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 2009 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25
  26 /*
  27  * Copyright (c) 2012 by Delphix. All rights reserved.
  28  */
  29
  30 #include <sys/zfs_context.h>
  31 #include <sys/vdev_impl.h>
  32 #include <sys/zio.h>
  33 #include <sys/avl.h>
  34
  35 /*
  36  * These tunables are for performance analysis.
  37  */
  38 /*
  39  * zfs_vdev_max_pending is the maximum number of i/os concurrently
  40  * pending to each device.  zfs_vdev_min_pending is the initial number
  41  * of i/os pending to each device (before it starts ramping up to
  42  * max_pending).
  43  */
  44 int zfs_vdev_max_pending = 10;
  45 int zfs_vdev_min_pending = 4;
  46
  47 /*
  48  * The deadlines are grouped into buckets based on zfs_vdev_time_shift:
  49  * deadline = pri + gethrtime() >> time_shift)
  50  */
  51 int zfs_vdev_time_shift = 29; /* each bucket is 0.537 seconds */
  52
  53 /* exponential I/O issue ramp-up rate */
  54 int zfs_vdev_ramp_rate = 2;
  55
  56 /*
  57  * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
  58  * For read I/Os, we also aggregate across small adjacency gaps; for writes
  59  * we include spans of optional I/Os to aid aggregation at the disk even when
  60  * they aren't able to help us aggregate at this level.
  61  */
  62 int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
  63 int zfs_vdev_read_gap_limit = 32 << 10;
  64 int zfs_vdev_write_gap_limit = 4 << 10;
  65
  66 /*
  67  * Virtual device vector for disk I/O scheduling.
  68  */
  69 int
  70 vdev_queue_deadline_compare(const void *x1, const void *x2)
  71 {
  72         const zio_t *z1 = x1;
  73         const zio_t *z2 = x2;
  74
  75         if (z1->io_deadline < z2->io_deadline)
  76                 return (-1);
  77         if (z1->io_deadline > z2->io_deadline)
  78                 return (1);
  79
  80         if (z1->io_offset < z2->io_offset)
  81                 return (-1);
  82         if (z1->io_offset > z2->io_offset)
  83                 return (1);
  84
  85         if (z1 < z2)
  86                 return (-1);
  87         if (z1 > z2)
  88                 return (1);
  89
  90         return (0);
  91 }
  92
  93 int
  94 vdev_queue_offset_compare(const void *x1, const void *x2)
  95 {
  96         const zio_t *z1 = x1;
  97         const zio_t *z2 = x2;
  98
  99         if (z1->io_offset < z2->io_offset)
 100                 return (-1);
 101         if (z1->io_offset > z2->io_offset)
 102                 return (1);
 103
 104         if (z1 < z2)
 105                 return (-1);
 106         if (z1 > z2)
 107                 return (1);
 108
 109         return (0);
 110 }
 111
 112 void
 113 vdev_queue_init(vdev_t *vd)
 114 {
 115         vdev_queue_t *vq = &vd->vdev_queue;
 116         int i;
 117
 118         mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
 119
 120         avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare,
 121             sizeof (zio_t), offsetof(struct zio, io_deadline_node));
 122
 123         avl_create(&vq->vq_read_tree, vdev_queue_offset_compare,
 124             sizeof (zio_t), offsetof(struct zio, io_offset_node));
 125
 126         avl_create(&vq->vq_write_tree, vdev_queue_offset_compare,
 127             sizeof (zio_t), offsetof(struct zio, io_offset_node));
 128
 129         avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare,
 130             sizeof (zio_t), offsetof(struct zio, io_offset_node));
 131
 132         /*
 133          * A list of buffers which can be used for aggregate I/O, this
 134          * avoids the need to allocate them on demand when memory is low.
