Paolo Valente | 70f2871 | 2013-05-09 19:10:02 +0200 | [diff] [blame] | 1 | /* |
| 2 | * BFQ: CGROUPS support. |
| 3 | * |
| 4 | * Based on ideas and code from CFQ: |
| 5 | * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> |
| 6 | * |
| 7 | * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it> |
| 8 | * Paolo Valente <paolo.valente@unimore.it> |
| 9 | * |
| 10 | * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it> |
| 11 | * |
| 12 | * Licensed under the GPL-2 as detailed in the accompanying COPYING.BFQ |
| 13 | * file. |
| 14 | */ |
| 15 | |
| 16 | #ifdef CONFIG_CGROUP_BFQIO |
| 17 | static struct bfqio_cgroup bfqio_root_cgroup = { |
| 18 | .weight = BFQ_DEFAULT_GRP_WEIGHT, |
| 19 | .ioprio = BFQ_DEFAULT_GRP_IOPRIO, |
| 20 | .ioprio_class = BFQ_DEFAULT_GRP_CLASS, |
| 21 | }; |
| 22 | |
| 23 | static inline void bfq_init_entity(struct bfq_entity *entity, |
| 24 | struct bfq_group *bfqg) |
| 25 | { |
| 26 | entity->weight = entity->new_weight; |
| 27 | entity->orig_weight = entity->new_weight; |
| 28 | entity->ioprio = entity->new_ioprio; |
| 29 | entity->ioprio_class = entity->new_ioprio_class; |
| 30 | entity->parent = bfqg->my_entity; |
| 31 | entity->sched_data = &bfqg->sched_data; |
| 32 | } |
| 33 | |
| 34 | static struct bfqio_cgroup *cgroup_to_bfqio(struct cgroup *cgroup) |
| 35 | { |
| 36 | return container_of(cgroup_subsys_state(cgroup, bfqio_subsys_id), |
| 37 | struct bfqio_cgroup, css); |
| 38 | } |
| 39 | |
| 40 | /* |
| 41 | * Search the bfq_group for bfqd into the hash table (by now only a list) |
| 42 | * of bgrp. Must be called under rcu_read_lock(). |
| 43 | */ |
| 44 | static struct bfq_group *bfqio_lookup_group(struct bfqio_cgroup *bgrp, |
| 45 | struct bfq_data *bfqd) |
| 46 | { |
| 47 | struct bfq_group *bfqg; |
| 48 | struct hlist_node *n; |
| 49 | void *key; |
| 50 | |
| 51 | hlist_for_each_entry_rcu(bfqg, n, &bgrp->group_data, group_node) { |
| 52 | key = rcu_dereference(bfqg->bfqd); |
| 53 | if (key == bfqd) |
| 54 | return bfqg; |
| 55 | } |
| 56 | |
| 57 | return NULL; |
| 58 | } |
| 59 | |
| 60 | static inline void bfq_group_init_entity(struct bfqio_cgroup *bgrp, |
| 61 | struct bfq_group *bfqg) |
| 62 | { |
| 63 | struct bfq_entity *entity = &bfqg->entity; |
| 64 | |
| 65 | /* |
| 66 | * If the weight of the entity has never been set via the sysfs |
| 67 | * interface, then bgrp->weight == 0. In this case we initialize |
| 68 | * the weight from the current ioprio value. Otherwise, the group |
| 69 | * weight, if set, has priority over the ioprio value. |
| 70 | */ |
| 71 | if (bgrp->weight == 0) { |
| 72 | entity->new_weight = bfq_ioprio_to_weight(bgrp->ioprio); |
| 73 | entity->new_ioprio = bgrp->ioprio; |
| 74 | } else { |
| 75 | if (bgrp->weight < BFQ_MIN_WEIGHT || |
| 76 | bgrp->weight > BFQ_MAX_WEIGHT) { |
| 77 | printk(KERN_CRIT "bfq_group_init_entity: " |
| 78 | "bgrp->weight %d\n", bgrp->weight); |
| 79 | BUG(); |
| 80 | } |
| 81 | entity->new_weight = bgrp->weight; |
| 82 | entity->new_ioprio = bfq_weight_to_ioprio(bgrp->weight); |
| 83 | } |
| 84 | entity->orig_weight = entity->weight = entity->new_weight; |
| 85 | entity->ioprio = entity->new_ioprio; |
| 86 | entity->ioprio_class = entity->new_ioprio_class = bgrp->ioprio_class; |
| 87 | entity->my_sched_data = &bfqg->sched_data; |
| 88 | bfqg->active_entities = 0; |
| 89 | } |
| 90 | |
| 91 | static inline void bfq_group_set_parent(struct bfq_group *bfqg, |
| 92 | struct bfq_group *parent) |
| 93 | { |
| 94 | struct bfq_entity *entity; |
| 95 | |
| 96 | BUG_ON(parent == NULL); |
| 97 | BUG_ON(bfqg == NULL); |
| 98 | |
| 99 | entity = &bfqg->entity; |
| 100 | entity->parent = parent->my_entity; |
| 101 | entity->sched_data = &parent->sched_data; |
| 102 | } |
| 103 | |
| 104 | /** |
| 105 | * bfq_group_chain_alloc - allocate a chain of groups. |
| 106 | * @bfqd: queue descriptor. |
| 107 | * @cgroup: the leaf cgroup this chain starts from. |
| 108 | * |
| 109 | * Allocate a chain of groups starting from the one belonging to |
| 110 | * @cgroup up to the root cgroup. Stop if a cgroup on the chain |
| 111 | * to the root has already an allocated group on @bfqd. |
| 112 | */ |
