1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5ea8dc4b6Seschrock * Common Development and Distribution License (the "License"). 6ea8dc4b6Seschrock * You may not use this file except in compliance with the License. 7fa9e4066Sahrens * 8fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 10fa9e4066Sahrens * See the License for the specific language governing permissions 11fa9e4066Sahrens * and limitations under the License. 12fa9e4066Sahrens * 13fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 14fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 16fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 17fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18fa9e4066Sahrens * 19fa9e4066Sahrens * CDDL HEADER END 20fa9e4066Sahrens */ 21fa9e4066Sahrens /* 22a3f829aeSBill Moore * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23fa9e4066Sahrens * Use is subject to license terms. 24fa9e4066Sahrens */ 25fa9e4066Sahrens 26283b8460SGeorge.Wilson /* 27*73527f44SAlex Reece * Copyright (c) 2012, 2014 by Delphix. All rights reserved. 28283b8460SGeorge.Wilson */ 29283b8460SGeorge.Wilson 30fa9e4066Sahrens #include <sys/zfs_context.h> 31fa9e4066Sahrens #include <sys/vdev_impl.h> 32c3a66015SMatthew Ahrens #include <sys/spa_impl.h> 33fa9e4066Sahrens #include <sys/zio.h> 34fa9e4066Sahrens #include <sys/avl.h> 3569962b56SMatthew Ahrens #include <sys/dsl_pool.h> 36fa9e4066Sahrens 37614409b5Sahrens /* 3869962b56SMatthew Ahrens * ZFS I/O Scheduler 3969962b56SMatthew Ahrens * --------------- 4069962b56SMatthew Ahrens * 4169962b56SMatthew Ahrens * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The 4269962b56SMatthew Ahrens * I/O scheduler determines when and in what order those operations are 4369962b56SMatthew Ahrens * issued. The I/O scheduler divides operations into five I/O classes 4469962b56SMatthew Ahrens * prioritized in the following order: sync read, sync write, async read, 4569962b56SMatthew Ahrens * async write, and scrub/resilver. Each queue defines the minimum and 4669962b56SMatthew Ahrens * maximum number of concurrent operations that may be issued to the device. 4769962b56SMatthew Ahrens * In addition, the device has an aggregate maximum. Note that the sum of the 4869962b56SMatthew Ahrens * per-queue minimums must not exceed the aggregate maximum, and if the 4969962b56SMatthew Ahrens * aggregate maximum is equal to or greater than the sum of the per-queue 5069962b56SMatthew Ahrens * maximums, the per-queue minimum has no effect. 5169962b56SMatthew Ahrens * 5269962b56SMatthew Ahrens * For many physical devices, throughput increases with the number of 5369962b56SMatthew Ahrens * concurrent operations, but latency typically suffers. Further, physical 5469962b56SMatthew Ahrens * devices typically have a limit at which more concurrent operations have no 5569962b56SMatthew Ahrens * effect on throughput or can actually cause it to decrease. 5669962b56SMatthew Ahrens * 5769962b56SMatthew Ahrens * The scheduler selects the next operation to issue by first looking for an 5869962b56SMatthew Ahrens * I/O class whose minimum has not been satisfied. Once all are satisfied and 5969962b56SMatthew Ahrens * the aggregate maximum has not been hit, the scheduler looks for classes 6069962b56SMatthew Ahrens * whose maximum has not been satisfied. Iteration through the I/O classes is 6169962b56SMatthew Ahrens * done in the order specified above. No further operations are issued if the 6269962b56SMatthew Ahrens * aggregate maximum number of concurrent operations has been hit or if there 6369962b56SMatthew Ahrens * are no operations queued for an I/O class that has not hit its maximum. 6469962b56SMatthew Ahrens * Every time an i/o is queued or an operation completes, the I/O scheduler 6569962b56SMatthew Ahrens * looks for new operations to issue. 6669962b56SMatthew Ahrens * 6769962b56SMatthew Ahrens * All I/O classes have a fixed maximum number of outstanding operations 6869962b56SMatthew Ahrens * except for the async write class. Asynchronous writes represent the data 6969962b56SMatthew Ahrens * that is committed to stable storage during the syncing stage for 7069962b56SMatthew Ahrens * transaction groups (see txg.c). Transaction groups enter the syncing state 7169962b56SMatthew Ahrens * periodically so the number of queued async writes will quickly burst up and 7269962b56SMatthew Ahrens * then bleed down to zero. Rather than servicing them as quickly as possible, 7369962b56SMatthew Ahrens * the I/O scheduler changes the maximum number of active async write i/os 7469962b56SMatthew Ahrens * according to the amount of dirty data in the pool (see dsl_pool.c). Since 7569962b56SMatthew Ahrens * both throughput and latency typically increase with the number of 7669962b56SMatthew Ahrens * concurrent operations issued to physical devices, reducing the burstiness 7769962b56SMatthew Ahrens * in the number of concurrent operations also stabilizes the response time of 7869962b56SMatthew Ahrens * operations from other -- and in particular synchronous -- queues. In broad 7969962b56SMatthew Ahrens * strokes, the I/O scheduler will issue more concurrent operations from the 8069962b56SMatthew Ahrens * async write queue as there's more dirty data in the pool. 8169962b56SMatthew Ahrens * 8269962b56SMatthew Ahrens * Async Writes 8369962b56SMatthew Ahrens * 8469962b56SMatthew Ahrens * The number of concurrent operations issued for the async write I/O class 8569962b56SMatthew Ahrens * follows a piece-wise linear function defined by a few adjustable points. 