blob: 8b782a63c29705661dc3ced3a13ead1dd4b42b22 [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
39/*
40 * Exit and free an icq. Called with both ioc and q locked.
41 */
42static void ioc_exit_icq(struct io_cq *icq)
43{
44 struct io_context *ioc = icq->ioc;
45 struct request_queue *q = icq->q;
46 struct elevator_type *et = q->elevator->type;
47
48 lockdep_assert_held(&ioc->lock);
49 lockdep_assert_held(q->queue_lock);
50
51 radix_tree_delete(&ioc->icq_tree, icq->q->id);
52 hlist_del_init(&icq->ioc_node);
53 list_del_init(&icq->q_node);
54
55 /*
56 * Both setting lookup hint to and clearing it from @icq are done
57 * under queue_lock. If it's not pointing to @icq now, it never
58 * will. Hint assignment itself can race safely.
59 */
60 if (rcu_dereference_raw(ioc->icq_hint) == icq)
61 rcu_assign_pointer(ioc->icq_hint, NULL);
62
Tejun Heo11a31222012-02-07 07:51:30 +010063 if (et->ops.elevator_exit_icq_fn)
Tejun Heo7e5a8792011-12-14 00:33:42 +010064 et->ops.elevator_exit_icq_fn(icq);
Tejun Heo7e5a8792011-12-14 00:33:42 +010065
66 /*
67 * @icq->q might have gone away by the time RCU callback runs
68 * making it impossible to determine icq_cache. Record it in @icq.
69 */
70 icq->__rcu_icq_cache = et->icq_cache;
71 call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
72}
73
Tejun Heob2efa052011-12-14 00:33:39 +010074/*
75 * Slow path for ioc release in put_io_context(). Performs double-lock
Tejun Heoc5869802011-12-14 00:33:41 +010076 * dancing to unlink all icq's and then frees ioc.
Tejun Heob2efa052011-12-14 00:33:39 +010077 */
78static void ioc_release_fn(struct work_struct *work)
79{
80 struct io_context *ioc = container_of(work, struct io_context,
81 release_work);
82 struct request_queue *last_q = NULL;
Tejun Heod8c66c52012-02-11 12:37:25 +010083 unsigned long flags;
Tejun Heob2efa052011-12-14 00:33:39 +010084
Tejun Heod8c66c52012-02-11 12:37:25 +010085 /*
86 * Exiting icq may call into put_io_context() through elevator
87 * which will trigger lockdep warning. The ioc's are guaranteed to
88 * be different, use a different locking subclass here. Use
89 * irqsave variant as there's no spin_lock_irq_nested().
90 */
91 spin_lock_irqsave_nested(&ioc->lock, flags, 1);
Tejun Heob2efa052011-12-14 00:33:39 +010092
Tejun Heoc5869802011-12-14 00:33:41 +010093 while (!hlist_empty(&ioc->icq_list)) {
94 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
95 struct io_cq, ioc_node);
96 struct request_queue *this_q = icq->q;
Tejun Heob2efa052011-12-14 00:33:39 +010097
98 if (this_q != last_q) {
99 /*
100 * Need to switch to @this_q. Once we release
101 * @ioc->lock, it can go away along with @cic.
102 * Hold on to it.
103 */
104 __blk_get_queue(this_q);
105
106 /*
107 * blk_put_queue() might sleep thanks to kobject
108 * idiocy. Always release both locks, put and
109 * restart.
