blob: 4464c823cff2a0f3228a2dd5d54da3c9cdcbbde4 [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>
9#include <linux/bootmem.h> /* for max_pfn/max_low_pfn */
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090010#include <linux/slab.h>
Jens Axboe86db1e22008-01-29 14:53:40 +010011
12#include "blk.h"
13
14/*
15 * For io context allocations
16 */
17static struct kmem_cache *iocontext_cachep;
18
Tejun Heo6e736be2011-12-14 00:33:38 +010019/**
20 * get_io_context - increment reference count to io_context
21 * @ioc: io_context to get
22 *
23 * Increment reference count to @ioc.
24 */
25void get_io_context(struct io_context *ioc)
26{
27 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
28 atomic_long_inc(&ioc->refcount);
29}
30EXPORT_SYMBOL(get_io_context);
31
Tejun Heo7e5a8792011-12-14 00:33:42 +010032static 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
Tejun Heo621032a2012-02-15 09:45:53 +010039/* Exit an icq. Called with both ioc and q locked. */
Tejun Heo7e5a8792011-12-14 00:33:42 +010040static void ioc_exit_icq(struct io_cq *icq)
41{
Tejun Heo621032a2012-02-15 09:45:53 +010042 struct elevator_type *et = icq->q->elevator->type;
43
44 if (icq->flags & ICQ_EXITED)
45 return;
46
47 if (et->ops.elevator_exit_icq_fn)
48 et->ops.elevator_exit_icq_fn(icq);
49
50 icq->flags |= ICQ_EXITED;
51}
52
53/* Release an icq. Called with both ioc and q locked. */
54static void ioc_destroy_icq(struct io_cq *icq)
55{
Tejun Heo7e5a8792011-12-14 00:33:42 +010056 struct io_context *ioc = icq->ioc;
57 struct request_queue *q = icq->q;
58 struct elevator_type *et = q->elevator->type;
59
60 lockdep_assert_held(&ioc->lock);
61 lockdep_assert_held(q->queue_lock);
62
63 radix_tree_delete(&ioc->icq_tree, icq->q->id);
64 hlist_del_init(&icq->ioc_node);
65 list_del_init(&icq->q_node);
66
67 /*
68 * Both setting lookup hint to and clearing it from @icq are done
69 * under queue_lock. If it's not pointing to @icq now, it never
70 * will. Hint assignment itself can race safely.
71 */
72 if (rcu_dereference_raw(ioc->icq_hint) == icq)
73 rcu_assign_pointer(ioc->icq_hint, NULL);
74
Tejun Heo621032a2012-02-15 09:45:53 +010075 ioc_exit_icq(icq);
Tejun Heo7e5a8792011-12-14 00:33:42 +010076
77 /*
78 * @icq->q might have gone away by the time RCU callback runs
79 * making it impossible to determine icq_cache. Record it in @icq.
80 */
81 icq->__rcu_icq_cache = et->icq_cache;
82 call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
83}
84
Tejun Heob2efa052011-12-14 00:33:39 +010085/*
86 * Slow path for ioc release in put_io_context(). Performs double-lock
Tejun Heoc5869802011-12-14 00:33:41 +010087 * dancing to unlink all icq's and then frees ioc.
Tejun Heob2efa052011-12-14 00:33:39 +010088 */
89static void ioc_release_fn(struct work_struct *work)
90{
91 struct io_context *ioc = container_of(work, struct io_context,
92 release_work);
Tejun Heod8c66c52012-02-11 12:37:25 +010093 unsigned long flags;
Tejun Heob2efa052011-12-14 00:33:39 +010094
Tejun Heod8c66c52012-02-11 12:37:25 +010095 /*
96 * Exiting icq may call into put_io_context() through elevator
97 * which will trigger lockdep warning. The ioc's are guaranteed to
98 * be different, use a different locking subclass here. Use
99 * irqsave variant as there's no spin_lock_irq_nested().
