blob: 381cb50a673c33ed86f6f5d597331137f2881ce3 [file] [log] [blame]
Jens Axboe86db1e22008-01-29 14:53:40 +01001/*
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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +09009#include <linux/slab.h>
Jens Axboe86db1e22008-01-29 14:53:40 +010010
11#include "blk.h"
12
13/*
14 * For io context allocations
15 */
16static struct kmem_cache *iocontext_cachep;
17
Tejun Heo6e736be2011-12-14 00:33:38 +010018/**
19 * get_io_context - increment reference count to io_context
20 * @ioc: io_context to get
21 *
22 * Increment reference count to @ioc.
23 */
24void get_io_context(struct io_context *ioc)
25{
26 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
27 atomic_long_inc(&ioc->refcount);
28}
29EXPORT_SYMBOL(get_io_context);
30
Tejun Heo7e5a8792011-12-14 00:33:42 +010031static void icq_free_icq_rcu(struct rcu_head *head)
32{
33 struct io_cq *icq = container_of(head, struct io_cq, __rcu_head);
34
35 kmem_cache_free(icq->__rcu_icq_cache, icq);
36}
37
Tejun Heo621032a2012-02-15 09:45:53 +010038/* Exit an icq. Called with both ioc and q locked. */
Tejun Heo7e5a8792011-12-14 00:33:42 +010039static void ioc_exit_icq(struct io_cq *icq)
40{
Tejun Heo621032a2012-02-15 09:45:53 +010041 struct elevator_type *et = icq->q->elevator->type;
42
43 if (icq->flags & ICQ_EXITED)
44 return;
45
46 if (et->ops.elevator_exit_icq_fn)
47 et->ops.elevator_exit_icq_fn(icq);
48
49 icq->flags |= ICQ_EXITED;
50}
51
52/* Release an icq. Called with both ioc and q locked. */
53static void ioc_destroy_icq(struct io_cq *icq)
54{
Tejun Heo7e5a8792011-12-14 00:33:42 +010055 struct io_context *ioc = icq->ioc;
56 struct request_queue *q = icq->q;
57 struct elevator_type *et = q->elevator->type;
58
59 lockdep_assert_held(&ioc->lock);
60 lockdep_assert_held(q->queue_lock);
61
62 radix_tree_delete(&ioc->icq_tree, icq->q->id);
63 hlist_del_init(&icq->ioc_node);
64 list_del_init(&icq->q_node);
65
66 /*
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.
70 */
Paul E. McKenneyec6c6762014-02-17 13:35:57 -080071 if (rcu_access_pointer(ioc->icq_hint) == icq)
Tejun Heo7e5a8792011-12-14 00:33:42 +010072 rcu_assign_pointer(ioc->icq_hint, NULL);
73
Tejun Heo621032a2012-02-15 09:45:53 +010074 ioc_exit_icq(icq);
Tejun Heo7e5a8792011-12-14 00:33:42 +010075
76 /*
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.
79 */
80 icq->__rcu_icq_cache = et->icq_cache;
81 call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
82}
83
Tejun Heob2efa052011-12-14 00:33:39 +010084/*
85 * Slow path for ioc release in put_io_context(). Performs double-lock
Tejun Heoc5869802011-12-14 00:33:41 +010086 * dancing to unlink all icq's and then frees ioc.
Tejun Heob2efa052011-12-14 00:33:39 +010087 */
88static void ioc_release_fn(struct work_struct *work)
89{
90 struct io_context *ioc = container_of(work, struct io_context,
91 release_work);
Tejun Heod8c66c52012-02-11 12:37:25 +010092 unsigned long flags;
Tejun Heob2efa052011-12-14 00:33:39 +010093
Tejun Heod8c66c52012-02-11 12:37:25 +010094 /*
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().
99 */
100 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
Tejun Heob2efa052011-12-14 00:33:39 +0100101
Tejun Heoc5869802011-12-14 00:33:41 +0100102 while (!hlist_empty(&ioc->icq_list)) {
103 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
104 struct io_cq, ioc_node);
Tejun Heo2274b022012-02-15 09:45:52 +0100105 struct request_queue *q = icq->q;
Tejun Heob2efa052011-12-14 00:33:39 +0100106
Tejun Heo2274b022012-02-15 09:45:52 +0100107 if (spin_trylock(q->queue_lock)) {
Tejun Heo621032a2012-02-15 09:45:53 +0100108 ioc_destroy_icq(icq);
Tejun Heo2274b022012-02-15 09:45:52 +0100109 spin_unlock(q->queue_lock);
110 } else {
111 spin_unlock_irqrestore(&ioc->lock, flags);
112 cpu_relax();
113 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
Tejun Heob2efa052011-12-14 00:33:39 +0100114 }
Jens Axboeffc4e752008-02-19 10:02:29 +0100115 }
Tejun Heob2efa052011-12-14 00:33:39 +0100116
Tejun Heo2274b022012-02-15 09:45:52 +0100117 spin_unlock_irqrestore(&ioc->lock, flags);
Tejun Heob2efa052011-12-14 00:33:39 +0100118
119 kmem_cache_free(iocontext_cachep, ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +0100120}
121
Tejun Heo42ec57a2011-12-14 00:33:37 +0100122/**
123 * put_io_context - put a reference of io_context
124 * @ioc: io_context to put
125 *
126 * Decrement reference count of @ioc and release it if the count reaches
Tejun Heo11a31222012-02-07 07:51:30 +0100127 * zero.
