sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu_zfetch.c

   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 2006 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25
  26 #pragma ident   "%Z%%M% %I%     %E% SMI"
  27
  28 #include <sys/zfs_context.h>
  29 #include <sys/dnode.h>
  30 #include <sys/dmu_objset.h>
  31 #include <sys/dmu_zfetch.h>
  32 #include <sys/dmu.h>
  33 #include <sys/dbuf.h>
  34
  35 /*
  36  * I'm against tune-ables, but these should probably exist as tweakable globals
  37  * until we can get this working the way we want it to.
  38  */
  39
  40 int zfs_prefetch_disable = 0;
  41
  42 /* max # of streams per zfetch */
  43 uint32_t        zfetch_max_streams = 8;
  44 /* min time before stream reclaim */
  45 uint32_t        zfetch_min_sec_reap = 2;
  46 /* max number of blocks to fetch at a time */
  47 uint32_t        zfetch_block_cap = 256;
  48 /* number of bytes in a array_read at which we stop prefetching (1Mb) */
  49 uint64_t        zfetch_array_rd_sz = 1024 * 1024;
  50
  51 SYSCTL_DECL(_vfs_zfs);
  52 SYSCTL_INT(_vfs_zfs, OID_AUTO, prefetch_disable, CTLFLAG_RW,
  53     &zfs_prefetch_disable, 0, "Disable prefetch");
  54 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zfetch, CTLFLAG_RW, 0, "ZFS ZFETCH");
  55 TUNABLE_INT("vfs.zfs.zfetch.max_streams", &zfetch_max_streams);
  56 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_streams, CTLFLAG_RW,
  57     &zfetch_max_streams, 0, "Max # of streams per zfetch");
  58 TUNABLE_INT("vfs.zfs.zfetch.min_sec_reap", &zfetch_min_sec_reap);
  59 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, min_sec_reap, CTLFLAG_RDTUN,
  60     &zfetch_min_sec_reap, 0, "Min time before stream reclaim");
  61 TUNABLE_INT("vfs.zfs.zfetch.block_cap", &zfetch_block_cap);
  62 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, block_cap, CTLFLAG_RDTUN,
  63     &zfetch_block_cap, 0, "Max number of blocks to fetch at a time");
  64 TUNABLE_QUAD("vfs.zfs.zfetch.array_rd_sz", &zfetch_array_rd_sz);
  65 SYSCTL_QUAD(_vfs_zfs_zfetch, OID_AUTO, array_rd_sz, CTLFLAG_RDTUN,
  66     &zfetch_array_rd_sz, 0,
  67     "Number of bytes in a array_read at which we stop prefetching");
  68
  69 /* forward decls for static routines */
  70 static int              dmu_zfetch_colinear(zfetch_t *, zstream_t *);
  71 static void             dmu_zfetch_dofetch(zfetch_t *, zstream_t *);
  72 static uint64_t         dmu_zfetch_fetch(dnode_t *, uint64_t, uint64_t);
  73 static uint64_t         dmu_zfetch_fetchsz(dnode_t *, uint64_t, uint64_t);
  74 static int              dmu_zfetch_find(zfetch_t *, zstream_t *, int);
  75 static int              dmu_zfetch_stream_insert(zfetch_t *, zstream_t *);
  76 static zstream_t        *dmu_zfetch_stream_reclaim(zfetch_t *);
  77 static void             dmu_zfetch_stream_remove(zfetch_t *, zstream_t *);
  78 static int              dmu_zfetch_streams_equal(zstream_t *, zstream_t *);
  79
  80 /*
  81  * Given a zfetch structure and a zstream structure, determine whether the
  82  * blocks to be read are part of a co-linear pair of existing prefetch
  83  * streams.  If a set is found, coalesce the streams, removing one, and
  84  * configure the prefetch so it looks for a strided access pattern.
  85  *
  86  * In other words: if we find two sequential access streams that are
  87  * the same length and distance N appart, and this read is N from the
  88  * last stream, then we are probably in a strided access pattern.  So
  89  * combine the two sequential streams into a single strided stream.
  90  *
  91  * If no co-linear streams are found, return NULL.
