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 /* 275b062782SMatthew Ahrens * Copyright (c) 2012, 2017 by Delphix. All rights reserved. 28c3d26abcSMatthew Ahrens * Copyright (c) 2014 Integros [integros.com] 29283b8460SGeorge.Wilson */ 30283b8460SGeorge.Wilson 31fa9e4066Sahrens #include <sys/zfs_context.h> 32fa9e4066Sahrens #include <sys/vdev_impl.h> 33c3a66015SMatthew Ahrens #include <sys/spa_impl.h> 34fa9e4066Sahrens #include <sys/zio.h> 35fa9e4066Sahrens #include <sys/avl.h> 3669962b56SMatthew Ahrens #include <sys/dsl_pool.h> 370f7643c7SGeorge Wilson #include <sys/metaslab_impl.h> 38770499e1SDan Kimmel #include <sys/abd.h> 39fa9e4066Sahrens 40614409b5Sahrens /* 4169962b56SMatthew Ahrens * ZFS I/O Scheduler 4269962b56SMatthew Ahrens * --------------- 4369962b56SMatthew Ahrens * 4469962b56SMatthew Ahrens * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The 4569962b56SMatthew Ahrens * I/O scheduler determines when and in what order those operations are 4669962b56SMatthew Ahrens * issued. The I/O scheduler divides operations into five I/O classes 4769962b56SMatthew Ahrens * prioritized in the following order: sync read, sync write, async read, 4869962b56SMatthew Ahrens * async write, and scrub/resilver. Each queue defines the minimum and 4969962b56SMatthew Ahrens * maximum number of concurrent operations that may be issued to the device. 5069962b56SMatthew Ahrens * In addition, the device has an aggregate maximum. Note that the sum of the 5169962b56SMatthew Ahrens * per-queue minimums must not exceed the aggregate maximum, and if the 5269962b56SMatthew Ahrens * aggregate maximum is equal to or greater than the sum of the per-queue 5369962b56SMatthew Ahrens * maximums, the per-queue minimum has no effect. 5469962b56SMatthew Ahrens * 5569962b56SMatthew Ahrens * For many physical devices, throughput increases with the number of 5669962b56SMatthew Ahrens * concurrent operations, but latency typically suffers. Further, physical 5769962b56SMatthew Ahrens * devices typically have a limit at which more concurrent operations have no 5869962b56SMatthew Ahrens * effect on throughput or can actually cause it to decrease. 5969962b56SMatthew Ahrens * 6069962b56SMatthew Ahrens * The scheduler selects the next operation to issue by first looking for an 6169962b56SMatthew Ahrens * I/O class whose minimum has not been satisfied. Once all are satisfied and 6269962b56SMatthew Ahrens * the aggregate maximum has not been hit, the scheduler looks for classes 6369962b56SMatthew Ahrens * whose maximum has not been satisfied. Iteration through the I/O classes is 6469962b56SMatthew Ahrens * done in the order specified above. No further operations are issued if the 6569962b56SMatthew Ahrens * aggregate maximum number of concurrent operations has been hit or if there 6669962b56SMatthew Ahrens * are no operations queued for an I/O class that has not hit its maximum. 6769962b56SMatthew Ahrens * Every time an i/o is queued or an operation completes, the I/O scheduler 6869962b56SMatthew Ahrens * looks for new operations to issue. 6969962b56SMatthew Ahrens * 7069962b56SMatthew Ahrens * All I/O classes have a fixed maximum number of outstanding operations 7169962b56SMatthew Ahrens * except for the async write class. Asynchronous writes represent the data 7269962b56SMatthew Ahrens * that is committed to stable storage during the syncing stage for 7369962b56SMatthew Ahrens * transaction groups (see txg.c). Transaction groups enter the syncing state 7469962b56SMatthew Ahrens * periodically so the number of queued async writes will quickly burst up and 7569962b56SMatthew Ahrens * then bleed down to zero. Rather than servicing them as quickly as possible, 7669962b56SMatthew Ahrens * the I/O scheduler changes the maximum number of active async write i/os 7769962b56SMatthew Ahrens * according to the amount of dirty data in the pool (see dsl_pool.c). Since 7869962b56SMatthew Ahrens * both throughput and latency typically increase with the number of 7969962b56SMatthew Ahrens * concurrent operations issued to physical devices, reducing the burstiness 8069962b56SMatthew Ahrens * in the number of concurrent operations also stabilizes the response time of 8169962b56SMatthew Ahrens * operations from other -- and in particular synchronous -- queues. In broad 8269962b56SMatthew Ahrens * strokes, the I/O scheduler will issue more concurrent operations from the 8369962b56SMatthew Ahrens * async write queue as there's more dirty data in the pool. 8469962b56SMatthew Ahrens * 8569962b56SMatthew Ahrens * Async Writes 8669962b56SMatthew Ahrens * 8769962b56SMatthew Ahrens * The number of concurrent operations issued for the async write I/O class 8869962b56SMatthew Ahrens * follows a piece-wise linear function defined by a few adjustable points. 