blob: 27a06e00eaec4312c5596ad58ebaa7fcc8c438ef [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 Heob2efa052011-12-14 00:33:39 +010032/*
33 * Releasing ioc may nest into another put_io_context() leading to nested
34 * fast path release. As the ioc's can't be the same, this is okay but
35 * makes lockdep whine. Keep track of nesting and use it as subclass.
36 */
37#ifdef CONFIG_LOCKDEP
38#define ioc_release_depth(q) ((q) ? (q)->ioc_release_depth : 0)
39#define ioc_release_depth_inc(q) (q)->ioc_release_depth++
40#define ioc_release_depth_dec(q) (q)->ioc_release_depth--
41#else
42#define ioc_release_depth(q) 0
43#define ioc_release_depth_inc(q) do { } while (0)
44#define ioc_release_depth_dec(q) do { } while (0)
45#endif
Jens Axboe86db1e22008-01-29 14:53:40 +010046
Tejun Heo7e5a8792011-12-14 00:33:42 +010047static void icq_free_icq_rcu(struct rcu_head *head)
48{
49 struct io_cq *icq = container_of(head, struct io_cq, __rcu_head);
50
51 kmem_cache_free(icq->__rcu_icq_cache, icq);
52}
53
54/*
55 * Exit and free an icq. Called with both ioc and q locked.
56 */
57static void ioc_exit_icq(struct io_cq *icq)
58{
59 struct io_context *ioc = icq->ioc;
60 struct request_queue *q = icq->q;
61 struct elevator_type *et = q->elevator->type;
62
63 lockdep_assert_held(&ioc->lock);
64 lockdep_assert_held(q->queue_lock);
65
66 radix_tree_delete(&ioc->icq_tree, icq->q->id);
67 hlist_del_init(&icq->ioc_node);
68 list_del_init(&icq->q_node);
69
70 /*
71 * Both setting lookup hint to and clearing it from @icq are done
72 * under queue_lock. If it's not pointing to @icq now, it never
73 * will. Hint assignment itself can race safely.
74 */
75 if (rcu_dereference_raw(ioc->icq_hint) == icq)
76 rcu_assign_pointer(ioc->icq_hint, NULL);
77
78 if (et->ops.elevator_exit_icq_fn) {
79 ioc_release_depth_inc(q);
80 et->ops.elevator_exit_icq_fn(icq);
81 ioc_release_depth_dec(q);
82 }
83
84 /*
85 * @icq->q might have gone away by the time RCU callback runs
86 * making it impossible to determine icq_cache. Record it in @icq.
87 */
88 icq->__rcu_icq_cache = et->icq_cache;
89 call_rcu(&icq->__rcu_head, icq_free_icq_rcu);
90}
91
Tejun Heob2efa052011-12-14 00:33:39 +010092/*
93 * Slow path for ioc release in put_io_context(). Performs double-lock
Tejun Heoc5869802011-12-14 00:33:41 +010094 * dancing to unlink all icq's and then frees ioc.
Tejun Heob2efa052011-12-14 00:33:39 +010095 */
96static void ioc_release_fn(struct work_struct *work)
97{
98 struct io_context *ioc = container_of(work, struct io_context,
99 release_work);
100 struct request_queue *last_q = NULL;
101
102 spin_lock_irq(&ioc->lock);
103
Tejun Heoc5869802011-12-14 00:33:41 +0100104 while (!hlist_empty(&ioc->icq_list)) {
105 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
106 struct io_cq, ioc_node);
107 struct request_queue *this_q = icq->q;
Tejun Heob2efa052011-12-14 00:33:39 +0100108
109 if (this_q != last_q) {
110 /*
111 * Need to switch to @this_q. Once we release
112 * @ioc->lock, it can go away along with @cic.
113 * Hold on to it.
114 */
115 __blk_get_queue(this_q);
116
117 /*
118 * blk_put_queue() might sleep thanks to kobject
119 * idiocy. Always release both locks, put and
120 * restart.
