| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Functions related to io context handling |
| 4 | */ |
| 5 | #include <linux/kernel.h> |
| 6 | #include <linux/module.h> |
| 7 | #include <linux/init.h> |
| 8 | #include <linux/bio.h> |
| 9 | #include <linux/blkdev.h> |
| 10 | #include <linux/slab.h> |
| 11 | #include <linux/security.h> |
| 12 | #include <linux/sched/task.h> |
| 13 | |
| 14 | #include "blk.h" |
| 15 | #include "blk-mq-sched.h" |
| 16 | |
| 17 | /* |
| 18 | * For io context allocations |
| 19 | */ |
| 20 | static struct kmem_cache *iocontext_cachep; |
| 21 | |
| 22 | #ifdef CONFIG_BLK_ICQ |
| 23 | /** |
| 24 | * get_io_context - increment reference count to io_context |
| 25 | * @ioc: io_context to get |
| 26 | * |
| 27 | * Increment reference count to @ioc. |
| 28 | */ |
| 29 | static void get_io_context(struct io_context *ioc) |
| 30 | { |
| 31 | BUG_ON(atomic_long_read(&ioc->refcount) <= 0); |
| 32 | atomic_long_inc(&ioc->refcount); |
| 33 | } |
| 34 | |
| 35 | static void icq_free_icq_rcu(struct rcu_head *head) |
| 36 | { |
| 37 | struct io_cq *icq = container_of(head, struct io_cq, __rcu_head); |
| 38 | |
| 39 | kmem_cache_free(icq->__rcu_icq_cache, icq); |
| 40 | } |
| 41 | |
| 42 | /* |
| 43 | * Exit an icq. Called with ioc locked for blk-mq, and with both ioc |
| 44 | * and queue locked for legacy. |
| 45 | */ |
| 46 | static void ioc_exit_icq(struct io_cq *icq) |
| 47 | { |
| 48 | struct elevator_type *et = icq->q->elevator->type; |
| 49 | |
| 50 | if (icq->flags & ICQ_EXITED) |
| 51 | return; |
| 52 | |
| 53 | if (et->ops.exit_icq) |
| 54 | et->ops.exit_icq(icq); |
| 55 | |
| 56 | icq->flags |= ICQ_EXITED; |
| 57 | } |
| 58 | |
| 59 | static void ioc_exit_icqs(struct io_context *ioc) |
| 60 | { |
| 61 | struct io_cq *icq; |
| 62 | |
| 63 | spin_lock_irq(&ioc->lock); |
| 64 | hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) |
| 65 | ioc_exit_icq(icq); |
| 66 | spin_unlock_irq(&ioc->lock); |
| 67 | } |
| 68 | |
| 69 | /* |
| 70 | * Release an icq. Called with ioc locked for blk-mq, and with both ioc |
| 71 | * and queue locked for legacy. |
| 72 | */ |
| 73 | static void ioc_destroy_icq(struct io_cq *icq) |
| 74 | { |
| 75 | struct io_context *ioc = icq->ioc; |
| 76 | struct request_queue *q = icq->q; |
| 77 | struct elevator_type *et = q->elevator->type; |
| 78 | |
| 79 | lockdep_assert_held(&ioc->lock); |
| 80 | lockdep_assert_held(&q->queue_lock); |
| 81 | |
| 82 | if (icq->flags & ICQ_DESTROYED) |
| 83 | return; |
| 84 | |
| 85 | radix_tree_delete(&ioc->icq_tree, icq->q->id); |
| 86 | hlist_del_init(&icq->ioc_node); |
| 87 | list_del_init(&icq->q_node); |
| 88 | |
| 89 | /* |
| 90 | * Both setting lookup hint to and clearing it from @icq are done |
| 91 | * under queue_lock. If it's not pointing to @icq now, it never |
| 92 | * will. Hint assignment itself can race safely. |
| 93 | */ |
| 94 | if (rcu_access_pointer(ioc->icq_hint) == icq) |
