module/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 /*
  27  * Copyright (c) 2013, 2015 by Delphix. All rights reserved.
  28  */
  29
  30 #include <sys/zfs_context.h>
  31 #include <sys/dnode.h>
  32 #include <sys/dmu_objset.h>
  33 #include <sys/dmu_zfetch.h>
  34 #include <sys/dmu.h>
  35 #include <sys/dbuf.h>
  36 #include <sys/kstat.h>
  37
  38 /*
  39  * This tunable disables predictive prefetch.  Note that it leaves "prescient"
  40  * prefetch (e.g. prefetch for zfs send) intact.  Unlike predictive prefetch,
  41  * prescient prefetch never issues i/os that end up not being needed,
  42  * so it can't hurt performance.
  43  */
  44
  45 int zfs_prefetch_disable = B_FALSE;
  46
  47 /* max # of streams per zfetch */
  48 unsigned int    zfetch_max_streams = 8;
  49 /* min time before stream reclaim */
  50 unsigned int    zfetch_min_sec_reap = 2;
  51 /* max bytes to prefetch per stream (default 8MB) */
  52 unsigned int    zfetch_max_distance = 8 * 1024 * 1024;
  53 /* max bytes to prefetch indirects for per stream (default 64MB) */
  54 unsigned int    zfetch_max_idistance = 64 * 1024 * 1024;
  55 /* max number of bytes in an array_read in which we allow prefetching (1MB) */
  56 unsigned long   zfetch_array_rd_sz = 1024 * 1024;
  57
  58 typedef struct zfetch_stats {
  59         kstat_named_t zfetchstat_hits;
  60         kstat_named_t zfetchstat_misses;
  61         kstat_named_t zfetchstat_max_streams;
  62 } zfetch_stats_t;
  63
  64 static zfetch_stats_t zfetch_stats = {
  65         { "hits",                       KSTAT_DATA_UINT64 },
  66         { "misses",                     KSTAT_DATA_UINT64 },
  67         { "max_streams",                KSTAT_DATA_UINT64 },
  68 };
  69
  70 #define ZFETCHSTAT_BUMP(stat) \
  71         atomic_inc_64(&zfetch_stats.stat.value.ui64);
  72
  73 kstat_t         *zfetch_ksp;
  74
  75 void
  76 zfetch_init(void)
  77 {
  78         zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
  79             KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
  80             KSTAT_FLAG_VIRTUAL);
  81
  82         if (zfetch_ksp != NULL) {
  83                 zfetch_ksp->ks_data = &zfetch_stats;
  84                 kstat_install(zfetch_ksp);
  85         }
  86 }
  87
  88 void
  89 zfetch_fini(void)
  90 {
  91         if (zfetch_ksp != NULL) {
  92                 kstat_delete(zfetch_ksp);
  93                 zfetch_ksp = NULL;
  94         }
  95 }
  96
  97 /*
  98  * This takes a pointer to a zfetch structure and a dnode.  It performs the
  99  * necessary setup for the zfetch structure, grokking data from the
 100  * associated dnode.
 101  */
 102 void
 103 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
 104 {
 105         if (zf == NULL)
 106                 return;
 107
 108         zf->zf_dnode = dno;
 109
 110         list_create(&zf->zf_stream, sizeof (zstream_t),
 111             offsetof(zstream_t, zs_node));
 112
 113         rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
 114 }
 115
 116 static void
 117 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
 118 {
 119         ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
 120         list_remove(&zf->zf_stream, zs);
 121         mutex_destroy(&zs->zs_lock);
 122         kmem_free(zs, sizeof (*zs));
 123 }
 124
 125 /*
 126  * Clean-up state associated with a zfetch structure (e.g. destroy the
 127  * streams).  This doesn't free the zfetch_t itself, that's left to the caller.
 128  */
 129 void
 130 dmu_zfetch_fini(zfetch_t *zf)
 131 {
 132         zstream_t *zs;
 133
 134         ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
 135
 136         rw_enter(&zf->zf_rwlock, RW_WRITER);
 137         while ((zs = list_head(&zf->zf_stream)) != NULL)
 138                 dmu_zfetch_stream_remove(zf, zs);
 139         rw_exit(&zf->zf_rwlock);
 140         list_destroy(&zf->zf_stream);
 141         rw_destroy(&zf->zf_rwlock);
 142
 143         zf->zf_dnode = NULL;
 144 }
 145
 146 /*
 147  * If there aren't too many streams already, create a new stream.
 148  * The "blkid" argument is the next block that we expect this stream to access.
 149  * While we're here, clean up old streams (which haven't been
 150  * accessed for at least zfetch_min_sec_reap seconds).