8969962b56SMatthew Ahrens * 9069962b56SMatthew Ahrens * | o---------| <-- zfs_vdev_async_write_max_active 9169962b56SMatthew Ahrens * ^ | /^ | 9269962b56SMatthew Ahrens * | | / | | 9369962b56SMatthew Ahrens * active | / | | 9469962b56SMatthew Ahrens * I/O | / | | 9569962b56SMatthew Ahrens * count | / | | 9669962b56SMatthew Ahrens * | / | | 9769962b56SMatthew Ahrens * |------------o | | <-- zfs_vdev_async_write_min_active 9869962b56SMatthew Ahrens * 0|____________^______|_________| 9969962b56SMatthew Ahrens * 0% | | 100% of zfs_dirty_data_max 10069962b56SMatthew Ahrens * | | 10169962b56SMatthew Ahrens * | `-- zfs_vdev_async_write_active_max_dirty_percent 10269962b56SMatthew Ahrens * `--------- zfs_vdev_async_write_active_min_dirty_percent 10369962b56SMatthew Ahrens * 10469962b56SMatthew Ahrens * Until the amount of dirty data exceeds a minimum percentage of the dirty 10569962b56SMatthew Ahrens * data allowed in the pool, the I/O scheduler will limit the number of 10669962b56SMatthew Ahrens * concurrent operations to the minimum. As that threshold is crossed, the 10769962b56SMatthew Ahrens * number of concurrent operations issued increases linearly to the maximum at 10869962b56SMatthew Ahrens * the specified maximum percentage of the dirty data allowed in the pool. 10969962b56SMatthew Ahrens * 11069962b56SMatthew Ahrens * Ideally, the amount of dirty data on a busy pool will stay in the sloped 11169962b56SMatthew Ahrens * part of the function between zfs_vdev_async_write_active_min_dirty_percent 11269962b56SMatthew Ahrens * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the 11369962b56SMatthew Ahrens * maximum percentage, this indicates that the rate of incoming data is 11469962b56SMatthew Ahrens * greater than the rate that the backend storage can handle. In this case, we 11569962b56SMatthew Ahrens * must further throttle incoming writes (see dmu_tx_delay() for details). 116614409b5Sahrens */ 117f7170741SWill Andrews 118614409b5Sahrens /* 11969962b56SMatthew Ahrens * The maximum number of i/os active to each device. Ideally, this will be >= 12069962b56SMatthew Ahrens * the sum of each queue's max_active. It must be at least the sum of each 12169962b56SMatthew Ahrens * queue's min_active. 122614409b5Sahrens */ 12369962b56SMatthew Ahrens uint32_t zfs_vdev_max_active = 1000; 124614409b5Sahrens 125c55e05cbSMatthew Ahrens /* 12669962b56SMatthew Ahrens * Per-queue limits on the number of i/os active to each device. If the 12769962b56SMatthew Ahrens * sum of the queue's max_active is < zfs_vdev_max_active, then the 12869962b56SMatthew Ahrens * min_active comes into play. We will send min_active from each queue, 12969962b56SMatthew Ahrens * and then select from queues in the order defined by zio_priority_t. 13069962b56SMatthew Ahrens * 13169962b56SMatthew Ahrens * In general, smaller max_active's will lead to lower latency of synchronous 13269962b56SMatthew Ahrens * operations. Larger max_active's may lead to higher overall throughput, 13369962b56SMatthew Ahrens * depending on underlying storage. 13469962b56SMatthew Ahrens * 13569962b56SMatthew Ahrens * The ratio of the queues' max_actives determines the balance of performance 13669962b56SMatthew Ahrens * between reads, writes, and scrubs. E.g., increasing 13769962b56SMatthew Ahrens * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete 13869962b56SMatthew Ahrens * more quickly, but reads and writes to have higher latency and lower 13969962b56SMatthew Ahrens * throughput. 140c55e05cbSMatthew Ahrens */ 14169962b56SMatthew Ahrens uint32_t zfs_vdev_sync_read_min_active = 10; 14269962b56SMatthew Ahrens uint32_t zfs_vdev_sync_read_max_active = 10; 14369962b56SMatthew Ahrens uint32_t zfs_vdev_sync_write_min_active = 10; 14469962b56SMatthew Ahrens uint32_t zfs_vdev_sync_write_max_active = 10; 14569962b56SMatthew Ahrens uint32_t zfs_vdev_async_read_min_active = 1; 14669962b56SMatthew Ahrens uint32_t zfs_vdev_async_read_max_active = 3; 14769962b56SMatthew Ahrens uint32_t zfs_vdev_async_write_min_active = 1; 14869962b56SMatthew Ahrens uint32_t zfs_vdev_async_write_max_active = 10; 14969962b56SMatthew Ahrens uint32_t zfs_vdev_scrub_min_active = 1; 15069962b56SMatthew Ahrens uint32_t zfs_vdev_scrub_max_active = 2; 151*5cabbc6bSPrashanth Sreenivasa uint32_t zfs_vdev_removal_min_active = 1; 152*5cabbc6bSPrashanth Sreenivasa uint32_t zfs_vdev_removal_max_active = 2; 153614409b5Sahrens 15469962b56SMatthew Ahrens /* 15569962b56SMatthew Ahrens * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent 15669962b56SMatthew Ahrens * dirty data, use zfs_vdev_async_write_min_active. When it has more than 15769962b56SMatthew Ahrens * zfs_vdev_async_write_active_max_dirty_percent, use 15869962b56SMatthew Ahrens * zfs_vdev_async_write_max_active. The value is linearly interpolated 15969962b56SMatthew Ahrens * between min and max. 16069962b56SMatthew Ahrens */ 16169962b56SMatthew Ahrens int zfs_vdev_async_write_active_min_dirty_percent = 30; 16269962b56SMatthew Ahrens int zfs_vdev_async_write_active_max_dirty_percent = 60; 163614409b5Sahrens 164614409b5Sahrens /* 165f94275ceSAdam Leventhal * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. 166f94275ceSAdam Leventhal * For read I/Os, we also aggregate across small adjacency gaps; for writes 167f94275ceSAdam Leventhal * we include spans of optional I/Os to aid aggregation at the disk even when 168f94275ceSAdam Leventhal * they aren't able to help us aggregate at this level. 