Jens Axboe86db1e22008-01-29 14:53:40 +0100128 */
Tejun Heo11a31222012-02-07 07:51:30 +0100129void put_io_context(struct io_context *ioc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100130{
Tejun Heob2efa052011-12-14 00:33:39 +0100131 unsigned long flags;
Xiaotian Fengff8c1472012-03-14 15:34:48 +0100132 bool free_ioc = false;
Tejun Heob2efa052011-12-14 00:33:39 +0100133
Jens Axboe86db1e22008-01-29 14:53:40 +0100134 if (ioc == NULL)
Tejun Heo42ec57a2011-12-14 00:33:37 +0100135 return;
Jens Axboe86db1e22008-01-29 14:53:40 +0100136
Tejun Heo42ec57a2011-12-14 00:33:37 +0100137 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
Jens Axboe86db1e22008-01-29 14:53:40 +0100138
Tejun Heob2efa052011-12-14 00:33:39 +0100139 /*
Tejun Heo11a31222012-02-07 07:51:30 +0100140 * Releasing ioc requires reverse order double locking and we may
141 * already be holding a queue_lock. Do it asynchronously from wq.
Tejun Heob2efa052011-12-14 00:33:39 +0100142 */
Tejun Heo11a31222012-02-07 07:51:30 +0100143 if (atomic_long_dec_and_test(&ioc->refcount)) {
144 spin_lock_irqsave(&ioc->lock, flags);
145 if (!hlist_empty(&ioc->icq_list))
Viresh Kumar695588f2013-04-24 17:12:56 +0530146 queue_work(system_power_efficient_wq,
147 &ioc->release_work);
Xiaotian Fengff8c1472012-03-14 15:34:48 +0100148 else
149 free_ioc = true;
Tejun Heo11a31222012-02-07 07:51:30 +0100150 spin_unlock_irqrestore(&ioc->lock, flags);
Tejun Heob2efa052011-12-14 00:33:39 +0100151 }
Xiaotian Fengff8c1472012-03-14 15:34:48 +0100152
153 if (free_ioc)
154 kmem_cache_free(iocontext_cachep, ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +0100155}
156EXPORT_SYMBOL(put_io_context);
157
Tejun Heof6e8d012012-03-05 13:15:26 -0800158/**
159 * put_io_context_active - put active reference on ioc
160 * @ioc: ioc of interest
161 *
162 * Undo get_io_context_active(). If active reference reaches zero after
163 * put, @ioc can never issue further IOs and ioscheds are notified.
164 */
165void put_io_context_active(struct io_context *ioc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100166{
Tejun Heo621032a2012-02-15 09:45:53 +0100167 unsigned long flags;
Tejun Heof6e8d012012-03-05 13:15:26 -0800168 struct io_cq *icq;
Jens Axboe86db1e22008-01-29 14:53:40 +0100169
Tejun Heof6e8d012012-03-05 13:15:26 -0800170 if (!atomic_dec_and_test(&ioc->active_ref)) {
Tejun Heo621032a2012-02-15 09:45:53 +0100171 put_io_context(ioc);
172 return;
173 }
174
175 /*
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.