8669962b56SMatthew Ahrens * 8769962b56SMatthew Ahrens * | o---------| <-- zfs_vdev_async_write_max_active 8869962b56SMatthew Ahrens * ^ | /^ | 8969962b56SMatthew Ahrens * | | / | | 9069962b56SMatthew Ahrens * active | / | | 9169962b56SMatthew Ahrens * I/O | / | | 9269962b56SMatthew Ahrens * count | / | | 9369962b56SMatthew Ahrens * | / | | 9469962b56SMatthew Ahrens * |------------o | | <-- zfs_vdev_async_write_min_active 9569962b56SMatthew Ahrens * 0|____________^______|_________| 9669962b56SMatthew Ahrens * 0% | | 100% of zfs_dirty_data_max 9769962b56SMatthew Ahrens * | | 9869962b56SMatthew Ahrens * | `-- zfs_vdev_async_write_active_max_dirty_percent 9969962b56SMatthew Ahrens * `--------- zfs_vdev_async_write_active_min_dirty_percent 10069962b56SMatthew Ahrens * 10169962b56SMatthew Ahrens * Until the amount of dirty data exceeds a minimum percentage of the dirty 10269962b56SMatthew Ahrens * data allowed in the pool, the I/O scheduler will limit the number of 10369962b56SMatthew Ahrens * concurrent operations to the minimum. As that threshold is crossed, the 10469962b56SMatthew Ahrens * number of concurrent operations issued increases linearly to the maximum at 10569962b56SMatthew Ahrens * the specified maximum percentage of the dirty data allowed in the pool. 10669962b56SMatthew Ahrens * 10769962b56SMatthew Ahrens * Ideally, the amount of dirty data on a busy pool will stay in the sloped 10869962b56SMatthew Ahrens * part of the function between zfs_vdev_async_write_active_min_dirty_percent 10969962b56SMatthew Ahrens * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the 11069962b56SMatthew Ahrens * maximum percentage, this indicates that the rate of incoming data is 11169962b56SMatthew Ahrens * greater than the rate that the backend storage can handle. In this case, we 11269962b56SMatthew Ahrens * must further throttle incoming writes (see dmu_tx_delay() for details). 113614409b5Sahrens */ 114f7170741SWill Andrews 115614409b5Sahrens /* 11669962b56SMatthew Ahrens * The maximum number of i/os active to each device. Ideally, this will be >= 11769962b56SMatthew Ahrens * the sum of each queue's max_active. It must be at least the sum of each 11869962b56SMatthew Ahrens * queue's min_active. 119614409b5Sahrens */ 12069962b56SMatthew Ahrens uint32_t zfs_vdev_max_active = 1000; 121614409b5Sahrens 122c55e05cbSMatthew Ahrens /* 12369962b56SMatthew Ahrens * Per-queue limits on the number of i/os active to each device. If the 12469962b56SMatthew Ahrens * sum of the queue's max_active is < zfs_vdev_max_active, then the 12569962b56SMatthew Ahrens * min_active comes into play. We will send min_active from each queue, 12669962b56SMatthew Ahrens * and then select from queues in the order defined by zio_priority_t. 12769962b56SMatthew Ahrens * 12869962b56SMatthew Ahrens * In general, smaller max_active's will lead to lower latency of synchronous 12969962b56SMatthew Ahrens * operations. Larger max_active's may lead to higher overall throughput, 13069962b56SMatthew Ahrens * depending on underlying storage. 13169962b56SMatthew Ahrens * 13269962b56SMatthew Ahrens * The ratio of the queues' max_actives determines the balance of performance 13369962b56SMatthew Ahrens * between reads, writes, and scrubs. E.g., increasing 13469962b56SMatthew Ahrens * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete 13569962b56SMatthew Ahrens * more quickly, but reads and writes to have higher latency and lower 13669962b56SMatthew Ahrens * throughput. 137c55e05cbSMatthew Ahrens */ 13869962b56SMatthew Ahrens uint32_t zfs_vdev_sync_read_min_active = 10; 13969962b56SMatthew Ahrens uint32_t zfs_vdev_sync_read_max_active = 10; 14069962b56SMatthew Ahrens uint32_t zfs_vdev_sync_write_min_active = 10; 14169962b56SMatthew Ahrens uint32_t zfs_vdev_sync_write_max_active = 10; 14269962b56SMatthew Ahrens uint32_t zfs_vdev_async_read_min_active = 1; 14369962b56SMatthew Ahrens uint32_t zfs_vdev_async_read_max_active = 3; 14469962b56SMatthew Ahrens uint32_t zfs_vdev_async_write_min_active = 1; 14569962b56SMatthew Ahrens uint32_t zfs_vdev_async_write_max_active = 10; 14669962b56SMatthew Ahrens uint32_t zfs_vdev_scrub_min_active = 1; 14769962b56SMatthew Ahrens uint32_t zfs_vdev_scrub_max_active = 2; 148614409b5Sahrens 14969962b56SMatthew Ahrens /* 15069962b56SMatthew Ahrens * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent 15169962b56SMatthew Ahrens * dirty data, use zfs_vdev_async_write_min_active. When it has more than 15269962b56SMatthew Ahrens * zfs_vdev_async_write_active_max_dirty_percent, use 15369962b56SMatthew Ahrens * zfs_vdev_async_write_max_active. The value is linearly interpolated 15469962b56SMatthew Ahrens * between min and max. 15569962b56SMatthew Ahrens */ 15669962b56SMatthew Ahrens int zfs_vdev_async_write_active_min_dirty_percent = 30; 15769962b56SMatthew Ahrens int zfs_vdev_async_write_active_max_dirty_percent = 60; 158614409b5Sahrens 159614409b5Sahrens /* 160f94275ceSAdam Leventhal * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. 