| 113 | static struct bfq_group *bfq_group_chain_alloc(struct bfq_data *bfqd, |
| 114 | struct cgroup *cgroup) |
| 115 | { |
| 116 | struct bfqio_cgroup *bgrp; |
| 117 | struct bfq_group *bfqg, *prev = NULL, *leaf = NULL; |
| 118 | |
| 119 | for (; cgroup != NULL; cgroup = cgroup->parent) { |
| 120 | bgrp = cgroup_to_bfqio(cgroup); |
| 121 | |
| 122 | bfqg = bfqio_lookup_group(bgrp, bfqd); |
| 123 | if (bfqg != NULL) { |
| 124 | /* |
| 125 | * All the cgroups in the path from there to the |
| 126 | * root must have a bfq_group for bfqd, so we don't |
| 127 | * need any more allocations. |
| 128 | */ |
| 129 | break; |
| 130 | } |
| 131 | |
| 132 | bfqg = kzalloc(sizeof(*bfqg), GFP_ATOMIC); |
| 133 | if (bfqg == NULL) |
| 134 | goto cleanup; |
| 135 | |
| 136 | bfq_group_init_entity(bgrp, bfqg); |
| 137 | bfqg->my_entity = &bfqg->entity; |
| 138 | |
| 139 | if (leaf == NULL) { |
| 140 | leaf = bfqg; |
| 141 | prev = leaf; |
| 142 | } else { |
| 143 | bfq_group_set_parent(prev, bfqg); |
| 144 | /* |
| 145 | * Build a list of allocated nodes using the bfqd |
| 146 | * filed, that is still unused and will be |
| 147 | * initialized only after the node will be |
| 148 | * connected. |
| 149 | */ |
| 150 | prev->bfqd = bfqg; |
| 151 | prev = bfqg; |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | return leaf; |
| 156 | |
| 157 | cleanup: |
| 158 | while (leaf != NULL) { |
| 159 | prev = leaf; |
| 160 | leaf = leaf->bfqd; |
| 161 | kfree(prev); |
| 162 | } |
| 163 | |
| 164 | return NULL; |
| 165 | } |
| 166 | |
| 167 | /** |
| 168 | * bfq_group_chain_link - link an allocated group chain to a cgroup |
| 169 | * hierarchy. |
| 170 | * @bfqd: the queue descriptor. |
| 171 | * @cgroup: the leaf cgroup to start from. |
| 172 | * @leaf: the leaf group (to be associated to @cgroup). |
| 173 | * |
| 174 | * Try to link a chain of groups to a cgroup hierarchy, connecting the |
| 175 | * nodes bottom-up, so we can be sure that when we find a cgroup in the |
| 176 | * hierarchy that already as a group associated to @bfqd all the nodes |
| 177 | * in the path to the root cgroup have one too. |
| 178 | * |
| 179 | * On locking: the queue lock protects the hierarchy (there is a hierarchy |
| 180 | * per device) while the bfqio_cgroup lock protects the list of groups |
| 181 | * belonging to the same cgroup. |
| 182 | */ |
| 183 | static void bfq_group_chain_link(struct bfq_data *bfqd, struct cgroup *cgroup, |
| 184 | struct bfq_group *leaf) |
| 185 | { |
| 186 | struct bfqio_cgroup *bgrp; |
| 187 | struct bfq_group *bfqg, *next, *prev = NULL; |
| 188 | unsigned long flags; |
| 189 | |
| 190 | assert_spin_locked(bfqd->queue->queue_lock); |
| 191 | |
| 192 | for (; cgroup != NULL && leaf != NULL; cgroup = cgroup->parent) { |
| 193 | bgrp = cgroup_to_bfqio(cgroup); |
| 194 | next = leaf->bfqd; |
| 195 | |
| 196 | bfqg = bfqio_lookup_group(bgrp, bfqd); |
| 197 | BUG_ON(bfqg != NULL); |
| 198 | |
| 199 | spin_lock_irqsave(&bgrp->lock, flags); |
| 200 | |
| 201 | rcu_assign_pointer(leaf->bfqd, bfqd); |
| 202 | hlist_add_head_rcu(&leaf->group_node, &bgrp->group_data); |
| 203 | hlist_add_head(&leaf->bfqd_node, &bfqd->group_list); |
| 204 | |
| 205 | spin_unlock_irqrestore(&bgrp->lock, flags); |
| 206 | |
| 207 | prev = leaf; |
| 208 | leaf = next; |
| 209 | } |
| 210 | |
| 211 | BUG_ON(cgroup == NULL && leaf != NULL); |
| 212 | if (cgroup != NULL && prev != NULL) { |
| 213 | bgrp = cgroup_to_bfqio(cgroup); |
| 214 | bfqg = bfqio_lookup_group(bgrp, bfqd); |
| 215 | bfq_group_set_parent(prev, bfqg); |
| 216 | } |
| 217 | } |
| 218 | |
| 219 | /** |
| 220 | * bfq_find_alloc_group - return the group associated to @bfqd in @cgroup. |
| 221 | * @bfqd: queue descriptor. |
| 222 | * @cgroup: cgroup being searched for. |
| 223 | * |
| 224 | * Return a group associated to @bfqd in @cgroup, allocating one if |
| 225 | * necessary. When a group is returned all the cgroups in the path |
| 226 | * to the root have a group associated to @bfqd. |
| 227 | * |
| 228 | * If the allocation fails, return the root group: this breaks guarantees |
| 229 | * but is a safe fallback. If this loss becomes a problem it can be |