| 95 | rcu_assign_pointer(ioc->icq_hint, NULL); |
| 96 | |
| 97 | ioc_exit_icq(icq); |
| 98 | |
| 99 | /* |
| 100 | * @icq->q might have gone away by the time RCU callback runs |
| 101 | * making it impossible to determine icq_cache. Record it in @icq. |
| 102 | */ |
| 103 | icq->__rcu_icq_cache = et->icq_cache; |
| 104 | icq->flags |= ICQ_DESTROYED; |
| 105 | call_rcu(&icq->__rcu_head, icq_free_icq_rcu); |
| 106 | } |
| 107 | |
| 108 | /* |
| 109 | * Slow path for ioc release in put_io_context(). Performs double-lock |
| 110 | * dancing to unlink all icq's and then frees ioc. |
| 111 | */ |
| 112 | static void ioc_release_fn(struct work_struct *work) |
| 113 | { |
| 114 | struct io_context *ioc = container_of(work, struct io_context, |
| 115 | release_work); |
| 116 | spin_lock_irq(&ioc->lock); |
| 117 | |
| 118 | while (!hlist_empty(&ioc->icq_list)) { |
| 119 | struct io_cq *icq = hlist_entry(ioc->icq_list.first, |
| 120 | struct io_cq, ioc_node); |
| 121 | struct request_queue *q = icq->q; |
| 122 | |
| 123 | if (spin_trylock(&q->queue_lock)) { |
| 124 | ioc_destroy_icq(icq); |
| 125 | spin_unlock(&q->queue_lock); |
| 126 | } else { |
| 127 | /* Make sure q and icq cannot be freed. */ |
| 128 | rcu_read_lock(); |
| 129 | |
| 130 | /* Re-acquire the locks in the correct order. */ |
| 131 | spin_unlock(&ioc->lock); |
| 132 | spin_lock(&q->queue_lock); |
| 133 | spin_lock(&ioc->lock); |
| 134 | |
| 135 | ioc_destroy_icq(icq); |
| 136 | |
| 137 | spin_unlock(&q->queue_lock); |
| 138 | rcu_read_unlock(); |
| 139 | } |
| 140 | } |
| 141 | |
| 142 | spin_unlock_irq(&ioc->lock); |
| 143 | |
| 144 | kmem_cache_free(iocontext_cachep, ioc); |
| 145 | } |
| 146 | |
| 147 | /* |
| 148 | * Releasing icqs requires reverse order double locking and we may already be |
| 149 | * holding a queue_lock. Do it asynchronously from a workqueue. |
| 150 | */ |
| 151 | static bool ioc_delay_free(struct io_context *ioc) |
| 152 | { |
| 153 | unsigned long flags; |
| 154 | |
| 155 | spin_lock_irqsave(&ioc->lock, flags); |
| 156 | if (!hlist_empty(&ioc->icq_list)) { |
| 157 | queue_work(system_power_efficient_wq, &ioc->release_work); |
| 158 | spin_unlock_irqrestore(&ioc->lock, flags); |
| 159 | return true; |
| 160 | } |
| 161 | spin_unlock_irqrestore(&ioc->lock, flags); |
| 162 | return false; |
| 163 | } |
| 164 | |
| 165 | /** |
| 166 | * ioc_clear_queue - break any ioc association with the specified queue |
| 167 | * @q: request_queue being cleared |
| 168 | * |
| 169 | * Walk @q->icq_list and exit all io_cq's. |
| 170 | */ |
| 171 | void ioc_clear_queue(struct request_queue *q) |
| 172 | { |
| 173 | spin_lock_irq(&q->queue_lock); |
| 174 | while (!list_empty(&q->icq_list)) { |
| 175 | struct io_cq *icq = |
| 176 | list_first_entry(&q->icq_list, struct io_cq, q_node); |
| 177 | |
| 178 | /* |
| 179 | * Other context won't hold ioc lock to wait for queue_lock, see |