 135          */
 136         list_create(&vq->vq_io_list, sizeof (vdev_io_t),
 137             offsetof(vdev_io_t, vi_node));
 138
 139         for (i = 0; i < zfs_vdev_max_pending; i++)
 140                 list_insert_tail(&vq->vq_io_list, zio_vdev_alloc());
 141 }
 142
 143 void
 144 vdev_queue_fini(vdev_t *vd)
 145 {
 146         vdev_queue_t *vq = &vd->vdev_queue;
 147         vdev_io_t *vi;
 148
 149         avl_destroy(&vq->vq_deadline_tree);
 150         avl_destroy(&vq->vq_read_tree);
 151         avl_destroy(&vq->vq_write_tree);
 152         avl_destroy(&vq->vq_pending_tree);
 153
 154         while ((vi = list_head(&vq->vq_io_list)) != NULL) {
 155                 list_remove(&vq->vq_io_list, vi);
 156                 zio_vdev_free(vi);
 157         }
 158
 159         list_destroy(&vq->vq_io_list);
 160
 161         mutex_destroy(&vq->vq_lock);
 162 }
 163
 164 static void
 165 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
 166 {
 167         avl_add(&vq->vq_deadline_tree, zio);
 168         avl_add(zio->io_vdev_tree, zio);
 169 }
 170
 171 static void
 172 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
 173 {
 174         avl_remove(&vq->vq_deadline_tree, zio);
 175         avl_remove(zio->io_vdev_tree, zio);
 176 }
 177
 178 static void
 179 vdev_queue_agg_io_done(zio_t *aio)
 180 {
 181         vdev_queue_t *vq = &aio->io_vd->vdev_queue;
 182         vdev_io_t *vi = aio->io_data;
 183         zio_t *pio;
 184
 185         while ((pio = zio_walk_parents(aio)) != NULL)
 186                 if (aio->io_type == ZIO_TYPE_READ)
 187                         bcopy((char *)aio->io_data + (pio->io_offset -
 188                             aio->io_offset), pio->io_data, pio->io_size);
 189
 190         mutex_enter(&vq->vq_lock);
 191         list_insert_tail(&vq->vq_io_list, vi);
 192         mutex_exit(&vq->vq_lock);
 193 }
 194
 195 /*
 196  * Compute the range spanned by two i/os, which is the endpoint of the last
 197  * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
 198  * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
 199  * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
 200  */
 201 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
 202 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
 203
 204 static zio_t *
 205 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit)
 206 {
 207         zio_t *fio, *lio, *aio, *dio, *nio, *mio;
 208         avl_tree_t *t;
 209         vdev_io_t *vi;
 210         int flags;
 211         uint64_t maxspan = MIN(zfs_vdev_aggregation_limit, SPA_MAXBLOCKSIZE);
 212         uint64_t maxgap;
 213         int stretch;
 214
 215 again:
 216         ASSERT(MUTEX_HELD(&vq->vq_lock));
 217
 218         if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit ||
 219             avl_numnodes(&vq->vq_deadline_tree) == 0)
 220                 return (NULL);
 221
 222         fio = lio = avl_first(&vq->vq_deadline_tree);
 223
 224         t = fio->io_vdev_tree;
 225         flags = fio->io_flags & ZIO_FLAG_AGG_INHERIT;
 226         maxgap = (t == &vq->vq_read_tree) ? zfs_vdev_read_gap_limit : 0;
 227
 228         vi = list_head(&vq->vq_io_list);
 229         if (vi == NULL) {
 230                 vi = zio_vdev_alloc();
 231                 list_insert_head(&vq->vq_io_list, vi);
 232         }
 233
 234         if (!(flags & ZIO_FLAG_DONT_AGGREGATE)) {
 235                 /*
 236                  * We can aggregate I/Os that are sufficiently adjacent and of
 237                  * the same flavor, as expressed by the AGG_INHERIT flags.
 238                  * The latter requirement is necessary so that certain
 239                  * attributes of the I/O, such as whether it's a normal I/O
 240                  * or a scrub/resilver, can be preserved in the aggregate.
 241                  * We can include optional I/Os, but don't allow them
 242                  * to begin a range as they add no benefit in that situation.
 243                  */
 244
 245                 /*
 246                  * We keep track of the last non-optional I/O.
 247                  */
 248                 mio = (fio->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : fio;
 249
 250                 /*
 251                  * Walk backwards through sufficiently contiguous I/Os
 252                  * recording the last non-option I/O.
 253                  */
 254                 while ((dio = AVL_PREV(t, fio)) != NULL &&
 255                     (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
 256                     IO_SPAN(dio, lio) <= maxspan &&
 257                     IO_GAP(dio, fio) <= maxgap) {
 258                         fio = dio;
 259                         if (mio == NULL && !(fio->io_flags & ZIO_FLAG_OPTIONAL))
 260                                 mio = fio;
 261                 }
 262
 263                 /*
 264                  * Skip any initial optional I/Os.