| 230 | * mitigated using the equivalent weight (given by the product of the |
| 231 | * weights of the groups in the path from @group to the root) in the |
| 232 | * root scheduler. |
| 233 | * |
| 234 | * We allocate all the missing nodes in the path from the leaf cgroup |
| 235 | * to the root and we connect the nodes only after all the allocations |
| 236 | * have been successful. |
| 237 | */ |
| 238 | static struct bfq_group *bfq_find_alloc_group(struct bfq_data *bfqd, |
| 239 | struct cgroup *cgroup) |
| 240 | { |
| 241 | struct bfqio_cgroup *bgrp = cgroup_to_bfqio(cgroup); |
| 242 | struct bfq_group *bfqg; |
| 243 | |
| 244 | bfqg = bfqio_lookup_group(bgrp, bfqd); |
| 245 | if (bfqg != NULL) |
| 246 | return bfqg; |
| 247 | |
| 248 | bfqg = bfq_group_chain_alloc(bfqd, cgroup); |
| 249 | if (bfqg != NULL) |
| 250 | bfq_group_chain_link(bfqd, cgroup, bfqg); |
| 251 | else |
| 252 | bfqg = bfqd->root_group; |
| 253 | |
| 254 | return bfqg; |
| 255 | } |
| 256 | |
| 257 | /** |
| 258 | * bfq_bfqq_move - migrate @bfqq to @bfqg. |
| 259 | * @bfqd: queue descriptor. |
| 260 | * @bfqq: the queue to move. |
| 261 | * @entity: @bfqq's entity. |
| 262 | * @bfqg: the group to move to. |
| 263 | * |
| 264 | * Move @bfqq to @bfqg, deactivating it from its old group and reactivating |
| 265 | * it on the new one. Avoid putting the entity on the old group idle tree. |
| 266 | * |
| 267 | * Must be called under the queue lock; the cgroup owning @bfqg must |
| 268 | * not disappear (by now this just means that we are called under |
| 269 | * rcu_read_lock()). |
| 270 | */ |
| 271 | static void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
| 272 | struct bfq_entity *entity, struct bfq_group *bfqg) |
| 273 | { |
| 274 | int busy, resume; |
| 275 | |
| 276 | busy = bfq_bfqq_busy(bfqq); |
| 277 | resume = !RB_EMPTY_ROOT(&bfqq->sort_list); |
| 278 | |
| 279 | BUG_ON(resume && !entity->on_st); |
| 280 | BUG_ON(busy && !resume && entity->on_st && |
| 281 | bfqq != bfqd->in_service_queue); |
| 282 | |
| 283 | if (busy) { |
| 284 | BUG_ON(atomic_read(&bfqq->ref) < 2); |
| 285 | |
| 286 | if (!resume) |
| 287 | bfq_del_bfqq_busy(bfqd, bfqq, 0); |
| 288 | else |
| 289 | bfq_deactivate_bfqq(bfqd, bfqq, 0); |
| 290 | } else if (entity->on_st) |
| 291 | bfq_put_idle_entity(bfq_entity_service_tree(entity), entity); |
| 292 | |
| 293 | /* |
| 294 | * Here we use a reference to bfqg. We don't need a refcounter |
| 295 | * as the cgroup reference will not be dropped, so that its |
| 296 | * destroy() callback will not be invoked. |
| 297 | */ |
| 298 | entity->parent = bfqg->my_entity; |
| 299 | entity->sched_data = &bfqg->sched_data; |
| 300 | |
| 301 | if (busy && resume) |
| 302 | bfq_activate_bfqq(bfqd, bfqq); |
| 303 | |
| 304 | if (bfqd->in_service_queue == NULL && !bfqd->rq_in_driver) |
| 305 | bfq_schedule_dispatch(bfqd); |
| 306 | } |
| 307 | |
| 308 | /** |
| 309 | * __bfq_bic_change_cgroup - move @bic to @cgroup. |
| 310 | * @bfqd: the queue descriptor. |
| 311 | * @bic: the bic to move. |
| 312 | * @cgroup: the cgroup to move to. |
| 313 | * |
| 314 | * Move bic to cgroup, assuming that bfqd->queue is locked; the caller |
| 315 | * has to make sure that the reference to cgroup is valid across the call. |
| 316 | * |
| 317 | * NOTE: an alternative approach might have been to store the current |
| 318 | * cgroup in bfqq and getting a reference to it, reducing the lookup |
| 319 | * time here, at the price of slightly more complex code. |
| 320 | */ |
| 321 | static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd, |
| 322 | struct bfq_io_cq *bic, |
| 323 | struct cgroup *cgroup) |
| 324 | { |
| 325 | struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0); |
| 326 | struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1); |
| 327 | struct bfq_entity *entity; |
| 328 | struct bfq_group *bfqg; |
| 329 | struct bfqio_cgroup *bgrp; |
| 330 | |
| 331 | bgrp = cgroup_to_bfqio(cgroup); |
| 332 | |
| 333 | bfqg = bfq_find_alloc_group(bfqd, cgroup); |
| 334 | if (async_bfqq != NULL) { |
| 335 | entity = &async_bfqq->entity; |
| 336 | |
| 337 | if (entity->sched_data != &bfqg->sched_data) { |
