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 by Delphix. All rights reserved.
  28  * Copyright (c) 2013 Steven Hartland. All rights reserved.
  29  */
  30 
  31 #include <sys/zfs_context.h>
  32 #include <sys/vdev_impl.h>
  33 #include <sys/spa_impl.h>
  34 #include <sys/zio.h>
  35 #include <sys/avl.h>
  36 #include <sys/dsl_pool.h>
  37 
  38 /*
  39  * ZFS I/O Scheduler
  40  * ---------------
  41  *
  42  * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios.  The
  43  * I/O scheduler determines when and in what order those operations are
  44  * issued.  The I/O scheduler divides operations into five I/O classes
  45  * prioritized in the following order: sync read, sync write, async read,
  46  * async write, and scrub/resilver.  Each queue defines the minimum and
  47  * maximum number of concurrent operations that may be issued to the device.
  48  * In addition, the device has an aggregate maximum. Note that the sum of the
  49  * per-queue minimums must not exceed the aggregate maximum, and if the
  50  * aggregate maximum is equal to or greater than the sum of the per-queue
  51  * maximums, the per-queue minimum has no effect.
  52  *
  53  * For many physical devices, throughput increases with the number of
  54  * concurrent operations, but latency typically suffers. Further, physical
  55  * devices typically have a limit at which more concurrent operations have no
  56  * effect on throughput or can actually cause it to decrease.
  57  *
  58  * The scheduler selects the next operation to issue by first looking for an
  59  * I/O class whose minimum has not been satisfied. Once all are satisfied and
  60  * the aggregate maximum has not been hit, the scheduler looks for classes
  61  * whose maximum has not been satisfied. Iteration through the I/O classes is
  62  * done in the order specified above. No further operations are issued if the
  63  * aggregate maximum number of concurrent operations has been hit or if there
  64  * are no operations queued for an I/O class that has not hit its maximum.
  65  * Every time an i/o is queued or an operation completes, the I/O scheduler
  66  * looks for new operations to issue.
  67  *
  68  * All I/O classes have a fixed maximum number of outstanding operations
  69  * except for the async write class. Asynchronous writes represent the data
  70  * that is committed to stable storage during the syncing stage for
  71  * transaction groups (see txg.c). Transaction groups enter the syncing state
  72  * periodically so the number of queued async writes will quickly burst up and
  73  * then bleed down to zero. Rather than servicing them as quickly as possible,
  74  * the I/O scheduler changes the maximum number of active async write i/os
  75  * according to the amount of dirty data in the pool (see dsl_pool.c). Since
  76  * both throughput and latency typically increase with the number of
  77  * concurrent operations issued to physical devices, reducing the burstiness
  78  * in the number of concurrent operations also stabilizes the response time of
  79  * operations from other -- and in particular synchronous -- queues. In broad
  80  * strokes, the I/O scheduler will issue more concurrent operations from the
  81  * async write queue as there's more dirty data in the pool.
  82  *
  83  * Async Writes
  84  *
  85  * The number of concurrent operations issued for the async write I/O class
  86  * follows a piece-wise linear function defined by a few adjustable points.
  87  *
  88  *        |                   o---------| <-- zfs_vdev_async_write_max_active
  89  *   ^    |                  /^         |
  90  *   |    |                 / |         |
  91  * active |                /  |         |
  92  *  I/O   |               /   |         |
  93  * count  |              /    |         |
  94  *        |             /     |         |
  95  *        |------------o      |         | <-- zfs_vdev_async_write_min_active
  96  *       0|____________^______|_________|
  97  *        0%           |      |       100% of zfs_dirty_data_max
  98  *                     |      |
  99  *                     |      `-- zfs_vdev_async_write_active_max_dirty_percent
 100  *                     `--------- zfs_vdev_async_write_active_min_dirty_percent
 101  *
 102  * Until the amount of dirty data exceeds a minimum percentage of the dirty
 103  * data allowed in the pool, the I/O scheduler will limit the number of
 104  * concurrent operations to the minimum. As that threshold is crossed, the
 105  * number of concurrent operations issued increases linearly to the maximum at
 106  * the specified maximum percentage of the dirty data allowed in the pool.