100 */
101 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
Tejun Heob2efa052011-12-14 00:33:39 +0100102
Tejun Heoc5869802011-12-14 00:33:41 +0100103 while (!hlist_empty(&ioc->icq_list)) {
104 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
105 struct io_cq, ioc_node);
Tejun Heo2274b022012-02-15 09:45:52 +0100106 struct request_queue *q = icq->q;
Tejun Heob2efa052011-12-14 00:33:39 +0100107
Tejun Heo2274b022012-02-15 09:45:52 +0100108 if (spin_trylock(q->queue_lock)) {
Tejun Heo621032a2012-02-15 09:45:53 +0100109 ioc_destroy_icq(icq);
Tejun Heo2274b022012-02-15 09:45:52 +0100110 spin_unlock(q->queue_lock);
111 } else {
112 spin_unlock_irqrestore(&ioc->lock, flags);
113 cpu_relax();
114 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
Tejun Heob2efa052011-12-14 00:33:39 +0100115 }
Jens Axboeffc4e752008-02-19 10:02:29 +0100116 }
Tejun Heob2efa052011-12-14 00:33:39 +0100117
Tejun Heo2274b022012-02-15 09:45:52 +0100118 spin_unlock_irqrestore(&ioc->lock, flags);
Tejun Heob2efa052011-12-14 00:33:39 +0100119
120 kmem_cache_free(iocontext_cachep, ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +0100121}
122
Tejun Heo42ec57a2011-12-14 00:33:37 +0100123/**
124 * put_io_context - put a reference of io_context
125 * @ioc: io_context to put
126 *
127 * Decrement reference count of @ioc and release it if the count reaches
Tejun Heo11a31222012-02-07 07:51:30 +0100128 * zero.
Jens Axboe86db1e22008-01-29 14:53:40 +0100129 */
Tejun Heo11a31222012-02-07 07:51:30 +0100130void put_io_context(struct io_context *ioc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100131{
Tejun Heob2efa052011-12-14 00:33:39 +0100132 unsigned long flags;
Xiaotian Fengff8c1472012-03-14 15:34:48 +0100133 bool free_ioc = false;
Tejun Heob2efa052011-12-14 00:33:39 +0100134
Jens Axboe86db1e22008-01-29 14:53:40 +0100135 if (ioc == NULL)
Tejun Heo42ec57a2011-12-14 00:33:37 +0100136 return;
Jens Axboe86db1e22008-01-29 14:53:40 +0100137
Tejun Heo42ec57a2011-12-14 00:33:37 +0100138 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
Jens Axboe86db1e22008-01-29 14:53:40 +0100139
Tejun Heob2efa052011-12-14 00:33:39 +0100140 /*
Tejun Heo11a31222012-02-07 07:51:30 +0100141 * Releasing ioc requires reverse order double locking and we may
142 * already be holding a queue_lock. Do it asynchronously from wq.
Tejun Heob2efa052011-12-14 00:33:39 +0100143 */
Tejun Heo11a31222012-02-07 07:51:30 +0100144 if (atomic_long_dec_and_test(&ioc->refcount)) {
145 spin_lock_irqsave(&ioc->lock, flags);
146 if (!hlist_empty(&ioc->icq_list))
Viresh Kumar695588f2013-04-24 17:12:56 +0530147 queue_work(system_power_efficient_wq,
148 &ioc->release_work);
Xiaotian Fengff8c1472012-03-14 15:34:48 +0100149 else
150 free_ioc = true;
Tejun Heo11a31222012-02-07 07:51:30 +0100151 spin_unlock_irqrestore(&ioc->lock, flags);
Tejun Heob2efa052011-12-14 00:33:39 +0100152 }
Xiaotian Fengff8c1472012-03-14 15:34:48 +0100153
154 if (free_ioc)
155 kmem_cache_free(iocontext_cachep, ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +0100156}
157EXPORT_SYMBOL(put_io_context);
158
Tejun Heof6e8d012012-03-05 13:15:26 -0800159/**
160 * put_io_context_active - put active reference on ioc
161 * @ioc: ioc of interest
162 *
163 * Undo get_io_context_active(). If active reference reaches zero after
164 * put, @ioc can never issue further IOs and ioscheds are notified.