| 338 | bic_set_bfqq(bic, NULL, 0); |
| 339 | bfq_log_bfqq(bfqd, async_bfqq, |
| 340 | "bic_change_group: %p %d", |
| 341 | async_bfqq, atomic_read(&async_bfqq->ref)); |
| 342 | bfq_put_queue(async_bfqq); |
| 343 | } |
| 344 | } |
| 345 | |
| 346 | if (sync_bfqq != NULL) { |
| 347 | entity = &sync_bfqq->entity; |
| 348 | if (entity->sched_data != &bfqg->sched_data) |
| 349 | bfq_bfqq_move(bfqd, sync_bfqq, entity, bfqg); |
| 350 | } |
| 351 | |
| 352 | return bfqg; |
| 353 | } |
| 354 | |
| 355 | /** |
| 356 | * bfq_bic_change_cgroup - move @bic to @cgroup. |
| 357 | * @bic: the bic being migrated. |
| 358 | * @cgroup: the destination cgroup. |
| 359 | * |
| 360 | * When the task owning @bic is moved to @cgroup, @bic is immediately |
| 361 | * moved into its new parent group. |
| 362 | */ |
| 363 | static void bfq_bic_change_cgroup(struct bfq_io_cq *bic, |
| 364 | struct cgroup *cgroup) |
| 365 | { |
| 366 | struct bfq_data *bfqd; |
| 367 | unsigned long uninitialized_var(flags); |
| 368 | |
| 369 | bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), |
| 370 | &flags); |
| 371 | if (bfqd != NULL) { |
| 372 | __bfq_bic_change_cgroup(bfqd, bic, cgroup); |
| 373 | bfq_put_bfqd_unlock(bfqd, &flags); |
| 374 | } |
| 375 | } |
| 376 | |
| 377 | /** |
| 378 | * bfq_bic_update_cgroup - update the cgroup of @bic. |
| 379 | * @bic: the @bic to update. |
| 380 | * |
| 381 | * Make sure that @bic is enqueued in the cgroup of the current task. |
| 382 | * We need this in addition to moving bics during the cgroup attach |
| 383 | * phase because the task owning @bic could be at its first disk |
| 384 | * access or we may end up in the root cgroup as the result of a |
| 385 | * memory allocation failure and here we try to move to the right |
| 386 | * group. |
| 387 | * |
| 388 | * Must be called under the queue lock. It is safe to use the returned |
| 389 | * value even after the rcu_read_unlock() as the migration/destruction |
| 390 | * paths act under the queue lock too. IOW it is impossible to race with |
| 391 | * group migration/destruction and end up with an invalid group as: |
| 392 | * a) here cgroup has not yet been destroyed, nor its destroy callback |
| 393 | * has started execution, as current holds a reference to it, |
| 394 | * b) if it is destroyed after rcu_read_unlock() [after current is |
| 395 | * migrated to a different cgroup] its attach() callback will have |
| 396 | * taken care of remove all the references to the old cgroup data. |
| 397 | */ |
| 398 | static struct bfq_group *bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
| 399 | { |
| 400 | struct bfq_data *bfqd = bic_to_bfqd(bic); |
| 401 | struct bfq_group *bfqg; |
| 402 | struct cgroup *cgroup; |
| 403 | |
| 404 | BUG_ON(bfqd == NULL); |
| 405 | |
| 406 | rcu_read_lock(); |
| 407 | cgroup = task_cgroup(current, bfqio_subsys_id); |
| 408 | bfqg = __bfq_bic_change_cgroup(bfqd, bic, cgroup); |
| 409 | rcu_read_unlock(); |
| 410 | |
| 411 | return bfqg; |
| 412 | } |
| 413 | |
| 414 | /** |
| 415 | * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st. |
| 416 | * @st: the service tree being flushed. |
| 417 | */ |
| 418 | static inline void bfq_flush_idle_tree(struct bfq_service_tree *st) |
| 419 | { |
| 420 | struct bfq_entity *entity = st->first_idle; |
| 421 | |
| 422 | for (; entity != NULL; entity = st->first_idle) |
| 423 | __bfq_deactivate_entity(entity, 0); |
| 424 | } |
| 425 | |
| 426 | /** |
| 427 | * bfq_reparent_leaf_entity - move leaf entity to the root_group. |
| 428 | * @bfqd: the device data structure with the root group. |
| 429 | * @entity: the entity to move. |
| 430 | */ |
| 431 | static inline void bfq_reparent_leaf_entity(struct bfq_data *bfqd, |
| 432 | struct bfq_entity *entity) |
| 433 | { |
| 434 | struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
| 435 | |
| 436 | BUG_ON(bfqq == NULL); |
| 437 | bfq_bfqq_move(bfqd, bfqq, entity, bfqd->root_group); |
| 438 | return; |
| 439 | } |
| 440 | |
| 441 | /** |
| 442 | * bfq_reparent_active_entities - move to the root group all active |
| 443 | * entities. |
| 444 | * @bfqd: the device data structure with the root group. |
