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, 2014 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 /* number of bytes in a array_read at which we stop prefetching (1MB) */
  54 unsigned long   zfetch_array_rd_sz = 1024 * 1024;
  55
  56 typedef struct zfetch_stats {
  57         kstat_named_t zfetchstat_hits;
  58         kstat_named_t zfetchstat_misses;
  59         kstat_named_t zfetchstat_max_streams;
  60 } zfetch_stats_t;
  61
  62 static zfetch_stats_t zfetch_stats = {
  63         { "hits",                       KSTAT_DATA_UINT64 },
  64         { "misses",                     KSTAT_DATA_UINT64 },
  65         { "max_streams",                KSTAT_DATA_UINT64 },
  66 };
  67
  68 #define ZFETCHSTAT_BUMP(stat) \
  69         atomic_inc_64(&zfetch_stats.stat.value.ui64);
  70
  71 kstat_t         *zfetch_ksp;
  72
  73 void
  74 zfetch_init(void)
  75 {
  76         zfetch_ksp = kstat_create("zfs", 0, "zfetchstats", "misc",
  77             KSTAT_TYPE_NAMED, sizeof (zfetch_stats) / sizeof (kstat_named_t),
  78             KSTAT_FLAG_VIRTUAL);
  79
  80         if (zfetch_ksp != NULL) {
  81                 zfetch_ksp->ks_data = &zfetch_stats;
  82                 kstat_install(zfetch_ksp);
  83         }
  84 }
  85
  86 void
  87 zfetch_fini(void)
  88 {
  89         if (zfetch_ksp != NULL) {
  90                 kstat_delete(zfetch_ksp);
  91                 zfetch_ksp = NULL;
  92         }
  93 }
  94
  95 /*
  96  * This takes a pointer to a zfetch structure and a dnode.  It performs the
  97  * necessary setup for the zfetch structure, grokking data from the
  98  * associated dnode.
  99  */
 100 void
 101 dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
 102 {
 103         if (zf == NULL)
 104                 return;
 105
 106         zf->zf_dnode = dno;
 107
 108         list_create(&zf->zf_stream, sizeof (zstream_t),
 109             offsetof(zstream_t, zs_node));
 110
 111         rw_init(&zf->zf_rwlock, NULL, RW_DEFAULT, NULL);
 112 }
 113
 114 static void
 115 dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
 116 {
 117         ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
 118         list_remove(&zf->zf_stream, zs);
 119         mutex_destroy(&zs->zs_lock);
 120         kmem_free(zs, sizeof (*zs));
 121 }
 122
 123 /*
 124  * Clean-up state associated with a zfetch structure (e.g. destroy the
 125  * streams).  This doesn't free the zfetch_t itself, that's left to the caller.
 126  */
 127 void
 128 dmu_zfetch_fini(zfetch_t *zf)
 129 {
 130         zstream_t *zs;
 131
 132         ASSERT(!RW_LOCK_HELD(&zf->zf_rwlock));
 133
 134         rw_enter(&zf->zf_rwlock, RW_WRITER);
 135         while ((zs = list_head(&zf->zf_stream)) != NULL)
 136                 dmu_zfetch_stream_remove(zf, zs);
 137         rw_exit(&zf->zf_rwlock);
 138         list_destroy(&zf->zf_stream);
 139         rw_destroy(&zf->zf_rwlock);
 140
 141         zf->zf_dnode = NULL;
 142 }
 143
 144 /*
 145  * If there aren't too many streams already, create a new stream.
 146  * The "blkid" argument is the next block that we expect this stream to access.
 147  * While we're here, clean up old streams (which haven't been
 148  * accessed for at least zfetch_min_sec_reap seconds).
 149  */
 150 static void
 151 dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
 152 {
 153         zstream_t *zs;
 154         zstream_t *zs_next;
 155         int numstreams = 0;
 156         uint32_t max_streams;
 157
 158         ASSERT(RW_WRITE_HELD(&zf->zf_rwlock));
 159
 160         /*
 161          * Clean up old streams.
 162          */
 163         for (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         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         zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
 190         zs->zs_blkid = blkid;
 191         zs->zs_pf_blkid = blkid;
 192         zs->zs_atime = gethrtime();
 193         mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
 194
 195         list_insert_head(&zf->zf_stream, zs);
 196 }
 197
 198 /*
 199  * This is the prefetch entry point.  It calls all of the other dmu_zfetch
 200  * routines to create, delete, find, or operate upon prefetch streams.