161f94275ceSAdam Leventhal * For read I/Os, we also aggregate across small adjacency gaps; for writes 162f94275ceSAdam Leventhal * we include spans of optional I/Os to aid aggregation at the disk even when 163f94275ceSAdam Leventhal * they aren't able to help us aggregate at this level. 164614409b5Sahrens */ 165614409b5Sahrens int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE; 1666f708f7cSJeff Bonwick int zfs_vdev_read_gap_limit = 32 << 10; 167f94275ceSAdam Leventhal int zfs_vdev_write_gap_limit = 4 << 10; 168614409b5Sahrens 169fa9e4066Sahrens int 17069962b56SMatthew Ahrens vdev_queue_offset_compare(const void *x1, const void *x2) 171fa9e4066Sahrens { 172fa9e4066Sahrens const zio_t *z1 = x1; 173fa9e4066Sahrens const zio_t *z2 = x2; 174fa9e4066Sahrens 175fa9e4066Sahrens if (z1->io_offset < z2->io_offset) 176fa9e4066Sahrens return (-1); 177fa9e4066Sahrens if (z1->io_offset > z2->io_offset) 178fa9e4066Sahrens return (1); 179fa9e4066Sahrens 180fa9e4066Sahrens if (z1 < z2) 181fa9e4066Sahrens return (-1); 182fa9e4066Sahrens if (z1 > z2) 183fa9e4066Sahrens return (1); 184fa9e4066Sahrens 185fa9e4066Sahrens return (0); 186fa9e4066Sahrens } 187fa9e4066Sahrens 188fa9e4066Sahrens int 18969962b56SMatthew Ahrens vdev_queue_timestamp_compare(const void *x1, const void *x2) 190fa9e4066Sahrens { 191fa9e4066Sahrens const zio_t *z1 = x1; 192fa9e4066Sahrens const zio_t *z2 = x2; 193fa9e4066Sahrens 19469962b56SMatthew Ahrens if (z1->io_timestamp < z2->io_timestamp) 195fa9e4066Sahrens return (-1); 19669962b56SMatthew Ahrens if (z1->io_timestamp > z2->io_timestamp) 197fa9e4066Sahrens return (1); 198fa9e4066Sahrens 199fa9e4066Sahrens if (z1 < z2) 200fa9e4066Sahrens return (-1); 201fa9e4066Sahrens if (z1 > z2) 202fa9e4066Sahrens return (1); 203fa9e4066Sahrens 204fa9e4066Sahrens return (0); 205fa9e4066Sahrens } 206fa9e4066Sahrens 207fa9e4066Sahrens void 208fa9e4066Sahrens vdev_queue_init(vdev_t *vd) 209fa9e4066Sahrens { 210fa9e4066Sahrens vdev_queue_t *vq = &vd->vdev_queue; 211fa9e4066Sahrens 212fa9e4066Sahrens mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 21369962b56SMatthew Ahrens vq->vq_vdev = vd; 214fa9e4066Sahrens 21569962b56SMatthew Ahrens avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 21669962b56SMatthew Ahrens sizeof (zio_t), offsetof(struct zio, io_queue_node)); 217fa9e4066Sahrens 21869962b56SMatthew Ahrens for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 21969962b56SMatthew Ahrens /* 22069962b56SMatthew Ahrens * The synchronous i/o queues are FIFO rather than LBA ordered. 22169962b56SMatthew Ahrens * This provides more consistent latency for these i/os, and 22269962b56SMatthew Ahrens * they tend to not be tightly clustered anyway so there is 22369962b56SMatthew Ahrens * little to no throughput loss. 22469962b56SMatthew Ahrens */ 22569962b56SMatthew Ahrens boolean_t fifo = (p == ZIO_PRIORITY_SYNC_READ || 22669962b56SMatthew Ahrens p == ZIO_PRIORITY_SYNC_WRITE); 22769962b56SMatthew Ahrens avl_create(&vq->vq_class[p].vqc_queued_tree, 22869962b56SMatthew Ahrens fifo ? vdev_queue_timestamp_compare : 22969962b56SMatthew Ahrens vdev_queue_offset_compare, 23069962b56SMatthew Ahrens sizeof (zio_t), offsetof(struct zio, io_queue_node)); 23169962b56SMatthew Ahrens } 232fa9e4066Sahrens } 233fa9e4066Sahrens 234fa9e4066Sahrens void 235fa9e4066Sahrens vdev_queue_fini(vdev_t *vd) 236fa9e4066Sahrens { 237fa9e4066Sahrens vdev_queue_t *vq = &vd->vdev_queue; 238fa9e4066Sahrens 23969962b56SMatthew Ahrens for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 24069962b56SMatthew Ahrens avl_destroy(&vq->vq_class[p].vqc_queued_tree); 24169962b56SMatthew Ahrens avl_destroy(&vq->vq_active_tree); 242fa9e4066Sahrens 243fa9e4066Sahrens mutex_destroy(&vq->vq_lock); 244fa9e4066Sahrens } 245fa9e4066Sahrens 246ea8dc4b6Seschrock static void 247ea8dc4b6Seschrock vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 248ea8dc4b6Seschrock { 249c3a66015SMatthew Ahrens spa_t *spa = zio->io_spa; 25069962b56SMatthew Ahrens ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 25169962b56SMatthew Ahrens avl_add(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 252c3a66015SMatthew Ahrens 25369962b56SMatthew Ahrens mutex_enter(&spa->spa_iokstat_lock); 25469962b56SMatthew Ahrens spa->spa_queue_stats[zio->io_priority].spa_queued++; 25569962b56SMatthew Ahrens if (spa->spa_iokstat != NULL) 