165 */
166void put_io_context_active(struct io_context *ioc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100167{
Tejun Heo621032a2012-02-15 09:45:53 +0100168 unsigned long flags;
Tejun Heof6e8d012012-03-05 13:15:26 -0800169 struct io_cq *icq;
Jens Axboe86db1e22008-01-29 14:53:40 +0100170
Tejun Heof6e8d012012-03-05 13:15:26 -0800171 if (!atomic_dec_and_test(&ioc->active_ref)) {
Tejun Heo621032a2012-02-15 09:45:53 +0100172 put_io_context(ioc);
173 return;
174 }
175
176 /*
177 * Need ioc lock to walk icq_list and q lock to exit icq. Perform
178 * reverse double locking. Read comment in ioc_release_fn() for
179 * explanation on the nested locking annotation.
180 */
181retry:
182 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800183 hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) {
Tejun Heo621032a2012-02-15 09:45:53 +0100184 if (icq->flags & ICQ_EXITED)
185 continue;
186 if (spin_trylock(icq->q->queue_lock)) {
187 ioc_exit_icq(icq);
188 spin_unlock(icq->q->queue_lock);
189 } else {
190 spin_unlock_irqrestore(&ioc->lock, flags);
191 cpu_relax();
192 goto retry;
193 }
194 }
195 spin_unlock_irqrestore(&ioc->lock, flags);
196
Tejun Heo11a31222012-02-07 07:51:30 +0100197 put_io_context(ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +0100198}
199
Tejun Heof6e8d012012-03-05 13:15:26 -0800200/* Called by the exiting task */
201void exit_io_context(struct task_struct *task)
202{
203 struct io_context *ioc;
204
205 task_lock(task);
206 ioc = task->io_context;
207 task->io_context = NULL;
208 task_unlock(task);
209
210 atomic_dec(&ioc->nr_tasks);
211 put_io_context_active(ioc);
212}
213
Tejun Heo7e5a8792011-12-14 00:33:42 +0100214/**
215 * ioc_clear_queue - break any ioc association with the specified queue
216 * @q: request_queue being cleared
217 *
218 * Walk @q->icq_list and exit all io_cq's. Must be called with @q locked.
219 */
220void ioc_clear_queue(struct request_queue *q)
221{
222 lockdep_assert_held(q->queue_lock);
223
224 while (!list_empty(&q->icq_list)) {
225 struct io_cq *icq = list_entry(q->icq_list.next,
226 struct io_cq, q_node);
227 struct io_context *ioc = icq->ioc;
228
229 spin_lock(&ioc->lock);
Tejun Heo621032a2012-02-15 09:45:53 +0100230 ioc_destroy_icq(icq);
Tejun Heo7e5a8792011-12-14 00:33:42 +0100231 spin_unlock(&ioc->lock);
232 }
233}
234
Tejun Heo24acfc32012-03-05 13:15:24 -0800235int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node)
Jens Axboe86db1e22008-01-29 14:53:40 +0100236{
Paul Bolledf415652011-06-06 05:11:34 +0200237 struct io_context *ioc;
Eric Dumazet3c9c7082012-05-31 13:39:05 +0200238 int ret;
Jens Axboe86db1e22008-01-29 14:53:40 +0100239
Tejun Heo42ec57a2011-12-14 00:33:37 +0100240 ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
241 node);
242 if (unlikely(!ioc))
Tejun Heo24acfc32012-03-05 13:15:24 -0800243 return -ENOMEM;
Tejun Heo42ec57a2011-12-14 00:33:37 +0100244
245 /* initialize */
246 atomic_long_set(&ioc->refcount, 1);
Olof Johansson4638a832012-08-01 12:17:27 +0200247 atomic_set(&ioc->nr_tasks, 1);
Tejun Heof6e8d012012-03-05 13:15:26 -0800248 atomic_set(&ioc->active_ref, 1);
Tejun Heo42ec57a2011-12-14 00:33:37 +0100249 spin_lock_init(&ioc->lock);
Tejun Heoc5869802011-12-14 00:33:41 +0100250 INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
251 INIT_HLIST_HEAD(&ioc->icq_list);
Tejun Heob2efa052011-12-14 00:33:39 +0100252 INIT_WORK(&ioc->release_work, ioc_release_fn);
Jens Axboe86db1e22008-01-29 14:53:40 +0100253
Tejun Heofd638362011-12-25 14:29:14 +0100254 /*
255 * Try to install. ioc shouldn't be installed if someone else
256 * already did or @task, which isn't %current, is exiting. Note
257 * that we need to allow ioc creation on exiting %current as exit
258 * path may issue IOs from e.g. exit_files(). The exit path is
259 * responsible for not issuing IO after exit_io_context().
