1fa2654db0a62c1c3f07b92fb8b33eff473d1b41
[linux-2.6.git] / block / blk-cgroup.c
1 /*
2  * Common Block IO controller cgroup interface
3  *
4  * Based on ideas and code from CFQ, CFS and BFQ:
5  * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
6  *
7  * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8  *                    Paolo Valente <paolo.valente@unimore.it>
9  *
10  * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
11  *                    Nauman Rafique <nauman@google.com>
12  */
13 #include <linux/ioprio.h>
14 #include <linux/seq_file.h>
15 #include <linux/kdev_t.h>
16 #include <linux/module.h>
17 #include <linux/err.h>
18 #include "blk-cgroup.h"
19
20 static DEFINE_SPINLOCK(blkio_list_lock);
21 static LIST_HEAD(blkio_list);
22
23 struct blkio_cgroup blkio_root_cgroup = { .weight = 2*BLKIO_WEIGHT_DEFAULT };
24 EXPORT_SYMBOL_GPL(blkio_root_cgroup);
25
26 bool blkiocg_css_tryget(struct blkio_cgroup *blkcg)
27 {
28         if (!css_tryget(&blkcg->css))
29                 return false;
30         return true;
31 }
32 EXPORT_SYMBOL_GPL(blkiocg_css_tryget);
33
34 void blkiocg_css_put(struct blkio_cgroup *blkcg)
35 {
36         css_put(&blkcg->css);
37 }
38 EXPORT_SYMBOL_GPL(blkiocg_css_put);
39
40 struct blkio_cgroup *cgroup_to_blkio_cgroup(struct cgroup *cgroup)
41 {
42         return container_of(cgroup_subsys_state(cgroup, blkio_subsys_id),
43                             struct blkio_cgroup, css);
44 }
45 EXPORT_SYMBOL_GPL(cgroup_to_blkio_cgroup);
46
47 void blkiocg_update_blkio_group_stats(struct blkio_group *blkg,
48                         unsigned long time, unsigned long sectors)
49 {
50         blkg->time += time;
51         blkg->sectors += sectors;
52 }
53 EXPORT_SYMBOL_GPL(blkiocg_update_blkio_group_stats);
54
55 void blkiocg_add_blkio_group(struct blkio_cgroup *blkcg,
56                         struct blkio_group *blkg, void *key, dev_t dev)
57 {
58         unsigned long flags;
59
60         spin_lock_irqsave(&blkcg->lock, flags);
61         rcu_assign_pointer(blkg->key, key);
62         blkg->blkcg_id = css_id(&blkcg->css);
63         hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
64         spin_unlock_irqrestore(&blkcg->lock, flags);
65 #ifdef CONFIG_DEBUG_BLK_CGROUP
66         /* Need to take css reference ? */
67         cgroup_path(blkcg->css.cgroup, blkg->path, sizeof(blkg->path));
68 #endif
69         blkg->dev = dev;
70 }
71 EXPORT_SYMBOL_GPL(blkiocg_add_blkio_group);
72
73 static void __blkiocg_del_blkio_group(struct blkio_group *blkg)
74 {
75         hlist_del_init_rcu(&blkg->blkcg_node);
76         blkg->blkcg_id = 0;
77 }
78
79 /*
80  * returns 0 if blkio_group was still on cgroup list. Otherwise returns 1
81  * indicating that blk_group was unhashed by the time we got to it.