256c3a66015SMatthew Ahrens kstat_waitq_enter(spa->spa_iokstat->ks_data); 25769962b56SMatthew Ahrens mutex_exit(&spa->spa_iokstat_lock); 258ea8dc4b6Seschrock } 259ea8dc4b6Seschrock 260ea8dc4b6Seschrock static void 261ea8dc4b6Seschrock vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 262ea8dc4b6Seschrock { 263c3a66015SMatthew Ahrens spa_t *spa = zio->io_spa; 26469962b56SMatthew Ahrens ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 26569962b56SMatthew Ahrens avl_remove(&vq->vq_class[zio->io_priority].vqc_queued_tree, zio); 266c3a66015SMatthew Ahrens 26769962b56SMatthew Ahrens mutex_enter(&spa->spa_iokstat_lock); 26869962b56SMatthew Ahrens ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 26969962b56SMatthew Ahrens spa->spa_queue_stats[zio->io_priority].spa_queued--; 27069962b56SMatthew Ahrens if (spa->spa_iokstat != NULL) 271c3a66015SMatthew Ahrens kstat_waitq_exit(spa->spa_iokstat->ks_data); 27269962b56SMatthew Ahrens mutex_exit(&spa->spa_iokstat_lock); 273c3a66015SMatthew Ahrens } 274c3a66015SMatthew Ahrens 275c3a66015SMatthew Ahrens static void 276c3a66015SMatthew Ahrens vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 277c3a66015SMatthew Ahrens { 278c3a66015SMatthew Ahrens spa_t *spa = zio->io_spa; 27969962b56SMatthew Ahrens ASSERT(MUTEX_HELD(&vq->vq_lock)); 28069962b56SMatthew Ahrens ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 28169962b56SMatthew Ahrens vq->vq_class[zio->io_priority].vqc_active++; 28269962b56SMatthew Ahrens avl_add(&vq->vq_active_tree, zio); 28369962b56SMatthew Ahrens 28469962b56SMatthew Ahrens mutex_enter(&spa->spa_iokstat_lock); 28569962b56SMatthew Ahrens spa->spa_queue_stats[zio->io_priority].spa_active++; 28669962b56SMatthew Ahrens if (spa->spa_iokstat != NULL) 287c3a66015SMatthew Ahrens kstat_runq_enter(spa->spa_iokstat->ks_data); 28869962b56SMatthew Ahrens mutex_exit(&spa->spa_iokstat_lock); 289c3a66015SMatthew Ahrens } 290c3a66015SMatthew Ahrens 291c3a66015SMatthew Ahrens static void 292c3a66015SMatthew Ahrens vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 293c3a66015SMatthew Ahrens { 294c3a66015SMatthew Ahrens spa_t *spa = zio->io_spa; 29569962b56SMatthew Ahrens ASSERT(MUTEX_HELD(&vq->vq_lock)); 29669962b56SMatthew Ahrens ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 29769962b56SMatthew Ahrens vq->vq_class[zio->io_priority].vqc_active--; 29869962b56SMatthew Ahrens avl_remove(&vq->vq_active_tree, zio); 29969962b56SMatthew Ahrens 30069962b56SMatthew Ahrens mutex_enter(&spa->spa_iokstat_lock); 30169962b56SMatthew Ahrens ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 30269962b56SMatthew Ahrens spa->spa_queue_stats[zio->io_priority].spa_active--; 303c3a66015SMatthew Ahrens if (spa->spa_iokstat != NULL) { 304c3a66015SMatthew Ahrens kstat_io_t *ksio = spa->spa_iokstat->ks_data; 305c3a66015SMatthew Ahrens 306c3a66015SMatthew Ahrens kstat_runq_exit(spa->spa_iokstat->ks_data); 307c3a66015SMatthew Ahrens if (zio->io_type == ZIO_TYPE_READ) { 308c3a66015SMatthew Ahrens ksio->reads++; 309c3a66015SMatthew Ahrens ksio->nread += zio->io_size; 310c3a66015SMatthew Ahrens } else if (zio->io_type == ZIO_TYPE_WRITE) { 311c3a66015SMatthew Ahrens ksio->writes++; 312c3a66015SMatthew Ahrens ksio->nwritten += zio->io_size; 313c3a66015SMatthew Ahrens } 314c3a66015SMatthew Ahrens } 31569962b56SMatthew Ahrens mutex_exit(&spa->spa_iokstat_lock); 316ea8dc4b6Seschrock } 317ea8dc4b6Seschrock 318fa9e4066Sahrens static void 319fa9e4066Sahrens vdev_queue_agg_io_done(zio_t *aio) 320fa9e4066Sahrens { 32169962b56SMatthew Ahrens if (aio->io_type == ZIO_TYPE_READ) { 32269962b56SMatthew Ahrens zio_t *pio; 32369962b56SMatthew Ahrens while ((pio = zio_walk_parents(aio)) != NULL) { 324a3f829aeSBill Moore bcopy((char *)aio->io_data + (pio->io_offset - 325a3f829aeSBill Moore aio->io_offset), pio->io_data, pio->io_size); 32669962b56SMatthew Ahrens } 32769962b56SMatthew Ahrens } 328fa9e4066Sahrens 329fa9e4066Sahrens zio_buf_free(aio->io_data, aio->io_size); 330fa9e4066Sahrens } 331fa9e4066Sahrens 33269962b56SMatthew Ahrens static int 33369962b56SMatthew Ahrens vdev_queue_class_min_active(zio_priority_t p) 33469962b56SMatthew Ahrens { 33569962b56SMatthew Ahrens switch (p) { 33669962b56SMatthew Ahrens case ZIO_PRIORITY_SYNC_READ: 33769962b56SMatthew Ahrens return (zfs_vdev_sync_read_min_active); 33869962b56SMatthew Ahrens case ZIO_PRIORITY_SYNC_WRITE: 33969962b56SMatthew Ahrens return (zfs_vdev_sync_write_min_active); 34069962b56SMatthew Ahrens case ZIO_PRIORITY_ASYNC_READ: 34169962b56SMatthew Ahrens return (zfs_vdev_async_read_min_active); 34269962b56SMatthew Ahrens case ZIO_PRIORITY_ASYNC_WRITE: 34369962b56SMatthew Ahrens return (zfs_vdev_async_write_min_active); 34469962b56SMatthew Ahrens case ZIO_PRIORITY_SCRUB: 34569962b56SMatthew Ahrens return (zfs_vdev_scrub_min_active); 34669962b56SMatthew Ahrens default: 34769962b56SMatthew Ahrens panic("invalid priority %u", p); 34869962b56SMatthew Ahrens return (0); 34969962b56SMatthew Ahrens } 