179 */
180retry:
181 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800182 hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) {
Tejun Heo621032a2012-02-15 09:45:53 +0100183 if (icq->flags & ICQ_EXITED)
184 continue;
185 if (spin_trylock(icq->q->queue_lock)) {
186 ioc_exit_icq(icq);
187 spin_unlock(icq->q->queue_lock);
188 } else {
189 spin_unlock_irqrestore(&ioc->lock, flags);
190 cpu_relax();
191 goto retry;
192 }
193 }
194 spin_unlock_irqrestore(&ioc->lock, flags);
195
Tejun Heo11a31222012-02-07 07:51:30 +0100196 put_io_context(ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +0100197}
198
Tejun Heof6e8d012012-03-05 13:15:26 -0800199/* Called by the exiting task */
200void exit_io_context(struct task_struct *task)
201{
202 struct io_context *ioc;
203
204 task_lock(task);
205 ioc = task->io_context;
206 task->io_context = NULL;
207 task_unlock(task);
208
209 atomic_dec(&ioc->nr_tasks);
210 put_io_context_active(ioc);
211}
212
Tejun Heo7e5a8792011-12-14 00:33:42 +0100213/**
214 * ioc_clear_queue - break any ioc association with the specified queue
215 * @q: request_queue being cleared
216 *
217 * Walk @q->icq_list and exit all io_cq's. Must be called with @q locked.
218 */
219void ioc_clear_queue(struct request_queue *q)
220{
221 lockdep_assert_held(q->queue_lock);
222
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;
227
228 spin_lock(&ioc->lock);
Tejun Heo621032a2012-02-15 09:45:53 +0100229 ioc_destroy_icq(icq);
Tejun Heo7e5a8792011-12-14 00:33:42 +0100230 spin_unlock(&ioc->lock);
231 }
232}
233
Tejun Heo24acfc32012-03-05 13:15:24 -0800234int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node)
Jens Axboe86db1e22008-01-29 14:53:40 +0100235{
Paul Bolledf415652011-06-06 05:11:34 +0200236 struct io_context *ioc;
Eric Dumazet3c9c7082012-05-31 13:39:05 +0200237 int ret;
Jens Axboe86db1e22008-01-29 14:53:40 +0100238
Tejun Heo42ec57a2011-12-14 00:33:37 +0100239 ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
240 node);
241 if (unlikely(!ioc))
Tejun Heo24acfc32012-03-05 13:15:24 -0800242 return -ENOMEM;
Tejun Heo42ec57a2011-12-14 00:33:37 +0100243
244 /* initialize */
245 atomic_long_set(&ioc->refcount, 1);
Olof Johansson4638a832012-08-01 12:17:27 +0200246 atomic_set(&ioc->nr_tasks, 1);
Tejun Heof6e8d012012-03-05 13:15:26 -0800247 atomic_set(&ioc->active_ref, 1);
Tejun Heo42ec57a2011-12-14 00:33:37 +0100248 spin_lock_init(&ioc->lock);
Tejun Heoc5869802011-12-14 00:33:41 +0100249 INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
250 INIT_HLIST_HEAD(&ioc->icq_list);
Tejun Heob2efa052011-12-14 00:33:39 +0100251 INIT_WORK(&ioc->release_work, ioc_release_fn);
Jens Axboe86db1e22008-01-29 14:53:40 +0100252
Tejun Heofd638362011-12-25 14:29:14 +0100253 /*
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().
259 */
Tejun Heo6e736be2011-12-14 00:33:38 +0100260 task_lock(task);
Tejun Heofd638362011-12-25 14:29:14 +0100261 if (!task->io_context &&
262 (task == current || !(task->flags & PF_EXITING)))
Tejun Heo6e736be2011-12-14 00:33:38 +0100263 task->io_context = ioc;
Tejun Heof2dbd762011-12-14 00:33:40 +0100264 else
Tejun Heo6e736be2011-12-14 00:33:38 +0100265 kmem_cache_free(iocontext_cachep, ioc);
Eric Dumazet3c9c7082012-05-31 13:39:05 +0200266
267 ret = task->io_context ? 0 : -EBUSY;
268
Tejun Heo6e736be2011-12-14 00:33:38 +0100269 task_unlock(task);
Tejun Heo24acfc32012-03-05 13:15:24 -0800270
Eric Dumazet3c9c7082012-05-31 13:39:05 +0200271 return ret;
Jens Axboe86db1e22008-01-29 14:53:40 +0100272}
Jens Axboe86db1e22008-01-29 14:53:40 +0100273
Tejun Heo6e736be2011-12-14 00:33:38 +0100274/**
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
Jens Axboe86db1e22008-01-29 14:53:40 +0100279 *
Tejun Heo6e736be2011-12-14 00:33:38 +0100280 * 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
282 * incremented.
283 *
284 * This function always goes through task_lock() and it's better to use
Tejun Heof2dbd762011-12-14 00:33:40 +0100285 * %current->io_context + get_io_context() for %current.
