vdev_queue.c revision 321574
14Srgrimes/* 2549Srgrimes * CDDL HEADER START 34Srgrimes * 44Srgrimes * The contents of this file are subject to the terms of the 54Srgrimes * Common Development and Distribution License (the "License"). 64Srgrimes * You may not use this file except in compliance with the License. 74Srgrimes * 84Srgrimes * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 94Srgrimes * or http://www.opensolaris.org/os/licensing. 104Srgrimes * See the License for the specific language governing permissions 114Srgrimes * and limitations under the License. 124Srgrimes * 134Srgrimes * When distributing Covered Code, include this CDDL HEADER in each 144Srgrimes * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 154Srgrimes * If applicable, add the following below this CDDL HEADER, with the 164Srgrimes * fields enclosed by brackets "[]" replaced with your own identifying 174Srgrimes * information: Portions Copyright [yyyy] [name of copyright owner] 184Srgrimes * 194Srgrimes * CDDL HEADER END 204Srgrimes */ 214Srgrimes/* 224Srgrimes * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 234Srgrimes * Use is subject to license terms. 244Srgrimes */ 254Srgrimes 264Srgrimes/* 274Srgrimes * Copyright (c) 2012, 2017 by Delphix. All rights reserved. 284Srgrimes * Copyright (c) 2014 Integros [integros.com] 294Srgrimes */ 304Srgrimes 314Srgrimes#include <sys/zfs_context.h> 324Srgrimes#include <sys/vdev_impl.h> 334Srgrimes#include <sys/spa_impl.h> 344Srgrimes#include <sys/zio.h> 354Srgrimes#include <sys/avl.h> 364Srgrimes#include <sys/dsl_pool.h> 37549Srgrimes#include <sys/metaslab_impl.h> 3825164Speter 394Srgrimes/* 404Srgrimes * ZFS I/O Scheduler 41556Srgrimes * --------------- 4213226Swollman * 4313228Swollman * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The 4413265Swollman * I/O scheduler determines when and in what order those operations are 4514328Speter * issued. The I/O scheduler divides operations into six I/O classes 4614825Swollman * prioritized in the following order: sync read, sync write, async read, 4725164Speter * async write, scrub/resilver and trim. Each queue defines the minimum and 4818252Sphk * maximum number of concurrent operations that may be issued to the device. 494Srgrimes * In addition, the device has an aggregate maximum. Note that the sum of the 501549Srgrimes * per-queue minimums must not exceed the aggregate maximum, and if the 511549Srgrimes * aggregate maximum is equal to or greater than the sum of the per-queue 5211390Sbde * maximums, the per-queue minimum has no effect. 531549Srgrimes * 541549Srgrimes * For many physical devices, throughput increases with the number of 551549Srgrimes * concurrent operations, but latency typically suffers. Further, physical 561549Srgrimes * devices typically have a limit at which more concurrent operations have no 571549Srgrimes * effect on throughput or can actually cause it to decrease. 581549Srgrimes * 591549Srgrimes * The scheduler selects the next operation to issue by first looking for an 601549Srgrimes * I/O class whose minimum has not been satisfied. Once all are satisfied and 611549Srgrimes * the aggregate maximum has not been hit, the scheduler looks for classes 621549Srgrimes * whose maximum has not been satisfied. Iteration through the I/O classes is 637090Sbde * done in the order specified above. No further operations are issued if the 641549Srgrimes * aggregate maximum number of concurrent operations has been hit or if there 652254Ssos * are no operations queued for an I/O class that has not hit its maximum. 661549Srgrimes * Every time an I/O is queued or an operation completes, the I/O scheduler 671549Srgrimes * looks for new operations to issue. 6812662Sdg * 694Srgrimes * All I/O classes have a fixed maximum number of outstanding operations 70556Srgrimes * except for the async write class. Asynchronous writes represent the data 712056Swollman * that is committed to stable storage during the syncing stage for 72556Srgrimes * transaction groups (see txg.c). Transaction groups enter the syncing state 73556Srgrimes * periodically so the number of queued async writes will quickly burst up and 74990Sdg * then bleed down to zero. Rather than servicing them as quickly as possible, 752056Swollman * the I/O scheduler changes the maximum number of active async write I/Os 76990Sdg * according to the amount of dirty data in the pool (see dsl_pool.c). Since 77990Sdg * both throughput and latency typically increase with the number of 78990Sdg * concurrent operations issued to physical devices, reducing the burstiness 792056Swollman * in the number of concurrent operations also stabilizes the response time of 80990Sdg * operations from other -- and in particular synchronous -- queues. In broad 81990Sdg * strokes, the I/O scheduler will issue more concurrent operations from the 822056Swollman * async write queue as there's more dirty data in the pool. 8312662Sdg * 8412662Sdg * Async Writes 8522521Sdyson * 862056Swollman * The number of concurrent operations issued for the async write I/O class 8712662Sdg * follows a piece-wise linear function defined by a few adjustable points. 