82  */
83 int blkiocg_del_blkio_group(struct blkio_group *blkg)
84 {
85         struct blkio_cgroup *blkcg;
86         unsigned long flags;
87         struct cgroup_subsys_state *css;
88         int ret = 1;
89
90         rcu_read_lock();
91         css = css_lookup(&blkio_subsys, blkg->blkcg_id);
92         if (!css)
93                 goto out;
94
95         blkcg = container_of(css, struct blkio_cgroup, css);
96         spin_lock_irqsave(&blkcg->lock, flags);
97         if (!hlist_unhashed(&blkg->blkcg_node)) {
98                 __blkiocg_del_blkio_group(blkg);
99                 ret = 0;
100         }
101         spin_unlock_irqrestore(&blkcg->lock, flags);
102 out:
103         rcu_read_unlock();
104         return ret;
105 }
106 EXPORT_SYMBOL_GPL(blkiocg_del_blkio_group);
107
108 /* called under rcu_read_lock(). */
109 struct blkio_group *blkiocg_lookup_group(struct blkio_cgroup *blkcg, void *key)
110 {
111         struct blkio_group *blkg;
112         struct hlist_node *n;
113         void *__key;
114
115         hlist_for_each_entry_rcu(blkg, n, &blkcg->blkg_list, blkcg_node) {
116                 __key = blkg->key;
117                 if (__key == key)
118                         return blkg;
119         }
120
121         return NULL;
122 }
123 EXPORT_SYMBOL_GPL(blkiocg_lookup_group);
124
125 #define SHOW_FUNCTION(__VAR)                                            \
126 static u64 blkiocg_##__VAR##_read(struct cgroup *cgroup,                \
127                                        struct cftype *cftype)           \
128 {                                                                       \
129         struct blkio_cgroup *blkcg;                                     \
130                                                                         \
131         blkcg = cgroup_to_blkio_cgroup(cgroup);                         \
132         return (u64)blkcg->__VAR;                                       \
133 }
134
135 SHOW_FUNCTION(weight);
136 #undef SHOW_FUNCTION
137
138 static int
139 blkiocg_weight_write(struct cgroup *cgroup, struct cftype *cftype, u64 val)
140 {
141         struct blkio_cgroup *blkcg;
142         struct blkio_group *blkg;
143         struct hlist_node *n;
144         struct blkio_policy_type *blkiop;
145
146         if (val < BLKIO_WEIGHT_MIN || val > BLKIO_WEIGHT_MAX)
147                 return -EINVAL;
148
149         blkcg = cgroup_to_blkio_cgroup(cgroup);
150         spin_lock_irq(&blkcg->lock);
151         blkcg->weight = (unsigned int)val;
152         hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
153                 spin_lock(&blkio_list_lock);
154                 list_for_each_entry(blkiop, &blkio_list, list)
155                         blkiop->ops.blkio_update_group_weight_fn(blkg,
156                                         blkcg->weight);
157                 spin_unlock(&blkio_list_lock);
158         }
159         spin_unlock_irq(&blkcg->lock);
160         return 0;
161 }
162
163 #define SHOW_FUNCTION_PER_GROUP(__VAR)                                  \
164 static int blkiocg_##__VAR##_read(struct cgroup *cgroup,                \
165                         struct cftype *cftype, struct seq_file *m)      \
166 {                                                                       \
167         struct blkio_cgroup *blkcg;                                     \
168         struct blkio_group *blkg;                                       \
169         struct hlist_node *n;                                           \
170                                                                         \
171         if (!cgroup_lock_live_group(cgroup))                            \
172                 return -ENODEV;                                         \
173                                                                         \
174         blkcg = cgroup_to_blkio_cgroup(cgroup);                         \
175         rcu_read_lock();                                                \
176         hlist_for_each_entry_rcu(blkg, n, &blkcg->blkg_list, blkcg_node) {\
177                 if (blkg->dev)                                          \
178                         seq_printf(m, "%u:%u %lu\n", MAJOR(blkg->dev),  \
179                                  MINOR(blkg->dev), blkg->__VAR);        \
180         }                                                               \
181         rcu_read_unlock();                                              \
182         cgroup_unlock();                                                \
183         return 0;                                                       \
184 }
185
186 SHOW_FUNCTION_PER_GROUP(time);
187 SHOW_FUNCTION_PER_GROUP(sectors);
188 #ifdef CONFIG_DEBUG_BLK_CGROUP
189 SHOW_FUNCTION_PER_GROUP(dequeue);
190 #endif
191 #undef SHOW_FUNCTION_PER_GROUP
192
193 #ifdef CONFIG_DEBUG_BLK_CGROUP
194 void blkiocg_update_blkio_group_dequeue_stats(struct blkio_group *blkg,
195                         unsigned long dequeue)
196 {
197         blkg->dequeue += dequeue;
198 }
199 EXPORT_SYMBOL_GPL(blkiocg_update_blkio_group_dequeue_stats);
200 #endif
201
202 struct cftype blkio_files[] = {
203         {
204                 .name = "weight",
205                 .read_u64 = blkiocg_weight_read,
206                 .write_u64 = blkiocg_weight_write,
207         },
208         {
209                 .name = "time",
210                 .read_seq_string = blkiocg_time_read,
211         },
212         {
213                 .name = "sectors",
214                 .read_seq_string = blkiocg_sectors_read,
215         },
216 #ifdef CONFIG_DEBUG_BLK_CGROUP
217        {
218                 .name = "dequeue",
219                 .read_seq_string = blkiocg_dequeue_read,
220        },
221 #endif
222 };
223
224 static int blkiocg_populate(struct cgroup_subsys *subsys, struct cgroup *cgroup)
225 {
226         return cgroup_add_files(cgroup, subsys, blkio_files,
227                                 ARRAY_SIZE(blkio_files));
228 }
229
230 static void blkiocg_destroy(struct cgroup_subsys *subsys, struct cgroup *cgroup)
231 {
232         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
233         unsigned long flags;
234         struct blkio_group *blkg;
235         void *key;
236         struct blkio_policy_type *blkiop;
237
238         rcu_read_lock();
239 remove_entry:
240         spin_lock_irqsave(&blkcg->lock, flags);
241
242         if (hlist_empty(&blkcg->blkg_list)) {
243                 spin_unlock_irqrestore(&blkcg->lock, flags);
244                 goto done;
245         }
246
247         blkg = hlist_entry(blkcg->blkg_list.first, struct blkio_group,
248                                 blkcg_node);
249         key = rcu_dereference(blkg->key);
250         __blkiocg_del_blkio_group(blkg);
251
252         spin_unlock_irqrestore(&blkcg->lock, flags);
253
254         /*
255          * This blkio_group is being unlinked as associated cgroup is going
256          * away. Let all the IO controlling policies know about this event.