  92  */
  93 static int
  94 dmu_zfetch_colinear(zfetch_t *zf, zstream_t *zh)
  95 {
  96         zstream_t       *z_walk;
  97         zstream_t       *z_comp;
  98
  99         if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER))
 100                 return (0);
 101
 102         if (zh == NULL) {
 103                 rw_exit(&zf->zf_rwlock);
 104                 return (0);
 105         }
 106
 107         for (z_walk = list_head(&zf->zf_stream); z_walk;
 108             z_walk = list_next(&zf->zf_stream, z_walk)) {
 109                 for (z_comp = list_next(&zf->zf_stream, z_walk); z_comp;
 110                     z_comp = list_next(&zf->zf_stream, z_comp)) {
 111                         int64_t         diff;
 112
 113                         if (z_walk->zst_len != z_walk->zst_stride ||
 114                             z_comp->zst_len != z_comp->zst_stride) {
 115                                 continue;
 116                         }
 117
 118                         diff = z_comp->zst_offset - z_walk->zst_offset;
 119                         if (z_comp->zst_offset + diff == zh->zst_offset) {
 120                                 z_walk->zst_offset = zh->zst_offset;
 121                                 z_walk->zst_direction = diff < 0 ? -1 : 1;
 122                                 z_walk->zst_stride =
 123                                     diff * z_walk->zst_direction;
 124                                 z_walk->zst_ph_offset =
 125                                     zh->zst_offset + z_walk->zst_stride;
 126                                 dmu_zfetch_stream_remove(zf, z_comp);
 127                                 mutex_destroy(&z_comp->zst_lock);
 128                                 kmem_free(z_comp, sizeof (zstream_t));
 129
 130                                 dmu_zfetch_dofetch(zf, z_walk);
 131
 132                                 rw_exit(&zf->zf_rwlock);
 133                                 return (1);
 134                         }
 135
 136                         diff = z_walk->zst_offset - z_comp->zst_offset;
 137                         if (z_walk->zst_offset + diff == zh->zst_offset) {
 138                                 z_walk->zst_offset = zh->zst_offset;
 139                                 z_walk->zst_direction = diff < 0 ? -1 : 1;
 140                                 z_walk->zst_stride =
 141                                     diff * z_walk->zst_direction;
 142                                 z_walk->zst_ph_offset =
 143                                     zh->zst_offset + z_walk->zst_stride;
 144                                 dmu_zfetch_stream_remove(zf, z_comp);
 145                                 mutex_destroy(&z_comp->zst_lock);
 146                                 kmem_free(z_comp, sizeof (zstream_t));
 147
 148                                 dmu_zfetch_dofetch(zf, z_walk);
 149
 150                                 rw_exit(&zf->zf_rwlock);
 151                                 return (1);
 152                         }
 153                 }
 154         }
 155
 156         rw_exit(&zf->zf_rwlock);
 157         return (0);
 158 }
 159
 160 /*
 161  * Given a zstream_t, determine the bounds of the prefetch.  Then call the
 162  * routine that actually prefetches the individual blocks.
 163  */
 164 static void
 165 dmu_zfetch_dofetch(zfetch_t *zf, zstream_t *zs)
 166 {
 167         uint64_t        prefetch_tail;
 168         uint64_t        prefetch_limit;
 169         uint64_t        prefetch_ofst;
 170         uint64_t        prefetch_len;
 171         uint64_t        blocks_fetched;
 172
 173         zs->zst_stride = MAX((int64_t)zs->zst_stride, zs->zst_len);
 174         zs->zst_cap = MIN(zfetch_block_cap, 2 * zs->zst_cap);
 175
 176         prefetch_tail = MAX((int64_t)zs->zst_ph_offset,
 177             (int64_t)(zs->zst_offset + zs->zst_stride));
 178         /*
 179          * XXX: use a faster division method?
 180          */
 181         prefetch_limit = zs->zst_offset + zs->zst_len +
 182             (zs->zst_cap * zs->zst_stride) / zs->zst_len;
 183
 184         while (prefetch_tail < prefetch_limit) {
 185                 prefetch_ofst = zs->zst_offset + zs->zst_direction *
 186                     (prefetch_tail - zs->zst_offset);
 187
 188                 prefetch_len = zs->zst_len;
 189
 190                 /*
 191                  * Don't prefetch beyond the end of the file, if working
 192                  * backwards.
