]>
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> | |
9 | #include <linux/bootmem.h> /* for max_pfn/max_low_pfn */ | |
5a0e3ad6 | 10 | #include <linux/slab.h> |
86db1e29 JA |
11 | |
12 | #include "blk.h" | |
13 | ||
14 | /* | |
15 | * For io context allocations | |
16 | */ | |
17 | static struct kmem_cache *iocontext_cachep; | |
18 | ||
6e736be7 TH |
19 | /** |
20 | * get_io_context - increment reference count to io_context | |
21 | * @ioc: io_context to get | |
22 | * | |
23 | * Increment reference count to @ioc. | |
24 | */ | |
25 | void get_io_context(struct io_context *ioc) | |
26 | { | |
27 | BUG_ON(atomic_long_read(&ioc->refcount) <= 0); | |
28 | atomic_long_inc(&ioc->refcount); | |
29 | } | |
30 | EXPORT_SYMBOL(get_io_context); | |
31 | ||
b2efa052 TH |
32 | /* |
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 | |
46 | ||
7e5a8794 TH |
47 | static 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 | */ | |
57 | static 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 | ||
b2efa052 TH |
92 | /* |
93 | * Slow path for ioc release in put_io_context(). Performs double-lock | |
c5869807 | 94 | * dancing to unlink all icq's and then frees ioc. |
b2efa052 TH |
95 | */ |
96 | static void ioc_release_fn(struct work_struct *work) | |
86db1e29 | 97 | { |
b2efa052 TH |
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 | ||
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); | |
107 | struct request_queue *this_q = icq->q; | |
b2efa052 TH |
108 | |
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 | } | |
7e5a8794 | 135 | ioc_exit_icq(icq); |
b2efa052 | 136 | } |
ffc4e759 | 137 | |
b2efa052 TH |
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); | |
ffc4e759 | 144 | } |
b2efa052 TH |
145 | |
146 | kmem_cache_free(iocontext_cachep, ioc); | |
86db1e29 JA |
147 | } |
148 | ||
42ec57a8 TH |
149 | /** |
150 | * put_io_context - put a reference of io_context | |
151 | * @ioc: io_context to put | |
b2efa052 | 152 | * @locked_q: request_queue the caller is holding queue_lock of (hint) |
42ec57a8 TH |
153 | * |
154 | * Decrement reference count of @ioc and release it if the count reaches | |
b2efa052 TH |
155 | * 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. | |
86db1e29 | 158 | */ |
b2efa052 | 159 | void put_io_context(struct io_context *ioc, struct request_queue *locked_q) |
86db1e29 | 160 | { |
b2efa052 TH |
161 | struct request_queue *last_q = locked_q; |
162 | unsigned long flags; | |
163 | ||
86db1e29 | 164 | if (ioc == NULL) |
42ec57a8 | 165 | return; |
86db1e29 | 166 | |
42ec57a8 | 167 | BUG_ON(atomic_long_read(&ioc->refcount) <= 0); |
b2efa052 TH |
168 | if (locked_q) |
169 | lockdep_assert_held(locked_q->queue_lock); | |
86db1e29 | 170 | |
42ec57a8 TH |
171 | if (!atomic_long_dec_and_test(&ioc->refcount)) |
172 | return; | |
86db1e29 | 173 | |
b2efa052 | 174 | /* |
c5869807 | 175 | * Destroy @ioc. This is a bit messy because icq's are chained |
b2efa052 | 176 | * from both ioc and queue, and ioc->lock nests inside queue_lock. |
c5869807 TH |
177 | * 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. | |
b2efa052 TH |
179 | * 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)); | |
196 | ||
c5869807 TH |
197 | 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; | |
b2efa052 TH |
201 | |