882056Swollman * 8912662Sdg * | o---------| <-- zfs_vdev_async_write_max_active 909507Sdg * ^ | /^ | 9112662Sdg * | | / | | 924Srgrimes * active | / | | 9312662Sdg * I/O | / | | 942056Swollman * count | / | | 952056Swollman * | / | | 96556Srgrimes * |------------o | | <-- zfs_vdev_async_write_min_active 977090Sbde * 0|____________^______|_________| 987090Sbde * 0% | | 100% of zfs_dirty_data_max 992772Swollman * | | 1002772Swollman * | `-- zfs_vdev_async_write_active_max_dirty_percent 1012056Swollman * `--------- zfs_vdev_async_write_active_min_dirty_percent 1024193Sbde * 1032056Swollman * Until the amount of dirty data exceeds a minimum percentage of the dirty 1042056Swollman * data allowed in the pool, the I/O scheduler will limit the number of 1054193Sbde * concurrent operations to the minimum. As that threshold is crossed, the 1062056Swollman * number of concurrent operations issued increases linearly to the maximum at 1072056Swollman * the specified maximum percentage of the dirty data allowed in the pool. 1082056Swollman * 1094819Sphk * Ideally, the amount of dirty data on a busy pool will stay in the sloped 1107090Sbde * part of the function between zfs_vdev_async_write_active_min_dirty_percent 11125164Speter * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the 11225164Speter * maximum percentage, this indicates that the rate of incoming data is 11325164Speter * greater than the rate that the backend storage can handle. In this case, we 11414825Swollman * must further throttle incoming writes (see dmu_tx_delay() for details). 11514825Swollman */ 11614825Swollman 117556Srgrimes/* 1184193Sbde * The maximum number of I/Os active to each device. Ideally, this will be >= 1192056Swollman * the sum of each queue's max_active. It must be at least the sum of each 12011875Smarkm * queue's min_active. 1214Srgrimes */ 12211390Sbdeuint32_t zfs_vdev_max_active = 1000; 12311390Sbde 12411390Sbde/* 12511390Sbde * Per-queue limits on the number of I/Os active to each device. If the 12612929Sdg * sum of the queue's max_active is < zfs_vdev_max_active, then the 12710358Sjulian * min_active comes into play. We will send min_active from each queue, 12824112Skato * and then select from queues in the order defined by zio_priority_t. 12917014Swollman * 13024112Skato * In general, smaller max_active's will lead to lower latency of synchronous 13124112Skato * operations. Larger max_active's may lead to higher overall throughput, 13224112Skato * depending on underlying storage. 13325083Sjdp * 13413085Sdg * The ratio of the queues' max_actives determines the balance of performance 13510653Sdg * between reads, writes, and scrubs. E.g., increasing 13610358Sjulian * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete 13710358Sjulian * more quickly, but reads and writes to have higher latency and lower 1384Srgrimes * throughput. 1392426Sdg */ 1402426Sdguint32_t zfs_vdev_sync_read_min_active = 10; 1412426Sdguint32_t zfs_vdev_sync_read_max_active = 10; 1421298Sdguint32_t zfs_vdev_sync_write_min_active = 10; 1431298Sdguint32_t zfs_vdev_sync_write_max_active = 10; 1441298Sdguint32_t zfs_vdev_async_read_min_active = 1; 1451298Sdguint32_t zfs_vdev_async_read_max_active = 3; 1461298Sdguint32_t zfs_vdev_async_write_min_active = 1; 1472426Sdguint32_t zfs_vdev_async_write_max_active = 10; 1482426Sdguint32_t zfs_vdev_scrub_min_active = 1; 1492426Sdguint32_t zfs_vdev_scrub_max_active = 2; 1501549Srgrimesuint32_t zfs_vdev_trim_min_active = 1; 151556Srgrimes/* 15215565Sphk * TRIM max active is large in comparison to the other values due to the fact 153556Srgrimes * that TRIM IOs are coalesced at the device layer. This value is set such 1542818Sache * that a typical SSD can process the queued IOs in a single request. 15512623Sphk */ 15615392Sphkuint32_t zfs_vdev_trim_max_active = 64; 15712623Sphk 15812623Sphk 15912623Sphk/* 16012623Sphk * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent 16112623Sphk * dirty data, use zfs_vdev_async_write_min_active. When it has more than 16212623Sphk * zfs_vdev_async_write_active_max_dirty_percent, use 16312623Sphk * zfs_vdev_async_write_max_active. The value is linearly interpolated 16412623Sphk * between min and max. 16512623Sphk */ 16612623Sphkint zfs_vdev_async_write_active_min_dirty_percent = 30; 16712623Sphkint zfs_vdev_async_write_active_max_dirty_percent = 60; 16812623Sphk 16912623Sphk/* 17012623Sphk * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O. 17112623Sphk * For read I/Os, we also aggregate across small adjacency gaps; for writes 17212623Sphk * we include spans of optional I/Os to aid aggregation at the disk even when 17312623Sphk * they aren't able to help us aggregate at this level. 