| 1 | /* |
| 2 | * Functions related to io context handling |
| 3 | */ |
| 4 | #include <linux/kernel.h> |
| 5 | #include <linux/module.h> |
| 6 | #include <linux/init.h> |
| 7 | #include <linux/bio.h> |
| 8 | #include <linux/blkdev.h> |
| 9 | #include <linux/slab.h> |
| 10 | #include <linux/sched/task.h> |
| 11 | |
| 12 | #include "blk.h" |
| 13 | |
| 14 | /* |
| 15 | * For io context allocations |
| 16 | */ |
| 17 | static struct kmem_cache *iocontext_cachep; |
| 18 | |
| 19 | /** |
| 20 | * get_io_context - increment reference count to io_context |
| 21 | * @ioc: io_context to get |
| 22 | * |
| 23 | * Increment reference count to @ioc. |
| 24 | */ |
| 25 | void get_io_context(struct io_context *ioc) |
| 26 | { |
| 27 | BUG_ON(atomic_long_read(&ioc->refcount) <= 0); |
| 28 | atomic_long_inc(&ioc->refcount); |
| 29 | } |
| 30 | EXPORT_SYMBOL(get_io_context); |
| 31 | |
| 32 | static void icq_free_icq_rcu(struct rcu_head *head) |
| 33 | { |
| 34 | struct io_cq *icq = container_of(head, struct io_cq, __rcu_head); |
| 35 | |
| 36 | kmem_cache_free(icq->__rcu_icq_cache, icq); |
| 37 | } |
| 38 | |
| 39 | /* |
| 40 | * Exit an icq. Called with ioc locked for blk-mq, and with both ioc |
| 41 | * and queue locked for legacy. |
| 42 | */ |
| 43 | static void ioc_exit_icq(struct io_cq *icq) |
| 44 | { |
| 45 | struct elevator_type *et = icq->q->elevator->type; |
| 46 | |
| 47 | if (icq->flags & ICQ_EXITED) |
| 48 | return; |
| 49 | |
| 50 | if (et->uses_mq && et->ops.mq.exit_icq) |
| 51 | et->ops.mq.exit_icq(icq); |
| 52 | else if (!et->uses_mq && et->ops.sq.elevator_exit_icq_fn) |
| 53 | et->ops.sq.elevator_exit_icq_fn(icq); |
| 54 | |
| 55 | icq->flags |= ICQ_EXITED; |
| 56 | } |
| 57 | |
| 58 | /* |
| 59 | * Release an icq. Called with ioc locked for blk-mq, and with both ioc |
| 60 | * and queue locked for legacy. |
| 61 | */ |
| 62 | static void ioc_destroy_icq(struct io_cq *icq) |
| 63 | { |
| 64 | struct io_context *ioc = icq->ioc; |
| 65 | struct request_queue *q = icq->q; |
| 66 | struct elevator_type *et = q->elevator->type; |
| 67 | |
| 68 | lockdep_assert_held(&ioc->lock); |
| 69 | |
| 70 | radix_tree_delete(&ioc->icq_tree, icq->q->id); |
| 71 | hlist_del_init(&icq->ioc_node); |
| 72 | list_del_init(&icq->q_node); |
| 73 | |
| 74 | /* |
| 75 | * Both setting lookup hint to and clearing it from @icq are done |
| 76 | * under queue_lock. If it's not pointing to @icq now, it never |
| 77 | * will. Hint assignment itself can race safely. |
| 78 | */ |
| 79 | if (rcu_access_pointer(ioc->icq_hint) == icq) |
| 80 | rcu_assign_pointer(ioc->icq_hint, NULL); |
| 81 | |
| 82 | ioc_exit_icq(icq); |
| 83 | |
| 84 | /* |
| 85 | * @icq->q might have gone away by the time RCU callback runs |
| 86 | * making it impossible to determine icq_cache. Record it in @icq. |
| 87 | */ |
| 88 | icq->__rcu_icq_cache = et->icq_cache; |
| 89 | call_rcu(&icq->__rcu_head, icq_free_icq_rcu); |
| 90 | } |
| 91 | |
| 92 | /* |
| 93 | * Slow path for ioc release in put_io_context(). Performs double-lock |
| 94 | * dancing to unlink all icq's and then frees ioc. |
| 95 | */ |
| 96 | static void ioc_release_fn(struct work_struct *work) |
| 97 | { |
| 98 | struct io_context *ioc = container_of(work, struct io_context, |
| 99 | release_work); |
| 100 | unsigned long flags; |
| 101 | |
| 102 | /* |
| 103 | * Exiting icq may call into put_io_context() through elevator |
| 104 | * which will trigger lockdep warning. The ioc's are guaranteed to |
| 105 | * be different, use a different locking subclass here. Use |
| 106 | * irqsave variant as there's no spin_lock_irq_nested(). |
| 107 | */ |
| 108 | spin_lock_irqsave_nested(&ioc->lock, flags, 1); |
| 109 | |