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 | } | |
7e5a8794 | 212 | ioc_exit_icq(icq); |
b2efa052 | 213 | } |
86db1e29 | 214 | |
b2efa052 TH |
215 | if (last_q && last_q != locked_q) |
216 | spin_unlock(last_q->queue_lock); | |
86db1e29 | 217 | |
b2efa052 | 218 | spin_unlock_irqrestore(&ioc->lock, flags); |
ffc4e759 | 219 | |
c5869807 TH |
220 | /* if no icq is left, we're done; otherwise, kick release_work */ |
221 | if (hlist_empty(&ioc->icq_list)) | |
b2efa052 TH |
222 | kmem_cache_free(iocontext_cachep, ioc); |
223 | else | |
224 | schedule_work(&ioc->release_work); | |
86db1e29 | 225 | } |
b2efa052 | 226 | EXPORT_SYMBOL(put_io_context); |
86db1e29 | 227 | |
27667c99 | 228 | /* Called by the exiting task */ |
b69f2292 | 229 | void exit_io_context(struct task_struct *task) |
86db1e29 JA |
230 | { |
231 | struct io_context *ioc; | |
232 | ||
6e736be7 TH |
233 | /* PF_EXITING prevents new io_context from being attached to @task */ |
234 | WARN_ON_ONCE(!(current->flags & PF_EXITING)); | |
235 | ||
b69f2292 LR |
236 | task_lock(task); |
237 | ioc = task->io_context; | |
238 | task->io_context = NULL; | |
239 | task_unlock(task); | |
86db1e29 | 240 | |
b2efa052 TH |
241 | atomic_dec(&ioc->nr_tasks); |
242 | put_io_context(ioc, NULL); | |
86db1e29 JA |
243 | } |
244 | ||
7e5a8794 TH |
245 | /** |
246 | * ioc_clear_queue - break any ioc association with the specified queue | |
247 | * @q: request_queue being cleared | |
248 | * | |
249 | * Walk @q->icq_list and exit all io_cq's. Must be called with @q locked. | |
250 | */ | |
251 | void ioc_clear_queue(struct request_queue *q) | |
252 | { | |
253 | lockdep_assert_held(q->queue_lock); | |
254 | ||
255 | while (!list_empty(&q->icq_list)) { | |
256 | struct io_cq *icq = list_entry(q->icq_list.next, | |
257 | struct io_cq, q_node); | |
258 | struct io_context *ioc = icq->ioc; | |
259 | ||
260 | spin_lock(&ioc->lock); | |
261 | ioc_exit_icq(icq); | |
262 | spin_unlock(&ioc->lock); | |
263 | } | |
264 | } | |
265 | ||
f2dbd76a TH |
266 | void create_io_context_slowpath(struct task_struct *task, gfp_t gfp_flags, |
267 | int node) | |
86db1e29 | 268 | { |
df415656 | 269 | struct io_context *ioc; |
86db1e29 | 270 | |
42ec57a8 TH |
271 | ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO, |
272 | node); | |
273 | if (unlikely(!ioc)) | |
f2dbd76a | 274 | return; |
42ec57a8 TH |
275 | |
276 | /* initialize */ | |
277 | atomic_long_set(&ioc->refcount, 1); | |
278 | atomic_set(&ioc->nr_tasks, 1); | |
279 | spin_lock_init(&ioc->lock); | |
c5869807 TH |
280 | INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH); |
281 | INIT_HLIST_HEAD(&ioc->icq_list); | |
b2efa052 | 282 | INIT_WORK(&ioc->release_work, ioc_release_fn); |
86db1e29 | 283 | |
6e736be7 TH |
284 | /* try to install, somebody might already have beaten us to it */ |
285 | task_lock(task); | |
f2dbd76a | 286 | if (!task->io_context && !(task->flags & PF_EXITING)) |
6e736be7 | 287 | task->io_context = ioc; |
f2dbd76a | 288 | else |
6e736be7 | 289 | kmem_cache_free(iocontext_cachep, ioc); |
6e736be7 | 290 | task_unlock(task); |
86db1e29 | 291 | } |
86db1e29 | 292 | |
6e736be7 TH |
293 | /** |
294 | * get_task_io_context - get io_context of a task | |
295 | * @task: task of interest | |
296 | * @gfp_flags: allocation flags, used if allocation is necessary | |