| 445 | * @bfqg: the group to move from. |
| 446 | * @st: the service tree with the entities. |
| 447 | * |
| 448 | * Needs queue_lock to be taken and reference to be valid over the call. |
| 449 | */ |
| 450 | static inline void bfq_reparent_active_entities(struct bfq_data *bfqd, |
| 451 | struct bfq_group *bfqg, |
| 452 | struct bfq_service_tree *st) |
| 453 | { |
| 454 | struct rb_root *active = &st->active; |
| 455 | struct bfq_entity *entity = NULL; |
| 456 | |
| 457 | if (!RB_EMPTY_ROOT(&st->active)) |
| 458 | entity = bfq_entity_of(rb_first(active)); |
| 459 | |
| 460 | for (; entity != NULL; entity = bfq_entity_of(rb_first(active))) |
| 461 | bfq_reparent_leaf_entity(bfqd, entity); |
| 462 | |
| 463 | if (bfqg->sched_data.in_service_entity != NULL) |
| 464 | bfq_reparent_leaf_entity(bfqd, |
| 465 | bfqg->sched_data.in_service_entity); |
| 466 | |
| 467 | return; |
| 468 | } |
| 469 | |
| 470 | /** |
| 471 | * bfq_destroy_group - destroy @bfqg. |
| 472 | * @bgrp: the bfqio_cgroup containing @bfqg. |
| 473 | * @bfqg: the group being destroyed. |
| 474 | * |
| 475 | * Destroy @bfqg, making sure that it is not referenced from its parent. |
| 476 | */ |
| 477 | static void bfq_destroy_group(struct bfqio_cgroup *bgrp, struct bfq_group *bfqg) |
| 478 | { |
| 479 | struct bfq_data *bfqd; |
| 480 | struct bfq_service_tree *st; |
| 481 | struct bfq_entity *entity = bfqg->my_entity; |
| 482 | unsigned long uninitialized_var(flags); |
| 483 | int i; |
| 484 | |
| 485 | hlist_del(&bfqg->group_node); |
| 486 | |
| 487 | /* |
| 488 | * Empty all service_trees belonging to this group before |
| 489 | * deactivating the group itself. |
| 490 | */ |
| 491 | for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) { |
| 492 | st = bfqg->sched_data.service_tree + i; |
| 493 | |
| 494 | /* |
| 495 | * The idle tree may still contain bfq_queues belonging |
| 496 | * to exited task because they never migrated to a different |
| 497 | * cgroup from the one being destroyed now. No one else |
| 498 | * can access them so it's safe to act without any lock. |
| 499 | */ |
| 500 | bfq_flush_idle_tree(st); |
| 501 | |
| 502 | /* |
| 503 | * It may happen that some queues are still active |
| 504 | * (busy) upon group destruction (if the corresponding |
| 505 | * processes have been forced to terminate). We move |
| 506 | * all the leaf entities corresponding to these queues |
| 507 | * to the root_group. |
| 508 | * Also, it may happen that the group has an entity |
| 509 | * in service, which is disconnected from the active |
| 510 | * tree: it must be moved, too. |
| 511 | * There is no need to put the sync queues, as the |
| 512 | * scheduler has taken no reference. |
| 513 | */ |
| 514 | bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
| 515 | if (bfqd != NULL) { |
| 516 | bfq_reparent_active_entities(bfqd, bfqg, st); |
| 517 | bfq_put_bfqd_unlock(bfqd, &flags); |
| 518 | } |
| 519 | BUG_ON(!RB_EMPTY_ROOT(&st->active)); |
| 520 | BUG_ON(!RB_EMPTY_ROOT(&st->idle)); |
| 521 | } |
| 522 | BUG_ON(bfqg->sched_data.next_in_service != NULL); |
| 523 | BUG_ON(bfqg->sched_data.in_service_entity != NULL); |
| 524 | |
| 525 | /* |
| 526 | * We may race with device destruction, take extra care when |
| 527 | * dereferencing bfqg->bfqd. |
| 528 | */ |
| 529 | bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
| 530 | if (bfqd != NULL) { |
| 531 | hlist_del(&bfqg->bfqd_node); |
| 532 | __bfq_deactivate_entity(entity, 0); |
| 533 | bfq_put_async_queues(bfqd, bfqg); |
| 534 | bfq_put_bfqd_unlock(bfqd, &flags); |
| 535 | } |
| 536 | BUG_ON(entity->tree != NULL); |
| 537 | |
| 538 | /* |
| 539 | * No need to defer the kfree() to the end of the RCU grace |
| 540 | * period: we are called from the destroy() callback of our |
| 541 | * cgroup, so we can be sure that no one is a) still using |
| 542 | * this cgroup or b) doing lookups in it. |
| 543 | */ |
| 544 | kfree(bfqg); |
| 545 | } |
| 546 | |
| 547 | static void bfq_end_wr_async(struct bfq_data *bfqd) |
| 548 | { |
| 549 | struct hlist_node *pos, *n; |
| 550 | struct bfq_group *bfqg; |
| 551 | |
| 552 | hlist_for_each_entry_safe(bfqg, pos, n, &bfqd->group_list, bfqd_node) |