169614409b5Sahrens */ 170b5152584SMatthew Ahrens int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE; 1716f708f7cSJeff Bonwick int zfs_vdev_read_gap_limit = 32 << 10; 172f94275ceSAdam Leventhal int zfs_vdev_write_gap_limit = 4 << 10; 173614409b5Sahrens 1740f7643c7SGeorge Wilson /* 1750f7643c7SGeorge Wilson * Define the queue depth percentage for each top-level. This percentage is 1760f7643c7SGeorge Wilson * used in conjunction with zfs_vdev_async_max_active to determine how many 1770f7643c7SGeorge Wilson * allocations a specific top-level vdev should handle. Once the queue depth 1780f7643c7SGeorge Wilson * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100 1790f7643c7SGeorge Wilson * then allocator will stop allocating blocks on that top-level device. 1800f7643c7SGeorge Wilson * The default kernel setting is 1000% which will yield 100 allocations per 1810f7643c7SGeorge Wilson * device. For userland testing, the default setting is 300% which equates 1820f7643c7SGeorge Wilson * to 30 allocations per device. 1830f7643c7SGeorge Wilson */ 1840f7643c7SGeorge Wilson #ifdef _KERNEL 1850f7643c7SGeorge Wilson int zfs_vdev_queue_depth_pct = 1000; 1860f7643c7SGeorge Wilson #else 1870f7643c7SGeorge Wilson int zfs_vdev_queue_depth_pct = 300; 1880f7643c7SGeorge Wilson #endif 1890f7643c7SGeorge Wilson 1900f7643c7SGeorge Wilson 191fa9e4066Sahrens int 19269962b56SMatthew Ahrens vdev_queue_offset_compare(const void *x1, const void *x2) 193fa9e4066Sahrens { 194fa9e4066Sahrens const zio_t *z1 = x1; 195fa9e4066Sahrens const zio_t *z2 = x2; 196fa9e4066Sahrens 197fa9e4066Sahrens if (z1->io_offset < z2->io_offset) 198fa9e4066Sahrens return (-1); 199fa9e4066Sahrens if (z1->io_offset > z2->io_offset) 200fa9e4066Sahrens return (1); 201fa9e4066Sahrens 202fa9e4066Sahrens if (z1 < z2) 203fa9e4066Sahrens return (-1); 204fa9e4066Sahrens if (z1 > z2) 205fa9e4066Sahrens return (1); 206fa9e4066Sahrens 207fa9e4066Sahrens return (0); 208fa9e4066Sahrens } 209fa9e4066Sahrens 210fe319232SJustin T. Gibbs static inline avl_tree_t * 211fe319232SJustin T. Gibbs vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) 212fe319232SJustin T. Gibbs { 213fe319232SJustin T. Gibbs return (&vq->vq_class[p].vqc_queued_tree); 214fe319232SJustin T. Gibbs } 215fe319232SJustin T. Gibbs 216fe319232SJustin T. Gibbs static inline avl_tree_t * 217fe319232SJustin T. Gibbs vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) 218fe319232SJustin T. Gibbs { 219fe319232SJustin T. Gibbs ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE); 220fe319232SJustin T. Gibbs if (t == ZIO_TYPE_READ) 221fe319232SJustin T. Gibbs return (&vq->vq_read_offset_tree); 222fe319232SJustin T. Gibbs else 223fe319232SJustin T. Gibbs return (&vq->vq_write_offset_tree); 224fe319232SJustin T. Gibbs } 225fe319232SJustin T. Gibbs 226fa9e4066Sahrens int 22769962b56SMatthew Ahrens vdev_queue_timestamp_compare(const void *x1, const void *x2) 228fa9e4066Sahrens { 229fa9e4066Sahrens const zio_t *z1 = x1; 230fa9e4066Sahrens const zio_t *z2 = x2; 231fa9e4066Sahrens 23269962b56SMatthew Ahrens if (z1->io_timestamp < z2->io_timestamp) 233fa9e4066Sahrens return (-1); 23469962b56SMatthew Ahrens if (z1->io_timestamp > z2->io_timestamp) 235fa9e4066Sahrens return (1); 236fa9e4066Sahrens 237fa9e4066Sahrens if (z1 < z2) 238fa9e4066Sahrens return (-1); 239fa9e4066Sahrens if (z1 > z2) 240fa9e4066Sahrens return (1); 241fa9e4066Sahrens 242fa9e4066Sahrens return (0); 243fa9e4066Sahrens } 244fa9e4066Sahrens 245fa9e4066Sahrens void 246fa9e4066Sahrens vdev_queue_init(vdev_t *vd) 247fa9e4066Sahrens { 248fa9e4066Sahrens vdev_queue_t *vq = &vd->vdev_queue; 249fa9e4066Sahrens 250fa9e4066Sahrens mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 25169962b56SMatthew Ahrens vq->vq_vdev = vd; 252fa9e4066Sahrens 25369962b56SMatthew Ahrens avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 25469962b56SMatthew Ahrens sizeof (zio_t), offsetof(struct zio, io_queue_node)); 255fe319232SJustin T. Gibbs avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), 256fe319232SJustin T. Gibbs vdev_queue_offset_compare, sizeof (zio_t), 257fe319232SJustin T. Gibbs offsetof(struct zio, io_offset_node)); 258fe319232SJustin T. Gibbs avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), 259fe319232SJustin T. Gibbs vdev_queue_offset_compare, sizeof (zio_t), 260fe319232SJustin T. Gibbs offsetof(struct zio, io_offset_node)); 261fa9e4066Sahrens 26269962b56SMatthew Ahrens for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 263fe319232SJustin T. Gibbs int (*compfn) (const void *, const void *); 264fe319232SJustin T. Gibbs 26569962b56SMatthew Ahrens /* 266fe319232SJustin T. Gibbs * The synchronous i/o queues are dispatched in FIFO rather 267fe319232SJustin T. Gibbs * than LBA order. This provides more consistent latency for 268fe319232SJustin T. Gibbs * these i/os. 26969962b56SMatthew Ahrens */ 270fe319232SJustin T. Gibbs if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) 271fe319232SJustin T. Gibbs compfn = vdev_queue_timestamp_compare; 272fe319232SJustin T. Gibbs else 273fe319232SJustin T. Gibbs compfn = vdev_queue_offset_compare; 274fe319232SJustin T. Gibbs 275fe319232SJustin T. Gibbs avl_create(vdev_queue_class_tree(vq, p), compfn, 27669962b56SMatthew Ahrens sizeof (zio_t), offsetof(struct zio, io_queue_node)); 27769962b56SMatthew Ahrens } 278fa9e4066Sahrens } 279fa9e4066Sahrens 280fa9e4066Sahrens void 281fa9e4066Sahrens vdev_queue_fini(vdev_t *vd) 282fa9e4066Sahrens { 283fa9e4066Sahrens vdev_queue_t *vq = &vd->vdev_queue; 284fa9e4066Sahrens 28569962b56SMatthew Ahrens for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 