121 */
122 if (last_q) {
123 spin_unlock(last_q->queue_lock);
124 spin_unlock_irq(&ioc->lock);
125 blk_put_queue(last_q);
126 } else {
127 spin_unlock_irq(&ioc->lock);
128 }
129
130 last_q = this_q;
131 spin_lock_irq(this_q->queue_lock);
132 spin_lock(&ioc->lock);
133 continue;
134 }
Tejun Heo7e5a8792011-12-14 00:33:42 +0100135 ioc_exit_icq(icq);
Jens Axboeffc4e752008-02-19 10:02:29 +0100136 }
Tejun Heob2efa052011-12-14 00:33:39 +0100137
138 if (last_q) {
139 spin_unlock(last_q->queue_lock);
140 spin_unlock_irq(&ioc->lock);
141 blk_put_queue(last_q);
142 } else {
143 spin_unlock_irq(&ioc->lock);
144 }
145
146 kmem_cache_free(iocontext_cachep, ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +0100147}
148
Tejun Heo42ec57a2011-12-14 00:33:37 +0100149/**
150 * put_io_context - put a reference of io_context
151 * @ioc: io_context to put
Tejun Heob2efa052011-12-14 00:33:39 +0100152 * @locked_q: request_queue the caller is holding queue_lock of (hint)
Tejun Heo42ec57a2011-12-14 00:33:37 +0100153 *
154 * Decrement reference count of @ioc and release it if the count reaches
Tejun Heob2efa052011-12-14 00:33:39 +0100155 * zero. If the caller is holding queue_lock of a queue, it can indicate
156 * that with @locked_q. This is an optimization hint and the caller is
157 * allowed to pass in %NULL even when it's holding a queue_lock.
Jens Axboe86db1e22008-01-29 14:53:40 +0100158 */
Tejun Heob2efa052011-12-14 00:33:39 +0100159void put_io_context(struct io_context *ioc, struct request_queue *locked_q)
Jens Axboe86db1e22008-01-29 14:53:40 +0100160{
Tejun Heob2efa052011-12-14 00:33:39 +0100161 struct request_queue *last_q = locked_q;
162 unsigned long flags;
163
Jens Axboe86db1e22008-01-29 14:53:40 +0100164 if (ioc == NULL)
Tejun Heo42ec57a2011-12-14 00:33:37 +0100165 return;
Jens Axboe86db1e22008-01-29 14:53:40 +0100166
Tejun Heo42ec57a2011-12-14 00:33:37 +0100167 BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
Tejun Heob2efa052011-12-14 00:33:39 +0100168 if (locked_q)
169 lockdep_assert_held(locked_q->queue_lock);
Jens Axboe86db1e22008-01-29 14:53:40 +0100170
Tejun Heo42ec57a2011-12-14 00:33:37 +0100171 if (!atomic_long_dec_and_test(&ioc->refcount))
172 return;
Jens Axboe86db1e22008-01-29 14:53:40 +0100173
Tejun Heob2efa052011-12-14 00:33:39 +0100174 /*
Tejun Heoc5869802011-12-14 00:33:41 +0100175 * Destroy @ioc. This is a bit messy because icq's are chained
Tejun Heob2efa052011-12-14 00:33:39 +0100176 * from both ioc and queue, and ioc->lock nests inside queue_lock.
Tejun Heoc5869802011-12-14 00:33:41 +0100177 * The inner ioc->lock should be held to walk our icq_list and then
178 * for each icq the outer matching queue_lock should be grabbed.
Tejun Heob2efa052011-12-14 00:33:39 +0100179 * ie. We need to do reverse-order double lock dancing.
180 *
181 * Another twist is that we are often called with one of the
182 * matching queue_locks held as indicated by @locked_q, which
183 * prevents performing double-lock dance for other queues.
184 *
185 * So, we do it in two stages. The fast path uses the queue_lock
186 * the caller is holding and, if other queues need to be accessed,
187 * uses trylock to avoid introducing locking dependency. This can
188 * handle most cases, especially if @ioc was performing IO on only
189 * single device.
190 *
191 * If trylock doesn't cut it, we defer to @ioc->release_work which
192 * can do all the double-locking dancing.