 107  *
 108  * Ideally, the amount of dirty data on a busy pool will stay in the sloped
 109  * part of the function between zfs_vdev_async_write_active_min_dirty_percent
 110  * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
 111  * maximum percentage, this indicates that the rate of incoming data is
 112  * greater than the rate that the backend storage can handle. In this case, we
 113  * must further throttle incoming writes (see dmu_tx_delay() for details).
 114  */
 115 
 116 /*
 117  * The maximum number of i/os active to each device.  Ideally, this will be >=
 118  * the sum of each queue's max_active.  It must be at least the sum of each
 119  * queue's min_active.
 120  */
 121 uint32_t zfs_vdev_max_active = 1000;
 122 
 123 /*
 124  * Per-queue limits on the number of i/os active to each device.  If the
 125  * sum of the queue's max_active is < zfs_vdev_max_active, then the
 126  * min_active comes into play.  We will send min_active from each queue,
 127  * and then select from queues in the order defined by zio_priority_t.
 128  *
 129  * In general, smaller max_active's will lead to lower latency of synchronous
 130  * operations.  Larger max_active's may lead to higher overall throughput,
 131  * depending on underlying storage.
 132  *
 133  * The ratio of the queues' max_actives determines the balance of performance
 134  * between reads, writes, and scrubs.  E.g., increasing
 135  * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
 136  * more quickly, but reads and writes to have higher latency and lower
 137  * throughput.
 138  */
 139 uint32_t zfs_vdev_sync_read_min_active = 10;
 140 uint32_t zfs_vdev_sync_read_max_active = 10;
 141 uint32_t zfs_vdev_sync_write_min_active = 10;
 142 uint32_t zfs_vdev_sync_write_max_active = 10;
 143 uint32_t zfs_vdev_async_read_min_active = 1;
 144 uint32_t zfs_vdev_async_read_max_active = 3;
 145 uint32_t zfs_vdev_async_write_min_active = 1;
 146 uint32_t zfs_vdev_async_write_max_active = 10;
 147 uint32_t zfs_vdev_scrub_min_active = 1;
 148 uint32_t zfs_vdev_scrub_max_active = 2;
 149 
 150 /*
 151  * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
 152  * dirty data, use zfs_vdev_async_write_min_active.  When it has more than
 153  * zfs_vdev_async_write_active_max_dirty_percent, use
 154  * zfs_vdev_async_write_max_active. The value is linearly interpolated
 155  * between min and max.
 156  */
 157 int zfs_vdev_async_write_active_min_dirty_percent = 30;
 158 int zfs_vdev_async_write_active_max_dirty_percent = 60;
 159 
 160 /*
 161  * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
 162  * For read I/Os, we also aggregate across small adjacency gaps; for writes
 163  * we include spans of optional I/Os to aid aggregation at the disk even when
 164  * they aren't able to help us aggregate at this level.
 165  */
 166 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
 167 int zfs_vdev_read_gap_limit = 32 << 10;
 168 int zfs_vdev_write_gap_limit = 4 << 10;
 169 
 170 int
 171 vdev_queue_offset_compare(const void *x1, const void *x2)
 172 {
 173         const zio_t *z1 = x1;
 174         const zio_t *z2 = x2;
 175 
 176         if (z1->io_offset < z2->io_offset)
 177                 return (-1);
 178         if (z1->io_offset > z2->io_offset)
 179                 return (1);
 180 
 181         if (z1 < z2)
 182                 return (-1);
 183         if (z1 > z2)
 184                 return (1);
 185 
 186         return (0);
 187 }
 188 
 189 int
 190 vdev_queue_timestamp_compare(const void *x1, const void *x2)
 191 {
 192         const zio_t *z1 = x1;
 193         const zio_t *z2 = x2;
 194 
 195         if (z1->io_timestamp < z2->io_timestamp)
 196                 return (-1);
 197         if (z1->io_timestamp > z2->io_timestamp)
 198                 return (1);
 199 
 200         if (z1 < z2)
 201                 return (-1);
 202         if (z1 > z2)
 203                 return (1);
 204 
 205         return (0);
 206 }
 207 
 208 void
 209 vdev_queue_init(vdev_t *vd)
 210 {
 211         vdev_queue_t *vq = &vd->vdev_queue;
 212 
 213         mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
 214         vq->vq_vdev = vd;
 215 
 216         avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
 217             sizeof (zio_t), offsetof(struct zio, io_queue_node));
 218 
 219         for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
 220                 /*
 221                  * The synchronous i/o queues are FIFO rather than LBA ordered.
