blkio: Export disk time and sectors used by a group to user space
[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 "blk-cgroup.h"
17
18 extern void cfq_unlink_blkio_group(void *, struct blkio_group *);
19
20 struct blkio_cgroup blkio_root_cgroup = { .weight = 2*BLKIO_WEIGHT_DEFAULT };
21
22 struct blkio_cgroup *cgroup_to_blkio_cgroup(struct cgroup *cgroup)
23 {
24         return container_of(cgroup_subsys_state(cgroup, blkio_subsys_id),
25                             struct blkio_cgroup, css);
26 }
27
28 void blkiocg_update_blkio_group_stats(struct blkio_group *blkg,
29                         unsigned long time, unsigned long sectors)
30 {
31         blkg->time += time;
32         blkg->sectors += sectors;
33 }
34
35 void blkiocg_add_blkio_group(struct blkio_cgroup *blkcg,
36                         struct blkio_group *blkg, void *key, dev_t dev)
37 {
38         unsigned long flags;
39
40         spin_lock_irqsave(&blkcg->lock, flags);
41         rcu_assign_pointer(blkg->key, key);
42         blkg->blkcg_id = css_id(&blkcg->css);
43         hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
44         spin_unlock_irqrestore(&blkcg->lock, flags);
45 #ifdef CONFIG_DEBUG_BLK_CGROUP
46         /* Need to take css reference ? */
47         cgroup_path(blkcg->css.cgroup, blkg->path, sizeof(blkg->path));
48 #endif
49         blkg->dev = dev;
50 }
51
52 static void __blkiocg_del_blkio_group(struct blkio_group *blkg)
53 {
54         hlist_del_init_rcu(&blkg->blkcg_node);
55         blkg->blkcg_id = 0;
56 }
57
58 /*
59  * returns 0 if blkio_group was still on cgroup list. Otherwise returns 1
60  * indicating that blk_group was unhashed by the time we got to it.
61  */
62 int blkiocg_del_blkio_group(struct blkio_group *blkg)
63 {
64         struct blkio_cgroup *blkcg;
65         unsigned long flags;
66         struct cgroup_subsys_state *css;
67         int ret = 1;
68
69         rcu_read_lock();
70         css = css_lookup(&blkio_subsys, blkg->blkcg_id);
71         if (!css)
72                 goto out;
73
74         blkcg = container_of(css, struct blkio_cgroup, css);
75         spin_lock_irqsave(&blkcg->lock, flags);
76         if (!hlist_unhashed(&blkg->blkcg_node)) {
77                 __blkiocg_del_blkio_group(blkg);
78                 ret = 0;
79         }
80         spin_unlock_irqrestore(&blkcg->lock, flags);
81 out:
82         rcu_read_unlock();
83         return ret;
84 }
85
86 /* called under rcu_read_lock(). */
87 struct blkio_group *blkiocg_lookup_group(struct blkio_cgroup *blkcg, void *key)
88 {
89         struct blkio_group *blkg;
90         struct hlist_node *n;
91         void *__key;
92
93         hlist_for_each_entry_rcu(blkg, n, &blkcg->blkg_list, blkcg_node) {
94                 __key = blkg->key;
95                 if (__key == key)
96                         return blkg;
97         }
98
99         return NULL;
100 }
101
102 #define SHOW_FUNCTION(__VAR)                                            \
103 static u64 blkiocg_##__VAR##_read(struct cgroup *cgroup,                \
104                                        struct cftype *cftype)           \
105 {                                                                       \
106         struct blkio_cgroup *blkcg;                                     \
107                                                                         \
108         blkcg = cgroup_to_blkio_cgroup(cgroup);                         \
109         return (u64)blkcg->__VAR;                                       \
110 }
111
112 SHOW_FUNCTION(weight);
113 #undef SHOW_FUNCTION
114
115 static int
116 blkiocg_weight_write(struct cgroup *cgroup, struct cftype *cftype, u64 val)
117 {
118         struct blkio_cgroup *blkcg;
119
120         if (val < BLKIO_WEIGHT_MIN || val > BLKIO_WEIGHT_MAX)
121                 return -EINVAL;
122
123         blkcg = cgroup_to_blkio_cgroup(cgroup);
124         blkcg->weight = (unsigned int)val;
125         return 0;
126 }
127
128 #define SHOW_FUNCTION_PER_GROUP(__VAR)                                  \
129 static int blkiocg_##__VAR##_read(struct cgroup *cgroup,                \
130                         struct cftype *cftype, struct seq_file *m)      \
131 {                                                                       \
132         struct blkio_cgroup *blkcg;                                     \
133         struct blkio_group *blkg;                                       \
134         struct hlist_node *n;                                           \
135                                                                         \
136         if (!cgroup_lock_live_group(cgroup))                            \
137                 return -ENODEV;                                         \
138                                                                         \
139         blkcg = cgroup_to_blkio_cgroup(cgroup);                         \
140         rcu_read_lock();                                                \
141         hlist_for_each_entry_rcu(blkg, n, &blkcg->blkg_list, blkcg_node) {\
142                 if (blkg->dev)                                          \
143                         seq_printf(m, "%u:%u %lu\n", MAJOR(blkg->dev),  \
144                                  MINOR(blkg->dev), blkg->__VAR);        \
145         }                                                               \
146         rcu_read_unlock();                                              \
