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