257          *
258          * Currently this is static call to one io controlling policy. Once
259          * we have more policies in place, we need some dynamic registration
260          * of callback function.
261          */
262         spin_lock(&blkio_list_lock);
263         list_for_each_entry(blkiop, &blkio_list, list)
264                 blkiop->ops.blkio_unlink_group_fn(key, blkg);
265         spin_unlock(&blkio_list_lock);
266         goto remove_entry;
267 done:
268         free_css_id(&blkio_subsys, &blkcg->css);
269         rcu_read_unlock();
270         kfree(blkcg);
271 }
272
273 static struct cgroup_subsys_state *
274 blkiocg_create(struct cgroup_subsys *subsys, struct cgroup *cgroup)
275 {
276         struct blkio_cgroup *blkcg, *parent_blkcg;
277
278         if (!cgroup->parent) {
279                 blkcg = &blkio_root_cgroup;
280                 goto done;
281         }
282
283         /* Currently we do not support hierarchy deeper than two level (0,1) */
284         parent_blkcg = cgroup_to_blkio_cgroup(cgroup->parent);
285         if (css_depth(&parent_blkcg->css) > 0)
286                 return ERR_PTR(-EINVAL);
287
288         blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
289         if (!blkcg)
290                 return ERR_PTR(-ENOMEM);
291
292         blkcg->weight = BLKIO_WEIGHT_DEFAULT;
293 done:
294         spin_lock_init(&blkcg->lock);
295         INIT_HLIST_HEAD(&blkcg->blkg_list);
296
297         return &blkcg->css;
298 }
299
300 /*
301  * We cannot support shared io contexts, as we have no mean to support
302  * two tasks with the same ioc in two different groups without major rework
303  * of the main cic data structures.  For now we allow a task to change
304  * its cgroup only if it's the only owner of its ioc.
305  */
306 static int blkiocg_can_attach(struct cgroup_subsys *subsys,
307                                 struct cgroup *cgroup, struct task_struct *tsk,
308                                 bool threadgroup)
309 {
310         struct io_context *ioc;
311         int ret = 0;
312
313         /* task_lock() is needed to avoid races with exit_io_context() */
314         task_lock(tsk);
315         ioc = tsk->io_context;
316         if (ioc && atomic_read(&ioc->nr_tasks) > 1)
317                 ret = -EINVAL;
318         task_unlock(tsk);
319
320         return ret;
321 }
322
323 static void blkiocg_attach(struct cgroup_subsys *subsys, struct cgroup *cgroup,
324                                 struct cgroup *prev, struct task_struct *tsk,
325                                 bool threadgroup)
326 {
327         struct io_context *ioc;
328
329         task_lock(tsk);
330         ioc = tsk->io_context;
331         if (ioc)
332                 ioc->cgroup_changed = 1;
333         task_unlock(tsk);
334 }
335
336 struct cgroup_subsys blkio_subsys = {
337         .name = "blkio",
338         .create = blkiocg_create,
339         .can_attach = blkiocg_can_attach,
340         .attach = blkiocg_attach,
341         .destroy = blkiocg_destroy,
342         .populate = blkiocg_populate,
343         .subsys_id = blkio_subsys_id,
344         .use_id = 1,
345 };
346
347 void blkio_policy_register(struct blkio_policy_type *blkiop)
348 {
349         spin_lock(&blkio_list_lock);
350         list_add_tail(&blkiop->list, &blkio_list);
351         spin_unlock(&blkio_list_lock);
352 }
353 EXPORT_SYMBOL_GPL(blkio_policy_register);
354
355 void blkio_policy_unregister(struct blkio_policy_type *blkiop)
356 {
357         spin_lock(&blkio_list_lock);
358         list_del_init(&blkiop->list);
359         spin_unlock(&blkio_list_lock);
360 }
361 EXPORT_SYMBOL_GPL(blkio_policy_unregister);