]>
Commit | Line | Data |
---|---|---|
86db1e29 JA |
1 | /* |
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> | |
5a0e3ad6 | 9 | #include <linux/slab.h> |
86db1e29 JA |
10 | |
11 | #include "blk.h" | |
12 | ||
13 | /* | |
14 | * For io context allocations | |
15 | */ | |
16 | static struct kmem_cache *iocontext_cachep; | |
17 | ||
6e736be7 TH |
18 | /** |
19 | * get_io_context - increment reference count to io_context | |
20 | * @ioc: io_context to get | |
21 | * | |
22 | * Increment reference count to @ioc. | |
23 | */ | |
24 | void get_io_context(struct io_context *ioc) | |
25 | { | |
26 | BUG_ON(atomic_long_read(&ioc->refcount) <= 0); | |
27 | atomic_long_inc(&ioc->refcount); | |
28 | } | |
29 | EXPORT_SYMBOL(get_io_context); | |
30 | ||
7e5a8794 TH |
31 | static 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 | ||
621032ad | 38 | /* Exit an icq. Called with both ioc and q locked. */ |
7e5a8794 | 39 | static void ioc_exit_icq(struct io_cq *icq) |
621032ad TH |
40 | { |
41 | struct elevator_type *et = icq->q->elevator->type; | |
42 | ||
43 | if (icq->flags & ICQ_EXITED) | |
44 | return; | |
45 | ||
bd166ef1 JA |
46 | if (et->uses_mq && et->ops.mq.exit_icq) |
47 | et->ops.mq.exit_icq(icq); | |
48 | else if (!et->uses_mq && et->ops.sq.elevator_exit_icq_fn) | |
c51ca6cf | 49 | et->ops.sq.elevator_exit_icq_fn(icq); |
621032ad TH |
50 | |
51 | icq->flags |= ICQ_EXITED; | |
52 | } | |
53 | ||
54 | /* Release an icq. Called with both ioc and q locked. */ | |
55 | static void ioc_destroy_icq(struct io_cq *icq) | |
7e5a8794 TH |
56 | { |
57 | struct io_context *ioc = icq->ioc; | |
58 | struct request_queue *q = icq->q; | |
59 | struct elevator_type *et = q->elevator->type; | |
60 | ||
61 | lockdep_assert_held(&ioc->lock); | |
62 | lockdep_assert_held(q->queue_lock); | |
63 | ||
64 | radix_tree_delete(&ioc->icq_tree, icq->q->id); | |
65 | hlist_del_init(&icq->ioc_node); | |
66 | list_del_init(&icq->q_node); | |
67 | ||
68 | /* | |
69 | * Both setting lookup hint to and clearing it from @icq are done | |
70 | * under queue_lock. If it's not pointing to @icq now, it never | |
71 | * will. Hint assignment itself can race safely. | |
72 | */ | |
ec6c676a | 73 | if (rcu_access_pointer(ioc->icq_hint) == icq) |
7e5a8794 TH |
74 | rcu_assign_pointer(ioc->icq_hint, NULL); |
75 | ||
621032ad | 76 | ioc_exit_icq(icq); |
7e5a8794 TH |
77 | |
78 | /* | |
79 | * @icq->q might have gone away by the time RCU callback runs | |
80 | * making it impossible to determine icq_cache. Record it in @icq. | |
81 | */ | |
82 | icq->__rcu_icq_cache = et->icq_cache; | |
83 | call_rcu(&icq->__rcu_head, icq_free_icq_rcu); | |
84 | } | |
85 | ||
b2efa052 TH |
86 | /* |
87 | * Slow path for ioc release in put_io_context(). Performs double-lock | |
c5869807 | 88 | * dancing to unlink all icq's and then frees ioc. |
b2efa052 TH |
89 | */ |
90 | static void ioc_release_fn(struct work_struct *work) | |
86db1e29 | 91 | { |
b2efa052 TH |
92 | struct io_context *ioc = container_of(work, struct io_context, |