 193                  */
 194                 if ((zs->zst_direction == ZFETCH_BACKWARD) &&
 195                     (prefetch_ofst > prefetch_tail)) {
 196                         prefetch_len += prefetch_ofst;
 197                         prefetch_ofst = 0;
 198                 }
 199
 200                 /* don't prefetch more than we're supposed to */
 201                 if (prefetch_len > zs->zst_len)
 202                         break;
 203
 204                 blocks_fetched = dmu_zfetch_fetch(zf->zf_dnode,
 205                     prefetch_ofst, zs->zst_len);
 206
 207                 prefetch_tail += zs->zst_stride;
 208                 /* stop if we've run out of stuff to prefetch */
 209                 if (blocks_fetched < zs->zst_len)
 210                         break;
 211         }
 212         zs->zst_ph_offset = prefetch_tail;
 213         zs->zst_last = LBOLT;
 214 }
 215
 216 /*
 217  * This takes a pointer to a zfetch structure and a dnode.  It performs the
 218  * necessary setup for the zfetch structure, grokking data from the
 219  * associated dnode.
 220  */
 221 void
 222 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
 223 {
 224         if (zf == NULL) {
 225                 return;
 226         }
 227
 228         zf->zf_dnode = dno;
 229         zf->zf_stream_cnt = 0;
 230         zf->zf_alloc_fail = 0;
 231
 232         list_create(&zf->zf_stream, sizeof (zstream_t),
 233             offsetof(zstream_t, zst_node));
 234
 235         rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
 236 }
 237
 238 /*
 239  * This function computes the actual size, in blocks, that can be prefetched,
 240  * and fetches it.
 241  */
 242 static uint64_t
 243 dmu_zfetch_fetch(dnode_t *dn, uint64_t blkid, uint64_t nblks)
 244 {
 245         uint64_t        fetchsz;
 246         uint64_t        i;
 247
 248         fetchsz = dmu_zfetch_fetchsz(dn, blkid, nblks);
 249
 250         for (i = 0; i < fetchsz; i++) {
 251                 dbuf_prefetch(dn, blkid + i);
 252         }
 253
 254         return (fetchsz);
 255 }
 256
 257 /*
 258  * this function returns the number of blocks that would be prefetched, based
 259  * upon the supplied dnode, blockid, and nblks.  This is used so that we can
 260  * update streams in place, and then prefetch with their old value after the
 261  * fact.  This way, we can delay the prefetch, but subsequent accesses to the
 262  * stream won't result in the same data being prefetched multiple times.
 263  */
 264 static uint64_t
 265 dmu_zfetch_fetchsz(dnode_t *dn, uint64_t blkid, uint64_t nblks)
 266 {
 267         uint64_t        fetchsz;
 268
 269         if (blkid > dn->dn_maxblkid) {
 270                 return (0);
 271         }
 272
 273         /* compute fetch size */
 274         if (blkid + nblks + 1 > dn->dn_maxblkid) {
 275                 fetchsz = (dn->dn_maxblkid - blkid) + 1;
 276                 ASSERT(blkid + fetchsz - 1 <= dn->dn_maxblkid);
 277         } else {
 278                 fetchsz = nblks;
 279         }
 280
 281
 282         return (fetchsz);
 283 }
 284
 285 /*
 286  * given a zfetch and a zsearch structure, see if there is an associated zstream
 287  * for this block read.  If so, it starts a prefetch for the stream it
 288  * located and returns true, otherwise it returns false
 289  */
 290 static int
 291 dmu_zfetch_find(zfetch_t *zf, zstream_t *zh, int prefetched)
 292 {
 293         zstream_t       *zs;
 294         int64_t         diff;
 295         int             reset = !prefetched;
 296         int             rc = 0;
 297
 298         if (zh == NULL)
 299                 return (0);
 300
 301         /*
 302          * XXX: This locking strategy is a bit coarse; however, it's impact has
 303          * yet to be tested.  If this turns out to be an issue, it can be
 304          * modified in a number of different ways.
 305          */
 306
 307         rw_enter(&zf->zf_rwlock, RW_READER);
 308 top:
 309
 310         for (zs = list_head(&zf->zf_stream); zs;
 311             zs = list_next(&zf->zf_stream, zs)) {
 312
 313                 /*
 314                  * XXX - should this be an assert?
 315                  */
 316                 if (zs->zst_len == 0) {
 317                         /* bogus stream */
 318                         continue;
 319                 }
 320
 321                 /*
 322                  * We hit this case when we are in a strided prefetch stream:
 323                  * we will read "len" blocks before "striding".
 324                  */
 325                 if (zh->zst_offset >= zs->zst_offset &&
 326                     zh->zst_offset < zs->zst_offset + zs->zst_len) {
 327                         /* already fetched */
 328                         rc = 1;
 329                         goto out;
 330                 }
 331
 332                 /*
 333                  * This is the forward sequential read case: we increment
 334                  * len by one each time we hit here, so we will enter this
 335                  * case on every read.