35069962b56SMatthew Ahrens } 35169962b56SMatthew Ahrens 35269962b56SMatthew Ahrens static int 353*73527f44SAlex Reece vdev_queue_max_async_writes(spa_t *spa) 35469962b56SMatthew Ahrens { 35569962b56SMatthew Ahrens int writes; 356*73527f44SAlex Reece uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 35769962b56SMatthew Ahrens uint64_t min_bytes = zfs_dirty_data_max * 35869962b56SMatthew Ahrens zfs_vdev_async_write_active_min_dirty_percent / 100; 35969962b56SMatthew Ahrens uint64_t max_bytes = zfs_dirty_data_max * 36069962b56SMatthew Ahrens zfs_vdev_async_write_active_max_dirty_percent / 100; 36169962b56SMatthew Ahrens 362*73527f44SAlex Reece /* 363*73527f44SAlex Reece * Sync tasks correspond to interactive user actions. To reduce the 364*73527f44SAlex Reece * execution time of those actions we push data out as fast as possible. 365*73527f44SAlex Reece */ 366*73527f44SAlex Reece if (spa_has_pending_synctask(spa)) { 367*73527f44SAlex Reece return (zfs_vdev_async_write_max_active); 368*73527f44SAlex Reece } 369*73527f44SAlex Reece 37069962b56SMatthew Ahrens if (dirty < min_bytes) 37169962b56SMatthew Ahrens return (zfs_vdev_async_write_min_active); 37269962b56SMatthew Ahrens if (dirty > max_bytes) 37369962b56SMatthew Ahrens return (zfs_vdev_async_write_max_active); 37469962b56SMatthew Ahrens 37569962b56SMatthew Ahrens /* 37669962b56SMatthew Ahrens * linear interpolation: 37769962b56SMatthew Ahrens * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 37869962b56SMatthew Ahrens * move right by min_bytes 37969962b56SMatthew Ahrens * move up by min_writes 38069962b56SMatthew Ahrens */ 38169962b56SMatthew Ahrens writes = (dirty - min_bytes) * 38269962b56SMatthew Ahrens (zfs_vdev_async_write_max_active - 38369962b56SMatthew Ahrens zfs_vdev_async_write_min_active) / 38469962b56SMatthew Ahrens (max_bytes - min_bytes) + 38569962b56SMatthew Ahrens zfs_vdev_async_write_min_active; 38669962b56SMatthew Ahrens ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 38769962b56SMatthew Ahrens ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 38869962b56SMatthew Ahrens return (writes); 38969962b56SMatthew Ahrens } 39069962b56SMatthew Ahrens 39169962b56SMatthew Ahrens static int 39269962b56SMatthew Ahrens vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 39369962b56SMatthew Ahrens { 39469962b56SMatthew Ahrens switch (p) { 39569962b56SMatthew Ahrens case ZIO_PRIORITY_SYNC_READ: 39669962b56SMatthew Ahrens return (zfs_vdev_sync_read_max_active); 39769962b56SMatthew Ahrens case ZIO_PRIORITY_SYNC_WRITE: 39869962b56SMatthew Ahrens return (zfs_vdev_sync_write_max_active); 39969962b56SMatthew Ahrens case ZIO_PRIORITY_ASYNC_READ: 40069962b56SMatthew Ahrens return (zfs_vdev_async_read_max_active); 40169962b56SMatthew Ahrens case ZIO_PRIORITY_ASYNC_WRITE: 402*73527f44SAlex Reece return (vdev_queue_max_async_writes(spa)); 40369962b56SMatthew Ahrens case ZIO_PRIORITY_SCRUB: 40469962b56SMatthew Ahrens return (zfs_vdev_scrub_max_active); 40569962b56SMatthew Ahrens default: 40669962b56SMatthew Ahrens panic("invalid priority %u", p); 40769962b56SMatthew Ahrens return (0); 40869962b56SMatthew Ahrens } 40969962b56SMatthew Ahrens } 41069962b56SMatthew Ahrens 41169962b56SMatthew Ahrens /* 41269962b56SMatthew Ahrens * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 41369962b56SMatthew Ahrens * there is no eligible class. 41469962b56SMatthew Ahrens */ 41569962b56SMatthew Ahrens static zio_priority_t 41669962b56SMatthew Ahrens vdev_queue_class_to_issue(vdev_queue_t *vq) 41769962b56SMatthew Ahrens { 41869962b56SMatthew Ahrens spa_t *spa = vq->vq_vdev->vdev_spa; 41969962b56SMatthew Ahrens zio_priority_t p; 42069962b56SMatthew Ahrens 42169962b56SMatthew Ahrens if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 42269962b56SMatthew Ahrens return (ZIO_PRIORITY_NUM_QUEUEABLE); 42369962b56SMatthew Ahrens 42469962b56SMatthew Ahrens /* find a queue that has not reached its minimum # outstanding i/os */ 42569962b56SMatthew Ahrens for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 42669962b56SMatthew Ahrens if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 42769962b56SMatthew Ahrens vq->vq_class[p].vqc_active < 42869962b56SMatthew Ahrens vdev_queue_class_min_active(p)) 42969962b56SMatthew Ahrens return (p); 43069962b56SMatthew Ahrens } 43169962b56SMatthew Ahrens 43269962b56SMatthew Ahrens /* 43369962b56SMatthew Ahrens * If we haven't found a queue, look for one that hasn't reached its 43469962b56SMatthew Ahrens * maximum # outstanding i/os. 43569962b56SMatthew Ahrens */ 43669962b56SMatthew Ahrens for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 43769962b56SMatthew Ahrens if (avl_numnodes(&vq->vq_class[p].vqc_queued_tree) > 0 && 43869962b56SMatthew Ahrens vq->vq_class[p].vqc_active < 43969962b56SMatthew Ahrens vdev_queue_class_max_active(spa, p)) 44069962b56SMatthew Ahrens return (p); 44169962b56SMatthew Ahrens } 44269962b56SMatthew Ahrens 44369962b56SMatthew Ahrens /* No eligible queued