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 2009 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25
  26 #include <sys/zfs_context.h>
  27 #include <sys/dnode.h>
  28 #include <sys/dmu_objset.h>
  29 #include <sys/dmu_zfetch.h>
  30 #include <sys/dmu.h>
  31 #include <sys/dbuf.h>
  32 #include <sys/kstat.h>
  33
  34 /*
  35  * I'm against tune-ables, but these should probably exist as tweakable globals
  36  * until we can get this working the way we want it to.
  37  */
  38
  39 int zfs_prefetch_disable = 0;
  40
  41 /* max # of streams per zfetch */
  42 uint32_t        zfetch_max_streams = 8;
  43 /* min time before stream reclaim */
  44 uint32_t        zfetch_min_sec_reap = 2;
  45 /* max number of blocks to fetch at a time */
  46 uint32_t        zfetch_block_cap = 256;
  47 /* number of bytes in a array_read at which we stop prefetching (1Mb) */
  48 uint64_t        zfetch_array_rd_sz = 1024 * 1024;
  49
  50 SYSCTL_DECL(_vfs_zfs);
  51 SYSCTL_INT(_vfs_zfs, OID_AUTO, prefetch_disable, CTLFLAG_RW,
  52     &zfs_prefetch_disable, 0, "Disable prefetch");
  53 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zfetch, CTLFLAG_RW, 0, "ZFS ZFETCH");
  54 TUNABLE_INT("vfs.zfs.zfetch.max_streams", &zfetch_max_streams);
  55 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, max_streams, CTLFLAG_RW,
  56     &zfetch_max_streams, 0, "Max # of streams per zfetch");
  57 TUNABLE_INT("vfs.zfs.zfetch.min_sec_reap", &zfetch_min_sec_reap);
  58 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, min_sec_reap, CTLFLAG_RDTUN,
  59     &zfetch_min_sec_reap, 0, "Min time before stream reclaim");
  60 TUNABLE_INT("vfs.zfs.zfetch.block_cap", &zfetch_block_cap);
  61 SYSCTL_UINT(_vfs_zfs_zfetch, OID_AUTO, block_cap, CTLFLAG_RDTUN,
  62     &zfetch_block_cap, 0, "Max number of blocks to fetch at a time");
  63 TUNABLE_QUAD("vfs.zfs.zfetch.array_rd_sz", &zfetch_array_rd_sz);
  64 SYSCTL_QUAD(_vfs_zfs_zfetch, OID_AUTO, array_rd_sz, CTLFLAG_RDTUN,
  65     &zfetch_array_rd_sz, 0,
  66     "Number of bytes in a array_read at which we stop prefetching");
  67
  68 /* forward decls for static routines */
  69 static int              dmu_zfetch_colinear(zfetch_t *, zstream_t *);
  70 static void             dmu_zfetch_dofetch(zfetch_t *, zstream_t *);
  71 static uint64_t         dmu_zfetch_fetch(dnode_t *, uint64_t, uint64_t);
  72 static uint64_t         dmu_zfetch_fetchsz(dnode_t *, uint64_t, uint64_t);
  73 static int              dmu_zfetch_find(zfetch_t *, zstream_t *, int);
  74 static int              dmu_zfetch_stream_insert(zfetch_t *, zstream_t *);
  75 static zstream_t        *dmu_zfetch_stream_reclaim(zfetch_t *);
  76 static void             dmu_zfetch_stream_remove(zfetch_t *, zstream_t *);
  77 static int              dmu_zfetch_streams_equal(zstream_t *, zstream_t *);
  78
  79 typedef struct zfetch_stats {
  80         kstat_named_t zfetchstat_hits;
  81         kstat_named_t zfetchstat_misses;
  82         kstat_named_t zfetchstat_colinear_hits;
  83         kstat_named_t zfetchstat_colinear_misses;
  84         kstat_named_t zfetchstat_stride_hits;
  85         kstat_named_t zfetchstat_stride_misses;
  86         kstat_named_t zfetchstat_reclaim_successes;
  87         kstat_named_t zfetchstat_reclaim_failures;
  88         kstat_named_t zfetchstat_stream_resets;
  89         kstat_named_t zfetchstat_stream_noresets;
  90         kstat_named_t zfetchstat_bogus_streams;
  91 } zfetch_stats_t;
  92
  93 static zfetch_stats_t zfetch_stats = {
  94         { "hits",                       KSTAT_DATA_UINT64 },
  95         { "misses",                     KSTAT_DATA_UINT64 },
  96         { "colinear_hits",              KSTAT_DATA_UINT64 },
  97         { "colinear_misses",            KSTAT_DATA_UINT64 },
  98         { "stride_hits",                KSTAT_DATA_UINT64 },
  99         { "stride_misses",              KSTAT_DATA_UINT64 },
 100         { "reclaim_successes",          KSTAT_DATA_UINT64 },
 101         { "reclaim_failures",           KSTAT_DATA_UINT64 },
 102         { "streams_resets",             KSTAT_DATA_UINT64 },
 103         { "streams_noresets",           KSTAT_DATA_UINT64 },
 104         { "bogus_streams",              KSTAT_DATA_UINT64 },