| 553 | bfq_end_wr_async_queues(bfqd, bfqg); |
| 554 | bfq_end_wr_async_queues(bfqd, bfqd->root_group); |
| 555 | } |
| 556 | |
| 557 | /** |
| 558 | * bfq_disconnect_groups - disconnect @bfqd from all its groups. |
| 559 | * @bfqd: the device descriptor being exited. |
| 560 | * |
| 561 | * When the device exits we just make sure that no lookup can return |
| 562 | * the now unused group structures. They will be deallocated on cgroup |
| 563 | * destruction. |
| 564 | */ |
| 565 | static void bfq_disconnect_groups(struct bfq_data *bfqd) |
| 566 | { |
| 567 | struct hlist_node *pos, *n; |
| 568 | struct bfq_group *bfqg; |
| 569 | |
| 570 | bfq_log(bfqd, "disconnect_groups beginning"); |
| 571 | hlist_for_each_entry_safe(bfqg, pos, n, &bfqd->group_list, bfqd_node) { |
| 572 | hlist_del(&bfqg->bfqd_node); |
| 573 | |
| 574 | __bfq_deactivate_entity(bfqg->my_entity, 0); |
| 575 | |
| 576 | /* |
| 577 | * Don't remove from the group hash, just set an |
| 578 | * invalid key. No lookups can race with the |
| 579 | * assignment as bfqd is being destroyed; this |
| 580 | * implies also that new elements cannot be added |
| 581 | * to the list. |
| 582 | */ |
| 583 | rcu_assign_pointer(bfqg->bfqd, NULL); |
| 584 | |
| 585 | bfq_log(bfqd, "disconnect_groups: put async for group %p", |
| 586 | bfqg); |
| 587 | bfq_put_async_queues(bfqd, bfqg); |
| 588 | } |
| 589 | } |
| 590 | |
| 591 | static inline void bfq_free_root_group(struct bfq_data *bfqd) |
| 592 | { |
| 593 | struct bfqio_cgroup *bgrp = &bfqio_root_cgroup; |
| 594 | struct bfq_group *bfqg = bfqd->root_group; |
| 595 | |
| 596 | bfq_put_async_queues(bfqd, bfqg); |
| 597 | |
| 598 | spin_lock_irq(&bgrp->lock); |
| 599 | hlist_del_rcu(&bfqg->group_node); |
| 600 | spin_unlock_irq(&bgrp->lock); |
| 601 | |
| 602 | /* |
| 603 | * No need to synchronize_rcu() here: since the device is gone |
| 604 | * there cannot be any read-side access to its root_group. |
| 605 | */ |
| 606 | kfree(bfqg); |
| 607 | } |
| 608 | |
| 609 | static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
| 610 | { |
| 611 | struct bfq_group *bfqg; |
| 612 | struct bfqio_cgroup *bgrp; |
| 613 | int i; |
| 614 | |
| 615 | bfqg = kzalloc_node(sizeof(*bfqg), GFP_KERNEL, node); |
| 616 | if (bfqg == NULL) |
| 617 | return NULL; |
| 618 | |
| 619 | bfqg->entity.parent = NULL; |
| 620 | for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
| 621 | bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
| 622 | |
| 623 | bgrp = &bfqio_root_cgroup; |
| 624 | spin_lock_irq(&bgrp->lock); |
| 625 | rcu_assign_pointer(bfqg->bfqd, bfqd); |
| 626 | hlist_add_head_rcu(&bfqg->group_node, &bgrp->group_data); |
| 627 | spin_unlock_irq(&bgrp->lock); |
| 628 | |
| 629 | return bfqg; |
| 630 | } |
| 631 | |
| 632 | #define SHOW_FUNCTION(__VAR) \ |
| 633 | static u64 bfqio_cgroup_##__VAR##_read(struct cgroup *cgroup, \ |
| 634 | struct cftype *cftype) \ |
| 635 | { \ |
| 636 | struct bfqio_cgroup *bgrp; \ |
| 637 | u64 ret; \ |
| 638 | \ |
| 639 | if (!cgroup_lock_live_group(cgroup)) \ |
| 640 | return -ENODEV; \ |
| 641 | \ |
| 642 | bgrp = cgroup_to_bfqio(cgroup); \ |
| 643 | spin_lock_irq(&bgrp->lock); \ |
| 644 | ret = bgrp->__VAR; \ |
| 645 | spin_unlock_irq(&bgrp->lock); \ |
| 646 | \ |
| 647 | cgroup_unlock(); \ |
| 648 | \ |
| 649 | return ret; \ |
| 650 | } |
| 651 | |
| 652 | SHOW_FUNCTION(weight); |
| 653 | SHOW_FUNCTION(ioprio); |
| 654 | SHOW_FUNCTION(ioprio_class); |
| 655 | #undef SHOW_FUNCTION |
| 656 | |
| 657 | #define STORE_FUNCTION(__VAR, __MIN, __MAX) \ |
| 658 | static int bfqio_cgroup_##__VAR##_write(struct cgroup *cgroup, \ |
| 659 | struct cftype *cftype, \ |
| 660 | u64 val) \ |
| 661 | { \ |
| 662 | struct bfqio_cgroup *bgrp; \ |
| 663 | struct bfq_group *bfqg; \ |
| 664 | struct hlist_node *n; \ |
| 665 | \ |
| 666 | if (val < (__MIN) || val > (__MAX)) \ |
| 667 | return -EINVAL; \ |
| 668 | \ |
| 669 | if (!cgroup_lock_live_group(cgroup)) \ |
| 670 | return -ENODEV; \ |
| 671 | \ |
| 672 | bgrp = cgroup_to_bfqio(cgroup); \ |
| 673 | \ |
| 674 | spin_lock_irq(&bgrp->lock); \ |
| 675 | bgrp->__VAR = (unsigned short)val; \ |