i/os */ 44469962b56SMatthew Ahrens return (ZIO_PRIORITY_NUM_QUEUEABLE); 44569962b56SMatthew Ahrens } 44669962b56SMatthew Ahrens 4476f708f7cSJeff Bonwick /* 4486f708f7cSJeff Bonwick * Compute the range spanned by two i/os, which is the endpoint of the last 4496f708f7cSJeff Bonwick * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 4506f708f7cSJeff Bonwick * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 4516f708f7cSJeff Bonwick * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 4526f708f7cSJeff Bonwick */ 4536f708f7cSJeff Bonwick #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 4546f708f7cSJeff Bonwick #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 455fa9e4066Sahrens 456fa9e4066Sahrens static zio_t * 45769962b56SMatthew Ahrens vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 458fa9e4066Sahrens { 45969962b56SMatthew Ahrens zio_t *first, *last, *aio, *dio, *mandatory, *nio; 46069962b56SMatthew Ahrens uint64_t maxgap = 0; 46169962b56SMatthew Ahrens uint64_t size; 46269962b56SMatthew Ahrens boolean_t stretch = B_FALSE; 46369962b56SMatthew Ahrens vdev_queue_class_t *vqc = &vq->vq_class[zio->io_priority]; 46469962b56SMatthew Ahrens avl_tree_t *t = &vqc->vqc_queued_tree; 46569962b56SMatthew Ahrens enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 46669962b56SMatthew Ahrens 46769962b56SMatthew Ahrens if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 46869962b56SMatthew Ahrens return (NULL); 469fa9e4066Sahrens 47069962b56SMatthew Ahrens /* 47169962b56SMatthew Ahrens * The synchronous i/o queues are not sorted by LBA, so we can't 47269962b56SMatthew Ahrens * find adjacent i/os. These i/os tend to not be tightly clustered, 47369962b56SMatthew Ahrens * or too large to aggregate, so this has little impact on performance. 47469962b56SMatthew Ahrens */ 47569962b56SMatthew Ahrens if (zio->io_priority == ZIO_PRIORITY_SYNC_READ || 47669962b56SMatthew Ahrens zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) 477fa9e4066Sahrens return (NULL); 478fa9e4066Sahrens 47969962b56SMatthew Ahrens first = last = zio; 480fa9e4066Sahrens 48169962b56SMatthew Ahrens if (zio->io_type == ZIO_TYPE_READ) 48269962b56SMatthew Ahrens maxgap = zfs_vdev_read_gap_limit; 4838ad4d6ddSJeff Bonwick 48469962b56SMatthew Ahrens /* 48569962b56SMatthew Ahrens * We can aggregate I/Os that are sufficiently adjacent and of 48669962b56SMatthew Ahrens * the same flavor, as expressed by the AGG_INHERIT flags. 48769962b56SMatthew Ahrens * The latter requirement is necessary so that certain 48869962b56SMatthew Ahrens * attributes of the I/O, such as whether it's a normal I/O 48969962b56SMatthew Ahrens * or a scrub/resilver, can be preserved in the aggregate. 49069962b56SMatthew Ahrens * We can include optional I/Os, but don't allow them 49169962b56SMatthew Ahrens * to begin a range as they add no benefit in that situation. 49269962b56SMatthew Ahrens */ 493f94275ceSAdam Leventhal 49469962b56SMatthew Ahrens /* 49569962b56SMatthew Ahrens * We keep track of the last non-optional I/O. 49669962b56SMatthew Ahrens */ 49769962b56SMatthew Ahrens mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 498f94275ceSAdam Leventhal 49969962b56SMatthew Ahrens /* 50069962b56SMatthew Ahrens * Walk backwards through sufficiently contiguous I/Os 50169962b56SMatthew Ahrens * recording the last non-option I/O. 50269962b56SMatthew Ahrens */ 50369962b56SMatthew Ahrens while ((dio = AVL_PREV(t, first)) != NULL && 50469962b56SMatthew Ahrens (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 50569962b56SMatthew Ahrens IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 50669962b56SMatthew Ahrens IO_GAP(dio, first) <= maxgap) { 50769962b56SMatthew Ahrens first = dio; 50869962b56SMatthew Ahrens if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 50969962b56SMatthew Ahrens mandatory = first; 51069962b56SMatthew Ahrens } 511f94275ceSAdam Leventhal 51269962b56SMatthew Ahrens /* 51369962b56SMatthew Ahrens * Skip any initial optional I/Os. 51469962b56SMatthew Ahrens */ 51569962b56SMatthew Ahrens while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 51669962b56SMatthew Ahrens first = AVL_NEXT(t, first); 51769962b56SMatthew Ahrens ASSERT(first != NULL); 51869962b56SMatthew Ahrens } 5196f708f7cSJeff Bonwick 52069962b56SMatthew Ahrens /* 52169962b56SMatthew Ahrens * Walk forward through sufficiently contiguous I/Os. 52269962b56SMatthew Ahrens */ 52369962b56SMatthew Ahrens while ((dio = AVL_NEXT(t, last)) != NULL && 52469962b56SMatthew Ahrens (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 52569962b56SMatthew Ahrens IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit && 52669962b56SMatthew Ahrens IO_GAP(last, dio) <= maxgap) { 52769962b56SMatthew Ahrens last = dio; 52869962b56SMatthew Ahrens if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 52969962b56SMatthew Ahrens mandatory = last; 53069962b56SMatthew Ahrens } 531f94275ceSAdam Leventhal 53269962b56SMatthew Ahrens /* 53369962b56SMatthew Ahrens * Now that we've established the range of the I/O aggregation 53469962b56SMatthew Ahrens * we must decide what to do with trailing optional I/Os. 53569962b56SMatthew Ahrens * For reads, there's nothing to do. While we are unable to 53669962b56SMatthew Ahrens * aggregate further, it's possible that a trailing optional 53769962b56SMatthew Ahrens * I/O would allow the underlying device to aggregate with 53869962b56SMatthew Ahrens * subsequent I/Os. We must therefore determine if the next 53969962b56SMatthew Ahrens * non-optional I/O is close enough to make aggregation 54069962b56SMatthew Ahrens * worthwhile. 