 336                  */
 337                 if (zh->zst_offset == zs->zst_offset + zs->zst_len) {
 338
 339                         reset = !prefetched && zs->zst_len > 1;
 340
 341                         if (mutex_tryenter(&zs->zst_lock) == 0) {
 342                                 rc = 1;
 343                                 goto out;
 344                         }
 345                         if (zh->zst_offset != zs->zst_offset + zs->zst_len) {
 346                                 mutex_exit(&zs->zst_lock);
 347                                 goto top;
 348                         }
 349                         zs->zst_len += zh->zst_len;
 350                         diff = zs->zst_len - zfetch_block_cap;
 351                         if (diff > 0) {
 352                                 zs->zst_offset += diff;
 353                                 zs->zst_len = zs->zst_len > diff ?
 354                                     zs->zst_len - diff : 0;
 355                         }
 356                         zs->zst_direction = ZFETCH_FORWARD;
 357
 358                         break;
 359
 360                 /*
 361                  * Same as above, but reading backwards through the file.
 362                  */
 363                 } else if (zh->zst_offset == zs->zst_offset - zh->zst_len) {
 364                         /* backwards sequential access */
 365
 366                         reset = !prefetched && zs->zst_len > 1;
 367
 368                         if (mutex_tryenter(&zs->zst_lock) == 0) {
 369                                 rc = 1;
 370                                 goto out;
 371                         }
 372                         if (zh->zst_offset != zs->zst_offset - zh->zst_len) {
 373                                 mutex_exit(&zs->zst_lock);
 374                                 goto top;
 375                         }
 376
 377                         zs->zst_offset = zs->zst_offset > zh->zst_len ?
 378                             zs->zst_offset - zh->zst_len : 0;
 379                         zs->zst_ph_offset = zs->zst_ph_offset > zh->zst_len ?
 380                             zs->zst_ph_offset - zh->zst_len : 0;
 381                         zs->zst_len += zh->zst_len;
 382
 383                         diff = zs->zst_len - zfetch_block_cap;
 384                         if (diff > 0) {
 385                                 zs->zst_ph_offset = zs->zst_ph_offset > diff ?
 386                                     zs->zst_ph_offset - diff : 0;
 387                                 zs->zst_len = zs->zst_len > diff ?
 388                                     zs->zst_len - diff : zs->zst_len;
 389                         }
 390                         zs->zst_direction = ZFETCH_BACKWARD;
 391
 392                         break;
 393
 394                 } else if ((zh->zst_offset - zs->zst_offset - zs->zst_stride <
 395                     zs->zst_len) && (zs->zst_len != zs->zst_stride)) {
 396                         /* strided forward access */
 397
 398                         if (mutex_tryenter(&zs->zst_lock) == 0) {
 399                                 rc = 1;
 400                                 goto out;
 401                         }
 402                         if ((zh->zst_offset - zs->zst_offset - zs->zst_stride >=
 403                             zs->zst_len) || (zs->zst_len == zs->zst_stride)) {
 404                                 mutex_exit(&zs->zst_lock);
 405                                 goto top;
 406                         }
 407
 408                         zs->zst_offset += zs->zst_stride;
 409                         zs->zst_direction = ZFETCH_FORWARD;
 410
 411                         break;
 412
 413                 } else if ((zh->zst_offset - zs->zst_offset + zs->zst_stride <
 414                     zs->zst_len) && (zs->zst_len != zs->zst_stride)) {
 415                         /* strided reverse access */
 416
 417                         if (mutex_tryenter(&zs->zst_lock) == 0) {
 418                                 rc = 1;
 419                                 goto out;
 420                         }
 421                         if ((zh->zst_offset - zs->zst_offset + zs->zst_stride >=
 422                             zs->zst_len) || (zs->zst_len == zs->zst_stride)) {
 423                                 mutex_exit(&zs->zst_lock);
 424                                 goto top;
 425                         }
 426
 427                         zs->zst_offset = zs->zst_offset > zs->zst_stride ?
 428                             zs->zst_offset - zs->zst_stride : 0;
 429                         zs->zst_ph_offset = (zs->zst_ph_offset >
 430                             (2 * zs->zst_stride)) ?