 105 };
 106
 107 #define ZFETCHSTAT_INCR(stat, val) \
 108         atomic_add_64(&zfetch_stats.stat.value.ui64, (val));
 109
 110 #define ZFETCHSTAT_BUMP(stat)           ZFETCHSTAT_INCR(stat, 1);
 111
 112 kstat_t         *zfetch_ksp;
 113
 114 /*
 115  * Given a zfetch structure and a zstream structure, determine whether the
 116  * blocks to be read are part of a co-linear pair of existing prefetch
 117  * streams.  If a set is found, coalesce the streams, removing one, and
 118  * configure the prefetch so it looks for a strided access pattern.
 119  *
 120  * In other words: if we find two sequential access streams that are
 121  * the same length and distance N appart, and this read is N from the
 122  * last stream, then we are probably in a strided access pattern.  So
 123  * combine the two sequential streams into a single strided stream.
 124  *
 125  * If no co-linear streams are found, return NULL.
 126  */
 127 static int
 128 dmu_zfetch_colinear(zfetch_t *zf, zstream_t *zh)
 129 {
 130         zstream_t       *z_walk;
 131         zstream_t       *z_comp;
 132
 133         if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER))
 134                 return (0);
 135
 136         if (zh == NULL) {
 137                 rw_exit(&zf->zf_rwlock);
 138                 return (0);
 139         }
 140
 141         for (z_walk = list_head(&zf->zf_stream); z_walk;
 142             z_walk = list_next(&zf->zf_stream, z_walk)) {
 143                 for (z_comp = list_next(&zf->zf_stream, z_walk); z_comp;
 144                     z_comp = list_next(&zf->zf_stream, z_comp)) {
 145                         int64_t         diff;
 146
 147                         if (z_walk->zst_len != z_walk->zst_stride ||
 148                             z_comp->zst_len != z_comp->zst_stride) {
 149                                 continue;
 150                         }
 151
 152                         diff = z_comp->zst_offset - z_walk->zst_offset;
 153                         if (z_comp->zst_offset + diff == zh->zst_offset) {
 154                                 z_walk->zst_offset = zh->zst_offset;
 155                                 z_walk->zst_direction = diff < 0 ? -1 : 1;
 156                                 z_walk->zst_stride =
 157                                     diff * z_walk->zst_direction;
 158                                 z_walk->zst_ph_offset =
 159                                     zh->zst_offset + z_walk->zst_stride;
 160                                 dmu_zfetch_stream_remove(zf, z_comp);
 161                                 mutex_destroy(&z_comp->zst_lock);
 162                                 kmem_free(z_comp, sizeof (zstream_t));
 163
 164                                 dmu_zfetch_dofetch(zf, z_walk);
 165
 166                                 rw_exit(&zf->zf_rwlock);
 167                                 return (1);
 168                         }
 169
 170                         diff = z_walk->zst_offset - z_comp->zst_offset;
 171                         if (z_walk->zst_offset + diff == zh->zst_offset) {
 172                                 z_walk->zst_offset = zh->zst_offset;
 173                                 z_walk->zst_direction = diff < 0 ? -1 : 1;
 174                                 z_walk->zst_stride =
 175                                     diff * z_walk->zst_direction;
 176                                 z_walk->zst_ph_offset =
 177                                     zh->zst_offset + z_walk->zst_stride;
 178                                 dmu_zfetch_stream_remove(zf, z_comp);
 179                                 mutex_destroy(&z_comp->zst_lock);
 180                                 kmem_free(z_comp, sizeof (zstream_t));
 181
 182                                 dmu_zfetch_dofetch(zf, z_walk);
 183
 184                                 rw_exit(&zf->zf_rwlock);
 185                                 return (1);
 186                         }
 187                 }
 188         }
 189
 190         rw_exit(&zf->zf_rwlock);
 191         return (0);
 192 }
 193
 194 /*
 195  * Given a zstream_t, determine the bounds of the prefetch.  Then call the
 196  * routine that actually prefetches the individual blocks.