286fe319232SJustin T. Gibbs avl_destroy(vdev_queue_class_tree(vq, p)); 28769962b56SMatthew Ahrens avl_destroy(&vq->vq_active_tree); 288fe319232SJustin T. Gibbs avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); 289fe319232SJustin T. Gibbs avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); 290fa9e4066Sahrens 291fa9e4066Sahrens mutex_destroy(&vq->vq_lock); 292fa9e4066Sahrens } 293fa9e4066Sahrens 294ea8dc4b6Seschrock static void 295ea8dc4b6Seschrock vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 296ea8dc4b6Seschrock { 297c3a66015SMatthew Ahrens spa_t *spa = zio->io_spa; 2980f7643c7SGeorge Wilson 29969962b56SMatthew Ahrens ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 300fe319232SJustin T. Gibbs avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); 301fe319232SJustin T. Gibbs avl_add(vdev_queue_type_tree(vq, zio->io_type), zio); 302c3a66015SMatthew Ahrens 30369962b56SMatthew Ahrens mutex_enter(&spa->spa_iokstat_lock); 30469962b56SMatthew Ahrens spa->spa_queue_stats[zio->io_priority].spa_queued++; 30569962b56SMatthew Ahrens if (spa->spa_iokstat != NULL) 306c3a66015SMatthew Ahrens kstat_waitq_enter(spa->spa_iokstat->ks_data); 30769962b56SMatthew Ahrens mutex_exit(&spa->spa_iokstat_lock); 308ea8dc4b6Seschrock } 309ea8dc4b6Seschrock 310ea8dc4b6Seschrock static void 311ea8dc4b6Seschrock vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 312ea8dc4b6Seschrock { 313c3a66015SMatthew Ahrens spa_t *spa = zio->io_spa; 3140f7643c7SGeorge Wilson 31569962b56SMatthew Ahrens ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 316fe319232SJustin T. Gibbs avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); 317fe319232SJustin T. Gibbs avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio); 318c3a66015SMatthew Ahrens 31969962b56SMatthew Ahrens mutex_enter(&spa->spa_iokstat_lock); 32069962b56SMatthew Ahrens ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 32169962b56SMatthew Ahrens spa->spa_queue_stats[zio->io_priority].spa_queued--; 32269962b56SMatthew Ahrens if (spa->spa_iokstat != NULL) 323c3a66015SMatthew Ahrens kstat_waitq_exit(spa->spa_iokstat->ks_data); 32469962b56SMatthew Ahrens mutex_exit(&spa->spa_iokstat_lock); 325c3a66015SMatthew Ahrens } 326c3a66015SMatthew Ahrens 327c3a66015SMatthew Ahrens static void 328c3a66015SMatthew Ahrens vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 329c3a66015SMatthew Ahrens { 330c3a66015SMatthew Ahrens spa_t *spa = zio->io_spa; 33169962b56SMatthew Ahrens ASSERT(MUTEX_HELD(&vq->vq_lock)); 33269962b56SMatthew Ahrens ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 33369962b56SMatthew Ahrens vq->vq_class[zio->io_priority].vqc_active++; 33469962b56SMatthew Ahrens avl_add(&vq->vq_active_tree, zio); 33569962b56SMatthew Ahrens 33669962b56SMatthew Ahrens mutex_enter(&spa->spa_iokstat_lock); 33769962b56SMatthew Ahrens spa->spa_queue_stats[zio->io_priority].spa_active++; 33869962b56SMatthew Ahrens if (spa->spa_iokstat != NULL) 339c3a66015SMatthew Ahrens kstat_runq_enter(spa->spa_iokstat->ks_data); 34069962b56SMatthew Ahrens mutex_exit(&spa->spa_iokstat_lock); 341c3a66015SMatthew Ahrens } 342c3a66015SMatthew Ahrens 343c3a66015SMatthew Ahrens static void 344c3a66015SMatthew Ahrens vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 345c3a66015SMatthew Ahrens { 346c3a66015SMatthew Ahrens spa_t *spa = zio->io_spa; 34769962b56SMatthew Ahrens ASSERT(MUTEX_HELD(&vq->vq_lock)); 34869962b56SMatthew Ahrens ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 34969962b56SMatthew Ahrens vq->vq_class[zio->io_priority].vqc_active--; 35069962b56SMatthew Ahrens avl_remove(&vq->vq_active_tree, zio); 35169962b56SMatthew Ahrens 35269962b56SMatthew Ahrens mutex_enter(&spa->spa_iokstat_lock); 35369962b56SMatthew Ahrens ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 35469962b56SMatthew Ahrens spa->spa_queue_stats[zio->io_priority].spa_active--; 355c3a66015SMatthew Ahrens if (spa->spa_iokstat != NULL) { 356c3a66015SMatthew Ahrens kstat_io_t *ksio = spa->spa_iokstat->ks_data; 357c3a66015SMatthew Ahrens 358c3a66015SMatthew Ahrens kstat_runq_exit(spa->spa_iokstat->ks_data); 359c3a66015SMatthew Ahrens if (zio->io_type == ZIO_TYPE_READ) { 360c3a66015SMatthew Ahrens ksio->reads++; 361c3a66015SMatthew Ahrens ksio->nread += zio->io_size; 362c3a66015SMatthew Ahrens } else if (zio->io_type == ZIO_TYPE_WRITE) { 363c3a66015SMatthew Ahrens ksio->writes++; 364c3a66015SMatthew Ahrens ksio->nwritten += zio->io_size; 365c3a66015SMatthew Ahrens } 366c3a66015SMatthew Ahrens } 36769962b56SMatthew Ahrens mutex_exit(&spa->spa_iokstat_lock); 368ea8dc4b6Seschrock } 369ea8dc4b6Seschrock 370fa9e4066Sahrens static void 371fa9e4066Sahrens vdev_queue_agg_io_done(zio_t *aio) 372fa9e4066Sahrens { 37369962b56SMatthew Ahrens if (aio->io_type == ZIO_TYPE_READ) { 37469962b56SMatthew Ahrens zio_t *pio; 3750f7643c7SGeorge Wilson zio_link_t *zl = NULL; 3760f7643c7SGeorge Wilson while ((pio = zio_walk_parents(aio, &zl)) != NULL) { 377770499e1SDan Kimmel abd_copy_off(pio->io_abd, aio->io_abd, 378770499e1SDan Kimmel 0, pio->io_offset - aio->io_offset, pio->io_size); 37969962b56SMatthew Ahrens } 38069962b56SMatthew Ahrens } 381fa9e4066Sahrens 382770499e1SDan Kimmel abd_free(aio->io_abd); 383fa9e4066Sahrens } 384fa9e4066Sahrens 38569962b56SMatthew Ahrens static int 38669962b56SMatthew