2 * Functions related to io context handling
4 #include <linux/kernel.h>
5 #include <linux/module.h>
6 #include <linux/init.h>
8 #include <linux/blkdev.h>
9 #include <linux/slab.h>
10 #include <linux/sched/task.h>
15 * For io context allocations
17 static struct kmem_cache *iocontext_cachep;
20 * get_io_context - increment reference count to io_context
21 * @ioc: io_context to get
23 * Increment reference count to @ioc.
25 void get_io_context(struct io_context *ioc)
27 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
28 atomic_long_inc(&ioc->refcount);
30 EXPORT_SYMBOL(get_io_context);
32 static void icq_free_icq_rcu(struct rcu_head *head)
34 struct io_cq *icq = container_of(head, struct io_cq, __rcu_head);
36 kmem_cache_free(icq->__rcu_icq_cache, icq);
40 * Exit an icq. Called with both ioc and q locked for sq, only ioc locked for
43 static void ioc_exit_icq(struct io_cq *icq)
45 struct elevator_type *et = icq->q->elevator->type;
47 if (icq->flags & ICQ_EXITED)
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);
55 icq->flags |= ICQ_EXITED;
58 /* Release an icq. Called with both ioc and q locked. */
59 static void ioc_destroy_icq(struct io_cq *icq)
61 struct io_context *ioc = icq->ioc;
62 struct request_queue *q = icq->q;
63 struct elevator_type *et = q->elevator->type;
65 lockdep_assert_held(&ioc->lock);
66 lockdep_assert_held(q->queue_lock);
68 radix_tree_delete(&ioc->icq_tree, icq->q->id);
69 hlist_del_init(&icq->ioc_node);
70 list_del_init(&icq->q_node);
73 * Both setting lookup hint to and clearing it from @icq are done
74 * under queue_lock. If it's not pointing to @icq now, it never
75 * will. Hint assignment itself can race safely.
77 if (rcu_access_pointer(ioc->icq_hint) == icq)
78 rcu_assign_pointer(ioc->icq_hint, NULL);
83 * @icq->q might have gone away by the time RCU callback runs
84 * making it impossible to determine icq_cache. Record it in @icq.
86 icq->__rcu_icq_cache = et->icq_cache;
87 call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
91 * Slow path for ioc release in put_io_context(). Performs double-lock
92 * dancing to unlink all icq's and then frees ioc.
94 static void ioc_release_fn(struct work_struct *work)
96 struct io_context *ioc = container_of(work, struct io_context,
101 * Exiting icq may call into put_io_context() through elevator
102 * which will trigger lockdep warning. The ioc's are guaranteed to
103 * be different, use a different locking subclass here. Use
104 * irqsave variant as there's no spin_lock_irq_nested().
106 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
108 while (!hlist_empty(&ioc->icq_list)) {
109 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
110 struct io_cq, ioc_node);
111 struct request_queue *q = icq->q;
113 if (spin_trylock(q->queue_lock)) {
114 ioc_destroy_icq(icq);
115 spin_unlock(q->queue_lock);
117 spin_unlock_irqrestore(&ioc->lock, flags);
119 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
123 spin_unlock_irqrestore(&ioc->lock, flags);
125 kmem_cache_free(iocontext_cachep, ioc);
129 * put_io_context - put a reference of io_context
130 * @ioc: io_context to put
132 * Decrement reference count of @ioc and release it if the count reaches
135 void put_io_context(struct io_context *ioc)
138 bool free_ioc = false;
143 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
146 * Releasing ioc requires reverse order double locking and we may
147 * already be holding a queue_lock. Do it asynchronously from wq.
149 if (atomic_long_dec_and_test(&ioc->refcount)) {
150 spin_lock_irqsave(&ioc->lock, flags);
151 if (!hlist_empty(&ioc->icq_list))
152 queue_work(system_power_efficient_wq,
156 spin_unlock_irqrestore(&ioc->lock, flags);
160 kmem_cache_free(iocontext_cachep, ioc);
162 EXPORT_SYMBOL(put_io_context);
165 * put_io_context_active - put active reference on ioc
166 * @ioc: ioc of interest
168 * Undo get_io_context_active(). If active reference reaches zero after
169 * put, @ioc can never issue further IOs and ioscheds are notified.
171 void put_io_context_active(struct io_context *ioc)
173 struct elevator_type *et;
177 if (!atomic_dec_and_test(&ioc->active_ref)) {
183 * Need ioc lock to walk icq_list and q lock to exit icq. Perform
184 * reverse double locking. Read comment in ioc_release_fn() for
185 * explanation on the nested locking annotation.
188 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
189 hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) {
190 if (icq->flags & ICQ_EXITED)
193 et = icq->q->elevator->type;
197 if (spin_trylock(icq->q->queue_lock)) {
199 spin_unlock(icq->q->queue_lock);
201 spin_unlock_irqrestore(&ioc->lock, flags);
207 spin_unlock_irqrestore(&ioc->lock, flags);
212 /* Called by the exiting task */
213 void exit_io_context(struct task_struct *task)
215 struct io_context *ioc;
218 ioc = task->io_context;
219 task->io_context = NULL;
222 atomic_dec(&ioc->nr_tasks);
223 put_io_context_active(ioc);
227 * ioc_clear_queue - break any ioc association with the specified queue
228 * @q: request_queue being cleared
230 * Walk @q->icq_list and exit all io_cq's. Must be called with @q locked.