| 180 | * details in ioc_release_fn(). |
| 181 | */ |
| 182 | spin_lock(&icq->ioc->lock); |
| 183 | ioc_destroy_icq(icq); |
| 184 | spin_unlock(&icq->ioc->lock); |
| 185 | } |
| 186 | spin_unlock_irq(&q->queue_lock); |
| 187 | } |
| 188 | #else /* CONFIG_BLK_ICQ */ |
| 189 | static inline void ioc_exit_icqs(struct io_context *ioc) |
| 190 | { |
| 191 | } |
| 192 | static inline bool ioc_delay_free(struct io_context *ioc) |
| 193 | { |
| 194 | return false; |
| 195 | } |
| 196 | #endif /* CONFIG_BLK_ICQ */ |
| 197 | |
| 198 | /** |
| 199 | * put_io_context - put a reference of io_context |
| 200 | * @ioc: io_context to put |
| 201 | * |
| 202 | * Decrement reference count of @ioc and release it if the count reaches |
| 203 | * zero. |
| 204 | */ |
| 205 | void put_io_context(struct io_context *ioc) |
| 206 | { |
| 207 | BUG_ON(atomic_long_read(&ioc->refcount) <= 0); |
| 208 | if (atomic_long_dec_and_test(&ioc->refcount) && !ioc_delay_free(ioc)) |
| 209 | kmem_cache_free(iocontext_cachep, ioc); |
| 210 | } |
| 211 | EXPORT_SYMBOL_GPL(put_io_context); |
| 212 | |
| 213 | /* Called by the exiting task */ |
| 214 | void exit_io_context(struct task_struct *task) |
| 215 | { |
| 216 | struct io_context *ioc; |
| 217 | |
| 218 | task_lock(task); |
| 219 | ioc = task->io_context; |
| 220 | task->io_context = NULL; |
| 221 | task_unlock(task); |
| 222 | |
| 223 | if (atomic_dec_and_test(&ioc->active_ref)) { |
| 224 | ioc_exit_icqs(ioc); |
| 225 | put_io_context(ioc); |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | static struct io_context *alloc_io_context(gfp_t gfp_flags, int node) |
| 230 | { |
| 231 | struct io_context *ioc; |
| 232 | |
| 233 | ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO, |
| 234 | node); |
| 235 | if (unlikely(!ioc)) |
| 236 | return NULL; |
| 237 | |
| 238 | atomic_long_set(&ioc->refcount, 1); |
| 239 | atomic_set(&ioc->active_ref, 1); |
| 240 | #ifdef CONFIG_BLK_ICQ |
| 241 | spin_lock_init(&ioc->lock); |
| 242 | INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC); |
| 243 | INIT_HLIST_HEAD(&ioc->icq_list); |
| 244 | INIT_WORK(&ioc->release_work, ioc_release_fn); |
| 245 | #endif |
| 246 | ioc->ioprio = IOPRIO_DEFAULT; |
| 247 | |
| 248 | return ioc; |
| 249 | } |
| 250 | |
| 251 | int set_task_ioprio(struct task_struct *task, int ioprio) |
| 252 | { |
| 253 | int err; |
| 254 | const struct cred *cred = current_cred(), *tcred; |
| 255 | |
| 256 | rcu_read_lock(); |
| 257 | tcred = __task_cred(task); |
| 258 | if (!uid_eq(tcred->uid, cred->euid) && |
| 259 | !uid_eq(tcred->uid, cred->uid) && !capable(CAP_SYS_NICE)) { |
| 260 | rcu_read_unlock(); |
| 261 | return -EPERM; |
| 262 | } |
| 263 | rcu_read_unlock(); |
| 264 | |
| 265 | err = security_task_setioprio(task, ioprio); |
| 266 | if (err) |
| 267 | return err; |
| 268 | |
| 269 | task_lock(task); |
| 270 | if (unlikely(!task->io_context)) { |