Ahrens vdev_queue_class_min_active(zio_priority_t p) 38769962b56SMatthew Ahrens { 38869962b56SMatthew Ahrens switch (p) { 38969962b56SMatthew Ahrens case ZIO_PRIORITY_SYNC_READ: 39069962b56SMatthew Ahrens return (zfs_vdev_sync_read_min_active); 39169962b56SMatthew Ahrens case ZIO_PRIORITY_SYNC_WRITE: 39269962b56SMatthew Ahrens return (zfs_vdev_sync_write_min_active); 39369962b56SMatthew Ahrens case ZIO_PRIORITY_ASYNC_READ: 39469962b56SMatthew Ahrens return (zfs_vdev_async_read_min_active); 39569962b56SMatthew Ahrens case ZIO_PRIORITY_ASYNC_WRITE: 39669962b56SMatthew Ahrens return (zfs_vdev_async_write_min_active); 39769962b56SMatthew Ahrens case ZIO_PRIORITY_SCRUB: 39869962b56SMatthew Ahrens return (zfs_vdev_scrub_min_active); 399*5cabbc6bSPrashanth Sreenivasa case ZIO_PRIORITY_REMOVAL: 400*5cabbc6bSPrashanth Sreenivasa return (zfs_vdev_removal_min_active); 40169962b56SMatthew Ahrens default: 40269962b56SMatthew Ahrens panic("invalid priority %u", p); 40369962b56SMatthew Ahrens return (0); 40469962b56SMatthew Ahrens } 40569962b56SMatthew Ahrens } 40669962b56SMatthew Ahrens 40769962b56SMatthew Ahrens static int 40873527f44SAlex Reece vdev_queue_max_async_writes(spa_t *spa) 40969962b56SMatthew Ahrens { 41069962b56SMatthew Ahrens int writes; 41173527f44SAlex Reece uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 41269962b56SMatthew Ahrens uint64_t min_bytes = zfs_dirty_data_max * 41369962b56SMatthew Ahrens zfs_vdev_async_write_active_min_dirty_percent / 100; 41469962b56SMatthew Ahrens uint64_t max_bytes = zfs_dirty_data_max * 41569962b56SMatthew Ahrens zfs_vdev_async_write_active_max_dirty_percent / 100; 41669962b56SMatthew Ahrens 41773527f44SAlex Reece /* 41873527f44SAlex Reece * Sync tasks correspond to interactive user actions. To reduce the 41973527f44SAlex Reece * execution time of those actions we push data out as fast as possible. 42073527f44SAlex Reece */ 42173527f44SAlex Reece if (spa_has_pending_synctask(spa)) { 42273527f44SAlex Reece return (zfs_vdev_async_write_max_active); 42373527f44SAlex Reece } 42473527f44SAlex Reece 42569962b56SMatthew Ahrens if (dirty < min_bytes) 42669962b56SMatthew Ahrens return (zfs_vdev_async_write_min_active); 42769962b56SMatthew Ahrens if (dirty > max_bytes) 42869962b56SMatthew Ahrens return (zfs_vdev_async_write_max_active); 42969962b56SMatthew Ahrens 43069962b56SMatthew Ahrens /* 43169962b56SMatthew Ahrens * linear interpolation: 43269962b56SMatthew Ahrens * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 43369962b56SMatthew Ahrens * move right by min_bytes 43469962b56SMatthew Ahrens * move up by min_writes 43569962b56SMatthew Ahrens */ 43669962b56SMatthew Ahrens writes = (dirty - min_bytes) * 43769962b56SMatthew Ahrens (zfs_vdev_async_write_max_active - 43869962b56SMatthew Ahrens zfs_vdev_async_write_min_active) / 43969962b56SMatthew Ahrens (max_bytes - min_bytes) + 44069962b56SMatthew Ahrens zfs_vdev_async_write_min_active; 44169962b56SMatthew Ahrens ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 44269962b56SMatthew Ahrens ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 44369962b56SMatthew Ahrens return (writes); 44469962b56SMatthew Ahrens } 44569962b56SMatthew Ahrens 44669962b56SMatthew Ahrens static int 44769962b56SMatthew Ahrens vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 44869962b56SMatthew Ahrens { 44969962b56SMatthew Ahrens switch (p) { 45069962b56SMatthew Ahrens case ZIO_PRIORITY_SYNC_READ: 45169962b56SMatthew Ahrens return (zfs_vdev_sync_read_max_active); 45269962b56SMatthew Ahrens case ZIO_PRIORITY_SYNC_WRITE: 45369962b56SMatthew Ahrens return (zfs_vdev_sync_write_max_active); 45469962b56SMatthew Ahrens case ZIO_PRIORITY_ASYNC_READ: 45569962b56SMatthew Ahrens return (zfs_vdev_async_read_max_active); 45669962b56SMatthew Ahrens case ZIO_PRIORITY_ASYNC_WRITE: 45773527f44SAlex Reece return (vdev_queue_max_async_writes(spa)); 45869962b56SMatthew Ahrens case ZIO_PRIORITY_SCRUB: 45969962b56SMatthew Ahrens return (zfs_vdev_scrub_max_active); 460*5cabbc6bSPrashanth Sreenivasa case ZIO_PRIORITY_REMOVAL: 461*5cabbc6bSPrashanth Sreenivasa return (zfs_vdev_removal_max_active); 46269962b56SMatthew Ahrens default: 46369962b56SMatthew Ahrens panic("invalid priority %u", p); 46469962b56SMatthew Ahrens return (0); 46569962b56SMatthew Ahrens } 46669962b56SMatthew Ahrens } 46769962b56SMatthew Ahrens 46869962b56SMatthew Ahrens /* 46969962b56SMatthew Ahrens * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 47069962b56SMatthew Ahrens * there is no eligible class. 