| 676 | hlist_for_each_entry(bfqg, n, &bgrp->group_data, group_node) { \ |
| 677 | /* \ |
| 678 | * Setting the ioprio_changed flag of the entity \ |
| 679 | * to 1 with new_##__VAR == ##__VAR would re-set \ |
| 680 | * the value of the weight to its ioprio mapping. \ |
| 681 | * Set the flag only if necessary. \ |
| 682 | */ \ |
| 683 | if ((unsigned short)val != bfqg->entity.new_##__VAR) { \ |
| 684 | bfqg->entity.new_##__VAR = (unsigned short)val; \ |
| 685 | /* \ |
| 686 | * Make sure that the above new value has been \ |
| 687 | * stored in bfqg->entity.new_##__VAR before \ |
| 688 | * setting the ioprio_changed flag. In fact, \ |
| 689 | * this flag may be read asynchronously (in \ |
| 690 | * critical sections protected by a different \ |
| 691 | * lock than that held here), and finding this \ |
| 692 | * flag set may cause the execution of the code \ |
| 693 | * for updating parameters whose value may \ |
| 694 | * depend also on bfqg->entity.new_##__VAR (in \ |
| 695 | * __bfq_entity_update_weight_prio). \ |
| 696 | * This barrier makes sure that the new value \ |
| 697 | * of bfqg->entity.new_##__VAR is correctly \ |
| 698 | * seen in that code. \ |
| 699 | */ \ |
| 700 | smp_wmb(); \ |
| 701 | bfqg->entity.ioprio_changed = 1; \ |
| 702 | } \ |
| 703 | } \ |
| 704 | spin_unlock_irq(&bgrp->lock); \ |
| 705 | \ |
| 706 | cgroup_unlock(); \ |
| 707 | \ |
| 708 | return 0; \ |
| 709 | } |
| 710 | |
| 711 | STORE_FUNCTION(weight, BFQ_MIN_WEIGHT, BFQ_MAX_WEIGHT); |
| 712 | STORE_FUNCTION(ioprio, 0, IOPRIO_BE_NR - 1); |
| 713 | STORE_FUNCTION(ioprio_class, IOPRIO_CLASS_RT, IOPRIO_CLASS_IDLE); |
| 714 | #undef STORE_FUNCTION |
| 715 | |
| 716 | static struct cftype bfqio_files[] = { |
| 717 | { |
| 718 | .name = "weight", |
| 719 | .read_u64 = bfqio_cgroup_weight_read, |
| 720 | .write_u64 = bfqio_cgroup_weight_write, |
| 721 | }, |
| 722 | { |
| 723 | .name = "ioprio", |
| 724 | .read_u64 = bfqio_cgroup_ioprio_read, |
| 725 | .write_u64 = bfqio_cgroup_ioprio_write, |
| 726 | }, |
| 727 | { |
| 728 | .name = "ioprio_class", |
| 729 | .read_u64 = bfqio_cgroup_ioprio_class_read, |
| 730 | .write_u64 = bfqio_cgroup_ioprio_class_write, |
| 731 | }, |
| 732 | }; |
| 733 | |
| 734 | static int bfqio_populate(struct cgroup_subsys *subsys, struct cgroup *cgroup) |
| 735 | { |
| 736 | return cgroup_add_files(cgroup, subsys, bfqio_files, |
| 737 | ARRAY_SIZE(bfqio_files)); |
| 738 | } |
| 739 | |
| 740 | static struct cgroup_subsys_state *bfqio_create(struct cgroup *cgroup) |
| 741 | { |
| 742 | struct bfqio_cgroup *bgrp; |
| 743 | |
| 744 | if (cgroup->parent != NULL) { |
| 745 | bgrp = kzalloc(sizeof(*bgrp), GFP_KERNEL); |
| 746 | if (bgrp == NULL) |
| 747 | return ERR_PTR(-ENOMEM); |
| 748 | } else |
| 749 | bgrp = &bfqio_root_cgroup; |
| 750 | |
| 751 | spin_lock_init(&bgrp->lock); |
| 752 | INIT_HLIST_HEAD(&bgrp->group_data); |
| 753 | bgrp->ioprio = BFQ_DEFAULT_GRP_IOPRIO; |
| 754 | bgrp->ioprio_class = BFQ_DEFAULT_GRP_CLASS; |
| 755 | |
| 756 | return &bgrp->css; |
| 757 | } |
| 758 | |
| 759 | /* |
| 760 | * We cannot support shared io contexts, as we have no means to support |
| 761 | * two tasks with the same ioc in two different groups without major rework |
| 762 | * of the main bic/bfqq data structures. By now we allow a task to change |
| 763 | * its cgroup only if it's the only owner of its ioc; the drawback of this |
| 764 | * behavior is that a group containing a task that forked using CLONE_IO |
| 765 | * will not be destroyed until the tasks sharing the ioc die. |
| 766 | */ |
| 767 | static int bfqio_can_attach(struct cgroup *cgroup, struct cgroup_taskset *tset) |
| 768 | { |
| 769 | struct task_struct *task; |
| 770 | struct io_context *ioc; |
| 771 | int ret = 0; |
| 772 | |
| 773 | cgroup_taskset_for_each(task, cgroup, tset) { |
| 774 | /* task_lock() is needed to avoid races with exit_io_context() */ |
| 775 | task_lock(task); |
| 776 | ioc = task->io_context; |
| 777 | if (ioc != NULL && atomic_read(&ioc->nr_tasks) > 1) |
| 778 | /* |
| 779 | * ioc == NULL means that the task is either too |
| 780 | * young or exiting: if it has still no ioc the |
| 781 | * ioc can't be shared, if the task is exiting the |
