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 | ||
47 | /* | |
48 | * Slow path for ioc release in put_io_context(). Performs double-lock | |
c5869807 | 49 | * dancing to unlink all icq's and then frees ioc. |
b2efa052 TH |
50 | */ |
51 | static void ioc_release_fn(struct work_struct *work) | |
86db1e29 | 52 | { |
b2efa052 TH |
53 | struct io_context *ioc = container_of(work, struct io_context, |
54 | release_work); | |
55 | struct request_queue *last_q = NULL; | |
56 | ||
57 | spin_lock_irq(&ioc->lock); | |
58 | ||
c5869807 TH |
59 | while (!hlist_empty(&ioc->icq_list)) { |
60 | struct io_cq *icq = hlist_entry(ioc->icq_list.first, | |
61 | struct io_cq, ioc_node); | |
62 | struct request_queue *this_q = icq->q; | |
b2efa052 TH |
63 | |
64 | if (this_q != last_q) { | |
65 | /* | |
66 | * Need to switch to @this_q. Once we release | |
67 | * @ioc->lock, it can go away along with @cic. | |
68 | * Hold on to it. | |
69 | */ | |
70 | __blk_get_queue(this_q); | |
71 | ||
72 | /* | |
73 | * blk_put_queue() might sleep thanks to kobject | |
74 | * idiocy. Always release both locks, put and | |
75 | * restart. | |
76 | */ | |
77 | if (last_q) { | |
78 | spin_unlock(last_q->queue_lock); | |
79 | spin_unlock_irq(&ioc->lock); | |
80 | blk_put_queue(last_q); | |
81 | } else { | |
82 | spin_unlock_irq(&ioc->lock); | |
83 | } | |
84 | ||
85 | last_q = this_q; | |
86 | spin_lock_irq(this_q->queue_lock); | |
87 | spin_lock(&ioc->lock); | |
88 | continue; | |
89 | } | |
90 | ioc_release_depth_inc(this_q); | |
c5869807 TH |
91 | icq->exit(icq); |
92 | icq->release(icq); | |
b2efa052 TH |
93 | ioc_release_depth_dec(this_q); |
94 | } | |
ffc4e759 | 95 | |
b2efa052 TH |
96 | if (last_q) { |
97 | spin_unlock(last_q->queue_lock); | |
98 | spin_unlock_irq(&ioc->lock); | |
99 | blk_put_queue(last_q); | |
100 | } else { | |
101 | spin_unlock_irq(&ioc->lock); | |
ffc4e759 | 102 | } |
b2efa052 TH |
103 | |
104 | kmem_cache_free(iocontext_cachep, ioc); | |
86db1e29 JA |
105 | } |
106 | ||
42ec57a8 TH |
107 | /** |
108 | * put_io_context - put a reference of io_context | |
109 | * @ioc: io_context to put | |
b2efa052 | 110 | * @locked_q: request_queue the caller is holding queue_lock of (hint) |
42ec57a8 TH |
111 | * |
112 | * Decrement reference count of @ioc and release it if the count reaches | |
b2efa052 TH |
113 | * zero. If the caller is holding queue_lock of a queue, it can indicate |
114 | * that with @locked_q. This is an optimization hint and the caller is | |
115 | * allowed to pass in %NULL even when it's holding a queue_lock. | |
86db1e29 | 116 | */ |
b2efa052 | 117 | void put_io_context(struct io_context *ioc, struct request_queue *locked_q) |
86db1e29 | 118 | { |
b2efa052 TH |
119 | struct request_queue *last_q = locked_q; |
120 | unsigned long flags; | |
121 | ||
86db1e29 | 122 | if (ioc == NULL) |
42ec57a8 | 123 | return; |
86db1e29 | 124 | |
42ec57a8 | 125 | BUG_ON(atomic_long_read(&ioc->refcount) <= 0); |
b2efa052 TH |
126 | if (locked_q) |
127 | lockdep_assert_held(locked_q->queue_lock); | |
86db1e29 | 128 | |
42ec57a8 TH |
129 | if (!atomic_long_dec_and_test(&ioc->refcount)) |
130 | return; | |
86db1e29 | 131 | |
b2efa052 | 132 | /* |
c5869807 | 133 | * Destroy @ioc. This is a bit messy because icq's are chained |
b2efa052 | 134 | * from both ioc and queue, and ioc->lock nests inside queue_lock. |
c5869807 TH |
135 | * The inner ioc->lock should be held to walk our icq_list and then |
136 | * for each icq the outer matching queue_lock should be grabbed. | |
b2efa052 TH |
137 | * ie. We need to do reverse-order double lock dancing. |
138 | * | |