47169962b56SMatthew Ahrens */ 47269962b56SMatthew Ahrens static zio_priority_t 47369962b56SMatthew Ahrens vdev_queue_class_to_issue(vdev_queue_t *vq) 47469962b56SMatthew Ahrens { 47569962b56SMatthew Ahrens spa_t *spa = vq->vq_vdev->vdev_spa; 47669962b56SMatthew Ahrens zio_priority_t p; 47769962b56SMatthew Ahrens 47869962b56SMatthew Ahrens if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 47969962b56SMatthew Ahrens return (ZIO_PRIORITY_NUM_QUEUEABLE); 48069962b56SMatthew Ahrens 48169962b56SMatthew Ahrens /* find a queue that has not reached its minimum # outstanding i/os */ 48269962b56SMatthew Ahrens for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 483fe319232SJustin T. Gibbs if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 48469962b56SMatthew Ahrens vq->vq_class[p].vqc_active < 48569962b56SMatthew Ahrens vdev_queue_class_min_active(p)) 48669962b56SMatthew Ahrens return (p); 48769962b56SMatthew Ahrens } 48869962b56SMatthew Ahrens 48969962b56SMatthew Ahrens /* 49069962b56SMatthew Ahrens * If we haven't found a queue, look for one that hasn't reached its 49169962b56SMatthew Ahrens * maximum # outstanding i/os. 49269962b56SMatthew Ahrens */ 49369962b56SMatthew Ahrens for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 494fe319232SJustin T. Gibbs if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 49569962b56SMatthew Ahrens vq->vq_class[p].vqc_active < 49669962b56SMatthew Ahrens vdev_queue_class_max_active(spa, p)) 49769962b56SMatthew Ahrens return (p); 49869962b56SMatthew Ahrens } 49969962b56SMatthew Ahrens 50069962b56SMatthew Ahrens /* No eligible queued i/os */ 50169962b56SMatthew Ahrens return (ZIO_PRIORITY_NUM_QUEUEABLE); 50269962b56SMatthew Ahrens } 50369962b56SMatthew Ahrens 5046f708f7cSJeff Bonwick /* 5056f708f7cSJeff Bonwick * Compute the range spanned by two i/os, which is the endpoint of the last 5066f708f7cSJeff Bonwick * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 5076f708f7cSJeff Bonwick * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 5086f708f7cSJeff Bonwick * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 5096f708f7cSJeff Bonwick */ 5106f708f7cSJeff Bonwick #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 5116f708f7cSJeff Bonwick #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 512fa9e4066Sahrens 513fa9e4066Sahrens static zio_t * 51469962b56SMatthew Ahrens vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 515fa9e4066Sahrens { 51669962b56SMatthew Ahrens zio_t *first, *last, *aio, *dio, *mandatory, *nio; 51769962b56SMatthew Ahrens uint64_t maxgap = 0; 51869962b56SMatthew Ahrens uint64_t size; 51969962b56SMatthew Ahrens boolean_t stretch = B_FALSE; 520fe319232SJustin T. Gibbs avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type); 52169962b56SMatthew Ahrens enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 52269962b56SMatthew Ahrens 52369962b56SMatthew Ahrens if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 52469962b56SMatthew Ahrens return (NULL); 525fa9e4066Sahrens 52669962b56SMatthew Ahrens first = last = zio; 527fa9e4066Sahrens 52869962b56SMatthew Ahrens if (zio->io_type == ZIO_TYPE_READ) 52969962b56SMatthew Ahrens maxgap = zfs_vdev_read_gap_limit; 5308ad4d6ddSJeff Bonwick 53169962b56SMatthew Ahrens /* 53269962b56SMatthew Ahrens * We can aggregate I/Os that are sufficiently adjacent and of 53369962b56SMatthew Ahrens * the same flavor, as expressed by the AGG_INHERIT flags. 53469962b56SMatthew Ahrens * The latter requirement is necessary so that certain 53569962b56SMatthew Ahrens * attributes of the I/O, such as whether it's a normal I/O 53669962b56SMatthew Ahrens * or a scrub/resilver, can be preserved in the aggregate. 53769962b56SMatthew Ahrens * We can include optional I/Os, but don't allow them 53869962b56SMatthew Ahrens * to begin a range as they add no benefit in that situation. 53969962b56SMatthew Ahrens */ 540f94275ceSAdam Leventhal 54169962b56SMatthew Ahrens /* 54269962b56SMatthew Ahrens * We keep track of the last non-optional I/O. 54369962b56SMatthew Ahrens */ 54469962b56SMatthew Ahrens mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 545f94275ceSAdam Leventhal 54669962b56SMatthew Ahrens /* 54769962b56SMatthew Ahrens * Walk backwards through sufficiently contiguous I/Os 5485b062782SMatthew Ahrens * recording the last non-optional I/O. 54969962b56SMatthew Ahrens */ 55069962b56SMatthew Ahrens while ((dio = AVL_PREV(t, first)) != NULL && 55169962b56SMatthew Ahrens (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 55269962b56SMatthew Ahrens IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 553*5cabbc6bSPrashanth Sreenivasa IO_GAP(dio, first) <= maxgap && 554*5cabbc6bSPrashanth Sreenivasa dio->io_type == zio->io_type) { 55569962b56SMatthew Ahrens first = dio; 55669962b56SMatthew Ahrens if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 55769962b56SMatthew Ahrens mandatory = first; 55869962b56SMatthew Ahrens } 559f94275ceSAdam Leventhal 56069962b56SMatthew Ahrens /* 56169962b56SMatthew Ahrens * Skip any initial optional I/Os. 56269962b56SMatthew Ahrens */ 56369962b56SMatthew Ahrens while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 56469962b56SMatthew Ahrens first = AVL_NEXT(t, first); 56569962b56SMatthew Ahrens ASSERT(first != NULL); 56669962b56SMatthew Ahrens } 5676f708f7cSJeff Bonwick 56869962b56SMatthew Ahrens /* 56969962b56SMatthew Ahrens * Walk forward through sufficiently contiguous I/Os. 5705b062782SMatthew Ahrens * The aggregation limit does not apply to optional i/os, so that 5715b062782SMatthew Ahrens * we can issue contiguous writes even if they are larger than the 5725b062782SMatthew Ahrens * aggregation limit. 