module/zfs/vdev_queue.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 2008 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25
  26 #include <sys/zfs_context.h>
  27 #include <sys/spa.h>
  28 #include <sys/vdev_impl.h>
  29 #include <sys/zio.h>
  30 #include <sys/avl.h>
  31
  32 /*
  33  * These tunables are for performance analysis.
  34  */
  35 /*
  36  * zfs_vdev_max_pending is the maximum number of i/os concurrently
  37  * pending to each device.  zfs_vdev_min_pending is the initial number
  38  * of i/os pending to each device (before it starts ramping up to
  39  * max_pending).
  40  */
  41 int zfs_vdev_max_pending = 35;
  42 int zfs_vdev_min_pending = 4;
  43
  44 /* deadline = pri + (lbolt >> time_shift) */
  45 int zfs_vdev_time_shift = 6;
  46
  47 /* exponential I/O issue ramp-up rate */
  48 int zfs_vdev_ramp_rate = 2;
  49
  50 /*
  51  * i/os will be aggregated into a single large i/o up to
  52  * zfs_vdev_aggregation_limit bytes long.
  53  */
  54 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
  55
  56 /*
  57  * Virtual device vector for disk I/O scheduling.
  58  */
  59 int
  60 vdev_queue_deadline_compare(const void *x1, const void *x2)
  61 {
  62         const zio_t *z1 = x1;
  63         const zio_t *z2 = x2;
  64
  65         if (z1->io_deadline < z2->io_deadline)
  66                 return (-1);
  67         if (z1->io_deadline > z2->io_deadline)
  68                 return (1);
  69
  70         if (z1->io_offset < z2->io_offset)
  71                 return (-1);
  72         if (z1->io_offset > z2->io_offset)
  73                 return (1);
  74
  75         if (z1 < z2)
  76                 return (-1);
  77         if (z1 > z2)
  78                 return (1);
  79
  80         return (0);
  81 }
  82
  83 int
  84 vdev_queue_offset_compare(const void *x1, const void *x2)
  85 {
  86         const zio_t *z1 = x1;
  87         const zio_t *z2 = x2;
  88
  89         if (z1->io_offset < z2->io_offset)
  90                 return (-1);
  91         if (z1->io_offset > z2->io_offset)
  92                 return (1);
  93
  94         if (z1 < z2)
  95                 return (-1);
  96         if (z1 > z2)
  97                 return (1);
  98
  99         return (0);
 100 }
 101
 102 void
 103 vdev_queue_init(vdev_t *vd)
 104 {
 105         vdev_queue_t *vq = &vd->vdev_queue;
 106
 107         mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
 108
 109         avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
 110             sizeof (zio_t), offsetof(struct zio, io_deadline_node));
 111
 112         avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
 113             sizeof (zio_t), offsetof(struct zio, io_offset_node));
 114
 115         avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
 116             sizeof (zio_t), offsetof(struct zio, io_offset_node));
 117
 118         avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
 119             sizeof (zio_t), offsetof(struct zio, io_offset_node));
 120 }
 121
 122 void
 123 vdev_queue_fini(vdev_t *vd)
 124 {
 125         vdev_queue_t *vq = &vd->vdev_queue;
 126
 127         avl_destroy(&vq->vq_deadline_tree);
 128         avl_destroy(&vq->vq_read_tree);
 129         avl_destroy(&vq->vq_write_tree);
 130         avl_destroy(&vq->vq_pending_tree);
 131
 132         mutex_destroy(&vq->vq_lock);
 133 }
 134
 135 static void
 136 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
 137 {
 138         avl_add(&vq->vq_deadline_tree, zio);
 139         avl_add(zio->io_vdev_tree, zio);
 140 }
 141
 142 static void
 143 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
 144 {
 145         avl_remove(&vq->vq_deadline_tree, zio);
 146         avl_remove(zio->io_vdev_tree, zio);
 147 }
 148
 149 static void
 150 vdev_queue_agg_io_done(zio_t *aio)
 151 {
 152         zio_t *dio;
 153         uint64_t offset = 0;
 154
 155         while ((dio = aio->io_delegate_list) != NULL) {
 156                 if (aio->io_type == ZIO_TYPE_READ)
 157                         bcopy((char *)aio->io_data + offset, dio->io_data,
 158                             dio->io_size);
 159                 offset += dio->io_size;
 160                 aio->io_delegate_list = dio->io_delegate_next;
 161                 dio->io_delegate_next = NULL;
 162                 dio->io_error = aio->io_error;
 163                 zio_execute(dio);
 164         }
 165         ASSERT3U(offset, ==, aio->io_size);
 166
 167         zio_buf_free(aio->io_data, aio->io_size);
 168 }
 169
 170 #define IS_ADJACENT(io, nio) \
 171         ((io)->io_offset + (io)->io_size == (nio)->io_offset)
 172
 173 static zio_t *
 174 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