17412623Sphk */ 17512623Sphkint zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE; 17612623Sphkint zfs_vdev_read_gap_limit = 32 << 10; 17712623Sphkint zfs_vdev_write_gap_limit = 4 << 10; 17812623Sphk 17912722Sphk/* 18012722Sphk * Define the queue depth percentage for each top-level. This percentage is 1814Srgrimes * used in conjunction with zfs_vdev_async_max_active to determine how many 1824Srgrimes * allocations a specific top-level vdev should handle. Once the queue depth 1834Srgrimes * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100 1849744Sdg * then allocator will stop allocating blocks on that top-level device. 185974Sdg * The default kernel setting is 1000% which will yield 100 allocations per 1869578Sdg * device. For userland testing, the default setting is 300% which equates 1879578Sdg * to 30 allocations per device. 1889578Sdg */ 18917559Swollman#ifdef _KERNEL 1908426Swollmanint zfs_vdev_queue_depth_pct = 1000; 19112722Sphk#else 1921549Srgrimesint zfs_vdev_queue_depth_pct = 300; 19312722Sphk#endif 19417559Swollman 195556Srgrimes 1961549Srgrimes#ifdef __FreeBSD__ 1971549Srgrimes#ifdef _KERNEL 19810358SjulianSYSCTL_DECL(_vfs_zfs_vdev); 19911390Sbde 20011390Sbdestatic int sysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS); 2014SrgrimesSYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_min_dirty_percent, 2024Srgrimes CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 2034Srgrimes sysctl_zfs_async_write_active_min_dirty_percent, "I", 2043489Sphk "Percentage of async write dirty data below which " 205798Swollman "async_write_min_active is used."); 20612243Sphk 2073502Sphkstatic int sysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS); 2084SrgrimesSYSCTL_PROC(_vfs_zfs_vdev, OID_AUTO, async_write_active_max_dirty_percent, 2094819Sphk CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int), 2104819Sphk sysctl_zfs_async_write_active_max_dirty_percent, "I", 2114819Sphk "Percentage of async write dirty data above which " 2124Srgrimes "async_write_max_active is used."); 2134Srgrimes 2144SrgrimesSYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, max_active, CTLFLAG_RWTUN, 215556Srgrimes &zfs_vdev_max_active, 0, 21625164Speter "The maximum number of I/Os of all types active for each device."); 21725164Speter 21825164Speter#define ZFS_VDEV_QUEUE_KNOB_MIN(name) \ 21917014SwollmanSYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _min_active, CTLFLAG_RWTUN,\ 2202014Swollman &zfs_vdev_ ## name ## _min_active, 0, \ 22124112Skato "Initial number of I/O requests of type " #name \ 22224112Skato " active for each device"); 22317014Swollman 22417014Swollman#define ZFS_VDEV_QUEUE_KNOB_MAX(name) \ 22517014SwollmanSYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, name ## _max_active, CTLFLAG_RWTUN,\ 22610616Sdg &zfs_vdev_ ## name ## _max_active, 0, \ 2279578Sdg "Maximum number of I/O requests of type " #name \ 2289578Sdg " active for each device"); 2299578Sdg 2309578SdgZFS_VDEV_QUEUE_KNOB_MIN(sync_read); 2319578SdgZFS_VDEV_QUEUE_KNOB_MAX(sync_read); 2324SrgrimesZFS_VDEV_QUEUE_KNOB_MIN(sync_write); 2339578SdgZFS_VDEV_QUEUE_KNOB_MAX(sync_write); 2349578SdgZFS_VDEV_QUEUE_KNOB_MIN(async_read); 2359578SdgZFS_VDEV_QUEUE_KNOB_MAX(async_read); 2369578SdgZFS_VDEV_QUEUE_KNOB_MIN(async_write); 2379578SdgZFS_VDEV_QUEUE_KNOB_MAX(async_write); 2389578SdgZFS_VDEV_QUEUE_KNOB_MIN(scrub); 2399578SdgZFS_VDEV_QUEUE_KNOB_MAX(scrub); 2409578SdgZFS_VDEV_QUEUE_KNOB_MIN(trim); 2419578SdgZFS_VDEV_QUEUE_KNOB_MAX(trim); 2429578Sdg 24312290Sphk#undef ZFS_VDEV_QUEUE_KNOB 2449578Sdg 2459578SdgSYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, aggregation_limit, CTLFLAG_RWTUN, 2469578Sdg &zfs_vdev_aggregation_limit, 0, 2479578Sdg "I/O requests are aggregated up to this size"); 2484SrgrimesSYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, read_gap_limit, CTLFLAG_RWTUN, 24917559Swollman &zfs_vdev_read_gap_limit, 0, 25017559Swollman "Acceptable gap between two reads being aggregated"); 25117559SwollmanSYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, write_gap_limit, CTLFLAG_RWTUN, 25217559Swollman &zfs_vdev_write_gap_limit, 0, 25317559Swollman "Acceptable gap between two writes being aggregated"); 2544SrgrimesSYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, queue_depth_pct, CTLFLAG_RWTUN, 2554Srgrimes &zfs_vdev_queue_depth_pct, 0, 2564Srgrimes "Queue depth percentage for each top-level"); 2574Srgrimes 2584Srgrimesstatic int 2594Srgrimessysctl_zfs_async_write_active_min_dirty_percent(SYSCTL_HANDLER_ARGS) 2604Srgrimes{ 2614Srgrimes int val, err; 2624Srgrimes 2634Srgrimes val = zfs_vdev_async_write_active_min_dirty_percent; 2644Srgrimes err = sysctl_handle_int(oidp, &val, 0, req); 2654Srgrimes if (err != 0 || req->newptr == NULL) 2664Srgrimes return (err); 2674Srgrimes 2684Srgrimes if (val < 0 || val > 100 || 2694Srgrimes val >= zfs_vdev_async_write_active_max_dirty_percent) 2704Srgrimes return (EINVAL); 2714Srgrimes 2724Srgrimes zfs_vdev_async_write_active_min_dirty_percent = val; 2734Srgrimes 2744Srgrimes return (0); 2754Srgrimes} 2764Srgrimes 2774Srgrimesstatic int 2784Srgrimessysctl_zfs_async_write_active_max_dirty_percent(SYSCTL_HANDLER_ARGS) 2794Srgrimes{ 280990Sdg int val, err; 281990Sdg 282990Sdg val = zfs_vdev_async_write_active_max_dirty_percent; 283990Sdg err = sysctl_handle_int(oidp, &val, 0, req); 284990Sdg if (err != 0 || req->newptr == NULL) 285990Sdg return (err); 286990Sdg 287990Sdg if (val < 0 || val > 100 || 288990Sdg val <= zfs_vdev_async_write_active_min_dirty_percent) 289990Sdg return (EINVAL); 290990Sdg 291990Sdg zfs_vdev_async_write_active_max_dirty_percent = val; 2924Srgrimes 2935837Sdg return (0); 2946327Sdg} 2955837Sdg#endif 29620578Sdg#endif 2975837Sdg 29819828Sdysonint 2995455Sdgvdev_queue_offset_compare(const void *x1, const void *x2) 3004Srgrimes{ 3014Srgrimes const zio_t *z1 = x1; 3024Srgrimes const zio_t *z2 = x2; 3032422Sdg 3044Srgrimes if (z1->io_offset < z2->io_offset) 3051298Sdg return (-1); 3061298Sdg if (z1->io_offset > z2->io_offset) 3071298Sdg return (1); 3085455Sdg 3095455Sdg if (z1 < z2) 3105455Sdg return (-1); 31124852Sdyson if (z1 > z2) 31224852Sdyson return (1); 3135455Sdg 31415583Sphk return (0); 3151298Sdg} 3161298Sdg 3171298Sdgstatic inline avl_tree_t * 3181298Sdgvdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) 3191298Sdg{ 3204Srgrimes return (&vq->vq_class[p].vqc_queued_tree); 3214Srgrimes} 3224Srgrimes 3234Srgrimesstatic inline avl_tree_t * 3244Srgrimesvdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) 3254Srgrimes{ 3264Srgrimes if (t == ZIO_TYPE_READ) 3274Srgrimes return (&vq->vq_read_offset_tree); 3284Srgrimes else if (t == ZIO_TYPE_WRITE) 3291321Sdg return (&vq->vq_write_offset_tree); 3304Srgrimes else 3314Srgrimes return (NULL); 3324Srgrimes} 3334Srgrimes 3344Srgrimesint 335556Srgrimesvdev_queue_timestamp_compare(const void *x1, const void *x2) 3362426Sdg{ 3371549Srgrimes const zio_t *z1 = x1; 33820146Sdyson const zio_t *z2 = x2; 3391887Sdg 3402426Sdg if (z1->io_timestamp < z2->io_timestamp) 3412426Sdg return (-1); 3422426Sdg if (z1->io_timestamp > z2->io_timestamp) 34320146Sdyson return (1); 3442426Sdg 3451887Sdg if (z1->io_offset < z2->io_offset) 34620146Sdyson return (-1); 3471887Sdg if (z1->io_offset > z2->io_offset) 3481887Sdg return (1); 3495455Sdg 3505455Sdg if (z1 < z2) 3515455Sdg return (-1); 3525455Sdg if (z1 > z2) 3531549Srgrimes return (1); 35425164Speter 35525164Speter return (0); 35625164Speter} 35725164Speter 35825164Spetervoid 35925164Spetervdev_queue_init(vdev_t *vd) 36025164Speter{ 36125164Speter vdev_queue_t *vq = &vd->vdev_queue; 36225164Speter 36325164Speter mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 36425164Speter vq->vq_vdev = vd; 36525164Speter 36625164Speter avl_create(&vq->vq_active_tree, vdev_queue_offset_compare, 36725164Speter sizeof (zio_t), offsetof(struct zio, io_queue_node)); 36825164Speter avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ), 36925164Speter vdev_queue_offset_compare, sizeof (zio_t), 37025164Speter offsetof(struct zio, io_offset_node)); 3714Srgrimes avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), 3724Srgrimes vdev_queue_offset_compare, sizeof (zio_t), 3734Srgrimes offsetof(struct zio, io_offset_node)); 3744Srgrimes 37515722Swollman for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 37621737Sdg int (*compfn) (const void *, const void *); 37721737Sdg 37821737Sdg /* 37921737Sdg * The synchronous i/o queues are dispatched in FIFO rather 38021737Sdg * than LBA order. This provides more consistent latency for 38121737Sdg * these i/os. 38221737Sdg */ 38321737Sdg if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) 38415722Swollman compfn = vdev_queue_timestamp_compare; 38515722Swollman else 3864Srgrimes compfn = vdev_queue_offset_compare; 3874Srgrimes 3884Srgrimes avl_create(vdev_queue_class_tree(vq, p), compfn, 3894Srgrimes sizeof (zio_t), offsetof(struct zio, io_queue_node)); 3904Srgrimes } 3914Srgrimes 3924475Sbde vq->vq_lastoffset = 0; 39320471Sjkh} 39420471Sjkh 39520471Sjkhvoid 39620471Sjkhvdev_queue_fini(vdev_t *vd) 39720471Sjkh{ 39818702Sjkh vdev_queue_t *vq = &vd->vdev_queue; 3993907Sjkh 40010666Sbde for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) 40110666Sbde avl_destroy(vdev_queue_class_tree(vq, p)); 40220471Sjkh avl_destroy(&vq->vq_active_tree); 4034Srgrimes avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); 4042422Sdg avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); 4054Srgrimes 4061298Sdg mutex_destroy(&vq->vq_lock); 4071298Sdg} 4081298Sdg 4091298Sdgstatic void 41010782Sdgvdev_queue_io_add(vdev_queue_t *vq, zio_t *zio) 41110616Sdg{ 41210616Sdg spa_t *spa = zio->io_spa; 4131298Sdg avl_tree_t *qtt; 4141298Sdg 4154Srgrimes ASSERT(MUTEX_HELD(&vq->vq_lock)); 4164Srgrimes ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 4174Srgrimes avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio); 4181887Sdg qtt = vdev_queue_type_tree(vq, zio->io_type); 4194Srgrimes if (qtt) 4204Srgrimes avl_add(qtt, zio); 42117559Swollman 42217559Swollman#ifdef illumos 42317559Swollman mutex_enter(&spa->spa_iokstat_lock); 42417559Swollman spa->spa_queue_stats[zio->io_priority].spa_queued++; 42517559Swollman if (spa->spa_iokstat != NULL) 42617559Swollman kstat_waitq_enter(spa->spa_iokstat->ks_data); 42717559Swollman mutex_exit(&spa->spa_iokstat_lock); 42817559Swollman#endif 42917559Swollman} 43017559Swollman 43117559Swollmanstatic void 