110 */
111 if (last_q) {
112 spin_unlock(last_q->queue_lock);
Tejun Heod8c66c52012-02-11 12:37:25 +0100113 spin_unlock_irqrestore(&ioc->lock, flags);
Tejun Heob2efa052011-12-14 00:33:39 +0100114 blk_put_queue(last_q);
115 } else {
Tejun Heod8c66c52012-02-11 12:37:25 +0100116 spin_unlock_irqrestore(&ioc->lock, flags);
Tejun Heob2efa052011-12-14 00:33:39 +0100117 }
118
119 last_q = this_q;
Tejun Heod8c66c52012-02-11 12:37:25 +0100120 spin_lock_irqsave(this_q->queue_lock, flags);
121 spin_lock_nested(&ioc->lock, 1);
Tejun Heob2efa052011-12-14 00:33:39 +0100122 continue;
123 }
Tejun Heo7e5a8792011-12-14 00:33:42 +0100124 ioc_exit_icq(icq);
Jens Axboeffc4e752008-02-19 10:02:29 +0100125 }
Tejun Heob2efa052011-12-14 00:33:39 +0100126
127 if (last_q) {
128 spin_unlock(last_q->queue_lock);
Tejun Heod8c66c52012-02-11 12:37:25 +0100129 spin_unlock_irqrestore(&ioc->lock, flags);
Tejun Heob2efa052011-12-14 00:33:39 +0100130 blk_put_queue(last_q);
131 } else {
Tejun Heod8c66c52012-02-11 12:37:25 +0100132 spin_unlock_irqrestore(&ioc->lock, flags);
Tejun Heob2efa052011-12-14 00:33:39 +0100133 }
134
135 kmem_cache_free(iocontext_cachep, ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +0100136}
137
Tejun Heo42ec57a2011-12-14 00:33:37 +0100138/**
139 * put_io_context - put a reference of io_context
140 * @ioc: io_context to put
141 *
142 * Decrement reference count of @ioc and release it if the count reaches
Tejun Heo11a31222012-02-07 07:51:30 +0100143 * zero.
Jens Axboe86db1e22008-01-29 14:53:40 +0100144 */
Tejun Heo11a31222012-02-07 07:51:30 +0100145void put_io_context(struct io_context *ioc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100146{
Tejun Heob2efa052011-12-14 00:33:39 +0100147 unsigned long flags;
148
Jens Axboe86db1e22008-01-29 14:53:40 +0100149 if (ioc == NULL)
Tejun Heo42ec57a2011-12-14 00:33:37 +0100150 return;
Jens Axboe86db1e22008-01-29 14:53:40 +0100151
Tejun Heo42ec57a2011-12-14 00:33:37 +0100152 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
Jens Axboe86db1e22008-01-29 14:53:40 +0100153
Tejun Heob2efa052011-12-14 00:33:39 +0100154 /*
Tejun Heo11a31222012-02-07 07:51:30 +0100155 * Releasing ioc requires reverse order double locking and we may
156 * already be holding a queue_lock. Do it asynchronously from wq.
Tejun Heob2efa052011-12-14 00:33:39 +0100157 */
Tejun Heo11a31222012-02-07 07:51:30 +0100158 if (atomic_long_dec_and_test(&ioc->refcount)) {
159 spin_lock_irqsave(&ioc->lock, flags);
160 if (!hlist_empty(&ioc->icq_list))
161 schedule_work(&ioc->release_work);
162 spin_unlock_irqrestore(&ioc->lock, flags);
Tejun Heob2efa052011-12-14 00:33:39 +0100163 }
Jens Axboe86db1e22008-01-29 14:53:40 +0100164}
165EXPORT_SYMBOL(put_io_context);
166
Bart Van Assche27667c92010-12-21 15:07:45 +0100167/* Called by the exiting task */
Louis Rillingb69f2292009-12-04 14:52:42 +0100168void exit_io_context(struct task_struct *task)
Jens Axboe86db1e22008-01-29 14:53:40 +0100169{
170 struct io_context *ioc;
171
Louis Rillingb69f2292009-12-04 14:52:42 +0100172 task_lock(task);
173 ioc = task->io_context;
174 task->io_context = NULL;
175 task_unlock(task);
Jens Axboe86db1e22008-01-29 14:53:40 +0100176
Tejun Heob2efa052011-12-14 00:33:39 +0100177 atomic_dec(&ioc->nr_tasks);
Tejun Heo11a31222012-02-07 07:51:30 +0100178 put_io_context(ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +0100179}
180
Tejun Heo7e5a8792011-12-14 00:33:42 +0100181/**
182 * ioc_clear_queue - break any ioc association with the specified queue
183 * @q: request_queue being cleared
184 *
185 * Walk @q->icq_list and exit all io_cq's. Must be called with @q locked.