 175 {
 176         zio_t *fio, *lio, *aio, *dio;
 177         avl_tree_t *tree;
 178         uint64_t size;
 179
 180         ASSERT(MUTEX_HELD(&vq->vq_lock));
 181
 182         if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
 183             avl_numnodes(&vq->vq_deadline_tree) == 0)
 184                 return (NULL);
 185
 186         fio = lio = avl_first(&vq->vq_deadline_tree);
 187
 188         tree = fio->io_vdev_tree;
 189         size = fio->io_size;
 190
 191         while ((dio = AVL_PREV(tree, fio)) != NULL && IS_ADJACENT(dio, fio) &&
 192             !((dio->io_flags | fio->io_flags) & ZIO_FLAG_DONT_AGGREGATE) &&
 193             size + dio->io_size <= zfs_vdev_aggregation_limit) {
 194                 dio->io_delegate_next = fio;
 195                 fio = dio;
 196                 size += dio->io_size;
 197         }
 198
 199         while ((dio = AVL_NEXT(tree, lio)) != NULL && IS_ADJACENT(lio, dio) &&
 200             !((lio->io_flags | dio->io_flags) & ZIO_FLAG_DONT_AGGREGATE) &&
 201             size + dio->io_size <= zfs_vdev_aggregation_limit) {
 202                 lio->io_delegate_next = dio;
 203                 lio = dio;
 204                 size += dio->io_size;
 205         }
 206
 207         if (fio != lio) {
 208                 char *buf = zio_buf_alloc(size);
 209                 uint64_t offset = 0;
 210
 211                 ASSERT(size <= zfs_vdev_aggregation_limit);
 212
 213                 aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
 214                     buf, size, fio->io_type, ZIO_PRIORITY_NOW,
 215                     ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
 216                     vdev_queue_agg_io_done, NULL);
 217
 218                 aio->io_delegate_list = fio;
 219
 220                 for (dio = fio; dio != NULL; dio = dio->io_delegate_next) {
 221                         ASSERT(dio->io_type == aio->io_type);
 222                         ASSERT(dio->io_vdev_tree == tree);
 223                         if (dio->io_type == ZIO_TYPE_WRITE)
 224                                 bcopy(dio->io_data, buf + offset, dio->io_size);
 225                         offset += dio->io_size;
 226                         vdev_queue_io_remove(vq, dio);
 227                         zio_vdev_io_bypass(dio);
 228                 }
 229
 230                 ASSERT(offset == size);
 231
 232                 avl_add(&vq->vq_pending_tree, aio);
 233
 234                 return (aio);
 235         }
 236
 237         ASSERT(fio->io_vdev_tree == tree);
 238         vdev_queue_io_remove(vq, fio);
 239
 240         avl_add(&vq->vq_pending_tree, fio);
 241
 242         return (fio);
 243 }
 244
 245 zio_t *
 246 vdev_queue_io(zio_t *zio)
 247 {
 248         vdev_queue_t *vq = &zio->io_vd->vdev_queue;
 249         zio_t *nio;
 250
 251         ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
 252
 253         if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
 254                 return (zio);
 255
 256         zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
 257
 258         if (zio->io_type == ZIO_TYPE_READ)
 259                 zio->io_vdev_tree = &vq->vq_read_tree;
 260         else
 261                 zio->io_vdev_tree = &vq->vq_write_tree;
 262
 263         mutex_enter(&vq->vq_lock);
 264
 265         zio->io_deadline = (lbolt64 >> zfs_vdev_time_shift) + zio->io_priority;
 266
 267         vdev_queue_io_add(vq, zio);
 268
 269         nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
 270
 271         mutex_exit(&vq->vq_lock);
 272
 273         if (nio == NULL)
 274                 return (NULL);
 275
 276         if (nio->io_done == vdev_queue_agg_io_done) {
 277                 zio_nowait(nio);
 278                 return (NULL);
 279         }
 280
 281         return (nio);
 282 }
 283
 284 void
 285 vdev_queue_io_done(zio_t *zio)
 286 {
 287         vdev_queue_t *vq = &zio->io_vd->vdev_queue;
 288
 289         mutex_enter(&vq->vq_lock);
 290
 291         avl_remove(&vq->vq_pending_tree, zio);
 292
 293         for (int i = 0; i < zfs_vdev_ramp_rate; i++) {
 294                 zio_t *nio = vdev_queue_io_to_issue(vq, zfs_vdev_max_pending);
 295                 if (nio == NULL)
 296                         break;
 297                 mutex_exit(&vq->vq_lock);
 298                 if (nio->io_done == vdev_queue_agg_io_done) {
 299                         zio_nowait(nio);
 300                 } else {
 301                         zio_vdev_io_reissue(nio);
 302                         zio_execute(nio);
 303                 }
 304                 mutex_enter(&vq->vq_lock);
 305         }
 306
 307         mutex_exit(&vq->vq_lock);
 308 }