43217559Swollmanvdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio) 43317559Swollman{ 43417559Swollman spa_t *spa = zio->io_spa; 43517559Swollman avl_tree_t *qtt; 43617559Swollman 43717559Swollman ASSERT(MUTEX_HELD(&vq->vq_lock)); 43817559Swollman ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 43917559Swollman avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio); 44017559Swollman qtt = vdev_queue_type_tree(vq, zio->io_type); 44117559Swollman if (qtt) 44217559Swollman avl_remove(qtt, zio); 44317559Swollman 44417559Swollman#ifdef illumos 44517559Swollman mutex_enter(&spa->spa_iokstat_lock); 44617559Swollman ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0); 44717559Swollman spa->spa_queue_stats[zio->io_priority].spa_queued--; 4484Srgrimes if (spa->spa_iokstat != NULL) 4494Srgrimes kstat_waitq_exit(spa->spa_iokstat->ks_data); 4504Srgrimes mutex_exit(&spa->spa_iokstat_lock); 4514Srgrimes#endif 45214331Speter} 4534Srgrimes 4544Srgrimesstatic void 4554Srgrimesvdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio) 4564Srgrimes{ 4574Srgrimes spa_t *spa = zio->io_spa; 4584Srgrimes ASSERT(MUTEX_HELD(&vq->vq_lock)); 4594Srgrimes ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 4604Srgrimes vq->vq_class[zio->io_priority].vqc_active++; 4614Srgrimes avl_add(&vq->vq_active_tree, zio); 46214503Shsu 4634Srgrimes#ifdef illumos 4644Srgrimes mutex_enter(&spa->spa_iokstat_lock); 4654Srgrimes spa->spa_queue_stats[zio->io_priority].spa_active++; 4664Srgrimes if (spa->spa_iokstat != NULL) 4676846Sdg kstat_runq_enter(spa->spa_iokstat->ks_data); 4681549Srgrimes mutex_exit(&spa->spa_iokstat_lock); 4693306Sphk#endif 4704Srgrimes} 4711549Srgrimes 47214331Speterstatic void 4734Srgrimesvdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio) 47414331Speter{ 4754Srgrimes spa_t *spa = zio->io_spa; 47614331Speter ASSERT(MUTEX_HELD(&vq->vq_lock)); 4771549Srgrimes ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 4785999Sats vq->vq_class[zio->io_priority].vqc_active--; 4791549Srgrimes avl_remove(&vq->vq_active_tree, zio); 48014331Speter 4814Srgrimes#ifdef illumos 48214331Speter mutex_enter(&spa->spa_iokstat_lock); 4834Srgrimes ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0); 4844Srgrimes spa->spa_queue_stats[zio->io_priority].spa_active--; 4851127Sdg if (spa->spa_iokstat != NULL) { 4861127Sdg kstat_io_t *ksio = spa->spa_iokstat->ks_data; 4871127Sdg 4881127Sdg kstat_runq_exit(spa->spa_iokstat->ks_data); 4891127Sdg if (zio->io_type == ZIO_TYPE_READ) { 4901127Sdg ksio->reads++; 4911127Sdg ksio->nread += zio->io_size; 4924Srgrimes } else if (zio->io_type == ZIO_TYPE_WRITE) { 4934Srgrimes ksio->writes++; 4944Srgrimes ksio->nwritten += zio->io_size; 4954Srgrimes } 4964Srgrimes } 4974Srgrimes mutex_exit(&spa->spa_iokstat_lock); 4984Srgrimes#endif 4994Srgrimes} 5004Srgrimes 5014Srgrimesstatic void 5024Srgrimesvdev_queue_agg_io_done(zio_t *aio) 5034Srgrimes{ 5044Srgrimes if (aio->io_type == ZIO_TYPE_READ) { 5058876Srgrimes zio_t *pio; 5064Srgrimes zio_link_t *zl = NULL; 5074Srgrimes while ((pio = zio_walk_parents(aio, &zl)) != NULL) { 5082254Ssos bcopy((char *)aio->io_data + (pio->io_offset - 5092254Ssos aio->io_offset), pio->io_data, pio->io_size); 5102254Ssos } 5112254Ssos } 5122254Ssos 5132254Ssos zio_buf_free(aio->io_data, aio->io_size); 5146846Sdg} 5156846Sdg 5166846Sdgstatic int 5176846Sdgvdev_queue_class_min_active(zio_priority_t p) 5186846Sdg{ 5194Srgrimes switch (p) { 5204Srgrimes case ZIO_PRIORITY_SYNC_READ: 5216846Sdg return (zfs_vdev_sync_read_min_active); 5226846Sdg case ZIO_PRIORITY_SYNC_WRITE: 5236846Sdg return (zfs_vdev_sync_write_min_active); 5246846Sdg case ZIO_PRIORITY_ASYNC_READ: 5256846Sdg return (zfs_vdev_async_read_min_active); 5266846Sdg case ZIO_PRIORITY_ASYNC_WRITE: 5276846Sdg return (zfs_vdev_async_write_min_active); 5286846Sdg case ZIO_PRIORITY_SCRUB: 5296846Sdg return (zfs_vdev_scrub_min_active); 5306846Sdg case ZIO_PRIORITY_TRIM: 5316846Sdg return (zfs_vdev_trim_min_active); 532924Sdg default: 5334Srgrimes panic("invalid priority %u", p); 5344Srgrimes return (0); 5354Srgrimes } 5366846Sdg} 5376846Sdg 5386846Sdgstatic __noinline int 5396846Sdgvdev_queue_max_async_writes(spa_t *spa) 5406846Sdg{ 5416846Sdg int writes; 5426846Sdg uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total; 5436846Sdg uint64_t min_bytes = zfs_dirty_data_max * 5446846Sdg zfs_vdev_async_write_active_min_dirty_percent / 100; 5456846Sdg uint64_t max_bytes = zfs_dirty_data_max * 5466846Sdg zfs_vdev_async_write_active_max_dirty_percent / 100; 5476846Sdg 5486846Sdg /* 5496846Sdg * Sync tasks correspond to interactive user actions. To reduce the 5506846Sdg * execution time of those actions we push data out as fast as possible. 