54169962b56SMatthew Ahrens */ 54269962b56SMatthew Ahrens if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 54369962b56SMatthew Ahrens zio_t *nio = last; 54469962b56SMatthew Ahrens while ((dio = AVL_NEXT(t, nio)) != NULL && 54569962b56SMatthew Ahrens IO_GAP(nio, dio) == 0 && 54669962b56SMatthew Ahrens IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 54769962b56SMatthew Ahrens nio = dio; 54869962b56SMatthew Ahrens if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 54969962b56SMatthew Ahrens stretch = B_TRUE; 55069962b56SMatthew Ahrens break; 551f94275ceSAdam Leventhal } 552f94275ceSAdam Leventhal } 55369962b56SMatthew Ahrens } 554f94275ceSAdam Leventhal 55569962b56SMatthew Ahrens if (stretch) { 55669962b56SMatthew Ahrens /* This may be a no-op. */ 55769962b56SMatthew Ahrens dio = AVL_NEXT(t, last); 55869962b56SMatthew Ahrens dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 55969962b56SMatthew Ahrens } else { 56069962b56SMatthew Ahrens while (last != mandatory && last != first) { 56169962b56SMatthew Ahrens ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 56269962b56SMatthew Ahrens last = AVL_PREV(t, last); 56369962b56SMatthew Ahrens ASSERT(last != NULL); 564f94275ceSAdam Leventhal } 565fa9e4066Sahrens } 566fa9e4066Sahrens 56769962b56SMatthew Ahrens if (first == last) 56869962b56SMatthew Ahrens return (NULL); 56969962b56SMatthew Ahrens 57069962b56SMatthew Ahrens size = IO_SPAN(first, last); 57169962b56SMatthew Ahrens ASSERT3U(size, <=, zfs_vdev_aggregation_limit); 57269962b56SMatthew Ahrens 57369962b56SMatthew Ahrens aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 57469962b56SMatthew Ahrens zio_buf_alloc(size), size, first->io_type, zio->io_priority, 57569962b56SMatthew Ahrens flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 57669962b56SMatthew Ahrens vdev_queue_agg_io_done, NULL); 57769962b56SMatthew Ahrens aio->io_timestamp = first->io_timestamp; 57869962b56SMatthew Ahrens 57969962b56SMatthew Ahrens nio = first; 58069962b56SMatthew Ahrens do { 58169962b56SMatthew Ahrens dio = nio; 58269962b56SMatthew Ahrens nio = AVL_NEXT(t, dio); 58369962b56SMatthew Ahrens ASSERT3U(dio->io_type, ==, aio->io_type); 58469962b56SMatthew Ahrens 58569962b56SMatthew Ahrens if (dio->io_flags & ZIO_FLAG_NODATA) { 58669962b56SMatthew Ahrens ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 58769962b56SMatthew Ahrens bzero((char *)aio->io_data + (dio->io_offset - 58869962b56SMatthew Ahrens aio->io_offset), dio->io_size); 58969962b56SMatthew Ahrens } else if (dio->io_type == ZIO_TYPE_WRITE) { 59069962b56SMatthew Ahrens bcopy(dio->io_data, (char *)aio->io_data + 59169962b56SMatthew Ahrens (dio->io_offset - aio->io_offset), 59269962b56SMatthew Ahrens dio->io_size); 59369962b56SMatthew Ahrens } 594a3f829aeSBill Moore 59569962b56SMatthew Ahrens zio_add_child(dio, aio); 59669962b56SMatthew Ahrens vdev_queue_io_remove(vq, dio); 59769962b56SMatthew Ahrens zio_vdev_io_bypass(dio); 59869962b56SMatthew Ahrens zio_execute(dio); 59969962b56SMatthew Ahrens } while (dio != last); 60069962b56SMatthew Ahrens 60169962b56SMatthew Ahrens return (aio); 60269962b56SMatthew Ahrens } 60369962b56SMatthew Ahrens 60469962b56SMatthew Ahrens static zio_t * 60569962b56SMatthew Ahrens vdev_queue_io_to_issue(vdev_queue_t *vq) 60669962b56SMatthew Ahrens { 60769962b56SMatthew Ahrens zio_t *zio, *aio; 60869962b56SMatthew Ahrens zio_priority_t p; 60969962b56SMatthew Ahrens avl_index_t idx; 61069962b56SMatthew Ahrens vdev_queue_class_t *vqc; 61169962b56SMatthew Ahrens zio_t search; 61269962b56SMatthew Ahrens 61369962b56SMatthew Ahrens again: 61469962b56SMatthew Ahrens ASSERT(MUTEX_HELD(&vq->vq_lock)); 615fa9e4066Sahrens 61669962b56SMatthew Ahrens p = vdev_queue_class_to_issue(vq); 617fa9e4066Sahrens 61869962b56SMatthew Ahrens if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 61969962b56SMatthew Ahrens /* No eligible queued i/os */ 62069962b56SMatthew Ahrens return (NULL); 621fa9e4066Sahrens } 622fa9e4066Sahrens 62369962b56SMatthew Ahrens /* 62469962b56SMatthew Ahrens * For LBA-ordered queues (async / scrub), issue the i/o which follows 62569962b56SMatthew Ahrens * the most recently issued i/o in LBA (offset) order. 62669962b56SMatthew Ahrens * 62769962b56SMatthew Ahrens * For FIFO queues (sync), issue the i/o with the lowest timestamp. 