 222                  * This provides more consistent latency for these i/os, and
 223                  * they tend to not be tightly clustered anyway so there is
 224                  * little to no throughput loss.
 225                  */
 226                 boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ ||
 227                     p == ZIO_PRIORITY_SYNC_WRITE);
 228                 avl_create(&vq->vq_class[p].vqc_queued_tree,
 229                     fifo ? vdev_queue_timestamp_compare :
 230                     vdev_queue_offset_compare,
 231                     sizeof (zio_t), offsetof(struct zio, io_queue_node));
 232         }
 233 
 234         vq->vq_last_queued_offset = 0;
 235 }
 236 
 237 void
 238 vdev_queue_fini(vdev_t *vd)
 239 {
 240         vdev_queue_t *vq = &vd->vdev_queue;
 241 
 242         for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
 243                 avl_destroy(&vq->vq_class[p].vqc_queued_tree);
 244         avl_destroy(&vq->vq_active_tree);
 245 
 246         mutex_destroy(&vq->vq_lock);
 247 }
 248 
 249 static void
 250 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
 251 {
 252         spa_t *spa = zio->io_spa;
 253         ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
 254         avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio);
 255 
 256         mutex_enter(&spa->spa_iokstat_lock);
 257         spa->spa_queue_stats[zio->io_priority].spa_queued++;
 258         if (spa->spa_iokstat != NULL)
 259                 kstat_waitq_enter(spa->spa_iokstat->ks_data);
 260         mutex_exit(&spa->spa_iokstat_lock);
 261 }
 262 
 263 static void
 264 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
 265 {
 266         spa_t *spa = zio->io_spa;
 267         ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
 268         avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio);
 269 
 270         mutex_enter(&spa->spa_iokstat_lock);
 271         ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
 272         spa->spa_queue_stats[zio->io_priority].spa_queued--;
 273         if (spa->spa_iokstat != NULL)
 274                 kstat_waitq_exit(spa->spa_iokstat->ks_data);
 275         mutex_exit(&spa->spa_iokstat_lock);
 276 }
 277 
 278 static void
 279 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
 280 {
 281         spa_t *spa = zio->io_spa;
 282         ASSERT(MUTEX_HELD(&vq->vq_lock));
 283         ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
 284         vq->vq_class[zio->io_priority].vqc_active++;
 285         avl_add(&vq->vq_active_tree, zio);
 286 
 287         mutex_enter(&spa->spa_iokstat_lock);
 288         spa->spa_queue_stats[zio->io_priority].spa_active++;
 289         if (spa->spa_iokstat != NULL)
 290                 kstat_runq_enter(spa->spa_iokstat->ks_data);
 291         mutex_exit(&spa->spa_iokstat_lock);
 292 }
 293 
 294 static void
 295 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
 296 {
 297         spa_t *spa = zio->io_spa;
 298         ASSERT(MUTEX_HELD(&vq->vq_lock));
 299         ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
 300         vq->vq_class[zio->io_priority].vqc_active--;
 301         avl_remove(&vq->vq_active_tree, zio);
 302 
 303         mutex_enter(&spa->spa_iokstat_lock);
 304         ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
 305         spa->spa_queue_stats[zio->io_priority].spa_active--;
 306         if (spa->spa_iokstat != NULL) {
 307                 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
 308 
 309                 kstat_runq_exit(spa->spa_iokstat->ks_data);
 310                 if (zio->io_type == ZIO_TYPE_READ) {
 311                         ksio->reads++;