93 | release_work); | |
d8c66c5d | 94 | unsigned long flags; |
b2efa052 | 95 | |
d8c66c5d TH |
96 | /* |
97 | * Exiting icq may call into put_io_context() through elevator | |
98 | * which will trigger lockdep warning. The ioc's are guaranteed to | |
99 | * be different, use a different locking subclass here. Use | |
100 | * irqsave variant as there's no spin_lock_irq_nested(). | |
101 | */ | |
102 | spin_lock_irqsave_nested(&ioc->lock, flags, 1); | |
b2efa052 | 103 | |
c5869807 TH |
104 | while (!hlist_empty(&ioc->icq_list)) { |
105 | struct io_cq *icq = hlist_entry(ioc->icq_list.first, | |
106 | struct io_cq, ioc_node); | |
2274b029 TH |
107 | struct request_queue *q = icq->q; |
108 | ||
109 | if (spin_trylock(q->queue_lock)) { | |
621032ad | 110 | ioc_destroy_icq(icq); |
2274b029 TH |
111 | spin_unlock(q->queue_lock); |
112 | } else { | |
113 | spin_unlock_irqrestore(&ioc->lock, flags); | |
114 | cpu_relax(); | |
115 | spin_lock_irqsave_nested(&ioc->lock, flags, 1); | |
b2efa052 | 116 | } |
b2efa052 | 117 | } |
ffc4e759 | 118 | |
2274b029 | 119 | spin_unlock_irqrestore(&ioc->lock, flags); |
b2efa052 TH |
120 | |
121 | kmem_cache_free(iocontext_cachep, ioc); | |
86db1e29 JA |
122 | } |
123 | ||
42ec57a8 TH |
124 | /** |
125 | * put_io_context - put a reference of io_context | |
126 | * @ioc: io_context to put | |
127 | * | |
128 | * Decrement reference count of @ioc and release it if the count reaches | |
11a3122f | 129 | * zero. |
86db1e29 | 130 | */ |
11a3122f | 131 | void put_io_context(struct io_context *ioc) |
86db1e29 | 132 | { |
b2efa052 | 133 | unsigned long flags; |
ff8c1474 | 134 | bool free_ioc = false; |
b2efa052 | 135 | |
86db1e29 | 136 | if (ioc == NULL) |
42ec57a8 | 137 | return; |
86db1e29 | 138 | |
42ec57a8 | 139 | BUG_ON(atomic_long_read(&ioc->refcount) <= 0); |
86db1e29 | 140 | |
b2efa052 | 141 | /* |
11a3122f TH |
142 | * Releasing ioc requires reverse order double locking and we may |
143 | * already be holding a queue_lock. Do it asynchronously from wq. | |
b2efa052 | 144 | */ |
11a3122f TH |
145 | if (atomic_long_dec_and_test(&ioc->refcount)) { |
146 | spin_lock_irqsave(&ioc->lock, flags); | |
147 | if (!hlist_empty(&ioc->icq_list)) | |
695588f9 VK |
148 | queue_work(system_power_efficient_wq, |
149 | &ioc->release_work); | |
ff8c1474 XF |
150 | else |
151 | free_ioc = true; | |
11a3122f | 152 | spin_unlock_irqrestore(&ioc->lock, flags); |
b2efa052 | 153 | } |
ff8c1474 XF |
154 | |
155 | if (free_ioc) | |
156 | kmem_cache_free(iocontext_cachep, ioc); | |
86db1e29 | 157 | } |
b2efa052 | 158 | EXPORT_SYMBOL(put_io_context); |
86db1e29 | 159 | |
f6e8d01b TH |
160 | /** |
161 | * put_io_context_active - put active reference on ioc | |
162 | * @ioc: ioc of interest | |
163 | * | |
164 | * Undo get_io_context_active(). If active reference reaches zero after | |
165 | * put, @ioc can never issue further IOs and ioscheds are notified. | |
166 | */ | |
167 | void put_io_context_active(struct io_context *ioc) | |
86db1e29 | 168 | { |