62869962b56SMatthew Ahrens */ 62969962b56SMatthew Ahrens vqc = &vq->vq_class[p]; 63069962b56SMatthew Ahrens search.io_timestamp = 0; 63169962b56SMatthew Ahrens search.io_offset = vq->vq_last_offset + 1; 63269962b56SMatthew Ahrens VERIFY3P(avl_find(&vqc->vqc_queued_tree, &search, &idx), ==, NULL); 63369962b56SMatthew Ahrens zio = avl_nearest(&vqc->vqc_queued_tree, idx, AVL_AFTER); 63469962b56SMatthew Ahrens if (zio == NULL) 63569962b56SMatthew Ahrens zio = avl_first(&vqc->vqc_queued_tree); 63669962b56SMatthew Ahrens ASSERT3U(zio->io_priority, ==, p); 63769962b56SMatthew Ahrens 63869962b56SMatthew Ahrens aio = vdev_queue_aggregate(vq, zio); 63969962b56SMatthew Ahrens if (aio != NULL) 64069962b56SMatthew Ahrens zio = aio; 64169962b56SMatthew Ahrens else 64269962b56SMatthew Ahrens vdev_queue_io_remove(vq, zio); 643fa9e4066Sahrens 644f94275ceSAdam Leventhal /* 645f94275ceSAdam Leventhal * If the I/O is or was optional and therefore has no data, we need to 646f94275ceSAdam Leventhal * simply discard it. We need to drop the vdev queue's lock to avoid a 647f94275ceSAdam Leventhal * deadlock that we could encounter since this I/O will complete 648f94275ceSAdam Leventhal * immediately. 649f94275ceSAdam Leventhal */ 65069962b56SMatthew Ahrens if (zio->io_flags & ZIO_FLAG_NODATA) { 651f94275ceSAdam Leventhal mutex_exit(&vq->vq_lock); 65269962b56SMatthew Ahrens zio_vdev_io_bypass(zio); 65369962b56SMatthew Ahrens zio_execute(zio); 654f94275ceSAdam Leventhal mutex_enter(&vq->vq_lock); 655f94275ceSAdam Leventhal goto again; 656f94275ceSAdam Leventhal } 657f94275ceSAdam Leventhal 65869962b56SMatthew Ahrens vdev_queue_pending_add(vq, zio); 65969962b56SMatthew Ahrens vq->vq_last_offset = zio->io_offset; 660fa9e4066Sahrens 66169962b56SMatthew Ahrens return (zio); 662fa9e4066Sahrens } 663fa9e4066Sahrens 664fa9e4066Sahrens zio_t * 665fa9e4066Sahrens vdev_queue_io(zio_t *zio) 666fa9e4066Sahrens { 667fa9e4066Sahrens vdev_queue_t *vq = &zio->io_vd->vdev_queue; 668fa9e4066Sahrens zio_t *nio; 669fa9e4066Sahrens 670fa9e4066Sahrens if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 671fa9e4066Sahrens return (zio); 672fa9e4066Sahrens 67369962b56SMatthew Ahrens /* 67469962b56SMatthew Ahrens * Children i/os inherent their parent's priority, which might 67569962b56SMatthew Ahrens * not match the child's i/o type. Fix it up here. 67669962b56SMatthew Ahrens */ 67769962b56SMatthew Ahrens if (zio->io_type == ZIO_TYPE_READ) { 67869962b56SMatthew Ahrens if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 67969962b56SMatthew Ahrens zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 68069962b56SMatthew Ahrens zio->io_priority != ZIO_PRIORITY_SCRUB) 68169962b56SMatthew Ahrens zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 68269962b56SMatthew Ahrens } else { 68369962b56SMatthew Ahrens ASSERT(zio->io_type == ZIO_TYPE_WRITE); 68469962b56SMatthew Ahrens if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 68569962b56SMatthew Ahrens zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 68669962b56SMatthew Ahrens zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 68769962b56SMatthew Ahrens } 688fa9e4066Sahrens 68969962b56SMatthew Ahrens zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 690fa9e4066Sahrens 691fa9e4066Sahrens mutex_enter(&vq->vq_lock); 692c55e05cbSMatthew Ahrens zio->io_timestamp = gethrtime(); 693ea8dc4b6Seschrock vdev_queue_io_add(vq, zio); 69469962b56SMatthew Ahrens nio = vdev_queue_io_to_issue(vq); 695fa9e4066Sahrens mutex_exit(&vq->vq_lock); 696fa9e4066Sahrens 697e05725b1Sbonwick if (nio == NULL) 698e05725b1Sbonwick return (NULL); 699e05725b1Sbonwick 700e05725b1Sbonwick if (nio->io_done == vdev_queue_agg_io_done) { 701e05725b1Sbonwick zio_nowait(nio); 702e05725b1Sbonwick return (NULL); 703e05725b1Sbonwick } 704fa9e4066Sahrens 705e05725b1Sbonwick return (nio); 706fa9e4066Sahrens } 707fa9e4066Sahrens 708fa9e4066Sahrens void 709fa9e4066Sahrens vdev_queue_io_done(zio_t *zio) 710fa9e4066Sahrens { 711fa9e4066Sahrens vdev_queue_t *vq = &zio->io_vd->vdev_queue; 71269962b56SMatthew Ahrens zio_t *nio; 713fa9e4066Sahrens 714283b8460SGeorge.Wilson if (zio_injection_enabled) 715283b8460SGeorge.Wilson delay(SEC_TO_TICK(zio_handle_io_delay(zio))); 716283b8460SGeorge.Wilson 717fa9e4066Sahrens mutex_enter(&vq->vq_lock); 718fa9e4066Sahrens 719c3a66015SMatthew Ahrens vdev_queue_pending_remove(vq, zio); 720fa9e4066Sahrens 721c55e05cbSMatthew Ahrens vq->vq_io_complete_ts = gethrtime(); 722283b8460SGeorge.Wilson 72369962b56SMatthew Ahrens while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 724fa9e4066Sahrens mutex_exit(&vq->vq_lock); 725e05725b1Sbonwick if (nio->io_done == vdev_queue_agg_io_done) { 726e05725b1Sbonwick zio_nowait(nio); 727e05725b1Sbonwick } else { 728fa9e4066Sahrens zio_vdev_io_reissue(nio); 729e05725b1Sbonwick zio_execute(nio); 730e05725b1Sbonwick } 731fa9e4066Sahrens mutex_enter(&vq->vq_lock); 732fa9e4066Sahrens } 733fa9e4066Sahrens 734fa9e4066Sahrens mutex_exit(&vq->vq_lock); 735fa9e4066Sahrens } 736