 312                         ksio->nread += zio->io_size;
 313                 } else if (zio->io_type == ZIO_TYPE_WRITE) {
 314                         ksio->writes++;
 315                         ksio->nwritten += zio->io_size;
 316                 }
 317         }
 318         mutex_exit(&spa->spa_iokstat_lock);
 319 }
 320 
 321 static void
 322 vdev_queue_agg_io_done(zio_t *aio)
 323 {
 324         if (aio->io_type == ZIO_TYPE_READ) {
 325                 zio_t *pio;
 326                 while ((pio = zio_walk_parents(aio)) != NULL) {
 327                         bcopy((char *)aio->io_data + (pio->io_offset -
 328                             aio->io_offset), pio->io_data, pio->io_size);
 329                 }
 330         }
 331 
 332         zio_buf_free(aio->io_data, aio->io_size);
 333 }
 334 
 335 static int
 336 vdev_queue_class_min_active(zio_priority_t p)
 337 {
 338         switch (p) {
 339         case ZIO_PRIORITY_SYNC_READ:
 340                 return (zfs_vdev_sync_read_min_active);
 341         case ZIO_PRIORITY_SYNC_WRITE:
 342                 return (zfs_vdev_sync_write_min_active);
 343         case ZIO_PRIORITY_ASYNC_READ:
 344                 return (zfs_vdev_async_read_min_active);
 345         case ZIO_PRIORITY_ASYNC_WRITE:
 346                 return (zfs_vdev_async_write_min_active);
 347         case ZIO_PRIORITY_SCRUB:
 348                 return (zfs_vdev_scrub_min_active);
 349         default:
 350                 panic("invalid priority %u", p);
 351                 return (0);
 352         }
 353 }
 354 
 355 static int
 356 vdev_queue_max_async_writes(uint64_t dirty)
 357 {
 358         int writes;
 359         uint64_t min_bytes = zfs_dirty_data_max *
 360             zfs_vdev_async_write_active_min_dirty_percent / 100;
 361         uint64_t max_bytes = zfs_dirty_data_max *
 362             zfs_vdev_async_write_active_max_dirty_percent / 100;
 363 
 364         if (dirty < min_bytes)
 365                 return (zfs_vdev_async_write_min_active);
 366         if (dirty > max_bytes)
 367                 return (zfs_vdev_async_write_max_active);
 368 
 369         /*
 370          * linear interpolation:
 371          * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
 372          * move right by min_bytes
 373          * move up by min_writes
 374          */
 375         writes = (dirty - min_bytes) *
 376             (zfs_vdev_async_write_max_active -
 377             zfs_vdev_async_write_min_active) /
 378             (max_bytes - min_bytes) +
 379             zfs_vdev_async_write_min_active;
 380         ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
 381         ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
 382         return (writes);
 383 }
 384 
 385 static int
 386 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
 387 {
 388         switch (p) {
 389         case ZIO_PRIORITY_SYNC_READ:
 390                 return (zfs_vdev_sync_read_max_active);
 391         case ZIO_PRIORITY_SYNC_WRITE:
 392                 return (zfs_vdev_sync_write_max_active);
 393         case ZIO_PRIORITY_ASYNC_READ:
 394                 return (zfs_vdev_async_read_max_active);
 395         case ZIO_PRIORITY_ASYNC_WRITE:
 396                 return (vdev_queue_max_async_writes(
 397                     spa->spa_dsl_pool->dp_dirty_total));
 398         case ZIO_PRIORITY_SCRUB:
 399                 return (zfs_vdev_scrub_max_active);
 400         default:
 401                 panic("invalid priority %u", p);
 402                 return (0);
 403         }
 404 }
 405 
 406 /*
 407  * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
 408  * there is no eligible class.