139 | * Another twist is that we are often called with one of the | |
140 | * matching queue_locks held as indicated by @locked_q, which | |
141 | * prevents performing double-lock dance for other queues. | |
142 | * | |
143 | * So, we do it in two stages. The fast path uses the queue_lock | |
144 | * the caller is holding and, if other queues need to be accessed, | |
145 | * uses trylock to avoid introducing locking dependency. This can | |
146 | * handle most cases, especially if @ioc was performing IO on only | |
147 | * single device. | |
148 | * | |
149 | * If trylock doesn't cut it, we defer to @ioc->release_work which | |
150 | * can do all the double-locking dancing. | |
151 | */ | |
152 | spin_lock_irqsave_nested(&ioc->lock, flags, | |
153 | ioc_release_depth(locked_q)); | |
154 | ||
c5869807 TH |
155 | while (!hlist_empty(&ioc->icq_list)) { |
156 | struct io_cq *icq = hlist_entry(ioc->icq_list.first, | |
157 | struct io_cq, ioc_node); | |
158 | struct request_queue *this_q = icq->q; | |
b2efa052 TH |
159 | |
160 | if (this_q != last_q) { | |
161 | if (last_q && last_q != locked_q) | |
162 | spin_unlock(last_q->queue_lock); | |
163 | last_q = NULL; | |
164 | ||
165 | if (!spin_trylock(this_q->queue_lock)) | |
166 | break; | |
167 | last_q = this_q; | |
168 | continue; | |
169 | } | |
170 | ioc_release_depth_inc(this_q); | |
c5869807 TH |
171 | icq->exit(icq); |
172 | icq->release(icq); | |
b2efa052 TH |
173 | ioc_release_depth_dec(this_q); |
174 | } | |
86db1e29 | 175 | |
b2efa052 TH |
176 | if (last_q && last_q != locked_q) |
177 | spin_unlock(last_q->queue_lock); | |
86db1e29 | 178 | |
b2efa052 | 179 | spin_unlock_irqrestore(&ioc->lock, flags); |
ffc4e759 | 180 | |
c5869807 TH |
181 | /* if no icq is left, we're done; otherwise, kick release_work */ |
182 | if (hlist_empty(&ioc->icq_list)) | |
b2efa052 TH |
183 | kmem_cache_free(iocontext_cachep, ioc); |
184 | else | |
185 | schedule_work(&ioc->release_work); | |
86db1e29 | 186 | } |
b2efa052 | 187 | EXPORT_SYMBOL(put_io_context); |
86db1e29 | 188 | |
27667c99 | 189 | /* Called by the exiting task */ |
b69f2292 | 190 | void exit_io_context(struct task_struct *task) |
86db1e29 JA |
191 | { |
192 | struct io_context *ioc; | |
193 | ||
6e736be7 TH |
194 | /* PF_EXITING prevents new io_context from being attached to @task */ |
195 | WARN_ON_ONCE(!(current->flags & PF_EXITING)); | |
196 | ||
b69f2292 LR |
197 | task_lock(task); |
198 | ioc = task->io_context; | |
199 | task->io_context = NULL; | |
200 | task_unlock(task); | |
86db1e29 | 201 | |
b2efa052 TH |
202 | atomic_dec(&ioc->nr_tasks); |
203 | put_io_context(ioc, NULL); | |
86db1e29 JA |
204 | } |
205 | ||
f2dbd76a TH |
206 | void create_io_context_slowpath(struct task_struct *task, gfp_t gfp_flags, |
207 | int node) | |
86db1e29 | 208 | { |
df415656 | 209 | struct io_context *ioc; |
86db1e29 | 210 | |
42ec57a8 TH |
211 | ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO, |
212 | node); | |
213 | if (unlikely(!ioc)) | |
f2dbd76a | 214 | return; |
42ec57a8 TH |
215 | |
216 | /* initialize */ | |
217 | atomic_long_set(&ioc->refcount, 1); | |
218 | atomic_set(&ioc->nr_tasks, 1); | |
219 | spin_lock_init(&ioc->lock); | |
c5869807 TH |
220 | INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH); |
221 | INIT_HLIST_HEAD(&ioc->icq_list); | |
b2efa052 | 222 | INIT_WORK(&ioc->release_work, ioc_release_fn); |
86db1e29 | 223 | |
6e736be7 TH |
224 | /* try to install, somebody might already have beaten us to it */ |
225 | task_lock(task); | |
f2dbd76a | 226 | if (!task->io_context && !(task->flags & PF_EXITING)) |
6e736be7 | 227 | task->io_context = ioc; |