| 110 | while (!hlist_empty(&ioc->icq_list)) { |
| 111 | struct io_cq *icq = hlist_entry(ioc->icq_list.first, |
| 112 | struct io_cq, ioc_node); |
| 113 | struct request_queue *q = icq->q; |
| 114 | |
| 115 | if (spin_trylock(q->queue_lock)) { |
| 116 | ioc_destroy_icq(icq); |
| 117 | spin_unlock(q->queue_lock); |
| 118 | } else { |
| 119 | spin_unlock_irqrestore(&ioc->lock, flags); |
| 120 | cpu_relax(); |
| 121 | spin_lock_irqsave_nested(&ioc->lock, flags, 1); |
| 122 | } |
| 123 | } |
| 124 | |
| 125 | spin_unlock_irqrestore(&ioc->lock, flags); |
| 126 | |
| 127 | kmem_cache_free(iocontext_cachep, ioc); |
| 128 | } |
| 129 | |
| 130 | /** |
| 131 | * put_io_context - put a reference of io_context |
| 132 | * @ioc: io_context to put |
| 133 | * |
| 134 | * Decrement reference count of @ioc and release it if the count reaches |
| 135 | * zero. |
| 136 | */ |
| 137 | void put_io_context(struct io_context *ioc) |
| 138 | { |
| 139 | unsigned long flags; |
| 140 | bool free_ioc = false; |
| 141 | |
| 142 | if (ioc == NULL) |
| 143 | return; |
| 144 | |
| 145 | BUG_ON(atomic_long_read(&ioc->refcount) <= 0); |
| 146 | |
| 147 | /* |
| 148 | * Releasing ioc requires reverse order double locking and we may |
| 149 | * already be holding a queue_lock. Do it asynchronously from wq. |
| 150 | */ |
| 151 | if (atomic_long_dec_and_test(&ioc->refcount)) { |
| 152 | spin_lock_irqsave(&ioc->lock, flags); |
| 153 | if (!hlist_empty(&ioc->icq_list)) |
| 154 | queue_work(system_power_efficient_wq, |
| 155 | &ioc->release_work); |
| 156 | else |
| 157 | free_ioc = true; |
| 158 | spin_unlock_irqrestore(&ioc->lock, flags); |
| 159 | } |
| 160 | |
| 161 | if (free_ioc) |
| 162 | kmem_cache_free(iocontext_cachep, ioc); |
| 163 | } |
| 164 | EXPORT_SYMBOL(put_io_context); |
| 165 | |
| 166 | /** |
| 167 | * put_io_context_active - put active reference on ioc |
| 168 | * @ioc: ioc of interest |
| 169 | * |
| 170 | * Undo get_io_context_active(). If active reference reaches zero after |
| 171 | * put, @ioc can never issue further IOs and ioscheds are notified. |
| 172 | */ |
| 173 | void put_io_context_active(struct io_context *ioc) |
| 174 | { |
| 175 | struct elevator_type *et; |
| 176 | unsigned long flags; |
| 177 | struct io_cq *icq; |
| 178 | |
| 179 | if (!atomic_dec_and_test(&ioc->active_ref)) { |
| 180 | put_io_context(ioc); |
| 181 | return; |
| 182 | } |
| 183 | |
| 184 | /* |
| 185 | * Need ioc lock to walk icq_list and q lock to exit icq. Perform |
| 186 | * reverse double locking. Read comment in ioc_release_fn() for |
| 187 | * explanation on the nested locking annotation. |
| 188 | */ |
| 189 | retry: |
| 190 | spin_lock_irqsave_nested(&ioc->lock, flags, 1); |
| 191 | hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) { |
| 192 | if (icq->flags & ICQ_EXITED) |
| 193 | continue; |
| 194 | |
| 195 | et = icq->q->elevator->type; |
| 196 | if (et->uses_mq) { |
| 197 | ioc_exit_icq(icq); |
| 198 | } else { |
| 199 | if (spin_trylock(icq->q->queue_lock)) { |
| 200 | ioc_exit_icq(icq); |
| 201 | spin_unlock(icq->q->queue_lock); |
| 202 | } else { |
| 203 | spin_unlock_irqrestore(&ioc->lock, flags); |
| 204 | cpu_relax(); |
| 205 | goto retry; |
| 206 | } |
| 207 | } |
| 208 | } |
| 209 | spin_unlock_irqrestore(&ioc->lock, flags); |
| 210 | |
| 211 | put_io_context(ioc); |
| 212 | } |
| 213 | |
| 214 | /* Called by the exiting task */ |
| 215 | void exit_io_context(struct task_struct *task) |
| 216 | { |
| 217 | struct io_context *ioc; |
| 218 | |
| 219 | task_lock(task); |
| 220 | ioc = task->io_context; |
| 221 | task->io_context = NULL; |
| 222 | task_unlock(task); |
| 223 | |
| 224 | atomic_dec(&ioc->nr_tasks); |
| 225 | put_io_context_active(ioc); |