193 */
194 spin_lock_irqsave_nested(&ioc->lock, flags,
195 ioc_release_depth(locked_q));
Tejun Heo42ec57a2011-12-14 00:33:37 +0100196
Tejun Heoc5869802011-12-14 00:33:41 +0100197 while (!hlist_empty(&ioc->icq_list)) {
198 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
199 struct io_cq, ioc_node);
200 struct request_queue *this_q = icq->q;
Tejun Heob2efa052011-12-14 00:33:39 +0100201
202 if (this_q != last_q) {
203 if (last_q && last_q != locked_q)
204 spin_unlock(last_q->queue_lock);
205 last_q = NULL;
206
207 if (!spin_trylock(this_q->queue_lock))
208 break;
209 last_q = this_q;
210 continue;
211 }
Tejun Heo7e5a8792011-12-14 00:33:42 +0100212 ioc_exit_icq(icq);
Tejun Heob2efa052011-12-14 00:33:39 +0100213 }
214
215 if (last_q && last_q != locked_q)
216 spin_unlock(last_q->queue_lock);
217
218 spin_unlock_irqrestore(&ioc->lock, flags);
219
Tejun Heoc5869802011-12-14 00:33:41 +0100220 /* if no icq is left, we're done; otherwise, kick release_work */
221 if (hlist_empty(&ioc->icq_list))
Tejun Heob2efa052011-12-14 00:33:39 +0100222 kmem_cache_free(iocontext_cachep, ioc);
223 else
224 schedule_work(&ioc->release_work);
Jens Axboe86db1e22008-01-29 14:53:40 +0100225}
226EXPORT_SYMBOL(put_io_context);
227
Bart Van Assche27667c92010-12-21 15:07:45 +0100228/* Called by the exiting task */
Louis Rillingb69f2292009-12-04 14:52:42 +0100229void exit_io_context(struct task_struct *task)
Jens Axboe86db1e22008-01-29 14:53:40 +0100230{
231 struct io_context *ioc;
232
Louis Rillingb69f2292009-12-04 14:52:42 +0100233 task_lock(task);
234 ioc = task->io_context;
235 task->io_context = NULL;
236 task_unlock(task);
Jens Axboe86db1e22008-01-29 14:53:40 +0100237
Tejun Heob2efa052011-12-14 00:33:39 +0100238 atomic_dec(&ioc->nr_tasks);
239 put_io_context(ioc, NULL);
Jens Axboe86db1e22008-01-29 14:53:40 +0100240}
241
Tejun Heo7e5a8792011-12-14 00:33:42 +0100242/**
243 * ioc_clear_queue - break any ioc association with the specified queue
244 * @q: request_queue being cleared
245 *
246 * Walk @q->icq_list and exit all io_cq's. Must be called with @q locked.
247 */
248void ioc_clear_queue(struct request_queue *q)
249{
250 lockdep_assert_held(q->queue_lock);
251
252 while (!list_empty(&q->icq_list)) {
253 struct io_cq *icq = list_entry(q->icq_list.next,
254 struct io_cq, q_node);
255 struct io_context *ioc = icq->ioc;
256
257 spin_lock(&ioc->lock);
258 ioc_exit_icq(icq);
259 spin_unlock(&ioc->lock);
260 }
261}
262
Tejun Heof2dbd762011-12-14 00:33:40 +0100263void create_io_context_slowpath(struct task_struct *task, gfp_t gfp_flags,
264 int node)
Jens Axboe86db1e22008-01-29 14:53:40 +0100265{
Paul Bolledf415652011-06-06 05:11:34 +0200266 struct io_context *ioc;
Jens Axboe86db1e22008-01-29 14:53:40 +0100267
Tejun Heo42ec57a2011-12-14 00:33:37 +0100268 ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
269 node);
270 if (unlikely(!ioc))
Tejun Heof2dbd762011-12-14 00:33:40 +0100271 return;
Tejun Heo42ec57a2011-12-14 00:33:37 +0100272
273 /* initialize */
274 atomic_long_set(&ioc->refcount, 1);
275 atomic_set(&ioc->nr_tasks, 1);
276 spin_lock_init(&ioc->lock);
Tejun Heoc5869802011-12-14 00:33:41 +0100277 INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
278 INIT_HLIST_HEAD(&ioc->icq_list);
Tejun Heob2efa052011-12-14 00:33:39 +0100279 INIT_WORK(&ioc->release_work, ioc_release_fn);
Jens Axboe86db1e22008-01-29 14:53:40 +0100280
Tejun Heofd638362011-12-25 14:29:14 +0100281 /*
282 * Try to install. ioc shouldn't be installed if someone else
283 * already did or @task, which isn't %current, is exiting. Note
284 * that we need to allow ioc creation on exiting %current as exit
285 * path may issue IOs from e.g. exit_files(). The exit path is
286 * responsible for not issuing IO after exit_io_context().