186 */
187void ioc_clear_queue(struct request_queue *q)
188{
189 lockdep_assert_held(q->queue_lock);
190
191 while (!list_empty(&q->icq_list)) {
192 struct io_cq *icq = list_entry(q->icq_list.next,
193 struct io_cq, q_node);
194 struct io_context *ioc = icq->ioc;
195
196 spin_lock(&ioc->lock);
197 ioc_exit_icq(icq);
198 spin_unlock(&ioc->lock);
199 }
200}
201
Tejun Heof2dbd762011-12-14 00:33:40 +0100202void create_io_context_slowpath(struct task_struct *task, gfp_t gfp_flags,
203 int node)
Jens Axboe86db1e22008-01-29 14:53:40 +0100204{
Paul Bolledf415652011-06-06 05:11:34 +0200205 struct io_context *ioc;
Jens Axboe86db1e22008-01-29 14:53:40 +0100206
Tejun Heo42ec57a2011-12-14 00:33:37 +0100207 ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
208 node);
209 if (unlikely(!ioc))
Tejun Heof2dbd762011-12-14 00:33:40 +0100210 return;
Tejun Heo42ec57a2011-12-14 00:33:37 +0100211
212 /* initialize */
213 atomic_long_set(&ioc->refcount, 1);
214 atomic_set(&ioc->nr_tasks, 1);
215 spin_lock_init(&ioc->lock);
Tejun Heoc5869802011-12-14 00:33:41 +0100216 INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
217 INIT_HLIST_HEAD(&ioc->icq_list);
Tejun Heob2efa052011-12-14 00:33:39 +0100218 INIT_WORK(&ioc->release_work, ioc_release_fn);
Jens Axboe86db1e22008-01-29 14:53:40 +0100219
Tejun Heofd638362011-12-25 14:29:14 +0100220 /*
221 * Try to install. ioc shouldn't be installed if someone else
222 * already did or @task, which isn't %current, is exiting. Note
223 * that we need to allow ioc creation on exiting %current as exit
224 * path may issue IOs from e.g. exit_files(). The exit path is
225 * responsible for not issuing IO after exit_io_context().
226 */
Tejun Heo6e736be2011-12-14 00:33:38 +0100227 task_lock(task);
Tejun Heofd638362011-12-25 14:29:14 +0100228 if (!task->io_context &&
229 (task == current || !(task->flags & PF_EXITING)))
Tejun Heo6e736be2011-12-14 00:33:38 +0100230 task->io_context = ioc;
Tejun Heof2dbd762011-12-14 00:33:40 +0100231 else
Tejun Heo6e736be2011-12-14 00:33:38 +0100232 kmem_cache_free(iocontext_cachep, ioc);
Tejun Heo6e736be2011-12-14 00:33:38 +0100233 task_unlock(task);
Jens Axboe86db1e22008-01-29 14:53:40 +0100234}
Jens Axboe86db1e22008-01-29 14:53:40 +0100235
Tejun Heo6e736be2011-12-14 00:33:38 +0100236/**
237 * get_task_io_context - get io_context of a task
238 * @task: task of interest
239 * @gfp_flags: allocation flags, used if allocation is necessary
240 * @node: allocation node, used if allocation is necessary
Jens Axboe86db1e22008-01-29 14:53:40 +0100241 *
Tejun Heo6e736be2011-12-14 00:33:38 +0100242 * Return io_context of @task. If it doesn't exist, it is created with
243 * @gfp_flags and @node. The returned io_context has its reference count
244 * incremented.
245 *
246 * This function always goes through task_lock() and it's better to use
Tejun Heof2dbd762011-12-14 00:33:40 +0100247 * %current->io_context + get_io_context() for %current.