| 271 | struct io_context *ioc; |
| 272 | |
| 273 | task_unlock(task); |
| 274 | |
| 275 | ioc = alloc_io_context(GFP_ATOMIC, NUMA_NO_NODE); |
| 276 | if (!ioc) |
| 277 | return -ENOMEM; |
| 278 | |
| 279 | task_lock(task); |
| 280 | if (task->flags & PF_EXITING) { |
| 281 | kmem_cache_free(iocontext_cachep, ioc); |
| 282 | goto out; |
| 283 | } |
| 284 | if (task->io_context) |
| 285 | kmem_cache_free(iocontext_cachep, ioc); |
| 286 | else |
| 287 | task->io_context = ioc; |
| 288 | } |
| 289 | task->io_context->ioprio = ioprio; |
| 290 | out: |
| 291 | task_unlock(task); |
| 292 | return 0; |
| 293 | } |
| 294 | EXPORT_SYMBOL_GPL(set_task_ioprio); |
| 295 | |
| 296 | int __copy_io(unsigned long clone_flags, struct task_struct *tsk) |
| 297 | { |
| 298 | struct io_context *ioc = current->io_context; |
| 299 | |
| 300 | /* |
| 301 | * Share io context with parent, if CLONE_IO is set |
| 302 | */ |
| 303 | if (clone_flags & CLONE_IO) { |
| 304 | atomic_inc(&ioc->active_ref); |
| 305 | tsk->io_context = ioc; |
| 306 | } else if (ioprio_valid(ioc->ioprio)) { |
| 307 | tsk->io_context = alloc_io_context(GFP_KERNEL, NUMA_NO_NODE); |
| 308 | if (!tsk->io_context) |
| 309 | return -ENOMEM; |
| 310 | tsk->io_context->ioprio = ioc->ioprio; |
| 311 | } |
| 312 | |
| 313 | return 0; |
| 314 | } |
| 315 | |
| 316 | #ifdef CONFIG_BLK_ICQ |
| 317 | /** |
| 318 | * ioc_lookup_icq - lookup io_cq from ioc |
| 319 | * @q: the associated request_queue |
| 320 | * |
| 321 | * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called |
| 322 | * with @q->queue_lock held. |
| 323 | */ |
| 324 | struct io_cq *ioc_lookup_icq(struct request_queue *q) |
| 325 | { |
| 326 | struct io_context *ioc = current->io_context; |
| 327 | struct io_cq *icq; |
| 328 | |
| 329 | lockdep_assert_held(&q->queue_lock); |
| 330 | |
| 331 | /* |
| 332 | * icq's are indexed from @ioc using radix tree and hint pointer, |
| 333 | * both of which are protected with RCU. All removals are done |
| 334 | * holding both q and ioc locks, and we're holding q lock - if we |
| 335 | * find a icq which points to us, it's guaranteed to be valid. |
| 336 | */ |
| 337 | rcu_read_lock(); |
| 338 | icq = rcu_dereference(ioc->icq_hint); |
| 339 | if (icq && icq->q == q) |
| 340 | goto out; |
| 341 | |
| 342 | icq = radix_tree_lookup(&ioc->icq_tree, q->id); |
| 343 | if (icq && icq->q == q) |
| 344 | rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */ |
| 345 | else |
| 346 | icq = NULL; |
| 347 | out: |
| 348 | rcu_read_unlock(); |
| 349 | return icq; |
| 350 | } |
| 351 | EXPORT_SYMBOL(ioc_lookup_icq); |
| 352 | |
| 353 | /** |
| 354 | * ioc_create_icq - create and link io_cq |
| 355 | * @q: request_queue of interest |
| 356 | * |
| 357 | * Make sure io_cq linking @ioc and @q exists. If icq doesn't exist, they |
| 358 | * will be created using @gfp_mask. |
| 359 | * |
| 360 | * The caller is responsible for ensuring @ioc won't go away and @q is |