260 */
Tejun Heo6e736be2011-12-14 00:33:38 +0100261 task_lock(task);
Tejun Heofd638362011-12-25 14:29:14 +0100262 if (!task->io_context &&
263 (task == current || !(task->flags & PF_EXITING)))
Tejun Heo6e736be2011-12-14 00:33:38 +0100264 task->io_context = ioc;
Tejun Heof2dbd762011-12-14 00:33:40 +0100265 else
Tejun Heo6e736be2011-12-14 00:33:38 +0100266 kmem_cache_free(iocontext_cachep, ioc);
Eric Dumazet3c9c7082012-05-31 13:39:05 +0200267
268 ret = task->io_context ? 0 : -EBUSY;
269
Tejun Heo6e736be2011-12-14 00:33:38 +0100270 task_unlock(task);
Tejun Heo24acfc32012-03-05 13:15:24 -0800271
Eric Dumazet3c9c7082012-05-31 13:39:05 +0200272 return ret;
Jens Axboe86db1e22008-01-29 14:53:40 +0100273}
Jens Axboe86db1e22008-01-29 14:53:40 +0100274
Tejun Heo6e736be2011-12-14 00:33:38 +0100275/**
276 * get_task_io_context - get io_context of a task
277 * @task: task of interest
278 * @gfp_flags: allocation flags, used if allocation is necessary
279 * @node: allocation node, used if allocation is necessary
Jens Axboe86db1e22008-01-29 14:53:40 +0100280 *
Tejun Heo6e736be2011-12-14 00:33:38 +0100281 * Return io_context of @task. If it doesn't exist, it is created with
282 * @gfp_flags and @node. The returned io_context has its reference count
283 * incremented.
284 *
285 * This function always goes through task_lock() and it's better to use
Tejun Heof2dbd762011-12-14 00:33:40 +0100286 * %current->io_context + get_io_context() for %current.
Jens Axboe86db1e22008-01-29 14:53:40 +0100287 */
Tejun Heo6e736be2011-12-14 00:33:38 +0100288struct io_context *get_task_io_context(struct task_struct *task,
289 gfp_t gfp_flags, int node)
Jens Axboe86db1e22008-01-29 14:53:40 +0100290{
Tejun Heo6e736be2011-12-14 00:33:38 +0100291 struct io_context *ioc;
Jens Axboe86db1e22008-01-29 14:53:40 +0100292
Tejun Heo6e736be2011-12-14 00:33:38 +0100293 might_sleep_if(gfp_flags & __GFP_WAIT);
Jens Axboe86db1e22008-01-29 14:53:40 +0100294
Tejun Heof2dbd762011-12-14 00:33:40 +0100295 do {
296 task_lock(task);
297 ioc = task->io_context;
298 if (likely(ioc)) {
299 get_io_context(ioc);
300 task_unlock(task);
301 return ioc;
302 }
Tejun Heo6e736be2011-12-14 00:33:38 +0100303 task_unlock(task);
Tejun Heo24acfc32012-03-05 13:15:24 -0800304 } while (!create_task_io_context(task, gfp_flags, node));
Tejun Heo6e736be2011-12-14 00:33:38 +0100305
Tejun Heof2dbd762011-12-14 00:33:40 +0100306 return NULL;
Jens Axboe86db1e22008-01-29 14:53:40 +0100307}
Tejun Heo6e736be2011-12-14 00:33:38 +0100308EXPORT_SYMBOL(get_task_io_context);
Jens Axboe86db1e22008-01-29 14:53:40 +0100309
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100310/**
311 * ioc_lookup_icq - lookup io_cq from ioc
312 * @ioc: the associated io_context
313 * @q: the associated request_queue
314 *
315 * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called
316 * with @q->queue_lock held.