 409  */
 410 static zio_priority_t
 411 vdev_queue_class_to_issue(vdev_queue_t *vq)
 412 {
 413         spa_t *spa = vq->vq_vdev->vdev_spa;
 414         zio_priority_t p;
 415 
 416         if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
 417                 return (ZIO_PRIORITY_NUM_QUEUEABLE);
 418 
 419         /* find a queue that has not reached its minimum # outstanding i/os */
 420         for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
 421                 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 &&
 422                     vq->vq_class[p].vqc_active <
 423                     vdev_queue_class_min_active(p))
 424                         return (p);
 425         }
 426 
 427         /*
 428          * If we haven't found a queue, look for one that hasn't reached its
 429          * maximum # outstanding i/os.
 430          */
 431         for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
 432                 if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 &&
 433                     vq->vq_class[p].vqc_active <
 434                     vdev_queue_class_max_active(spa, p))
 435                         return (p);
 436         }
 437 
 438         /* No eligible queued i/os */
 439         return (ZIO_PRIORITY_NUM_QUEUEABLE);
 440 }
 441 
 442 /*
 443  * Compute the range spanned by two i/os, which is the endpoint of the last
 444  * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
 445  * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
 446  * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
 447  */
 448 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
 449 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
 450 
 451 static zio_t *
 452 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
 453 {
 454         zio_t *first, *last, *aio, *dio, *mandatory, *nio;
 455         uint64_t maxgap = 0;
 456         uint64_t size;
 457         boolean_t stretch = B_FALSE;
 458         vdev_queue_class_t *vqc = &vq->vq_class[zio->io_priority];
 459         avl_tree_t *t = &vqc->vqc_queued_tree;
 460         enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
 461 
 462         if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
 463                 return (NULL);
 464 
 465         /*
 466          * The synchronous i/o queues are not sorted by LBA, so we can't
 467          * find adjacent i/os.  These i/os tend to not be tightly clustered,
 468          * or too large to aggregate, so this has little impact on performance.
 469          */
 470         if (zio->io_priority == ZIO_PRIORITY_SYNC_READ ||
 471             zio->io_priority == ZIO_PRIORITY_SYNC_WRITE)
 472                 return (NULL);
 473 
 474         first = last = zio;
 475 
 476         if (zio->io_type == ZIO_TYPE_READ)
 477                 maxgap = zfs_vdev_read_gap_limit;
 478 
 479         /*
 480          * We can aggregate I/Os that are sufficiently adjacent and of
 481          * the same flavor, as expressed by the AGG_INHERIT flags.
 482          * The latter requirement is necessary so that certain
 483          * attributes of the I/O, such as whether it's a normal I/O
 484          * or a scrub/resilver, can be preserved in the aggregate.
 485          * We can include optional I/Os, but don't allow them
 486          * to begin a range as they add no benefit in that situation.
 487          */
 488 
 489         /*
 490          * We keep track of the last non-optional I/O.
 491          */
 492         mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
 493 
 494         /*
 495          * Walk backwards through sufficiently contiguous I/Os
 496          * recording the last non-option I/O.
 497          */
 498         while ((dio = AVL_PREV(t, first)) != NULL &&
 499             (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
 500             IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
 501             IO_GAP(dio, first) <= maxgap) {
 502                 first = dio;
 503                 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
 504                         mandatory = first;
 505         }
 506 
 507         /*
 508          * Skip any initial optional I/Os.
 509          */
 510         while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
 511                 first = AVL_NEXT(t, first);
 512                 ASSERT(first != NULL);
 513         }
 514 
 515         /*
 516          * Walk forward through sufficiently contiguous I/Os.
 517          */
 518         while ((dio = AVL_NEXT(t, last)) != NULL &&
 519             (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
 520             IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit &&
 521             IO_GAP(last, dio) <= maxgap) {
 522                 last = dio;
 523                 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
 524                         mandatory = last;
 525         }
 526 
 527         /*
 528          * Now that we've established the range of the I/O aggregation
 529          * we must decide what to do with trailing optional I/Os.