 197  */
 198 static void
 199 dmu_zfetch_dofetch(zfetch_t *zf, zstream_t *zs)
 200 {
 201         uint64_t        prefetch_tail;
 202         uint64_t        prefetch_limit;
 203         uint64_t        prefetch_ofst;
 204         uint64_t        prefetch_len;
 205         uint64_t        blocks_fetched;
 206
 207         zs->zst_stride = MAX((int64_t)zs->zst_stride, zs->zst_len);
 208         zs->zst_cap = MIN(zfetch_block_cap, 2 * zs->zst_cap);
 209
 210         prefetch_tail = MAX((int64_t)zs->zst_ph_offset,
 211             (int64_t)(zs->zst_offset + zs->zst_stride));
 212         /*
 213          * XXX: use a faster division method?
 214          */
 215         prefetch_limit = zs->zst_offset + zs->zst_len +
 216             (zs->zst_cap * zs->zst_stride) / zs->zst_len;
 217
 218         while (prefetch_tail < prefetch_limit) {
 219                 prefetch_ofst = zs->zst_offset + zs->zst_direction *
 220                     (prefetch_tail - zs->zst_offset);
 221
 222                 prefetch_len = zs->zst_len;
 223
 224                 /*
 225                  * Don't prefetch beyond the end of the file, if working
 226                  * backwards.
 227                  */
 228                 if ((zs->zst_direction == ZFETCH_BACKWARD) &&
 229                     (prefetch_ofst > prefetch_tail)) {
 230                         prefetch_len += prefetch_ofst;
 231                         prefetch_ofst = 0;
 232                 }
 233
 234                 /* don't prefetch more than we're supposed to */
 235                 if (prefetch_len > zs->zst_len)
 236                         break;
 237
 238                 blocks_fetched = dmu_zfetch_fetch(zf->zf_dnode,
 239                     prefetch_ofst, zs->zst_len);
 240
 241                 prefetch_tail += zs->zst_stride;
 242                 /* stop if we've run out of stuff to prefetch */
 243                 if (blocks_fetched < zs->zst_len)
 244                         break;
 245         }
 246         zs->zst_ph_offset = prefetch_tail;
 247         zs->zst_last = LBOLT;
 248 }
 249
 250 void
 251 zfetch_init(void)
 252 {
 253
 254         zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
 255             KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
 256             KSTAT_FLAG_VIRTUAL);
 257
 258         if (zfetch_ksp != NULL) {
 259                 zfetch_ksp->ks_data = &zfetch_stats;
 260                 kstat_install(zfetch_ksp);
 261         }
 262 }
 263
 264 void
 265 zfetch_fini(void)
 266 {
 267         if (zfetch_ksp != NULL) {
 268                 kstat_delete(zfetch_ksp);
 269                 zfetch_ksp = NULL;
 270         }
 271 }
 272
 273 /*
 274  * This takes a pointer to a zfetch structure and a dnode.  It performs the
 275  * necessary setup for the zfetch structure, grokking data from the
 276  * associated dnode.
 277  */
 278 void
 279 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
 280 {
 281         if (zf == NULL) {
 282                 return;
 283         }
 284
 285         zf->zf_dnode = dno;
 286         zf->zf_stream_cnt = 0;
 287         zf->zf_alloc_fail = 0;
 288
 289         list_create(&zf->zf_stream, sizeof (zstream_t),
 290             offsetof(zstream_t, zst_node));
 291
 292         rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
 293 }
 294
 295 /*
 296  * This function computes the actual size, in blocks, that can be prefetched,
 297  * and fetches it.