Jens Axboe86db1e22008-01-29 14:53:40 +0100248 */
Tejun Heo6e736be2011-12-14 00:33:38 +0100249struct io_context *get_task_io_context(struct task_struct *task,
250 gfp_t gfp_flags, int node)
Jens Axboe86db1e22008-01-29 14:53:40 +0100251{
Tejun Heo6e736be2011-12-14 00:33:38 +0100252 struct io_context *ioc;
Jens Axboe86db1e22008-01-29 14:53:40 +0100253
Tejun Heo6e736be2011-12-14 00:33:38 +0100254 might_sleep_if(gfp_flags & __GFP_WAIT);
Jens Axboe86db1e22008-01-29 14:53:40 +0100255
Tejun Heof2dbd762011-12-14 00:33:40 +0100256 do {
257 task_lock(task);
258 ioc = task->io_context;
259 if (likely(ioc)) {
260 get_io_context(ioc);
261 task_unlock(task);
262 return ioc;
263 }
Tejun Heo6e736be2011-12-14 00:33:38 +0100264 task_unlock(task);
Tejun Heof2dbd762011-12-14 00:33:40 +0100265 } while (create_io_context(task, gfp_flags, node));
Tejun Heo6e736be2011-12-14 00:33:38 +0100266
Tejun Heof2dbd762011-12-14 00:33:40 +0100267 return NULL;
Jens Axboe86db1e22008-01-29 14:53:40 +0100268}
Tejun Heo6e736be2011-12-14 00:33:38 +0100269EXPORT_SYMBOL(get_task_io_context);
Jens Axboe86db1e22008-01-29 14:53:40 +0100270
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100271/**
272 * ioc_lookup_icq - lookup io_cq from ioc
273 * @ioc: the associated io_context
274 * @q: the associated request_queue
275 *
276 * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called
277 * with @q->queue_lock held.
278 */
279struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
280{
281 struct io_cq *icq;
282
283 lockdep_assert_held(q->queue_lock);
284
285 /*
286 * icq's are indexed from @ioc using radix tree and hint pointer,
287 * both of which are protected with RCU. All removals are done
288 * holding both q and ioc locks, and we're holding q lock - if we
289 * find a icq which points to us, it's guaranteed to be valid.
290 */
291 rcu_read_lock();
292 icq = rcu_dereference(ioc->icq_hint);
293 if (icq && icq->q == q)
294 goto out;
295
296 icq = radix_tree_lookup(&ioc->icq_tree, q->id);
297 if (icq && icq->q == q)
298 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
299 else
300 icq = NULL;
301out:
302 rcu_read_unlock();
303 return icq;
304}
305EXPORT_SYMBOL(ioc_lookup_icq);
306
Tejun Heof1f8cc92011-12-14 00:33:42 +0100307/**
308 * ioc_create_icq - create and link io_cq
309 * @q: request_queue of interest
310 * @gfp_mask: allocation mask
311 *
312 * Make sure io_cq linking %current->io_context and @q exists. If either
313 * io_context and/or icq don't exist, they will be created using @gfp_mask.
314 *
315 * The caller is responsible for ensuring @ioc won't go away and @q is
316 * alive and will stay alive until this function returns.