297 | * @node: allocation node, used if allocation is necessary | |
298 | * | |
299 | * Return io_context of @task. If it doesn't exist, it is created with | |
300 | * @gfp_flags and @node. The returned io_context has its reference count | |
301 | * incremented. | |
86db1e29 | 302 | * |
6e736be7 | 303 | * This function always goes through task_lock() and it's better to use |
f2dbd76a | 304 | * %current->io_context + get_io_context() for %current. |
86db1e29 | 305 | */ |
6e736be7 TH |
306 | struct io_context *get_task_io_context(struct task_struct *task, |
307 | gfp_t gfp_flags, int node) | |
86db1e29 | 308 | { |
6e736be7 | 309 | struct io_context *ioc; |
86db1e29 | 310 | |
6e736be7 TH |
311 | might_sleep_if(gfp_flags & __GFP_WAIT); |
312 | ||
f2dbd76a TH |
313 | do { |
314 | task_lock(task); | |
315 | ioc = task->io_context; | |
316 | if (likely(ioc)) { | |
317 | get_io_context(ioc); | |
318 | task_unlock(task); | |
319 | return ioc; | |
320 | } | |
6e736be7 | 321 | task_unlock(task); |
f2dbd76a | 322 | } while (create_io_context(task, gfp_flags, node)); |
6e736be7 | 323 | |
f2dbd76a | 324 | return NULL; |
86db1e29 | 325 | } |
6e736be7 | 326 | EXPORT_SYMBOL(get_task_io_context); |
86db1e29 | 327 | |
47fdd4ca TH |
328 | /** |
329 | * ioc_lookup_icq - lookup io_cq from ioc | |
330 | * @ioc: the associated io_context | |
331 | * @q: the associated request_queue | |
332 | * | |
333 | * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called | |
334 | * with @q->queue_lock held. | |
335 | */ | |
336 | struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q) | |
337 | { | |
338 | struct io_cq *icq; | |
339 | ||
340 | lockdep_assert_held(q->queue_lock); | |
341 | ||
342 | /* | |
343 | * icq's are indexed from @ioc using radix tree and hint pointer, | |
344 | * both of which are protected with RCU. All removals are done | |
345 | * holding both q and ioc locks, and we're holding q lock - if we | |
346 | * find a icq which points to us, it's guaranteed to be valid. | |
347 | */ | |
348 | rcu_read_lock(); | |
349 | icq = rcu_dereference(ioc->icq_hint); | |
350 | if (icq && icq->q == q) | |
351 | goto out; | |
352 | ||
353 | icq = radix_tree_lookup(&ioc->icq_tree, q->id); | |
354 | if (icq && icq->q == q) | |
355 | rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */ | |
356 | else | |
357 | icq = NULL; | |
358 | out: | |
359 | rcu_read_unlock(); | |
360 | return icq; | |
361 | } | |
362 | EXPORT_SYMBOL(ioc_lookup_icq); | |
363 | ||
f1f8cc94 TH |
364 | /** |
365 | * ioc_create_icq - create and link io_cq | |
366 | * @q: request_queue of interest | |
367 | * @gfp_mask: allocation mask | |
368 | * | |
369 | * Make sure io_cq linking %current->io_context and @q exists. If either | |
370 | * io_context and/or icq don't exist, they will be created using @gfp_mask. | |
371 | * | |
372 | * The caller is responsible for ensuring @ioc won't go away and @q is | |
373 | * alive and will stay alive until this function returns. | |
374 | */ | |
375 | struct io_cq *ioc_create_icq(struct request_queue *q, gfp_t gfp_mask) | |
376 | { | |
377 | struct elevator_type *et = q->elevator->type; | |
378 | struct io_context *ioc; | |
379 | struct io_cq *icq; | |
380 | ||
381 | /* allocate stuff */ | |
382 | ioc = create_io_context(current, gfp_mask, q->node); | |