Jens Axboe86db1e22008-01-29 14:53:40 +0100286 */
Tejun Heo6e736be2011-12-14 00:33:38 +0100287struct io_context *get_task_io_context(struct task_struct *task,
288 gfp_t gfp_flags, int node)
Jens Axboe86db1e22008-01-29 14:53:40 +0100289{
Tejun Heo6e736be2011-12-14 00:33:38 +0100290 struct io_context *ioc;
Jens Axboe86db1e22008-01-29 14:53:40 +0100291
Mel Gormand0164ad2015-11-06 16:28:21 -0800292 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
Jens Axboe86db1e22008-01-29 14:53:40 +0100293
Tejun Heof2dbd762011-12-14 00:33:40 +0100294 do {
295 task_lock(task);
296 ioc = task->io_context;
297 if (likely(ioc)) {
298 get_io_context(ioc);
299 task_unlock(task);
300 return ioc;
301 }
Tejun Heo6e736be2011-12-14 00:33:38 +0100302 task_unlock(task);
Tejun Heo24acfc32012-03-05 13:15:24 -0800303 } while (!create_task_io_context(task, gfp_flags, node));
Tejun Heo6e736be2011-12-14 00:33:38 +0100304
Tejun Heof2dbd762011-12-14 00:33:40 +0100305 return NULL;
Jens Axboe86db1e22008-01-29 14:53:40 +0100306}
Tejun Heo6e736be2011-12-14 00:33:38 +0100307EXPORT_SYMBOL(get_task_io_context);
Jens Axboe86db1e22008-01-29 14:53:40 +0100308
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100309/**
310 * ioc_lookup_icq - lookup io_cq from ioc
311 * @ioc: the associated io_context
312 * @q: the associated request_queue
313 *
314 * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called
315 * with @q->queue_lock held.
316 */
317struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
318{
319 struct io_cq *icq;
320
321 lockdep_assert_held(q->queue_lock);
322
323 /*
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.
328 */
329 rcu_read_lock();
330 icq = rcu_dereference(ioc->icq_hint);
331 if (icq && icq->q == q)
332 goto out;
333
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 */
337 else
338 icq = NULL;
339out:
340 rcu_read_unlock();
341 return icq;
342}
343EXPORT_SYMBOL(ioc_lookup_icq);
344
Tejun Heof1f8cc92011-12-14 00:33:42 +0100345/**
346 * ioc_create_icq - create and link io_cq
Tejun Heo24acfc32012-03-05 13:15:24 -0800347 * @ioc: io_context of interest
Tejun Heof1f8cc92011-12-14 00:33:42 +0100348 * @q: request_queue of interest
349 * @gfp_mask: allocation mask
350 *
Tejun Heo24acfc32012-03-05 13:15:24 -0800351 * Make sure io_cq linking @ioc and @q exists. If icq doesn't exist, they
352 * will be created using @gfp_mask.
Tejun Heof1f8cc92011-12-14 00:33:42 +0100353 *
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.
356 */
Tejun Heo24acfc32012-03-05 13:15:24 -0800357struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q,
358 gfp_t gfp_mask)
Tejun Heof1f8cc92011-12-14 00:33:42 +0100359{
360 struct elevator_type *et = q->elevator->type;
Tejun Heof1f8cc92011-12-14 00:33:42 +0100361 struct io_cq *icq;
362
363 /* allocate stuff */
Tejun Heof1f8cc92011-12-14 00:33:42 +0100364 icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
365 q->node);
366 if (!icq)
367 return NULL;
368
Jan Kara5e4c0d972013-09-11 14:26:05 -0700369 if (radix_tree_maybe_preload(gfp_mask) < 0) {
Tejun Heof1f8cc92011-12-14 00:33:42 +0100370 kmem_cache_free(et->icq_cache, icq);
371 return NULL;
372 }
373
374 icq->ioc = ioc;
375 icq->q = q;
376 INIT_LIST_HEAD(&icq->q_node);
377 INIT_HLIST_NODE(&icq->ioc_node);
378
379 /* lock both q and ioc and try to link @icq */
380 spin_lock_irq(q->queue_lock);
381 spin_lock(&ioc->lock);
382
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);
388 } else {
389 kmem_cache_free(et->icq_cache, icq);
390 icq = ioc_lookup_icq(ioc, q);
391 if (!icq)
392 printk(KERN_ERR "cfq: icq link failed!\n");
393 }
394
395 spin_unlock(&ioc->lock);
396 spin_unlock_irq(q->queue_lock);
397 radix_tree_preload_end();
398 return icq;
399}
400
Adrian Bunk13341592008-02-18 13:45:53 +0100401static int __init blk_ioc_init(void)
Jens Axboe86db1e22008-01-29 14:53:40 +0100402{
403 iocontext_cachep = kmem_cache_create("blkdev_ioc",
404 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
405 return 0;
406}
407subsys_initcall(blk_ioc_init);