 201  */
 202 void
 203 dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks)
 204 {
 205         zstream_t *zs;
 206         int64_t pf_start;
 207         int pf_nblks;
 208         int i;
 209
 210         if (zfs_prefetch_disable)
 211                 return;
 212
 213         /*
 214          * As a fast path for small (single-block) files, ignore access
 215          * to the first block.
 216          */
 217         if (blkid == 0)
 218                 return;
 219
 220         rw_enter(&zf->zf_rwlock, RW_READER);
 221
 222         for (zs = list_head(&zf->zf_stream); zs != NULL;
 223             zs = list_next(&zf->zf_stream, zs)) {
 224                 if (blkid == zs->zs_blkid) {
 225                         mutex_enter(&zs->zs_lock);
 226                         /*
 227                          * zs_blkid could have changed before we
 228                          * acquired zs_lock; re-check them here.
 229                          */
 230                         if (blkid != zs->zs_blkid) {
 231                                 mutex_exit(&zs->zs_lock);
 232                                 continue;
 233                         }
 234                         break;
 235                 }
 236         }
 237
 238         if (zs == NULL) {
 239                 /*
 240                  * This access is not part of any existing stream.  Create
 241                  * a new stream for it.
 242                  */
 243                 ZFETCHSTAT_BUMP(zfetchstat_misses);
 244                 if (rw_tryupgrade(&zf->zf_rwlock))
 245                         dmu_zfetch_stream_create(zf, blkid + nblks);
 246                 rw_exit(&zf->zf_rwlock);
 247                 return;
 248         }
 249
 250         /*
 251          * This access was to a block that we issued a prefetch for on
 252          * behalf of this stream. Issue further prefetches for this stream.
 253          *
 254          * Normally, we start prefetching where we stopped
 255          * prefetching last (zs_pf_blkid).  But when we get our first
 256          * hit on this stream, zs_pf_blkid == zs_blkid, we don't
 257          * want to prefetch to block we just accessed.  In this case,
 258          * start just after the block we just accessed.
 259          */
 260         pf_start = MAX(zs->zs_pf_blkid, blkid + nblks);
 261
 262         /*
 263          * Double our amount of prefetched data, but don't let the
 264          * prefetch get further ahead than zfetch_max_distance.
 265          */
 266         pf_nblks =
 267             MIN((int64_t)zs->zs_pf_blkid - zs->zs_blkid + nblks,
 268             zs->zs_blkid + nblks +
 269             (zfetch_max_distance >> zf->zf_dnode->dn_datablkshift) - pf_start);
 270
 271         zs->zs_pf_blkid = pf_start + pf_nblks;
 272         zs->zs_atime = gethrtime();
 273         zs->zs_blkid = blkid + nblks;
 274
 275         /*
 276          * dbuf_prefetch() issues the prefetch i/o
 277          * asynchronously, but it may need to wait for an
 278          * indirect block to be read from disk.  Therefore
 279          * we do not want to hold any locks while we call it.
 280          */
 281         mutex_exit(&zs->zs_lock);
 282         rw_exit(&zf->zf_rwlock);
 283         for (i = 0; i < pf_nblks; i++) {
 284                 dbuf_prefetch(zf->zf_dnode, 0, pf_start + i,
 285                     ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH);
 286         }
 287         ZFETCHSTAT_BUMP(zfetchstat_hits);
 288 }
 289
 290 #if defined(_KERNEL) && defined(HAVE_SPL)
 291 module_param(zfs_prefetch_disable, int, 0644);
 292 MODULE_PARM_DESC(zfs_prefetch_disable, "Disable all ZFS prefetching");
 293
 294 module_param(zfetch_max_streams, uint, 0644);
 295 MODULE_PARM_DESC(zfetch_max_streams, "Max number of streams per zfetch");
 296
 297 module_param(zfetch_min_sec_reap, uint, 0644);
 298 MODULE_PARM_DESC(zfetch_min_sec_reap, "Min time before stream reclaim");
 299
 300 module_param(zfetch_max_distance, uint, 0644);
 301 MODULE_PARM_DESC(zfetch_max_distance,
 302         "Max bytes to prefetch per stream (default 8MB)");
 303
 304 module_param(zfetch_array_rd_sz, ulong, 0644);
 305 MODULE_PARM_DESC(zfetch_array_rd_sz, "Number of bytes in a array_read");
 306 #endif