147         cgroup_unlock();                                                \
148         return 0;                                                       \
149 }
150
151 SHOW_FUNCTION_PER_GROUP(time);
152 SHOW_FUNCTION_PER_GROUP(sectors);
153 #ifdef CONFIG_DEBUG_BLK_CGROUP
154 SHOW_FUNCTION_PER_GROUP(dequeue);
155 #endif
156 #undef SHOW_FUNCTION_PER_GROUP
157
158 #ifdef CONFIG_DEBUG_BLK_CGROUP
159 void blkiocg_update_blkio_group_dequeue_stats(struct blkio_group *blkg,
160                         unsigned long dequeue)
161 {
162         blkg->dequeue += dequeue;
163 }
164 #endif
165
166 struct cftype blkio_files[] = {
167         {
168                 .name = "weight",
169                 .read_u64 = blkiocg_weight_read,
170                 .write_u64 = blkiocg_weight_write,
171         },
172         {
173                 .name = "time",
174                 .read_seq_string = blkiocg_time_read,
175         },
176         {
177                 .name = "sectors",
178                 .read_seq_string = blkiocg_sectors_read,
179         },
180 #ifdef CONFIG_DEBUG_BLK_CGROUP
181        {
182                 .name = "dequeue",
183                 .read_seq_string = blkiocg_dequeue_read,
184        },
185 #endif
186 };
187
188 static int blkiocg_populate(struct cgroup_subsys *subsys, struct cgroup *cgroup)
189 {
190         return cgroup_add_files(cgroup, subsys, blkio_files,
191                                 ARRAY_SIZE(blkio_files));
192 }
193
194 static void blkiocg_destroy(struct cgroup_subsys *subsys, struct cgroup *cgroup)
195 {
196         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
197         unsigned long flags;
198         struct blkio_group *blkg;
199         void *key;
200
201         rcu_read_lock();
202 remove_entry:
203         spin_lock_irqsave(&blkcg->lock, flags);
204
205         if (hlist_empty(&blkcg->blkg_list)) {
206                 spin_unlock_irqrestore(&blkcg->lock, flags);
207                 goto done;
208         }
209
210         blkg = hlist_entry(blkcg->blkg_list.first, struct blkio_group,
211                                 blkcg_node);
212         key = rcu_dereference(blkg->key);
213         __blkiocg_del_blkio_group(blkg);
214
215         spin_unlock_irqrestore(&blkcg->lock, flags);
216
217         /*
218          * This blkio_group is being unlinked as associated cgroup is going
219          * away. Let all the IO controlling policies know about this event.
220          *
221          * Currently this is static call to one io controlling policy. Once
222          * we have more policies in place, we need some dynamic registration
223          * of callback function.
224          */
225         cfq_unlink_blkio_group(key, blkg);
226         goto remove_entry;
227 done:
228         free_css_id(&blkio_subsys, &blkcg->css);
229         rcu_read_unlock();
230         kfree(blkcg);
231 }
232
233 static struct cgroup_subsys_state *
234 blkiocg_create(struct cgroup_subsys *subsys, struct cgroup *cgroup)
235 {
236         struct blkio_cgroup *blkcg, *parent_blkcg;
237
238         if (!cgroup->parent) {
239                 blkcg = &blkio_root_cgroup;
240                 goto done;
241         }
242
243         /* Currently we do not support hierarchy deeper than two level (0,1) */
244         parent_blkcg = cgroup_to_blkio_cgroup(cgroup->parent);
245         if (css_depth(&parent_blkcg->css) > 0)
246                 return ERR_PTR(-EINVAL);
247
248         blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
249         if (!blkcg)
250                 return ERR_PTR(-ENOMEM);
251
252         blkcg->weight = BLKIO_WEIGHT_DEFAULT;
253 done:
254         spin_lock_init(&blkcg->lock);
255         INIT_HLIST_HEAD(&blkcg->blkg_list);
256
257         return &blkcg->css;
258 }
259
260 /*
261  * We cannot support shared io contexts, as we have no mean to support
262  * two tasks with the same ioc in two different groups without major rework
263  * of the main cic data structures.  For now we allow a task to change
264  * its cgroup only if it's the only owner of its ioc.
265  */
266 static int blkiocg_can_attach(struct cgroup_subsys *subsys,
267                                 struct cgroup *cgroup, struct task_struct *tsk,
268                                 bool threadgroup)
269 {
270         struct io_context *ioc;
271         int ret = 0;
272
273         /* task_lock() is needed to avoid races with exit_io_context() */
274         task_lock(tsk);
275         ioc = tsk->io_context;
276         if (ioc && atomic_read(&ioc->nr_tasks) > 1)
277                 ret = -EINVAL;
278         task_unlock(tsk);
279
280         return ret;
281 }
282
283 static void blkiocg_attach(struct cgroup_subsys *subsys, struct cgroup *cgroup,
284                                 struct cgroup *prev, struct task_struct *tsk,
285                                 bool threadgroup)
286 {
287         struct io_context *ioc;
288
289         task_lock(tsk);
290         ioc = tsk->io_context;
291         if (ioc)
292                 ioc->cgroup_changed = 1;
293         task_unlock(tsk);
294 }
295
296 struct cgroup_subsys blkio_subsys = {
297         .name = "blkio",
298         .create = blkiocg_create,
299         .can_attach = blkiocg_can_attach,
300         .attach = blkiocg_attach,
301         .destroy = blkiocg_destroy,
302         .populate = blkiocg_populate,
303         .subsys_id = blkio_subsys_id,
304         .use_id = 1,
305 };