| 226 | } |
| 227 | |
| 228 | static void __ioc_clear_queue(struct list_head *icq_list) |
| 229 | { |
| 230 | unsigned long flags; |
| 231 | |
| 232 | while (!list_empty(icq_list)) { |
| 233 | struct io_cq *icq = list_entry(icq_list->next, |
| 234 | struct io_cq, q_node); |
| 235 | struct io_context *ioc = icq->ioc; |
| 236 | |
| 237 | spin_lock_irqsave(&ioc->lock, flags); |
| 238 | ioc_destroy_icq(icq); |
| 239 | spin_unlock_irqrestore(&ioc->lock, flags); |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | /** |
| 244 | * ioc_clear_queue - break any ioc association with the specified queue |
| 245 | * @q: request_queue being cleared |
| 246 | * |
| 247 | * Walk @q->icq_list and exit all io_cq's. |
| 248 | */ |
| 249 | void ioc_clear_queue(struct request_queue *q) |
| 250 | { |
| 251 | LIST_HEAD(icq_list); |
| 252 | |
| 253 | spin_lock_irq(q->queue_lock); |
| 254 | list_splice_init(&q->icq_list, &icq_list); |
| 255 | |
| 256 | if (q->mq_ops) { |
| 257 | spin_unlock_irq(q->queue_lock); |
| 258 | __ioc_clear_queue(&icq_list); |
| 259 | } else { |
| 260 | __ioc_clear_queue(&icq_list); |
| 261 | spin_unlock_irq(q->queue_lock); |
| 262 | } |
| 263 | } |
| 264 | |
| 265 | int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node) |
| 266 | { |
| 267 | struct io_context *ioc; |
| 268 | int ret; |
| 269 | |
| 270 | ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO, |
| 271 | node); |
| 272 | if (unlikely(!ioc)) |
| 273 | return -ENOMEM; |
| 274 | |
| 275 | /* initialize */ |
| 276 | atomic_long_set(&ioc->refcount, 1); |
| 277 | atomic_set(&ioc->nr_tasks, 1); |
| 278 | atomic_set(&ioc->active_ref, 1); |
| 279 | spin_lock_init(&ioc->lock); |
| 280 | INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH); |
| 281 | INIT_HLIST_HEAD(&ioc->icq_list); |
| 282 | INIT_WORK(&ioc->release_work, ioc_release_fn); |
| 283 | |
| 284 | /* |
| 285 | * Try to install. ioc shouldn't be installed if someone else |
| 286 | * already did or @task, which isn't %current, is exiting. Note |
| 287 | * that we need to allow ioc creation on exiting %current as exit |
| 288 | * path may issue IOs from e.g. exit_files(). The exit path is |
| 289 | * responsible for not issuing IO after exit_io_context(). |
| 290 | */ |
| 291 | task_lock(task); |
| 292 | if (!task->io_context && |
| 293 | (task == current || !(task->flags & PF_EXITING))) |
| 294 | task->io_context = ioc; |
| 295 | else |
| 296 | kmem_cache_free(iocontext_cachep, ioc); |
| 297 | |
| 298 | ret = task->io_context ? 0 : -EBUSY; |
| 299 | |
| 300 | task_unlock(task); |
| 301 | |
| 302 | return ret; |
| 303 | } |
| 304 | |
| 305 | /** |
| 306 | * get_task_io_context - get io_context of a task |
| 307 | * @task: task of interest |
| 308 | * @gfp_flags: allocation flags, used if allocation is necessary |
| 309 | * @node: allocation node, used if allocation is necessary |
| 310 | * |
| 311 | * Return io_context of @task. If it doesn't exist, it is created with |
| 312 | * @gfp_flags and @node. The returned io_context has its reference count |
| 313 | * incremented. |
| 314 | * |
| 315 | * This function always goes through task_lock() and it's better to use |
| 316 | * %current->io_context + get_io_context() for %current. |
| 317 | */ |
| 318 | struct io_context *get_task_io_context(struct task_struct *task, |
| 319 | gfp_t gfp_flags, int node) |
| 320 | { |
| 321 | struct io_context *ioc; |
| 322 | |
| 323 | might_sleep_if(gfpflags_allow_blocking(gfp_flags)); |
| 324 | |
| 325 | do { |
| 326 | task_lock(task); |
| 327 | ioc = task->io_context; |
| 328 | if (likely(ioc)) { |
| 329 | get_io_context(ioc); |
| 330 | task_unlock(task); |
| 331 | return ioc; |
| 332 | } |
| 333 | task_unlock(task); |