 298  */
 299 static uint64_t
 300 dmu_zfetch_fetch(dnode_t *dn, uint64_t blkid, uint64_t nblks)
 301 {
 302         uint64_t        fetchsz;
 303         uint64_t        i;
 304
 305         fetchsz = dmu_zfetch_fetchsz(dn, blkid, nblks);
 306
 307         for (i = 0; i < fetchsz; i++) {
 308                 dbuf_prefetch(dn, blkid + i);
 309         }
 310
 311         return (fetchsz);
 312 }
 313
 314 /*
 315  * this function returns the number of blocks that would be prefetched, based
 316  * upon the supplied dnode, blockid, and nblks.  This is used so that we can
 317  * update streams in place, and then prefetch with their old value after the
 318  * fact.  This way, we can delay the prefetch, but subsequent accesses to the
 319  * stream won't result in the same data being prefetched multiple times.
 320  */
 321 static uint64_t
 322 dmu_zfetch_fetchsz(dnode_t *dn, uint64_t blkid, uint64_t nblks)
 323 {
 324         uint64_t        fetchsz;
 325
 326         if (blkid > dn->dn_maxblkid) {
 327                 return (0);
 328         }
 329
 330         /* compute fetch size */
 331         if (blkid + nblks + 1 > dn->dn_maxblkid) {
 332                 fetchsz = (dn->dn_maxblkid - blkid) + 1;
 333                 ASSERT(blkid + fetchsz - 1 <= dn->dn_maxblkid);
 334         } else {
 335                 fetchsz = nblks;
 336         }
 337
 338
 339         return (fetchsz);
 340 }
 341
 342 /*
 343  * given a zfetch and a zstream structure, see if there is an associated zstream
 344  * for this block read.  If so, it starts a prefetch for the stream it
 345  * located and returns true, otherwise it returns false
 346  */
 347 static int
 348 dmu_zfetch_find(zfetch_t *zf, zstream_t *zh, int prefetched)
 349 {
 350         zstream_t       *zs;
 351         int64_t         diff;
 352         int             reset = !prefetched;
 353         int             rc = 0;
 354
 355         if (zh == NULL)
 356                 return (0);
 357
 358         /*
 359          * XXX: This locking strategy is a bit coarse; however, it's impact has
 360          * yet to be tested.  If this turns out to be an issue, it can be
 361          * modified in a number of different ways.
 362          */
 363
 364         rw_enter(&zf->zf_rwlock, RW_READER);
 365 top:
 366
 367         for (zs = list_head(&zf->zf_stream); zs;
 368             zs = list_next(&zf->zf_stream, zs)) {
 369
 370                 /*
 371                  * XXX - should this be an assert?
 372                  */
 373                 if (zs->zst_len == 0) {
 374                         /* bogus stream */
 375                         ZFETCHSTAT_BUMP(zfetchstat_bogus_streams);
 376                         continue;
 377                 }
 378
 379                 /*
 380                  * We hit this case when we are in a strided prefetch stream:
 381                  * we will read "len" blocks before "striding".
 382                  */
 383                 if (zh->zst_offset >= zs->zst_offset &&
 384                     zh->zst_offset < zs->zst_offset + zs->zst_len) {
 385                         if (prefetched) {
 386                                 /* already fetched */
 387                                 ZFETCHSTAT_BUMP(zfetchstat_stride_hits);
 388                                 rc = 1;
 389                                 goto out;
 390                         } else {
 391                                 ZFETCHSTAT_BUMP(zfetchstat_stride_misses);
 392                         }
 393                 }
 394
 395                 /*
 396                  * This is the forward sequential read case: we increment
 397                  * len by one each time we hit here, so we will enter this
 398                  * case on every read.
 399                  */
 400                 if (zh->zst_offset == zs->zst_offset + zs->zst_len) {
 401
 402                         reset = !prefetched && zs->zst_len > 1;
 403
 404                         if (mutex_tryenter(&zs->zst_lock) == 0) {
 405                                 rc = 1;
 406                                 goto out;
 407                         }
 408                         if (zh->zst_offset != zs->zst_offset + zs->zst_len) {
 409                                 mutex_exit(&zs->zst_lock);
 410                                 goto top;
 411                         }
 412                         zs->zst_len += zh->zst_len;
 413                         diff = zs->zst_len - zfetch_block_cap;
 414                         if (diff > 0) {
 415                                 zs->zst_offset += diff;
 416                                 zs->zst_len = zs->zst_len > diff ?
 417                                     zs->zst_len - diff : 0;
 418                         }
 419                         zs->zst_direction = ZFETCH_FORWARD;
 420
 421                         break;
 422
 423                 /*
 424                  * Same as above, but reading backwards through the file.