621032ad | 169 | unsigned long flags; |
f6e8d01b | 170 | struct io_cq *icq; |
86db1e29 | 171 | |
f6e8d01b | 172 | if (!atomic_dec_and_test(&ioc->active_ref)) { |
621032ad TH |
173 | put_io_context(ioc); |
174 | return; | |
175 | } | |
176 | ||
177 | /* | |
178 | * Need ioc lock to walk icq_list and q lock to exit icq. Perform | |
179 | * reverse double locking. Read comment in ioc_release_fn() for | |
180 | * explanation on the nested locking annotation. | |
181 | */ | |
182 | retry: | |
183 | spin_lock_irqsave_nested(&ioc->lock, flags, 1); | |
b67bfe0d | 184 | hlist_for_each_entry(icq, &ioc->icq_list, ioc_node) { |
621032ad TH |
185 | if (icq->flags & ICQ_EXITED) |
186 | continue; | |
187 | if (spin_trylock(icq->q->queue_lock)) { | |
188 | ioc_exit_icq(icq); | |
189 | spin_unlock(icq->q->queue_lock); | |
190 | } else { | |
191 | spin_unlock_irqrestore(&ioc->lock, flags); | |
192 | cpu_relax(); | |
193 | goto retry; | |
194 | } | |
195 | } | |
196 | spin_unlock_irqrestore(&ioc->lock, flags); | |
197 | ||
11a3122f | 198 | put_io_context(ioc); |
86db1e29 JA |
199 | } |
200 | ||
f6e8d01b TH |
201 | /* Called by the exiting task */ |
202 | void exit_io_context(struct task_struct *task) | |
203 | { | |
204 | struct io_context *ioc; | |
205 | ||
206 | task_lock(task); | |
207 | ioc = task->io_context; | |
208 | task->io_context = NULL; | |
209 | task_unlock(task); | |
210 | ||
211 | atomic_dec(&ioc->nr_tasks); | |
212 | put_io_context_active(ioc); | |
213 | } | |
214 | ||
7e5a8794 TH |
215 | /** |
216 | * ioc_clear_queue - break any ioc association with the specified queue | |
217 | * @q: request_queue being cleared | |
218 | * | |
219 | * Walk @q->icq_list and exit all io_cq's. Must be called with @q locked. | |
220 | */ | |
221 | void ioc_clear_queue(struct request_queue *q) | |
222 | { | |
223 | lockdep_assert_held(q->queue_lock); | |
224 | ||
225 | while (!list_empty(&q->icq_list)) { | |
226 | struct io_cq *icq = list_entry(q->icq_list.next, | |
227 | struct io_cq, q_node); | |
228 | struct io_context *ioc = icq->ioc; | |
229 | ||
230 | spin_lock(&ioc->lock); | |
621032ad | 231 | ioc_destroy_icq(icq); |
7e5a8794 TH |
232 | spin_unlock(&ioc->lock); |
233 | } | |
234 | } | |
235 | ||
24acfc34 | 236 | int create_task_io_context(struct task_struct *task, gfp_t gfp_flags, int node) |
86db1e29 | 237 | { |
df415656 | 238 | struct io_context *ioc; |
3c9c708c | 239 | int ret; |
86db1e29 | 240 | |
42ec57a8 TH |
241 | ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO, |
242 | node); | |
243 | if (unlikely(!ioc)) | |
24acfc34 | 244 | return -ENOMEM; |
42ec57a8 TH |
245 | |
246 | /* initialize */ | |
247 | atomic_long_set(&ioc->refcount, 1); | |
4638a83e | 248 | atomic_set(&ioc->nr_tasks, 1); |
f6e8d01b | 249 | atomic_set(&ioc->active_ref, 1); |
42ec57a8 | 250 | spin_lock_init(&ioc->lock); |
c5869807 TH |
251 | INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH); |
252 | INIT_HLIST_HEAD(&ioc->icq_list); | |
b2efa052 | 253 | INIT_WORK(&ioc->release_work, ioc_release_fn); |
86db1e29 | 254 | |
