *may* exist on this object.
1348  */
1349 void
1350 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1351 {
1352         uint64_t object;
1353         sa_os_t *sa = tx->tx_objset->os_sa;
1354 
1355         ASSERT(hdl != NULL);
1356 
1357         object = sa_handle_object(hdl);
1358 
1359         dmu_tx_hold_bonus(tx, object);
1360 
1361         if (tx->tx_objset->os_sa->sa_master_obj == 0)
1362                 return;
1363 
1364         if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1365             tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1366                 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1367                 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1368                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1369                 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1370         }
1371 
1372         dmu_tx_sa_registration_hold(sa, tx);
1373 
1374         if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1375                 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1376 
1377         if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1378                 ASSERT(tx->tx_txg == 0);
1379                 dmu_tx_hold_spill(tx, object);
1380         } else {
1381                 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1382                 dnode_t *dn;
1383 
1384                 DB_DNODE_ENTER(db);
1385                 dn = DB_DNODE(db);
1386                 if (dn->dn_have_spill) {
1387                         ASSERT(tx->tx_txg == 0);
1388                         dmu_tx_hold_spill(tx, object);
1389                 }
1390                 DB_DNODE_EXIT(db);
1391         }
1392 }