 265                  */
 266                 while ((fio->io_flags & ZIO_FLAG_OPTIONAL) && fio != lio) {
 267                         fio = AVL_NEXT(t, fio);
 268                         ASSERT(fio != NULL);
 269                 }
 270
 271                 /*
 272                  * Walk forward through sufficiently contiguous I/Os.
 273                  */
 274                 while ((dio = AVL_NEXT(t, lio)) != NULL &&
 275                     (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
 276                     IO_SPAN(fio, dio) <= maxspan &&
 277                     IO_GAP(lio, dio) <= maxgap) {
 278                         lio = dio;
 279                         if (!(lio->io_flags & ZIO_FLAG_OPTIONAL))
 280                                 mio = lio;
 281                 }
 282
 283                 /*
 284                  * Now that we've established the range of the I/O aggregation
 285                  * we must decide what to do with trailing optional I/Os.
 286                  * For reads, there's nothing to do. While we are unable to
 287                  * aggregate further, it's possible that a trailing optional
 288                  * I/O would allow the underlying device to aggregate with
 289                  * subsequent I/Os. We must therefore determine if the next
 290                  * non-optional I/O is close enough to make aggregation
 291                  * worthwhile.
 292                  */
 293                 stretch = B_FALSE;
 294                 if (t != &vq->vq_read_tree && mio != NULL) {
 295                         nio = lio;
 296                         while ((dio = AVL_NEXT(t, nio)) != NULL &&
 297                             IO_GAP(nio, dio) == 0 &&
 298                             IO_GAP(mio, dio) <= zfs_vdev_write_gap_limit) {
 299                                 nio = dio;
 300                                 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
 301                                         stretch = B_TRUE;
 302                                         break;
 303                                 }
 304                         }
 305                 }
 306
 307                 if (stretch) {
 308                         /* This may be a no-op. */
 309                         VERIFY((dio = AVL_NEXT(t, lio)) != NULL);
 310                         dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
 311                 } else {
 312                         while (lio != mio && lio != fio) {
 313                                 ASSERT(lio->io_flags & ZIO_FLAG_OPTIONAL);
 314                                 lio = AVL_PREV(t, lio);
 315                                 ASSERT(lio != NULL);
 316                         }
 317                 }
 318         }
 319
 320         if (fio != lio) {
 321                 uint64_t size = IO_SPAN(fio, lio);
 322                 ASSERT(size <= maxspan);
 323                 ASSERT(vi != NULL);
 324
 325                 aio = zio_vdev_delegated_io(fio->io_vd, fio->io_offset,
 326                     vi, size, fio->io_type, ZIO_PRIORITY_AGG,
 327                     flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
 328                     vdev_queue_agg_io_done, NULL);
 329                 aio->io_timestamp = fio->io_timestamp;
 330
 331                 nio = fio;
 332                 do {
 333                         dio = nio;
 334                         nio = AVL_NEXT(t, dio);
 335                         ASSERT(dio->io_type == aio->io_type);
 336                         ASSERT(dio->io_vdev_tree == t);
 337
 338                         if (dio->io_flags & ZIO_FLAG_NODATA) {
 339                                 ASSERT(dio->io_type == ZIO_TYPE_WRITE);
 340                                 bzero((char *)aio->io_data + (dio->io_offset -
 341                                     aio->io_offset), dio->io_size);
 342                         } else if (dio->io_type == ZIO_TYPE_WRITE) {
 343                                 bcopy(dio->io_data, (char *)aio->io_data +
 344                                     (dio->io_offset - aio->io_offset),
 345                                     dio->io_size);
 346                         }
 347
 348                         zio_add_child(dio, aio);
 349                         vdev_queue_io_remove(vq, dio);
 350                         zio_vdev_io_bypass(dio);
 351                         zio_execute(dio);
 352                 } while (dio != lio);
 353
 354                 avl_add(&vq->vq_pending_tree, aio);
 355                 list_remove(&vq->vq_io_list, vi);
 356
 357                 return (aio);
 358         }
 359
 360         ASSERT(fio->io_vdev_tree == t);
 361         vdev_queue_io_remove(vq, fio);
 362
 363         /*
 364          * If the I/O is or was optional and therefore has no data, we need to
 365          * simply discard it. We need to drop the vdev queue's lock to avoid a
 366          * deadlock that we could encounter since this I/O will complete
 367          * immediately.