317 */
318struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
319{
320 struct io_cq *icq;
321
322 lockdep_assert_held(q->queue_lock);
323
324 /*
325 * icq's are indexed from @ioc using radix tree and hint pointer,
326 * both of which are protected with RCU. All removals are done
327 * holding both q and ioc locks, and we're holding q lock - if we
328 * find a icq which points to us, it's guaranteed to be valid.
329 */
330 rcu_read_lock();
331 icq = rcu_dereference(ioc->icq_hint);
332 if (icq && icq->q == q)
333 goto out;
334
335 icq = radix_tree_lookup(&ioc->icq_tree, q->id);
336 if (icq && icq->q == q)
337 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
338 else
339 icq = NULL;
340out:
341 rcu_read_unlock();
342 return icq;
343}
344EXPORT_SYMBOL(ioc_lookup_icq);
345
Tejun Heof1f8cc92011-12-14 00:33:42 +0100346/**
347 * ioc_create_icq - create and link io_cq
Tejun Heo24acfc32012-03-05 13:15:24 -0800348 * @ioc: io_context of interest
Tejun Heof1f8cc92011-12-14 00:33:42 +0100349 * @q: request_queue of interest
350 * @gfp_mask: allocation mask
351 *
Tejun Heo24acfc32012-03-05 13:15:24 -0800352 * Make sure io_cq linking @ioc and @q exists. If icq doesn't exist, they
353 * will be created using @gfp_mask.
Tejun Heof1f8cc92011-12-14 00:33:42 +0100354 *
355 * The caller is responsible for ensuring @ioc won't go away and @q is
356 * alive and will stay alive until this function returns.
357 */
Tejun Heo24acfc32012-03-05 13:15:24 -0800358struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q,
359 gfp_t gfp_mask)
Tejun Heof1f8cc92011-12-14 00:33:42 +0100360{
361 struct elevator_type *et = q->elevator->type;
Tejun Heof1f8cc92011-12-14 00:33:42 +0100362 struct io_cq *icq;
363
364 /* allocate stuff */
Tejun Heof1f8cc92011-12-14 00:33:42 +0100365 icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
366 q->node);
367 if (!icq)
368 return NULL;
369
370 if (radix_tree_preload(gfp_mask) < 0) {
371 kmem_cache_free(et->icq_cache, icq);
372 return NULL;
373 }
374
375 icq->ioc = ioc;
376 icq->q = q;
377 INIT_LIST_HEAD(&icq->q_node);
378 INIT_HLIST_NODE(&icq->ioc_node);
379
380 /* lock both q and ioc and try to link @icq */
381 spin_lock_irq(q->queue_lock);
382 spin_lock(&ioc->lock);
383
384 if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
385 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
386 list_add(&icq->q_node, &q->icq_list);
387 if (et->ops.elevator_init_icq_fn)
388 et->ops.elevator_init_icq_fn(icq);
389 } else {
390 kmem_cache_free(et->icq_cache, icq);
391 icq = ioc_lookup_icq(ioc, q);
392 if (!icq)
393 printk(KERN_ERR "cfq: icq link failed!\n");
394 }
395
396 spin_unlock(&ioc->lock);
397 spin_unlock_irq(q->queue_lock);
398 radix_tree_preload_end();
399 return icq;
400}
401
Adrian Bunk13341592008-02-18 13:45:53 +0100402static int __init blk_ioc_init(void)
Jens Axboe86db1e22008-01-29 14:53:40 +0100403{
404 iocontext_cachep = kmem_cache_create("blkdev_ioc",
405 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
406 return 0;
407}
408subsys_initcall(blk_ioc_init);