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>
14 * For io context allocations
16 static struct kmem_cache *iocontext_cachep;
19 * get_io_context - increment reference count to io_context
20 * @ioc: io_context to get
22 * Increment reference count to @ioc.
24 void get_io_context(struct io_context *ioc)
26 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
27 atomic_long_inc(&ioc->refcount);
29 EXPORT_SYMBOL(get_io_context);
31 static void icq_free_icq_rcu(struct rcu_head *head)
33 struct io_cq *icq = container_of(head, struct io_cq, __rcu_head);
35 kmem_cache_free(icq->__rcu_icq_cache, icq);
38 /* Exit an icq. Called with both ioc and q locked. */
39 static void ioc_exit_icq(struct io_cq *icq)
41 struct elevator_type *et = icq->q->elevator->type;
43 if (icq->flags & ICQ_EXITED)
46 if (et->ops.elevator_exit_icq_fn)
47 et->ops.elevator_exit_icq_fn(icq);
49 icq->flags |= ICQ_EXITED;
52 /* Release an icq. Called with both ioc and q locked. */
53 static void ioc_destroy_icq(struct io_cq *icq)
55 struct io_context *ioc = icq->ioc;
56 struct request_queue *q = icq->q;
57 struct elevator_type *et = q->elevator->type;
59 lockdep_assert_held(&ioc->lock);
60 lockdep_assert_held(q->queue_lock);
62 radix_tree_delete(&ioc->icq_tree, icq->q->id);
63 hlist_del_init(&icq->ioc_node);
64 list_del_init(&icq->q_node);
67 * Both setting lookup hint to and clearing it from @icq are done
68 * under queue_lock. If it's not pointing to @icq now, it never
69 * will. Hint assignment itself can race safely.
71 if (rcu_access_pointer(ioc->icq_hint) == icq)
72 rcu_assign_pointer(ioc->icq_hint, NULL);
77 * @icq->q might have gone away by the time RCU callback runs
78 * making it impossible to determine icq_cache. Record it in @icq.
80 icq->__rcu_icq_cache = et->icq_cache;
81 call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
85 * Slow path for ioc release in put_io_context(). Performs double-lock
86 * dancing to unlink all icq's and then frees ioc.
88 static void ioc_release_fn(struct work_struct *work)
90 struct io_context *ioc = container_of(work, struct io_context,
95 * Exiting icq may call into put_io_context() through elevator
96 * which will trigger lockdep warning. The ioc's are guaranteed to
97 * be different, use a different locking subclass here. Use
98 * irqsave variant as there's no spin_lock_irq_nested().
100 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
102 while (!hlist_empty(&ioc->icq_list)) {
103 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
104 struct io_cq, ioc_node);
105 struct request_queue *q = icq->q;
107 if (spin_trylock(q->queue_lock)) {
108 ioc_destroy_icq(icq);
109 spin_unlock(q->queue_lock);
111 spin_unlock_irqrestore(&ioc->lock, flags);
113 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
117 spin_unlock_irqrestore(&ioc->lock, flags);
119 kmem_cache_free(iocontext_cachep, ioc);
123 * put_io_context - put a reference of io_context
124 * @ioc: io_context to put
126 * Decrement reference count of @ioc and release it if the count reaches
129 void put_io_context(struct io_context *ioc)
132 bool free_ioc = false;
137 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
140 * Releasing ioc requires reverse order double locking and we may
141 * already be holding a queue_lock. Do it asynchronously from wq.
143 if (atomic_long_dec_and_test(&ioc->refcount)) {
144 spin_lock_irqsave(&ioc->lock, flags);
145 if (!hlist_empty(&ioc->icq_list))
146 queue_work(system_power_efficient_wq,
150 spin_unlock_irqrestore(&ioc->lock, flags);
154 kmem_cache_free(iocontext_cachep, ioc);
156 EXPORT_SYMBOL(put_io_context);
159 * put_io_context_active - put active reference on ioc
160 * @ioc: ioc of interest
162 * Undo get_io_context_active(). If active reference reaches zero after
163 * put, @ioc can never issue further IOs and ioscheds are notified.
165 void put_io_context_active(struct io_context *ioc)
170 if (!atomic_dec_and_test(&ioc->active_ref)) {
176 * Need ioc lock to walk icq_list and q lock to exit icq. Perform
177 * reverse double locking. Read comment in ioc_release_fn() for
178 * explanation on the nested locking annotation.
181 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
182 hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) {
183 if (icq->flags & ICQ_EXITED)
185 if (spin_trylock(icq->q->queue_lock)) {
187 spin_unlock(icq->q->queue_lock);
189 spin_unlock_irqrestore(&ioc->lock, flags);
194 spin_unlock_irqrestore(&ioc->lock, flags);
199 /* Called by the exiting task */
200 void exit_io_context(struct task_struct *task)
202 struct io_context *ioc;
205 ioc = task->io_context;
206 task->io_context = NULL;
209 atomic_dec(&ioc->nr_tasks);
210 put_io_context_active(ioc);
214 * ioc_clear_queue - break any ioc association with the specified queue
215 * @q: request_queue being cleared
217 * Walk @q->icq_list and exit all io_cq's. Must be called with @q locked.