5516846Sdg */ 5526846Sdg if (spa_has_pending_synctask(spa)) { 5536846Sdg return (zfs_vdev_async_write_max_active); 554924Sdg } 55514331Speter 5561208Shsu if (dirty < min_bytes) 5571066Sdg return (zfs_vdev_async_write_min_active); 5581066Sdg if (dirty > max_bytes) 5591066Sdg return (zfs_vdev_async_write_max_active); 5601066Sdg 5614Srgrimes /* 5624Srgrimes * linear interpolation: 5634Srgrimes * slope = (max_writes - min_writes) / (max_bytes - min_bytes) 5644Srgrimes * move right by min_bytes 5654Srgrimes * move up by min_writes 5664Srgrimes */ 5674Srgrimes writes = (dirty - min_bytes) * 5684Srgrimes (zfs_vdev_async_write_max_active - 5694Srgrimes zfs_vdev_async_write_min_active) / 5705675Sbde (max_bytes - min_bytes) + 5714Srgrimes zfs_vdev_async_write_min_active; 572798Swollman ASSERT3U(writes, >=, zfs_vdev_async_write_min_active); 5734Srgrimes ASSERT3U(writes, <=, zfs_vdev_async_write_max_active); 5744Srgrimes return (writes); 57511390Sbde} 57611390Sbde 57711390Sbdestatic int 5784Srgrimesvdev_queue_class_max_active(spa_t *spa, zio_priority_t p) 5794Srgrimes{ 5804Srgrimes switch (p) { 5814Srgrimes case ZIO_PRIORITY_SYNC_READ: 5821549Srgrimes return (zfs_vdev_sync_read_max_active); 5831208Shsu case ZIO_PRIORITY_SYNC_WRITE: 5844Srgrimes return (zfs_vdev_sync_write_max_active); 5854Srgrimes case ZIO_PRIORITY_ASYNC_READ: 586924Sdg return (zfs_vdev_async_read_max_active); 5874Srgrimes case ZIO_PRIORITY_ASYNC_WRITE: 5884Srgrimes return (vdev_queue_max_async_writes(spa)); 5894Srgrimes case ZIO_PRIORITY_SCRUB: 5904Srgrimes return (zfs_vdev_scrub_max_active); 5914Srgrimes case ZIO_PRIORITY_TRIM: 5924Srgrimes return (zfs_vdev_trim_max_active); 5934Srgrimes default: 5944Srgrimes panic("invalid priority %u", p); 59520998Sdyson return (0); 59624283Smpp } 5974Srgrimes} 5985603Sbde 5995675Sbde/* 6005603Sbde * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if 6015603Sbde * there is no eligible class. 6021208Shsu */ 6035675Sbdestatic zio_priority_t 6045675Sbdevdev_queue_class_to_issue(vdev_queue_t *vq) 6055675Sbde{ 6065675Sbde spa_t *spa = vq->vq_vdev->vdev_spa; 6075675Sbde zio_priority_t p; 6085675Sbde 6095675Sbde ASSERT(MUTEX_HELD(&vq->vq_lock)); 6105675Sbde 6115675Sbde if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active) 6125675Sbde return (ZIO_PRIORITY_NUM_QUEUEABLE); 6135675Sbde 6141208Shsu /* find a queue that has not reached its minimum # outstanding i/os */ 6155603Sbde for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 6161208Shsu if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 6171208Shsu vq->vq_class[p].vqc_active < 6181208Shsu vdev_queue_class_min_active(p)) 6191208Shsu return (p); 6201208Shsu } 6215603Sbde 6225603Sbde /* 6235603Sbde * If we haven't found a queue, look for one that hasn't reached its 6241208Shsu * maximum # outstanding i/os. 6255603Sbde */ 6265603Sbde for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { 6271208Shsu if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 && 6285603Sbde vq->vq_class[p].vqc_active < 6291208Shsu vdev_queue_class_max_active(spa, p)) 6301208Shsu return (p); 6311208Shsu } 6321208Shsu 6331208Shsu /* No eligible queued i/os */ 6344Srgrimes return (ZIO_PRIORITY_NUM_QUEUEABLE); 6351051Sdg} 6361051Sdg 6371051Sdg/* 6381051Sdg * Compute the range spanned by two i/os, which is the endpoint of the last 6391051Sdg * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset). 6401051Sdg * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio); 6411051Sdg * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0. 6421051Sdg */ 6431051Sdg#define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset) 6441051Sdg#define IO_GAP(fio, lio) (-IO_SPAN(lio, fio)) 6451051Sdg 6464Srgrimesstatic zio_t * 64720998Sdysonvdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) 6484Srgrimes{ 6491208Shsu zio_t *first, *last, *aio, *dio, *mandatory, *nio; 6501549Srgrimes void *abuf; 65114331Speter uint64_t maxgap = 0; 6521549Srgrimes uint64_t size; 65314331Speter boolean_t stretch; 65420016Sbde avl_tree_t *t; 655924Sdg enum zio_flag flags; 656924Sdg 657924Sdg ASSERT(MUTEX_HELD(&vq->vq_lock)); 6581208Shsu 6594Srgrimes if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE) 6604Srgrimes return (NULL); 6614Srgrimes 66212722Sphk first = last = zio; 66323070Salex 66412722Sphk if (zio->io_type == ZIO_TYPE_READ) 66523070Salex maxgap = zfs_vdev_read_gap_limit; 66617677Sjulian 66712722Sphk /* 66817677Sjulian * We can aggregate I/Os that are sufficiently adjacent and of 66917677Sjulian * the same flavor, as expressed by the AGG_INHERIT flags. 6704Srgrimes * The latter requirement is necessary so that certain 6714Srgrimes * attributes of the I/O, such as whether it's a normal I/O 6724Srgrimes * or a scrub/resilver, can be preserved in the aggregate. 6734Srgrimes * We can include optional I/Os, but don't allow them 67419274Sjulian * to begin a range as they add no benefit in that situation. 