317 */
318struct io_cq *ioc_create_icq(struct request_queue *q, gfp_t gfp_mask)
319{
320 struct elevator_type *et = q->elevator->type;
321 struct io_context *ioc;
322 struct io_cq *icq;
323
324 /* allocate stuff */
325 ioc = create_io_context(current, gfp_mask, q->node);
326 if (!ioc)
327 return NULL;
328
329 icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
330 q->node);
331 if (!icq)
332 return NULL;
333
334 if (radix_tree_preload(gfp_mask) < 0) {
335 kmem_cache_free(et->icq_cache, icq);
336 return NULL;
337 }
338
339 icq->ioc = ioc;
340 icq->q = q;
341 INIT_LIST_HEAD(&icq->q_node);
342 INIT_HLIST_NODE(&icq->ioc_node);
343
344 /* lock both q and ioc and try to link @icq */
345 spin_lock_irq(q->queue_lock);
346 spin_lock(&ioc->lock);
347
348 if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
349 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
350 list_add(&icq->q_node, &q->icq_list);
351 if (et->ops.elevator_init_icq_fn)
352 et->ops.elevator_init_icq_fn(icq);
353 } else {
354 kmem_cache_free(et->icq_cache, icq);
355 icq = ioc_lookup_icq(ioc, q);
356 if (!icq)
357 printk(KERN_ERR "cfq: icq link failed!\n");
358 }
359
360 spin_unlock(&ioc->lock);
361 spin_unlock_irq(q->queue_lock);
362 radix_tree_preload_end();
363 return icq;
364}
365
Tejun Heodc869002011-12-14 00:33:38 +0100366void ioc_set_changed(struct io_context *ioc, int which)
367{
Tejun Heoc5869802011-12-14 00:33:41 +0100368 struct io_cq *icq;
Tejun Heodc869002011-12-14 00:33:38 +0100369 struct hlist_node *n;
370
Tejun Heoc5869802011-12-14 00:33:41 +0100371 hlist_for_each_entry(icq, n, &ioc->icq_list, ioc_node)
372 set_bit(which, &icq->changed);
Tejun Heodc869002011-12-14 00:33:38 +0100373}
374
375/**
376 * ioc_ioprio_changed - notify ioprio change
377 * @ioc: io_context of interest
378 * @ioprio: new ioprio
379 *
Tejun Heoc5869802011-12-14 00:33:41 +0100380 * @ioc's ioprio has changed to @ioprio. Set %ICQ_IOPRIO_CHANGED for all
381 * icq's. iosched is responsible for checking the bit and applying it on
Tejun Heodc869002011-12-14 00:33:38 +0100382 * request issue path.
383 */
384void ioc_ioprio_changed(struct io_context *ioc, int ioprio)
385{
386 unsigned long flags;
387
388 spin_lock_irqsave(&ioc->lock, flags);
389 ioc->ioprio = ioprio;
Tejun Heoc5869802011-12-14 00:33:41 +0100390 ioc_set_changed(ioc, ICQ_IOPRIO_CHANGED);
Tejun Heodc869002011-12-14 00:33:38 +0100391 spin_unlock_irqrestore(&ioc->lock, flags);
392}
393
394/**
395 * ioc_cgroup_changed - notify cgroup change
396 * @ioc: io_context of interest
397 *
Tejun Heoc5869802011-12-14 00:33:41 +0100398 * @ioc's cgroup has changed. Set %ICQ_CGROUP_CHANGED for all icq's.
Tejun Heodc869002011-12-14 00:33:38 +0100399 * iosched is responsible for checking the bit and applying it on request
400 * issue path.
401 */
402void ioc_cgroup_changed(struct io_context *ioc)
403{
404 unsigned long flags;
405
406 spin_lock_irqsave(&ioc->lock, flags);
Tejun Heoc5869802011-12-14 00:33:41 +0100407 ioc_set_changed(ioc, ICQ_CGROUP_CHANGED);
Tejun Heodc869002011-12-14 00:33:38 +0100408 spin_unlock_irqrestore(&ioc->lock, flags);
409}
Jens Axboe64c42992011-12-19 10:36:44 +0100410EXPORT_SYMBOL(ioc_cgroup_changed);
Tejun Heodc869002011-12-14 00:33:38 +0100411
Adrian Bunk13341592008-02-18 13:45:53 +0100412static int __init blk_ioc_init(void)
Jens Axboe86db1e22008-01-29 14:53:40 +0100413{
414 iocontext_cachep = kmem_cache_create("blkdev_ioc",
415 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
416 return 0;
417}
418subsys_initcall(blk_ioc_init);