| 334 | } while (!create_task_io_context(task, gfp_flags, node)); |
| 335 | |
| 336 | return NULL; |
| 337 | } |
| 338 | EXPORT_SYMBOL(get_task_io_context); |
| 339 | |
| 340 | /** |
| 341 | * ioc_lookup_icq - lookup io_cq from ioc |
| 342 | * @ioc: the associated io_context |
| 343 | * @q: the associated request_queue |
| 344 | * |
| 345 | * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called |
| 346 | * with @q->queue_lock held. |
| 347 | */ |
| 348 | struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q) |
| 349 | { |
| 350 | struct io_cq *icq; |
| 351 | |
| 352 | lockdep_assert_held(q->queue_lock); |
| 353 | |
| 354 | /* |
| 355 | * icq's are indexed from @ioc using radix tree and hint pointer, |
| 356 | * both of which are protected with RCU. All removals are done |
| 357 | * holding both q and ioc locks, and we're holding q lock - if we |
| 358 | * find a icq which points to us, it's guaranteed to be valid. |
| 359 | */ |
| 360 | rcu_read_lock(); |
| 361 | icq = rcu_dereference(ioc->icq_hint); |
| 362 | if (icq && icq->q == q) |
| 363 | goto out; |
| 364 | |
| 365 | icq = radix_tree_lookup(&ioc->icq_tree, q->id); |
| 366 | if (icq && icq->q == q) |
| 367 | rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */ |
| 368 | else |
| 369 | icq = NULL; |
| 370 | out: |
| 371 | rcu_read_unlock(); |
| 372 | return icq; |
| 373 | } |
| 374 | EXPORT_SYMBOL(ioc_lookup_icq); |
| 375 | |
| 376 | /** |
| 377 | * ioc_create_icq - create and link io_cq |
| 378 | * @ioc: io_context of interest |
| 379 | * @q: request_queue of interest |
| 380 | * @gfp_mask: allocation mask |
| 381 | * |
| 382 | * Make sure io_cq linking @ioc and @q exists. If icq doesn't exist, they |
| 383 | * will be created using @gfp_mask. |
| 384 | * |
| 385 | * The caller is responsible for ensuring @ioc won't go away and @q is |
| 386 | * alive and will stay alive until this function returns. |
| 387 | */ |
| 388 | struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q, |
| 389 | gfp_t gfp_mask) |
| 390 | { |
| 391 | struct elevator_type *et = q->elevator->type; |
| 392 | struct io_cq *icq; |
| 393 | |
| 394 | /* allocate stuff */ |
| 395 | icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO, |
| 396 | q->node); |
| 397 | if (!icq) |
| 398 | return NULL; |
| 399 | |
| 400 | if (radix_tree_maybe_preload(gfp_mask) < 0) { |
| 401 | kmem_cache_free(et->icq_cache, icq); |
| 402 | return NULL; |
| 403 | } |
| 404 | |
| 405 | icq->ioc = ioc; |
| 406 | icq->q = q; |
| 407 | INIT_LIST_HEAD(&icq->q_node); |
| 408 | INIT_HLIST_NODE(&icq->ioc_node); |
| 409 | |
| 410 | /* lock both q and ioc and try to link @icq */ |
| 411 | spin_lock_irq(q->queue_lock); |
| 412 | spin_lock(&ioc->lock); |
| 413 | |
| 414 | if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) { |
| 415 | hlist_add_head(&icq->ioc_node, &ioc->icq_list); |
| 416 | list_add(&icq->q_node, &q->icq_list); |
| 417 | if (et->uses_mq && et->ops.mq.init_icq) |
| 418 | et->ops.mq.init_icq(icq); |
| 419 | else if (!et->uses_mq && et->ops.sq.elevator_init_icq_fn) |
| 420 | et->ops.sq.elevator_init_icq_fn(icq); |
| 421 | } else { |
| 422 | kmem_cache_free(et->icq_cache, icq); |
| 423 | icq = ioc_lookup_icq(ioc, q); |
| 424 | if (!icq) |
| 425 | printk(KERN_ERR "cfq: icq link failed!\n"); |
| 426 | } |
| 427 | |
| 428 | spin_unlock(&ioc->lock); |
| 429 | spin_unlock_irq(q->queue_lock); |
| 430 | radix_tree_preload_end(); |
| 431 | return icq; |
| 432 | } |
| 433 | |
| 434 | static int __init blk_ioc_init(void) |
| 435 | { |
| 436 | iocontext_cachep = kmem_cache_create("blkdev_ioc", |
| 437 | sizeof(struct io_context), 0, SLAB_PANIC, NULL); |
| 438 | return 0; |
| 439 | } |
| 440 | subsys_initcall(blk_ioc_init); |