 425                  */
 426                 } else if (zh->zst_offset == zs->zst_offset - zh->zst_len) {
 427                         /* backwards sequential access */
 428
 429                         reset = !prefetched && zs->zst_len > 1;
 430
 431                         if (mutex_tryenter(&zs->zst_lock) == 0) {
 432                                 rc = 1;
 433                                 goto out;
 434                         }
 435                         if (zh->zst_offset != zs->zst_offset - zh->zst_len) {
 436                                 mutex_exit(&zs->zst_lock);
 437                                 goto top;
 438                         }
 439
 440                         zs->zst_offset = zs->zst_offset > zh->zst_len ?
 441                             zs->zst_offset - zh->zst_len : 0;
 442                         zs->zst_ph_offset = zs->zst_ph_offset > zh->zst_len ?
 443                             zs->zst_ph_offset - zh->zst_len : 0;
 444                         zs->zst_len += zh->zst_len;
 445
 446                         diff = zs->zst_len - zfetch_block_cap;
 447                         if (diff > 0) {
 448                                 zs->zst_ph_offset = zs->zst_ph_offset > diff ?
 449                                     zs->zst_ph_offset - diff : 0;
 450                                 zs->zst_len = zs->zst_len > diff ?
 451                                     zs->zst_len - diff : zs->zst_len;
 452                         }
 453                         zs->zst_direction = ZFETCH_BACKWARD;
 454
 455                         break;
 456
 457                 } else if ((zh->zst_offset - zs->zst_offset - zs->zst_stride <
 458                     zs->zst_len) && (zs->zst_len != zs->zst_stride)) {
 459                         /* strided forward access */
 460
 461                         if (mutex_tryenter(&zs->zst_lock) == 0) {
 462                                 rc = 1;
 463                                 goto out;
 464                         }
 465                         if ((zh->zst_offset - zs->zst_offset - zs->zst_stride >=
 466                             zs->zst_len) || (zs->zst_len == zs->zst_stride)) {
 467                                 mutex_exit(&zs->zst_lock);
 468                                 goto top;
 469                         }
 470
 471                         zs->zst_offset += zs->zst_stride;
 472                         zs->zst_direction = ZFETCH_FORWARD;
 473
 474                         break;
 475
 476                 } else if ((zh->zst_offset - zs->zst_offset + zs->zst_stride <
 477                     zs->zst_len) && (zs->zst_len != zs->zst_stride)) {
 478                         /* strided reverse access */
 479
 480                         if (mutex_tryenter(&zs->zst_lock) == 0) {
 481                                 rc = 1;
 482                                 goto out;
 483                         }
 484                         if ((zh->zst_offset - zs->zst_offset + zs->zst_stride >=
 485                             zs->zst_len) || (zs->zst_len == zs->zst_stride)) {
 486                                 mutex_exit(&zs->zst_lock);
 487                                 goto top;
 488                         }
 489
 490                         zs->zst_offset = zs->zst_offset > zs->zst_stride ?
 491                             zs->zst_offset - zs->zst_stride : 0;
 492                         zs->zst_ph_offset = (zs->zst_ph_offset >
 493                             (2 * zs->zst_stride)) ?
 494                             (zs->zst_ph_offset - (2 * zs->zst_stride)) : 0;
 495                         zs->zst_direction = ZFETCH_BACKWARD;
 496
 497                         break;
 498                 }
 499         }
 500
 501         if (zs) {
 502                 if (reset) {
 503                         zstream_t *remove = zs;
 504
 505                         ZFETCHSTAT_BUMP(zfetchstat_stream_resets);
 506                         rc = 0;
 507                         mutex_exit(&zs->zst_lock);
 508                         rw_exit(&zf->zf_rwlock);
 509                         rw_enter(&zf->zf_rwlock, RW_WRITER);
 510                         /*
 511                          * Relocate the stream, in case someone removes
 512                          * it while we were acquiring the WRITER lock.