 151  */
 152 static void
 153 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
 154 {
 155         zstream_t *zs_next;
 156         int numstreams = 0;
 157
 158         ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
 159
 160         /*
 161          * Clean up old streams.
 162          */
 163         for (zstream_t *zs = list_head(&zf->zf_stream);
 164             zs != NULL; zs = zs_next) {
 165                 zs_next = list_next(&zf->zf_stream, zs);
 166                 if (((gethrtime() - zs->zs_atime) / NANOSEC) >
 167                     zfetch_min_sec_reap)
 168                         dmu_zfetch_stream_remove(zf, zs);
 169                 else
 170                         numstreams++;
 171         }
 172
 173         /*
 174          * The maximum number of streams is normally zfetch_max_streams,
 175          * but for small files we lower it such that it's at least possible
 176          * for all the streams to be non-overlapping.
 177          *
 178          * If we are already at the maximum number of streams for this file,
 179          * even after removing old streams, then don't create this stream.
 180          */
 181         uint32_t max_streams = MAX(1, MIN(zfetch_max_streams,
 182             zf->zf_dnode->dn_maxblkid * zf->zf_dnode->dn_datablksz /
 183             zfetch_max_distance));
 184         if (numstreams >= max_streams) {
 185                 ZFETCHSTAT_BUMP(zfetchstat_max_streams);
 186                 return;
 187         }
 188
 189         zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
 190         zs->zs_blkid = blkid;
 191         zs->zs_pf_blkid = blkid;
 192         zs->zs_ipf_blkid = blkid;
 193         zs->zs_atime = gethrtime();
 194         mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
 195
 196         list_insert_head(&zf->zf_stream, zs);
 197 }
 198
 199 /*
 200  * This is the predictive prefetch entry point.  It associates dnode access
 201  * specified with blkid and nblks arguments with prefetch stream, predicts
 202  * further accesses based on that stats and initiates speculative prefetch.
 203  * fetch_data argument specifies whether actual data blocks should be fetched:
 204  *   FALSE -- prefetch only indirect blocks for predicted data blocks;
 205  *   TRUE -- prefetch predicted data blocks plus following indirect blocks.
 206  */
 207 void
 208 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data)
 209 {
 210         zstream_t *zs;
 211         int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
 212         int64_t pf_ahead_blks, max_blks;
 213         int epbs, max_dist_blks, pf_nblks, ipf_nblks;
 214         uint64_t end_of_access_blkid;
 215         end_of_access_blkid = blkid + nblks;
 216         spa_t *spa = zf->zf_dnode->dn_objset->os_spa;
 217         krw_t rw = RW_READER;
 218
 219         if (zfs_prefetch_disable)
 220                 return;
 221         /*
 222          * If we haven't yet loaded the indirect vdevs' mappings, we
 223          * can only read from blocks that we carefully ensure are on
 224          * concrete vdevs (or previously-loaded indirect vdevs).  So we
 225          * can't allow the predictive prefetcher to attempt reads of other
 226          * blocks (e.g. of the MOS's dnode object).
 227          */
 228         if (!spa_indirect_vdevs_loaded(spa))
 229                 return;
 230
 231         /*
 232          * As a fast path for small (single-block) files, ignore access
 233          * to the first block.
 234          */
 235         if (blkid == 0)
 236                 return;
 237
 238 retry:
 239         rw_enter(&zf->zf_rwlock, rw);
 240
 241         /*
 242          * Find matching prefetch stream.  Depending on whether the accesses
 243          * are block-aligned, first block of the new access may either follow
 244          * the last block of the previous access, or be equal to it.
 245          */
 246         for (zs = list_head(&zf->zf_stream); zs != NULL;
 247             zs = list_next(&zf->zf_stream, zs)) {
 248                 if (blkid == zs->zs_blkid || blkid + 1 == zs->zs_blkid) {
 249                         mutex_enter(&zs->zs_lock);
 250                         /*
 251                          * zs_blkid could have changed before we
 252                          * acquired zs_lock; re-check them here.
 253                          */
 254                         if (blkid == zs->zs_blkid) {
 255                                 break;
 256                         } else if (blkid + 1 == zs->zs_blkid) {
 257                                 blkid++;
 258                                 nblks--;
 259                                 if (nblks == 0) {
 260                                         /* Already prefetched this before. */
 261                                         mutex_exit(&zs->zs_lock);
 262                                         rw_exit(&zf->zf_rwlock);
 263                                         return;
 264                                 }
 265                                 break;
 266                         }
 267                         mutex_exit(&zs->zs_lock);
 268                 }
 269         }
 270
 271         if (zs == NULL) {
 272                 /*
 273                  * This access is not part of any existing stream.  Create
 274                  * a new stream for it.