 368          */
 369         if (fio->io_flags & ZIO_FLAG_NODATA) {
 370                 mutex_exit(&vq->vq_lock);
 371                 zio_vdev_io_bypass(fio);
 372                 zio_execute(fio);
 373                 mutex_enter(&vq->vq_lock);
 374                 goto again;
 375         }
 376
 377         avl_add(&vq->vq_pending_tree, fio);
 378
 379         return (fio);
 380 }
 381
 382 zio_t *
 383 vdev_queue_io(zio_t *zio)
 384 {
 385         vdev_queue_t *vq = &zio->io_vd->vdev_queue;
 386         zio_t *nio;
 387
 388         ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
 389
 390         if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
 391                 return (zio);
 392
 393         zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
 394
 395         if (zio->io_type == ZIO_TYPE_READ)
 396                 zio->io_vdev_tree = &vq->vq_read_tree;
 397         else
 398                 zio->io_vdev_tree = &vq->vq_write_tree;
 399
 400         mutex_enter(&vq->vq_lock);
 401
 402         zio->io_timestamp = gethrtime();
 403         zio->io_deadline = (zio->io_timestamp >> zfs_vdev_time_shift) +
 404             zio->io_priority;
 405
 406         vdev_queue_io_add(vq, zio);
 407
 408         nio = vdev_queue_io_to_issue(vq, zfs_vdev_min_pending);
 409
 410         mutex_exit(&vq->vq_lock);
 411
 412         if (nio == NULL)
 413                 return (NULL);
 414
 415         if (nio->io_done == vdev_queue_agg_io_done) {
 416                 zio_nowait(nio);
 417                 return (NULL);
 418         }
 419
 420         return (nio);
 421 }
 422
 423 void
 424 vdev_queue_io_done(zio_t *zio)
 425 {
 426         vdev_queue_t *vq = &zio->io_vd->vdev_queue;
 427         int i;
 428
 429         if (zio_injection_enabled)
 430                 delay(SEC_TO_TICK(zio_handle_io_delay(zio)));
 431
 432         mutex_enter(&vq->vq_lock);
 433
 434         avl_remove(&vq->vq_pending_tree, zio);
 435
 436         zio->io_delta = gethrtime() - zio->io_timestamp;
 437         vq->vq_io_complete_ts = gethrtime();
 438         vq->vq_io_delta_ts = vq->vq_io_complete_ts - zio->io_timestamp;
 439
 440         for (i = 0; i < zfs_vdev_ramp_rate; i++) {
 441                 zio_t *nio = vdev_queue_io_to_issue(vq, zfs_vdev_max_pending);
 442                 if (nio == NULL)
 443                         break;
 444                 mutex_exit(&vq->vq_lock);
 445                 if (nio->io_done == vdev_queue_agg_io_done) {
 446                         zio_nowait(nio);
 447                 } else {
 448                         zio_vdev_io_reissue(nio);
 449                         zio_execute(nio);
 450                 }
 451                 mutex_enter(&vq->vq_lock);
 452         }
 453
 454         mutex_exit(&vq->vq_lock);
 455 }
 456
 457 #if defined(_KERNEL) && defined(HAVE_SPL)
 458 module_param(zfs_vdev_max_pending, int, 0644);
 459 MODULE_PARM_DESC(zfs_vdev_max_pending, "Max pending per-vdev I/Os");
 460
 461 module_param(zfs_vdev_min_pending, int, 0644);
 462 MODULE_PARM_DESC(zfs_vdev_min_pending, "Min pending per-vdev I/Os");
 463
 464 module_param(zfs_vdev_aggregation_limit, int, 0644);
 465 MODULE_PARM_DESC(zfs_vdev_aggregation_limit, "Max vdev I/O aggregation size");
 466
 467 module_param(zfs_vdev_time_shift, int, 0644);
 468 MODULE_PARM_DESC(zfs_vdev_time_shift, "Deadline time shift for vdev I/O");
 469
 470 module_param(zfs_vdev_ramp_rate, int, 0644);
 471 MODULE_PARM_DESC(zfs_vdev_ramp_rate, "Exponential I/O issue ramp-up rate");
 472
 473 module_param(zfs_vdev_read_gap_limit, int, 0644);
 474 MODULE_PARM_DESC(zfs_vdev_read_gap_limit, "Aggregate read I/O over gap");
 475
 476 module_param(zfs_vdev_write_gap_limit, int, 0644);
 477 MODULE_PARM_DESC(zfs_vdev_write_gap_limit, "Aggregate write I/O over gap");
 478 #endif