fd638368 TH |
255 | /* |
256 | * Try to install. ioc shouldn't be installed if someone else | |
257 | * already did or @task, which isn't %current, is exiting. Note | |
258 | * that we need to allow ioc creation on exiting %current as exit | |
259 | * path may issue IOs from e.g. exit_files(). The exit path is | |
260 | * responsible for not issuing IO after exit_io_context(). | |
261 | */ | |
6e736be7 | 262 | task_lock(task); |
fd638368 TH |
263 | if (!task->io_context && |
264 | (task == current || !(task->flags & PF_EXITING))) | |
6e736be7 | 265 | task->io_context = ioc; |
f2dbd76a | 266 | else |
6e736be7 | 267 | kmem_cache_free(iocontext_cachep, ioc); |
3c9c708c ED |
268 | |
269 | ret = task->io_context ? 0 : -EBUSY; | |
270 | ||
6e736be7 | 271 | task_unlock(task); |
24acfc34 | 272 | |
3c9c708c | 273 | return ret; |
86db1e29 | 274 | } |
86db1e29 | 275 | |
6e736be7 TH |
276 | /** |
277 | * get_task_io_context - get io_context of a task | |
278 | * @task: task of interest | |
279 | * @gfp_flags: allocation flags, used if allocation is necessary | |
280 | * @node: allocation node, used if allocation is necessary | |
281 | * | |
282 | * Return io_context of @task. If it doesn't exist, it is created with | |
283 | * @gfp_flags and @node. The returned io_context has its reference count | |
284 | * incremented. | |
86db1e29 | 285 | * |
6e736be7 | 286 | * This function always goes through task_lock() and it's better to use |
f2dbd76a | 287 | * %current->io_context + get_io_context() for %current. |
86db1e29 | 288 | */ |
6e736be7 TH |
289 | struct io_context *get_task_io_context(struct task_struct *task, |
290 | gfp_t gfp_flags, int node) | |
86db1e29 | 291 | { |
6e736be7 | 292 | struct io_context *ioc; |
86db1e29 | 293 | |
d0164adc | 294 | might_sleep_if(gfpflags_allow_blocking(gfp_flags)); |
6e736be7 | 295 | |
f2dbd76a TH |
296 | do { |
297 | task_lock(task); | |
298 | ioc = task->io_context; | |
299 | if (likely(ioc)) { | |
300 | get_io_context(ioc); | |
301 | task_unlock(task); | |
302 | return ioc; | |
303 | } | |
6e736be7 | 304 | task_unlock(task); |
24acfc34 | 305 | } while (!create_task_io_context(task, gfp_flags, node)); |
6e736be7 | 306 | |
f2dbd76a | 307 | return NULL; |
86db1e29 | 308 | } |
6e736be7 | 309 | EXPORT_SYMBOL(get_task_io_context); |
86db1e29 | 310 | |
47fdd4ca TH |
311 | /** |
312 | * ioc_lookup_icq - lookup io_cq from ioc | |
313 | * @ioc: the associated io_context | |
314 | * @q: the associated request_queue | |
315 | * | |
316 | * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called | |
317 | * with @q->queue_lock held. | |
318 | */ | |
319 | struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q) | |
320 | { | |
321 | struct io_cq *icq; | |
322 | ||
323 | lockdep_assert_held(q->queue_lock); | |
324 | ||
325 | /* | |
326 | * icq's are indexed from @ioc using radix tree and hint pointer, | |
327 | * both of which are protected with RCU. All removals are done | |
328 | * holding both q and ioc locks, and we're holding q lock - if we | |