383 | if (!ioc) | |
384 | return NULL; | |
385 | ||
386 | icq = kmem_cache_alloc_node(et->icq_cache, gfp_mask | __GFP_ZERO, | |
387 | q->node); | |
388 | if (!icq) | |
389 | return NULL; | |
390 | ||
391 | if (radix_tree_preload(gfp_mask) < 0) { | |
392 | kmem_cache_free(et->icq_cache, icq); | |
393 | return NULL; | |
394 | } | |
395 | ||
396 | icq->ioc = ioc; | |
397 | icq->q = q; | |
398 | INIT_LIST_HEAD(&icq->q_node); | |
399 | INIT_HLIST_NODE(&icq->ioc_node); | |
400 | ||
401 | /* lock both q and ioc and try to link @icq */ | |
402 | spin_lock_irq(q->queue_lock); | |
403 | spin_lock(&ioc->lock); | |
404 | ||
405 | if (likely(!radix_tree_insert(&ioc->icq_tree, q->id, icq))) { | |
406 | hlist_add_head(&icq->ioc_node, &ioc->icq_list); | |
407 | list_add(&icq->q_node, &q->icq_list); | |
408 | if (et->ops.elevator_init_icq_fn) | |
409 | et->ops.elevator_init_icq_fn(icq); | |
410 | } else { | |
411 | kmem_cache_free(et->icq_cache, icq); | |
412 | icq = ioc_lookup_icq(ioc, q); | |
413 | if (!icq) | |
414 | printk(KERN_ERR "cfq: icq link failed!\n"); | |
415 | } | |
416 | ||
417 | spin_unlock(&ioc->lock); | |
418 | spin_unlock_irq(q->queue_lock); | |
419 | radix_tree_preload_end(); | |
420 | return icq; | |
421 | } | |
422 | ||
dc86900e TH |
423 | void ioc_set_changed(struct io_context *ioc, int which) |
424 | { | |
c5869807 | 425 | struct io_cq *icq; |
dc86900e TH |
426 | struct hlist_node *n; |
427 | ||
c5869807 TH |
428 | hlist_for_each_entry(icq, n, &ioc->icq_list, ioc_node) |
429 | set_bit(which, &icq->changed); | |
dc86900e TH |
430 | } |
431 | ||
432 | /** | |
433 | * ioc_ioprio_changed - notify ioprio change | |
434 | * @ioc: io_context of interest | |
435 | * @ioprio: new ioprio | |
436 | * | |
c5869807 TH |
437 | * @ioc's ioprio has changed to @ioprio. Set %ICQ_IOPRIO_CHANGED for all |
438 | * icq's. iosched is responsible for checking the bit and applying it on | |
dc86900e TH |
439 | * request issue path. |
440 | */ | |
441 | void ioc_ioprio_changed(struct io_context *ioc, int ioprio) | |
442 | { | |
443 | unsigned long flags; | |
444 | ||
445 | spin_lock_irqsave(&ioc->lock, flags); | |
446 | ioc->ioprio = ioprio; | |
c5869807 | 447 | ioc_set_changed(ioc, ICQ_IOPRIO_CHANGED); |
dc86900e TH |
448 | spin_unlock_irqrestore(&ioc->lock, flags); |
449 | } | |
450 | ||
451 | /** | |
452 | * ioc_cgroup_changed - notify cgroup change | |
453 | * @ioc: io_context of interest | |
454 | * | |
c5869807 | 455 | * @ioc's cgroup has changed. Set %ICQ_CGROUP_CHANGED for all icq's. |
dc86900e TH |
456 | * iosched is responsible for checking the bit and applying it on request |
457 | * issue path. | |
458 | */ | |
459 | void ioc_cgroup_changed(struct io_context *ioc) | |
460 | { | |
461 | unsigned long flags; | |
462 | ||
463 | spin_lock_irqsave(&ioc->lock, flags); | |
c5869807 | 464 | ioc_set_changed(ioc, ICQ_CGROUP_CHANGED); |
dc86900e TH |
465 | spin_unlock_irqrestore(&ioc->lock, flags); |
466 | } | |
64c42998 | 467 | EXPORT_SYMBOL(ioc_cgroup_changed); |
dc86900e | 468 | |
13341598 | 469 | static int __init blk_ioc_init(void) |
86db1e29 JA |
470 | { |
471 | iocontext_cachep = kmem_cache_create("blkdev_ioc", | |
472 | sizeof(struct io_context), 0, SLAB_PANIC, NULL); | |
473 | return 0; | |
474 | } | |
475 | subsys_initcall(blk_ioc_init); |