 530          * For reads, there's nothing to do. While we are unable to
 531          * aggregate further, it's possible that a trailing optional
 532          * I/O would allow the underlying device to aggregate with
 533          * subsequent I/Os. We must therefore determine if the next
 534          * non-optional I/O is close enough to make aggregation
 535          * worthwhile.
 536          */
 537         if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
 538                 zio_t *nio = last;
 539                 while ((dio = AVL_NEXT(t, nio)) != NULL &&
 540                     IO_GAP(nio, dio) == 0 &&
 541                     IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
 542                         nio = dio;
 543                         if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
 544                                 stretch = B_TRUE;
 545                                 break;
 546                         }
 547                 }
 548         }
 549 
 550         if (stretch) {
 551                 /* This may be a no-op. */
 552                 dio = AVL_NEXT(t, last);
 553                 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
 554         } else {
 555                 while (last != mandatory && last != first) {
 556                         ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
 557                         last = AVL_PREV(t, last);
 558                         ASSERT(last != NULL);
 559                 }
 560         }
 561 
 562         if (first == last)
 563                 return (NULL);
 564 
 565         size = IO_SPAN(first, last);
 566         ASSERT3U(size, <=, zfs_vdev_aggregation_limit);
 567 
 568         aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
 569             zio_buf_alloc(size), size, first->io_type, zio->io_priority,
 570             flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
 571             vdev_queue_agg_io_done, NULL);
 572         aio->io_timestamp = first->io_timestamp;
 573 
 574         nio = first;
 575         do {
 576                 dio = nio;
 577                 nio = AVL_NEXT(t, dio);
 578                 ASSERT3U(dio->io_type, ==, aio->io_type);
 579 
 580                 if (dio->io_flags & ZIO_FLAG_NODATA) {
 581                         ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
 582                         bzero((char *)aio->io_data + (dio->io_offset -
 583                             aio->io_offset), dio->io_size);
 584                 } else if (dio->io_type == ZIO_TYPE_WRITE) {
 585                         bcopy(dio->io_data, (char *)aio->io_data +
 586                             (dio->io_offset - aio->io_offset),
 587                             dio->io_size);
 588                 }
 589 
 590                 zio_add_child(dio, aio);
 591                 vdev_queue_io_remove(vq, dio);
 592                 zio_vdev_io_bypass(dio);
 593                 zio_execute(dio);
 594         } while (dio != last);
 595 
 596         return (aio);
 597 }
 598 
 599 static zio_t *
 600 vdev_queue_io_to_issue(vdev_queue_t *vq)
 601 {
 602         zio_t *zio, *aio;
 603         zio_priority_t p;
 604         avl_index_t idx;
 605         vdev_queue_class_t *vqc;
 606         zio_t search;
 607 
 608 again:
 609         ASSERT(MUTEX_HELD(&vq->vq_lock));
 610 
 611         p = vdev_queue_class_to_issue(vq);
 612 
 613         if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
 614                 /* No eligible queued i/os */
 615                 return (NULL);
 616         }
 617 
 618         /*
 619          * For LBA-ordered queues (async / scrub), issue the i/o which follows
 620          * the most recently issued i/o in LBA (offset) order.
 621          *
 622          * For FIFO queues (sync), issue the i/o with the lowest timestamp.
 623          */
 624         vqc = &vq->vq_class[p];
 625         search.io_timestamp = 0;
 626         search.io_offset = vq->vq_last_offset + 1;
 627         VERIFY3P(avl_find(&vqc->vqc_queued_tree, &search, &idx), ==, NULL);
 628         zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER);
 629         if (zio == NULL)
 630                 zio = avl_first(&vqc->vqc_queued_tree);
 631         ASSERT3U(zio->io_priority, ==, p);
 632 
 633         aio = vdev_queue_aggregate(vq, zio);
 634         if (aio != NULL)
 635                 zio = aio;
 636         else
 637                 vdev_queue_io_remove(vq, zio);
 638 
 639         /*
 640          * If the I/O is or was optional and therefore has no data, we need to
 641          * simply discard it. We need to drop the vdev queue's lock to avoid a
 642          * deadlock that we could encounter since this I/O will complete
 643          * immediately.
