1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2005 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #pragma ident "%Z%%M% %I% %E% SMI" 28 29 #include <sys/zfs_context.h> 30 #include <sys/spa.h> 31 #include <sys/vdev_impl.h> 32 #include <sys/zio.h> 33 #include <sys/avl.h> 34 35 /* 36 * Virtual device vector for disk I/O scheduling. 37 */ 38 int 39 vdev_queue_deadline_compare(const void *x1, const void *x2) 40 { 41 const zio_t *z1 = x1; 42 const zio_t *z2 = x2; 43 44 if (z1->io_deadline < z2->io_deadline) 45 return (-1); 46 if (z1->io_deadline > z2->io_deadline) 47 return (1); 48 49 if (z1->io_offset < z2->io_offset) 50 return (-1); 51 if (z1->io_offset > z2->io_offset) 52 return (1); 53 54 if (z1 < z2) 55 return (-1); 56 if (z1 > z2) 57 return (1); 58 59 return (0); 60 } 61 62 int 63 vdev_queue_offset_compare(const void *x1, const void *x2) 64 { 65 const zio_t *z1 = x1; 66 const zio_t *z2 = x2; 67 68 if (z1->io_offset < z2->io_offset) 69 return (-1); 70 if (z1->io_offset > z2->io_offset) 71 return (1); 72 73 if (z1 < z2) 74 return (-1); 75 if (z1 > z2) 76 return (1); 77 78 return (0); 79 } 80 81 void 82 vdev_queue_init(vdev_t *vd) 83 { 84 vdev_queue_t *vq = &vd->vdev_queue; 85 86 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL); 87 88 avl_create(&vq->vq_deadline_tree, vdev_queue_deadline_compare, 89 sizeof (zio_t), offsetof(struct zio, io_deadline_node)); 90 91 avl_create(&vq->vq_read_tree, vdev_queue_offset_compare, 92 sizeof (zio_t), offsetof(struct zio, io_offset_node)); 93 94 avl_create(&vq->vq_write_tree, vdev_queue_offset_compare, 95 sizeof (zio_t), offsetof(struct zio, io_offset_node)); 96 97 avl_create(&vq->vq_pending_tree, vdev_queue_offset_compare, 98 sizeof (zio_t), offsetof(struct zio, io_offset_node)); 99 } 100 101 void 102 vdev_queue_fini(vdev_t *vd) 103 { 104 vdev_queue_t *vq = &vd->vdev_queue; 105 106 avl_destroy(&vq->vq_deadline_tree); 107 avl_destroy(&vq->vq_read_tree); 108 avl_destroy(&vq->vq_write_tree); 109 avl_destroy(&vq->vq_pending_tree); 110 111 mutex_destroy(&vq->vq_lock); 112 } 113 114 static void 115 vdev_queue_agg_io_done(zio_t *aio) 116 { 117 zio_t *dio; 118 uint64_t offset = 0; 119 120 while ((dio = aio->io_delegate_list) != NULL) { 121 if (aio->io_type == ZIO_TYPE_READ) 122 bcopy((char *)aio->io_data + offset, dio->io_data, 123 dio->io_size); 124 offset += dio->io_size; 125 aio->io_delegate_list = dio->io_delegate_next; 126 dio->io_delegate_next = NULL; 127 dio->io_error = aio->io_error; 128 zio_next_stage(dio); 129 } 130 ASSERT3U(offset, ==, aio->io_size); 131 132 zio_buf_free(aio->io_data, aio->io_size); 133 } 134 135 #define IS_ADJACENT(io, nio) \ 136 ((io)->io_offset + (io)->io_size == (nio)->io_offset) 137 138 typedef void zio_issue_func_t(zio_t *); 139 140 static zio_t * 141 vdev_queue_io_to_issue(vdev_queue_t *vq, uint64_t pending_limit, 142 zio_issue_func_t **funcp) 143 { 144 zio_t *fio, *lio, *aio, *dio; 145 avl_tree_t *tree; 146 uint64_t size; 147 148 ASSERT(MUTEX_HELD(&vq->vq_lock)); 149 150 *funcp = NULL; 151 152 if (avl_numnodes(&vq->vq_pending_tree) >= pending_limit || 153 avl_numnodes(&vq->vq_deadline_tree) == 0) 154 return (NULL); 155 156 fio = lio = avl_first(&vq->vq_deadline_tree); 157 158 tree = fio->io_vdev_tree; 159 size = fio->io_size; 160 161 while ((dio = AVL_PREV(tree, fio)) != NULL && IS_ADJACENT(dio, fio) && 162 size + dio->io_size <= vq->vq_agg_limit) { 163 dio->io_delegate_next = fio; 164 fio = dio; 165 size += dio->io_size; 166 } 167 168 while ((dio = AVL_NEXT(tree, lio)) != NULL && IS_ADJACENT(lio, dio) && 169 size + dio->io_size <= vq->vq_agg_limit) { 170 lio->io_delegate_next = dio; 171 lio = dio; 172 size += dio->io_size; 173 } 174 175 if (fio != lio) { 176 char *buf = zio_buf_alloc(size); 177 uint64_t offset = 0; 178 int nagg = 0; 179 180 ASSERT(size <= vq->vq_agg_limit); 181 182 aio = zio_vdev_child_io(fio, NULL, fio->io_vd, 183 fio->io_offset, buf, size, fio->io_type, 184 ZIO_PRIORITY_NOW, ZIO_FLAG_DONT_QUEUE | 185 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_PROPAGATE, 186 vdev_queue_agg_io_done, NULL); 187 188 aio->io_delegate_list = fio; 189 190 for (dio = fio; dio != NULL; dio = dio->io_delegate_next) { 191 ASSERT(dio->io_type == aio->io_type); 192 if (dio->io_type == ZIO_TYPE_WRITE) 193 bcopy(dio->io_data, buf + offset, dio->io_size); 194 offset += dio->io_size; 195 avl_remove(&vq->vq_deadline_tree, dio); 196 avl_remove(tree, dio); 197 zio_vdev_io_bypass(dio); 198 nagg++; 199 } 200 201 ASSERT(offset == size); 202 203 dprintf("%5s T=%llu off=%8llx agg=%3d " 204 "old=%5llx new=%5llx\n", 205 zio_type_name[fio->io_type], 206 fio->io_deadline, fio->io_offset, nagg, fio->io_size, size); 207 208 avl_add(&vq->vq_pending_tree, aio); 209 210 *funcp = zio_nowait; 211 return (aio); 212 } 213 214 avl_remove(&vq->vq_deadline_tree, fio); 215 avl_remove(tree, fio); 216 217 avl_add(&vq->vq_pending_tree, fio); 218 219 *funcp = zio_next_stage; 220 221 return (fio); 222 } 223 224 zio_t * 225 vdev_queue_io(zio_t *zio) 226 { 227 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 228 zio_t *nio; 229 zio_issue_func_t *func; 230 231 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); 232 233 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE) 234 return (zio); 235 236 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; 237 238 if (zio->io_type == ZIO_TYPE_READ) 239 zio->io_vdev_tree = &vq->vq_read_tree; 240 else 241 zio->io_vdev_tree = &vq->vq_write_tree; 242 243 mutex_enter(&vq->vq_lock); 244 245 zio->io_deadline = (zio->io_timestamp >> vq->vq_time_shift) + 246 zio->io_priority; 247 248 avl_add(&vq->vq_deadline_tree, zio); 249 avl_add(zio->io_vdev_tree, zio); 250 251 nio = vdev_queue_io_to_issue(vq, vq->vq_min_pending, &func); 252 253 mutex_exit(&vq->vq_lock); 254 255 if (nio == NULL || func != zio_nowait) 256 return (nio); 257 258 func(nio); 259 return (NULL); 260 } 261 262 void 263 vdev_queue_io_done(zio_t *zio) 264 { 265 vdev_queue_t *vq = &zio->io_vd->vdev_queue; 266 zio_t *nio; 267 zio_issue_func_t *func; 268 int i; 269 270 mutex_enter(&vq->vq_lock); 271 272 avl_remove(&vq->vq_pending_tree, zio); 273 274 for (i = 0; i < vq->vq_ramp_rate; i++) { 275 nio = vdev_queue_io_to_issue(vq, vq->vq_max_pending, &func); 276 if (nio == NULL) 277 break; 278 mutex_exit(&vq->vq_lock); 279 if (func == zio_next_stage) 280 zio_vdev_io_reissue(nio); 281 func(nio); 282 mutex_enter(&vq->vq_lock); 283 } 284 285 mutex_exit(&vq->vq_lock); 286 } 287