329 | * find a icq which points to us, it's guaranteed to be valid. | |
330 | */ | |
331 | rcu_read_lock(); | |
332 | icq = rcu_dereference(ioc->icq_hint); | |
333 | if (icq && icq->q == q) | |
334 | goto out; | |
335 | ||
336 | icq = radix_tree_lookup(&ioc->icq_tree, q->id); | |
337 | if (icq && icq->q == q) | |
338 | rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */ | |
339 | else | |
340 | icq = NULL; | |
341 | out: | |
342 | rcu_read_unlock(); | |
343 | return icq; | |
344 | } | |
345 | EXPORT_SYMBOL(ioc_lookup_icq); | |
346 | ||
f1f8cc94 TH |
347 | /** |
348 | * ioc_create_icq - create and link io_cq | |
24acfc34 | 349 | * @ioc: io_context of interest |
f1f8cc94 TH |
350 | * @q: request_queue of interest |
351 | * @gfp_mask: allocation mask | |
352 | * | |
24acfc34 TH |
353 | * Make sure io_cq linking @ioc and @q exists. If icq doesn't exist, they |
354 | * will be created using @gfp_mask. | |
f1f8cc94 TH |
355 | * |
356 | * The caller is responsible for ensuring @ioc won't go away and @q is | |
357 | * alive and will stay alive until this function returns. | |
358 | */ | |
24acfc34 TH |
359 | struct io_cq *ioc_create_icq(struct io_context *ioc, struct request_queue *q, |
360 | gfp_t gfp_mask) | |
f1f8cc94 TH |
361 | { |
362 | struct elevator_type *et = q->elevator->type; | |
f1f8cc94 TH |
363 | struct io_cq *icq; |
364 | ||
365 | /* allocate stuff */ | |
f1f8cc94 TH |
366 | icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO, |
367 | q->node); | |
368 | if (!icq) | |
369 | return NULL; | |
370 | ||
5e4c0d97 | 371 | if (radix_tree_maybe_preload(gfp_mask) < 0) { |
f1f8cc94 TH |
372 | kmem_cache_free(et->icq_cache, icq); |
373 | return NULL; | |
374 | } | |
375 | ||
376 | icq->ioc = ioc; | |
377 | icq->q = q; | |
378 | INIT_LIST_HEAD(&icq->q_node); | |
379 | INIT_HLIST_NODE(&icq->ioc_node); | |
380 | ||
381 | /* lock both q and ioc and try to link @icq */ | |
382 | spin_lock_irq(q->queue_lock); | |
383 | spin_lock(&ioc->lock); | |
384 | ||
385 | if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) { | |
386 | hlist_add_head(&icq->ioc_node, &ioc->icq_list); | |
387 | list_add(&icq->q_node, &q->icq_list); | |
bd166ef1 JA |
388 | if (et->uses_mq && et->ops.mq.init_icq) |
389 | et->ops.mq.init_icq(icq); | |
390 | else if (!et->uses_mq && et->ops.sq.elevator_init_icq_fn) | |
c51ca6cf | 391 | et->ops.sq.elevator_init_icq_fn(icq); |
f1f8cc94 TH |
392 | } else { |
393 | kmem_cache_free(et->icq_cache, icq); | |
394 | icq = ioc_lookup_icq(ioc, q); | |
395 | if (!icq) | |
396 | printk(KERN_ERR "cfq: icq link failed!\n"); | |
397 | } | |
398 | ||
399 | spin_unlock(&ioc->lock); | |
400 | spin_unlock_irq(q->queue_lock); | |
401 | radix_tree_preload_end(); | |
402 | return icq; | |
403 | } | |
404 | ||
13341598 | 405 | static int __init blk_ioc_init(void) |
86db1e29 JA |
406 | { |
407 | iocontext_cachep = kmem_cache_create("blkdev_ioc", | |
408 | sizeof(struct io_context), 0, SLAB_PANIC, NULL); | |
409 | return 0; | |
410 | } | |
411 | subsys_initcall(blk_ioc_init); |