57369962b56SMatthew Ahrens */ 57469962b56SMatthew Ahrens while ((dio = AVL_NEXT(t, last)) != NULL && 57569962b56SMatthew Ahrens (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 5765b062782SMatthew Ahrens (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit || 5775b062782SMatthew Ahrens (dio->io_flags & ZIO_FLAG_OPTIONAL)) && 578*5cabbc6bSPrashanth Sreenivasa IO_GAP(last, dio) <= maxgap && 579*5cabbc6bSPrashanth Sreenivasa dio->io_type == zio->io_type) { 58069962b56SMatthew Ahrens last = dio; 58169962b56SMatthew Ahrens if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 58269962b56SMatthew Ahrens mandatory = last; 58369962b56SMatthew Ahrens } 584f94275ceSAdam Leventhal 58569962b56SMatthew Ahrens /* 58669962b56SMatthew Ahrens * Now that we've established the range of the I/O aggregation 58769962b56SMatthew Ahrens * we must decide what to do with trailing optional I/Os. 58869962b56SMatthew Ahrens * For reads, there's nothing to do. While we are unable to 58969962b56SMatthew Ahrens * aggregate further, it's possible that a trailing optional 59069962b56SMatthew Ahrens * I/O would allow the underlying device to aggregate with 59169962b56SMatthew Ahrens * subsequent I/Os. We must therefore determine if the next 59269962b56SMatthew Ahrens * non-optional I/O is close enough to make aggregation 59369962b56SMatthew Ahrens * worthwhile. 59469962b56SMatthew Ahrens */ 59569962b56SMatthew Ahrens if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 59669962b56SMatthew Ahrens zio_t *nio = last; 59769962b56SMatthew Ahrens while ((dio = AVL_NEXT(t, nio)) != NULL && 59869962b56SMatthew Ahrens IO_GAP(nio, dio) == 0 && 59969962b56SMatthew Ahrens IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 60069962b56SMatthew Ahrens nio = dio; 60169962b56SMatthew Ahrens if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 60269962b56SMatthew Ahrens stretch = B_TRUE; 60369962b56SMatthew Ahrens break; 604f94275ceSAdam Leventhal } 605f94275ceSAdam Leventhal } 60669962b56SMatthew Ahrens } 607f94275ceSAdam Leventhal 60869962b56SMatthew Ahrens if (stretch) { 6095b062782SMatthew Ahrens /* 6105b062782SMatthew Ahrens * We are going to include an optional io in our aggregated 6115b062782SMatthew Ahrens * span, thus closing the write gap. Only mandatory i/os can 6125b062782SMatthew Ahrens * start aggregated spans, so make sure that the next i/o 6135b062782SMatthew Ahrens * after our span is mandatory. 6145b062782SMatthew Ahrens */ 61569962b56SMatthew Ahrens dio = AVL_NEXT(t, last); 61669962b56SMatthew Ahrens dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 61769962b56SMatthew Ahrens } else { 6185b062782SMatthew Ahrens /* do not include the optional i/o */ 61969962b56SMatthew Ahrens while (last != mandatory && last != first) { 62069962b56SMatthew Ahrens ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 62169962b56SMatthew Ahrens last = AVL_PREV(t, last); 62269962b56SMatthew Ahrens ASSERT(last != NULL); 623f94275ceSAdam Leventhal } 624fa9e4066Sahrens } 625fa9e4066Sahrens 62669962b56SMatthew Ahrens if (first == last) 62769962b56SMatthew Ahrens return (NULL); 62869962b56SMatthew Ahrens 62969962b56SMatthew Ahrens size = IO_SPAN(first, last); 6305b062782SMatthew Ahrens ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 63169962b56SMatthew Ahrens 63269962b56SMatthew Ahrens aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 633770499e1SDan Kimmel abd_alloc_for_io(size, B_TRUE), size, first->io_type, 634770499e1SDan Kimmel zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 63569962b56SMatthew Ahrens vdev_queue_agg_io_done, NULL); 63669962b56SMatthew Ahrens aio->io_timestamp = first->io_timestamp; 63769962b56SMatthew Ahrens 63869962b56SMatthew Ahrens nio = first; 63969962b56SMatthew Ahrens do { 64069962b56SMatthew Ahrens dio = nio; 64169962b56SMatthew Ahrens nio = AVL_NEXT(t, dio); 64269962b56SMatthew Ahrens ASSERT3U(dio->io_type, ==, aio->io_type); 64369962b56SMatthew Ahrens 64469962b56SMatthew Ahrens if (dio->io_flags & ZIO_FLAG_NODATA) { 64569962b56SMatthew Ahrens ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 646770499e1SDan Kimmel abd_zero_off(aio->io_abd, 647770499e1SDan Kimmel dio->io_offset - aio->io_offset, dio->io_size); 64869962b56SMatthew Ahrens } else if (dio->io_type == ZIO_TYPE_WRITE) { 649770499e1SDan Kimmel abd_copy_off(aio->io_abd, dio->io_abd, 650770499e1SDan Kimmel dio->io_offset - aio->io_offset, 0, dio->io_size); 65169962b56SMatthew Ahrens } 652a3f829aeSBill Moore 65369962b56SMatthew Ahrens zio_add_child(dio, aio); 65469962b56SMatthew Ahrens vdev_queue_io_remove(vq, dio); 65569962b56SMatthew Ahrens zio_vdev_io_bypass(dio); 65669962b56SMatthew Ahrens zio_execute(dio); 65769962b56SMatthew Ahrens } while (dio != last); 65869962b56SMatthew Ahrens 65969962b56SMatthew Ahrens return (aio); 66069962b56SMatthew Ahrens } 66169962b56SMatthew Ahrens 66269962b56SMatthew Ahrens static zio_t * 66369962b56SMatthew Ahrens vdev_queue_io_to_issue(vdev_queue_t *vq) 66469962b56SMatthew Ahrens { 66569962b56SMatthew Ahrens zio_t *zio, *aio; 66669962b56SMatthew Ahrens zio_priority_t p; 66769962b56SMatthew Ahrens avl_index_t idx; 668fe319232SJustin T. Gibbs avl_tree_t *tree; 66969962b56SMatthew Ahrens zio_t search; 67069962b56SMatthew Ahrens 67169962b56SMatthew Ahrens again: 67269962b56SMatthew Ahrens ASSERT(MUTEX_HELD(&vq->vq_lock)); 673fa9e4066Sahrens 67469962b56SMatthew Ahrens p = vdev_queue_class_to_issue(vq); 675fa9e4066Sahrens 67669962b56SMatthew Ahrens if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 67769962b56SMatthew Ahrens /* No eligible queued i/os */ 67869962b56SMatthew Ahrens return (NULL); 679fa9e4066Sahrens } 680fa9e4066Sahrens 68169962b56SMatthew Ahrens /* 68269962b56SMatthew Ahrens * For LBA-ordered queues (async / scrub), issue the i/o which follows 68369962b56SMatthew Ahrens * the most recently issued i/o in LBA (offset) order. 