67519274Sjulian */ 67619274Sjulian 67719274Sjulian /* 67819274Sjulian * We keep track of the last non-optional I/O. 67919274Sjulian */ 68019274Sjulian mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first; 68119274Sjulian 68219274Sjulian /* 68319274Sjulian * Walk backwards through sufficiently contiguous I/Os 6844Srgrimes * recording the last non-optional I/O. 6854Srgrimes */ 686556Srgrimes flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT; 687924Sdg t = vdev_queue_type_tree(vq, zio->io_type); 6884Srgrimes while (t != NULL && (dio = AVL_PREV(t, first)) != NULL && 6894Srgrimes (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 690924Sdg IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit && 6914Srgrimes IO_GAP(dio, first) <= maxgap) { 6924201Sdg first = dio; 6934Srgrimes if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL)) 69414348Sjkh mandatory = first; 69514348Sjkh } 69614348Sjkh 69714348Sjkh /* 69814348Sjkh * Skip any initial optional I/Os. 69914348Sjkh */ 70014348Sjkh while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) { 70114348Sjkh first = AVL_NEXT(t, first); 70214348Sjkh ASSERT(first != NULL); 70314348Sjkh } 70414348Sjkh 70514348Sjkh /* 70614348Sjkh * Walk forward through sufficiently contiguous I/Os. 7074201Sdg * The aggregation limit does not apply to optional i/os, so that 7084201Sdg * we can issue contiguous writes even if they are larger than the 7094201Sdg * aggregation limit. 7105603Sbde */ 7114201Sdg while ((dio = AVL_NEXT(t, last)) != NULL && 7124201Sdg (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags && 7134201Sdg (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit || 7144201Sdg (dio->io_flags & ZIO_FLAG_OPTIONAL)) && 7154201Sdg IO_GAP(last, dio) <= maxgap) { 71621975Sbde last = dio; 71721975Sbde if (!(last->io_flags & ZIO_FLAG_OPTIONAL)) 71821975Sbde mandatory = last; 71921975Sbde } 72021975Sbde 72121975Sbde /* 72221975Sbde * Now that we've established the range of the I/O aggregation 72321975Sbde * we must decide what to do with trailing optional I/Os. 72421975Sbde * For reads, there's nothing to do. While we are unable to 72521975Sbde * aggregate further, it's possible that a trailing optional 72621975Sbde * I/O would allow the underlying device to aggregate with 72721975Sbde * subsequent I/Os. We must therefore determine if the next 72821975Sbde * non-optional I/O is close enough to make aggregation 72921975Sbde * worthwhile. 73021975Sbde */ 73121975Sbde stretch = B_FALSE; 73221975Sbde if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) { 73321975Sbde zio_t *nio = last; 73421975Sbde while ((dio = AVL_NEXT(t, nio)) != NULL && 73521975Sbde IO_GAP(nio, dio) == 0 && 73621975Sbde IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) { 73721975Sbde nio = dio; 7384Srgrimes if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) { 73921975Sbde stretch = B_TRUE; 7404Srgrimes break; 7414Srgrimes } 74212186Sphk } 74312186Sphk } 7441549Srgrimes 7454038Sache if (stretch) { 74612186Sphk /* 74712243Sphk * We are going to include an optional io in our aggregated 74812243Sphk * span, thus closing the write gap. Only mandatory i/os can 74912186Sphk * start aggregated spans, so make sure that the next i/o 75012186Sphk * after our span is mandatory. 75112186Sphk */ 7521549Srgrimes dio = AVL_NEXT(t, last); 75312623Sphk dio->io_flags &= ~ZIO_FLAG_OPTIONAL; 75412623Sphk } else { 7551549Srgrimes /* do not include the optional i/o */ 75612186Sphk while (last != mandatory && last != first) { 75712186Sphk ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL); 75812186Sphk last = AVL_PREV(t, last); 75912186Sphk ASSERT(last != NULL); 76012186Sphk } 76112186Sphk } 76215045Sache 76315045Sache if (first == last) 76415045Sache return (NULL); 7651549Srgrimes 7664Srgrimes size = IO_SPAN(first, last); 7674Srgrimes ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 7684Srgrimes 7694Srgrimes abuf = zio_buf_alloc_nowait(size); 7704Srgrimes if (abuf == NULL) 7714Srgrimes return (NULL); 7724Srgrimes 7734501Sbde aio = zio_vdev_delegated_io(first->io_vd, first->io_offset, 7744501Sbde abuf, size, first->io_type, zio->io_priority, 77525164Speter flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE, 77625164Speter vdev_queue_agg_io_done, NULL); 77725164Speter aio->io_timestamp = first->io_timestamp; 7784475Sbde 77925164Speter nio = first; 7801051Sdg do { 7814501Sbde dio = nio; 78225164Speter nio = AVL_NEXT(t, dio); 78325164Speter ASSERT3U(dio->io_type, ==, aio->io_type); 78425164Speter 78525164Speter if (dio->io_flags & ZIO_FLAG_NODATA) { 78625164Speter ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE); 78725164Speter bzero((char *)aio->io_data + (dio->io_offset - 78825164Speter aio->io_offset), dio->io_size); 78925164Speter } else if (dio->io_type == ZIO_TYPE_WRITE) { 79025164Speter bcopy(dio->io_data, (char *)aio->io_data + 79124690Speter (dio->io_offset - aio->io_offset), 79225164Speter dio->io_size); 7934Srgrimes } 79412929Sdg 79512977Sbde zio_add_child(dio, aio); 7964Srgrimes vdev_queue_io_remove(vq, dio); 7974Srgrimes zio_vdev_io_bypass(dio); 7984Srgrimes zio_execute(dio); 79924691Speter } while (dio != last); 80024691Speter 80124691Speter return (aio); 80224691Speter} 80324691Speter 804556Srgrimesstatic zio_t * 80525164Spetervdev_queue_io_to_issue(vdev_queue_t *vq) 80625164Speter{ 80725164Speter zio_t *zio, *aio; 80825164Speter zio_priority_t p; 80925164Speter avl_index_t idx; 8103489Sphk avl_tree_t *tree; 8114Srgrimes zio_t search; 812556Srgrimes 8134Srgrimesagain: 8144Srgrimes ASSERT(MUTEX_HELD(&vq->vq_lock)); 8154Srgrimes 816556Srgrimes p = vdev_queue_class_to_issue(vq); 8174Srgrimes 8184Srgrimes if (p == ZIO_PRIORITY_NUM_QUEUEABLE) { 8193489Sphk /* No eligible queued i/os */ 8204Srgrimes return (NULL); 8214Srgrimes } 8224Srgrimes 8234Srgrimes /* 8244Srgrimes * For LBA-ordered queues (async / scrub), issue the i/o which follows 825556Srgrimes * the most recently issued i/o in LBA (offset) order. 