 644          */
 645         if (zio->io_flags & ZIO_FLAG_NODATA) {
 646                 mutex_exit(&vq->vq_lock);
 647                 zio_vdev_io_bypass(zio);
 648                 zio_execute(zio);
 649                 mutex_enter(&vq->vq_lock);
 650                 goto again;
 651         }
 652 
 653         vdev_queue_pending_add(vq, zio);
 654         vq->vq_last_offset = zio->io_offset;
 655 
 656         return (zio);
 657 }
 658 
 659 zio_t *
 660 vdev_queue_io(zio_t *zio)
 661 {
 662         vdev_queue_t *vq = &zio->io_vd->vdev_queue;
 663         zio_t *nio;
 664 
 665         if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
 666                 return (zio);
 667 
 668         /*
 669          * Children i/os inherent their parent's priority, which might
 670          * not match the child's i/o type.  Fix it up here.
 671          */
 672         if (zio->io_type == ZIO_TYPE_READ) {
 673                 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
 674                     zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
 675                     zio->io_priority != ZIO_PRIORITY_SCRUB)
 676                         zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
 677         } else {
 678                 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
 679                 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
 680                     zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE)
 681                         zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
 682         }
 683 
 684         zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
 685 
 686         mutex_enter(&vq->vq_lock);
 687         zio->io_timestamp = gethrtime();
 688         vdev_queue_io_add(vq, zio);
 689         nio = vdev_queue_io_to_issue(vq);
 690         mutex_exit(&vq->vq_lock);
 691 
 692         if (nio == NULL)
 693                 return (NULL);
 694 
 695         if (nio->io_done == vdev_queue_agg_io_done) {
 696                 zio_nowait(nio);
 697                 return (NULL);
 698         }
 699 
 700         return (nio);
 701 }
 702 
 703 void
 704 vdev_queue_io_done(zio_t *zio)
 705 {
 706         vdev_queue_t *vq = &zio->io_vd->vdev_queue;
 707         zio_t *nio;
 708 
 709         if (zio_injection_enabled)
 710                 delay(SEC_TO_TICK(zio_handle_io_delay(zio)));
 711 
 712         mutex_enter(&vq->vq_lock);
 713 
 714         vdev_queue_pending_remove(vq, zio);
 715 
 716         vq->vq_io_complete_ts = gethrtime();
 717 
 718         while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
 719                 mutex_exit(&vq->vq_lock);
 720                 if (nio->io_done == vdev_queue_agg_io_done) {
 721                         zio_nowait(nio);
 722                 } else {
 723                         zio_vdev_io_reissue(nio);
 724                         zio_execute(nio);
 725                 }
 726                 mutex_enter(&vq->vq_lock);
 727         }
 728 
 729         mutex_exit(&vq->vq_lock);
 730 }
 731 
 732 /*
 733  * As these three methods are only used for load calculations we're not
 734  * concerned if we get an incorrect value on 32bit platforms due to lack of
 735  * vq_lock mutex use here, instead we prefer to keep it lock free for
 736  * performance.
 737  */
 738 int
 739 vdev_queue_length(vdev_t *vd)
 740 {
 741         vdev_queue_t *vq = &vd->vdev_queue;
 742         int len = 0;
 743 
 744         for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
 745                 len += avl_numnodes(&vq->vq_class[p].vqc_queued_tree);
 746 
 747         return (len);
 748 }
 749 
 750 uint64_t
 751 vdev_queue_last_queued_offset(vdev_t *vd)
 752 {
 753         return (vd->vdev_queue.vq_last_queued_offset);
 754 }
 755 
 756 void
 757 vdev_queue_register_last_queued_offset(vdev_t *vd, zio_t *zio)
 758 {
 759         vd->vdev_queue.vq_last_queued_offset = zio->io_offset + zio->io_size;
 760 }