287 */
Tejun Heo6e736be2011-12-14 00:33:38 +0100288 task_lock(task);
Tejun Heofd638362011-12-25 14:29:14 +0100289 if (!task->io_context &&
290 (task == current || !(task->flags & PF_EXITING)))
Tejun Heo6e736be2011-12-14 00:33:38 +0100291 task->io_context = ioc;
Tejun Heof2dbd762011-12-14 00:33:40 +0100292 else
Tejun Heo6e736be2011-12-14 00:33:38 +0100293 kmem_cache_free(iocontext_cachep, ioc);
Tejun Heo6e736be2011-12-14 00:33:38 +0100294 task_unlock(task);
Jens Axboe86db1e22008-01-29 14:53:40 +0100295}
Jens Axboe86db1e22008-01-29 14:53:40 +0100296
Tejun Heo6e736be2011-12-14 00:33:38 +0100297/**
298 * get_task_io_context - get io_context of a task
299 * @task: task of interest
300 * @gfp_flags: allocation flags, used if allocation is necessary
301 * @node: allocation node, used if allocation is necessary
Jens Axboe86db1e22008-01-29 14:53:40 +0100302 *
Tejun Heo6e736be2011-12-14 00:33:38 +0100303 * Return io_context of @task. If it doesn't exist, it is created with
304 * @gfp_flags and @node. The returned io_context has its reference count
305 * incremented.
306 *
307 * This function always goes through task_lock() and it's better to use
Tejun Heof2dbd762011-12-14 00:33:40 +0100308 * %current->io_context + get_io_context() for %current.
Jens Axboe86db1e22008-01-29 14:53:40 +0100309 */
Tejun Heo6e736be2011-12-14 00:33:38 +0100310struct io_context *get_task_io_context(struct task_struct *task,
311 gfp_t gfp_flags, int node)
Jens Axboe86db1e22008-01-29 14:53:40 +0100312{
Tejun Heo6e736be2011-12-14 00:33:38 +0100313 struct io_context *ioc;
Jens Axboe86db1e22008-01-29 14:53:40 +0100314
Tejun Heo6e736be2011-12-14 00:33:38 +0100315 might_sleep_if(gfp_flags & __GFP_WAIT);
Jens Axboe86db1e22008-01-29 14:53:40 +0100316
Tejun Heof2dbd762011-12-14 00:33:40 +0100317 do {
318 task_lock(task);
319 ioc = task->io_context;
320 if (likely(ioc)) {
321 get_io_context(ioc);
322 task_unlock(task);
323 return ioc;
324 }
Tejun Heo6e736be2011-12-14 00:33:38 +0100325 task_unlock(task);
Tejun Heof2dbd762011-12-14 00:33:40 +0100326 } while (create_io_context(task, gfp_flags, node));
Tejun Heo6e736be2011-12-14 00:33:38 +0100327
Tejun Heof2dbd762011-12-14 00:33:40 +0100328 return NULL;
Jens Axboe86db1e22008-01-29 14:53:40 +0100329}
Tejun Heo6e736be2011-12-14 00:33:38 +0100330EXPORT_SYMBOL(get_task_io_context);
Jens Axboe86db1e22008-01-29 14:53:40 +0100331
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100332/**
333 * ioc_lookup_icq - lookup io_cq from ioc
334 * @ioc: the associated io_context
335 * @q: the associated request_queue
336 *
337 * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called
338 * with @q->queue_lock held.
339 */
340struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
341{
342 struct io_cq *icq;
343
344 lockdep_assert_held(q->queue_lock);
345
346 /*
347 * icq's are indexed from @ioc using radix tree and hint pointer,
348 * both of which are protected with RCU. All removals are done
349 * holding both q and ioc locks, and we're holding q lock - if we
350 * find a icq which points to us, it's guaranteed to be valid.
351 */
352 rcu_read_lock();
353 icq = rcu_dereference(ioc->icq_hint);
354 if (icq && icq->q == q)
355 goto out;
356
357 icq = radix_tree_lookup(&ioc->icq_tree, q->id);
358 if (icq && icq->q == q)
359 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
360 else
361 icq = NULL;
362out:
363 rcu_read_unlock();
364 return icq;
365}
366EXPORT_SYMBOL(ioc_lookup_icq);
367
Tejun Heof1f8cc92011-12-14 00:33:42 +0100368/**
369 * ioc_create_icq - create and link io_cq
370 * @q: request_queue of interest
371 * @gfp_mask: allocation mask
372 *
373 * Make sure io_cq linking %current->io_context and @q exists. If either
374 * io_context and/or icq don't exist, they will be created using @gfp_mask.
375 *
376 * The caller is responsible for ensuring @ioc won't go away and @q is
377 * alive and will stay alive until this function returns.