219 void ioc_clear_queue(struct request_queue *q)
221 lockdep_assert_held(q->queue_lock);
223 while (!list_empty(&q->icq_list)) {
224 struct io_cq *icq = list_entry(q->icq_list.next,
225 struct io_cq, q_node);
226 struct io_context *ioc = icq->ioc;
228 spin_lock(&ioc->lock);
229 ioc_destroy_icq(icq);
230 spin_unlock(&ioc->lock);
234 int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node)
236 struct io_context *ioc;
239 ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
245 atomic_long_set(&ioc->refcount, 1);
246 atomic_set(&ioc->nr_tasks, 1);
247 atomic_set(&ioc->active_ref, 1);
248 spin_lock_init(&ioc->lock);
249 INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
250 INIT_HLIST_HEAD(&ioc->icq_list);
251 INIT_WORK(&ioc->release_work, ioc_release_fn);
254 * Try to install. ioc shouldn't be installed if someone else
255 * already did or @task, which isn't %current, is exiting. Note
256 * that we need to allow ioc creation on exiting %current as exit
257 * path may issue IOs from e.g. exit_files(). The exit path is
258 * responsible for not issuing IO after exit_io_context().
261 if (!task->io_context &&
262 (task == current || !(task->flags & PF_EXITING)))
263 task->io_context = ioc;
265 kmem_cache_free(iocontext_cachep, ioc);
267 ret = task->io_context ? 0 : -EBUSY;
275 * get_task_io_context - get io_context of a task
276 * @task: task of interest
277 * @gfp_flags: allocation flags, used if allocation is necessary
278 * @node: allocation node, used if allocation is necessary
280 * Return io_context of @task. If it doesn't exist, it is created with
281 * @gfp_flags and @node. The returned io_context has its reference count
284 * This function always goes through task_lock() and it's better to use
285 * %current->io_context + get_io_context() for %current.
287 struct io_context *get_task_io_context(struct task_struct *task,
288 gfp_t gfp_flags, int node)
290 struct io_context *ioc;
292 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
296 ioc = task->io_context;
303 } while (!create_task_io_context(task, gfp_flags, node));
307 EXPORT_SYMBOL(get_task_io_context);
310 * ioc_lookup_icq - lookup io_cq from ioc
311 * @ioc: the associated io_context
312 * @q: the associated request_queue
314 * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called
315 * with @q->queue_lock held.
317 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
321 lockdep_assert_held(q->queue_lock);
324 * icq's are indexed from @ioc using radix tree and hint pointer,
325 * both of which are protected with RCU. All removals are done
326 * holding both q and ioc locks, and we're holding q lock - if we
327 * find a icq which points to us, it's guaranteed to be valid.
330 icq = rcu_dereference(ioc->icq_hint);
331 if (icq && icq->q == q)
334 icq = radix_tree_lookup(&ioc->icq_tree, q->id);
335 if (icq && icq->q == q)
336 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
343 EXPORT_SYMBOL(ioc_lookup_icq);
346 * ioc_create_icq - create and link io_cq
347 * @ioc: io_context of interest
348 * @q: request_queue of interest
349 * @gfp_mask: allocation mask
351 * Make sure io_cq linking @ioc and @q exists. If icq doesn't exist, they
352 * will be created using @gfp_mask.
354 * The caller is responsible for ensuring @ioc won't go away and @q is
355 * alive and will stay alive until this function returns.
357 struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q,
360 struct elevator_type *et = q->elevator->type;
364 icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
369 if (radix_tree_maybe_preload(gfp_mask) < 0) {
370 kmem_cache_free(et->icq_cache, icq);
376 INIT_LIST_HEAD(&icq->q_node);
377 INIT_HLIST_NODE(&icq->ioc_node);
379 /* lock both q and ioc and try to link @icq */
380 spin_lock_irq(q->queue_lock);
381 spin_lock(&ioc->lock);
383 if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
384 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
385 list_add(&icq->q_node, &q->icq_list);
386 if (et->ops.elevator_init_icq_fn)
387 et->ops.elevator_init_icq_fn(icq);
389 kmem_cache_free(et->icq_cache, icq);
390 icq = ioc_lookup_icq(ioc, q);
392 printk(KERN_ERR "cfq: icq link failed!\n");
395 spin_unlock(&ioc->lock);
396 spin_unlock_irq(q->queue_lock);
397 radix_tree_preload_end();
401 static int __init blk_ioc_init(void)
403 iocontext_cachep = kmem_cache_create("blkdev_ioc",
404 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
407 subsys_initcall(blk_ioc_init);