232 void ioc_clear_queue(struct request_queue *q)
234 lockdep_assert_held(q->queue_lock);
236 while (!list_empty(&q->icq_list)) {
237 struct io_cq *icq = list_entry(q->icq_list.next,
238 struct io_cq, q_node);
239 struct io_context *ioc = icq->ioc;
241 spin_lock(&ioc->lock);
242 ioc_destroy_icq(icq);
243 spin_unlock(&ioc->lock);
247 int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node)
249 struct io_context *ioc;
252 ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
258 atomic_long_set(&ioc->refcount, 1);
259 atomic_set(&ioc->nr_tasks, 1);
260 atomic_set(&ioc->active_ref, 1);
261 spin_lock_init(&ioc->lock);
262 INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
263 INIT_HLIST_HEAD(&ioc->icq_list);
264 INIT_WORK(&ioc->release_work, ioc_release_fn);
267 * Try to install. ioc shouldn't be installed if someone else
268 * already did or @task, which isn't %current, is exiting. Note
269 * that we need to allow ioc creation on exiting %current as exit
270 * path may issue IOs from e.g. exit_files(). The exit path is
271 * responsible for not issuing IO after exit_io_context().
274 if (!task->io_context &&
275 (task == current || !(task->flags & PF_EXITING)))
276 task->io_context = ioc;
278 kmem_cache_free(iocontext_cachep, ioc);
280 ret = task->io_context ? 0 : -EBUSY;
288 * get_task_io_context - get io_context of a task
289 * @task: task of interest
290 * @gfp_flags: allocation flags, used if allocation is necessary
291 * @node: allocation node, used if allocation is necessary
293 * Return io_context of @task. If it doesn't exist, it is created with
294 * @gfp_flags and @node. The returned io_context has its reference count
297 * This function always goes through task_lock() and it's better to use
298 * %current->io_context + get_io_context() for %current.
300 struct io_context *get_task_io_context(struct task_struct *task,
301 gfp_t gfp_flags, int node)
303 struct io_context *ioc;
305 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
309 ioc = task->io_context;
316 } while (!create_task_io_context(task, gfp_flags, node));
320 EXPORT_SYMBOL(get_task_io_context);
323 * ioc_lookup_icq - lookup io_cq from ioc
324 * @ioc: the associated io_context
325 * @q: the associated request_queue
327 * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called
328 * with @q->queue_lock held.
330 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
334 lockdep_assert_held(q->queue_lock);
337 * icq's are indexed from @ioc using radix tree and hint pointer,
338 * both of which are protected with RCU. All removals are done
339 * holding both q and ioc locks, and we're holding q lock - if we
340 * find a icq which points to us, it's guaranteed to be valid.
343 icq = rcu_dereference(ioc->icq_hint);
344 if (icq && icq->q == q)
347 icq = radix_tree_lookup(&ioc->icq_tree, q->id);
348 if (icq && icq->q == q)
349 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
356 EXPORT_SYMBOL(ioc_lookup_icq);
359 * ioc_create_icq - create and link io_cq
360 * @ioc: io_context of interest
361 * @q: request_queue of interest
362 * @gfp_mask: allocation mask
364 * Make sure io_cq linking @ioc and @q exists. If icq doesn't exist, they
365 * will be created using @gfp_mask.
367 * The caller is responsible for ensuring @ioc won't go away and @q is
368 * alive and will stay alive until this function returns.
370 struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q,
373 struct elevator_type *et = q->elevator->type;
377 icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
382 if (radix_tree_maybe_preload(gfp_mask) < 0) {
383 kmem_cache_free(et->icq_cache, icq);
389 INIT_LIST_HEAD(&icq->q_node);
390 INIT_HLIST_NODE(&icq->ioc_node);
392 /* lock both q and ioc and try to link @icq */
393 spin_lock_irq(q->queue_lock);
394 spin_lock(&ioc->lock);
396 if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
397 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
398 list_add(&icq->q_node, &q->icq_list);
399 if (et->uses_mq && et->ops.mq.init_icq)
400 et->ops.mq.init_icq(icq);
401 else if (!et->uses_mq && et->ops.sq.elevator_init_icq_fn)
402 et->ops.sq.elevator_init_icq_fn(icq);
404 kmem_cache_free(et->icq_cache, icq);
405 icq = ioc_lookup_icq(ioc, q);
407 printk(KERN_ERR "cfq: icq link failed!\n");
410 spin_unlock(&ioc->lock);
411 spin_unlock_irq(q->queue_lock);
412 radix_tree_preload_end();
416 static int __init blk_ioc_init(void)
418 iocontext_cachep = kmem_cache_create("blkdev_ioc",
419 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
422 subsys_initcall(blk_ioc_init);