 431                             (zs->zst_ph_offset - (2 * zs->zst_stride)) : 0;
 432                         zs->zst_direction = ZFETCH_BACKWARD;
 433
 434                         break;
 435                 }
 436         }
 437
 438         if (zs) {
 439                 if (reset) {
 440                         zstream_t *remove = zs;
 441
 442                         rc = 0;
 443                         mutex_exit(&zs->zst_lock);
 444                         rw_exit(&zf->zf_rwlock);
 445                         rw_enter(&zf->zf_rwlock, RW_WRITER);
 446                         /*
 447                          * Relocate the stream, in case someone removes
 448                          * it while we were acquiring the WRITER lock.
 449                          */
 450                         for (zs = list_head(&zf->zf_stream); zs;
 451                             zs = list_next(&zf->zf_stream, zs)) {
 452                                 if (zs == remove) {
 453                                         dmu_zfetch_stream_remove(zf, zs);
 454                                         mutex_destroy(&zs->zst_lock);
 455                                         kmem_free(zs, sizeof (zstream_t));
 456                                         break;
 457                                 }
 458                         }
 459                 } else {
 460                         rc = 1;
 461                         dmu_zfetch_dofetch(zf, zs);
 462                         mutex_exit(&zs->zst_lock);
 463                 }
 464         }
 465 out:
 466         rw_exit(&zf->zf_rwlock);
 467         return (rc);
 468 }
 469
 470 /*
 471  * Clean-up state associated with a zfetch structure.  This frees allocated
 472  * structure members, empties the zf_stream tree, and generally makes things
 473  * nice.  This doesn't free the zfetch_t itself, that's left to the caller.
 474  */
 475 void
 476 dmu_zfetch_rele(zfetch_t *zf)
 477 {
 478         zstream_t       *zs;
 479         zstream_t       *zs_next;
 480
 481         ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
 482
 483         for (zs = list_head(&zf->zf_stream); zs; zs = zs_next) {
 484                 zs_next = list_next(&zf->zf_stream, zs);
 485
 486                 list_remove(&zf->zf_stream, zs);
 487                 mutex_destroy(&zs->zst_lock);
 488                 kmem_free(zs, sizeof (zstream_t));
 489         }
 490         list_destroy(&zf->zf_stream);
 491         rw_destroy(&zf->zf_rwlock);
 492
 493         zf->zf_dnode = NULL;
 494 }
 495
 496 /*
 497  * Given a zfetch and zstream structure, insert the zstream structure into the
 498  * AVL tree contained within the zfetch structure.  Peform the appropriate
 499  * book-keeping.  It is possible that another thread has inserted a stream which
 500  * matches one that we are about to insert, so we must be sure to check for this
 501  * case.  If one is found, return failure, and let the caller cleanup the
 502  * duplicates.
 503  */
 504 static int
 505 dmu_zfetch_stream_insert(zfetch_t *zf, zstream_t *zs)
 506 {
 507         zstream_t       *zs_walk;
 508         zstream_t       *zs_next;
 509
 510         ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
 511
 512         for (zs_walk = list_head(&zf->zf_stream); zs_walk; zs_walk = zs_next) {
 513                 zs_next = list_next(&zf->zf_stream, zs_walk);
 514
 515                 if (dmu_zfetch_streams_equal(zs_walk, zs)) {
 516                     return (0);
 517                 }
 518         }
 519
 520         list_insert_head(&zf->zf_stream, zs);
 521         zf->zf_stream_cnt++;
 522
 523         return (1);
 524 }
 525
 526
 527 /*
 528  * Walk the list of zstreams in the given zfetch, find an old one (by time), and
 529  * reclaim it for use by the caller.
 530  */
 531 static zstream_t *
 532 dmu_zfetch_stream_reclaim(zfetch_t *zf)
 533 {
 534         zstream_t       *zs;
 535
 536         if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER))
 537                 return (0);
 538
 539         for (zs = list_head(&zf->zf_stream); zs;
 540             zs = list_next(&zf->zf_stream, zs)) {
 541
 542                 if (((LBOLT - zs->zst_last) / hz) > zfetch_min_sec_reap)
 543                         break;
 544         }
 545
 546         if (zs) {
 547                 dmu_zfetch_stream_remove(zf, zs);
 548                 mutex_destroy(&zs->zst_lock);
 549                 bzero(zs, sizeof (zstream_t));
 550         } else {
 551                 zf->zf_alloc_fail++;
 552         }
 553         rw_exit(&zf->zf_rwlock);
 554
 555         return (zs);
 556 }
 557
 558 /*
 559  * Given a zfetch and zstream structure, remove the zstream structure from its
 560  * container in the zfetch structure.  Perform the appropriate book-keeping.