68469962b56SMatthew Ahrens * 68569962b56SMatthew Ahrens * For FIFO queues (sync), issue the i/o with the lowest timestamp. 68669962b56SMatthew Ahrens */ 687fe319232SJustin T. Gibbs tree = vdev_queue_class_tree(vq, p); 68869962b56SMatthew Ahrens search.io_timestamp = 0; 68969962b56SMatthew Ahrens search.io_offset = vq->vq_last_offset + 1; 690fe319232SJustin T. Gibbs VERIFY3P(avl_find(tree, &search, &idx), ==, NULL); 691fe319232SJustin T. Gibbs zio = avl_nearest(tree, idx, AVL_AFTER); 69269962b56SMatthew Ahrens if (zio == NULL) 693fe319232SJustin T. Gibbs zio = avl_first(tree); 69469962b56SMatthew Ahrens ASSERT3U(zio->io_priority, ==, p); 69569962b56SMatthew Ahrens 69669962b56SMatthew Ahrens aio = vdev_queue_aggregate(vq, zio); 69769962b56SMatthew Ahrens if (aio != NULL) 69869962b56SMatthew Ahrens zio = aio; 69969962b56SMatthew Ahrens else 70069962b56SMatthew Ahrens vdev_queue_io_remove(vq, zio); 701fa9e4066Sahrens 702f94275ceSAdam Leventhal /* 703f94275ceSAdam Leventhal * If the I/O is or was optional and therefore has no data, we need to 704f94275ceSAdam Leventhal * simply discard it. We need to drop the vdev queue's lock to avoid a 705f94275ceSAdam Leventhal * deadlock that we could encounter since this I/O will complete 706f94275ceSAdam Leventhal * immediately. 707f94275ceSAdam Leventhal */ 70869962b56SMatthew Ahrens if (zio->io_flags & ZIO_FLAG_NODATA) { 709f94275ceSAdam Leventhal mutex_exit(&vq->vq_lock); 71069962b56SMatthew Ahrens zio_vdev_io_bypass(zio); 71169962b56SMatthew Ahrens zio_execute(zio); 712f94275ceSAdam Leventhal mutex_enter(&vq->vq_lock); 713f94275ceSAdam Leventhal goto again; 714f94275ceSAdam Leventhal } 715f94275ceSAdam Leventhal 71669962b56SMatthew Ahrens vdev_queue_pending_add(vq, zio); 71769962b56SMatthew Ahrens vq->vq_last_offset = zio->io_offset; 718fa9e4066Sahrens 71969962b56SMatthew Ahrens return (zio); 720fa9e4066Sahrens } 721fa9e4066Sahrens 722fa9e4066Sahrens zio_t * 723fa9e4066Sahrens vdev_queue_io(zio_t *zio) 724fa9e4066Sahrens { 725fa9e4066Sahrens vdev_queue_t *vq = &zio->io_vd->vdev_queue; 726fa9e4066Sahrens zio_t *nio; 727fa9e4066Sahrens 728fa9e4066Sahrens if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 729fa9e4066Sahrens return (zio); 730fa9e4066Sahrens 73169962b56SMatthew Ahrens /* 73269962b56SMatthew Ahrens * Children i/os inherent their parent's priority, which might 73369962b56SMatthew Ahrens * not match the child's i/o type. Fix it up here. 73469962b56SMatthew Ahrens */ 73569962b56SMatthew Ahrens if (zio->io_type == ZIO_TYPE_READ) { 73669962b56SMatthew Ahrens if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 73769962b56SMatthew Ahrens zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 738*5cabbc6bSPrashanth Sreenivasa zio->io_priority != ZIO_PRIORITY_SCRUB && 739*5cabbc6bSPrashanth Sreenivasa zio->io_priority != ZIO_PRIORITY_REMOVAL) 74069962b56SMatthew Ahrens zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 74169962b56SMatthew Ahrens } else { 74269962b56SMatthew Ahrens ASSERT(zio->io_type == ZIO_TYPE_WRITE); 74369962b56SMatthew Ahrens if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 744*5cabbc6bSPrashanth Sreenivasa zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE && 745*5cabbc6bSPrashanth Sreenivasa zio->io_priority != ZIO_PRIORITY_REMOVAL) 74669962b56SMatthew Ahrens zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 74769962b56SMatthew Ahrens } 748fa9e4066Sahrens 74969962b56SMatthew Ahrens zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 750fa9e4066Sahrens 751fa9e4066Sahrens mutex_enter(&vq->vq_lock); 752c55e05cbSMatthew Ahrens zio->io_timestamp = gethrtime(); 753ea8dc4b6Seschrock vdev_queue_io_add(vq, zio); 75469962b56SMatthew Ahrens nio = vdev_queue_io_to_issue(vq); 755fa9e4066Sahrens mutex_exit(&vq->vq_lock); 756fa9e4066Sahrens 757e05725b1Sbonwick if (nio == NULL) 758e05725b1Sbonwick return (NULL); 759e05725b1Sbonwick 760e05725b1Sbonwick if (nio->io_done == vdev_queue_agg_io_done) { 761e05725b1Sbonwick zio_nowait(nio); 762e05725b1Sbonwick return (NULL); 763e05725b1Sbonwick } 764fa9e4066Sahrens 765e05725b1Sbonwick return (nio); 766fa9e4066Sahrens } 767fa9e4066Sahrens 768fa9e4066Sahrens void 769fa9e4066Sahrens vdev_queue_io_done(zio_t *zio) 770fa9e4066Sahrens { 771fa9e4066Sahrens vdev_queue_t *vq = &zio->io_vd->vdev_queue; 77269962b56SMatthew Ahrens zio_t *nio; 773fa9e4066Sahrens 774fa9e4066Sahrens mutex_enter(&vq->vq_lock); 775fa9e4066Sahrens 776c3a66015SMatthew Ahrens vdev_queue_pending_remove(vq, zio); 777fa9e4066Sahrens 778c55e05cbSMatthew Ahrens vq->vq_io_complete_ts = gethrtime(); 779283b8460SGeorge.Wilson 78069962b56SMatthew Ahrens while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 781fa9e4066Sahrens mutex_exit(&vq->vq_lock); 782e05725b1Sbonwick if (nio->io_done == vdev_queue_agg_io_done) { 783e05725b1Sbonwick zio_nowait(nio); 784e05725b1Sbonwick } else { 785fa9e4066Sahrens zio_vdev_io_reissue(nio); 786e05725b1Sbonwick zio_execute(nio); 787e05725b1Sbonwick } 788fa9e4066Sahrens mutex_enter(&vq->vq_lock); 789fa9e4066Sahrens } 790fa9e4066Sahrens 791fa9e4066Sahrens mutex_exit(&vq->vq_lock); 792fa9e4066Sahrens } 793