| 782 | * attach will fail anyway, no matter what we |
| 783 | * return here. |
| 784 | */ |
| 785 | ret = -EINVAL; |
| 786 | task_unlock(task); |
| 787 | if (ret) |
| 788 | break; |
| 789 | } |
| 790 | |
| 791 | return ret; |
| 792 | } |
| 793 | |
| 794 | static void bfqio_attach(struct cgroup *cgroup, struct cgroup_taskset *tset) |
| 795 | { |
| 796 | struct task_struct *task; |
| 797 | struct io_context *ioc; |
| 798 | struct io_cq *icq; |
| 799 | struct hlist_node *n; |
| 800 | |
| 801 | /* |
| 802 | * IMPORTANT NOTE: The move of more than one process at a time to a |
| 803 | * new group has not yet been tested. |
| 804 | */ |
| 805 | cgroup_taskset_for_each(task, cgroup, tset) { |
| 806 | ioc = get_task_io_context(task, GFP_ATOMIC, NUMA_NO_NODE); |
| 807 | if (ioc) { |
| 808 | /* |
| 809 | * Handle cgroup change here. |
| 810 | */ |
| 811 | rcu_read_lock(); |
| 812 | hlist_for_each_entry_rcu(icq, n, &ioc->icq_list, ioc_node) |
| 813 | if (!strncmp( |
| 814 | icq->q->elevator->type->elevator_name, |
| 815 | "bfq", ELV_NAME_MAX)) |
| 816 | bfq_bic_change_cgroup(icq_to_bic(icq), |
| 817 | cgroup); |
| 818 | rcu_read_unlock(); |
| 819 | put_io_context(ioc); |
| 820 | } |
| 821 | } |
| 822 | } |
| 823 | |
| 824 | static void bfqio_destroy(struct cgroup *cgroup) |
| 825 | { |
| 826 | struct bfqio_cgroup *bgrp = cgroup_to_bfqio(cgroup); |
| 827 | struct hlist_node *n, *tmp; |
| 828 | struct bfq_group *bfqg; |
| 829 | |
| 830 | /* |
| 831 | * Since we are destroying the cgroup, there are no more tasks |
| 832 | * referencing it, and all the RCU grace periods that may have |
| 833 | * referenced it are ended (as the destruction of the parent |
| 834 | * cgroup is RCU-safe); bgrp->group_data will not be accessed by |
| 835 | * anything else and we don't need any synchronization. |
| 836 | */ |
| 837 | hlist_for_each_entry_safe(bfqg, n, tmp, &bgrp->group_data, group_node) |
| 838 | bfq_destroy_group(bgrp, bfqg); |
| 839 | |
| 840 | BUG_ON(!hlist_empty(&bgrp->group_data)); |
| 841 | |
| 842 | kfree(bgrp); |
| 843 | } |
| 844 | |
| 845 | struct cgroup_subsys bfqio_subsys = { |
| 846 | .name = "bfqio", |
| 847 | .create = bfqio_create, |
| 848 | .can_attach = bfqio_can_attach, |
| 849 | .attach = bfqio_attach, |
| 850 | .destroy = bfqio_destroy, |
| 851 | .populate = bfqio_populate, |
| 852 | .subsys_id = bfqio_subsys_id, |
| 853 | }; |
| 854 | #else |
| 855 | static inline void bfq_init_entity(struct bfq_entity *entity, |
| 856 | struct bfq_group *bfqg) |
| 857 | { |
| 858 | entity->weight = entity->new_weight; |
| 859 | entity->orig_weight = entity->new_weight; |
| 860 | entity->ioprio = entity->new_ioprio; |
| 861 | entity->ioprio_class = entity->new_ioprio_class; |
| 862 | entity->sched_data = &bfqg->sched_data; |
| 863 | } |
| 864 | |
| 865 | static inline struct bfq_group * |
| 866 | bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
| 867 | { |
| 868 | struct bfq_data *bfqd = bic_to_bfqd(bic); |
| 869 | return bfqd->root_group; |
| 870 | } |
| 871 | |
| 872 | static inline void bfq_bfqq_move(struct bfq_data *bfqd, |
| 873 | struct bfq_queue *bfqq, |
| 874 | struct bfq_entity *entity, |
| 875 | struct bfq_group *bfqg) |
| 876 | { |
| 877 | } |
| 878 | |
| 879 | static void bfq_end_wr_async(struct bfq_data *bfqd) |
| 880 | { |
| 881 | bfq_end_wr_async_queues(bfqd, bfqd->root_group); |
| 882 | } |
| 883 | |
| 884 | static inline void bfq_disconnect_groups(struct bfq_data *bfqd) |
| 885 | { |
| 886 | bfq_put_async_queues(bfqd, bfqd->root_group); |
| 887 | } |
| 888 | |
| 889 | static inline void bfq_free_root_group(struct bfq_data *bfqd) |
| 890 | { |
| 891 | kfree(bfqd->root_group); |
| 892 | } |
| 893 | |
| 894 | static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
| 895 | { |
| 896 | struct bfq_group *bfqg; |
| 897 | int i; |
| 898 | |
| 899 | bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node); |
| 900 | if (bfqg == NULL) |
| 901 | return NULL; |
| 902 | |
| 903 | for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
| 904 | bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
| 905 | |
| 906 | return bfqg; |
| 907 | } |
| 908 | #endif |