| 361 | * alive and will stay alive until this function returns. |
| 362 | */ |
| 363 | static struct io_cq *ioc_create_icq(struct request_queue *q) |
| 364 | { |
| 365 | struct io_context *ioc = current->io_context; |
| 366 | struct elevator_type *et = q->elevator->type; |
| 367 | struct io_cq *icq; |
| 368 | |
| 369 | /* allocate stuff */ |
| 370 | icq = kmem_cache_alloc_node(et->icq_cache, GFP_ATOMIC | __GFP_ZERO, |
| 371 | q->node); |
| 372 | if (!icq) |
| 373 | return NULL; |
| 374 | |
| 375 | if (radix_tree_maybe_preload(GFP_ATOMIC) < 0) { |
| 376 | kmem_cache_free(et->icq_cache, icq); |
| 377 | return NULL; |
| 378 | } |
| 379 | |
| 380 | icq->ioc = ioc; |
| 381 | icq->q = q; |
| 382 | INIT_LIST_HEAD(&icq->q_node); |
| 383 | INIT_HLIST_NODE(&icq->ioc_node); |
| 384 | |
| 385 | /* lock both q and ioc and try to link @icq */ |
| 386 | spin_lock_irq(&q->queue_lock); |
| 387 | spin_lock(&ioc->lock); |
| 388 | |
| 389 | if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) { |
| 390 | hlist_add_head(&icq->ioc_node, &ioc->icq_list); |
| 391 | list_add(&icq->q_node, &q->icq_list); |
| 392 | if (et->ops.init_icq) |
| 393 | et->ops.init_icq(icq); |
| 394 | } else { |
| 395 | kmem_cache_free(et->icq_cache, icq); |
| 396 | icq = ioc_lookup_icq(q); |
| 397 | if (!icq) |
| 398 | printk(KERN_ERR "cfq: icq link failed!\n"); |
| 399 | } |
| 400 | |
| 401 | spin_unlock(&ioc->lock); |
| 402 | spin_unlock_irq(&q->queue_lock); |
| 403 | radix_tree_preload_end(); |
| 404 | return icq; |
| 405 | } |
| 406 | |
| 407 | struct io_cq *ioc_find_get_icq(struct request_queue *q) |
| 408 | { |
| 409 | struct io_context *ioc = current->io_context; |
| 410 | struct io_cq *icq = NULL; |
| 411 | |
| 412 | if (unlikely(!ioc)) { |
| 413 | ioc = alloc_io_context(GFP_ATOMIC, q->node); |
| 414 | if (!ioc) |
| 415 | return NULL; |
| 416 | |
| 417 | task_lock(current); |
| 418 | if (current->io_context) { |
| 419 | kmem_cache_free(iocontext_cachep, ioc); |
| 420 | ioc = current->io_context; |
| 421 | } else { |
| 422 | current->io_context = ioc; |
| 423 | } |
| 424 | |
| 425 | get_io_context(ioc); |
| 426 | task_unlock(current); |
| 427 | } else { |
| 428 | get_io_context(ioc); |
| 429 | |
| 430 | spin_lock_irq(&q->queue_lock); |
| 431 | icq = ioc_lookup_icq(q); |
| 432 | spin_unlock_irq(&q->queue_lock); |
| 433 | } |
| 434 | |
| 435 | if (!icq) { |
| 436 | icq = ioc_create_icq(q); |
| 437 | if (!icq) { |
| 438 | put_io_context(ioc); |
| 439 | return NULL; |
| 440 | } |
| 441 | } |
| 442 | return icq; |
| 443 | } |
| 444 | EXPORT_SYMBOL_GPL(ioc_find_get_icq); |
| 445 | #endif /* CONFIG_BLK_ICQ */ |
| 446 | |
| 447 | static int __init blk_ioc_init(void) |
| 448 | { |
| 449 | iocontext_cachep = kmem_cache_create("blkdev_ioc", |
| 450 | sizeof(struct io_context), 0, SLAB_PANIC, NULL); |
| 451 | return 0; |
| 452 | } |
| 453 | subsys_initcall(blk_ioc_init); |