378 */
379struct io_cq *ioc_create_icq(struct request_queue *q, gfp_t gfp_mask)
380{
381 struct elevator_type *et = q->elevator->type;
382 struct io_context *ioc;
383 struct io_cq *icq;
384
385 /* allocate stuff */
386 ioc = create_io_context(current, gfp_mask, q->node);
387 if (!ioc)
388 return NULL;
389
390 icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO,
391 q->node);
392 if (!icq)
393 return NULL;
394
395 if (radix_tree_preload(gfp_mask) < 0) {
396 kmem_cache_free(et->icq_cache, icq);
397 return NULL;
398 }
399
400 icq->ioc = ioc;
401 icq->q = q;
402 INIT_LIST_HEAD(&icq->q_node);
403 INIT_HLIST_NODE(&icq->ioc_node);
404
405 /* lock both q and ioc and try to link @icq */
406 spin_lock_irq(q->queue_lock);
407 spin_lock(&ioc->lock);
408
409 if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) {
410 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
411 list_add(&icq->q_node, &q->icq_list);
412 if (et->ops.elevator_init_icq_fn)
413 et->ops.elevator_init_icq_fn(icq);
414 } else {
415 kmem_cache_free(et->icq_cache, icq);
416 icq = ioc_lookup_icq(ioc, q);
417 if (!icq)
418 printk(KERN_ERR "cfq: icq link failed!\n");
419 }
420
421 spin_unlock(&ioc->lock);
422 spin_unlock_irq(q->queue_lock);
423 radix_tree_preload_end();
424 return icq;
425}
426
Tejun Heodc869002011-12-14 00:33:38 +0100427void ioc_set_changed(struct io_context *ioc, int which)
428{
Tejun Heoc5869802011-12-14 00:33:41 +0100429 struct io_cq *icq;
Tejun Heodc869002011-12-14 00:33:38 +0100430 struct hlist_node *n;
431
Tejun Heoc5869802011-12-14 00:33:41 +0100432 hlist_for_each_entry(icq, n, &ioc->icq_list, ioc_node)
433 set_bit(which, &icq->changed);
Tejun Heodc869002011-12-14 00:33:38 +0100434}
435
436/**
437 * ioc_ioprio_changed - notify ioprio change
438 * @ioc: io_context of interest
439 * @ioprio: new ioprio
440 *
Tejun Heoc5869802011-12-14 00:33:41 +0100441 * @ioc's ioprio has changed to @ioprio. Set %ICQ_IOPRIO_CHANGED for all
442 * icq's. iosched is responsible for checking the bit and applying it on
Tejun Heodc869002011-12-14 00:33:38 +0100443 * request issue path.
444 */
445void ioc_ioprio_changed(struct io_context *ioc, int ioprio)
446{
447 unsigned long flags;
448
449 spin_lock_irqsave(&ioc->lock, flags);
450 ioc->ioprio = ioprio;
Tejun Heoc5869802011-12-14 00:33:41 +0100451 ioc_set_changed(ioc, ICQ_IOPRIO_CHANGED);
Tejun Heodc869002011-12-14 00:33:38 +0100452 spin_unlock_irqrestore(&ioc->lock, flags);
453}
454
455/**
456 * ioc_cgroup_changed - notify cgroup change
457 * @ioc: io_context of interest
458 *
Tejun Heoc5869802011-12-14 00:33:41 +0100459 * @ioc's cgroup has changed. Set %ICQ_CGROUP_CHANGED for all icq's.
Tejun Heodc869002011-12-14 00:33:38 +0100460 * iosched is responsible for checking the bit and applying it on request
461 * issue path.
462 */
463void ioc_cgroup_changed(struct io_context *ioc)
464{
465 unsigned long flags;
466
467 spin_lock_irqsave(&ioc->lock, flags);
Tejun Heoc5869802011-12-14 00:33:41 +0100468 ioc_set_changed(ioc, ICQ_CGROUP_CHANGED);
Tejun Heodc869002011-12-14 00:33:38 +0100469 spin_unlock_irqrestore(&ioc->lock, flags);
470}
Jens Axboe64c42992011-12-19 10:36:44 +0100471EXPORT_SYMBOL(ioc_cgroup_changed);
Tejun Heodc869002011-12-14 00:33:38 +0100472
Adrian Bunk13341592008-02-18 13:45:53 +0100473static int __init blk_ioc_init(void)
Jens Axboe86db1e22008-01-29 14:53:40 +0100474{
475 iocontext_cachep = kmem_cache_create("blkdev_ioc",
476 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
477 return 0;
478}
479subsys_initcall(blk_ioc_init);