 513                          */
 514                         for (zs = list_head(&zf->zf_stream); zs;
 515                             zs = list_next(&zf->zf_stream, zs)) {
 516                                 if (zs == remove) {
 517                                         dmu_zfetch_stream_remove(zf, zs);
 518                                         mutex_destroy(&zs->zst_lock);
 519                                         kmem_free(zs, sizeof (zstream_t));
 520                                         break;
 521                                 }
 522                         }
 523                 } else {
 524                         ZFETCHSTAT_BUMP(zfetchstat_stream_noresets);
 525                         rc = 1;
 526                         dmu_zfetch_dofetch(zf, zs);
 527                         mutex_exit(&zs->zst_lock);
 528                 }
 529         }
 530 out:
 531         rw_exit(&zf->zf_rwlock);
 532         return (rc);
 533 }
 534
 535 /*
 536  * Clean-up state associated with a zfetch structure.  This frees allocated
 537  * structure members, empties the zf_stream tree, and generally makes things
 538  * nice.  This doesn't free the zfetch_t itself, that's left to the caller.
 539  */
 540 void
 541 dmu_zfetch_rele(zfetch_t *zf)
 542 {
 543         zstream_t       *zs;
 544         zstream_t       *zs_next;
 545
 546         ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
 547
 548         for (zs = list_head(&zf->zf_stream); zs; zs = zs_next) {
 549                 zs_next = list_next(&zf->zf_stream, zs);
 550
 551                 list_remove(&zf->zf_stream, zs);
 552                 mutex_destroy(&zs->zst_lock);
 553                 kmem_free(zs, sizeof (zstream_t));
 554         }
 555         list_destroy(&zf->zf_stream);
 556         rw_destroy(&zf->zf_rwlock);
 557
 558         zf->zf_dnode = NULL;
 559 }
 560
 561 /*
 562  * Given a zfetch and zstream structure, insert the zstream structure into the
 563  * AVL tree contained within the zfetch structure.  Peform the appropriate
 564  * book-keeping.  It is possible that another thread has inserted a stream which
 565  * matches one that we are about to insert, so we must be sure to check for this
 566  * case.  If one is found, return failure, and let the caller cleanup the
 567  * duplicates.
 568  */
 569 static int
 570 dmu_zfetch_stream_insert(zfetch_t *zf, zstream_t *zs)
 571 {
 572         zstream_t       *zs_walk;
 573         zstream_t       *zs_next;
 574
 575         ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
 576
 577         for (zs_walk = list_head(&zf->zf_stream); zs_walk; zs_walk = zs_next) {
 578                 zs_next = list_next(&zf->zf_stream, zs_walk);
 579
 580                 if (dmu_zfetch_streams_equal(zs_walk, zs)) {
 581                         return (0);
 582                 }
 583         }
 584
 585         list_insert_head(&zf->zf_stream, zs);
 586         zf->zf_stream_cnt++;
 587         return (1);
 588 }
 589
 590
 591 /*
 592  * Walk the list of zstreams in the given zfetch, find an old one (by time), and
 593  * reclaim it for use by the caller.
 594  */
 595 static zstream_t *
 596 dmu_zfetch_stream_reclaim(zfetch_t *zf)
 597 {
 598         zstream_t       *zs;
 599
 600         if (! rw_tryenter(&zf->zf_rwlock, RW_WRITER))
 601                 return (0);
 602
 603         for (zs = list_head(&zf->zf_stream); zs;
 604             zs = list_next(&zf->zf_stream, zs)) {
 605
 606                 if (((LBOLT - zs->zst_last) / hz) > zfetch_min_sec_reap)
 607                         break;
 608         }
 609
 610         if (zs) {
 611                 dmu_zfetch_stream_remove(zf, zs);
 612                 mutex_destroy(&zs->zst_lock);
 613                 bzero(zs, sizeof (zstream_t));
 614         } else {
 615                 zf->zf_alloc_fail++;
 616         }
 617         rw_exit(&zf->zf_rwlock);
 618
 619         return (zs);
 620 }
 621
 622 /*
 623  * Given a zfetch and zstream structure, remove the zstream structure from its
 624  * container in the zfetch structure.  Perform the appropriate book-keeping.
 625  */
 626 static void
 627 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
 628 {
 629         ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
 630
 631         list_remove(&zf->zf_stream, zs);
 632         zf->zf_stream_cnt--;
 633 }
 634
 635 static int
 636 dmu_zfetch_streams_equal(zstream_t *zs1, zstream_t *zs2)
 637 {
 638         if (zs1->zst_offset != zs2->zst_offset)
 639                 return (0);
 640
 641         if (zs1->zst_len != zs2->zst_len)
 642                 return (0);
 643
 644         if (zs1->zst_stride != zs2->zst_stride)
 645                 return (0);
 646
 647         if (zs1->zst_ph_offset != zs2->zst_ph_offset)
 648                 return (0);
 649
 650         if (zs1->zst_cap != zs2->zst_cap)
 651                 return (0);
 652
 653         if (zs1->zst_direction != zs2->zst_direction)
 654                 return (0);
 655
 656         return (1);
 657 }
 658
 659 /*
 660  * This is the prefetch entry point.  It calls all of the other dmu_zfetch
 661  * routines to create, delete, find, or operate upon prefetch streams.