 561  */
 562 static void
 563 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
 564 {
 565         ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
 566
 567         list_remove(&zf->zf_stream, zs);
 568         zf->zf_stream_cnt--;
 569 }
 570
 571 static int
 572 dmu_zfetch_streams_equal(zstream_t *zs1, zstream_t *zs2)
 573 {
 574         if (zs1->zst_offset != zs2->zst_offset)
 575                 return (0);
 576
 577         if (zs1->zst_len != zs2->zst_len)
 578                 return (0);
 579
 580         if (zs1->zst_stride != zs2->zst_stride)
 581                 return (0);
 582
 583         if (zs1->zst_ph_offset != zs2->zst_ph_offset)
 584                 return (0);
 585
 586         if (zs1->zst_cap != zs2->zst_cap)
 587                 return (0);
 588
 589         if (zs1->zst_direction != zs2->zst_direction)
 590                 return (0);
 591
 592         return (1);
 593 }
 594
 595 /*
 596  * This is the prefetch entry point.  It calls all of the other dmu_zfetch
 597  * routines to create, delete, find, or operate upon prefetch streams.
 598  */
 599 void
 600 dmu_zfetch(zfetch_t *zf, uint64_t offset, uint64_t size, int prefetched)
 601 {
 602         zstream_t       zst;
 603         zstream_t       *newstream;
 604         int             fetched;
 605         int             inserted;
 606         unsigned int    blkshft;
 607         uint64_t        blksz;
 608
 609         if (zfs_prefetch_disable)
 610                 return;
 611
 612         /* files that aren't ln2 blocksz are only one block -- nothing to do */
 613         if (!zf->zf_dnode->dn_datablkshift)
 614                 return;
 615
 616         /* convert offset and size, into blockid and nblocks */
 617         blkshft = zf->zf_dnode->dn_datablkshift;
 618         blksz = (1 << blkshft);
 619
 620         bzero(&zst, sizeof (zstream_t));
 621         zst.zst_offset = offset >> blkshft;
 622         zst.zst_len = (P2ROUNDUP(offset + size, blksz) -
 623             P2ALIGN(offset, blksz)) >> blkshft;
 624
 625         fetched = dmu_zfetch_find(zf, &zst, prefetched);
 626         if (!fetched) {
 627                 fetched = dmu_zfetch_colinear(zf, &zst);
 628         }
 629
 630         if (!fetched) {
 631                 newstream = dmu_zfetch_stream_reclaim(zf);
 632
 633                 /*
 634                  * we still couldn't find a stream, drop the lock, and allocate
 635                  * one if possible.  Otherwise, give up and go home.
 636                  */
 637                 if (newstream == NULL) {
 638                         uint64_t        maxblocks;
 639                         uint32_t        max_streams;
 640                         uint32_t        cur_streams;
 641
 642                         cur_streams = zf->zf_stream_cnt;
 643                         maxblocks = zf->zf_dnode->dn_maxblkid;
 644
 645                         max_streams = MIN(zfetch_max_streams,
 646                             (maxblocks / zfetch_block_cap));
 647                         if (max_streams == 0) {
 648                                 max_streams++;
 649                         }
 650
 651                         if (cur_streams >= max_streams) {
 652                                 return;
 653                         }
 654
 655                         newstream = kmem_zalloc(sizeof (zstream_t), KM_SLEEP);
 656                 }
 657
 658                 newstream->zst_offset = zst.zst_offset;
 659                 newstream->zst_len = zst.zst_len;
 660                 newstream->zst_stride = zst.zst_len;
 661                 newstream->zst_ph_offset = zst.zst_len + zst.zst_offset;
 662                 newstream->zst_cap = zst.zst_len;
 663                 newstream->zst_direction = ZFETCH_FORWARD;
 664                 newstream->zst_last = LBOLT;
 665
 666                 mutex_init(&newstream->zst_lock, NULL, MUTEX_DEFAULT, NULL);
 667
 668                 rw_enter(&zf->zf_rwlock, RW_WRITER);
 669                 inserted = dmu_zfetch_stream_insert(zf, newstream);
 670                 rw_exit(&zf->zf_rwlock);
 671
 672                 if (!inserted) {
 673                         mutex_destroy(&newstream->zst_lock);
 674                         kmem_free(newstream, sizeof (zstream_t));
 675                 }
 676         }
 677 }