 275                  */
 276                 ZFETCHSTAT_BUMP(zfetchstat_misses);
 277                 if (rw == RW_READER && !rw_tryupgrade(&zf->zf_rwlock)) {
 278                         rw_exit(&zf->zf_rwlock);
 279                         rw = RW_WRITER;
 280                         goto retry;
 281                 }
 282
 283                 dmu_zfetch_stream_create(zf, end_of_access_blkid);
 284                 rw_exit(&zf->zf_rwlock);
 285                 return;
 286         }
 287
 288         /*
 289          * This access was to a block that we issued a prefetch for on
 290          * behalf of this stream. Issue further prefetches for this stream.
 291          *
 292          * Normally, we start prefetching where we stopped
 293          * prefetching last (zs_pf_blkid).  But when we get our first
 294          * hit on this stream, zs_pf_blkid == zs_blkid, we don't
 295          * want to prefetch the block we just accessed.  In this case,
 296          * start just after the block we just accessed.
 297          */
 298         pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
 299
 300         /*
 301          * Double our amount of prefetched data, but don't let the
 302          * prefetch get further ahead than zfetch_max_distance.
 303          */
 304         if (fetch_data) {
 305                 max_dist_blks =
 306                     zfetch_max_distance >> zf->zf_dnode->dn_datablkshift;
 307                 /*
 308                  * Previously, we were (zs_pf_blkid - blkid) ahead.  We
 309                  * want to now be double that, so read that amount again,
 310                  * plus the amount we are catching up by (i.e. the amount
 311                  * read just now).
 312                  */
 313                 pf_ahead_blks = zs->zs_pf_blkid - blkid + nblks;
 314                 max_blks = max_dist_blks - (pf_start - end_of_access_blkid);
 315                 pf_nblks = MIN(pf_ahead_blks, max_blks);
 316         } else {
 317                 pf_nblks = 0;
 318         }
 319
 320         zs->zs_pf_blkid = pf_start + pf_nblks;
 321
 322         /*
 323          * Do the same for indirects, starting from where we stopped last,
 324          * or where we will stop reading data blocks (and the indirects
 325          * that point to them).
 326          */
 327         ipf_start = MAX(zs->zs_ipf_blkid, zs->zs_pf_blkid);
 328         max_dist_blks = zfetch_max_idistance >> zf->zf_dnode->dn_datablkshift;
 329         /*
 330          * We want to double our distance ahead of the data prefetch
 331          * (or reader, if we are not prefetching data).  Previously, we
 332          * were (zs_ipf_blkid - blkid) ahead.  To double that, we read
 333          * that amount again, plus the amount we are catching up by
 334          * (i.e. the amount read now + the amount of data prefetched now).
 335          */
 336         pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
 337         max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
 338         ipf_nblks = MIN(pf_ahead_blks, max_blks);
 339         zs->zs_ipf_blkid = ipf_start + ipf_nblks;
 340
 341         epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
 342         ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
 343         ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
 344
 345         zs->zs_atime = gethrtime();
 346         zs->zs_blkid = end_of_access_blkid;
 347         mutex_exit(&zs->zs_lock);
 348         rw_exit(&zf->zf_rwlock);
 349
 350         /*
 351          * dbuf_prefetch() is asynchronous (even when it needs to read
 352          * indirect blocks), but we still prefer to drop our locks before
 353          * calling it to reduce the time we hold them.
 354          */
 355
 356         for (int i = 0; i < pf_nblks; i++) {
 357                 dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
 358                     ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
 359         }
 360         for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) {
 361                 dbuf_prefetch(zf->zf_dnode, 1, iblk,
 362                     ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
 363         }
 364         ZFETCHSTAT_BUMP(zfetchstat_hits);
 365 }
 366
 367 #if defined(_KERNEL)
 368 /* BEGIN CSTYLED */
 369 module_param(zfs_prefetch_disable, int, 0644);
 370 MODULE_PARM_DESC(zfs_prefetch_disable, "Disable all ZFS prefetching");
 371
 372 module_param(zfetch_max_streams, uint, 0644);
 373 MODULE_PARM_DESC(zfetch_max_streams, "Max number of streams per zfetch");
 374
 375 module_param(zfetch_min_sec_reap, uint, 0644);
 376 MODULE_PARM_DESC(zfetch_min_sec_reap, "Min time before stream reclaim");
 377
 378 module_param(zfetch_max_distance, uint, 0644);
 379 MODULE_PARM_DESC(zfetch_max_distance,
 380         "Max bytes to prefetch per stream (default 8MB)");
 381
 382 module_param(zfetch_array_rd_sz, ulong, 0644);
 383 MODULE_PARM_DESC(zfetch_array_rd_sz, "Number of bytes in a array_read");
 384 /* END CSTYLED */
 385 #endif