 662  */
 663 void
 664 dmu_zfetch(zfetch_t *zf, uint64_t offset, uint64_t size, int prefetched)
 665 {
 666         zstream_t       zst;
 667         zstream_t       *newstream;
 668         int             fetched;
 669         int             inserted;
 670         unsigned int    blkshft;
 671         uint64_t        blksz;
 672
 673         if (zfs_prefetch_disable)
 674                 return;
 675
 676         /* files that aren't ln2 blocksz are only one block -- nothing to do */
 677         if (!zf->zf_dnode->dn_datablkshift)
 678                 return;
 679
 680         /* convert offset and size, into blockid and nblocks */
 681         blkshft = zf->zf_dnode->dn_datablkshift;
 682         blksz = (1 << blkshft);
 683
 684         bzero(&zst, sizeof (zstream_t));
 685         zst.zst_offset = offset >> blkshft;
 686         zst.zst_len = (P2ROUNDUP(offset + size, blksz) -
 687             P2ALIGN(offset, blksz)) >> blkshft;
 688
 689         fetched = dmu_zfetch_find(zf, &zst, prefetched);
 690         if (fetched) {
 691                 ZFETCHSTAT_BUMP(zfetchstat_hits);
 692         } else {
 693                 ZFETCHSTAT_BUMP(zfetchstat_misses);
 694                 if (fetched = dmu_zfetch_colinear(zf, &zst)) {
 695                         ZFETCHSTAT_BUMP(zfetchstat_colinear_hits);
 696                 } else {
 697                         ZFETCHSTAT_BUMP(zfetchstat_colinear_misses);
 698                 }
 699         }
 700
 701         if (!fetched) {
 702                 newstream = dmu_zfetch_stream_reclaim(zf);
 703
 704                 /*
 705                  * we still couldn't find a stream, drop the lock, and allocate
 706                  * one if possible.  Otherwise, give up and go home.
 707                  */
 708                 if (newstream) {
 709                         ZFETCHSTAT_BUMP(zfetchstat_reclaim_successes);
 710                 } else {
 711                         uint64_t        maxblocks;
 712                         uint32_t        max_streams;
 713                         uint32_t        cur_streams;
 714
 715                         ZFETCHSTAT_BUMP(zfetchstat_reclaim_failures);
 716                         cur_streams = zf->zf_stream_cnt;
 717                         maxblocks = zf->zf_dnode->dn_maxblkid;
 718
 719                         max_streams = MIN(zfetch_max_streams,
 720                             (maxblocks / zfetch_block_cap));
 721                         if (max_streams == 0) {
 722                                 max_streams++;
 723                         }
 724
 725                         if (cur_streams >= max_streams) {
 726                                 return;
 727                         }
 728                         newstream = kmem_zalloc(sizeof (zstream_t), KM_SLEEP);
 729                 }
 730
 731                 newstream->zst_offset = zst.zst_offset;
 732                 newstream->zst_len = zst.zst_len;
 733                 newstream->zst_stride = zst.zst_len;
 734                 newstream->zst_ph_offset = zst.zst_len + zst.zst_offset;
 735                 newstream->zst_cap = zst.zst_len;
 736                 newstream->zst_direction = ZFETCH_FORWARD;
 737                 newstream->zst_last = LBOLT;
 738
 739                 mutex_init(&newstream->zst_lock, NULL, MUTEX_DEFAULT, NULL);
 740
 741                 rw_enter(&zf->zf_rwlock, RW_WRITER);
 742                 inserted = dmu_zfetch_stream_insert(zf, newstream);
 743                 rw_exit(&zf->zf_rwlock);
 744
 745                 if (!inserted) {
 746                         mutex_destroy(&newstream->zst_lock);
 747                         kmem_free(newstream, sizeof (zstream_t));
 748                 }
 749         }
 750 }