f2dbd76a | 228 | else |
6e736be7 | 229 | kmem_cache_free(iocontext_cachep, ioc); |
6e736be7 | 230 | task_unlock(task); |
86db1e29 | 231 | } |
f2dbd76a | 232 | EXPORT_SYMBOL(create_io_context_slowpath); |
86db1e29 | 233 | |
6e736be7 TH |
234 | /** |
235 | * get_task_io_context - get io_context of a task | |
236 | * @task: task of interest | |
237 | * @gfp_flags: allocation flags, used if allocation is necessary | |
238 | * @node: allocation node, used if allocation is necessary | |
239 | * | |
240 | * Return io_context of @task. If it doesn't exist, it is created with | |
241 | * @gfp_flags and @node. The returned io_context has its reference count | |
242 | * incremented. | |
86db1e29 | 243 | * |
6e736be7 | 244 | * This function always goes through task_lock() and it's better to use |
f2dbd76a | 245 | * %current->io_context + get_io_context() for %current. |
86db1e29 | 246 | */ |
6e736be7 TH |
247 | struct io_context *get_task_io_context(struct task_struct *task, |
248 | gfp_t gfp_flags, int node) | |
86db1e29 | 249 | { |
6e736be7 | 250 | struct io_context *ioc; |
86db1e29 | 251 | |
6e736be7 TH |
252 | might_sleep_if(gfp_flags & __GFP_WAIT); |
253 | ||
f2dbd76a TH |
254 | do { |
255 | task_lock(task); | |
256 | ioc = task->io_context; | |
257 | if (likely(ioc)) { | |
258 | get_io_context(ioc); | |
259 | task_unlock(task); | |
260 | return ioc; | |
261 | } | |
6e736be7 | 262 | task_unlock(task); |
f2dbd76a | 263 | } while (create_io_context(task, gfp_flags, node)); |
6e736be7 | 264 | |
f2dbd76a | 265 | return NULL; |
86db1e29 | 266 | } |
6e736be7 | 267 | EXPORT_SYMBOL(get_task_io_context); |
86db1e29 | 268 | |
dc86900e TH |
269 | void ioc_set_changed(struct io_context *ioc, int which) |
270 | { | |
c5869807 | 271 | struct io_cq *icq; |
dc86900e TH |
272 | struct hlist_node *n; |
273 | ||
c5869807 TH |
274 | hlist_for_each_entry(icq, n, &ioc->icq_list, ioc_node) |
275 | set_bit(which, &icq->changed); | |
dc86900e TH |
276 | } |
277 | ||
278 | /** | |
279 | * ioc_ioprio_changed - notify ioprio change | |
280 | * @ioc: io_context of interest | |
281 | * @ioprio: new ioprio | |
282 | * | |
c5869807 TH |
283 | * @ioc's ioprio has changed to @ioprio. Set %ICQ_IOPRIO_CHANGED for all |
284 | * icq's. iosched is responsible for checking the bit and applying it on | |
dc86900e TH |
285 | * request issue path. |
286 | */ | |
287 | void ioc_ioprio_changed(struct io_context *ioc, int ioprio) | |
288 | { | |
289 | unsigned long flags; | |
290 | ||
291 | spin_lock_irqsave(&ioc->lock, flags); | |
292 | ioc->ioprio = ioprio; | |
c5869807 | 293 | ioc_set_changed(ioc, ICQ_IOPRIO_CHANGED); |
dc86900e TH |
294 | spin_unlock_irqrestore(&ioc->lock, flags); |
295 | } | |
296 | ||
297 | /** | |
298 | * ioc_cgroup_changed - notify cgroup change | |
299 | * @ioc: io_context of interest | |
300 | * | |
c5869807 | 301 | * @ioc's cgroup has changed. Set %ICQ_CGROUP_CHANGED for all icq's. |
dc86900e TH |
302 | * iosched is responsible for checking the bit and applying it on request |
303 | * issue path. | |
304 | */ | |
305 | void ioc_cgroup_changed(struct io_context *ioc) | |
306 | { | |
307 | unsigned long flags; | |
308 | ||
309 | spin_lock_irqsave(&ioc->lock, flags); | |
c5869807 | 310 | ioc_set_changed(ioc, ICQ_CGROUP_CHANGED); |
dc86900e TH |
311 | spin_unlock_irqrestore(&ioc->lock, flags); |
312 | } | |
313 | ||
13341598 | 314 | static int __init blk_ioc_init(void) |
86db1e29 JA |
315 | { |
316 | iocontext_cachep = kmem_cache_create("blkdev_ioc", | |
317 | sizeof(struct io_context), 0, SLAB_PANIC, NULL); | |
318 | return 0; | |
319 | } | |
320 | subsys_initcall(blk_ioc_init); |