8264Srgrimes * 8274Srgrimes * For FIFO queues (sync), issue the i/o with the lowest timestamp. 8283489Sphk */ 8294Srgrimes tree = vdev_queue_class_tree(vq, p); 8304Srgrimes search.io_timestamp = 0; 8314Srgrimes search.io_offset = vq->vq_last_offset + 1; 8324Srgrimes VERIFY3P(avl_find(tree, &search, &idx), ==, NULL); 8334Srgrimes zio = avl_nearest(tree, idx, AVL_AFTER); 834556Srgrimes if (zio == NULL) 8354Srgrimes zio = avl_first(tree); 8364Srgrimes ASSERT3U(zio->io_priority, ==, p); 8373489Sphk 8383489Sphk aio = vdev_queue_aggregate(vq, zio); 8394Srgrimes if (aio != NULL) 8404Srgrimes zio = aio; 84124690Speter else 8424Srgrimes vdev_queue_io_remove(vq, zio); 843556Srgrimes 8444Srgrimes /* 8454Srgrimes * If the I/O is or was optional and therefore has no data, we need to 8463489Sphk * simply discard it. We need to drop the vdev queue's lock to avoid a 8474Srgrimes * deadlock that we could encounter since this I/O will complete 8484Srgrimes * immediately. 8494Srgrimes */ 8504Srgrimes if (zio->io_flags & ZIO_FLAG_NODATA) { 8514Srgrimes mutex_exit(&vq->vq_lock); 852556Srgrimes zio_vdev_io_bypass(zio); 8534Srgrimes zio_execute(zio); 8544Srgrimes mutex_enter(&vq->vq_lock); 8553489Sphk goto again; 85612929Sdg } 85712929Sdg 8584Srgrimes vdev_queue_pending_add(vq, zio); 8594Srgrimes vq->vq_last_offset = zio->io_offset; 8604Srgrimes 861556Srgrimes return (zio); 8624Srgrimes} 8634Srgrimes 8643489Sphkzio_t * 86525164Spetervdev_queue_io(zio_t *zio) 86625164Speter{ 86725164Speter vdev_queue_t *vq = &zio->io_vd->vdev_queue; 86825164Speter zio_t *nio; 86925164Speter 87025164Speter if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 87112929Sdg return (zio); 8724Srgrimes 8734Srgrimes /* 8744Srgrimes * Children i/os inherent their parent's priority, which might 875556Srgrimes * not match the child's i/o type. Fix it up here. 8764Srgrimes */ 8771051Sdg if (zio->io_type == ZIO_TYPE_READ) { 8783489Sphk if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && 8793258Sdg zio->io_priority != ZIO_PRIORITY_ASYNC_READ && 8801051Sdg zio->io_priority != ZIO_PRIORITY_SCRUB) 8811051Sdg zio->io_priority = ZIO_PRIORITY_ASYNC_READ; 8821051Sdg } else if (zio->io_type == ZIO_TYPE_WRITE) { 8831051Sdg if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 8841051Sdg zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE) 8851051Sdg zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; 8861051Sdg } else { 8873489Sphk ASSERT(zio->io_type == ZIO_TYPE_FREE); 8883258Sdg zio->io_priority = ZIO_PRIORITY_TRIM; 8893489Sphk } 8903258Sdg 8913258Sdg zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 8923258Sdg 8933258Sdg mutex_enter(&vq->vq_lock); 8943258Sdg zio->io_timestamp = gethrtime(); 8953489Sphk vdev_queue_io_add(vq, zio); 8963489Sphk nio = vdev_queue_io_to_issue(vq); 8973258Sdg mutex_exit(&vq->vq_lock); 8983489Sphk 8993258Sdg if (nio == NULL) 9003258Sdg return (NULL); 9013258Sdg 9023258Sdg if (nio->io_done == vdev_queue_agg_io_done) { 9033258Sdg zio_nowait(nio); 9043489Sphk return (NULL); 9053489Sphk } 9063258Sdg 9073489Sphk return (nio); 9083258Sdg} 9093258Sdg 9103258Sdgvoid 9113258Sdgvdev_queue_io_done(zio_t *zio) 9123258Sdg{ 9133489Sphk vdev_queue_t *vq = &zio->io_vd->vdev_queue; 9141051Sdg zio_t *nio; 9154Srgrimes 91612722Sphk mutex_enter(&vq->vq_lock); 9174Srgrimes 9184Srgrimes vdev_queue_pending_remove(vq, zio); 9194Srgrimes 9204Srgrimes vq->vq_io_complete_ts = gethrtime(); 9214Srgrimes 9224Srgrimes while ((nio = vdev_queue_io_to_issue(vq)) != NULL) { 923556Srgrimes mutex_exit(&vq->vq_lock); 9244Srgrimes if (nio->io_done == vdev_queue_agg_io_done) { 9254Srgrimes zio_nowait(nio); 9264Srgrimes } else { 9274Srgrimes zio_vdev_io_reissue(nio); 9284Srgrimes zio_execute(nio); 9294Srgrimes } 9304Srgrimes mutex_enter(&vq->vq_lock); 9314Srgrimes } 932556Srgrimes 9334Srgrimes mutex_exit(&vq->vq_lock); 9344Srgrimes} 9354Srgrimes 9364Srgrimes/* 9374Srgrimes * As these three methods are only used for load calculations we're not concerned 9384Srgrimes * if we get an incorrect value on 32bit platforms due to lack of vq_lock mutex 9394Srgrimes * use here, instead we prefer to keep it lock free for performance. 9404Srgrimes */ 941556Srgrimesint 9424Srgrimesvdev_queue_length(vdev_t *vd) 9434Srgrimes{ 9444Srgrimes return (avl_numnodes(&vd->vdev_queue.vq_active_tree)); 9454Srgrimes} 9464Srgrimes 9474Srgrimesuint64_t 9484Srgrimesvdev_queue_lastoffset(vdev_t *vd) 9494Srgrimes{ 950556Srgrimes return (vd->vdev_queue.vq_lastoffset); 9514Srgrimes} 9524Srgrimes 9534Srgrimesvoid 9544Srgrimesvdev_queue_register_lastoffset(vdev_t *vd, zio_t *zio) 9554Srgrimes{ 9564Srgrimes vd->vdev_queue.vq_lastoffset = zio->io_offset + zio->io_size; 9574Srgrimes} 9584Srgrimes